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 https://opensource.org/licenses/CDDL-1.0. 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/multilist.h> 42 #include <sys/abd.h> 43 #include <sys/zil.h> 44 #include <sys/fm/fs/zfs.h> 45 #ifdef _KERNEL 46 #include <sys/shrinker.h> 47 #include <sys/vmsystm.h> 48 #include <sys/zpl.h> 49 #include <linux/page_compat.h> 50 #include <linux/notifier.h> 51 #include <linux/memory.h> 52 #endif 53 #include <sys/callb.h> 54 #include <sys/kstat.h> 55 #include <sys/zthr.h> 56 #include <zfs_fletcher.h> 57 #include <sys/arc_impl.h> 58 #include <sys/trace_zfs.h> 59 #include <sys/aggsum.h> 60 61 /* 62 * This is a limit on how many pages the ARC shrinker makes available for 63 * eviction in response to one page allocation attempt. Note that in 64 * practice, the kernel's shrinker can ask us to evict up to about 4x this 65 * for one allocation attempt. 66 * 67 * The default limit of 10,000 (in practice, 160MB per allocation attempt 68 * with 4K pages) limits the amount of time spent attempting to reclaim ARC 69 * memory to less than 100ms per allocation attempt, even with a small 70 * average compressed block size of ~8KB. 71 * 72 * See also the comment in arc_shrinker_count(). 73 * Set to 0 to disable limit. 74 */ 75 int zfs_arc_shrinker_limit = 10000; 76 77 #ifdef CONFIG_MEMORY_HOTPLUG 78 static struct notifier_block arc_hotplug_callback_mem_nb; 79 #endif 80 81 /* 82 * Return a default max arc size based on the amount of physical memory. 83 */ 84 uint64_t 85 arc_default_max(uint64_t min, uint64_t allmem) 86 { 87 /* Default to 1/2 of all memory. */ 88 return (MAX(allmem / 2, min)); 89 } 90 91 #ifdef _KERNEL 92 /* 93 * Return maximum amount of memory that we could possibly use. Reduced 94 * to half of all memory in user space which is primarily used for testing. 95 */ 96 uint64_t 97 arc_all_memory(void) 98 { 99 #ifdef CONFIG_HIGHMEM 100 return (ptob(zfs_totalram_pages - zfs_totalhigh_pages)); 101 #else 102 return (ptob(zfs_totalram_pages)); 103 #endif /* CONFIG_HIGHMEM */ 104 } 105 106 /* 107 * Return the amount of memory that is considered free. In user space 108 * which is primarily used for testing we pretend that free memory ranges 109 * from 0-20% of all memory. 110 */ 111 uint64_t 112 arc_free_memory(void) 113 { 114 #ifdef CONFIG_HIGHMEM 115 struct sysinfo si; 116 si_meminfo(&si); 117 return (ptob(si.freeram - si.freehigh)); 118 #else 119 return (ptob(nr_free_pages() + 120 nr_inactive_file_pages())); 121 #endif /* CONFIG_HIGHMEM */ 122 } 123 124 /* 125 * Return the amount of memory that can be consumed before reclaim will be 126 * needed. Positive if there is sufficient free memory, negative indicates 127 * the amount of memory that needs to be freed up. 128 */ 129 int64_t 130 arc_available_memory(void) 131 { 132 return (arc_free_memory() - arc_sys_free); 133 } 134 135 static uint64_t 136 arc_evictable_memory(void) 137 { 138 int64_t asize = aggsum_value(&arc_sums.arcstat_size); 139 uint64_t arc_clean = 140 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_DATA]) + 141 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) + 142 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_DATA]) + 143 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 144 uint64_t arc_dirty = MAX((int64_t)asize - (int64_t)arc_clean, 0); 145 146 /* 147 * Scale reported evictable memory in proportion to page cache, cap 148 * at specified min/max. 149 */ 150 uint64_t min = (ptob(nr_file_pages()) / 100) * zfs_arc_pc_percent; 151 min = MAX(arc_c_min, MIN(arc_c_max, min)); 152 153 if (arc_dirty >= min) 154 return (arc_clean); 155 156 return (MAX((int64_t)asize - (int64_t)min, 0)); 157 } 158 159 /* 160 * The _count() function returns the number of free-able objects. 161 * The _scan() function returns the number of objects that were freed. 162 */ 163 static unsigned long 164 arc_shrinker_count(struct shrinker *shrink, struct shrink_control *sc) 165 { 166 /* 167 * __GFP_FS won't be set if we are called from ZFS code (see 168 * kmem_flags_convert(), which removes it). To avoid a deadlock, we 169 * don't allow evicting in this case. We return 0 rather than 170 * SHRINK_STOP so that the shrinker logic doesn't accumulate a 171 * deficit against us. 172 */ 173 if (!(sc->gfp_mask & __GFP_FS)) { 174 return (0); 175 } 176 177 /* 178 * This code is reached in the "direct reclaim" case, where the 179 * kernel (outside ZFS) is trying to allocate a page, and the system 180 * is low on memory. 181 * 182 * The kernel's shrinker code doesn't understand how many pages the 183 * ARC's callback actually frees, so it may ask the ARC to shrink a 184 * lot for one page allocation. This is problematic because it may 185 * take a long time, thus delaying the page allocation, and because 186 * it may force the ARC to unnecessarily shrink very small. 187 * 188 * Therefore, we limit the amount of data that we say is evictable, 189 * which limits the amount that the shrinker will ask us to evict for 190 * one page allocation attempt. 191 * 192 * In practice, we may be asked to shrink 4x the limit to satisfy one 193 * page allocation, before the kernel's shrinker code gives up on us. 194 * When that happens, we rely on the kernel code to find the pages 195 * that we freed before invoking the OOM killer. This happens in 196 * __alloc_pages_slowpath(), which retries and finds the pages we 197 * freed when it calls get_page_from_freelist(). 198 * 199 * See also the comment above zfs_arc_shrinker_limit. 200 */ 201 int64_t limit = zfs_arc_shrinker_limit != 0 ? 202 zfs_arc_shrinker_limit : INT64_MAX; 203 return (MIN(limit, btop((int64_t)arc_evictable_memory()))); 204 } 205 206 static unsigned long 207 arc_shrinker_scan(struct shrinker *shrink, struct shrink_control *sc) 208 { 209 ASSERT((sc->gfp_mask & __GFP_FS) != 0); 210 211 /* The arc is considered warm once reclaim has occurred */ 212 if (unlikely(arc_warm == B_FALSE)) 213 arc_warm = B_TRUE; 214 215 /* 216 * Evict the requested number of pages by reducing arc_c and waiting 217 * for the requested amount of data to be evicted. 218 */ 219 arc_reduce_target_size(ptob(sc->nr_to_scan)); 220 arc_wait_for_eviction(ptob(sc->nr_to_scan), B_FALSE); 221 if (current->reclaim_state != NULL) 222 #ifdef HAVE_RECLAIM_STATE_RECLAIMED 223 current->reclaim_state->reclaimed += sc->nr_to_scan; 224 #else 225 current->reclaim_state->reclaimed_slab += sc->nr_to_scan; 226 #endif 227 228 /* 229 * We are experiencing memory pressure which the arc_evict_zthr was 230 * unable to keep up with. Set arc_no_grow to briefly pause arc 231 * growth to avoid compounding the memory pressure. 232 */ 233 arc_no_grow = B_TRUE; 234 235 /* 236 * When direct reclaim is observed it usually indicates a rapid 237 * increase in memory pressure. This occurs because the kswapd 238 * threads were unable to asynchronously keep enough free memory 239 * available. 240 */ 241 if (current_is_kswapd()) { 242 ARCSTAT_BUMP(arcstat_memory_indirect_count); 243 } else { 244 ARCSTAT_BUMP(arcstat_memory_direct_count); 245 } 246 247 return (sc->nr_to_scan); 248 } 249 250 SPL_SHRINKER_DECLARE(arc_shrinker, 251 arc_shrinker_count, arc_shrinker_scan, DEFAULT_SEEKS); 252 253 int 254 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg) 255 { 256 uint64_t free_memory = arc_free_memory(); 257 258 if (free_memory > arc_all_memory() * arc_lotsfree_percent / 100) 259 return (0); 260 261 if (txg > spa->spa_lowmem_last_txg) { 262 spa->spa_lowmem_last_txg = txg; 263 spa->spa_lowmem_page_load = 0; 264 } 265 /* 266 * If we are in pageout, we know that memory is already tight, 267 * the arc is already going to be evicting, so we just want to 268 * continue to let page writes occur as quickly as possible. 269 */ 270 if (current_is_kswapd()) { 271 if (spa->spa_lowmem_page_load > 272 MAX(arc_sys_free / 4, free_memory) / 4) { 273 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim); 274 return (SET_ERROR(ERESTART)); 275 } 276 /* Note: reserve is inflated, so we deflate */ 277 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8); 278 return (0); 279 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) { 280 /* memory is low, delay before restarting */ 281 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 282 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim); 283 return (SET_ERROR(EAGAIN)); 284 } 285 spa->spa_lowmem_page_load = 0; 286 return (0); 287 } 288 289 static void 290 arc_set_sys_free(uint64_t allmem) 291 { 292 /* 293 * The ARC tries to keep at least this much memory available for the 294 * system. This gives the ARC time to shrink in response to memory 295 * pressure, before running completely out of memory and invoking the 296 * direct-reclaim ARC shrinker. 297 * 298 * This should be more than twice high_wmark_pages(), so that 299 * arc_wait_for_eviction() will wait until at least the 300 * high_wmark_pages() are free (see arc_evict_state_impl()). 301 * 302 * Note: Even when the system is very low on memory, the kernel's 303 * shrinker code may only ask for one "batch" of pages (512KB) to be 304 * evicted. If concurrent allocations consume these pages, there may 305 * still be insufficient free pages, and the OOM killer takes action. 306 * 307 * By setting arc_sys_free large enough, and having 308 * arc_wait_for_eviction() wait until there is at least arc_sys_free/2 309 * free memory, it is much less likely that concurrent allocations can 310 * consume all the memory that was evicted before checking for 311 * OOM. 312 * 313 * It's hard to iterate the zones from a linux kernel module, which 314 * makes it difficult to determine the watermark dynamically. Instead 315 * we compute the maximum high watermark for this system, based 316 * on the amount of memory, assuming default parameters on Linux kernel 317 * 5.3. 318 */ 319 320 /* 321 * Base wmark_low is 4 * the square root of Kbytes of RAM. 322 */ 323 long wmark = 4 * int_sqrt(allmem/1024) * 1024; 324 325 /* 326 * Clamp to between 128K and 64MB. 327 */ 328 wmark = MAX(wmark, 128 * 1024); 329 wmark = MIN(wmark, 64 * 1024 * 1024); 330 331 /* 332 * watermark_boost can increase the wmark by up to 150%. 333 */ 334 wmark += wmark * 150 / 100; 335 336 /* 337 * arc_sys_free needs to be more than 2x the watermark, because 338 * arc_wait_for_eviction() waits for half of arc_sys_free. Bump this up 339 * to 3x to ensure we're above it. 340 */ 341 arc_sys_free = wmark * 3 + allmem / 32; 342 } 343 344 void 345 arc_lowmem_init(void) 346 { 347 uint64_t allmem = arc_all_memory(); 348 349 /* 350 * Register a shrinker to support synchronous (direct) memory 351 * reclaim from the arc. This is done to prevent kswapd from 352 * swapping out pages when it is preferable to shrink the arc. 353 */ 354 spl_register_shrinker(&arc_shrinker); 355 arc_set_sys_free(allmem); 356 } 357 358 void 359 arc_lowmem_fini(void) 360 { 361 spl_unregister_shrinker(&arc_shrinker); 362 } 363 364 int 365 param_set_arc_u64(const char *buf, zfs_kernel_param_t *kp) 366 { 367 int error; 368 369 error = spl_param_set_u64(buf, kp); 370 if (error < 0) 371 return (SET_ERROR(error)); 372 373 arc_tuning_update(B_TRUE); 374 375 return (0); 376 } 377 378 int 379 param_set_arc_min(const char *buf, zfs_kernel_param_t *kp) 380 { 381 return (param_set_arc_u64(buf, kp)); 382 } 383 384 int 385 param_set_arc_max(const char *buf, zfs_kernel_param_t *kp) 386 { 387 return (param_set_arc_u64(buf, kp)); 388 } 389 390 int 391 param_set_arc_int(const char *buf, zfs_kernel_param_t *kp) 392 { 393 int error; 394 395 error = param_set_int(buf, kp); 396 if (error < 0) 397 return (SET_ERROR(error)); 398 399 arc_tuning_update(B_TRUE); 400 401 return (0); 402 } 403 404 #ifdef CONFIG_MEMORY_HOTPLUG 405 static int 406 arc_hotplug_callback(struct notifier_block *self, unsigned long action, 407 void *arg) 408 { 409 (void) self, (void) arg; 410 uint64_t allmem = arc_all_memory(); 411 if (action != MEM_ONLINE) 412 return (NOTIFY_OK); 413 414 arc_set_limits(allmem); 415 416 #ifdef __LP64__ 417 if (zfs_dirty_data_max_max == 0) 418 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024, 419 allmem * zfs_dirty_data_max_max_percent / 100); 420 #else 421 if (zfs_dirty_data_max_max == 0) 422 zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024, 423 allmem * zfs_dirty_data_max_max_percent / 100); 424 #endif 425 426 arc_set_sys_free(allmem); 427 return (NOTIFY_OK); 428 } 429 #endif 430 431 void 432 arc_register_hotplug(void) 433 { 434 #ifdef CONFIG_MEMORY_HOTPLUG 435 arc_hotplug_callback_mem_nb.notifier_call = arc_hotplug_callback; 436 /* There is no significance to the value 100 */ 437 arc_hotplug_callback_mem_nb.priority = 100; 438 register_memory_notifier(&arc_hotplug_callback_mem_nb); 439 #endif 440 } 441 442 void 443 arc_unregister_hotplug(void) 444 { 445 #ifdef CONFIG_MEMORY_HOTPLUG 446 unregister_memory_notifier(&arc_hotplug_callback_mem_nb); 447 #endif 448 } 449 #else /* _KERNEL */ 450 int64_t 451 arc_available_memory(void) 452 { 453 int64_t lowest = INT64_MAX; 454 455 /* Every 100 calls, free a small amount */ 456 if (random_in_range(100) == 0) 457 lowest = -1024; 458 459 return (lowest); 460 } 461 462 int 463 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg) 464 { 465 (void) spa, (void) reserve, (void) txg; 466 return (0); 467 } 468 469 uint64_t 470 arc_all_memory(void) 471 { 472 return (ptob(physmem) / 2); 473 } 474 475 uint64_t 476 arc_free_memory(void) 477 { 478 return (random_in_range(arc_all_memory() * 20 / 100)); 479 } 480 481 void 482 arc_register_hotplug(void) 483 { 484 } 485 486 void 487 arc_unregister_hotplug(void) 488 { 489 } 490 #endif /* _KERNEL */ 491 492 /* 493 * Helper function for arc_prune_async() it is responsible for safely 494 * handling the execution of a registered arc_prune_func_t. 495 */ 496 static void 497 arc_prune_task(void *ptr) 498 { 499 arc_prune_t *ap = (arc_prune_t *)ptr; 500 arc_prune_func_t *func = ap->p_pfunc; 501 502 if (func != NULL) 503 func(ap->p_adjust, ap->p_private); 504 505 zfs_refcount_remove(&ap->p_refcnt, func); 506 } 507 508 /* 509 * Notify registered consumers they must drop holds on a portion of the ARC 510 * buffered they reference. This provides a mechanism to ensure the ARC can 511 * honor the metadata limit and reclaim otherwise pinned ARC buffers. This 512 * is analogous to dnlc_reduce_cache() but more generic. 513 * 514 * This operation is performed asynchronously so it may be safely called 515 * in the context of the arc_reclaim_thread(). A reference is taken here 516 * for each registered arc_prune_t and the arc_prune_task() is responsible 517 * for releasing it once the registered arc_prune_func_t has completed. 518 */ 519 void 520 arc_prune_async(uint64_t adjust) 521 { 522 arc_prune_t *ap; 523 524 mutex_enter(&arc_prune_mtx); 525 for (ap = list_head(&arc_prune_list); ap != NULL; 526 ap = list_next(&arc_prune_list, ap)) { 527 528 if (zfs_refcount_count(&ap->p_refcnt) >= 2) 529 continue; 530 531 zfs_refcount_add(&ap->p_refcnt, ap->p_pfunc); 532 ap->p_adjust = adjust; 533 if (taskq_dispatch(arc_prune_taskq, arc_prune_task, 534 ap, TQ_SLEEP) == TASKQID_INVALID) { 535 zfs_refcount_remove(&ap->p_refcnt, ap->p_pfunc); 536 continue; 537 } 538 ARCSTAT_BUMP(arcstat_prune); 539 } 540 mutex_exit(&arc_prune_mtx); 541 } 542 543 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, shrinker_limit, INT, ZMOD_RW, 544 "Limit on number of pages that ARC shrinker can reclaim at once"); 545