1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * linux/mm/swap_state.c 4 * 5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 6 * Swap reorganised 29.12.95, Stephen Tweedie 7 * 8 * Rewritten to use page cache, (C) 1998 Stephen Tweedie 9 */ 10 #include <linux/mm.h> 11 #include <linux/gfp.h> 12 #include <linux/kernel_stat.h> 13 #include <linux/mempolicy.h> 14 #include <linux/swap.h> 15 #include <linux/swapops.h> 16 #include <linux/init.h> 17 #include <linux/pagemap.h> 18 #include <linux/pagevec.h> 19 #include <linux/backing-dev.h> 20 #include <linux/blkdev.h> 21 #include <linux/migrate.h> 22 #include <linux/vmalloc.h> 23 #include <linux/huge_mm.h> 24 #include <linux/shmem_fs.h> 25 #include "internal.h" 26 #include "swap.h" 27 28 /* 29 * swapper_space is a fiction, retained to simplify the path through 30 * vmscan's shrink_folio_list. 31 */ 32 static const struct address_space_operations swap_aops = { 33 .dirty_folio = noop_dirty_folio, 34 #ifdef CONFIG_MIGRATION 35 .migrate_folio = migrate_folio, 36 #endif 37 }; 38 39 struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly; 40 static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly; 41 static bool enable_vma_readahead __read_mostly = true; 42 43 #define SWAP_RA_ORDER_CEILING 5 44 45 #define SWAP_RA_WIN_SHIFT (PAGE_SHIFT / 2) 46 #define SWAP_RA_HITS_MASK ((1UL << SWAP_RA_WIN_SHIFT) - 1) 47 #define SWAP_RA_HITS_MAX SWAP_RA_HITS_MASK 48 #define SWAP_RA_WIN_MASK (~PAGE_MASK & ~SWAP_RA_HITS_MASK) 49 50 #define SWAP_RA_HITS(v) ((v) & SWAP_RA_HITS_MASK) 51 #define SWAP_RA_WIN(v) (((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT) 52 #define SWAP_RA_ADDR(v) ((v) & PAGE_MASK) 53 54 #define SWAP_RA_VAL(addr, win, hits) \ 55 (((addr) & PAGE_MASK) | \ 56 (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) | \ 57 ((hits) & SWAP_RA_HITS_MASK)) 58 59 /* Initial readahead hits is 4 to start up with a small window */ 60 #define GET_SWAP_RA_VAL(vma) \ 61 (atomic_long_read(&(vma)->swap_readahead_info) ? : 4) 62 63 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4); 64 65 void show_swap_cache_info(void) 66 { 67 printk("%lu pages in swap cache\n", total_swapcache_pages()); 68 printk("Free swap = %ldkB\n", K(get_nr_swap_pages())); 69 printk("Total swap = %lukB\n", K(total_swap_pages)); 70 } 71 72 void *get_shadow_from_swap_cache(swp_entry_t entry) 73 { 74 struct address_space *address_space = swap_address_space(entry); 75 pgoff_t idx = swap_cache_index(entry); 76 void *shadow; 77 78 shadow = xa_load(&address_space->i_pages, idx); 79 if (xa_is_value(shadow)) 80 return shadow; 81 return NULL; 82 } 83 84 /* 85 * add_to_swap_cache resembles filemap_add_folio on swapper_space, 86 * but sets SwapCache flag and 'swap' instead of mapping and index. 87 */ 88 int add_to_swap_cache(struct folio *folio, swp_entry_t entry, 89 gfp_t gfp, void **shadowp) 90 { 91 struct address_space *address_space = swap_address_space(entry); 92 pgoff_t idx = swap_cache_index(entry); 93 XA_STATE_ORDER(xas, &address_space->i_pages, idx, folio_order(folio)); 94 unsigned long i, nr = folio_nr_pages(folio); 95 void *old; 96 97 xas_set_update(&xas, workingset_update_node); 98 99 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 100 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); 101 VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio); 102 103 folio_ref_add(folio, nr); 104 folio_set_swapcache(folio); 105 folio->swap = entry; 106 107 do { 108 xas_lock_irq(&xas); 109 xas_create_range(&xas); 110 if (xas_error(&xas)) 111 goto unlock; 112 for (i = 0; i < nr; i++) { 113 VM_BUG_ON_FOLIO(xas.xa_index != idx + i, folio); 114 if (shadowp) { 115 old = xas_load(&xas); 116 if (xa_is_value(old)) 117 *shadowp = old; 118 } 119 xas_store(&xas, folio); 120 xas_next(&xas); 121 } 122 address_space->nrpages += nr; 123 __node_stat_mod_folio(folio, NR_FILE_PAGES, nr); 124 __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, nr); 125 unlock: 126 xas_unlock_irq(&xas); 127 } while (xas_nomem(&xas, gfp)); 128 129 if (!xas_error(&xas)) 130 return 0; 131 132 folio_clear_swapcache(folio); 133 folio_ref_sub(folio, nr); 134 return xas_error(&xas); 135 } 136 137 /* 138 * This must be called only on folios that have 139 * been verified to be in the swap cache. 140 */ 141 void __delete_from_swap_cache(struct folio *folio, 142 swp_entry_t entry, void *shadow) 143 { 144 struct address_space *address_space = swap_address_space(entry); 145 int i; 146 long nr = folio_nr_pages(folio); 147 pgoff_t idx = swap_cache_index(entry); 148 XA_STATE(xas, &address_space->i_pages, idx); 149 150 xas_set_update(&xas, workingset_update_node); 151 152 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 153 VM_BUG_ON_FOLIO(!folio_test_swapcache(folio), folio); 154 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio); 155 156 for (i = 0; i < nr; i++) { 157 void *entry = xas_store(&xas, shadow); 158 VM_BUG_ON_PAGE(entry != folio, entry); 159 xas_next(&xas); 160 } 161 folio->swap.val = 0; 162 folio_clear_swapcache(folio); 163 address_space->nrpages -= nr; 164 __node_stat_mod_folio(folio, NR_FILE_PAGES, -nr); 165 __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, -nr); 166 } 167 168 /* 169 * This must be called only on folios that have 170 * been verified to be in the swap cache and locked. 171 * It will never put the folio into the free list, 172 * the caller has a reference on the folio. 173 */ 174 void delete_from_swap_cache(struct folio *folio) 175 { 176 swp_entry_t entry = folio->swap; 177 struct address_space *address_space = swap_address_space(entry); 178 179 xa_lock_irq(&address_space->i_pages); 180 __delete_from_swap_cache(folio, entry, NULL); 181 xa_unlock_irq(&address_space->i_pages); 182 183 put_swap_folio(folio, entry); 184 folio_ref_sub(folio, folio_nr_pages(folio)); 185 } 186 187 void clear_shadow_from_swap_cache(int type, unsigned long begin, 188 unsigned long end) 189 { 190 unsigned long curr = begin; 191 void *old; 192 193 for (;;) { 194 swp_entry_t entry = swp_entry(type, curr); 195 unsigned long index = curr & SWAP_ADDRESS_SPACE_MASK; 196 struct address_space *address_space = swap_address_space(entry); 197 XA_STATE(xas, &address_space->i_pages, index); 198 199 xas_set_update(&xas, workingset_update_node); 200 201 xa_lock_irq(&address_space->i_pages); 202 xas_for_each(&xas, old, min(index + (end - curr), SWAP_ADDRESS_SPACE_PAGES)) { 203 if (!xa_is_value(old)) 204 continue; 205 xas_store(&xas, NULL); 206 } 207 xa_unlock_irq(&address_space->i_pages); 208 209 /* search the next swapcache until we meet end */ 210 curr = ALIGN((curr + 1), SWAP_ADDRESS_SPACE_PAGES); 211 if (curr > end) 212 break; 213 } 214 } 215 216 /* 217 * If we are the only user, then try to free up the swap cache. 218 * 219 * Its ok to check the swapcache flag without the folio lock 220 * here because we are going to recheck again inside 221 * folio_free_swap() _with_ the lock. 222 * - Marcelo 223 */ 224 void free_swap_cache(struct folio *folio) 225 { 226 if (folio_test_swapcache(folio) && !folio_mapped(folio) && 227 folio_trylock(folio)) { 228 folio_free_swap(folio); 229 folio_unlock(folio); 230 } 231 } 232 233 /* 234 * Perform a free_page(), also freeing any swap cache associated with 235 * this page if it is the last user of the page. 236 */ 237 void free_page_and_swap_cache(struct page *page) 238 { 239 struct folio *folio = page_folio(page); 240 241 free_swap_cache(folio); 242 if (!is_huge_zero_folio(folio)) 243 folio_put(folio); 244 } 245 246 /* 247 * Passed an array of pages, drop them all from swapcache and then release 248 * them. They are removed from the LRU and freed if this is their last use. 249 */ 250 void free_pages_and_swap_cache(struct encoded_page **pages, int nr) 251 { 252 struct folio_batch folios; 253 unsigned int refs[PAGEVEC_SIZE]; 254 255 folio_batch_init(&folios); 256 for (int i = 0; i < nr; i++) { 257 struct folio *folio = page_folio(encoded_page_ptr(pages[i])); 258 259 free_swap_cache(folio); 260 refs[folios.nr] = 1; 261 if (unlikely(encoded_page_flags(pages[i]) & 262 ENCODED_PAGE_BIT_NR_PAGES_NEXT)) 263 refs[folios.nr] = encoded_nr_pages(pages[++i]); 264 265 if (folio_batch_add(&folios, folio) == 0) 266 folios_put_refs(&folios, refs); 267 } 268 if (folios.nr) 269 folios_put_refs(&folios, refs); 270 } 271 272 static inline bool swap_use_vma_readahead(void) 273 { 274 return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap); 275 } 276 277 /* 278 * Lookup a swap entry in the swap cache. A found folio will be returned 279 * unlocked and with its refcount incremented - we rely on the kernel 280 * lock getting page table operations atomic even if we drop the folio 281 * lock before returning. 282 * 283 * Caller must lock the swap device or hold a reference to keep it valid. 284 */ 285 struct folio *swap_cache_get_folio(swp_entry_t entry, 286 struct vm_area_struct *vma, unsigned long addr) 287 { 288 struct folio *folio; 289 290 folio = filemap_get_folio(swap_address_space(entry), swap_cache_index(entry)); 291 if (!IS_ERR(folio)) { 292 bool vma_ra = swap_use_vma_readahead(); 293 bool readahead; 294 295 /* 296 * At the moment, we don't support PG_readahead for anon THP 297 * so let's bail out rather than confusing the readahead stat. 298 */ 299 if (unlikely(folio_test_large(folio))) 300 return folio; 301 302 readahead = folio_test_clear_readahead(folio); 303 if (vma && vma_ra) { 304 unsigned long ra_val; 305 int win, hits; 306 307 ra_val = GET_SWAP_RA_VAL(vma); 308 win = SWAP_RA_WIN(ra_val); 309 hits = SWAP_RA_HITS(ra_val); 310 if (readahead) 311 hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX); 312 atomic_long_set(&vma->swap_readahead_info, 313 SWAP_RA_VAL(addr, win, hits)); 314 } 315 316 if (readahead) { 317 count_vm_event(SWAP_RA_HIT); 318 if (!vma || !vma_ra) 319 atomic_inc(&swapin_readahead_hits); 320 } 321 } else { 322 folio = NULL; 323 } 324 325 return folio; 326 } 327 328 /** 329 * filemap_get_incore_folio - Find and get a folio from the page or swap caches. 330 * @mapping: The address_space to search. 331 * @index: The page cache index. 332 * 333 * This differs from filemap_get_folio() in that it will also look for the 334 * folio in the swap cache. 335 * 336 * Return: The found folio or %NULL. 337 */ 338 struct folio *filemap_get_incore_folio(struct address_space *mapping, 339 pgoff_t index) 340 { 341 swp_entry_t swp; 342 struct swap_info_struct *si; 343 struct folio *folio = filemap_get_entry(mapping, index); 344 345 if (!folio) 346 return ERR_PTR(-ENOENT); 347 if (!xa_is_value(folio)) 348 return folio; 349 if (!shmem_mapping(mapping)) 350 return ERR_PTR(-ENOENT); 351 352 swp = radix_to_swp_entry(folio); 353 /* There might be swapin error entries in shmem mapping. */ 354 if (non_swap_entry(swp)) 355 return ERR_PTR(-ENOENT); 356 /* Prevent swapoff from happening to us */ 357 si = get_swap_device(swp); 358 if (!si) 359 return ERR_PTR(-ENOENT); 360 index = swap_cache_index(swp); 361 folio = filemap_get_folio(swap_address_space(swp), index); 362 put_swap_device(si); 363 return folio; 364 } 365 366 struct folio *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, 367 struct mempolicy *mpol, pgoff_t ilx, bool *new_page_allocated, 368 bool skip_if_exists) 369 { 370 struct swap_info_struct *si = swp_swap_info(entry); 371 struct folio *folio; 372 struct folio *new_folio = NULL; 373 struct folio *result = NULL; 374 void *shadow = NULL; 375 376 *new_page_allocated = false; 377 for (;;) { 378 int err; 379 /* 380 * First check the swap cache. Since this is normally 381 * called after swap_cache_get_folio() failed, re-calling 382 * that would confuse statistics. 383 */ 384 folio = filemap_get_folio(swap_address_space(entry), 385 swap_cache_index(entry)); 386 if (!IS_ERR(folio)) 387 goto got_folio; 388 389 /* 390 * Just skip read ahead for unused swap slot. 391 */ 392 if (!swap_entry_swapped(si, entry)) 393 goto put_and_return; 394 395 /* 396 * Get a new folio to read into from swap. Allocate it now if 397 * new_folio not exist, before marking swap_map SWAP_HAS_CACHE, 398 * when -EEXIST will cause any racers to loop around until we 399 * add it to cache. 400 */ 401 if (!new_folio) { 402 new_folio = folio_alloc_mpol(gfp_mask, 0, mpol, ilx, numa_node_id()); 403 if (!new_folio) 404 goto put_and_return; 405 } 406 407 /* 408 * Swap entry may have been freed since our caller observed it. 409 */ 410 err = swapcache_prepare(entry, 1); 411 if (!err) 412 break; 413 else if (err != -EEXIST) 414 goto put_and_return; 415 416 /* 417 * Protect against a recursive call to __read_swap_cache_async() 418 * on the same entry waiting forever here because SWAP_HAS_CACHE 419 * is set but the folio is not the swap cache yet. This can 420 * happen today if mem_cgroup_swapin_charge_folio() below 421 * triggers reclaim through zswap, which may call 422 * __read_swap_cache_async() in the writeback path. 423 */ 424 if (skip_if_exists) 425 goto put_and_return; 426 427 /* 428 * We might race against __delete_from_swap_cache(), and 429 * stumble across a swap_map entry whose SWAP_HAS_CACHE 430 * has not yet been cleared. Or race against another 431 * __read_swap_cache_async(), which has set SWAP_HAS_CACHE 432 * in swap_map, but not yet added its folio to swap cache. 433 */ 434 schedule_timeout_uninterruptible(1); 435 } 436 437 /* 438 * The swap entry is ours to swap in. Prepare the new folio. 439 */ 440 __folio_set_locked(new_folio); 441 __folio_set_swapbacked(new_folio); 442 443 if (mem_cgroup_swapin_charge_folio(new_folio, NULL, gfp_mask, entry)) 444 goto fail_unlock; 445 446 /* May fail (-ENOMEM) if XArray node allocation failed. */ 447 if (add_to_swap_cache(new_folio, entry, gfp_mask & GFP_RECLAIM_MASK, &shadow)) 448 goto fail_unlock; 449 450 memcg1_swapin(entry, 1); 451 452 if (shadow) 453 workingset_refault(new_folio, shadow); 454 455 /* Caller will initiate read into locked new_folio */ 456 folio_add_lru(new_folio); 457 *new_page_allocated = true; 458 folio = new_folio; 459 got_folio: 460 result = folio; 461 goto put_and_return; 462 463 fail_unlock: 464 put_swap_folio(new_folio, entry); 465 folio_unlock(new_folio); 466 put_and_return: 467 if (!(*new_page_allocated) && new_folio) 468 folio_put(new_folio); 469 return result; 470 } 471 472 /* 473 * Locate a page of swap in physical memory, reserving swap cache space 474 * and reading the disk if it is not already cached. 475 * A failure return means that either the page allocation failed or that 476 * the swap entry is no longer in use. 477 * 478 * get/put_swap_device() aren't needed to call this function, because 479 * __read_swap_cache_async() call them and swap_read_folio() holds the 480 * swap cache folio lock. 481 */ 482 struct folio *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, 483 struct vm_area_struct *vma, unsigned long addr, 484 struct swap_iocb **plug) 485 { 486 struct swap_info_struct *si; 487 bool page_allocated; 488 struct mempolicy *mpol; 489 pgoff_t ilx; 490 struct folio *folio; 491 492 si = get_swap_device(entry); 493 if (!si) 494 return NULL; 495 496 mpol = get_vma_policy(vma, addr, 0, &ilx); 497 folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx, 498 &page_allocated, false); 499 mpol_cond_put(mpol); 500 501 if (page_allocated) 502 swap_read_folio(folio, plug); 503 504 put_swap_device(si); 505 return folio; 506 } 507 508 static unsigned int __swapin_nr_pages(unsigned long prev_offset, 509 unsigned long offset, 510 int hits, 511 int max_pages, 512 int prev_win) 513 { 514 unsigned int pages, last_ra; 515 516 /* 517 * This heuristic has been found to work well on both sequential and 518 * random loads, swapping to hard disk or to SSD: please don't ask 519 * what the "+ 2" means, it just happens to work well, that's all. 520 */ 521 pages = hits + 2; 522 if (pages == 2) { 523 /* 524 * We can have no readahead hits to judge by: but must not get 525 * stuck here forever, so check for an adjacent offset instead 526 * (and don't even bother to check whether swap type is same). 527 */ 528 if (offset != prev_offset + 1 && offset != prev_offset - 1) 529 pages = 1; 530 } else { 531 unsigned int roundup = 4; 532 while (roundup < pages) 533 roundup <<= 1; 534 pages = roundup; 535 } 536 537 if (pages > max_pages) 538 pages = max_pages; 539 540 /* Don't shrink readahead too fast */ 541 last_ra = prev_win / 2; 542 if (pages < last_ra) 543 pages = last_ra; 544 545 return pages; 546 } 547 548 static unsigned long swapin_nr_pages(unsigned long offset) 549 { 550 static unsigned long prev_offset; 551 unsigned int hits, pages, max_pages; 552 static atomic_t last_readahead_pages; 553 554 max_pages = 1 << READ_ONCE(page_cluster); 555 if (max_pages <= 1) 556 return 1; 557 558 hits = atomic_xchg(&swapin_readahead_hits, 0); 559 pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits, 560 max_pages, 561 atomic_read(&last_readahead_pages)); 562 if (!hits) 563 WRITE_ONCE(prev_offset, offset); 564 atomic_set(&last_readahead_pages, pages); 565 566 return pages; 567 } 568 569 /** 570 * swap_cluster_readahead - swap in pages in hope we need them soon 571 * @entry: swap entry of this memory 572 * @gfp_mask: memory allocation flags 573 * @mpol: NUMA memory allocation policy to be applied 574 * @ilx: NUMA interleave index, for use only when MPOL_INTERLEAVE 575 * 576 * Returns the struct folio for entry and addr, after queueing swapin. 577 * 578 * Primitive swap readahead code. We simply read an aligned block of 579 * (1 << page_cluster) entries in the swap area. This method is chosen 580 * because it doesn't cost us any seek time. We also make sure to queue 581 * the 'original' request together with the readahead ones... 582 * 583 * Note: it is intentional that the same NUMA policy and interleave index 584 * are used for every page of the readahead: neighbouring pages on swap 585 * are fairly likely to have been swapped out from the same node. 586 */ 587 struct folio *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask, 588 struct mempolicy *mpol, pgoff_t ilx) 589 { 590 struct folio *folio; 591 unsigned long entry_offset = swp_offset(entry); 592 unsigned long offset = entry_offset; 593 unsigned long start_offset, end_offset; 594 unsigned long mask; 595 struct swap_info_struct *si = swp_swap_info(entry); 596 struct blk_plug plug; 597 struct swap_iocb *splug = NULL; 598 bool page_allocated; 599 600 mask = swapin_nr_pages(offset) - 1; 601 if (!mask) 602 goto skip; 603 604 /* Read a page_cluster sized and aligned cluster around offset. */ 605 start_offset = offset & ~mask; 606 end_offset = offset | mask; 607 if (!start_offset) /* First page is swap header. */ 608 start_offset++; 609 if (end_offset >= si->max) 610 end_offset = si->max - 1; 611 612 blk_start_plug(&plug); 613 for (offset = start_offset; offset <= end_offset ; offset++) { 614 /* Ok, do the async read-ahead now */ 615 folio = __read_swap_cache_async( 616 swp_entry(swp_type(entry), offset), 617 gfp_mask, mpol, ilx, &page_allocated, false); 618 if (!folio) 619 continue; 620 if (page_allocated) { 621 swap_read_folio(folio, &splug); 622 if (offset != entry_offset) { 623 folio_set_readahead(folio); 624 count_vm_event(SWAP_RA); 625 } 626 } 627 folio_put(folio); 628 } 629 blk_finish_plug(&plug); 630 swap_read_unplug(splug); 631 lru_add_drain(); /* Push any new pages onto the LRU now */ 632 skip: 633 /* The page was likely read above, so no need for plugging here */ 634 folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx, 635 &page_allocated, false); 636 if (unlikely(page_allocated)) 637 swap_read_folio(folio, NULL); 638 return folio; 639 } 640 641 int init_swap_address_space(unsigned int type, unsigned long nr_pages) 642 { 643 struct address_space *spaces, *space; 644 unsigned int i, nr; 645 646 nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES); 647 spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL); 648 if (!spaces) 649 return -ENOMEM; 650 for (i = 0; i < nr; i++) { 651 space = spaces + i; 652 xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ); 653 atomic_set(&space->i_mmap_writable, 0); 654 space->a_ops = &swap_aops; 655 /* swap cache doesn't use writeback related tags */ 656 mapping_set_no_writeback_tags(space); 657 } 658 nr_swapper_spaces[type] = nr; 659 swapper_spaces[type] = spaces; 660 661 return 0; 662 } 663 664 void exit_swap_address_space(unsigned int type) 665 { 666 int i; 667 struct address_space *spaces = swapper_spaces[type]; 668 669 for (i = 0; i < nr_swapper_spaces[type]; i++) 670 VM_WARN_ON_ONCE(!mapping_empty(&spaces[i])); 671 kvfree(spaces); 672 nr_swapper_spaces[type] = 0; 673 swapper_spaces[type] = NULL; 674 } 675 676 static int swap_vma_ra_win(struct vm_fault *vmf, unsigned long *start, 677 unsigned long *end) 678 { 679 struct vm_area_struct *vma = vmf->vma; 680 unsigned long ra_val; 681 unsigned long faddr, prev_faddr, left, right; 682 unsigned int max_win, hits, prev_win, win; 683 684 max_win = 1 << min(READ_ONCE(page_cluster), SWAP_RA_ORDER_CEILING); 685 if (max_win == 1) 686 return 1; 687 688 faddr = vmf->address; 689 ra_val = GET_SWAP_RA_VAL(vma); 690 prev_faddr = SWAP_RA_ADDR(ra_val); 691 prev_win = SWAP_RA_WIN(ra_val); 692 hits = SWAP_RA_HITS(ra_val); 693 win = __swapin_nr_pages(PFN_DOWN(prev_faddr), PFN_DOWN(faddr), hits, 694 max_win, prev_win); 695 atomic_long_set(&vma->swap_readahead_info, SWAP_RA_VAL(faddr, win, 0)); 696 if (win == 1) 697 return 1; 698 699 if (faddr == prev_faddr + PAGE_SIZE) 700 left = faddr; 701 else if (prev_faddr == faddr + PAGE_SIZE) 702 left = faddr - (win << PAGE_SHIFT) + PAGE_SIZE; 703 else 704 left = faddr - (((win - 1) / 2) << PAGE_SHIFT); 705 right = left + (win << PAGE_SHIFT); 706 if ((long)left < 0) 707 left = 0; 708 *start = max3(left, vma->vm_start, faddr & PMD_MASK); 709 *end = min3(right, vma->vm_end, (faddr & PMD_MASK) + PMD_SIZE); 710 711 return win; 712 } 713 714 /** 715 * swap_vma_readahead - swap in pages in hope we need them soon 716 * @targ_entry: swap entry of the targeted memory 717 * @gfp_mask: memory allocation flags 718 * @mpol: NUMA memory allocation policy to be applied 719 * @targ_ilx: NUMA interleave index, for use only when MPOL_INTERLEAVE 720 * @vmf: fault information 721 * 722 * Returns the struct folio for entry and addr, after queueing swapin. 723 * 724 * Primitive swap readahead code. We simply read in a few pages whose 725 * virtual addresses are around the fault address in the same vma. 726 * 727 * Caller must hold read mmap_lock if vmf->vma is not NULL. 728 * 729 */ 730 static struct folio *swap_vma_readahead(swp_entry_t targ_entry, gfp_t gfp_mask, 731 struct mempolicy *mpol, pgoff_t targ_ilx, struct vm_fault *vmf) 732 { 733 struct blk_plug plug; 734 struct swap_iocb *splug = NULL; 735 struct folio *folio; 736 pte_t *pte = NULL, pentry; 737 int win; 738 unsigned long start, end, addr; 739 swp_entry_t entry; 740 pgoff_t ilx; 741 bool page_allocated; 742 743 win = swap_vma_ra_win(vmf, &start, &end); 744 if (win == 1) 745 goto skip; 746 747 ilx = targ_ilx - PFN_DOWN(vmf->address - start); 748 749 blk_start_plug(&plug); 750 for (addr = start; addr < end; ilx++, addr += PAGE_SIZE) { 751 if (!pte++) { 752 pte = pte_offset_map(vmf->pmd, addr); 753 if (!pte) 754 break; 755 } 756 pentry = ptep_get_lockless(pte); 757 if (!is_swap_pte(pentry)) 758 continue; 759 entry = pte_to_swp_entry(pentry); 760 if (unlikely(non_swap_entry(entry))) 761 continue; 762 pte_unmap(pte); 763 pte = NULL; 764 folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx, 765 &page_allocated, false); 766 if (!folio) 767 continue; 768 if (page_allocated) { 769 swap_read_folio(folio, &splug); 770 if (addr != vmf->address) { 771 folio_set_readahead(folio); 772 count_vm_event(SWAP_RA); 773 } 774 } 775 folio_put(folio); 776 } 777 if (pte) 778 pte_unmap(pte); 779 blk_finish_plug(&plug); 780 swap_read_unplug(splug); 781 lru_add_drain(); 782 skip: 783 /* The folio was likely read above, so no need for plugging here */ 784 folio = __read_swap_cache_async(targ_entry, gfp_mask, mpol, targ_ilx, 785 &page_allocated, false); 786 if (unlikely(page_allocated)) 787 swap_read_folio(folio, NULL); 788 return folio; 789 } 790 791 /** 792 * swapin_readahead - swap in pages in hope we need them soon 793 * @entry: swap entry of this memory 794 * @gfp_mask: memory allocation flags 795 * @vmf: fault information 796 * 797 * Returns the struct folio for entry and addr, after queueing swapin. 798 * 799 * It's a main entry function for swap readahead. By the configuration, 800 * it will read ahead blocks by cluster-based(ie, physical disk based) 801 * or vma-based(ie, virtual address based on faulty address) readahead. 802 */ 803 struct folio *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, 804 struct vm_fault *vmf) 805 { 806 struct mempolicy *mpol; 807 pgoff_t ilx; 808 struct folio *folio; 809 810 mpol = get_vma_policy(vmf->vma, vmf->address, 0, &ilx); 811 folio = swap_use_vma_readahead() ? 812 swap_vma_readahead(entry, gfp_mask, mpol, ilx, vmf) : 813 swap_cluster_readahead(entry, gfp_mask, mpol, ilx); 814 mpol_cond_put(mpol); 815 816 return folio; 817 } 818 819 #ifdef CONFIG_SYSFS 820 static ssize_t vma_ra_enabled_show(struct kobject *kobj, 821 struct kobj_attribute *attr, char *buf) 822 { 823 return sysfs_emit(buf, "%s\n", str_true_false(enable_vma_readahead)); 824 } 825 static ssize_t vma_ra_enabled_store(struct kobject *kobj, 826 struct kobj_attribute *attr, 827 const char *buf, size_t count) 828 { 829 ssize_t ret; 830 831 ret = kstrtobool(buf, &enable_vma_readahead); 832 if (ret) 833 return ret; 834 835 return count; 836 } 837 static struct kobj_attribute vma_ra_enabled_attr = __ATTR_RW(vma_ra_enabled); 838 839 static struct attribute *swap_attrs[] = { 840 &vma_ra_enabled_attr.attr, 841 NULL, 842 }; 843 844 static const struct attribute_group swap_attr_group = { 845 .attrs = swap_attrs, 846 }; 847 848 static int __init swap_init_sysfs(void) 849 { 850 int err; 851 struct kobject *swap_kobj; 852 853 swap_kobj = kobject_create_and_add("swap", mm_kobj); 854 if (!swap_kobj) { 855 pr_err("failed to create swap kobject\n"); 856 return -ENOMEM; 857 } 858 err = sysfs_create_group(swap_kobj, &swap_attr_group); 859 if (err) { 860 pr_err("failed to register swap group\n"); 861 goto delete_obj; 862 } 863 return 0; 864 865 delete_obj: 866 kobject_put(swap_kobj); 867 return err; 868 } 869 subsys_initcall(swap_init_sysfs); 870 #endif 871