1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
27 */
28
29 /*
30 * DVA-based Adjustable Replacement Cache
31 *
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
36 *
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
47 *
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
53 *
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
59 * tight.
60 *
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
69 *
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
72 */
73
74 /*
75 * The locking model:
76 *
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal arc algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
83 * arc list locks.
84 *
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
88 *
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
92 *
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
95 *
96 * Each arc state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an arc list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
101 *
102 * Arc buffers may have an associated eviction callback function.
103 * This function will be invoked prior to removing the buffer (e.g.
104 * in arc_do_user_evicts()). Note however that the data associated
105 * with the buffer may be evicted prior to the callback. The callback
106 * must be made with *no locks held* (to prevent deadlock). Additionally,
107 * the users of callbacks must ensure that their private data is
108 * protected from simultaneous callbacks from arc_clear_callback()
109 * and arc_do_user_evicts().
110 *
111 * Note that the majority of the performance stats are manipulated
112 * with atomic operations.
113 *
114 * The L2ARC uses the l2ad_mtx on each vdev for the following:
115 *
116 * - L2ARC buflist creation
117 * - L2ARC buflist eviction
118 * - L2ARC write completion, which walks L2ARC buflists
119 * - ARC header destruction, as it removes from L2ARC buflists
120 * - ARC header release, as it removes from L2ARC buflists
121 */
122
123 /*
124 * ARC operation:
125 *
126 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
127 * This structure can point either to a block that is still in the cache or to
128 * one that is only accessible in an L2 ARC device, or it can provide
129 * information about a block that was recently evicted. If a block is
130 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
131 * information to retrieve it from the L2ARC device. This information is
132 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
133 * that is in this state cannot access the data directly.
134 *
135 * Blocks that are actively being referenced or have not been evicted
136 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
137 * the arc_buf_hdr_t that will point to the data block in memory. A block can
138 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
139 * caches data in two ways -- in a list of arc buffers (arc_buf_t) and
140 * also in the arc_buf_hdr_t's private physical data block pointer (b_pdata).
141 * Each arc buffer (arc_buf_t) is being actively accessed by a specific ARC
142 * consumer, and always contains uncompressed data. The ARC will provide
143 * references to this data and will keep it cached until it is no longer in
144 * use. Typically, the arc will try to cache only the L1ARC's physical data
145 * block and will aggressively evict any arc_buf_t that is no longer referenced.
146 * The amount of memory consumed by the arc_buf_t's can be seen via the
147 * "overhead_size" kstat.
148 *
149 *
150 * arc_buf_hdr_t
151 * +-----------+
152 * | |
153 * | |
154 * | |
155 * +-----------+
156 * l2arc_buf_hdr_t| |
157 * | |
158 * +-----------+
159 * l1arc_buf_hdr_t| |
160 * | | arc_buf_t
161 * | b_buf +------------>+---------+ arc_buf_t
162 * | | |b_next +---->+---------+
163 * | b_pdata +-+ |---------| |b_next +-->NULL
164 * +-----------+ | | | +---------+
165 * | |b_data +-+ | |
166 * | +---------+ | |b_data +-+
167 * +->+------+ | +---------+ |
168 * (potentially) | | | |
169 * compressed | | | |
170 * data +------+ | v
171 * +->+------+ +------+
172 * uncompressed | | | |
173 * data | | | |
174 * +------+ +------+
175 *
176 * The L1ARC's data pointer, however, may or may not be uncompressed. The
177 * ARC has the ability to store the physical data (b_pdata) associated with
178 * the DVA of the arc_buf_hdr_t. Since the b_pdata is a copy of the on-disk
179 * physical block, it will match its on-disk compression characteristics.
180 * If the block on-disk is compressed, then the physical data block
181 * in the cache will also be compressed and vice-versa. This behavior
182 * can be disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
183 * compressed ARC functionality is disabled, the b_pdata will point to an
184 * uncompressed version of the on-disk data.
185 *
186 * When a consumer reads a block, the ARC must first look to see if the
187 * arc_buf_hdr_t is cached. If the hdr is cached and already has an arc_buf_t,
188 * then an additional arc_buf_t is allocated and the uncompressed data is
189 * bcopied from the existing arc_buf_t. If the hdr is cached but does not
190 * have an arc_buf_t, then the ARC allocates a new arc_buf_t and decompresses
191 * the b_pdata contents into the arc_buf_t's b_data. If the arc_buf_hdr_t's
192 * b_pdata is not compressed, then the block is shared with the newly
193 * allocated arc_buf_t. This block sharing only occurs with one arc_buf_t
194 * in the arc buffer chain. Sharing the block reduces the memory overhead
195 * required when the hdr is caching uncompressed blocks or the compressed
196 * arc functionality has been disabled via 'zfs_compressed_arc_enabled'.
197 *
198 * The diagram below shows an example of an uncompressed ARC hdr that is
199 * sharing its data with an arc_buf_t:
200 *
201 * arc_buf_hdr_t
202 * +-----------+
203 * | |
204 * | |
205 * | |
206 * +-----------+
207 * l2arc_buf_hdr_t| |
208 * | |
209 * +-----------+
210 * l1arc_buf_hdr_t| |
211 * | | arc_buf_t (shared)
212 * | b_buf +------------>+---------+ arc_buf_t
213 * | | |b_next +---->+---------+
214 * | b_pdata +-+ |---------| |b_next +-->NULL
215 * +-----------+ | | | +---------+
216 * | |b_data +-+ | |
217 * | +---------+ | |b_data +-+
218 * +->+------+ | +---------+ |
219 * | | | |
220 * uncompressed | | | |
221 * data +------+ | |
222 * ^ +->+------+ |
223 * | uncompressed | | |
224 * | data | | |
225 * | +------+ |
226 * +---------------------------------+
227 *
228 * Writing to the arc requires that the ARC first discard the b_pdata
229 * since the physical block is about to be rewritten. The new data contents
230 * will be contained in the arc_buf_t (uncompressed). As the I/O pipeline
231 * performs the write, it may compress the data before writing it to disk.
232 * The ARC will be called with the transformed data and will bcopy the
233 * transformed on-disk block into a newly allocated b_pdata.
234 *
235 * When the L2ARC is in use, it will also take advantage of the b_pdata. The
236 * L2ARC will always write the contents of b_pdata to the L2ARC. This means
237 * that when compressed arc is enabled that the L2ARC blocks are identical
238 * to the on-disk block in the main data pool. This provides a significant
239 * advantage since the ARC can leverage the bp's checksum when reading from the
240 * L2ARC to determine if the contents are valid. However, if the compressed
241 * arc is disabled, then the L2ARC's block must be transformed to look
242 * like the physical block in the main data pool before comparing the
243 * checksum and determining its validity.
244 */
245
246 #include <sys/spa.h>
247 #include <sys/zio.h>
248 #include <sys/spa_impl.h>
249 #include <sys/zio_compress.h>
250 #include <sys/zio_checksum.h>
251 #include <sys/zfs_context.h>
252 #include <sys/arc.h>
253 #include <sys/refcount.h>
254 #include <sys/vdev.h>
255 #include <sys/vdev_impl.h>
256 #include <sys/dsl_pool.h>
257 #include <sys/multilist.h>
258 #ifdef _KERNEL
259 #include <sys/dnlc.h>
260 #include <sys/racct.h>
261 #endif
262 #include <sys/callb.h>
263 #include <sys/kstat.h>
264 #include <sys/trim_map.h>
265 #include <zfs_fletcher.h>
266 #include <sys/sdt.h>
267
268 #include <machine/vmparam.h>
269
270 #ifdef illumos
271 #ifndef _KERNEL
272 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
273 boolean_t arc_watch = B_FALSE;
274 int arc_procfd;
275 #endif
276 #endif /* illumos */
277
278 #ifdef __NetBSD__
279 #include <uvm/uvm.h>
280 #ifndef btop
281 #define btop(x) ((x) / PAGE_SIZE)
282 #endif
283 #ifndef ptob
284 #define ptob(x) ((x) * PAGE_SIZE)
285 #endif
286 //#define needfree (uvm_availmem() < uvmexp.freetarg ? uvmexp.freetarg : 0)
287 #define buf_init arc_buf_init
288 #define freemem uvm_availmem(false)
289 #define minfree uvmexp.freemin
290 #define desfree uvmexp.freetarg
291 #define zfs_arc_free_target desfree
292 #define lotsfree (desfree * 2)
293 #define availrmem desfree
294 #define swapfs_minfree 0
295 #define swapfs_reserve 0
296 #undef curproc
297 #define curproc curlwp
298 #define proc_pageout uvm.pagedaemon_lwp
299
300 static void *zio_arena;
301
302 #include <sys/callback.h>
303 /* Structures used for memory and kva space reclaim. */
304 static struct callback_entry arc_kva_reclaim_entry;
305
306 #endif /* __NetBSD__ */
307
308 static kmutex_t arc_reclaim_lock;
309 static kcondvar_t arc_reclaim_thread_cv;
310 static boolean_t arc_reclaim_thread_exit;
311 static kcondvar_t arc_reclaim_waiters_cv;
312
313 #ifdef __FreeBSD__
314 static kmutex_t arc_dnlc_evicts_lock;
315 static kcondvar_t arc_dnlc_evicts_cv;
316 static boolean_t arc_dnlc_evicts_thread_exit;
317
318 uint_t arc_reduce_dnlc_percent = 3;
319 #endif
320
321 /*
322 * The number of headers to evict in arc_evict_state_impl() before
323 * dropping the sublist lock and evicting from another sublist. A lower
324 * value means we're more likely to evict the "correct" header (i.e. the
325 * oldest header in the arc state), but comes with higher overhead
326 * (i.e. more invocations of arc_evict_state_impl()).
327 */
328 int zfs_arc_evict_batch_limit = 10;
329
330 /*
331 * The number of sublists used for each of the arc state lists. If this
332 * is not set to a suitable value by the user, it will be configured to
333 * the number of CPUs on the system in arc_init().
334 */
335 int zfs_arc_num_sublists_per_state = 0;
336
337 /* number of seconds before growing cache again */
338 static int arc_grow_retry = 60;
339
340 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
341 int zfs_arc_overflow_shift = 8;
342
343 /* shift of arc_c for calculating both min and max arc_p */
344 static int arc_p_min_shift = 4;
345
346 /* log2(fraction of arc to reclaim) */
347 static int arc_shrink_shift = 7;
348
349 /*
350 * log2(fraction of ARC which must be free to allow growing).
351 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
352 * when reading a new block into the ARC, we will evict an equal-sized block
353 * from the ARC.
354 *
355 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
356 * we will still not allow it to grow.
357 */
358 int arc_no_grow_shift = 5;
359
360
361 /*
362 * minimum lifespan of a prefetch block in clock ticks
363 * (initialized in arc_init())
364 */
365 static int arc_min_prefetch_lifespan;
366
367 /*
368 * If this percent of memory is free, don't throttle.
369 */
370 int arc_lotsfree_percent = 10;
371
372 static int arc_dead;
373 extern boolean_t zfs_prefetch_disable;
374
375 /*
376 * The arc has filled available memory and has now warmed up.
377 */
378 static boolean_t arc_warm;
379
380 /*
381 * These tunables are for performance analysis.
382 */
383 uint64_t zfs_arc_max;
384 uint64_t zfs_arc_min;
385 uint64_t zfs_arc_meta_limit = 0;
386 uint64_t zfs_arc_meta_min = 0;
387 int zfs_arc_grow_retry = 0;
388 int zfs_arc_shrink_shift = 0;
389 int zfs_arc_p_min_shift = 0;
390 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
391
392 /* Absolute min for arc min / max is 16MB. */
393 static uint64_t arc_abs_min = 16 << 20;
394
395 boolean_t zfs_compressed_arc_enabled = B_TRUE;
396
397 #if defined(__FreeBSD__) && defined(_KERNEL)
398 u_int zfs_arc_free_target = 0;
399
400 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
401 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
402 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
403 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
404
405 static void
arc_free_target_init(void * unused __unused)406 arc_free_target_init(void *unused __unused)
407 {
408
409 zfs_arc_free_target = vm_pageout_wakeup_thresh;
410 }
411 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
412 arc_free_target_init, NULL);
413
414 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
415 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
416 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
417 SYSCTL_DECL(_vfs_zfs);
418 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
419 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
420 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
421 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
422 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
423 &zfs_arc_average_blocksize, 0,
424 "ARC average blocksize");
425 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
426 &arc_shrink_shift, 0,
427 "log2(fraction of arc to reclaim)");
428 SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
429 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC");
430
431 /*
432 * We don't have a tunable for arc_free_target due to the dependency on
433 * pagedaemon initialisation.
434 */
435 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
436 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
437 sysctl_vfs_zfs_arc_free_target, "IU",
438 "Desired number of free pages below which ARC triggers reclaim");
439
440 static int
sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)441 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
442 {
443 u_int val;
444 int err;
445
446 val = zfs_arc_free_target;
447 err = sysctl_handle_int(oidp, &val, 0, req);
448 if (err != 0 || req->newptr == NULL)
449 return (err);
450
451 if (val < minfree)
452 return (EINVAL);
453 if (val > vm_cnt.v_page_count)
454 return (EINVAL);
455
456 zfs_arc_free_target = val;
457
458 return (0);
459 }
460
461 /*
462 * Must be declared here, before the definition of corresponding kstat
463 * macro which uses the same names will confuse the compiler.
464 */
465 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
466 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
467 sysctl_vfs_zfs_arc_meta_limit, "QU",
468 "ARC metadata limit");
469 #endif
470
471 /*
472 * Note that buffers can be in one of 6 states:
473 * ARC_anon - anonymous (discussed below)
474 * ARC_mru - recently used, currently cached
475 * ARC_mru_ghost - recentely used, no longer in cache
476 * ARC_mfu - frequently used, currently cached
477 * ARC_mfu_ghost - frequently used, no longer in cache
478 * ARC_l2c_only - exists in L2ARC but not other states
479 * When there are no active references to the buffer, they are
480 * are linked onto a list in one of these arc states. These are
481 * the only buffers that can be evicted or deleted. Within each
482 * state there are multiple lists, one for meta-data and one for
483 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
484 * etc.) is tracked separately so that it can be managed more
485 * explicitly: favored over data, limited explicitly.
486 *
487 * Anonymous buffers are buffers that are not associated with
488 * a DVA. These are buffers that hold dirty block copies
489 * before they are written to stable storage. By definition,
490 * they are "ref'd" and are considered part of arc_mru
491 * that cannot be freed. Generally, they will aquire a DVA
492 * as they are written and migrate onto the arc_mru list.
493 *
494 * The ARC_l2c_only state is for buffers that are in the second
495 * level ARC but no longer in any of the ARC_m* lists. The second
496 * level ARC itself may also contain buffers that are in any of
497 * the ARC_m* states - meaning that a buffer can exist in two
498 * places. The reason for the ARC_l2c_only state is to keep the
499 * buffer header in the hash table, so that reads that hit the
500 * second level ARC benefit from these fast lookups.
501 */
502
503 typedef struct arc_state {
504 /*
505 * list of evictable buffers
506 */
507 multilist_t arcs_list[ARC_BUFC_NUMTYPES];
508 /*
509 * total amount of evictable data in this state
510 */
511 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
512 /*
513 * total amount of data in this state; this includes: evictable,
514 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
515 */
516 refcount_t arcs_size;
517 } arc_state_t;
518
519 /* The 6 states: */
520 static arc_state_t ARC_anon;
521 static arc_state_t ARC_mru;
522 static arc_state_t ARC_mru_ghost;
523 static arc_state_t ARC_mfu;
524 static arc_state_t ARC_mfu_ghost;
525 static arc_state_t ARC_l2c_only;
526
527 typedef struct arc_stats {
528 kstat_named_t arcstat_hits;
529 kstat_named_t arcstat_misses;
530 kstat_named_t arcstat_demand_data_hits;
531 kstat_named_t arcstat_demand_data_misses;
532 kstat_named_t arcstat_demand_metadata_hits;
533 kstat_named_t arcstat_demand_metadata_misses;
534 kstat_named_t arcstat_prefetch_data_hits;
535 kstat_named_t arcstat_prefetch_data_misses;
536 kstat_named_t arcstat_prefetch_metadata_hits;
537 kstat_named_t arcstat_prefetch_metadata_misses;
538 kstat_named_t arcstat_mru_hits;
539 kstat_named_t arcstat_mru_ghost_hits;
540 kstat_named_t arcstat_mfu_hits;
541 kstat_named_t arcstat_mfu_ghost_hits;
542 kstat_named_t arcstat_allocated;
543 kstat_named_t arcstat_deleted;
544 /*
545 * Number of buffers that could not be evicted because the hash lock
546 * was held by another thread. The lock may not necessarily be held
547 * by something using the same buffer, since hash locks are shared
548 * by multiple buffers.
549 */
550 kstat_named_t arcstat_mutex_miss;
551 /*
552 * Number of buffers skipped because they have I/O in progress, are
553 * indrect prefetch buffers that have not lived long enough, or are
554 * not from the spa we're trying to evict from.
555 */
556 kstat_named_t arcstat_evict_skip;
557 /*
558 * Number of times arc_evict_state() was unable to evict enough
559 * buffers to reach it's target amount.
560 */
561 kstat_named_t arcstat_evict_not_enough;
562 kstat_named_t arcstat_evict_l2_cached;
563 kstat_named_t arcstat_evict_l2_eligible;
564 kstat_named_t arcstat_evict_l2_ineligible;
565 kstat_named_t arcstat_evict_l2_skip;
566 kstat_named_t arcstat_hash_elements;
567 kstat_named_t arcstat_hash_elements_max;
568 kstat_named_t arcstat_hash_collisions;
569 kstat_named_t arcstat_hash_chains;
570 kstat_named_t arcstat_hash_chain_max;
571 kstat_named_t arcstat_p;
572 kstat_named_t arcstat_c;
573 kstat_named_t arcstat_c_min;
574 kstat_named_t arcstat_c_max;
575 kstat_named_t arcstat_size;
576 /*
577 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pdata.
578 * Note that the compressed bytes may match the uncompressed bytes
579 * if the block is either not compressed or compressed arc is disabled.
580 */
581 kstat_named_t arcstat_compressed_size;
582 /*
583 * Uncompressed size of the data stored in b_pdata. If compressed
584 * arc is disabled then this value will be identical to the stat
585 * above.
586 */
587 kstat_named_t arcstat_uncompressed_size;
588 /*
589 * Number of bytes stored in all the arc_buf_t's. This is classified
590 * as "overhead" since this data is typically short-lived and will
591 * be evicted from the arc when it becomes unreferenced unless the
592 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
593 * values have been set (see comment in dbuf.c for more information).
594 */
595 kstat_named_t arcstat_overhead_size;
596 /*
597 * Number of bytes consumed by internal ARC structures necessary
598 * for tracking purposes; these structures are not actually
599 * backed by ARC buffers. This includes arc_buf_hdr_t structures
600 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
601 * caches), and arc_buf_t structures (allocated via arc_buf_t
602 * cache).
603 */
604 kstat_named_t arcstat_hdr_size;
605 /*
606 * Number of bytes consumed by ARC buffers of type equal to
607 * ARC_BUFC_DATA. This is generally consumed by buffers backing
608 * on disk user data (e.g. plain file contents).
609 */
610 kstat_named_t arcstat_data_size;
611 /*
612 * Number of bytes consumed by ARC buffers of type equal to
613 * ARC_BUFC_METADATA. This is generally consumed by buffers
614 * backing on disk data that is used for internal ZFS
615 * structures (e.g. ZAP, dnode, indirect blocks, etc).
616 */
617 kstat_named_t arcstat_metadata_size;
618 /*
619 * Number of bytes consumed by various buffers and structures
620 * not actually backed with ARC buffers. This includes bonus
621 * buffers (allocated directly via zio_buf_* functions),
622 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
623 * cache), and dnode_t structures (allocated via dnode_t cache).
624 */
625 kstat_named_t arcstat_other_size;
626 /*
627 * Total number of bytes consumed by ARC buffers residing in the
628 * arc_anon state. This includes *all* buffers in the arc_anon
629 * state; e.g. data, metadata, evictable, and unevictable buffers
630 * are all included in this value.
631 */
632 kstat_named_t arcstat_anon_size;
633 /*
634 * Number of bytes consumed by ARC buffers that meet the
635 * following criteria: backing buffers of type ARC_BUFC_DATA,
636 * residing in the arc_anon state, and are eligible for eviction
637 * (e.g. have no outstanding holds on the buffer).
638 */
639 kstat_named_t arcstat_anon_evictable_data;
640 /*
641 * Number of bytes consumed by ARC buffers that meet the
642 * following criteria: backing buffers of type ARC_BUFC_METADATA,
643 * residing in the arc_anon state, and are eligible for eviction
644 * (e.g. have no outstanding holds on the buffer).
645 */
646 kstat_named_t arcstat_anon_evictable_metadata;
647 /*
648 * Total number of bytes consumed by ARC buffers residing in the
649 * arc_mru state. This includes *all* buffers in the arc_mru
650 * state; e.g. data, metadata, evictable, and unevictable buffers
651 * are all included in this value.
652 */
653 kstat_named_t arcstat_mru_size;
654 /*
655 * Number of bytes consumed by ARC buffers that meet the
656 * following criteria: backing buffers of type ARC_BUFC_DATA,
657 * residing in the arc_mru state, and are eligible for eviction
658 * (e.g. have no outstanding holds on the buffer).
659 */
660 kstat_named_t arcstat_mru_evictable_data;
661 /*
662 * Number of bytes consumed by ARC buffers that meet the
663 * following criteria: backing buffers of type ARC_BUFC_METADATA,
664 * residing in the arc_mru state, and are eligible for eviction
665 * (e.g. have no outstanding holds on the buffer).
666 */
667 kstat_named_t arcstat_mru_evictable_metadata;
668 /*
669 * Total number of bytes that *would have been* consumed by ARC
670 * buffers in the arc_mru_ghost state. The key thing to note
671 * here, is the fact that this size doesn't actually indicate
672 * RAM consumption. The ghost lists only consist of headers and
673 * don't actually have ARC buffers linked off of these headers.
674 * Thus, *if* the headers had associated ARC buffers, these
675 * buffers *would have* consumed this number of bytes.
676 */
677 kstat_named_t arcstat_mru_ghost_size;
678 /*
679 * Number of bytes that *would have been* consumed by ARC
680 * buffers that are eligible for eviction, of type
681 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
682 */
683 kstat_named_t arcstat_mru_ghost_evictable_data;
684 /*
685 * Number of bytes that *would have been* consumed by ARC
686 * buffers that are eligible for eviction, of type
687 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
688 */
689 kstat_named_t arcstat_mru_ghost_evictable_metadata;
690 /*
691 * Total number of bytes consumed by ARC buffers residing in the
692 * arc_mfu state. This includes *all* buffers in the arc_mfu
693 * state; e.g. data, metadata, evictable, and unevictable buffers
694 * are all included in this value.
695 */
696 kstat_named_t arcstat_mfu_size;
697 /*
698 * Number of bytes consumed by ARC buffers that are eligible for
699 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
700 * state.
701 */
702 kstat_named_t arcstat_mfu_evictable_data;
703 /*
704 * Number of bytes consumed by ARC buffers that are eligible for
705 * eviction, of type ARC_BUFC_METADATA, and reside in the
706 * arc_mfu state.
707 */
708 kstat_named_t arcstat_mfu_evictable_metadata;
709 /*
710 * Total number of bytes that *would have been* consumed by ARC
711 * buffers in the arc_mfu_ghost state. See the comment above
712 * arcstat_mru_ghost_size for more details.
713 */
714 kstat_named_t arcstat_mfu_ghost_size;
715 /*
716 * Number of bytes that *would have been* consumed by ARC
717 * buffers that are eligible for eviction, of type
718 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
719 */
720 kstat_named_t arcstat_mfu_ghost_evictable_data;
721 /*
722 * Number of bytes that *would have been* consumed by ARC
723 * buffers that are eligible for eviction, of type
724 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
725 */
726 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
727 kstat_named_t arcstat_l2_hits;
728 kstat_named_t arcstat_l2_misses;
729 kstat_named_t arcstat_l2_feeds;
730 kstat_named_t arcstat_l2_rw_clash;
731 kstat_named_t arcstat_l2_read_bytes;
732 kstat_named_t arcstat_l2_write_bytes;
733 kstat_named_t arcstat_l2_writes_sent;
734 kstat_named_t arcstat_l2_writes_done;
735 kstat_named_t arcstat_l2_writes_error;
736 kstat_named_t arcstat_l2_writes_lock_retry;
737 kstat_named_t arcstat_l2_evict_lock_retry;
738 kstat_named_t arcstat_l2_evict_reading;
739 kstat_named_t arcstat_l2_evict_l1cached;
740 kstat_named_t arcstat_l2_free_on_write;
741 kstat_named_t arcstat_l2_abort_lowmem;
742 kstat_named_t arcstat_l2_cksum_bad;
743 kstat_named_t arcstat_l2_io_error;
744 kstat_named_t arcstat_l2_size;
745 kstat_named_t arcstat_l2_asize;
746 kstat_named_t arcstat_l2_hdr_size;
747 kstat_named_t arcstat_l2_write_trylock_fail;
748 kstat_named_t arcstat_l2_write_passed_headroom;
749 kstat_named_t arcstat_l2_write_spa_mismatch;
750 kstat_named_t arcstat_l2_write_in_l2;
751 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
752 kstat_named_t arcstat_l2_write_not_cacheable;
753 kstat_named_t arcstat_l2_write_full;
754 kstat_named_t arcstat_l2_write_buffer_iter;
755 kstat_named_t arcstat_l2_write_pios;
756 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
757 kstat_named_t arcstat_l2_write_buffer_list_iter;
758 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
759 kstat_named_t arcstat_memory_throttle_count;
760 kstat_named_t arcstat_meta_used;
761 kstat_named_t arcstat_meta_limit;
762 kstat_named_t arcstat_meta_max;
763 kstat_named_t arcstat_meta_min;
764 kstat_named_t arcstat_sync_wait_for_async;
765 kstat_named_t arcstat_demand_hit_predictive_prefetch;
766 } arc_stats_t;
767
768 static arc_stats_t arc_stats = {
769 { "hits", KSTAT_DATA_UINT64 },
770 { "misses", KSTAT_DATA_UINT64 },
771 { "demand_data_hits", KSTAT_DATA_UINT64 },
772 { "demand_data_misses", KSTAT_DATA_UINT64 },
773 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
774 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
775 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
776 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
777 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
778 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
779 { "mru_hits", KSTAT_DATA_UINT64 },
780 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
781 { "mfu_hits", KSTAT_DATA_UINT64 },
782 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
783 { "allocated", KSTAT_DATA_UINT64 },
784 { "deleted", KSTAT_DATA_UINT64 },
785 { "mutex_miss", KSTAT_DATA_UINT64 },
786 { "evict_skip", KSTAT_DATA_UINT64 },
787 { "evict_not_enough", KSTAT_DATA_UINT64 },
788 { "evict_l2_cached", KSTAT_DATA_UINT64 },
789 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
790 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
791 { "evict_l2_skip", KSTAT_DATA_UINT64 },
792 { "hash_elements", KSTAT_DATA_UINT64 },
793 { "hash_elements_max", KSTAT_DATA_UINT64 },
794 { "hash_collisions", KSTAT_DATA_UINT64 },
795 { "hash_chains", KSTAT_DATA_UINT64 },
796 { "hash_chain_max", KSTAT_DATA_UINT64 },
797 { "p", KSTAT_DATA_UINT64 },
798 { "c", KSTAT_DATA_UINT64 },
799 { "c_min", KSTAT_DATA_UINT64 },
800 { "c_max", KSTAT_DATA_UINT64 },
801 { "size", KSTAT_DATA_UINT64 },
802 { "compressed_size", KSTAT_DATA_UINT64 },
803 { "uncompressed_size", KSTAT_DATA_UINT64 },
804 { "overhead_size", KSTAT_DATA_UINT64 },
805 { "hdr_size", KSTAT_DATA_UINT64 },
806 { "data_size", KSTAT_DATA_UINT64 },
807 { "metadata_size", KSTAT_DATA_UINT64 },
808 { "other_size", KSTAT_DATA_UINT64 },
809 { "anon_size", KSTAT_DATA_UINT64 },
810 { "anon_evictable_data", KSTAT_DATA_UINT64 },
811 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
812 { "mru_size", KSTAT_DATA_UINT64 },
813 { "mru_evictable_data", KSTAT_DATA_UINT64 },
814 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
815 { "mru_ghost_size", KSTAT_DATA_UINT64 },
816 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
817 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
818 { "mfu_size", KSTAT_DATA_UINT64 },
819 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
820 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
821 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
822 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
823 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
824 { "l2_hits", KSTAT_DATA_UINT64 },
825 { "l2_misses", KSTAT_DATA_UINT64 },
826 { "l2_feeds", KSTAT_DATA_UINT64 },
827 { "l2_rw_clash", KSTAT_DATA_UINT64 },
828 { "l2_read_bytes", KSTAT_DATA_UINT64 },
829 { "l2_write_bytes", KSTAT_DATA_UINT64 },
830 { "l2_writes_sent", KSTAT_DATA_UINT64 },
831 { "l2_writes_done", KSTAT_DATA_UINT64 },
832 { "l2_writes_error", KSTAT_DATA_UINT64 },
833 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
834 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
835 { "l2_evict_reading", KSTAT_DATA_UINT64 },
836 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
837 { "l2_free_on_write", KSTAT_DATA_UINT64 },
838 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
839 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
840 { "l2_io_error", KSTAT_DATA_UINT64 },
841 { "l2_size", KSTAT_DATA_UINT64 },
842 { "l2_asize", KSTAT_DATA_UINT64 },
843 { "l2_hdr_size", KSTAT_DATA_UINT64 },
844 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
845 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
846 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
847 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
848 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
849 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
850 { "l2_write_full", KSTAT_DATA_UINT64 },
851 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
852 { "l2_write_pios", KSTAT_DATA_UINT64 },
853 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
854 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
855 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
856 { "memory_throttle_count", KSTAT_DATA_UINT64 },
857 { "arc_meta_used", KSTAT_DATA_UINT64 },
858 { "arc_meta_limit", KSTAT_DATA_UINT64 },
859 { "arc_meta_max", KSTAT_DATA_UINT64 },
860 { "arc_meta_min", KSTAT_DATA_UINT64 },
861 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
862 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
863 };
864
865 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
866
867 #define ARCSTAT_INCR(stat, val) \
868 atomic_add_64(&arc_stats.stat.value.ui64, (val))
869
870 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
871 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
872
873 #define ARCSTAT_MAX(stat, val) { \
874 uint64_t m; \
875 while ((val) > (m = arc_stats.stat.value.ui64) && \
876 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
877 continue; \
878 }
879
880 #define ARCSTAT_MAXSTAT(stat) \
881 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
882
883 /*
884 * We define a macro to allow ARC hits/misses to be easily broken down by
885 * two separate conditions, giving a total of four different subtypes for
886 * each of hits and misses (so eight statistics total).
887 */
888 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
889 if (cond1) { \
890 if (cond2) { \
891 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
892 } else { \
893 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
894 } \
895 } else { \
896 if (cond2) { \
897 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
898 } else { \
899 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
900 } \
901 }
902
903 kstat_t *arc_ksp;
904 static arc_state_t *arc_anon;
905 static arc_state_t *arc_mru;
906 static arc_state_t *arc_mru_ghost;
907 static arc_state_t *arc_mfu;
908 static arc_state_t *arc_mfu_ghost;
909 static arc_state_t *arc_l2c_only;
910
911 /*
912 * There are several ARC variables that are critical to export as kstats --
913 * but we don't want to have to grovel around in the kstat whenever we wish to
914 * manipulate them. For these variables, we therefore define them to be in
915 * terms of the statistic variable. This assures that we are not introducing
916 * the possibility of inconsistency by having shadow copies of the variables,
917 * while still allowing the code to be readable.
918 */
919 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
920 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
921 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
922 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
923 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
924 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
925 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
926 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
927 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
928
929 /* compressed size of entire arc */
930 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
931 /* uncompressed size of entire arc */
932 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
933 /* number of bytes in the arc from arc_buf_t's */
934 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
935
936 static int arc_no_grow; /* Don't try to grow cache size */
937 static uint64_t arc_tempreserve;
938 static uint64_t arc_loaned_bytes;
939
940 typedef struct arc_callback arc_callback_t;
941
942 struct arc_callback {
943 void *acb_private;
944 arc_done_func_t *acb_done;
945 arc_buf_t *acb_buf;
946 zio_t *acb_zio_dummy;
947 arc_callback_t *acb_next;
948 };
949
950 typedef struct arc_write_callback arc_write_callback_t;
951
952 struct arc_write_callback {
953 void *awcb_private;
954 arc_done_func_t *awcb_ready;
955 arc_done_func_t *awcb_children_ready;
956 arc_done_func_t *awcb_physdone;
957 arc_done_func_t *awcb_done;
958 arc_buf_t *awcb_buf;
959 };
960
961 /*
962 * ARC buffers are separated into multiple structs as a memory saving measure:
963 * - Common fields struct, always defined, and embedded within it:
964 * - L2-only fields, always allocated but undefined when not in L2ARC
965 * - L1-only fields, only allocated when in L1ARC
966 *
967 * Buffer in L1 Buffer only in L2
968 * +------------------------+ +------------------------+
969 * | arc_buf_hdr_t | | arc_buf_hdr_t |
970 * | | | |
971 * | | | |
972 * | | | |
973 * +------------------------+ +------------------------+
974 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
975 * | (undefined if L1-only) | | |
976 * +------------------------+ +------------------------+
977 * | l1arc_buf_hdr_t |
978 * | |
979 * | |
980 * | |
981 * | |
982 * +------------------------+
983 *
984 * Because it's possible for the L2ARC to become extremely large, we can wind
985 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
986 * is minimized by only allocating the fields necessary for an L1-cached buffer
987 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
988 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
989 * words in pointers. arc_hdr_realloc() is used to switch a header between
990 * these two allocation states.
991 */
992 typedef struct l1arc_buf_hdr {
993 kmutex_t b_freeze_lock;
994 zio_cksum_t *b_freeze_cksum;
995 #ifdef ZFS_DEBUG
996 /*
997 * used for debugging wtih kmem_flags - by allocating and freeing
998 * b_thawed when the buffer is thawed, we get a record of the stack
999 * trace that thawed it.
1000 */
1001 void *b_thawed;
1002 #endif
1003
1004 arc_buf_t *b_buf;
1005 uint32_t b_bufcnt;
1006 /* for waiting on writes to complete */
1007 kcondvar_t b_cv;
1008 uint8_t b_byteswap;
1009
1010 /* protected by arc state mutex */
1011 arc_state_t *b_state;
1012 multilist_node_t b_arc_node;
1013
1014 /* updated atomically */
1015 clock_t b_arc_access;
1016
1017 /* self protecting */
1018 refcount_t b_refcnt;
1019
1020 arc_callback_t *b_acb;
1021 void *b_pdata;
1022 } l1arc_buf_hdr_t;
1023
1024 typedef struct l2arc_dev l2arc_dev_t;
1025
1026 typedef struct l2arc_buf_hdr {
1027 /* protected by arc_buf_hdr mutex */
1028 l2arc_dev_t *b_dev; /* L2ARC device */
1029 uint64_t b_daddr; /* disk address, offset byte */
1030
1031 list_node_t b_l2node;
1032 } l2arc_buf_hdr_t;
1033
1034 struct arc_buf_hdr {
1035 /* protected by hash lock */
1036 dva_t b_dva;
1037 uint64_t b_birth;
1038
1039 arc_buf_contents_t b_type;
1040 arc_buf_hdr_t *b_hash_next;
1041 arc_flags_t b_flags;
1042
1043 /*
1044 * This field stores the size of the data buffer after
1045 * compression, and is set in the arc's zio completion handlers.
1046 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1047 *
1048 * While the block pointers can store up to 32MB in their psize
1049 * field, we can only store up to 32MB minus 512B. This is due
1050 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1051 * a field of zeros represents 512B in the bp). We can't use a
1052 * bias of 1 since we need to reserve a psize of zero, here, to
1053 * represent holes and embedded blocks.
1054 *
1055 * This isn't a problem in practice, since the maximum size of a
1056 * buffer is limited to 16MB, so we never need to store 32MB in
1057 * this field. Even in the upstream illumos code base, the
1058 * maximum size of a buffer is limited to 16MB.
1059 */
1060 uint16_t b_psize;
1061
1062 /*
1063 * This field stores the size of the data buffer before
1064 * compression, and cannot change once set. It is in units
1065 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1066 */
1067 uint16_t b_lsize; /* immutable */
1068 uint64_t b_spa; /* immutable */
1069
1070 /* L2ARC fields. Undefined when not in L2ARC. */
1071 l2arc_buf_hdr_t b_l2hdr;
1072 /* L1ARC fields. Undefined when in l2arc_only state */
1073 l1arc_buf_hdr_t b_l1hdr;
1074 };
1075
1076 #if defined(__FreeBSD__) && defined(_KERNEL)
1077 static int
sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)1078 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1079 {
1080 uint64_t val;
1081 int err;
1082
1083 val = arc_meta_limit;
1084 err = sysctl_handle_64(oidp, &val, 0, req);
1085 if (err != 0 || req->newptr == NULL)
1086 return (err);
1087
1088 if (val <= 0 || val > arc_c_max)
1089 return (EINVAL);
1090
1091 arc_meta_limit = val;
1092 return (0);
1093 }
1094
1095 static int
sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)1096 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1097 {
1098 uint64_t val;
1099 int err;
1100
1101 val = zfs_arc_max;
1102 err = sysctl_handle_64(oidp, &val, 0, req);
1103 if (err != 0 || req->newptr == NULL)
1104 return (err);
1105
1106 if (zfs_arc_max == 0) {
1107 /* Loader tunable so blindly set */
1108 zfs_arc_max = val;
1109 return (0);
1110 }
1111
1112 if (val < arc_abs_min || val > kmem_size())
1113 return (EINVAL);
1114 if (val < arc_c_min)
1115 return (EINVAL);
1116 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1117 return (EINVAL);
1118
1119 arc_c_max = val;
1120
1121 arc_c = arc_c_max;
1122 arc_p = (arc_c >> 1);
1123
1124 if (zfs_arc_meta_limit == 0) {
1125 /* limit meta-data to 1/4 of the arc capacity */
1126 arc_meta_limit = arc_c_max / 4;
1127 }
1128
1129 /* if kmem_flags are set, lets try to use less memory */
1130 if (kmem_debugging())
1131 arc_c = arc_c / 2;
1132
1133 zfs_arc_max = arc_c;
1134
1135 return (0);
1136 }
1137
1138 static int
sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)1139 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1140 {
1141 uint64_t val;
1142 int err;
1143
1144 val = zfs_arc_min;
1145 err = sysctl_handle_64(oidp, &val, 0, req);
1146 if (err != 0 || req->newptr == NULL)
1147 return (err);
1148
1149 if (zfs_arc_min == 0) {
1150 /* Loader tunable so blindly set */
1151 zfs_arc_min = val;
1152 return (0);
1153 }
1154
1155 if (val < arc_abs_min || val > arc_c_max)
1156 return (EINVAL);
1157
1158 arc_c_min = val;
1159
1160 if (zfs_arc_meta_min == 0)
1161 arc_meta_min = arc_c_min / 2;
1162
1163 if (arc_c < arc_c_min)
1164 arc_c = arc_c_min;
1165
1166 zfs_arc_min = arc_c_min;
1167
1168 return (0);
1169 }
1170 #endif
1171
1172 #define GHOST_STATE(state) \
1173 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1174 (state) == arc_l2c_only)
1175
1176 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1177 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1178 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1179 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1180 #define HDR_COMPRESSION_ENABLED(hdr) \
1181 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1182
1183 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1184 #define HDR_L2_READING(hdr) \
1185 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1186 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1187 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1188 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1189 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1190 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1191
1192 #define HDR_ISTYPE_METADATA(hdr) \
1193 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1194 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1195
1196 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1197 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1198
1199 /* For storing compression mode in b_flags */
1200 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1201
1202 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1203 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1204 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1205 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1206
1207 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1208
1209 /*
1210 * Other sizes
1211 */
1212
1213 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1214 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1215
1216 /*
1217 * Hash table routines
1218 */
1219
1220 #define HT_LOCK_PAD CACHE_LINE_SIZE
1221
1222 struct ht_lock {
1223 kmutex_t ht_lock;
1224 #ifdef _KERNEL
1225 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1226 #endif
1227 };
1228
1229 #define BUF_LOCKS 256
1230 typedef struct buf_hash_table {
1231 uint64_t ht_mask;
1232 arc_buf_hdr_t **ht_table;
1233 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1234 } buf_hash_table_t;
1235
1236 static buf_hash_table_t buf_hash_table;
1237
1238 #define BUF_HASH_INDEX(spa, dva, birth) \
1239 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1240 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1241 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1242 #define HDR_LOCK(hdr) \
1243 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1244
1245 uint64_t zfs_crc64_table[256];
1246
1247 /*
1248 * Level 2 ARC
1249 */
1250
1251 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1252 #define L2ARC_HEADROOM 2 /* num of writes */
1253 /*
1254 * If we discover during ARC scan any buffers to be compressed, we boost
1255 * our headroom for the next scanning cycle by this percentage multiple.
1256 */
1257 #define L2ARC_HEADROOM_BOOST 200
1258 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1259 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1260
1261 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1262 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1263
1264 /* L2ARC Performance Tunables */
1265 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1266 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1267 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1268 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1269 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1270 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1271 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1272 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1273 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1274
1275 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1276 &l2arc_write_max, 0, "max write size");
1277 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1278 &l2arc_write_boost, 0, "extra write during warmup");
1279 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1280 &l2arc_headroom, 0, "number of dev writes");
1281 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1282 &l2arc_feed_secs, 0, "interval seconds");
1283 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1284 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1285
1286 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1287 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1288 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1289 &l2arc_feed_again, 0, "turbo warmup");
1290 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1291 &l2arc_norw, 0, "no reads during writes");
1292
1293 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1294 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1295 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1296 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1297 "size of anonymous state");
1298 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1299 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1300 "size of anonymous state");
1301
1302 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1303 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1304 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1305 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1306 "size of metadata in mru state");
1307 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1308 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1309 "size of data in mru state");
1310
1311 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1312 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1313 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1314 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1315 "size of metadata in mru ghost state");
1316 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1317 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1318 "size of data in mru ghost state");
1319
1320 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1321 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1322 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1323 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1324 "size of metadata in mfu state");
1325 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1326 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1327 "size of data in mfu state");
1328
1329 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1330 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1331 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1332 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1333 "size of metadata in mfu ghost state");
1334 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1335 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1336 "size of data in mfu ghost state");
1337
1338 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1339 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1340
1341 /*
1342 * L2ARC Internals
1343 */
1344 struct l2arc_dev {
1345 vdev_t *l2ad_vdev; /* vdev */
1346 spa_t *l2ad_spa; /* spa */
1347 uint64_t l2ad_hand; /* next write location */
1348 uint64_t l2ad_start; /* first addr on device */
1349 uint64_t l2ad_end; /* last addr on device */
1350 boolean_t l2ad_first; /* first sweep through */
1351 boolean_t l2ad_writing; /* currently writing */
1352 kmutex_t l2ad_mtx; /* lock for buffer list */
1353 list_t l2ad_buflist; /* buffer list */
1354 list_node_t l2ad_node; /* device list node */
1355 refcount_t l2ad_alloc; /* allocated bytes */
1356 };
1357
1358 static list_t L2ARC_dev_list; /* device list */
1359 static list_t *l2arc_dev_list; /* device list pointer */
1360 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1361 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1362 static list_t L2ARC_free_on_write; /* free after write buf list */
1363 static list_t *l2arc_free_on_write; /* free after write list ptr */
1364 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1365 static uint64_t l2arc_ndev; /* number of devices */
1366
1367 typedef struct l2arc_read_callback {
1368 arc_buf_hdr_t *l2rcb_hdr; /* read buffer */
1369 blkptr_t l2rcb_bp; /* original blkptr */
1370 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1371 int l2rcb_flags; /* original flags */
1372 void *l2rcb_data; /* temporary buffer */
1373 } l2arc_read_callback_t;
1374
1375 typedef struct l2arc_write_callback {
1376 l2arc_dev_t *l2wcb_dev; /* device info */
1377 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1378 } l2arc_write_callback_t;
1379
1380 typedef struct l2arc_data_free {
1381 /* protected by l2arc_free_on_write_mtx */
1382 void *l2df_data;
1383 size_t l2df_size;
1384 arc_buf_contents_t l2df_type;
1385 list_node_t l2df_list_node;
1386 } l2arc_data_free_t;
1387
1388 static kmutex_t l2arc_feed_thr_lock;
1389 static kcondvar_t l2arc_feed_thr_cv;
1390 static uint8_t l2arc_thread_exit;
1391
1392 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1393 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1394 static void arc_hdr_free_pdata(arc_buf_hdr_t *hdr);
1395 static void arc_hdr_alloc_pdata(arc_buf_hdr_t *);
1396 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1397 static boolean_t arc_is_overflowing();
1398 static void arc_buf_watch(arc_buf_t *);
1399
1400 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1401 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1402 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1403 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1404
1405 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1406 static void l2arc_read_done(zio_t *);
1407
1408 static void
l2arc_trim(const arc_buf_hdr_t * hdr)1409 l2arc_trim(const arc_buf_hdr_t *hdr)
1410 {
1411 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1412
1413 ASSERT(HDR_HAS_L2HDR(hdr));
1414 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1415
1416 if (HDR_GET_PSIZE(hdr) != 0) {
1417 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1418 HDR_GET_PSIZE(hdr), 0);
1419 }
1420 }
1421
1422 static uint64_t
buf_hash(uint64_t spa,const dva_t * dva,uint64_t birth)1423 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1424 {
1425 uint8_t *vdva = (uint8_t *)dva;
1426 uint64_t crc = -1ULL;
1427 int i;
1428
1429 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1430
1431 for (i = 0; i < sizeof (dva_t); i++)
1432 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1433
1434 crc ^= (spa>>8) ^ birth;
1435
1436 return (crc);
1437 }
1438
1439 #define HDR_EMPTY(hdr) \
1440 ((hdr)->b_dva.dva_word[0] == 0 && \
1441 (hdr)->b_dva.dva_word[1] == 0)
1442
1443 #define HDR_EQUAL(spa, dva, birth, hdr) \
1444 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1445 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1446 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1447
1448 static void
buf_discard_identity(arc_buf_hdr_t * hdr)1449 buf_discard_identity(arc_buf_hdr_t *hdr)
1450 {
1451 hdr->b_dva.dva_word[0] = 0;
1452 hdr->b_dva.dva_word[1] = 0;
1453 hdr->b_birth = 0;
1454 }
1455
1456 static arc_buf_hdr_t *
buf_hash_find(uint64_t spa,const blkptr_t * bp,kmutex_t ** lockp)1457 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1458 {
1459 const dva_t *dva = BP_IDENTITY(bp);
1460 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1461 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1462 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1463 arc_buf_hdr_t *hdr;
1464
1465 mutex_enter(hash_lock);
1466 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1467 hdr = hdr->b_hash_next) {
1468 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1469 *lockp = hash_lock;
1470 return (hdr);
1471 }
1472 }
1473 mutex_exit(hash_lock);
1474 *lockp = NULL;
1475 return (NULL);
1476 }
1477
1478 /*
1479 * Insert an entry into the hash table. If there is already an element
1480 * equal to elem in the hash table, then the already existing element
1481 * will be returned and the new element will not be inserted.
1482 * Otherwise returns NULL.
1483 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1484 */
1485 static arc_buf_hdr_t *
buf_hash_insert(arc_buf_hdr_t * hdr,kmutex_t ** lockp)1486 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1487 {
1488 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1489 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1490 arc_buf_hdr_t *fhdr;
1491 uint32_t i;
1492
1493 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1494 ASSERT(hdr->b_birth != 0);
1495 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1496
1497 if (lockp != NULL) {
1498 *lockp = hash_lock;
1499 mutex_enter(hash_lock);
1500 } else {
1501 ASSERT(MUTEX_HELD(hash_lock));
1502 }
1503
1504 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1505 fhdr = fhdr->b_hash_next, i++) {
1506 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1507 return (fhdr);
1508 }
1509
1510 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1511 buf_hash_table.ht_table[idx] = hdr;
1512 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1513
1514 /* collect some hash table performance data */
1515 if (i > 0) {
1516 ARCSTAT_BUMP(arcstat_hash_collisions);
1517 if (i == 1)
1518 ARCSTAT_BUMP(arcstat_hash_chains);
1519
1520 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1521 }
1522
1523 ARCSTAT_BUMP(arcstat_hash_elements);
1524 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1525
1526 return (NULL);
1527 }
1528
1529 static void
buf_hash_remove(arc_buf_hdr_t * hdr)1530 buf_hash_remove(arc_buf_hdr_t *hdr)
1531 {
1532 arc_buf_hdr_t *fhdr, **hdrp;
1533 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1534
1535 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1536 ASSERT(HDR_IN_HASH_TABLE(hdr));
1537
1538 hdrp = &buf_hash_table.ht_table[idx];
1539 while ((fhdr = *hdrp) != hdr) {
1540 ASSERT3P(fhdr, !=, NULL);
1541 hdrp = &fhdr->b_hash_next;
1542 }
1543 *hdrp = hdr->b_hash_next;
1544 hdr->b_hash_next = NULL;
1545 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1546
1547 /* collect some hash table performance data */
1548 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1549
1550 if (buf_hash_table.ht_table[idx] &&
1551 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1552 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1553 }
1554
1555 /*
1556 * Global data structures and functions for the buf kmem cache.
1557 */
1558 static kmem_cache_t *hdr_full_cache;
1559 static kmem_cache_t *hdr_l2only_cache;
1560 static kmem_cache_t *buf_cache;
1561
1562 static void
buf_fini(void)1563 buf_fini(void)
1564 {
1565 int i;
1566
1567 kmem_free(buf_hash_table.ht_table,
1568 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1569 for (i = 0; i < BUF_LOCKS; i++)
1570 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1571 kmem_cache_destroy(hdr_full_cache);
1572 kmem_cache_destroy(hdr_l2only_cache);
1573 kmem_cache_destroy(buf_cache);
1574 }
1575
1576 /*
1577 * Constructor callback - called when the cache is empty
1578 * and a new buf is requested.
1579 */
1580 /* ARGSUSED */
1581 static int
hdr_full_cons(void * vbuf,void * unused,int kmflag)1582 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1583 {
1584 arc_buf_hdr_t *hdr = vbuf;
1585
1586 bzero(hdr, HDR_FULL_SIZE);
1587 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1588 refcount_create(&hdr->b_l1hdr.b_refcnt);
1589 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1590 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1591 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1592
1593 return (0);
1594 }
1595
1596 /* ARGSUSED */
1597 static int
hdr_l2only_cons(void * vbuf,void * unused,int kmflag)1598 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1599 {
1600 arc_buf_hdr_t *hdr = vbuf;
1601
1602 bzero(hdr, HDR_L2ONLY_SIZE);
1603 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1604
1605 return (0);
1606 }
1607
1608 /* ARGSUSED */
1609 static int
buf_cons(void * vbuf,void * unused,int kmflag)1610 buf_cons(void *vbuf, void *unused, int kmflag)
1611 {
1612 arc_buf_t *buf = vbuf;
1613
1614 bzero(buf, sizeof (arc_buf_t));
1615 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1616 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1617
1618 return (0);
1619 }
1620
1621 /*
1622 * Destructor callback - called when a cached buf is
1623 * no longer required.
1624 */
1625 /* ARGSUSED */
1626 static void
hdr_full_dest(void * vbuf,void * unused)1627 hdr_full_dest(void *vbuf, void *unused)
1628 {
1629 arc_buf_hdr_t *hdr = vbuf;
1630
1631 ASSERT(HDR_EMPTY(hdr));
1632 cv_destroy(&hdr->b_l1hdr.b_cv);
1633 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1634 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1635 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1636 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1637 }
1638
1639 /* ARGSUSED */
1640 static void
hdr_l2only_dest(void * vbuf,void * unused)1641 hdr_l2only_dest(void *vbuf, void *unused)
1642 {
1643 arc_buf_hdr_t *hdr = vbuf;
1644
1645 ASSERT(HDR_EMPTY(hdr));
1646 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1647 }
1648
1649 /* ARGSUSED */
1650 static void
buf_dest(void * vbuf,void * unused)1651 buf_dest(void *vbuf, void *unused)
1652 {
1653 arc_buf_t *buf = vbuf;
1654
1655 mutex_destroy(&buf->b_evict_lock);
1656 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1657 }
1658
1659 /*
1660 * Reclaim callback -- invoked when memory is low.
1661 */
1662 /* ARGSUSED */
1663 static void
hdr_recl(void * unused)1664 hdr_recl(void *unused)
1665 {
1666 dprintf("hdr_recl called\n");
1667 /*
1668 * umem calls the reclaim func when we destroy the buf cache,
1669 * which is after we do arc_fini().
1670 */
1671 if (!arc_dead)
1672 cv_signal(&arc_reclaim_thread_cv);
1673 }
1674
1675 static void
buf_init(void)1676 buf_init(void)
1677 {
1678 uint64_t *ct;
1679 uint64_t hsize = 1ULL << 12;
1680 int i, j;
1681
1682 /*
1683 * The hash table is big enough to fill all of physical memory
1684 * with an average block size of zfs_arc_average_blocksize (default 8K).
1685 * By default, the table will take up
1686 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1687 */
1688 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1689 hsize <<= 1;
1690 retry:
1691 buf_hash_table.ht_mask = hsize - 1;
1692 buf_hash_table.ht_table =
1693 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1694 if (buf_hash_table.ht_table == NULL) {
1695 ASSERT(hsize > (1ULL << 8));
1696 hsize >>= 1;
1697 goto retry;
1698 }
1699
1700 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1701 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1702 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1703 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1704 NULL, NULL, 0);
1705 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1706 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1707
1708 for (i = 0; i < 256; i++)
1709 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1710 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1711
1712 for (i = 0; i < BUF_LOCKS; i++) {
1713 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1714 NULL, MUTEX_DEFAULT, NULL);
1715 }
1716 }
1717
1718 #define ARC_MINTIME (hz>>4) /* 62 ms */
1719
1720 static inline boolean_t
arc_buf_is_shared(arc_buf_t * buf)1721 arc_buf_is_shared(arc_buf_t *buf)
1722 {
1723 boolean_t shared = (buf->b_data != NULL &&
1724 buf->b_data == buf->b_hdr->b_l1hdr.b_pdata);
1725 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1726 return (shared);
1727 }
1728
1729 static inline void
arc_cksum_free(arc_buf_hdr_t * hdr)1730 arc_cksum_free(arc_buf_hdr_t *hdr)
1731 {
1732 ASSERT(HDR_HAS_L1HDR(hdr));
1733 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1734 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1735 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1736 hdr->b_l1hdr.b_freeze_cksum = NULL;
1737 }
1738 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1739 }
1740
1741 static void
arc_cksum_verify(arc_buf_t * buf)1742 arc_cksum_verify(arc_buf_t *buf)
1743 {
1744 arc_buf_hdr_t *hdr = buf->b_hdr;
1745 zio_cksum_t zc;
1746
1747 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1748 return;
1749
1750 ASSERT(HDR_HAS_L1HDR(hdr));
1751
1752 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1753 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1754 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1755 return;
1756 }
1757 fletcher_2_native(buf->b_data, HDR_GET_LSIZE(hdr), NULL, &zc);
1758 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1759 panic("buffer modified while frozen!");
1760 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1761 }
1762
1763 static boolean_t
arc_cksum_is_equal(arc_buf_hdr_t * hdr,zio_t * zio)1764 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1765 {
1766 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1767 boolean_t valid_cksum;
1768
1769 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1770 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1771
1772 /*
1773 * We rely on the blkptr's checksum to determine if the block
1774 * is valid or not. When compressed arc is enabled, the l2arc
1775 * writes the block to the l2arc just as it appears in the pool.
1776 * This allows us to use the blkptr's checksum to validate the
1777 * data that we just read off of the l2arc without having to store
1778 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1779 * arc is disabled, then the data written to the l2arc is always
1780 * uncompressed and won't match the block as it exists in the main
1781 * pool. When this is the case, we must first compress it if it is
1782 * compressed on the main pool before we can validate the checksum.
1783 */
1784 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1785 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1786 uint64_t lsize = HDR_GET_LSIZE(hdr);
1787 uint64_t csize;
1788
1789 void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr));
1790 csize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1791 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1792 if (csize < HDR_GET_PSIZE(hdr)) {
1793 /*
1794 * Compressed blocks are always a multiple of the
1795 * smallest ashift in the pool. Ideally, we would
1796 * like to round up the csize to the next
1797 * spa_min_ashift but that value may have changed
1798 * since the block was last written. Instead,
1799 * we rely on the fact that the hdr's psize
1800 * was set to the psize of the block when it was
1801 * last written. We set the csize to that value
1802 * and zero out any part that should not contain
1803 * data.
1804 */
1805 bzero((char *)cbuf + csize, HDR_GET_PSIZE(hdr) - csize);
1806 csize = HDR_GET_PSIZE(hdr);
1807 }
1808 zio_push_transform(zio, cbuf, csize, HDR_GET_PSIZE(hdr), NULL);
1809 }
1810
1811 /*
1812 * Block pointers always store the checksum for the logical data.
1813 * If the block pointer has the gang bit set, then the checksum
1814 * it represents is for the reconstituted data and not for an
1815 * individual gang member. The zio pipeline, however, must be able to
1816 * determine the checksum of each of the gang constituents so it
1817 * treats the checksum comparison differently than what we need
1818 * for l2arc blocks. This prevents us from using the
1819 * zio_checksum_error() interface directly. Instead we must call the
1820 * zio_checksum_error_impl() so that we can ensure the checksum is
1821 * generated using the correct checksum algorithm and accounts for the
1822 * logical I/O size and not just a gang fragment.
1823 */
1824 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1825 BP_GET_CHECKSUM(zio->io_bp), zio->io_data, zio->io_size,
1826 zio->io_offset, NULL) == 0);
1827 zio_pop_transforms(zio);
1828 return (valid_cksum);
1829 }
1830
1831 static void
arc_cksum_compute(arc_buf_t * buf)1832 arc_cksum_compute(arc_buf_t *buf)
1833 {
1834 arc_buf_hdr_t *hdr = buf->b_hdr;
1835
1836 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1837 return;
1838
1839 ASSERT(HDR_HAS_L1HDR(hdr));
1840 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1841 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1842 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1843 return;
1844 }
1845 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1846 KM_SLEEP);
1847 fletcher_2_native(buf->b_data, HDR_GET_LSIZE(hdr), NULL,
1848 hdr->b_l1hdr.b_freeze_cksum);
1849 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1850 #ifdef illumos
1851 arc_buf_watch(buf);
1852 #endif
1853 }
1854
1855 #ifdef illumos
1856 #ifndef _KERNEL
1857 typedef struct procctl {
1858 long cmd;
1859 prwatch_t prwatch;
1860 } procctl_t;
1861 #endif
1862
1863 /* ARGSUSED */
1864 static void
arc_buf_unwatch(arc_buf_t * buf)1865 arc_buf_unwatch(arc_buf_t *buf)
1866 {
1867 #ifndef _KERNEL
1868 if (arc_watch) {
1869 int result;
1870 procctl_t ctl;
1871 ctl.cmd = PCWATCH;
1872 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1873 ctl.prwatch.pr_size = 0;
1874 ctl.prwatch.pr_wflags = 0;
1875 result = write(arc_procfd, &ctl, sizeof (ctl));
1876 ASSERT3U(result, ==, sizeof (ctl));
1877 }
1878 #endif
1879 }
1880
1881 /* ARGSUSED */
1882 static void
arc_buf_watch(arc_buf_t * buf)1883 arc_buf_watch(arc_buf_t *buf)
1884 {
1885 #ifndef _KERNEL
1886 if (arc_watch) {
1887 int result;
1888 procctl_t ctl;
1889 ctl.cmd = PCWATCH;
1890 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1891 ctl.prwatch.pr_size = HDR_GET_LSIZE(buf->b_hdr);
1892 ctl.prwatch.pr_wflags = WA_WRITE;
1893 result = write(arc_procfd, &ctl, sizeof (ctl));
1894 ASSERT3U(result, ==, sizeof (ctl));
1895 }
1896 #endif
1897 }
1898 #endif /* illumos */
1899
1900 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t * hdr)1901 arc_buf_type(arc_buf_hdr_t *hdr)
1902 {
1903 arc_buf_contents_t type;
1904 if (HDR_ISTYPE_METADATA(hdr)) {
1905 type = ARC_BUFC_METADATA;
1906 } else {
1907 type = ARC_BUFC_DATA;
1908 }
1909 VERIFY3U(hdr->b_type, ==, type);
1910 return (type);
1911 }
1912
1913 static uint32_t
arc_bufc_to_flags(arc_buf_contents_t type)1914 arc_bufc_to_flags(arc_buf_contents_t type)
1915 {
1916 switch (type) {
1917 case ARC_BUFC_DATA:
1918 /* metadata field is 0 if buffer contains normal data */
1919 return (0);
1920 case ARC_BUFC_METADATA:
1921 return (ARC_FLAG_BUFC_METADATA);
1922 default:
1923 break;
1924 }
1925 panic("undefined ARC buffer type!");
1926 return ((uint32_t)-1);
1927 }
1928
1929 void
arc_buf_thaw(arc_buf_t * buf)1930 arc_buf_thaw(arc_buf_t *buf)
1931 {
1932 arc_buf_hdr_t *hdr = buf->b_hdr;
1933
1934 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1935 if (hdr->b_l1hdr.b_state != arc_anon)
1936 panic("modifying non-anon buffer!");
1937 if (HDR_IO_IN_PROGRESS(hdr))
1938 panic("modifying buffer while i/o in progress!");
1939 arc_cksum_verify(buf);
1940 }
1941
1942 ASSERT(HDR_HAS_L1HDR(hdr));
1943 arc_cksum_free(hdr);
1944
1945 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1946 #ifdef ZFS_DEBUG
1947 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1948 if (hdr->b_l1hdr.b_thawed != NULL)
1949 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1950 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1951 }
1952 #endif
1953
1954 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1955
1956 #ifdef illumos
1957 arc_buf_unwatch(buf);
1958 #endif
1959 }
1960
1961 void
arc_buf_freeze(arc_buf_t * buf)1962 arc_buf_freeze(arc_buf_t *buf)
1963 {
1964 arc_buf_hdr_t *hdr = buf->b_hdr;
1965 kmutex_t *hash_lock;
1966
1967 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1968 return;
1969
1970 hash_lock = HDR_LOCK(hdr);
1971 mutex_enter(hash_lock);
1972
1973 ASSERT(HDR_HAS_L1HDR(hdr));
1974 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
1975 hdr->b_l1hdr.b_state == arc_anon);
1976 arc_cksum_compute(buf);
1977 mutex_exit(hash_lock);
1978
1979 }
1980
1981 /*
1982 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1983 * the following functions should be used to ensure that the flags are
1984 * updated in a thread-safe way. When manipulating the flags either
1985 * the hash_lock must be held or the hdr must be undiscoverable. This
1986 * ensures that we're not racing with any other threads when updating
1987 * the flags.
1988 */
1989 static inline void
arc_hdr_set_flags(arc_buf_hdr_t * hdr,arc_flags_t flags)1990 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1991 {
1992 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1993 hdr->b_flags |= flags;
1994 }
1995
1996 static inline void
arc_hdr_clear_flags(arc_buf_hdr_t * hdr,arc_flags_t flags)1997 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1998 {
1999 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2000 hdr->b_flags &= ~flags;
2001 }
2002
2003 /*
2004 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2005 * done in a special way since we have to clear and set bits
2006 * at the same time. Consumers that wish to set the compression bits
2007 * must use this function to ensure that the flags are updated in
2008 * thread-safe manner.
2009 */
2010 static void
arc_hdr_set_compress(arc_buf_hdr_t * hdr,enum zio_compress cmp)2011 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2012 {
2013 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2014
2015 /*
2016 * Holes and embedded blocks will always have a psize = 0 so
2017 * we ignore the compression of the blkptr and set the
2018 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2019 * Holes and embedded blocks remain anonymous so we don't
2020 * want to uncompress them. Mark them as uncompressed.
2021 */
2022 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2023 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2024 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2025 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2026 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2027 } else {
2028 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2029 HDR_SET_COMPRESS(hdr, cmp);
2030 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2031 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2032 }
2033 }
2034
2035 static int
arc_decompress(arc_buf_t * buf)2036 arc_decompress(arc_buf_t *buf)
2037 {
2038 arc_buf_hdr_t *hdr = buf->b_hdr;
2039 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2040 int error;
2041
2042 if (arc_buf_is_shared(buf)) {
2043 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2044 } else if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) {
2045 /*
2046 * The arc_buf_hdr_t is either not compressed or is
2047 * associated with an embedded block or a hole in which
2048 * case they remain anonymous.
2049 */
2050 IMPLY(HDR_COMPRESSION_ENABLED(hdr), HDR_GET_PSIZE(hdr) == 0 ||
2051 HDR_GET_PSIZE(hdr) == HDR_GET_LSIZE(hdr));
2052 ASSERT(!HDR_SHARED_DATA(hdr));
2053 bcopy(hdr->b_l1hdr.b_pdata, buf->b_data, HDR_GET_LSIZE(hdr));
2054 } else {
2055 ASSERT(!HDR_SHARED_DATA(hdr));
2056 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2057 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2058 hdr->b_l1hdr.b_pdata, buf->b_data, HDR_GET_PSIZE(hdr),
2059 HDR_GET_LSIZE(hdr));
2060 if (error != 0) {
2061 zfs_dbgmsg("hdr %p, compress %d, psize %d, lsize %d",
2062 hdr, HDR_GET_COMPRESS(hdr), HDR_GET_PSIZE(hdr),
2063 HDR_GET_LSIZE(hdr));
2064 return (SET_ERROR(EIO));
2065 }
2066 }
2067 if (bswap != DMU_BSWAP_NUMFUNCS) {
2068 ASSERT(!HDR_SHARED_DATA(hdr));
2069 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2070 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2071 }
2072 arc_cksum_compute(buf);
2073 return (0);
2074 }
2075
2076 /*
2077 * Return the size of the block, b_pdata, that is stored in the arc_buf_hdr_t.
2078 */
2079 static uint64_t
arc_hdr_size(arc_buf_hdr_t * hdr)2080 arc_hdr_size(arc_buf_hdr_t *hdr)
2081 {
2082 uint64_t size;
2083
2084 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2085 HDR_GET_PSIZE(hdr) > 0) {
2086 size = HDR_GET_PSIZE(hdr);
2087 } else {
2088 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2089 size = HDR_GET_LSIZE(hdr);
2090 }
2091 return (size);
2092 }
2093
2094 /*
2095 * Increment the amount of evictable space in the arc_state_t's refcount.
2096 * We account for the space used by the hdr and the arc buf individually
2097 * so that we can add and remove them from the refcount individually.
2098 */
2099 static void
arc_evictable_space_increment(arc_buf_hdr_t * hdr,arc_state_t * state)2100 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2101 {
2102 arc_buf_contents_t type = arc_buf_type(hdr);
2103 uint64_t lsize = HDR_GET_LSIZE(hdr);
2104
2105 ASSERT(HDR_HAS_L1HDR(hdr));
2106
2107 if (GHOST_STATE(state)) {
2108 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2109 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2110 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2111 (void) refcount_add_many(&state->arcs_esize[type], lsize, hdr);
2112 return;
2113 }
2114
2115 ASSERT(!GHOST_STATE(state));
2116 if (hdr->b_l1hdr.b_pdata != NULL) {
2117 (void) refcount_add_many(&state->arcs_esize[type],
2118 arc_hdr_size(hdr), hdr);
2119 }
2120 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2121 buf = buf->b_next) {
2122 if (arc_buf_is_shared(buf)) {
2123 ASSERT(ARC_BUF_LAST(buf));
2124 continue;
2125 }
2126 (void) refcount_add_many(&state->arcs_esize[type], lsize, buf);
2127 }
2128 }
2129
2130 /*
2131 * Decrement the amount of evictable space in the arc_state_t's refcount.
2132 * We account for the space used by the hdr and the arc buf individually
2133 * so that we can add and remove them from the refcount individually.
2134 */
2135 static void
arc_evitable_space_decrement(arc_buf_hdr_t * hdr,arc_state_t * state)2136 arc_evitable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2137 {
2138 arc_buf_contents_t type = arc_buf_type(hdr);
2139 uint64_t lsize = HDR_GET_LSIZE(hdr);
2140
2141 ASSERT(HDR_HAS_L1HDR(hdr));
2142
2143 if (GHOST_STATE(state)) {
2144 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2145 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2146 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2147 (void) refcount_remove_many(&state->arcs_esize[type],
2148 lsize, hdr);
2149 return;
2150 }
2151
2152 ASSERT(!GHOST_STATE(state));
2153 if (hdr->b_l1hdr.b_pdata != NULL) {
2154 (void) refcount_remove_many(&state->arcs_esize[type],
2155 arc_hdr_size(hdr), hdr);
2156 }
2157 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2158 buf = buf->b_next) {
2159 if (arc_buf_is_shared(buf)) {
2160 ASSERT(ARC_BUF_LAST(buf));
2161 continue;
2162 }
2163 (void) refcount_remove_many(&state->arcs_esize[type],
2164 lsize, buf);
2165 }
2166 }
2167
2168 /*
2169 * Add a reference to this hdr indicating that someone is actively
2170 * referencing that memory. When the refcount transitions from 0 to 1,
2171 * we remove it from the respective arc_state_t list to indicate that
2172 * it is not evictable.
2173 */
2174 static void
add_reference(arc_buf_hdr_t * hdr,void * tag)2175 add_reference(arc_buf_hdr_t *hdr, void *tag)
2176 {
2177 ASSERT(HDR_HAS_L1HDR(hdr));
2178 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2179 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2180 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2181 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2182 }
2183
2184 arc_state_t *state = hdr->b_l1hdr.b_state;
2185
2186 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2187 (state != arc_anon)) {
2188 /* We don't use the L2-only state list. */
2189 if (state != arc_l2c_only) {
2190 multilist_remove(&state->arcs_list[arc_buf_type(hdr)],
2191 hdr);
2192 arc_evitable_space_decrement(hdr, state);
2193 }
2194 /* remove the prefetch flag if we get a reference */
2195 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2196 }
2197 }
2198
2199 /*
2200 * Remove a reference from this hdr. When the reference transitions from
2201 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2202 * list making it eligible for eviction.
2203 */
2204 static int
remove_reference(arc_buf_hdr_t * hdr,kmutex_t * hash_lock,void * tag)2205 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2206 {
2207 int cnt;
2208 arc_state_t *state = hdr->b_l1hdr.b_state;
2209
2210 ASSERT(HDR_HAS_L1HDR(hdr));
2211 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2212 ASSERT(!GHOST_STATE(state));
2213
2214 /*
2215 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2216 * check to prevent usage of the arc_l2c_only list.
2217 */
2218 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2219 (state != arc_anon)) {
2220 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
2221 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2222 arc_evictable_space_increment(hdr, state);
2223 }
2224 return (cnt);
2225 }
2226
2227 /*
2228 * Move the supplied buffer to the indicated state. The hash lock
2229 * for the buffer must be held by the caller.
2230 */
2231 static void
arc_change_state(arc_state_t * new_state,arc_buf_hdr_t * hdr,kmutex_t * hash_lock)2232 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2233 kmutex_t *hash_lock)
2234 {
2235 arc_state_t *old_state;
2236 int64_t refcnt;
2237 uint32_t bufcnt;
2238 boolean_t update_old, update_new;
2239 arc_buf_contents_t buftype = arc_buf_type(hdr);
2240
2241 /*
2242 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2243 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2244 * L1 hdr doesn't always exist when we change state to arc_anon before
2245 * destroying a header, in which case reallocating to add the L1 hdr is
2246 * pointless.
2247 */
2248 if (HDR_HAS_L1HDR(hdr)) {
2249 old_state = hdr->b_l1hdr.b_state;
2250 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2251 bufcnt = hdr->b_l1hdr.b_bufcnt;
2252 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pdata != NULL);
2253 } else {
2254 old_state = arc_l2c_only;
2255 refcnt = 0;
2256 bufcnt = 0;
2257 update_old = B_FALSE;
2258 }
2259 update_new = update_old;
2260
2261 ASSERT(MUTEX_HELD(hash_lock));
2262 ASSERT3P(new_state, !=, old_state);
2263 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2264 ASSERT(old_state != arc_anon || bufcnt <= 1);
2265
2266 /*
2267 * If this buffer is evictable, transfer it from the
2268 * old state list to the new state list.
2269 */
2270 if (refcnt == 0) {
2271 if (old_state != arc_anon && old_state != arc_l2c_only) {
2272 ASSERT(HDR_HAS_L1HDR(hdr));
2273 multilist_remove(&old_state->arcs_list[buftype], hdr);
2274
2275 if (GHOST_STATE(old_state)) {
2276 ASSERT0(bufcnt);
2277 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2278 update_old = B_TRUE;
2279 }
2280 arc_evitable_space_decrement(hdr, old_state);
2281 }
2282 if (new_state != arc_anon && new_state != arc_l2c_only) {
2283
2284 /*
2285 * An L1 header always exists here, since if we're
2286 * moving to some L1-cached state (i.e. not l2c_only or
2287 * anonymous), we realloc the header to add an L1hdr
2288 * beforehand.
2289 */
2290 ASSERT(HDR_HAS_L1HDR(hdr));
2291 multilist_insert(&new_state->arcs_list[buftype], hdr);
2292
2293 if (GHOST_STATE(new_state)) {
2294 ASSERT0(bufcnt);
2295 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2296 update_new = B_TRUE;
2297 }
2298 arc_evictable_space_increment(hdr, new_state);
2299 }
2300 }
2301
2302 ASSERT(!HDR_EMPTY(hdr));
2303 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2304 buf_hash_remove(hdr);
2305
2306 /* adjust state sizes (ignore arc_l2c_only) */
2307
2308 if (update_new && new_state != arc_l2c_only) {
2309 ASSERT(HDR_HAS_L1HDR(hdr));
2310 if (GHOST_STATE(new_state)) {
2311 ASSERT0(bufcnt);
2312
2313 /*
2314 * When moving a header to a ghost state, we first
2315 * remove all arc buffers. Thus, we'll have a
2316 * bufcnt of zero, and no arc buffer to use for
2317 * the reference. As a result, we use the arc
2318 * header pointer for the reference.
2319 */
2320 (void) refcount_add_many(&new_state->arcs_size,
2321 HDR_GET_LSIZE(hdr), hdr);
2322 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2323 } else {
2324 uint32_t buffers = 0;
2325
2326 /*
2327 * Each individual buffer holds a unique reference,
2328 * thus we must remove each of these references one
2329 * at a time.
2330 */
2331 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2332 buf = buf->b_next) {
2333 ASSERT3U(bufcnt, !=, 0);
2334 buffers++;
2335
2336 /*
2337 * When the arc_buf_t is sharing the data
2338 * block with the hdr, the owner of the
2339 * reference belongs to the hdr. Only
2340 * add to the refcount if the arc_buf_t is
2341 * not shared.
2342 */
2343 if (arc_buf_is_shared(buf)) {
2344 ASSERT(ARC_BUF_LAST(buf));
2345 continue;
2346 }
2347
2348 (void) refcount_add_many(&new_state->arcs_size,
2349 HDR_GET_LSIZE(hdr), buf);
2350 }
2351 ASSERT3U(bufcnt, ==, buffers);
2352
2353 if (hdr->b_l1hdr.b_pdata != NULL) {
2354 (void) refcount_add_many(&new_state->arcs_size,
2355 arc_hdr_size(hdr), hdr);
2356 } else {
2357 ASSERT(GHOST_STATE(old_state));
2358 }
2359 }
2360 }
2361
2362 if (update_old && old_state != arc_l2c_only) {
2363 ASSERT(HDR_HAS_L1HDR(hdr));
2364 if (GHOST_STATE(old_state)) {
2365 ASSERT0(bufcnt);
2366
2367 /*
2368 * When moving a header off of a ghost state,
2369 * the header will not contain any arc buffers.
2370 * We use the arc header pointer for the reference
2371 * which is exactly what we did when we put the
2372 * header on the ghost state.
2373 */
2374
2375 (void) refcount_remove_many(&old_state->arcs_size,
2376 HDR_GET_LSIZE(hdr), hdr);
2377 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2378 } else {
2379 uint32_t buffers = 0;
2380
2381 /*
2382 * Each individual buffer holds a unique reference,
2383 * thus we must remove each of these references one
2384 * at a time.
2385 */
2386 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2387 buf = buf->b_next) {
2388 ASSERT3P(bufcnt, !=, 0);
2389 buffers++;
2390
2391 /*
2392 * When the arc_buf_t is sharing the data
2393 * block with the hdr, the owner of the
2394 * reference belongs to the hdr. Only
2395 * add to the refcount if the arc_buf_t is
2396 * not shared.
2397 */
2398 if (arc_buf_is_shared(buf)) {
2399 ASSERT(ARC_BUF_LAST(buf));
2400 continue;
2401 }
2402
2403 (void) refcount_remove_many(
2404 &old_state->arcs_size, HDR_GET_LSIZE(hdr),
2405 buf);
2406 }
2407 ASSERT3U(bufcnt, ==, buffers);
2408 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2409 (void) refcount_remove_many(
2410 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2411 }
2412 }
2413
2414 if (HDR_HAS_L1HDR(hdr))
2415 hdr->b_l1hdr.b_state = new_state;
2416
2417 /*
2418 * L2 headers should never be on the L2 state list since they don't
2419 * have L1 headers allocated.
2420 */
2421 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2422 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2423 }
2424
2425 void
arc_space_consume(uint64_t space,arc_space_type_t type)2426 arc_space_consume(uint64_t space, arc_space_type_t type)
2427 {
2428 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2429
2430 switch (type) {
2431 case ARC_SPACE_DATA:
2432 ARCSTAT_INCR(arcstat_data_size, space);
2433 break;
2434 case ARC_SPACE_META:
2435 ARCSTAT_INCR(arcstat_metadata_size, space);
2436 break;
2437 case ARC_SPACE_OTHER:
2438 ARCSTAT_INCR(arcstat_other_size, space);
2439 break;
2440 case ARC_SPACE_HDRS:
2441 ARCSTAT_INCR(arcstat_hdr_size, space);
2442 break;
2443 case ARC_SPACE_L2HDRS:
2444 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2445 break;
2446 }
2447
2448 if (type != ARC_SPACE_DATA)
2449 ARCSTAT_INCR(arcstat_meta_used, space);
2450
2451 atomic_add_64(&arc_size, space);
2452 }
2453
2454 void
arc_space_return(uint64_t space,arc_space_type_t type)2455 arc_space_return(uint64_t space, arc_space_type_t type)
2456 {
2457 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2458
2459 switch (type) {
2460 case ARC_SPACE_DATA:
2461 ARCSTAT_INCR(arcstat_data_size, -space);
2462 break;
2463 case ARC_SPACE_META:
2464 ARCSTAT_INCR(arcstat_metadata_size, -space);
2465 break;
2466 case ARC_SPACE_OTHER:
2467 ARCSTAT_INCR(arcstat_other_size, -space);
2468 break;
2469 case ARC_SPACE_HDRS:
2470 ARCSTAT_INCR(arcstat_hdr_size, -space);
2471 break;
2472 case ARC_SPACE_L2HDRS:
2473 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2474 break;
2475 }
2476
2477 if (type != ARC_SPACE_DATA) {
2478 ASSERT(arc_meta_used >= space);
2479 if (arc_meta_max < arc_meta_used)
2480 arc_meta_max = arc_meta_used;
2481 ARCSTAT_INCR(arcstat_meta_used, -space);
2482 }
2483
2484 ASSERT(arc_size >= space);
2485 atomic_add_64(&arc_size, -space);
2486 }
2487
2488 /*
2489 * Allocate an initial buffer for this hdr, subsequent buffers will
2490 * use arc_buf_clone().
2491 */
2492 static arc_buf_t *
arc_buf_alloc_impl(arc_buf_hdr_t * hdr,void * tag)2493 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag)
2494 {
2495 arc_buf_t *buf;
2496
2497 ASSERT(HDR_HAS_L1HDR(hdr));
2498 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2499 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2500 hdr->b_type == ARC_BUFC_METADATA);
2501
2502 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2503 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2504 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2505
2506 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2507 buf->b_hdr = hdr;
2508 buf->b_data = NULL;
2509 buf->b_next = NULL;
2510
2511 add_reference(hdr, tag);
2512
2513 /*
2514 * We're about to change the hdr's b_flags. We must either
2515 * hold the hash_lock or be undiscoverable.
2516 */
2517 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2518
2519 /*
2520 * If the hdr's data can be shared (no byteswapping, hdr is
2521 * uncompressed, hdr's data is not currently being written to the
2522 * L2ARC write) then we share the data buffer and set the appropriate
2523 * bit in the hdr's b_flags to indicate the hdr is sharing it's
2524 * b_pdata with the arc_buf_t. Otherwise, we allocate a new buffer to
2525 * store the buf's data.
2526 */
2527 if (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2528 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF && !HDR_L2_WRITING(hdr)) {
2529 buf->b_data = hdr->b_l1hdr.b_pdata;
2530 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2531 } else {
2532 buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2533 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2534 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2535 }
2536 VERIFY3P(buf->b_data, !=, NULL);
2537
2538 hdr->b_l1hdr.b_buf = buf;
2539 hdr->b_l1hdr.b_bufcnt += 1;
2540
2541 return (buf);
2542 }
2543
2544 /*
2545 * Used when allocating additional buffers.
2546 */
2547 static arc_buf_t *
arc_buf_clone(arc_buf_t * from)2548 arc_buf_clone(arc_buf_t *from)
2549 {
2550 arc_buf_t *buf;
2551 arc_buf_hdr_t *hdr = from->b_hdr;
2552 uint64_t size = HDR_GET_LSIZE(hdr);
2553
2554 ASSERT(HDR_HAS_L1HDR(hdr));
2555 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2556
2557 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2558 buf->b_hdr = hdr;
2559 buf->b_data = NULL;
2560 buf->b_next = hdr->b_l1hdr.b_buf;
2561 hdr->b_l1hdr.b_buf = buf;
2562 buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2563 bcopy(from->b_data, buf->b_data, size);
2564 hdr->b_l1hdr.b_bufcnt += 1;
2565
2566 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2567 return (buf);
2568 }
2569
2570 static char *arc_onloan_tag = "onloan";
2571
2572 /*
2573 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2574 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2575 * buffers must be returned to the arc before they can be used by the DMU or
2576 * freed.
2577 */
2578 arc_buf_t *
arc_loan_buf(spa_t * spa,int size)2579 arc_loan_buf(spa_t *spa, int size)
2580 {
2581 arc_buf_t *buf;
2582
2583 buf = arc_alloc_buf(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
2584
2585 atomic_add_64(&arc_loaned_bytes, size);
2586 return (buf);
2587 }
2588
2589 /*
2590 * Return a loaned arc buffer to the arc.
2591 */
2592 void
arc_return_buf(arc_buf_t * buf,void * tag)2593 arc_return_buf(arc_buf_t *buf, void *tag)
2594 {
2595 arc_buf_hdr_t *hdr = buf->b_hdr;
2596
2597 ASSERT3P(buf->b_data, !=, NULL);
2598 ASSERT(HDR_HAS_L1HDR(hdr));
2599 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2600 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2601
2602 atomic_add_64(&arc_loaned_bytes, -HDR_GET_LSIZE(hdr));
2603 }
2604
2605 /* Detach an arc_buf from a dbuf (tag) */
2606 void
arc_loan_inuse_buf(arc_buf_t * buf,void * tag)2607 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2608 {
2609 arc_buf_hdr_t *hdr = buf->b_hdr;
2610
2611 ASSERT3P(buf->b_data, !=, NULL);
2612 ASSERT(HDR_HAS_L1HDR(hdr));
2613 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2614 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2615
2616 atomic_add_64(&arc_loaned_bytes, HDR_GET_LSIZE(hdr));
2617 }
2618
2619 static void
l2arc_free_data_on_write(void * data,size_t size,arc_buf_contents_t type)2620 l2arc_free_data_on_write(void *data, size_t size, arc_buf_contents_t type)
2621 {
2622 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2623
2624 df->l2df_data = data;
2625 df->l2df_size = size;
2626 df->l2df_type = type;
2627 mutex_enter(&l2arc_free_on_write_mtx);
2628 list_insert_head(l2arc_free_on_write, df);
2629 mutex_exit(&l2arc_free_on_write_mtx);
2630 }
2631
2632 static void
arc_hdr_free_on_write(arc_buf_hdr_t * hdr)2633 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2634 {
2635 arc_state_t *state = hdr->b_l1hdr.b_state;
2636 arc_buf_contents_t type = arc_buf_type(hdr);
2637 uint64_t size = arc_hdr_size(hdr);
2638
2639 /* protected by hash lock, if in the hash table */
2640 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2641 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2642 ASSERT(state != arc_anon && state != arc_l2c_only);
2643
2644 (void) refcount_remove_many(&state->arcs_esize[type],
2645 size, hdr);
2646 }
2647 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2648 if (type == ARC_BUFC_METADATA) {
2649 arc_space_return(size, ARC_SPACE_META);
2650 } else {
2651 ASSERT(type == ARC_BUFC_DATA);
2652 arc_space_return(size, ARC_SPACE_DATA);
2653 }
2654
2655 l2arc_free_data_on_write(hdr->b_l1hdr.b_pdata, size, type);
2656 }
2657
2658 /*
2659 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2660 * data buffer, we transfer the refcount ownership to the hdr and update
2661 * the appropriate kstats.
2662 */
2663 static void
arc_share_buf(arc_buf_hdr_t * hdr,arc_buf_t * buf)2664 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2665 {
2666 arc_state_t *state = hdr->b_l1hdr.b_state;
2667
2668 ASSERT(!HDR_SHARED_DATA(hdr));
2669 ASSERT(!arc_buf_is_shared(buf));
2670 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2671 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2672
2673 /*
2674 * Start sharing the data buffer. We transfer the
2675 * refcount ownership to the hdr since it always owns
2676 * the refcount whenever an arc_buf_t is shared.
2677 */
2678 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2679 hdr->b_l1hdr.b_pdata = buf->b_data;
2680 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2681
2682 /*
2683 * Since we've transferred ownership to the hdr we need
2684 * to increment its compressed and uncompressed kstats and
2685 * decrement the overhead size.
2686 */
2687 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2688 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2689 ARCSTAT_INCR(arcstat_overhead_size, -HDR_GET_LSIZE(hdr));
2690 }
2691
2692 static void
arc_unshare_buf(arc_buf_hdr_t * hdr,arc_buf_t * buf)2693 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2694 {
2695 arc_state_t *state = hdr->b_l1hdr.b_state;
2696
2697 ASSERT(HDR_SHARED_DATA(hdr));
2698 ASSERT(arc_buf_is_shared(buf));
2699 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2700 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2701
2702 /*
2703 * We are no longer sharing this buffer so we need
2704 * to transfer its ownership to the rightful owner.
2705 */
2706 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
2707 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2708 hdr->b_l1hdr.b_pdata = NULL;
2709
2710 /*
2711 * Since the buffer is no longer shared between
2712 * the arc buf and the hdr, count it as overhead.
2713 */
2714 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2715 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2716 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2717 }
2718
2719 /*
2720 * Free up buf->b_data and if 'remove' is set, then pull the
2721 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2722 */
2723 static void
arc_buf_destroy_impl(arc_buf_t * buf,boolean_t remove)2724 arc_buf_destroy_impl(arc_buf_t *buf, boolean_t remove)
2725 {
2726 arc_buf_t **bufp;
2727 arc_buf_hdr_t *hdr = buf->b_hdr;
2728 uint64_t size = HDR_GET_LSIZE(hdr);
2729 boolean_t destroyed_buf_is_shared = arc_buf_is_shared(buf);
2730
2731 /*
2732 * Free up the data associated with the buf but only
2733 * if we're not sharing this with the hdr. If we are sharing
2734 * it with the hdr, then hdr will have performed the allocation
2735 * so allow it to do the free.
2736 */
2737 if (buf->b_data != NULL) {
2738 /*
2739 * We're about to change the hdr's b_flags. We must either
2740 * hold the hash_lock or be undiscoverable.
2741 */
2742 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2743
2744 arc_cksum_verify(buf);
2745 #ifdef illumos
2746 arc_buf_unwatch(buf);
2747 #endif
2748
2749 if (destroyed_buf_is_shared) {
2750 ASSERT(ARC_BUF_LAST(buf));
2751 ASSERT(HDR_SHARED_DATA(hdr));
2752 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2753 } else {
2754 arc_free_data_buf(hdr, buf->b_data, size, buf);
2755 ARCSTAT_INCR(arcstat_overhead_size, -size);
2756 }
2757 buf->b_data = NULL;
2758
2759 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
2760 hdr->b_l1hdr.b_bufcnt -= 1;
2761 }
2762
2763 /* only remove the buf if requested */
2764 if (!remove)
2765 return;
2766
2767 /* remove the buf from the hdr list */
2768 arc_buf_t *lastbuf = NULL;
2769 bufp = &hdr->b_l1hdr.b_buf;
2770 while (*bufp != NULL) {
2771 if (*bufp == buf)
2772 *bufp = buf->b_next;
2773
2774 /*
2775 * If we've removed a buffer in the middle of
2776 * the list then update the lastbuf and update
2777 * bufp.
2778 */
2779 if (*bufp != NULL) {
2780 lastbuf = *bufp;
2781 bufp = &(*bufp)->b_next;
2782 }
2783 }
2784 buf->b_next = NULL;
2785 ASSERT3P(lastbuf, !=, buf);
2786
2787 /*
2788 * If the current arc_buf_t is sharing its data
2789 * buffer with the hdr, then reassign the hdr's
2790 * b_pdata to share it with the new buffer at the end
2791 * of the list. The shared buffer is always the last one
2792 * on the hdr's buffer list.
2793 */
2794 if (destroyed_buf_is_shared && lastbuf != NULL) {
2795 ASSERT(ARC_BUF_LAST(buf));
2796 ASSERT(ARC_BUF_LAST(lastbuf));
2797 VERIFY(!arc_buf_is_shared(lastbuf));
2798
2799 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2800 arc_hdr_free_pdata(hdr);
2801
2802 /*
2803 * We must setup a new shared block between the
2804 * last buffer and the hdr. The data would have
2805 * been allocated by the arc buf so we need to transfer
2806 * ownership to the hdr since it's now being shared.
2807 */
2808 arc_share_buf(hdr, lastbuf);
2809 } else if (HDR_SHARED_DATA(hdr)) {
2810 ASSERT(arc_buf_is_shared(lastbuf));
2811 }
2812
2813 if (hdr->b_l1hdr.b_bufcnt == 0)
2814 arc_cksum_free(hdr);
2815
2816 /* clean up the buf */
2817 buf->b_hdr = NULL;
2818 kmem_cache_free(buf_cache, buf);
2819 }
2820
2821 static void
arc_hdr_alloc_pdata(arc_buf_hdr_t * hdr)2822 arc_hdr_alloc_pdata(arc_buf_hdr_t *hdr)
2823 {
2824 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2825 ASSERT(HDR_HAS_L1HDR(hdr));
2826 ASSERT(!HDR_SHARED_DATA(hdr));
2827
2828 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2829 hdr->b_l1hdr.b_pdata = arc_get_data_buf(hdr, arc_hdr_size(hdr), hdr);
2830 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2831 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2832
2833 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2834 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2835 }
2836
2837 static void
arc_hdr_free_pdata(arc_buf_hdr_t * hdr)2838 arc_hdr_free_pdata(arc_buf_hdr_t *hdr)
2839 {
2840 ASSERT(HDR_HAS_L1HDR(hdr));
2841 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2842
2843 /*
2844 * If the hdr is currently being written to the l2arc then
2845 * we defer freeing the data by adding it to the l2arc_free_on_write
2846 * list. The l2arc will free the data once it's finished
2847 * writing it to the l2arc device.
2848 */
2849 if (HDR_L2_WRITING(hdr)) {
2850 arc_hdr_free_on_write(hdr);
2851 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2852 } else {
2853 arc_free_data_buf(hdr, hdr->b_l1hdr.b_pdata,
2854 arc_hdr_size(hdr), hdr);
2855 }
2856 hdr->b_l1hdr.b_pdata = NULL;
2857 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2858
2859 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2860 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2861 }
2862
2863 static arc_buf_hdr_t *
arc_hdr_alloc(uint64_t spa,int32_t psize,int32_t lsize,enum zio_compress compress,arc_buf_contents_t type)2864 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
2865 enum zio_compress compress, arc_buf_contents_t type)
2866 {
2867 arc_buf_hdr_t *hdr;
2868
2869 ASSERT3U(lsize, >, 0);
2870 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
2871
2872 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2873 ASSERT(HDR_EMPTY(hdr));
2874 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2875 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
2876 HDR_SET_PSIZE(hdr, psize);
2877 HDR_SET_LSIZE(hdr, lsize);
2878 hdr->b_spa = spa;
2879 hdr->b_type = type;
2880 hdr->b_flags = 0;
2881 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
2882 arc_hdr_set_compress(hdr, compress);
2883
2884 hdr->b_l1hdr.b_state = arc_anon;
2885 hdr->b_l1hdr.b_arc_access = 0;
2886 hdr->b_l1hdr.b_bufcnt = 0;
2887 hdr->b_l1hdr.b_buf = NULL;
2888
2889 /*
2890 * Allocate the hdr's buffer. This will contain either
2891 * the compressed or uncompressed data depending on the block
2892 * it references and compressed arc enablement.
2893 */
2894 arc_hdr_alloc_pdata(hdr);
2895 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2896
2897 return (hdr);
2898 }
2899
2900 /*
2901 * Transition between the two allocation states for the arc_buf_hdr struct.
2902 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2903 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2904 * version is used when a cache buffer is only in the L2ARC in order to reduce
2905 * memory usage.
2906 */
2907 static arc_buf_hdr_t *
arc_hdr_realloc(arc_buf_hdr_t * hdr,kmem_cache_t * old,kmem_cache_t * new)2908 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
2909 {
2910 ASSERT(HDR_HAS_L2HDR(hdr));
2911
2912 arc_buf_hdr_t *nhdr;
2913 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2914
2915 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
2916 (old == hdr_l2only_cache && new == hdr_full_cache));
2917
2918 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
2919
2920 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
2921 buf_hash_remove(hdr);
2922
2923 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
2924
2925 if (new == hdr_full_cache) {
2926 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2927 /*
2928 * arc_access and arc_change_state need to be aware that a
2929 * header has just come out of L2ARC, so we set its state to
2930 * l2c_only even though it's about to change.
2931 */
2932 nhdr->b_l1hdr.b_state = arc_l2c_only;
2933
2934 /* Verify previous threads set to NULL before freeing */
2935 ASSERT3P(nhdr->b_l1hdr.b_pdata, ==, NULL);
2936 } else {
2937 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2938 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2939 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2940
2941 /*
2942 * If we've reached here, We must have been called from
2943 * arc_evict_hdr(), as such we should have already been
2944 * removed from any ghost list we were previously on
2945 * (which protects us from racing with arc_evict_state),
2946 * thus no locking is needed during this check.
2947 */
2948 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2949
2950 /*
2951 * A buffer must not be moved into the arc_l2c_only
2952 * state if it's not finished being written out to the
2953 * l2arc device. Otherwise, the b_l1hdr.b_pdata field
2954 * might try to be accessed, even though it was removed.
2955 */
2956 VERIFY(!HDR_L2_WRITING(hdr));
2957 VERIFY3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2958
2959 #ifdef ZFS_DEBUG
2960 if (hdr->b_l1hdr.b_thawed != NULL) {
2961 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2962 hdr->b_l1hdr.b_thawed = NULL;
2963 }
2964 #endif
2965
2966 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2967 }
2968 /*
2969 * The header has been reallocated so we need to re-insert it into any
2970 * lists it was on.
2971 */
2972 (void) buf_hash_insert(nhdr, NULL);
2973
2974 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
2975
2976 mutex_enter(&dev->l2ad_mtx);
2977
2978 /*
2979 * We must place the realloc'ed header back into the list at
2980 * the same spot. Otherwise, if it's placed earlier in the list,
2981 * l2arc_write_buffers() could find it during the function's
2982 * write phase, and try to write it out to the l2arc.
2983 */
2984 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
2985 list_remove(&dev->l2ad_buflist, hdr);
2986
2987 mutex_exit(&dev->l2ad_mtx);
2988
2989 /*
2990 * Since we're using the pointer address as the tag when
2991 * incrementing and decrementing the l2ad_alloc refcount, we
2992 * must remove the old pointer (that we're about to destroy) and
2993 * add the new pointer to the refcount. Otherwise we'd remove
2994 * the wrong pointer address when calling arc_hdr_destroy() later.
2995 */
2996
2997 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
2998 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
2999
3000 buf_discard_identity(hdr);
3001 kmem_cache_free(old, hdr);
3002
3003 return (nhdr);
3004 }
3005
3006 /*
3007 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3008 * The buf is returned thawed since we expect the consumer to modify it.
3009 */
3010 arc_buf_t *
arc_alloc_buf(spa_t * spa,int32_t size,void * tag,arc_buf_contents_t type)3011 arc_alloc_buf(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
3012 {
3013 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3014 ZIO_COMPRESS_OFF, type);
3015 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3016 arc_buf_t *buf = arc_buf_alloc_impl(hdr, tag);
3017 arc_buf_thaw(buf);
3018 return (buf);
3019 }
3020
3021 static void
arc_hdr_l2hdr_destroy(arc_buf_hdr_t * hdr)3022 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3023 {
3024 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3025 l2arc_dev_t *dev = l2hdr->b_dev;
3026 uint64_t asize = arc_hdr_size(hdr);
3027
3028 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3029 ASSERT(HDR_HAS_L2HDR(hdr));
3030
3031 list_remove(&dev->l2ad_buflist, hdr);
3032
3033 ARCSTAT_INCR(arcstat_l2_asize, -asize);
3034 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr));
3035
3036 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3037
3038 (void) refcount_remove_many(&dev->l2ad_alloc, asize, hdr);
3039 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3040 }
3041
3042 static void
arc_hdr_destroy(arc_buf_hdr_t * hdr)3043 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3044 {
3045 if (HDR_HAS_L1HDR(hdr)) {
3046 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3047 hdr->b_l1hdr.b_bufcnt > 0);
3048 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3049 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3050 }
3051 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3052 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3053
3054 if (!HDR_EMPTY(hdr))
3055 buf_discard_identity(hdr);
3056
3057 if (HDR_HAS_L2HDR(hdr)) {
3058 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3059 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3060
3061 if (!buflist_held)
3062 mutex_enter(&dev->l2ad_mtx);
3063
3064 /*
3065 * Even though we checked this conditional above, we
3066 * need to check this again now that we have the
3067 * l2ad_mtx. This is because we could be racing with
3068 * another thread calling l2arc_evict() which might have
3069 * destroyed this header's L2 portion as we were waiting
3070 * to acquire the l2ad_mtx. If that happens, we don't
3071 * want to re-destroy the header's L2 portion.
3072 */
3073 if (HDR_HAS_L2HDR(hdr)) {
3074 l2arc_trim(hdr);
3075 arc_hdr_l2hdr_destroy(hdr);
3076 }
3077
3078 if (!buflist_held)
3079 mutex_exit(&dev->l2ad_mtx);
3080 }
3081
3082 if (HDR_HAS_L1HDR(hdr)) {
3083 arc_cksum_free(hdr);
3084
3085 while (hdr->b_l1hdr.b_buf != NULL)
3086 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf, B_TRUE);
3087
3088 #ifdef ZFS_DEBUG
3089 if (hdr->b_l1hdr.b_thawed != NULL) {
3090 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3091 hdr->b_l1hdr.b_thawed = NULL;
3092 }
3093 #endif
3094
3095 if (hdr->b_l1hdr.b_pdata != NULL) {
3096 arc_hdr_free_pdata(hdr);
3097 }
3098 }
3099
3100 ASSERT3P(hdr->b_hash_next, ==, NULL);
3101 if (HDR_HAS_L1HDR(hdr)) {
3102 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3103 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3104 kmem_cache_free(hdr_full_cache, hdr);
3105 } else {
3106 kmem_cache_free(hdr_l2only_cache, hdr);
3107 }
3108 }
3109
3110 void
arc_buf_destroy(arc_buf_t * buf,void * tag)3111 arc_buf_destroy(arc_buf_t *buf, void* tag)
3112 {
3113 arc_buf_hdr_t *hdr = buf->b_hdr;
3114 kmutex_t *hash_lock = HDR_LOCK(hdr);
3115
3116 if (hdr->b_l1hdr.b_state == arc_anon) {
3117 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3118 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3119 VERIFY0(remove_reference(hdr, NULL, tag));
3120 arc_hdr_destroy(hdr);
3121 return;
3122 }
3123
3124 mutex_enter(hash_lock);
3125 ASSERT3P(hdr, ==, buf->b_hdr);
3126 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3127 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3128 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3129 ASSERT3P(buf->b_data, !=, NULL);
3130
3131 (void) remove_reference(hdr, hash_lock, tag);
3132 arc_buf_destroy_impl(buf, B_TRUE);
3133 mutex_exit(hash_lock);
3134 }
3135
3136 int32_t
arc_buf_size(arc_buf_t * buf)3137 arc_buf_size(arc_buf_t *buf)
3138 {
3139 return (HDR_GET_LSIZE(buf->b_hdr));
3140 }
3141
3142 /*
3143 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3144 * state of the header is dependent on its state prior to entering this
3145 * function. The following transitions are possible:
3146 *
3147 * - arc_mru -> arc_mru_ghost
3148 * - arc_mfu -> arc_mfu_ghost
3149 * - arc_mru_ghost -> arc_l2c_only
3150 * - arc_mru_ghost -> deleted
3151 * - arc_mfu_ghost -> arc_l2c_only
3152 * - arc_mfu_ghost -> deleted
3153 */
3154 static int64_t
arc_evict_hdr(arc_buf_hdr_t * hdr,kmutex_t * hash_lock)3155 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3156 {
3157 arc_state_t *evicted_state, *state;
3158 int64_t bytes_evicted = 0;
3159
3160 ASSERT(MUTEX_HELD(hash_lock));
3161 ASSERT(HDR_HAS_L1HDR(hdr));
3162
3163 state = hdr->b_l1hdr.b_state;
3164 if (GHOST_STATE(state)) {
3165 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3166 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3167
3168 /*
3169 * l2arc_write_buffers() relies on a header's L1 portion
3170 * (i.e. its b_pdata field) during its write phase.
3171 * Thus, we cannot push a header onto the arc_l2c_only
3172 * state (removing it's L1 piece) until the header is
3173 * done being written to the l2arc.
3174 */
3175 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3176 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3177 return (bytes_evicted);
3178 }
3179
3180 ARCSTAT_BUMP(arcstat_deleted);
3181 bytes_evicted += HDR_GET_LSIZE(hdr);
3182
3183 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3184
3185 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
3186 if (HDR_HAS_L2HDR(hdr)) {
3187 ASSERT(hdr->b_l1hdr.b_pdata == NULL);
3188 /*
3189 * This buffer is cached on the 2nd Level ARC;
3190 * don't destroy the header.
3191 */
3192 arc_change_state(arc_l2c_only, hdr, hash_lock);
3193 /*
3194 * dropping from L1+L2 cached to L2-only,
3195 * realloc to remove the L1 header.
3196 */
3197 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3198 hdr_l2only_cache);
3199 } else {
3200 ASSERT(hdr->b_l1hdr.b_pdata == NULL);
3201 arc_change_state(arc_anon, hdr, hash_lock);
3202 arc_hdr_destroy(hdr);
3203 }
3204 return (bytes_evicted);
3205 }
3206
3207 ASSERT(state == arc_mru || state == arc_mfu);
3208 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3209
3210 /* prefetch buffers have a minimum lifespan */
3211 if (HDR_IO_IN_PROGRESS(hdr) ||
3212 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3213 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3214 arc_min_prefetch_lifespan)) {
3215 ARCSTAT_BUMP(arcstat_evict_skip);
3216 return (bytes_evicted);
3217 }
3218
3219 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3220 while (hdr->b_l1hdr.b_buf) {
3221 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3222 if (!mutex_tryenter(&buf->b_evict_lock)) {
3223 ARCSTAT_BUMP(arcstat_mutex_miss);
3224 break;
3225 }
3226 if (buf->b_data != NULL)
3227 bytes_evicted += HDR_GET_LSIZE(hdr);
3228 mutex_exit(&buf->b_evict_lock);
3229 arc_buf_destroy_impl(buf, B_TRUE);
3230 }
3231
3232 if (HDR_HAS_L2HDR(hdr)) {
3233 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3234 } else {
3235 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3236 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3237 HDR_GET_LSIZE(hdr));
3238 } else {
3239 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3240 HDR_GET_LSIZE(hdr));
3241 }
3242 }
3243
3244 if (hdr->b_l1hdr.b_bufcnt == 0) {
3245 arc_cksum_free(hdr);
3246
3247 bytes_evicted += arc_hdr_size(hdr);
3248
3249 /*
3250 * If this hdr is being evicted and has a compressed
3251 * buffer then we discard it here before we change states.
3252 * This ensures that the accounting is updated correctly
3253 * in arc_free_data_buf().
3254 */
3255 arc_hdr_free_pdata(hdr);
3256
3257 arc_change_state(evicted_state, hdr, hash_lock);
3258 ASSERT(HDR_IN_HASH_TABLE(hdr));
3259 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3260 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3261 }
3262
3263 return (bytes_evicted);
3264 }
3265
3266 static uint64_t
arc_evict_state_impl(multilist_t * ml,int idx,arc_buf_hdr_t * marker,uint64_t spa,int64_t bytes)3267 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3268 uint64_t spa, int64_t bytes)
3269 {
3270 multilist_sublist_t *mls;
3271 uint64_t bytes_evicted = 0;
3272 arc_buf_hdr_t *hdr;
3273 kmutex_t *hash_lock;
3274 int evict_count = 0;
3275
3276 ASSERT3P(marker, !=, NULL);
3277 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3278
3279 mls = multilist_sublist_lock(ml, idx);
3280
3281 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3282 hdr = multilist_sublist_prev(mls, marker)) {
3283 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3284 (evict_count >= zfs_arc_evict_batch_limit))
3285 break;
3286
3287 /*
3288 * To keep our iteration location, move the marker
3289 * forward. Since we're not holding hdr's hash lock, we
3290 * must be very careful and not remove 'hdr' from the
3291 * sublist. Otherwise, other consumers might mistake the
3292 * 'hdr' as not being on a sublist when they call the
3293 * multilist_link_active() function (they all rely on
3294 * the hash lock protecting concurrent insertions and
3295 * removals). multilist_sublist_move_forward() was
3296 * specifically implemented to ensure this is the case
3297 * (only 'marker' will be removed and re-inserted).
3298 */
3299 multilist_sublist_move_forward(mls, marker);
3300
3301 /*
3302 * The only case where the b_spa field should ever be
3303 * zero, is the marker headers inserted by
3304 * arc_evict_state(). It's possible for multiple threads
3305 * to be calling arc_evict_state() concurrently (e.g.
3306 * dsl_pool_close() and zio_inject_fault()), so we must
3307 * skip any markers we see from these other threads.
3308 */
3309 if (hdr->b_spa == 0)
3310 continue;
3311
3312 /* we're only interested in evicting buffers of a certain spa */
3313 if (spa != 0 && hdr->b_spa != spa) {
3314 ARCSTAT_BUMP(arcstat_evict_skip);
3315 continue;
3316 }
3317
3318 hash_lock = HDR_LOCK(hdr);
3319
3320 /*
3321 * We aren't calling this function from any code path
3322 * that would already be holding a hash lock, so we're
3323 * asserting on this assumption to be defensive in case
3324 * this ever changes. Without this check, it would be
3325 * possible to incorrectly increment arcstat_mutex_miss
3326 * below (e.g. if the code changed such that we called
3327 * this function with a hash lock held).
3328 */
3329 ASSERT(!MUTEX_HELD(hash_lock));
3330
3331 if (mutex_tryenter(hash_lock)) {
3332 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3333 mutex_exit(hash_lock);
3334
3335 bytes_evicted += evicted;
3336
3337 /*
3338 * If evicted is zero, arc_evict_hdr() must have
3339 * decided to skip this header, don't increment
3340 * evict_count in this case.
3341 */
3342 if (evicted != 0)
3343 evict_count++;
3344
3345 /*
3346 * If arc_size isn't overflowing, signal any
3347 * threads that might happen to be waiting.
3348 *
3349 * For each header evicted, we wake up a single
3350 * thread. If we used cv_broadcast, we could
3351 * wake up "too many" threads causing arc_size
3352 * to significantly overflow arc_c; since
3353 * arc_get_data_buf() doesn't check for overflow
3354 * when it's woken up (it doesn't because it's
3355 * possible for the ARC to be overflowing while
3356 * full of un-evictable buffers, and the
3357 * function should proceed in this case).
3358 *
3359 * If threads are left sleeping, due to not
3360 * using cv_broadcast, they will be woken up
3361 * just before arc_reclaim_thread() sleeps.
3362 */
3363 mutex_enter(&arc_reclaim_lock);
3364 if (!arc_is_overflowing())
3365 cv_signal(&arc_reclaim_waiters_cv);
3366 mutex_exit(&arc_reclaim_lock);
3367 } else {
3368 ARCSTAT_BUMP(arcstat_mutex_miss);
3369 }
3370 }
3371
3372 multilist_sublist_unlock(mls);
3373
3374 return (bytes_evicted);
3375 }
3376
3377 /*
3378 * Evict buffers from the given arc state, until we've removed the
3379 * specified number of bytes. Move the removed buffers to the
3380 * appropriate evict state.
3381 *
3382 * This function makes a "best effort". It skips over any buffers
3383 * it can't get a hash_lock on, and so, may not catch all candidates.
3384 * It may also return without evicting as much space as requested.
3385 *
3386 * If bytes is specified using the special value ARC_EVICT_ALL, this
3387 * will evict all available (i.e. unlocked and evictable) buffers from
3388 * the given arc state; which is used by arc_flush().
3389 */
3390 static uint64_t
arc_evict_state(arc_state_t * state,uint64_t spa,int64_t bytes,arc_buf_contents_t type)3391 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3392 arc_buf_contents_t type)
3393 {
3394 uint64_t total_evicted = 0;
3395 multilist_t *ml = &state->arcs_list[type];
3396 int num_sublists;
3397 arc_buf_hdr_t **markers;
3398
3399 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3400
3401 num_sublists = multilist_get_num_sublists(ml);
3402
3403 /*
3404 * If we've tried to evict from each sublist, made some
3405 * progress, but still have not hit the target number of bytes
3406 * to evict, we want to keep trying. The markers allow us to
3407 * pick up where we left off for each individual sublist, rather
3408 * than starting from the tail each time.
3409 */
3410 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3411 for (int i = 0; i < num_sublists; i++) {
3412 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3413
3414 /*
3415 * A b_spa of 0 is used to indicate that this header is
3416 * a marker. This fact is used in arc_adjust_type() and
3417 * arc_evict_state_impl().
3418 */
3419 markers[i]->b_spa = 0;
3420
3421 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3422 multilist_sublist_insert_tail(mls, markers[i]);
3423 multilist_sublist_unlock(mls);
3424 }
3425
3426 /*
3427 * While we haven't hit our target number of bytes to evict, or
3428 * we're evicting all available buffers.
3429 */
3430 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3431 /*
3432 * Start eviction using a randomly selected sublist,
3433 * this is to try and evenly balance eviction across all
3434 * sublists. Always starting at the same sublist
3435 * (e.g. index 0) would cause evictions to favor certain
3436 * sublists over others.
3437 */
3438 int sublist_idx = multilist_get_random_index(ml);
3439 uint64_t scan_evicted = 0;
3440
3441 for (int i = 0; i < num_sublists; i++) {
3442 uint64_t bytes_remaining;
3443 uint64_t bytes_evicted;
3444
3445 if (bytes == ARC_EVICT_ALL)
3446 bytes_remaining = ARC_EVICT_ALL;
3447 else if (total_evicted < bytes)
3448 bytes_remaining = bytes - total_evicted;
3449 else
3450 break;
3451
3452 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3453 markers[sublist_idx], spa, bytes_remaining);
3454
3455 scan_evicted += bytes_evicted;
3456 total_evicted += bytes_evicted;
3457
3458 /* we've reached the end, wrap to the beginning */
3459 if (++sublist_idx >= num_sublists)
3460 sublist_idx = 0;
3461 }
3462
3463 /*
3464 * If we didn't evict anything during this scan, we have
3465 * no reason to believe we'll evict more during another
3466 * scan, so break the loop.
3467 */
3468 if (scan_evicted == 0) {
3469 /* This isn't possible, let's make that obvious */
3470 ASSERT3S(bytes, !=, 0);
3471
3472 /*
3473 * When bytes is ARC_EVICT_ALL, the only way to
3474 * break the loop is when scan_evicted is zero.
3475 * In that case, we actually have evicted enough,
3476 * so we don't want to increment the kstat.
3477 */
3478 if (bytes != ARC_EVICT_ALL) {
3479 ASSERT3S(total_evicted, <, bytes);
3480 ARCSTAT_BUMP(arcstat_evict_not_enough);
3481 }
3482
3483 break;
3484 }
3485 }
3486
3487 for (int i = 0; i < num_sublists; i++) {
3488 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3489 multilist_sublist_remove(mls, markers[i]);
3490 multilist_sublist_unlock(mls);
3491
3492 kmem_cache_free(hdr_full_cache, markers[i]);
3493 }
3494 kmem_free(markers, sizeof (*markers) * num_sublists);
3495
3496 return (total_evicted);
3497 }
3498
3499 /*
3500 * Flush all "evictable" data of the given type from the arc state
3501 * specified. This will not evict any "active" buffers (i.e. referenced).
3502 *
3503 * When 'retry' is set to B_FALSE, the function will make a single pass
3504 * over the state and evict any buffers that it can. Since it doesn't
3505 * continually retry the eviction, it might end up leaving some buffers
3506 * in the ARC due to lock misses.
3507 *
3508 * When 'retry' is set to B_TRUE, the function will continually retry the
3509 * eviction until *all* evictable buffers have been removed from the
3510 * state. As a result, if concurrent insertions into the state are
3511 * allowed (e.g. if the ARC isn't shutting down), this function might
3512 * wind up in an infinite loop, continually trying to evict buffers.
3513 */
3514 static uint64_t
arc_flush_state(arc_state_t * state,uint64_t spa,arc_buf_contents_t type,boolean_t retry)3515 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3516 boolean_t retry)
3517 {
3518 uint64_t evicted = 0;
3519
3520 while (refcount_count(&state->arcs_esize[type]) != 0) {
3521 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3522
3523 if (!retry)
3524 break;
3525 }
3526
3527 return (evicted);
3528 }
3529
3530 /*
3531 * Evict the specified number of bytes from the state specified,
3532 * restricting eviction to the spa and type given. This function
3533 * prevents us from trying to evict more from a state's list than
3534 * is "evictable", and to skip evicting altogether when passed a
3535 * negative value for "bytes". In contrast, arc_evict_state() will
3536 * evict everything it can, when passed a negative value for "bytes".
3537 */
3538 static uint64_t
arc_adjust_impl(arc_state_t * state,uint64_t spa,int64_t bytes,arc_buf_contents_t type)3539 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3540 arc_buf_contents_t type)
3541 {
3542 int64_t delta;
3543
3544 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3545 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3546 return (arc_evict_state(state, spa, delta, type));
3547 }
3548
3549 return (0);
3550 }
3551
3552 /*
3553 * Evict metadata buffers from the cache, such that arc_meta_used is
3554 * capped by the arc_meta_limit tunable.
3555 */
3556 static uint64_t
arc_adjust_meta(void)3557 arc_adjust_meta(void)
3558 {
3559 uint64_t total_evicted = 0;
3560 int64_t target;
3561
3562 /*
3563 * If we're over the meta limit, we want to evict enough
3564 * metadata to get back under the meta limit. We don't want to
3565 * evict so much that we drop the MRU below arc_p, though. If
3566 * we're over the meta limit more than we're over arc_p, we
3567 * evict some from the MRU here, and some from the MFU below.
3568 */
3569 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3570 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3571 refcount_count(&arc_mru->arcs_size) - arc_p));
3572
3573 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3574
3575 /*
3576 * Similar to the above, we want to evict enough bytes to get us
3577 * below the meta limit, but not so much as to drop us below the
3578 * space alloted to the MFU (which is defined as arc_c - arc_p).
3579 */
3580 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3581 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3582
3583 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3584
3585 return (total_evicted);
3586 }
3587
3588 /*
3589 * Return the type of the oldest buffer in the given arc state
3590 *
3591 * This function will select a random sublist of type ARC_BUFC_DATA and
3592 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3593 * is compared, and the type which contains the "older" buffer will be
3594 * returned.
3595 */
3596 static arc_buf_contents_t
arc_adjust_type(arc_state_t * state)3597 arc_adjust_type(arc_state_t *state)
3598 {
3599 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
3600 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
3601 int data_idx = multilist_get_random_index(data_ml);
3602 int meta_idx = multilist_get_random_index(meta_ml);
3603 multilist_sublist_t *data_mls;
3604 multilist_sublist_t *meta_mls;
3605 arc_buf_contents_t type;
3606 arc_buf_hdr_t *data_hdr;
3607 arc_buf_hdr_t *meta_hdr;
3608
3609 /*
3610 * We keep the sublist lock until we're finished, to prevent
3611 * the headers from being destroyed via arc_evict_state().
3612 */
3613 data_mls = multilist_sublist_lock(data_ml, data_idx);
3614 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3615
3616 /*
3617 * These two loops are to ensure we skip any markers that
3618 * might be at the tail of the lists due to arc_evict_state().
3619 */
3620
3621 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3622 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3623 if (data_hdr->b_spa != 0)
3624 break;
3625 }
3626
3627 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3628 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3629 if (meta_hdr->b_spa != 0)
3630 break;
3631 }
3632
3633 if (data_hdr == NULL && meta_hdr == NULL) {
3634 type = ARC_BUFC_DATA;
3635 } else if (data_hdr == NULL) {
3636 ASSERT3P(meta_hdr, !=, NULL);
3637 type = ARC_BUFC_METADATA;
3638 } else if (meta_hdr == NULL) {
3639 ASSERT3P(data_hdr, !=, NULL);
3640 type = ARC_BUFC_DATA;
3641 } else {
3642 ASSERT3P(data_hdr, !=, NULL);
3643 ASSERT3P(meta_hdr, !=, NULL);
3644
3645 /* The headers can't be on the sublist without an L1 header */
3646 ASSERT(HDR_HAS_L1HDR(data_hdr));
3647 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3648
3649 if (data_hdr->b_l1hdr.b_arc_access <
3650 meta_hdr->b_l1hdr.b_arc_access) {
3651 type = ARC_BUFC_DATA;
3652 } else {
3653 type = ARC_BUFC_METADATA;
3654 }
3655 }
3656
3657 multilist_sublist_unlock(meta_mls);
3658 multilist_sublist_unlock(data_mls);
3659
3660 return (type);
3661 }
3662
3663 /*
3664 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3665 */
3666 static uint64_t
arc_adjust(void)3667 arc_adjust(void)
3668 {
3669 uint64_t total_evicted = 0;
3670 uint64_t bytes;
3671 int64_t target;
3672
3673 /*
3674 * If we're over arc_meta_limit, we want to correct that before
3675 * potentially evicting data buffers below.
3676 */
3677 total_evicted += arc_adjust_meta();
3678
3679 /*
3680 * Adjust MRU size
3681 *
3682 * If we're over the target cache size, we want to evict enough
3683 * from the list to get back to our target size. We don't want
3684 * to evict too much from the MRU, such that it drops below
3685 * arc_p. So, if we're over our target cache size more than
3686 * the MRU is over arc_p, we'll evict enough to get back to
3687 * arc_p here, and then evict more from the MFU below.
3688 */
3689 target = MIN((int64_t)(arc_size - arc_c),
3690 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3691 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
3692
3693 /*
3694 * If we're below arc_meta_min, always prefer to evict data.
3695 * Otherwise, try to satisfy the requested number of bytes to
3696 * evict from the type which contains older buffers; in an
3697 * effort to keep newer buffers in the cache regardless of their
3698 * type. If we cannot satisfy the number of bytes from this
3699 * type, spill over into the next type.
3700 */
3701 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3702 arc_meta_used > arc_meta_min) {
3703 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3704 total_evicted += bytes;
3705
3706 /*
3707 * If we couldn't evict our target number of bytes from
3708 * metadata, we try to get the rest from data.
3709 */
3710 target -= bytes;
3711
3712 total_evicted +=
3713 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3714 } else {
3715 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3716 total_evicted += bytes;
3717
3718 /*
3719 * If we couldn't evict our target number of bytes from
3720 * data, we try to get the rest from metadata.
3721 */
3722 target -= bytes;
3723
3724 total_evicted +=
3725 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3726 }
3727
3728 /*
3729 * Adjust MFU size
3730 *
3731 * Now that we've tried to evict enough from the MRU to get its
3732 * size back to arc_p, if we're still above the target cache
3733 * size, we evict the rest from the MFU.
3734 */
3735 target = arc_size - arc_c;
3736
3737 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3738 arc_meta_used > arc_meta_min) {
3739 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3740 total_evicted += bytes;
3741
3742 /*
3743 * If we couldn't evict our target number of bytes from
3744 * metadata, we try to get the rest from data.
3745 */
3746 target -= bytes;
3747
3748 total_evicted +=
3749 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3750 } else {
3751 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3752 total_evicted += bytes;
3753
3754 /*
3755 * If we couldn't evict our target number of bytes from
3756 * data, we try to get the rest from data.
3757 */
3758 target -= bytes;
3759
3760 total_evicted +=
3761 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3762 }
3763
3764 /*
3765 * Adjust ghost lists
3766 *
3767 * In addition to the above, the ARC also defines target values
3768 * for the ghost lists. The sum of the mru list and mru ghost
3769 * list should never exceed the target size of the cache, and
3770 * the sum of the mru list, mfu list, mru ghost list, and mfu
3771 * ghost list should never exceed twice the target size of the
3772 * cache. The following logic enforces these limits on the ghost
3773 * caches, and evicts from them as needed.
3774 */
3775 target = refcount_count(&arc_mru->arcs_size) +
3776 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3777
3778 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3779 total_evicted += bytes;
3780
3781 target -= bytes;
3782
3783 total_evicted +=
3784 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3785
3786 /*
3787 * We assume the sum of the mru list and mfu list is less than
3788 * or equal to arc_c (we enforced this above), which means we
3789 * can use the simpler of the two equations below:
3790 *
3791 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3792 * mru ghost + mfu ghost <= arc_c
3793 */
3794 target = refcount_count(&arc_mru_ghost->arcs_size) +
3795 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3796
3797 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3798 total_evicted += bytes;
3799
3800 target -= bytes;
3801
3802 total_evicted +=
3803 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3804
3805 return (total_evicted);
3806 }
3807
3808 void
arc_flush(spa_t * spa,boolean_t retry)3809 arc_flush(spa_t *spa, boolean_t retry)
3810 {
3811 uint64_t guid = 0;
3812
3813 /*
3814 * If retry is B_TRUE, a spa must not be specified since we have
3815 * no good way to determine if all of a spa's buffers have been
3816 * evicted from an arc state.
3817 */
3818 ASSERT(!retry || spa == 0);
3819
3820 if (spa != NULL)
3821 guid = spa_load_guid(spa);
3822
3823 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3824 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3825
3826 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3827 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3828
3829 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3830 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3831
3832 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3833 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3834 }
3835
3836 void
arc_shrink(int64_t to_free)3837 arc_shrink(int64_t to_free)
3838 {
3839 if (arc_c > arc_c_min) {
3840 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
3841 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
3842 if (arc_c > arc_c_min + to_free)
3843 atomic_add_64(&arc_c, -to_free);
3844 else
3845 arc_c = arc_c_min;
3846
3847 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3848 if (arc_c > arc_size)
3849 arc_c = MAX(arc_size, arc_c_min);
3850 if (arc_p > arc_c)
3851 arc_p = (arc_c >> 1);
3852
3853 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
3854 arc_p);
3855
3856 ASSERT(arc_c >= arc_c_min);
3857 ASSERT((int64_t)arc_p >= 0);
3858 }
3859
3860 if (arc_size > arc_c) {
3861 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
3862 uint64_t, arc_c);
3863 (void) arc_adjust();
3864 }
3865 }
3866
3867 static long needfree = 0;
3868
3869 typedef enum free_memory_reason_t {
3870 FMR_UNKNOWN,
3871 FMR_NEEDFREE,
3872 FMR_LOTSFREE,
3873 FMR_SWAPFS_MINFREE,
3874 FMR_PAGES_PP_MAXIMUM,
3875 FMR_HEAP_ARENA,
3876 FMR_ZIO_ARENA,
3877 FMR_ZIO_FRAG,
3878 } free_memory_reason_t;
3879
3880 int64_t last_free_memory;
3881 free_memory_reason_t last_free_reason;
3882
3883 /*
3884 * Additional reserve of pages for pp_reserve.
3885 */
3886 int64_t arc_pages_pp_reserve = 64;
3887
3888 /*
3889 * Additional reserve of pages for swapfs.
3890 */
3891 int64_t arc_swapfs_reserve = 64;
3892
3893 /*
3894 * Return the amount of memory that can be consumed before reclaim will be
3895 * needed. Positive if there is sufficient free memory, negative indicates
3896 * the amount of memory that needs to be freed up.
3897 */
3898 static int64_t
arc_available_memory(void)3899 arc_available_memory(void)
3900 {
3901 int64_t lowest = INT64_MAX;
3902 int64_t n;
3903 free_memory_reason_t r = FMR_UNKNOWN;
3904
3905 #ifdef _KERNEL
3906 if (needfree > 0) {
3907 n = PAGESIZE * (-needfree);
3908 if (n < lowest) {
3909 lowest = n;
3910 r = FMR_NEEDFREE;
3911 }
3912 }
3913
3914 /*
3915 * Cooperate with pagedaemon when it's time for it to scan
3916 * and reclaim some pages.
3917 */
3918 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
3919 if (n < lowest) {
3920 lowest = n;
3921 r = FMR_LOTSFREE;
3922 }
3923
3924 #ifdef illumos
3925 /*
3926 * check that we're out of range of the pageout scanner. It starts to
3927 * schedule paging if freemem is less than lotsfree and needfree.
3928 * lotsfree is the high-water mark for pageout, and needfree is the
3929 * number of needed free pages. We add extra pages here to make sure
3930 * the scanner doesn't start up while we're freeing memory.
3931 */
3932 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3933 if (n < lowest) {
3934 lowest = n;
3935 r = FMR_LOTSFREE;
3936 }
3937
3938 /*
3939 * check to make sure that swapfs has enough space so that anon
3940 * reservations can still succeed. anon_resvmem() checks that the
3941 * availrmem is greater than swapfs_minfree, and the number of reserved
3942 * swap pages. We also add a bit of extra here just to prevent
3943 * circumstances from getting really dire.
3944 */
3945 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3946 desfree - arc_swapfs_reserve);
3947 if (n < lowest) {
3948 lowest = n;
3949 r = FMR_SWAPFS_MINFREE;
3950 }
3951
3952
3953 /*
3954 * Check that we have enough availrmem that memory locking (e.g., via
3955 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3956 * stores the number of pages that cannot be locked; when availrmem
3957 * drops below pages_pp_maximum, page locking mechanisms such as
3958 * page_pp_lock() will fail.)
3959 */
3960 n = PAGESIZE * (availrmem - pages_pp_maximum -
3961 arc_pages_pp_reserve);
3962 if (n < lowest) {
3963 lowest = n;
3964 r = FMR_PAGES_PP_MAXIMUM;
3965 }
3966
3967 #endif /* illumos */
3968 #if !defined(_LP64)
3969 /*
3970 * If we're on an i386 platform, it's possible that we'll exhaust the
3971 * kernel heap space before we ever run out of available physical
3972 * memory. Most checks of the size of the heap_area compare against
3973 * tune.t_minarmem, which is the minimum available real memory that we
3974 * can have in the system. However, this is generally fixed at 25 pages
3975 * which is so low that it's useless. In this comparison, we seek to
3976 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3977 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3978 * free)
3979 */
3980 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
3981 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3982 if (n < lowest) {
3983 lowest = n;
3984 r = FMR_HEAP_ARENA;
3985 }
3986 #define zio_arena NULL
3987 #else
3988 #define zio_arena heap_arena
3989 #endif
3990
3991 /*
3992 * If zio data pages are being allocated out of a separate heap segment,
3993 * then enforce that the size of available vmem for this arena remains
3994 * above about 1/16th free.
3995 *
3996 * Note: The 1/16th arena free requirement was put in place
3997 * to aggressively evict memory from the arc in order to avoid
3998 * memory fragmentation issues.
3999 */
4000 if (zio_arena != NULL) {
4001 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4002 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
4003 if (n < lowest) {
4004 lowest = n;
4005 r = FMR_ZIO_ARENA;
4006 }
4007 }
4008
4009 #if __FreeBSD__
4010 /*
4011 * Above limits know nothing about real level of KVA fragmentation.
4012 * Start aggressive reclamation if too little sequential KVA left.
4013 */
4014 if (lowest > 0) {
4015 n = (vmem_size(heap_arena, VMEM_MAXFREE) < SPA_MAXBLOCKSIZE) ?
4016 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) :
4017 INT64_MAX;
4018 if (n < lowest) {
4019 lowest = n;
4020 r = FMR_ZIO_FRAG;
4021 }
4022 }
4023 #endif
4024
4025 #else /* _KERNEL */
4026 /* Every 100 calls, free a small amount */
4027 if (spa_get_random(100) == 0)
4028 lowest = -1024;
4029 #endif /* _KERNEL */
4030
4031 last_free_memory = lowest;
4032 last_free_reason = r;
4033 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4034 return (lowest);
4035 }
4036
4037
4038 /*
4039 * Determine if the system is under memory pressure and is asking
4040 * to reclaim memory. A return value of B_TRUE indicates that the system
4041 * is under memory pressure and that the arc should adjust accordingly.
4042 */
4043 static boolean_t
arc_reclaim_needed(void)4044 arc_reclaim_needed(void)
4045 {
4046 return (arc_available_memory() < 0);
4047 }
4048
4049 extern kmem_cache_t *zio_buf_cache[];
4050 extern kmem_cache_t *zio_data_buf_cache[];
4051 extern kmem_cache_t *range_seg_cache;
4052
4053 static __noinline void
arc_kmem_reap_now(void)4054 arc_kmem_reap_now(void)
4055 {
4056 size_t i;
4057 kmem_cache_t *prev_cache = NULL;
4058 kmem_cache_t *prev_data_cache = NULL;
4059
4060 DTRACE_PROBE(arc__kmem_reap_start);
4061 #ifdef _KERNEL
4062 if (arc_meta_used >= arc_meta_limit) {
4063 /*
4064 * We are exceeding our meta-data cache limit.
4065 * Purge some DNLC entries to release holds on meta-data.
4066 */
4067 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4068 }
4069 #if defined(__i386)
4070 /*
4071 * Reclaim unused memory from all kmem caches.
4072 */
4073 kmem_reap();
4074 #endif
4075 #endif
4076
4077 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4078 if (zio_buf_cache[i] != prev_cache) {
4079 prev_cache = zio_buf_cache[i];
4080 kmem_cache_reap_now(zio_buf_cache[i]);
4081 }
4082 if (zio_data_buf_cache[i] != prev_data_cache) {
4083 prev_data_cache = zio_data_buf_cache[i];
4084 kmem_cache_reap_now(zio_data_buf_cache[i]);
4085 }
4086 }
4087 kmem_cache_reap_now(buf_cache);
4088 kmem_cache_reap_now(hdr_full_cache);
4089 kmem_cache_reap_now(hdr_l2only_cache);
4090 kmem_cache_reap_now(range_seg_cache);
4091
4092 #ifdef illumos
4093 if (zio_arena != NULL) {
4094 /*
4095 * Ask the vmem arena to reclaim unused memory from its
4096 * quantum caches.
4097 */
4098 vmem_qcache_reap(zio_arena);
4099 }
4100 #endif
4101 DTRACE_PROBE(arc__kmem_reap_end);
4102 }
4103
4104 /*
4105 * Threads can block in arc_get_data_buf() waiting for this thread to evict
4106 * enough data and signal them to proceed. When this happens, the threads in
4107 * arc_get_data_buf() are sleeping while holding the hash lock for their
4108 * particular arc header. Thus, we must be careful to never sleep on a
4109 * hash lock in this thread. This is to prevent the following deadlock:
4110 *
4111 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
4112 * waiting for the reclaim thread to signal it.
4113 *
4114 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4115 * fails, and goes to sleep forever.
4116 *
4117 * This possible deadlock is avoided by always acquiring a hash lock
4118 * using mutex_tryenter() from arc_reclaim_thread().
4119 */
4120 static void
arc_reclaim_thread(void * dummy __unused)4121 arc_reclaim_thread(void *dummy __unused)
4122 {
4123 hrtime_t growtime = 0;
4124 callb_cpr_t cpr;
4125
4126 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4127
4128 mutex_enter(&arc_reclaim_lock);
4129 while (!arc_reclaim_thread_exit) {
4130 uint64_t evicted = 0;
4131
4132 /*
4133 * This is necessary in order for the mdb ::arc dcmd to
4134 * show up to date information. Since the ::arc command
4135 * does not call the kstat's update function, without
4136 * this call, the command may show stale stats for the
4137 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4138 * with this change, the data might be up to 1 second
4139 * out of date; but that should suffice. The arc_state_t
4140 * structures can be queried directly if more accurate
4141 * information is needed.
4142 */
4143 if (arc_ksp != NULL)
4144 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4145
4146 mutex_exit(&arc_reclaim_lock);
4147
4148 /*
4149 * We call arc_adjust() before (possibly) calling
4150 * arc_kmem_reap_now(), so that we can wake up
4151 * arc_get_data_buf() sooner.
4152 */
4153 evicted = arc_adjust();
4154
4155 int64_t free_memory = arc_available_memory();
4156 if (free_memory < 0) {
4157
4158 arc_no_grow = B_TRUE;
4159 arc_warm = B_TRUE;
4160
4161 /*
4162 * Wait at least zfs_grow_retry (default 60) seconds
4163 * before considering growing.
4164 */
4165 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4166
4167 arc_kmem_reap_now();
4168
4169 /*
4170 * If we are still low on memory, shrink the ARC
4171 * so that we have arc_shrink_min free space.
4172 */
4173 free_memory = arc_available_memory();
4174
4175 int64_t to_free =
4176 (arc_c >> arc_shrink_shift) - free_memory;
4177 if (to_free > 0) {
4178 #ifdef _KERNEL
4179 to_free = MAX(to_free, ptob(needfree));
4180 #endif
4181 arc_shrink(to_free);
4182 }
4183 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4184 arc_no_grow = B_TRUE;
4185 } else if (gethrtime() >= growtime) {
4186 arc_no_grow = B_FALSE;
4187 }
4188
4189 mutex_enter(&arc_reclaim_lock);
4190
4191 /*
4192 * If evicted is zero, we couldn't evict anything via
4193 * arc_adjust(). This could be due to hash lock
4194 * collisions, but more likely due to the majority of
4195 * arc buffers being unevictable. Therefore, even if
4196 * arc_size is above arc_c, another pass is unlikely to
4197 * be helpful and could potentially cause us to enter an
4198 * infinite loop.
4199 */
4200 if (arc_size <= arc_c || evicted == 0) {
4201 #ifdef _KERNEL
4202 needfree = 0;
4203 #endif
4204 /*
4205 * We're either no longer overflowing, or we
4206 * can't evict anything more, so we should wake
4207 * up any threads before we go to sleep.
4208 */
4209 cv_broadcast(&arc_reclaim_waiters_cv);
4210
4211 /*
4212 * Block until signaled, or after one second (we
4213 * might need to perform arc_kmem_reap_now()
4214 * even if we aren't being signalled)
4215 */
4216 CALLB_CPR_SAFE_BEGIN(&cpr);
4217 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4218 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4219 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4220 }
4221 }
4222
4223 arc_reclaim_thread_exit = B_FALSE;
4224 cv_broadcast(&arc_reclaim_thread_cv);
4225 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4226 thread_exit();
4227 }
4228
4229 #ifdef __FreeBSD__
4230
4231 static u_int arc_dnlc_evicts_arg;
4232 extern struct vfsops zfs_vfsops;
4233
4234 static void
arc_dnlc_evicts_thread(void * dummy __unused)4235 arc_dnlc_evicts_thread(void *dummy __unused)
4236 {
4237 callb_cpr_t cpr;
4238 u_int percent;
4239
4240 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4241
4242 mutex_enter(&arc_dnlc_evicts_lock);
4243 while (!arc_dnlc_evicts_thread_exit) {
4244 CALLB_CPR_SAFE_BEGIN(&cpr);
4245 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4246 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4247 if (arc_dnlc_evicts_arg != 0) {
4248 percent = arc_dnlc_evicts_arg;
4249 mutex_exit(&arc_dnlc_evicts_lock);
4250 #ifdef _KERNEL
4251 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4252 #endif
4253 mutex_enter(&arc_dnlc_evicts_lock);
4254 /*
4255 * Clear our token only after vnlru_free()
4256 * pass is done, to avoid false queueing of
4257 * the requests.
4258 */
4259 arc_dnlc_evicts_arg = 0;
4260 }
4261 }
4262 arc_dnlc_evicts_thread_exit = FALSE;
4263 cv_broadcast(&arc_dnlc_evicts_cv);
4264 CALLB_CPR_EXIT(&cpr);
4265 thread_exit();
4266 }
4267
4268 void
dnlc_reduce_cache(void * arg)4269 dnlc_reduce_cache(void *arg)
4270 {
4271 u_int percent;
4272
4273 percent = (u_int)(uintptr_t)arg;
4274 mutex_enter(&arc_dnlc_evicts_lock);
4275 if (arc_dnlc_evicts_arg == 0) {
4276 arc_dnlc_evicts_arg = percent;
4277 cv_broadcast(&arc_dnlc_evicts_cv);
4278 }
4279 mutex_exit(&arc_dnlc_evicts_lock);
4280 }
4281
4282 #endif
4283
4284 /*
4285 * Adapt arc info given the number of bytes we are trying to add and
4286 * the state that we are comming from. This function is only called
4287 * when we are adding new content to the cache.
4288 */
4289 static void
arc_adapt(int bytes,arc_state_t * state)4290 arc_adapt(int bytes, arc_state_t *state)
4291 {
4292 int mult;
4293 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4294 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4295 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4296
4297 if (state == arc_l2c_only)
4298 return;
4299
4300 ASSERT(bytes > 0);
4301 /*
4302 * Adapt the target size of the MRU list:
4303 * - if we just hit in the MRU ghost list, then increase
4304 * the target size of the MRU list.
4305 * - if we just hit in the MFU ghost list, then increase
4306 * the target size of the MFU list by decreasing the
4307 * target size of the MRU list.
4308 */
4309 if (state == arc_mru_ghost) {
4310 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4311 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4312
4313 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4314 } else if (state == arc_mfu_ghost) {
4315 uint64_t delta;
4316
4317 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4318 mult = MIN(mult, 10);
4319
4320 delta = MIN(bytes * mult, arc_p);
4321 arc_p = MAX(arc_p_min, arc_p - delta);
4322 }
4323 ASSERT((int64_t)arc_p >= 0);
4324
4325 if (arc_reclaim_needed()) {
4326 cv_signal(&arc_reclaim_thread_cv);
4327 return;
4328 }
4329
4330 if (arc_no_grow)
4331 return;
4332
4333 if (arc_c >= arc_c_max)
4334 return;
4335
4336 /*
4337 * If we're within (2 * maxblocksize) bytes of the target
4338 * cache size, increment the target cache size
4339 */
4340 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
4341 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
4342 atomic_add_64(&arc_c, (int64_t)bytes);
4343 if (arc_c > arc_c_max)
4344 arc_c = arc_c_max;
4345 else if (state == arc_anon)
4346 atomic_add_64(&arc_p, (int64_t)bytes);
4347 if (arc_p > arc_c)
4348 arc_p = arc_c;
4349 }
4350 ASSERT((int64_t)arc_p >= 0);
4351 }
4352
4353 /*
4354 * Check if arc_size has grown past our upper threshold, determined by
4355 * zfs_arc_overflow_shift.
4356 */
4357 static boolean_t
arc_is_overflowing(void)4358 arc_is_overflowing(void)
4359 {
4360 /* Always allow at least one block of overflow */
4361 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4362 arc_c >> zfs_arc_overflow_shift);
4363
4364 return (arc_size >= arc_c + overflow);
4365 }
4366
4367 /*
4368 * Allocate a block and return it to the caller. If we are hitting the
4369 * hard limit for the cache size, we must sleep, waiting for the eviction
4370 * thread to catch up. If we're past the target size but below the hard
4371 * limit, we'll only signal the reclaim thread and continue on.
4372 */
4373 static void *
arc_get_data_buf(arc_buf_hdr_t * hdr,uint64_t size,void * tag)4374 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4375 {
4376 void *datap = NULL;
4377 arc_state_t *state = hdr->b_l1hdr.b_state;
4378 arc_buf_contents_t type = arc_buf_type(hdr);
4379
4380 arc_adapt(size, state);
4381
4382 /*
4383 * If arc_size is currently overflowing, and has grown past our
4384 * upper limit, we must be adding data faster than the evict
4385 * thread can evict. Thus, to ensure we don't compound the
4386 * problem by adding more data and forcing arc_size to grow even
4387 * further past it's target size, we halt and wait for the
4388 * eviction thread to catch up.
4389 *
4390 * It's also possible that the reclaim thread is unable to evict
4391 * enough buffers to get arc_size below the overflow limit (e.g.
4392 * due to buffers being un-evictable, or hash lock collisions).
4393 * In this case, we want to proceed regardless if we're
4394 * overflowing; thus we don't use a while loop here.
4395 */
4396 if (arc_is_overflowing()) {
4397 mutex_enter(&arc_reclaim_lock);
4398
4399 /*
4400 * Now that we've acquired the lock, we may no longer be
4401 * over the overflow limit, lets check.
4402 *
4403 * We're ignoring the case of spurious wake ups. If that
4404 * were to happen, it'd let this thread consume an ARC
4405 * buffer before it should have (i.e. before we're under
4406 * the overflow limit and were signalled by the reclaim
4407 * thread). As long as that is a rare occurrence, it
4408 * shouldn't cause any harm.
4409 */
4410 if (arc_is_overflowing()) {
4411 cv_signal(&arc_reclaim_thread_cv);
4412 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4413 }
4414
4415 mutex_exit(&arc_reclaim_lock);
4416 }
4417
4418 VERIFY3U(hdr->b_type, ==, type);
4419 if (type == ARC_BUFC_METADATA) {
4420 datap = zio_buf_alloc(size);
4421 arc_space_consume(size, ARC_SPACE_META);
4422 } else {
4423 ASSERT(type == ARC_BUFC_DATA);
4424 datap = zio_data_buf_alloc(size);
4425 arc_space_consume(size, ARC_SPACE_DATA);
4426 }
4427
4428 /*
4429 * Update the state size. Note that ghost states have a
4430 * "ghost size" and so don't need to be updated.
4431 */
4432 if (!GHOST_STATE(state)) {
4433
4434 (void) refcount_add_many(&state->arcs_size, size, tag);
4435
4436 /*
4437 * If this is reached via arc_read, the link is
4438 * protected by the hash lock. If reached via
4439 * arc_buf_alloc, the header should not be accessed by
4440 * any other thread. And, if reached via arc_read_done,
4441 * the hash lock will protect it if it's found in the
4442 * hash table; otherwise no other thread should be
4443 * trying to [add|remove]_reference it.
4444 */
4445 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4446 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4447 (void) refcount_add_many(&state->arcs_esize[type],
4448 size, tag);
4449 }
4450
4451 /*
4452 * If we are growing the cache, and we are adding anonymous
4453 * data, and we have outgrown arc_p, update arc_p
4454 */
4455 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
4456 (refcount_count(&arc_anon->arcs_size) +
4457 refcount_count(&arc_mru->arcs_size) > arc_p))
4458 arc_p = MIN(arc_c, arc_p + size);
4459 }
4460 ARCSTAT_BUMP(arcstat_allocated);
4461 return (datap);
4462 }
4463
4464 /*
4465 * Free the arc data buffer.
4466 */
4467 static void
arc_free_data_buf(arc_buf_hdr_t * hdr,void * data,uint64_t size,void * tag)4468 arc_free_data_buf(arc_buf_hdr_t *hdr, void *data, uint64_t size, void *tag)
4469 {
4470 arc_state_t *state = hdr->b_l1hdr.b_state;
4471 arc_buf_contents_t type = arc_buf_type(hdr);
4472
4473 /* protected by hash lock, if in the hash table */
4474 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4475 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4476 ASSERT(state != arc_anon && state != arc_l2c_only);
4477
4478 (void) refcount_remove_many(&state->arcs_esize[type],
4479 size, tag);
4480 }
4481 (void) refcount_remove_many(&state->arcs_size, size, tag);
4482
4483 VERIFY3U(hdr->b_type, ==, type);
4484 if (type == ARC_BUFC_METADATA) {
4485 zio_buf_free(data, size);
4486 arc_space_return(size, ARC_SPACE_META);
4487 } else {
4488 ASSERT(type == ARC_BUFC_DATA);
4489 zio_data_buf_free(data, size);
4490 arc_space_return(size, ARC_SPACE_DATA);
4491 }
4492 }
4493
4494 /*
4495 * This routine is called whenever a buffer is accessed.
4496 * NOTE: the hash lock is dropped in this function.
4497 */
4498 static void
arc_access(arc_buf_hdr_t * hdr,kmutex_t * hash_lock)4499 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4500 {
4501 clock_t now;
4502
4503 ASSERT(MUTEX_HELD(hash_lock));
4504 ASSERT(HDR_HAS_L1HDR(hdr));
4505
4506 if (hdr->b_l1hdr.b_state == arc_anon) {
4507 /*
4508 * This buffer is not in the cache, and does not
4509 * appear in our "ghost" list. Add the new buffer
4510 * to the MRU state.
4511 */
4512
4513 ASSERT0(hdr->b_l1hdr.b_arc_access);
4514 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4515 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4516 arc_change_state(arc_mru, hdr, hash_lock);
4517
4518 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4519 now = ddi_get_lbolt();
4520
4521 /*
4522 * If this buffer is here because of a prefetch, then either:
4523 * - clear the flag if this is a "referencing" read
4524 * (any subsequent access will bump this into the MFU state).
4525 * or
4526 * - move the buffer to the head of the list if this is
4527 * another prefetch (to make it less likely to be evicted).
4528 */
4529 if (HDR_PREFETCH(hdr)) {
4530 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4531 /* link protected by hash lock */
4532 ASSERT(multilist_link_active(
4533 &hdr->b_l1hdr.b_arc_node));
4534 } else {
4535 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4536 ARCSTAT_BUMP(arcstat_mru_hits);
4537 }
4538 hdr->b_l1hdr.b_arc_access = now;
4539 return;
4540 }
4541
4542 /*
4543 * This buffer has been "accessed" only once so far,
4544 * but it is still in the cache. Move it to the MFU
4545 * state.
4546 */
4547 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4548 /*
4549 * More than 125ms have passed since we
4550 * instantiated this buffer. Move it to the
4551 * most frequently used state.
4552 */
4553 hdr->b_l1hdr.b_arc_access = now;
4554 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4555 arc_change_state(arc_mfu, hdr, hash_lock);
4556 }
4557 ARCSTAT_BUMP(arcstat_mru_hits);
4558 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4559 arc_state_t *new_state;
4560 /*
4561 * This buffer has been "accessed" recently, but
4562 * was evicted from the cache. Move it to the
4563 * MFU state.
4564 */
4565
4566 if (HDR_PREFETCH(hdr)) {
4567 new_state = arc_mru;
4568 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4569 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4570 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4571 } else {
4572 new_state = arc_mfu;
4573 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4574 }
4575
4576 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4577 arc_change_state(new_state, hdr, hash_lock);
4578
4579 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4580 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
4581 /*
4582 * This buffer has been accessed more than once and is
4583 * still in the cache. Keep it in the MFU state.
4584 *
4585 * NOTE: an add_reference() that occurred when we did
4586 * the arc_read() will have kicked this off the list.
4587 * If it was a prefetch, we will explicitly move it to
4588 * the head of the list now.
4589 */
4590 if ((HDR_PREFETCH(hdr)) != 0) {
4591 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4592 /* link protected by hash_lock */
4593 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4594 }
4595 ARCSTAT_BUMP(arcstat_mfu_hits);
4596 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4597 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4598 arc_state_t *new_state = arc_mfu;
4599 /*
4600 * This buffer has been accessed more than once but has
4601 * been evicted from the cache. Move it back to the
4602 * MFU state.
4603 */
4604
4605 if (HDR_PREFETCH(hdr)) {
4606 /*
4607 * This is a prefetch access...
4608 * move this block back to the MRU state.
4609 */
4610 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4611 new_state = arc_mru;
4612 }
4613
4614 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4615 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4616 arc_change_state(new_state, hdr, hash_lock);
4617
4618 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4619 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4620 /*
4621 * This buffer is on the 2nd Level ARC.
4622 */
4623
4624 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4625 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4626 arc_change_state(arc_mfu, hdr, hash_lock);
4627 } else {
4628 ASSERT(!"invalid arc state");
4629 }
4630 }
4631
4632 /* a generic arc_done_func_t which you can use */
4633 /* ARGSUSED */
4634 void
arc_bcopy_func(zio_t * zio,arc_buf_t * buf,void * arg)4635 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4636 {
4637 if (zio == NULL || zio->io_error == 0)
4638 bcopy(buf->b_data, arg, HDR_GET_LSIZE(buf->b_hdr));
4639 arc_buf_destroy(buf, arg);
4640 }
4641
4642 /* a generic arc_done_func_t */
4643 void
arc_getbuf_func(zio_t * zio,arc_buf_t * buf,void * arg)4644 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4645 {
4646 arc_buf_t **bufp = arg;
4647 if (zio && zio->io_error) {
4648 arc_buf_destroy(buf, arg);
4649 *bufp = NULL;
4650 } else {
4651 *bufp = buf;
4652 ASSERT(buf->b_data);
4653 }
4654 }
4655
4656 static void
arc_hdr_verify(arc_buf_hdr_t * hdr,blkptr_t * bp)4657 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
4658 {
4659 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
4660 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
4661 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
4662 } else {
4663 if (HDR_COMPRESSION_ENABLED(hdr)) {
4664 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
4665 BP_GET_COMPRESS(bp));
4666 }
4667 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
4668 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
4669 }
4670 }
4671
4672 static void
arc_read_done(zio_t * zio)4673 arc_read_done(zio_t *zio)
4674 {
4675 arc_buf_hdr_t *hdr = zio->io_private;
4676 arc_buf_t *abuf = NULL; /* buffer we're assigning to callback */
4677 kmutex_t *hash_lock = NULL;
4678 arc_callback_t *callback_list, *acb;
4679 int freeable = B_FALSE;
4680
4681 /*
4682 * The hdr was inserted into hash-table and removed from lists
4683 * prior to starting I/O. We should find this header, since
4684 * it's in the hash table, and it should be legit since it's
4685 * not possible to evict it during the I/O. The only possible
4686 * reason for it not to be found is if we were freed during the
4687 * read.
4688 */
4689 if (HDR_IN_HASH_TABLE(hdr)) {
4690 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
4691 ASSERT3U(hdr->b_dva.dva_word[0], ==,
4692 BP_IDENTITY(zio->io_bp)->dva_word[0]);
4693 ASSERT3U(hdr->b_dva.dva_word[1], ==,
4694 BP_IDENTITY(zio->io_bp)->dva_word[1]);
4695
4696 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
4697 &hash_lock);
4698
4699 ASSERT((found == hdr &&
4700 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
4701 (found == hdr && HDR_L2_READING(hdr)));
4702 ASSERT3P(hash_lock, !=, NULL);
4703 }
4704
4705 if (zio->io_error == 0) {
4706 /* byteswap if necessary */
4707 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
4708 if (BP_GET_LEVEL(zio->io_bp) > 0) {
4709 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
4710 } else {
4711 hdr->b_l1hdr.b_byteswap =
4712 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
4713 }
4714 } else {
4715 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
4716 }
4717 }
4718
4719 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
4720 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
4721 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
4722
4723 callback_list = hdr->b_l1hdr.b_acb;
4724 ASSERT3P(callback_list, !=, NULL);
4725
4726 if (hash_lock && zio->io_error == 0 &&
4727 hdr->b_l1hdr.b_state == arc_anon) {
4728 /*
4729 * Only call arc_access on anonymous buffers. This is because
4730 * if we've issued an I/O for an evicted buffer, we've already
4731 * called arc_access (to prevent any simultaneous readers from
4732 * getting confused).
4733 */
4734 arc_access(hdr, hash_lock);
4735 }
4736
4737 /* create copies of the data buffer for the callers */
4738 for (acb = callback_list; acb; acb = acb->acb_next) {
4739 if (acb->acb_done != NULL) {
4740 /*
4741 * If we're here, then this must be a demand read
4742 * since prefetch requests don't have callbacks.
4743 * If a read request has a callback (i.e. acb_done is
4744 * not NULL), then we decompress the data for the
4745 * first request and clone the rest. This avoids
4746 * having to waste cpu resources decompressing data
4747 * that nobody is explicitly waiting to read.
4748 */
4749 if (abuf == NULL) {
4750 acb->acb_buf = arc_buf_alloc_impl(hdr,
4751 acb->acb_private);
4752 if (zio->io_error == 0) {
4753 zio->io_error =
4754 arc_decompress(acb->acb_buf);
4755 }
4756 abuf = acb->acb_buf;
4757 } else {
4758 add_reference(hdr, acb->acb_private);
4759 acb->acb_buf = arc_buf_clone(abuf);
4760 }
4761 }
4762 }
4763 hdr->b_l1hdr.b_acb = NULL;
4764 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
4765 if (abuf == NULL) {
4766 /*
4767 * This buffer didn't have a callback so it must
4768 * be a prefetch.
4769 */
4770 ASSERT(HDR_PREFETCH(hdr));
4771 ASSERT0(hdr->b_l1hdr.b_bufcnt);
4772 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
4773 }
4774
4775 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4776 callback_list != NULL);
4777
4778 if (zio->io_error == 0) {
4779 arc_hdr_verify(hdr, zio->io_bp);
4780 } else {
4781 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
4782 if (hdr->b_l1hdr.b_state != arc_anon)
4783 arc_change_state(arc_anon, hdr, hash_lock);
4784 if (HDR_IN_HASH_TABLE(hdr))
4785 buf_hash_remove(hdr);
4786 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4787 }
4788
4789 /*
4790 * Broadcast before we drop the hash_lock to avoid the possibility
4791 * that the hdr (and hence the cv) might be freed before we get to
4792 * the cv_broadcast().
4793 */
4794 cv_broadcast(&hdr->b_l1hdr.b_cv);
4795
4796 if (hash_lock != NULL) {
4797 mutex_exit(hash_lock);
4798 } else {
4799 /*
4800 * This block was freed while we waited for the read to
4801 * complete. It has been removed from the hash table and
4802 * moved to the anonymous state (so that it won't show up
4803 * in the cache).
4804 */
4805 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4806 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4807 }
4808
4809 /* execute each callback and free its structure */
4810 while ((acb = callback_list) != NULL) {
4811 if (acb->acb_done)
4812 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4813
4814 if (acb->acb_zio_dummy != NULL) {
4815 acb->acb_zio_dummy->io_error = zio->io_error;
4816 zio_nowait(acb->acb_zio_dummy);
4817 }
4818
4819 callback_list = acb->acb_next;
4820 kmem_free(acb, sizeof (arc_callback_t));
4821 }
4822
4823 if (freeable)
4824 arc_hdr_destroy(hdr);
4825 }
4826
4827 /*
4828 * "Read" the block at the specified DVA (in bp) via the
4829 * cache. If the block is found in the cache, invoke the provided
4830 * callback immediately and return. Note that the `zio' parameter
4831 * in the callback will be NULL in this case, since no IO was
4832 * required. If the block is not in the cache pass the read request
4833 * on to the spa with a substitute callback function, so that the
4834 * requested block will be added to the cache.
4835 *
4836 * If a read request arrives for a block that has a read in-progress,
4837 * either wait for the in-progress read to complete (and return the
4838 * results); or, if this is a read with a "done" func, add a record
4839 * to the read to invoke the "done" func when the read completes,
4840 * and return; or just return.
4841 *
4842 * arc_read_done() will invoke all the requested "done" functions
4843 * for readers of this block.
4844 */
4845 int
arc_read(zio_t * pio,spa_t * spa,const blkptr_t * bp,arc_done_func_t * done,void * private,zio_priority_t priority,int zio_flags,arc_flags_t * arc_flags,const zbookmark_phys_t * zb)4846 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
4847 void *private, zio_priority_t priority, int zio_flags,
4848 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
4849 {
4850 arc_buf_hdr_t *hdr = NULL;
4851 kmutex_t *hash_lock = NULL;
4852 zio_t *rzio;
4853 uint64_t guid = spa_load_guid(spa);
4854
4855 ASSERT(!BP_IS_EMBEDDED(bp) ||
4856 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
4857
4858 top:
4859 if (!BP_IS_EMBEDDED(bp)) {
4860 /*
4861 * Embedded BP's have no DVA and require no I/O to "read".
4862 * Create an anonymous arc buf to back it.
4863 */
4864 hdr = buf_hash_find(guid, bp, &hash_lock);
4865 }
4866
4867 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pdata != NULL) {
4868 arc_buf_t *buf = NULL;
4869 *arc_flags |= ARC_FLAG_CACHED;
4870
4871 if (HDR_IO_IN_PROGRESS(hdr)) {
4872
4873 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
4874 priority == ZIO_PRIORITY_SYNC_READ) {
4875 /*
4876 * This sync read must wait for an
4877 * in-progress async read (e.g. a predictive
4878 * prefetch). Async reads are queued
4879 * separately at the vdev_queue layer, so
4880 * this is a form of priority inversion.
4881 * Ideally, we would "inherit" the demand
4882 * i/o's priority by moving the i/o from
4883 * the async queue to the synchronous queue,
4884 * but there is currently no mechanism to do
4885 * so. Track this so that we can evaluate
4886 * the magnitude of this potential performance
4887 * problem.
4888 *
4889 * Note that if the prefetch i/o is already
4890 * active (has been issued to the device),
4891 * the prefetch improved performance, because
4892 * we issued it sooner than we would have
4893 * without the prefetch.
4894 */
4895 DTRACE_PROBE1(arc__sync__wait__for__async,
4896 arc_buf_hdr_t *, hdr);
4897 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
4898 }
4899 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4900 arc_hdr_clear_flags(hdr,
4901 ARC_FLAG_PREDICTIVE_PREFETCH);
4902 }
4903
4904 if (*arc_flags & ARC_FLAG_WAIT) {
4905 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4906 mutex_exit(hash_lock);
4907 goto top;
4908 }
4909 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4910
4911 if (done) {
4912 arc_callback_t *acb = NULL;
4913
4914 acb = kmem_zalloc(sizeof (arc_callback_t),
4915 KM_SLEEP);
4916 acb->acb_done = done;
4917 acb->acb_private = private;
4918 if (pio != NULL)
4919 acb->acb_zio_dummy = zio_null(pio,
4920 spa, NULL, NULL, NULL, zio_flags);
4921
4922 ASSERT3P(acb->acb_done, !=, NULL);
4923 acb->acb_next = hdr->b_l1hdr.b_acb;
4924 hdr->b_l1hdr.b_acb = acb;
4925 mutex_exit(hash_lock);
4926 return (0);
4927 }
4928 mutex_exit(hash_lock);
4929 return (0);
4930 }
4931
4932 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4933 hdr->b_l1hdr.b_state == arc_mfu);
4934
4935 if (done) {
4936 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4937 /*
4938 * This is a demand read which does not have to
4939 * wait for i/o because we did a predictive
4940 * prefetch i/o for it, which has completed.
4941 */
4942 DTRACE_PROBE1(
4943 arc__demand__hit__predictive__prefetch,
4944 arc_buf_hdr_t *, hdr);
4945 ARCSTAT_BUMP(
4946 arcstat_demand_hit_predictive_prefetch);
4947 arc_hdr_clear_flags(hdr,
4948 ARC_FLAG_PREDICTIVE_PREFETCH);
4949 }
4950 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
4951
4952 /*
4953 * If this block is already in use, create a new
4954 * copy of the data so that we will be guaranteed
4955 * that arc_release() will always succeed.
4956 */
4957 buf = hdr->b_l1hdr.b_buf;
4958 if (buf == NULL) {
4959 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4960 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
4961 buf = arc_buf_alloc_impl(hdr, private);
4962 VERIFY0(arc_decompress(buf));
4963 } else {
4964 add_reference(hdr, private);
4965 buf = arc_buf_clone(buf);
4966 }
4967 ASSERT3P(buf->b_data, !=, NULL);
4968
4969 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4970 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4971 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
4972 }
4973 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4974 arc_access(hdr, hash_lock);
4975 if (*arc_flags & ARC_FLAG_L2CACHE)
4976 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
4977 mutex_exit(hash_lock);
4978 ARCSTAT_BUMP(arcstat_hits);
4979 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4980 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4981 data, metadata, hits);
4982
4983 if (done)
4984 done(NULL, buf, private);
4985 } else {
4986 uint64_t lsize = BP_GET_LSIZE(bp);
4987 uint64_t psize = BP_GET_PSIZE(bp);
4988 arc_callback_t *acb;
4989 vdev_t *vd = NULL;
4990 uint64_t addr = 0;
4991 boolean_t devw = B_FALSE;
4992 uint64_t size;
4993
4994 if (hdr == NULL) {
4995 /* this block is not in the cache */
4996 arc_buf_hdr_t *exists = NULL;
4997 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4998 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
4999 BP_GET_COMPRESS(bp), type);
5000
5001 if (!BP_IS_EMBEDDED(bp)) {
5002 hdr->b_dva = *BP_IDENTITY(bp);
5003 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5004 exists = buf_hash_insert(hdr, &hash_lock);
5005 }
5006 if (exists != NULL) {
5007 /* somebody beat us to the hash insert */
5008 mutex_exit(hash_lock);
5009 buf_discard_identity(hdr);
5010 arc_hdr_destroy(hdr);
5011 goto top; /* restart the IO request */
5012 }
5013 } else {
5014 /*
5015 * This block is in the ghost cache. If it was L2-only
5016 * (and thus didn't have an L1 hdr), we realloc the
5017 * header to add an L1 hdr.
5018 */
5019 if (!HDR_HAS_L1HDR(hdr)) {
5020 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5021 hdr_full_cache);
5022 }
5023 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
5024 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5025 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5026 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5027 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5028
5029 /*
5030 * This is a delicate dance that we play here.
5031 * This hdr is in the ghost list so we access it
5032 * to move it out of the ghost list before we
5033 * initiate the read. If it's a prefetch then
5034 * it won't have a callback so we'll remove the
5035 * reference that arc_buf_alloc_impl() created. We
5036 * do this after we've called arc_access() to
5037 * avoid hitting an assert in remove_reference().
5038 */
5039 arc_access(hdr, hash_lock);
5040 arc_hdr_alloc_pdata(hdr);
5041 }
5042 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
5043 size = arc_hdr_size(hdr);
5044
5045 /*
5046 * If compression is enabled on the hdr, then will do
5047 * RAW I/O and will store the compressed data in the hdr's
5048 * data block. Otherwise, the hdr's data block will contain
5049 * the uncompressed data.
5050 */
5051 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5052 zio_flags |= ZIO_FLAG_RAW;
5053 }
5054
5055 if (*arc_flags & ARC_FLAG_PREFETCH)
5056 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5057 if (*arc_flags & ARC_FLAG_L2CACHE)
5058 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5059 if (BP_GET_LEVEL(bp) > 0)
5060 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5061 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5062 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5063 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5064
5065 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5066 acb->acb_done = done;
5067 acb->acb_private = private;
5068
5069 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5070 hdr->b_l1hdr.b_acb = acb;
5071 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5072
5073 if (HDR_HAS_L2HDR(hdr) &&
5074 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5075 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5076 addr = hdr->b_l2hdr.b_daddr;
5077 /*
5078 * Lock out device removal.
5079 */
5080 if (vdev_is_dead(vd) ||
5081 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5082 vd = NULL;
5083 }
5084
5085 if (priority == ZIO_PRIORITY_ASYNC_READ)
5086 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5087 else
5088 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5089
5090 if (hash_lock != NULL)
5091 mutex_exit(hash_lock);
5092
5093 /*
5094 * At this point, we have a level 1 cache miss. Try again in
5095 * L2ARC if possible.
5096 */
5097 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5098
5099 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5100 uint64_t, lsize, zbookmark_phys_t *, zb);
5101 ARCSTAT_BUMP(arcstat_misses);
5102 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5103 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5104 data, metadata, misses);
5105 #ifdef __FreeBSD__
5106 #ifdef _KERNEL
5107 #ifdef RACCT
5108 if (racct_enable) {
5109 PROC_LOCK(curproc);
5110 racct_add_force(curproc, RACCT_READBPS, size);
5111 racct_add_force(curproc, RACCT_READIOPS, 1);
5112 PROC_UNLOCK(curproc);
5113 }
5114 #endif /* RACCT */
5115 curthread->td_ru.ru_inblock++;
5116 #endif
5117 #endif
5118
5119 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5120 /*
5121 * Read from the L2ARC if the following are true:
5122 * 1. The L2ARC vdev was previously cached.
5123 * 2. This buffer still has L2ARC metadata.
5124 * 3. This buffer isn't currently writing to the L2ARC.
5125 * 4. The L2ARC entry wasn't evicted, which may
5126 * also have invalidated the vdev.
5127 * 5. This isn't prefetch and l2arc_noprefetch is set.
5128 */
5129 if (HDR_HAS_L2HDR(hdr) &&
5130 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5131 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5132 l2arc_read_callback_t *cb;
5133 void* b_data;
5134
5135 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5136 ARCSTAT_BUMP(arcstat_l2_hits);
5137
5138 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5139 KM_SLEEP);
5140 cb->l2rcb_hdr = hdr;
5141 cb->l2rcb_bp = *bp;
5142 cb->l2rcb_zb = *zb;
5143 cb->l2rcb_flags = zio_flags;
5144 uint64_t asize = vdev_psize_to_asize(vd, size);
5145 if (asize != size) {
5146 b_data = zio_data_buf_alloc(asize);
5147 cb->l2rcb_data = b_data;
5148 } else {
5149 b_data = hdr->b_l1hdr.b_pdata;
5150 }
5151
5152 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5153 addr + asize < vd->vdev_psize -
5154 VDEV_LABEL_END_SIZE);
5155
5156 /*
5157 * l2arc read. The SCL_L2ARC lock will be
5158 * released by l2arc_read_done().
5159 * Issue a null zio if the underlying buffer
5160 * was squashed to zero size by compression.
5161 */
5162 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5163 ZIO_COMPRESS_EMPTY);
5164 rzio = zio_read_phys(pio, vd, addr,
5165 asize, b_data,
5166 ZIO_CHECKSUM_OFF,
5167 l2arc_read_done, cb, priority,
5168 zio_flags | ZIO_FLAG_DONT_CACHE |
5169 ZIO_FLAG_CANFAIL |
5170 ZIO_FLAG_DONT_PROPAGATE |
5171 ZIO_FLAG_DONT_RETRY, B_FALSE);
5172 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5173 zio_t *, rzio);
5174 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5175
5176 if (*arc_flags & ARC_FLAG_NOWAIT) {
5177 zio_nowait(rzio);
5178 return (0);
5179 }
5180
5181 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5182 if (zio_wait(rzio) == 0)
5183 return (0);
5184
5185 /* l2arc read error; goto zio_read() */
5186 } else {
5187 DTRACE_PROBE1(l2arc__miss,
5188 arc_buf_hdr_t *, hdr);
5189 ARCSTAT_BUMP(arcstat_l2_misses);
5190 if (HDR_L2_WRITING(hdr))
5191 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5192 spa_config_exit(spa, SCL_L2ARC, vd);
5193 }
5194 } else {
5195 if (vd != NULL)
5196 spa_config_exit(spa, SCL_L2ARC, vd);
5197 if (l2arc_ndev != 0) {
5198 DTRACE_PROBE1(l2arc__miss,
5199 arc_buf_hdr_t *, hdr);
5200 ARCSTAT_BUMP(arcstat_l2_misses);
5201 }
5202 }
5203
5204 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pdata, size,
5205 arc_read_done, hdr, priority, zio_flags, zb);
5206
5207 if (*arc_flags & ARC_FLAG_WAIT)
5208 return (zio_wait(rzio));
5209
5210 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5211 zio_nowait(rzio);
5212 }
5213 return (0);
5214 }
5215
5216 /*
5217 * Notify the arc that a block was freed, and thus will never be used again.
5218 */
5219 void
arc_freed(spa_t * spa,const blkptr_t * bp)5220 arc_freed(spa_t *spa, const blkptr_t *bp)
5221 {
5222 arc_buf_hdr_t *hdr;
5223 kmutex_t *hash_lock;
5224 uint64_t guid = spa_load_guid(spa);
5225
5226 ASSERT(!BP_IS_EMBEDDED(bp));
5227
5228 hdr = buf_hash_find(guid, bp, &hash_lock);
5229 if (hdr == NULL)
5230 return;
5231
5232 /*
5233 * We might be trying to free a block that is still doing I/O
5234 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5235 * dmu_sync-ed block). If this block is being prefetched, then it
5236 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5237 * until the I/O completes. A block may also have a reference if it is
5238 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5239 * have written the new block to its final resting place on disk but
5240 * without the dedup flag set. This would have left the hdr in the MRU
5241 * state and discoverable. When the txg finally syncs it detects that
5242 * the block was overridden in open context and issues an override I/O.
5243 * Since this is a dedup block, the override I/O will determine if the
5244 * block is already in the DDT. If so, then it will replace the io_bp
5245 * with the bp from the DDT and allow the I/O to finish. When the I/O
5246 * reaches the done callback, dbuf_write_override_done, it will
5247 * check to see if the io_bp and io_bp_override are identical.
5248 * If they are not, then it indicates that the bp was replaced with
5249 * the bp in the DDT and the override bp is freed. This allows
5250 * us to arrive here with a reference on a block that is being
5251 * freed. So if we have an I/O in progress, or a reference to
5252 * this hdr, then we don't destroy the hdr.
5253 */
5254 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5255 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5256 arc_change_state(arc_anon, hdr, hash_lock);
5257 arc_hdr_destroy(hdr);
5258 mutex_exit(hash_lock);
5259 } else {
5260 mutex_exit(hash_lock);
5261 }
5262
5263 }
5264
5265 /*
5266 * Release this buffer from the cache, making it an anonymous buffer. This
5267 * must be done after a read and prior to modifying the buffer contents.
5268 * If the buffer has more than one reference, we must make
5269 * a new hdr for the buffer.
5270 */
5271 void
arc_release(arc_buf_t * buf,void * tag)5272 arc_release(arc_buf_t *buf, void *tag)
5273 {
5274 arc_buf_hdr_t *hdr = buf->b_hdr;
5275
5276 /*
5277 * It would be nice to assert that if it's DMU metadata (level >
5278 * 0 || it's the dnode file), then it must be syncing context.
5279 * But we don't know that information at this level.
5280 */
5281
5282 mutex_enter(&buf->b_evict_lock);
5283
5284 ASSERT(HDR_HAS_L1HDR(hdr));
5285
5286 /*
5287 * We don't grab the hash lock prior to this check, because if
5288 * the buffer's header is in the arc_anon state, it won't be
5289 * linked into the hash table.
5290 */
5291 if (hdr->b_l1hdr.b_state == arc_anon) {
5292 mutex_exit(&buf->b_evict_lock);
5293 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5294 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5295 ASSERT(!HDR_HAS_L2HDR(hdr));
5296 ASSERT(HDR_EMPTY(hdr));
5297 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5298 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5299 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5300
5301 hdr->b_l1hdr.b_arc_access = 0;
5302
5303 /*
5304 * If the buf is being overridden then it may already
5305 * have a hdr that is not empty.
5306 */
5307 buf_discard_identity(hdr);
5308 arc_buf_thaw(buf);
5309
5310 return;
5311 }
5312
5313 kmutex_t *hash_lock = HDR_LOCK(hdr);
5314 mutex_enter(hash_lock);
5315
5316 /*
5317 * This assignment is only valid as long as the hash_lock is
5318 * held, we must be careful not to reference state or the
5319 * b_state field after dropping the lock.
5320 */
5321 arc_state_t *state = hdr->b_l1hdr.b_state;
5322 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5323 ASSERT3P(state, !=, arc_anon);
5324
5325 /* this buffer is not on any list */
5326 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
5327
5328 if (HDR_HAS_L2HDR(hdr)) {
5329 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5330
5331 /*
5332 * We have to recheck this conditional again now that
5333 * we're holding the l2ad_mtx to prevent a race with
5334 * another thread which might be concurrently calling
5335 * l2arc_evict(). In that case, l2arc_evict() might have
5336 * destroyed the header's L2 portion as we were waiting
5337 * to acquire the l2ad_mtx.
5338 */
5339 if (HDR_HAS_L2HDR(hdr)) {
5340 l2arc_trim(hdr);
5341 arc_hdr_l2hdr_destroy(hdr);
5342 }
5343
5344 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5345 }
5346
5347 /*
5348 * Do we have more than one buf?
5349 */
5350 if (hdr->b_l1hdr.b_bufcnt > 1) {
5351 arc_buf_hdr_t *nhdr;
5352 arc_buf_t **bufp;
5353 uint64_t spa = hdr->b_spa;
5354 uint64_t psize = HDR_GET_PSIZE(hdr);
5355 uint64_t lsize = HDR_GET_LSIZE(hdr);
5356 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5357 arc_buf_contents_t type = arc_buf_type(hdr);
5358 VERIFY3U(hdr->b_type, ==, type);
5359
5360 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5361 (void) remove_reference(hdr, hash_lock, tag);
5362
5363 if (arc_buf_is_shared(buf)) {
5364 ASSERT(HDR_SHARED_DATA(hdr));
5365 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5366 ASSERT(ARC_BUF_LAST(buf));
5367 }
5368
5369 /*
5370 * Pull the data off of this hdr and attach it to
5371 * a new anonymous hdr. Also find the last buffer
5372 * in the hdr's buffer list.
5373 */
5374 arc_buf_t *lastbuf = NULL;
5375 bufp = &hdr->b_l1hdr.b_buf;
5376 while (*bufp != NULL) {
5377 if (*bufp == buf) {
5378 *bufp = buf->b_next;
5379 }
5380
5381 /*
5382 * If we've removed a buffer in the middle of
5383 * the list then update the lastbuf and update
5384 * bufp.
5385 */
5386 if (*bufp != NULL) {
5387 lastbuf = *bufp;
5388 bufp = &(*bufp)->b_next;
5389 }
5390 }
5391 buf->b_next = NULL;
5392 ASSERT3P(lastbuf, !=, buf);
5393 ASSERT3P(lastbuf, !=, NULL);
5394
5395 /*
5396 * If the current arc_buf_t and the hdr are sharing their data
5397 * buffer, then we must stop sharing that block, transfer
5398 * ownership and setup sharing with a new arc_buf_t at the end
5399 * of the hdr's b_buf list.
5400 */
5401 if (arc_buf_is_shared(buf)) {
5402 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5403 ASSERT(ARC_BUF_LAST(lastbuf));
5404 VERIFY(!arc_buf_is_shared(lastbuf));
5405
5406 /*
5407 * First, sever the block sharing relationship between
5408 * buf and the arc_buf_hdr_t. Then, setup a new
5409 * block sharing relationship with the last buffer
5410 * on the arc_buf_t list.
5411 */
5412 arc_unshare_buf(hdr, buf);
5413 arc_share_buf(hdr, lastbuf);
5414 VERIFY3P(lastbuf->b_data, !=, NULL);
5415 } else if (HDR_SHARED_DATA(hdr)) {
5416 ASSERT(arc_buf_is_shared(lastbuf));
5417 }
5418 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
5419 ASSERT3P(state, !=, arc_l2c_only);
5420
5421 (void) refcount_remove_many(&state->arcs_size,
5422 HDR_GET_LSIZE(hdr), buf);
5423
5424 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5425 ASSERT3P(state, !=, arc_l2c_only);
5426 (void) refcount_remove_many(&state->arcs_esize[type],
5427 HDR_GET_LSIZE(hdr), buf);
5428 }
5429
5430 hdr->b_l1hdr.b_bufcnt -= 1;
5431 arc_cksum_verify(buf);
5432 #ifdef illumos
5433 arc_buf_unwatch(buf);
5434 #endif
5435
5436 mutex_exit(hash_lock);
5437
5438 /*
5439 * Allocate a new hdr. The new hdr will contain a b_pdata
5440 * buffer which will be freed in arc_write().
5441 */
5442 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5443 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5444 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5445 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5446 VERIFY3U(nhdr->b_type, ==, type);
5447 ASSERT(!HDR_SHARED_DATA(nhdr));
5448
5449 nhdr->b_l1hdr.b_buf = buf;
5450 nhdr->b_l1hdr.b_bufcnt = 1;
5451 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5452 buf->b_hdr = nhdr;
5453
5454 mutex_exit(&buf->b_evict_lock);
5455 (void) refcount_add_many(&arc_anon->arcs_size,
5456 HDR_GET_LSIZE(nhdr), buf);
5457 } else {
5458 mutex_exit(&buf->b_evict_lock);
5459 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5460 /* protected by hash lock, or hdr is on arc_anon */
5461 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5462 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5463 arc_change_state(arc_anon, hdr, hash_lock);
5464 hdr->b_l1hdr.b_arc_access = 0;
5465 mutex_exit(hash_lock);
5466
5467 buf_discard_identity(hdr);
5468 arc_buf_thaw(buf);
5469 }
5470 }
5471
5472 int
arc_released(arc_buf_t * buf)5473 arc_released(arc_buf_t *buf)
5474 {
5475 int released;
5476
5477 mutex_enter(&buf->b_evict_lock);
5478 released = (buf->b_data != NULL &&
5479 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5480 mutex_exit(&buf->b_evict_lock);
5481 return (released);
5482 }
5483
5484 #ifdef ZFS_DEBUG
5485 int
arc_referenced(arc_buf_t * buf)5486 arc_referenced(arc_buf_t *buf)
5487 {
5488 int referenced;
5489
5490 mutex_enter(&buf->b_evict_lock);
5491 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5492 mutex_exit(&buf->b_evict_lock);
5493 return (referenced);
5494 }
5495 #endif
5496
5497 static void
arc_write_ready(zio_t * zio)5498 arc_write_ready(zio_t *zio)
5499 {
5500 arc_write_callback_t *callback = zio->io_private;
5501 arc_buf_t *buf = callback->awcb_buf;
5502 arc_buf_hdr_t *hdr = buf->b_hdr;
5503 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5504
5505 ASSERT(HDR_HAS_L1HDR(hdr));
5506 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5507 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5508
5509 /*
5510 * If we're reexecuting this zio because the pool suspended, then
5511 * cleanup any state that was previously set the first time the
5512 * callback as invoked.
5513 */
5514 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5515 arc_cksum_free(hdr);
5516 #ifdef illumos
5517 arc_buf_unwatch(buf);
5518 #endif
5519 if (hdr->b_l1hdr.b_pdata != NULL) {
5520 if (arc_buf_is_shared(buf)) {
5521 ASSERT(HDR_SHARED_DATA(hdr));
5522
5523 arc_unshare_buf(hdr, buf);
5524 } else {
5525 arc_hdr_free_pdata(hdr);
5526 }
5527 }
5528 }
5529 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
5530 ASSERT(!HDR_SHARED_DATA(hdr));
5531 ASSERT(!arc_buf_is_shared(buf));
5532
5533 callback->awcb_ready(zio, buf, callback->awcb_private);
5534
5535 if (HDR_IO_IN_PROGRESS(hdr))
5536 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
5537
5538 arc_cksum_compute(buf);
5539 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5540
5541 enum zio_compress compress;
5542 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5543 compress = ZIO_COMPRESS_OFF;
5544 } else {
5545 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
5546 compress = BP_GET_COMPRESS(zio->io_bp);
5547 }
5548 HDR_SET_PSIZE(hdr, psize);
5549 arc_hdr_set_compress(hdr, compress);
5550
5551 /*
5552 * If the hdr is compressed, then copy the compressed
5553 * zio contents into arc_buf_hdr_t. Otherwise, copy the original
5554 * data buf into the hdr. Ideally, we would like to always copy the
5555 * io_data into b_pdata but the user may have disabled compressed
5556 * arc thus the on-disk block may or may not match what we maintain
5557 * in the hdr's b_pdata field.
5558 */
5559 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5560 ASSERT(BP_GET_COMPRESS(zio->io_bp) != ZIO_COMPRESS_OFF);
5561 ASSERT3U(psize, >, 0);
5562 arc_hdr_alloc_pdata(hdr);
5563 bcopy(zio->io_data, hdr->b_l1hdr.b_pdata, psize);
5564 } else {
5565 ASSERT3P(buf->b_data, ==, zio->io_orig_data);
5566 ASSERT3U(zio->io_orig_size, ==, HDR_GET_LSIZE(hdr));
5567 ASSERT3U(hdr->b_l1hdr.b_byteswap, ==, DMU_BSWAP_NUMFUNCS);
5568 ASSERT(!HDR_SHARED_DATA(hdr));
5569 ASSERT(!arc_buf_is_shared(buf));
5570 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5571 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
5572
5573 /*
5574 * This hdr is not compressed so we're able to share
5575 * the arc_buf_t data buffer with the hdr.
5576 */
5577 arc_share_buf(hdr, buf);
5578 VERIFY0(bcmp(zio->io_orig_data, hdr->b_l1hdr.b_pdata,
5579 HDR_GET_LSIZE(hdr)));
5580 }
5581 arc_hdr_verify(hdr, zio->io_bp);
5582 }
5583
5584 static void
arc_write_children_ready(zio_t * zio)5585 arc_write_children_ready(zio_t *zio)
5586 {
5587 arc_write_callback_t *callback = zio->io_private;
5588 arc_buf_t *buf = callback->awcb_buf;
5589
5590 callback->awcb_children_ready(zio, buf, callback->awcb_private);
5591 }
5592
5593 /*
5594 * The SPA calls this callback for each physical write that happens on behalf
5595 * of a logical write. See the comment in dbuf_write_physdone() for details.
5596 */
5597 static void
arc_write_physdone(zio_t * zio)5598 arc_write_physdone(zio_t *zio)
5599 {
5600 arc_write_callback_t *cb = zio->io_private;
5601 if (cb->awcb_physdone != NULL)
5602 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5603 }
5604
5605 static void
arc_write_done(zio_t * zio)5606 arc_write_done(zio_t *zio)
5607 {
5608 arc_write_callback_t *callback = zio->io_private;
5609 arc_buf_t *buf = callback->awcb_buf;
5610 arc_buf_hdr_t *hdr = buf->b_hdr;
5611
5612 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5613
5614 if (zio->io_error == 0) {
5615 arc_hdr_verify(hdr, zio->io_bp);
5616
5617 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5618 buf_discard_identity(hdr);
5619 } else {
5620 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5621 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5622 }
5623 } else {
5624 ASSERT(HDR_EMPTY(hdr));
5625 }
5626
5627 /*
5628 * If the block to be written was all-zero or compressed enough to be
5629 * embedded in the BP, no write was performed so there will be no
5630 * dva/birth/checksum. The buffer must therefore remain anonymous
5631 * (and uncached).
5632 */
5633 if (!HDR_EMPTY(hdr)) {
5634 arc_buf_hdr_t *exists;
5635 kmutex_t *hash_lock;
5636
5637 ASSERT(zio->io_error == 0);
5638
5639 arc_cksum_verify(buf);
5640
5641 exists = buf_hash_insert(hdr, &hash_lock);
5642 if (exists != NULL) {
5643 /*
5644 * This can only happen if we overwrite for
5645 * sync-to-convergence, because we remove
5646 * buffers from the hash table when we arc_free().
5647 */
5648 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5649 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5650 panic("bad overwrite, hdr=%p exists=%p",
5651 (void *)hdr, (void *)exists);
5652 ASSERT(refcount_is_zero(
5653 &exists->b_l1hdr.b_refcnt));
5654 arc_change_state(arc_anon, exists, hash_lock);
5655 mutex_exit(hash_lock);
5656 arc_hdr_destroy(exists);
5657 exists = buf_hash_insert(hdr, &hash_lock);
5658 ASSERT3P(exists, ==, NULL);
5659 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5660 /* nopwrite */
5661 ASSERT(zio->io_prop.zp_nopwrite);
5662 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5663 panic("bad nopwrite, hdr=%p exists=%p",
5664 (void *)hdr, (void *)exists);
5665 } else {
5666 /* Dedup */
5667 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
5668 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5669 ASSERT(BP_GET_DEDUP(zio->io_bp));
5670 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5671 }
5672 }
5673 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5674 /* if it's not anon, we are doing a scrub */
5675 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
5676 arc_access(hdr, hash_lock);
5677 mutex_exit(hash_lock);
5678 } else {
5679 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5680 }
5681
5682 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5683 callback->awcb_done(zio, buf, callback->awcb_private);
5684
5685 kmem_free(callback, sizeof (arc_write_callback_t));
5686 }
5687
5688 zio_t *
arc_write(zio_t * pio,spa_t * spa,uint64_t txg,blkptr_t * bp,arc_buf_t * buf,boolean_t l2arc,const zio_prop_t * zp,arc_done_func_t * ready,arc_done_func_t * children_ready,arc_done_func_t * physdone,arc_done_func_t * done,void * private,zio_priority_t priority,int zio_flags,const zbookmark_phys_t * zb)5689 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
5690 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
5691 arc_done_func_t *children_ready, arc_done_func_t *physdone,
5692 arc_done_func_t *done, void *private, zio_priority_t priority,
5693 int zio_flags, const zbookmark_phys_t *zb)
5694 {
5695 arc_buf_hdr_t *hdr = buf->b_hdr;
5696 arc_write_callback_t *callback;
5697 zio_t *zio;
5698
5699 ASSERT3P(ready, !=, NULL);
5700 ASSERT3P(done, !=, NULL);
5701 ASSERT(!HDR_IO_ERROR(hdr));
5702 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5703 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5704 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
5705 if (l2arc)
5706 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5707 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
5708 callback->awcb_ready = ready;
5709 callback->awcb_children_ready = children_ready;
5710 callback->awcb_physdone = physdone;
5711 callback->awcb_done = done;
5712 callback->awcb_private = private;
5713 callback->awcb_buf = buf;
5714
5715 /*
5716 * The hdr's b_pdata is now stale, free it now. A new data block
5717 * will be allocated when the zio pipeline calls arc_write_ready().
5718 */
5719 if (hdr->b_l1hdr.b_pdata != NULL) {
5720 /*
5721 * If the buf is currently sharing the data block with
5722 * the hdr then we need to break that relationship here.
5723 * The hdr will remain with a NULL data pointer and the
5724 * buf will take sole ownership of the block.
5725 */
5726 if (arc_buf_is_shared(buf)) {
5727 ASSERT(ARC_BUF_LAST(buf));
5728 arc_unshare_buf(hdr, buf);
5729 } else {
5730 arc_hdr_free_pdata(hdr);
5731 }
5732 VERIFY3P(buf->b_data, !=, NULL);
5733 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
5734 }
5735 ASSERT(!arc_buf_is_shared(buf));
5736 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
5737
5738 zio = zio_write(pio, spa, txg, bp, buf->b_data, HDR_GET_LSIZE(hdr), zp,
5739 arc_write_ready,
5740 (children_ready != NULL) ? arc_write_children_ready : NULL,
5741 arc_write_physdone, arc_write_done, callback,
5742 priority, zio_flags, zb);
5743
5744 return (zio);
5745 }
5746
5747 static int
arc_memory_throttle(uint64_t reserve,uint64_t txg)5748 arc_memory_throttle(uint64_t reserve, uint64_t txg)
5749 {
5750 #ifdef _KERNEL
5751 uint64_t available_memory = ptob(freemem);
5752 static uint64_t page_load = 0;
5753 static uint64_t last_txg = 0;
5754
5755 #if !defined(_LP64)
5756 available_memory =
5757 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
5758 #endif
5759
5760 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
5761 return (0);
5762
5763 if (txg > last_txg) {
5764 last_txg = txg;
5765 page_load = 0;
5766 }
5767 /*
5768 * If we are in pageout, we know that memory is already tight,
5769 * the arc is already going to be evicting, so we just want to
5770 * continue to let page writes occur as quickly as possible.
5771 */
5772 if (curlwp == uvm.pagedaemon_lwp) {
5773 if (page_load > MAX(ptob(minfree), available_memory) / 4)
5774 return (SET_ERROR(ERESTART));
5775 /* Note: reserve is inflated, so we deflate */
5776 page_load += reserve / 8;
5777 return (0);
5778 } else if (page_load > 0 && arc_reclaim_needed()) {
5779 /* memory is low, delay before restarting */
5780 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
5781 return (SET_ERROR(EAGAIN));
5782 }
5783 page_load = 0;
5784 #endif
5785 return (0);
5786 }
5787
5788 void
arc_tempreserve_clear(uint64_t reserve)5789 arc_tempreserve_clear(uint64_t reserve)
5790 {
5791 atomic_add_64(&arc_tempreserve, -reserve);
5792 ASSERT((int64_t)arc_tempreserve >= 0);
5793 }
5794
5795 int
arc_tempreserve_space(uint64_t reserve,uint64_t txg)5796 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
5797 {
5798 int error;
5799 uint64_t anon_size;
5800
5801 if (reserve > arc_c/4 && !arc_no_grow) {
5802 arc_c = MIN(arc_c_max, reserve * 4);
5803 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
5804 }
5805 if (reserve > arc_c)
5806 return (SET_ERROR(ENOMEM));
5807
5808 /*
5809 * Don't count loaned bufs as in flight dirty data to prevent long
5810 * network delays from blocking transactions that are ready to be
5811 * assigned to a txg.
5812 */
5813 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
5814 arc_loaned_bytes), 0);
5815
5816 /*
5817 * Writes will, almost always, require additional memory allocations
5818 * in order to compress/encrypt/etc the data. We therefore need to
5819 * make sure that there is sufficient available memory for this.
5820 */
5821 error = arc_memory_throttle(reserve, txg);
5822 if (error != 0)
5823 return (error);
5824
5825 /*
5826 * Throttle writes when the amount of dirty data in the cache
5827 * gets too large. We try to keep the cache less than half full
5828 * of dirty blocks so that our sync times don't grow too large.
5829 * Note: if two requests come in concurrently, we might let them
5830 * both succeed, when one of them should fail. Not a huge deal.
5831 */
5832
5833 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
5834 anon_size > arc_c / 4) {
5835 uint64_t meta_esize =
5836 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5837 uint64_t data_esize =
5838 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5839 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5840 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5841 arc_tempreserve >> 10, meta_esize >> 10,
5842 data_esize >> 10, reserve >> 10, arc_c >> 10);
5843 return (SET_ERROR(ERESTART));
5844 }
5845 atomic_add_64(&arc_tempreserve, reserve);
5846 return (0);
5847 }
5848
5849 static void
arc_kstat_update_state(arc_state_t * state,kstat_named_t * size,kstat_named_t * evict_data,kstat_named_t * evict_metadata)5850 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
5851 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
5852 {
5853 size->value.ui64 = refcount_count(&state->arcs_size);
5854 evict_data->value.ui64 =
5855 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
5856 evict_metadata->value.ui64 =
5857 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
5858 }
5859
5860 static int
arc_kstat_update(kstat_t * ksp,int rw)5861 arc_kstat_update(kstat_t *ksp, int rw)
5862 {
5863 arc_stats_t *as = ksp->ks_data;
5864
5865 if (rw == KSTAT_WRITE) {
5866 return (EACCES);
5867 } else {
5868 arc_kstat_update_state(arc_anon,
5869 &as->arcstat_anon_size,
5870 &as->arcstat_anon_evictable_data,
5871 &as->arcstat_anon_evictable_metadata);
5872 arc_kstat_update_state(arc_mru,
5873 &as->arcstat_mru_size,
5874 &as->arcstat_mru_evictable_data,
5875 &as->arcstat_mru_evictable_metadata);
5876 arc_kstat_update_state(arc_mru_ghost,
5877 &as->arcstat_mru_ghost_size,
5878 &as->arcstat_mru_ghost_evictable_data,
5879 &as->arcstat_mru_ghost_evictable_metadata);
5880 arc_kstat_update_state(arc_mfu,
5881 &as->arcstat_mfu_size,
5882 &as->arcstat_mfu_evictable_data,
5883 &as->arcstat_mfu_evictable_metadata);
5884 arc_kstat_update_state(arc_mfu_ghost,
5885 &as->arcstat_mfu_ghost_size,
5886 &as->arcstat_mfu_ghost_evictable_data,
5887 &as->arcstat_mfu_ghost_evictable_metadata);
5888 }
5889
5890 return (0);
5891 }
5892
5893 /*
5894 * This function *must* return indices evenly distributed between all
5895 * sublists of the multilist. This is needed due to how the ARC eviction
5896 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5897 * distributed between all sublists and uses this assumption when
5898 * deciding which sublist to evict from and how much to evict from it.
5899 */
5900 unsigned int
arc_state_multilist_index_func(multilist_t * ml,void * obj)5901 arc_state_multilist_index_func(multilist_t *ml, void *obj)
5902 {
5903 arc_buf_hdr_t *hdr = obj;
5904
5905 /*
5906 * We rely on b_dva to generate evenly distributed index
5907 * numbers using buf_hash below. So, as an added precaution,
5908 * let's make sure we never add empty buffers to the arc lists.
5909 */
5910 ASSERT(!HDR_EMPTY(hdr));
5911
5912 /*
5913 * The assumption here, is the hash value for a given
5914 * arc_buf_hdr_t will remain constant throughout it's lifetime
5915 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
5916 * Thus, we don't need to store the header's sublist index
5917 * on insertion, as this index can be recalculated on removal.
5918 *
5919 * Also, the low order bits of the hash value are thought to be
5920 * distributed evenly. Otherwise, in the case that the multilist
5921 * has a power of two number of sublists, each sublists' usage
5922 * would not be evenly distributed.
5923 */
5924 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
5925 multilist_get_num_sublists(ml));
5926 }
5927
5928 #ifdef _KERNEL
5929 #ifdef __FreeBSD__
5930 static eventhandler_tag arc_event_lowmem = NULL;
5931 #endif
5932
5933 static void
arc_lowmem(void * arg __unused,int howto __unused)5934 arc_lowmem(void *arg __unused, int howto __unused)
5935 {
5936
5937 mutex_enter(&arc_reclaim_lock);
5938 /* XXX: Memory deficit should be passed as argument. */
5939 needfree = btoc(arc_c >> arc_shrink_shift);
5940 DTRACE_PROBE(arc__needfree);
5941 cv_signal(&arc_reclaim_thread_cv);
5942
5943 /*
5944 * It is unsafe to block here in arbitrary threads, because we can come
5945 * here from ARC itself and may hold ARC locks and thus risk a deadlock
5946 * with ARC reclaim thread.
5947 */
5948 if (curlwp == uvm.pagedaemon_lwp)
5949 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5950 mutex_exit(&arc_reclaim_lock);
5951 }
5952 #endif
5953
5954 static void
arc_state_init(void)5955 arc_state_init(void)
5956 {
5957 arc_anon = &ARC_anon;
5958 arc_mru = &ARC_mru;
5959 arc_mru_ghost = &ARC_mru_ghost;
5960 arc_mfu = &ARC_mfu;
5961 arc_mfu_ghost = &ARC_mfu_ghost;
5962 arc_l2c_only = &ARC_l2c_only;
5963
5964 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5965 sizeof (arc_buf_hdr_t),
5966 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5967 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5968 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5969 sizeof (arc_buf_hdr_t),
5970 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5971 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5972 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5973 sizeof (arc_buf_hdr_t),
5974 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5975 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5976 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5977 sizeof (arc_buf_hdr_t),
5978 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5979 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5980 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5981 sizeof (arc_buf_hdr_t),
5982 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5983 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5984 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5985 sizeof (arc_buf_hdr_t),
5986 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5987 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5988 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5989 sizeof (arc_buf_hdr_t),
5990 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5991 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5992 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5993 sizeof (arc_buf_hdr_t),
5994 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5995 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5996 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5997 sizeof (arc_buf_hdr_t),
5998 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5999 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
6000 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
6001 sizeof (arc_buf_hdr_t),
6002 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6003 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
6004
6005 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6006 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6007 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6008 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6009 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6010 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6011 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6012 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6013 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6014 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6015 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6016 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6017
6018 refcount_create(&arc_anon->arcs_size);
6019 refcount_create(&arc_mru->arcs_size);
6020 refcount_create(&arc_mru_ghost->arcs_size);
6021 refcount_create(&arc_mfu->arcs_size);
6022 refcount_create(&arc_mfu_ghost->arcs_size);
6023 refcount_create(&arc_l2c_only->arcs_size);
6024 }
6025
6026 static void
arc_state_fini(void)6027 arc_state_fini(void)
6028 {
6029 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6030 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6031 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6032 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6033 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6034 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6035 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6036 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6037 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6038 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6039 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6040 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6041
6042 refcount_destroy(&arc_anon->arcs_size);
6043 refcount_destroy(&arc_mru->arcs_size);
6044 refcount_destroy(&arc_mru_ghost->arcs_size);
6045 refcount_destroy(&arc_mfu->arcs_size);
6046 refcount_destroy(&arc_mfu_ghost->arcs_size);
6047 refcount_destroy(&arc_l2c_only->arcs_size);
6048
6049 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
6050 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6051 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6052 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6053 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
6054 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6055 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
6056 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6057 }
6058
6059 uint64_t
arc_max_bytes(void)6060 arc_max_bytes(void)
6061 {
6062 return (arc_c_max);
6063 }
6064
6065 void
arc_init(void)6066 arc_init(void)
6067 {
6068 int i, prefetch_tunable_set = 0;
6069
6070 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
6071 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
6072 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
6073
6074 #ifdef __FreeBSD__
6075 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
6076 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
6077 #endif
6078
6079 /* Convert seconds to clock ticks */
6080 arc_min_prefetch_lifespan = 1 * hz;
6081
6082 /* Start out with 1/8 of all memory */
6083 arc_c = kmem_size() / 8;
6084
6085 #ifdef illumos
6086 #ifdef _KERNEL
6087 /*
6088 * On architectures where the physical memory can be larger
6089 * than the addressable space (intel in 32-bit mode), we may
6090 * need to limit the cache to 1/8 of VM size.
6091 */
6092 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
6093 #endif
6094 #endif /* illumos */
6095 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
6096 arc_c_min = MAX(arc_c / 4, arc_abs_min);
6097 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
6098 if (arc_c * 8 >= 1 << 30)
6099 arc_c_max = (arc_c * 8) - (1 << 30);
6100 else
6101 arc_c_max = arc_c_min;
6102 arc_c_max = MAX(arc_c * 5, arc_c_max);
6103
6104 /*
6105 * In userland, there's only the memory pressure that we artificially
6106 * create (see arc_available_memory()). Don't let arc_c get too
6107 * small, because it can cause transactions to be larger than
6108 * arc_c, causing arc_tempreserve_space() to fail.
6109 */
6110 #ifndef _KERNEL
6111 arc_c_min = arc_c_max / 2;
6112 #endif
6113
6114 #ifdef _KERNEL
6115 /*
6116 * Allow the tunables to override our calculations if they are
6117 * reasonable.
6118 */
6119 if (zfs_arc_max > arc_abs_min && zfs_arc_max < kmem_size()) {
6120 arc_c_max = zfs_arc_max;
6121 arc_c_min = MIN(arc_c_min, arc_c_max);
6122 }
6123 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
6124 arc_c_min = zfs_arc_min;
6125 #endif
6126
6127 arc_c = arc_c_max;
6128 arc_p = (arc_c >> 1);
6129 arc_size = 0;
6130
6131 /* limit meta-data to 1/4 of the arc capacity */
6132 arc_meta_limit = arc_c_max / 4;
6133
6134 /* Allow the tunable to override if it is reasonable */
6135 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6136 arc_meta_limit = zfs_arc_meta_limit;
6137
6138 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6139 arc_c_min = arc_meta_limit / 2;
6140
6141 if (zfs_arc_meta_min > 0) {
6142 arc_meta_min = zfs_arc_meta_min;
6143 } else {
6144 arc_meta_min = arc_c_min / 2;
6145 }
6146
6147 if (zfs_arc_grow_retry > 0)
6148 arc_grow_retry = zfs_arc_grow_retry;
6149
6150 if (zfs_arc_shrink_shift > 0)
6151 arc_shrink_shift = zfs_arc_shrink_shift;
6152
6153 /*
6154 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6155 */
6156 if (arc_no_grow_shift >= arc_shrink_shift)
6157 arc_no_grow_shift = arc_shrink_shift - 1;
6158
6159 if (zfs_arc_p_min_shift > 0)
6160 arc_p_min_shift = zfs_arc_p_min_shift;
6161
6162 if (zfs_arc_num_sublists_per_state < 1)
6163 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1);
6164
6165 /* if kmem_flags are set, lets try to use less memory */
6166 if (kmem_debugging())
6167 arc_c = arc_c / 2;
6168 if (arc_c < arc_c_min)
6169 arc_c = arc_c_min;
6170
6171 zfs_arc_min = arc_c_min;
6172 zfs_arc_max = arc_c_max;
6173
6174 arc_state_init();
6175 buf_init();
6176
6177 arc_reclaim_thread_exit = B_FALSE;
6178 #ifdef __FreeBSD__
6179 arc_dnlc_evicts_thread_exit = FALSE;
6180 #endif
6181
6182 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6183 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6184
6185 if (arc_ksp != NULL) {
6186 arc_ksp->ks_data = &arc_stats;
6187 arc_ksp->ks_update = arc_kstat_update;
6188 kstat_install(arc_ksp);
6189 }
6190
6191 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6192 TS_RUN, minclsyspri);
6193
6194 #ifdef __FreeBSD__
6195 #ifdef _KERNEL
6196 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
6197 EVENTHANDLER_PRI_FIRST);
6198 #endif
6199
6200 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
6201 TS_RUN, minclsyspri);
6202 #endif
6203
6204 arc_dead = B_FALSE;
6205 arc_warm = B_FALSE;
6206
6207 /*
6208 * Calculate maximum amount of dirty data per pool.
6209 *
6210 * If it has been set by /etc/system, take that.
6211 * Otherwise, use a percentage of physical memory defined by
6212 * zfs_dirty_data_max_percent (default 10%) with a cap at
6213 * zfs_dirty_data_max_max (default 4GB).
6214 */
6215 if (zfs_dirty_data_max == 0) {
6216 zfs_dirty_data_max = ptob(physmem) *
6217 zfs_dirty_data_max_percent / 100;
6218 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6219 zfs_dirty_data_max_max);
6220 }
6221
6222 #ifdef _KERNEL
6223 #ifdef __FreeBSD__
6224 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
6225 prefetch_tunable_set = 1;
6226
6227 #ifdef __i386__
6228 if (prefetch_tunable_set == 0) {
6229 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
6230 "-- to enable,\n");
6231 printf(" add \"vfs.zfs.prefetch_disable=0\" "
6232 "to /boot/loader.conf.\n");
6233 zfs_prefetch_disable = 1;
6234 }
6235 #else
6236 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
6237 prefetch_tunable_set == 0) {
6238 printf("ZFS NOTICE: Prefetch is disabled by default if less "
6239 "than 4GB of RAM is present;\n"
6240 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
6241 "to /boot/loader.conf.\n");
6242 zfs_prefetch_disable = 1;
6243 }
6244 #endif
6245 #endif
6246 /* Warn about ZFS memory and address space requirements. */
6247 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
6248 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
6249 "expect unstable behavior.\n");
6250 }
6251 if (kmem_size() < 512 * (1 << 20)) {
6252 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
6253 "expect unstable behavior.\n");
6254 #ifdef __FreeBSD__
6255 printf(" Consider tuning vm.kmem_size and "
6256 "vm.kmem_size_max\n");
6257 printf(" in /boot/loader.conf.\n");
6258 #endif
6259 }
6260 #endif
6261 }
6262
6263 void
arc_fini(void)6264 arc_fini(void)
6265 {
6266 mutex_enter(&arc_reclaim_lock);
6267 arc_reclaim_thread_exit = B_TRUE;
6268 /*
6269 * The reclaim thread will set arc_reclaim_thread_exit back to
6270 * B_FALSE when it is finished exiting; we're waiting for that.
6271 */
6272 while (arc_reclaim_thread_exit) {
6273 cv_signal(&arc_reclaim_thread_cv);
6274 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6275 }
6276 mutex_exit(&arc_reclaim_lock);
6277
6278 /* Use B_TRUE to ensure *all* buffers are evicted */
6279 arc_flush(NULL, B_TRUE);
6280
6281 #ifdef __FreeBSD__
6282 mutex_enter(&arc_dnlc_evicts_lock);
6283 arc_dnlc_evicts_thread_exit = TRUE;
6284
6285 /*
6286 * The user evicts thread will set arc_user_evicts_thread_exit
6287 * to FALSE when it is finished exiting; we're waiting for that.
6288 */
6289 while (arc_dnlc_evicts_thread_exit) {
6290 cv_signal(&arc_dnlc_evicts_cv);
6291 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
6292 }
6293 mutex_exit(&arc_dnlc_evicts_lock);
6294
6295 mutex_destroy(&arc_dnlc_evicts_lock);
6296 cv_destroy(&arc_dnlc_evicts_cv);
6297 #endif
6298
6299 arc_dead = B_TRUE;
6300
6301 if (arc_ksp != NULL) {
6302 kstat_delete(arc_ksp);
6303 arc_ksp = NULL;
6304 }
6305
6306 mutex_destroy(&arc_reclaim_lock);
6307 cv_destroy(&arc_reclaim_thread_cv);
6308 cv_destroy(&arc_reclaim_waiters_cv);
6309
6310 arc_state_fini();
6311 buf_fini();
6312
6313 ASSERT0(arc_loaned_bytes);
6314
6315 #ifdef __FreeBSD__
6316 #ifdef _KERNEL
6317 if (arc_event_lowmem != NULL)
6318 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
6319 #endif
6320 #endif
6321 }
6322
6323 /*
6324 * Level 2 ARC
6325 *
6326 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6327 * It uses dedicated storage devices to hold cached data, which are populated
6328 * using large infrequent writes. The main role of this cache is to boost
6329 * the performance of random read workloads. The intended L2ARC devices
6330 * include short-stroked disks, solid state disks, and other media with
6331 * substantially faster read latency than disk.
6332 *
6333 * +-----------------------+
6334 * | ARC |
6335 * +-----------------------+
6336 * | ^ ^
6337 * | | |
6338 * l2arc_feed_thread() arc_read()
6339 * | | |
6340 * | l2arc read |
6341 * V | |
6342 * +---------------+ |
6343 * | L2ARC | |
6344 * +---------------+ |
6345 * | ^ |
6346 * l2arc_write() | |
6347 * | | |
6348 * V | |
6349 * +-------+ +-------+
6350 * | vdev | | vdev |
6351 * | cache | | cache |
6352 * +-------+ +-------+
6353 * +=========+ .-----.
6354 * : L2ARC : |-_____-|
6355 * : devices : | Disks |
6356 * +=========+ `-_____-'
6357 *
6358 * Read requests are satisfied from the following sources, in order:
6359 *
6360 * 1) ARC
6361 * 2) vdev cache of L2ARC devices
6362 * 3) L2ARC devices
6363 * 4) vdev cache of disks
6364 * 5) disks
6365 *
6366 * Some L2ARC device types exhibit extremely slow write performance.
6367 * To accommodate for this there are some significant differences between
6368 * the L2ARC and traditional cache design:
6369 *
6370 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6371 * the ARC behave as usual, freeing buffers and placing headers on ghost
6372 * lists. The ARC does not send buffers to the L2ARC during eviction as
6373 * this would add inflated write latencies for all ARC memory pressure.
6374 *
6375 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6376 * It does this by periodically scanning buffers from the eviction-end of
6377 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6378 * not already there. It scans until a headroom of buffers is satisfied,
6379 * which itself is a buffer for ARC eviction. If a compressible buffer is
6380 * found during scanning and selected for writing to an L2ARC device, we
6381 * temporarily boost scanning headroom during the next scan cycle to make
6382 * sure we adapt to compression effects (which might significantly reduce
6383 * the data volume we write to L2ARC). The thread that does this is
6384 * l2arc_feed_thread(), illustrated below; example sizes are included to
6385 * provide a better sense of ratio than this diagram:
6386 *
6387 * head --> tail
6388 * +---------------------+----------+
6389 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6390 * +---------------------+----------+ | o L2ARC eligible
6391 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6392 * +---------------------+----------+ |
6393 * 15.9 Gbytes ^ 32 Mbytes |
6394 * headroom |
6395 * l2arc_feed_thread()
6396 * |
6397 * l2arc write hand <--[oooo]--'
6398 * | 8 Mbyte
6399 * | write max
6400 * V
6401 * +==============================+
6402 * L2ARC dev |####|#|###|###| |####| ... |
6403 * +==============================+
6404 * 32 Gbytes
6405 *
6406 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6407 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6408 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6409 * safe to say that this is an uncommon case, since buffers at the end of
6410 * the ARC lists have moved there due to inactivity.
6411 *
6412 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6413 * then the L2ARC simply misses copying some buffers. This serves as a
6414 * pressure valve to prevent heavy read workloads from both stalling the ARC
6415 * with waits and clogging the L2ARC with writes. This also helps prevent
6416 * the potential for the L2ARC to churn if it attempts to cache content too
6417 * quickly, such as during backups of the entire pool.
6418 *
6419 * 5. After system boot and before the ARC has filled main memory, there are
6420 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6421 * lists can remain mostly static. Instead of searching from tail of these
6422 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6423 * for eligible buffers, greatly increasing its chance of finding them.
6424 *
6425 * The L2ARC device write speed is also boosted during this time so that
6426 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6427 * there are no L2ARC reads, and no fear of degrading read performance
6428 * through increased writes.
6429 *
6430 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6431 * the vdev queue can aggregate them into larger and fewer writes. Each
6432 * device is written to in a rotor fashion, sweeping writes through
6433 * available space then repeating.
6434 *
6435 * 7. The L2ARC does not store dirty content. It never needs to flush
6436 * write buffers back to disk based storage.
6437 *
6438 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6439 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6440 *
6441 * The performance of the L2ARC can be tweaked by a number of tunables, which
6442 * may be necessary for different workloads:
6443 *
6444 * l2arc_write_max max write bytes per interval
6445 * l2arc_write_boost extra write bytes during device warmup
6446 * l2arc_noprefetch skip caching prefetched buffers
6447 * l2arc_headroom number of max device writes to precache
6448 * l2arc_headroom_boost when we find compressed buffers during ARC
6449 * scanning, we multiply headroom by this
6450 * percentage factor for the next scan cycle,
6451 * since more compressed buffers are likely to
6452 * be present
6453 * l2arc_feed_secs seconds between L2ARC writing
6454 *
6455 * Tunables may be removed or added as future performance improvements are
6456 * integrated, and also may become zpool properties.
6457 *
6458 * There are three key functions that control how the L2ARC warms up:
6459 *
6460 * l2arc_write_eligible() check if a buffer is eligible to cache
6461 * l2arc_write_size() calculate how much to write
6462 * l2arc_write_interval() calculate sleep delay between writes
6463 *
6464 * These three functions determine what to write, how much, and how quickly
6465 * to send writes.
6466 */
6467
6468 static boolean_t
l2arc_write_eligible(uint64_t spa_guid,arc_buf_hdr_t * hdr)6469 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6470 {
6471 /*
6472 * A buffer is *not* eligible for the L2ARC if it:
6473 * 1. belongs to a different spa.
6474 * 2. is already cached on the L2ARC.
6475 * 3. has an I/O in progress (it may be an incomplete read).
6476 * 4. is flagged not eligible (zfs property).
6477 */
6478 if (hdr->b_spa != spa_guid) {
6479 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
6480 return (B_FALSE);
6481 }
6482 if (HDR_HAS_L2HDR(hdr)) {
6483 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
6484 return (B_FALSE);
6485 }
6486 if (HDR_IO_IN_PROGRESS(hdr)) {
6487 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
6488 return (B_FALSE);
6489 }
6490 if (!HDR_L2CACHE(hdr)) {
6491 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
6492 return (B_FALSE);
6493 }
6494
6495 return (B_TRUE);
6496 }
6497
6498 static uint64_t
l2arc_write_size(void)6499 l2arc_write_size(void)
6500 {
6501 uint64_t size;
6502
6503 /*
6504 * Make sure our globals have meaningful values in case the user
6505 * altered them.
6506 */
6507 size = l2arc_write_max;
6508 if (size == 0) {
6509 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
6510 "be greater than zero, resetting it to the default (%d)",
6511 L2ARC_WRITE_SIZE);
6512 size = l2arc_write_max = L2ARC_WRITE_SIZE;
6513 }
6514
6515 if (arc_warm == B_FALSE)
6516 size += l2arc_write_boost;
6517
6518 return (size);
6519
6520 }
6521
6522 static clock_t
l2arc_write_interval(clock_t began,uint64_t wanted,uint64_t wrote)6523 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
6524 {
6525 clock_t interval, next, now;
6526
6527 /*
6528 * If the ARC lists are busy, increase our write rate; if the
6529 * lists are stale, idle back. This is achieved by checking
6530 * how much we previously wrote - if it was more than half of
6531 * what we wanted, schedule the next write much sooner.
6532 */
6533 if (l2arc_feed_again && wrote > (wanted / 2))
6534 interval = (hz * l2arc_feed_min_ms) / 1000;
6535 else
6536 interval = hz * l2arc_feed_secs;
6537
6538 now = ddi_get_lbolt();
6539 next = MAX(now, MIN(now + interval, began + interval));
6540
6541 return (next);
6542 }
6543
6544 /*
6545 * Cycle through L2ARC devices. This is how L2ARC load balances.
6546 * If a device is returned, this also returns holding the spa config lock.
6547 */
6548 static l2arc_dev_t *
l2arc_dev_get_next(void)6549 l2arc_dev_get_next(void)
6550 {
6551 l2arc_dev_t *first, *next = NULL;
6552
6553 /*
6554 * Lock out the removal of spas (spa_namespace_lock), then removal
6555 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6556 * both locks will be dropped and a spa config lock held instead.
6557 */
6558 mutex_enter(&spa_namespace_lock);
6559 mutex_enter(&l2arc_dev_mtx);
6560
6561 /* if there are no vdevs, there is nothing to do */
6562 if (l2arc_ndev == 0)
6563 goto out;
6564
6565 first = NULL;
6566 next = l2arc_dev_last;
6567 do {
6568 /* loop around the list looking for a non-faulted vdev */
6569 if (next == NULL) {
6570 next = list_head(l2arc_dev_list);
6571 } else {
6572 next = list_next(l2arc_dev_list, next);
6573 if (next == NULL)
6574 next = list_head(l2arc_dev_list);
6575 }
6576
6577 /* if we have come back to the start, bail out */
6578 if (first == NULL)
6579 first = next;
6580 else if (next == first)
6581 break;
6582
6583 } while (vdev_is_dead(next->l2ad_vdev));
6584
6585 /* if we were unable to find any usable vdevs, return NULL */
6586 if (vdev_is_dead(next->l2ad_vdev))
6587 next = NULL;
6588
6589 l2arc_dev_last = next;
6590
6591 out:
6592 mutex_exit(&l2arc_dev_mtx);
6593
6594 /*
6595 * Grab the config lock to prevent the 'next' device from being
6596 * removed while we are writing to it.
6597 */
6598 if (next != NULL)
6599 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6600 mutex_exit(&spa_namespace_lock);
6601
6602 return (next);
6603 }
6604
6605 /*
6606 * Free buffers that were tagged for destruction.
6607 */
6608 static void
l2arc_do_free_on_write()6609 l2arc_do_free_on_write()
6610 {
6611 list_t *buflist;
6612 l2arc_data_free_t *df, *df_prev;
6613
6614 mutex_enter(&l2arc_free_on_write_mtx);
6615 buflist = l2arc_free_on_write;
6616
6617 for (df = list_tail(buflist); df; df = df_prev) {
6618 df_prev = list_prev(buflist, df);
6619 ASSERT3P(df->l2df_data, !=, NULL);
6620 if (df->l2df_type == ARC_BUFC_METADATA) {
6621 zio_buf_free(df->l2df_data, df->l2df_size);
6622 } else {
6623 ASSERT(df->l2df_type == ARC_BUFC_DATA);
6624 zio_data_buf_free(df->l2df_data, df->l2df_size);
6625 }
6626 list_remove(buflist, df);
6627 kmem_free(df, sizeof (l2arc_data_free_t));
6628 }
6629
6630 mutex_exit(&l2arc_free_on_write_mtx);
6631 }
6632
6633 /*
6634 * A write to a cache device has completed. Update all headers to allow
6635 * reads from these buffers to begin.
6636 */
6637 static void
l2arc_write_done(zio_t * zio)6638 l2arc_write_done(zio_t *zio)
6639 {
6640 l2arc_write_callback_t *cb;
6641 l2arc_dev_t *dev;
6642 list_t *buflist;
6643 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6644 kmutex_t *hash_lock;
6645 int64_t bytes_dropped = 0;
6646
6647 cb = zio->io_private;
6648 ASSERT3P(cb, !=, NULL);
6649 dev = cb->l2wcb_dev;
6650 ASSERT3P(dev, !=, NULL);
6651 head = cb->l2wcb_head;
6652 ASSERT3P(head, !=, NULL);
6653 buflist = &dev->l2ad_buflist;
6654 ASSERT3P(buflist, !=, NULL);
6655 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
6656 l2arc_write_callback_t *, cb);
6657
6658 if (zio->io_error != 0)
6659 ARCSTAT_BUMP(arcstat_l2_writes_error);
6660
6661 /*
6662 * All writes completed, or an error was hit.
6663 */
6664 top:
6665 mutex_enter(&dev->l2ad_mtx);
6666 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
6667 hdr_prev = list_prev(buflist, hdr);
6668
6669 hash_lock = HDR_LOCK(hdr);
6670
6671 /*
6672 * We cannot use mutex_enter or else we can deadlock
6673 * with l2arc_write_buffers (due to swapping the order
6674 * the hash lock and l2ad_mtx are taken).
6675 */
6676 if (!mutex_tryenter(hash_lock)) {
6677 /*
6678 * Missed the hash lock. We must retry so we
6679 * don't leave the ARC_FLAG_L2_WRITING bit set.
6680 */
6681 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
6682
6683 /*
6684 * We don't want to rescan the headers we've
6685 * already marked as having been written out, so
6686 * we reinsert the head node so we can pick up
6687 * where we left off.
6688 */
6689 list_remove(buflist, head);
6690 list_insert_after(buflist, hdr, head);
6691
6692 mutex_exit(&dev->l2ad_mtx);
6693
6694 /*
6695 * We wait for the hash lock to become available
6696 * to try and prevent busy waiting, and increase
6697 * the chance we'll be able to acquire the lock
6698 * the next time around.
6699 */
6700 mutex_enter(hash_lock);
6701 mutex_exit(hash_lock);
6702 goto top;
6703 }
6704
6705 /*
6706 * We could not have been moved into the arc_l2c_only
6707 * state while in-flight due to our ARC_FLAG_L2_WRITING
6708 * bit being set. Let's just ensure that's being enforced.
6709 */
6710 ASSERT(HDR_HAS_L1HDR(hdr));
6711
6712 if (zio->io_error != 0) {
6713 /*
6714 * Error - drop L2ARC entry.
6715 */
6716 list_remove(buflist, hdr);
6717 l2arc_trim(hdr);
6718 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
6719
6720 ARCSTAT_INCR(arcstat_l2_asize, -arc_hdr_size(hdr));
6721 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr));
6722
6723 bytes_dropped += arc_hdr_size(hdr);
6724 (void) refcount_remove_many(&dev->l2ad_alloc,
6725 arc_hdr_size(hdr), hdr);
6726 }
6727
6728 /*
6729 * Allow ARC to begin reads and ghost list evictions to
6730 * this L2ARC entry.
6731 */
6732 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
6733
6734 mutex_exit(hash_lock);
6735 }
6736
6737 atomic_inc_64(&l2arc_writes_done);
6738 list_remove(buflist, head);
6739 ASSERT(!HDR_HAS_L1HDR(head));
6740 kmem_cache_free(hdr_l2only_cache, head);
6741 mutex_exit(&dev->l2ad_mtx);
6742
6743 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
6744
6745 l2arc_do_free_on_write();
6746
6747 kmem_free(cb, sizeof (l2arc_write_callback_t));
6748 }
6749
6750 /*
6751 * A read to a cache device completed. Validate buffer contents before
6752 * handing over to the regular ARC routines.
6753 */
6754 static void
l2arc_read_done(zio_t * zio)6755 l2arc_read_done(zio_t *zio)
6756 {
6757 l2arc_read_callback_t *cb;
6758 arc_buf_hdr_t *hdr;
6759 kmutex_t *hash_lock;
6760 boolean_t valid_cksum;
6761
6762 ASSERT3P(zio->io_vd, !=, NULL);
6763 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
6764
6765 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
6766
6767 cb = zio->io_private;
6768 ASSERT3P(cb, !=, NULL);
6769 hdr = cb->l2rcb_hdr;
6770 ASSERT3P(hdr, !=, NULL);
6771
6772 hash_lock = HDR_LOCK(hdr);
6773 mutex_enter(hash_lock);
6774 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6775
6776 /*
6777 * If the data was read into a temporary buffer,
6778 * move it and free the buffer.
6779 */
6780 if (cb->l2rcb_data != NULL) {
6781 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
6782 if (zio->io_error == 0) {
6783 bcopy(cb->l2rcb_data, hdr->b_l1hdr.b_pdata,
6784 arc_hdr_size(hdr));
6785 }
6786
6787 /*
6788 * The following must be done regardless of whether
6789 * there was an error:
6790 * - free the temporary buffer
6791 * - point zio to the real ARC buffer
6792 * - set zio size accordingly
6793 * These are required because zio is either re-used for
6794 * an I/O of the block in the case of the error
6795 * or the zio is passed to arc_read_done() and it
6796 * needs real data.
6797 */
6798 zio_data_buf_free(cb->l2rcb_data, zio->io_size);
6799 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
6800 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_pdata;
6801 }
6802
6803 ASSERT3P(zio->io_data, !=, NULL);
6804
6805 /*
6806 * Check this survived the L2ARC journey.
6807 */
6808 ASSERT3P(zio->io_data, ==, hdr->b_l1hdr.b_pdata);
6809 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
6810 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
6811
6812 valid_cksum = arc_cksum_is_equal(hdr, zio);
6813 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
6814 mutex_exit(hash_lock);
6815 zio->io_private = hdr;
6816 arc_read_done(zio);
6817 } else {
6818 mutex_exit(hash_lock);
6819 /*
6820 * Buffer didn't survive caching. Increment stats and
6821 * reissue to the original storage device.
6822 */
6823 if (zio->io_error != 0) {
6824 ARCSTAT_BUMP(arcstat_l2_io_error);
6825 } else {
6826 zio->io_error = SET_ERROR(EIO);
6827 }
6828 if (!valid_cksum)
6829 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
6830
6831 /*
6832 * If there's no waiter, issue an async i/o to the primary
6833 * storage now. If there *is* a waiter, the caller must
6834 * issue the i/o in a context where it's OK to block.
6835 */
6836 if (zio->io_waiter == NULL) {
6837 zio_t *pio = zio_unique_parent(zio);
6838
6839 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6840
6841 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
6842 hdr->b_l1hdr.b_pdata, zio->io_size, arc_read_done,
6843 hdr, zio->io_priority, cb->l2rcb_flags,
6844 &cb->l2rcb_zb));
6845 }
6846 }
6847
6848 kmem_free(cb, sizeof (l2arc_read_callback_t));
6849 }
6850
6851 /*
6852 * This is the list priority from which the L2ARC will search for pages to
6853 * cache. This is used within loops (0..3) to cycle through lists in the
6854 * desired order. This order can have a significant effect on cache
6855 * performance.
6856 *
6857 * Currently the metadata lists are hit first, MFU then MRU, followed by
6858 * the data lists. This function returns a locked list, and also returns
6859 * the lock pointer.
6860 */
6861 static multilist_sublist_t *
l2arc_sublist_lock(int list_num)6862 l2arc_sublist_lock(int list_num)
6863 {
6864 multilist_t *ml = NULL;
6865 unsigned int idx;
6866
6867 ASSERT(list_num >= 0 && list_num <= 3);
6868
6869 switch (list_num) {
6870 case 0:
6871 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
6872 break;
6873 case 1:
6874 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
6875 break;
6876 case 2:
6877 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
6878 break;
6879 case 3:
6880 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
6881 break;
6882 }
6883
6884 /*
6885 * Return a randomly-selected sublist. This is acceptable
6886 * because the caller feeds only a little bit of data for each
6887 * call (8MB). Subsequent calls will result in different
6888 * sublists being selected.
6889 */
6890 idx = multilist_get_random_index(ml);
6891 return (multilist_sublist_lock(ml, idx));
6892 }
6893
6894 /*
6895 * Evict buffers from the device write hand to the distance specified in
6896 * bytes. This distance may span populated buffers, it may span nothing.
6897 * This is clearing a region on the L2ARC device ready for writing.
6898 * If the 'all' boolean is set, every buffer is evicted.
6899 */
6900 static void
l2arc_evict(l2arc_dev_t * dev,uint64_t distance,boolean_t all)6901 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6902 {
6903 list_t *buflist;
6904 arc_buf_hdr_t *hdr, *hdr_prev;
6905 kmutex_t *hash_lock;
6906 uint64_t taddr;
6907
6908 buflist = &dev->l2ad_buflist;
6909
6910 if (!all && dev->l2ad_first) {
6911 /*
6912 * This is the first sweep through the device. There is
6913 * nothing to evict.
6914 */
6915 return;
6916 }
6917
6918 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6919 /*
6920 * When nearing the end of the device, evict to the end
6921 * before the device write hand jumps to the start.
6922 */
6923 taddr = dev->l2ad_end;
6924 } else {
6925 taddr = dev->l2ad_hand + distance;
6926 }
6927 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6928 uint64_t, taddr, boolean_t, all);
6929
6930 top:
6931 mutex_enter(&dev->l2ad_mtx);
6932 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6933 hdr_prev = list_prev(buflist, hdr);
6934
6935 hash_lock = HDR_LOCK(hdr);
6936
6937 /*
6938 * We cannot use mutex_enter or else we can deadlock
6939 * with l2arc_write_buffers (due to swapping the order
6940 * the hash lock and l2ad_mtx are taken).
6941 */
6942 if (!mutex_tryenter(hash_lock)) {
6943 /*
6944 * Missed the hash lock. Retry.
6945 */
6946 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
6947 mutex_exit(&dev->l2ad_mtx);
6948 mutex_enter(hash_lock);
6949 mutex_exit(hash_lock);
6950 goto top;
6951 }
6952
6953 if (HDR_L2_WRITE_HEAD(hdr)) {
6954 /*
6955 * We hit a write head node. Leave it for
6956 * l2arc_write_done().
6957 */
6958 list_remove(buflist, hdr);
6959 mutex_exit(hash_lock);
6960 continue;
6961 }
6962
6963 if (!all && HDR_HAS_L2HDR(hdr) &&
6964 (hdr->b_l2hdr.b_daddr >= taddr ||
6965 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
6966 /*
6967 * We've evicted to the target address,
6968 * or the end of the device.
6969 */
6970 mutex_exit(hash_lock);
6971 break;
6972 }
6973
6974 ASSERT(HDR_HAS_L2HDR(hdr));
6975 if (!HDR_HAS_L1HDR(hdr)) {
6976 ASSERT(!HDR_L2_READING(hdr));
6977 /*
6978 * This doesn't exist in the ARC. Destroy.
6979 * arc_hdr_destroy() will call list_remove()
6980 * and decrement arcstat_l2_size.
6981 */
6982 arc_change_state(arc_anon, hdr, hash_lock);
6983 arc_hdr_destroy(hdr);
6984 } else {
6985 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
6986 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
6987 /*
6988 * Invalidate issued or about to be issued
6989 * reads, since we may be about to write
6990 * over this location.
6991 */
6992 if (HDR_L2_READING(hdr)) {
6993 ARCSTAT_BUMP(arcstat_l2_evict_reading);
6994 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
6995 }
6996
6997 /* Ensure this header has finished being written */
6998 ASSERT(!HDR_L2_WRITING(hdr));
6999
7000 arc_hdr_l2hdr_destroy(hdr);
7001 }
7002 mutex_exit(hash_lock);
7003 }
7004 mutex_exit(&dev->l2ad_mtx);
7005 }
7006
7007 /*
7008 * Find and write ARC buffers to the L2ARC device.
7009 *
7010 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7011 * for reading until they have completed writing.
7012 * The headroom_boost is an in-out parameter used to maintain headroom boost
7013 * state between calls to this function.
7014 *
7015 * Returns the number of bytes actually written (which may be smaller than
7016 * the delta by which the device hand has changed due to alignment).
7017 */
7018 static uint64_t
l2arc_write_buffers(spa_t * spa,l2arc_dev_t * dev,uint64_t target_sz)7019 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
7020 {
7021 arc_buf_hdr_t *hdr, *hdr_prev, *head;
7022 uint64_t write_asize, write_psize, write_sz, headroom;
7023 boolean_t full;
7024 l2arc_write_callback_t *cb;
7025 zio_t *pio, *wzio;
7026 uint64_t guid = spa_load_guid(spa);
7027 int try;
7028
7029 ASSERT3P(dev->l2ad_vdev, !=, NULL);
7030
7031 pio = NULL;
7032 write_sz = write_asize = write_psize = 0;
7033 full = B_FALSE;
7034 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7035 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7036
7037 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
7038 /*
7039 * Copy buffers for L2ARC writing.
7040 */
7041 for (try = 0; try <= 3; try++) {
7042 multilist_sublist_t *mls = l2arc_sublist_lock(try);
7043 uint64_t passed_sz = 0;
7044
7045 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
7046
7047 /*
7048 * L2ARC fast warmup.
7049 *
7050 * Until the ARC is warm and starts to evict, read from the
7051 * head of the ARC lists rather than the tail.
7052 */
7053 if (arc_warm == B_FALSE)
7054 hdr = multilist_sublist_head(mls);
7055 else
7056 hdr = multilist_sublist_tail(mls);
7057 if (hdr == NULL)
7058 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
7059
7060 headroom = target_sz * l2arc_headroom;
7061 if (zfs_compressed_arc_enabled)
7062 headroom = (headroom * l2arc_headroom_boost) / 100;
7063
7064 for (; hdr; hdr = hdr_prev) {
7065 kmutex_t *hash_lock;
7066
7067 if (arc_warm == B_FALSE)
7068 hdr_prev = multilist_sublist_next(mls, hdr);
7069 else
7070 hdr_prev = multilist_sublist_prev(mls, hdr);
7071 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
7072 HDR_GET_LSIZE(hdr));
7073
7074 hash_lock = HDR_LOCK(hdr);
7075 if (!mutex_tryenter(hash_lock)) {
7076 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
7077 /*
7078 * Skip this buffer rather than waiting.
7079 */
7080 continue;
7081 }
7082
7083 passed_sz += HDR_GET_LSIZE(hdr);
7084 if (passed_sz > headroom) {
7085 /*
7086 * Searched too far.
7087 */
7088 mutex_exit(hash_lock);
7089 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
7090 break;
7091 }
7092
7093 if (!l2arc_write_eligible(guid, hdr)) {
7094 mutex_exit(hash_lock);
7095 continue;
7096 }
7097
7098 /*
7099 * We rely on the L1 portion of the header below, so
7100 * it's invalid for this header to have been evicted out
7101 * of the ghost cache, prior to being written out. The
7102 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7103 */
7104 ASSERT(HDR_HAS_L1HDR(hdr));
7105
7106 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
7107 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
7108 ASSERT3U(arc_hdr_size(hdr), >, 0);
7109 uint64_t size = arc_hdr_size(hdr);
7110 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
7111 size);
7112
7113 if ((write_psize + asize) > target_sz) {
7114 full = B_TRUE;
7115 mutex_exit(hash_lock);
7116 ARCSTAT_BUMP(arcstat_l2_write_full);
7117 break;
7118 }
7119
7120 if (pio == NULL) {
7121 /*
7122 * Insert a dummy header on the buflist so
7123 * l2arc_write_done() can find where the
7124 * write buffers begin without searching.
7125 */
7126 mutex_enter(&dev->l2ad_mtx);
7127 list_insert_head(&dev->l2ad_buflist, head);
7128 mutex_exit(&dev->l2ad_mtx);
7129
7130 cb = kmem_alloc(
7131 sizeof (l2arc_write_callback_t), KM_SLEEP);
7132 cb->l2wcb_dev = dev;
7133 cb->l2wcb_head = head;
7134 pio = zio_root(spa, l2arc_write_done, cb,
7135 ZIO_FLAG_CANFAIL);
7136 ARCSTAT_BUMP(arcstat_l2_write_pios);
7137 }
7138
7139 hdr->b_l2hdr.b_dev = dev;
7140 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7141 arc_hdr_set_flags(hdr,
7142 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7143
7144 mutex_enter(&dev->l2ad_mtx);
7145 list_insert_head(&dev->l2ad_buflist, hdr);
7146 mutex_exit(&dev->l2ad_mtx);
7147
7148 (void) refcount_add_many(&dev->l2ad_alloc, size, hdr);
7149
7150 /*
7151 * Normally the L2ARC can use the hdr's data, but if
7152 * we're sharing data between the hdr and one of its
7153 * bufs, L2ARC needs its own copy of the data so that
7154 * the ZIO below can't race with the buf consumer. To
7155 * ensure that this copy will be available for the
7156 * lifetime of the ZIO and be cleaned up afterwards, we
7157 * add it to the l2arc_free_on_write queue.
7158 */
7159 void *to_write;
7160 if (!HDR_SHARED_DATA(hdr) && size == asize) {
7161 to_write = hdr->b_l1hdr.b_pdata;
7162 } else {
7163 arc_buf_contents_t type = arc_buf_type(hdr);
7164 if (type == ARC_BUFC_METADATA) {
7165 to_write = zio_buf_alloc(asize);
7166 } else {
7167 ASSERT3U(type, ==, ARC_BUFC_DATA);
7168 to_write = zio_data_buf_alloc(asize);
7169 }
7170
7171 bcopy(hdr->b_l1hdr.b_pdata, to_write, size);
7172 if (asize != size)
7173 bzero(to_write + size, asize - size);
7174 l2arc_free_data_on_write(to_write, asize, type);
7175 }
7176 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7177 hdr->b_l2hdr.b_daddr, asize, to_write,
7178 ZIO_CHECKSUM_OFF, NULL, hdr,
7179 ZIO_PRIORITY_ASYNC_WRITE,
7180 ZIO_FLAG_CANFAIL, B_FALSE);
7181
7182 write_sz += HDR_GET_LSIZE(hdr);
7183 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7184 zio_t *, wzio);
7185
7186 write_asize += size;
7187 write_psize += asize;
7188 dev->l2ad_hand += asize;
7189
7190 mutex_exit(hash_lock);
7191
7192 (void) zio_nowait(wzio);
7193 }
7194
7195 multilist_sublist_unlock(mls);
7196
7197 if (full == B_TRUE)
7198 break;
7199 }
7200
7201 /* No buffers selected for writing? */
7202 if (pio == NULL) {
7203 ASSERT0(write_sz);
7204 ASSERT(!HDR_HAS_L1HDR(head));
7205 kmem_cache_free(hdr_l2only_cache, head);
7206 return (0);
7207 }
7208
7209 ASSERT3U(write_psize, <=, target_sz);
7210 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7211 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
7212 ARCSTAT_INCR(arcstat_l2_size, write_sz);
7213 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
7214 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
7215
7216 /*
7217 * Bump device hand to the device start if it is approaching the end.
7218 * l2arc_evict() will already have evicted ahead for this case.
7219 */
7220 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7221 dev->l2ad_hand = dev->l2ad_start;
7222 dev->l2ad_first = B_FALSE;
7223 }
7224
7225 dev->l2ad_writing = B_TRUE;
7226 (void) zio_wait(pio);
7227 dev->l2ad_writing = B_FALSE;
7228
7229 return (write_asize);
7230 }
7231
7232 /*
7233 * This thread feeds the L2ARC at regular intervals. This is the beating
7234 * heart of the L2ARC.
7235 */
7236 static void
l2arc_feed_thread(void * dummy __unused)7237 l2arc_feed_thread(void *dummy __unused)
7238 {
7239 callb_cpr_t cpr;
7240 l2arc_dev_t *dev;
7241 spa_t *spa;
7242 uint64_t size, wrote;
7243 clock_t begin, next = ddi_get_lbolt() + hz;
7244
7245 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7246
7247 mutex_enter(&l2arc_feed_thr_lock);
7248
7249 while (l2arc_thread_exit == 0) {
7250 CALLB_CPR_SAFE_BEGIN(&cpr);
7251 #ifdef __NetBSD__
7252 clock_t now = ddi_get_lbolt();
7253 if (next > now)
7254 (void) cv_timedwait(&l2arc_feed_thr_cv,
7255 &l2arc_feed_thr_lock, next - now);
7256 #else
7257 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7258 next - ddi_get_lbolt());
7259 #endif
7260 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7261 next = ddi_get_lbolt() + hz;
7262
7263 /*
7264 * Quick check for L2ARC devices.
7265 */
7266 mutex_enter(&l2arc_dev_mtx);
7267 if (l2arc_ndev == 0) {
7268 mutex_exit(&l2arc_dev_mtx);
7269 continue;
7270 }
7271 mutex_exit(&l2arc_dev_mtx);
7272 begin = ddi_get_lbolt();
7273
7274 /*
7275 * This selects the next l2arc device to write to, and in
7276 * doing so the next spa to feed from: dev->l2ad_spa. This
7277 * will return NULL if there are now no l2arc devices or if
7278 * they are all faulted.
7279 *
7280 * If a device is returned, its spa's config lock is also
7281 * held to prevent device removal. l2arc_dev_get_next()
7282 * will grab and release l2arc_dev_mtx.
7283 */
7284 if ((dev = l2arc_dev_get_next()) == NULL)
7285 continue;
7286
7287 spa = dev->l2ad_spa;
7288 ASSERT3P(spa, !=, NULL);
7289
7290 /*
7291 * If the pool is read-only then force the feed thread to
7292 * sleep a little longer.
7293 */
7294 if (!spa_writeable(spa)) {
7295 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7296 spa_config_exit(spa, SCL_L2ARC, dev);
7297 continue;
7298 }
7299
7300 /*
7301 * Avoid contributing to memory pressure.
7302 */
7303 if (arc_reclaim_needed()) {
7304 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7305 spa_config_exit(spa, SCL_L2ARC, dev);
7306 continue;
7307 }
7308
7309 ARCSTAT_BUMP(arcstat_l2_feeds);
7310
7311 size = l2arc_write_size();
7312
7313 /*
7314 * Evict L2ARC buffers that will be overwritten.
7315 */
7316 l2arc_evict(dev, size, B_FALSE);
7317
7318 /*
7319 * Write ARC buffers.
7320 */
7321 wrote = l2arc_write_buffers(spa, dev, size);
7322
7323 /*
7324 * Calculate interval between writes.
7325 */
7326 next = l2arc_write_interval(begin, size, wrote);
7327 spa_config_exit(spa, SCL_L2ARC, dev);
7328 }
7329
7330 l2arc_thread_exit = 0;
7331 cv_broadcast(&l2arc_feed_thr_cv);
7332 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7333 thread_exit();
7334 }
7335
7336 boolean_t
l2arc_vdev_present(vdev_t * vd)7337 l2arc_vdev_present(vdev_t *vd)
7338 {
7339 l2arc_dev_t *dev;
7340
7341 mutex_enter(&l2arc_dev_mtx);
7342 for (dev = list_head(l2arc_dev_list); dev != NULL;
7343 dev = list_next(l2arc_dev_list, dev)) {
7344 if (dev->l2ad_vdev == vd)
7345 break;
7346 }
7347 mutex_exit(&l2arc_dev_mtx);
7348
7349 return (dev != NULL);
7350 }
7351
7352 /*
7353 * Add a vdev for use by the L2ARC. By this point the spa has already
7354 * validated the vdev and opened it.
7355 */
7356 void
l2arc_add_vdev(spa_t * spa,vdev_t * vd)7357 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7358 {
7359 l2arc_dev_t *adddev;
7360
7361 ASSERT(!l2arc_vdev_present(vd));
7362
7363 vdev_ashift_optimize(vd);
7364
7365 /*
7366 * Create a new l2arc device entry.
7367 */
7368 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7369 adddev->l2ad_spa = spa;
7370 adddev->l2ad_vdev = vd;
7371 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7372 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7373 adddev->l2ad_hand = adddev->l2ad_start;
7374 adddev->l2ad_first = B_TRUE;
7375 adddev->l2ad_writing = B_FALSE;
7376
7377 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7378 /*
7379 * This is a list of all ARC buffers that are still valid on the
7380 * device.
7381 */
7382 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7383 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7384
7385 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7386 refcount_create(&adddev->l2ad_alloc);
7387
7388 /*
7389 * Add device to global list
7390 */
7391 mutex_enter(&l2arc_dev_mtx);
7392 list_insert_head(l2arc_dev_list, adddev);
7393 atomic_inc_64(&l2arc_ndev);
7394 mutex_exit(&l2arc_dev_mtx);
7395 }
7396
7397 /*
7398 * Remove a vdev from the L2ARC.
7399 */
7400 void
l2arc_remove_vdev(vdev_t * vd)7401 l2arc_remove_vdev(vdev_t *vd)
7402 {
7403 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7404
7405 /*
7406 * Find the device by vdev
7407 */
7408 mutex_enter(&l2arc_dev_mtx);
7409 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7410 nextdev = list_next(l2arc_dev_list, dev);
7411 if (vd == dev->l2ad_vdev) {
7412 remdev = dev;
7413 break;
7414 }
7415 }
7416 ASSERT3P(remdev, !=, NULL);
7417
7418 /*
7419 * Remove device from global list
7420 */
7421 list_remove(l2arc_dev_list, remdev);
7422 l2arc_dev_last = NULL; /* may have been invalidated */
7423 atomic_dec_64(&l2arc_ndev);
7424 mutex_exit(&l2arc_dev_mtx);
7425
7426 /*
7427 * Clear all buflists and ARC references. L2ARC device flush.
7428 */
7429 l2arc_evict(remdev, 0, B_TRUE);
7430 list_destroy(&remdev->l2ad_buflist);
7431 mutex_destroy(&remdev->l2ad_mtx);
7432 refcount_destroy(&remdev->l2ad_alloc);
7433 kmem_free(remdev, sizeof (l2arc_dev_t));
7434 }
7435
7436 void
l2arc_init(void)7437 l2arc_init(void)
7438 {
7439 l2arc_thread_exit = 0;
7440 l2arc_ndev = 0;
7441 l2arc_writes_sent = 0;
7442 l2arc_writes_done = 0;
7443
7444 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7445 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7446 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7447 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7448
7449 l2arc_dev_list = &L2ARC_dev_list;
7450 l2arc_free_on_write = &L2ARC_free_on_write;
7451 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7452 offsetof(l2arc_dev_t, l2ad_node));
7453 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7454 offsetof(l2arc_data_free_t, l2df_list_node));
7455 }
7456
7457 void
l2arc_fini(void)7458 l2arc_fini(void)
7459 {
7460 /*
7461 * This is called from dmu_fini(), which is called from spa_fini();
7462 * Because of this, we can assume that all l2arc devices have
7463 * already been removed when the pools themselves were removed.
7464 */
7465
7466 l2arc_do_free_on_write();
7467
7468 mutex_destroy(&l2arc_feed_thr_lock);
7469 cv_destroy(&l2arc_feed_thr_cv);
7470 mutex_destroy(&l2arc_dev_mtx);
7471 mutex_destroy(&l2arc_free_on_write_mtx);
7472
7473 list_destroy(l2arc_dev_list);
7474 list_destroy(l2arc_free_on_write);
7475 }
7476
7477 void
l2arc_start(void)7478 l2arc_start(void)
7479 {
7480 if (!(spa_mode_global & FWRITE))
7481 return;
7482
7483 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7484 TS_RUN, minclsyspri);
7485 }
7486
7487 void
l2arc_stop(void)7488 l2arc_stop(void)
7489 {
7490 if (!(spa_mode_global & FWRITE))
7491 return;
7492
7493 mutex_enter(&l2arc_feed_thr_lock);
7494 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7495 l2arc_thread_exit = 1;
7496 while (l2arc_thread_exit != 0)
7497 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7498 mutex_exit(&l2arc_feed_thr_lock);
7499 }
7500