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 2009 Sun Microsystems, Inc.  All rights reserved.
23 * Use is subject to license terms.
24 */
25
26/*
27 * Copyright (c) 2011, 2019 by Delphix. All rights reserved.
28 */
29
30#ifndef _SYS_METASLAB_IMPL_H
31#define	_SYS_METASLAB_IMPL_H
32
33#include <sys/metaslab.h>
34#include <sys/space_map.h>
35#include <sys/range_tree.h>
36#include <sys/vdev.h>
37#include <sys/txg.h>
38#include <sys/avl.h>
39#include <sys/multilist.h>
40
41#ifdef	__cplusplus
42extern "C" {
43#endif
44
45/*
46 * Metaslab allocation tracing record.
47 */
48typedef struct metaslab_alloc_trace {
49	list_node_t			mat_list_node;
50	metaslab_group_t		*mat_mg;
51	metaslab_t			*mat_msp;
52	uint64_t			mat_size;
53	uint64_t			mat_weight;
54	uint32_t			mat_dva_id;
55	uint64_t			mat_offset;
56	int					mat_allocator;
57} metaslab_alloc_trace_t;
58
59/*
60 * Used by the metaslab allocation tracing facility to indicate
61 * error conditions. These errors are stored to the offset member
62 * of the metaslab_alloc_trace_t record and displayed by mdb.
63 */
64typedef enum trace_alloc_type {
65	TRACE_ALLOC_FAILURE	= -1ULL,
66	TRACE_TOO_SMALL		= -2ULL,
67	TRACE_FORCE_GANG	= -3ULL,
68	TRACE_NOT_ALLOCATABLE	= -4ULL,
69	TRACE_GROUP_FAILURE	= -5ULL,
70	TRACE_ENOSPC		= -6ULL,
71	TRACE_CONDENSING	= -7ULL,
72	TRACE_VDEV_ERROR	= -8ULL,
73	TRACE_DISABLED		= -9ULL,
74} trace_alloc_type_t;
75
76#define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
77#define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
78#define	METASLAB_WEIGHT_CLAIM		(1ULL << 61)
79#define	METASLAB_WEIGHT_TYPE		(1ULL << 60)
80#define	METASLAB_ACTIVE_MASK		\
81	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY | \
82	METASLAB_WEIGHT_CLAIM)
83
84/*
85 * The metaslab weight is used to encode the amount of free space in a
86 * metaslab, such that the "best" metaslab appears first when sorting the
87 * metaslabs by weight. The weight (and therefore the "best" metaslab) can
88 * be determined in two different ways: by computing a weighted sum of all
89 * the free space in the metaslab (a space based weight) or by counting only
90 * the free segments of the largest size (a segment based weight). We prefer
91 * the segment based weight because it reflects how the free space is
92 * comprised, but we cannot always use it -- legacy pools do not have the
93 * space map histogram information necessary to determine the largest
94 * contiguous regions. Pools that have the space map histogram determine
95 * the segment weight by looking at each bucket in the histogram and
96 * determining the free space whose size in bytes is in the range:
97 *	[2^i, 2^(i+1))
98 * We then encode the largest index, i, that contains regions into the
99 * segment-weighted value.
100 *
101 * Space-based weight:
102 *
103 *      64      56      48      40      32      24      16      8       0
104 *      +-------+-------+-------+-------+-------+-------+-------+-------+
105 *      |PSC1|                  weighted-free space                     |
106 *      +-------+-------+-------+-------+-------+-------+-------+-------+
107 *
108 *	PS - indicates primary and secondary activation
109 *	C - indicates activation for claimed block zio
110 *	space - the fragmentation-weighted space
111 *
112 * Segment-based weight:
113 *
114 *      64      56      48      40      32      24      16      8       0
115 *      +-------+-------+-------+-------+-------+-------+-------+-------+
116 *      |PSC0| idx|            count of segments in region              |
117 *      +-------+-------+-------+-------+-------+-------+-------+-------+
118 *
119 *	PS - indicates primary and secondary activation
120 *	C - indicates activation for claimed block zio
121 *	idx - index for the highest bucket in the histogram
122 *	count - number of segments in the specified bucket
123 */
124#define	WEIGHT_GET_ACTIVE(weight)		BF64_GET((weight), 61, 3)
125#define	WEIGHT_SET_ACTIVE(weight, x)		BF64_SET((weight), 61, 3, x)
126
127#define	WEIGHT_IS_SPACEBASED(weight)		\
128	((weight) == 0 || BF64_GET((weight), 60, 1))
129#define	WEIGHT_SET_SPACEBASED(weight)		BF64_SET((weight), 60, 1, 1)
130
131/*
132 * These macros are only applicable to segment-based weighting.
133 */
134#define	WEIGHT_GET_INDEX(weight)		BF64_GET((weight), 54, 6)
135#define	WEIGHT_SET_INDEX(weight, x)		BF64_SET((weight), 54, 6, x)
136#define	WEIGHT_GET_COUNT(weight)		BF64_GET((weight), 0, 54)
137#define	WEIGHT_SET_COUNT(weight, x)		BF64_SET((weight), 0, 54, x)
138
139/*
140 * A metaslab class encompasses a category of allocatable top-level vdevs.
141 * Each top-level vdev is associated with a metaslab group which defines
142 * the allocatable region for that vdev. Examples of these categories include
143 * "normal" for data block allocations (i.e. main pool allocations) or "log"
144 * for allocations designated for intent log devices (i.e. slog devices).
145 * When a block allocation is requested from the SPA it is associated with a
146 * metaslab_class_t, and only top-level vdevs (i.e. metaslab groups) belonging
147 * to the class can be used to satisfy that request. Allocations are done
148 * by traversing the metaslab groups that are linked off of the mc_rotor field.
149 * This rotor points to the next metaslab group where allocations will be
150 * attempted. Allocating a block is a 3 step process -- select the metaslab
151 * group, select the metaslab, and then allocate the block. The metaslab
152 * class defines the low-level block allocator that will be used as the
153 * final step in allocation. These allocators are pluggable allowing each class
154 * to use a block allocator that best suits that class.
155 */
156struct metaslab_class {
157	kmutex_t		mc_lock;
158	spa_t			*mc_spa;
159	metaslab_group_t	*mc_rotor;
160	metaslab_ops_t		*mc_ops;
161	uint64_t		mc_aliquot;
162
163	/*
164	 * Track the number of metaslab groups that have been initialized
165	 * and can accept allocations. An initialized metaslab group is
166	 * one has been completely added to the config (i.e. we have
167	 * updated the MOS config and the space has been added to the pool).
168	 */
169	uint64_t		mc_groups;
170
171	/*
172	 * Toggle to enable/disable the allocation throttle.
173	 */
174	boolean_t		mc_alloc_throttle_enabled;
175
176	/*
177	 * The allocation throttle works on a reservation system. Whenever
178	 * an asynchronous zio wants to perform an allocation it must
179	 * first reserve the number of blocks that it wants to allocate.
180	 * If there aren't sufficient slots available for the pending zio
181	 * then that I/O is throttled until more slots free up. The current
182	 * number of reserved allocations is maintained by the mc_alloc_slots
183	 * refcount. The mc_alloc_max_slots value determines the maximum
184	 * number of allocations that the system allows. Gang blocks are
185	 * allowed to reserve slots even if we've reached the maximum
186	 * number of allocations allowed.
187	 */
188	uint64_t		*mc_alloc_max_slots;
189	zfs_refcount_t		*mc_alloc_slots;
190
191	uint64_t		mc_alloc_groups; /* # of allocatable groups */
192
193	uint64_t		mc_alloc;	/* total allocated space */
194	uint64_t		mc_deferred;	/* total deferred frees */
195	uint64_t		mc_space;	/* total space (alloc + free) */
196	uint64_t		mc_dspace;	/* total deflated space */
197	uint64_t		mc_histogram[RANGE_TREE_HISTOGRAM_SIZE];
198
199	/*
200	 * List of all loaded metaslabs in the class, sorted in order of most
201	 * recent use.
202	 */
203	multilist_t		*mc_metaslab_txg_list;
204};
205
206/*
207 * Metaslab groups encapsulate all the allocatable regions (i.e. metaslabs)
208 * of a top-level vdev. They are linked togther to form a circular linked
209 * list and can belong to only one metaslab class. Metaslab groups may become
210 * ineligible for allocations for a number of reasons such as limited free
211 * space, fragmentation, or going offline. When this happens the allocator will
212 * simply find the next metaslab group in the linked list and attempt
213 * to allocate from that group instead.
214 */
215struct metaslab_group {
216	kmutex_t		mg_lock;
217	metaslab_t		**mg_primaries;
218	metaslab_t		**mg_secondaries;
219	avl_tree_t		mg_metaslab_tree;
220	uint64_t		mg_aliquot;
221	boolean_t		mg_allocatable;		/* can we allocate? */
222	uint64_t		mg_ms_ready;
223
224	/*
225	 * A metaslab group is considered to be initialized only after
226	 * we have updated the MOS config and added the space to the pool.
227	 * We only allow allocation attempts to a metaslab group if it
228	 * has been initialized.
229	 */
230	boolean_t		mg_initialized;
231
232	uint64_t		mg_free_capacity;	/* percentage free */
233	int64_t			mg_bias;
234	int64_t			mg_activation_count;
235	metaslab_class_t	*mg_class;
236	vdev_t			*mg_vd;
237	taskq_t			*mg_taskq;
238	metaslab_group_t	*mg_prev;
239	metaslab_group_t	*mg_next;
240
241	/*
242	 * In order for the allocation throttle to function properly, we cannot
243	 * have too many IOs going to each disk by default; the throttle
244	 * operates by allocating more work to disks that finish quickly, so
245	 * allocating larger chunks to each disk reduces its effectiveness.
246	 * However, if the number of IOs going to each allocator is too small,
247	 * we will not perform proper aggregation at the vdev_queue layer,
248	 * also resulting in decreased performance. Therefore, we will use a
249	 * ramp-up strategy.
250	 *
251	 * Each allocator in each metaslab group has a current queue depth
252	 * (mg_alloc_queue_depth[allocator]) and a current max queue depth
253	 * (mg_cur_max_alloc_queue_depth[allocator]), and each metaslab group
254	 * has an absolute max queue depth (mg_max_alloc_queue_depth).  We
255	 * add IOs to an allocator until the mg_alloc_queue_depth for that
256	 * allocator hits the cur_max. Every time an IO completes for a given
257	 * allocator on a given metaslab group, we increment its cur_max until
258	 * it reaches mg_max_alloc_queue_depth. The cur_max resets every txg to
259	 * help protect against disks that decrease in performance over time.
260	 *
261	 * It's possible for an allocator to handle more allocations than
262	 * its max. This can occur when gang blocks are required or when other
263	 * groups are unable to handle their share of allocations.
264	 */
265	uint64_t		mg_max_alloc_queue_depth;
266	uint64_t		*mg_cur_max_alloc_queue_depth;
267	zfs_refcount_t		*mg_alloc_queue_depth;
268	int			mg_allocators;
269	/*
270	 * A metalab group that can no longer allocate the minimum block
271	 * size will set mg_no_free_space. Once a metaslab group is out
272	 * of space then its share of work must be distributed to other
273	 * groups.
274	 */
275	boolean_t		mg_no_free_space;
276
277	uint64_t		mg_allocations;
278	uint64_t		mg_failed_allocations;
279	uint64_t		mg_fragmentation;
280	uint64_t		mg_histogram[RANGE_TREE_HISTOGRAM_SIZE];
281
282	int			mg_ms_disabled;
283	boolean_t		mg_disabled_updating;
284	kmutex_t		mg_ms_disabled_lock;
285	kcondvar_t		mg_ms_disabled_cv;
286};
287
288/*
289 * This value defines the number of elements in the ms_lbas array. The value
290 * of 64 was chosen as it covers all power of 2 buckets up to UINT64_MAX.
291 * This is the equivalent of highbit(UINT64_MAX).
292 */
293#define	MAX_LBAS	64
294
295/*
296 * Each metaslab maintains a set of in-core trees to track metaslab
297 * operations.  The in-core free tree (ms_allocatable) contains the list of
298 * free segments which are eligible for allocation.  As blocks are
299 * allocated, the allocated segment are removed from the ms_allocatable and
300 * added to a per txg allocation tree (ms_allocating).  As blocks are
301 * freed, they are added to the free tree (ms_freeing).  These trees
302 * allow us to process all allocations and frees in syncing context
303 * where it is safe to update the on-disk space maps.  An additional set
304 * of in-core trees is maintained to track deferred frees
305 * (ms_defer).  Once a block is freed it will move from the
306 * ms_freed to the ms_defer tree.  A deferred free means that a block
307 * has been freed but cannot be used by the pool until TXG_DEFER_SIZE
308 * transactions groups later.  For example, a block that is freed in txg
309 * 50 will not be available for reallocation until txg 52 (50 +
310 * TXG_DEFER_SIZE).  This provides a safety net for uberblock rollback.
311 * A pool could be safely rolled back TXG_DEFERS_SIZE transactions
312 * groups and ensure that no block has been reallocated.
313 *
314 * The simplified transition diagram looks like this:
315 *
316 *
317 *      ALLOCATE
318 *         |
319 *         V
320 *    free segment (ms_allocatable) -> ms_allocating[4] -> (write to space map)
321 *         ^
322 *         |                        ms_freeing <--- FREE
323 *         |                             |
324 *         |                             v
325 *         |                         ms_freed
326 *         |                             |
327 *         +-------- ms_defer[2] <-------+-------> (write to space map)
328 *
329 *
330 * Each metaslab's space is tracked in a single space map in the MOS,
331 * which is only updated in syncing context.  Each time we sync a txg,
332 * we append the allocs and frees from that txg to the space map.  The
333 * pool space is only updated once all metaslabs have finished syncing.
334 *
335 * To load the in-core free tree we read the space map from disk.  This
336 * object contains a series of alloc and free records that are combined
337 * to make up the list of all free segments in this metaslab.  These
338 * segments are represented in-core by the ms_allocatable and are stored
339 * in an AVL tree.
340 *
341 * As the space map grows (as a result of the appends) it will
342 * eventually become space-inefficient.  When the metaslab's in-core
343 * free tree is zfs_condense_pct/100 times the size of the minimal
344 * on-disk representation, we rewrite it in its minimized form.  If a
345 * metaslab needs to condense then we must set the ms_condensing flag to
346 * ensure that allocations are not performed on the metaslab that is
347 * being written.
348 */
349struct metaslab {
350	/*
351	 * This is the main lock of the metaslab and its purpose is to
352	 * coordinate our allocations and frees [e.g metaslab_block_alloc(),
353	 * metaslab_free_concrete(), ..etc] with our various syncing
354	 * procedures [e.g. metaslab_sync(), metaslab_sync_done(), ..etc].
355	 *
356	 * The lock is also used during some miscellaneous operations like
357	 * using the metaslab's histogram for the metaslab group's histogram
358	 * aggregation, or marking the metaslab for initialization.
359	 */
360	kmutex_t	ms_lock;
361
362	/*
363	 * Acquired together with the ms_lock whenever we expect to
364	 * write to metaslab data on-disk (i.e flushing entries to
365	 * the metaslab's space map). It helps coordinate readers of
366	 * the metaslab's space map [see spa_vdev_remove_thread()]
367	 * with writers [see metaslab_sync() or metaslab_flush()].
368	 *
369	 * Note that metaslab_load(), even though a reader, uses
370	 * a completely different mechanism to deal with the reading
371	 * of the metaslab's space map based on ms_synced_length. That
372	 * said, the function still uses the ms_sync_lock after it
373	 * has read the ms_sm [see relevant comment in metaslab_load()
374	 * as to why].
375	 */
376	kmutex_t	ms_sync_lock;
377
378	kcondvar_t	ms_load_cv;
379	space_map_t	*ms_sm;
380	uint64_t	ms_id;
381	uint64_t	ms_start;
382	uint64_t	ms_size;
383	uint64_t	ms_fragmentation;
384
385	range_tree_t	*ms_allocating[TXG_SIZE];
386	range_tree_t	*ms_allocatable;
387	uint64_t	ms_allocated_this_txg;
388	uint64_t	ms_allocating_total;
389
390	/*
391	 * The following range trees are accessed only from syncing context.
392	 * ms_free*tree only have entries while syncing, and are empty
393	 * between syncs.
394	 */
395	range_tree_t	*ms_freeing;	/* to free this syncing txg */
396	range_tree_t	*ms_freed;	/* already freed this syncing txg */
397	range_tree_t	*ms_defer[TXG_DEFER_SIZE];
398	range_tree_t	*ms_checkpointing; /* to add to the checkpoint */
399
400	/*
401	 * The ms_trim tree is the set of allocatable segments which are
402	 * eligible for trimming. (When the metaslab is loaded, it's a
403	 * subset of ms_allocatable.)  It's kept in-core as long as the
404	 * autotrim property is set and is not vacated when the metaslab
405	 * is unloaded.  Its purpose is to aggregate freed ranges to
406	 * facilitate efficient trimming.
407	 */
408	range_tree_t	*ms_trim;
409
410	boolean_t	ms_condensing;	/* condensing? */
411	boolean_t	ms_condense_wanted;
412
413	/*
414	 * The number of consumers which have disabled the metaslab.
415	 */
416	uint64_t	ms_disabled;
417
418	/*
419	 * We must always hold the ms_lock when modifying ms_loaded
420	 * and ms_loading.
421	 */
422	boolean_t	ms_loaded;
423	boolean_t	ms_loading;
424	kcondvar_t	ms_flush_cv;
425	boolean_t	ms_flushing;
426
427	/*
428	 * The following histograms count entries that are in the
429	 * metaslab's space map (and its histogram) but are not in
430	 * ms_allocatable yet, because they are in ms_freed, ms_freeing,
431	 * or ms_defer[].
432	 *
433	 * When the metaslab is not loaded, its ms_weight needs to
434	 * reflect what is allocatable (i.e. what will be part of
435	 * ms_allocatable if it is loaded).  The weight is computed from
436	 * the spacemap histogram, but that includes ranges that are
437	 * not yet allocatable (because they are in ms_freed,
438	 * ms_freeing, or ms_defer[]).  Therefore, when calculating the
439	 * weight, we need to remove those ranges.
440	 *
441	 * The ranges in the ms_freed and ms_defer[] range trees are all
442	 * present in the spacemap.  However, the spacemap may have
443	 * multiple entries to represent a contiguous range, because it
444	 * is written across multiple sync passes, but the changes of
445	 * all sync passes are consolidated into the range trees.
446	 * Adjacent ranges that are freed in different sync passes of
447	 * one txg will be represented separately (as 2 or more entries)
448	 * in the space map (and its histogram), but these adjacent
449	 * ranges will be consolidated (represented as one entry) in the
450	 * ms_freed/ms_defer[] range trees (and their histograms).
451	 *
452	 * When calculating the weight, we can not simply subtract the
453	 * range trees' histograms from the spacemap's histogram,
454	 * because the range trees' histograms may have entries in
455	 * higher buckets than the spacemap, due to consolidation.
456	 * Instead we must subtract the exact entries that were added to
457	 * the spacemap's histogram.  ms_synchist and ms_deferhist[]
458	 * represent these exact entries, so we can subtract them from
459	 * the spacemap's histogram when calculating ms_weight.
460	 *
461	 * ms_synchist represents the same ranges as ms_freeing +
462	 * ms_freed, but without consolidation across sync passes.
463	 *
464	 * ms_deferhist[i] represents the same ranges as ms_defer[i],
465	 * but without consolidation across sync passes.
466	 */
467	uint64_t	ms_synchist[SPACE_MAP_HISTOGRAM_SIZE];
468	uint64_t	ms_deferhist[TXG_DEFER_SIZE][SPACE_MAP_HISTOGRAM_SIZE];
469
470	/*
471	 * Tracks the exact amount of allocated space of this metaslab
472	 * (and specifically the metaslab's space map) up to the most
473	 * recently completed sync pass [see usage in metaslab_sync()].
474	 */
475	uint64_t	ms_allocated_space;
476	int64_t		ms_deferspace;	/* sum of ms_defermap[] space	*/
477	uint64_t	ms_weight;	/* weight vs. others in group	*/
478	uint64_t	ms_activation_weight;	/* activation weight	*/
479
480	/*
481	 * Track of whenever a metaslab is selected for loading or allocation.
482	 * We use this value to determine how long the metaslab should
483	 * stay cached.
484	 */
485	uint64_t	ms_selected_txg;
486	/*
487	 * ms_load/unload_time can be used for performance monitoring
488	 * (e.g. by dtrace or mdb).
489	 */
490	hrtime_t	ms_load_time;	/* time last loaded */
491	hrtime_t	ms_unload_time;	/* time last unloaded */
492	hrtime_t	ms_selected_time; /* time last allocated from */
493
494	uint64_t	ms_alloc_txg;	/* last successful alloc (debug only) */
495	uint64_t	ms_max_size;	/* maximum allocatable size	*/
496
497	/*
498	 * -1 if it's not active in an allocator, otherwise set to the allocator
499	 * this metaslab is active for.
500	 */
501	int		ms_allocator;
502	boolean_t	ms_primary; /* Only valid if ms_allocator is not -1 */
503
504	/*
505	 * The metaslab block allocators can optionally use a size-ordered
506	 * range tree and/or an array of LBAs. Not all allocators use
507	 * this functionality. The ms_allocatable_by_size should always
508	 * contain the same number of segments as the ms_allocatable. The
509	 * only difference is that the ms_allocatable_by_size is ordered by
510	 * segment sizes.
511	 */
512	avl_tree_t	ms_allocatable_by_size;
513	avl_tree_t	ms_unflushed_frees_by_size;
514	uint64_t	ms_lbas[MAX_LBAS];
515
516	metaslab_group_t *ms_group;	/* metaslab group		*/
517	avl_node_t	ms_group_node;	/* node in metaslab group tree	*/
518	txg_node_t	ms_txg_node;	/* per-txg dirty metaslab links	*/
519	avl_node_t	ms_spa_txg_node; /* node in spa_metaslabs_by_txg */
520	/*
521	 * Node in metaslab class's selected txg list
522	 */
523	multilist_node_t	ms_class_txg_node;
524
525	/*
526	 * Allocs and frees that are committed to the vdev log spacemap but
527	 * not yet to this metaslab's spacemap.
528	 */
529	range_tree_t	*ms_unflushed_allocs;
530	range_tree_t	*ms_unflushed_frees;
531
532	/*
533	 * We have flushed entries up to but not including this TXG. In
534	 * other words, all changes from this TXG and onward should not
535	 * be in this metaslab's space map and must be read from the
536	 * log space maps.
537	 */
538	uint64_t	ms_unflushed_txg;
539
540	/* updated every time we are done syncing the metaslab's space map */
541	uint64_t	ms_synced_length;
542
543	boolean_t	ms_new;
544};
545
546typedef struct metaslab_unflushed_phys {
547	/* on-disk counterpart of ms_unflushed_txg */
548	uint64_t	msp_unflushed_txg;
549} metaslab_unflushed_phys_t;
550
551#ifdef	__cplusplus
552}
553#endif
554
555#endif	/* _SYS_METASLAB_IMPL_H */
556