1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * linux/mm/swap.c 4 * 5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 6 */ 7 8 /* 9 * This file contains the default values for the operation of the 10 * Linux VM subsystem. Fine-tuning documentation can be found in 11 * Documentation/admin-guide/sysctl/vm.rst. 12 * Started 18.12.91 13 * Swap aging added 23.2.95, Stephen Tweedie. 14 * Buffermem limits added 12.3.98, Rik van Riel. 15 */ 16 17 #include <linux/mm.h> 18 #include <linux/sched.h> 19 #include <linux/kernel_stat.h> 20 #include <linux/swap.h> 21 #include <linux/mman.h> 22 #include <linux/pagemap.h> 23 #include <linux/pagevec.h> 24 #include <linux/init.h> 25 #include <linux/export.h> 26 #include <linux/mm_inline.h> 27 #include <linux/percpu_counter.h> 28 #include <linux/memremap.h> 29 #include <linux/percpu.h> 30 #include <linux/cpu.h> 31 #include <linux/notifier.h> 32 #include <linux/backing-dev.h> 33 #include <linux/memcontrol.h> 34 #include <linux/gfp.h> 35 #include <linux/uio.h> 36 #include <linux/hugetlb.h> 37 #include <linux/page_idle.h> 38 #include <linux/local_lock.h> 39 #include <linux/buffer_head.h> 40 41 #include "internal.h" 42 43 #define CREATE_TRACE_POINTS 44 #include <trace/events/pagemap.h> 45 46 /* How many pages do we try to swap or page in/out together? */ 47 int page_cluster; 48 49 /* Protecting only lru_rotate.fbatch which requires disabling interrupts */ 50 struct lru_rotate { 51 local_lock_t lock; 52 struct folio_batch fbatch; 53 }; 54 static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = { 55 .lock = INIT_LOCAL_LOCK(lock), 56 }; 57 58 /* 59 * The following folio batches are grouped together because they are protected 60 * by disabling preemption (and interrupts remain enabled). 61 */ 62 struct cpu_fbatches { 63 local_lock_t lock; 64 struct folio_batch lru_add; 65 struct folio_batch lru_deactivate_file; 66 struct folio_batch lru_deactivate; 67 struct folio_batch lru_lazyfree; 68 #ifdef CONFIG_SMP 69 struct folio_batch activate; 70 #endif 71 }; 72 static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = { 73 .lock = INIT_LOCAL_LOCK(lock), 74 }; 75 76 /* 77 * This path almost never happens for VM activity - pages are normally freed 78 * via pagevecs. But it gets used by networking - and for compound pages. 79 */ 80 static void __page_cache_release(struct folio *folio) 81 { 82 if (folio_test_lru(folio)) { 83 struct lruvec *lruvec; 84 unsigned long flags; 85 86 lruvec = folio_lruvec_lock_irqsave(folio, &flags); 87 lruvec_del_folio(lruvec, folio); 88 __folio_clear_lru_flags(folio); 89 unlock_page_lruvec_irqrestore(lruvec, flags); 90 } 91 /* See comment on folio_test_mlocked in release_pages() */ 92 if (unlikely(folio_test_mlocked(folio))) { 93 long nr_pages = folio_nr_pages(folio); 94 95 __folio_clear_mlocked(folio); 96 zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages); 97 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages); 98 } 99 } 100 101 static void __folio_put_small(struct folio *folio) 102 { 103 __page_cache_release(folio); 104 mem_cgroup_uncharge(folio); 105 free_unref_page(&folio->page, 0); 106 } 107 108 static void __folio_put_large(struct folio *folio) 109 { 110 /* 111 * __page_cache_release() is supposed to be called for thp, not for 112 * hugetlb. This is because hugetlb page does never have PageLRU set 113 * (it's never listed to any LRU lists) and no memcg routines should 114 * be called for hugetlb (it has a separate hugetlb_cgroup.) 115 */ 116 if (!folio_test_hugetlb(folio)) 117 __page_cache_release(folio); 118 destroy_large_folio(folio); 119 } 120 121 void __folio_put(struct folio *folio) 122 { 123 if (unlikely(folio_is_zone_device(folio))) 124 free_zone_device_page(&folio->page); 125 else if (unlikely(folio_test_large(folio))) 126 __folio_put_large(folio); 127 else 128 __folio_put_small(folio); 129 } 130 EXPORT_SYMBOL(__folio_put); 131 132 /** 133 * put_pages_list() - release a list of pages 134 * @pages: list of pages threaded on page->lru 135 * 136 * Release a list of pages which are strung together on page.lru. 137 */ 138 void put_pages_list(struct list_head *pages) 139 { 140 struct folio *folio, *next; 141 142 list_for_each_entry_safe(folio, next, pages, lru) { 143 if (!folio_put_testzero(folio)) { 144 list_del(&folio->lru); 145 continue; 146 } 147 if (folio_test_large(folio)) { 148 list_del(&folio->lru); 149 __folio_put_large(folio); 150 continue; 151 } 152 /* LRU flag must be clear because it's passed using the lru */ 153 } 154 155 free_unref_page_list(pages); 156 INIT_LIST_HEAD(pages); 157 } 158 EXPORT_SYMBOL(put_pages_list); 159 160 /* 161 * get_kernel_pages() - pin kernel pages in memory 162 * @kiov: An array of struct kvec structures 163 * @nr_segs: number of segments to pin 164 * @write: pinning for read/write, currently ignored 165 * @pages: array that receives pointers to the pages pinned. 166 * Should be at least nr_segs long. 167 * 168 * Returns number of pages pinned. This may be fewer than the number requested. 169 * If nr_segs is 0 or negative, returns 0. If no pages were pinned, returns 0. 170 * Each page returned must be released with a put_page() call when it is 171 * finished with. 172 */ 173 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write, 174 struct page **pages) 175 { 176 int seg; 177 178 for (seg = 0; seg < nr_segs; seg++) { 179 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE)) 180 return seg; 181 182 pages[seg] = kmap_to_page(kiov[seg].iov_base); 183 get_page(pages[seg]); 184 } 185 186 return seg; 187 } 188 EXPORT_SYMBOL_GPL(get_kernel_pages); 189 190 typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio); 191 192 static void lru_add_fn(struct lruvec *lruvec, struct folio *folio) 193 { 194 int was_unevictable = folio_test_clear_unevictable(folio); 195 long nr_pages = folio_nr_pages(folio); 196 197 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio); 198 199 /* 200 * Is an smp_mb__after_atomic() still required here, before 201 * folio_evictable() tests the mlocked flag, to rule out the possibility 202 * of stranding an evictable folio on an unevictable LRU? I think 203 * not, because __munlock_page() only clears the mlocked flag 204 * while the LRU lock is held. 205 * 206 * (That is not true of __page_cache_release(), and not necessarily 207 * true of release_pages(): but those only clear the mlocked flag after 208 * folio_put_testzero() has excluded any other users of the folio.) 209 */ 210 if (folio_evictable(folio)) { 211 if (was_unevictable) 212 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages); 213 } else { 214 folio_clear_active(folio); 215 folio_set_unevictable(folio); 216 /* 217 * folio->mlock_count = !!folio_test_mlocked(folio)? 218 * But that leaves __mlock_page() in doubt whether another 219 * actor has already counted the mlock or not. Err on the 220 * safe side, underestimate, let page reclaim fix it, rather 221 * than leaving a page on the unevictable LRU indefinitely. 222 */ 223 folio->mlock_count = 0; 224 if (!was_unevictable) 225 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages); 226 } 227 228 lruvec_add_folio(lruvec, folio); 229 trace_mm_lru_insertion(folio); 230 } 231 232 static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn) 233 { 234 int i; 235 struct lruvec *lruvec = NULL; 236 unsigned long flags = 0; 237 238 for (i = 0; i < folio_batch_count(fbatch); i++) { 239 struct folio *folio = fbatch->folios[i]; 240 241 /* block memcg migration while the folio moves between lru */ 242 if (move_fn != lru_add_fn && !folio_test_clear_lru(folio)) 243 continue; 244 245 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags); 246 move_fn(lruvec, folio); 247 248 folio_set_lru(folio); 249 } 250 251 if (lruvec) 252 unlock_page_lruvec_irqrestore(lruvec, flags); 253 folios_put(fbatch->folios, folio_batch_count(fbatch)); 254 folio_batch_init(fbatch); 255 } 256 257 static void folio_batch_add_and_move(struct folio_batch *fbatch, 258 struct folio *folio, move_fn_t move_fn) 259 { 260 if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) && 261 !lru_cache_disabled()) 262 return; 263 folio_batch_move_lru(fbatch, move_fn); 264 } 265 266 static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio) 267 { 268 if (!folio_test_unevictable(folio)) { 269 lruvec_del_folio(lruvec, folio); 270 folio_clear_active(folio); 271 lruvec_add_folio_tail(lruvec, folio); 272 __count_vm_events(PGROTATED, folio_nr_pages(folio)); 273 } 274 } 275 276 /* 277 * Writeback is about to end against a folio which has been marked for 278 * immediate reclaim. If it still appears to be reclaimable, move it 279 * to the tail of the inactive list. 280 * 281 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races. 282 */ 283 void folio_rotate_reclaimable(struct folio *folio) 284 { 285 if (!folio_test_locked(folio) && !folio_test_dirty(folio) && 286 !folio_test_unevictable(folio) && folio_test_lru(folio)) { 287 struct folio_batch *fbatch; 288 unsigned long flags; 289 290 folio_get(folio); 291 local_lock_irqsave(&lru_rotate.lock, flags); 292 fbatch = this_cpu_ptr(&lru_rotate.fbatch); 293 folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn); 294 local_unlock_irqrestore(&lru_rotate.lock, flags); 295 } 296 } 297 298 void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages) 299 { 300 do { 301 unsigned long lrusize; 302 303 /* 304 * Hold lruvec->lru_lock is safe here, since 305 * 1) The pinned lruvec in reclaim, or 306 * 2) From a pre-LRU page during refault (which also holds the 307 * rcu lock, so would be safe even if the page was on the LRU 308 * and could move simultaneously to a new lruvec). 309 */ 310 spin_lock_irq(&lruvec->lru_lock); 311 /* Record cost event */ 312 if (file) 313 lruvec->file_cost += nr_pages; 314 else 315 lruvec->anon_cost += nr_pages; 316 317 /* 318 * Decay previous events 319 * 320 * Because workloads change over time (and to avoid 321 * overflow) we keep these statistics as a floating 322 * average, which ends up weighing recent refaults 323 * more than old ones. 324 */ 325 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) + 326 lruvec_page_state(lruvec, NR_ACTIVE_ANON) + 327 lruvec_page_state(lruvec, NR_INACTIVE_FILE) + 328 lruvec_page_state(lruvec, NR_ACTIVE_FILE); 329 330 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) { 331 lruvec->file_cost /= 2; 332 lruvec->anon_cost /= 2; 333 } 334 spin_unlock_irq(&lruvec->lru_lock); 335 } while ((lruvec = parent_lruvec(lruvec))); 336 } 337 338 void lru_note_cost_folio(struct folio *folio) 339 { 340 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio), 341 folio_nr_pages(folio)); 342 } 343 344 static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio) 345 { 346 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) { 347 long nr_pages = folio_nr_pages(folio); 348 349 lruvec_del_folio(lruvec, folio); 350 folio_set_active(folio); 351 lruvec_add_folio(lruvec, folio); 352 trace_mm_lru_activate(folio); 353 354 __count_vm_events(PGACTIVATE, nr_pages); 355 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE, 356 nr_pages); 357 } 358 } 359 360 #ifdef CONFIG_SMP 361 static void folio_activate_drain(int cpu) 362 { 363 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu); 364 365 if (folio_batch_count(fbatch)) 366 folio_batch_move_lru(fbatch, folio_activate_fn); 367 } 368 369 void folio_activate(struct folio *folio) 370 { 371 if (folio_test_lru(folio) && !folio_test_active(folio) && 372 !folio_test_unevictable(folio)) { 373 struct folio_batch *fbatch; 374 375 folio_get(folio); 376 local_lock(&cpu_fbatches.lock); 377 fbatch = this_cpu_ptr(&cpu_fbatches.activate); 378 folio_batch_add_and_move(fbatch, folio, folio_activate_fn); 379 local_unlock(&cpu_fbatches.lock); 380 } 381 } 382 383 #else 384 static inline void folio_activate_drain(int cpu) 385 { 386 } 387 388 void folio_activate(struct folio *folio) 389 { 390 struct lruvec *lruvec; 391 392 if (folio_test_clear_lru(folio)) { 393 lruvec = folio_lruvec_lock_irq(folio); 394 folio_activate_fn(lruvec, folio); 395 unlock_page_lruvec_irq(lruvec); 396 folio_set_lru(folio); 397 } 398 } 399 #endif 400 401 static void __lru_cache_activate_folio(struct folio *folio) 402 { 403 struct folio_batch *fbatch; 404 int i; 405 406 local_lock(&cpu_fbatches.lock); 407 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add); 408 409 /* 410 * Search backwards on the optimistic assumption that the folio being 411 * activated has just been added to this batch. Note that only 412 * the local batch is examined as a !LRU folio could be in the 413 * process of being released, reclaimed, migrated or on a remote 414 * batch that is currently being drained. Furthermore, marking 415 * a remote batch's folio active potentially hits a race where 416 * a folio is marked active just after it is added to the inactive 417 * list causing accounting errors and BUG_ON checks to trigger. 418 */ 419 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) { 420 struct folio *batch_folio = fbatch->folios[i]; 421 422 if (batch_folio == folio) { 423 folio_set_active(folio); 424 break; 425 } 426 } 427 428 local_unlock(&cpu_fbatches.lock); 429 } 430 431 #ifdef CONFIG_LRU_GEN 432 static void folio_inc_refs(struct folio *folio) 433 { 434 unsigned long new_flags, old_flags = READ_ONCE(folio->flags); 435 436 if (folio_test_unevictable(folio)) 437 return; 438 439 if (!folio_test_referenced(folio)) { 440 folio_set_referenced(folio); 441 return; 442 } 443 444 if (!folio_test_workingset(folio)) { 445 folio_set_workingset(folio); 446 return; 447 } 448 449 /* see the comment on MAX_NR_TIERS */ 450 do { 451 new_flags = old_flags & LRU_REFS_MASK; 452 if (new_flags == LRU_REFS_MASK) 453 break; 454 455 new_flags += BIT(LRU_REFS_PGOFF); 456 new_flags |= old_flags & ~LRU_REFS_MASK; 457 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags)); 458 } 459 #else 460 static void folio_inc_refs(struct folio *folio) 461 { 462 } 463 #endif /* CONFIG_LRU_GEN */ 464 465 /* 466 * Mark a page as having seen activity. 467 * 468 * inactive,unreferenced -> inactive,referenced 469 * inactive,referenced -> active,unreferenced 470 * active,unreferenced -> active,referenced 471 * 472 * When a newly allocated page is not yet visible, so safe for non-atomic ops, 473 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page). 474 */ 475 void folio_mark_accessed(struct folio *folio) 476 { 477 if (lru_gen_enabled()) { 478 folio_inc_refs(folio); 479 return; 480 } 481 482 if (!folio_test_referenced(folio)) { 483 folio_set_referenced(folio); 484 } else if (folio_test_unevictable(folio)) { 485 /* 486 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But, 487 * this list is never rotated or maintained, so marking an 488 * unevictable page accessed has no effect. 489 */ 490 } else if (!folio_test_active(folio)) { 491 /* 492 * If the folio is on the LRU, queue it for activation via 493 * cpu_fbatches.activate. Otherwise, assume the folio is in a 494 * folio_batch, mark it active and it'll be moved to the active 495 * LRU on the next drain. 496 */ 497 if (folio_test_lru(folio)) 498 folio_activate(folio); 499 else 500 __lru_cache_activate_folio(folio); 501 folio_clear_referenced(folio); 502 workingset_activation(folio); 503 } 504 if (folio_test_idle(folio)) 505 folio_clear_idle(folio); 506 } 507 EXPORT_SYMBOL(folio_mark_accessed); 508 509 /** 510 * folio_add_lru - Add a folio to an LRU list. 511 * @folio: The folio to be added to the LRU. 512 * 513 * Queue the folio for addition to the LRU. The decision on whether 514 * to add the page to the [in]active [file|anon] list is deferred until the 515 * folio_batch is drained. This gives a chance for the caller of folio_add_lru() 516 * have the folio added to the active list using folio_mark_accessed(). 517 */ 518 void folio_add_lru(struct folio *folio) 519 { 520 struct folio_batch *fbatch; 521 522 VM_BUG_ON_FOLIO(folio_test_active(folio) && 523 folio_test_unevictable(folio), folio); 524 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio); 525 526 /* see the comment in lru_gen_add_folio() */ 527 if (lru_gen_enabled() && !folio_test_unevictable(folio) && 528 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC)) 529 folio_set_active(folio); 530 531 folio_get(folio); 532 local_lock(&cpu_fbatches.lock); 533 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add); 534 folio_batch_add_and_move(fbatch, folio, lru_add_fn); 535 local_unlock(&cpu_fbatches.lock); 536 } 537 EXPORT_SYMBOL(folio_add_lru); 538 539 /** 540 * lru_cache_add_inactive_or_unevictable 541 * @page: the page to be added to LRU 542 * @vma: vma in which page is mapped for determining reclaimability 543 * 544 * Place @page on the inactive or unevictable LRU list, depending on its 545 * evictability. 546 */ 547 void lru_cache_add_inactive_or_unevictable(struct page *page, 548 struct vm_area_struct *vma) 549 { 550 VM_BUG_ON_PAGE(PageLRU(page), page); 551 552 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED)) 553 mlock_new_page(page); 554 else 555 lru_cache_add(page); 556 } 557 558 /* 559 * If the folio cannot be invalidated, it is moved to the 560 * inactive list to speed up its reclaim. It is moved to the 561 * head of the list, rather than the tail, to give the flusher 562 * threads some time to write it out, as this is much more 563 * effective than the single-page writeout from reclaim. 564 * 565 * If the folio isn't mapped and dirty/writeback, the folio 566 * could be reclaimed asap using the reclaim flag. 567 * 568 * 1. active, mapped folio -> none 569 * 2. active, dirty/writeback folio -> inactive, head, reclaim 570 * 3. inactive, mapped folio -> none 571 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim 572 * 5. inactive, clean -> inactive, tail 573 * 6. Others -> none 574 * 575 * In 4, it moves to the head of the inactive list so the folio is 576 * written out by flusher threads as this is much more efficient 577 * than the single-page writeout from reclaim. 578 */ 579 static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio) 580 { 581 bool active = folio_test_active(folio); 582 long nr_pages = folio_nr_pages(folio); 583 584 if (folio_test_unevictable(folio)) 585 return; 586 587 /* Some processes are using the folio */ 588 if (folio_mapped(folio)) 589 return; 590 591 lruvec_del_folio(lruvec, folio); 592 folio_clear_active(folio); 593 folio_clear_referenced(folio); 594 595 if (folio_test_writeback(folio) || folio_test_dirty(folio)) { 596 /* 597 * Setting the reclaim flag could race with 598 * folio_end_writeback() and confuse readahead. But the 599 * race window is _really_ small and it's not a critical 600 * problem. 601 */ 602 lruvec_add_folio(lruvec, folio); 603 folio_set_reclaim(folio); 604 } else { 605 /* 606 * The folio's writeback ended while it was in the batch. 607 * We move that folio to the tail of the inactive list. 608 */ 609 lruvec_add_folio_tail(lruvec, folio); 610 __count_vm_events(PGROTATED, nr_pages); 611 } 612 613 if (active) { 614 __count_vm_events(PGDEACTIVATE, nr_pages); 615 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, 616 nr_pages); 617 } 618 } 619 620 static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio) 621 { 622 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) { 623 long nr_pages = folio_nr_pages(folio); 624 625 lruvec_del_folio(lruvec, folio); 626 folio_clear_active(folio); 627 folio_clear_referenced(folio); 628 lruvec_add_folio(lruvec, folio); 629 630 __count_vm_events(PGDEACTIVATE, nr_pages); 631 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, 632 nr_pages); 633 } 634 } 635 636 static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio) 637 { 638 if (folio_test_anon(folio) && folio_test_swapbacked(folio) && 639 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) { 640 long nr_pages = folio_nr_pages(folio); 641 642 lruvec_del_folio(lruvec, folio); 643 folio_clear_active(folio); 644 folio_clear_referenced(folio); 645 /* 646 * Lazyfree folios are clean anonymous folios. They have 647 * the swapbacked flag cleared, to distinguish them from normal 648 * anonymous folios 649 */ 650 folio_clear_swapbacked(folio); 651 lruvec_add_folio(lruvec, folio); 652 653 __count_vm_events(PGLAZYFREE, nr_pages); 654 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE, 655 nr_pages); 656 } 657 } 658 659 /* 660 * Drain pages out of the cpu's folio_batch. 661 * Either "cpu" is the current CPU, and preemption has already been 662 * disabled; or "cpu" is being hot-unplugged, and is already dead. 663 */ 664 void lru_add_drain_cpu(int cpu) 665 { 666 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu); 667 struct folio_batch *fbatch = &fbatches->lru_add; 668 669 if (folio_batch_count(fbatch)) 670 folio_batch_move_lru(fbatch, lru_add_fn); 671 672 fbatch = &per_cpu(lru_rotate.fbatch, cpu); 673 /* Disabling interrupts below acts as a compiler barrier. */ 674 if (data_race(folio_batch_count(fbatch))) { 675 unsigned long flags; 676 677 /* No harm done if a racing interrupt already did this */ 678 local_lock_irqsave(&lru_rotate.lock, flags); 679 folio_batch_move_lru(fbatch, lru_move_tail_fn); 680 local_unlock_irqrestore(&lru_rotate.lock, flags); 681 } 682 683 fbatch = &fbatches->lru_deactivate_file; 684 if (folio_batch_count(fbatch)) 685 folio_batch_move_lru(fbatch, lru_deactivate_file_fn); 686 687 fbatch = &fbatches->lru_deactivate; 688 if (folio_batch_count(fbatch)) 689 folio_batch_move_lru(fbatch, lru_deactivate_fn); 690 691 fbatch = &fbatches->lru_lazyfree; 692 if (folio_batch_count(fbatch)) 693 folio_batch_move_lru(fbatch, lru_lazyfree_fn); 694 695 folio_activate_drain(cpu); 696 } 697 698 /** 699 * deactivate_file_folio() - Deactivate a file folio. 700 * @folio: Folio to deactivate. 701 * 702 * This function hints to the VM that @folio is a good reclaim candidate, 703 * for example if its invalidation fails due to the folio being dirty 704 * or under writeback. 705 * 706 * Context: Caller holds a reference on the folio. 707 */ 708 void deactivate_file_folio(struct folio *folio) 709 { 710 struct folio_batch *fbatch; 711 712 /* Deactivating an unevictable folio will not accelerate reclaim */ 713 if (folio_test_unevictable(folio)) 714 return; 715 716 folio_get(folio); 717 local_lock(&cpu_fbatches.lock); 718 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file); 719 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn); 720 local_unlock(&cpu_fbatches.lock); 721 } 722 723 /* 724 * deactivate_page - deactivate a page 725 * @page: page to deactivate 726 * 727 * deactivate_page() moves @page to the inactive list if @page was on the active 728 * list and was not an unevictable page. This is done to accelerate the reclaim 729 * of @page. 730 */ 731 void deactivate_page(struct page *page) 732 { 733 struct folio *folio = page_folio(page); 734 735 if (folio_test_lru(folio) && !folio_test_unevictable(folio) && 736 (folio_test_active(folio) || lru_gen_enabled())) { 737 struct folio_batch *fbatch; 738 739 folio_get(folio); 740 local_lock(&cpu_fbatches.lock); 741 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate); 742 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn); 743 local_unlock(&cpu_fbatches.lock); 744 } 745 } 746 747 /** 748 * mark_page_lazyfree - make an anon page lazyfree 749 * @page: page to deactivate 750 * 751 * mark_page_lazyfree() moves @page to the inactive file list. 752 * This is done to accelerate the reclaim of @page. 753 */ 754 void mark_page_lazyfree(struct page *page) 755 { 756 struct folio *folio = page_folio(page); 757 758 if (folio_test_lru(folio) && folio_test_anon(folio) && 759 folio_test_swapbacked(folio) && !folio_test_swapcache(folio) && 760 !folio_test_unevictable(folio)) { 761 struct folio_batch *fbatch; 762 763 folio_get(folio); 764 local_lock(&cpu_fbatches.lock); 765 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree); 766 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn); 767 local_unlock(&cpu_fbatches.lock); 768 } 769 } 770 771 void lru_add_drain(void) 772 { 773 local_lock(&cpu_fbatches.lock); 774 lru_add_drain_cpu(smp_processor_id()); 775 local_unlock(&cpu_fbatches.lock); 776 mlock_page_drain_local(); 777 } 778 779 /* 780 * It's called from per-cpu workqueue context in SMP case so 781 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on 782 * the same cpu. It shouldn't be a problem in !SMP case since 783 * the core is only one and the locks will disable preemption. 784 */ 785 static void lru_add_and_bh_lrus_drain(void) 786 { 787 local_lock(&cpu_fbatches.lock); 788 lru_add_drain_cpu(smp_processor_id()); 789 local_unlock(&cpu_fbatches.lock); 790 invalidate_bh_lrus_cpu(); 791 mlock_page_drain_local(); 792 } 793 794 void lru_add_drain_cpu_zone(struct zone *zone) 795 { 796 local_lock(&cpu_fbatches.lock); 797 lru_add_drain_cpu(smp_processor_id()); 798 drain_local_pages(zone); 799 local_unlock(&cpu_fbatches.lock); 800 mlock_page_drain_local(); 801 } 802 803 #ifdef CONFIG_SMP 804 805 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work); 806 807 static void lru_add_drain_per_cpu(struct work_struct *dummy) 808 { 809 lru_add_and_bh_lrus_drain(); 810 } 811 812 static bool cpu_needs_drain(unsigned int cpu) 813 { 814 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu); 815 816 /* Check these in order of likelihood that they're not zero */ 817 return folio_batch_count(&fbatches->lru_add) || 818 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) || 819 folio_batch_count(&fbatches->lru_deactivate_file) || 820 folio_batch_count(&fbatches->lru_deactivate) || 821 folio_batch_count(&fbatches->lru_lazyfree) || 822 folio_batch_count(&fbatches->activate) || 823 need_mlock_page_drain(cpu) || 824 has_bh_in_lru(cpu, NULL); 825 } 826 827 /* 828 * Doesn't need any cpu hotplug locking because we do rely on per-cpu 829 * kworkers being shut down before our page_alloc_cpu_dead callback is 830 * executed on the offlined cpu. 831 * Calling this function with cpu hotplug locks held can actually lead 832 * to obscure indirect dependencies via WQ context. 833 */ 834 static inline void __lru_add_drain_all(bool force_all_cpus) 835 { 836 /* 837 * lru_drain_gen - Global pages generation number 838 * 839 * (A) Definition: global lru_drain_gen = x implies that all generations 840 * 0 < n <= x are already *scheduled* for draining. 841 * 842 * This is an optimization for the highly-contended use case where a 843 * user space workload keeps constantly generating a flow of pages for 844 * each CPU. 845 */ 846 static unsigned int lru_drain_gen; 847 static struct cpumask has_work; 848 static DEFINE_MUTEX(lock); 849 unsigned cpu, this_gen; 850 851 /* 852 * Make sure nobody triggers this path before mm_percpu_wq is fully 853 * initialized. 854 */ 855 if (WARN_ON(!mm_percpu_wq)) 856 return; 857 858 /* 859 * Guarantee folio_batch counter stores visible by this CPU 860 * are visible to other CPUs before loading the current drain 861 * generation. 862 */ 863 smp_mb(); 864 865 /* 866 * (B) Locally cache global LRU draining generation number 867 * 868 * The read barrier ensures that the counter is loaded before the mutex 869 * is taken. It pairs with smp_mb() inside the mutex critical section 870 * at (D). 871 */ 872 this_gen = smp_load_acquire(&lru_drain_gen); 873 874 mutex_lock(&lock); 875 876 /* 877 * (C) Exit the draining operation if a newer generation, from another 878 * lru_add_drain_all(), was already scheduled for draining. Check (A). 879 */ 880 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus)) 881 goto done; 882 883 /* 884 * (D) Increment global generation number 885 * 886 * Pairs with smp_load_acquire() at (B), outside of the critical 887 * section. Use a full memory barrier to guarantee that the 888 * new global drain generation number is stored before loading 889 * folio_batch counters. 890 * 891 * This pairing must be done here, before the for_each_online_cpu loop 892 * below which drains the page vectors. 893 * 894 * Let x, y, and z represent some system CPU numbers, where x < y < z. 895 * Assume CPU #z is in the middle of the for_each_online_cpu loop 896 * below and has already reached CPU #y's per-cpu data. CPU #x comes 897 * along, adds some pages to its per-cpu vectors, then calls 898 * lru_add_drain_all(). 899 * 900 * If the paired barrier is done at any later step, e.g. after the 901 * loop, CPU #x will just exit at (C) and miss flushing out all of its 902 * added pages. 903 */ 904 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1); 905 smp_mb(); 906 907 cpumask_clear(&has_work); 908 for_each_online_cpu(cpu) { 909 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu); 910 911 if (cpu_needs_drain(cpu)) { 912 INIT_WORK(work, lru_add_drain_per_cpu); 913 queue_work_on(cpu, mm_percpu_wq, work); 914 __cpumask_set_cpu(cpu, &has_work); 915 } 916 } 917 918 for_each_cpu(cpu, &has_work) 919 flush_work(&per_cpu(lru_add_drain_work, cpu)); 920 921 done: 922 mutex_unlock(&lock); 923 } 924 925 void lru_add_drain_all(void) 926 { 927 __lru_add_drain_all(false); 928 } 929 #else 930 void lru_add_drain_all(void) 931 { 932 lru_add_drain(); 933 } 934 #endif /* CONFIG_SMP */ 935 936 atomic_t lru_disable_count = ATOMIC_INIT(0); 937 938 /* 939 * lru_cache_disable() needs to be called before we start compiling 940 * a list of pages to be migrated using isolate_lru_page(). 941 * It drains pages on LRU cache and then disable on all cpus until 942 * lru_cache_enable is called. 943 * 944 * Must be paired with a call to lru_cache_enable(). 945 */ 946 void lru_cache_disable(void) 947 { 948 atomic_inc(&lru_disable_count); 949 /* 950 * Readers of lru_disable_count are protected by either disabling 951 * preemption or rcu_read_lock: 952 * 953 * preempt_disable, local_irq_disable [bh_lru_lock()] 954 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT] 955 * preempt_disable [local_lock !CONFIG_PREEMPT_RT] 956 * 957 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on 958 * preempt_disable() regions of code. So any CPU which sees 959 * lru_disable_count = 0 will have exited the critical 960 * section when synchronize_rcu() returns. 961 */ 962 synchronize_rcu_expedited(); 963 #ifdef CONFIG_SMP 964 __lru_add_drain_all(true); 965 #else 966 lru_add_and_bh_lrus_drain(); 967 #endif 968 } 969 970 /** 971 * release_pages - batched put_page() 972 * @pages: array of pages to release 973 * @nr: number of pages 974 * 975 * Decrement the reference count on all the pages in @pages. If it 976 * fell to zero, remove the page from the LRU and free it. 977 */ 978 void release_pages(struct page **pages, int nr) 979 { 980 int i; 981 LIST_HEAD(pages_to_free); 982 struct lruvec *lruvec = NULL; 983 unsigned long flags = 0; 984 unsigned int lock_batch; 985 986 for (i = 0; i < nr; i++) { 987 struct folio *folio = page_folio(pages[i]); 988 989 /* 990 * Make sure the IRQ-safe lock-holding time does not get 991 * excessive with a continuous string of pages from the 992 * same lruvec. The lock is held only if lruvec != NULL. 993 */ 994 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) { 995 unlock_page_lruvec_irqrestore(lruvec, flags); 996 lruvec = NULL; 997 } 998 999 if (is_huge_zero_page(&folio->page)) 1000 continue; 1001 1002 if (folio_is_zone_device(folio)) { 1003 if (lruvec) { 1004 unlock_page_lruvec_irqrestore(lruvec, flags); 1005 lruvec = NULL; 1006 } 1007 if (put_devmap_managed_page(&folio->page)) 1008 continue; 1009 if (folio_put_testzero(folio)) 1010 free_zone_device_page(&folio->page); 1011 continue; 1012 } 1013 1014 if (!folio_put_testzero(folio)) 1015 continue; 1016 1017 if (folio_test_large(folio)) { 1018 if (lruvec) { 1019 unlock_page_lruvec_irqrestore(lruvec, flags); 1020 lruvec = NULL; 1021 } 1022 __folio_put_large(folio); 1023 continue; 1024 } 1025 1026 if (folio_test_lru(folio)) { 1027 struct lruvec *prev_lruvec = lruvec; 1028 1029 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, 1030 &flags); 1031 if (prev_lruvec != lruvec) 1032 lock_batch = 0; 1033 1034 lruvec_del_folio(lruvec, folio); 1035 __folio_clear_lru_flags(folio); 1036 } 1037 1038 /* 1039 * In rare cases, when truncation or holepunching raced with 1040 * munlock after VM_LOCKED was cleared, Mlocked may still be 1041 * found set here. This does not indicate a problem, unless 1042 * "unevictable_pgs_cleared" appears worryingly large. 1043 */ 1044 if (unlikely(folio_test_mlocked(folio))) { 1045 __folio_clear_mlocked(folio); 1046 zone_stat_sub_folio(folio, NR_MLOCK); 1047 count_vm_event(UNEVICTABLE_PGCLEARED); 1048 } 1049 1050 list_add(&folio->lru, &pages_to_free); 1051 } 1052 if (lruvec) 1053 unlock_page_lruvec_irqrestore(lruvec, flags); 1054 1055 mem_cgroup_uncharge_list(&pages_to_free); 1056 free_unref_page_list(&pages_to_free); 1057 } 1058 EXPORT_SYMBOL(release_pages); 1059 1060 /* 1061 * The pages which we're about to release may be in the deferred lru-addition 1062 * queues. That would prevent them from really being freed right now. That's 1063 * OK from a correctness point of view but is inefficient - those pages may be 1064 * cache-warm and we want to give them back to the page allocator ASAP. 1065 * 1066 * So __pagevec_release() will drain those queues here. 1067 * folio_batch_move_lru() calls folios_put() directly to avoid 1068 * mutual recursion. 1069 */ 1070 void __pagevec_release(struct pagevec *pvec) 1071 { 1072 if (!pvec->percpu_pvec_drained) { 1073 lru_add_drain(); 1074 pvec->percpu_pvec_drained = true; 1075 } 1076 release_pages(pvec->pages, pagevec_count(pvec)); 1077 pagevec_reinit(pvec); 1078 } 1079 EXPORT_SYMBOL(__pagevec_release); 1080 1081 /** 1082 * folio_batch_remove_exceptionals() - Prune non-folios from a batch. 1083 * @fbatch: The batch to prune 1084 * 1085 * find_get_entries() fills a batch with both folios and shadow/swap/DAX 1086 * entries. This function prunes all the non-folio entries from @fbatch 1087 * without leaving holes, so that it can be passed on to folio-only batch 1088 * operations. 1089 */ 1090 void folio_batch_remove_exceptionals(struct folio_batch *fbatch) 1091 { 1092 unsigned int i, j; 1093 1094 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) { 1095 struct folio *folio = fbatch->folios[i]; 1096 if (!xa_is_value(folio)) 1097 fbatch->folios[j++] = folio; 1098 } 1099 fbatch->nr = j; 1100 } 1101 1102 unsigned pagevec_lookup_range_tag(struct pagevec *pvec, 1103 struct address_space *mapping, pgoff_t *index, pgoff_t end, 1104 xa_mark_t tag) 1105 { 1106 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag, 1107 PAGEVEC_SIZE, pvec->pages); 1108 return pagevec_count(pvec); 1109 } 1110 EXPORT_SYMBOL(pagevec_lookup_range_tag); 1111 1112 /* 1113 * Perform any setup for the swap system 1114 */ 1115 void __init swap_setup(void) 1116 { 1117 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT); 1118 1119 /* Use a smaller cluster for small-memory machines */ 1120 if (megs < 16) 1121 page_cluster = 2; 1122 else 1123 page_cluster = 3; 1124 /* 1125 * Right now other parts of the system means that we 1126 * _really_ don't want to cluster much more 1127 */ 1128 } 1129