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