xref: /illumos-gate/usr/src/uts/common/fs/zfs/sys/metaslab_impl.h (revision 33c72b7598992897b94815b1f47b7b8077e53808)
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, 2018 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 
40 #ifdef	__cplusplus
41 extern "C" {
42 #endif
43 
44 /*
45  * Metaslab allocation tracing record.
46  */
47 typedef struct metaslab_alloc_trace {
48 	list_node_t			mat_list_node;
49 	metaslab_group_t		*mat_mg;
50 	metaslab_t			*mat_msp;
51 	uint64_t			mat_size;
52 	uint64_t			mat_weight;
53 	uint32_t			mat_dva_id;
54 	uint64_t			mat_offset;
55 	int					mat_allocator;
56 } metaslab_alloc_trace_t;
57 
58 /*
59  * Used by the metaslab allocation tracing facility to indicate
60  * error conditions. These errors are stored to the offset member
61  * of the metaslab_alloc_trace_t record and displayed by mdb.
62  */
63 typedef enum trace_alloc_type {
64 	TRACE_ALLOC_FAILURE	= -1ULL,
65 	TRACE_TOO_SMALL		= -2ULL,
66 	TRACE_FORCE_GANG	= -3ULL,
67 	TRACE_NOT_ALLOCATABLE	= -4ULL,
68 	TRACE_GROUP_FAILURE	= -5ULL,
69 	TRACE_ENOSPC		= -6ULL,
70 	TRACE_CONDENSING	= -7ULL,
71 	TRACE_VDEV_ERROR	= -8ULL,
72 	TRACE_INITIALIZING	= -9ULL
73 } trace_alloc_type_t;
74 
75 #define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
76 #define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
77 #define	METASLAB_WEIGHT_CLAIM		(1ULL << 61)
78 #define	METASLAB_WEIGHT_TYPE		(1ULL << 60)
79 #define	METASLAB_ACTIVE_MASK		\
80 	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY | \
81 	METASLAB_WEIGHT_CLAIM)
82 
83 /*
84  * The metaslab weight is used to encode the amount of free space in a
85  * metaslab, such that the "best" metaslab appears first when sorting the
86  * metaslabs by weight. The weight (and therefore the "best" metaslab) can
87  * be determined in two different ways: by computing a weighted sum of all
88  * the free space in the metaslab (a space based weight) or by counting only
89  * the free segments of the largest size (a segment based weight). We prefer
90  * the segment based weight because it reflects how the free space is
91  * comprised, but we cannot always use it -- legacy pools do not have the
92  * space map histogram information necessary to determine the largest
93  * contiguous regions. Pools that have the space map histogram determine
94  * the segment weight by looking at each bucket in the histogram and
95  * determining the free space whose size in bytes is in the range:
96  *	[2^i, 2^(i+1))
97  * We then encode the largest index, i, that contains regions into the
98  * segment-weighted value.
99  *
100  * Space-based weight:
101  *
102  *      64      56      48      40      32      24      16      8       0
103  *      +-------+-------+-------+-------+-------+-------+-------+-------+
104  *      |PSC1|                  weighted-free space                     |
105  *      +-------+-------+-------+-------+-------+-------+-------+-------+
106  *
107  *	PS - indicates primary and secondary activation
108  *	C - indicates activation for claimed block zio
109  *	space - the fragmentation-weighted space
110  *
111  * Segment-based weight:
112  *
113  *      64      56      48      40      32      24      16      8       0
114  *      +-------+-------+-------+-------+-------+-------+-------+-------+
115  *      |PSC0| idx|            count of segments in region              |
116  *      +-------+-------+-------+-------+-------+-------+-------+-------+
117  *
118  *	PS - indicates primary and secondary activation
119  *	C - indicates activation for claimed block zio
120  *	idx - index for the highest bucket in the histogram
121  *	count - number of segments in the specified bucket
122  */
123 #define	WEIGHT_GET_ACTIVE(weight)		BF64_GET((weight), 61, 3)
124 #define	WEIGHT_SET_ACTIVE(weight, x)		BF64_SET((weight), 61, 3, x)
125 
126 #define	WEIGHT_IS_SPACEBASED(weight)		\
127 	((weight) == 0 || BF64_GET((weight), 60, 1))
128 #define	WEIGHT_SET_SPACEBASED(weight)		BF64_SET((weight), 60, 1, 1)
129 
130 /*
131  * These macros are only applicable to segment-based weighting.
132  */
133 #define	WEIGHT_GET_INDEX(weight)		BF64_GET((weight), 54, 6)
134 #define	WEIGHT_SET_INDEX(weight, x)		BF64_SET((weight), 54, 6, x)
135 #define	WEIGHT_GET_COUNT(weight)		BF64_GET((weight), 0, 54)
136 #define	WEIGHT_SET_COUNT(weight, x)		BF64_SET((weight), 0, 54, x)
137 
138 /*
139  * A metaslab class encompasses a category of allocatable top-level vdevs.
140  * Each top-level vdev is associated with a metaslab group which defines
141  * the allocatable region for that vdev. Examples of these categories include
142  * "normal" for data block allocations (i.e. main pool allocations) or "log"
143  * for allocations designated for intent log devices (i.e. slog devices).
144  * When a block allocation is requested from the SPA it is associated with a
145  * metaslab_class_t, and only top-level vdevs (i.e. metaslab groups) belonging
146  * to the class can be used to satisfy that request. Allocations are done
147  * by traversing the metaslab groups that are linked off of the mc_rotor field.
148  * This rotor points to the next metaslab group where allocations will be
149  * attempted. Allocating a block is a 3 step process -- select the metaslab
150  * group, select the metaslab, and then allocate the block. The metaslab
151  * class defines the low-level block allocator that will be used as the
152  * final step in allocation. These allocators are pluggable allowing each class
153  * to use a block allocator that best suits that class.
154  */
155 struct metaslab_class {
156 	kmutex_t		mc_lock;
157 	spa_t			*mc_spa;
158 	metaslab_group_t	*mc_rotor;
159 	metaslab_ops_t		*mc_ops;
160 	uint64_t		mc_aliquot;
161 
162 	/*
163 	 * Track the number of metaslab groups that have been initialized
164 	 * and can accept allocations. An initialized metaslab group is
165 	 * one has been completely added to the config (i.e. we have
166 	 * updated the MOS config and the space has been added to the pool).
167 	 */
168 	uint64_t		mc_groups;
169 
170 	/*
171 	 * Toggle to enable/disable the allocation throttle.
172 	 */
173 	boolean_t		mc_alloc_throttle_enabled;
174 
175 	/*
176 	 * The allocation throttle works on a reservation system. Whenever
177 	 * an asynchronous zio wants to perform an allocation it must
178 	 * first reserve the number of blocks that it wants to allocate.
179 	 * If there aren't sufficient slots available for the pending zio
180 	 * then that I/O is throttled until more slots free up. The current
181 	 * number of reserved allocations is maintained by the mc_alloc_slots
182 	 * refcount. The mc_alloc_max_slots value determines the maximum
183 	 * number of allocations that the system allows. Gang blocks are
184 	 * allowed to reserve slots even if we've reached the maximum
185 	 * number of allocations allowed.
186 	 */
187 	uint64_t		*mc_alloc_max_slots;
188 	zfs_refcount_t		*mc_alloc_slots;
189 
190 	uint64_t		mc_alloc_groups; /* # of allocatable groups */
191 
192 	uint64_t		mc_alloc;	/* total allocated space */
193 	uint64_t		mc_deferred;	/* total deferred frees */
194 	uint64_t		mc_space;	/* total space (alloc + free) */
195 	uint64_t		mc_dspace;	/* total deflated space */
196 	uint64_t		mc_histogram[RANGE_TREE_HISTOGRAM_SIZE];
197 };
198 
199 /*
200  * Metaslab groups encapsulate all the allocatable regions (i.e. metaslabs)
201  * of a top-level vdev. They are linked togther to form a circular linked
202  * list and can belong to only one metaslab class. Metaslab groups may become
203  * ineligible for allocations for a number of reasons such as limited free
204  * space, fragmentation, or going offline. When this happens the allocator will
205  * simply find the next metaslab group in the linked list and attempt
206  * to allocate from that group instead.
207  */
208 struct metaslab_group {
209 	kmutex_t		mg_lock;
210 	metaslab_t		**mg_primaries;
211 	metaslab_t		**mg_secondaries;
212 	avl_tree_t		mg_metaslab_tree;
213 	uint64_t		mg_aliquot;
214 	boolean_t		mg_allocatable;		/* can we allocate? */
215 	uint64_t		mg_ms_ready;
216 
217 	/*
218 	 * A metaslab group is considered to be initialized only after
219 	 * we have updated the MOS config and added the space to the pool.
220 	 * We only allow allocation attempts to a metaslab group if it
221 	 * has been initialized.
222 	 */
223 	boolean_t		mg_initialized;
224 
225 	uint64_t		mg_free_capacity;	/* percentage free */
226 	int64_t			mg_bias;
227 	int64_t			mg_activation_count;
228 	metaslab_class_t	*mg_class;
229 	vdev_t			*mg_vd;
230 	taskq_t			*mg_taskq;
231 	metaslab_group_t	*mg_prev;
232 	metaslab_group_t	*mg_next;
233 
234 	/*
235 	 * In order for the allocation throttle to function properly, we cannot
236 	 * have too many IOs going to each disk by default; the throttle
237 	 * operates by allocating more work to disks that finish quickly, so
238 	 * allocating larger chunks to each disk reduces its effectiveness.
239 	 * However, if the number of IOs going to each allocator is too small,
240 	 * we will not perform proper aggregation at the vdev_queue layer,
241 	 * also resulting in decreased performance. Therefore, we will use a
242 	 * ramp-up strategy.
243 	 *
244 	 * Each allocator in each metaslab group has a current queue depth
245 	 * (mg_alloc_queue_depth[allocator]) and a current max queue depth
246 	 * (mg_cur_max_alloc_queue_depth[allocator]), and each metaslab group
247 	 * has an absolute max queue depth (mg_max_alloc_queue_depth).  We
248 	 * add IOs to an allocator until the mg_alloc_queue_depth for that
249 	 * allocator hits the cur_max. Every time an IO completes for a given
250 	 * allocator on a given metaslab group, we increment its cur_max until
251 	 * it reaches mg_max_alloc_queue_depth. The cur_max resets every txg to
252 	 * help protect against disks that decrease in performance over time.
253 	 *
254 	 * It's possible for an allocator to handle more allocations than
255 	 * its max. This can occur when gang blocks are required or when other
256 	 * groups are unable to handle their share of allocations.
257 	 */
258 	uint64_t		mg_max_alloc_queue_depth;
259 	uint64_t		*mg_cur_max_alloc_queue_depth;
260 	zfs_refcount_t		*mg_alloc_queue_depth;
261 	int			mg_allocators;
262 	/*
263 	 * A metalab group that can no longer allocate the minimum block
264 	 * size will set mg_no_free_space. Once a metaslab group is out
265 	 * of space then its share of work must be distributed to other
266 	 * groups.
267 	 */
268 	boolean_t		mg_no_free_space;
269 
270 	uint64_t		mg_allocations;
271 	uint64_t		mg_failed_allocations;
272 	uint64_t		mg_fragmentation;
273 	uint64_t		mg_histogram[RANGE_TREE_HISTOGRAM_SIZE];
274 
275 	int			mg_ms_initializing;
276 	boolean_t		mg_initialize_updating;
277 	kmutex_t		mg_ms_initialize_lock;
278 	kcondvar_t		mg_ms_initialize_cv;
279 };
280 
281 /*
282  * This value defines the number of elements in the ms_lbas array. The value
283  * of 64 was chosen as it covers all power of 2 buckets up to UINT64_MAX.
284  * This is the equivalent of highbit(UINT64_MAX).
285  */
286 #define	MAX_LBAS	64
287 
288 /*
289  * Each metaslab maintains a set of in-core trees to track metaslab
290  * operations.  The in-core free tree (ms_allocatable) contains the list of
291  * free segments which are eligible for allocation.  As blocks are
292  * allocated, the allocated segment are removed from the ms_allocatable and
293  * added to a per txg allocation tree (ms_allocating).  As blocks are
294  * freed, they are added to the free tree (ms_freeing).  These trees
295  * allow us to process all allocations and frees in syncing context
296  * where it is safe to update the on-disk space maps.  An additional set
297  * of in-core trees is maintained to track deferred frees
298  * (ms_defer).  Once a block is freed it will move from the
299  * ms_freed to the ms_defer tree.  A deferred free means that a block
300  * has been freed but cannot be used by the pool until TXG_DEFER_SIZE
301  * transactions groups later.  For example, a block that is freed in txg
302  * 50 will not be available for reallocation until txg 52 (50 +
303  * TXG_DEFER_SIZE).  This provides a safety net for uberblock rollback.
304  * A pool could be safely rolled back TXG_DEFERS_SIZE transactions
305  * groups and ensure that no block has been reallocated.
306  *
307  * The simplified transition diagram looks like this:
308  *
309  *
310  *      ALLOCATE
311  *         |
312  *         V
313  *    free segment (ms_allocatable) -> ms_allocating[4] -> (write to space map)
314  *         ^
315  *         |                        ms_freeing <--- FREE
316  *         |                             |
317  *         |                             v
318  *         |                         ms_freed
319  *         |                             |
320  *         +-------- ms_defer[2] <-------+-------> (write to space map)
321  *
322  *
323  * Each metaslab's space is tracked in a single space map in the MOS,
324  * which is only updated in syncing context.  Each time we sync a txg,
325  * we append the allocs and frees from that txg to the space map.  The
326  * pool space is only updated once all metaslabs have finished syncing.
327  *
328  * To load the in-core free tree we read the space map from disk.  This
329  * object contains a series of alloc and free records that are combined
330  * to make up the list of all free segments in this metaslab.  These
331  * segments are represented in-core by the ms_allocatable and are stored
332  * in an AVL tree.
333  *
334  * As the space map grows (as a result of the appends) it will
335  * eventually become space-inefficient.  When the metaslab's in-core
336  * free tree is zfs_condense_pct/100 times the size of the minimal
337  * on-disk representation, we rewrite it in its minimized form.  If a
338  * metaslab needs to condense then we must set the ms_condensing flag to
339  * ensure that allocations are not performed on the metaslab that is
340  * being written.
341  */
342 struct metaslab {
343 	/*
344 	 * This is the main lock of the metaslab and its purpose is to
345 	 * coordinate our allocations and frees [e.g metaslab_block_alloc(),
346 	 * metaslab_free_concrete(), ..etc] with our various syncing
347 	 * procedures [e.g. metaslab_sync(), metaslab_sync_done(), ..etc].
348 	 *
349 	 * The lock is also used during some miscellaneous operations like
350 	 * using the metaslab's histogram for the metaslab group's histogram
351 	 * aggregation, or marking the metaslab for initialization.
352 	 */
353 	kmutex_t	ms_lock;
354 
355 	/*
356 	 * Acquired together with the ms_lock whenever we expect to
357 	 * write to metaslab data on-disk (i.e flushing entries to
358 	 * the metaslab's space map). It helps coordinate readers of
359 	 * the metaslab's space map [see spa_vdev_remove_thread()]
360 	 * with writers [see metaslab_sync()].
361 	 *
362 	 * Note that metaslab_load(), even though a reader, uses
363 	 * a completely different mechanism to deal with the reading
364 	 * of the metaslab's space map based on ms_synced_length. That
365 	 * said, the function still uses the ms_sync_lock after it
366 	 * has read the ms_sm [see relevant comment in metaslab_load()
367 	 * as to why].
368 	 */
369 	kmutex_t	ms_sync_lock;
370 
371 	kcondvar_t	ms_load_cv;
372 	space_map_t	*ms_sm;
373 	uint64_t	ms_id;
374 	uint64_t	ms_start;
375 	uint64_t	ms_size;
376 	uint64_t	ms_fragmentation;
377 
378 	range_tree_t	*ms_allocating[TXG_SIZE];
379 	range_tree_t	*ms_allocatable;
380 	uint64_t	ms_allocated_this_txg;
381 
382 	/*
383 	 * The following range trees are accessed only from syncing context.
384 	 * ms_free*tree only have entries while syncing, and are empty
385 	 * between syncs.
386 	 */
387 	range_tree_t	*ms_freeing;	/* to free this syncing txg */
388 	range_tree_t	*ms_freed;	/* already freed this syncing txg */
389 	range_tree_t	*ms_defer[TXG_DEFER_SIZE];
390 	range_tree_t	*ms_checkpointing; /* to add to the checkpoint */
391 
392 	boolean_t	ms_condensing;	/* condensing? */
393 	boolean_t	ms_condense_wanted;
394 	uint64_t	ms_condense_checked_txg;
395 
396 	uint64_t	ms_initializing; /* leaves initializing this ms */
397 
398 	/*
399 	 * We must always hold the ms_lock when modifying ms_loaded
400 	 * and ms_loading.
401 	 */
402 	boolean_t	ms_loaded;
403 	boolean_t	ms_loading;
404 
405 	/*
406 	 * The following histograms count entries that are in the
407 	 * metaslab's space map (and its histogram) but are not in
408 	 * ms_allocatable yet, because they are in ms_freed, ms_freeing,
409 	 * or ms_defer[].
410 	 *
411 	 * When the metaslab is not loaded, its ms_weight needs to
412 	 * reflect what is allocatable (i.e. what will be part of
413 	 * ms_allocatable if it is loaded).  The weight is computed from
414 	 * the spacemap histogram, but that includes ranges that are
415 	 * not yet allocatable (because they are in ms_freed,
416 	 * ms_freeing, or ms_defer[]).  Therefore, when calculating the
417 	 * weight, we need to remove those ranges.
418 	 *
419 	 * The ranges in the ms_freed and ms_defer[] range trees are all
420 	 * present in the spacemap.  However, the spacemap may have
421 	 * multiple entries to represent a contiguous range, because it
422 	 * is written across multiple sync passes, but the changes of
423 	 * all sync passes are consolidated into the range trees.
424 	 * Adjacent ranges that are freed in different sync passes of
425 	 * one txg will be represented separately (as 2 or more entries)
426 	 * in the space map (and its histogram), but these adjacent
427 	 * ranges will be consolidated (represented as one entry) in the
428 	 * ms_freed/ms_defer[] range trees (and their histograms).
429 	 *
430 	 * When calculating the weight, we can not simply subtract the
431 	 * range trees' histograms from the spacemap's histogram,
432 	 * because the range trees' histograms may have entries in
433 	 * higher buckets than the spacemap, due to consolidation.
434 	 * Instead we must subtract the exact entries that were added to
435 	 * the spacemap's histogram.  ms_synchist and ms_deferhist[]
436 	 * represent these exact entries, so we can subtract them from
437 	 * the spacemap's histogram when calculating ms_weight.
438 	 *
439 	 * ms_synchist represents the same ranges as ms_freeing +
440 	 * ms_freed, but without consolidation across sync passes.
441 	 *
442 	 * ms_deferhist[i] represents the same ranges as ms_defer[i],
443 	 * but without consolidation across sync passes.
444 	 */
445 	uint64_t	ms_synchist[SPACE_MAP_HISTOGRAM_SIZE];
446 	uint64_t	ms_deferhist[TXG_DEFER_SIZE][SPACE_MAP_HISTOGRAM_SIZE];
447 
448 	/*
449 	 * Tracks the exact amount of allocated space of this metaslab
450 	 * (and specifically the metaslab's space map) up to the most
451 	 * recently completed sync pass [see usage in metaslab_sync()].
452 	 */
453 	uint64_t	ms_allocated_space;
454 	int64_t		ms_deferspace;	/* sum of ms_defermap[] space	*/
455 	uint64_t	ms_weight;	/* weight vs. others in group	*/
456 	uint64_t	ms_activation_weight;	/* activation weight	*/
457 
458 	/*
459 	 * Track of whenever a metaslab is selected for loading or allocation.
460 	 * We use this value to determine how long the metaslab should
461 	 * stay cached.
462 	 */
463 	uint64_t	ms_selected_txg;
464 
465 	uint64_t	ms_alloc_txg;	/* last successful alloc (debug only) */
466 	uint64_t	ms_max_size;	/* maximum allocatable size	*/
467 
468 	/*
469 	 * -1 if it's not active in an allocator, otherwise set to the allocator
470 	 * this metaslab is active for.
471 	 */
472 	int		ms_allocator;
473 	boolean_t	ms_primary; /* Only valid if ms_allocator is not -1 */
474 
475 	/*
476 	 * The metaslab block allocators can optionally use a size-ordered
477 	 * range tree and/or an array of LBAs. Not all allocators use
478 	 * this functionality. The ms_allocatable_by_size should always
479 	 * contain the same number of segments as the ms_allocatable. The
480 	 * only difference is that the ms_allocatable_by_size is ordered by
481 	 * segment sizes.
482 	 */
483 	avl_tree_t	ms_allocatable_by_size;
484 	uint64_t	ms_lbas[MAX_LBAS];
485 
486 	metaslab_group_t *ms_group;	/* metaslab group		*/
487 	avl_node_t	ms_group_node;	/* node in metaslab group tree	*/
488 	txg_node_t	ms_txg_node;	/* per-txg dirty metaslab links	*/
489 
490 	/* updated every time we are done syncing the metaslab's space map */
491 	uint64_t	ms_synced_length;
492 
493 	boolean_t	ms_new;
494 };
495 
496 #ifdef	__cplusplus
497 }
498 #endif
499 
500 #endif	/* _SYS_METASLAB_IMPL_H */
501