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 #include <linux/version.h> 53 #endif 54 #include <sys/callb.h> 55 #include <sys/kstat.h> 56 #include <sys/zthr.h> 57 #include <zfs_fletcher.h> 58 #include <sys/arc_impl.h> 59 #include <sys/trace_zfs.h> 60 #include <sys/aggsum.h> 61 62 #ifdef _KERNEL 63 /* 64 * This is a limit on how many pages the ARC shrinker makes available for 65 * eviction in response to one page allocation attempt. Note that in 66 * practice, the kernel's shrinker can ask us to evict up to about 4x this 67 * for one allocation attempt. 68 * 69 * The default limit of 10,000 (in practice, 160MB per allocation attempt 70 * with 4K pages) limits the amount of time spent attempting to reclaim ARC 71 * memory to less than 100ms per allocation attempt, even with a small 72 * average compressed block size of ~8KB. 73 * 74 * See also the comment in arc_shrinker_count(). 75 * Set to 0 to disable limit. 76 */ 77 static int zfs_arc_shrinker_limit = 10000; 78 79 /* 80 * Relative cost of ARC eviction, AKA number of seeks needed to restore evicted 81 * page. Bigger values make ARC more precious and evictions smaller comparing 82 * to other kernel subsystems. Value of 4 means parity with page cache, 83 * according to my reading of kernel's do_shrink_slab() and other code. 84 */ 85 static int zfs_arc_shrinker_seeks = DEFAULT_SEEKS; 86 87 #ifdef CONFIG_MEMORY_HOTPLUG 88 static struct notifier_block arc_hotplug_callback_mem_nb; 89 #endif 90 #endif 91 92 /* 93 * Return a default max arc size based on the amount of physical memory. 94 * This may be overridden by tuning the zfs_arc_max module parameter. 95 */ 96 uint64_t 97 arc_default_max(uint64_t min, uint64_t allmem) 98 { 99 uint64_t size; 100 101 if (allmem >= 1 << 30) 102 size = allmem - (1 << 30); 103 else 104 size = min; 105 return (MAX(allmem * 5 / 8, size)); 106 } 107 108 #ifdef _KERNEL 109 /* 110 * Return maximum amount of memory that we could possibly use. Reduced 111 * to half of all memory in user space which is primarily used for testing. 112 */ 113 uint64_t 114 arc_all_memory(void) 115 { 116 #ifdef CONFIG_HIGHMEM 117 return (ptob(zfs_totalram_pages - zfs_totalhigh_pages)); 118 #else 119 return (ptob(zfs_totalram_pages)); 120 #endif /* CONFIG_HIGHMEM */ 121 } 122 123 /* 124 * Return the amount of memory that is considered free. In user space 125 * which is primarily used for testing we pretend that free memory ranges 126 * from 0-20% of all memory. 127 */ 128 uint64_t 129 arc_free_memory(void) 130 { 131 #ifdef CONFIG_HIGHMEM 132 struct sysinfo si; 133 si_meminfo(&si); 134 return (ptob(si.freeram - si.freehigh)); 135 #else 136 return (ptob(nr_free_pages() + 137 nr_inactive_file_pages())); 138 #endif /* CONFIG_HIGHMEM */ 139 } 140 141 /* 142 * Return the amount of memory that can be consumed before reclaim will be 143 * needed. Positive if there is sufficient free memory, negative indicates 144 * the amount of memory that needs to be freed up. 145 */ 146 int64_t 147 arc_available_memory(void) 148 { 149 return (arc_free_memory() - arc_sys_free); 150 } 151 152 static uint64_t 153 arc_evictable_memory(void) 154 { 155 int64_t asize = aggsum_value(&arc_sums.arcstat_size); 156 uint64_t arc_clean = 157 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_DATA]) + 158 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) + 159 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_DATA]) + 160 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 161 uint64_t arc_dirty = MAX((int64_t)asize - (int64_t)arc_clean, 0); 162 163 /* 164 * Scale reported evictable memory in proportion to page cache, cap 165 * at specified min/max. 166 */ 167 uint64_t min = (ptob(nr_file_pages()) / 100) * zfs_arc_pc_percent; 168 min = MAX(arc_c_min, MIN(arc_c_max, min)); 169 170 if (arc_dirty >= min) 171 return (arc_clean); 172 173 return (MAX((int64_t)asize - (int64_t)min, 0)); 174 } 175 176 /* 177 * The _count() function returns the number of free-able objects. 178 * The _scan() function returns the number of objects that were freed. 179 */ 180 static unsigned long 181 arc_shrinker_count(struct shrinker *shrink, struct shrink_control *sc) 182 { 183 /* 184 * The kernel's shrinker code may not understand how many pages the 185 * ARC's callback actually frees, so it may ask the ARC to shrink a 186 * lot for one page allocation. This is problematic because it may 187 * take a long time, thus delaying the page allocation, and because 188 * it may force the ARC to unnecessarily shrink very small. 189 * 190 * Therefore, we limit the amount of data that we say is evictable, 191 * which limits the amount that the shrinker will ask us to evict for 192 * one page allocation attempt. 193 * 194 * In practice, we may be asked to shrink 4x the limit to satisfy one 195 * page allocation, before the kernel's shrinker code gives up on us. 196 * When that happens, we rely on the kernel code to find the pages 197 * that we freed before invoking the OOM killer. This happens in 198 * __alloc_pages_slowpath(), which retries and finds the pages we 199 * freed when it calls get_page_from_freelist(). 200 * 201 * See also the comment above zfs_arc_shrinker_limit. 202 */ 203 int64_t can_free = btop(arc_evictable_memory()); 204 int64_t limit = zfs_arc_shrinker_limit != 0 ? 205 zfs_arc_shrinker_limit : INT64_MAX; 206 return (MIN(can_free, limit)); 207 } 208 209 static unsigned long 210 arc_shrinker_scan(struct shrinker *shrink, struct shrink_control *sc) 211 { 212 /* The arc is considered warm once reclaim has occurred */ 213 if (unlikely(arc_warm == B_FALSE)) 214 arc_warm = B_TRUE; 215 216 /* 217 * We are experiencing memory pressure which the arc_evict_zthr was 218 * unable to keep up with. Set arc_no_grow to briefly pause ARC 219 * growth to avoid compounding the memory pressure. 220 */ 221 arc_no_grow = B_TRUE; 222 223 /* 224 * Evict the requested number of pages by reducing arc_c and waiting 225 * for the requested amount of data to be evicted. To avoid deadlock 226 * do not wait for eviction if we may be called from ZFS itself (see 227 * kmem_flags_convert() removing __GFP_FS). It may cause excessive 228 * eviction later if many evictions are accumulated, but just skipping 229 * the eviction is not good either if most of memory is used by ARC. 230 */ 231 uint64_t to_free = arc_reduce_target_size(ptob(sc->nr_to_scan)); 232 if (sc->gfp_mask & __GFP_FS) 233 arc_wait_for_eviction(to_free, B_FALSE, B_FALSE); 234 if (current->reclaim_state != NULL) 235 #ifdef HAVE_RECLAIM_STATE_RECLAIMED 236 current->reclaim_state->reclaimed += btop(to_free); 237 #else 238 current->reclaim_state->reclaimed_slab += btop(to_free); 239 #endif 240 241 /* 242 * When direct reclaim is observed it usually indicates a rapid 243 * increase in memory pressure. This occurs because the kswapd 244 * threads were unable to asynchronously keep enough free memory 245 * available. 246 */ 247 if (current_is_kswapd()) { 248 ARCSTAT_BUMP(arcstat_memory_indirect_count); 249 } else { 250 ARCSTAT_BUMP(arcstat_memory_direct_count); 251 } 252 253 return (btop(to_free)); 254 } 255 256 static struct shrinker *arc_shrinker = NULL; 257 258 int 259 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg) 260 { 261 uint64_t free_memory = arc_free_memory(); 262 263 if (free_memory > arc_all_memory() * arc_lotsfree_percent / 100) 264 return (0); 265 266 if (txg > spa->spa_lowmem_last_txg) { 267 spa->spa_lowmem_last_txg = txg; 268 spa->spa_lowmem_page_load = 0; 269 } 270 /* 271 * If we are in pageout, we know that memory is already tight, 272 * the arc is already going to be evicting, so we just want to 273 * continue to let page writes occur as quickly as possible. 274 */ 275 if (current_is_kswapd()) { 276 if (spa->spa_lowmem_page_load > 277 MAX(arc_sys_free / 4, free_memory) / 4) { 278 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim); 279 return (SET_ERROR(ERESTART)); 280 } 281 /* Note: reserve is inflated, so we deflate */ 282 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8); 283 return (0); 284 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) { 285 /* memory is low, delay before restarting */ 286 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 287 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim); 288 return (SET_ERROR(EAGAIN)); 289 } 290 spa->spa_lowmem_page_load = 0; 291 return (0); 292 } 293 294 static void 295 arc_set_sys_free(uint64_t allmem) 296 { 297 /* 298 * The ARC tries to keep at least this much memory available for the 299 * system. This gives the ARC time to shrink in response to memory 300 * pressure, before running completely out of memory and invoking the 301 * direct-reclaim ARC shrinker. 302 * 303 * This should be more than twice high_wmark_pages(), so that 304 * arc_wait_for_eviction() will wait until at least the 305 * high_wmark_pages() are free (see arc_evict_state_impl()). 306 * 307 * Note: If concurrent allocations consume these pages, there may 308 * still be insufficient free pages, and the OOM killer takes action. 309 * 310 * By setting arc_sys_free large enough, and having 311 * arc_wait_for_eviction() wait until there is at least arc_sys_free/2 312 * free memory, it is much less likely that concurrent allocations can 313 * consume all the memory that was evicted before checking for 314 * OOM. 315 * 316 * It's hard to iterate the zones from a linux kernel module, which 317 * makes it difficult to determine the watermark dynamically. Instead 318 * we compute the maximum high watermark for this system, based 319 * on the amount of memory, using the same method as the kernel uses 320 * to calculate its internal `min_free_kbytes` variable. See 321 * torvalds/linux@ee8eb9a5fe86 for the change in the upper clamp value 322 * from 64M to 256M. 323 */ 324 325 /* 326 * Base wmark_low is 4 * the square root of Kbytes of RAM. 327 */ 328 long wmark = int_sqrt(allmem / 1024 * 16) * 1024; 329 330 /* 331 * Clamp to between 128K and 256/64MB. 332 */ 333 wmark = MAX(wmark, 128 * 1024); 334 #if LINUX_VERSION_CODE >= KERNEL_VERSION(5, 7, 0) 335 wmark = MIN(wmark, 256 * 1024 * 1024); 336 #else 337 wmark = MIN(wmark, 64 * 1024 * 1024); 338 #endif 339 340 /* 341 * watermark_boost can increase the wmark by up to 150%. 342 */ 343 wmark += wmark * 150 / 100; 344 345 /* 346 * arc_sys_free needs to be more than 2x the watermark, because 347 * arc_wait_for_eviction() waits for half of arc_sys_free. Bump this up 348 * to 3x to ensure we're above it. 349 */ 350 arc_sys_free = wmark * 3 + allmem / 32; 351 } 352 353 void 354 arc_lowmem_init(void) 355 { 356 uint64_t allmem = arc_all_memory(); 357 358 /* 359 * Register a shrinker to support synchronous (direct) memory 360 * reclaim from the arc. This is done to prevent kswapd from 361 * swapping out pages when it is preferable to shrink the arc. 362 */ 363 arc_shrinker = spl_register_shrinker("zfs-arc-shrinker", 364 arc_shrinker_count, arc_shrinker_scan, zfs_arc_shrinker_seeks); 365 VERIFY(arc_shrinker); 366 367 arc_set_sys_free(allmem); 368 } 369 370 void 371 arc_lowmem_fini(void) 372 { 373 spl_unregister_shrinker(arc_shrinker); 374 arc_shrinker = NULL; 375 } 376 377 int 378 param_set_arc_u64(const char *buf, zfs_kernel_param_t *kp) 379 { 380 int error; 381 382 error = spl_param_set_u64(buf, kp); 383 if (error < 0) 384 return (SET_ERROR(error)); 385 386 arc_tuning_update(B_TRUE); 387 388 return (0); 389 } 390 391 int 392 param_set_arc_min(const char *buf, zfs_kernel_param_t *kp) 393 { 394 return (param_set_arc_u64(buf, kp)); 395 } 396 397 int 398 param_set_arc_max(const char *buf, zfs_kernel_param_t *kp) 399 { 400 return (param_set_arc_u64(buf, kp)); 401 } 402 403 int 404 param_set_arc_int(const char *buf, zfs_kernel_param_t *kp) 405 { 406 int error; 407 408 error = param_set_int(buf, kp); 409 if (error < 0) 410 return (SET_ERROR(error)); 411 412 arc_tuning_update(B_TRUE); 413 414 return (0); 415 } 416 417 #ifdef CONFIG_MEMORY_HOTPLUG 418 static int 419 arc_hotplug_callback(struct notifier_block *self, unsigned long action, 420 void *arg) 421 { 422 (void) self, (void) arg; 423 uint64_t allmem = arc_all_memory(); 424 if (action != MEM_ONLINE) 425 return (NOTIFY_OK); 426 427 arc_set_limits(allmem); 428 429 #ifdef __LP64__ 430 if (zfs_dirty_data_max_max == 0) 431 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024, 432 allmem * zfs_dirty_data_max_max_percent / 100); 433 #else 434 if (zfs_dirty_data_max_max == 0) 435 zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024, 436 allmem * zfs_dirty_data_max_max_percent / 100); 437 #endif 438 439 arc_set_sys_free(allmem); 440 return (NOTIFY_OK); 441 } 442 #endif 443 444 void 445 arc_register_hotplug(void) 446 { 447 #ifdef CONFIG_MEMORY_HOTPLUG 448 arc_hotplug_callback_mem_nb.notifier_call = arc_hotplug_callback; 449 /* There is no significance to the value 100 */ 450 arc_hotplug_callback_mem_nb.priority = 100; 451 register_memory_notifier(&arc_hotplug_callback_mem_nb); 452 #endif 453 } 454 455 void 456 arc_unregister_hotplug(void) 457 { 458 #ifdef CONFIG_MEMORY_HOTPLUG 459 unregister_memory_notifier(&arc_hotplug_callback_mem_nb); 460 #endif 461 } 462 #else /* _KERNEL */ 463 int64_t 464 arc_available_memory(void) 465 { 466 int64_t lowest = INT64_MAX; 467 468 /* Every 100 calls, free a small amount */ 469 if (random_in_range(100) == 0) 470 lowest = -1024; 471 472 return (lowest); 473 } 474 475 int 476 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg) 477 { 478 (void) spa, (void) reserve, (void) txg; 479 return (0); 480 } 481 482 uint64_t 483 arc_all_memory(void) 484 { 485 return (ptob(physmem) / 2); 486 } 487 488 uint64_t 489 arc_free_memory(void) 490 { 491 return (random_in_range(arc_all_memory() * 20 / 100)); 492 } 493 494 void 495 arc_register_hotplug(void) 496 { 497 } 498 499 void 500 arc_unregister_hotplug(void) 501 { 502 } 503 #endif /* _KERNEL */ 504 505 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, shrinker_limit, INT, ZMOD_RW, 506 "Limit on number of pages that ARC shrinker can reclaim at once"); 507 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, shrinker_seeks, INT, ZMOD_RD, 508 "Relative cost of ARC eviction vs other kernel subsystems"); 509