xref: /illumos-gate/usr/src/uts/common/fs/zfs/abd.c (revision 6c448ad8)
1 /*
2  * This file and its contents are supplied under the terms of the
3  * Common Development and Distribution License ("CDDL"), version 1.0.
4  * You may only use this file in accordance with the terms of version
5  * 1.0 of the CDDL.
6  *
7  * A full copy of the text of the CDDL should have accompanied this
8  * source.  A copy of the CDDL is also available via the Internet at
9  * http://www.illumos.org/license/CDDL.
10  */
11 
12 /*
13  * Copyright (c) 2014 by Chunwei Chen. All rights reserved.
14  * Copyright (c) 2019 by Delphix. All rights reserved.
15  * Copyright 2020 Joyent, Inc.
16  * Copyright 2023 RackTop Systems, Inc.
17  */
18 
19 /*
20  * ARC buffer data (ABD).
21  *
22  * ABDs are an abstract data structure for the ARC which can use two
23  * different ways of storing the underlying data:
24  *
25  * (a) Linear buffer. In this case, all the data in the ABD is stored in one
26  *     contiguous buffer in memory (from a zio_[data_]buf_* kmem cache).
27  *
28  *         +-------------------+
29  *         | ABD (linear)      |
30  *         |   abd_flags = ... |
31  *         |   abd_size = ...  |     +--------------------------------+
32  *         |   abd_buf ------------->| raw buffer of size abd_size    |
33  *         +-------------------+     +--------------------------------+
34  *              no abd_chunks
35  *
36  * (b) Scattered buffer. In this case, the data in the ABD is split into
37  *     equal-sized chunks (from the abd_chunk_cache kmem_cache), with pointers
38  *     to the chunks recorded in an array at the end of the ABD structure.
39  *
40  *         +-------------------+
41  *         | ABD (scattered)   |
42  *         |   abd_flags = ... |
43  *         |   abd_size = ...  |
44  *         |   abd_offset = 0  |                           +-----------+
45  *         |   abd_chunks[0] ----------------------------->| chunk 0   |
46  *         |   abd_chunks[1] ---------------------+        +-----------+
47  *         |   ...             |                  |        +-----------+
48  *         |   abd_chunks[N-1] ---------+         +------->| chunk 1   |
49  *         +-------------------+        |                  +-----------+
50  *                                      |                      ...
51  *                                      |                  +-----------+
52  *                                      +----------------->| chunk N-1 |
53  *                                                         +-----------+
54  *
55  * Using a large proportion of scattered ABDs decreases ARC fragmentation since
56  * when we are at the limit of allocatable space, using equal-size chunks will
57  * allow us to quickly reclaim enough space for a new large allocation (assuming
58  * it is also scattered).
59  *
60  * In addition to directly allocating a linear or scattered ABD, it is also
61  * possible to create an ABD by requesting the "sub-ABD" starting at an offset
62  * within an existing ABD. In linear buffers this is simple (set abd_buf of
63  * the new ABD to the starting point within the original raw buffer), but
64  * scattered ABDs are a little more complex. The new ABD makes a copy of the
65  * relevant abd_chunks pointers (but not the underlying data). However, to
66  * provide arbitrary rather than only chunk-aligned starting offsets, it also
67  * tracks an abd_offset field which represents the starting point of the data
68  * within the first chunk in abd_chunks. For both linear and scattered ABDs,
69  * creating an offset ABD marks the original ABD as the offset's parent, and the
70  * original ABD's abd_children refcount is incremented. This data allows us to
71  * ensure the root ABD isn't deleted before its children.
72  *
73  * Most consumers should never need to know what type of ABD they're using --
74  * the ABD public API ensures that it's possible to transparently switch from
75  * using a linear ABD to a scattered one when doing so would be beneficial.
76  *
77  * If you need to use the data within an ABD directly, if you know it's linear
78  * (because you allocated it) you can use abd_to_buf() to access the underlying
79  * raw buffer. Otherwise, you should use one of the abd_borrow_buf* functions
80  * which will allocate a raw buffer if necessary. Use the abd_return_buf*
81  * functions to return any raw buffers that are no longer necessary when you're
82  * done using them.
83  *
84  * There are a variety of ABD APIs that implement basic buffer operations:
85  * compare, copy, read, write, and fill with zeroes. If you need a custom
86  * function which progressively accesses the whole ABD, use the abd_iterate_*
87  * functions.
88  */
89 
90 #include <sys/abd.h>
91 #include <sys/param.h>
92 #include <sys/zio.h>
93 #include <sys/zfs_context.h>
94 #include <sys/zfs_znode.h>
95 
96 typedef struct abd_stats {
97 	kstat_named_t abdstat_struct_size;
98 	kstat_named_t abdstat_scatter_cnt;
99 	kstat_named_t abdstat_scatter_data_size;
100 	kstat_named_t abdstat_scatter_chunk_waste;
101 	kstat_named_t abdstat_linear_cnt;
102 	kstat_named_t abdstat_linear_data_size;
103 } abd_stats_t;
104 
105 static abd_stats_t abd_stats = {
106 	/* Amount of memory occupied by all of the abd_t struct allocations */
107 	{ "struct_size",			KSTAT_DATA_UINT64 },
108 	/*
109 	 * The number of scatter ABDs which are currently allocated, excluding
110 	 * ABDs which don't own their data (for instance the ones which were
111 	 * allocated through abd_get_offset()).
112 	 */
113 	{ "scatter_cnt",			KSTAT_DATA_UINT64 },
114 	/* Amount of data stored in all scatter ABDs tracked by scatter_cnt */
115 	{ "scatter_data_size",			KSTAT_DATA_UINT64 },
116 	/*
117 	 * The amount of space wasted at the end of the last chunk across all
118 	 * scatter ABDs tracked by scatter_cnt.
119 	 */
120 	{ "scatter_chunk_waste",		KSTAT_DATA_UINT64 },
121 	/*
122 	 * The number of linear ABDs which are currently allocated, excluding
123 	 * ABDs which don't own their data (for instance the ones which were
124 	 * allocated through abd_get_offset() and abd_get_from_buf()). If an
125 	 * ABD takes ownership of its buf then it will become tracked.
126 	 */
127 	{ "linear_cnt",				KSTAT_DATA_UINT64 },
128 	/* Amount of data stored in all linear ABDs tracked by linear_cnt */
129 	{ "linear_data_size",			KSTAT_DATA_UINT64 },
130 };
131 
132 #define	ABDSTAT(stat)		(abd_stats.stat.value.ui64)
133 #define	ABDSTAT_INCR(stat, val) \
134 	atomic_add_64(&abd_stats.stat.value.ui64, (val))
135 #define	ABDSTAT_BUMP(stat)	ABDSTAT_INCR(stat, 1)
136 #define	ABDSTAT_BUMPDOWN(stat)	ABDSTAT_INCR(stat, -1)
137 
138 /*
139  * It is possible to make all future ABDs be linear by setting this to B_FALSE.
140  * Otherwise, ABDs are allocated scattered by default unless the caller uses
141  * abd_alloc_linear().
142  */
143 boolean_t zfs_abd_scatter_enabled = B_TRUE;
144 
145 /*
146  * zfs_abd_scatter_min_size is the minimum allocation size to use scatter
147  * ABD's.  Smaller allocations will use linear ABD's which uses
148  * zio_[data_]buf_alloc().
149  *
150  * Scatter ABD's use at least one page each, so sub-page allocations waste
151  * some space when allocated as scatter (e.g. 2KB scatter allocation wastes
152  * half of each page).  Using linear ABD's for small allocations means that
153  * they will be put on slabs which contain many allocations.  This can
154  * improve memory efficiency, but it also makes it much harder for ARC
155  * evictions to actually free pages, because all the buffers on one slab need
156  * to be freed in order for the slab (and underlying pages) to be freed.
157  * Typically, 512B and 1KB kmem caches have 16 buffers per slab, so it's
158  * possible for them to actually waste more memory than scatter (one page per
159  * buf = wasting 3/4 or 7/8th; one buf per slab = wasting 15/16th).
160  *
161  * Spill blocks are typically 512B and are heavily used on systems running
162  * selinux with the default dnode size and the `xattr=sa` property set.
163  *
164  * By default we use linear allocations for 512B and 1KB, and scatter
165  * allocations for larger (1.5KB and up).
166  */
167 int zfs_abd_scatter_min_size = 512 * 3;
168 
169 /*
170  * The size of the chunks ABD allocates. Because the sizes allocated from the
171  * kmem_cache can't change, this tunable can only be modified at boot. Changing
172  * it at runtime would cause ABD iteration to work incorrectly for ABDs which
173  * were allocated with the old size, so a safeguard has been put in place which
174  * will cause the machine to panic if you change it and try to access the data
175  * within a scattered ABD.
176  */
177 size_t zfs_abd_chunk_size = 4096;
178 
179 #ifdef _KERNEL
180 extern vmem_t *zio_alloc_arena;
181 #endif
182 
183 kmem_cache_t *abd_chunk_cache;
184 static kstat_t *abd_ksp;
185 
186 extern inline boolean_t abd_is_linear(abd_t *abd);
187 extern inline void abd_copy(abd_t *dabd, abd_t *sabd, size_t size);
188 extern inline void abd_copy_from_buf(abd_t *abd, const void *buf, size_t size);
189 extern inline void abd_copy_to_buf(void* buf, abd_t *abd, size_t size);
190 extern inline int abd_cmp_buf(abd_t *abd, const void *buf, size_t size);
191 extern inline void abd_zero(abd_t *abd, size_t size);
192 
193 static void *
abd_alloc_chunk()194 abd_alloc_chunk()
195 {
196 	void *c = kmem_cache_alloc(abd_chunk_cache, KM_PUSHPAGE);
197 	ASSERT3P(c, !=, NULL);
198 	return (c);
199 }
200 
201 static void
abd_free_chunk(void * c)202 abd_free_chunk(void *c)
203 {
204 	kmem_cache_free(abd_chunk_cache, c);
205 }
206 
207 void
abd_init(void)208 abd_init(void)
209 {
210 	vmem_t *data_alloc_arena = NULL;
211 
212 #ifdef _KERNEL
213 	data_alloc_arena = zio_alloc_arena;
214 #endif
215 
216 	/*
217 	 * Since ABD chunks do not appear in crash dumps, we pass KMC_NOTOUCH
218 	 * so that no allocator metadata is stored with the buffers.
219 	 */
220 	abd_chunk_cache = kmem_cache_create("abd_chunk", zfs_abd_chunk_size, 64,
221 	    NULL, NULL, NULL, NULL, data_alloc_arena, KMC_NOTOUCH);
222 
223 	abd_ksp = kstat_create("zfs", 0, "abdstats", "misc", KSTAT_TYPE_NAMED,
224 	    sizeof (abd_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
225 	if (abd_ksp != NULL) {
226 		abd_ksp->ks_data = &abd_stats;
227 		kstat_install(abd_ksp);
228 	}
229 }
230 
231 void
abd_fini(void)232 abd_fini(void)
233 {
234 	if (abd_ksp != NULL) {
235 		kstat_delete(abd_ksp);
236 		abd_ksp = NULL;
237 	}
238 
239 	kmem_cache_destroy(abd_chunk_cache);
240 	abd_chunk_cache = NULL;
241 }
242 
243 static inline size_t
abd_chunkcnt_for_bytes(size_t size)244 abd_chunkcnt_for_bytes(size_t size)
245 {
246 	return (P2ROUNDUP(size, zfs_abd_chunk_size) / zfs_abd_chunk_size);
247 }
248 
249 static inline size_t
abd_scatter_chunkcnt(abd_t * abd)250 abd_scatter_chunkcnt(abd_t *abd)
251 {
252 	ASSERT(!abd_is_linear(abd));
253 	return (abd_chunkcnt_for_bytes(
254 	    abd->abd_u.abd_scatter.abd_offset + abd->abd_size));
255 }
256 
257 static inline void
abd_verify(abd_t * abd)258 abd_verify(abd_t *abd)
259 {
260 	ASSERT3U(abd->abd_size, >, 0);
261 	ASSERT3U(abd->abd_size, <=, SPA_MAXBLOCKSIZE);
262 	ASSERT3U(abd->abd_flags, ==, abd->abd_flags & (ABD_FLAG_LINEAR |
263 	    ABD_FLAG_OWNER | ABD_FLAG_META));
264 	IMPLY(abd->abd_parent != NULL, !(abd->abd_flags & ABD_FLAG_OWNER));
265 	IMPLY(abd->abd_flags & ABD_FLAG_META, abd->abd_flags & ABD_FLAG_OWNER);
266 	if (abd_is_linear(abd)) {
267 		ASSERT3P(abd->abd_u.abd_linear.abd_buf, !=, NULL);
268 	} else {
269 		ASSERT3U(abd->abd_u.abd_scatter.abd_offset, <,
270 		    zfs_abd_chunk_size);
271 		size_t n = abd_scatter_chunkcnt(abd);
272 		for (int i = 0; i < n; i++) {
273 			ASSERT3P(
274 			    abd->abd_u.abd_scatter.abd_chunks[i], !=, NULL);
275 		}
276 	}
277 }
278 
279 static inline abd_t *
abd_alloc_struct(size_t chunkcnt)280 abd_alloc_struct(size_t chunkcnt)
281 {
282 	size_t size = offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]);
283 	abd_t *abd = kmem_alloc(size, KM_PUSHPAGE);
284 	ASSERT3P(abd, !=, NULL);
285 	ABDSTAT_INCR(abdstat_struct_size, size);
286 
287 	return (abd);
288 }
289 
290 static inline void
abd_free_struct(abd_t * abd)291 abd_free_struct(abd_t *abd)
292 {
293 	size_t chunkcnt = abd_is_linear(abd) ? 0 : abd_scatter_chunkcnt(abd);
294 	int size = offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]);
295 	kmem_free(abd, size);
296 	ABDSTAT_INCR(abdstat_struct_size, -size);
297 }
298 
299 /*
300  * Allocate an ABD, along with its own underlying data buffers. Use this if you
301  * don't care whether the ABD is linear or not.
302  */
303 abd_t *
abd_alloc(size_t size,boolean_t is_metadata)304 abd_alloc(size_t size, boolean_t is_metadata)
305 {
306 	/* see the comment above zfs_abd_scatter_min_size */
307 	if (!zfs_abd_scatter_enabled || size < zfs_abd_scatter_min_size)
308 		return (abd_alloc_linear(size, is_metadata));
309 
310 	VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
311 
312 	size_t n = abd_chunkcnt_for_bytes(size);
313 	abd_t *abd = abd_alloc_struct(n);
314 
315 	abd->abd_flags = ABD_FLAG_OWNER;
316 	if (is_metadata) {
317 		abd->abd_flags |= ABD_FLAG_META;
318 	}
319 	abd->abd_size = size;
320 	abd->abd_parent = NULL;
321 	zfs_refcount_create(&abd->abd_children);
322 
323 	abd->abd_u.abd_scatter.abd_offset = 0;
324 	abd->abd_u.abd_scatter.abd_chunk_size = zfs_abd_chunk_size;
325 
326 	for (int i = 0; i < n; i++) {
327 		void *c = abd_alloc_chunk();
328 		ASSERT3P(c, !=, NULL);
329 		abd->abd_u.abd_scatter.abd_chunks[i] = c;
330 	}
331 
332 	ABDSTAT_BUMP(abdstat_scatter_cnt);
333 	ABDSTAT_INCR(abdstat_scatter_data_size, size);
334 	ABDSTAT_INCR(abdstat_scatter_chunk_waste,
335 	    n * zfs_abd_chunk_size - size);
336 
337 	return (abd);
338 }
339 
340 static void
abd_free_scatter(abd_t * abd)341 abd_free_scatter(abd_t *abd)
342 {
343 	size_t n = abd_scatter_chunkcnt(abd);
344 	for (int i = 0; i < n; i++) {
345 		abd_free_chunk(abd->abd_u.abd_scatter.abd_chunks[i]);
346 	}
347 
348 	zfs_refcount_destroy(&abd->abd_children);
349 	ABDSTAT_BUMPDOWN(abdstat_scatter_cnt);
350 	ABDSTAT_INCR(abdstat_scatter_data_size, -(int)abd->abd_size);
351 	ABDSTAT_INCR(abdstat_scatter_chunk_waste,
352 	    abd->abd_size - n * zfs_abd_chunk_size);
353 
354 	abd_free_struct(abd);
355 }
356 
357 /*
358  * Allocate an ABD that must be linear, along with its own underlying data
359  * buffer. Only use this when it would be very annoying to write your ABD
360  * consumer with a scattered ABD.
361  */
362 abd_t *
abd_alloc_linear(size_t size,boolean_t is_metadata)363 abd_alloc_linear(size_t size, boolean_t is_metadata)
364 {
365 	abd_t *abd = abd_alloc_struct(0);
366 
367 	VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
368 
369 	abd->abd_flags = ABD_FLAG_LINEAR | ABD_FLAG_OWNER;
370 	if (is_metadata) {
371 		abd->abd_flags |= ABD_FLAG_META;
372 	}
373 	abd->abd_size = size;
374 	abd->abd_parent = NULL;
375 	zfs_refcount_create(&abd->abd_children);
376 
377 	if (is_metadata) {
378 		abd->abd_u.abd_linear.abd_buf = zio_buf_alloc(size);
379 	} else {
380 		abd->abd_u.abd_linear.abd_buf = zio_data_buf_alloc(size);
381 	}
382 
383 	ABDSTAT_BUMP(abdstat_linear_cnt);
384 	ABDSTAT_INCR(abdstat_linear_data_size, size);
385 
386 	return (abd);
387 }
388 
389 static void
abd_free_linear(abd_t * abd)390 abd_free_linear(abd_t *abd)
391 {
392 	if (abd->abd_flags & ABD_FLAG_META) {
393 		zio_buf_free(abd->abd_u.abd_linear.abd_buf, abd->abd_size);
394 	} else {
395 		zio_data_buf_free(abd->abd_u.abd_linear.abd_buf, abd->abd_size);
396 	}
397 
398 	zfs_refcount_destroy(&abd->abd_children);
399 	ABDSTAT_BUMPDOWN(abdstat_linear_cnt);
400 	ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size);
401 
402 	abd_free_struct(abd);
403 }
404 
405 /*
406  * Free an ABD. Only use this on ABDs allocated with abd_alloc() or
407  * abd_alloc_linear().
408  */
409 void
abd_free(abd_t * abd)410 abd_free(abd_t *abd)
411 {
412 	abd_verify(abd);
413 	ASSERT3P(abd->abd_parent, ==, NULL);
414 	ASSERT(abd->abd_flags & ABD_FLAG_OWNER);
415 	if (abd_is_linear(abd))
416 		abd_free_linear(abd);
417 	else
418 		abd_free_scatter(abd);
419 }
420 
421 /*
422  * Allocate an ABD of the same format (same metadata flag, same scatterize
423  * setting) as another ABD.
424  */
425 abd_t *
abd_alloc_sametype(abd_t * sabd,size_t size)426 abd_alloc_sametype(abd_t *sabd, size_t size)
427 {
428 	boolean_t is_metadata = (sabd->abd_flags & ABD_FLAG_META) != 0;
429 	if (abd_is_linear(sabd)) {
430 		return (abd_alloc_linear(size, is_metadata));
431 	} else {
432 		return (abd_alloc(size, is_metadata));
433 	}
434 }
435 
436 /*
437  * If we're going to use this ABD for doing I/O using the block layer, the
438  * consumer of the ABD data doesn't care if it's scattered or not, and we don't
439  * plan to store this ABD in memory for a long period of time, we should
440  * allocate the ABD type that requires the least data copying to do the I/O.
441  *
442  * Currently this is linear ABDs, however if ldi_strategy() can ever issue I/Os
443  * using a scatter/gather list we should switch to that and replace this call
444  * with vanilla abd_alloc().
445  */
446 abd_t *
abd_alloc_for_io(size_t size,boolean_t is_metadata)447 abd_alloc_for_io(size_t size, boolean_t is_metadata)
448 {
449 	return (abd_alloc_linear(size, is_metadata));
450 }
451 
452 /*
453  * Allocate a new ABD to point to offset off of sabd. It shares the underlying
454  * buffer data with sabd. Use abd_put() to free. sabd must not be freed while
455  * any derived ABDs exist.
456  */
457 /* ARGSUSED */
458 static inline abd_t *
abd_get_offset_impl(abd_t * sabd,size_t off,size_t size)459 abd_get_offset_impl(abd_t *sabd, size_t off, size_t size)
460 {
461 	abd_t *abd;
462 
463 	abd_verify(sabd);
464 	ASSERT3U(off, <=, sabd->abd_size);
465 
466 	if (abd_is_linear(sabd)) {
467 		abd = abd_alloc_struct(0);
468 
469 		/*
470 		 * Even if this buf is filesystem metadata, we only track that
471 		 * if we own the underlying data buffer, which is not true in
472 		 * this case. Therefore, we don't ever use ABD_FLAG_META here.
473 		 */
474 		abd->abd_flags = ABD_FLAG_LINEAR;
475 
476 		abd->abd_u.abd_linear.abd_buf =
477 		    (char *)sabd->abd_u.abd_linear.abd_buf + off;
478 	} else {
479 		size_t new_offset = sabd->abd_u.abd_scatter.abd_offset + off;
480 		size_t chunkcnt = abd_scatter_chunkcnt(sabd) -
481 		    (new_offset / zfs_abd_chunk_size);
482 
483 		abd = abd_alloc_struct(chunkcnt);
484 
485 		/*
486 		 * Even if this buf is filesystem metadata, we only track that
487 		 * if we own the underlying data buffer, which is not true in
488 		 * this case. Therefore, we don't ever use ABD_FLAG_META here.
489 		 */
490 		abd->abd_flags = 0;
491 
492 		abd->abd_u.abd_scatter.abd_offset =
493 		    new_offset % zfs_abd_chunk_size;
494 		abd->abd_u.abd_scatter.abd_chunk_size = zfs_abd_chunk_size;
495 
496 		/* Copy the scatterlist starting at the correct offset */
497 		(void) memcpy(&abd->abd_u.abd_scatter.abd_chunks,
498 		    &sabd->abd_u.abd_scatter.abd_chunks[new_offset /
499 		    zfs_abd_chunk_size],
500 		    chunkcnt * sizeof (void *));
501 	}
502 
503 	abd->abd_size = sabd->abd_size - off;
504 	abd->abd_parent = sabd;
505 	zfs_refcount_create(&abd->abd_children);
506 	(void) zfs_refcount_add_many(&sabd->abd_children, abd->abd_size, abd);
507 
508 	return (abd);
509 }
510 
511 abd_t *
abd_get_offset(abd_t * sabd,size_t off)512 abd_get_offset(abd_t *sabd, size_t off)
513 {
514 	size_t size = sabd->abd_size > off ? sabd->abd_size - off : 0;
515 
516 	VERIFY3U(size, >, 0);
517 
518 	return (abd_get_offset_impl(sabd, off, size));
519 }
520 
521 abd_t *
abd_get_offset_size(abd_t * sabd,size_t off,size_t size)522 abd_get_offset_size(abd_t *sabd, size_t off, size_t size)
523 {
524 	ASSERT3U(off + size, <=, sabd->abd_size);
525 
526 	return (abd_get_offset_impl(sabd, off, size));
527 }
528 
529 
530 /*
531  * Allocate a linear ABD structure for buf. You must free this with abd_put()
532  * since the resulting ABD doesn't own its own buffer.
533  */
534 abd_t *
abd_get_from_buf(void * buf,size_t size)535 abd_get_from_buf(void *buf, size_t size)
536 {
537 	abd_t *abd = abd_alloc_struct(0);
538 
539 	VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
540 
541 	/*
542 	 * Even if this buf is filesystem metadata, we only track that if we
543 	 * own the underlying data buffer, which is not true in this case.
544 	 * Therefore, we don't ever use ABD_FLAG_META here.
545 	 */
546 	abd->abd_flags = ABD_FLAG_LINEAR;
547 	abd->abd_size = size;
548 	abd->abd_parent = NULL;
549 	zfs_refcount_create(&abd->abd_children);
550 
551 	abd->abd_u.abd_linear.abd_buf = buf;
552 
553 	return (abd);
554 }
555 
556 /*
557  * Free an ABD allocated from abd_get_offset() or abd_get_from_buf(). Will not
558  * free the underlying scatterlist or buffer.
559  */
560 void
abd_put(abd_t * abd)561 abd_put(abd_t *abd)
562 {
563 	abd_verify(abd);
564 	ASSERT(!(abd->abd_flags & ABD_FLAG_OWNER));
565 
566 	if (abd->abd_parent != NULL) {
567 		(void) zfs_refcount_remove_many(&abd->abd_parent->abd_children,
568 		    abd->abd_size, abd);
569 	}
570 
571 	zfs_refcount_destroy(&abd->abd_children);
572 	abd_free_struct(abd);
573 }
574 
575 /*
576  * Get the raw buffer associated with a linear ABD.
577  */
578 void *
abd_to_buf(abd_t * abd)579 abd_to_buf(abd_t *abd)
580 {
581 	ASSERT(abd_is_linear(abd));
582 	abd_verify(abd);
583 	return (abd->abd_u.abd_linear.abd_buf);
584 }
585 
586 /*
587  * Borrow a raw buffer from an ABD without copying the contents of the ABD
588  * into the buffer. If the ABD is scattered, this will allocate a raw buffer
589  * whose contents are undefined. To copy over the existing data in the ABD, use
590  * abd_borrow_buf_copy() instead.
591  */
592 void *
abd_borrow_buf(abd_t * abd,size_t n)593 abd_borrow_buf(abd_t *abd, size_t n)
594 {
595 	void *buf;
596 	abd_verify(abd);
597 	ASSERT3U(abd->abd_size, >=, n);
598 	if (abd_is_linear(abd)) {
599 		buf = abd_to_buf(abd);
600 	} else if ((abd->abd_flags & ABD_FLAG_META) != 0) {
601 		buf = zio_buf_alloc(n);
602 	} else {
603 		buf = zio_data_buf_alloc(n);
604 	}
605 	(void) zfs_refcount_add_many(&abd->abd_children, n, buf);
606 
607 	return (buf);
608 }
609 
610 void *
abd_borrow_buf_copy(abd_t * abd,size_t n)611 abd_borrow_buf_copy(abd_t *abd, size_t n)
612 {
613 	void *buf = abd_borrow_buf(abd, n);
614 	if (!abd_is_linear(abd)) {
615 		abd_copy_to_buf(buf, abd, n);
616 	}
617 	return (buf);
618 }
619 
620 /*
621  * Return a borrowed raw buffer to an ABD. If the ABD is scattered, this will
622  * not change the contents of the ABD and will ASSERT that you didn't modify
623  * the buffer since it was borrowed. If you want any changes you made to buf to
624  * be copied back to abd, use abd_return_buf_copy() instead.
625  */
626 void
abd_return_buf(abd_t * abd,void * buf,size_t n)627 abd_return_buf(abd_t *abd, void *buf, size_t n)
628 {
629 	abd_verify(abd);
630 	ASSERT3U(abd->abd_size, >=, n);
631 	if (abd_is_linear(abd)) {
632 		ASSERT3P(buf, ==, abd_to_buf(abd));
633 	} else if ((abd->abd_flags & ABD_FLAG_META) != 0) {
634 		ASSERT0(abd_cmp_buf(abd, buf, n));
635 		zio_buf_free(buf, n);
636 	} else {
637 		ASSERT0(abd_cmp_buf(abd, buf, n));
638 		zio_data_buf_free(buf, n);
639 	}
640 	(void) zfs_refcount_remove_many(&abd->abd_children, n, buf);
641 }
642 
643 void
abd_return_buf_copy(abd_t * abd,void * buf,size_t n)644 abd_return_buf_copy(abd_t *abd, void *buf, size_t n)
645 {
646 	if (!abd_is_linear(abd)) {
647 		abd_copy_from_buf(abd, buf, n);
648 	}
649 	abd_return_buf(abd, buf, n);
650 }
651 
652 /*
653  * Give this ABD ownership of the buffer that it's storing. Can only be used on
654  * linear ABDs which were allocated via abd_get_from_buf(), or ones allocated
655  * with abd_alloc_linear() which subsequently released ownership of their buf
656  * with abd_release_ownership_of_buf().
657  */
658 void
abd_take_ownership_of_buf(abd_t * abd,boolean_t is_metadata)659 abd_take_ownership_of_buf(abd_t *abd, boolean_t is_metadata)
660 {
661 	ASSERT(abd_is_linear(abd));
662 	ASSERT(!(abd->abd_flags & ABD_FLAG_OWNER));
663 	abd_verify(abd);
664 
665 	abd->abd_flags |= ABD_FLAG_OWNER;
666 	if (is_metadata) {
667 		abd->abd_flags |= ABD_FLAG_META;
668 	}
669 
670 	ABDSTAT_BUMP(abdstat_linear_cnt);
671 	ABDSTAT_INCR(abdstat_linear_data_size, abd->abd_size);
672 }
673 
674 void
abd_release_ownership_of_buf(abd_t * abd)675 abd_release_ownership_of_buf(abd_t *abd)
676 {
677 	ASSERT(abd_is_linear(abd));
678 	ASSERT(abd->abd_flags & ABD_FLAG_OWNER);
679 	abd_verify(abd);
680 
681 	abd->abd_flags &= ~ABD_FLAG_OWNER;
682 	/* Disable this flag since we no longer own the data buffer */
683 	abd->abd_flags &= ~ABD_FLAG_META;
684 
685 	ABDSTAT_BUMPDOWN(abdstat_linear_cnt);
686 	ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size);
687 }
688 
689 struct abd_iter {
690 	abd_t		*iter_abd;	/* ABD being iterated through */
691 	size_t		iter_pos;	/* position (relative to abd_offset) */
692 	void		*iter_mapaddr;	/* addr corresponding to iter_pos */
693 	size_t		iter_mapsize;	/* length of data valid at mapaddr */
694 };
695 
696 static inline size_t
abd_iter_scatter_chunk_offset(struct abd_iter * aiter)697 abd_iter_scatter_chunk_offset(struct abd_iter *aiter)
698 {
699 	ASSERT(!abd_is_linear(aiter->iter_abd));
700 	return ((aiter->iter_abd->abd_u.abd_scatter.abd_offset +
701 	    aiter->iter_pos) % zfs_abd_chunk_size);
702 }
703 
704 static inline size_t
abd_iter_scatter_chunk_index(struct abd_iter * aiter)705 abd_iter_scatter_chunk_index(struct abd_iter *aiter)
706 {
707 	ASSERT(!abd_is_linear(aiter->iter_abd));
708 	return ((aiter->iter_abd->abd_u.abd_scatter.abd_offset +
709 	    aiter->iter_pos) / zfs_abd_chunk_size);
710 }
711 
712 /*
713  * Initialize the abd_iter.
714  */
715 static void
abd_iter_init(struct abd_iter * aiter,abd_t * abd)716 abd_iter_init(struct abd_iter *aiter, abd_t *abd)
717 {
718 	abd_verify(abd);
719 	aiter->iter_abd = abd;
720 	aiter->iter_pos = 0;
721 	aiter->iter_mapaddr = NULL;
722 	aiter->iter_mapsize = 0;
723 }
724 
725 /*
726  * Advance the iterator by a certain amount. Cannot be called when a chunk is
727  * in use. This can be safely called when the aiter has already exhausted, in
728  * which case this does nothing.
729  */
730 static void
abd_iter_advance(struct abd_iter * aiter,size_t amount)731 abd_iter_advance(struct abd_iter *aiter, size_t amount)
732 {
733 	ASSERT3P(aiter->iter_mapaddr, ==, NULL);
734 	ASSERT0(aiter->iter_mapsize);
735 
736 	/* There's nothing left to advance to, so do nothing */
737 	if (aiter->iter_pos == aiter->iter_abd->abd_size)
738 		return;
739 
740 	aiter->iter_pos += amount;
741 }
742 
743 /*
744  * Map the current chunk into aiter. This can be safely called when the aiter
745  * has already exhausted, in which case this does nothing.
746  */
747 static void
abd_iter_map(struct abd_iter * aiter)748 abd_iter_map(struct abd_iter *aiter)
749 {
750 	void *paddr;
751 	size_t offset = 0;
752 
753 	ASSERT3P(aiter->iter_mapaddr, ==, NULL);
754 	ASSERT0(aiter->iter_mapsize);
755 
756 	/* Panic if someone has changed zfs_abd_chunk_size */
757 	IMPLY(!abd_is_linear(aiter->iter_abd), zfs_abd_chunk_size ==
758 	    aiter->iter_abd->abd_u.abd_scatter.abd_chunk_size);
759 
760 	/* There's nothing left to iterate over, so do nothing */
761 	if (aiter->iter_pos == aiter->iter_abd->abd_size)
762 		return;
763 
764 	if (abd_is_linear(aiter->iter_abd)) {
765 		offset = aiter->iter_pos;
766 		aiter->iter_mapsize = aiter->iter_abd->abd_size - offset;
767 		paddr = aiter->iter_abd->abd_u.abd_linear.abd_buf;
768 	} else {
769 		size_t index = abd_iter_scatter_chunk_index(aiter);
770 		offset = abd_iter_scatter_chunk_offset(aiter);
771 		aiter->iter_mapsize = MIN(zfs_abd_chunk_size - offset,
772 		    aiter->iter_abd->abd_size - aiter->iter_pos);
773 		paddr = aiter->iter_abd->abd_u.abd_scatter.abd_chunks[index];
774 	}
775 	aiter->iter_mapaddr = (char *)paddr + offset;
776 }
777 
778 /*
779  * Unmap the current chunk from aiter. This can be safely called when the aiter
780  * has already exhausted, in which case this does nothing.
781  */
782 static void
abd_iter_unmap(struct abd_iter * aiter)783 abd_iter_unmap(struct abd_iter *aiter)
784 {
785 	/* There's nothing left to unmap, so do nothing */
786 	if (aiter->iter_pos == aiter->iter_abd->abd_size)
787 		return;
788 
789 	ASSERT3P(aiter->iter_mapaddr, !=, NULL);
790 	ASSERT3U(aiter->iter_mapsize, >, 0);
791 
792 	aiter->iter_mapaddr = NULL;
793 	aiter->iter_mapsize = 0;
794 }
795 
796 int
abd_iterate_func(abd_t * abd,size_t off,size_t size,abd_iter_func_t * func,void * private)797 abd_iterate_func(abd_t *abd, size_t off, size_t size,
798     abd_iter_func_t *func, void *private)
799 {
800 	int ret = 0;
801 	struct abd_iter aiter;
802 
803 	abd_verify(abd);
804 	ASSERT3U(off + size, <=, abd->abd_size);
805 
806 	abd_iter_init(&aiter, abd);
807 	abd_iter_advance(&aiter, off);
808 
809 	while (size > 0) {
810 		abd_iter_map(&aiter);
811 
812 		size_t len = MIN(aiter.iter_mapsize, size);
813 		ASSERT3U(len, >, 0);
814 
815 		ret = func(aiter.iter_mapaddr, len, private);
816 
817 		abd_iter_unmap(&aiter);
818 
819 		if (ret != 0)
820 			break;
821 
822 		size -= len;
823 		abd_iter_advance(&aiter, len);
824 	}
825 
826 	return (ret);
827 }
828 
829 struct buf_arg {
830 	void *arg_buf;
831 };
832 
833 static int
abd_copy_to_buf_off_cb(void * buf,size_t size,void * private)834 abd_copy_to_buf_off_cb(void *buf, size_t size, void *private)
835 {
836 	struct buf_arg *ba_ptr = private;
837 
838 	(void) memcpy(ba_ptr->arg_buf, buf, size);
839 	ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size;
840 
841 	return (0);
842 }
843 
844 /*
845  * Copy abd to buf. (off is the offset in abd.)
846  */
847 void
abd_copy_to_buf_off(void * buf,abd_t * abd,size_t off,size_t size)848 abd_copy_to_buf_off(void *buf, abd_t *abd, size_t off, size_t size)
849 {
850 	struct buf_arg ba_ptr = { buf };
851 
852 	(void) abd_iterate_func(abd, off, size, abd_copy_to_buf_off_cb,
853 	    &ba_ptr);
854 }
855 
856 static int
abd_cmp_buf_off_cb(void * buf,size_t size,void * private)857 abd_cmp_buf_off_cb(void *buf, size_t size, void *private)
858 {
859 	int ret;
860 	struct buf_arg *ba_ptr = private;
861 
862 	ret = memcmp(buf, ba_ptr->arg_buf, size);
863 	ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size;
864 
865 	return (ret);
866 }
867 
868 /*
869  * Compare the contents of abd to buf. (off is the offset in abd.)
870  */
871 int
abd_cmp_buf_off(abd_t * abd,const void * buf,size_t off,size_t size)872 abd_cmp_buf_off(abd_t *abd, const void *buf, size_t off, size_t size)
873 {
874 	struct buf_arg ba_ptr = { (void *) buf };
875 
876 	return (abd_iterate_func(abd, off, size, abd_cmp_buf_off_cb, &ba_ptr));
877 }
878 
879 static int
abd_copy_from_buf_off_cb(void * buf,size_t size,void * private)880 abd_copy_from_buf_off_cb(void *buf, size_t size, void *private)
881 {
882 	struct buf_arg *ba_ptr = private;
883 
884 	(void) memcpy(buf, ba_ptr->arg_buf, size);
885 	ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size;
886 
887 	return (0);
888 }
889 
890 /*
891  * Copy from buf to abd. (off is the offset in abd.)
892  */
893 void
abd_copy_from_buf_off(abd_t * abd,const void * buf,size_t off,size_t size)894 abd_copy_from_buf_off(abd_t *abd, const void *buf, size_t off, size_t size)
895 {
896 	struct buf_arg ba_ptr = { (void *) buf };
897 
898 	(void) abd_iterate_func(abd, off, size, abd_copy_from_buf_off_cb,
899 	    &ba_ptr);
900 }
901 
902 /*ARGSUSED*/
903 static int
abd_zero_off_cb(void * buf,size_t size,void * private)904 abd_zero_off_cb(void *buf, size_t size, void *private)
905 {
906 	(void) memset(buf, 0, size);
907 	return (0);
908 }
909 
910 /*
911  * Zero out the abd from a particular offset to the end.
912  */
913 void
abd_zero_off(abd_t * abd,size_t off,size_t size)914 abd_zero_off(abd_t *abd, size_t off, size_t size)
915 {
916 	(void) abd_iterate_func(abd, off, size, abd_zero_off_cb, NULL);
917 }
918 
919 /*
920  * Iterate over two ABDs and call func incrementally on the two ABDs' data in
921  * equal-sized chunks (passed to func as raw buffers). func could be called many
922  * times during this iteration.
923  */
924 int
abd_iterate_func2(abd_t * dabd,abd_t * sabd,size_t doff,size_t soff,size_t size,abd_iter_func2_t * func,void * private)925 abd_iterate_func2(abd_t *dabd, abd_t *sabd, size_t doff, size_t soff,
926     size_t size, abd_iter_func2_t *func, void *private)
927 {
928 	int ret = 0;
929 	struct abd_iter daiter, saiter;
930 
931 	abd_verify(dabd);
932 	abd_verify(sabd);
933 
934 	ASSERT3U(doff + size, <=, dabd->abd_size);
935 	ASSERT3U(soff + size, <=, sabd->abd_size);
936 
937 	abd_iter_init(&daiter, dabd);
938 	abd_iter_init(&saiter, sabd);
939 	abd_iter_advance(&daiter, doff);
940 	abd_iter_advance(&saiter, soff);
941 
942 	while (size > 0) {
943 		abd_iter_map(&daiter);
944 		abd_iter_map(&saiter);
945 
946 		size_t dlen = MIN(daiter.iter_mapsize, size);
947 		size_t slen = MIN(saiter.iter_mapsize, size);
948 		size_t len = MIN(dlen, slen);
949 		ASSERT(dlen > 0 || slen > 0);
950 
951 		ret = func(daiter.iter_mapaddr, saiter.iter_mapaddr, len,
952 		    private);
953 
954 		abd_iter_unmap(&saiter);
955 		abd_iter_unmap(&daiter);
956 
957 		if (ret != 0)
958 			break;
959 
960 		size -= len;
961 		abd_iter_advance(&daiter, len);
962 		abd_iter_advance(&saiter, len);
963 	}
964 
965 	return (ret);
966 }
967 
968 /*ARGSUSED*/
969 static int
abd_copy_off_cb(void * dbuf,void * sbuf,size_t size,void * private)970 abd_copy_off_cb(void *dbuf, void *sbuf, size_t size, void *private)
971 {
972 	(void) memcpy(dbuf, sbuf, size);
973 	return (0);
974 }
975 
976 /*
977  * Copy from sabd to dabd starting from soff and doff.
978  */
979 void
abd_copy_off(abd_t * dabd,abd_t * sabd,size_t doff,size_t soff,size_t size)980 abd_copy_off(abd_t *dabd, abd_t *sabd, size_t doff, size_t soff, size_t size)
981 {
982 	(void) abd_iterate_func2(dabd, sabd, doff, soff, size,
983 	    abd_copy_off_cb, NULL);
984 }
985 
986 /*ARGSUSED*/
987 static int
abd_cmp_cb(void * bufa,void * bufb,size_t size,void * private)988 abd_cmp_cb(void *bufa, void *bufb, size_t size, void *private)
989 {
990 	return (memcmp(bufa, bufb, size));
991 }
992 
993 /*
994  * Compares the first size bytes of two ABDs.
995  */
996 int
abd_cmp(abd_t * dabd,abd_t * sabd,size_t size)997 abd_cmp(abd_t *dabd, abd_t *sabd, size_t size)
998 {
999 	return (abd_iterate_func2(dabd, sabd, 0, 0, size, abd_cmp_cb, NULL));
1000 }
1001 
1002 /*
1003  * Iterate over code ABDs and a data ABD and call @func_raidz_gen.
1004  *
1005  * @cabds          parity ABDs, must have equal size
1006  * @dabd           data ABD. Can be NULL (in this case @dsize = 0)
1007  * @func_raidz_gen should be implemented so that its behaviour
1008  *                 is the same when taking linear and when taking scatter
1009  */
1010 void
abd_raidz_gen_iterate(abd_t ** cabds,abd_t * dabd,ssize_t csize,ssize_t dsize,const unsigned parity,void (* func_raidz_gen)(void **,const void *,size_t,size_t))1011 abd_raidz_gen_iterate(abd_t **cabds, abd_t *dabd,
1012     ssize_t csize, ssize_t dsize, const unsigned parity,
1013     void (*func_raidz_gen)(void **, const void *, size_t, size_t))
1014 {
1015 	int i;
1016 	ssize_t len, dlen;
1017 	struct abd_iter caiters[3];
1018 	struct abd_iter daiter = {0};
1019 	void *caddrs[3];
1020 
1021 	ASSERT3U(parity, <=, 3);
1022 
1023 	for (i = 0; i < parity; i++)
1024 		abd_iter_init(&caiters[i], cabds[i]);
1025 
1026 	if (dabd)
1027 		abd_iter_init(&daiter, dabd);
1028 
1029 	ASSERT3S(dsize, >=, 0);
1030 
1031 #ifdef _KERNEL
1032 	kpreempt_disable();
1033 #endif
1034 	while (csize > 0) {
1035 		len = csize;
1036 
1037 		if (dabd && dsize > 0)
1038 			abd_iter_map(&daiter);
1039 
1040 		for (i = 0; i < parity; i++) {
1041 			abd_iter_map(&caiters[i]);
1042 			caddrs[i] = caiters[i].iter_mapaddr;
1043 		}
1044 
1045 		switch (parity) {
1046 			case 3:
1047 				len = MIN(caiters[2].iter_mapsize, len);
1048 				/* falls through */
1049 			case 2:
1050 				len = MIN(caiters[1].iter_mapsize, len);
1051 				/* falls through */
1052 			case 1:
1053 				len = MIN(caiters[0].iter_mapsize, len);
1054 		}
1055 
1056 		/* must be progressive */
1057 		ASSERT3S(len, >, 0);
1058 
1059 		if (dabd && dsize > 0) {
1060 			/* this needs precise iter.length */
1061 			len = MIN(daiter.iter_mapsize, len);
1062 			len = MIN(dsize, len);
1063 			dlen = len;
1064 		} else
1065 			dlen = 0;
1066 
1067 		/* must be progressive */
1068 		ASSERT3S(len, >, 0);
1069 		/*
1070 		 * The iterated function likely will not do well if each
1071 		 * segment except the last one is not multiple of 512 (raidz).
1072 		 */
1073 		ASSERT3U(((uint64_t)len & 511ULL), ==, 0);
1074 
1075 		func_raidz_gen(caddrs, daiter.iter_mapaddr, len, dlen);
1076 
1077 		for (i = parity-1; i >= 0; i--) {
1078 			abd_iter_unmap(&caiters[i]);
1079 			abd_iter_advance(&caiters[i], len);
1080 		}
1081 
1082 		if (dabd && dsize > 0) {
1083 			abd_iter_unmap(&daiter);
1084 			abd_iter_advance(&daiter, dlen);
1085 			dsize -= dlen;
1086 		}
1087 
1088 		csize -= len;
1089 
1090 		ASSERT3S(dsize, >=, 0);
1091 		ASSERT3S(csize, >=, 0);
1092 	}
1093 #ifdef _KERNEL
1094 	kpreempt_enable();
1095 #endif
1096 }
1097 
1098 /*
1099  * Iterate over code ABDs and data reconstruction target ABDs and call
1100  * @func_raidz_rec. Function maps at most 6 pages atomically.
1101  *
1102  * @cabds           parity ABDs, must have equal size
1103  * @tabds           rec target ABDs, at most 3
1104  * @tsize           size of data target columns
1105  * @func_raidz_rec  expects syndrome data in target columns. Function
1106  *                  reconstructs data and overwrites target columns.
1107  */
1108 void
abd_raidz_rec_iterate(abd_t ** cabds,abd_t ** tabds,ssize_t tsize,const unsigned parity,void (* func_raidz_rec)(void ** t,const size_t tsize,void ** c,const unsigned * mul),const unsigned * mul)1109 abd_raidz_rec_iterate(abd_t **cabds, abd_t **tabds,
1110     ssize_t tsize, const unsigned parity,
1111     void (*func_raidz_rec)(void **t, const size_t tsize, void **c,
1112     const unsigned *mul),
1113     const unsigned *mul)
1114 {
1115 	int i;
1116 	ssize_t len;
1117 	struct abd_iter citers[3];
1118 	struct abd_iter xiters[3];
1119 	void *caddrs[3], *xaddrs[3];
1120 
1121 	ASSERT3U(parity, <=, 3);
1122 
1123 	for (i = 0; i < parity; i++) {
1124 		abd_iter_init(&citers[i], cabds[i]);
1125 		abd_iter_init(&xiters[i], tabds[i]);
1126 	}
1127 
1128 #ifdef _KERNEL
1129 	kpreempt_disable();
1130 #endif
1131 	while (tsize > 0) {
1132 
1133 		for (i = 0; i < parity; i++) {
1134 			abd_iter_map(&citers[i]);
1135 			abd_iter_map(&xiters[i]);
1136 			caddrs[i] = citers[i].iter_mapaddr;
1137 			xaddrs[i] = xiters[i].iter_mapaddr;
1138 		}
1139 
1140 		len = tsize;
1141 		switch (parity) {
1142 			case 3:
1143 				len = MIN(xiters[2].iter_mapsize, len);
1144 				len = MIN(citers[2].iter_mapsize, len);
1145 				/* falls through */
1146 			case 2:
1147 				len = MIN(xiters[1].iter_mapsize, len);
1148 				len = MIN(citers[1].iter_mapsize, len);
1149 				/* falls through */
1150 			case 1:
1151 				len = MIN(xiters[0].iter_mapsize, len);
1152 				len = MIN(citers[0].iter_mapsize, len);
1153 		}
1154 		/* must be progressive */
1155 		ASSERT3S(len, >, 0);
1156 		/*
1157 		 * The iterated function likely will not do well if each
1158 		 * segment except the last one is not multiple of 512 (raidz).
1159 		 */
1160 		ASSERT3U(((uint64_t)len & 511ULL), ==, 0);
1161 
1162 		func_raidz_rec(xaddrs, len, caddrs, mul);
1163 
1164 		for (i = parity-1; i >= 0; i--) {
1165 			abd_iter_unmap(&xiters[i]);
1166 			abd_iter_unmap(&citers[i]);
1167 			abd_iter_advance(&xiters[i], len);
1168 			abd_iter_advance(&citers[i], len);
1169 		}
1170 
1171 		tsize -= len;
1172 		ASSERT3S(tsize, >=, 0);
1173 	}
1174 #ifdef _KERNEL
1175 	kpreempt_enable();
1176 #endif
1177 }
1178