| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Workingset detection |
| 4 | * |
| 5 | * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner |
| 6 | */ |
| 7 | |
| 8 | #include <linux/memcontrol.h> |
| 9 | #include <linux/mm_inline.h> |
| 10 | #include <linux/writeback.h> |
| 11 | #include <linux/shmem_fs.h> |
| 12 | #include <linux/pagemap.h> |
| 13 | #include <linux/atomic.h> |
| 14 | #include <linux/module.h> |
| 15 | #include <linux/swap.h> |
| 16 | #include <linux/dax.h> |
| 17 | #include <linux/fs.h> |
| 18 | #include <linux/mm.h> |
| 19 | #include "internal.h" |
| 20 | |
| 21 | /* |
| 22 | * Double CLOCK lists |
| 23 | * |
| 24 | * Per node, two clock lists are maintained for file pages: the |
| 25 | * inactive and the active list. Freshly faulted pages start out at |
| 26 | * the head of the inactive list and page reclaim scans pages from the |
| 27 | * tail. Pages that are accessed multiple times on the inactive list |
| 28 | * are promoted to the active list, to protect them from reclaim, |
| 29 | * whereas active pages are demoted to the inactive list when the |
| 30 | * active list grows too big. |
| 31 | * |
| 32 | * fault ------------------------+ |
| 33 | * | |
| 34 | * +--------------+ | +-------------+ |
| 35 | * reclaim <- | inactive | <-+-- demotion | active | <--+ |
| 36 | * +--------------+ +-------------+ | |
| 37 | * | | |
| 38 | * +-------------- promotion ------------------+ |
| 39 | * |
| 40 | * |
| 41 | * Access frequency and refault distance |
| 42 | * |
| 43 | * A workload is thrashing when its pages are frequently used but they |
| 44 | * are evicted from the inactive list every time before another access |
| 45 | * would have promoted them to the active list. |
| 46 | * |
| 47 | * In cases where the average access distance between thrashing pages |
| 48 | * is bigger than the size of memory there is nothing that can be |
| 49 | * done - the thrashing set could never fit into memory under any |
| 50 | * circumstance. |
| 51 | * |
| 52 | * However, the average access distance could be bigger than the |
| 53 | * inactive list, yet smaller than the size of memory. In this case, |
| 54 | * the set could fit into memory if it weren't for the currently |
| 55 | * active pages - which may be used more, hopefully less frequently: |
| 56 | * |
| 57 | * +-memory available to cache-+ |
| 58 | * | | |
| 59 | * +-inactive------+-active----+ |
| 60 | * a b | c d e f g h i | J K L M N | |
| 61 | * +---------------+-----------+ |
| 62 | * |
| 63 | * It is prohibitively expensive to accurately track access frequency |
| 64 | * of pages. But a reasonable approximation can be made to measure |
| 65 | * thrashing on the inactive list, after which refaulting pages can be |
| 66 | * activated optimistically to compete with the existing active pages. |
| 67 | * |
| 68 | * Approximating inactive page access frequency - Observations: |
| 69 | * |
| 70 | * 1. When a page is accessed for the first time, it is added to the |
| 71 | * head of the inactive list, slides every existing inactive page |
| 72 | * towards the tail by one slot, and pushes the current tail page |
| 73 | * out of memory. |
| 74 | * |
| 75 | * 2. When a page is accessed for the second time, it is promoted to |
| 76 | * the active list, shrinking the inactive list by one slot. This |
| 77 | * also slides all inactive pages that were faulted into the cache |
| 78 | * more recently than the activated page towards the tail of the |
| 79 | * inactive list. |
| 80 | * |
| 81 | * Thus: |
| 82 | * |
| 83 | * 1. The sum of evictions and activations between any two points in |
| 84 | * time indicate the minimum number of inactive pages accessed in |
| 85 | * between. |
| 86 | * |
| 87 | * 2. Moving one inactive page N page slots towards the tail of the |
| 88 | * list requires at least N inactive page accesses. |
| 89 | * |
| 90 | * Combining these: |
| 91 | * |
| 92 | * 1. When a page is finally evicted from memory, the number of |
| 93 | * inactive pages accessed while the page was in cache is at least |
| 94 | * the number of page slots on the inactive list. |
| 95 | * |
| 96 | * 2. In addition, measuring the sum of evictions and activations (E) |
| 97 | * at the time of a page's eviction, and comparing it to another |
| 98 | * reading (R) at the time the page faults back into memory tells |
| 99 | * the minimum number of accesses while the page was not cached. |
| 100 | * This is called the refault distance. |
| 101 | * |
| 102 | * Because the first access of the page was the fault and the second |
| 103 | * access the refault, we combine the in-cache distance with the |
| 104 | * out-of-cache distance to get the complete minimum access distance |
| 105 | * of this page: |
| 106 | * |
| 107 | * NR_inactive + (R - E) |
| 108 | * |
| 109 | * And knowing the minimum access distance of a page, we can easily |
| 110 | * tell if the page would be able to stay in cache assuming all page |
| 111 | * slots in the cache were available: |
| 112 | * |
| 113 | * NR_inactive + (R - E) <= NR_inactive + NR_active |
| 114 | * |
| 115 | * If we have swap we should consider about NR_inactive_anon and |
| 116 | * NR_active_anon, so for page cache and anonymous respectively: |
| 117 | * |
| 118 | * NR_inactive_file + (R - E) <= NR_inactive_file + NR_active_file |
| 119 | * + NR_inactive_anon + NR_active_anon |
| 120 | * |
| 121 | * NR_inactive_anon + (R - E) <= NR_inactive_anon + NR_active_anon |
| 122 | * + NR_inactive_file + NR_active_file |
| 123 | * |
| 124 | * Which can be further simplified to: |
| 125 | * |
| 126 | * (R - E) <= NR_active_file + NR_inactive_anon + NR_active_anon |
| 127 | * |
| 128 | * (R - E) <= NR_active_anon + NR_inactive_file + NR_active_file |
| 129 | * |
| 130 | * Put into words, the refault distance (out-of-cache) can be seen as |
| 131 | * a deficit in inactive list space (in-cache). If the inactive list |
| 132 | * had (R - E) more page slots, the page would not have been evicted |
| 133 | * in between accesses, but activated instead. And on a full system, |
| 134 | * the only thing eating into inactive list space is active pages. |
| 135 | * |
| 136 | * |
| 137 | * Refaulting inactive pages |
| 138 | * |
| 139 | * All that is known about the active list is that the pages have been |
| 140 | * accessed more than once in the past. This means that at any given |
| 141 | * time there is actually a good chance that pages on the active list |
| 142 | * are no longer in active use. |
| 143 | * |
| 144 | * So when a refault distance of (R - E) is observed and there are at |
| 145 | * least (R - E) pages in the userspace workingset, the refaulting page |
| 146 | * is activated optimistically in the hope that (R - E) pages are actually |
| 147 | * used less frequently than the refaulting page - or even not used at |
| 148 | * all anymore. |
| 149 | * |
| 150 | * That means if inactive cache is refaulting with a suitable refault |
| 151 | * distance, we assume the cache workingset is transitioning and put |
| 152 | * pressure on the current workingset. |
| 153 | * |
| 154 | * If this is wrong and demotion kicks in, the pages which are truly |
| 155 | * used more frequently will be reactivated while the less frequently |
| 156 | * used once will be evicted from memory. |
| 157 | * |
| 158 | * But if this is right, the stale pages will be pushed out of memory |
| 159 | * and the used pages get to stay in cache. |
| 160 | * |
| 161 | * Refaulting active pages |
| 162 | * |
| 163 | * If on the other hand the refaulting pages have recently been |
| 164 | * deactivated, it means that the active list is no longer protecting |
| 165 | * actively used cache from reclaim. The cache is NOT transitioning to |
| 166 | * a different workingset; the existing workingset is thrashing in the |
| 167 | * space allocated to the page cache. |
| 168 | * |
| 169 | * |
| 170 | * Implementation |
| 171 | * |
| 172 | * For each node's LRU lists, a counter for inactive evictions and |
| 173 | * activations is maintained (node->nonresident_age). |
| 174 | * |
| 175 | * On eviction, a snapshot of this counter (along with some bits to |
| 176 | * identify the node) is stored in the now empty page cache |
| 177 | * slot of the evicted page. This is called a shadow entry. |
| 178 | * |
| 179 | * On cache misses for which there are shadow entries, an eligible |
| 180 | * refault distance will immediately activate the refaulting page. |
| 181 | */ |
| 182 | |
| 183 | #define WORKINGSET_SHIFT 1 |
| 184 | #define EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \ |
| 185 | WORKINGSET_SHIFT + NODES_SHIFT + \ |
| 186 | MEM_CGROUP_ID_SHIFT) |
| 187 | #define EVICTION_MASK (~0UL >> EVICTION_SHIFT) |
| 188 | |
| 189 | /* |
| 190 | * Eviction timestamps need to be able to cover the full range of |
| 191 | * actionable refaults. However, bits are tight in the xarray |
| 192 | * entry, and after storing the identifier for the lruvec there might |
| 193 | * not be enough left to represent every single actionable refault. In |
| 194 | * that case, we have to sacrifice granularity for distance, and group |
| 195 | * evictions into coarser buckets by shaving off lower timestamp bits. |
| 196 | */ |
| 197 | static unsigned int bucket_order __read_mostly; |
| 198 | |
| 199 | static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction, |
| 200 | bool workingset) |
| 201 | { |
| 202 | eviction &= EVICTION_MASK; |
| 203 | eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid; |
| 204 | eviction = (eviction << NODES_SHIFT) | pgdat->node_id; |
| 205 | eviction = (eviction << WORKINGSET_SHIFT) | workingset; |
| 206 | |
| 207 | return xa_mk_value(eviction); |
| 208 | } |
| 209 | |
| 210 | static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat, |
| 211 | unsigned long *evictionp, bool *workingsetp) |
| 212 | { |
| 213 | unsigned long entry = xa_to_value(shadow); |
| 214 | int memcgid, nid; |
| 215 | bool workingset; |
| 216 | |
| 217 | workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1); |
| 218 | entry >>= WORKINGSET_SHIFT; |
| 219 | nid = entry & ((1UL << NODES_SHIFT) - 1); |
| 220 | entry >>= NODES_SHIFT; |
| 221 | memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1); |
| 222 | entry >>= MEM_CGROUP_ID_SHIFT; |
| 223 | |
| 224 | *memcgidp = memcgid; |
| 225 | *pgdat = NODE_DATA(nid); |
| 226 | *evictionp = entry; |
| 227 | *workingsetp = workingset; |
| 228 | } |
| 229 | |
| 230 | #ifdef CONFIG_LRU_GEN |
| 231 | |
| 232 | static void *lru_gen_eviction(struct folio *folio) |
| 233 | { |
| 234 | int hist; |
| 235 | unsigned long token; |
| 236 | unsigned long min_seq; |
| 237 | struct lruvec *lruvec; |
| 238 | struct lru_gen_folio *lrugen; |
| 239 | int type = folio_is_file_lru(folio); |
| 240 | int delta = folio_nr_pages(folio); |
| 241 | int refs = folio_lru_refs(folio); |
| 242 | int tier = lru_tier_from_refs(refs); |
| 243 | struct mem_cgroup *memcg = folio_memcg(folio); |
| 244 | struct pglist_data *pgdat = folio_pgdat(folio); |
| 245 | |
| 246 | BUILD_BUG_ON(LRU_GEN_WIDTH + LRU_REFS_WIDTH > BITS_PER_LONG - EVICTION_SHIFT); |
| 247 | |
| 248 | lruvec = mem_cgroup_lruvec(memcg, pgdat); |
| 249 | lrugen = &lruvec->lrugen; |
| 250 | min_seq = READ_ONCE(lrugen->min_seq[type]); |
| 251 | token = (min_seq << LRU_REFS_WIDTH) | max(refs - 1, 0); |
| 252 | |
| 253 | hist = lru_hist_from_seq(min_seq); |
| 254 | atomic_long_add(delta, &lrugen->evicted[hist][type][tier]); |
| 255 | |
| 256 | return pack_shadow(mem_cgroup_id(memcg), pgdat, token, refs); |
| 257 | } |
| 258 | |
| 259 | /* |
| 260 | * Tests if the shadow entry is for a folio that was recently evicted. |
| 261 | * Fills in @lruvec, @token, @workingset with the values unpacked from shadow. |
| 262 | */ |
| 263 | static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec, |
| 264 | unsigned long *token, bool *workingset) |
| 265 | { |
| 266 | int memcg_id; |
| 267 | unsigned long min_seq; |
| 268 | struct mem_cgroup *memcg; |
| 269 | struct pglist_data *pgdat; |
| 270 | |
| 271 | unpack_shadow(shadow, &memcg_id, &pgdat, token, workingset); |
| 272 | |
| 273 | memcg = mem_cgroup_from_id(memcg_id); |
| 274 | *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
| 275 | |
| 276 | min_seq = READ_ONCE((*lruvec)->lrugen.min_seq[file]); |
| 277 | return (*token >> LRU_REFS_WIDTH) == (min_seq & (EVICTION_MASK >> LRU_REFS_WIDTH)); |
| 278 | } |
| 279 | |
| 280 | static void lru_gen_refault(struct folio *folio, void *shadow) |
| 281 | { |
| 282 | bool recent; |
| 283 | int hist, tier, refs; |
| 284 | bool workingset; |
| 285 | unsigned long token; |
| 286 | struct lruvec *lruvec; |
| 287 | struct lru_gen_folio *lrugen; |
| 288 | int type = folio_is_file_lru(folio); |
| 289 | int delta = folio_nr_pages(folio); |
| 290 | |
| 291 | rcu_read_lock(); |
| 292 | |
| 293 | recent = lru_gen_test_recent(shadow, type, &lruvec, &token, &workingset); |
| 294 | if (lruvec != folio_lruvec(folio)) |
| 295 | goto unlock; |
| 296 | |
| 297 | mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + type, delta); |
| 298 | |
| 299 | if (!recent) |
| 300 | goto unlock; |
| 301 | |
| 302 | lrugen = &lruvec->lrugen; |
| 303 | |
| 304 | hist = lru_hist_from_seq(READ_ONCE(lrugen->min_seq[type])); |
| 305 | /* see the comment in folio_lru_refs() */ |
| 306 | refs = (token & (BIT(LRU_REFS_WIDTH) - 1)) + workingset; |
| 307 | tier = lru_tier_from_refs(refs); |
| 308 | |
| 309 | atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]); |
| 310 | mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta); |
| 311 | |
| 312 | /* |
| 313 | * Count the following two cases as stalls: |
| 314 | * 1. For pages accessed through page tables, hotter pages pushed out |
| 315 | * hot pages which refaulted immediately. |
| 316 | * 2. For pages accessed multiple times through file descriptors, |
| 317 | * they would have been protected by sort_folio(). |
| 318 | */ |
| 319 | if (lru_gen_in_fault() || refs >= BIT(LRU_REFS_WIDTH) - 1) { |
| 320 | set_mask_bits(&folio->flags, 0, LRU_REFS_MASK | BIT(PG_workingset)); |
| 321 | mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + type, delta); |
| 322 | } |
| 323 | unlock: |
| 324 | rcu_read_unlock(); |
| 325 | } |
| 326 | |
| 327 | #else /* !CONFIG_LRU_GEN */ |
| 328 | |
| 329 | static void *lru_gen_eviction(struct folio *folio) |
| 330 | { |
| 331 | return NULL; |
| 332 | } |
| 333 | |
| 334 | static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec, |
| 335 | unsigned long *token, bool *workingset) |
| 336 | { |
| 337 | return false; |
| 338 | } |
| 339 | |
| 340 | static void lru_gen_refault(struct folio *folio, void *shadow) |
| 341 | { |
| 342 | } |
| 343 | |
| 344 | #endif /* CONFIG_LRU_GEN */ |
| 345 | |
| 346 | /** |
| 347 | * workingset_age_nonresident - age non-resident entries as LRU ages |
| 348 | * @lruvec: the lruvec that was aged |
| 349 | * @nr_pages: the number of pages to count |
| 350 | * |
| 351 | * As in-memory pages are aged, non-resident pages need to be aged as |
| 352 | * well, in order for the refault distances later on to be comparable |
| 353 | * to the in-memory dimensions. This function allows reclaim and LRU |
| 354 | * operations to drive the non-resident aging along in parallel. |
| 355 | */ |
| 356 | void workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages) |
| 357 | { |
| 358 | /* |
| 359 | * Reclaiming a cgroup means reclaiming all its children in a |
| 360 | * round-robin fashion. That means that each cgroup has an LRU |
| 361 | * order that is composed of the LRU orders of its child |
| 362 | * cgroups; and every page has an LRU position not just in the |
| 363 | * cgroup that owns it, but in all of that group's ancestors. |
| 364 | * |
| 365 | * So when the physical inactive list of a leaf cgroup ages, |
| 366 | * the virtual inactive lists of all its parents, including |
| 367 | * the root cgroup's, age as well. |
| 368 | */ |
| 369 | do { |
| 370 | atomic_long_add(nr_pages, &lruvec->nonresident_age); |
| 371 | } while ((lruvec = parent_lruvec(lruvec))); |
| 372 | } |
| 373 | |
| 374 | /** |
| 375 | * workingset_eviction - note the eviction of a folio from memory |
| 376 | * @target_memcg: the cgroup that is causing the reclaim |
| 377 | * @folio: the folio being evicted |
| 378 | * |
| 379 | * Return: a shadow entry to be stored in @folio->mapping->i_pages in place |
| 380 | * of the evicted @folio so that a later refault can be detected. |
| 381 | */ |
| 382 | void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg) |
| 383 | { |
| 384 | struct pglist_data *pgdat = folio_pgdat(folio); |
| 385 | unsigned long eviction; |
| 386 | struct lruvec *lruvec; |
| 387 | int memcgid; |
| 388 | |
| 389 | /* Folio is fully exclusive and pins folio's memory cgroup pointer */ |
| 390 | VM_BUG_ON_FOLIO(folio_test_lru(folio), folio); |
| 391 | VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); |
| 392 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
| 393 | |
| 394 | if (lru_gen_enabled()) |
| 395 | return lru_gen_eviction(folio); |
| 396 | |
| 397 | lruvec = mem_cgroup_lruvec(target_memcg, pgdat); |
| 398 | /* XXX: target_memcg can be NULL, go through lruvec */ |
| 399 | memcgid = mem_cgroup_id(lruvec_memcg(lruvec)); |
| 400 | eviction = atomic_long_read(&lruvec->nonresident_age); |
| 401 | eviction >>= bucket_order; |
| 402 | workingset_age_nonresident(lruvec, folio_nr_pages(folio)); |
| 403 | return pack_shadow(memcgid, pgdat, eviction, |
| 404 | folio_test_workingset(folio)); |
| 405 | } |
| 406 | |
| 407 | /** |
| 408 | * workingset_test_recent - tests if the shadow entry is for a folio that was |
| 409 | * recently evicted. Also fills in @workingset with the value unpacked from |
| 410 | * shadow. |
| 411 | * @shadow: the shadow entry to be tested. |
| 412 | * @file: whether the corresponding folio is from the file lru. |
| 413 | * @workingset: where the workingset value unpacked from shadow should |
| 414 | * be stored. |
| 415 | * @flush: whether to flush cgroup rstat. |
| 416 | * |
| 417 | * Return: true if the shadow is for a recently evicted folio; false otherwise. |
| 418 | */ |
| 419 | bool workingset_test_recent(void *shadow, bool file, bool *workingset, |
| 420 | bool flush) |
| 421 | { |
| 422 | struct mem_cgroup *eviction_memcg; |
| 423 | struct lruvec *eviction_lruvec; |
| 424 | unsigned long refault_distance; |
| 425 | unsigned long workingset_size; |
| 426 | unsigned long refault; |
| 427 | int memcgid; |
| 428 | struct pglist_data *pgdat; |
| 429 | unsigned long eviction; |
| 430 | |
| 431 | rcu_read_lock(); |
| 432 | |
| 433 | if (lru_gen_enabled()) { |
| 434 | bool recent = lru_gen_test_recent(shadow, file, |
| 435 | &eviction_lruvec, &eviction, workingset); |
| 436 | |
| 437 | rcu_read_unlock(); |
| 438 | return recent; |
| 439 | } |
| 440 | |
| 441 | |
| 442 | unpack_shadow(shadow, &memcgid, &pgdat, &eviction, workingset); |
| 443 | eviction <<= bucket_order; |
| 444 | |
| 445 | /* |
| 446 | * Look up the memcg associated with the stored ID. It might |
| 447 | * have been deleted since the folio's eviction. |
| 448 | * |
| 449 | * Note that in rare events the ID could have been recycled |
| 450 | * for a new cgroup that refaults a shared folio. This is |
| 451 | * impossible to tell from the available data. However, this |
| 452 | * should be a rare and limited disturbance, and activations |
| 453 | * are always speculative anyway. Ultimately, it's the aging |
| 454 | * algorithm's job to shake out the minimum access frequency |
| 455 | * for the active cache. |
| 456 | * |
| 457 | * XXX: On !CONFIG_MEMCG, this will always return NULL; it |
| 458 | * would be better if the root_mem_cgroup existed in all |
| 459 | * configurations instead. |
| 460 | */ |
| 461 | eviction_memcg = mem_cgroup_from_id(memcgid); |
| 462 | if (!mem_cgroup_disabled() && |
| 463 | (!eviction_memcg || !mem_cgroup_tryget(eviction_memcg))) { |
| 464 | rcu_read_unlock(); |
| 465 | return false; |
| 466 | } |
| 467 | |
| 468 | rcu_read_unlock(); |
| 469 | |
| 470 | /* |
| 471 | * Flush stats (and potentially sleep) outside the RCU read section. |
| 472 | * |
| 473 | * Note that workingset_test_recent() itself might be called in RCU read |
| 474 | * section (for e.g, in cachestat) - these callers need to skip flushing |
| 475 | * stats (via the flush argument). |
| 476 | * |
| 477 | * XXX: With per-memcg flushing and thresholding, is ratelimiting |
| 478 | * still needed here? |
| 479 | */ |
| 480 | if (flush) |
| 481 | mem_cgroup_flush_stats_ratelimited(eviction_memcg); |
| 482 | |
| 483 | eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat); |
| 484 | refault = atomic_long_read(&eviction_lruvec->nonresident_age); |
| 485 | |
| 486 | /* |
| 487 | * Calculate the refault distance |
| 488 | * |
| 489 | * The unsigned subtraction here gives an accurate distance |
| 490 | * across nonresident_age overflows in most cases. There is a |
| 491 | * special case: usually, shadow entries have a short lifetime |
| 492 | * and are either refaulted or reclaimed along with the inode |
| 493 | * before they get too old. But it is not impossible for the |
| 494 | * nonresident_age to lap a shadow entry in the field, which |
| 495 | * can then result in a false small refault distance, leading |
| 496 | * to a false activation should this old entry actually |
| 497 | * refault again. However, earlier kernels used to deactivate |
| 498 | * unconditionally with *every* reclaim invocation for the |
| 499 | * longest time, so the occasional inappropriate activation |
| 500 | * leading to pressure on the active list is not a problem. |
| 501 | */ |
| 502 | refault_distance = (refault - eviction) & EVICTION_MASK; |
| 503 | |
| 504 | /* |
| 505 | * Compare the distance to the existing workingset size. We |
| 506 | * don't activate pages that couldn't stay resident even if |
| 507 | * all the memory was available to the workingset. Whether |
| 508 | * workingset competition needs to consider anon or not depends |
| 509 | * on having free swap space. |
| 510 | */ |
| 511 | workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE); |
| 512 | if (!file) { |
| 513 | workingset_size += lruvec_page_state(eviction_lruvec, |
| 514 | NR_INACTIVE_FILE); |
| 515 | } |
| 516 | if (mem_cgroup_get_nr_swap_pages(eviction_memcg) > 0) { |
| 517 | workingset_size += lruvec_page_state(eviction_lruvec, |
| 518 | NR_ACTIVE_ANON); |
| 519 | if (file) { |
| 520 | workingset_size += lruvec_page_state(eviction_lruvec, |
| 521 | NR_INACTIVE_ANON); |
| 522 | } |
| 523 | } |
| 524 | |
| 525 | mem_cgroup_put(eviction_memcg); |
| 526 | return refault_distance <= workingset_size; |
| 527 | } |
| 528 | |
| 529 | /** |
| 530 | * workingset_refault - Evaluate the refault of a previously evicted folio. |
| 531 | * @folio: The freshly allocated replacement folio. |
| 532 | * @shadow: Shadow entry of the evicted folio. |
| 533 | * |
| 534 | * Calculates and evaluates the refault distance of the previously |
| 535 | * evicted folio in the context of the node and the memcg whose memory |
| 536 | * pressure caused the eviction. |
| 537 | */ |
| 538 | void workingset_refault(struct folio *folio, void *shadow) |
| 539 | { |
| 540 | bool file = folio_is_file_lru(folio); |
| 541 | struct pglist_data *pgdat; |
| 542 | struct mem_cgroup *memcg; |
| 543 | struct lruvec *lruvec; |
| 544 | bool workingset; |
| 545 | long nr; |
| 546 | |
| 547 | if (lru_gen_enabled()) { |
| 548 | lru_gen_refault(folio, shadow); |
| 549 | return; |
| 550 | } |
| 551 | |
| 552 | /* |
| 553 | * The activation decision for this folio is made at the level |
| 554 | * where the eviction occurred, as that is where the LRU order |
| 555 | * during folio reclaim is being determined. |
| 556 | * |
| 557 | * However, the cgroup that will own the folio is the one that |
| 558 | * is actually experiencing the refault event. Make sure the folio is |
| 559 | * locked to guarantee folio_memcg() stability throughout. |
| 560 | */ |
| 561 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
| 562 | nr = folio_nr_pages(folio); |
| 563 | memcg = folio_memcg(folio); |
| 564 | pgdat = folio_pgdat(folio); |
| 565 | lruvec = mem_cgroup_lruvec(memcg, pgdat); |
| 566 | |
| 567 | mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr); |
| 568 | |
| 569 | if (!workingset_test_recent(shadow, file, &workingset, true)) |
| 570 | return; |
| 571 | |
| 572 | folio_set_active(folio); |
| 573 | workingset_age_nonresident(lruvec, nr); |
| 574 | mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr); |
| 575 | |
| 576 | /* Folio was active prior to eviction */ |
| 577 | if (workingset) { |
| 578 | folio_set_workingset(folio); |
| 579 | /* |
| 580 | * XXX: Move to folio_add_lru() when it supports new vs |
| 581 | * putback |
| 582 | */ |
| 583 | lru_note_cost_refault(folio); |
| 584 | mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr); |
| 585 | } |
| 586 | } |
| 587 | |
| 588 | /** |
| 589 | * workingset_activation - note a page activation |
| 590 | * @folio: Folio that is being activated. |
| 591 | */ |
| 592 | void workingset_activation(struct folio *folio) |
| 593 | { |
| 594 | struct mem_cgroup *memcg; |
| 595 | |
| 596 | rcu_read_lock(); |
| 597 | /* |
| 598 | * Filter non-memcg pages here, e.g. unmap can call |
| 599 | * mark_page_accessed() on VDSO pages. |
| 600 | * |
| 601 | * XXX: See workingset_refault() - this should return |
| 602 | * root_mem_cgroup even for !CONFIG_MEMCG. |
| 603 | */ |
| 604 | memcg = folio_memcg_rcu(folio); |
| 605 | if (!mem_cgroup_disabled() && !memcg) |
| 606 | goto out; |
| 607 | workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio)); |
| 608 | out: |
| 609 | rcu_read_unlock(); |
| 610 | } |
| 611 | |
| 612 | /* |
| 613 | * Shadow entries reflect the share of the working set that does not |
| 614 | * fit into memory, so their number depends on the access pattern of |
| 615 | * the workload. In most cases, they will refault or get reclaimed |
| 616 | * along with the inode, but a (malicious) workload that streams |
| 617 | * through files with a total size several times that of available |
| 618 | * memory, while preventing the inodes from being reclaimed, can |
| 619 | * create excessive amounts of shadow nodes. To keep a lid on this, |
| 620 | * track shadow nodes and reclaim them when they grow way past the |
| 621 | * point where they would still be useful. |
| 622 | */ |
| 623 | |
| 624 | struct list_lru shadow_nodes; |
| 625 | |
| 626 | void workingset_update_node(struct xa_node *node) |
| 627 | { |
| 628 | struct address_space *mapping; |
| 629 | struct page *page = virt_to_page(node); |
| 630 | |
| 631 | /* |
| 632 | * Track non-empty nodes that contain only shadow entries; |
| 633 | * unlink those that contain pages or are being freed. |
| 634 | * |
| 635 | * Avoid acquiring the list_lru lock when the nodes are |
| 636 | * already where they should be. The list_empty() test is safe |
| 637 | * as node->private_list is protected by the i_pages lock. |
| 638 | */ |
| 639 | mapping = container_of(node->array, struct address_space, i_pages); |
| 640 | lockdep_assert_held(&mapping->i_pages.xa_lock); |
| 641 | |
| 642 | if (node->count && node->count == node->nr_values) { |
| 643 | if (list_empty(&node->private_list)) { |
| 644 | list_lru_add_obj(&shadow_nodes, &node->private_list); |
| 645 | __inc_node_page_state(page, WORKINGSET_NODES); |
| 646 | } |
| 647 | } else { |
| 648 | if (!list_empty(&node->private_list)) { |
| 649 | list_lru_del_obj(&shadow_nodes, &node->private_list); |
| 650 | __dec_node_page_state(page, WORKINGSET_NODES); |
| 651 | } |
| 652 | } |
| 653 | } |
| 654 | |
| 655 | static unsigned long count_shadow_nodes(struct shrinker *shrinker, |
| 656 | struct shrink_control *sc) |
| 657 | { |
| 658 | unsigned long max_nodes; |
| 659 | unsigned long nodes; |
| 660 | unsigned long pages; |
| 661 | |
| 662 | nodes = list_lru_shrink_count(&shadow_nodes, sc); |
| 663 | if (!nodes) |
| 664 | return SHRINK_EMPTY; |
| 665 | |
| 666 | /* |
| 667 | * Approximate a reasonable limit for the nodes |
| 668 | * containing shadow entries. We don't need to keep more |
| 669 | * shadow entries than possible pages on the active list, |
| 670 | * since refault distances bigger than that are dismissed. |
| 671 | * |
| 672 | * The size of the active list converges toward 100% of |
| 673 | * overall page cache as memory grows, with only a tiny |
| 674 | * inactive list. Assume the total cache size for that. |
| 675 | * |
| 676 | * Nodes might be sparsely populated, with only one shadow |
| 677 | * entry in the extreme case. Obviously, we cannot keep one |
| 678 | * node for every eligible shadow entry, so compromise on a |
| 679 | * worst-case density of 1/8th. Below that, not all eligible |
| 680 | * refaults can be detected anymore. |
| 681 | * |
| 682 | * On 64-bit with 7 xa_nodes per page and 64 slots |
| 683 | * each, this will reclaim shadow entries when they consume |
| 684 | * ~1.8% of available memory: |
| 685 | * |
| 686 | * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE |
| 687 | */ |
| 688 | #ifdef CONFIG_MEMCG |
| 689 | if (sc->memcg) { |
| 690 | struct lruvec *lruvec; |
| 691 | int i; |
| 692 | |
| 693 | mem_cgroup_flush_stats_ratelimited(sc->memcg); |
| 694 | lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid)); |
| 695 | for (pages = 0, i = 0; i < NR_LRU_LISTS; i++) |
| 696 | pages += lruvec_page_state_local(lruvec, |
| 697 | NR_LRU_BASE + i); |
| 698 | pages += lruvec_page_state_local( |
| 699 | lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT; |
| 700 | pages += lruvec_page_state_local( |
| 701 | lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT; |
| 702 | } else |
| 703 | #endif |
| 704 | pages = node_present_pages(sc->nid); |
| 705 | |
| 706 | max_nodes = pages >> (XA_CHUNK_SHIFT - 3); |
| 707 | |
| 708 | if (nodes <= max_nodes) |
| 709 | return 0; |
| 710 | return nodes - max_nodes; |
| 711 | } |
| 712 | |
| 713 | static enum lru_status shadow_lru_isolate(struct list_head *item, |
| 714 | struct list_lru_one *lru, |
| 715 | spinlock_t *lru_lock, |
| 716 | void *arg) __must_hold(lru_lock) |
| 717 | { |
| 718 | struct xa_node *node = container_of(item, struct xa_node, private_list); |
| 719 | struct address_space *mapping; |
| 720 | int ret; |
| 721 | |
| 722 | /* |
| 723 | * Page cache insertions and deletions synchronously maintain |
| 724 | * the shadow node LRU under the i_pages lock and the |
| 725 | * lru_lock. Because the page cache tree is emptied before |
| 726 | * the inode can be destroyed, holding the lru_lock pins any |
| 727 | * address_space that has nodes on the LRU. |
| 728 | * |
| 729 | * We can then safely transition to the i_pages lock to |
| 730 | * pin only the address_space of the particular node we want |
| 731 | * to reclaim, take the node off-LRU, and drop the lru_lock. |
| 732 | */ |
| 733 | |
| 734 | mapping = container_of(node->array, struct address_space, i_pages); |
| 735 | |
| 736 | /* Coming from the list, invert the lock order */ |
| 737 | if (!xa_trylock(&mapping->i_pages)) { |
| 738 | spin_unlock_irq(lru_lock); |
| 739 | ret = LRU_RETRY; |
| 740 | goto out; |
| 741 | } |
| 742 | |
| 743 | /* For page cache we need to hold i_lock */ |
| 744 | if (mapping->host != NULL) { |
| 745 | if (!spin_trylock(&mapping->host->i_lock)) { |
| 746 | xa_unlock(&mapping->i_pages); |
| 747 | spin_unlock_irq(lru_lock); |
| 748 | ret = LRU_RETRY; |
| 749 | goto out; |
| 750 | } |
| 751 | } |
| 752 | |
| 753 | list_lru_isolate(lru, item); |
| 754 | __dec_node_page_state(virt_to_page(node), WORKINGSET_NODES); |
| 755 | |
| 756 | spin_unlock(lru_lock); |
| 757 | |
| 758 | /* |
| 759 | * The nodes should only contain one or more shadow entries, |
| 760 | * no pages, so we expect to be able to remove them all and |
| 761 | * delete and free the empty node afterwards. |
| 762 | */ |
| 763 | if (WARN_ON_ONCE(!node->nr_values)) |
| 764 | goto out_invalid; |
| 765 | if (WARN_ON_ONCE(node->count != node->nr_values)) |
| 766 | goto out_invalid; |
| 767 | xa_delete_node(node, workingset_update_node); |
| 768 | __inc_lruvec_kmem_state(node, WORKINGSET_NODERECLAIM); |
| 769 | |
| 770 | out_invalid: |
| 771 | xa_unlock_irq(&mapping->i_pages); |
| 772 | if (mapping->host != NULL) { |
| 773 | if (mapping_shrinkable(mapping)) |
| 774 | inode_add_lru(mapping->host); |
| 775 | spin_unlock(&mapping->host->i_lock); |
| 776 | } |
| 777 | ret = LRU_REMOVED_RETRY; |
| 778 | out: |
| 779 | cond_resched(); |
| 780 | spin_lock_irq(lru_lock); |
| 781 | return ret; |
| 782 | } |
| 783 | |
| 784 | static unsigned long scan_shadow_nodes(struct shrinker *shrinker, |
| 785 | struct shrink_control *sc) |
| 786 | { |
| 787 | /* list_lru lock nests inside the IRQ-safe i_pages lock */ |
| 788 | return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate, |
| 789 | NULL); |
| 790 | } |
| 791 | |
| 792 | /* |
| 793 | * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe |
| 794 | * i_pages lock. |
| 795 | */ |
| 796 | static struct lock_class_key shadow_nodes_key; |
| 797 | |
| 798 | static int __init workingset_init(void) |
| 799 | { |
| 800 | struct shrinker *workingset_shadow_shrinker; |
| 801 | unsigned int timestamp_bits; |
| 802 | unsigned int max_order; |
| 803 | int ret = -ENOMEM; |
| 804 | |
| 805 | BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT); |
| 806 | /* |
| 807 | * Calculate the eviction bucket size to cover the longest |
| 808 | * actionable refault distance, which is currently half of |
| 809 | * memory (totalram_pages/2). However, memory hotplug may add |
| 810 | * some more pages at runtime, so keep working with up to |
| 811 | * double the initial memory by using totalram_pages as-is. |
| 812 | */ |
| 813 | timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT; |
| 814 | max_order = fls_long(totalram_pages() - 1); |
| 815 | if (max_order > timestamp_bits) |
| 816 | bucket_order = max_order - timestamp_bits; |
| 817 | pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n", |
| 818 | timestamp_bits, max_order, bucket_order); |
| 819 | |
| 820 | workingset_shadow_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE | |
| 821 | SHRINKER_MEMCG_AWARE, |
| 822 | "mm-shadow"); |
| 823 | if (!workingset_shadow_shrinker) |
| 824 | goto err; |
| 825 | |
| 826 | ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key, |
| 827 | workingset_shadow_shrinker); |
| 828 | if (ret) |
| 829 | goto err_list_lru; |
| 830 | |
| 831 | workingset_shadow_shrinker->count_objects = count_shadow_nodes; |
| 832 | workingset_shadow_shrinker->scan_objects = scan_shadow_nodes; |
| 833 | /* ->count reports only fully expendable nodes */ |
| 834 | workingset_shadow_shrinker->seeks = 0; |
| 835 | |
| 836 | shrinker_register(workingset_shadow_shrinker); |
| 837 | return 0; |
| 838 | err_list_lru: |
| 839 | shrinker_free(workingset_shadow_shrinker); |
| 840 | err: |
| 841 | return ret; |
| 842 | } |
| 843 | module_init(workingset_init); |