xref: /illumos-gate/usr/src/uts/common/fs/zfs/arc.c (revision e419e0b9)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2019, Joyent, Inc.
24  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25  * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26  * Copyright 2017 Nexenta Systems, Inc.  All rights reserved.
27  * Copyright (c) 2011, 2019, Delphix. All rights reserved.
28  * Copyright (c) 2020, George Amanakis. All rights reserved.
29  * Copyright (c) 2020, The FreeBSD Foundation [1]
30  * Copyright 2024 Bill Sommerfeld <sommerfeld@hamachi.org>
31  *
32  * [1] Portions of this software were developed by Allan Jude
33  *     under sponsorship from the FreeBSD Foundation.
34  */
35 
36 /*
37  * DVA-based Adjustable Replacement Cache
38  *
39  * While much of the theory of operation used here is
40  * based on the self-tuning, low overhead replacement cache
41  * presented by Megiddo and Modha at FAST 2003, there are some
42  * significant differences:
43  *
44  * 1. The Megiddo and Modha model assumes any page is evictable.
45  * Pages in its cache cannot be "locked" into memory.  This makes
46  * the eviction algorithm simple: evict the last page in the list.
47  * This also make the performance characteristics easy to reason
48  * about.  Our cache is not so simple.  At any given moment, some
49  * subset of the blocks in the cache are un-evictable because we
50  * have handed out a reference to them.  Blocks are only evictable
51  * when there are no external references active.  This makes
52  * eviction far more problematic:  we choose to evict the evictable
53  * blocks that are the "lowest" in the list.
54  *
55  * There are times when it is not possible to evict the requested
56  * space.  In these circumstances we are unable to adjust the cache
57  * size.  To prevent the cache growing unbounded at these times we
58  * implement a "cache throttle" that slows the flow of new data
59  * into the cache until we can make space available.
60  *
61  * 2. The Megiddo and Modha model assumes a fixed cache size.
62  * Pages are evicted when the cache is full and there is a cache
63  * miss.  Our model has a variable sized cache.  It grows with
64  * high use, but also tries to react to memory pressure from the
65  * operating system: decreasing its size when system memory is
66  * tight.
67  *
68  * 3. The Megiddo and Modha model assumes a fixed page size. All
69  * elements of the cache are therefore exactly the same size.  So
70  * when adjusting the cache size following a cache miss, its simply
71  * a matter of choosing a single page to evict.  In our model, we
72  * have variable sized cache blocks (rangeing from 512 bytes to
73  * 128K bytes).  We therefore choose a set of blocks to evict to make
74  * space for a cache miss that approximates as closely as possible
75  * the space used by the new block.
76  *
77  * See also:  "ARC: A Self-Tuning, Low Overhead Replacement Cache"
78  * by N. Megiddo & D. Modha, FAST 2003
79  */
80 
81 /*
82  * The locking model:
83  *
84  * A new reference to a cache buffer can be obtained in two
85  * ways: 1) via a hash table lookup using the DVA as a key,
86  * or 2) via one of the ARC lists.  The arc_read() interface
87  * uses method 1, while the internal ARC algorithms for
88  * adjusting the cache use method 2.  We therefore provide two
89  * types of locks: 1) the hash table lock array, and 2) the
90  * ARC list locks.
91  *
92  * Buffers do not have their own mutexes, rather they rely on the
93  * hash table mutexes for the bulk of their protection (i.e. most
94  * fields in the arc_buf_hdr_t are protected by these mutexes).
95  *
96  * buf_hash_find() returns the appropriate mutex (held) when it
97  * locates the requested buffer in the hash table.  It returns
98  * NULL for the mutex if the buffer was not in the table.
99  *
100  * buf_hash_remove() expects the appropriate hash mutex to be
101  * already held before it is invoked.
102  *
103  * Each ARC state also has a mutex which is used to protect the
104  * buffer list associated with the state.  When attempting to
105  * obtain a hash table lock while holding an ARC list lock you
106  * must use: mutex_tryenter() to avoid deadlock.  Also note that
107  * the active state mutex must be held before the ghost state mutex.
108  *
109  * Note that the majority of the performance stats are manipulated
110  * with atomic operations.
111  *
112  * The L2ARC uses the l2ad_mtx on each vdev for the following:
113  *
114  *	- L2ARC buflist creation
115  *	- L2ARC buflist eviction
116  *	- L2ARC write completion, which walks L2ARC buflists
117  *	- ARC header destruction, as it removes from L2ARC buflists
118  *	- ARC header release, as it removes from L2ARC buflists
119  */
120 
121 /*
122  * ARC operation:
123  *
124  * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
125  * This structure can point either to a block that is still in the cache or to
126  * one that is only accessible in an L2 ARC device, or it can provide
127  * information about a block that was recently evicted. If a block is
128  * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
129  * information to retrieve it from the L2ARC device. This information is
130  * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
131  * that is in this state cannot access the data directly.
132  *
133  * Blocks that are actively being referenced or have not been evicted
134  * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
135  * the arc_buf_hdr_t that will point to the data block in memory. A block can
136  * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
137  * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
138  * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
139  *
140  * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
141  * ability to store the physical data (b_pabd) associated with the DVA of the
142  * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
143  * it will match its on-disk compression characteristics. This behavior can be
144  * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
145  * compressed ARC functionality is disabled, the b_pabd will point to an
146  * uncompressed version of the on-disk data.
147  *
148  * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
149  * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
150  * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
151  * consumer. The ARC will provide references to this data and will keep it
152  * cached until it is no longer in use. The ARC caches only the L1ARC's physical
153  * data block and will evict any arc_buf_t that is no longer referenced. The
154  * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
155  * "overhead_size" kstat.
156  *
157  * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
158  * compressed form. The typical case is that consumers will want uncompressed
159  * data, and when that happens a new data buffer is allocated where the data is
160  * decompressed for them to use. Currently the only consumer who wants
161  * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
162  * exists on disk. When this happens, the arc_buf_t's data buffer is shared
163  * with the arc_buf_hdr_t.
164  *
165  * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
166  * first one is owned by a compressed send consumer (and therefore references
167  * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
168  * used by any other consumer (and has its own uncompressed copy of the data
169  * buffer).
170  *
171  *   arc_buf_hdr_t
172  *   +-----------+
173  *   | fields    |
174  *   | common to |
175  *   | L1- and   |
176  *   | L2ARC     |
177  *   +-----------+
178  *   | l2arc_buf_hdr_t
179  *   |           |
180  *   +-----------+
181  *   | l1arc_buf_hdr_t
182  *   |           |              arc_buf_t
183  *   | b_buf     +------------>+-----------+      arc_buf_t
184  *   | b_pabd    +-+           |b_next     +---->+-----------+
185  *   +-----------+ |           |-----------|     |b_next     +-->NULL
186  *                 |           |b_comp = T |     +-----------+
187  *                 |           |b_data     +-+   |b_comp = F |
188  *                 |           +-----------+ |   |b_data     +-+
189  *                 +->+------+               |   +-----------+ |
190  *        compressed  |      |               |                 |
191  *           data     |      |<--------------+                 | uncompressed
192  *                    +------+          compressed,            |     data
193  *                                        shared               +-->+------+
194  *                                         data                    |      |
195  *                                                                 |      |
196  *                                                                 +------+
197  *
198  * When a consumer reads a block, the ARC must first look to see if the
199  * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
200  * arc_buf_t and either copies uncompressed data into a new data buffer from an
201  * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
202  * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
203  * hdr is compressed and the desired compression characteristics of the
204  * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
205  * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
206  * the last buffer in the hdr's b_buf list, however a shared compressed buf can
207  * be anywhere in the hdr's list.
208  *
209  * The diagram below shows an example of an uncompressed ARC hdr that is
210  * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
211  * the last element in the buf list):
212  *
213  *                arc_buf_hdr_t
214  *                +-----------+
215  *                |           |
216  *                |           |
217  *                |           |
218  *                +-----------+
219  * l2arc_buf_hdr_t|           |
220  *                |           |
221  *                +-----------+
222  * l1arc_buf_hdr_t|           |
223  *                |           |                 arc_buf_t    (shared)
224  *                |    b_buf  +------------>+---------+      arc_buf_t
225  *                |           |             |b_next   +---->+---------+
226  *                |  b_pabd   +-+           |---------|     |b_next   +-->NULL
227  *                +-----------+ |           |         |     +---------+
228  *                              |           |b_data   +-+   |         |
229  *                              |           +---------+ |   |b_data   +-+
230  *                              +->+------+             |   +---------+ |
231  *                                 |      |             |               |
232  *                   uncompressed  |      |             |               |
233  *                        data     +------+             |               |
234  *                                    ^                 +->+------+     |
235  *                                    |       uncompressed |      |     |
236  *                                    |           data     |      |     |
237  *                                    |                    +------+     |
238  *                                    +---------------------------------+
239  *
240  * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
241  * since the physical block is about to be rewritten. The new data contents
242  * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
243  * it may compress the data before writing it to disk. The ARC will be called
244  * with the transformed data and will bcopy the transformed on-disk block into
245  * a newly allocated b_pabd. Writes are always done into buffers which have
246  * either been loaned (and hence are new and don't have other readers) or
247  * buffers which have been released (and hence have their own hdr, if there
248  * were originally other readers of the buf's original hdr). This ensures that
249  * the ARC only needs to update a single buf and its hdr after a write occurs.
250  *
251  * When the L2ARC is in use, it will also take advantage of the b_pabd. The
252  * L2ARC will always write the contents of b_pabd to the L2ARC. This means
253  * that when compressed ARC is enabled that the L2ARC blocks are identical
254  * to the on-disk block in the main data pool. This provides a significant
255  * advantage since the ARC can leverage the bp's checksum when reading from the
256  * L2ARC to determine if the contents are valid. However, if the compressed
257  * ARC is disabled, then the L2ARC's block must be transformed to look
258  * like the physical block in the main data pool before comparing the
259  * checksum and determining its validity.
260  *
261  * The L1ARC has a slightly different system for storing encrypted data.
262  * Raw (encrypted + possibly compressed) data has a few subtle differences from
263  * data that is just compressed. The biggest difference is that it is not
264  * possible to decrypt encrypted data (or visa versa) if the keys aren't loaded.
265  * The other difference is that encryption cannot be treated as a suggestion.
266  * If a caller would prefer compressed data, but they actually wind up with
267  * uncompressed data the worst thing that could happen is there might be a
268  * performance hit. If the caller requests encrypted data, however, we must be
269  * sure they actually get it or else secret information could be leaked. Raw
270  * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
271  * may have both an encrypted version and a decrypted version of its data at
272  * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
273  * copied out of this header. To avoid complications with b_pabd, raw buffers
274  * cannot be shared.
275  */
276 
277 #include <sys/spa.h>
278 #include <sys/zio.h>
279 #include <sys/spa_impl.h>
280 #include <sys/zio_compress.h>
281 #include <sys/zio_checksum.h>
282 #include <sys/zfs_context.h>
283 #include <sys/arc.h>
284 #include <sys/refcount.h>
285 #include <sys/vdev.h>
286 #include <sys/vdev_impl.h>
287 #include <sys/dsl_pool.h>
288 #include <sys/zio_checksum.h>
289 #include <sys/multilist.h>
290 #include <sys/abd.h>
291 #include <sys/zil.h>
292 #include <sys/fm/fs/zfs.h>
293 #ifdef _KERNEL
294 #include <sys/vmsystm.h>
295 #include <vm/anon.h>
296 #include <sys/fs/swapnode.h>
297 #include <sys/dnlc.h>
298 #endif
299 #include <sys/callb.h>
300 #include <sys/kstat.h>
301 #include <sys/zthr.h>
302 #include <zfs_fletcher.h>
303 #include <sys/arc_impl.h>
304 #include <sys/aggsum.h>
305 #include <sys/cityhash.h>
306 #include <sys/param.h>
307 
308 #ifndef _KERNEL
309 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
310 boolean_t arc_watch = B_FALSE;
311 int arc_procfd;
312 #endif
313 
314 /*
315  * This thread's job is to keep enough free memory in the system, by
316  * calling arc_kmem_reap_now() plus arc_shrink(), which improves
317  * arc_available_memory().
318  */
319 static zthr_t		*arc_reap_zthr;
320 
321 /*
322  * This thread's job is to keep arc_size under arc_c, by calling
323  * arc_adjust(), which improves arc_is_overflowing().
324  */
325 static zthr_t		*arc_adjust_zthr;
326 
327 static kmutex_t		arc_adjust_lock;
328 static kcondvar_t	arc_adjust_waiters_cv;
329 static boolean_t	arc_adjust_needed = B_FALSE;
330 
331 uint_t arc_reduce_dnlc_percent = 3;
332 
333 /*
334  * The number of headers to evict in arc_evict_state_impl() before
335  * dropping the sublist lock and evicting from another sublist. A lower
336  * value means we're more likely to evict the "correct" header (i.e. the
337  * oldest header in the arc state), but comes with higher overhead
338  * (i.e. more invocations of arc_evict_state_impl()).
339  */
340 int zfs_arc_evict_batch_limit = 10;
341 
342 /* number of seconds before growing cache again */
343 int arc_grow_retry = 60;
344 
345 /*
346  * Minimum time between calls to arc_kmem_reap_soon().  Note that this will
347  * be converted to ticks, so with the default hz=100, a setting of 15 ms
348  * will actually wait 2 ticks, or 20ms.
349  */
350 int arc_kmem_cache_reap_retry_ms = 1000;
351 
352 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
353 int zfs_arc_overflow_shift = 8;
354 
355 /* shift of arc_c for calculating both min and max arc_p */
356 int arc_p_min_shift = 4;
357 
358 /* log2(fraction of arc to reclaim) */
359 int arc_shrink_shift = 7;
360 
361 /*
362  * log2(fraction of ARC which must be free to allow growing).
363  * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
364  * when reading a new block into the ARC, we will evict an equal-sized block
365  * from the ARC.
366  *
367  * This must be less than arc_shrink_shift, so that when we shrink the ARC,
368  * we will still not allow it to grow.
369  */
370 int			arc_no_grow_shift = 5;
371 
372 
373 /*
374  * minimum lifespan of a prefetch block in clock ticks
375  * (initialized in arc_init())
376  */
377 static int		zfs_arc_min_prefetch_ms = 1;
378 static int		zfs_arc_min_prescient_prefetch_ms = 6;
379 
380 /*
381  * If this percent of memory is free, don't throttle.
382  */
383 int arc_lotsfree_percent = 10;
384 
385 static boolean_t arc_initialized;
386 
387 /*
388  * The arc has filled available memory and has now warmed up.
389  */
390 static boolean_t arc_warm;
391 
392 /*
393  * log2 fraction of the zio arena to keep free.
394  */
395 int arc_zio_arena_free_shift = 2;
396 
397 /*
398  * These tunables are for performance analysis.
399  */
400 uint64_t zfs_arc_max;
401 uint64_t zfs_arc_min;
402 uint64_t zfs_arc_meta_limit = 0;
403 uint64_t zfs_arc_meta_min = 0;
404 int zfs_arc_grow_retry = 0;
405 int zfs_arc_shrink_shift = 0;
406 int zfs_arc_p_min_shift = 0;
407 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
408 
409 /*
410  * ARC dirty data constraints for arc_tempreserve_space() throttle
411  */
412 uint_t zfs_arc_dirty_limit_percent = 50;	/* total dirty data limit */
413 uint_t zfs_arc_anon_limit_percent = 25;		/* anon block dirty limit */
414 uint_t zfs_arc_pool_dirty_percent = 20;		/* each pool's anon allowance */
415 
416 boolean_t zfs_compressed_arc_enabled = B_TRUE;
417 
418 /* The 6 states: */
419 static arc_state_t ARC_anon;
420 static arc_state_t ARC_mru;
421 static arc_state_t ARC_mru_ghost;
422 static arc_state_t ARC_mfu;
423 static arc_state_t ARC_mfu_ghost;
424 static arc_state_t ARC_l2c_only;
425 
426 arc_stats_t arc_stats = {
427 	{ "hits",			KSTAT_DATA_UINT64 },
428 	{ "misses",			KSTAT_DATA_UINT64 },
429 	{ "demand_data_hits",		KSTAT_DATA_UINT64 },
430 	{ "demand_data_misses",		KSTAT_DATA_UINT64 },
431 	{ "demand_metadata_hits",	KSTAT_DATA_UINT64 },
432 	{ "demand_metadata_misses",	KSTAT_DATA_UINT64 },
433 	{ "prefetch_data_hits",		KSTAT_DATA_UINT64 },
434 	{ "prefetch_data_misses",	KSTAT_DATA_UINT64 },
435 	{ "prefetch_metadata_hits",	KSTAT_DATA_UINT64 },
436 	{ "prefetch_metadata_misses",	KSTAT_DATA_UINT64 },
437 	{ "mru_hits",			KSTAT_DATA_UINT64 },
438 	{ "mru_ghost_hits",		KSTAT_DATA_UINT64 },
439 	{ "mfu_hits",			KSTAT_DATA_UINT64 },
440 	{ "mfu_ghost_hits",		KSTAT_DATA_UINT64 },
441 	{ "deleted",			KSTAT_DATA_UINT64 },
442 	{ "mutex_miss",			KSTAT_DATA_UINT64 },
443 	{ "access_skip",		KSTAT_DATA_UINT64 },
444 	{ "evict_skip",			KSTAT_DATA_UINT64 },
445 	{ "evict_not_enough",		KSTAT_DATA_UINT64 },
446 	{ "evict_l2_cached",		KSTAT_DATA_UINT64 },
447 	{ "evict_l2_eligible",		KSTAT_DATA_UINT64 },
448 	{ "evict_l2_eligible_mfu",	KSTAT_DATA_UINT64 },
449 	{ "evict_l2_eligible_mru",	KSTAT_DATA_UINT64 },
450 	{ "evict_l2_ineligible",	KSTAT_DATA_UINT64 },
451 	{ "evict_l2_skip",		KSTAT_DATA_UINT64 },
452 	{ "hash_elements",		KSTAT_DATA_UINT64 },
453 	{ "hash_elements_max",		KSTAT_DATA_UINT64 },
454 	{ "hash_collisions",		KSTAT_DATA_UINT64 },
455 	{ "hash_chains",		KSTAT_DATA_UINT64 },
456 	{ "hash_chain_max",		KSTAT_DATA_UINT64 },
457 	{ "p",				KSTAT_DATA_UINT64 },
458 	{ "c",				KSTAT_DATA_UINT64 },
459 	{ "c_min",			KSTAT_DATA_UINT64 },
460 	{ "c_max",			KSTAT_DATA_UINT64 },
461 	{ "size",			KSTAT_DATA_UINT64 },
462 	{ "compressed_size",		KSTAT_DATA_UINT64 },
463 	{ "uncompressed_size",		KSTAT_DATA_UINT64 },
464 	{ "overhead_size",		KSTAT_DATA_UINT64 },
465 	{ "hdr_size",			KSTAT_DATA_UINT64 },
466 	{ "data_size",			KSTAT_DATA_UINT64 },
467 	{ "metadata_size",		KSTAT_DATA_UINT64 },
468 	{ "other_size",			KSTAT_DATA_UINT64 },
469 	{ "anon_size",			KSTAT_DATA_UINT64 },
470 	{ "anon_evictable_data",	KSTAT_DATA_UINT64 },
471 	{ "anon_evictable_metadata",	KSTAT_DATA_UINT64 },
472 	{ "mru_size",			KSTAT_DATA_UINT64 },
473 	{ "mru_evictable_data",		KSTAT_DATA_UINT64 },
474 	{ "mru_evictable_metadata",	KSTAT_DATA_UINT64 },
475 	{ "mru_ghost_size",		KSTAT_DATA_UINT64 },
476 	{ "mru_ghost_evictable_data",	KSTAT_DATA_UINT64 },
477 	{ "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
478 	{ "mfu_size",			KSTAT_DATA_UINT64 },
479 	{ "mfu_evictable_data",		KSTAT_DATA_UINT64 },
480 	{ "mfu_evictable_metadata",	KSTAT_DATA_UINT64 },
481 	{ "mfu_ghost_size",		KSTAT_DATA_UINT64 },
482 	{ "mfu_ghost_evictable_data",	KSTAT_DATA_UINT64 },
483 	{ "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
484 	{ "l2_hits",			KSTAT_DATA_UINT64 },
485 	{ "l2_misses",			KSTAT_DATA_UINT64 },
486 	{ "l2_prefetch_asize",		KSTAT_DATA_UINT64 },
487 	{ "l2_mru_asize",		KSTAT_DATA_UINT64 },
488 	{ "l2_mfu_asize",		KSTAT_DATA_UINT64 },
489 	{ "l2_bufc_data_asize",		KSTAT_DATA_UINT64 },
490 	{ "l2_bufc_metadata_asize",	KSTAT_DATA_UINT64 },
491 	{ "l2_feeds",			KSTAT_DATA_UINT64 },
492 	{ "l2_rw_clash",		KSTAT_DATA_UINT64 },
493 	{ "l2_read_bytes",		KSTAT_DATA_UINT64 },
494 	{ "l2_write_bytes",		KSTAT_DATA_UINT64 },
495 	{ "l2_writes_sent",		KSTAT_DATA_UINT64 },
496 	{ "l2_writes_done",		KSTAT_DATA_UINT64 },
497 	{ "l2_writes_error",		KSTAT_DATA_UINT64 },
498 	{ "l2_writes_lock_retry",	KSTAT_DATA_UINT64 },
499 	{ "l2_evict_lock_retry",	KSTAT_DATA_UINT64 },
500 	{ "l2_evict_reading",		KSTAT_DATA_UINT64 },
501 	{ "l2_evict_l1cached",		KSTAT_DATA_UINT64 },
502 	{ "l2_free_on_write",		KSTAT_DATA_UINT64 },
503 	{ "l2_abort_lowmem",		KSTAT_DATA_UINT64 },
504 	{ "l2_cksum_bad",		KSTAT_DATA_UINT64 },
505 	{ "l2_io_error",		KSTAT_DATA_UINT64 },
506 	{ "l2_size",			KSTAT_DATA_UINT64 },
507 	{ "l2_asize",			KSTAT_DATA_UINT64 },
508 	{ "l2_hdr_size",		KSTAT_DATA_UINT64 },
509 	{ "l2_log_blk_writes",		KSTAT_DATA_UINT64 },
510 	{ "l2_log_blk_avg_asize",	KSTAT_DATA_UINT64 },
511 	{ "l2_log_blk_asize",		KSTAT_DATA_UINT64 },
512 	{ "l2_log_blk_count",		KSTAT_DATA_UINT64 },
513 	{ "l2_data_to_meta_ratio",	KSTAT_DATA_UINT64 },
514 	{ "l2_rebuild_success",		KSTAT_DATA_UINT64 },
515 	{ "l2_rebuild_unsupported",	KSTAT_DATA_UINT64 },
516 	{ "l2_rebuild_io_errors",	KSTAT_DATA_UINT64 },
517 	{ "l2_rebuild_dh_errors",	KSTAT_DATA_UINT64 },
518 	{ "l2_rebuild_cksum_lb_errors",	KSTAT_DATA_UINT64 },
519 	{ "l2_rebuild_lowmem",		KSTAT_DATA_UINT64 },
520 	{ "l2_rebuild_size",		KSTAT_DATA_UINT64 },
521 	{ "l2_rebuild_asize",		KSTAT_DATA_UINT64 },
522 	{ "l2_rebuild_bufs",		KSTAT_DATA_UINT64 },
523 	{ "l2_rebuild_bufs_precached",	KSTAT_DATA_UINT64 },
524 	{ "l2_rebuild_log_blks",	KSTAT_DATA_UINT64 },
525 	{ "memory_throttle_count",	KSTAT_DATA_UINT64 },
526 	{ "arc_meta_used",		KSTAT_DATA_UINT64 },
527 	{ "arc_meta_limit",		KSTAT_DATA_UINT64 },
528 	{ "arc_meta_max",		KSTAT_DATA_UINT64 },
529 	{ "arc_meta_min",		KSTAT_DATA_UINT64 },
530 	{ "async_upgrade_sync",		KSTAT_DATA_UINT64 },
531 	{ "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
532 	{ "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
533 };
534 
535 #define	ARCSTAT_MAX(stat, val) {					\
536 	uint64_t m;							\
537 	while ((val) > (m = arc_stats.stat.value.ui64) &&		\
538 	    (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val))))	\
539 		continue;						\
540 }
541 
542 #define	ARCSTAT_MAXSTAT(stat) \
543 	ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
544 
545 /*
546  * We define a macro to allow ARC hits/misses to be easily broken down by
547  * two separate conditions, giving a total of four different subtypes for
548  * each of hits and misses (so eight statistics total).
549  */
550 #define	ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
551 	if (cond1) {							\
552 		if (cond2) {						\
553 			ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
554 		} else {						\
555 			ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
556 		}							\
557 	} else {							\
558 		if (cond2) {						\
559 			ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
560 		} else {						\
561 			ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
562 		}							\
563 	}
564 
565 /*
566  * This macro allows us to use kstats as floating averages. Each time we
567  * update this kstat, we first factor it and the update value by
568  * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
569  * average. This macro assumes that integer loads and stores are atomic, but
570  * is not safe for multiple writers updating the kstat in parallel (only the
571  * last writer's update will remain).
572  */
573 #define	ARCSTAT_F_AVG_FACTOR	3
574 #define	ARCSTAT_F_AVG(stat, value) \
575 	do { \
576 		uint64_t x = ARCSTAT(stat); \
577 		x = x - x / ARCSTAT_F_AVG_FACTOR + \
578 		    (value) / ARCSTAT_F_AVG_FACTOR; \
579 		ARCSTAT(stat) = x; \
580 		_NOTE(CONSTCOND) \
581 	} while (0)
582 
583 kstat_t			*arc_ksp;
584 static arc_state_t	*arc_anon;
585 static arc_state_t	*arc_mru;
586 static arc_state_t	*arc_mru_ghost;
587 static arc_state_t	*arc_mfu;
588 static arc_state_t	*arc_mfu_ghost;
589 static arc_state_t	*arc_l2c_only;
590 
591 /*
592  * There are also some ARC variables that we want to export, but that are
593  * updated so often that having the canonical representation be the statistic
594  * variable causes a performance bottleneck. We want to use aggsum_t's for these
595  * instead, but still be able to export the kstat in the same way as before.
596  * The solution is to always use the aggsum version, except in the kstat update
597  * callback.
598  */
599 aggsum_t arc_size;
600 aggsum_t arc_meta_used;
601 aggsum_t astat_data_size;
602 aggsum_t astat_metadata_size;
603 aggsum_t astat_hdr_size;
604 aggsum_t astat_other_size;
605 aggsum_t astat_l2_hdr_size;
606 
607 static int		arc_no_grow;	/* Don't try to grow cache size */
608 static hrtime_t		arc_growtime;
609 static uint64_t		arc_tempreserve;
610 static uint64_t		arc_loaned_bytes;
611 
612 #define	GHOST_STATE(state)	\
613 	((state) == arc_mru_ghost || (state) == arc_mfu_ghost ||	\
614 	(state) == arc_l2c_only)
615 
616 #define	HDR_IN_HASH_TABLE(hdr)	((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
617 #define	HDR_IO_IN_PROGRESS(hdr)	((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
618 #define	HDR_IO_ERROR(hdr)	((hdr)->b_flags & ARC_FLAG_IO_ERROR)
619 #define	HDR_PREFETCH(hdr)	((hdr)->b_flags & ARC_FLAG_PREFETCH)
620 #define	HDR_PRESCIENT_PREFETCH(hdr)	\
621 	((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
622 #define	HDR_COMPRESSION_ENABLED(hdr)	\
623 	((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
624 
625 #define	HDR_L2CACHE(hdr)	((hdr)->b_flags & ARC_FLAG_L2CACHE)
626 #define	HDR_L2_READING(hdr)	\
627 	(((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) &&	\
628 	((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
629 #define	HDR_L2_WRITING(hdr)	((hdr)->b_flags & ARC_FLAG_L2_WRITING)
630 #define	HDR_L2_EVICTED(hdr)	((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
631 #define	HDR_L2_WRITE_HEAD(hdr)	((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
632 #define	HDR_PROTECTED(hdr)	((hdr)->b_flags & ARC_FLAG_PROTECTED)
633 #define	HDR_NOAUTH(hdr)		((hdr)->b_flags & ARC_FLAG_NOAUTH)
634 #define	HDR_SHARED_DATA(hdr)	((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
635 
636 #define	HDR_ISTYPE_METADATA(hdr)	\
637 	((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
638 #define	HDR_ISTYPE_DATA(hdr)	(!HDR_ISTYPE_METADATA(hdr))
639 
640 #define	HDR_HAS_L1HDR(hdr)	((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
641 #define	HDR_HAS_L2HDR(hdr)	((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
642 #define	HDR_HAS_RABD(hdr)	\
643 	(HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) &&	\
644 	(hdr)->b_crypt_hdr.b_rabd != NULL)
645 #define	HDR_ENCRYPTED(hdr)	\
646 	(HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
647 #define	HDR_AUTHENTICATED(hdr)	\
648 	(HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
649 
650 /* For storing compression mode in b_flags */
651 #define	HDR_COMPRESS_OFFSET	(highbit64(ARC_FLAG_COMPRESS_0) - 1)
652 
653 #define	HDR_GET_COMPRESS(hdr)	((enum zio_compress)BF32_GET((hdr)->b_flags, \
654 	HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
655 #define	HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
656 	HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
657 
658 #define	ARC_BUF_LAST(buf)	((buf)->b_next == NULL)
659 #define	ARC_BUF_SHARED(buf)	((buf)->b_flags & ARC_BUF_FLAG_SHARED)
660 #define	ARC_BUF_COMPRESSED(buf)	((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
661 #define	ARC_BUF_ENCRYPTED(buf)	((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
662 
663 /*
664  * Other sizes
665  */
666 
667 #define	HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
668 #define	HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
669 #define	HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
670 
671 /*
672  * Hash table routines
673  */
674 
675 #define	HT_LOCK_PAD	64
676 
677 struct ht_lock {
678 	kmutex_t	ht_lock;
679 #ifdef _KERNEL
680 	unsigned char	pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
681 #endif
682 };
683 
684 #define	BUF_LOCKS 256
685 typedef struct buf_hash_table {
686 	uint64_t ht_mask;
687 	arc_buf_hdr_t **ht_table;
688 	struct ht_lock ht_locks[BUF_LOCKS];
689 } buf_hash_table_t;
690 
691 static buf_hash_table_t buf_hash_table;
692 
693 #define	BUF_HASH_INDEX(spa, dva, birth) \
694 	(buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
695 #define	BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
696 #define	BUF_HASH_LOCK(idx)	(&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
697 #define	HDR_LOCK(hdr) \
698 	(BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
699 
700 uint64_t zfs_crc64_table[256];
701 
702 /*
703  * Level 2 ARC
704  */
705 
706 #define	L2ARC_WRITE_SIZE	(8 * 1024 * 1024)	/* initial write max */
707 #define	L2ARC_HEADROOM		2			/* num of writes */
708 /*
709  * If we discover during ARC scan any buffers to be compressed, we boost
710  * our headroom for the next scanning cycle by this percentage multiple.
711  */
712 #define	L2ARC_HEADROOM_BOOST	200
713 #define	L2ARC_FEED_SECS		1		/* caching interval secs */
714 #define	L2ARC_FEED_MIN_MS	200		/* min caching interval ms */
715 
716 /*
717  * We can feed L2ARC from two states of ARC buffers, mru and mfu,
718  * and each of the state has two types: data and metadata.
719  */
720 #define	L2ARC_FEED_TYPES	4
721 
722 
723 #define	l2arc_writes_sent	ARCSTAT(arcstat_l2_writes_sent)
724 #define	l2arc_writes_done	ARCSTAT(arcstat_l2_writes_done)
725 
726 /* L2ARC Performance Tunables */
727 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE;	/* default max write size */
728 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE;	/* extra write during warmup */
729 uint64_t l2arc_headroom = L2ARC_HEADROOM;	/* number of dev writes */
730 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
731 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS;	/* interval seconds */
732 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS;	/* min interval milliseconds */
733 boolean_t l2arc_noprefetch = B_TRUE;		/* don't cache prefetch bufs */
734 boolean_t l2arc_feed_again = B_TRUE;		/* turbo warmup */
735 boolean_t l2arc_norw = B_TRUE;			/* no reads during writes */
736 int l2arc_meta_percent = 33;			/* limit on headers size */
737 
738 /*
739  * L2ARC Internals
740  */
741 static list_t L2ARC_dev_list;			/* device list */
742 static list_t *l2arc_dev_list;			/* device list pointer */
743 static kmutex_t l2arc_dev_mtx;			/* device list mutex */
744 static l2arc_dev_t *l2arc_dev_last;		/* last device used */
745 static list_t L2ARC_free_on_write;		/* free after write buf list */
746 static list_t *l2arc_free_on_write;		/* free after write list ptr */
747 static kmutex_t l2arc_free_on_write_mtx;	/* mutex for list */
748 static uint64_t l2arc_ndev;			/* number of devices */
749 
750 typedef struct l2arc_read_callback {
751 	arc_buf_hdr_t		*l2rcb_hdr;		/* read header */
752 	blkptr_t		l2rcb_bp;		/* original blkptr */
753 	zbookmark_phys_t	l2rcb_zb;		/* original bookmark */
754 	int			l2rcb_flags;		/* original flags */
755 	abd_t			*l2rcb_abd;		/* temporary buffer */
756 } l2arc_read_callback_t;
757 
758 typedef struct l2arc_data_free {
759 	/* protected by l2arc_free_on_write_mtx */
760 	abd_t		*l2df_abd;
761 	size_t		l2df_size;
762 	arc_buf_contents_t l2df_type;
763 	list_node_t	l2df_list_node;
764 } l2arc_data_free_t;
765 
766 static kmutex_t l2arc_feed_thr_lock;
767 static kcondvar_t l2arc_feed_thr_cv;
768 static uint8_t l2arc_thread_exit;
769 
770 static kmutex_t l2arc_rebuild_thr_lock;
771 static kcondvar_t l2arc_rebuild_thr_cv;
772 
773 enum arc_hdr_alloc_flags {
774 	ARC_HDR_ALLOC_RDATA = 0x1,
775 	ARC_HDR_DO_ADAPT = 0x2,
776 };
777 
778 
779 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *, boolean_t);
780 typedef enum arc_fill_flags {
781 	ARC_FILL_LOCKED		= 1 << 0, /* hdr lock is held */
782 	ARC_FILL_COMPRESSED	= 1 << 1, /* fill with compressed data */
783 	ARC_FILL_ENCRYPTED	= 1 << 2, /* fill with encrypted data */
784 	ARC_FILL_NOAUTH		= 1 << 3, /* don't attempt to authenticate */
785 	ARC_FILL_IN_PLACE	= 1 << 4  /* fill in place (special case) */
786 } arc_fill_flags_t;
787 
788 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
789 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *, boolean_t);
790 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
791 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
792 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
793 static void arc_hdr_free_pabd(arc_buf_hdr_t *, boolean_t);
794 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *, int);
795 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
796 static boolean_t arc_is_overflowing();
797 static void arc_buf_watch(arc_buf_t *);
798 static l2arc_dev_t *l2arc_vdev_get(vdev_t *vd);
799 
800 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
801 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
802 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
803 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
804 
805 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
806 static void l2arc_read_done(zio_t *);
807 static void l2arc_do_free_on_write(void);
808 static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
809     boolean_t state_only);
810 
811 #define	l2arc_hdr_arcstats_increment(hdr) \
812 	l2arc_hdr_arcstats_update((hdr), B_TRUE, B_FALSE)
813 #define	l2arc_hdr_arcstats_decrement(hdr) \
814 	l2arc_hdr_arcstats_update((hdr), B_FALSE, B_FALSE)
815 #define	l2arc_hdr_arcstats_increment_state(hdr) \
816 	l2arc_hdr_arcstats_update((hdr), B_TRUE, B_TRUE)
817 #define	l2arc_hdr_arcstats_decrement_state(hdr) \
818 	l2arc_hdr_arcstats_update((hdr), B_FALSE, B_TRUE)
819 
820 /*
821  * The arc_all_memory function is a ZoL enhancement that lives in their OSL
822  * code. In user-space code, which is used primarily for testing, we return
823  * half of all memory.
824  */
825 uint64_t
arc_all_memory(void)826 arc_all_memory(void)
827 {
828 #ifdef _KERNEL
829 	return (ptob(physmem));
830 #else
831 	return ((sysconf(_SC_PAGESIZE) * sysconf(_SC_PHYS_PAGES)) / 2);
832 #endif
833 }
834 
835 /*
836  * We use Cityhash for this. It's fast, and has good hash properties without
837  * requiring any large static buffers.
838  */
839 static uint64_t
buf_hash(uint64_t spa,const dva_t * dva,uint64_t birth)840 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
841 {
842 	return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
843 }
844 
845 #define	HDR_EMPTY(hdr)						\
846 	((hdr)->b_dva.dva_word[0] == 0 &&			\
847 	(hdr)->b_dva.dva_word[1] == 0)
848 
849 #define	HDR_EMPTY_OR_LOCKED(hdr)				\
850 	(HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))
851 
852 #define	HDR_EQUAL(spa, dva, birth, hdr)				\
853 	((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) &&	\
854 	((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) &&	\
855 	((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
856 
857 static void
buf_discard_identity(arc_buf_hdr_t * hdr)858 buf_discard_identity(arc_buf_hdr_t *hdr)
859 {
860 	hdr->b_dva.dva_word[0] = 0;
861 	hdr->b_dva.dva_word[1] = 0;
862 	hdr->b_birth = 0;
863 }
864 
865 static arc_buf_hdr_t *
buf_hash_find(uint64_t spa,const blkptr_t * bp,kmutex_t ** lockp)866 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
867 {
868 	const dva_t *dva = BP_IDENTITY(bp);
869 	uint64_t birth = BP_PHYSICAL_BIRTH(bp);
870 	uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
871 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
872 	arc_buf_hdr_t *hdr;
873 
874 	mutex_enter(hash_lock);
875 	for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
876 	    hdr = hdr->b_hash_next) {
877 		if (HDR_EQUAL(spa, dva, birth, hdr)) {
878 			*lockp = hash_lock;
879 			return (hdr);
880 		}
881 	}
882 	mutex_exit(hash_lock);
883 	*lockp = NULL;
884 	return (NULL);
885 }
886 
887 /*
888  * Insert an entry into the hash table.  If there is already an element
889  * equal to elem in the hash table, then the already existing element
890  * will be returned and the new element will not be inserted.
891  * Otherwise returns NULL.
892  * If lockp == NULL, the caller is assumed to already hold the hash lock.
893  */
894 static arc_buf_hdr_t *
buf_hash_insert(arc_buf_hdr_t * hdr,kmutex_t ** lockp)895 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
896 {
897 	uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
898 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
899 	arc_buf_hdr_t *fhdr;
900 	uint32_t i;
901 
902 	ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
903 	ASSERT(hdr->b_birth != 0);
904 	ASSERT(!HDR_IN_HASH_TABLE(hdr));
905 
906 	if (lockp != NULL) {
907 		*lockp = hash_lock;
908 		mutex_enter(hash_lock);
909 	} else {
910 		ASSERT(MUTEX_HELD(hash_lock));
911 	}
912 
913 	for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
914 	    fhdr = fhdr->b_hash_next, i++) {
915 		if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
916 			return (fhdr);
917 	}
918 
919 	hdr->b_hash_next = buf_hash_table.ht_table[idx];
920 	buf_hash_table.ht_table[idx] = hdr;
921 	arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
922 
923 	/* collect some hash table performance data */
924 	if (i > 0) {
925 		ARCSTAT_BUMP(arcstat_hash_collisions);
926 		if (i == 1)
927 			ARCSTAT_BUMP(arcstat_hash_chains);
928 
929 		ARCSTAT_MAX(arcstat_hash_chain_max, i);
930 	}
931 
932 	ARCSTAT_BUMP(arcstat_hash_elements);
933 	ARCSTAT_MAXSTAT(arcstat_hash_elements);
934 
935 	return (NULL);
936 }
937 
938 static void
buf_hash_remove(arc_buf_hdr_t * hdr)939 buf_hash_remove(arc_buf_hdr_t *hdr)
940 {
941 	arc_buf_hdr_t *fhdr, **hdrp;
942 	uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
943 
944 	ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
945 	ASSERT(HDR_IN_HASH_TABLE(hdr));
946 
947 	hdrp = &buf_hash_table.ht_table[idx];
948 	while ((fhdr = *hdrp) != hdr) {
949 		ASSERT3P(fhdr, !=, NULL);
950 		hdrp = &fhdr->b_hash_next;
951 	}
952 	*hdrp = hdr->b_hash_next;
953 	hdr->b_hash_next = NULL;
954 	arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
955 
956 	/* collect some hash table performance data */
957 	ARCSTAT_BUMPDOWN(arcstat_hash_elements);
958 
959 	if (buf_hash_table.ht_table[idx] &&
960 	    buf_hash_table.ht_table[idx]->b_hash_next == NULL)
961 		ARCSTAT_BUMPDOWN(arcstat_hash_chains);
962 }
963 
964 /*
965  * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU
966  *		metadata and data are cached from ARC into L2ARC.
967  */
968 int l2arc_mfuonly = 0;
969 
970 /*
971  * Global data structures and functions for the buf kmem cache.
972  */
973 
974 static kmem_cache_t *hdr_full_cache;
975 static kmem_cache_t *hdr_full_crypt_cache;
976 static kmem_cache_t *hdr_l2only_cache;
977 static kmem_cache_t *buf_cache;
978 
979 static void
buf_fini(void)980 buf_fini(void)
981 {
982 	int i;
983 
984 	kmem_free(buf_hash_table.ht_table,
985 	    (buf_hash_table.ht_mask + 1) * sizeof (void *));
986 	for (i = 0; i < BUF_LOCKS; i++)
987 		mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
988 	kmem_cache_destroy(hdr_full_cache);
989 	kmem_cache_destroy(hdr_full_crypt_cache);
990 	kmem_cache_destroy(hdr_l2only_cache);
991 	kmem_cache_destroy(buf_cache);
992 }
993 
994 /*
995  * Constructor callback - called when the cache is empty
996  * and a new buf is requested.
997  */
998 /* ARGSUSED */
999 static int
hdr_full_cons(void * vbuf,void * unused,int kmflag)1000 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1001 {
1002 	arc_buf_hdr_t *hdr = vbuf;
1003 
1004 	bzero(hdr, HDR_FULL_SIZE);
1005 	hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1006 	cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1007 	zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1008 	mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1009 	multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1010 	arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1011 
1012 	return (0);
1013 }
1014 
1015 /* ARGSUSED */
1016 static int
hdr_full_crypt_cons(void * vbuf,void * unused,int kmflag)1017 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
1018 {
1019 	arc_buf_hdr_t *hdr = vbuf;
1020 
1021 	(void) hdr_full_cons(vbuf, unused, kmflag);
1022 	bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr));
1023 	arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1024 
1025 	return (0);
1026 }
1027 
1028 /* ARGSUSED */
1029 static int
hdr_l2only_cons(void * vbuf,void * unused,int kmflag)1030 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1031 {
1032 	arc_buf_hdr_t *hdr = vbuf;
1033 
1034 	bzero(hdr, HDR_L2ONLY_SIZE);
1035 	arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1036 
1037 	return (0);
1038 }
1039 
1040 /* ARGSUSED */
1041 static int
buf_cons(void * vbuf,void * unused,int kmflag)1042 buf_cons(void *vbuf, void *unused, int kmflag)
1043 {
1044 	arc_buf_t *buf = vbuf;
1045 
1046 	bzero(buf, sizeof (arc_buf_t));
1047 	mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1048 	arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1049 
1050 	return (0);
1051 }
1052 
1053 /*
1054  * Destructor callback - called when a cached buf is
1055  * no longer required.
1056  */
1057 /* ARGSUSED */
1058 static void
hdr_full_dest(void * vbuf,void * unused)1059 hdr_full_dest(void *vbuf, void *unused)
1060 {
1061 	arc_buf_hdr_t *hdr = vbuf;
1062 
1063 	ASSERT(HDR_EMPTY(hdr));
1064 	cv_destroy(&hdr->b_l1hdr.b_cv);
1065 	zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1066 	mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1067 	ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1068 	arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1069 }
1070 
1071 /* ARGSUSED */
1072 static void
hdr_full_crypt_dest(void * vbuf,void * unused)1073 hdr_full_crypt_dest(void *vbuf, void *unused)
1074 {
1075 	arc_buf_hdr_t *hdr = vbuf;
1076 
1077 	hdr_full_dest(hdr, unused);
1078 	arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1079 }
1080 
1081 /* ARGSUSED */
1082 static void
hdr_l2only_dest(void * vbuf,void * unused)1083 hdr_l2only_dest(void *vbuf, void *unused)
1084 {
1085 	arc_buf_hdr_t *hdr = vbuf;
1086 
1087 	ASSERT(HDR_EMPTY(hdr));
1088 	arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1089 }
1090 
1091 /* ARGSUSED */
1092 static void
buf_dest(void * vbuf,void * unused)1093 buf_dest(void *vbuf, void *unused)
1094 {
1095 	arc_buf_t *buf = vbuf;
1096 
1097 	mutex_destroy(&buf->b_evict_lock);
1098 	arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1099 }
1100 
1101 /*
1102  * Reclaim callback -- invoked when memory is low.
1103  */
1104 /* ARGSUSED */
1105 static void
hdr_recl(void * unused)1106 hdr_recl(void *unused)
1107 {
1108 	dprintf("hdr_recl called\n");
1109 	/*
1110 	 * umem calls the reclaim func when we destroy the buf cache,
1111 	 * which is after we do arc_fini().
1112 	 */
1113 	if (arc_initialized)
1114 		zthr_wakeup(arc_reap_zthr);
1115 }
1116 
1117 static void
buf_init(void)1118 buf_init(void)
1119 {
1120 	uint64_t *ct;
1121 	uint64_t hsize = 1ULL << 12;
1122 	int i, j;
1123 
1124 	/*
1125 	 * The hash table is big enough to fill all of physical memory
1126 	 * with an average block size of zfs_arc_average_blocksize (default 8K).
1127 	 * By default, the table will take up
1128 	 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1129 	 */
1130 	while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1131 		hsize <<= 1;
1132 retry:
1133 	buf_hash_table.ht_mask = hsize - 1;
1134 	buf_hash_table.ht_table =
1135 	    kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1136 	if (buf_hash_table.ht_table == NULL) {
1137 		ASSERT(hsize > (1ULL << 8));
1138 		hsize >>= 1;
1139 		goto retry;
1140 	}
1141 
1142 	hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1143 	    0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1144 	hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1145 	    HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1146 	    hdr_recl, NULL, NULL, 0);
1147 	hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1148 	    HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1149 	    NULL, NULL, 0);
1150 	buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1151 	    0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1152 
1153 	for (i = 0; i < 256; i++)
1154 		for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1155 			*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1156 
1157 	for (i = 0; i < BUF_LOCKS; i++) {
1158 		mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1159 		    NULL, MUTEX_DEFAULT, NULL);
1160 	}
1161 }
1162 
1163 /*
1164  * This is the size that the buf occupies in memory. If the buf is compressed,
1165  * it will correspond to the compressed size. You should use this method of
1166  * getting the buf size unless you explicitly need the logical size.
1167  */
1168 int32_t
arc_buf_size(arc_buf_t * buf)1169 arc_buf_size(arc_buf_t *buf)
1170 {
1171 	return (ARC_BUF_COMPRESSED(buf) ?
1172 	    HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1173 }
1174 
1175 int32_t
arc_buf_lsize(arc_buf_t * buf)1176 arc_buf_lsize(arc_buf_t *buf)
1177 {
1178 	return (HDR_GET_LSIZE(buf->b_hdr));
1179 }
1180 
1181 /*
1182  * This function will return B_TRUE if the buffer is encrypted in memory.
1183  * This buffer can be decrypted by calling arc_untransform().
1184  */
1185 boolean_t
arc_is_encrypted(arc_buf_t * buf)1186 arc_is_encrypted(arc_buf_t *buf)
1187 {
1188 	return (ARC_BUF_ENCRYPTED(buf) != 0);
1189 }
1190 
1191 /*
1192  * Returns B_TRUE if the buffer represents data that has not had its MAC
1193  * verified yet.
1194  */
1195 boolean_t
arc_is_unauthenticated(arc_buf_t * buf)1196 arc_is_unauthenticated(arc_buf_t *buf)
1197 {
1198 	return (HDR_NOAUTH(buf->b_hdr) != 0);
1199 }
1200 
1201 void
arc_get_raw_params(arc_buf_t * buf,boolean_t * byteorder,uint8_t * salt,uint8_t * iv,uint8_t * mac)1202 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1203     uint8_t *iv, uint8_t *mac)
1204 {
1205 	arc_buf_hdr_t *hdr = buf->b_hdr;
1206 
1207 	ASSERT(HDR_PROTECTED(hdr));
1208 
1209 	bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
1210 	bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
1211 	bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
1212 	*byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1213 	    /* CONSTCOND */
1214 	    ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1215 }
1216 
1217 /*
1218  * Indicates how this buffer is compressed in memory. If it is not compressed
1219  * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1220  * arc_untransform() as long as it is also unencrypted.
1221  */
1222 enum zio_compress
arc_get_compression(arc_buf_t * buf)1223 arc_get_compression(arc_buf_t *buf)
1224 {
1225 	return (ARC_BUF_COMPRESSED(buf) ?
1226 	    HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1227 }
1228 
1229 #define	ARC_MINTIME	(hz>>4) /* 62 ms */
1230 
1231 /*
1232  * Return the compression algorithm used to store this data in the ARC. If ARC
1233  * compression is enabled or this is an encrypted block, this will be the same
1234  * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1235  */
1236 static inline enum zio_compress
arc_hdr_get_compress(arc_buf_hdr_t * hdr)1237 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1238 {
1239 	return (HDR_COMPRESSION_ENABLED(hdr) ?
1240 	    HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1241 }
1242 
1243 static inline boolean_t
arc_buf_is_shared(arc_buf_t * buf)1244 arc_buf_is_shared(arc_buf_t *buf)
1245 {
1246 	boolean_t shared = (buf->b_data != NULL &&
1247 	    buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1248 	    abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1249 	    buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1250 	IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1251 	IMPLY(shared, ARC_BUF_SHARED(buf));
1252 	IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1253 
1254 	/*
1255 	 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1256 	 * already being shared" requirement prevents us from doing that.
1257 	 */
1258 
1259 	return (shared);
1260 }
1261 
1262 /*
1263  * Free the checksum associated with this header. If there is no checksum, this
1264  * is a no-op.
1265  */
1266 static inline void
arc_cksum_free(arc_buf_hdr_t * hdr)1267 arc_cksum_free(arc_buf_hdr_t *hdr)
1268 {
1269 	ASSERT(HDR_HAS_L1HDR(hdr));
1270 
1271 	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1272 	if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1273 		kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1274 		hdr->b_l1hdr.b_freeze_cksum = NULL;
1275 	}
1276 	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1277 }
1278 
1279 /*
1280  * Return true iff at least one of the bufs on hdr is not compressed.
1281  * Encrypted buffers count as compressed.
1282  */
1283 static boolean_t
arc_hdr_has_uncompressed_buf(arc_buf_hdr_t * hdr)1284 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1285 {
1286 	ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr));
1287 
1288 	for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1289 		if (!ARC_BUF_COMPRESSED(b)) {
1290 			return (B_TRUE);
1291 		}
1292 	}
1293 	return (B_FALSE);
1294 }
1295 
1296 /*
1297  * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1298  * matches the checksum that is stored in the hdr. If there is no checksum,
1299  * or if the buf is compressed, this is a no-op.
1300  */
1301 static void
arc_cksum_verify(arc_buf_t * buf)1302 arc_cksum_verify(arc_buf_t *buf)
1303 {
1304 	arc_buf_hdr_t *hdr = buf->b_hdr;
1305 	zio_cksum_t zc;
1306 
1307 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1308 		return;
1309 
1310 	if (ARC_BUF_COMPRESSED(buf))
1311 		return;
1312 
1313 	ASSERT(HDR_HAS_L1HDR(hdr));
1314 
1315 	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1316 
1317 	if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1318 		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1319 		return;
1320 	}
1321 
1322 	fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1323 	if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1324 		panic("buffer modified while frozen!");
1325 	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1326 }
1327 
1328 /*
1329  * This function makes the assumption that data stored in the L2ARC
1330  * will be transformed exactly as it is in the main pool. Because of
1331  * this we can verify the checksum against the reading process's bp.
1332  */
1333 static boolean_t
arc_cksum_is_equal(arc_buf_hdr_t * hdr,zio_t * zio)1334 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1335 {
1336 	ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1337 	VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1338 
1339 	/*
1340 	 * Block pointers always store the checksum for the logical data.
1341 	 * If the block pointer has the gang bit set, then the checksum
1342 	 * it represents is for the reconstituted data and not for an
1343 	 * individual gang member. The zio pipeline, however, must be able to
1344 	 * determine the checksum of each of the gang constituents so it
1345 	 * treats the checksum comparison differently than what we need
1346 	 * for l2arc blocks. This prevents us from using the
1347 	 * zio_checksum_error() interface directly. Instead we must call the
1348 	 * zio_checksum_error_impl() so that we can ensure the checksum is
1349 	 * generated using the correct checksum algorithm and accounts for the
1350 	 * logical I/O size and not just a gang fragment.
1351 	 */
1352 	return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1353 	    BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1354 	    zio->io_offset, NULL) == 0);
1355 }
1356 
1357 /*
1358  * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1359  * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1360  * isn't modified later on. If buf is compressed or there is already a checksum
1361  * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1362  */
1363 static void
arc_cksum_compute(arc_buf_t * buf)1364 arc_cksum_compute(arc_buf_t *buf)
1365 {
1366 	arc_buf_hdr_t *hdr = buf->b_hdr;
1367 
1368 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1369 		return;
1370 
1371 	ASSERT(HDR_HAS_L1HDR(hdr));
1372 
1373 	mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1374 	if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1375 		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1376 		return;
1377 	}
1378 
1379 	ASSERT(!ARC_BUF_ENCRYPTED(buf));
1380 	ASSERT(!ARC_BUF_COMPRESSED(buf));
1381 	hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1382 	    KM_SLEEP);
1383 	fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1384 	    hdr->b_l1hdr.b_freeze_cksum);
1385 	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1386 	arc_buf_watch(buf);
1387 }
1388 
1389 #ifndef _KERNEL
1390 typedef struct procctl {
1391 	long cmd;
1392 	prwatch_t prwatch;
1393 } procctl_t;
1394 #endif
1395 
1396 /* ARGSUSED */
1397 static void
arc_buf_unwatch(arc_buf_t * buf)1398 arc_buf_unwatch(arc_buf_t *buf)
1399 {
1400 #ifndef _KERNEL
1401 	if (arc_watch) {
1402 		int result;
1403 		procctl_t ctl;
1404 		ctl.cmd = PCWATCH;
1405 		ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1406 		ctl.prwatch.pr_size = 0;
1407 		ctl.prwatch.pr_wflags = 0;
1408 		result = write(arc_procfd, &ctl, sizeof (ctl));
1409 		ASSERT3U(result, ==, sizeof (ctl));
1410 	}
1411 #endif
1412 }
1413 
1414 /* ARGSUSED */
1415 static void
arc_buf_watch(arc_buf_t * buf)1416 arc_buf_watch(arc_buf_t *buf)
1417 {
1418 #ifndef _KERNEL
1419 	if (arc_watch) {
1420 		int result;
1421 		procctl_t ctl;
1422 		ctl.cmd = PCWATCH;
1423 		ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1424 		ctl.prwatch.pr_size = arc_buf_size(buf);
1425 		ctl.prwatch.pr_wflags = WA_WRITE;
1426 		result = write(arc_procfd, &ctl, sizeof (ctl));
1427 		ASSERT3U(result, ==, sizeof (ctl));
1428 	}
1429 #endif
1430 }
1431 
1432 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t * hdr)1433 arc_buf_type(arc_buf_hdr_t *hdr)
1434 {
1435 	arc_buf_contents_t type;
1436 	if (HDR_ISTYPE_METADATA(hdr)) {
1437 		type = ARC_BUFC_METADATA;
1438 	} else {
1439 		type = ARC_BUFC_DATA;
1440 	}
1441 	VERIFY3U(hdr->b_type, ==, type);
1442 	return (type);
1443 }
1444 
1445 boolean_t
arc_is_metadata(arc_buf_t * buf)1446 arc_is_metadata(arc_buf_t *buf)
1447 {
1448 	return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1449 }
1450 
1451 static uint32_t
arc_bufc_to_flags(arc_buf_contents_t type)1452 arc_bufc_to_flags(arc_buf_contents_t type)
1453 {
1454 	switch (type) {
1455 	case ARC_BUFC_DATA:
1456 		/* metadata field is 0 if buffer contains normal data */
1457 		return (0);
1458 	case ARC_BUFC_METADATA:
1459 		return (ARC_FLAG_BUFC_METADATA);
1460 	default:
1461 		break;
1462 	}
1463 	panic("undefined ARC buffer type!");
1464 	return ((uint32_t)-1);
1465 }
1466 
1467 void
arc_buf_thaw(arc_buf_t * buf)1468 arc_buf_thaw(arc_buf_t *buf)
1469 {
1470 	arc_buf_hdr_t *hdr = buf->b_hdr;
1471 
1472 	ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1473 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1474 
1475 	arc_cksum_verify(buf);
1476 
1477 	/*
1478 	 * Compressed buffers do not manipulate the b_freeze_cksum.
1479 	 */
1480 	if (ARC_BUF_COMPRESSED(buf))
1481 		return;
1482 
1483 	ASSERT(HDR_HAS_L1HDR(hdr));
1484 	arc_cksum_free(hdr);
1485 
1486 	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1487 #ifdef ZFS_DEBUG
1488 	if (zfs_flags & ZFS_DEBUG_MODIFY) {
1489 		if (hdr->b_l1hdr.b_thawed != NULL)
1490 			kmem_free(hdr->b_l1hdr.b_thawed, 1);
1491 		hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1492 	}
1493 #endif
1494 
1495 	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1496 
1497 	arc_buf_unwatch(buf);
1498 }
1499 
1500 void
arc_buf_freeze(arc_buf_t * buf)1501 arc_buf_freeze(arc_buf_t *buf)
1502 {
1503 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1504 		return;
1505 
1506 	if (ARC_BUF_COMPRESSED(buf))
1507 		return;
1508 
1509 	ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1510 	arc_cksum_compute(buf);
1511 }
1512 
1513 /*
1514  * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1515  * the following functions should be used to ensure that the flags are
1516  * updated in a thread-safe way. When manipulating the flags either
1517  * the hash_lock must be held or the hdr must be undiscoverable. This
1518  * ensures that we're not racing with any other threads when updating
1519  * the flags.
1520  */
1521 static inline void
arc_hdr_set_flags(arc_buf_hdr_t * hdr,arc_flags_t flags)1522 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1523 {
1524 	ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1525 	hdr->b_flags |= flags;
1526 }
1527 
1528 static inline void
arc_hdr_clear_flags(arc_buf_hdr_t * hdr,arc_flags_t flags)1529 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1530 {
1531 	ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1532 	hdr->b_flags &= ~flags;
1533 }
1534 
1535 /*
1536  * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1537  * done in a special way since we have to clear and set bits
1538  * at the same time. Consumers that wish to set the compression bits
1539  * must use this function to ensure that the flags are updated in
1540  * thread-safe manner.
1541  */
1542 static void
arc_hdr_set_compress(arc_buf_hdr_t * hdr,enum zio_compress cmp)1543 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1544 {
1545 	ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1546 
1547 	/*
1548 	 * Holes and embedded blocks will always have a psize = 0 so
1549 	 * we ignore the compression of the blkptr and set the
1550 	 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
1551 	 * Holes and embedded blocks remain anonymous so we don't
1552 	 * want to uncompress them. Mark them as uncompressed.
1553 	 */
1554 	if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1555 		arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1556 		ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1557 	} else {
1558 		arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1559 		ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1560 	}
1561 
1562 	HDR_SET_COMPRESS(hdr, cmp);
1563 	ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1564 }
1565 
1566 /*
1567  * Looks for another buf on the same hdr which has the data decompressed, copies
1568  * from it, and returns true. If no such buf exists, returns false.
1569  */
1570 static boolean_t
arc_buf_try_copy_decompressed_data(arc_buf_t * buf)1571 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1572 {
1573 	arc_buf_hdr_t *hdr = buf->b_hdr;
1574 	boolean_t copied = B_FALSE;
1575 
1576 	ASSERT(HDR_HAS_L1HDR(hdr));
1577 	ASSERT3P(buf->b_data, !=, NULL);
1578 	ASSERT(!ARC_BUF_COMPRESSED(buf));
1579 
1580 	for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1581 	    from = from->b_next) {
1582 		/* can't use our own data buffer */
1583 		if (from == buf) {
1584 			continue;
1585 		}
1586 
1587 		if (!ARC_BUF_COMPRESSED(from)) {
1588 			bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1589 			copied = B_TRUE;
1590 			break;
1591 		}
1592 	}
1593 
1594 	/*
1595 	 * Note: With encryption support, the following assertion is no longer
1596 	 * necessarily valid. If we receive two back to back raw snapshots
1597 	 * (send -w), the second receive can use a hdr with a cksum already
1598 	 * calculated. This happens via:
1599 	 *    dmu_recv_stream() -> receive_read_record() -> arc_loan_raw_buf()
1600 	 * The rsend/send_mixed_raw test case exercises this code path.
1601 	 *
1602 	 * There were no decompressed bufs, so there should not be a
1603 	 * checksum on the hdr either.
1604 	 * EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1605 	 */
1606 
1607 	return (copied);
1608 }
1609 
1610 /*
1611  * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1612  */
1613 static uint64_t
arc_hdr_size(arc_buf_hdr_t * hdr)1614 arc_hdr_size(arc_buf_hdr_t *hdr)
1615 {
1616 	uint64_t size;
1617 
1618 	if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1619 	    HDR_GET_PSIZE(hdr) > 0) {
1620 		size = HDR_GET_PSIZE(hdr);
1621 	} else {
1622 		ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1623 		size = HDR_GET_LSIZE(hdr);
1624 	}
1625 	return (size);
1626 }
1627 
1628 static int
arc_hdr_authenticate(arc_buf_hdr_t * hdr,spa_t * spa,uint64_t dsobj)1629 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1630 {
1631 	int ret;
1632 	uint64_t csize;
1633 	uint64_t lsize = HDR_GET_LSIZE(hdr);
1634 	uint64_t psize = HDR_GET_PSIZE(hdr);
1635 	void *tmpbuf = NULL;
1636 	abd_t *abd = hdr->b_l1hdr.b_pabd;
1637 
1638 	ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1639 	ASSERT(HDR_AUTHENTICATED(hdr));
1640 	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1641 
1642 	/*
1643 	 * The MAC is calculated on the compressed data that is stored on disk.
1644 	 * However, if compressed arc is disabled we will only have the
1645 	 * decompressed data available to us now. Compress it into a temporary
1646 	 * abd so we can verify the MAC. The performance overhead of this will
1647 	 * be relatively low, since most objects in an encrypted objset will
1648 	 * be encrypted (instead of authenticated) anyway.
1649 	 */
1650 	if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1651 	    !HDR_COMPRESSION_ENABLED(hdr)) {
1652 		tmpbuf = zio_buf_alloc(lsize);
1653 		abd = abd_get_from_buf(tmpbuf, lsize);
1654 		abd_take_ownership_of_buf(abd, B_TRUE);
1655 
1656 		csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1657 		    hdr->b_l1hdr.b_pabd, tmpbuf, lsize);
1658 		ASSERT3U(csize, <=, psize);
1659 		abd_zero_off(abd, csize, psize - csize);
1660 	}
1661 
1662 	/*
1663 	 * Authentication is best effort. We authenticate whenever the key is
1664 	 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1665 	 */
1666 	if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1667 		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1668 		ASSERT3U(lsize, ==, psize);
1669 		ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1670 		    psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1671 	} else {
1672 		ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1673 		    hdr->b_crypt_hdr.b_mac);
1674 	}
1675 
1676 	if (ret == 0)
1677 		arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1678 	else if (ret != ENOENT)
1679 		goto error;
1680 
1681 	if (tmpbuf != NULL)
1682 		abd_free(abd);
1683 
1684 	return (0);
1685 
1686 error:
1687 	if (tmpbuf != NULL)
1688 		abd_free(abd);
1689 
1690 	return (ret);
1691 }
1692 
1693 /*
1694  * This function will take a header that only has raw encrypted data in
1695  * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1696  * b_l1hdr.b_pabd. If designated in the header flags, this function will
1697  * also decompress the data.
1698  */
1699 static int
arc_hdr_decrypt(arc_buf_hdr_t * hdr,spa_t * spa,const zbookmark_phys_t * zb)1700 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1701 {
1702 	int ret;
1703 	abd_t *cabd = NULL;
1704 	void *tmp = NULL;
1705 	boolean_t no_crypt = B_FALSE;
1706 	boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1707 
1708 	ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1709 	ASSERT(HDR_ENCRYPTED(hdr));
1710 
1711 	arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT);
1712 
1713 	ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1714 	    B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1715 	    hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1716 	    hdr->b_crypt_hdr.b_rabd, &no_crypt);
1717 	if (ret != 0)
1718 		goto error;
1719 
1720 	if (no_crypt) {
1721 		abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1722 		    HDR_GET_PSIZE(hdr));
1723 	}
1724 
1725 	/*
1726 	 * If this header has disabled arc compression but the b_pabd is
1727 	 * compressed after decrypting it, we need to decompress the newly
1728 	 * decrypted data.
1729 	 */
1730 	if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1731 	    !HDR_COMPRESSION_ENABLED(hdr)) {
1732 		/*
1733 		 * We want to make sure that we are correctly honoring the
1734 		 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1735 		 * and then loan a buffer from it, rather than allocating a
1736 		 * linear buffer and wrapping it in an abd later.
1737 		 */
1738 		cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr, B_TRUE);
1739 		tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1740 
1741 		ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1742 		    hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1743 		    HDR_GET_LSIZE(hdr));
1744 		if (ret != 0) {
1745 			abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1746 			goto error;
1747 		}
1748 
1749 		abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1750 		arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1751 		    arc_hdr_size(hdr), hdr);
1752 		hdr->b_l1hdr.b_pabd = cabd;
1753 	}
1754 
1755 	return (0);
1756 
1757 error:
1758 	arc_hdr_free_pabd(hdr, B_FALSE);
1759 	if (cabd != NULL)
1760 		arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
1761 
1762 	return (ret);
1763 }
1764 
1765 /*
1766  * This function is called during arc_buf_fill() to prepare the header's
1767  * abd plaintext pointer for use. This involves authenticated protected
1768  * data and decrypting encrypted data into the plaintext abd.
1769  */
1770 static int
arc_fill_hdr_crypt(arc_buf_hdr_t * hdr,kmutex_t * hash_lock,spa_t * spa,const zbookmark_phys_t * zb,boolean_t noauth)1771 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
1772     const zbookmark_phys_t *zb, boolean_t noauth)
1773 {
1774 	int ret;
1775 
1776 	ASSERT(HDR_PROTECTED(hdr));
1777 
1778 	if (hash_lock != NULL)
1779 		mutex_enter(hash_lock);
1780 
1781 	if (HDR_NOAUTH(hdr) && !noauth) {
1782 		/*
1783 		 * The caller requested authenticated data but our data has
1784 		 * not been authenticated yet. Verify the MAC now if we can.
1785 		 */
1786 		ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
1787 		if (ret != 0)
1788 			goto error;
1789 	} else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
1790 		/*
1791 		 * If we only have the encrypted version of the data, but the
1792 		 * unencrypted version was requested we take this opportunity
1793 		 * to store the decrypted version in the header for future use.
1794 		 */
1795 		ret = arc_hdr_decrypt(hdr, spa, zb);
1796 		if (ret != 0)
1797 			goto error;
1798 	}
1799 
1800 	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1801 
1802 	if (hash_lock != NULL)
1803 		mutex_exit(hash_lock);
1804 
1805 	return (0);
1806 
1807 error:
1808 	if (hash_lock != NULL)
1809 		mutex_exit(hash_lock);
1810 
1811 	return (ret);
1812 }
1813 
1814 /*
1815  * This function is used by the dbuf code to decrypt bonus buffers in place.
1816  * The dbuf code itself doesn't have any locking for decrypting a shared dnode
1817  * block, so we use the hash lock here to protect against concurrent calls to
1818  * arc_buf_fill().
1819  */
1820 /* ARGSUSED */
1821 static void
arc_buf_untransform_in_place(arc_buf_t * buf,kmutex_t * hash_lock)1822 arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock)
1823 {
1824 	arc_buf_hdr_t *hdr = buf->b_hdr;
1825 
1826 	ASSERT(HDR_ENCRYPTED(hdr));
1827 	ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1828 	ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1829 	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1830 
1831 	zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
1832 	    arc_buf_size(buf));
1833 	buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
1834 	buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1835 	hdr->b_crypt_hdr.b_ebufcnt -= 1;
1836 }
1837 
1838 /*
1839  * Given a buf that has a data buffer attached to it, this function will
1840  * efficiently fill the buf with data of the specified compression setting from
1841  * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1842  * are already sharing a data buf, no copy is performed.
1843  *
1844  * If the buf is marked as compressed but uncompressed data was requested, this
1845  * will allocate a new data buffer for the buf, remove that flag, and fill the
1846  * buf with uncompressed data. You can't request a compressed buf on a hdr with
1847  * uncompressed data, and (since we haven't added support for it yet) if you
1848  * want compressed data your buf must already be marked as compressed and have
1849  * the correct-sized data buffer.
1850  */
1851 static int
arc_buf_fill(arc_buf_t * buf,spa_t * spa,const zbookmark_phys_t * zb,arc_fill_flags_t flags)1852 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
1853     arc_fill_flags_t flags)
1854 {
1855 	int error = 0;
1856 	arc_buf_hdr_t *hdr = buf->b_hdr;
1857 	boolean_t hdr_compressed =
1858 	    (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
1859 	boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
1860 	boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
1861 	dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
1862 	kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
1863 
1864 	ASSERT3P(buf->b_data, !=, NULL);
1865 	IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
1866 	IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
1867 	IMPLY(encrypted, HDR_ENCRYPTED(hdr));
1868 	IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
1869 	IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
1870 	IMPLY(encrypted, !ARC_BUF_SHARED(buf));
1871 
1872 	/*
1873 	 * If the caller wanted encrypted data we just need to copy it from
1874 	 * b_rabd and potentially byteswap it. We won't be able to do any
1875 	 * further transforms on it.
1876 	 */
1877 	if (encrypted) {
1878 		ASSERT(HDR_HAS_RABD(hdr));
1879 		abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
1880 		    HDR_GET_PSIZE(hdr));
1881 		goto byteswap;
1882 	}
1883 
1884 	/*
1885 	 * Adjust encrypted and authenticated headers to accomodate
1886 	 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
1887 	 * allowed to fail decryption due to keys not being loaded
1888 	 * without being marked as an IO error.
1889 	 */
1890 	if (HDR_PROTECTED(hdr)) {
1891 		error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
1892 		    zb, !!(flags & ARC_FILL_NOAUTH));
1893 		if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
1894 			return (error);
1895 		} else if (error != 0) {
1896 			if (hash_lock != NULL)
1897 				mutex_enter(hash_lock);
1898 			arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
1899 			if (hash_lock != NULL)
1900 				mutex_exit(hash_lock);
1901 			return (error);
1902 		}
1903 	}
1904 
1905 	/*
1906 	 * There is a special case here for dnode blocks which are
1907 	 * decrypting their bonus buffers. These blocks may request to
1908 	 * be decrypted in-place. This is necessary because there may
1909 	 * be many dnodes pointing into this buffer and there is
1910 	 * currently no method to synchronize replacing the backing
1911 	 * b_data buffer and updating all of the pointers. Here we use
1912 	 * the hash lock to ensure there are no races. If the need
1913 	 * arises for other types to be decrypted in-place, they must
1914 	 * add handling here as well.
1915 	 */
1916 	if ((flags & ARC_FILL_IN_PLACE) != 0) {
1917 		ASSERT(!hdr_compressed);
1918 		ASSERT(!compressed);
1919 		ASSERT(!encrypted);
1920 
1921 		if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
1922 			ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1923 
1924 			if (hash_lock != NULL)
1925 				mutex_enter(hash_lock);
1926 			arc_buf_untransform_in_place(buf, hash_lock);
1927 			if (hash_lock != NULL)
1928 				mutex_exit(hash_lock);
1929 
1930 			/* Compute the hdr's checksum if necessary */
1931 			arc_cksum_compute(buf);
1932 		}
1933 
1934 		return (0);
1935 	}
1936 
1937 	if (hdr_compressed == compressed) {
1938 		if (!arc_buf_is_shared(buf)) {
1939 			abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
1940 			    arc_buf_size(buf));
1941 		}
1942 	} else {
1943 		ASSERT(hdr_compressed);
1944 		ASSERT(!compressed);
1945 		ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
1946 
1947 		/*
1948 		 * If the buf is sharing its data with the hdr, unlink it and
1949 		 * allocate a new data buffer for the buf.
1950 		 */
1951 		if (arc_buf_is_shared(buf)) {
1952 			ASSERT(ARC_BUF_COMPRESSED(buf));
1953 
1954 			/* We need to give the buf its own b_data */
1955 			buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
1956 			buf->b_data =
1957 			    arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1958 			arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
1959 
1960 			/* Previously overhead was 0; just add new overhead */
1961 			ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
1962 		} else if (ARC_BUF_COMPRESSED(buf)) {
1963 			/* We need to reallocate the buf's b_data */
1964 			arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
1965 			    buf);
1966 			buf->b_data =
1967 			    arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1968 
1969 			/* We increased the size of b_data; update overhead */
1970 			ARCSTAT_INCR(arcstat_overhead_size,
1971 			    HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
1972 		}
1973 
1974 		/*
1975 		 * Regardless of the buf's previous compression settings, it
1976 		 * should not be compressed at the end of this function.
1977 		 */
1978 		buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1979 
1980 		/*
1981 		 * Try copying the data from another buf which already has a
1982 		 * decompressed version. If that's not possible, it's time to
1983 		 * bite the bullet and decompress the data from the hdr.
1984 		 */
1985 		if (arc_buf_try_copy_decompressed_data(buf)) {
1986 			/* Skip byteswapping and checksumming (already done) */
1987 			ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
1988 			return (0);
1989 		} else {
1990 			error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1991 			    hdr->b_l1hdr.b_pabd, buf->b_data,
1992 			    HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
1993 
1994 			/*
1995 			 * Absent hardware errors or software bugs, this should
1996 			 * be impossible, but log it anyway so we can debug it.
1997 			 */
1998 			if (error != 0) {
1999 				zfs_dbgmsg(
2000 				    "hdr %p, compress %d, psize %d, lsize %d",
2001 				    hdr, arc_hdr_get_compress(hdr),
2002 				    HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2003 				if (hash_lock != NULL)
2004 					mutex_enter(hash_lock);
2005 				arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2006 				if (hash_lock != NULL)
2007 					mutex_exit(hash_lock);
2008 				return (SET_ERROR(EIO));
2009 			}
2010 		}
2011 	}
2012 
2013 byteswap:
2014 	/* Byteswap the buf's data if necessary */
2015 	if (bswap != DMU_BSWAP_NUMFUNCS) {
2016 		ASSERT(!HDR_SHARED_DATA(hdr));
2017 		ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2018 		dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2019 	}
2020 
2021 	/* Compute the hdr's checksum if necessary */
2022 	arc_cksum_compute(buf);
2023 
2024 	return (0);
2025 }
2026 
2027 /*
2028  * If this function is being called to decrypt an encrypted buffer or verify an
2029  * authenticated one, the key must be loaded and a mapping must be made
2030  * available in the keystore via spa_keystore_create_mapping() or one of its
2031  * callers.
2032  */
2033 int
arc_untransform(arc_buf_t * buf,spa_t * spa,const zbookmark_phys_t * zb,boolean_t in_place)2034 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2035     boolean_t in_place)
2036 {
2037 	int ret;
2038 	arc_fill_flags_t flags = 0;
2039 
2040 	if (in_place)
2041 		flags |= ARC_FILL_IN_PLACE;
2042 
2043 	ret = arc_buf_fill(buf, spa, zb, flags);
2044 	if (ret == ECKSUM) {
2045 		/*
2046 		 * Convert authentication and decryption errors to EIO
2047 		 * (and generate an ereport) before leaving the ARC.
2048 		 */
2049 		ret = SET_ERROR(EIO);
2050 		spa_log_error(spa, zb);
2051 		(void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2052 		    spa, NULL, zb, NULL, 0, 0);
2053 	}
2054 
2055 	return (ret);
2056 }
2057 
2058 /*
2059  * Increment the amount of evictable space in the arc_state_t's refcount.
2060  * We account for the space used by the hdr and the arc buf individually
2061  * so that we can add and remove them from the refcount individually.
2062  */
2063 static void
arc_evictable_space_increment(arc_buf_hdr_t * hdr,arc_state_t * state)2064 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2065 {
2066 	arc_buf_contents_t type = arc_buf_type(hdr);
2067 
2068 	ASSERT(HDR_HAS_L1HDR(hdr));
2069 
2070 	if (GHOST_STATE(state)) {
2071 		ASSERT0(hdr->b_l1hdr.b_bufcnt);
2072 		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2073 		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2074 		ASSERT(!HDR_HAS_RABD(hdr));
2075 		(void) zfs_refcount_add_many(&state->arcs_esize[type],
2076 		    HDR_GET_LSIZE(hdr), hdr);
2077 		return;
2078 	}
2079 
2080 	ASSERT(!GHOST_STATE(state));
2081 	if (hdr->b_l1hdr.b_pabd != NULL) {
2082 		(void) zfs_refcount_add_many(&state->arcs_esize[type],
2083 		    arc_hdr_size(hdr), hdr);
2084 	}
2085 	if (HDR_HAS_RABD(hdr)) {
2086 		(void) zfs_refcount_add_many(&state->arcs_esize[type],
2087 		    HDR_GET_PSIZE(hdr), hdr);
2088 	}
2089 	for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2090 	    buf = buf->b_next) {
2091 		if (arc_buf_is_shared(buf))
2092 			continue;
2093 		(void) zfs_refcount_add_many(&state->arcs_esize[type],
2094 		    arc_buf_size(buf), buf);
2095 	}
2096 }
2097 
2098 /*
2099  * Decrement the amount of evictable space in the arc_state_t's refcount.
2100  * We account for the space used by the hdr and the arc buf individually
2101  * so that we can add and remove them from the refcount individually.
2102  */
2103 static void
arc_evictable_space_decrement(arc_buf_hdr_t * hdr,arc_state_t * state)2104 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2105 {
2106 	arc_buf_contents_t type = arc_buf_type(hdr);
2107 
2108 	ASSERT(HDR_HAS_L1HDR(hdr));
2109 
2110 	if (GHOST_STATE(state)) {
2111 		ASSERT0(hdr->b_l1hdr.b_bufcnt);
2112 		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2113 		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2114 		ASSERT(!HDR_HAS_RABD(hdr));
2115 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
2116 		    HDR_GET_LSIZE(hdr), hdr);
2117 		return;
2118 	}
2119 
2120 	ASSERT(!GHOST_STATE(state));
2121 	if (hdr->b_l1hdr.b_pabd != NULL) {
2122 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
2123 		    arc_hdr_size(hdr), hdr);
2124 	}
2125 	if (HDR_HAS_RABD(hdr)) {
2126 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
2127 		    HDR_GET_PSIZE(hdr), hdr);
2128 	}
2129 	for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2130 	    buf = buf->b_next) {
2131 		if (arc_buf_is_shared(buf))
2132 			continue;
2133 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
2134 		    arc_buf_size(buf), buf);
2135 	}
2136 }
2137 
2138 /*
2139  * Add a reference to this hdr indicating that someone is actively
2140  * referencing that memory. When the refcount transitions from 0 to 1,
2141  * we remove it from the respective arc_state_t list to indicate that
2142  * it is not evictable.
2143  */
2144 static void
add_reference(arc_buf_hdr_t * hdr,void * tag)2145 add_reference(arc_buf_hdr_t *hdr, void *tag)
2146 {
2147 	ASSERT(HDR_HAS_L1HDR(hdr));
2148 	if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) {
2149 		ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2150 		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2151 		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2152 	}
2153 
2154 	arc_state_t *state = hdr->b_l1hdr.b_state;
2155 
2156 	if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2157 	    (state != arc_anon)) {
2158 		/* We don't use the L2-only state list. */
2159 		if (state != arc_l2c_only) {
2160 			multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2161 			    hdr);
2162 			arc_evictable_space_decrement(hdr, state);
2163 		}
2164 		/* remove the prefetch flag if we get a reference */
2165 		if (HDR_HAS_L2HDR(hdr))
2166 			l2arc_hdr_arcstats_decrement_state(hdr);
2167 		arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2168 		if (HDR_HAS_L2HDR(hdr))
2169 			l2arc_hdr_arcstats_increment_state(hdr);
2170 	}
2171 }
2172 
2173 /*
2174  * Remove a reference from this hdr. When the reference transitions from
2175  * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2176  * list making it eligible for eviction.
2177  */
2178 static int
remove_reference(arc_buf_hdr_t * hdr,kmutex_t * hash_lock,void * tag)2179 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2180 {
2181 	int cnt;
2182 	arc_state_t *state = hdr->b_l1hdr.b_state;
2183 
2184 	ASSERT(HDR_HAS_L1HDR(hdr));
2185 	ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2186 	ASSERT(!GHOST_STATE(state));
2187 
2188 	/*
2189 	 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2190 	 * check to prevent usage of the arc_l2c_only list.
2191 	 */
2192 	if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2193 	    (state != arc_anon)) {
2194 		multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2195 		ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2196 		arc_evictable_space_increment(hdr, state);
2197 	}
2198 	return (cnt);
2199 }
2200 
2201 /*
2202  * Move the supplied buffer to the indicated state. The hash lock
2203  * for the buffer must be held by the caller.
2204  */
2205 static void
arc_change_state(arc_state_t * new_state,arc_buf_hdr_t * hdr,kmutex_t * hash_lock)2206 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2207     kmutex_t *hash_lock)
2208 {
2209 	arc_state_t *old_state;
2210 	int64_t refcnt;
2211 	uint32_t bufcnt;
2212 	boolean_t update_old, update_new;
2213 	arc_buf_contents_t buftype = arc_buf_type(hdr);
2214 
2215 	/*
2216 	 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2217 	 * in arc_read() when bringing a buffer out of the L2ARC.  However, the
2218 	 * L1 hdr doesn't always exist when we change state to arc_anon before
2219 	 * destroying a header, in which case reallocating to add the L1 hdr is
2220 	 * pointless.
2221 	 */
2222 	if (HDR_HAS_L1HDR(hdr)) {
2223 		old_state = hdr->b_l1hdr.b_state;
2224 		refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2225 		bufcnt = hdr->b_l1hdr.b_bufcnt;
2226 
2227 		update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2228 		    HDR_HAS_RABD(hdr));
2229 	} else {
2230 		old_state = arc_l2c_only;
2231 		refcnt = 0;
2232 		bufcnt = 0;
2233 		update_old = B_FALSE;
2234 	}
2235 	update_new = update_old;
2236 
2237 	ASSERT(MUTEX_HELD(hash_lock));
2238 	ASSERT3P(new_state, !=, old_state);
2239 	ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2240 	ASSERT(old_state != arc_anon || bufcnt <= 1);
2241 
2242 	/*
2243 	 * If this buffer is evictable, transfer it from the
2244 	 * old state list to the new state list.
2245 	 */
2246 	if (refcnt == 0) {
2247 		if (old_state != arc_anon && old_state != arc_l2c_only) {
2248 			ASSERT(HDR_HAS_L1HDR(hdr));
2249 			multilist_remove(old_state->arcs_list[buftype], hdr);
2250 
2251 			if (GHOST_STATE(old_state)) {
2252 				ASSERT0(bufcnt);
2253 				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2254 				update_old = B_TRUE;
2255 			}
2256 			arc_evictable_space_decrement(hdr, old_state);
2257 		}
2258 		if (new_state != arc_anon && new_state != arc_l2c_only) {
2259 
2260 			/*
2261 			 * An L1 header always exists here, since if we're
2262 			 * moving to some L1-cached state (i.e. not l2c_only or
2263 			 * anonymous), we realloc the header to add an L1hdr
2264 			 * beforehand.
2265 			 */
2266 			ASSERT(HDR_HAS_L1HDR(hdr));
2267 			multilist_insert(new_state->arcs_list[buftype], hdr);
2268 
2269 			if (GHOST_STATE(new_state)) {
2270 				ASSERT0(bufcnt);
2271 				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2272 				update_new = B_TRUE;
2273 			}
2274 			arc_evictable_space_increment(hdr, new_state);
2275 		}
2276 	}
2277 
2278 	ASSERT(!HDR_EMPTY(hdr));
2279 	if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2280 		buf_hash_remove(hdr);
2281 
2282 	/* adjust state sizes (ignore arc_l2c_only) */
2283 
2284 	if (update_new && new_state != arc_l2c_only) {
2285 		ASSERT(HDR_HAS_L1HDR(hdr));
2286 		if (GHOST_STATE(new_state)) {
2287 			ASSERT0(bufcnt);
2288 
2289 			/*
2290 			 * When moving a header to a ghost state, we first
2291 			 * remove all arc buffers. Thus, we'll have a
2292 			 * bufcnt of zero, and no arc buffer to use for
2293 			 * the reference. As a result, we use the arc
2294 			 * header pointer for the reference.
2295 			 */
2296 			(void) zfs_refcount_add_many(&new_state->arcs_size,
2297 			    HDR_GET_LSIZE(hdr), hdr);
2298 			ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2299 			ASSERT(!HDR_HAS_RABD(hdr));
2300 		} else {
2301 			uint32_t buffers = 0;
2302 
2303 			/*
2304 			 * Each individual buffer holds a unique reference,
2305 			 * thus we must remove each of these references one
2306 			 * at a time.
2307 			 */
2308 			for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2309 			    buf = buf->b_next) {
2310 				ASSERT3U(bufcnt, !=, 0);
2311 				buffers++;
2312 
2313 				/*
2314 				 * When the arc_buf_t is sharing the data
2315 				 * block with the hdr, the owner of the
2316 				 * reference belongs to the hdr. Only
2317 				 * add to the refcount if the arc_buf_t is
2318 				 * not shared.
2319 				 */
2320 				if (arc_buf_is_shared(buf))
2321 					continue;
2322 
2323 				(void) zfs_refcount_add_many(
2324 				    &new_state->arcs_size,
2325 				    arc_buf_size(buf), buf);
2326 			}
2327 			ASSERT3U(bufcnt, ==, buffers);
2328 
2329 			if (hdr->b_l1hdr.b_pabd != NULL) {
2330 				(void) zfs_refcount_add_many(
2331 				    &new_state->arcs_size,
2332 				    arc_hdr_size(hdr), hdr);
2333 			}
2334 
2335 			if (HDR_HAS_RABD(hdr)) {
2336 				(void) zfs_refcount_add_many(
2337 				    &new_state->arcs_size,
2338 				    HDR_GET_PSIZE(hdr), hdr);
2339 			}
2340 		}
2341 	}
2342 
2343 	if (update_old && old_state != arc_l2c_only) {
2344 		ASSERT(HDR_HAS_L1HDR(hdr));
2345 		if (GHOST_STATE(old_state)) {
2346 			ASSERT0(bufcnt);
2347 			ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2348 			ASSERT(!HDR_HAS_RABD(hdr));
2349 
2350 			/*
2351 			 * When moving a header off of a ghost state,
2352 			 * the header will not contain any arc buffers.
2353 			 * We use the arc header pointer for the reference
2354 			 * which is exactly what we did when we put the
2355 			 * header on the ghost state.
2356 			 */
2357 
2358 			(void) zfs_refcount_remove_many(&old_state->arcs_size,
2359 			    HDR_GET_LSIZE(hdr), hdr);
2360 		} else {
2361 			uint32_t buffers = 0;
2362 
2363 			/*
2364 			 * Each individual buffer holds a unique reference,
2365 			 * thus we must remove each of these references one
2366 			 * at a time.
2367 			 */
2368 			for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2369 			    buf = buf->b_next) {
2370 				ASSERT3U(bufcnt, !=, 0);
2371 				buffers++;
2372 
2373 				/*
2374 				 * When the arc_buf_t is sharing the data
2375 				 * block with the hdr, the owner of the
2376 				 * reference belongs to the hdr. Only
2377 				 * add to the refcount if the arc_buf_t is
2378 				 * not shared.
2379 				 */
2380 				if (arc_buf_is_shared(buf))
2381 					continue;
2382 
2383 				(void) zfs_refcount_remove_many(
2384 				    &old_state->arcs_size, arc_buf_size(buf),
2385 				    buf);
2386 			}
2387 			ASSERT3U(bufcnt, ==, buffers);
2388 			ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2389 			    HDR_HAS_RABD(hdr));
2390 
2391 			if (hdr->b_l1hdr.b_pabd != NULL) {
2392 				(void) zfs_refcount_remove_many(
2393 				    &old_state->arcs_size, arc_hdr_size(hdr),
2394 				    hdr);
2395 			}
2396 
2397 			if (HDR_HAS_RABD(hdr)) {
2398 				(void) zfs_refcount_remove_many(
2399 				    &old_state->arcs_size, HDR_GET_PSIZE(hdr),
2400 				    hdr);
2401 			}
2402 		}
2403 	}
2404 
2405 	if (HDR_HAS_L1HDR(hdr)) {
2406 		hdr->b_l1hdr.b_state = new_state;
2407 
2408 		if (HDR_HAS_L2HDR(hdr) && new_state != arc_l2c_only) {
2409 			l2arc_hdr_arcstats_decrement_state(hdr);
2410 			hdr->b_l2hdr.b_arcs_state = new_state->arcs_state;
2411 			l2arc_hdr_arcstats_increment_state(hdr);
2412 		}
2413 	}
2414 
2415 	/*
2416 	 * L2 headers should never be on the L2 state list since they don't
2417 	 * have L1 headers allocated.
2418 	 */
2419 	ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2420 	    multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2421 }
2422 
2423 void
arc_space_consume(uint64_t space,arc_space_type_t type)2424 arc_space_consume(uint64_t space, arc_space_type_t type)
2425 {
2426 	ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2427 
2428 	switch (type) {
2429 	case ARC_SPACE_DATA:
2430 		aggsum_add(&astat_data_size, space);
2431 		break;
2432 	case ARC_SPACE_META:
2433 		aggsum_add(&astat_metadata_size, space);
2434 		break;
2435 	case ARC_SPACE_OTHER:
2436 		aggsum_add(&astat_other_size, space);
2437 		break;
2438 	case ARC_SPACE_HDRS:
2439 		aggsum_add(&astat_hdr_size, space);
2440 		break;
2441 	case ARC_SPACE_L2HDRS:
2442 		aggsum_add(&astat_l2_hdr_size, space);
2443 		break;
2444 	}
2445 
2446 	if (type != ARC_SPACE_DATA)
2447 		aggsum_add(&arc_meta_used, space);
2448 
2449 	aggsum_add(&arc_size, space);
2450 }
2451 
2452 void
arc_space_return(uint64_t space,arc_space_type_t type)2453 arc_space_return(uint64_t space, arc_space_type_t type)
2454 {
2455 	ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2456 
2457 	switch (type) {
2458 	case ARC_SPACE_DATA:
2459 		aggsum_add(&astat_data_size, -space);
2460 		break;
2461 	case ARC_SPACE_META:
2462 		aggsum_add(&astat_metadata_size, -space);
2463 		break;
2464 	case ARC_SPACE_OTHER:
2465 		aggsum_add(&astat_other_size, -space);
2466 		break;
2467 	case ARC_SPACE_HDRS:
2468 		aggsum_add(&astat_hdr_size, -space);
2469 		break;
2470 	case ARC_SPACE_L2HDRS:
2471 		aggsum_add(&astat_l2_hdr_size, -space);
2472 		break;
2473 	}
2474 
2475 	if (type != ARC_SPACE_DATA) {
2476 		ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2477 		/*
2478 		 * We use the upper bound here rather than the precise value
2479 		 * because the arc_meta_max value doesn't need to be
2480 		 * precise. It's only consumed by humans via arcstats.
2481 		 */
2482 		if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2483 			arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2484 		aggsum_add(&arc_meta_used, -space);
2485 	}
2486 
2487 	ASSERT(aggsum_compare(&arc_size, space) >= 0);
2488 	aggsum_add(&arc_size, -space);
2489 }
2490 
2491 /*
2492  * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2493  * with the hdr's b_pabd.
2494  */
2495 static boolean_t
arc_can_share(arc_buf_hdr_t * hdr,arc_buf_t * buf)2496 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2497 {
2498 	/*
2499 	 * The criteria for sharing a hdr's data are:
2500 	 * 1. the buffer is not encrypted
2501 	 * 2. the hdr's compression matches the buf's compression
2502 	 * 3. the hdr doesn't need to be byteswapped
2503 	 * 4. the hdr isn't already being shared
2504 	 * 5. the buf is either compressed or it is the last buf in the hdr list
2505 	 *
2506 	 * Criterion #5 maintains the invariant that shared uncompressed
2507 	 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2508 	 * might ask, "if a compressed buf is allocated first, won't that be the
2509 	 * last thing in the list?", but in that case it's impossible to create
2510 	 * a shared uncompressed buf anyway (because the hdr must be compressed
2511 	 * to have the compressed buf). You might also think that #3 is
2512 	 * sufficient to make this guarantee, however it's possible
2513 	 * (specifically in the rare L2ARC write race mentioned in
2514 	 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2515 	 * is sharable, but wasn't at the time of its allocation. Rather than
2516 	 * allow a new shared uncompressed buf to be created and then shuffle
2517 	 * the list around to make it the last element, this simply disallows
2518 	 * sharing if the new buf isn't the first to be added.
2519 	 */
2520 	ASSERT3P(buf->b_hdr, ==, hdr);
2521 	boolean_t hdr_compressed = arc_hdr_get_compress(hdr) !=
2522 	    ZIO_COMPRESS_OFF;
2523 	boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2524 	return (!ARC_BUF_ENCRYPTED(buf) &&
2525 	    buf_compressed == hdr_compressed &&
2526 	    hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2527 	    !HDR_SHARED_DATA(hdr) &&
2528 	    (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2529 }
2530 
2531 /*
2532  * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2533  * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2534  * copy was made successfully, or an error code otherwise.
2535  */
2536 static int
arc_buf_alloc_impl(arc_buf_hdr_t * hdr,spa_t * spa,const zbookmark_phys_t * zb,void * tag,boolean_t encrypted,boolean_t compressed,boolean_t noauth,boolean_t fill,arc_buf_t ** ret)2537 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2538     void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth,
2539     boolean_t fill, arc_buf_t **ret)
2540 {
2541 	arc_buf_t *buf;
2542 	arc_fill_flags_t flags = ARC_FILL_LOCKED;
2543 
2544 	ASSERT(HDR_HAS_L1HDR(hdr));
2545 	ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2546 	VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2547 	    hdr->b_type == ARC_BUFC_METADATA);
2548 	ASSERT3P(ret, !=, NULL);
2549 	ASSERT3P(*ret, ==, NULL);
2550 	IMPLY(encrypted, compressed);
2551 
2552 	buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2553 	buf->b_hdr = hdr;
2554 	buf->b_data = NULL;
2555 	buf->b_next = hdr->b_l1hdr.b_buf;
2556 	buf->b_flags = 0;
2557 
2558 	add_reference(hdr, tag);
2559 
2560 	/*
2561 	 * We're about to change the hdr's b_flags. We must either
2562 	 * hold the hash_lock or be undiscoverable.
2563 	 */
2564 	ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2565 
2566 	/*
2567 	 * Only honor requests for compressed bufs if the hdr is actually
2568 	 * compressed. This must be overriden if the buffer is encrypted since
2569 	 * encrypted buffers cannot be decompressed.
2570 	 */
2571 	if (encrypted) {
2572 		buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2573 		buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2574 		flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2575 	} else if (compressed &&
2576 	    arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2577 		buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2578 		flags |= ARC_FILL_COMPRESSED;
2579 	}
2580 
2581 	if (noauth) {
2582 		ASSERT0(encrypted);
2583 		flags |= ARC_FILL_NOAUTH;
2584 	}
2585 
2586 	/*
2587 	 * If the hdr's data can be shared then we share the data buffer and
2588 	 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2589 	 * allocate a new buffer to store the buf's data.
2590 	 *
2591 	 * There are two additional restrictions here because we're sharing
2592 	 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2593 	 * actively involved in an L2ARC write, because if this buf is used by
2594 	 * an arc_write() then the hdr's data buffer will be released when the
2595 	 * write completes, even though the L2ARC write might still be using it.
2596 	 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2597 	 * need to be ABD-aware.
2598 	 */
2599 	boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2600 	    hdr->b_l1hdr.b_pabd != NULL && abd_is_linear(hdr->b_l1hdr.b_pabd);
2601 
2602 	/* Set up b_data and sharing */
2603 	if (can_share) {
2604 		buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2605 		buf->b_flags |= ARC_BUF_FLAG_SHARED;
2606 		arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2607 	} else {
2608 		buf->b_data =
2609 		    arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2610 		ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2611 	}
2612 	VERIFY3P(buf->b_data, !=, NULL);
2613 
2614 	hdr->b_l1hdr.b_buf = buf;
2615 	hdr->b_l1hdr.b_bufcnt += 1;
2616 	if (encrypted)
2617 		hdr->b_crypt_hdr.b_ebufcnt += 1;
2618 
2619 	/*
2620 	 * If the user wants the data from the hdr, we need to either copy or
2621 	 * decompress the data.
2622 	 */
2623 	if (fill) {
2624 		ASSERT3P(zb, !=, NULL);
2625 		return (arc_buf_fill(buf, spa, zb, flags));
2626 	}
2627 
2628 	return (0);
2629 }
2630 
2631 static char *arc_onloan_tag = "onloan";
2632 
2633 static inline void
arc_loaned_bytes_update(int64_t delta)2634 arc_loaned_bytes_update(int64_t delta)
2635 {
2636 	atomic_add_64(&arc_loaned_bytes, delta);
2637 
2638 	/* assert that it did not wrap around */
2639 	ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2640 }
2641 
2642 /*
2643  * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2644  * flight data by arc_tempreserve_space() until they are "returned". Loaned
2645  * buffers must be returned to the arc before they can be used by the DMU or
2646  * freed.
2647  */
2648 arc_buf_t *
arc_loan_buf(spa_t * spa,boolean_t is_metadata,int size)2649 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2650 {
2651 	arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2652 	    is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2653 
2654 	arc_loaned_bytes_update(arc_buf_size(buf));
2655 
2656 	return (buf);
2657 }
2658 
2659 arc_buf_t *
arc_loan_compressed_buf(spa_t * spa,uint64_t psize,uint64_t lsize,enum zio_compress compression_type)2660 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2661     enum zio_compress compression_type)
2662 {
2663 	arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2664 	    psize, lsize, compression_type);
2665 
2666 	arc_loaned_bytes_update(arc_buf_size(buf));
2667 
2668 	return (buf);
2669 }
2670 
2671 arc_buf_t *
arc_loan_raw_buf(spa_t * spa,uint64_t dsobj,boolean_t byteorder,const uint8_t * salt,const uint8_t * iv,const uint8_t * mac,dmu_object_type_t ot,uint64_t psize,uint64_t lsize,enum zio_compress compression_type)2672 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2673     const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2674     dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2675     enum zio_compress compression_type)
2676 {
2677 	arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2678 	    byteorder, salt, iv, mac, ot, psize, lsize, compression_type);
2679 
2680 	atomic_add_64(&arc_loaned_bytes, psize);
2681 	return (buf);
2682 }
2683 
2684 /*
2685  * Performance tuning of L2ARC persistence:
2686  *
2687  * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
2688  *		an L2ARC device (either at pool import or later) will attempt
2689  *		to rebuild L2ARC buffer contents.
2690  * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
2691  *		whether log blocks are written to the L2ARC device. If the L2ARC
2692  *		device is less than 1GB, the amount of data l2arc_evict()
2693  *		evicts is significant compared to the amount of restored L2ARC
2694  *		data. In this case do not write log blocks in L2ARC in order
2695  *		not to waste space.
2696  */
2697 int l2arc_rebuild_enabled = B_TRUE;
2698 unsigned long l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;
2699 
2700 /* L2ARC persistence rebuild control routines. */
2701 void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
2702 static void l2arc_dev_rebuild_start(l2arc_dev_t *dev);
2703 static int l2arc_rebuild(l2arc_dev_t *dev);
2704 
2705 /* L2ARC persistence read I/O routines. */
2706 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
2707 static int l2arc_log_blk_read(l2arc_dev_t *dev,
2708     const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
2709     l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
2710     zio_t *this_io, zio_t **next_io);
2711 static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
2712     const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
2713 static void l2arc_log_blk_fetch_abort(zio_t *zio);
2714 
2715 /* L2ARC persistence block restoration routines. */
2716 static void l2arc_log_blk_restore(l2arc_dev_t *dev,
2717     const l2arc_log_blk_phys_t *lb, uint64_t lb_asize);
2718 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
2719     l2arc_dev_t *dev);
2720 
2721 /* L2ARC persistence write I/O routines. */
2722 static void l2arc_dev_hdr_update(l2arc_dev_t *dev);
2723 static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
2724     l2arc_write_callback_t *cb);
2725 
2726 /* L2ARC persistence auxilliary routines. */
2727 boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
2728     const l2arc_log_blkptr_t *lbp);
2729 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
2730     const arc_buf_hdr_t *ab);
2731 boolean_t l2arc_range_check_overlap(uint64_t bottom,
2732     uint64_t top, uint64_t check);
2733 static void l2arc_blk_fetch_done(zio_t *zio);
2734 static inline uint64_t
2735     l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);
2736 
2737 /*
2738  * Return a loaned arc buffer to the arc.
2739  */
2740 void
arc_return_buf(arc_buf_t * buf,void * tag)2741 arc_return_buf(arc_buf_t *buf, void *tag)
2742 {
2743 	arc_buf_hdr_t *hdr = buf->b_hdr;
2744 
2745 	ASSERT3P(buf->b_data, !=, NULL);
2746 	ASSERT(HDR_HAS_L1HDR(hdr));
2747 	(void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2748 	(void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2749 
2750 	arc_loaned_bytes_update(-arc_buf_size(buf));
2751 }
2752 
2753 /* Detach an arc_buf from a dbuf (tag) */
2754 void
arc_loan_inuse_buf(arc_buf_t * buf,void * tag)2755 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2756 {
2757 	arc_buf_hdr_t *hdr = buf->b_hdr;
2758 
2759 	ASSERT3P(buf->b_data, !=, NULL);
2760 	ASSERT(HDR_HAS_L1HDR(hdr));
2761 	(void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2762 	(void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2763 
2764 	arc_loaned_bytes_update(arc_buf_size(buf));
2765 }
2766 
2767 static void
l2arc_free_abd_on_write(abd_t * abd,size_t size,arc_buf_contents_t type)2768 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2769 {
2770 	l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2771 
2772 	df->l2df_abd = abd;
2773 	df->l2df_size = size;
2774 	df->l2df_type = type;
2775 	mutex_enter(&l2arc_free_on_write_mtx);
2776 	list_insert_head(l2arc_free_on_write, df);
2777 	mutex_exit(&l2arc_free_on_write_mtx);
2778 }
2779 
2780 static void
arc_hdr_free_on_write(arc_buf_hdr_t * hdr,boolean_t free_rdata)2781 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2782 {
2783 	arc_state_t *state = hdr->b_l1hdr.b_state;
2784 	arc_buf_contents_t type = arc_buf_type(hdr);
2785 	uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
2786 
2787 	/* protected by hash lock, if in the hash table */
2788 	if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2789 		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2790 		ASSERT(state != arc_anon && state != arc_l2c_only);
2791 
2792 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
2793 		    size, hdr);
2794 	}
2795 	(void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
2796 	if (type == ARC_BUFC_METADATA) {
2797 		arc_space_return(size, ARC_SPACE_META);
2798 	} else {
2799 		ASSERT(type == ARC_BUFC_DATA);
2800 		arc_space_return(size, ARC_SPACE_DATA);
2801 	}
2802 
2803 	if (free_rdata) {
2804 		l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
2805 	} else {
2806 		l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2807 	}
2808 }
2809 
2810 /*
2811  * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2812  * data buffer, we transfer the refcount ownership to the hdr and update
2813  * the appropriate kstats.
2814  */
2815 static void
arc_share_buf(arc_buf_hdr_t * hdr,arc_buf_t * buf)2816 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2817 {
2818 	/* LINTED */
2819 	arc_state_t *state = hdr->b_l1hdr.b_state;
2820 
2821 	ASSERT(arc_can_share(hdr, buf));
2822 	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2823 	ASSERT(!ARC_BUF_ENCRYPTED(buf));
2824 	ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2825 
2826 	/*
2827 	 * Start sharing the data buffer. We transfer the
2828 	 * refcount ownership to the hdr since it always owns
2829 	 * the refcount whenever an arc_buf_t is shared.
2830 	 */
2831 	zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
2832 	    arc_hdr_size(hdr), buf, hdr);
2833 	hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2834 	abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2835 	    HDR_ISTYPE_METADATA(hdr));
2836 	arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2837 	buf->b_flags |= ARC_BUF_FLAG_SHARED;
2838 
2839 	/*
2840 	 * Since we've transferred ownership to the hdr we need
2841 	 * to increment its compressed and uncompressed kstats and
2842 	 * decrement the overhead size.
2843 	 */
2844 	ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2845 	ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2846 	ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2847 }
2848 
2849 static void
arc_unshare_buf(arc_buf_hdr_t * hdr,arc_buf_t * buf)2850 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2851 {
2852 	/* LINTED */
2853 	arc_state_t *state = hdr->b_l1hdr.b_state;
2854 
2855 	ASSERT(arc_buf_is_shared(buf));
2856 	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2857 	ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2858 
2859 	/*
2860 	 * We are no longer sharing this buffer so we need
2861 	 * to transfer its ownership to the rightful owner.
2862 	 */
2863 	zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
2864 	    arc_hdr_size(hdr), hdr, buf);
2865 	arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2866 	abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2867 	abd_put(hdr->b_l1hdr.b_pabd);
2868 	hdr->b_l1hdr.b_pabd = NULL;
2869 	buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2870 
2871 	/*
2872 	 * Since the buffer is no longer shared between
2873 	 * the arc buf and the hdr, count it as overhead.
2874 	 */
2875 	ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2876 	ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2877 	ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2878 }
2879 
2880 /*
2881  * Remove an arc_buf_t from the hdr's buf list and return the last
2882  * arc_buf_t on the list. If no buffers remain on the list then return
2883  * NULL.
2884  */
2885 static arc_buf_t *
arc_buf_remove(arc_buf_hdr_t * hdr,arc_buf_t * buf)2886 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2887 {
2888 	arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
2889 	arc_buf_t *lastbuf = NULL;
2890 
2891 	ASSERT(HDR_HAS_L1HDR(hdr));
2892 	ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2893 
2894 	/*
2895 	 * Remove the buf from the hdr list and locate the last
2896 	 * remaining buffer on the list.
2897 	 */
2898 	while (*bufp != NULL) {
2899 		if (*bufp == buf)
2900 			*bufp = buf->b_next;
2901 
2902 		/*
2903 		 * If we've removed a buffer in the middle of
2904 		 * the list then update the lastbuf and update
2905 		 * bufp.
2906 		 */
2907 		if (*bufp != NULL) {
2908 			lastbuf = *bufp;
2909 			bufp = &(*bufp)->b_next;
2910 		}
2911 	}
2912 	buf->b_next = NULL;
2913 	ASSERT3P(lastbuf, !=, buf);
2914 	IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
2915 	IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
2916 	IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
2917 
2918 	return (lastbuf);
2919 }
2920 
2921 /*
2922  * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2923  * list and free it.
2924  */
2925 static void
arc_buf_destroy_impl(arc_buf_t * buf)2926 arc_buf_destroy_impl(arc_buf_t *buf)
2927 {
2928 	arc_buf_hdr_t *hdr = buf->b_hdr;
2929 
2930 	/*
2931 	 * Free up the data associated with the buf but only if we're not
2932 	 * sharing this with the hdr. If we are sharing it with the hdr, the
2933 	 * hdr is responsible for doing the free.
2934 	 */
2935 	if (buf->b_data != NULL) {
2936 		/*
2937 		 * We're about to change the hdr's b_flags. We must either
2938 		 * hold the hash_lock or be undiscoverable.
2939 		 */
2940 		ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2941 
2942 		arc_cksum_verify(buf);
2943 		arc_buf_unwatch(buf);
2944 
2945 		if (arc_buf_is_shared(buf)) {
2946 			arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2947 		} else {
2948 			uint64_t size = arc_buf_size(buf);
2949 			arc_free_data_buf(hdr, buf->b_data, size, buf);
2950 			ARCSTAT_INCR(arcstat_overhead_size, -size);
2951 		}
2952 		buf->b_data = NULL;
2953 
2954 		ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
2955 		hdr->b_l1hdr.b_bufcnt -= 1;
2956 
2957 		if (ARC_BUF_ENCRYPTED(buf)) {
2958 			hdr->b_crypt_hdr.b_ebufcnt -= 1;
2959 
2960 			/*
2961 			 * If we have no more encrypted buffers and we've
2962 			 * already gotten a copy of the decrypted data we can
2963 			 * free b_rabd to save some space.
2964 			 */
2965 			if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
2966 			    HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
2967 			    !HDR_IO_IN_PROGRESS(hdr)) {
2968 				arc_hdr_free_pabd(hdr, B_TRUE);
2969 			}
2970 		}
2971 	}
2972 
2973 	arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
2974 
2975 	if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
2976 		/*
2977 		 * If the current arc_buf_t is sharing its data buffer with the
2978 		 * hdr, then reassign the hdr's b_pabd to share it with the new
2979 		 * buffer at the end of the list. The shared buffer is always
2980 		 * the last one on the hdr's buffer list.
2981 		 *
2982 		 * There is an equivalent case for compressed bufs, but since
2983 		 * they aren't guaranteed to be the last buf in the list and
2984 		 * that is an exceedingly rare case, we just allow that space be
2985 		 * wasted temporarily. We must also be careful not to share
2986 		 * encrypted buffers, since they cannot be shared.
2987 		 */
2988 		if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
2989 			/* Only one buf can be shared at once */
2990 			VERIFY(!arc_buf_is_shared(lastbuf));
2991 			/* hdr is uncompressed so can't have compressed buf */
2992 			VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
2993 
2994 			ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2995 			arc_hdr_free_pabd(hdr, B_FALSE);
2996 
2997 			/*
2998 			 * We must setup a new shared block between the
2999 			 * last buffer and the hdr. The data would have
3000 			 * been allocated by the arc buf so we need to transfer
3001 			 * ownership to the hdr since it's now being shared.
3002 			 */
3003 			arc_share_buf(hdr, lastbuf);
3004 		}
3005 	} else if (HDR_SHARED_DATA(hdr)) {
3006 		/*
3007 		 * Uncompressed shared buffers are always at the end
3008 		 * of the list. Compressed buffers don't have the
3009 		 * same requirements. This makes it hard to
3010 		 * simply assert that the lastbuf is shared so
3011 		 * we rely on the hdr's compression flags to determine
3012 		 * if we have a compressed, shared buffer.
3013 		 */
3014 		ASSERT3P(lastbuf, !=, NULL);
3015 		ASSERT(arc_buf_is_shared(lastbuf) ||
3016 		    arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3017 	}
3018 
3019 	/*
3020 	 * Free the checksum if we're removing the last uncompressed buf from
3021 	 * this hdr.
3022 	 */
3023 	if (!arc_hdr_has_uncompressed_buf(hdr)) {
3024 		arc_cksum_free(hdr);
3025 	}
3026 
3027 	/* clean up the buf */
3028 	buf->b_hdr = NULL;
3029 	kmem_cache_free(buf_cache, buf);
3030 }
3031 
3032 static void
arc_hdr_alloc_pabd(arc_buf_hdr_t * hdr,int alloc_flags)3033 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr, int alloc_flags)
3034 {
3035 	uint64_t size;
3036 	boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0);
3037 	boolean_t do_adapt = ((alloc_flags & ARC_HDR_DO_ADAPT) != 0);
3038 
3039 	ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3040 	ASSERT(HDR_HAS_L1HDR(hdr));
3041 	ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3042 	IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3043 
3044 	if (alloc_rdata) {
3045 		size = HDR_GET_PSIZE(hdr);
3046 		ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3047 		hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr,
3048 		    do_adapt);
3049 		ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3050 	} else {
3051 		size = arc_hdr_size(hdr);
3052 		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3053 		hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr,
3054 		    do_adapt);
3055 		ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3056 	}
3057 
3058 	ARCSTAT_INCR(arcstat_compressed_size, size);
3059 	ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3060 }
3061 
3062 static void
arc_hdr_free_pabd(arc_buf_hdr_t * hdr,boolean_t free_rdata)3063 arc_hdr_free_pabd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3064 {
3065 	uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3066 
3067 	ASSERT(HDR_HAS_L1HDR(hdr));
3068 	ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3069 	IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3070 
3071 
3072 	/*
3073 	 * If the hdr is currently being written to the l2arc then
3074 	 * we defer freeing the data by adding it to the l2arc_free_on_write
3075 	 * list. The l2arc will free the data once it's finished
3076 	 * writing it to the l2arc device.
3077 	 */
3078 	if (HDR_L2_WRITING(hdr)) {
3079 		arc_hdr_free_on_write(hdr, free_rdata);
3080 		ARCSTAT_BUMP(arcstat_l2_free_on_write);
3081 	} else if (free_rdata) {
3082 		arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3083 	} else {
3084 		arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3085 		    size, hdr);
3086 	}
3087 
3088 	if (free_rdata) {
3089 		hdr->b_crypt_hdr.b_rabd = NULL;
3090 	} else {
3091 		hdr->b_l1hdr.b_pabd = NULL;
3092 	}
3093 
3094 	if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3095 		hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3096 
3097 	ARCSTAT_INCR(arcstat_compressed_size, -size);
3098 	ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3099 }
3100 
3101 static arc_buf_hdr_t *
arc_hdr_alloc(uint64_t spa,int32_t psize,int32_t lsize,boolean_t protected,enum zio_compress compression_type,arc_buf_contents_t type,boolean_t alloc_rdata)3102 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3103     boolean_t protected, enum zio_compress compression_type,
3104     arc_buf_contents_t type, boolean_t alloc_rdata)
3105 {
3106 	arc_buf_hdr_t *hdr;
3107 	int flags = ARC_HDR_DO_ADAPT;
3108 
3109 	VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3110 	if (protected) {
3111 		hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3112 	} else {
3113 		hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3114 	}
3115 	flags |= alloc_rdata ? ARC_HDR_ALLOC_RDATA : 0;
3116 	ASSERT(HDR_EMPTY(hdr));
3117 	ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3118 	ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3119 	HDR_SET_PSIZE(hdr, psize);
3120 	HDR_SET_LSIZE(hdr, lsize);
3121 	hdr->b_spa = spa;
3122 	hdr->b_type = type;
3123 	hdr->b_flags = 0;
3124 	arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3125 	arc_hdr_set_compress(hdr, compression_type);
3126 	if (protected)
3127 		arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3128 
3129 	hdr->b_l1hdr.b_state = arc_anon;
3130 	hdr->b_l1hdr.b_arc_access = 0;
3131 	hdr->b_l1hdr.b_bufcnt = 0;
3132 	hdr->b_l1hdr.b_buf = NULL;
3133 
3134 	/*
3135 	 * Allocate the hdr's buffer. This will contain either
3136 	 * the compressed or uncompressed data depending on the block
3137 	 * it references and compressed arc enablement.
3138 	 */
3139 	arc_hdr_alloc_pabd(hdr, flags);
3140 	ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3141 
3142 	return (hdr);
3143 }
3144 
3145 /*
3146  * Transition between the two allocation states for the arc_buf_hdr struct.
3147  * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3148  * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3149  * version is used when a cache buffer is only in the L2ARC in order to reduce
3150  * memory usage.
3151  */
3152 static arc_buf_hdr_t *
arc_hdr_realloc(arc_buf_hdr_t * hdr,kmem_cache_t * old,kmem_cache_t * new)3153 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3154 {
3155 	ASSERT(HDR_HAS_L2HDR(hdr));
3156 
3157 	arc_buf_hdr_t *nhdr;
3158 	l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3159 
3160 	ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3161 	    (old == hdr_l2only_cache && new == hdr_full_cache));
3162 
3163 	/*
3164 	 * if the caller wanted a new full header and the header is to be
3165 	 * encrypted we will actually allocate the header from the full crypt
3166 	 * cache instead. The same applies to freeing from the old cache.
3167 	 */
3168 	if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3169 		new = hdr_full_crypt_cache;
3170 	if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3171 		old = hdr_full_crypt_cache;
3172 
3173 	nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3174 
3175 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3176 	buf_hash_remove(hdr);
3177 
3178 	bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3179 
3180 	if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3181 		arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3182 		/*
3183 		 * arc_access and arc_change_state need to be aware that a
3184 		 * header has just come out of L2ARC, so we set its state to
3185 		 * l2c_only even though it's about to change.
3186 		 */
3187 		nhdr->b_l1hdr.b_state = arc_l2c_only;
3188 
3189 		/* Verify previous threads set to NULL before freeing */
3190 		ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3191 		ASSERT(!HDR_HAS_RABD(hdr));
3192 	} else {
3193 		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3194 		ASSERT0(hdr->b_l1hdr.b_bufcnt);
3195 		ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3196 
3197 		/*
3198 		 * If we've reached here, We must have been called from
3199 		 * arc_evict_hdr(), as such we should have already been
3200 		 * removed from any ghost list we were previously on
3201 		 * (which protects us from racing with arc_evict_state),
3202 		 * thus no locking is needed during this check.
3203 		 */
3204 		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3205 
3206 		/*
3207 		 * A buffer must not be moved into the arc_l2c_only
3208 		 * state if it's not finished being written out to the
3209 		 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3210 		 * might try to be accessed, even though it was removed.
3211 		 */
3212 		VERIFY(!HDR_L2_WRITING(hdr));
3213 		VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3214 		ASSERT(!HDR_HAS_RABD(hdr));
3215 
3216 #ifdef ZFS_DEBUG
3217 		if (hdr->b_l1hdr.b_thawed != NULL) {
3218 			kmem_free(hdr->b_l1hdr.b_thawed, 1);
3219 			hdr->b_l1hdr.b_thawed = NULL;
3220 		}
3221 #endif
3222 
3223 		arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3224 	}
3225 	/*
3226 	 * The header has been reallocated so we need to re-insert it into any
3227 	 * lists it was on.
3228 	 */
3229 	(void) buf_hash_insert(nhdr, NULL);
3230 
3231 	ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3232 
3233 	mutex_enter(&dev->l2ad_mtx);
3234 
3235 	/*
3236 	 * We must place the realloc'ed header back into the list at
3237 	 * the same spot. Otherwise, if it's placed earlier in the list,
3238 	 * l2arc_write_buffers() could find it during the function's
3239 	 * write phase, and try to write it out to the l2arc.
3240 	 */
3241 	list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3242 	list_remove(&dev->l2ad_buflist, hdr);
3243 
3244 	mutex_exit(&dev->l2ad_mtx);
3245 
3246 	/*
3247 	 * Since we're using the pointer address as the tag when
3248 	 * incrementing and decrementing the l2ad_alloc refcount, we
3249 	 * must remove the old pointer (that we're about to destroy) and
3250 	 * add the new pointer to the refcount. Otherwise we'd remove
3251 	 * the wrong pointer address when calling arc_hdr_destroy() later.
3252 	 */
3253 
3254 	(void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3255 	    hdr);
3256 	(void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr),
3257 	    nhdr);
3258 
3259 	buf_discard_identity(hdr);
3260 	kmem_cache_free(old, hdr);
3261 
3262 	return (nhdr);
3263 }
3264 
3265 /*
3266  * This function allows an L1 header to be reallocated as a crypt
3267  * header and vice versa. If we are going to a crypt header, the
3268  * new fields will be zeroed out.
3269  */
3270 static arc_buf_hdr_t *
arc_hdr_realloc_crypt(arc_buf_hdr_t * hdr,boolean_t need_crypt)3271 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3272 {
3273 	arc_buf_hdr_t *nhdr;
3274 	arc_buf_t *buf;
3275 	kmem_cache_t *ncache, *ocache;
3276 
3277 	ASSERT(HDR_HAS_L1HDR(hdr));
3278 	ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3279 	ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3280 	ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3281 	ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
3282 	ASSERT3P(hdr->b_hash_next, ==, NULL);
3283 
3284 	if (need_crypt) {
3285 		ncache = hdr_full_crypt_cache;
3286 		ocache = hdr_full_cache;
3287 	} else {
3288 		ncache = hdr_full_cache;
3289 		ocache = hdr_full_crypt_cache;
3290 	}
3291 
3292 	nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3293 
3294 	/*
3295 	 * Copy all members that aren't locks or condvars to the new header.
3296 	 * No lists are pointing to us (as we asserted above), so we don't
3297 	 * need to worry about the list nodes.
3298 	 */
3299 	nhdr->b_dva = hdr->b_dva;
3300 	nhdr->b_birth = hdr->b_birth;
3301 	nhdr->b_type = hdr->b_type;
3302 	nhdr->b_flags = hdr->b_flags;
3303 	nhdr->b_psize = hdr->b_psize;
3304 	nhdr->b_lsize = hdr->b_lsize;
3305 	nhdr->b_spa = hdr->b_spa;
3306 	nhdr->b_l2hdr.b_dev = hdr->b_l2hdr.b_dev;
3307 	nhdr->b_l2hdr.b_daddr = hdr->b_l2hdr.b_daddr;
3308 	nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3309 	nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3310 	nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3311 	nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3312 	nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3313 	nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3314 	nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3315 #ifdef ZFS_DEBUG
3316 	if (hdr->b_l1hdr.b_thawed != NULL) {
3317 		nhdr->b_l1hdr.b_thawed = hdr->b_l1hdr.b_thawed;
3318 		hdr->b_l1hdr.b_thawed = NULL;
3319 	}
3320 #endif
3321 
3322 	/*
3323 	 * This refcount_add() exists only to ensure that the individual
3324 	 * arc buffers always point to a header that is referenced, avoiding
3325 	 * a small race condition that could trigger ASSERTs.
3326 	 */
3327 	(void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3328 	nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3329 	for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
3330 		mutex_enter(&buf->b_evict_lock);
3331 		buf->b_hdr = nhdr;
3332 		mutex_exit(&buf->b_evict_lock);
3333 	}
3334 	zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3335 	(void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3336 	ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3337 
3338 	if (need_crypt) {
3339 		arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3340 	} else {
3341 		arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3342 	}
3343 
3344 	/* unset all members of the original hdr */
3345 	bzero(&hdr->b_dva, sizeof (dva_t));
3346 	hdr->b_birth = 0;
3347 	hdr->b_type = ARC_BUFC_INVALID;
3348 	hdr->b_flags = 0;
3349 	hdr->b_psize = 0;
3350 	hdr->b_lsize = 0;
3351 	hdr->b_spa = 0;
3352 	hdr->b_l2hdr.b_dev = NULL;
3353 	hdr->b_l2hdr.b_daddr = 0;
3354 	hdr->b_l1hdr.b_freeze_cksum = NULL;
3355 	hdr->b_l1hdr.b_buf = NULL;
3356 	hdr->b_l1hdr.b_bufcnt = 0;
3357 	hdr->b_l1hdr.b_byteswap = 0;
3358 	hdr->b_l1hdr.b_state = NULL;
3359 	hdr->b_l1hdr.b_arc_access = 0;
3360 	hdr->b_l1hdr.b_acb = NULL;
3361 	hdr->b_l1hdr.b_pabd = NULL;
3362 
3363 	if (ocache == hdr_full_crypt_cache) {
3364 		ASSERT(!HDR_HAS_RABD(hdr));
3365 		hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
3366 		hdr->b_crypt_hdr.b_ebufcnt = 0;
3367 		hdr->b_crypt_hdr.b_dsobj = 0;
3368 		bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3369 		bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3370 		bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3371 	}
3372 
3373 	buf_discard_identity(hdr);
3374 	kmem_cache_free(ocache, hdr);
3375 
3376 	return (nhdr);
3377 }
3378 
3379 /*
3380  * This function is used by the send / receive code to convert a newly
3381  * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3382  * is also used to allow the root objset block to be uupdated without altering
3383  * its embedded MACs. Both block types will always be uncompressed so we do not
3384  * have to worry about compression type or psize.
3385  */
3386 void
arc_convert_to_raw(arc_buf_t * buf,uint64_t dsobj,boolean_t byteorder,dmu_object_type_t ot,const uint8_t * salt,const uint8_t * iv,const uint8_t * mac)3387 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3388     dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3389     const uint8_t *mac)
3390 {
3391 	arc_buf_hdr_t *hdr = buf->b_hdr;
3392 
3393 	ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3394 	ASSERT(HDR_HAS_L1HDR(hdr));
3395 	ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3396 
3397 	buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3398 	if (!HDR_PROTECTED(hdr))
3399 		hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3400 	hdr->b_crypt_hdr.b_dsobj = dsobj;
3401 	hdr->b_crypt_hdr.b_ot = ot;
3402 	hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3403 	    DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3404 	if (!arc_hdr_has_uncompressed_buf(hdr))
3405 		arc_cksum_free(hdr);
3406 
3407 	if (salt != NULL)
3408 		bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3409 	if (iv != NULL)
3410 		bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3411 	if (mac != NULL)
3412 		bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3413 }
3414 
3415 /*
3416  * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3417  * The buf is returned thawed since we expect the consumer to modify it.
3418  */
3419 arc_buf_t *
arc_alloc_buf(spa_t * spa,void * tag,arc_buf_contents_t type,int32_t size)3420 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3421 {
3422 	arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3423 	    B_FALSE, ZIO_COMPRESS_OFF, type, B_FALSE);
3424 
3425 	arc_buf_t *buf = NULL;
3426 	VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3427 	    B_FALSE, B_FALSE, &buf));
3428 	arc_buf_thaw(buf);
3429 
3430 	return (buf);
3431 }
3432 
3433 /*
3434  * Allocates an ARC buf header that's in an evicted & L2-cached state.
3435  * This is used during l2arc reconstruction to make empty ARC buffers
3436  * which circumvent the regular disk->arc->l2arc path and instead come
3437  * into being in the reverse order, i.e. l2arc->arc.
3438  */
3439 arc_buf_hdr_t *
arc_buf_alloc_l2only(size_t size,arc_buf_contents_t type,l2arc_dev_t * dev,dva_t dva,uint64_t daddr,int32_t psize,uint64_t birth,enum zio_compress compress,boolean_t protected,boolean_t prefetch,arc_state_type_t arcs_state)3440 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
3441     dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
3442     enum zio_compress compress, boolean_t protected,
3443     boolean_t prefetch, arc_state_type_t arcs_state)
3444 {
3445 	arc_buf_hdr_t	*hdr;
3446 
3447 	ASSERT(size != 0);
3448 	hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP);
3449 	hdr->b_birth = birth;
3450 	hdr->b_type = type;
3451 	hdr->b_flags = 0;
3452 	arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
3453 	HDR_SET_LSIZE(hdr, size);
3454 	HDR_SET_PSIZE(hdr, psize);
3455 	arc_hdr_set_compress(hdr, compress);
3456 	if (protected)
3457 		arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3458 	if (prefetch)
3459 		arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
3460 	hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa);
3461 
3462 	hdr->b_dva = dva;
3463 
3464 	hdr->b_l2hdr.b_dev = dev;
3465 	hdr->b_l2hdr.b_daddr = daddr;
3466 	hdr->b_l2hdr.b_arcs_state = arcs_state;
3467 
3468 	return (hdr);
3469 }
3470 
3471 /*
3472  * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3473  * for bufs containing metadata.
3474  */
3475 arc_buf_t *
arc_alloc_compressed_buf(spa_t * spa,void * tag,uint64_t psize,uint64_t lsize,enum zio_compress compression_type)3476 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3477     enum zio_compress compression_type)
3478 {
3479 	ASSERT3U(lsize, >, 0);
3480 	ASSERT3U(lsize, >=, psize);
3481 	ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3482 	ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3483 
3484 	arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3485 	    B_FALSE, compression_type, ARC_BUFC_DATA, B_FALSE);
3486 
3487 	arc_buf_t *buf = NULL;
3488 	VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3489 	    B_TRUE, B_FALSE, B_FALSE, &buf));
3490 	arc_buf_thaw(buf);
3491 	ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3492 
3493 	if (!arc_buf_is_shared(buf)) {
3494 		/*
3495 		 * To ensure that the hdr has the correct data in it if we call
3496 		 * arc_untransform() on this buf before it's been written to
3497 		 * disk, it's easiest if we just set up sharing between the
3498 		 * buf and the hdr.
3499 		 */
3500 		ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3501 		arc_hdr_free_pabd(hdr, B_FALSE);
3502 		arc_share_buf(hdr, buf);
3503 	}
3504 
3505 	return (buf);
3506 }
3507 
3508 arc_buf_t *
arc_alloc_raw_buf(spa_t * spa,void * tag,uint64_t dsobj,boolean_t byteorder,const uint8_t * salt,const uint8_t * iv,const uint8_t * mac,dmu_object_type_t ot,uint64_t psize,uint64_t lsize,enum zio_compress compression_type)3509 arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
3510     const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
3511     dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3512     enum zio_compress compression_type)
3513 {
3514 	arc_buf_hdr_t *hdr;
3515 	arc_buf_t *buf;
3516 	arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3517 	    ARC_BUFC_METADATA : ARC_BUFC_DATA;
3518 
3519 	ASSERT3U(lsize, >, 0);
3520 	ASSERT3U(lsize, >=, psize);
3521 	ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3522 	ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3523 
3524 	hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3525 	    compression_type, type, B_TRUE);
3526 
3527 	hdr->b_crypt_hdr.b_dsobj = dsobj;
3528 	hdr->b_crypt_hdr.b_ot = ot;
3529 	hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3530 	    DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3531 	bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3532 	bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3533 	bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3534 
3535 	/*
3536 	 * This buffer will be considered encrypted even if the ot is not an
3537 	 * encrypted type. It will become authenticated instead in
3538 	 * arc_write_ready().
3539 	 */
3540 	buf = NULL;
3541 	VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3542 	    B_FALSE, B_FALSE, &buf));
3543 	arc_buf_thaw(buf);
3544 	ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3545 
3546 	return (buf);
3547 }
3548 
3549 static void
l2arc_hdr_arcstats_update(arc_buf_hdr_t * hdr,boolean_t incr,boolean_t state_only)3550 l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
3551     boolean_t state_only)
3552 {
3553 	l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3554 	l2arc_dev_t *dev = l2hdr->b_dev;
3555 	uint64_t lsize = HDR_GET_LSIZE(hdr);
3556 	uint64_t psize = HDR_GET_PSIZE(hdr);
3557 	uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3558 	arc_buf_contents_t type = hdr->b_type;
3559 	int64_t lsize_s;
3560 	int64_t psize_s;
3561 	int64_t asize_s;
3562 
3563 	if (incr) {
3564 		lsize_s = lsize;
3565 		psize_s = psize;
3566 		asize_s = asize;
3567 	} else {
3568 		lsize_s = -lsize;
3569 		psize_s = -psize;
3570 		asize_s = -asize;
3571 	}
3572 
3573 	/* If the buffer is a prefetch, count it as such. */
3574 	if (HDR_PREFETCH(hdr)) {
3575 		ARCSTAT_INCR(arcstat_l2_prefetch_asize, asize_s);
3576 	} else {
3577 		/*
3578 		 * We use the value stored in the L2 header upon initial
3579 		 * caching in L2ARC. This value will be updated in case
3580 		 * an MRU/MRU_ghost buffer transitions to MFU but the L2ARC
3581 		 * metadata (log entry) cannot currently be updated. Having
3582 		 * the ARC state in the L2 header solves the problem of a
3583 		 * possibly absent L1 header (apparent in buffers restored
3584 		 * from persistent L2ARC).
3585 		 */
3586 		switch (hdr->b_l2hdr.b_arcs_state) {
3587 			case ARC_STATE_MRU_GHOST:
3588 			case ARC_STATE_MRU:
3589 				ARCSTAT_INCR(arcstat_l2_mru_asize, asize_s);
3590 				break;
3591 			case ARC_STATE_MFU_GHOST:
3592 			case ARC_STATE_MFU:
3593 				ARCSTAT_INCR(arcstat_l2_mfu_asize, asize_s);
3594 				break;
3595 			default:
3596 				break;
3597 		}
3598 	}
3599 
3600 	if (state_only)
3601 		return;
3602 
3603 	ARCSTAT_INCR(arcstat_l2_psize, psize_s);
3604 	ARCSTAT_INCR(arcstat_l2_lsize, lsize_s);
3605 
3606 	switch (type) {
3607 		case ARC_BUFC_DATA:
3608 			ARCSTAT_INCR(arcstat_l2_bufc_data_asize, asize_s);
3609 			break;
3610 		case ARC_BUFC_METADATA:
3611 			ARCSTAT_INCR(arcstat_l2_bufc_metadata_asize, asize_s);
3612 			break;
3613 		default:
3614 			break;
3615 	}
3616 }
3617 
3618 
3619 static void
arc_hdr_l2hdr_destroy(arc_buf_hdr_t * hdr)3620 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3621 {
3622 	l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3623 	l2arc_dev_t *dev = l2hdr->b_dev;
3624 	uint64_t psize = HDR_GET_PSIZE(hdr);
3625 	uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3626 
3627 	ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3628 	ASSERT(HDR_HAS_L2HDR(hdr));
3629 
3630 	list_remove(&dev->l2ad_buflist, hdr);
3631 
3632 	l2arc_hdr_arcstats_decrement(hdr);
3633 	vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3634 
3635 	(void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3636 	    hdr);
3637 	arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3638 }
3639 
3640 static void
arc_hdr_destroy(arc_buf_hdr_t * hdr)3641 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3642 {
3643 	if (HDR_HAS_L1HDR(hdr)) {
3644 		ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3645 		    hdr->b_l1hdr.b_bufcnt > 0);
3646 		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3647 		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3648 	}
3649 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3650 	ASSERT(!HDR_IN_HASH_TABLE(hdr));
3651 
3652 	if (HDR_HAS_L2HDR(hdr)) {
3653 		l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3654 		boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3655 
3656 		if (!buflist_held)
3657 			mutex_enter(&dev->l2ad_mtx);
3658 
3659 		/*
3660 		 * Even though we checked this conditional above, we
3661 		 * need to check this again now that we have the
3662 		 * l2ad_mtx. This is because we could be racing with
3663 		 * another thread calling l2arc_evict() which might have
3664 		 * destroyed this header's L2 portion as we were waiting
3665 		 * to acquire the l2ad_mtx. If that happens, we don't
3666 		 * want to re-destroy the header's L2 portion.
3667 		 */
3668 		if (HDR_HAS_L2HDR(hdr))
3669 			arc_hdr_l2hdr_destroy(hdr);
3670 
3671 		if (!buflist_held)
3672 			mutex_exit(&dev->l2ad_mtx);
3673 	}
3674 
3675 	/*
3676 	 * The header's identity can only be safely discarded once it is no
3677 	 * longer discoverable.  This requires removing it from the hash table
3678 	 * and the l2arc header list.  After this point the hash lock can not
3679 	 * be used to protect the header.
3680 	 */
3681 	if (!HDR_EMPTY(hdr))
3682 		buf_discard_identity(hdr);
3683 
3684 	if (HDR_HAS_L1HDR(hdr)) {
3685 		arc_cksum_free(hdr);
3686 
3687 		while (hdr->b_l1hdr.b_buf != NULL)
3688 			arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3689 
3690 #ifdef ZFS_DEBUG
3691 		if (hdr->b_l1hdr.b_thawed != NULL) {
3692 			kmem_free(hdr->b_l1hdr.b_thawed, 1);
3693 			hdr->b_l1hdr.b_thawed = NULL;
3694 		}
3695 #endif
3696 
3697 		if (hdr->b_l1hdr.b_pabd != NULL)
3698 			arc_hdr_free_pabd(hdr, B_FALSE);
3699 
3700 		if (HDR_HAS_RABD(hdr))
3701 			arc_hdr_free_pabd(hdr, B_TRUE);
3702 	}
3703 
3704 	ASSERT3P(hdr->b_hash_next, ==, NULL);
3705 	if (HDR_HAS_L1HDR(hdr)) {
3706 		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3707 		ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3708 
3709 		if (!HDR_PROTECTED(hdr)) {
3710 			kmem_cache_free(hdr_full_cache, hdr);
3711 		} else {
3712 			kmem_cache_free(hdr_full_crypt_cache, hdr);
3713 		}
3714 	} else {
3715 		kmem_cache_free(hdr_l2only_cache, hdr);
3716 	}
3717 }
3718 
3719 void
arc_buf_destroy(arc_buf_t * buf,void * tag)3720 arc_buf_destroy(arc_buf_t *buf, void* tag)
3721 {
3722 	arc_buf_hdr_t *hdr = buf->b_hdr;
3723 
3724 	if (hdr->b_l1hdr.b_state == arc_anon) {
3725 		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3726 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3727 		VERIFY0(remove_reference(hdr, NULL, tag));
3728 		arc_hdr_destroy(hdr);
3729 		return;
3730 	}
3731 
3732 	kmutex_t *hash_lock = HDR_LOCK(hdr);
3733 	mutex_enter(hash_lock);
3734 
3735 	ASSERT3P(hdr, ==, buf->b_hdr);
3736 	ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3737 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3738 	ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3739 	ASSERT3P(buf->b_data, !=, NULL);
3740 
3741 	(void) remove_reference(hdr, hash_lock, tag);
3742 	arc_buf_destroy_impl(buf);
3743 	mutex_exit(hash_lock);
3744 }
3745 
3746 /*
3747  * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3748  * state of the header is dependent on its state prior to entering this
3749  * function. The following transitions are possible:
3750  *
3751  *    - arc_mru -> arc_mru_ghost
3752  *    - arc_mfu -> arc_mfu_ghost
3753  *    - arc_mru_ghost -> arc_l2c_only
3754  *    - arc_mru_ghost -> deleted
3755  *    - arc_mfu_ghost -> arc_l2c_only
3756  *    - arc_mfu_ghost -> deleted
3757  */
3758 static int64_t
arc_evict_hdr(arc_buf_hdr_t * hdr,kmutex_t * hash_lock)3759 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3760 {
3761 	arc_state_t *evicted_state, *state;
3762 	int64_t bytes_evicted = 0;
3763 	int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3764 	    zfs_arc_min_prescient_prefetch_ms : zfs_arc_min_prefetch_ms;
3765 
3766 	ASSERT(MUTEX_HELD(hash_lock));
3767 	ASSERT(HDR_HAS_L1HDR(hdr));
3768 
3769 	state = hdr->b_l1hdr.b_state;
3770 	if (GHOST_STATE(state)) {
3771 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3772 		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3773 
3774 		/*
3775 		 * l2arc_write_buffers() relies on a header's L1 portion
3776 		 * (i.e. its b_pabd field) during its write phase.
3777 		 * Thus, we cannot push a header onto the arc_l2c_only
3778 		 * state (removing its L1 piece) until the header is
3779 		 * done being written to the l2arc.
3780 		 */
3781 		if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3782 			ARCSTAT_BUMP(arcstat_evict_l2_skip);
3783 			return (bytes_evicted);
3784 		}
3785 
3786 		ARCSTAT_BUMP(arcstat_deleted);
3787 		bytes_evicted += HDR_GET_LSIZE(hdr);
3788 
3789 		DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3790 
3791 		if (HDR_HAS_L2HDR(hdr)) {
3792 			ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3793 			ASSERT(!HDR_HAS_RABD(hdr));
3794 			/*
3795 			 * This buffer is cached on the 2nd Level ARC;
3796 			 * don't destroy the header.
3797 			 */
3798 			arc_change_state(arc_l2c_only, hdr, hash_lock);
3799 			/*
3800 			 * dropping from L1+L2 cached to L2-only,
3801 			 * realloc to remove the L1 header.
3802 			 */
3803 			hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3804 			    hdr_l2only_cache);
3805 		} else {
3806 			arc_change_state(arc_anon, hdr, hash_lock);
3807 			arc_hdr_destroy(hdr);
3808 		}
3809 		return (bytes_evicted);
3810 	}
3811 
3812 	ASSERT(state == arc_mru || state == arc_mfu);
3813 	evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3814 
3815 	/* prefetch buffers have a minimum lifespan */
3816 	if (HDR_IO_IN_PROGRESS(hdr) ||
3817 	    ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3818 	    ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < min_lifetime * hz)) {
3819 		ARCSTAT_BUMP(arcstat_evict_skip);
3820 		return (bytes_evicted);
3821 	}
3822 
3823 	ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3824 	while (hdr->b_l1hdr.b_buf) {
3825 		arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3826 		if (!mutex_tryenter(&buf->b_evict_lock)) {
3827 			ARCSTAT_BUMP(arcstat_mutex_miss);
3828 			break;
3829 		}
3830 		if (buf->b_data != NULL)
3831 			bytes_evicted += HDR_GET_LSIZE(hdr);
3832 		mutex_exit(&buf->b_evict_lock);
3833 		arc_buf_destroy_impl(buf);
3834 	}
3835 
3836 	if (HDR_HAS_L2HDR(hdr)) {
3837 		ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3838 	} else {
3839 		if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3840 			ARCSTAT_INCR(arcstat_evict_l2_eligible,
3841 			    HDR_GET_LSIZE(hdr));
3842 
3843 			switch (state->arcs_state) {
3844 				case ARC_STATE_MRU:
3845 					ARCSTAT_INCR(
3846 					    arcstat_evict_l2_eligible_mru,
3847 					    HDR_GET_LSIZE(hdr));
3848 					break;
3849 				case ARC_STATE_MFU:
3850 					ARCSTAT_INCR(
3851 					    arcstat_evict_l2_eligible_mfu,
3852 					    HDR_GET_LSIZE(hdr));
3853 					break;
3854 				default:
3855 					break;
3856 			}
3857 		} else {
3858 			ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3859 			    HDR_GET_LSIZE(hdr));
3860 		}
3861 	}
3862 
3863 	if (hdr->b_l1hdr.b_bufcnt == 0) {
3864 		arc_cksum_free(hdr);
3865 
3866 		bytes_evicted += arc_hdr_size(hdr);
3867 
3868 		/*
3869 		 * If this hdr is being evicted and has a compressed
3870 		 * buffer then we discard it here before we change states.
3871 		 * This ensures that the accounting is updated correctly
3872 		 * in arc_free_data_impl().
3873 		 */
3874 		if (hdr->b_l1hdr.b_pabd != NULL)
3875 			arc_hdr_free_pabd(hdr, B_FALSE);
3876 
3877 		if (HDR_HAS_RABD(hdr))
3878 			arc_hdr_free_pabd(hdr, B_TRUE);
3879 
3880 		arc_change_state(evicted_state, hdr, hash_lock);
3881 		ASSERT(HDR_IN_HASH_TABLE(hdr));
3882 		arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3883 		DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3884 	}
3885 
3886 	return (bytes_evicted);
3887 }
3888 
3889 static uint64_t
arc_evict_state_impl(multilist_t * ml,int idx,arc_buf_hdr_t * marker,uint64_t spa,int64_t bytes)3890 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3891     uint64_t spa, int64_t bytes)
3892 {
3893 	multilist_sublist_t *mls;
3894 	uint64_t bytes_evicted = 0;
3895 	arc_buf_hdr_t *hdr;
3896 	kmutex_t *hash_lock;
3897 	int evict_count = 0;
3898 
3899 	ASSERT3P(marker, !=, NULL);
3900 	IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3901 
3902 	mls = multilist_sublist_lock(ml, idx);
3903 
3904 	for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3905 	    hdr = multilist_sublist_prev(mls, marker)) {
3906 		if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3907 		    (evict_count >= zfs_arc_evict_batch_limit))
3908 			break;
3909 
3910 		/*
3911 		 * To keep our iteration location, move the marker
3912 		 * forward. Since we're not holding hdr's hash lock, we
3913 		 * must be very careful and not remove 'hdr' from the
3914 		 * sublist. Otherwise, other consumers might mistake the
3915 		 * 'hdr' as not being on a sublist when they call the
3916 		 * multilist_link_active() function (they all rely on
3917 		 * the hash lock protecting concurrent insertions and
3918 		 * removals). multilist_sublist_move_forward() was
3919 		 * specifically implemented to ensure this is the case
3920 		 * (only 'marker' will be removed and re-inserted).
3921 		 */
3922 		multilist_sublist_move_forward(mls, marker);
3923 
3924 		/*
3925 		 * The only case where the b_spa field should ever be
3926 		 * zero, is the marker headers inserted by
3927 		 * arc_evict_state(). It's possible for multiple threads
3928 		 * to be calling arc_evict_state() concurrently (e.g.
3929 		 * dsl_pool_close() and zio_inject_fault()), so we must
3930 		 * skip any markers we see from these other threads.
3931 		 */
3932 		if (hdr->b_spa == 0)
3933 			continue;
3934 
3935 		/* we're only interested in evicting buffers of a certain spa */
3936 		if (spa != 0 && hdr->b_spa != spa) {
3937 			ARCSTAT_BUMP(arcstat_evict_skip);
3938 			continue;
3939 		}
3940 
3941 		hash_lock = HDR_LOCK(hdr);
3942 
3943 		/*
3944 		 * We aren't calling this function from any code path
3945 		 * that would already be holding a hash lock, so we're
3946 		 * asserting on this assumption to be defensive in case
3947 		 * this ever changes. Without this check, it would be
3948 		 * possible to incorrectly increment arcstat_mutex_miss
3949 		 * below (e.g. if the code changed such that we called
3950 		 * this function with a hash lock held).
3951 		 */
3952 		ASSERT(!MUTEX_HELD(hash_lock));
3953 
3954 		if (mutex_tryenter(hash_lock)) {
3955 			uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3956 			mutex_exit(hash_lock);
3957 
3958 			bytes_evicted += evicted;
3959 
3960 			/*
3961 			 * If evicted is zero, arc_evict_hdr() must have
3962 			 * decided to skip this header, don't increment
3963 			 * evict_count in this case.
3964 			 */
3965 			if (evicted != 0)
3966 				evict_count++;
3967 
3968 			/*
3969 			 * If arc_size isn't overflowing, signal any
3970 			 * threads that might happen to be waiting.
3971 			 *
3972 			 * For each header evicted, we wake up a single
3973 			 * thread. If we used cv_broadcast, we could
3974 			 * wake up "too many" threads causing arc_size
3975 			 * to significantly overflow arc_c; since
3976 			 * arc_get_data_impl() doesn't check for overflow
3977 			 * when it's woken up (it doesn't because it's
3978 			 * possible for the ARC to be overflowing while
3979 			 * full of un-evictable buffers, and the
3980 			 * function should proceed in this case).
3981 			 *
3982 			 * If threads are left sleeping, due to not
3983 			 * using cv_broadcast here, they will be woken
3984 			 * up via cv_broadcast in arc_adjust_cb() just
3985 			 * before arc_adjust_zthr sleeps.
3986 			 */
3987 			mutex_enter(&arc_adjust_lock);
3988 			if (!arc_is_overflowing())
3989 				cv_signal(&arc_adjust_waiters_cv);
3990 			mutex_exit(&arc_adjust_lock);
3991 		} else {
3992 			ARCSTAT_BUMP(arcstat_mutex_miss);
3993 		}
3994 	}
3995 
3996 	multilist_sublist_unlock(mls);
3997 
3998 	return (bytes_evicted);
3999 }
4000 
4001 /*
4002  * Evict buffers from the given arc state, until we've removed the
4003  * specified number of bytes. Move the removed buffers to the
4004  * appropriate evict state.
4005  *
4006  * This function makes a "best effort". It skips over any buffers
4007  * it can't get a hash_lock on, and so, may not catch all candidates.
4008  * It may also return without evicting as much space as requested.
4009  *
4010  * If bytes is specified using the special value ARC_EVICT_ALL, this
4011  * will evict all available (i.e. unlocked and evictable) buffers from
4012  * the given arc state; which is used by arc_flush().
4013  */
4014 static uint64_t
arc_evict_state(arc_state_t * state,uint64_t spa,int64_t bytes,arc_buf_contents_t type)4015 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
4016     arc_buf_contents_t type)
4017 {
4018 	uint64_t total_evicted = 0;
4019 	multilist_t *ml = state->arcs_list[type];
4020 	int num_sublists;
4021 	arc_buf_hdr_t **markers;
4022 
4023 	IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
4024 
4025 	num_sublists = multilist_get_num_sublists(ml);
4026 
4027 	/*
4028 	 * If we've tried to evict from each sublist, made some
4029 	 * progress, but still have not hit the target number of bytes
4030 	 * to evict, we want to keep trying. The markers allow us to
4031 	 * pick up where we left off for each individual sublist, rather
4032 	 * than starting from the tail each time.
4033 	 */
4034 	markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
4035 	for (int i = 0; i < num_sublists; i++) {
4036 		markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4037 
4038 		/*
4039 		 * A b_spa of 0 is used to indicate that this header is
4040 		 * a marker. This fact is used in arc_adjust_type() and
4041 		 * arc_evict_state_impl().
4042 		 */
4043 		markers[i]->b_spa = 0;
4044 
4045 		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4046 		multilist_sublist_insert_tail(mls, markers[i]);
4047 		multilist_sublist_unlock(mls);
4048 	}
4049 
4050 	/*
4051 	 * While we haven't hit our target number of bytes to evict, or
4052 	 * we're evicting all available buffers.
4053 	 */
4054 	while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
4055 		/*
4056 		 * Start eviction using a randomly selected sublist,
4057 		 * this is to try and evenly balance eviction across all
4058 		 * sublists. Always starting at the same sublist
4059 		 * (e.g. index 0) would cause evictions to favor certain
4060 		 * sublists over others.
4061 		 */
4062 		int sublist_idx = multilist_get_random_index(ml);
4063 		uint64_t scan_evicted = 0;
4064 
4065 		for (int i = 0; i < num_sublists; i++) {
4066 			uint64_t bytes_remaining;
4067 			uint64_t bytes_evicted;
4068 
4069 			if (bytes == ARC_EVICT_ALL)
4070 				bytes_remaining = ARC_EVICT_ALL;
4071 			else if (total_evicted < bytes)
4072 				bytes_remaining = bytes - total_evicted;
4073 			else
4074 				break;
4075 
4076 			bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4077 			    markers[sublist_idx], spa, bytes_remaining);
4078 
4079 			scan_evicted += bytes_evicted;
4080 			total_evicted += bytes_evicted;
4081 
4082 			/* we've reached the end, wrap to the beginning */
4083 			if (++sublist_idx >= num_sublists)
4084 				sublist_idx = 0;
4085 		}
4086 
4087 		/*
4088 		 * If we didn't evict anything during this scan, we have
4089 		 * no reason to believe we'll evict more during another
4090 		 * scan, so break the loop.
4091 		 */
4092 		if (scan_evicted == 0) {
4093 			/* This isn't possible, let's make that obvious */
4094 			ASSERT3S(bytes, !=, 0);
4095 
4096 			/*
4097 			 * When bytes is ARC_EVICT_ALL, the only way to
4098 			 * break the loop is when scan_evicted is zero.
4099 			 * In that case, we actually have evicted enough,
4100 			 * so we don't want to increment the kstat.
4101 			 */
4102 			if (bytes != ARC_EVICT_ALL) {
4103 				ASSERT3S(total_evicted, <, bytes);
4104 				ARCSTAT_BUMP(arcstat_evict_not_enough);
4105 			}
4106 
4107 			break;
4108 		}
4109 	}
4110 
4111 	for (int i = 0; i < num_sublists; i++) {
4112 		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4113 		multilist_sublist_remove(mls, markers[i]);
4114 		multilist_sublist_unlock(mls);
4115 
4116 		kmem_cache_free(hdr_full_cache, markers[i]);
4117 	}
4118 	kmem_free(markers, sizeof (*markers) * num_sublists);
4119 
4120 	return (total_evicted);
4121 }
4122 
4123 /*
4124  * Flush all "evictable" data of the given type from the arc state
4125  * specified. This will not evict any "active" buffers (i.e. referenced).
4126  *
4127  * When 'retry' is set to B_FALSE, the function will make a single pass
4128  * over the state and evict any buffers that it can. Since it doesn't
4129  * continually retry the eviction, it might end up leaving some buffers
4130  * in the ARC due to lock misses.
4131  *
4132  * When 'retry' is set to B_TRUE, the function will continually retry the
4133  * eviction until *all* evictable buffers have been removed from the
4134  * state. As a result, if concurrent insertions into the state are
4135  * allowed (e.g. if the ARC isn't shutting down), this function might
4136  * wind up in an infinite loop, continually trying to evict buffers.
4137  */
4138 static uint64_t
arc_flush_state(arc_state_t * state,uint64_t spa,arc_buf_contents_t type,boolean_t retry)4139 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4140     boolean_t retry)
4141 {
4142 	uint64_t evicted = 0;
4143 
4144 	while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4145 		evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4146 
4147 		if (!retry)
4148 			break;
4149 	}
4150 
4151 	return (evicted);
4152 }
4153 
4154 /*
4155  * Evict the specified number of bytes from the state specified,
4156  * restricting eviction to the spa and type given. This function
4157  * prevents us from trying to evict more from a state's list than
4158  * is "evictable", and to skip evicting altogether when passed a
4159  * negative value for "bytes". In contrast, arc_evict_state() will
4160  * evict everything it can, when passed a negative value for "bytes".
4161  */
4162 static uint64_t
arc_adjust_impl(arc_state_t * state,uint64_t spa,int64_t bytes,arc_buf_contents_t type)4163 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4164     arc_buf_contents_t type)
4165 {
4166 	int64_t delta;
4167 
4168 	if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4169 		delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4170 		    bytes);
4171 		return (arc_evict_state(state, spa, delta, type));
4172 	}
4173 
4174 	return (0);
4175 }
4176 
4177 /*
4178  * Evict metadata buffers from the cache, such that arc_meta_used is
4179  * capped by the arc_meta_limit tunable.
4180  */
4181 static uint64_t
arc_adjust_meta(uint64_t meta_used)4182 arc_adjust_meta(uint64_t meta_used)
4183 {
4184 	uint64_t total_evicted = 0;
4185 	int64_t target;
4186 
4187 	/*
4188 	 * If we're over the meta limit, we want to evict enough
4189 	 * metadata to get back under the meta limit. We don't want to
4190 	 * evict so much that we drop the MRU below arc_p, though. If
4191 	 * we're over the meta limit more than we're over arc_p, we
4192 	 * evict some from the MRU here, and some from the MFU below.
4193 	 */
4194 	target = MIN((int64_t)(meta_used - arc_meta_limit),
4195 	    (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4196 	    zfs_refcount_count(&arc_mru->arcs_size) - arc_p));
4197 
4198 	total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4199 
4200 	/*
4201 	 * Similar to the above, we want to evict enough bytes to get us
4202 	 * below the meta limit, but not so much as to drop us below the
4203 	 * space allotted to the MFU (which is defined as arc_c - arc_p).
4204 	 */
4205 	target = MIN((int64_t)(meta_used - arc_meta_limit),
4206 	    (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
4207 	    (arc_c - arc_p)));
4208 
4209 	total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4210 
4211 	return (total_evicted);
4212 }
4213 
4214 /*
4215  * Return the type of the oldest buffer in the given arc state
4216  *
4217  * This function will select a random sublist of type ARC_BUFC_DATA and
4218  * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4219  * is compared, and the type which contains the "older" buffer will be
4220  * returned.
4221  */
4222 static arc_buf_contents_t
arc_adjust_type(arc_state_t * state)4223 arc_adjust_type(arc_state_t *state)
4224 {
4225 	multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4226 	multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4227 	int data_idx = multilist_get_random_index(data_ml);
4228 	int meta_idx = multilist_get_random_index(meta_ml);
4229 	multilist_sublist_t *data_mls;
4230 	multilist_sublist_t *meta_mls;
4231 	arc_buf_contents_t type;
4232 	arc_buf_hdr_t *data_hdr;
4233 	arc_buf_hdr_t *meta_hdr;
4234 
4235 	/*
4236 	 * We keep the sublist lock until we're finished, to prevent
4237 	 * the headers from being destroyed via arc_evict_state().
4238 	 */
4239 	data_mls = multilist_sublist_lock(data_ml, data_idx);
4240 	meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4241 
4242 	/*
4243 	 * These two loops are to ensure we skip any markers that
4244 	 * might be at the tail of the lists due to arc_evict_state().
4245 	 */
4246 
4247 	for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4248 	    data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4249 		if (data_hdr->b_spa != 0)
4250 			break;
4251 	}
4252 
4253 	for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4254 	    meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4255 		if (meta_hdr->b_spa != 0)
4256 			break;
4257 	}
4258 
4259 	if (data_hdr == NULL && meta_hdr == NULL) {
4260 		type = ARC_BUFC_DATA;
4261 	} else if (data_hdr == NULL) {
4262 		ASSERT3P(meta_hdr, !=, NULL);
4263 		type = ARC_BUFC_METADATA;
4264 	} else if (meta_hdr == NULL) {
4265 		ASSERT3P(data_hdr, !=, NULL);
4266 		type = ARC_BUFC_DATA;
4267 	} else {
4268 		ASSERT3P(data_hdr, !=, NULL);
4269 		ASSERT3P(meta_hdr, !=, NULL);
4270 
4271 		/* The headers can't be on the sublist without an L1 header */
4272 		ASSERT(HDR_HAS_L1HDR(data_hdr));
4273 		ASSERT(HDR_HAS_L1HDR(meta_hdr));
4274 
4275 		if (data_hdr->b_l1hdr.b_arc_access <
4276 		    meta_hdr->b_l1hdr.b_arc_access) {
4277 			type = ARC_BUFC_DATA;
4278 		} else {
4279 			type = ARC_BUFC_METADATA;
4280 		}
4281 	}
4282 
4283 	multilist_sublist_unlock(meta_mls);
4284 	multilist_sublist_unlock(data_mls);
4285 
4286 	return (type);
4287 }
4288 
4289 /*
4290  * Evict buffers from the cache, such that arc_size is capped by arc_c.
4291  */
4292 static uint64_t
arc_adjust(void)4293 arc_adjust(void)
4294 {
4295 	uint64_t total_evicted = 0;
4296 	uint64_t bytes;
4297 	int64_t target;
4298 	uint64_t asize = aggsum_value(&arc_size);
4299 	uint64_t ameta = aggsum_value(&arc_meta_used);
4300 
4301 	/*
4302 	 * If we're over arc_meta_limit, we want to correct that before
4303 	 * potentially evicting data buffers below.
4304 	 */
4305 	total_evicted += arc_adjust_meta(ameta);
4306 
4307 	/*
4308 	 * Adjust MRU size
4309 	 *
4310 	 * If we're over the target cache size, we want to evict enough
4311 	 * from the list to get back to our target size. We don't want
4312 	 * to evict too much from the MRU, such that it drops below
4313 	 * arc_p. So, if we're over our target cache size more than
4314 	 * the MRU is over arc_p, we'll evict enough to get back to
4315 	 * arc_p here, and then evict more from the MFU below.
4316 	 */
4317 	target = MIN((int64_t)(asize - arc_c),
4318 	    (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4319 	    zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4320 
4321 	/*
4322 	 * If we're below arc_meta_min, always prefer to evict data.
4323 	 * Otherwise, try to satisfy the requested number of bytes to
4324 	 * evict from the type which contains older buffers; in an
4325 	 * effort to keep newer buffers in the cache regardless of their
4326 	 * type. If we cannot satisfy the number of bytes from this
4327 	 * type, spill over into the next type.
4328 	 */
4329 	if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4330 	    ameta > arc_meta_min) {
4331 		bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4332 		total_evicted += bytes;
4333 
4334 		/*
4335 		 * If we couldn't evict our target number of bytes from
4336 		 * metadata, we try to get the rest from data.
4337 		 */
4338 		target -= bytes;
4339 
4340 		total_evicted +=
4341 		    arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4342 	} else {
4343 		bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4344 		total_evicted += bytes;
4345 
4346 		/*
4347 		 * If we couldn't evict our target number of bytes from
4348 		 * data, we try to get the rest from metadata.
4349 		 */
4350 		target -= bytes;
4351 
4352 		total_evicted +=
4353 		    arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4354 	}
4355 
4356 	/*
4357 	 * Adjust MFU size
4358 	 *
4359 	 * Now that we've tried to evict enough from the MRU to get its
4360 	 * size back to arc_p, if we're still above the target cache
4361 	 * size, we evict the rest from the MFU.
4362 	 */
4363 	target = asize - arc_c;
4364 
4365 	if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4366 	    ameta > arc_meta_min) {
4367 		bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4368 		total_evicted += bytes;
4369 
4370 		/*
4371 		 * If we couldn't evict our target number of bytes from
4372 		 * metadata, we try to get the rest from data.
4373 		 */
4374 		target -= bytes;
4375 
4376 		total_evicted +=
4377 		    arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4378 	} else {
4379 		bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4380 		total_evicted += bytes;
4381 
4382 		/*
4383 		 * If we couldn't evict our target number of bytes from
4384 		 * data, we try to get the rest from data.
4385 		 */
4386 		target -= bytes;
4387 
4388 		total_evicted +=
4389 		    arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4390 	}
4391 
4392 	/*
4393 	 * Adjust ghost lists
4394 	 *
4395 	 * In addition to the above, the ARC also defines target values
4396 	 * for the ghost lists. The sum of the mru list and mru ghost
4397 	 * list should never exceed the target size of the cache, and
4398 	 * the sum of the mru list, mfu list, mru ghost list, and mfu
4399 	 * ghost list should never exceed twice the target size of the
4400 	 * cache. The following logic enforces these limits on the ghost
4401 	 * caches, and evicts from them as needed.
4402 	 */
4403 	target = zfs_refcount_count(&arc_mru->arcs_size) +
4404 	    zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4405 
4406 	bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4407 	total_evicted += bytes;
4408 
4409 	target -= bytes;
4410 
4411 	total_evicted +=
4412 	    arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4413 
4414 	/*
4415 	 * We assume the sum of the mru list and mfu list is less than
4416 	 * or equal to arc_c (we enforced this above), which means we
4417 	 * can use the simpler of the two equations below:
4418 	 *
4419 	 *	mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4420 	 *		    mru ghost + mfu ghost <= arc_c
4421 	 */
4422 	target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
4423 	    zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4424 
4425 	bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4426 	total_evicted += bytes;
4427 
4428 	target -= bytes;
4429 
4430 	total_evicted +=
4431 	    arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4432 
4433 	return (total_evicted);
4434 }
4435 
4436 void
arc_flush(spa_t * spa,boolean_t retry)4437 arc_flush(spa_t *spa, boolean_t retry)
4438 {
4439 	uint64_t guid = 0;
4440 
4441 	/*
4442 	 * If retry is B_TRUE, a spa must not be specified since we have
4443 	 * no good way to determine if all of a spa's buffers have been
4444 	 * evicted from an arc state.
4445 	 */
4446 	ASSERT(!retry || spa == 0);
4447 
4448 	if (spa != NULL)
4449 		guid = spa_load_guid(spa);
4450 
4451 	(void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4452 	(void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4453 
4454 	(void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4455 	(void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4456 
4457 	(void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4458 	(void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4459 
4460 	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4461 	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4462 }
4463 
4464 static void
arc_reduce_target_size(int64_t to_free)4465 arc_reduce_target_size(int64_t to_free)
4466 {
4467 	uint64_t asize = aggsum_value(&arc_size);
4468 	if (arc_c > arc_c_min) {
4469 
4470 		if (arc_c > arc_c_min + to_free)
4471 			atomic_add_64(&arc_c, -to_free);
4472 		else
4473 			arc_c = arc_c_min;
4474 
4475 		atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4476 		if (asize < arc_c)
4477 			arc_c = MAX(asize, arc_c_min);
4478 		if (arc_p > arc_c)
4479 			arc_p = (arc_c >> 1);
4480 		ASSERT(arc_c >= arc_c_min);
4481 		ASSERT((int64_t)arc_p >= 0);
4482 	}
4483 
4484 	if (asize > arc_c) {
4485 		/* See comment in arc_adjust_cb_check() on why lock+flag */
4486 		mutex_enter(&arc_adjust_lock);
4487 		arc_adjust_needed = B_TRUE;
4488 		mutex_exit(&arc_adjust_lock);
4489 		zthr_wakeup(arc_adjust_zthr);
4490 	}
4491 }
4492 
4493 typedef enum free_memory_reason_t {
4494 	FMR_UNKNOWN,
4495 	FMR_NEEDFREE,
4496 	FMR_LOTSFREE,
4497 	FMR_SWAPFS_MINFREE,
4498 	FMR_PAGES_PP_MAXIMUM,
4499 	FMR_HEAP_ARENA,
4500 	FMR_ZIO_ARENA,
4501 } free_memory_reason_t;
4502 
4503 int64_t last_free_memory;
4504 free_memory_reason_t last_free_reason;
4505 
4506 /*
4507  * Additional reserve of pages for pp_reserve.
4508  */
4509 int64_t arc_pages_pp_reserve = 64;
4510 
4511 /*
4512  * Additional reserve of pages for swapfs.
4513  */
4514 int64_t arc_swapfs_reserve = 64;
4515 
4516 /*
4517  * Return the amount of memory that can be consumed before reclaim will be
4518  * needed.  Positive if there is sufficient free memory, negative indicates
4519  * the amount of memory that needs to be freed up.
4520  */
4521 static int64_t
arc_available_memory(void)4522 arc_available_memory(void)
4523 {
4524 	int64_t lowest = INT64_MAX;
4525 	int64_t n;
4526 	free_memory_reason_t r = FMR_UNKNOWN;
4527 
4528 #ifdef _KERNEL
4529 	if (needfree > 0) {
4530 		n = PAGESIZE * (-needfree);
4531 		if (n < lowest) {
4532 			lowest = n;
4533 			r = FMR_NEEDFREE;
4534 		}
4535 	}
4536 
4537 	/*
4538 	 * check that we're out of range of the pageout scanner.  It starts to
4539 	 * schedule paging if freemem is less than lotsfree and needfree.
4540 	 * lotsfree is the high-water mark for pageout, and needfree is the
4541 	 * number of needed free pages.  We add extra pages here to make sure
4542 	 * the scanner doesn't start up while we're freeing memory.
4543 	 */
4544 	n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4545 	if (n < lowest) {
4546 		lowest = n;
4547 		r = FMR_LOTSFREE;
4548 	}
4549 
4550 	/*
4551 	 * check to make sure that swapfs has enough space so that anon
4552 	 * reservations can still succeed. anon_resvmem() checks that the
4553 	 * availrmem is greater than swapfs_minfree, and the number of reserved
4554 	 * swap pages.  We also add a bit of extra here just to prevent
4555 	 * circumstances from getting really dire.
4556 	 */
4557 	n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4558 	    desfree - arc_swapfs_reserve);
4559 	if (n < lowest) {
4560 		lowest = n;
4561 		r = FMR_SWAPFS_MINFREE;
4562 	}
4563 
4564 
4565 	/*
4566 	 * Check that we have enough availrmem that memory locking (e.g., via
4567 	 * mlock(3C) or memcntl(2)) can still succeed.  (pages_pp_maximum
4568 	 * stores the number of pages that cannot be locked; when availrmem
4569 	 * drops below pages_pp_maximum, page locking mechanisms such as
4570 	 * page_pp_lock() will fail.)
4571 	 */
4572 	n = PAGESIZE * (availrmem - pages_pp_maximum -
4573 	    arc_pages_pp_reserve);
4574 	if (n < lowest) {
4575 		lowest = n;
4576 		r = FMR_PAGES_PP_MAXIMUM;
4577 	}
4578 
4579 
4580 	/*
4581 	 * If zio data pages are being allocated out of a separate heap segment,
4582 	 * then enforce that the size of available vmem for this arena remains
4583 	 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4584 	 *
4585 	 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4586 	 * memory (in the zio_arena) free, which can avoid memory
4587 	 * fragmentation issues.
4588 	 */
4589 	if (zio_arena != NULL) {
4590 		n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4591 		    (vmem_size(zio_arena, VMEM_ALLOC) >>
4592 		    arc_zio_arena_free_shift);
4593 		if (n < lowest) {
4594 			lowest = n;
4595 			r = FMR_ZIO_ARENA;
4596 		}
4597 	}
4598 #else
4599 	/* Every 100 calls, free a small amount */
4600 	if (spa_get_random(100) == 0)
4601 		lowest = -1024;
4602 #endif
4603 
4604 	last_free_memory = lowest;
4605 	last_free_reason = r;
4606 
4607 	return (lowest);
4608 }
4609 
4610 
4611 /*
4612  * Determine if the system is under memory pressure and is asking
4613  * to reclaim memory. A return value of B_TRUE indicates that the system
4614  * is under memory pressure and that the arc should adjust accordingly.
4615  */
4616 static boolean_t
arc_reclaim_needed(void)4617 arc_reclaim_needed(void)
4618 {
4619 	return (arc_available_memory() < 0);
4620 }
4621 
4622 static void
arc_kmem_reap_soon(void)4623 arc_kmem_reap_soon(void)
4624 {
4625 	size_t			i;
4626 	kmem_cache_t		*prev_cache = NULL;
4627 	kmem_cache_t		*prev_data_cache = NULL;
4628 	extern kmem_cache_t	*zio_buf_cache[];
4629 	extern kmem_cache_t	*zio_data_buf_cache[];
4630 	extern kmem_cache_t	*zfs_btree_leaf_cache;
4631 	extern kmem_cache_t	*abd_chunk_cache;
4632 
4633 #ifdef _KERNEL
4634 	if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4635 		/*
4636 		 * We are exceeding our meta-data cache limit.
4637 		 * Purge some DNLC entries to release holds on meta-data.
4638 		 */
4639 		dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4640 	}
4641 #endif
4642 
4643 	for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4644 		if (zio_buf_cache[i] != prev_cache) {
4645 			prev_cache = zio_buf_cache[i];
4646 			kmem_cache_reap_soon(zio_buf_cache[i]);
4647 		}
4648 		if (zio_data_buf_cache[i] != prev_data_cache) {
4649 			prev_data_cache = zio_data_buf_cache[i];
4650 			kmem_cache_reap_soon(zio_data_buf_cache[i]);
4651 		}
4652 	}
4653 	kmem_cache_reap_soon(abd_chunk_cache);
4654 	kmem_cache_reap_soon(buf_cache);
4655 	kmem_cache_reap_soon(hdr_full_cache);
4656 	kmem_cache_reap_soon(hdr_l2only_cache);
4657 	kmem_cache_reap_soon(zfs_btree_leaf_cache);
4658 
4659 	if (zio_arena != NULL) {
4660 		/*
4661 		 * Ask the vmem arena to reclaim unused memory from its
4662 		 * quantum caches.
4663 		 */
4664 		vmem_qcache_reap(zio_arena);
4665 	}
4666 }
4667 
4668 /* ARGSUSED */
4669 static boolean_t
arc_adjust_cb_check(void * arg,zthr_t * zthr)4670 arc_adjust_cb_check(void *arg, zthr_t *zthr)
4671 {
4672 	/*
4673 	 * This is necessary in order for the mdb ::arc dcmd to
4674 	 * show up to date information. Since the ::arc command
4675 	 * does not call the kstat's update function, without
4676 	 * this call, the command may show stale stats for the
4677 	 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4678 	 * with this change, the data might be up to 1 second
4679 	 * out of date(the arc_adjust_zthr has a maximum sleep
4680 	 * time of 1 second); but that should suffice.  The
4681 	 * arc_state_t structures can be queried directly if more
4682 	 * accurate information is needed.
4683 	 */
4684 	if (arc_ksp != NULL)
4685 		arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4686 
4687 	/*
4688 	 * We have to rely on arc_get_data_impl() to tell us when to adjust,
4689 	 * rather than checking if we are overflowing here, so that we are
4690 	 * sure to not leave arc_get_data_impl() waiting on
4691 	 * arc_adjust_waiters_cv.  If we have become "not overflowing" since
4692 	 * arc_get_data_impl() checked, we need to wake it up.  We could
4693 	 * broadcast the CV here, but arc_get_data_impl() may have not yet
4694 	 * gone to sleep.  We would need to use a mutex to ensure that this
4695 	 * function doesn't broadcast until arc_get_data_impl() has gone to
4696 	 * sleep (e.g. the arc_adjust_lock).  However, the lock ordering of
4697 	 * such a lock would necessarily be incorrect with respect to the
4698 	 * zthr_lock, which is held before this function is called, and is
4699 	 * held by arc_get_data_impl() when it calls zthr_wakeup().
4700 	 */
4701 	return (arc_adjust_needed);
4702 }
4703 
4704 /*
4705  * Keep arc_size under arc_c by running arc_adjust which evicts data
4706  * from the ARC.
4707  */
4708 /* ARGSUSED */
4709 static void
arc_adjust_cb(void * arg,zthr_t * zthr)4710 arc_adjust_cb(void *arg, zthr_t *zthr)
4711 {
4712 	uint64_t evicted = 0;
4713 
4714 	/* Evict from cache */
4715 	evicted = arc_adjust();
4716 
4717 	/*
4718 	 * If evicted is zero, we couldn't evict anything
4719 	 * via arc_adjust(). This could be due to hash lock
4720 	 * collisions, but more likely due to the majority of
4721 	 * arc buffers being unevictable. Therefore, even if
4722 	 * arc_size is above arc_c, another pass is unlikely to
4723 	 * be helpful and could potentially cause us to enter an
4724 	 * infinite loop.  Additionally, zthr_iscancelled() is
4725 	 * checked here so that if the arc is shutting down, the
4726 	 * broadcast will wake any remaining arc adjust waiters.
4727 	 */
4728 	mutex_enter(&arc_adjust_lock);
4729 	arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) &&
4730 	    evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0;
4731 	if (!arc_adjust_needed) {
4732 		/*
4733 		 * We're either no longer overflowing, or we
4734 		 * can't evict anything more, so we should wake
4735 		 * up any waiters.
4736 		 */
4737 		cv_broadcast(&arc_adjust_waiters_cv);
4738 	}
4739 	mutex_exit(&arc_adjust_lock);
4740 }
4741 
4742 /* ARGSUSED */
4743 static boolean_t
arc_reap_cb_check(void * arg,zthr_t * zthr)4744 arc_reap_cb_check(void *arg, zthr_t *zthr)
4745 {
4746 	int64_t free_memory = arc_available_memory();
4747 
4748 	/*
4749 	 * If a kmem reap is already active, don't schedule more.  We must
4750 	 * check for this because kmem_cache_reap_soon() won't actually
4751 	 * block on the cache being reaped (this is to prevent callers from
4752 	 * becoming implicitly blocked by a system-wide kmem reap -- which,
4753 	 * on a system with many, many full magazines, can take minutes).
4754 	 */
4755 	if (!kmem_cache_reap_active() &&
4756 	    free_memory < 0) {
4757 		arc_no_grow = B_TRUE;
4758 		arc_warm = B_TRUE;
4759 		/*
4760 		 * Wait at least zfs_grow_retry (default 60) seconds
4761 		 * before considering growing.
4762 		 */
4763 		arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4764 		return (B_TRUE);
4765 	} else if (free_memory < arc_c >> arc_no_grow_shift) {
4766 		arc_no_grow = B_TRUE;
4767 	} else if (gethrtime() >= arc_growtime) {
4768 		arc_no_grow = B_FALSE;
4769 	}
4770 
4771 	return (B_FALSE);
4772 }
4773 
4774 /*
4775  * Keep enough free memory in the system by reaping the ARC's kmem
4776  * caches.  To cause more slabs to be reapable, we may reduce the
4777  * target size of the cache (arc_c), causing the arc_adjust_cb()
4778  * to free more buffers.
4779  */
4780 /* ARGSUSED */
4781 static void
arc_reap_cb(void * arg,zthr_t * zthr)4782 arc_reap_cb(void *arg, zthr_t *zthr)
4783 {
4784 	int64_t free_memory;
4785 
4786 	/*
4787 	 * Kick off asynchronous kmem_reap()'s of all our caches.
4788 	 */
4789 	arc_kmem_reap_soon();
4790 
4791 	/*
4792 	 * Wait at least arc_kmem_cache_reap_retry_ms between
4793 	 * arc_kmem_reap_soon() calls. Without this check it is possible to
4794 	 * end up in a situation where we spend lots of time reaping
4795 	 * caches, while we're near arc_c_min.  Waiting here also gives the
4796 	 * subsequent free memory check a chance of finding that the
4797 	 * asynchronous reap has already freed enough memory, and we don't
4798 	 * need to call arc_reduce_target_size().
4799 	 */
4800 	delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
4801 
4802 	/*
4803 	 * Reduce the target size as needed to maintain the amount of free
4804 	 * memory in the system at a fraction of the arc_size (1/128th by
4805 	 * default).  If oversubscribed (free_memory < 0) then reduce the
4806 	 * target arc_size by the deficit amount plus the fractional
4807 	 * amount.  If free memory is positive but less then the fractional
4808 	 * amount, reduce by what is needed to hit the fractional amount.
4809 	 */
4810 	free_memory = arc_available_memory();
4811 
4812 	int64_t to_free =
4813 	    (arc_c >> arc_shrink_shift) - free_memory;
4814 	if (to_free > 0) {
4815 #ifdef _KERNEL
4816 		to_free = MAX(to_free, ptob(needfree));
4817 #endif
4818 		arc_reduce_target_size(to_free);
4819 	}
4820 }
4821 
4822 /*
4823  * Adapt arc info given the number of bytes we are trying to add and
4824  * the state that we are coming from.  This function is only called
4825  * when we are adding new content to the cache.
4826  */
4827 static void
arc_adapt(int bytes,arc_state_t * state)4828 arc_adapt(int bytes, arc_state_t *state)
4829 {
4830 	int mult;
4831 	uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4832 	int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
4833 	int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);
4834 
4835 	ASSERT(bytes > 0);
4836 	/*
4837 	 * Adapt the target size of the MRU list:
4838 	 *	- if we just hit in the MRU ghost list, then increase
4839 	 *	  the target size of the MRU list.
4840 	 *	- if we just hit in the MFU ghost list, then increase
4841 	 *	  the target size of the MFU list by decreasing the
4842 	 *	  target size of the MRU list.
4843 	 */
4844 	if (state == arc_mru_ghost) {
4845 		mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4846 		mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4847 
4848 		arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4849 	} else if (state == arc_mfu_ghost) {
4850 		uint64_t delta;
4851 
4852 		mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4853 		mult = MIN(mult, 10);
4854 
4855 		delta = MIN(bytes * mult, arc_p);
4856 		arc_p = MAX(arc_p_min, arc_p - delta);
4857 	}
4858 	ASSERT((int64_t)arc_p >= 0);
4859 
4860 	/*
4861 	 * Wake reap thread if we do not have any available memory
4862 	 */
4863 	if (arc_reclaim_needed()) {
4864 		zthr_wakeup(arc_reap_zthr);
4865 		return;
4866 	}
4867 
4868 
4869 	if (arc_no_grow)
4870 		return;
4871 
4872 	if (arc_c >= arc_c_max)
4873 		return;
4874 
4875 	/*
4876 	 * If we're within (2 * maxblocksize) bytes of the target
4877 	 * cache size, increment the target cache size
4878 	 */
4879 	if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
4880 	    0) {
4881 		atomic_add_64(&arc_c, (int64_t)bytes);
4882 		if (arc_c > arc_c_max)
4883 			arc_c = arc_c_max;
4884 		else if (state == arc_anon)
4885 			atomic_add_64(&arc_p, (int64_t)bytes);
4886 		if (arc_p > arc_c)
4887 			arc_p = arc_c;
4888 	}
4889 	ASSERT((int64_t)arc_p >= 0);
4890 }
4891 
4892 /*
4893  * Check if arc_size has grown past our upper threshold, determined by
4894  * zfs_arc_overflow_shift.
4895  */
4896 static boolean_t
arc_is_overflowing(void)4897 arc_is_overflowing(void)
4898 {
4899 	/* Always allow at least one block of overflow */
4900 	uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4901 	    arc_c >> zfs_arc_overflow_shift);
4902 
4903 	/*
4904 	 * We just compare the lower bound here for performance reasons. Our
4905 	 * primary goals are to make sure that the arc never grows without
4906 	 * bound, and that it can reach its maximum size. This check
4907 	 * accomplishes both goals. The maximum amount we could run over by is
4908 	 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4909 	 * in the ARC. In practice, that's in the tens of MB, which is low
4910 	 * enough to be safe.
4911 	 */
4912 	return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
4913 }
4914 
4915 static abd_t *
arc_get_data_abd(arc_buf_hdr_t * hdr,uint64_t size,void * tag,boolean_t do_adapt)4916 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag,
4917     boolean_t do_adapt)
4918 {
4919 	arc_buf_contents_t type = arc_buf_type(hdr);
4920 
4921 	arc_get_data_impl(hdr, size, tag, do_adapt);
4922 	if (type == ARC_BUFC_METADATA) {
4923 		return (abd_alloc(size, B_TRUE));
4924 	} else {
4925 		ASSERT(type == ARC_BUFC_DATA);
4926 		return (abd_alloc(size, B_FALSE));
4927 	}
4928 }
4929 
4930 static void *
arc_get_data_buf(arc_buf_hdr_t * hdr,uint64_t size,void * tag)4931 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4932 {
4933 	arc_buf_contents_t type = arc_buf_type(hdr);
4934 
4935 	arc_get_data_impl(hdr, size, tag, B_TRUE);
4936 	if (type == ARC_BUFC_METADATA) {
4937 		return (zio_buf_alloc(size));
4938 	} else {
4939 		ASSERT(type == ARC_BUFC_DATA);
4940 		return (zio_data_buf_alloc(size));
4941 	}
4942 }
4943 
4944 /*
4945  * Allocate a block and return it to the caller. If we are hitting the
4946  * hard limit for the cache size, we must sleep, waiting for the eviction
4947  * thread to catch up. If we're past the target size but below the hard
4948  * limit, we'll only signal the reclaim thread and continue on.
4949  */
4950 static void
arc_get_data_impl(arc_buf_hdr_t * hdr,uint64_t size,void * tag,boolean_t do_adapt)4951 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag,
4952     boolean_t do_adapt)
4953 {
4954 	arc_state_t *state = hdr->b_l1hdr.b_state;
4955 	arc_buf_contents_t type = arc_buf_type(hdr);
4956 
4957 	if (do_adapt)
4958 		arc_adapt(size, state);
4959 
4960 	/*
4961 	 * If arc_size is currently overflowing, and has grown past our
4962 	 * upper limit, we must be adding data faster than the evict
4963 	 * thread can evict. Thus, to ensure we don't compound the
4964 	 * problem by adding more data and forcing arc_size to grow even
4965 	 * further past its target size, we halt and wait for the
4966 	 * eviction thread to catch up.
4967 	 *
4968 	 * It's also possible that the reclaim thread is unable to evict
4969 	 * enough buffers to get arc_size below the overflow limit (e.g.
4970 	 * due to buffers being un-evictable, or hash lock collisions).
4971 	 * In this case, we want to proceed regardless if we're
4972 	 * overflowing; thus we don't use a while loop here.
4973 	 */
4974 	if (arc_is_overflowing()) {
4975 		mutex_enter(&arc_adjust_lock);
4976 
4977 		/*
4978 		 * Now that we've acquired the lock, we may no longer be
4979 		 * over the overflow limit, lets check.
4980 		 *
4981 		 * We're ignoring the case of spurious wake ups. If that
4982 		 * were to happen, it'd let this thread consume an ARC
4983 		 * buffer before it should have (i.e. before we're under
4984 		 * the overflow limit and were signalled by the reclaim
4985 		 * thread). As long as that is a rare occurrence, it
4986 		 * shouldn't cause any harm.
4987 		 */
4988 		if (arc_is_overflowing()) {
4989 			arc_adjust_needed = B_TRUE;
4990 			zthr_wakeup(arc_adjust_zthr);
4991 			(void) cv_wait(&arc_adjust_waiters_cv,
4992 			    &arc_adjust_lock);
4993 		}
4994 		mutex_exit(&arc_adjust_lock);
4995 	}
4996 
4997 	VERIFY3U(hdr->b_type, ==, type);
4998 	if (type == ARC_BUFC_METADATA) {
4999 		arc_space_consume(size, ARC_SPACE_META);
5000 	} else {
5001 		arc_space_consume(size, ARC_SPACE_DATA);
5002 	}
5003 
5004 	/*
5005 	 * Update the state size.  Note that ghost states have a
5006 	 * "ghost size" and so don't need to be updated.
5007 	 */
5008 	if (!GHOST_STATE(state)) {
5009 
5010 		(void) zfs_refcount_add_many(&state->arcs_size, size, tag);
5011 
5012 		/*
5013 		 * If this is reached via arc_read, the link is
5014 		 * protected by the hash lock. If reached via
5015 		 * arc_buf_alloc, the header should not be accessed by
5016 		 * any other thread. And, if reached via arc_read_done,
5017 		 * the hash lock will protect it if it's found in the
5018 		 * hash table; otherwise no other thread should be
5019 		 * trying to [add|remove]_reference it.
5020 		 */
5021 		if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5022 			ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5023 			(void) zfs_refcount_add_many(&state->arcs_esize[type],
5024 			    size, tag);
5025 		}
5026 
5027 		/*
5028 		 * If we are growing the cache, and we are adding anonymous
5029 		 * data, and we have outgrown arc_p, update arc_p
5030 		 */
5031 		if (aggsum_compare(&arc_size, arc_c) < 0 &&
5032 		    hdr->b_l1hdr.b_state == arc_anon &&
5033 		    (zfs_refcount_count(&arc_anon->arcs_size) +
5034 		    zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
5035 			arc_p = MIN(arc_c, arc_p + size);
5036 	}
5037 }
5038 
5039 static void
arc_free_data_abd(arc_buf_hdr_t * hdr,abd_t * abd,uint64_t size,void * tag)5040 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5041 {
5042 	arc_free_data_impl(hdr, size, tag);
5043 	abd_free(abd);
5044 }
5045 
5046 static void
arc_free_data_buf(arc_buf_hdr_t * hdr,void * buf,uint64_t size,void * tag)5047 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5048 {
5049 	arc_buf_contents_t type = arc_buf_type(hdr);
5050 
5051 	arc_free_data_impl(hdr, size, tag);
5052 	if (type == ARC_BUFC_METADATA) {
5053 		zio_buf_free(buf, size);
5054 	} else {
5055 		ASSERT(type == ARC_BUFC_DATA);
5056 		zio_data_buf_free(buf, size);
5057 	}
5058 }
5059 
5060 /*
5061  * Free the arc data buffer.
5062  */
5063 static void
arc_free_data_impl(arc_buf_hdr_t * hdr,uint64_t size,void * tag)5064 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5065 {
5066 	arc_state_t *state = hdr->b_l1hdr.b_state;
5067 	arc_buf_contents_t type = arc_buf_type(hdr);
5068 
5069 	/* protected by hash lock, if in the hash table */
5070 	if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5071 		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5072 		ASSERT(state != arc_anon && state != arc_l2c_only);
5073 
5074 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
5075 		    size, tag);
5076 	}
5077 	(void) zfs_refcount_remove_many(&state->arcs_size, size, tag);
5078 
5079 	VERIFY3U(hdr->b_type, ==, type);
5080 	if (type == ARC_BUFC_METADATA) {
5081 		arc_space_return(size, ARC_SPACE_META);
5082 	} else {
5083 		ASSERT(type == ARC_BUFC_DATA);
5084 		arc_space_return(size, ARC_SPACE_DATA);
5085 	}
5086 }
5087 
5088 /*
5089  * This routine is called whenever a buffer is accessed.
5090  * NOTE: the hash lock is dropped in this function.
5091  */
5092 static void
arc_access(arc_buf_hdr_t * hdr,kmutex_t * hash_lock)5093 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5094 {
5095 	clock_t now;
5096 
5097 	ASSERT(MUTEX_HELD(hash_lock));
5098 	ASSERT(HDR_HAS_L1HDR(hdr));
5099 
5100 	if (hdr->b_l1hdr.b_state == arc_anon) {
5101 		/*
5102 		 * This buffer is not in the cache, and does not
5103 		 * appear in our "ghost" list.  Add the new buffer
5104 		 * to the MRU state.
5105 		 */
5106 
5107 		ASSERT0(hdr->b_l1hdr.b_arc_access);
5108 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5109 		DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5110 		arc_change_state(arc_mru, hdr, hash_lock);
5111 
5112 	} else if (hdr->b_l1hdr.b_state == arc_mru) {
5113 		now = ddi_get_lbolt();
5114 
5115 		/*
5116 		 * If this buffer is here because of a prefetch, then either:
5117 		 * - clear the flag if this is a "referencing" read
5118 		 *   (any subsequent access will bump this into the MFU state).
5119 		 * or
5120 		 * - move the buffer to the head of the list if this is
5121 		 *   another prefetch (to make it less likely to be evicted).
5122 		 */
5123 		if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5124 			if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5125 				/* link protected by hash lock */
5126 				ASSERT(multilist_link_active(
5127 				    &hdr->b_l1hdr.b_arc_node));
5128 			} else {
5129 				if (HDR_HAS_L2HDR(hdr))
5130 					l2arc_hdr_arcstats_decrement_state(hdr);
5131 				arc_hdr_clear_flags(hdr,
5132 				    ARC_FLAG_PREFETCH |
5133 				    ARC_FLAG_PRESCIENT_PREFETCH);
5134 				ARCSTAT_BUMP(arcstat_mru_hits);
5135 				if (HDR_HAS_L2HDR(hdr))
5136 					l2arc_hdr_arcstats_increment_state(hdr);
5137 			}
5138 			hdr->b_l1hdr.b_arc_access = now;
5139 			return;
5140 		}
5141 
5142 		/*
5143 		 * This buffer has been "accessed" only once so far,
5144 		 * but it is still in the cache. Move it to the MFU
5145 		 * state.
5146 		 */
5147 		if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
5148 			/*
5149 			 * More than 125ms have passed since we
5150 			 * instantiated this buffer.  Move it to the
5151 			 * most frequently used state.
5152 			 */
5153 			hdr->b_l1hdr.b_arc_access = now;
5154 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5155 			arc_change_state(arc_mfu, hdr, hash_lock);
5156 		}
5157 		ARCSTAT_BUMP(arcstat_mru_hits);
5158 	} else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5159 		arc_state_t	*new_state;
5160 		/*
5161 		 * This buffer has been "accessed" recently, but
5162 		 * was evicted from the cache.  Move it to the
5163 		 * MFU state.
5164 		 */
5165 		if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5166 			new_state = arc_mru;
5167 			if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5168 				if (HDR_HAS_L2HDR(hdr))
5169 					l2arc_hdr_arcstats_decrement_state(hdr);
5170 				arc_hdr_clear_flags(hdr,
5171 				    ARC_FLAG_PREFETCH |
5172 				    ARC_FLAG_PRESCIENT_PREFETCH);
5173 				if (HDR_HAS_L2HDR(hdr))
5174 					l2arc_hdr_arcstats_increment_state(hdr);
5175 			}
5176 			DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5177 		} else {
5178 			new_state = arc_mfu;
5179 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5180 		}
5181 
5182 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5183 		arc_change_state(new_state, hdr, hash_lock);
5184 
5185 		ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5186 	} else if (hdr->b_l1hdr.b_state == arc_mfu) {
5187 		/*
5188 		 * This buffer has been accessed more than once and is
5189 		 * still in the cache.  Keep it in the MFU state.
5190 		 *
5191 		 * NOTE: an add_reference() that occurred when we did
5192 		 * the arc_read() will have kicked this off the list.
5193 		 * If it was a prefetch, we will explicitly move it to
5194 		 * the head of the list now.
5195 		 */
5196 		ARCSTAT_BUMP(arcstat_mfu_hits);
5197 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5198 	} else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5199 		arc_state_t	*new_state = arc_mfu;
5200 		/*
5201 		 * This buffer has been accessed more than once but has
5202 		 * been evicted from the cache.  Move it back to the
5203 		 * MFU state.
5204 		 */
5205 
5206 		if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5207 			/*
5208 			 * This is a prefetch access...
5209 			 * move this block back to the MRU state.
5210 			 */
5211 			new_state = arc_mru;
5212 		}
5213 
5214 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5215 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5216 		arc_change_state(new_state, hdr, hash_lock);
5217 
5218 		ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5219 	} else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5220 		/*
5221 		 * This buffer is on the 2nd Level ARC.
5222 		 */
5223 
5224 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5225 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5226 		arc_change_state(arc_mfu, hdr, hash_lock);
5227 	} else {
5228 		ASSERT(!"invalid arc state");
5229 	}
5230 }
5231 
5232 /*
5233  * This routine is called by dbuf_hold() to update the arc_access() state
5234  * which otherwise would be skipped for entries in the dbuf cache.
5235  */
5236 void
arc_buf_access(arc_buf_t * buf)5237 arc_buf_access(arc_buf_t *buf)
5238 {
5239 	mutex_enter(&buf->b_evict_lock);
5240 	arc_buf_hdr_t *hdr = buf->b_hdr;
5241 
5242 	/*
5243 	 * Avoid taking the hash_lock when possible as an optimization.
5244 	 * The header must be checked again under the hash_lock in order
5245 	 * to handle the case where it is concurrently being released.
5246 	 */
5247 	if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5248 		mutex_exit(&buf->b_evict_lock);
5249 		return;
5250 	}
5251 
5252 	kmutex_t *hash_lock = HDR_LOCK(hdr);
5253 	mutex_enter(hash_lock);
5254 
5255 	if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5256 		mutex_exit(hash_lock);
5257 		mutex_exit(&buf->b_evict_lock);
5258 		ARCSTAT_BUMP(arcstat_access_skip);
5259 		return;
5260 	}
5261 
5262 	mutex_exit(&buf->b_evict_lock);
5263 
5264 	ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5265 	    hdr->b_l1hdr.b_state == arc_mfu);
5266 
5267 	DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5268 	arc_access(hdr, hash_lock);
5269 	mutex_exit(hash_lock);
5270 
5271 	ARCSTAT_BUMP(arcstat_hits);
5272 	ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5273 	    demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5274 }
5275 
5276 /* a generic arc_read_done_func_t which you can use */
5277 /* ARGSUSED */
5278 void
arc_bcopy_func(zio_t * zio,const zbookmark_phys_t * zb,const blkptr_t * bp,arc_buf_t * buf,void * arg)5279 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5280     arc_buf_t *buf, void *arg)
5281 {
5282 	if (buf == NULL)
5283 		return;
5284 
5285 	bcopy(buf->b_data, arg, arc_buf_size(buf));
5286 	arc_buf_destroy(buf, arg);
5287 }
5288 
5289 /* a generic arc_read_done_func_t */
5290 void
arc_getbuf_func(zio_t * zio,const zbookmark_phys_t * zb,const blkptr_t * bp,arc_buf_t * buf,void * arg)5291 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5292     arc_buf_t *buf, void *arg)
5293 {
5294 	arc_buf_t **bufp = arg;
5295 
5296 	if (buf == NULL) {
5297 		ASSERT(zio == NULL || zio->io_error != 0);
5298 		*bufp = NULL;
5299 	} else {
5300 		ASSERT(zio == NULL || zio->io_error == 0);
5301 		*bufp = buf;
5302 		ASSERT(buf->b_data != NULL);
5303 	}
5304 }
5305 
5306 static void
arc_hdr_verify(arc_buf_hdr_t * hdr,const blkptr_t * bp)5307 arc_hdr_verify(arc_buf_hdr_t *hdr, const blkptr_t *bp)
5308 {
5309 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5310 		ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5311 		ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5312 	} else {
5313 		if (HDR_COMPRESSION_ENABLED(hdr)) {
5314 			ASSERT3U(arc_hdr_get_compress(hdr), ==,
5315 			    BP_GET_COMPRESS(bp));
5316 		}
5317 		ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5318 		ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5319 		ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5320 	}
5321 }
5322 
5323 /*
5324  * XXX this should be changed to return an error, and callers
5325  * re-read from disk on failure (on nondebug bits).
5326  */
5327 static void
arc_hdr_verify_checksum(spa_t * spa,arc_buf_hdr_t * hdr,const blkptr_t * bp)5328 arc_hdr_verify_checksum(spa_t *spa, arc_buf_hdr_t *hdr, const blkptr_t *bp)
5329 {
5330 	arc_hdr_verify(hdr, bp);
5331 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
5332 		return;
5333 	int err = 0;
5334 	abd_t *abd = NULL;
5335 	if (BP_IS_ENCRYPTED(bp)) {
5336 		if (HDR_HAS_RABD(hdr)) {
5337 			abd = hdr->b_crypt_hdr.b_rabd;
5338 		}
5339 	} else if (HDR_COMPRESSION_ENABLED(hdr)) {
5340 		abd = hdr->b_l1hdr.b_pabd;
5341 	}
5342 	if (abd != NULL) {
5343 		/*
5344 		 * The offset is only used for labels, which are not
5345 		 * cached in the ARC, so it doesn't matter what we
5346 		 * pass for the offset parameter.
5347 		 */
5348 		int psize = HDR_GET_PSIZE(hdr);
5349 		err = zio_checksum_error_impl(spa, bp,
5350 		    BP_GET_CHECKSUM(bp), abd, psize, 0, NULL);
5351 		if (err != 0) {
5352 			/*
5353 			 * Use abd_copy_to_buf() rather than
5354 			 * abd_borrow_buf_copy() so that we are sure to
5355 			 * include the buf in crash dumps.
5356 			 */
5357 			void *buf = kmem_alloc(psize, KM_SLEEP);
5358 			abd_copy_to_buf(buf, abd, psize);
5359 			panic("checksum of cached data doesn't match BP "
5360 			    "err=%u hdr=%p bp=%p abd=%p buf=%p",
5361 			    err, (void *)hdr, (void *)bp, (void *)abd, buf);
5362 		}
5363 	}
5364 }
5365 
5366 static void
arc_read_done(zio_t * zio)5367 arc_read_done(zio_t *zio)
5368 {
5369 	blkptr_t	*bp = zio->io_bp;
5370 	arc_buf_hdr_t	*hdr = zio->io_private;
5371 	kmutex_t	*hash_lock = NULL;
5372 	arc_callback_t	*callback_list;
5373 	arc_callback_t	*acb;
5374 	boolean_t	freeable = B_FALSE;
5375 
5376 	/*
5377 	 * The hdr was inserted into hash-table and removed from lists
5378 	 * prior to starting I/O.  We should find this header, since
5379 	 * it's in the hash table, and it should be legit since it's
5380 	 * not possible to evict it during the I/O.  The only possible
5381 	 * reason for it not to be found is if we were freed during the
5382 	 * read.
5383 	 */
5384 	if (HDR_IN_HASH_TABLE(hdr)) {
5385 		ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5386 		ASSERT3U(hdr->b_dva.dva_word[0], ==,
5387 		    BP_IDENTITY(zio->io_bp)->dva_word[0]);
5388 		ASSERT3U(hdr->b_dva.dva_word[1], ==,
5389 		    BP_IDENTITY(zio->io_bp)->dva_word[1]);
5390 
5391 		arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5392 		    &hash_lock);
5393 
5394 		ASSERT((found == hdr &&
5395 		    DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5396 		    (found == hdr && HDR_L2_READING(hdr)));
5397 		ASSERT3P(hash_lock, !=, NULL);
5398 	}
5399 
5400 	if (BP_IS_PROTECTED(bp)) {
5401 		hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5402 		hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5403 		zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5404 		    hdr->b_crypt_hdr.b_iv);
5405 
5406 		if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5407 			void *tmpbuf;
5408 
5409 			tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5410 			    sizeof (zil_chain_t));
5411 			zio_crypt_decode_mac_zil(tmpbuf,
5412 			    hdr->b_crypt_hdr.b_mac);
5413 			abd_return_buf(zio->io_abd, tmpbuf,
5414 			    sizeof (zil_chain_t));
5415 		} else {
5416 			zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
5417 		}
5418 	}
5419 
5420 	if (zio->io_error == 0) {
5421 		/* byteswap if necessary */
5422 		if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5423 			if (BP_GET_LEVEL(zio->io_bp) > 0) {
5424 				hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5425 			} else {
5426 				hdr->b_l1hdr.b_byteswap =
5427 				    DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5428 			}
5429 		} else {
5430 			hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5431 		}
5432 	}
5433 
5434 	arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5435 
5436 	callback_list = hdr->b_l1hdr.b_acb;
5437 	ASSERT3P(callback_list, !=, NULL);
5438 
5439 	if (hash_lock && zio->io_error == 0 &&
5440 	    hdr->b_l1hdr.b_state == arc_anon) {
5441 		/*
5442 		 * Only call arc_access on anonymous buffers.  This is because
5443 		 * if we've issued an I/O for an evicted buffer, we've already
5444 		 * called arc_access (to prevent any simultaneous readers from
5445 		 * getting confused).
5446 		 */
5447 		arc_access(hdr, hash_lock);
5448 	}
5449 
5450 	/*
5451 	 * If a read request has a callback (i.e. acb_done is not NULL), then we
5452 	 * make a buf containing the data according to the parameters which were
5453 	 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5454 	 * aren't needlessly decompressing the data multiple times.
5455 	 */
5456 	int callback_cnt = 0;
5457 	for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5458 		if (!acb->acb_done)
5459 			continue;
5460 
5461 		callback_cnt++;
5462 
5463 		if (zio->io_error != 0)
5464 			continue;
5465 
5466 		int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5467 		    &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5468 		    acb->acb_compressed, acb->acb_noauth, B_TRUE,
5469 		    &acb->acb_buf);
5470 
5471 		/*
5472 		 * Assert non-speculative zios didn't fail because an
5473 		 * encryption key wasn't loaded
5474 		 */
5475 		ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5476 		    error != EACCES);
5477 
5478 		/*
5479 		 * If we failed to decrypt, report an error now (as the zio
5480 		 * layer would have done if it had done the transforms).
5481 		 */
5482 		if (error == ECKSUM) {
5483 			ASSERT(BP_IS_PROTECTED(bp));
5484 			error = SET_ERROR(EIO);
5485 			if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5486 				spa_log_error(zio->io_spa, &acb->acb_zb);
5487 				(void) zfs_ereport_post(
5488 				    FM_EREPORT_ZFS_AUTHENTICATION,
5489 				    zio->io_spa, NULL, &acb->acb_zb, zio, 0, 0);
5490 			}
5491 		}
5492 
5493 		if (error != 0) {
5494 			/*
5495 			 * Decompression failed.  Set io_error
5496 			 * so that when we call acb_done (below),
5497 			 * we will indicate that the read failed.
5498 			 * Note that in the unusual case where one
5499 			 * callback is compressed and another
5500 			 * uncompressed, we will mark all of them
5501 			 * as failed, even though the uncompressed
5502 			 * one can't actually fail.  In this case,
5503 			 * the hdr will not be anonymous, because
5504 			 * if there are multiple callbacks, it's
5505 			 * because multiple threads found the same
5506 			 * arc buf in the hash table.
5507 			 */
5508 			zio->io_error = error;
5509 		}
5510 	}
5511 
5512 	/*
5513 	 * If there are multiple callbacks, we must have the hash lock,
5514 	 * because the only way for multiple threads to find this hdr is
5515 	 * in the hash table.  This ensures that if there are multiple
5516 	 * callbacks, the hdr is not anonymous.  If it were anonymous,
5517 	 * we couldn't use arc_buf_destroy() in the error case below.
5518 	 */
5519 	ASSERT(callback_cnt < 2 || hash_lock != NULL);
5520 
5521 	hdr->b_l1hdr.b_acb = NULL;
5522 	arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5523 	if (callback_cnt == 0)
5524 		ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
5525 
5526 	ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5527 	    callback_list != NULL);
5528 
5529 	if (zio->io_error == 0) {
5530 		arc_hdr_verify(hdr, zio->io_bp);
5531 	} else {
5532 		arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5533 		if (hdr->b_l1hdr.b_state != arc_anon)
5534 			arc_change_state(arc_anon, hdr, hash_lock);
5535 		if (HDR_IN_HASH_TABLE(hdr))
5536 			buf_hash_remove(hdr);
5537 		freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5538 	}
5539 
5540 	/*
5541 	 * Broadcast before we drop the hash_lock to avoid the possibility
5542 	 * that the hdr (and hence the cv) might be freed before we get to
5543 	 * the cv_broadcast().
5544 	 */
5545 	cv_broadcast(&hdr->b_l1hdr.b_cv);
5546 
5547 	if (hash_lock != NULL) {
5548 		mutex_exit(hash_lock);
5549 	} else {
5550 		/*
5551 		 * This block was freed while we waited for the read to
5552 		 * complete.  It has been removed from the hash table and
5553 		 * moved to the anonymous state (so that it won't show up
5554 		 * in the cache).
5555 		 */
5556 		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5557 		freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5558 	}
5559 
5560 	/* execute each callback and free its structure */
5561 	while ((acb = callback_list) != NULL) {
5562 
5563 		if (acb->acb_done != NULL) {
5564 			if (zio->io_error != 0 && acb->acb_buf != NULL) {
5565 				/*
5566 				 * If arc_buf_alloc_impl() fails during
5567 				 * decompression, the buf will still be
5568 				 * allocated, and needs to be freed here.
5569 				 */
5570 				arc_buf_destroy(acb->acb_buf, acb->acb_private);
5571 				acb->acb_buf = NULL;
5572 			}
5573 			acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5574 			    acb->acb_buf, acb->acb_private);
5575 		}
5576 
5577 		if (acb->acb_zio_dummy != NULL) {
5578 			acb->acb_zio_dummy->io_error = zio->io_error;
5579 			zio_nowait(acb->acb_zio_dummy);
5580 		}
5581 
5582 		callback_list = acb->acb_next;
5583 		kmem_free(acb, sizeof (arc_callback_t));
5584 	}
5585 
5586 	if (freeable)
5587 		arc_hdr_destroy(hdr);
5588 }
5589 
5590 /*
5591  * "Read" the block at the specified DVA (in bp) via the
5592  * cache.  If the block is found in the cache, invoke the provided
5593  * callback immediately and return.  Note that the `zio' parameter
5594  * in the callback will be NULL in this case, since no IO was
5595  * required.  If the block is not in the cache pass the read request
5596  * on to the spa with a substitute callback function, so that the
5597  * requested block will be added to the cache.
5598  *
5599  * If a read request arrives for a block that has a read in-progress,
5600  * either wait for the in-progress read to complete (and return the
5601  * results); or, if this is a read with a "done" func, add a record
5602  * to the read to invoke the "done" func when the read completes,
5603  * and return; or just return.
5604  *
5605  * arc_read_done() will invoke all the requested "done" functions
5606  * for readers of this block.
5607  */
5608 int
arc_read(zio_t * pio,spa_t * spa,const blkptr_t * bp,arc_read_done_func_t * done,void * private,zio_priority_t priority,int zio_flags,arc_flags_t * arc_flags,const zbookmark_phys_t * zb)5609 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done,
5610     void *private, zio_priority_t priority, int zio_flags,
5611     arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5612 {
5613 	arc_buf_hdr_t *hdr = NULL;
5614 	kmutex_t *hash_lock = NULL;
5615 	zio_t *rzio;
5616 	uint64_t guid = spa_load_guid(spa);
5617 	boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5618 	boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5619 	    (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5620 	boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5621 	    (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5622 	int rc = 0;
5623 
5624 	ASSERT(!BP_IS_EMBEDDED(bp) ||
5625 	    BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5626 
5627 top:
5628 	if (!BP_IS_EMBEDDED(bp)) {
5629 		/*
5630 		 * Embedded BP's have no DVA and require no I/O to "read".
5631 		 * Create an anonymous arc buf to back it.
5632 		 */
5633 		hdr = buf_hash_find(guid, bp, &hash_lock);
5634 	}
5635 
5636 	/*
5637 	 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5638 	 * we maintain encrypted data seperately from compressed / uncompressed
5639 	 * data. If the user is requesting raw encrypted data and we don't have
5640 	 * that in the header we will read from disk to guarantee that we can
5641 	 * get it even if the encryption keys aren't loaded.
5642 	 */
5643 	if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
5644 	    (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
5645 		arc_buf_t *buf = NULL;
5646 		*arc_flags |= ARC_FLAG_CACHED;
5647 
5648 		if (HDR_IO_IN_PROGRESS(hdr)) {
5649 			zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5650 
5651 			ASSERT3P(head_zio, !=, NULL);
5652 			if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5653 			    priority == ZIO_PRIORITY_SYNC_READ) {
5654 				/*
5655 				 * This is a sync read that needs to wait for
5656 				 * an in-flight async read. Request that the
5657 				 * zio have its priority upgraded.
5658 				 */
5659 				zio_change_priority(head_zio, priority);
5660 				DTRACE_PROBE1(arc__async__upgrade__sync,
5661 				    arc_buf_hdr_t *, hdr);
5662 				ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5663 			}
5664 			if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5665 				arc_hdr_clear_flags(hdr,
5666 				    ARC_FLAG_PREDICTIVE_PREFETCH);
5667 			}
5668 
5669 			if (*arc_flags & ARC_FLAG_WAIT) {
5670 				cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5671 				mutex_exit(hash_lock);
5672 				goto top;
5673 			}
5674 			ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5675 
5676 			if (done) {
5677 				arc_callback_t *acb = NULL;
5678 
5679 				acb = kmem_zalloc(sizeof (arc_callback_t),
5680 				    KM_SLEEP);
5681 				acb->acb_done = done;
5682 				acb->acb_private = private;
5683 				acb->acb_compressed = compressed_read;
5684 				acb->acb_encrypted = encrypted_read;
5685 				acb->acb_noauth = noauth_read;
5686 				acb->acb_zb = *zb;
5687 				if (pio != NULL)
5688 					acb->acb_zio_dummy = zio_null(pio,
5689 					    spa, NULL, NULL, NULL, zio_flags);
5690 
5691 				ASSERT3P(acb->acb_done, !=, NULL);
5692 				acb->acb_zio_head = head_zio;
5693 				acb->acb_next = hdr->b_l1hdr.b_acb;
5694 				hdr->b_l1hdr.b_acb = acb;
5695 				mutex_exit(hash_lock);
5696 				return (0);
5697 			}
5698 			mutex_exit(hash_lock);
5699 			return (0);
5700 		}
5701 
5702 		ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5703 		    hdr->b_l1hdr.b_state == arc_mfu);
5704 
5705 		if (done) {
5706 			if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5707 				/*
5708 				 * This is a demand read which does not have to
5709 				 * wait for i/o because we did a predictive
5710 				 * prefetch i/o for it, which has completed.
5711 				 */
5712 				DTRACE_PROBE1(
5713 				    arc__demand__hit__predictive__prefetch,
5714 				    arc_buf_hdr_t *, hdr);
5715 				ARCSTAT_BUMP(
5716 				    arcstat_demand_hit_predictive_prefetch);
5717 				arc_hdr_clear_flags(hdr,
5718 				    ARC_FLAG_PREDICTIVE_PREFETCH);
5719 			}
5720 
5721 			if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5722 				ARCSTAT_BUMP(
5723 				    arcstat_demand_hit_prescient_prefetch);
5724 				arc_hdr_clear_flags(hdr,
5725 				    ARC_FLAG_PRESCIENT_PREFETCH);
5726 			}
5727 
5728 			ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5729 
5730 			arc_hdr_verify_checksum(spa, hdr, bp);
5731 
5732 			/* Get a buf with the desired data in it. */
5733 			rc = arc_buf_alloc_impl(hdr, spa, zb, private,
5734 			    encrypted_read, compressed_read, noauth_read,
5735 			    B_TRUE, &buf);
5736 			if (rc == ECKSUM) {
5737 				/*
5738 				 * Convert authentication and decryption errors
5739 				 * to EIO (and generate an ereport if needed)
5740 				 * before leaving the ARC.
5741 				 */
5742 				rc = SET_ERROR(EIO);
5743 				if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5744 					spa_log_error(spa, zb);
5745 					(void) zfs_ereport_post(
5746 					    FM_EREPORT_ZFS_AUTHENTICATION,
5747 					    spa, NULL, zb, NULL, 0, 0);
5748 				}
5749 			}
5750 			if (rc != 0) {
5751 				(void) remove_reference(hdr, hash_lock,
5752 				    private);
5753 				arc_buf_destroy_impl(buf);
5754 				buf = NULL;
5755 			}
5756 			/* assert any errors weren't due to unloaded keys */
5757 			ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5758 			    rc != EACCES);
5759 		} else if (*arc_flags & ARC_FLAG_PREFETCH &&
5760 		    zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5761 			if (HDR_HAS_L2HDR(hdr))
5762 				l2arc_hdr_arcstats_decrement_state(hdr);
5763 			arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5764 			if (HDR_HAS_L2HDR(hdr))
5765 				l2arc_hdr_arcstats_increment_state(hdr);
5766 		}
5767 		DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5768 		arc_access(hdr, hash_lock);
5769 		if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5770 			arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5771 		if (*arc_flags & ARC_FLAG_L2CACHE)
5772 			arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5773 		mutex_exit(hash_lock);
5774 		ARCSTAT_BUMP(arcstat_hits);
5775 		ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5776 		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5777 		    data, metadata, hits);
5778 
5779 		if (done)
5780 			done(NULL, zb, bp, buf, private);
5781 	} else {
5782 		uint64_t lsize = BP_GET_LSIZE(bp);
5783 		uint64_t psize = BP_GET_PSIZE(bp);
5784 		arc_callback_t *acb;
5785 		vdev_t *vd = NULL;
5786 		uint64_t addr = 0;
5787 		boolean_t devw = B_FALSE;
5788 		uint64_t size;
5789 		abd_t *hdr_abd;
5790 		int alloc_flags = encrypted_read ? ARC_HDR_ALLOC_RDATA : 0;
5791 
5792 		if (hdr == NULL) {
5793 			/* this block is not in the cache */
5794 			arc_buf_hdr_t *exists = NULL;
5795 			arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5796 			hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5797 			    BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), type,
5798 			    encrypted_read);
5799 
5800 			if (!BP_IS_EMBEDDED(bp)) {
5801 				hdr->b_dva = *BP_IDENTITY(bp);
5802 				hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5803 				exists = buf_hash_insert(hdr, &hash_lock);
5804 			}
5805 			if (exists != NULL) {
5806 				/* somebody beat us to the hash insert */
5807 				mutex_exit(hash_lock);
5808 				buf_discard_identity(hdr);
5809 				arc_hdr_destroy(hdr);
5810 				goto top; /* restart the IO request */
5811 			}
5812 		} else {
5813 			/*
5814 			 * This block is in the ghost cache or encrypted data
5815 			 * was requested and we didn't have it. If it was
5816 			 * L2-only (and thus didn't have an L1 hdr),
5817 			 * we realloc the header to add an L1 hdr.
5818 			 */
5819 			if (!HDR_HAS_L1HDR(hdr)) {
5820 				hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5821 				    hdr_full_cache);
5822 			}
5823 
5824 			if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
5825 				ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5826 				ASSERT(!HDR_HAS_RABD(hdr));
5827 				ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5828 				ASSERT0(zfs_refcount_count(
5829 				    &hdr->b_l1hdr.b_refcnt));
5830 				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5831 				ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5832 			} else if (HDR_IO_IN_PROGRESS(hdr)) {
5833 				/*
5834 				 * If this header already had an IO in progress
5835 				 * and we are performing another IO to fetch
5836 				 * encrypted data we must wait until the first
5837 				 * IO completes so as not to confuse
5838 				 * arc_read_done(). This should be very rare
5839 				 * and so the performance impact shouldn't
5840 				 * matter.
5841 				 */
5842 				cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5843 				mutex_exit(hash_lock);
5844 				goto top;
5845 			}
5846 
5847 			/*
5848 			 * This is a delicate dance that we play here.
5849 			 * This hdr might be in the ghost list so we access
5850 			 * it to move it out of the ghost list before we
5851 			 * initiate the read. If it's a prefetch then
5852 			 * it won't have a callback so we'll remove the
5853 			 * reference that arc_buf_alloc_impl() created. We
5854 			 * do this after we've called arc_access() to
5855 			 * avoid hitting an assert in remove_reference().
5856 			 */
5857 			arc_adapt(arc_hdr_size(hdr), hdr->b_l1hdr.b_state);
5858 			arc_access(hdr, hash_lock);
5859 			arc_hdr_alloc_pabd(hdr, alloc_flags);
5860 		}
5861 
5862 		if (encrypted_read) {
5863 			ASSERT(HDR_HAS_RABD(hdr));
5864 			size = HDR_GET_PSIZE(hdr);
5865 			hdr_abd = hdr->b_crypt_hdr.b_rabd;
5866 			zio_flags |= ZIO_FLAG_RAW;
5867 		} else {
5868 			ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5869 			size = arc_hdr_size(hdr);
5870 			hdr_abd = hdr->b_l1hdr.b_pabd;
5871 
5872 			if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
5873 				zio_flags |= ZIO_FLAG_RAW_COMPRESS;
5874 			}
5875 
5876 			/*
5877 			 * For authenticated bp's, we do not ask the ZIO layer
5878 			 * to authenticate them since this will cause the entire
5879 			 * IO to fail if the key isn't loaded. Instead, we
5880 			 * defer authentication until arc_buf_fill(), which will
5881 			 * verify the data when the key is available.
5882 			 */
5883 			if (BP_IS_AUTHENTICATED(bp))
5884 				zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
5885 		}
5886 
5887 		if (*arc_flags & ARC_FLAG_PREFETCH &&
5888 		    zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5889 			if (HDR_HAS_L2HDR(hdr))
5890 				l2arc_hdr_arcstats_decrement_state(hdr);
5891 			arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5892 			if (HDR_HAS_L2HDR(hdr))
5893 				l2arc_hdr_arcstats_increment_state(hdr);
5894 		}
5895 		if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5896 			arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5897 
5898 		if (*arc_flags & ARC_FLAG_L2CACHE)
5899 			arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5900 		if (BP_IS_AUTHENTICATED(bp))
5901 			arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
5902 		if (BP_GET_LEVEL(bp) > 0)
5903 			arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5904 		if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5905 			arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5906 		ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5907 
5908 		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5909 		acb->acb_done = done;
5910 		acb->acb_private = private;
5911 		acb->acb_compressed = compressed_read;
5912 		acb->acb_encrypted = encrypted_read;
5913 		acb->acb_noauth = noauth_read;
5914 		acb->acb_zb = *zb;
5915 
5916 		ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5917 		hdr->b_l1hdr.b_acb = acb;
5918 		arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5919 
5920 		if (HDR_HAS_L2HDR(hdr) &&
5921 		    (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5922 			devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5923 			addr = hdr->b_l2hdr.b_daddr;
5924 			/*
5925 			 * Lock out L2ARC device removal.
5926 			 */
5927 			if (vdev_is_dead(vd) ||
5928 			    !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5929 				vd = NULL;
5930 		}
5931 
5932 		/*
5933 		 * We count both async reads and scrub IOs as asynchronous so
5934 		 * that both can be upgraded in the event of a cache hit while
5935 		 * the read IO is still in-flight.
5936 		 */
5937 		if (priority == ZIO_PRIORITY_ASYNC_READ ||
5938 		    priority == ZIO_PRIORITY_SCRUB)
5939 			arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5940 		else
5941 			arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5942 
5943 		/*
5944 		 * At this point, we have a level 1 cache miss.  Try again in
5945 		 * L2ARC if possible.
5946 		 */
5947 		ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5948 
5949 		DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5950 		    uint64_t, lsize, zbookmark_phys_t *, zb);
5951 		ARCSTAT_BUMP(arcstat_misses);
5952 		ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5953 		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5954 		    data, metadata, misses);
5955 
5956 		if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5957 			/*
5958 			 * Read from the L2ARC if the following are true:
5959 			 * 1. The L2ARC vdev was previously cached.
5960 			 * 2. This buffer still has L2ARC metadata.
5961 			 * 3. This buffer isn't currently writing to the L2ARC.
5962 			 * 4. The L2ARC entry wasn't evicted, which may
5963 			 *    also have invalidated the vdev.
5964 			 * 5. This isn't prefetch or l2arc_noprefetch is 0.
5965 			 */
5966 			if (HDR_HAS_L2HDR(hdr) &&
5967 			    !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5968 			    !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5969 				l2arc_read_callback_t *cb;
5970 				abd_t *abd;
5971 				uint64_t asize;
5972 
5973 				DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5974 				ARCSTAT_BUMP(arcstat_l2_hits);
5975 
5976 				cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5977 				    KM_SLEEP);
5978 				cb->l2rcb_hdr = hdr;
5979 				cb->l2rcb_bp = *bp;
5980 				cb->l2rcb_zb = *zb;
5981 				cb->l2rcb_flags = zio_flags;
5982 
5983 				/*
5984 				 * When Compressed ARC is disabled, but the
5985 				 * L2ARC block is compressed, arc_hdr_size()
5986 				 * will have returned LSIZE rather than PSIZE.
5987 				 */
5988 				if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
5989 				    !HDR_COMPRESSION_ENABLED(hdr) &&
5990 				    HDR_GET_PSIZE(hdr) != 0) {
5991 					size = HDR_GET_PSIZE(hdr);
5992 				}
5993 
5994 				asize = vdev_psize_to_asize(vd, size);
5995 				if (asize != size) {
5996 					abd = abd_alloc_for_io(asize,
5997 					    HDR_ISTYPE_METADATA(hdr));
5998 					cb->l2rcb_abd = abd;
5999 				} else {
6000 					abd = hdr_abd;
6001 				}
6002 
6003 				ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6004 				    addr + asize <= vd->vdev_psize -
6005 				    VDEV_LABEL_END_SIZE);
6006 
6007 				/*
6008 				 * l2arc read.  The SCL_L2ARC lock will be
6009 				 * released by l2arc_read_done().
6010 				 * Issue a null zio if the underlying buffer
6011 				 * was squashed to zero size by compression.
6012 				 */
6013 				ASSERT3U(arc_hdr_get_compress(hdr), !=,
6014 				    ZIO_COMPRESS_EMPTY);
6015 				rzio = zio_read_phys(pio, vd, addr,
6016 				    asize, abd,
6017 				    ZIO_CHECKSUM_OFF,
6018 				    l2arc_read_done, cb, priority,
6019 				    zio_flags | ZIO_FLAG_DONT_CACHE |
6020 				    ZIO_FLAG_CANFAIL |
6021 				    ZIO_FLAG_DONT_PROPAGATE |
6022 				    ZIO_FLAG_DONT_RETRY, B_FALSE);
6023 				acb->acb_zio_head = rzio;
6024 
6025 				if (hash_lock != NULL)
6026 					mutex_exit(hash_lock);
6027 
6028 				DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6029 				    zio_t *, rzio);
6030 				ARCSTAT_INCR(arcstat_l2_read_bytes,
6031 				    HDR_GET_PSIZE(hdr));
6032 
6033 				if (*arc_flags & ARC_FLAG_NOWAIT) {
6034 					zio_nowait(rzio);
6035 					return (0);
6036 				}
6037 
6038 				ASSERT(*arc_flags & ARC_FLAG_WAIT);
6039 				if (zio_wait(rzio) == 0)
6040 					return (0);
6041 
6042 				/* l2arc read error; goto zio_read() */
6043 				if (hash_lock != NULL)
6044 					mutex_enter(hash_lock);
6045 			} else {
6046 				DTRACE_PROBE1(l2arc__miss,
6047 				    arc_buf_hdr_t *, hdr);
6048 				ARCSTAT_BUMP(arcstat_l2_misses);
6049 				if (HDR_L2_WRITING(hdr))
6050 					ARCSTAT_BUMP(arcstat_l2_rw_clash);
6051 				spa_config_exit(spa, SCL_L2ARC, vd);
6052 			}
6053 		} else {
6054 			if (vd != NULL)
6055 				spa_config_exit(spa, SCL_L2ARC, vd);
6056 			if (l2arc_ndev != 0) {
6057 				DTRACE_PROBE1(l2arc__miss,
6058 				    arc_buf_hdr_t *, hdr);
6059 				ARCSTAT_BUMP(arcstat_l2_misses);
6060 			}
6061 		}
6062 
6063 		rzio = zio_read(pio, spa, bp, hdr_abd, size,
6064 		    arc_read_done, hdr, priority, zio_flags, zb);
6065 		acb->acb_zio_head = rzio;
6066 
6067 		if (hash_lock != NULL)
6068 			mutex_exit(hash_lock);
6069 
6070 		if (*arc_flags & ARC_FLAG_WAIT)
6071 			return (zio_wait(rzio));
6072 
6073 		ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6074 		zio_nowait(rzio);
6075 	}
6076 	return (rc);
6077 }
6078 
6079 /*
6080  * Notify the arc that a block was freed, and thus will never be used again.
6081  */
6082 void
arc_freed(spa_t * spa,const blkptr_t * bp)6083 arc_freed(spa_t *spa, const blkptr_t *bp)
6084 {
6085 	arc_buf_hdr_t *hdr;
6086 	kmutex_t *hash_lock;
6087 	uint64_t guid = spa_load_guid(spa);
6088 
6089 	ASSERT(!BP_IS_EMBEDDED(bp));
6090 
6091 	hdr = buf_hash_find(guid, bp, &hash_lock);
6092 	if (hdr == NULL)
6093 		return;
6094 
6095 	/*
6096 	 * We might be trying to free a block that is still doing I/O
6097 	 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6098 	 * dmu_sync-ed block). If this block is being prefetched, then it
6099 	 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6100 	 * until the I/O completes. A block may also have a reference if it is
6101 	 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6102 	 * have written the new block to its final resting place on disk but
6103 	 * without the dedup flag set. This would have left the hdr in the MRU
6104 	 * state and discoverable. When the txg finally syncs it detects that
6105 	 * the block was overridden in open context and issues an override I/O.
6106 	 * Since this is a dedup block, the override I/O will determine if the
6107 	 * block is already in the DDT. If so, then it will replace the io_bp
6108 	 * with the bp from the DDT and allow the I/O to finish. When the I/O
6109 	 * reaches the done callback, dbuf_write_override_done, it will
6110 	 * check to see if the io_bp and io_bp_override are identical.
6111 	 * If they are not, then it indicates that the bp was replaced with
6112 	 * the bp in the DDT and the override bp is freed. This allows
6113 	 * us to arrive here with a reference on a block that is being
6114 	 * freed. So if we have an I/O in progress, or a reference to
6115 	 * this hdr, then we don't destroy the hdr.
6116 	 */
6117 	if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6118 	    zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6119 		arc_change_state(arc_anon, hdr, hash_lock);
6120 		arc_hdr_destroy(hdr);
6121 		mutex_exit(hash_lock);
6122 	} else {
6123 		mutex_exit(hash_lock);
6124 	}
6125 
6126 }
6127 
6128 /*
6129  * Release this buffer from the cache, making it an anonymous buffer.  This
6130  * must be done after a read and prior to modifying the buffer contents.
6131  * If the buffer has more than one reference, we must make
6132  * a new hdr for the buffer.
6133  */
6134 void
arc_release(arc_buf_t * buf,void * tag)6135 arc_release(arc_buf_t *buf, void *tag)
6136 {
6137 	/*
6138 	 * It would be nice to assert that if its DMU metadata (level >
6139 	 * 0 || it's the dnode file), then it must be syncing context.
6140 	 * But we don't know that information at this level.
6141 	 */
6142 
6143 	mutex_enter(&buf->b_evict_lock);
6144 
6145 	arc_buf_hdr_t *hdr = buf->b_hdr;
6146 
6147 	ASSERT(HDR_HAS_L1HDR(hdr));
6148 
6149 	/*
6150 	 * We don't grab the hash lock prior to this check, because if
6151 	 * the buffer's header is in the arc_anon state, it won't be
6152 	 * linked into the hash table.
6153 	 */
6154 	if (hdr->b_l1hdr.b_state == arc_anon) {
6155 		mutex_exit(&buf->b_evict_lock);
6156 		/*
6157 		 * If we are called from dmu_convert_mdn_block_to_raw(),
6158 		 * a write might be in progress.  This is OK because
6159 		 * the caller won't change the content of this buffer,
6160 		 * only the flags (via arc_convert_to_raw()).
6161 		 */
6162 		/* ASSERT(!HDR_IO_IN_PROGRESS(hdr)); */
6163 		ASSERT(!HDR_IN_HASH_TABLE(hdr));
6164 		ASSERT(!HDR_HAS_L2HDR(hdr));
6165 		ASSERT(HDR_EMPTY(hdr));
6166 
6167 		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6168 		ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6169 		ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6170 
6171 		hdr->b_l1hdr.b_arc_access = 0;
6172 
6173 		/*
6174 		 * If the buf is being overridden then it may already
6175 		 * have a hdr that is not empty.
6176 		 */
6177 		buf_discard_identity(hdr);
6178 		arc_buf_thaw(buf);
6179 
6180 		return;
6181 	}
6182 
6183 	kmutex_t *hash_lock = HDR_LOCK(hdr);
6184 	mutex_enter(hash_lock);
6185 
6186 	/*
6187 	 * Wait for any other IO for this hdr, as additional
6188 	 * buf(s) could be about to appear, in which case
6189 	 * we would not want to transition hdr to arc_anon.
6190 	 */
6191 	while (HDR_IO_IN_PROGRESS(hdr)) {
6192 		DTRACE_PROBE1(arc_release__io, arc_buf_hdr_t *, hdr);
6193 		cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6194 	}
6195 
6196 	/*
6197 	 * This assignment is only valid as long as the hash_lock is
6198 	 * held, we must be careful not to reference state or the
6199 	 * b_state field after dropping the lock.
6200 	 */
6201 	arc_state_t *state = hdr->b_l1hdr.b_state;
6202 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6203 	ASSERT3P(state, !=, arc_anon);
6204 
6205 	/* this buffer is not on any list */
6206 	ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6207 
6208 	if (HDR_HAS_L2HDR(hdr)) {
6209 		mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6210 
6211 		/*
6212 		 * We have to recheck this conditional again now that
6213 		 * we're holding the l2ad_mtx to prevent a race with
6214 		 * another thread which might be concurrently calling
6215 		 * l2arc_evict(). In that case, l2arc_evict() might have
6216 		 * destroyed the header's L2 portion as we were waiting
6217 		 * to acquire the l2ad_mtx.
6218 		 */
6219 		if (HDR_HAS_L2HDR(hdr))
6220 			arc_hdr_l2hdr_destroy(hdr);
6221 
6222 		mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6223 	}
6224 
6225 	/*
6226 	 * Do we have more than one buf?
6227 	 */
6228 	if (hdr->b_l1hdr.b_bufcnt > 1) {
6229 		arc_buf_hdr_t *nhdr;
6230 		uint64_t spa = hdr->b_spa;
6231 		uint64_t psize = HDR_GET_PSIZE(hdr);
6232 		uint64_t lsize = HDR_GET_LSIZE(hdr);
6233 		boolean_t protected = HDR_PROTECTED(hdr);
6234 		enum zio_compress compress = arc_hdr_get_compress(hdr);
6235 		arc_buf_contents_t type = arc_buf_type(hdr);
6236 		VERIFY3U(hdr->b_type, ==, type);
6237 
6238 		ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6239 		(void) remove_reference(hdr, hash_lock, tag);
6240 
6241 		if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6242 			ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6243 			ASSERT(ARC_BUF_LAST(buf));
6244 		}
6245 
6246 		/*
6247 		 * Pull the data off of this hdr and attach it to
6248 		 * a new anonymous hdr. Also find the last buffer
6249 		 * in the hdr's buffer list.
6250 		 */
6251 		arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6252 		ASSERT3P(lastbuf, !=, NULL);
6253 
6254 		/*
6255 		 * If the current arc_buf_t and the hdr are sharing their data
6256 		 * buffer, then we must stop sharing that block.
6257 		 */
6258 		if (arc_buf_is_shared(buf)) {
6259 			ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6260 			VERIFY(!arc_buf_is_shared(lastbuf));
6261 
6262 			/*
6263 			 * First, sever the block sharing relationship between
6264 			 * buf and the arc_buf_hdr_t.
6265 			 */
6266 			arc_unshare_buf(hdr, buf);
6267 
6268 			/*
6269 			 * Now we need to recreate the hdr's b_pabd. Since we
6270 			 * have lastbuf handy, we try to share with it, but if
6271 			 * we can't then we allocate a new b_pabd and copy the
6272 			 * data from buf into it.
6273 			 */
6274 			if (arc_can_share(hdr, lastbuf)) {
6275 				arc_share_buf(hdr, lastbuf);
6276 			} else {
6277 				arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT);
6278 				abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6279 				    buf->b_data, psize);
6280 			}
6281 			VERIFY3P(lastbuf->b_data, !=, NULL);
6282 		} else if (HDR_SHARED_DATA(hdr)) {
6283 			/*
6284 			 * Uncompressed shared buffers are always at the end
6285 			 * of the list. Compressed buffers don't have the
6286 			 * same requirements. This makes it hard to
6287 			 * simply assert that the lastbuf is shared so
6288 			 * we rely on the hdr's compression flags to determine
6289 			 * if we have a compressed, shared buffer.
6290 			 */
6291 			ASSERT(arc_buf_is_shared(lastbuf) ||
6292 			    arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6293 			ASSERT(!ARC_BUF_SHARED(buf));
6294 		}
6295 		ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6296 		ASSERT3P(state, !=, arc_l2c_only);
6297 
6298 		(void) zfs_refcount_remove_many(&state->arcs_size,
6299 		    arc_buf_size(buf), buf);
6300 
6301 		if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6302 			ASSERT3P(state, !=, arc_l2c_only);
6303 			(void) zfs_refcount_remove_many(
6304 			    &state->arcs_esize[type],
6305 			    arc_buf_size(buf), buf);
6306 		}
6307 
6308 		hdr->b_l1hdr.b_bufcnt -= 1;
6309 		if (ARC_BUF_ENCRYPTED(buf))
6310 			hdr->b_crypt_hdr.b_ebufcnt -= 1;
6311 
6312 		arc_cksum_verify(buf);
6313 		arc_buf_unwatch(buf);
6314 
6315 		/* if this is the last uncompressed buf free the checksum */
6316 		if (!arc_hdr_has_uncompressed_buf(hdr))
6317 			arc_cksum_free(hdr);
6318 
6319 		mutex_exit(hash_lock);
6320 
6321 		/*
6322 		 * Allocate a new hdr. The new hdr will contain a b_pabd
6323 		 * buffer which will be freed in arc_write().
6324 		 */
6325 		nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6326 		    compress, type, HDR_HAS_RABD(hdr));
6327 		ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6328 		ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6329 		ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6330 		VERIFY3U(nhdr->b_type, ==, type);
6331 		ASSERT(!HDR_SHARED_DATA(nhdr));
6332 
6333 		nhdr->b_l1hdr.b_buf = buf;
6334 		nhdr->b_l1hdr.b_bufcnt = 1;
6335 		if (ARC_BUF_ENCRYPTED(buf))
6336 			nhdr->b_crypt_hdr.b_ebufcnt = 1;
6337 		(void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6338 		buf->b_hdr = nhdr;
6339 
6340 		mutex_exit(&buf->b_evict_lock);
6341 		(void) zfs_refcount_add_many(&arc_anon->arcs_size,
6342 		    arc_buf_size(buf), buf);
6343 	} else {
6344 		mutex_exit(&buf->b_evict_lock);
6345 		ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6346 		/* protected by hash lock, or hdr is on arc_anon */
6347 		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6348 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6349 		arc_change_state(arc_anon, hdr, hash_lock);
6350 		hdr->b_l1hdr.b_arc_access = 0;
6351 
6352 		mutex_exit(hash_lock);
6353 		buf_discard_identity(hdr);
6354 		arc_buf_thaw(buf);
6355 	}
6356 }
6357 
6358 int
arc_released(arc_buf_t * buf)6359 arc_released(arc_buf_t *buf)
6360 {
6361 	int released;
6362 
6363 	mutex_enter(&buf->b_evict_lock);
6364 	released = (buf->b_data != NULL &&
6365 	    buf->b_hdr->b_l1hdr.b_state == arc_anon);
6366 	mutex_exit(&buf->b_evict_lock);
6367 	return (released);
6368 }
6369 
6370 #ifdef ZFS_DEBUG
6371 int
arc_referenced(arc_buf_t * buf)6372 arc_referenced(arc_buf_t *buf)
6373 {
6374 	int referenced;
6375 
6376 	mutex_enter(&buf->b_evict_lock);
6377 	referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6378 	mutex_exit(&buf->b_evict_lock);
6379 	return (referenced);
6380 }
6381 #endif
6382 
6383 static void
arc_write_ready(zio_t * zio)6384 arc_write_ready(zio_t *zio)
6385 {
6386 	arc_write_callback_t *callback = zio->io_private;
6387 	arc_buf_t *buf = callback->awcb_buf;
6388 	arc_buf_hdr_t *hdr = buf->b_hdr;
6389 	blkptr_t *bp = zio->io_bp;
6390 	uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6391 
6392 	ASSERT(HDR_HAS_L1HDR(hdr));
6393 	ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6394 	ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6395 
6396 	/*
6397 	 * If we're reexecuting this zio because the pool suspended, then
6398 	 * cleanup any state that was previously set the first time the
6399 	 * callback was invoked.
6400 	 */
6401 	if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6402 		arc_cksum_free(hdr);
6403 		arc_buf_unwatch(buf);
6404 		if (hdr->b_l1hdr.b_pabd != NULL) {
6405 			if (arc_buf_is_shared(buf)) {
6406 				arc_unshare_buf(hdr, buf);
6407 			} else {
6408 				arc_hdr_free_pabd(hdr, B_FALSE);
6409 			}
6410 		}
6411 
6412 		if (HDR_HAS_RABD(hdr))
6413 			arc_hdr_free_pabd(hdr, B_TRUE);
6414 	}
6415 	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6416 	ASSERT(!HDR_HAS_RABD(hdr));
6417 	ASSERT(!HDR_SHARED_DATA(hdr));
6418 	ASSERT(!arc_buf_is_shared(buf));
6419 
6420 	callback->awcb_ready(zio, buf, callback->awcb_private);
6421 
6422 	if (HDR_IO_IN_PROGRESS(hdr))
6423 		ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6424 
6425 	arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6426 
6427 	if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
6428 		hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
6429 
6430 	if (BP_IS_PROTECTED(bp)) {
6431 		/* ZIL blocks are written through zio_rewrite */
6432 		ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6433 		ASSERT(HDR_PROTECTED(hdr));
6434 
6435 		if (BP_SHOULD_BYTESWAP(bp)) {
6436 			if (BP_GET_LEVEL(bp) > 0) {
6437 				hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6438 			} else {
6439 				hdr->b_l1hdr.b_byteswap =
6440 				    DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6441 			}
6442 		} else {
6443 			hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6444 		}
6445 
6446 		hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6447 		hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6448 		zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6449 		    hdr->b_crypt_hdr.b_iv);
6450 		zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6451 	}
6452 
6453 	/*
6454 	 * If this block was written for raw encryption but the zio layer
6455 	 * ended up only authenticating it, adjust the buffer flags now.
6456 	 */
6457 	if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6458 		arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6459 		buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6460 		if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6461 			buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6462 	} else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6463 		buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6464 		buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6465 	}
6466 
6467 	/* this must be done after the buffer flags are adjusted */
6468 	arc_cksum_compute(buf);
6469 
6470 	enum zio_compress compress;
6471 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6472 		compress = ZIO_COMPRESS_OFF;
6473 	} else {
6474 		ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6475 		compress = BP_GET_COMPRESS(bp);
6476 	}
6477 	HDR_SET_PSIZE(hdr, psize);
6478 	arc_hdr_set_compress(hdr, compress);
6479 
6480 	if (zio->io_error != 0 || psize == 0)
6481 		goto out;
6482 
6483 	/*
6484 	 * Fill the hdr with data. If the buffer is encrypted we have no choice
6485 	 * but to copy the data into b_rabd. If the hdr is compressed, the data
6486 	 * we want is available from the zio, otherwise we can take it from
6487 	 * the buf.
6488 	 *
6489 	 * We might be able to share the buf's data with the hdr here. However,
6490 	 * doing so would cause the ARC to be full of linear ABDs if we write a
6491 	 * lot of shareable data. As a compromise, we check whether scattered
6492 	 * ABDs are allowed, and assume that if they are then the user wants
6493 	 * the ARC to be primarily filled with them regardless of the data being
6494 	 * written. Therefore, if they're allowed then we allocate one and copy
6495 	 * the data into it; otherwise, we share the data directly if we can.
6496 	 */
6497 	if (ARC_BUF_ENCRYPTED(buf)) {
6498 		ASSERT3U(psize, >, 0);
6499 		ASSERT(ARC_BUF_COMPRESSED(buf));
6500 		arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT|ARC_HDR_ALLOC_RDATA);
6501 		abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6502 	} else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6503 		/*
6504 		 * Ideally, we would always copy the io_abd into b_pabd, but the
6505 		 * user may have disabled compressed ARC, thus we must check the
6506 		 * hdr's compression setting rather than the io_bp's.
6507 		 */
6508 		if (BP_IS_ENCRYPTED(bp)) {
6509 			ASSERT3U(psize, >, 0);
6510 			arc_hdr_alloc_pabd(hdr,
6511 			    ARC_HDR_DO_ADAPT|ARC_HDR_ALLOC_RDATA);
6512 			abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6513 		} else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6514 		    !ARC_BUF_COMPRESSED(buf)) {
6515 			ASSERT3U(psize, >, 0);
6516 			arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT);
6517 			abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6518 		} else {
6519 			ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6520 			arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT);
6521 			abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6522 			    arc_buf_size(buf));
6523 		}
6524 	} else {
6525 		ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6526 		ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6527 		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6528 		arc_share_buf(hdr, buf);
6529 	}
6530 
6531 out:
6532 	arc_hdr_verify(hdr, bp);
6533 }
6534 
6535 static void
arc_write_children_ready(zio_t * zio)6536 arc_write_children_ready(zio_t *zio)
6537 {
6538 	arc_write_callback_t *callback = zio->io_private;
6539 	arc_buf_t *buf = callback->awcb_buf;
6540 
6541 	callback->awcb_children_ready(zio, buf, callback->awcb_private);
6542 }
6543 
6544 /*
6545  * The SPA calls this callback for each physical write that happens on behalf
6546  * of a logical write.  See the comment in dbuf_write_physdone() for details.
6547  */
6548 static void
arc_write_physdone(zio_t * zio)6549 arc_write_physdone(zio_t *zio)
6550 {
6551 	arc_write_callback_t *cb = zio->io_private;
6552 	if (cb->awcb_physdone != NULL)
6553 		cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6554 }
6555 
6556 static void
arc_write_done(zio_t * zio)6557 arc_write_done(zio_t *zio)
6558 {
6559 	arc_write_callback_t *callback = zio->io_private;
6560 	arc_buf_t *buf = callback->awcb_buf;
6561 	arc_buf_hdr_t *hdr = buf->b_hdr;
6562 
6563 	ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6564 
6565 	if (zio->io_error == 0) {
6566 		arc_hdr_verify(hdr, zio->io_bp);
6567 
6568 		if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6569 			buf_discard_identity(hdr);
6570 		} else {
6571 			hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6572 			hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6573 		}
6574 	} else {
6575 		ASSERT(HDR_EMPTY(hdr));
6576 	}
6577 
6578 	/*
6579 	 * If the block to be written was all-zero or compressed enough to be
6580 	 * embedded in the BP, no write was performed so there will be no
6581 	 * dva/birth/checksum.  The buffer must therefore remain anonymous
6582 	 * (and uncached).
6583 	 */
6584 	if (!HDR_EMPTY(hdr)) {
6585 		arc_buf_hdr_t *exists;
6586 		kmutex_t *hash_lock;
6587 
6588 		ASSERT3U(zio->io_error, ==, 0);
6589 
6590 		arc_cksum_verify(buf);
6591 
6592 		exists = buf_hash_insert(hdr, &hash_lock);
6593 		if (exists != NULL) {
6594 			/*
6595 			 * This can only happen if we overwrite for
6596 			 * sync-to-convergence, because we remove
6597 			 * buffers from the hash table when we arc_free().
6598 			 */
6599 			if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6600 				if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6601 					panic("bad overwrite, hdr=%p exists=%p",
6602 					    (void *)hdr, (void *)exists);
6603 				ASSERT(zfs_refcount_is_zero(
6604 				    &exists->b_l1hdr.b_refcnt));
6605 				arc_change_state(arc_anon, exists, hash_lock);
6606 				arc_hdr_destroy(exists);
6607 				mutex_exit(hash_lock);
6608 				exists = buf_hash_insert(hdr, &hash_lock);
6609 				ASSERT3P(exists, ==, NULL);
6610 			} else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6611 				/* nopwrite */
6612 				ASSERT(zio->io_prop.zp_nopwrite);
6613 				if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6614 					panic("bad nopwrite, hdr=%p exists=%p",
6615 					    (void *)hdr, (void *)exists);
6616 			} else {
6617 				/* Dedup */
6618 				ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6619 				ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6620 				ASSERT(BP_GET_DEDUP(zio->io_bp));
6621 				ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6622 			}
6623 		}
6624 		arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6625 		/* if it's not anon, we are doing a scrub */
6626 		if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6627 			arc_access(hdr, hash_lock);
6628 		mutex_exit(hash_lock);
6629 	} else {
6630 		arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6631 	}
6632 
6633 	ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6634 	callback->awcb_done(zio, buf, callback->awcb_private);
6635 
6636 	abd_put(zio->io_abd);
6637 	kmem_free(callback, sizeof (arc_write_callback_t));
6638 }
6639 
6640 zio_t *
arc_write(zio_t * pio,spa_t * spa,uint64_t txg,blkptr_t * bp,arc_buf_t * buf,boolean_t l2arc,const zio_prop_t * zp,arc_write_done_func_t * ready,arc_write_done_func_t * children_ready,arc_write_done_func_t * physdone,arc_write_done_func_t * done,void * private,zio_priority_t priority,int zio_flags,const zbookmark_phys_t * zb)6641 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6642     boolean_t l2arc, const zio_prop_t *zp, arc_write_done_func_t *ready,
6643     arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
6644     arc_write_done_func_t *done, void *private, zio_priority_t priority,
6645     int zio_flags, const zbookmark_phys_t *zb)
6646 {
6647 	arc_buf_hdr_t *hdr = buf->b_hdr;
6648 	arc_write_callback_t *callback;
6649 	zio_t *zio;
6650 	zio_prop_t localprop = *zp;
6651 
6652 	ASSERT3P(ready, !=, NULL);
6653 	ASSERT3P(done, !=, NULL);
6654 	ASSERT(!HDR_IO_ERROR(hdr));
6655 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6656 	ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6657 	ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6658 	if (l2arc)
6659 		arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6660 
6661 	if (ARC_BUF_ENCRYPTED(buf)) {
6662 		ASSERT(ARC_BUF_COMPRESSED(buf));
6663 		localprop.zp_encrypt = B_TRUE;
6664 		localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6665 		/* CONSTCOND */
6666 		localprop.zp_byteorder =
6667 		    (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
6668 		    ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
6669 		bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt,
6670 		    ZIO_DATA_SALT_LEN);
6671 		bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv,
6672 		    ZIO_DATA_IV_LEN);
6673 		bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac,
6674 		    ZIO_DATA_MAC_LEN);
6675 		if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
6676 			localprop.zp_nopwrite = B_FALSE;
6677 			localprop.zp_copies =
6678 			    MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
6679 		}
6680 		zio_flags |= ZIO_FLAG_RAW;
6681 	} else if (ARC_BUF_COMPRESSED(buf)) {
6682 		ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6683 		localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6684 		zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6685 	}
6686 
6687 	callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6688 	callback->awcb_ready = ready;
6689 	callback->awcb_children_ready = children_ready;
6690 	callback->awcb_physdone = physdone;
6691 	callback->awcb_done = done;
6692 	callback->awcb_private = private;
6693 	callback->awcb_buf = buf;
6694 
6695 	/*
6696 	 * The hdr's b_pabd is now stale, free it now. A new data block
6697 	 * will be allocated when the zio pipeline calls arc_write_ready().
6698 	 */
6699 	if (hdr->b_l1hdr.b_pabd != NULL) {
6700 		/*
6701 		 * If the buf is currently sharing the data block with
6702 		 * the hdr then we need to break that relationship here.
6703 		 * The hdr will remain with a NULL data pointer and the
6704 		 * buf will take sole ownership of the block.
6705 		 */
6706 		if (arc_buf_is_shared(buf)) {
6707 			arc_unshare_buf(hdr, buf);
6708 		} else {
6709 			arc_hdr_free_pabd(hdr, B_FALSE);
6710 		}
6711 		VERIFY3P(buf->b_data, !=, NULL);
6712 	}
6713 
6714 	if (HDR_HAS_RABD(hdr))
6715 		arc_hdr_free_pabd(hdr, B_TRUE);
6716 
6717 	if (!(zio_flags & ZIO_FLAG_RAW))
6718 		arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6719 
6720 	ASSERT(!arc_buf_is_shared(buf));
6721 	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6722 
6723 	zio = zio_write(pio, spa, txg, bp,
6724 	    abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6725 	    HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6726 	    (children_ready != NULL) ? arc_write_children_ready : NULL,
6727 	    arc_write_physdone, arc_write_done, callback,
6728 	    priority, zio_flags, zb);
6729 
6730 	return (zio);
6731 }
6732 
6733 static int
arc_memory_throttle(spa_t * spa,uint64_t reserve,uint64_t txg)6734 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
6735 {
6736 #ifdef _KERNEL
6737 	uint64_t available_memory = ptob(freemem);
6738 
6739 
6740 	if (freemem > physmem * arc_lotsfree_percent / 100)
6741 		return (0);
6742 
6743 	if (txg > spa->spa_lowmem_last_txg) {
6744 		spa->spa_lowmem_last_txg = txg;
6745 		spa->spa_lowmem_page_load = 0;
6746 	}
6747 	/*
6748 	 * If we are in pageout, we know that memory is already tight,
6749 	 * the arc is already going to be evicting, so we just want to
6750 	 * continue to let page writes occur as quickly as possible.
6751 	 */
6752 	if (curproc == proc_pageout) {
6753 		if (spa->spa_lowmem_page_load >
6754 		    MAX(ptob(minfree), available_memory) / 4)
6755 			return (SET_ERROR(ERESTART));
6756 		/* Note: reserve is inflated, so we deflate */
6757 		atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
6758 		return (0);
6759 	} else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
6760 		/* memory is low, delay before restarting */
6761 		ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6762 		return (SET_ERROR(EAGAIN));
6763 	}
6764 	spa->spa_lowmem_page_load = 0;
6765 #endif /* _KERNEL */
6766 	return (0);
6767 }
6768 
6769 /*
6770  * In more extreme cases, return B_TRUE if system memory is tight enough
6771  * that ZFS should defer work requiring new allocations.
6772  */
6773 boolean_t
arc_memory_is_low(void)6774 arc_memory_is_low(void)
6775 {
6776 #ifdef _KERNEL
6777 	if (freemem < minfree + needfree)
6778 		return (B_TRUE);
6779 #endif /* _KERNEL */
6780 	return (B_FALSE);
6781 }
6782 
6783 void
arc_tempreserve_clear(uint64_t reserve)6784 arc_tempreserve_clear(uint64_t reserve)
6785 {
6786 	atomic_add_64(&arc_tempreserve, -reserve);
6787 	ASSERT((int64_t)arc_tempreserve >= 0);
6788 }
6789 
6790 int
arc_tempreserve_space(spa_t * spa,uint64_t reserve,uint64_t txg)6791 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
6792 {
6793 	int error;
6794 	uint64_t anon_size;
6795 
6796 	if (reserve > arc_c/4 && !arc_no_grow)
6797 		arc_c = MIN(arc_c_max, reserve * 4);
6798 	if (reserve > arc_c)
6799 		return (SET_ERROR(ENOMEM));
6800 
6801 	/*
6802 	 * Don't count loaned bufs as in flight dirty data to prevent long
6803 	 * network delays from blocking transactions that are ready to be
6804 	 * assigned to a txg.
6805 	 */
6806 
6807 	/* assert that it has not wrapped around */
6808 	ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6809 
6810 	anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) -
6811 	    arc_loaned_bytes), 0);
6812 
6813 	/*
6814 	 * Writes will, almost always, require additional memory allocations
6815 	 * in order to compress/encrypt/etc the data.  We therefore need to
6816 	 * make sure that there is sufficient available memory for this.
6817 	 */
6818 	error = arc_memory_throttle(spa, reserve, txg);
6819 	if (error != 0)
6820 		return (error);
6821 
6822 	/*
6823 	 * Throttle writes when the amount of dirty data in the cache
6824 	 * gets too large.  We try to keep the cache less than half full
6825 	 * of dirty blocks so that our sync times don't grow too large.
6826 	 *
6827 	 * In the case of one pool being built on another pool, we want
6828 	 * to make sure we don't end up throttling the lower (backing)
6829 	 * pool when the upper pool is the majority contributor to dirty
6830 	 * data. To insure we make forward progress during throttling, we
6831 	 * also check the current pool's net dirty data and only throttle
6832 	 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
6833 	 * data in the cache.
6834 	 *
6835 	 * Note: if two requests come in concurrently, we might let them
6836 	 * both succeed, when one of them should fail.  Not a huge deal.
6837 	 */
6838 	uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6839 	uint64_t spa_dirty_anon = spa_dirty_data(spa);
6840 
6841 	if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
6842 	    anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
6843 	    spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6844 		uint64_t meta_esize =
6845 		    zfs_refcount_count(
6846 		    &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6847 		uint64_t data_esize =
6848 		    zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6849 		dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6850 		    "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6851 		    arc_tempreserve >> 10, meta_esize >> 10,
6852 		    data_esize >> 10, reserve >> 10, arc_c >> 10);
6853 		return (SET_ERROR(ERESTART));
6854 	}
6855 	atomic_add_64(&arc_tempreserve, reserve);
6856 	return (0);
6857 }
6858 
6859 static void
arc_kstat_update_state(arc_state_t * state,kstat_named_t * size,kstat_named_t * evict_data,kstat_named_t * evict_metadata)6860 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6861     kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6862 {
6863 	size->value.ui64 = zfs_refcount_count(&state->arcs_size);
6864 	evict_data->value.ui64 =
6865 	    zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6866 	evict_metadata->value.ui64 =
6867 	    zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6868 }
6869 
6870 static int
arc_kstat_update(kstat_t * ksp,int rw)6871 arc_kstat_update(kstat_t *ksp, int rw)
6872 {
6873 	arc_stats_t *as = ksp->ks_data;
6874 
6875 	if (rw == KSTAT_WRITE) {
6876 		return (EACCES);
6877 	} else {
6878 		arc_kstat_update_state(arc_anon,
6879 		    &as->arcstat_anon_size,
6880 		    &as->arcstat_anon_evictable_data,
6881 		    &as->arcstat_anon_evictable_metadata);
6882 		arc_kstat_update_state(arc_mru,
6883 		    &as->arcstat_mru_size,
6884 		    &as->arcstat_mru_evictable_data,
6885 		    &as->arcstat_mru_evictable_metadata);
6886 		arc_kstat_update_state(arc_mru_ghost,
6887 		    &as->arcstat_mru_ghost_size,
6888 		    &as->arcstat_mru_ghost_evictable_data,
6889 		    &as->arcstat_mru_ghost_evictable_metadata);
6890 		arc_kstat_update_state(arc_mfu,
6891 		    &as->arcstat_mfu_size,
6892 		    &as->arcstat_mfu_evictable_data,
6893 		    &as->arcstat_mfu_evictable_metadata);
6894 		arc_kstat_update_state(arc_mfu_ghost,
6895 		    &as->arcstat_mfu_ghost_size,
6896 		    &as->arcstat_mfu_ghost_evictable_data,
6897 		    &as->arcstat_mfu_ghost_evictable_metadata);
6898 
6899 		ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6900 		ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6901 		ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6902 		ARCSTAT(arcstat_metadata_size) =
6903 		    aggsum_value(&astat_metadata_size);
6904 		ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6905 		ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size);
6906 		ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6907 	}
6908 
6909 	return (0);
6910 }
6911 
6912 /*
6913  * This function *must* return indices evenly distributed between all
6914  * sublists of the multilist. This is needed due to how the ARC eviction
6915  * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6916  * distributed between all sublists and uses this assumption when
6917  * deciding which sublist to evict from and how much to evict from it.
6918  */
6919 unsigned int
arc_state_multilist_index_func(multilist_t * ml,void * obj)6920 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6921 {
6922 	arc_buf_hdr_t *hdr = obj;
6923 
6924 	/*
6925 	 * We rely on b_dva to generate evenly distributed index
6926 	 * numbers using buf_hash below. So, as an added precaution,
6927 	 * let's make sure we never add empty buffers to the arc lists.
6928 	 */
6929 	ASSERT(!HDR_EMPTY(hdr));
6930 
6931 	/*
6932 	 * The assumption here, is the hash value for a given
6933 	 * arc_buf_hdr_t will remain constant throughout its lifetime
6934 	 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
6935 	 * Thus, we don't need to store the header's sublist index
6936 	 * on insertion, as this index can be recalculated on removal.
6937 	 *
6938 	 * Also, the low order bits of the hash value are thought to be
6939 	 * distributed evenly. Otherwise, in the case that the multilist
6940 	 * has a power of two number of sublists, each sublists' usage
6941 	 * would not be evenly distributed.
6942 	 */
6943 	return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6944 	    multilist_get_num_sublists(ml));
6945 }
6946 
6947 static void
arc_state_init(void)6948 arc_state_init(void)
6949 {
6950 	arc_anon = &ARC_anon;
6951 	arc_mru = &ARC_mru;
6952 	arc_mru_ghost = &ARC_mru_ghost;
6953 	arc_mfu = &ARC_mfu;
6954 	arc_mfu_ghost = &ARC_mfu_ghost;
6955 	arc_l2c_only = &ARC_l2c_only;
6956 
6957 	arc_mru->arcs_list[ARC_BUFC_METADATA] =
6958 	    multilist_create(sizeof (arc_buf_hdr_t),
6959 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6960 	    arc_state_multilist_index_func);
6961 	arc_mru->arcs_list[ARC_BUFC_DATA] =
6962 	    multilist_create(sizeof (arc_buf_hdr_t),
6963 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6964 	    arc_state_multilist_index_func);
6965 	arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6966 	    multilist_create(sizeof (arc_buf_hdr_t),
6967 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6968 	    arc_state_multilist_index_func);
6969 	arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6970 	    multilist_create(sizeof (arc_buf_hdr_t),
6971 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6972 	    arc_state_multilist_index_func);
6973 	arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6974 	    multilist_create(sizeof (arc_buf_hdr_t),
6975 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6976 	    arc_state_multilist_index_func);
6977 	arc_mfu->arcs_list[ARC_BUFC_DATA] =
6978 	    multilist_create(sizeof (arc_buf_hdr_t),
6979 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6980 	    arc_state_multilist_index_func);
6981 	arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6982 	    multilist_create(sizeof (arc_buf_hdr_t),
6983 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6984 	    arc_state_multilist_index_func);
6985 	arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6986 	    multilist_create(sizeof (arc_buf_hdr_t),
6987 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6988 	    arc_state_multilist_index_func);
6989 	arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6990 	    multilist_create(sizeof (arc_buf_hdr_t),
6991 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6992 	    arc_state_multilist_index_func);
6993 	arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6994 	    multilist_create(sizeof (arc_buf_hdr_t),
6995 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6996 	    arc_state_multilist_index_func);
6997 
6998 	zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6999 	zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7000 	zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7001 	zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7002 	zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7003 	zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7004 	zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7005 	zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7006 	zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7007 	zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7008 	zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7009 	zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7010 
7011 	zfs_refcount_create(&arc_anon->arcs_size);
7012 	zfs_refcount_create(&arc_mru->arcs_size);
7013 	zfs_refcount_create(&arc_mru_ghost->arcs_size);
7014 	zfs_refcount_create(&arc_mfu->arcs_size);
7015 	zfs_refcount_create(&arc_mfu_ghost->arcs_size);
7016 	zfs_refcount_create(&arc_l2c_only->arcs_size);
7017 
7018 	aggsum_init(&arc_meta_used, 0);
7019 	aggsum_init(&arc_size, 0);
7020 	aggsum_init(&astat_data_size, 0);
7021 	aggsum_init(&astat_metadata_size, 0);
7022 	aggsum_init(&astat_hdr_size, 0);
7023 	aggsum_init(&astat_other_size, 0);
7024 	aggsum_init(&astat_l2_hdr_size, 0);
7025 
7026 	arc_anon->arcs_state = ARC_STATE_ANON;
7027 	arc_mru->arcs_state = ARC_STATE_MRU;
7028 	arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7029 	arc_mfu->arcs_state = ARC_STATE_MFU;
7030 	arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7031 	arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7032 }
7033 
7034 static void
arc_state_fini(void)7035 arc_state_fini(void)
7036 {
7037 	zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7038 	zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7039 	zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7040 	zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7041 	zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7042 	zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7043 	zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7044 	zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7045 	zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7046 	zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7047 	zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7048 	zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7049 
7050 	zfs_refcount_destroy(&arc_anon->arcs_size);
7051 	zfs_refcount_destroy(&arc_mru->arcs_size);
7052 	zfs_refcount_destroy(&arc_mru_ghost->arcs_size);
7053 	zfs_refcount_destroy(&arc_mfu->arcs_size);
7054 	zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
7055 	zfs_refcount_destroy(&arc_l2c_only->arcs_size);
7056 
7057 	multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
7058 	multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7059 	multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7060 	multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7061 	multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
7062 	multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7063 	multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
7064 	multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7065 	multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7066 	multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7067 
7068 	aggsum_fini(&arc_meta_used);
7069 	aggsum_fini(&arc_size);
7070 	aggsum_fini(&astat_data_size);
7071 	aggsum_fini(&astat_metadata_size);
7072 	aggsum_fini(&astat_hdr_size);
7073 	aggsum_fini(&astat_other_size);
7074 	aggsum_fini(&astat_l2_hdr_size);
7075 
7076 }
7077 
7078 uint64_t
arc_max_bytes(void)7079 arc_max_bytes(void)
7080 {
7081 	return (arc_c_max);
7082 }
7083 
7084 void
arc_init(void)7085 arc_init(void)
7086 {
7087 	/*
7088 	 * allmem is "all memory that we could possibly use".
7089 	 */
7090 #ifdef _KERNEL
7091 	uint64_t allmem = ptob(physmem - swapfs_minfree);
7092 #else
7093 	uint64_t allmem = (physmem * PAGESIZE) / 2;
7094 #endif
7095 	mutex_init(&arc_adjust_lock, NULL, MUTEX_DEFAULT, NULL);
7096 	cv_init(&arc_adjust_waiters_cv, NULL, CV_DEFAULT, NULL);
7097 
7098 	/*
7099 	 * Set the minimum cache size to 1/64 of all memory, with a hard
7100 	 * minimum of 64MB.
7101 	 */
7102 	arc_c_min = MAX(allmem / 64, 64 << 20);
7103 	/*
7104 	 * In a system with a lot of physical memory this will still result in
7105 	 * an ARC size floor that is quite large in absolute terms.  Cap the
7106 	 * growth of the value at 1GB.
7107 	 */
7108 	arc_c_min = MIN(arc_c_min, 1 << 30);
7109 
7110 	/* set max to 3/4 of all memory, or all but 1GB, whichever is more */
7111 	if (allmem >= 1 << 30)
7112 		arc_c_max = allmem - (1 << 30);
7113 	else
7114 		arc_c_max = arc_c_min;
7115 	arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
7116 
7117 	/*
7118 	 * In userland, there's only the memory pressure that we artificially
7119 	 * create (see arc_available_memory()).  Don't let arc_c get too
7120 	 * small, because it can cause transactions to be larger than
7121 	 * arc_c, causing arc_tempreserve_space() to fail.
7122 	 */
7123 #ifndef _KERNEL
7124 	arc_c_min = arc_c_max / 2;
7125 #endif
7126 
7127 	/*
7128 	 * Allow the tunables to override our calculations if they are
7129 	 * reasonable (ie. over 64MB)
7130 	 */
7131 	if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) {
7132 		arc_c_max = zfs_arc_max;
7133 		arc_c_min = MIN(arc_c_min, arc_c_max);
7134 	}
7135 	if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
7136 		arc_c_min = zfs_arc_min;
7137 
7138 	arc_c = arc_c_max;
7139 	arc_p = (arc_c >> 1);
7140 
7141 	/* limit meta-data to 1/4 of the arc capacity */
7142 	arc_meta_limit = arc_c_max / 4;
7143 
7144 #ifdef _KERNEL
7145 	/*
7146 	 * Metadata is stored in the kernel's heap.  Don't let us
7147 	 * use more than half the heap for the ARC.
7148 	 */
7149 	arc_meta_limit = MIN(arc_meta_limit,
7150 	    vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
7151 #endif
7152 
7153 	/* Allow the tunable to override if it is reasonable */
7154 	if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
7155 		arc_meta_limit = zfs_arc_meta_limit;
7156 
7157 	if (zfs_arc_meta_min > 0) {
7158 		arc_meta_min = zfs_arc_meta_min;
7159 	} else {
7160 		arc_meta_min = arc_c_min / 2;
7161 	}
7162 
7163 	if (zfs_arc_grow_retry > 0)
7164 		arc_grow_retry = zfs_arc_grow_retry;
7165 
7166 	if (zfs_arc_shrink_shift > 0)
7167 		arc_shrink_shift = zfs_arc_shrink_shift;
7168 
7169 	/*
7170 	 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
7171 	 */
7172 	if (arc_no_grow_shift >= arc_shrink_shift)
7173 		arc_no_grow_shift = arc_shrink_shift - 1;
7174 
7175 	if (zfs_arc_p_min_shift > 0)
7176 		arc_p_min_shift = zfs_arc_p_min_shift;
7177 
7178 	/* if kmem_flags are set, lets try to use less memory */
7179 	if (kmem_debugging())
7180 		arc_c = arc_c / 2;
7181 	if (arc_c < arc_c_min)
7182 		arc_c = arc_c_min;
7183 
7184 	arc_state_init();
7185 
7186 	/*
7187 	 * The arc must be "uninitialized", so that hdr_recl() (which is
7188 	 * registered by buf_init()) will not access arc_reap_zthr before
7189 	 * it is created.
7190 	 */
7191 	ASSERT(!arc_initialized);
7192 	buf_init();
7193 
7194 	arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7195 	    sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7196 
7197 	if (arc_ksp != NULL) {
7198 		arc_ksp->ks_data = &arc_stats;
7199 		arc_ksp->ks_update = arc_kstat_update;
7200 		kstat_install(arc_ksp);
7201 	}
7202 
7203 	arc_adjust_zthr = zthr_create(arc_adjust_cb_check,
7204 	    arc_adjust_cb, NULL);
7205 	arc_reap_zthr = zthr_create_timer(arc_reap_cb_check,
7206 	    arc_reap_cb, NULL, SEC2NSEC(1));
7207 
7208 	arc_initialized = B_TRUE;
7209 	arc_warm = B_FALSE;
7210 
7211 	/*
7212 	 * Calculate maximum amount of dirty data per pool.
7213 	 *
7214 	 * If it has been set by /etc/system, take that.
7215 	 * Otherwise, use a percentage of physical memory defined by
7216 	 * zfs_dirty_data_max_percent (default 10%) with a cap at
7217 	 * zfs_dirty_data_max_max (default 4GB).
7218 	 */
7219 	if (zfs_dirty_data_max == 0) {
7220 		zfs_dirty_data_max = physmem * PAGESIZE *
7221 		    zfs_dirty_data_max_percent / 100;
7222 		zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7223 		    zfs_dirty_data_max_max);
7224 	}
7225 }
7226 
7227 void
arc_fini(void)7228 arc_fini(void)
7229 {
7230 	/* Use B_TRUE to ensure *all* buffers are evicted */
7231 	arc_flush(NULL, B_TRUE);
7232 
7233 	arc_initialized = B_FALSE;
7234 
7235 	if (arc_ksp != NULL) {
7236 		kstat_delete(arc_ksp);
7237 		arc_ksp = NULL;
7238 	}
7239 
7240 	(void) zthr_cancel(arc_adjust_zthr);
7241 	zthr_destroy(arc_adjust_zthr);
7242 
7243 	(void) zthr_cancel(arc_reap_zthr);
7244 	zthr_destroy(arc_reap_zthr);
7245 
7246 	mutex_destroy(&arc_adjust_lock);
7247 	cv_destroy(&arc_adjust_waiters_cv);
7248 
7249 	/*
7250 	 * buf_fini() must proceed arc_state_fini() because buf_fin() may
7251 	 * trigger the release of kmem magazines, which can callback to
7252 	 * arc_space_return() which accesses aggsums freed in act_state_fini().
7253 	 */
7254 	buf_fini();
7255 	arc_state_fini();
7256 
7257 	ASSERT0(arc_loaned_bytes);
7258 }
7259 
7260 /*
7261  * Level 2 ARC
7262  *
7263  * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7264  * It uses dedicated storage devices to hold cached data, which are populated
7265  * using large infrequent writes.  The main role of this cache is to boost
7266  * the performance of random read workloads.  The intended L2ARC devices
7267  * include short-stroked disks, solid state disks, and other media with
7268  * substantially faster read latency than disk.
7269  *
7270  *                 +-----------------------+
7271  *                 |         ARC           |
7272  *                 +-----------------------+
7273  *                    |         ^     ^
7274  *                    |         |     |
7275  *      l2arc_feed_thread()    arc_read()
7276  *                    |         |     |
7277  *                    |  l2arc read   |
7278  *                    V         |     |
7279  *               +---------------+    |
7280  *               |     L2ARC     |    |
7281  *               +---------------+    |
7282  *                   |    ^           |
7283  *          l2arc_write() |           |
7284  *                   |    |           |
7285  *                   V    |           |
7286  *                 +-------+      +-------+
7287  *                 | vdev  |      | vdev  |
7288  *                 | cache |      | cache |
7289  *                 +-------+      +-------+
7290  *                 +=========+     .-----.
7291  *                 :  L2ARC  :    |-_____-|
7292  *                 : devices :    | Disks |
7293  *                 +=========+    `-_____-'
7294  *
7295  * Read requests are satisfied from the following sources, in order:
7296  *
7297  *	1) ARC
7298  *	2) vdev cache of L2ARC devices
7299  *	3) L2ARC devices
7300  *	4) vdev cache of disks
7301  *	5) disks
7302  *
7303  * Some L2ARC device types exhibit extremely slow write performance.
7304  * To accommodate for this there are some significant differences between
7305  * the L2ARC and traditional cache design:
7306  *
7307  * 1. There is no eviction path from the ARC to the L2ARC.  Evictions from
7308  * the ARC behave as usual, freeing buffers and placing headers on ghost
7309  * lists.  The ARC does not send buffers to the L2ARC during eviction as
7310  * this would add inflated write latencies for all ARC memory pressure.
7311  *
7312  * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7313  * It does this by periodically scanning buffers from the eviction-end of
7314  * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7315  * not already there. It scans until a headroom of buffers is satisfied,
7316  * which itself is a buffer for ARC eviction. If a compressible buffer is
7317  * found during scanning and selected for writing to an L2ARC device, we
7318  * temporarily boost scanning headroom during the next scan cycle to make
7319  * sure we adapt to compression effects (which might significantly reduce
7320  * the data volume we write to L2ARC). The thread that does this is
7321  * l2arc_feed_thread(), illustrated below; example sizes are included to
7322  * provide a better sense of ratio than this diagram:
7323  *
7324  *	       head -->                        tail
7325  *	        +---------------------+----------+
7326  *	ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->.   # already on L2ARC
7327  *	        +---------------------+----------+   |   o L2ARC eligible
7328  *	ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->|   : ARC buffer
7329  *	        +---------------------+----------+   |
7330  *	             15.9 Gbytes      ^ 32 Mbytes    |
7331  *	                           headroom          |
7332  *	                                      l2arc_feed_thread()
7333  *	                                             |
7334  *	                 l2arc write hand <--[oooo]--'
7335  *	                         |           8 Mbyte
7336  *	                         |          write max
7337  *	                         V
7338  *		  +==============================+
7339  *	L2ARC dev |####|#|###|###|    |####| ... |
7340  *	          +==============================+
7341  *	                     32 Gbytes
7342  *
7343  * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7344  * evicted, then the L2ARC has cached a buffer much sooner than it probably
7345  * needed to, potentially wasting L2ARC device bandwidth and storage.  It is
7346  * safe to say that this is an uncommon case, since buffers at the end of
7347  * the ARC lists have moved there due to inactivity.
7348  *
7349  * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7350  * then the L2ARC simply misses copying some buffers.  This serves as a
7351  * pressure valve to prevent heavy read workloads from both stalling the ARC
7352  * with waits and clogging the L2ARC with writes.  This also helps prevent
7353  * the potential for the L2ARC to churn if it attempts to cache content too
7354  * quickly, such as during backups of the entire pool.
7355  *
7356  * 5. After system boot and before the ARC has filled main memory, there are
7357  * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7358  * lists can remain mostly static.  Instead of searching from tail of these
7359  * lists as pictured, the l2arc_feed_thread() will search from the list heads
7360  * for eligible buffers, greatly increasing its chance of finding them.
7361  *
7362  * The L2ARC device write speed is also boosted during this time so that
7363  * the L2ARC warms up faster.  Since there have been no ARC evictions yet,
7364  * there are no L2ARC reads, and no fear of degrading read performance
7365  * through increased writes.
7366  *
7367  * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7368  * the vdev queue can aggregate them into larger and fewer writes.  Each
7369  * device is written to in a rotor fashion, sweeping writes through
7370  * available space then repeating.
7371  *
7372  * 7. The L2ARC does not store dirty content.  It never needs to flush
7373  * write buffers back to disk based storage.
7374  *
7375  * 8. If an ARC buffer is written (and dirtied) which also exists in the
7376  * L2ARC, the now stale L2ARC buffer is immediately dropped.
7377  *
7378  * The performance of the L2ARC can be tweaked by a number of tunables, which
7379  * may be necessary for different workloads:
7380  *
7381  *	l2arc_write_max		max write bytes per interval
7382  *	l2arc_write_boost	extra write bytes during device warmup
7383  *	l2arc_noprefetch	skip caching prefetched buffers
7384  *	l2arc_headroom		number of max device writes to precache
7385  *	l2arc_headroom_boost	when we find compressed buffers during ARC
7386  *				scanning, we multiply headroom by this
7387  *				percentage factor for the next scan cycle,
7388  *				since more compressed buffers are likely to
7389  *				be present
7390  *	l2arc_feed_secs		seconds between L2ARC writing
7391  *
7392  * Tunables may be removed or added as future performance improvements are
7393  * integrated, and also may become zpool properties.
7394  *
7395  * There are three key functions that control how the L2ARC warms up:
7396  *
7397  *	l2arc_write_eligible()	check if a buffer is eligible to cache
7398  *	l2arc_write_size()	calculate how much to write
7399  *	l2arc_write_interval()	calculate sleep delay between writes
7400  *
7401  * These three functions determine what to write, how much, and how quickly
7402  * to send writes.
7403  *
7404  * L2ARC persistence:
7405  *
7406  * When writing buffers to L2ARC, we periodically add some metadata to
7407  * make sure we can pick them up after reboot, thus dramatically reducing
7408  * the impact that any downtime has on the performance of storage systems
7409  * with large caches.
7410  *
7411  * The implementation works fairly simply by integrating the following two
7412  * modifications:
7413  *
7414  * *) When writing to the L2ARC, we occasionally write a "l2arc log block",
7415  *    which is an additional piece of metadata which describes what's been
7416  *    written. This allows us to rebuild the arc_buf_hdr_t structures of the
7417  *    main ARC buffers. There are 2 linked-lists of log blocks headed by
7418  *    dh_start_lbps[2]. We alternate which chain we append to, so they are
7419  *    time-wise and offset-wise interleaved, but that is an optimization rather
7420  *    than for correctness. The log block also includes a pointer to the
7421  *    previous block in its chain.
7422  *
7423  * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
7424  *    for our header bookkeeping purposes. This contains a device header,
7425  *    which contains our top-level reference structures. We update it each
7426  *    time we write a new log block, so that we're able to locate it in the
7427  *    L2ARC device. If this write results in an inconsistent device header
7428  *    (e.g. due to power failure), we detect this by verifying the header's
7429  *    checksum and simply fail to reconstruct the L2ARC after reboot.
7430  *
7431  * Implementation diagram:
7432  *
7433  * +=== L2ARC device (not to scale) ======================================+
7434  * |       ___two newest log block pointers__.__________                  |
7435  * |      /                                   \dh_start_lbps[1]           |
7436  * |	 /				       \         \dh_start_lbps[0]|
7437  * |.___/__.                                    V         V               |
7438  * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
7439  * ||   hdr|      ^         /^       /^        /         /                |
7440  * |+------+  ...--\-------/  \-----/--\------/         /                 |
7441  * |                \--------------/    \--------------/                  |
7442  * +======================================================================+
7443  *
7444  * As can be seen on the diagram, rather than using a simple linked list,
7445  * we use a pair of linked lists with alternating elements. This is a
7446  * performance enhancement due to the fact that we only find out the
7447  * address of the next log block access once the current block has been
7448  * completely read in. Obviously, this hurts performance, because we'd be
7449  * keeping the device's I/O queue at only a 1 operation deep, thus
7450  * incurring a large amount of I/O round-trip latency. Having two lists
7451  * allows us to fetch two log blocks ahead of where we are currently
7452  * rebuilding L2ARC buffers.
7453  *
7454  * On-device data structures:
7455  *
7456  * L2ARC device header:	l2arc_dev_hdr_phys_t
7457  * L2ARC log block:	l2arc_log_blk_phys_t
7458  *
7459  * L2ARC reconstruction:
7460  *
7461  * When writing data, we simply write in the standard rotary fashion,
7462  * evicting buffers as we go and simply writing new data over them (writing
7463  * a new log block every now and then). This obviously means that once we
7464  * loop around the end of the device, we will start cutting into an already
7465  * committed log block (and its referenced data buffers), like so:
7466  *
7467  *    current write head__       __old tail
7468  *                        \     /
7469  *                        V    V
7470  * <--|bufs |lb |bufs |lb |    |bufs |lb |bufs |lb |-->
7471  *                         ^    ^^^^^^^^^___________________________________
7472  *                         |                                                \
7473  *                   <<nextwrite>> may overwrite this blk and/or its bufs --'
7474  *
7475  * When importing the pool, we detect this situation and use it to stop
7476  * our scanning process (see l2arc_rebuild).
7477  *
7478  * There is one significant caveat to consider when rebuilding ARC contents
7479  * from an L2ARC device: what about invalidated buffers? Given the above
7480  * construction, we cannot update blocks which we've already written to amend
7481  * them to remove buffers which were invalidated. Thus, during reconstruction,
7482  * we might be populating the cache with buffers for data that's not on the
7483  * main pool anymore, or may have been overwritten!
7484  *
7485  * As it turns out, this isn't a problem. Every arc_read request includes
7486  * both the DVA and, crucially, the birth TXG of the BP the caller is
7487  * looking for. So even if the cache were populated by completely rotten
7488  * blocks for data that had been long deleted and/or overwritten, we'll
7489  * never actually return bad data from the cache, since the DVA with the
7490  * birth TXG uniquely identify a block in space and time - once created,
7491  * a block is immutable on disk. The worst thing we have done is wasted
7492  * some time and memory at l2arc rebuild to reconstruct outdated ARC
7493  * entries that will get dropped from the l2arc as it is being updated
7494  * with new blocks.
7495  *
7496  * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write
7497  * hand are not restored. This is done by saving the offset (in bytes)
7498  * l2arc_evict() has evicted to in the L2ARC device header and taking it
7499  * into account when restoring buffers.
7500  */
7501 
7502 static boolean_t
l2arc_write_eligible(uint64_t spa_guid,arc_buf_hdr_t * hdr)7503 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7504 {
7505 	/*
7506 	 * A buffer is *not* eligible for the L2ARC if it:
7507 	 * 1. belongs to a different spa.
7508 	 * 2. is already cached on the L2ARC.
7509 	 * 3. has an I/O in progress (it may be an incomplete read).
7510 	 * 4. is flagged not eligible (zfs property).
7511 	 * 5. is a prefetch and l2arc_noprefetch is set.
7512 	 */
7513 	if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
7514 	    HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr) ||
7515 	    (l2arc_noprefetch && HDR_PREFETCH(hdr)))
7516 		return (B_FALSE);
7517 
7518 	return (B_TRUE);
7519 }
7520 
7521 static uint64_t
l2arc_write_size(l2arc_dev_t * dev)7522 l2arc_write_size(l2arc_dev_t *dev)
7523 {
7524 	uint64_t size, dev_size;
7525 
7526 	/*
7527 	 * Make sure our globals have meaningful values in case the user
7528 	 * altered them.
7529 	 */
7530 	size = l2arc_write_max;
7531 	if (size == 0) {
7532 		cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7533 		    "be greater than zero, resetting it to the default (%d)",
7534 		    L2ARC_WRITE_SIZE);
7535 		size = l2arc_write_max = L2ARC_WRITE_SIZE;
7536 	}
7537 
7538 	if (arc_warm == B_FALSE)
7539 		size += l2arc_write_boost;
7540 
7541 	/*
7542 	 * Make sure the write size does not exceed the size of the cache
7543 	 * device. This is important in l2arc_evict(), otherwise infinite
7544 	 * iteration can occur.
7545 	 */
7546 	dev_size = dev->l2ad_end - dev->l2ad_start;
7547 	if ((size + l2arc_log_blk_overhead(size, dev)) >= dev_size) {
7548 		cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost "
7549 		    "plus the overhead of log blocks (persistent L2ARC, "
7550 		    "%" PRIu64 " bytes) exceeds the size of the cache device "
7551 		    "(guid %" PRIu64 "), resetting them to the default (%d)",
7552 		    l2arc_log_blk_overhead(size, dev),
7553 		    dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE);
7554 		size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE;
7555 
7556 		if (arc_warm == B_FALSE)
7557 			size += l2arc_write_boost;
7558 	}
7559 
7560 	return (size);
7561 
7562 }
7563 
7564 static clock_t
l2arc_write_interval(clock_t began,uint64_t wanted,uint64_t wrote)7565 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7566 {
7567 	clock_t interval, next, now;
7568 
7569 	/*
7570 	 * If the ARC lists are busy, increase our write rate; if the
7571 	 * lists are stale, idle back.  This is achieved by checking
7572 	 * how much we previously wrote - if it was more than half of
7573 	 * what we wanted, schedule the next write much sooner.
7574 	 */
7575 	if (l2arc_feed_again && wrote > (wanted / 2))
7576 		interval = (hz * l2arc_feed_min_ms) / 1000;
7577 	else
7578 		interval = hz * l2arc_feed_secs;
7579 
7580 	now = ddi_get_lbolt();
7581 	next = MAX(now, MIN(now + interval, began + interval));
7582 
7583 	return (next);
7584 }
7585 
7586 /*
7587  * Cycle through L2ARC devices.  This is how L2ARC load balances.
7588  * If a device is returned, this also returns holding the spa config lock.
7589  */
7590 static l2arc_dev_t *
l2arc_dev_get_next(void)7591 l2arc_dev_get_next(void)
7592 {
7593 	l2arc_dev_t *first, *next = NULL;
7594 
7595 	/*
7596 	 * Lock out the removal of spas (spa_namespace_lock), then removal
7597 	 * of cache devices (l2arc_dev_mtx).  Once a device has been selected,
7598 	 * both locks will be dropped and a spa config lock held instead.
7599 	 */
7600 	mutex_enter(&spa_namespace_lock);
7601 	mutex_enter(&l2arc_dev_mtx);
7602 
7603 	/* if there are no vdevs, there is nothing to do */
7604 	if (l2arc_ndev == 0)
7605 		goto out;
7606 
7607 	first = NULL;
7608 	next = l2arc_dev_last;
7609 	do {
7610 		/* loop around the list looking for a non-faulted vdev */
7611 		if (next == NULL) {
7612 			next = list_head(l2arc_dev_list);
7613 		} else {
7614 			next = list_next(l2arc_dev_list, next);
7615 			if (next == NULL)
7616 				next = list_head(l2arc_dev_list);
7617 		}
7618 
7619 		/* if we have come back to the start, bail out */
7620 		if (first == NULL)
7621 			first = next;
7622 		else if (next == first)
7623 			break;
7624 
7625 	} while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild);
7626 
7627 	/* if we were unable to find any usable vdevs, return NULL */
7628 	if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild)
7629 		next = NULL;
7630 
7631 	l2arc_dev_last = next;
7632 
7633 out:
7634 	mutex_exit(&l2arc_dev_mtx);
7635 
7636 	/*
7637 	 * Grab the config lock to prevent the 'next' device from being
7638 	 * removed while we are writing to it.
7639 	 */
7640 	if (next != NULL)
7641 		spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7642 	mutex_exit(&spa_namespace_lock);
7643 
7644 	return (next);
7645 }
7646 
7647 /*
7648  * Free buffers that were tagged for destruction.
7649  */
7650 static void
l2arc_do_free_on_write()7651 l2arc_do_free_on_write()
7652 {
7653 	list_t *buflist;
7654 	l2arc_data_free_t *df, *df_prev;
7655 
7656 	mutex_enter(&l2arc_free_on_write_mtx);
7657 	buflist = l2arc_free_on_write;
7658 
7659 	for (df = list_tail(buflist); df; df = df_prev) {
7660 		df_prev = list_prev(buflist, df);
7661 		ASSERT3P(df->l2df_abd, !=, NULL);
7662 		abd_free(df->l2df_abd);
7663 		list_remove(buflist, df);
7664 		kmem_free(df, sizeof (l2arc_data_free_t));
7665 	}
7666 
7667 	mutex_exit(&l2arc_free_on_write_mtx);
7668 }
7669 
7670 /*
7671  * A write to a cache device has completed.  Update all headers to allow
7672  * reads from these buffers to begin.
7673  */
7674 static void
l2arc_write_done(zio_t * zio)7675 l2arc_write_done(zio_t *zio)
7676 {
7677 	l2arc_write_callback_t	*cb;
7678 	l2arc_lb_abd_buf_t	*abd_buf;
7679 	l2arc_lb_ptr_buf_t	*lb_ptr_buf;
7680 	l2arc_dev_t		*dev;
7681 	l2arc_dev_hdr_phys_t	*l2dhdr;
7682 	list_t			*buflist;
7683 	arc_buf_hdr_t		*head, *hdr, *hdr_prev;
7684 	kmutex_t		*hash_lock;
7685 	int64_t			bytes_dropped = 0;
7686 
7687 	cb = zio->io_private;
7688 	ASSERT3P(cb, !=, NULL);
7689 	dev = cb->l2wcb_dev;
7690 	l2dhdr = dev->l2ad_dev_hdr;
7691 	ASSERT3P(dev, !=, NULL);
7692 	head = cb->l2wcb_head;
7693 	ASSERT3P(head, !=, NULL);
7694 	buflist = &dev->l2ad_buflist;
7695 	ASSERT3P(buflist, !=, NULL);
7696 	DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7697 	    l2arc_write_callback_t *, cb);
7698 
7699 	/*
7700 	 * All writes completed, or an error was hit.
7701 	 */
7702 top:
7703 	mutex_enter(&dev->l2ad_mtx);
7704 	for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7705 		hdr_prev = list_prev(buflist, hdr);
7706 
7707 		hash_lock = HDR_LOCK(hdr);
7708 
7709 		/*
7710 		 * We cannot use mutex_enter or else we can deadlock
7711 		 * with l2arc_write_buffers (due to swapping the order
7712 		 * the hash lock and l2ad_mtx are taken).
7713 		 */
7714 		if (!mutex_tryenter(hash_lock)) {
7715 			/*
7716 			 * Missed the hash lock. We must retry so we
7717 			 * don't leave the ARC_FLAG_L2_WRITING bit set.
7718 			 */
7719 			ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7720 
7721 			/*
7722 			 * We don't want to rescan the headers we've
7723 			 * already marked as having been written out, so
7724 			 * we reinsert the head node so we can pick up
7725 			 * where we left off.
7726 			 */
7727 			list_remove(buflist, head);
7728 			list_insert_after(buflist, hdr, head);
7729 
7730 			mutex_exit(&dev->l2ad_mtx);
7731 
7732 			/*
7733 			 * We wait for the hash lock to become available
7734 			 * to try and prevent busy waiting, and increase
7735 			 * the chance we'll be able to acquire the lock
7736 			 * the next time around.
7737 			 */
7738 			mutex_enter(hash_lock);
7739 			mutex_exit(hash_lock);
7740 			goto top;
7741 		}
7742 
7743 		/*
7744 		 * We could not have been moved into the arc_l2c_only
7745 		 * state while in-flight due to our ARC_FLAG_L2_WRITING
7746 		 * bit being set. Let's just ensure that's being enforced.
7747 		 */
7748 		ASSERT(HDR_HAS_L1HDR(hdr));
7749 
7750 		if (zio->io_error != 0) {
7751 			/*
7752 			 * Error - drop L2ARC entry.
7753 			 */
7754 			list_remove(buflist, hdr);
7755 			arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7756 
7757 			uint64_t psize = HDR_GET_PSIZE(hdr);
7758 			l2arc_hdr_arcstats_decrement(hdr);
7759 
7760 			bytes_dropped +=
7761 			    vdev_psize_to_asize(dev->l2ad_vdev, psize);
7762 			(void) zfs_refcount_remove_many(&dev->l2ad_alloc,
7763 			    arc_hdr_size(hdr), hdr);
7764 		}
7765 
7766 		/*
7767 		 * Allow ARC to begin reads and ghost list evictions to
7768 		 * this L2ARC entry.
7769 		 */
7770 		arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7771 
7772 		mutex_exit(hash_lock);
7773 	}
7774 
7775 	/*
7776 	 * Free the allocated abd buffers for writing the log blocks.
7777 	 * If the zio failed reclaim the allocated space and remove the
7778 	 * pointers to these log blocks from the log block pointer list
7779 	 * of the L2ARC device.
7780 	 */
7781 	while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) {
7782 		abd_free(abd_buf->abd);
7783 		zio_buf_free(abd_buf, sizeof (*abd_buf));
7784 		if (zio->io_error != 0) {
7785 			lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list);
7786 			/*
7787 			 * L2BLK_GET_PSIZE returns aligned size for log
7788 			 * blocks.
7789 			 */
7790 			uint64_t asize =
7791 			    L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop);
7792 			bytes_dropped += asize;
7793 			ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
7794 			ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
7795 			zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
7796 			    lb_ptr_buf);
7797 			zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
7798 			kmem_free(lb_ptr_buf->lb_ptr,
7799 			    sizeof (l2arc_log_blkptr_t));
7800 			kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
7801 		}
7802 	}
7803 	list_destroy(&cb->l2wcb_abd_list);
7804 
7805 	if (zio->io_error != 0) {
7806 		ARCSTAT_BUMP(arcstat_l2_writes_error);
7807 
7808 		/*
7809 		 * Restore the lbps array in the header to its previous state.
7810 		 * If the list of log block pointers is empty, zero out the
7811 		 * log block pointers in the device header.
7812 		 */
7813 		lb_ptr_buf = list_head(&dev->l2ad_lbptr_list);
7814 		for (int i = 0; i < 2; i++) {
7815 			if (lb_ptr_buf == NULL) {
7816 				/*
7817 				 * If the list is empty zero out the device
7818 				 * header. Otherwise zero out the second log
7819 				 * block pointer in the header.
7820 				 */
7821 				if (i == 0) {
7822 					bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
7823 				} else {
7824 					bzero(&l2dhdr->dh_start_lbps[i],
7825 					    sizeof (l2arc_log_blkptr_t));
7826 				}
7827 				break;
7828 			}
7829 			bcopy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[i],
7830 			    sizeof (l2arc_log_blkptr_t));
7831 			lb_ptr_buf = list_next(&dev->l2ad_lbptr_list,
7832 			    lb_ptr_buf);
7833 		}
7834 	}
7835 
7836 	atomic_inc_64(&l2arc_writes_done);
7837 	list_remove(buflist, head);
7838 	ASSERT(!HDR_HAS_L1HDR(head));
7839 	kmem_cache_free(hdr_l2only_cache, head);
7840 	mutex_exit(&dev->l2ad_mtx);
7841 
7842 	ASSERT(dev->l2ad_vdev != NULL);
7843 	vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7844 
7845 	l2arc_do_free_on_write();
7846 
7847 	kmem_free(cb, sizeof (l2arc_write_callback_t));
7848 }
7849 
7850 static int
l2arc_untransform(zio_t * zio,l2arc_read_callback_t * cb)7851 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
7852 {
7853 	int ret;
7854 	spa_t *spa = zio->io_spa;
7855 	arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
7856 	blkptr_t *bp = zio->io_bp;
7857 	uint8_t salt[ZIO_DATA_SALT_LEN];
7858 	uint8_t iv[ZIO_DATA_IV_LEN];
7859 	uint8_t mac[ZIO_DATA_MAC_LEN];
7860 	boolean_t no_crypt = B_FALSE;
7861 
7862 	/*
7863 	 * ZIL data is never be written to the L2ARC, so we don't need
7864 	 * special handling for its unique MAC storage.
7865 	 */
7866 	ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
7867 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
7868 	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7869 
7870 	/*
7871 	 * If the data was encrypted, decrypt it now. Note that
7872 	 * we must check the bp here and not the hdr, since the
7873 	 * hdr does not have its encryption parameters updated
7874 	 * until arc_read_done().
7875 	 */
7876 	if (BP_IS_ENCRYPTED(bp)) {
7877 		abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
7878 		    B_TRUE);
7879 
7880 		zio_crypt_decode_params_bp(bp, salt, iv);
7881 		zio_crypt_decode_mac_bp(bp, mac);
7882 
7883 		ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
7884 		    BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
7885 		    salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
7886 		    hdr->b_l1hdr.b_pabd, &no_crypt);
7887 		if (ret != 0) {
7888 			arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
7889 			goto error;
7890 		}
7891 
7892 		/*
7893 		 * If we actually performed decryption, replace b_pabd
7894 		 * with the decrypted data. Otherwise we can just throw
7895 		 * our decryption buffer away.
7896 		 */
7897 		if (!no_crypt) {
7898 			arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
7899 			    arc_hdr_size(hdr), hdr);
7900 			hdr->b_l1hdr.b_pabd = eabd;
7901 			zio->io_abd = eabd;
7902 		} else {
7903 			arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
7904 		}
7905 	}
7906 
7907 	/*
7908 	 * If the L2ARC block was compressed, but ARC compression
7909 	 * is disabled we decompress the data into a new buffer and
7910 	 * replace the existing data.
7911 	 */
7912 	if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
7913 	    !HDR_COMPRESSION_ENABLED(hdr)) {
7914 		abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
7915 		    B_TRUE);
7916 		void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
7917 
7918 		ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
7919 		    hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
7920 		    HDR_GET_LSIZE(hdr));
7921 		if (ret != 0) {
7922 			abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
7923 			arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
7924 			goto error;
7925 		}
7926 
7927 		abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
7928 		arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
7929 		    arc_hdr_size(hdr), hdr);
7930 		hdr->b_l1hdr.b_pabd = cabd;
7931 		zio->io_abd = cabd;
7932 		zio->io_size = HDR_GET_LSIZE(hdr);
7933 	}
7934 
7935 	return (0);
7936 
7937 error:
7938 	return (ret);
7939 }
7940 
7941 
7942 /*
7943  * A read to a cache device completed.  Validate buffer contents before
7944  * handing over to the regular ARC routines.
7945  */
7946 static void
l2arc_read_done(zio_t * zio)7947 l2arc_read_done(zio_t *zio)
7948 {
7949 	int tfm_error = 0;
7950 	l2arc_read_callback_t *cb = zio->io_private;
7951 	arc_buf_hdr_t *hdr;
7952 	kmutex_t *hash_lock;
7953 	boolean_t valid_cksum;
7954 	boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
7955 	    (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
7956 
7957 	ASSERT3P(zio->io_vd, !=, NULL);
7958 	ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7959 
7960 	spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7961 
7962 	ASSERT3P(cb, !=, NULL);
7963 	hdr = cb->l2rcb_hdr;
7964 	ASSERT3P(hdr, !=, NULL);
7965 
7966 	hash_lock = HDR_LOCK(hdr);
7967 	mutex_enter(hash_lock);
7968 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7969 
7970 	/*
7971 	 * If the data was read into a temporary buffer,
7972 	 * move it and free the buffer.
7973 	 */
7974 	if (cb->l2rcb_abd != NULL) {
7975 		ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7976 		if (zio->io_error == 0) {
7977 			if (using_rdata) {
7978 				abd_copy(hdr->b_crypt_hdr.b_rabd,
7979 				    cb->l2rcb_abd, arc_hdr_size(hdr));
7980 			} else {
7981 				abd_copy(hdr->b_l1hdr.b_pabd,
7982 				    cb->l2rcb_abd, arc_hdr_size(hdr));
7983 			}
7984 		}
7985 
7986 		/*
7987 		 * The following must be done regardless of whether
7988 		 * there was an error:
7989 		 * - free the temporary buffer
7990 		 * - point zio to the real ARC buffer
7991 		 * - set zio size accordingly
7992 		 * These are required because zio is either re-used for
7993 		 * an I/O of the block in the case of the error
7994 		 * or the zio is passed to arc_read_done() and it
7995 		 * needs real data.
7996 		 */
7997 		abd_free(cb->l2rcb_abd);
7998 		zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7999 
8000 		if (using_rdata) {
8001 			ASSERT(HDR_HAS_RABD(hdr));
8002 			zio->io_abd = zio->io_orig_abd =
8003 			    hdr->b_crypt_hdr.b_rabd;
8004 		} else {
8005 			ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8006 			zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8007 		}
8008 	}
8009 
8010 	ASSERT3P(zio->io_abd, !=, NULL);
8011 
8012 	/*
8013 	 * Check this survived the L2ARC journey.
8014 	 */
8015 	ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8016 	    (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8017 	zio->io_bp_copy = cb->l2rcb_bp;	/* XXX fix in L2ARC 2.0	*/
8018 	zio->io_bp = &zio->io_bp_copy;	/* XXX fix in L2ARC 2.0	*/
8019 
8020 	valid_cksum = arc_cksum_is_equal(hdr, zio);
8021 
8022 	/*
8023 	 * b_rabd will always match the data as it exists on disk if it is
8024 	 * being used. Therefore if we are reading into b_rabd we do not
8025 	 * attempt to untransform the data.
8026 	 */
8027 	if (valid_cksum && !using_rdata)
8028 		tfm_error = l2arc_untransform(zio, cb);
8029 
8030 	if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8031 	    !HDR_L2_EVICTED(hdr)) {
8032 		mutex_exit(hash_lock);
8033 		zio->io_private = hdr;
8034 		arc_read_done(zio);
8035 	} else {
8036 		/*
8037 		 * Buffer didn't survive caching.  Increment stats and
8038 		 * reissue to the original storage device.
8039 		 */
8040 		if (zio->io_error != 0) {
8041 			ARCSTAT_BUMP(arcstat_l2_io_error);
8042 		} else {
8043 			zio->io_error = SET_ERROR(EIO);
8044 		}
8045 		if (!valid_cksum || tfm_error != 0)
8046 			ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8047 
8048 		/*
8049 		 * If there's no waiter, issue an async i/o to the primary
8050 		 * storage now.  If there *is* a waiter, the caller must
8051 		 * issue the i/o in a context where it's OK to block.
8052 		 */
8053 		if (zio->io_waiter == NULL) {
8054 			zio_t *pio = zio_unique_parent(zio);
8055 			void *abd = (using_rdata) ?
8056 			    hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8057 
8058 			ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8059 
8060 			zio = zio_read(pio, zio->io_spa, zio->io_bp,
8061 			    abd, zio->io_size, arc_read_done,
8062 			    hdr, zio->io_priority, cb->l2rcb_flags,
8063 			    &cb->l2rcb_zb);
8064 
8065 			/*
8066 			 * Original ZIO will be freed, so we need to update
8067 			 * ARC header with the new ZIO pointer to be used
8068 			 * by zio_change_priority() in arc_read().
8069 			 */
8070 			for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
8071 			    acb != NULL; acb = acb->acb_next)
8072 				acb->acb_zio_head = zio;
8073 
8074 			mutex_exit(hash_lock);
8075 			zio_nowait(zio);
8076 		} else {
8077 			mutex_exit(hash_lock);
8078 		}
8079 	}
8080 
8081 	kmem_free(cb, sizeof (l2arc_read_callback_t));
8082 }
8083 
8084 /*
8085  * This is the list priority from which the L2ARC will search for pages to
8086  * cache.  This is used within loops (0..3) to cycle through lists in the
8087  * desired order.  This order can have a significant effect on cache
8088  * performance.
8089  *
8090  * Currently the metadata lists are hit first, MFU then MRU, followed by
8091  * the data lists.  This function returns a locked list, and also returns
8092  * the lock pointer.
8093  */
8094 static multilist_sublist_t *
l2arc_sublist_lock(int list_num)8095 l2arc_sublist_lock(int list_num)
8096 {
8097 	multilist_t *ml = NULL;
8098 	unsigned int idx;
8099 
8100 	ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8101 
8102 	switch (list_num) {
8103 	case 0:
8104 		ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
8105 		break;
8106 	case 1:
8107 		ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
8108 		break;
8109 	case 2:
8110 		ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
8111 		break;
8112 	case 3:
8113 		ml = arc_mru->arcs_list[ARC_BUFC_DATA];
8114 		break;
8115 	default:
8116 		return (NULL);
8117 	}
8118 
8119 	/*
8120 	 * Return a randomly-selected sublist. This is acceptable
8121 	 * because the caller feeds only a little bit of data for each
8122 	 * call (8MB). Subsequent calls will result in different
8123 	 * sublists being selected.
8124 	 */
8125 	idx = multilist_get_random_index(ml);
8126 	return (multilist_sublist_lock(ml, idx));
8127 }
8128 
8129 /*
8130  * Calculates the maximum overhead of L2ARC metadata log blocks for a given
8131  * L2ARC write size. l2arc_evict and l2arc_write_size need to include this
8132  * overhead in processing to make sure there is enough headroom available
8133  * when writing buffers.
8134  */
8135 static inline uint64_t
l2arc_log_blk_overhead(uint64_t write_sz,l2arc_dev_t * dev)8136 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev)
8137 {
8138 	if (dev->l2ad_log_entries == 0) {
8139 		return (0);
8140 	} else {
8141 		uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT;
8142 
8143 		uint64_t log_blocks = (log_entries +
8144 		    dev->l2ad_log_entries - 1) /
8145 		    dev->l2ad_log_entries;
8146 
8147 		return (vdev_psize_to_asize(dev->l2ad_vdev,
8148 		    sizeof (l2arc_log_blk_phys_t)) * log_blocks);
8149 	}
8150 }
8151 
8152 /*
8153  * Evict buffers from the device write hand to the distance specified in
8154  * bytes. This distance may span populated buffers, it may span nothing.
8155  * This is clearing a region on the L2ARC device ready for writing.
8156  * If the 'all' boolean is set, every buffer is evicted.
8157  */
8158 static void
l2arc_evict(l2arc_dev_t * dev,uint64_t distance,boolean_t all)8159 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
8160 {
8161 	list_t *buflist;
8162 	arc_buf_hdr_t *hdr, *hdr_prev;
8163 	kmutex_t *hash_lock;
8164 	uint64_t taddr;
8165 	l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev;
8166 	boolean_t rerun;
8167 
8168 	buflist = &dev->l2ad_buflist;
8169 
8170 	/*
8171 	 * We need to add in the worst case scenario of log block overhead.
8172 	 */
8173 	distance += l2arc_log_blk_overhead(distance, dev);
8174 
8175 top:
8176 	rerun = B_FALSE;
8177 	if (dev->l2ad_hand >= (dev->l2ad_end - distance)) {
8178 		/*
8179 		 * When there is no space to accommodate upcoming writes,
8180 		 * evict to the end. Then bump the write and evict hands
8181 		 * to the start and iterate. This iteration does not
8182 		 * happen indefinitely as we make sure in
8183 		 * l2arc_write_size() that when the write hand is reset,
8184 		 * the write size does not exceed the end of the device.
8185 		 */
8186 		rerun = B_TRUE;
8187 		taddr = dev->l2ad_end;
8188 	} else {
8189 		taddr = dev->l2ad_hand + distance;
8190 	}
8191 	DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
8192 	    uint64_t, taddr, boolean_t, all);
8193 
8194 	/*
8195 	 * This check has to be placed after deciding whether to iterate
8196 	 * (rerun).
8197 	 */
8198 	if (!all && dev->l2ad_first) {
8199 		/*
8200 		 * This is the first sweep through the device. There is
8201 		 * nothing to evict.
8202 		 */
8203 		goto out;
8204 	}
8205 
8206 	/*
8207 	 * When rebuilding L2ARC we retrieve the evict hand from the header of
8208 	 * the device. Of note, l2arc_evict() does not actually delete buffers
8209 	 * from the cache device, but keeping track of the evict hand will be
8210 	 * useful when TRIM is implemented.
8211 	 */
8212 	dev->l2ad_evict = MAX(dev->l2ad_evict, taddr);
8213 
8214 retry:
8215 	mutex_enter(&dev->l2ad_mtx);
8216 	/*
8217 	 * We have to account for evicted log blocks. Run vdev_space_update()
8218 	 * on log blocks whose offset (in bytes) is before the evicted offset
8219 	 * (in bytes) by searching in the list of pointers to log blocks
8220 	 * present in the L2ARC device.
8221 	 */
8222 	for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf;
8223 	    lb_ptr_buf = lb_ptr_buf_prev) {
8224 
8225 		lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf);
8226 
8227 		/* L2BLK_GET_PSIZE returns aligned size for log blocks */
8228 		uint64_t asize = L2BLK_GET_PSIZE(
8229 		    (lb_ptr_buf->lb_ptr)->lbp_prop);
8230 
8231 		/*
8232 		 * We don't worry about log blocks left behind (ie
8233 		 * lbp_payload_start < l2ad_hand) because l2arc_write_buffers()
8234 		 * will never write more than l2arc_evict() evicts.
8235 		 */
8236 		if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) {
8237 			break;
8238 		} else {
8239 			vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
8240 			ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8241 			ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8242 			zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8243 			    lb_ptr_buf);
8244 			zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8245 			list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf);
8246 			kmem_free(lb_ptr_buf->lb_ptr,
8247 			    sizeof (l2arc_log_blkptr_t));
8248 			kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8249 		}
8250 	}
8251 
8252 	for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
8253 		hdr_prev = list_prev(buflist, hdr);
8254 
8255 		ASSERT(!HDR_EMPTY(hdr));
8256 		hash_lock = HDR_LOCK(hdr);
8257 
8258 		/*
8259 		 * We cannot use mutex_enter or else we can deadlock
8260 		 * with l2arc_write_buffers (due to swapping the order
8261 		 * the hash lock and l2ad_mtx are taken).
8262 		 */
8263 		if (!mutex_tryenter(hash_lock)) {
8264 			/*
8265 			 * Missed the hash lock.  Retry.
8266 			 */
8267 			ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8268 			mutex_exit(&dev->l2ad_mtx);
8269 			mutex_enter(hash_lock);
8270 			mutex_exit(hash_lock);
8271 			goto retry;
8272 		}
8273 
8274 		/*
8275 		 * A header can't be on this list if it doesn't have L2 header.
8276 		 */
8277 		ASSERT(HDR_HAS_L2HDR(hdr));
8278 
8279 		/* Ensure this header has finished being written. */
8280 		ASSERT(!HDR_L2_WRITING(hdr));
8281 		ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8282 
8283 		if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict ||
8284 		    hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8285 			/*
8286 			 * We've evicted to the target address,
8287 			 * or the end of the device.
8288 			 */
8289 			mutex_exit(hash_lock);
8290 			break;
8291 		}
8292 
8293 		if (!HDR_HAS_L1HDR(hdr)) {
8294 			ASSERT(!HDR_L2_READING(hdr));
8295 			/*
8296 			 * This doesn't exist in the ARC.  Destroy.
8297 			 * arc_hdr_destroy() will call list_remove()
8298 			 * and decrement arcstat_l2_lsize.
8299 			 */
8300 			arc_change_state(arc_anon, hdr, hash_lock);
8301 			arc_hdr_destroy(hdr);
8302 		} else {
8303 			ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8304 			ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8305 			/*
8306 			 * Invalidate issued or about to be issued
8307 			 * reads, since we may be about to write
8308 			 * over this location.
8309 			 */
8310 			if (HDR_L2_READING(hdr)) {
8311 				ARCSTAT_BUMP(arcstat_l2_evict_reading);
8312 				arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
8313 			}
8314 
8315 			arc_hdr_l2hdr_destroy(hdr);
8316 		}
8317 		mutex_exit(hash_lock);
8318 	}
8319 	mutex_exit(&dev->l2ad_mtx);
8320 
8321 out:
8322 	/*
8323 	 * We need to check if we evict all buffers, otherwise we may iterate
8324 	 * unnecessarily.
8325 	 */
8326 	if (!all && rerun) {
8327 		/*
8328 		 * Bump device hand to the device start if it is approaching the
8329 		 * end. l2arc_evict() has already evicted ahead for this case.
8330 		 */
8331 		dev->l2ad_hand = dev->l2ad_start;
8332 		dev->l2ad_evict = dev->l2ad_start;
8333 		dev->l2ad_first = B_FALSE;
8334 		goto top;
8335 	}
8336 
8337 	ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end);
8338 	if (!dev->l2ad_first)
8339 		ASSERT3U(dev->l2ad_hand, <, dev->l2ad_evict);
8340 }
8341 
8342 /*
8343  * Handle any abd transforms that might be required for writing to the L2ARC.
8344  * If successful, this function will always return an abd with the data
8345  * transformed as it is on disk in a new abd of asize bytes.
8346  */
8347 static int
l2arc_apply_transforms(spa_t * spa,arc_buf_hdr_t * hdr,uint64_t asize,abd_t ** abd_out)8348 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
8349     abd_t **abd_out)
8350 {
8351 	int ret;
8352 	void *tmp = NULL;
8353 	abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
8354 	enum zio_compress compress = HDR_GET_COMPRESS(hdr);
8355 	uint64_t psize = HDR_GET_PSIZE(hdr);
8356 	uint64_t size = arc_hdr_size(hdr);
8357 	boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
8358 	boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
8359 	dsl_crypto_key_t *dck = NULL;
8360 	uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
8361 	boolean_t no_crypt = B_FALSE;
8362 
8363 	ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8364 	    !HDR_COMPRESSION_ENABLED(hdr)) ||
8365 	    HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
8366 	ASSERT3U(psize, <=, asize);
8367 
8368 	/*
8369 	 * If this data simply needs its own buffer, we simply allocate it
8370 	 * and copy the data. This may be done to eliminate a dependency on a
8371 	 * shared buffer or to reallocate the buffer to match asize.
8372 	 */
8373 	if (HDR_HAS_RABD(hdr) && asize != psize) {
8374 		ASSERT3U(asize, >=, psize);
8375 		to_write = abd_alloc_for_io(asize, ismd);
8376 		abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
8377 		if (psize != asize)
8378 			abd_zero_off(to_write, psize, asize - psize);
8379 		goto out;
8380 	}
8381 
8382 	if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
8383 	    !HDR_ENCRYPTED(hdr)) {
8384 		ASSERT3U(size, ==, psize);
8385 		to_write = abd_alloc_for_io(asize, ismd);
8386 		abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
8387 		if (size != asize)
8388 			abd_zero_off(to_write, size, asize - size);
8389 		goto out;
8390 	}
8391 
8392 	if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
8393 		cabd = abd_alloc_for_io(asize, ismd);
8394 		tmp = abd_borrow_buf(cabd, asize);
8395 
8396 		psize = zio_compress_data(compress, to_write, tmp, size);
8397 		ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
8398 		if (psize < asize)
8399 			bzero((char *)tmp + psize, asize - psize);
8400 		psize = HDR_GET_PSIZE(hdr);
8401 		abd_return_buf_copy(cabd, tmp, asize);
8402 		to_write = cabd;
8403 	}
8404 
8405 	if (HDR_ENCRYPTED(hdr)) {
8406 		eabd = abd_alloc_for_io(asize, ismd);
8407 
8408 		/*
8409 		 * If the dataset was disowned before the buffer
8410 		 * made it to this point, the key to re-encrypt
8411 		 * it won't be available. In this case we simply
8412 		 * won't write the buffer to the L2ARC.
8413 		 */
8414 		ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
8415 		    FTAG, &dck);
8416 		if (ret != 0)
8417 			goto error;
8418 
8419 		ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
8420 		    hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
8421 		    hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
8422 		    &no_crypt);
8423 		if (ret != 0)
8424 			goto error;
8425 
8426 		if (no_crypt)
8427 			abd_copy(eabd, to_write, psize);
8428 
8429 		if (psize != asize)
8430 			abd_zero_off(eabd, psize, asize - psize);
8431 
8432 		/* assert that the MAC we got here matches the one we saved */
8433 		ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
8434 		spa_keystore_dsl_key_rele(spa, dck, FTAG);
8435 
8436 		if (to_write == cabd)
8437 			abd_free(cabd);
8438 
8439 		to_write = eabd;
8440 	}
8441 
8442 out:
8443 	ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
8444 	*abd_out = to_write;
8445 	return (0);
8446 
8447 error:
8448 	if (dck != NULL)
8449 		spa_keystore_dsl_key_rele(spa, dck, FTAG);
8450 	if (cabd != NULL)
8451 		abd_free(cabd);
8452 	if (eabd != NULL)
8453 		abd_free(eabd);
8454 
8455 	*abd_out = NULL;
8456 	return (ret);
8457 }
8458 
8459 static void
l2arc_blk_fetch_done(zio_t * zio)8460 l2arc_blk_fetch_done(zio_t *zio)
8461 {
8462 	l2arc_read_callback_t *cb;
8463 
8464 	cb = zio->io_private;
8465 	if (cb->l2rcb_abd != NULL)
8466 		abd_put(cb->l2rcb_abd);
8467 	kmem_free(cb, sizeof (l2arc_read_callback_t));
8468 }
8469 
8470 /*
8471  * Find and write ARC buffers to the L2ARC device.
8472  *
8473  * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
8474  * for reading until they have completed writing.
8475  * The headroom_boost is an in-out parameter used to maintain headroom boost
8476  * state between calls to this function.
8477  *
8478  * Returns the number of bytes actually written (which may be smaller than
8479  * the delta by which the device hand has changed due to alignment and the
8480  * writing of log blocks).
8481  */
8482 static uint64_t
l2arc_write_buffers(spa_t * spa,l2arc_dev_t * dev,uint64_t target_sz)8483 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
8484 {
8485 	arc_buf_hdr_t		*hdr, *hdr_prev, *head;
8486 	uint64_t		write_asize, write_psize, write_lsize, headroom;
8487 	boolean_t		full;
8488 	l2arc_write_callback_t	*cb = NULL;
8489 	zio_t			*pio, *wzio;
8490 	uint64_t		guid = spa_load_guid(spa);
8491 	l2arc_dev_hdr_phys_t	*l2dhdr = dev->l2ad_dev_hdr;
8492 
8493 	ASSERT3P(dev->l2ad_vdev, !=, NULL);
8494 
8495 	pio = NULL;
8496 	write_lsize = write_asize = write_psize = 0;
8497 	full = B_FALSE;
8498 	head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
8499 	arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
8500 
8501 	/*
8502 	 * Copy buffers for L2ARC writing.
8503 	 */
8504 	for (int try = 0; try < L2ARC_FEED_TYPES; try++) {
8505 		/*
8506 		 * If try == 1 or 3, we cache MRU metadata and data
8507 		 * respectively.
8508 		 */
8509 		if (l2arc_mfuonly) {
8510 			if (try == 1 || try == 3)
8511 				continue;
8512 		}
8513 
8514 		multilist_sublist_t *mls = l2arc_sublist_lock(try);
8515 		uint64_t passed_sz = 0;
8516 
8517 		VERIFY3P(mls, !=, NULL);
8518 
8519 		/*
8520 		 * L2ARC fast warmup.
8521 		 *
8522 		 * Until the ARC is warm and starts to evict, read from the
8523 		 * head of the ARC lists rather than the tail.
8524 		 */
8525 		if (arc_warm == B_FALSE)
8526 			hdr = multilist_sublist_head(mls);
8527 		else
8528 			hdr = multilist_sublist_tail(mls);
8529 
8530 		headroom = target_sz * l2arc_headroom;
8531 		if (zfs_compressed_arc_enabled)
8532 			headroom = (headroom * l2arc_headroom_boost) / 100;
8533 
8534 		for (; hdr; hdr = hdr_prev) {
8535 			kmutex_t *hash_lock;
8536 			abd_t *to_write = NULL;
8537 
8538 			if (arc_warm == B_FALSE)
8539 				hdr_prev = multilist_sublist_next(mls, hdr);
8540 			else
8541 				hdr_prev = multilist_sublist_prev(mls, hdr);
8542 
8543 			hash_lock = HDR_LOCK(hdr);
8544 			if (!mutex_tryenter(hash_lock)) {
8545 				/*
8546 				 * Skip this buffer rather than waiting.
8547 				 */
8548 				continue;
8549 			}
8550 
8551 			passed_sz += HDR_GET_LSIZE(hdr);
8552 			if (l2arc_headroom != 0 && passed_sz > headroom) {
8553 				/*
8554 				 * Searched too far.
8555 				 */
8556 				mutex_exit(hash_lock);
8557 				break;
8558 			}
8559 
8560 			if (!l2arc_write_eligible(guid, hdr)) {
8561 				mutex_exit(hash_lock);
8562 				continue;
8563 			}
8564 
8565 			ASSERT(HDR_HAS_L1HDR(hdr));
8566 
8567 			ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8568 			ASSERT3U(arc_hdr_size(hdr), >, 0);
8569 			ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8570 			    HDR_HAS_RABD(hdr));
8571 			uint64_t psize = HDR_GET_PSIZE(hdr);
8572 			uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
8573 			    psize);
8574 
8575 			if ((write_asize + asize) > target_sz) {
8576 				full = B_TRUE;
8577 				mutex_exit(hash_lock);
8578 				break;
8579 			}
8580 
8581 			/*
8582 			 * We rely on the L1 portion of the header below, so
8583 			 * it's invalid for this header to have been evicted out
8584 			 * of the ghost cache, prior to being written out. The
8585 			 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8586 			 */
8587 			arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
8588 
8589 			/*
8590 			 * If this header has b_rabd, we can use this since it
8591 			 * must always match the data exactly as it exists on
8592 			 * disk. Otherwise, the L2ARC can normally use the
8593 			 * hdr's data, but if we're sharing data between the
8594 			 * hdr and one of its bufs, L2ARC needs its own copy of
8595 			 * the data so that the ZIO below can't race with the
8596 			 * buf consumer. To ensure that this copy will be
8597 			 * available for the lifetime of the ZIO and be cleaned
8598 			 * up afterwards, we add it to the l2arc_free_on_write
8599 			 * queue. If we need to apply any transforms to the
8600 			 * data (compression, encryption) we will also need the
8601 			 * extra buffer.
8602 			 */
8603 			if (HDR_HAS_RABD(hdr) && psize == asize) {
8604 				to_write = hdr->b_crypt_hdr.b_rabd;
8605 			} else if ((HDR_COMPRESSION_ENABLED(hdr) ||
8606 			    HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
8607 			    !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
8608 			    psize == asize) {
8609 				to_write = hdr->b_l1hdr.b_pabd;
8610 			} else {
8611 				int ret;
8612 				arc_buf_contents_t type = arc_buf_type(hdr);
8613 
8614 				ret = l2arc_apply_transforms(spa, hdr, asize,
8615 				    &to_write);
8616 				if (ret != 0) {
8617 					arc_hdr_clear_flags(hdr,
8618 					    ARC_FLAG_L2_WRITING);
8619 					mutex_exit(hash_lock);
8620 					continue;
8621 				}
8622 
8623 				l2arc_free_abd_on_write(to_write, asize, type);
8624 			}
8625 
8626 			if (pio == NULL) {
8627 				/*
8628 				 * Insert a dummy header on the buflist so
8629 				 * l2arc_write_done() can find where the
8630 				 * write buffers begin without searching.
8631 				 */
8632 				mutex_enter(&dev->l2ad_mtx);
8633 				list_insert_head(&dev->l2ad_buflist, head);
8634 				mutex_exit(&dev->l2ad_mtx);
8635 
8636 				cb = kmem_alloc(
8637 				    sizeof (l2arc_write_callback_t), KM_SLEEP);
8638 				cb->l2wcb_dev = dev;
8639 				cb->l2wcb_head = head;
8640 				/*
8641 				 * Create a list to save allocated abd buffers
8642 				 * for l2arc_log_blk_commit().
8643 				 */
8644 				list_create(&cb->l2wcb_abd_list,
8645 				    sizeof (l2arc_lb_abd_buf_t),
8646 				    offsetof(l2arc_lb_abd_buf_t, node));
8647 				pio = zio_root(spa, l2arc_write_done, cb,
8648 				    ZIO_FLAG_CANFAIL);
8649 			}
8650 
8651 			hdr->b_l2hdr.b_dev = dev;
8652 			hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
8653 			hdr->b_l2hdr.b_arcs_state =
8654 			    hdr->b_l1hdr.b_state->arcs_state;
8655 			arc_hdr_set_flags(hdr,
8656 			    ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
8657 
8658 			mutex_enter(&dev->l2ad_mtx);
8659 			list_insert_head(&dev->l2ad_buflist, hdr);
8660 			mutex_exit(&dev->l2ad_mtx);
8661 
8662 			(void) zfs_refcount_add_many(&dev->l2ad_alloc,
8663 			    arc_hdr_size(hdr), hdr);
8664 
8665 			wzio = zio_write_phys(pio, dev->l2ad_vdev,
8666 			    hdr->b_l2hdr.b_daddr, asize, to_write,
8667 			    ZIO_CHECKSUM_OFF, NULL, hdr,
8668 			    ZIO_PRIORITY_ASYNC_WRITE,
8669 			    ZIO_FLAG_CANFAIL, B_FALSE);
8670 
8671 			write_lsize += HDR_GET_LSIZE(hdr);
8672 			DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
8673 			    zio_t *, wzio);
8674 
8675 			write_psize += psize;
8676 			write_asize += asize;
8677 			dev->l2ad_hand += asize;
8678 			l2arc_hdr_arcstats_increment(hdr);
8679 			vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
8680 
8681 			mutex_exit(hash_lock);
8682 
8683 			/*
8684 			 * Append buf info to current log and commit if full.
8685 			 * arcstat_l2_{size,asize} kstats are updated
8686 			 * internally.
8687 			 */
8688 			if (l2arc_log_blk_insert(dev, hdr))
8689 				l2arc_log_blk_commit(dev, pio, cb);
8690 
8691 			(void) zio_nowait(wzio);
8692 		}
8693 
8694 		multilist_sublist_unlock(mls);
8695 
8696 		if (full == B_TRUE)
8697 			break;
8698 	}
8699 
8700 	/* No buffers selected for writing? */
8701 	if (pio == NULL) {
8702 		ASSERT0(write_lsize);
8703 		ASSERT(!HDR_HAS_L1HDR(head));
8704 		kmem_cache_free(hdr_l2only_cache, head);
8705 
8706 		/*
8707 		 * Although we did not write any buffers l2ad_evict may
8708 		 * have advanced.
8709 		 */
8710 		if (dev->l2ad_evict != l2dhdr->dh_evict)
8711 			l2arc_dev_hdr_update(dev);
8712 
8713 		return (0);
8714 	}
8715 
8716 	if (!dev->l2ad_first)
8717 		ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
8718 
8719 	ASSERT3U(write_asize, <=, target_sz);
8720 	ARCSTAT_BUMP(arcstat_l2_writes_sent);
8721 	ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
8722 
8723 	dev->l2ad_writing = B_TRUE;
8724 	(void) zio_wait(pio);
8725 	dev->l2ad_writing = B_FALSE;
8726 
8727 	/*
8728 	 * Update the device header after the zio completes as
8729 	 * l2arc_write_done() may have updated the memory holding the log block
8730 	 * pointers in the device header.
8731 	 */
8732 	l2arc_dev_hdr_update(dev);
8733 
8734 	return (write_asize);
8735 }
8736 
8737 static boolean_t
l2arc_hdr_limit_reached(void)8738 l2arc_hdr_limit_reached(void)
8739 {
8740 	int64_t s = aggsum_upper_bound(&astat_l2_hdr_size);
8741 
8742 	return (arc_reclaim_needed() || (s > arc_meta_limit * 3 / 4) ||
8743 	    (s > (arc_warm ? arc_c : arc_c_max) * l2arc_meta_percent / 100));
8744 }
8745 
8746 /*
8747  * This thread feeds the L2ARC at regular intervals.  This is the beating
8748  * heart of the L2ARC.
8749  */
8750 /* ARGSUSED */
8751 static void
l2arc_feed_thread(void * unused)8752 l2arc_feed_thread(void *unused)
8753 {
8754 	callb_cpr_t cpr;
8755 	l2arc_dev_t *dev;
8756 	spa_t *spa;
8757 	uint64_t size, wrote;
8758 	clock_t begin, next = ddi_get_lbolt();
8759 
8760 	CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
8761 
8762 	mutex_enter(&l2arc_feed_thr_lock);
8763 
8764 	while (l2arc_thread_exit == 0) {
8765 		CALLB_CPR_SAFE_BEGIN(&cpr);
8766 		(void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
8767 		    next);
8768 		CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
8769 		next = ddi_get_lbolt() + hz;
8770 
8771 		/*
8772 		 * Quick check for L2ARC devices.
8773 		 */
8774 		mutex_enter(&l2arc_dev_mtx);
8775 		if (l2arc_ndev == 0) {
8776 			mutex_exit(&l2arc_dev_mtx);
8777 			continue;
8778 		}
8779 		mutex_exit(&l2arc_dev_mtx);
8780 		begin = ddi_get_lbolt();
8781 
8782 		/*
8783 		 * This selects the next l2arc device to write to, and in
8784 		 * doing so the next spa to feed from: dev->l2ad_spa.   This
8785 		 * will return NULL if there are now no l2arc devices or if
8786 		 * they are all faulted.
8787 		 *
8788 		 * If a device is returned, its spa's config lock is also
8789 		 * held to prevent device removal.  l2arc_dev_get_next()
8790 		 * will grab and release l2arc_dev_mtx.
8791 		 */
8792 		if ((dev = l2arc_dev_get_next()) == NULL)
8793 			continue;
8794 
8795 		spa = dev->l2ad_spa;
8796 		ASSERT3P(spa, !=, NULL);
8797 
8798 		/*
8799 		 * If the pool is read-only then force the feed thread to
8800 		 * sleep a little longer.
8801 		 */
8802 		if (!spa_writeable(spa)) {
8803 			next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
8804 			spa_config_exit(spa, SCL_L2ARC, dev);
8805 			continue;
8806 		}
8807 
8808 		/*
8809 		 * Avoid contributing to memory pressure.
8810 		 */
8811 		if (l2arc_hdr_limit_reached()) {
8812 			ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
8813 			spa_config_exit(spa, SCL_L2ARC, dev);
8814 			continue;
8815 		}
8816 
8817 		ARCSTAT_BUMP(arcstat_l2_feeds);
8818 
8819 		size = l2arc_write_size(dev);
8820 
8821 		/*
8822 		 * Evict L2ARC buffers that will be overwritten.
8823 		 */
8824 		l2arc_evict(dev, size, B_FALSE);
8825 
8826 		/*
8827 		 * Write ARC buffers.
8828 		 */
8829 		wrote = l2arc_write_buffers(spa, dev, size);
8830 
8831 		/*
8832 		 * Calculate interval between writes.
8833 		 */
8834 		next = l2arc_write_interval(begin, size, wrote);
8835 		spa_config_exit(spa, SCL_L2ARC, dev);
8836 	}
8837 
8838 	l2arc_thread_exit = 0;
8839 	cv_broadcast(&l2arc_feed_thr_cv);
8840 	CALLB_CPR_EXIT(&cpr);		/* drops l2arc_feed_thr_lock */
8841 	thread_exit();
8842 }
8843 
8844 boolean_t
l2arc_vdev_present(vdev_t * vd)8845 l2arc_vdev_present(vdev_t *vd)
8846 {
8847 	return (l2arc_vdev_get(vd) != NULL);
8848 }
8849 
8850 /*
8851  * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
8852  * the vdev_t isn't an L2ARC device.
8853  */
8854 static l2arc_dev_t *
l2arc_vdev_get(vdev_t * vd)8855 l2arc_vdev_get(vdev_t *vd)
8856 {
8857 	l2arc_dev_t	*dev;
8858 
8859 	mutex_enter(&l2arc_dev_mtx);
8860 	for (dev = list_head(l2arc_dev_list); dev != NULL;
8861 	    dev = list_next(l2arc_dev_list, dev)) {
8862 		if (dev->l2ad_vdev == vd)
8863 			break;
8864 	}
8865 	mutex_exit(&l2arc_dev_mtx);
8866 
8867 	return (dev);
8868 }
8869 
8870 /*
8871  * Add a vdev for use by the L2ARC.  By this point the spa has already
8872  * validated the vdev and opened it.
8873  */
8874 void
l2arc_add_vdev(spa_t * spa,vdev_t * vd)8875 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
8876 {
8877 	l2arc_dev_t		*adddev;
8878 	uint64_t		l2dhdr_asize;
8879 
8880 	ASSERT(!l2arc_vdev_present(vd));
8881 
8882 	/*
8883 	 * Create a new l2arc device entry.
8884 	 */
8885 	adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
8886 	adddev->l2ad_spa = spa;
8887 	adddev->l2ad_vdev = vd;
8888 	/* leave extra size for an l2arc device header */
8889 	l2dhdr_asize = adddev->l2ad_dev_hdr_asize =
8890 	    MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift);
8891 	adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize;
8892 	adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
8893 	ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
8894 	adddev->l2ad_hand = adddev->l2ad_start;
8895 	adddev->l2ad_evict = adddev->l2ad_start;
8896 	adddev->l2ad_first = B_TRUE;
8897 	adddev->l2ad_writing = B_FALSE;
8898 	adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP);
8899 
8900 	mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
8901 	/*
8902 	 * This is a list of all ARC buffers that are still valid on the
8903 	 * device.
8904 	 */
8905 	list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
8906 	    offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
8907 
8908 	/*
8909 	 * This is a list of pointers to log blocks that are still present
8910 	 * on the device.
8911 	 */
8912 	list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t),
8913 	    offsetof(l2arc_lb_ptr_buf_t, node));
8914 
8915 	vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
8916 	zfs_refcount_create(&adddev->l2ad_alloc);
8917 	zfs_refcount_create(&adddev->l2ad_lb_asize);
8918 	zfs_refcount_create(&adddev->l2ad_lb_count);
8919 
8920 	/*
8921 	 * Add device to global list
8922 	 */
8923 	mutex_enter(&l2arc_dev_mtx);
8924 	list_insert_head(l2arc_dev_list, adddev);
8925 	atomic_inc_64(&l2arc_ndev);
8926 	mutex_exit(&l2arc_dev_mtx);
8927 
8928 	/*
8929 	 * Decide if vdev is eligible for L2ARC rebuild
8930 	 */
8931 	l2arc_rebuild_vdev(adddev->l2ad_vdev, B_FALSE);
8932 }
8933 
8934 void
l2arc_rebuild_vdev(vdev_t * vd,boolean_t reopen)8935 l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen)
8936 {
8937 	l2arc_dev_t		*dev = NULL;
8938 	l2arc_dev_hdr_phys_t	*l2dhdr;
8939 	uint64_t		l2dhdr_asize;
8940 	spa_t			*spa;
8941 
8942 	dev = l2arc_vdev_get(vd);
8943 	ASSERT3P(dev, !=, NULL);
8944 	spa = dev->l2ad_spa;
8945 	l2dhdr = dev->l2ad_dev_hdr;
8946 	l2dhdr_asize = dev->l2ad_dev_hdr_asize;
8947 
8948 	/*
8949 	 * The L2ARC has to hold at least the payload of one log block for
8950 	 * them to be restored (persistent L2ARC). The payload of a log block
8951 	 * depends on the amount of its log entries. We always write log blocks
8952 	 * with 1022 entries. How many of them are committed or restored depends
8953 	 * on the size of the L2ARC device. Thus the maximum payload of
8954 	 * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device
8955 	 * is less than that, we reduce the amount of committed and restored
8956 	 * log entries per block so as to enable persistence.
8957 	 */
8958 	if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) {
8959 		dev->l2ad_log_entries = 0;
8960 	} else {
8961 		dev->l2ad_log_entries = MIN((dev->l2ad_end -
8962 		    dev->l2ad_start) >> SPA_MAXBLOCKSHIFT,
8963 		    L2ARC_LOG_BLK_MAX_ENTRIES);
8964 	}
8965 
8966 	/*
8967 	 * Read the device header, if an error is returned do not rebuild L2ARC.
8968 	 */
8969 	if (l2arc_dev_hdr_read(dev) == 0 && dev->l2ad_log_entries > 0) {
8970 		/*
8971 		 * If we are onlining a cache device (vdev_reopen) that was
8972 		 * still present (l2arc_vdev_present()) and rebuild is enabled,
8973 		 * we should evict all ARC buffers and pointers to log blocks
8974 		 * and reclaim their space before restoring its contents to
8975 		 * L2ARC.
8976 		 */
8977 		if (reopen) {
8978 			if (!l2arc_rebuild_enabled) {
8979 				return;
8980 			} else {
8981 				l2arc_evict(dev, 0, B_TRUE);
8982 				/* start a new log block */
8983 				dev->l2ad_log_ent_idx = 0;
8984 				dev->l2ad_log_blk_payload_asize = 0;
8985 				dev->l2ad_log_blk_payload_start = 0;
8986 			}
8987 		}
8988 		/*
8989 		 * Just mark the device as pending for a rebuild. We won't
8990 		 * be starting a rebuild in line here as it would block pool
8991 		 * import. Instead spa_load_impl will hand that off to an
8992 		 * async task which will call l2arc_spa_rebuild_start.
8993 		 */
8994 		dev->l2ad_rebuild = B_TRUE;
8995 	} else if (spa_writeable(spa)) {
8996 		/*
8997 		 * In this case create a new header. We zero out the memory
8998 		 * holding the header to reset dh_start_lbps.
8999 		 */
9000 		bzero(l2dhdr, l2dhdr_asize);
9001 		l2arc_dev_hdr_update(dev);
9002 	}
9003 }
9004 
9005 /*
9006  * Remove a vdev from the L2ARC.
9007  */
9008 void
l2arc_remove_vdev(vdev_t * vd)9009 l2arc_remove_vdev(vdev_t *vd)
9010 {
9011 	l2arc_dev_t *remdev = NULL;
9012 
9013 	/*
9014 	 * Find the device by vdev
9015 	 */
9016 	remdev = l2arc_vdev_get(vd);
9017 	ASSERT3P(remdev, !=, NULL);
9018 
9019 	/*
9020 	 * Cancel any ongoing or scheduled rebuild.
9021 	 */
9022 	mutex_enter(&l2arc_rebuild_thr_lock);
9023 	if (remdev->l2ad_rebuild_began == B_TRUE) {
9024 		remdev->l2ad_rebuild_cancel = B_TRUE;
9025 		while (remdev->l2ad_rebuild == B_TRUE)
9026 			cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock);
9027 	}
9028 	mutex_exit(&l2arc_rebuild_thr_lock);
9029 
9030 	/*
9031 	 * Remove device from global list
9032 	 */
9033 	mutex_enter(&l2arc_dev_mtx);
9034 	list_remove(l2arc_dev_list, remdev);
9035 	l2arc_dev_last = NULL;		/* may have been invalidated */
9036 	atomic_dec_64(&l2arc_ndev);
9037 	mutex_exit(&l2arc_dev_mtx);
9038 
9039 	/*
9040 	 * Clear all buflists and ARC references.  L2ARC device flush.
9041 	 */
9042 	l2arc_evict(remdev, 0, B_TRUE);
9043 	list_destroy(&remdev->l2ad_buflist);
9044 	ASSERT(list_is_empty(&remdev->l2ad_lbptr_list));
9045 	list_destroy(&remdev->l2ad_lbptr_list);
9046 	mutex_destroy(&remdev->l2ad_mtx);
9047 	zfs_refcount_destroy(&remdev->l2ad_alloc);
9048 	zfs_refcount_destroy(&remdev->l2ad_lb_asize);
9049 	zfs_refcount_destroy(&remdev->l2ad_lb_count);
9050 	kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
9051 	kmem_free(remdev, sizeof (l2arc_dev_t));
9052 }
9053 
9054 void
l2arc_init(void)9055 l2arc_init(void)
9056 {
9057 	l2arc_thread_exit = 0;
9058 	l2arc_ndev = 0;
9059 	l2arc_writes_sent = 0;
9060 	l2arc_writes_done = 0;
9061 
9062 	mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9063 	cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
9064 	mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9065 	cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL);
9066 	mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
9067 	mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
9068 
9069 	l2arc_dev_list = &L2ARC_dev_list;
9070 	l2arc_free_on_write = &L2ARC_free_on_write;
9071 	list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
9072 	    offsetof(l2arc_dev_t, l2ad_node));
9073 	list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
9074 	    offsetof(l2arc_data_free_t, l2df_list_node));
9075 }
9076 
9077 void
l2arc_fini(void)9078 l2arc_fini(void)
9079 {
9080 	/*
9081 	 * This is called from dmu_fini(), which is called from spa_fini();
9082 	 * Because of this, we can assume that all l2arc devices have
9083 	 * already been removed when the pools themselves were removed.
9084 	 */
9085 
9086 	l2arc_do_free_on_write();
9087 
9088 	mutex_destroy(&l2arc_feed_thr_lock);
9089 	cv_destroy(&l2arc_feed_thr_cv);
9090 	mutex_destroy(&l2arc_rebuild_thr_lock);
9091 	cv_destroy(&l2arc_rebuild_thr_cv);
9092 	mutex_destroy(&l2arc_dev_mtx);
9093 	mutex_destroy(&l2arc_free_on_write_mtx);
9094 
9095 	list_destroy(l2arc_dev_list);
9096 	list_destroy(l2arc_free_on_write);
9097 }
9098 
9099 void
l2arc_start(void)9100 l2arc_start(void)
9101 {
9102 	if (!(spa_mode_global & FWRITE))
9103 		return;
9104 
9105 	(void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
9106 	    TS_RUN, minclsyspri);
9107 }
9108 
9109 void
l2arc_stop(void)9110 l2arc_stop(void)
9111 {
9112 	if (!(spa_mode_global & FWRITE))
9113 		return;
9114 
9115 	mutex_enter(&l2arc_feed_thr_lock);
9116 	cv_signal(&l2arc_feed_thr_cv);	/* kick thread out of startup */
9117 	l2arc_thread_exit = 1;
9118 	while (l2arc_thread_exit != 0)
9119 		cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
9120 	mutex_exit(&l2arc_feed_thr_lock);
9121 }
9122 
9123 /*
9124  * Punches out rebuild threads for the L2ARC devices in a spa. This should
9125  * be called after pool import from the spa async thread, since starting
9126  * these threads directly from spa_import() will make them part of the
9127  * "zpool import" context and delay process exit (and thus pool import).
9128  */
9129 void
l2arc_spa_rebuild_start(spa_t * spa)9130 l2arc_spa_rebuild_start(spa_t *spa)
9131 {
9132 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
9133 
9134 	/*
9135 	 * Locate the spa's l2arc devices and kick off rebuild threads.
9136 	 */
9137 	for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
9138 		l2arc_dev_t *dev =
9139 		    l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
9140 		if (dev == NULL) {
9141 			/* Don't attempt a rebuild if the vdev is UNAVAIL */
9142 			continue;
9143 		}
9144 		mutex_enter(&l2arc_rebuild_thr_lock);
9145 		if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
9146 			dev->l2ad_rebuild_began = B_TRUE;
9147 			(void) thread_create(NULL, 0,
9148 			    (void (*)(void *))l2arc_dev_rebuild_start,
9149 			    dev, 0, &p0, TS_RUN, minclsyspri);
9150 		}
9151 		mutex_exit(&l2arc_rebuild_thr_lock);
9152 	}
9153 }
9154 
9155 /*
9156  * Main entry point for L2ARC rebuilding.
9157  */
9158 static void
l2arc_dev_rebuild_start(l2arc_dev_t * dev)9159 l2arc_dev_rebuild_start(l2arc_dev_t *dev)
9160 {
9161 	VERIFY(!dev->l2ad_rebuild_cancel);
9162 	VERIFY(dev->l2ad_rebuild);
9163 	(void) l2arc_rebuild(dev);
9164 	mutex_enter(&l2arc_rebuild_thr_lock);
9165 	dev->l2ad_rebuild_began = B_FALSE;
9166 	dev->l2ad_rebuild = B_FALSE;
9167 	mutex_exit(&l2arc_rebuild_thr_lock);
9168 
9169 	thread_exit();
9170 }
9171 
9172 /*
9173  * This function implements the actual L2ARC metadata rebuild. It:
9174  * starts reading the log block chain and restores each block's contents
9175  * to memory (reconstructing arc_buf_hdr_t's).
9176  *
9177  * Operation stops under any of the following conditions:
9178  *
9179  * 1) We reach the end of the log block chain.
9180  * 2) We encounter *any* error condition (cksum errors, io errors)
9181  */
9182 static int
l2arc_rebuild(l2arc_dev_t * dev)9183 l2arc_rebuild(l2arc_dev_t *dev)
9184 {
9185 	vdev_t			*vd = dev->l2ad_vdev;
9186 	spa_t			*spa = vd->vdev_spa;
9187 	int			err = 0;
9188 	l2arc_dev_hdr_phys_t	*l2dhdr = dev->l2ad_dev_hdr;
9189 	l2arc_log_blk_phys_t	*this_lb, *next_lb;
9190 	zio_t			*this_io = NULL, *next_io = NULL;
9191 	l2arc_log_blkptr_t	lbps[2];
9192 	l2arc_lb_ptr_buf_t	*lb_ptr_buf;
9193 	boolean_t		lock_held;
9194 
9195 	this_lb = kmem_zalloc(sizeof (*this_lb), KM_SLEEP);
9196 	next_lb = kmem_zalloc(sizeof (*next_lb), KM_SLEEP);
9197 
9198 	/*
9199 	 * We prevent device removal while issuing reads to the device,
9200 	 * then during the rebuilding phases we drop this lock again so
9201 	 * that a spa_unload or device remove can be initiated - this is
9202 	 * safe, because the spa will signal us to stop before removing
9203 	 * our device and wait for us to stop.
9204 	 */
9205 	spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
9206 	lock_held = B_TRUE;
9207 
9208 	/*
9209 	 * Retrieve the persistent L2ARC device state.
9210 	 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9211 	 */
9212 	dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start);
9213 	dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr +
9214 	    L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop),
9215 	    dev->l2ad_start);
9216 	dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST);
9217 
9218 	/*
9219 	 * In case the zfs module parameter l2arc_rebuild_enabled is false
9220 	 * we do not start the rebuild process.
9221 	 */
9222 	if (!l2arc_rebuild_enabled)
9223 		goto out;
9224 
9225 	/* Prepare the rebuild process */
9226 	bcopy(l2dhdr->dh_start_lbps, lbps, sizeof (lbps));
9227 
9228 	/* Start the rebuild process */
9229 	for (;;) {
9230 		if (!l2arc_log_blkptr_valid(dev, &lbps[0]))
9231 			break;
9232 
9233 		if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1],
9234 		    this_lb, next_lb, this_io, &next_io)) != 0)
9235 			goto out;
9236 
9237 		/*
9238 		 * Our memory pressure valve. If the system is running low
9239 		 * on memory, rather than swamping memory with new ARC buf
9240 		 * hdrs, we opt not to rebuild the L2ARC. At this point,
9241 		 * however, we have already set up our L2ARC dev to chain in
9242 		 * new metadata log blocks, so the user may choose to offline/
9243 		 * online the L2ARC dev at a later time (or re-import the pool)
9244 		 * to reconstruct it (when there's less memory pressure).
9245 		 */
9246 		if (l2arc_hdr_limit_reached()) {
9247 			ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
9248 			cmn_err(CE_NOTE, "System running low on memory, "
9249 			    "aborting L2ARC rebuild.");
9250 			err = SET_ERROR(ENOMEM);
9251 			goto out;
9252 		}
9253 
9254 		spa_config_exit(spa, SCL_L2ARC, vd);
9255 		lock_held = B_FALSE;
9256 
9257 		/*
9258 		 * Now that we know that the next_lb checks out alright, we
9259 		 * can start reconstruction from this log block.
9260 		 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9261 		 */
9262 		uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop);
9263 		l2arc_log_blk_restore(dev, this_lb, asize);
9264 
9265 		/*
9266 		 * log block restored, include its pointer in the list of
9267 		 * pointers to log blocks present in the L2ARC device.
9268 		 */
9269 		lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
9270 		lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t),
9271 		    KM_SLEEP);
9272 		bcopy(&lbps[0], lb_ptr_buf->lb_ptr,
9273 		    sizeof (l2arc_log_blkptr_t));
9274 		mutex_enter(&dev->l2ad_mtx);
9275 		list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf);
9276 		ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
9277 		ARCSTAT_BUMP(arcstat_l2_log_blk_count);
9278 		zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
9279 		zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
9280 		mutex_exit(&dev->l2ad_mtx);
9281 		vdev_space_update(vd, asize, 0, 0);
9282 
9283 		/* BEGIN CSTYLED */
9284 		/*
9285 		 * Protection against loops of log blocks:
9286 		 *
9287 		 *				       l2ad_hand  l2ad_evict
9288 		 *                                         V          V
9289 		 * l2ad_start |=======================================| l2ad_end
9290 		 *             -----|||----|||---|||----|||
9291 		 *                  (3)    (2)   (1)    (0)
9292 		 *             ---|||---|||----|||---|||
9293 		 *		  (7)   (6)    (5)   (4)
9294 		 *
9295 		 * In this situation the pointer of log block (4) passes
9296 		 * l2arc_log_blkptr_valid() but the log block should not be
9297 		 * restored as it is overwritten by the payload of log block
9298 		 * (0). Only log blocks (0)-(3) should be restored. We check
9299 		 * whether l2ad_evict lies in between the payload starting
9300 		 * offset of the next log block (lbps[1].lbp_payload_start)
9301 		 * and the payload starting offset of the present log block
9302 		 * (lbps[0].lbp_payload_start). If true and this isn't the
9303 		 * first pass, we are looping from the beginning and we should
9304 		 * stop.
9305 		 */
9306 		/* END CSTYLED */
9307 		if (l2arc_range_check_overlap(lbps[1].lbp_payload_start,
9308 		    lbps[0].lbp_payload_start, dev->l2ad_evict) &&
9309 		    !dev->l2ad_first)
9310 			goto out;
9311 
9312 		for (;;) {
9313 			mutex_enter(&l2arc_rebuild_thr_lock);
9314 			if (dev->l2ad_rebuild_cancel) {
9315 				dev->l2ad_rebuild = B_FALSE;
9316 				cv_signal(&l2arc_rebuild_thr_cv);
9317 				mutex_exit(&l2arc_rebuild_thr_lock);
9318 				err = SET_ERROR(ECANCELED);
9319 				goto out;
9320 			}
9321 			mutex_exit(&l2arc_rebuild_thr_lock);
9322 			if (spa_config_tryenter(spa, SCL_L2ARC, vd,
9323 			    RW_READER)) {
9324 				lock_held = B_TRUE;
9325 				break;
9326 			}
9327 			/*
9328 			 * L2ARC config lock held by somebody in writer,
9329 			 * possibly due to them trying to remove us. They'll
9330 			 * likely to want us to shut down, so after a little
9331 			 * delay, we check l2ad_rebuild_cancel and retry
9332 			 * the lock again.
9333 			 */
9334 			delay(1);
9335 		}
9336 
9337 		/*
9338 		 * Continue with the next log block.
9339 		 */
9340 		lbps[0] = lbps[1];
9341 		lbps[1] = this_lb->lb_prev_lbp;
9342 		PTR_SWAP(this_lb, next_lb);
9343 		this_io = next_io;
9344 		next_io = NULL;
9345 	}
9346 
9347 	if (this_io != NULL)
9348 		l2arc_log_blk_fetch_abort(this_io);
9349 out:
9350 	if (next_io != NULL)
9351 		l2arc_log_blk_fetch_abort(next_io);
9352 	kmem_free(this_lb, sizeof (*this_lb));
9353 	kmem_free(next_lb, sizeof (*next_lb));
9354 
9355 	if (!l2arc_rebuild_enabled) {
9356 		spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9357 		    "disabled");
9358 	} else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) {
9359 		ARCSTAT_BUMP(arcstat_l2_rebuild_success);
9360 		spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9361 		    "successful, restored %llu blocks",
9362 		    (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
9363 	} else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) {
9364 		/*
9365 		 * No error but also nothing restored, meaning the lbps array
9366 		 * in the device header points to invalid/non-present log
9367 		 * blocks. Reset the header.
9368 		 */
9369 		spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9370 		    "no valid log blocks");
9371 		bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
9372 		l2arc_dev_hdr_update(dev);
9373 	} else if (err == ECANCELED) {
9374 		/*
9375 		 * In case the rebuild was canceled do not log to spa history
9376 		 * log as the pool may be in the process of being removed.
9377 		 */
9378 		zfs_dbgmsg("L2ARC rebuild aborted, restored %llu blocks",
9379 		    zfs_refcount_count(&dev->l2ad_lb_count));
9380 	} else if (err != 0) {
9381 		spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9382 		    "aborted, restored %llu blocks",
9383 		    (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
9384 	}
9385 
9386 	if (lock_held)
9387 		spa_config_exit(spa, SCL_L2ARC, vd);
9388 
9389 	return (err);
9390 }
9391 
9392 /*
9393  * Attempts to read the device header on the provided L2ARC device and writes
9394  * it to `hdr'. On success, this function returns 0, otherwise the appropriate
9395  * error code is returned.
9396  */
9397 static int
l2arc_dev_hdr_read(l2arc_dev_t * dev)9398 l2arc_dev_hdr_read(l2arc_dev_t *dev)
9399 {
9400 	int			err;
9401 	uint64_t		guid;
9402 	l2arc_dev_hdr_phys_t	*l2dhdr = dev->l2ad_dev_hdr;
9403 	const uint64_t		l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9404 	abd_t			*abd;
9405 
9406 	guid = spa_guid(dev->l2ad_vdev->vdev_spa);
9407 
9408 	abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
9409 
9410 	err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
9411 	    VDEV_LABEL_START_SIZE, l2dhdr_asize, abd,
9412 	    ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
9413 	    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
9414 	    ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
9415 	    ZIO_FLAG_SPECULATIVE, B_FALSE));
9416 
9417 	abd_put(abd);
9418 
9419 	if (err != 0) {
9420 		ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors);
9421 		zfs_dbgmsg("L2ARC IO error (%d) while reading device header, "
9422 		    "vdev guid: %llu", err, dev->l2ad_vdev->vdev_guid);
9423 		return (err);
9424 	}
9425 
9426 	if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC))
9427 		byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr));
9428 
9429 	if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC ||
9430 	    l2dhdr->dh_spa_guid != guid ||
9431 	    l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid ||
9432 	    l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION ||
9433 	    l2dhdr->dh_log_entries != dev->l2ad_log_entries ||
9434 	    l2dhdr->dh_end != dev->l2ad_end ||
9435 	    !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end,
9436 	    l2dhdr->dh_evict)) {
9437 		/*
9438 		 * Attempt to rebuild a device containing no actual dev hdr
9439 		 * or containing a header from some other pool or from another
9440 		 * version of persistent L2ARC.
9441 		 */
9442 		ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
9443 		return (SET_ERROR(ENOTSUP));
9444 	}
9445 
9446 	return (0);
9447 }
9448 
9449 /*
9450  * Reads L2ARC log blocks from storage and validates their contents.
9451  *
9452  * This function implements a simple fetcher to make sure that while
9453  * we're processing one buffer the L2ARC is already fetching the next
9454  * one in the chain.
9455  *
9456  * The arguments this_lp and next_lp point to the current and next log block
9457  * address in the block chain. Similarly, this_lb and next_lb hold the
9458  * l2arc_log_blk_phys_t's of the current and next L2ARC blk.
9459  *
9460  * The `this_io' and `next_io' arguments are used for block fetching.
9461  * When issuing the first blk IO during rebuild, you should pass NULL for
9462  * `this_io'. This function will then issue a sync IO to read the block and
9463  * also issue an async IO to fetch the next block in the block chain. The
9464  * fetched IO is returned in `next_io'. On subsequent calls to this
9465  * function, pass the value returned in `next_io' from the previous call
9466  * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO.
9467  * Prior to the call, you should initialize your `next_io' pointer to be
9468  * NULL. If no fetch IO was issued, the pointer is left set at NULL.
9469  *
9470  * On success, this function returns 0, otherwise it returns an appropriate
9471  * error code. On error the fetching IO is aborted and cleared before
9472  * returning from this function. Therefore, if we return `success', the
9473  * caller can assume that we have taken care of cleanup of fetch IOs.
9474  */
9475 static int
l2arc_log_blk_read(l2arc_dev_t * dev,const l2arc_log_blkptr_t * this_lbp,const l2arc_log_blkptr_t * next_lbp,l2arc_log_blk_phys_t * this_lb,l2arc_log_blk_phys_t * next_lb,zio_t * this_io,zio_t ** next_io)9476 l2arc_log_blk_read(l2arc_dev_t *dev,
9477     const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
9478     l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
9479     zio_t *this_io, zio_t **next_io)
9480 {
9481 	int		err = 0;
9482 	zio_cksum_t	cksum;
9483 	abd_t		*abd = NULL;
9484 	uint64_t	asize;
9485 
9486 	ASSERT(this_lbp != NULL && next_lbp != NULL);
9487 	ASSERT(this_lb != NULL && next_lb != NULL);
9488 	ASSERT(next_io != NULL && *next_io == NULL);
9489 	ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
9490 
9491 	/*
9492 	 * Check to see if we have issued the IO for this log block in a
9493 	 * previous run. If not, this is the first call, so issue it now.
9494 	 */
9495 	if (this_io == NULL) {
9496 		this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp,
9497 		    this_lb);
9498 	}
9499 
9500 	/*
9501 	 * Peek to see if we can start issuing the next IO immediately.
9502 	 */
9503 	if (l2arc_log_blkptr_valid(dev, next_lbp)) {
9504 		/*
9505 		 * Start issuing IO for the next log block early - this
9506 		 * should help keep the L2ARC device busy while we
9507 		 * decompress and restore this log block.
9508 		 */
9509 		*next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp,
9510 		    next_lb);
9511 	}
9512 
9513 	/* Wait for the IO to read this log block to complete */
9514 	if ((err = zio_wait(this_io)) != 0) {
9515 		ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
9516 		zfs_dbgmsg("L2ARC IO error (%d) while reading log block, "
9517 		    "offset: %llu, vdev guid: %llu", err, this_lbp->lbp_daddr,
9518 		    dev->l2ad_vdev->vdev_guid);
9519 		goto cleanup;
9520 	}
9521 
9522 	/*
9523 	 * Make sure the buffer checks out.
9524 	 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9525 	 */
9526 	asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop);
9527 	fletcher_4_native(this_lb, asize, NULL, &cksum);
9528 	if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
9529 		ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors);
9530 		zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, "
9531 		    "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu",
9532 		    this_lbp->lbp_daddr, dev->l2ad_vdev->vdev_guid,
9533 		    dev->l2ad_hand, dev->l2ad_evict);
9534 		err = SET_ERROR(ECKSUM);
9535 		goto cleanup;
9536 	}
9537 
9538 	/* Now we can take our time decoding this buffer */
9539 	switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) {
9540 	case ZIO_COMPRESS_OFF:
9541 		break;
9542 	case ZIO_COMPRESS_LZ4:
9543 		abd = abd_alloc_for_io(asize, B_TRUE);
9544 		abd_copy_from_buf_off(abd, this_lb, 0, asize);
9545 		if ((err = zio_decompress_data(
9546 		    L2BLK_GET_COMPRESS((this_lbp)->lbp_prop),
9547 		    abd, this_lb, asize, sizeof (*this_lb))) != 0) {
9548 			err = SET_ERROR(EINVAL);
9549 			goto cleanup;
9550 		}
9551 		break;
9552 	default:
9553 		err = SET_ERROR(EINVAL);
9554 		goto cleanup;
9555 	}
9556 	if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
9557 		byteswap_uint64_array(this_lb, sizeof (*this_lb));
9558 	if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
9559 		err = SET_ERROR(EINVAL);
9560 		goto cleanup;
9561 	}
9562 cleanup:
9563 	/* Abort an in-flight fetch I/O in case of error */
9564 	if (err != 0 && *next_io != NULL) {
9565 		l2arc_log_blk_fetch_abort(*next_io);
9566 		*next_io = NULL;
9567 	}
9568 	if (abd != NULL)
9569 		abd_free(abd);
9570 	return (err);
9571 }
9572 
9573 /*
9574  * Restores the payload of a log block to ARC. This creates empty ARC hdr
9575  * entries which only contain an l2arc hdr, essentially restoring the
9576  * buffers to their L2ARC evicted state. This function also updates space
9577  * usage on the L2ARC vdev to make sure it tracks restored buffers.
9578  */
9579 static void
l2arc_log_blk_restore(l2arc_dev_t * dev,const l2arc_log_blk_phys_t * lb,uint64_t lb_asize)9580 l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb,
9581     uint64_t lb_asize)
9582 {
9583 	uint64_t	size = 0, asize = 0;
9584 	uint64_t	log_entries = dev->l2ad_log_entries;
9585 
9586 	/*
9587 	 * Usually arc_adapt() is called only for data, not headers, but
9588 	 * since we may allocate significant amount of memory here, let ARC
9589 	 * grow its arc_c.
9590 	 */
9591 	arc_adapt(log_entries * HDR_L2ONLY_SIZE, arc_l2c_only);
9592 
9593 	for (int i = log_entries - 1; i >= 0; i--) {
9594 		/*
9595 		 * Restore goes in the reverse temporal direction to preserve
9596 		 * correct temporal ordering of buffers in the l2ad_buflist.
9597 		 * l2arc_hdr_restore also does a list_insert_tail instead of
9598 		 * list_insert_head on the l2ad_buflist:
9599 		 *
9600 		 *		LIST	l2ad_buflist		LIST
9601 		 *		HEAD  <------ (time) ------	TAIL
9602 		 * direction	+-----+-----+-----+-----+-----+    direction
9603 		 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
9604 		 * fill		+-----+-----+-----+-----+-----+
9605 		 *		^				^
9606 		 *		|				|
9607 		 *		|				|
9608 		 *	l2arc_feed_thread		l2arc_rebuild
9609 		 *	will place new bufs here	restores bufs here
9610 		 *
9611 		 * During l2arc_rebuild() the device is not used by
9612 		 * l2arc_feed_thread() as dev->l2ad_rebuild is set to true.
9613 		 */
9614 		size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop);
9615 		asize += vdev_psize_to_asize(dev->l2ad_vdev,
9616 		    L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop));
9617 		l2arc_hdr_restore(&lb->lb_entries[i], dev);
9618 	}
9619 
9620 	/*
9621 	 * Record rebuild stats:
9622 	 *	size		Logical size of restored buffers in the L2ARC
9623 	 *	asize		Aligned size of restored buffers in the L2ARC
9624 	 */
9625 	ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
9626 	ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize);
9627 	ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries);
9628 	ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize);
9629 	ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize);
9630 	ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
9631 }
9632 
9633 /*
9634  * Restores a single ARC buf hdr from a log entry. The ARC buffer is put
9635  * into a state indicating that it has been evicted to L2ARC.
9636  */
9637 static void
l2arc_hdr_restore(const l2arc_log_ent_phys_t * le,l2arc_dev_t * dev)9638 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
9639 {
9640 	arc_buf_hdr_t		*hdr, *exists;
9641 	kmutex_t		*hash_lock;
9642 	arc_buf_contents_t	type = L2BLK_GET_TYPE((le)->le_prop);
9643 	uint64_t		asize;
9644 
9645 	/*
9646 	 * Do all the allocation before grabbing any locks, this lets us
9647 	 * sleep if memory is full and we don't have to deal with failed
9648 	 * allocations.
9649 	 */
9650 	hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type,
9651 	    dev, le->le_dva, le->le_daddr,
9652 	    L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
9653 	    L2BLK_GET_COMPRESS((le)->le_prop),
9654 	    L2BLK_GET_PROTECTED((le)->le_prop),
9655 	    L2BLK_GET_PREFETCH((le)->le_prop),
9656 	    L2BLK_GET_STATE((le)->le_prop));
9657 	asize = vdev_psize_to_asize(dev->l2ad_vdev,
9658 	    L2BLK_GET_PSIZE((le)->le_prop));
9659 
9660 	/*
9661 	 * vdev_space_update() has to be called before arc_hdr_destroy() to
9662 	 * avoid underflow since the latter also calls vdev_space_update().
9663 	 */
9664 	l2arc_hdr_arcstats_increment(hdr);
9665 	vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9666 
9667 	mutex_enter(&dev->l2ad_mtx);
9668 	list_insert_tail(&dev->l2ad_buflist, hdr);
9669 	(void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
9670 	mutex_exit(&dev->l2ad_mtx);
9671 
9672 	exists = buf_hash_insert(hdr, &hash_lock);
9673 	if (exists) {
9674 		/* Buffer was already cached, no need to restore it. */
9675 		arc_hdr_destroy(hdr);
9676 		/*
9677 		 * If the buffer is already cached, check whether it has
9678 		 * L2ARC metadata. If not, enter them and update the flag.
9679 		 * This is important is case of onlining a cache device, since
9680 		 * we previously evicted all L2ARC metadata from ARC.
9681 		 */
9682 		if (!HDR_HAS_L2HDR(exists)) {
9683 			arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
9684 			exists->b_l2hdr.b_dev = dev;
9685 			exists->b_l2hdr.b_daddr = le->le_daddr;
9686 			exists->b_l2hdr.b_arcs_state =
9687 			    L2BLK_GET_STATE((le)->le_prop);
9688 			mutex_enter(&dev->l2ad_mtx);
9689 			list_insert_tail(&dev->l2ad_buflist, exists);
9690 			(void) zfs_refcount_add_many(&dev->l2ad_alloc,
9691 			    arc_hdr_size(exists), exists);
9692 			mutex_exit(&dev->l2ad_mtx);
9693 			l2arc_hdr_arcstats_increment(exists);
9694 			vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9695 		}
9696 		ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
9697 	}
9698 
9699 	mutex_exit(hash_lock);
9700 }
9701 
9702 /*
9703  * Starts an asynchronous read IO to read a log block. This is used in log
9704  * block reconstruction to start reading the next block before we are done
9705  * decoding and reconstructing the current block, to keep the l2arc device
9706  * nice and hot with read IO to process.
9707  * The returned zio will contain newly allocated memory buffers for the IO
9708  * data which should then be freed by the caller once the zio is no longer
9709  * needed (i.e. due to it having completed). If you wish to abort this
9710  * zio, you should do so using l2arc_log_blk_fetch_abort, which takes
9711  * care of disposing of the allocated buffers correctly.
9712  */
9713 static zio_t *
l2arc_log_blk_fetch(vdev_t * vd,const l2arc_log_blkptr_t * lbp,l2arc_log_blk_phys_t * lb)9714 l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
9715     l2arc_log_blk_phys_t *lb)
9716 {
9717 	uint32_t		asize;
9718 	zio_t			*pio;
9719 	l2arc_read_callback_t	*cb;
9720 
9721 	/* L2BLK_GET_PSIZE returns aligned size for log blocks */
9722 	asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
9723 	ASSERT(asize <= sizeof (l2arc_log_blk_phys_t));
9724 
9725 	cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP);
9726 	cb->l2rcb_abd = abd_get_from_buf(lb, asize);
9727 	pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb,
9728 	    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
9729 	    ZIO_FLAG_DONT_RETRY);
9730 	(void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize,
9731 	    cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL,
9732 	    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
9733 	    ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
9734 
9735 	return (pio);
9736 }
9737 
9738 /*
9739  * Aborts a zio returned from l2arc_log_blk_fetch and frees the data
9740  * buffers allocated for it.
9741  */
9742 static void
l2arc_log_blk_fetch_abort(zio_t * zio)9743 l2arc_log_blk_fetch_abort(zio_t *zio)
9744 {
9745 	(void) zio_wait(zio);
9746 }
9747 
9748 /*
9749  * Creates a zio to update the device header on an l2arc device.
9750  */
9751 static void
l2arc_dev_hdr_update(l2arc_dev_t * dev)9752 l2arc_dev_hdr_update(l2arc_dev_t *dev)
9753 {
9754 	l2arc_dev_hdr_phys_t	*l2dhdr = dev->l2ad_dev_hdr;
9755 	const uint64_t		l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9756 	abd_t			*abd;
9757 	int			err;
9758 
9759 	VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER));
9760 
9761 	l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC;
9762 	l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION;
9763 	l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
9764 	l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid;
9765 	l2dhdr->dh_log_entries = dev->l2ad_log_entries;
9766 	l2dhdr->dh_evict = dev->l2ad_evict;
9767 	l2dhdr->dh_start = dev->l2ad_start;
9768 	l2dhdr->dh_end = dev->l2ad_end;
9769 	l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize);
9770 	l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count);
9771 	l2dhdr->dh_flags = 0;
9772 	if (dev->l2ad_first)
9773 		l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
9774 
9775 	abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
9776 
9777 	err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev,
9778 	    VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL,
9779 	    NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE));
9780 
9781 	abd_put(abd);
9782 
9783 	if (err != 0) {
9784 		zfs_dbgmsg("L2ARC IO error (%d) while writing device header, "
9785 		    "vdev guid: %llu", err, dev->l2ad_vdev->vdev_guid);
9786 	}
9787 }
9788 
9789 /*
9790  * Commits a log block to the L2ARC device. This routine is invoked from
9791  * l2arc_write_buffers when the log block fills up.
9792  * This function allocates some memory to temporarily hold the serialized
9793  * buffer to be written. This is then released in l2arc_write_done.
9794  */
9795 static void
l2arc_log_blk_commit(l2arc_dev_t * dev,zio_t * pio,l2arc_write_callback_t * cb)9796 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb)
9797 {
9798 	l2arc_log_blk_phys_t	*lb = &dev->l2ad_log_blk;
9799 	l2arc_dev_hdr_phys_t	*l2dhdr = dev->l2ad_dev_hdr;
9800 	uint64_t		psize, asize;
9801 	zio_t			*wzio;
9802 	l2arc_lb_abd_buf_t	*abd_buf;
9803 	uint8_t			*tmpbuf;
9804 	l2arc_lb_ptr_buf_t	*lb_ptr_buf;
9805 
9806 	VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries);
9807 
9808 	tmpbuf = zio_buf_alloc(sizeof (*lb));
9809 	abd_buf = zio_buf_alloc(sizeof (*abd_buf));
9810 	abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb));
9811 	lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
9812 	lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP);
9813 
9814 	/* link the buffer into the block chain */
9815 	lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1];
9816 	lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
9817 
9818 	/*
9819 	 * l2arc_log_blk_commit() may be called multiple times during a single
9820 	 * l2arc_write_buffers() call. Save the allocated abd buffers in a list
9821 	 * so we can free them in l2arc_write_done() later on.
9822 	 */
9823 	list_insert_tail(&cb->l2wcb_abd_list, abd_buf);
9824 
9825 	/* try to compress the buffer */
9826 	psize = zio_compress_data(ZIO_COMPRESS_LZ4,
9827 	    abd_buf->abd, tmpbuf, sizeof (*lb));
9828 
9829 	/* a log block is never entirely zero */
9830 	ASSERT(psize != 0);
9831 	asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
9832 	ASSERT(asize <= sizeof (*lb));
9833 
9834 	/*
9835 	 * Update the start log block pointer in the device header to point
9836 	 * to the log block we're about to write.
9837 	 */
9838 	l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0];
9839 	l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
9840 	l2dhdr->dh_start_lbps[0].lbp_payload_asize =
9841 	    dev->l2ad_log_blk_payload_asize;
9842 	l2dhdr->dh_start_lbps[0].lbp_payload_start =
9843 	    dev->l2ad_log_blk_payload_start;
9844 	_NOTE(CONSTCOND)
9845 	L2BLK_SET_LSIZE(
9846 	    (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb));
9847 	L2BLK_SET_PSIZE(
9848 	    (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize);
9849 	L2BLK_SET_CHECKSUM(
9850 	    (&l2dhdr->dh_start_lbps[0])->lbp_prop,
9851 	    ZIO_CHECKSUM_FLETCHER_4);
9852 	if (asize < sizeof (*lb)) {
9853 		/* compression succeeded */
9854 		bzero(tmpbuf + psize, asize - psize);
9855 		L2BLK_SET_COMPRESS(
9856 		    (&l2dhdr->dh_start_lbps[0])->lbp_prop,
9857 		    ZIO_COMPRESS_LZ4);
9858 	} else {
9859 		/* compression failed */
9860 		bcopy(lb, tmpbuf, sizeof (*lb));
9861 		L2BLK_SET_COMPRESS(
9862 		    (&l2dhdr->dh_start_lbps[0])->lbp_prop,
9863 		    ZIO_COMPRESS_OFF);
9864 	}
9865 
9866 	/* checksum what we're about to write */
9867 	fletcher_4_native(tmpbuf, asize, NULL,
9868 	    &l2dhdr->dh_start_lbps[0].lbp_cksum);
9869 
9870 	abd_put(abd_buf->abd);
9871 
9872 	/* perform the write itself */
9873 	abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb));
9874 	abd_take_ownership_of_buf(abd_buf->abd, B_TRUE);
9875 	wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
9876 	    asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL,
9877 	    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
9878 	DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
9879 	(void) zio_nowait(wzio);
9880 
9881 	dev->l2ad_hand += asize;
9882 	/*
9883 	 * Include the committed log block's pointer  in the list of pointers
9884 	 * to log blocks present in the L2ARC device.
9885 	 */
9886 	bcopy(&l2dhdr->dh_start_lbps[0], lb_ptr_buf->lb_ptr,
9887 	    sizeof (l2arc_log_blkptr_t));
9888 	mutex_enter(&dev->l2ad_mtx);
9889 	list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf);
9890 	ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
9891 	ARCSTAT_BUMP(arcstat_l2_log_blk_count);
9892 	zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
9893 	zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
9894 	mutex_exit(&dev->l2ad_mtx);
9895 	vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9896 
9897 	/* bump the kstats */
9898 	ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
9899 	ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
9900 	ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize);
9901 	ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
9902 	    dev->l2ad_log_blk_payload_asize / asize);
9903 
9904 	/* start a new log block */
9905 	dev->l2ad_log_ent_idx = 0;
9906 	dev->l2ad_log_blk_payload_asize = 0;
9907 	dev->l2ad_log_blk_payload_start = 0;
9908 }
9909 
9910 /*
9911  * Validates an L2ARC log block address to make sure that it can be read
9912  * from the provided L2ARC device.
9913  */
9914 boolean_t
l2arc_log_blkptr_valid(l2arc_dev_t * dev,const l2arc_log_blkptr_t * lbp)9915 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
9916 {
9917 	/* L2BLK_GET_PSIZE returns aligned size for log blocks */
9918 	uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
9919 	uint64_t end = lbp->lbp_daddr + asize - 1;
9920 	uint64_t start = lbp->lbp_payload_start;
9921 	boolean_t evicted = B_FALSE;
9922 
9923 	/* BEGIN CSTYLED */
9924 	/*
9925 	 * A log block is valid if all of the following conditions are true:
9926 	 * - it fits entirely (including its payload) between l2ad_start and
9927 	 *   l2ad_end
9928 	 * - it has a valid size
9929 	 * - neither the log block itself nor part of its payload was evicted
9930 	 *   by l2arc_evict():
9931 	 *
9932 	 *		l2ad_hand          l2ad_evict
9933 	 *		|			 |	lbp_daddr
9934 	 *		|     start		 |	|  end
9935 	 *		|     |			 |	|  |
9936 	 *		V     V		         V	V  V
9937 	 *   l2ad_start ============================================ l2ad_end
9938 	 *                    --------------------------||||
9939 	 *				^		 ^
9940 	 *				|		log block
9941 	 *				payload
9942 	 */
9943 	/* END CSTYLED */
9944 	evicted =
9945 	    l2arc_range_check_overlap(start, end, dev->l2ad_hand) ||
9946 	    l2arc_range_check_overlap(start, end, dev->l2ad_evict) ||
9947 	    l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) ||
9948 	    l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end);
9949 
9950 	return (start >= dev->l2ad_start && end <= dev->l2ad_end &&
9951 	    asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) &&
9952 	    (!evicted || dev->l2ad_first));
9953 }
9954 
9955 /*
9956  * Inserts ARC buffer header `hdr' into the current L2ARC log block on
9957  * the device. The buffer being inserted must be present in L2ARC.
9958  * Returns B_TRUE if the L2ARC log block is full and needs to be committed
9959  * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
9960  */
9961 static boolean_t
l2arc_log_blk_insert(l2arc_dev_t * dev,const arc_buf_hdr_t * hdr)9962 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
9963 {
9964 	l2arc_log_blk_phys_t	*lb = &dev->l2ad_log_blk;
9965 	l2arc_log_ent_phys_t	*le;
9966 
9967 	if (dev->l2ad_log_entries == 0)
9968 		return (B_FALSE);
9969 
9970 	int index = dev->l2ad_log_ent_idx++;
9971 
9972 	ASSERT3S(index, <, dev->l2ad_log_entries);
9973 	ASSERT(HDR_HAS_L2HDR(hdr));
9974 
9975 	le = &lb->lb_entries[index];
9976 	bzero(le, sizeof (*le));
9977 	le->le_dva = hdr->b_dva;
9978 	le->le_birth = hdr->b_birth;
9979 	le->le_daddr = hdr->b_l2hdr.b_daddr;
9980 	if (index == 0)
9981 		dev->l2ad_log_blk_payload_start = le->le_daddr;
9982 	L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr));
9983 	L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr));
9984 	L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr));
9985 	L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
9986 	L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
9987 	L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
9988 	L2BLK_SET_STATE((le)->le_prop, hdr->b_l1hdr.b_state->arcs_state);
9989 
9990 	dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
9991 	    HDR_GET_PSIZE(hdr));
9992 
9993 	return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries);
9994 }
9995 
9996 /*
9997  * Checks whether a given L2ARC device address sits in a time-sequential
9998  * range. The trick here is that the L2ARC is a rotary buffer, so we can't
9999  * just do a range comparison, we need to handle the situation in which the
10000  * range wraps around the end of the L2ARC device. Arguments:
10001  *	bottom -- Lower end of the range to check (written to earlier).
10002  *	top    -- Upper end of the range to check (written to later).
10003  *	check  -- The address for which we want to determine if it sits in
10004  *		  between the top and bottom.
10005  *
10006  * The 3-way conditional below represents the following cases:
10007  *
10008  *	bottom < top : Sequentially ordered case:
10009  *	  <check>--------+-------------------+
10010  *	                 |  (overlap here?)  |
10011  *	 L2ARC dev       V                   V
10012  *	 |---------------<bottom>============<top>--------------|
10013  *
10014  *	bottom > top: Looped-around case:
10015  *	                      <check>--------+------------------+
10016  *	                                     |  (overlap here?) |
10017  *	 L2ARC dev                           V                  V
10018  *	 |===============<top>---------------<bottom>===========|
10019  *	 ^               ^
10020  *	 |  (or here?)   |
10021  *	 +---------------+---------<check>
10022  *
10023  *	top == bottom : Just a single address comparison.
10024  */
10025 boolean_t
l2arc_range_check_overlap(uint64_t bottom,uint64_t top,uint64_t check)10026 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
10027 {
10028 	if (bottom < top)
10029 		return (bottom <= check && check <= top);
10030 	else if (bottom > top)
10031 		return (check <= top || bottom <= check);
10032 	else
10033 		return (check == top);
10034 }
10035