xref: /illumos-gate/usr/src/uts/common/fs/zfs/sys/metaslab_impl.h (revision e2f4f3dab373b605b62ac175115f264a5c417eb6)
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, 2015 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 } metaslab_alloc_trace_t;
56 
57 /*
58  * Used by the metaslab allocation tracing facility to indicate
59  * error conditions. These errors are stored to the offset member
60  * of the metaslab_alloc_trace_t record and displayed by mdb.
61  */
62 typedef enum trace_alloc_type {
63 	TRACE_ALLOC_FAILURE	= -1ULL,
64 	TRACE_TOO_SMALL		= -2ULL,
65 	TRACE_FORCE_GANG	= -3ULL,
66 	TRACE_NOT_ALLOCATABLE	= -4ULL,
67 	TRACE_GROUP_FAILURE	= -5ULL,
68 	TRACE_ENOSPC		= -6ULL,
69 	TRACE_CONDENSING	= -7ULL,
70 	TRACE_VDEV_ERROR	= -8ULL
71 } trace_alloc_type_t;
72 
73 #define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
74 #define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
75 #define	METASLAB_WEIGHT_TYPE		(1ULL << 61)
76 #define	METASLAB_ACTIVE_MASK		\
77 	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
78 
79 /*
80  * The metaslab weight is used to encode the amount of free space in a
81  * metaslab, such that the "best" metaslab appears first when sorting the
82  * metaslabs by weight. The weight (and therefore the "best" metaslab) can
83  * be determined in two different ways: by computing a weighted sum of all
84  * the free space in the metaslab (a space based weight) or by counting only
85  * the free segments of the largest size (a segment based weight). We prefer
86  * the segment based weight because it reflects how the free space is
87  * comprised, but we cannot always use it -- legacy pools do not have the
88  * space map histogram information necessary to determine the largest
89  * contiguous regions. Pools that have the space map histogram determine
90  * the segment weight by looking at each bucket in the histogram and
91  * determining the free space whose size in bytes is in the range:
92  *	[2^i, 2^(i+1))
93  * We then encode the largest index, i, that contains regions into the
94  * segment-weighted value.
95  *
96  * Space-based weight:
97  *
98  *      64      56      48      40      32      24      16      8       0
99  *      +-------+-------+-------+-------+-------+-------+-------+-------+
100  *      |PS1|                   weighted-free space                     |
101  *      +-------+-------+-------+-------+-------+-------+-------+-------+
102  *
103  *	PS - indicates primary and secondary activation
104  *	space - the fragmentation-weighted space
105  *
106  * Segment-based weight:
107  *
108  *      64      56      48      40      32      24      16      8       0
109  *      +-------+-------+-------+-------+-------+-------+-------+-------+
110  *      |PS0| idx|             count of segments in region              |
111  *      +-------+-------+-------+-------+-------+-------+-------+-------+
112  *
113  *	PS - indicates primary and secondary activation
114  *	idx - index for the highest bucket in the histogram
115  *	count - number of segments in the specified bucket
116  */
117 #define	WEIGHT_GET_ACTIVE(weight)		BF64_GET((weight), 62, 2)
118 #define	WEIGHT_SET_ACTIVE(weight, x)		BF64_SET((weight), 62, 2, x)
119 
120 #define	WEIGHT_IS_SPACEBASED(weight)		\
121 	((weight) == 0 || BF64_GET((weight), 61, 1))
122 #define	WEIGHT_SET_SPACEBASED(weight)		BF64_SET((weight), 61, 1, 1)
123 
124 /*
125  * These macros are only applicable to segment-based weighting.
126  */
127 #define	WEIGHT_GET_INDEX(weight)		BF64_GET((weight), 55, 6)
128 #define	WEIGHT_SET_INDEX(weight, x)		BF64_SET((weight), 55, 6, x)
129 #define	WEIGHT_GET_COUNT(weight)		BF64_GET((weight), 0, 55)
130 #define	WEIGHT_SET_COUNT(weight, x)		BF64_SET((weight), 0, 55, x)
131 
132 /*
133  * A metaslab class encompasses a category of allocatable top-level vdevs.
134  * Each top-level vdev is associated with a metaslab group which defines
135  * the allocatable region for that vdev. Examples of these categories include
136  * "normal" for data block allocations (i.e. main pool allocations) or "log"
137  * for allocations designated for intent log devices (i.e. slog devices).
138  * When a block allocation is requested from the SPA it is associated with a
139  * metaslab_class_t, and only top-level vdevs (i.e. metaslab groups) belonging
140  * to the class can be used to satisfy that request. Allocations are done
141  * by traversing the metaslab groups that are linked off of the mc_rotor field.
142  * This rotor points to the next metaslab group where allocations will be
143  * attempted. Allocating a block is a 3 step process -- select the metaslab
144  * group, select the metaslab, and then allocate the block. The metaslab
145  * class defines the low-level block allocator that will be used as the
146  * final step in allocation. These allocators are pluggable allowing each class
147  * to use a block allocator that best suits that class.
148  */
149 struct metaslab_class {
150 	kmutex_t		mc_lock;
151 	spa_t			*mc_spa;
152 	metaslab_group_t	*mc_rotor;
153 	metaslab_ops_t		*mc_ops;
154 	uint64_t		mc_aliquot;
155 
156 	/*
157 	 * Track the number of metaslab groups that have been initialized
158 	 * and can accept allocations. An initialized metaslab group is
159 	 * one has been completely added to the config (i.e. we have
160 	 * updated the MOS config and the space has been added to the pool).
161 	 */
162 	uint64_t		mc_groups;
163 
164 	/*
165 	 * Toggle to enable/disable the allocation throttle.
166 	 */
167 	boolean_t		mc_alloc_throttle_enabled;
168 
169 	/*
170 	 * The allocation throttle works on a reservation system. Whenever
171 	 * an asynchronous zio wants to perform an allocation it must
172 	 * first reserve the number of blocks that it wants to allocate.
173 	 * If there aren't sufficient slots available for the pending zio
174 	 * then that I/O is throttled until more slots free up. The current
175 	 * number of reserved allocations is maintained by the mc_alloc_slots
176 	 * refcount. The mc_alloc_max_slots value determines the maximum
177 	 * number of allocations that the system allows. Gang blocks are
178 	 * allowed to reserve slots even if we've reached the maximum
179 	 * number of allocations allowed.
180 	 */
181 	uint64_t		mc_alloc_max_slots;
182 	refcount_t		mc_alloc_slots;
183 
184 	uint64_t		mc_alloc_groups; /* # of allocatable groups */
185 
186 	uint64_t		mc_alloc;	/* total allocated space */
187 	uint64_t		mc_deferred;	/* total deferred frees */
188 	uint64_t		mc_space;	/* total space (alloc + free) */
189 	uint64_t		mc_dspace;	/* total deflated space */
190 	uint64_t		mc_histogram[RANGE_TREE_HISTOGRAM_SIZE];
191 };
192 
193 /*
194  * Metaslab groups encapsulate all the allocatable regions (i.e. metaslabs)
195  * of a top-level vdev. They are linked togther to form a circular linked
196  * list and can belong to only one metaslab class. Metaslab groups may become
197  * ineligible for allocations for a number of reasons such as limited free
198  * space, fragmentation, or going offline. When this happens the allocator will
199  * simply find the next metaslab group in the linked list and attempt
200  * to allocate from that group instead.
201  */
202 struct metaslab_group {
203 	kmutex_t		mg_lock;
204 	avl_tree_t		mg_metaslab_tree;
205 	uint64_t		mg_aliquot;
206 	boolean_t		mg_allocatable;		/* can we allocate? */
207 
208 	/*
209 	 * A metaslab group is considered to be initialized only after
210 	 * we have updated the MOS config and added the space to the pool.
211 	 * We only allow allocation attempts to a metaslab group if it
212 	 * has been initialized.
213 	 */
214 	boolean_t		mg_initialized;
215 
216 	uint64_t		mg_free_capacity;	/* percentage free */
217 	int64_t			mg_bias;
218 	int64_t			mg_activation_count;
219 	metaslab_class_t	*mg_class;
220 	vdev_t			*mg_vd;
221 	taskq_t			*mg_taskq;
222 	metaslab_group_t	*mg_prev;
223 	metaslab_group_t	*mg_next;
224 
225 	/*
226 	 * Each metaslab group can handle mg_max_alloc_queue_depth allocations
227 	 * which are tracked by mg_alloc_queue_depth. It's possible for a
228 	 * metaslab group to handle more allocations than its max. This
229 	 * can occur when gang blocks are required or when other groups
230 	 * are unable to handle their share of allocations.
231 	 */
232 	uint64_t		mg_max_alloc_queue_depth;
233 	refcount_t		mg_alloc_queue_depth;
234 
235 	/*
236 	 * A metalab group that can no longer allocate the minimum block
237 	 * size will set mg_no_free_space. Once a metaslab group is out
238 	 * of space then its share of work must be distributed to other
239 	 * groups.
240 	 */
241 	boolean_t		mg_no_free_space;
242 
243 	uint64_t		mg_allocations;
244 	uint64_t		mg_failed_allocations;
245 	uint64_t		mg_fragmentation;
246 	uint64_t		mg_histogram[RANGE_TREE_HISTOGRAM_SIZE];
247 };
248 
249 /*
250  * This value defines the number of elements in the ms_lbas array. The value
251  * of 64 was chosen as it covers all power of 2 buckets up to UINT64_MAX.
252  * This is the equivalent of highbit(UINT64_MAX).
253  */
254 #define	MAX_LBAS	64
255 
256 /*
257  * Each metaslab maintains a set of in-core trees to track metaslab
258  * operations.  The in-core free tree (ms_tree) contains the list of
259  * free segments which are eligible for allocation.  As blocks are
260  * allocated, the allocated segments are removed from the ms_tree and
261  * added to a per txg allocation tree (ms_alloctree).  This allows us to
262  * process all allocations in syncing context where it is safe to update
263  * the on-disk space maps.  Frees are also processed in syncing context.
264  * Most frees are generated from syncing context, and those that are not
265  * are held in the spa_free_bplist for processing in syncing context.
266  * An additional set of in-core trees is maintained to track deferred
267  * frees (ms_defertree).  Once a block is freed it will move from the
268  * ms_freedtree to the ms_defertree.  A deferred free means that a block
269  * has been freed but cannot be used by the pool until TXG_DEFER_SIZE
270  * transactions groups later.  For example, a block that is freed in txg
271  * 50 will not be available for reallocation until txg 52 (50 +
272  * TXG_DEFER_SIZE).  This provides a safety net for uberblock rollback.
273  * A pool could be safely rolled back TXG_DEFERS_SIZE transactions
274  * groups and ensure that no block has been reallocated.
275  *
276  * The simplified transition diagram looks like this:
277  *
278  *
279  *      ALLOCATE
280  *         |
281  *         V
282  *    free segment (ms_tree) -----> ms_alloctree[4] ----> (write to space map)
283  *         ^
284  *         |                           ms_freeingtree <--- FREE
285  *         |                                 |
286  *         |                                 v
287  *         |                           ms_freedtree
288  *         |                                 |
289  *         +-------- ms_defertree[2] <-------+---------> (write to space map)
290  *
291  *
292  * Each metaslab's space is tracked in a single space map in the MOS,
293  * which is only updated in syncing context.  Each time we sync a txg,
294  * we append the allocs and frees from that txg to the space map.  The
295  * pool space is only updated once all metaslabs have finished syncing.
296  *
297  * To load the in-core free tree we read the space map from disk.  This
298  * object contains a series of alloc and free records that are combined
299  * to make up the list of all free segments in this metaslab.  These
300  * segments are represented in-core by the ms_tree and are stored in an
301  * AVL tree.
302  *
303  * As the space map grows (as a result of the appends) it will
304  * eventually become space-inefficient.  When the metaslab's in-core
305  * free tree is zfs_condense_pct/100 times the size of the minimal
306  * on-disk representation, we rewrite it in its minimized form.  If a
307  * metaslab needs to condense then we must set the ms_condensing flag to
308  * ensure that allocations are not performed on the metaslab that is
309  * being written.
310  */
311 struct metaslab {
312 	kmutex_t	ms_lock;
313 	kcondvar_t	ms_load_cv;
314 	space_map_t	*ms_sm;
315 	uint64_t	ms_id;
316 	uint64_t	ms_start;
317 	uint64_t	ms_size;
318 	uint64_t	ms_fragmentation;
319 
320 	range_tree_t	*ms_alloctree[TXG_SIZE];
321 	range_tree_t	*ms_tree;
322 
323 	/*
324 	 * The following range trees are accessed only from syncing context.
325 	 * ms_free*tree only have entries while syncing, and are empty
326 	 * between syncs.
327 	 */
328 	range_tree_t	*ms_freeingtree; /* to free this syncing txg */
329 	range_tree_t	*ms_freedtree; /* already freed this syncing txg */
330 	range_tree_t	*ms_defertree[TXG_DEFER_SIZE];
331 
332 	boolean_t	ms_condensing;	/* condensing? */
333 	boolean_t	ms_condense_wanted;
334 
335 	/*
336 	 * We must hold both ms_lock and ms_group->mg_lock in order to
337 	 * modify ms_loaded.
338 	 */
339 	boolean_t	ms_loaded;
340 	boolean_t	ms_loading;
341 
342 	int64_t		ms_deferspace;	/* sum of ms_defermap[] space	*/
343 	uint64_t	ms_weight;	/* weight vs. others in group	*/
344 	uint64_t	ms_activation_weight;	/* activation weight	*/
345 
346 	/*
347 	 * Track of whenever a metaslab is selected for loading or allocation.
348 	 * We use this value to determine how long the metaslab should
349 	 * stay cached.
350 	 */
351 	uint64_t	ms_selected_txg;
352 
353 	uint64_t	ms_alloc_txg;	/* last successful alloc (debug only) */
354 	uint64_t	ms_max_size;	/* maximum allocatable size	*/
355 
356 	/*
357 	 * The metaslab block allocators can optionally use a size-ordered
358 	 * range tree and/or an array of LBAs. Not all allocators use
359 	 * this functionality. The ms_size_tree should always contain the
360 	 * same number of segments as the ms_tree. The only difference
361 	 * is that the ms_size_tree is ordered by segment sizes.
362 	 */
363 	avl_tree_t	ms_size_tree;
364 	uint64_t	ms_lbas[MAX_LBAS];
365 
366 	metaslab_group_t *ms_group;	/* metaslab group		*/
367 	avl_node_t	ms_group_node;	/* node in metaslab group tree	*/
368 	txg_node_t	ms_txg_node;	/* per-txg dirty metaslab links	*/
369 };
370 
371 #ifdef	__cplusplus
372 }
373 #endif
374 
375 #endif	/* _SYS_METASLAB_IMPL_H */
376