module/os/linux/zfs/arc_os.c

   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) 2018, Joyent, Inc.
  24  * Copyright (c) 2011, 2019 by Delphix. All rights reserved.
  25  * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
  26  * Copyright 2017 Nexenta Systems, Inc.  All rights reserved.
  27  */
  28
  29 #include <sys/spa.h>
  30 #include <sys/zio.h>
  31 #include <sys/spa_impl.h>
  32 #include <sys/zio_compress.h>
  33 #include <sys/zio_checksum.h>
  34 #include <sys/zfs_context.h>
  35 #include <sys/arc.h>
  36 #include <sys/zfs_refcount.h>
  37 #include <sys/vdev.h>
  38 #include <sys/vdev_trim.h>
  39 #include <sys/vdev_impl.h>
  40 #include <sys/dsl_pool.h>
  41 #include <sys/zio_checksum.h>
  42 #include <sys/multilist.h>
  43 #include <sys/abd.h>
  44 #include <sys/zil.h>
  45 #include <sys/fm/fs/zfs.h>
  46 #ifdef _KERNEL
  47 #include <sys/shrinker.h>
  48 #include <sys/vmsystm.h>
  49 #include <sys/zpl.h>
  50 #include <linux/page_compat.h>
  51 #endif
  52 #include <sys/callb.h>
  53 #include <sys/kstat.h>
  54 #include <sys/zthr.h>
  55 #include <zfs_fletcher.h>
  56 #include <sys/arc_impl.h>
  57 #include <sys/trace_zfs.h>
  58 #include <sys/aggsum.h>
  59
  60 /*
  61  * This is a limit on how many pages the ARC shrinker makes available for
  62  * eviction in response to one page allocation attempt.  Note that in
  63  * practice, the kernel's shrinker can ask us to evict up to about 4x this
  64  * for one allocation attempt.
  65  *
  66  * The default limit of 10,000 (in practice, 160MB per allocation attempt
  67  * with 4K pages) limits the amount of time spent attempting to reclaim ARC
  68  * memory to less than 100ms per allocation attempt, even with a small
  69  * average compressed block size of ~8KB.
  70  *
  71  * See also the comment in arc_shrinker_count().
  72  * Set to 0 to disable limit.
  73  */
  74 int zfs_arc_shrinker_limit = 10000;
  75
  76
  77 /*
  78  * Return a default max arc size based on the amount of physical memory.
  79  */
  80 uint64_t
  81 arc_default_max(uint64_t min, uint64_t allmem)
  82 {
  83         /* Default to 1/2 of all memory. */
  84         return (MAX(allmem / 2, min));
  85 }
  86
  87 #ifdef _KERNEL
  88 /*
  89  * Return maximum amount of memory that we could possibly use.  Reduced
  90  * to half of all memory in user space which is primarily used for testing.
  91  */
  92 uint64_t
  93 arc_all_memory(void)
  94 {
  95 #ifdef CONFIG_HIGHMEM
  96         return (ptob(zfs_totalram_pages - zfs_totalhigh_pages));
  97 #else
  98         return (ptob(zfs_totalram_pages));
  99 #endif /* CONFIG_HIGHMEM */
 100 }
 101
 102 /*
 103  * Return the amount of memory that is considered free.  In user space
 104  * which is primarily used for testing we pretend that free memory ranges
 105  * from 0-20% of all memory.
 106  */
 107 uint64_t
 108 arc_free_memory(void)
 109 {
 110 #ifdef CONFIG_HIGHMEM
 111         struct sysinfo si;
 112         si_meminfo(&si);
 113         return (ptob(si.freeram - si.freehigh));
 114 #else
 115         return (ptob(nr_free_pages() +
 116             nr_inactive_file_pages() +
 117             nr_slab_reclaimable_pages()));
 118 #endif /* CONFIG_HIGHMEM */
 119 }
 120
 121 /*
 122  * Return the amount of memory that can be consumed before reclaim will be
 123  * needed.  Positive if there is sufficient free memory, negative indicates
 124  * the amount of memory that needs to be freed up.
 125  */
 126 int64_t
 127 arc_available_memory(void)
 128 {
 129         return (arc_free_memory() - arc_sys_free);
 130 }
 131
 132 static uint64_t
 133 arc_evictable_memory(void)
 134 {
 135         int64_t asize = aggsum_value(&arc_size);
 136         uint64_t arc_clean =
 137             zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_DATA]) +
 138             zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) +
 139             zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_DATA]) +
 140             zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
 141         uint64_t arc_dirty = MAX((int64_t)asize - (int64_t)arc_clean, 0);
 142
 143         /*
 144          * Scale reported evictable memory in proportion to page cache, cap
 145          * at specified min/max.
 146          */
 147         uint64_t min = (ptob(nr_file_pages()) / 100) * zfs_arc_pc_percent;
 148         min = MAX(arc_c_min, MIN(arc_c_max, min));
 149
 150         if (arc_dirty >= min)
 151                 return (arc_clean);
 152
 153         return (MAX((int64_t)asize - (int64_t)min, 0));
 154 }
 155
 156 /*
 157  * The _count() function returns the number of free-able objects.
 158  * The _scan() function returns the number of objects that were freed.
 159  */
 160 static unsigned long
 161 arc_shrinker_count(struct shrinker *shrink, struct shrink_control *sc)
 162 {
 163         /*
 164          * __GFP_FS won't be set if we are called from ZFS code (see
 165          * kmem_flags_convert(), which removes it).  To avoid a deadlock, we
 166          * don't allow evicting in this case.  We return 0 rather than
 167          * SHRINK_STOP so that the shrinker logic doesn't accumulate a
 168          * deficit against us.
 169          */
 170         if (!(sc->gfp_mask & __GFP_FS)) {
 171                 return (0);
 172         }
 173
 174         /*
 175          * This code is reached in the "direct reclaim" case, where the
 176          * kernel (outside ZFS) is trying to allocate a page, and the system
 177          * is low on memory.
 178          *
 179          * The kernel's shrinker code doesn't understand how many pages the
 180          * ARC's callback actually frees, so it may ask the ARC to shrink a
 181          * lot for one page allocation. This is problematic because it may
 182          * take a long time, thus delaying the page allocation, and because
 183          * it may force the ARC to unnecessarily shrink very small.
 184          *
 185          * Therefore, we limit the amount of data that we say is evictable,
 186          * which limits the amount that the shrinker will ask us to evict for
 187          * one page allocation attempt.
 188          *
 189          * In practice, we may be asked to shrink 4x the limit to satisfy one
 190          * page allocation, before the kernel's shrinker code gives up on us.
 191          * When that happens, we rely on the kernel code to find the pages
 192          * that we freed before invoking the OOM killer.  This happens in
 193          * __alloc_pages_slowpath(), which retries and finds the pages we
 194          * freed when it calls get_page_from_freelist().
 195          *
 196          * See also the comment above zfs_arc_shrinker_limit.
 197          */
 198         int64_t limit = zfs_arc_shrinker_limit != 0 ?
 199             zfs_arc_shrinker_limit : INT64_MAX;
 200         return (MIN(limit, btop((int64_t)arc_evictable_memory())));
 201 }
 202
 203 static unsigned long
 204 arc_shrinker_scan(struct shrinker *shrink, struct shrink_control *sc)
 205 {
 206         ASSERT((sc->gfp_mask & __GFP_FS) != 0);
 207
 208         /* The arc is considered warm once reclaim has occurred */
 209         if (unlikely(arc_warm == B_FALSE))
 210                 arc_warm = B_TRUE;
 211
 212         /*
 213          * Evict the requested number of pages by reducing arc_c and waiting
 214          * for the requested amount of data to be evicted.
 215          */
 216         arc_reduce_target_size(ptob(sc->nr_to_scan));
 217         arc_wait_for_eviction(ptob(sc->nr_to_scan));
 218         if (current->reclaim_state != NULL)
 219                 current->reclaim_state->reclaimed_slab += sc->nr_to_scan;
 220
 221         /*
 222          * We are experiencing memory pressure which the arc_evict_zthr was
 223          * unable to keep up with. Set arc_no_grow to briefly pause arc
 224          * growth to avoid compounding the memory pressure.
 225          */
 226         arc_no_grow = B_TRUE;
 227
 228         /*
 229          * When direct reclaim is observed it usually indicates a rapid
 230          * increase in memory pressure.  This occurs because the kswapd
 231          * threads were unable to asynchronously keep enough free memory
 232          * available.
 233          */
 234         if (current_is_kswapd()) {
 235                 ARCSTAT_BUMP(arcstat_memory_indirect_count);
 236         } else {
 237                 ARCSTAT_BUMP(arcstat_memory_direct_count);
 238         }
 239
 240         return (sc->nr_to_scan);
 241 }
 242
 243 SPL_SHRINKER_DECLARE(arc_shrinker,
 244     arc_shrinker_count, arc_shrinker_scan, DEFAULT_SEEKS);
 245
 246 int
 247 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
 248 {
 249         uint64_t free_memory = arc_free_memory();
 250
 251         if (free_memory > arc_all_memory() * arc_lotsfree_percent / 100)
 252                 return (0);
 253
 254         if (txg > spa->spa_lowmem_last_txg) {
 255                 spa->spa_lowmem_last_txg = txg;
 256                 spa->spa_lowmem_page_load = 0;
 257         }
 258         /*
 259          * If we are in pageout, we know that memory is already tight,
 260          * the arc is already going to be evicting, so we just want to
 261          * continue to let page writes occur as quickly as possible.
 262          */
 263         if (current_is_kswapd()) {
 264                 if (spa->spa_lowmem_page_load >
 265                     MAX(arc_sys_free / 4, free_memory) / 4) {
 266                         DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
 267                         return (SET_ERROR(ERESTART));
 268                 }
 269                 /* Note: reserve is inflated, so we deflate */
 270                 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
 271                 return (0);
 272         } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
 273                 /* memory is low, delay before restarting */
 274                 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
 275                 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
 276                 return (SET_ERROR(EAGAIN));
 277         }
 278         spa->spa_lowmem_page_load = 0;
 279         return (0);
 280 }
 281
 282 void
 283 arc_lowmem_init(void)
 284 {
 285         uint64_t allmem = arc_all_memory();
 286
 287         /*
 288          * Register a shrinker to support synchronous (direct) memory
 289          * reclaim from the arc.  This is done to prevent kswapd from
 290          * swapping out pages when it is preferable to shrink the arc.
 291          */
 292         spl_register_shrinker(&arc_shrinker);
 293
 294         /*
 295          * The ARC tries to keep at least this much memory available for the
 296          * system.  This gives the ARC time to shrink in response to memory
 297          * pressure, before running completely out of memory and invoking the
 298          * direct-reclaim ARC shrinker.
 299          *
 300          * This should be more than twice high_wmark_pages(), so that
 301          * arc_wait_for_eviction() will wait until at least the
 302          * high_wmark_pages() are free (see arc_evict_state_impl()).
 303          *
 304          * Note: Even when the system is very low on memory, the kernel's
 305          * shrinker code may only ask for one "batch" of pages (512KB) to be
 306          * evicted.  If concurrent allocations consume these pages, there may
 307          * still be insufficient free pages, and the OOM killer takes action.
 308          *
 309          * By setting arc_sys_free large enough, and having
 310          * arc_wait_for_eviction() wait until there is at least arc_sys_free/2
 311          * free memory, it is much less likely that concurrent allocations can
 312          * consume all the memory that was evicted before checking for
 313          * OOM.
 314          *
 315          * It's hard to iterate the zones from a linux kernel module, which
 316          * makes it difficult to determine the watermark dynamically. Instead
 317          * we compute the maximum high watermark for this system, based
 318          * on the amount of memory, assuming default parameters on Linux kernel
 319          * 5.3.
 320          */
 321
 322         /*
 323          * Base wmark_low is 4 * the square root of Kbytes of RAM.
 324          */
 325         long wmark = 4 * int_sqrt(allmem/1024) * 1024;
 326
 327         /*
 328          * Clamp to between 128K and 64MB.
 329          */
 330         wmark = MAX(wmark, 128 * 1024);
 331         wmark = MIN(wmark, 64 * 1024 * 1024);
 332
 333         /*
 334          * watermark_boost can increase the wmark by up to 150%.
 335          */
 336         wmark += wmark * 150 / 100;
 337
 338         /*
 339          * arc_sys_free needs to be more than 2x the watermark, because
 340          * arc_wait_for_eviction() waits for half of arc_sys_free.  Bump this up
 341          * to 3x to ensure we're above it.
 342          */
 343         arc_sys_free = wmark * 3 + allmem / 32;
 344 }
 345
 346 void
 347 arc_lowmem_fini(void)
 348 {
 349         spl_unregister_shrinker(&arc_shrinker);
 350 }
 351
 352 int
 353 param_set_arc_long(const char *buf, zfs_kernel_param_t *kp)
 354 {
 355         int error;
 356
 357         error = param_set_long(buf, kp);
 358         if (error < 0)
 359                 return (SET_ERROR(error));
 360
 361         arc_tuning_update(B_TRUE);
 362
 363         return (0);
 364 }
 365
 366 int
 367 param_set_arc_int(const char *buf, zfs_kernel_param_t *kp)
 368 {
 369         int error;
 370
 371         error = param_set_int(buf, kp);
 372         if (error < 0)
 373                 return (SET_ERROR(error));
 374
 375         arc_tuning_update(B_TRUE);
 376
 377         return (0);
 378 }
 379 #else /* _KERNEL */
 380 int64_t
 381 arc_available_memory(void)
 382 {
 383         int64_t lowest = INT64_MAX;
 384
 385         /* Every 100 calls, free a small amount */
 386         if (spa_get_random(100) == 0)
 387                 lowest = -1024;
 388
 389         return (lowest);
 390 }
 391
 392 int
 393 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
 394 {
 395         return (0);
 396 }
 397
 398 uint64_t
 399 arc_all_memory(void)
 400 {
 401         return (ptob(physmem) / 2);
 402 }
 403
 404 uint64_t
 405 arc_free_memory(void)
 406 {
 407         return (spa_get_random(arc_all_memory() * 20 / 100));
 408 }
 409 #endif /* _KERNEL */
 410
 411 /*
 412  * Helper function for arc_prune_async() it is responsible for safely
 413  * handling the execution of a registered arc_prune_func_t.
 414  */
 415 static void
 416 arc_prune_task(void *ptr)
 417 {
 418         arc_prune_t *ap = (arc_prune_t *)ptr;
 419         arc_prune_func_t *func = ap->p_pfunc;
 420
 421         if (func != NULL)
 422                 func(ap->p_adjust, ap->p_private);
 423
 424         zfs_refcount_remove(&ap->p_refcnt, func);
 425 }
 426
 427 /*
 428  * Notify registered consumers they must drop holds on a portion of the ARC
 429  * buffered they reference.  This provides a mechanism to ensure the ARC can
 430  * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers.  This
 431  * is analogous to dnlc_reduce_cache() but more generic.
 432  *
 433  * This operation is performed asynchronously so it may be safely called
 434  * in the context of the arc_reclaim_thread().  A reference is taken here
 435  * for each registered arc_prune_t and the arc_prune_task() is responsible
 436  * for releasing it once the registered arc_prune_func_t has completed.
 437  */
 438 void
 439 arc_prune_async(int64_t adjust)
 440 {
 441         arc_prune_t *ap;
 442
 443         mutex_enter(&arc_prune_mtx);
 444         for (ap = list_head(&arc_prune_list); ap != NULL;
 445             ap = list_next(&arc_prune_list, ap)) {
 446
 447                 if (zfs_refcount_count(&ap->p_refcnt) >= 2)
 448                         continue;
 449
 450                 zfs_refcount_add(&ap->p_refcnt, ap->p_pfunc);
 451                 ap->p_adjust = adjust;
 452                 if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
 453                     ap, TQ_SLEEP) == TASKQID_INVALID) {
 454                         zfs_refcount_remove(&ap->p_refcnt, ap->p_pfunc);
 455                         continue;
 456                 }
 457                 ARCSTAT_BUMP(arcstat_prune);
 458         }
 459         mutex_exit(&arc_prune_mtx);
 460 }
 461
 462 /* BEGIN CSTYLED */
 463 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, shrinker_limit, INT, ZMOD_RW,
 464         "Limit on number of pages that ARC shrinker can reclaim at once");
 465 /* END CSTYLED */