Loading...
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/swap.h>
14#include <linux/swapops.h>
15#include <linux/init.h>
16#include <linux/pagemap.h>
17#include <linux/backing-dev.h>
18#include <linux/blkdev.h>
19#include <linux/pagevec.h>
20#include <linux/migrate.h>
21#include <linux/vmalloc.h>
22#include <linux/swap_slots.h>
23#include <linux/huge_mm.h>
24#include <linux/shmem_fs.h>
25#include "internal.h"
26
27/*
28 * swapper_space is a fiction, retained to simplify the path through
29 * vmscan's shrink_page_list.
30 */
31static const struct address_space_operations swap_aops = {
32 .writepage = swap_writepage,
33 .set_page_dirty = swap_set_page_dirty,
34#ifdef CONFIG_MIGRATION
35 .migratepage = migrate_page,
36#endif
37};
38
39struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly;
40static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly;
41static bool enable_vma_readahead __read_mostly = true;
42
43#define SWAP_RA_WIN_SHIFT (PAGE_SHIFT / 2)
44#define SWAP_RA_HITS_MASK ((1UL << SWAP_RA_WIN_SHIFT) - 1)
45#define SWAP_RA_HITS_MAX SWAP_RA_HITS_MASK
46#define SWAP_RA_WIN_MASK (~PAGE_MASK & ~SWAP_RA_HITS_MASK)
47
48#define SWAP_RA_HITS(v) ((v) & SWAP_RA_HITS_MASK)
49#define SWAP_RA_WIN(v) (((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
50#define SWAP_RA_ADDR(v) ((v) & PAGE_MASK)
51
52#define SWAP_RA_VAL(addr, win, hits) \
53 (((addr) & PAGE_MASK) | \
54 (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) | \
55 ((hits) & SWAP_RA_HITS_MASK))
56
57/* Initial readahead hits is 4 to start up with a small window */
58#define GET_SWAP_RA_VAL(vma) \
59 (atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
60
61#define INC_CACHE_INFO(x) data_race(swap_cache_info.x++)
62#define ADD_CACHE_INFO(x, nr) data_race(swap_cache_info.x += (nr))
63
64static struct {
65 unsigned long add_total;
66 unsigned long del_total;
67 unsigned long find_success;
68 unsigned long find_total;
69} swap_cache_info;
70
71static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
72
73void show_swap_cache_info(void)
74{
75 printk("%lu pages in swap cache\n", total_swapcache_pages());
76 printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
77 swap_cache_info.add_total, swap_cache_info.del_total,
78 swap_cache_info.find_success, swap_cache_info.find_total);
79 printk("Free swap = %ldkB\n",
80 get_nr_swap_pages() << (PAGE_SHIFT - 10));
81 printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
82}
83
84void *get_shadow_from_swap_cache(swp_entry_t entry)
85{
86 struct address_space *address_space = swap_address_space(entry);
87 pgoff_t idx = swp_offset(entry);
88 struct page *page;
89
90 page = xa_load(&address_space->i_pages, idx);
91 if (xa_is_value(page))
92 return page;
93 return NULL;
94}
95
96/*
97 * add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
98 * but sets SwapCache flag and private instead of mapping and index.
99 */
100int add_to_swap_cache(struct page *page, swp_entry_t entry,
101 gfp_t gfp, void **shadowp)
102{
103 struct address_space *address_space = swap_address_space(entry);
104 pgoff_t idx = swp_offset(entry);
105 XA_STATE_ORDER(xas, &address_space->i_pages, idx, compound_order(page));
106 unsigned long i, nr = thp_nr_pages(page);
107 void *old;
108
109 VM_BUG_ON_PAGE(!PageLocked(page), page);
110 VM_BUG_ON_PAGE(PageSwapCache(page), page);
111 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
112
113 page_ref_add(page, nr);
114 SetPageSwapCache(page);
115
116 do {
117 xas_lock_irq(&xas);
118 xas_create_range(&xas);
119 if (xas_error(&xas))
120 goto unlock;
121 for (i = 0; i < nr; i++) {
122 VM_BUG_ON_PAGE(xas.xa_index != idx + i, page);
123 old = xas_load(&xas);
124 if (xa_is_value(old)) {
125 if (shadowp)
126 *shadowp = old;
127 }
128 set_page_private(page + i, entry.val + i);
129 xas_store(&xas, page);
130 xas_next(&xas);
131 }
132 address_space->nrpages += nr;
133 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
134 __mod_lruvec_page_state(page, NR_SWAPCACHE, nr);
135 ADD_CACHE_INFO(add_total, nr);
136unlock:
137 xas_unlock_irq(&xas);
138 } while (xas_nomem(&xas, gfp));
139
140 if (!xas_error(&xas))
141 return 0;
142
143 ClearPageSwapCache(page);
144 page_ref_sub(page, nr);
145 return xas_error(&xas);
146}
147
148/*
149 * This must be called only on pages that have
150 * been verified to be in the swap cache.
151 */
152void __delete_from_swap_cache(struct page *page,
153 swp_entry_t entry, void *shadow)
154{
155 struct address_space *address_space = swap_address_space(entry);
156 int i, nr = thp_nr_pages(page);
157 pgoff_t idx = swp_offset(entry);
158 XA_STATE(xas, &address_space->i_pages, idx);
159
160 VM_BUG_ON_PAGE(!PageLocked(page), page);
161 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
162 VM_BUG_ON_PAGE(PageWriteback(page), page);
163
164 for (i = 0; i < nr; i++) {
165 void *entry = xas_store(&xas, shadow);
166 VM_BUG_ON_PAGE(entry != page, entry);
167 set_page_private(page + i, 0);
168 xas_next(&xas);
169 }
170 ClearPageSwapCache(page);
171 address_space->nrpages -= nr;
172 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
173 __mod_lruvec_page_state(page, NR_SWAPCACHE, -nr);
174 ADD_CACHE_INFO(del_total, nr);
175}
176
177/**
178 * add_to_swap - allocate swap space for a page
179 * @page: page we want to move to swap
180 *
181 * Allocate swap space for the page and add the page to the
182 * swap cache. Caller needs to hold the page lock.
183 */
184int add_to_swap(struct page *page)
185{
186 swp_entry_t entry;
187 int err;
188
189 VM_BUG_ON_PAGE(!PageLocked(page), page);
190 VM_BUG_ON_PAGE(!PageUptodate(page), page);
191
192 entry = get_swap_page(page);
193 if (!entry.val)
194 return 0;
195
196 /*
197 * XArray node allocations from PF_MEMALLOC contexts could
198 * completely exhaust the page allocator. __GFP_NOMEMALLOC
199 * stops emergency reserves from being allocated.
200 *
201 * TODO: this could cause a theoretical memory reclaim
202 * deadlock in the swap out path.
203 */
204 /*
205 * Add it to the swap cache.
206 */
207 err = add_to_swap_cache(page, entry,
208 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN, NULL);
209 if (err)
210 /*
211 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
212 * clear SWAP_HAS_CACHE flag.
213 */
214 goto fail;
215 /*
216 * Normally the page will be dirtied in unmap because its pte should be
217 * dirty. A special case is MADV_FREE page. The page's pte could have
218 * dirty bit cleared but the page's SwapBacked bit is still set because
219 * clearing the dirty bit and SwapBacked bit has no lock protected. For
220 * such page, unmap will not set dirty bit for it, so page reclaim will
221 * not write the page out. This can cause data corruption when the page
222 * is swap in later. Always setting the dirty bit for the page solves
223 * the problem.
224 */
225 set_page_dirty(page);
226
227 return 1;
228
229fail:
230 put_swap_page(page, entry);
231 return 0;
232}
233
234/*
235 * This must be called only on pages that have
236 * been verified to be in the swap cache and locked.
237 * It will never put the page into the free list,
238 * the caller has a reference on the page.
239 */
240void delete_from_swap_cache(struct page *page)
241{
242 swp_entry_t entry = { .val = page_private(page) };
243 struct address_space *address_space = swap_address_space(entry);
244
245 xa_lock_irq(&address_space->i_pages);
246 __delete_from_swap_cache(page, entry, NULL);
247 xa_unlock_irq(&address_space->i_pages);
248
249 put_swap_page(page, entry);
250 page_ref_sub(page, thp_nr_pages(page));
251}
252
253void clear_shadow_from_swap_cache(int type, unsigned long begin,
254 unsigned long end)
255{
256 unsigned long curr = begin;
257 void *old;
258
259 for (;;) {
260 swp_entry_t entry = swp_entry(type, curr);
261 struct address_space *address_space = swap_address_space(entry);
262 XA_STATE(xas, &address_space->i_pages, curr);
263
264 xa_lock_irq(&address_space->i_pages);
265 xas_for_each(&xas, old, end) {
266 if (!xa_is_value(old))
267 continue;
268 xas_store(&xas, NULL);
269 }
270 xa_unlock_irq(&address_space->i_pages);
271
272 /* search the next swapcache until we meet end */
273 curr >>= SWAP_ADDRESS_SPACE_SHIFT;
274 curr++;
275 curr <<= SWAP_ADDRESS_SPACE_SHIFT;
276 if (curr > end)
277 break;
278 }
279}
280
281/*
282 * If we are the only user, then try to free up the swap cache.
283 *
284 * Its ok to check for PageSwapCache without the page lock
285 * here because we are going to recheck again inside
286 * try_to_free_swap() _with_ the lock.
287 * - Marcelo
288 */
289void free_swap_cache(struct page *page)
290{
291 if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
292 try_to_free_swap(page);
293 unlock_page(page);
294 }
295}
296
297/*
298 * Perform a free_page(), also freeing any swap cache associated with
299 * this page if it is the last user of the page.
300 */
301void free_page_and_swap_cache(struct page *page)
302{
303 free_swap_cache(page);
304 if (!is_huge_zero_page(page))
305 put_page(page);
306}
307
308/*
309 * Passed an array of pages, drop them all from swapcache and then release
310 * them. They are removed from the LRU and freed if this is their last use.
311 */
312void free_pages_and_swap_cache(struct page **pages, int nr)
313{
314 struct page **pagep = pages;
315 int i;
316
317 lru_add_drain();
318 for (i = 0; i < nr; i++)
319 free_swap_cache(pagep[i]);
320 release_pages(pagep, nr);
321}
322
323static inline bool swap_use_vma_readahead(void)
324{
325 return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
326}
327
328/*
329 * Lookup a swap entry in the swap cache. A found page will be returned
330 * unlocked and with its refcount incremented - we rely on the kernel
331 * lock getting page table operations atomic even if we drop the page
332 * lock before returning.
333 */
334struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma,
335 unsigned long addr)
336{
337 struct page *page;
338 struct swap_info_struct *si;
339
340 si = get_swap_device(entry);
341 if (!si)
342 return NULL;
343 page = find_get_page(swap_address_space(entry), swp_offset(entry));
344 put_swap_device(si);
345
346 INC_CACHE_INFO(find_total);
347 if (page) {
348 bool vma_ra = swap_use_vma_readahead();
349 bool readahead;
350
351 INC_CACHE_INFO(find_success);
352 /*
353 * At the moment, we don't support PG_readahead for anon THP
354 * so let's bail out rather than confusing the readahead stat.
355 */
356 if (unlikely(PageTransCompound(page)))
357 return page;
358
359 readahead = TestClearPageReadahead(page);
360 if (vma && vma_ra) {
361 unsigned long ra_val;
362 int win, hits;
363
364 ra_val = GET_SWAP_RA_VAL(vma);
365 win = SWAP_RA_WIN(ra_val);
366 hits = SWAP_RA_HITS(ra_val);
367 if (readahead)
368 hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
369 atomic_long_set(&vma->swap_readahead_info,
370 SWAP_RA_VAL(addr, win, hits));
371 }
372
373 if (readahead) {
374 count_vm_event(SWAP_RA_HIT);
375 if (!vma || !vma_ra)
376 atomic_inc(&swapin_readahead_hits);
377 }
378 }
379
380 return page;
381}
382
383/**
384 * find_get_incore_page - Find and get a page from the page or swap caches.
385 * @mapping: The address_space to search.
386 * @index: The page cache index.
387 *
388 * This differs from find_get_page() in that it will also look for the
389 * page in the swap cache.
390 *
391 * Return: The found page or %NULL.
392 */
393struct page *find_get_incore_page(struct address_space *mapping, pgoff_t index)
394{
395 swp_entry_t swp;
396 struct swap_info_struct *si;
397 struct page *page = pagecache_get_page(mapping, index,
398 FGP_ENTRY | FGP_HEAD, 0);
399
400 if (!page)
401 return page;
402 if (!xa_is_value(page))
403 return find_subpage(page, index);
404 if (!shmem_mapping(mapping))
405 return NULL;
406
407 swp = radix_to_swp_entry(page);
408 /* Prevent swapoff from happening to us */
409 si = get_swap_device(swp);
410 if (!si)
411 return NULL;
412 page = find_get_page(swap_address_space(swp), swp_offset(swp));
413 put_swap_device(si);
414 return page;
415}
416
417struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
418 struct vm_area_struct *vma, unsigned long addr,
419 bool *new_page_allocated)
420{
421 struct swap_info_struct *si;
422 struct page *page;
423 void *shadow = NULL;
424
425 *new_page_allocated = false;
426
427 for (;;) {
428 int err;
429 /*
430 * First check the swap cache. Since this is normally
431 * called after lookup_swap_cache() failed, re-calling
432 * that would confuse statistics.
433 */
434 si = get_swap_device(entry);
435 if (!si)
436 return NULL;
437 page = find_get_page(swap_address_space(entry),
438 swp_offset(entry));
439 put_swap_device(si);
440 if (page)
441 return page;
442
443 /*
444 * Just skip read ahead for unused swap slot.
445 * During swap_off when swap_slot_cache is disabled,
446 * we have to handle the race between putting
447 * swap entry in swap cache and marking swap slot
448 * as SWAP_HAS_CACHE. That's done in later part of code or
449 * else swap_off will be aborted if we return NULL.
450 */
451 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
452 return NULL;
453
454 /*
455 * Get a new page to read into from swap. Allocate it now,
456 * before marking swap_map SWAP_HAS_CACHE, when -EEXIST will
457 * cause any racers to loop around until we add it to cache.
458 */
459 page = alloc_page_vma(gfp_mask, vma, addr);
460 if (!page)
461 return NULL;
462
463 /*
464 * Swap entry may have been freed since our caller observed it.
465 */
466 err = swapcache_prepare(entry);
467 if (!err)
468 break;
469
470 put_page(page);
471 if (err != -EEXIST)
472 return NULL;
473
474 /*
475 * We might race against __delete_from_swap_cache(), and
476 * stumble across a swap_map entry whose SWAP_HAS_CACHE
477 * has not yet been cleared. Or race against another
478 * __read_swap_cache_async(), which has set SWAP_HAS_CACHE
479 * in swap_map, but not yet added its page to swap cache.
480 */
481 cond_resched();
482 }
483
484 /*
485 * The swap entry is ours to swap in. Prepare the new page.
486 */
487
488 __SetPageLocked(page);
489 __SetPageSwapBacked(page);
490
491 if (mem_cgroup_swapin_charge_page(page, NULL, gfp_mask, entry))
492 goto fail_unlock;
493
494 /* May fail (-ENOMEM) if XArray node allocation failed. */
495 if (add_to_swap_cache(page, entry, gfp_mask & GFP_RECLAIM_MASK, &shadow))
496 goto fail_unlock;
497
498 mem_cgroup_swapin_uncharge_swap(entry);
499
500 if (shadow)
501 workingset_refault(page, shadow);
502
503 /* Caller will initiate read into locked page */
504 lru_cache_add(page);
505 *new_page_allocated = true;
506 return page;
507
508fail_unlock:
509 put_swap_page(page, entry);
510 unlock_page(page);
511 put_page(page);
512 return NULL;
513}
514
515/*
516 * Locate a page of swap in physical memory, reserving swap cache space
517 * and reading the disk if it is not already cached.
518 * A failure return means that either the page allocation failed or that
519 * the swap entry is no longer in use.
520 */
521struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
522 struct vm_area_struct *vma, unsigned long addr, bool do_poll)
523{
524 bool page_was_allocated;
525 struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
526 vma, addr, &page_was_allocated);
527
528 if (page_was_allocated)
529 swap_readpage(retpage, do_poll);
530
531 return retpage;
532}
533
534static unsigned int __swapin_nr_pages(unsigned long prev_offset,
535 unsigned long offset,
536 int hits,
537 int max_pages,
538 int prev_win)
539{
540 unsigned int pages, last_ra;
541
542 /*
543 * This heuristic has been found to work well on both sequential and
544 * random loads, swapping to hard disk or to SSD: please don't ask
545 * what the "+ 2" means, it just happens to work well, that's all.
546 */
547 pages = hits + 2;
548 if (pages == 2) {
549 /*
550 * We can have no readahead hits to judge by: but must not get
551 * stuck here forever, so check for an adjacent offset instead
552 * (and don't even bother to check whether swap type is same).
553 */
554 if (offset != prev_offset + 1 && offset != prev_offset - 1)
555 pages = 1;
556 } else {
557 unsigned int roundup = 4;
558 while (roundup < pages)
559 roundup <<= 1;
560 pages = roundup;
561 }
562
563 if (pages > max_pages)
564 pages = max_pages;
565
566 /* Don't shrink readahead too fast */
567 last_ra = prev_win / 2;
568 if (pages < last_ra)
569 pages = last_ra;
570
571 return pages;
572}
573
574static unsigned long swapin_nr_pages(unsigned long offset)
575{
576 static unsigned long prev_offset;
577 unsigned int hits, pages, max_pages;
578 static atomic_t last_readahead_pages;
579
580 max_pages = 1 << READ_ONCE(page_cluster);
581 if (max_pages <= 1)
582 return 1;
583
584 hits = atomic_xchg(&swapin_readahead_hits, 0);
585 pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits,
586 max_pages,
587 atomic_read(&last_readahead_pages));
588 if (!hits)
589 WRITE_ONCE(prev_offset, offset);
590 atomic_set(&last_readahead_pages, pages);
591
592 return pages;
593}
594
595/**
596 * swap_cluster_readahead - swap in pages in hope we need them soon
597 * @entry: swap entry of this memory
598 * @gfp_mask: memory allocation flags
599 * @vmf: fault information
600 *
601 * Returns the struct page for entry and addr, after queueing swapin.
602 *
603 * Primitive swap readahead code. We simply read an aligned block of
604 * (1 << page_cluster) entries in the swap area. This method is chosen
605 * because it doesn't cost us any seek time. We also make sure to queue
606 * the 'original' request together with the readahead ones...
607 *
608 * This has been extended to use the NUMA policies from the mm triggering
609 * the readahead.
610 *
611 * Caller must hold read mmap_lock if vmf->vma is not NULL.
612 */
613struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
614 struct vm_fault *vmf)
615{
616 struct page *page;
617 unsigned long entry_offset = swp_offset(entry);
618 unsigned long offset = entry_offset;
619 unsigned long start_offset, end_offset;
620 unsigned long mask;
621 struct swap_info_struct *si = swp_swap_info(entry);
622 struct blk_plug plug;
623 bool do_poll = true, page_allocated;
624 struct vm_area_struct *vma = vmf->vma;
625 unsigned long addr = vmf->address;
626
627 mask = swapin_nr_pages(offset) - 1;
628 if (!mask)
629 goto skip;
630
631 do_poll = false;
632 /* Read a page_cluster sized and aligned cluster around offset. */
633 start_offset = offset & ~mask;
634 end_offset = offset | mask;
635 if (!start_offset) /* First page is swap header. */
636 start_offset++;
637 if (end_offset >= si->max)
638 end_offset = si->max - 1;
639
640 blk_start_plug(&plug);
641 for (offset = start_offset; offset <= end_offset ; offset++) {
642 /* Ok, do the async read-ahead now */
643 page = __read_swap_cache_async(
644 swp_entry(swp_type(entry), offset),
645 gfp_mask, vma, addr, &page_allocated);
646 if (!page)
647 continue;
648 if (page_allocated) {
649 swap_readpage(page, false);
650 if (offset != entry_offset) {
651 SetPageReadahead(page);
652 count_vm_event(SWAP_RA);
653 }
654 }
655 put_page(page);
656 }
657 blk_finish_plug(&plug);
658
659 lru_add_drain(); /* Push any new pages onto the LRU now */
660skip:
661 return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll);
662}
663
664int init_swap_address_space(unsigned int type, unsigned long nr_pages)
665{
666 struct address_space *spaces, *space;
667 unsigned int i, nr;
668
669 nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
670 spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL);
671 if (!spaces)
672 return -ENOMEM;
673 for (i = 0; i < nr; i++) {
674 space = spaces + i;
675 xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ);
676 atomic_set(&space->i_mmap_writable, 0);
677 space->a_ops = &swap_aops;
678 /* swap cache doesn't use writeback related tags */
679 mapping_set_no_writeback_tags(space);
680 }
681 nr_swapper_spaces[type] = nr;
682 swapper_spaces[type] = spaces;
683
684 return 0;
685}
686
687void exit_swap_address_space(unsigned int type)
688{
689 int i;
690 struct address_space *spaces = swapper_spaces[type];
691
692 for (i = 0; i < nr_swapper_spaces[type]; i++)
693 VM_WARN_ON_ONCE(!mapping_empty(&spaces[i]));
694 kvfree(spaces);
695 nr_swapper_spaces[type] = 0;
696 swapper_spaces[type] = NULL;
697}
698
699static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
700 unsigned long faddr,
701 unsigned long lpfn,
702 unsigned long rpfn,
703 unsigned long *start,
704 unsigned long *end)
705{
706 *start = max3(lpfn, PFN_DOWN(vma->vm_start),
707 PFN_DOWN(faddr & PMD_MASK));
708 *end = min3(rpfn, PFN_DOWN(vma->vm_end),
709 PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
710}
711
712static void swap_ra_info(struct vm_fault *vmf,
713 struct vma_swap_readahead *ra_info)
714{
715 struct vm_area_struct *vma = vmf->vma;
716 unsigned long ra_val;
717 unsigned long faddr, pfn, fpfn;
718 unsigned long start, end;
719 pte_t *pte, *orig_pte;
720 unsigned int max_win, hits, prev_win, win, left;
721#ifndef CONFIG_64BIT
722 pte_t *tpte;
723#endif
724
725 max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
726 SWAP_RA_ORDER_CEILING);
727 if (max_win == 1) {
728 ra_info->win = 1;
729 return;
730 }
731
732 faddr = vmf->address;
733 orig_pte = pte = pte_offset_map(vmf->pmd, faddr);
734
735 fpfn = PFN_DOWN(faddr);
736 ra_val = GET_SWAP_RA_VAL(vma);
737 pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val));
738 prev_win = SWAP_RA_WIN(ra_val);
739 hits = SWAP_RA_HITS(ra_val);
740 ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits,
741 max_win, prev_win);
742 atomic_long_set(&vma->swap_readahead_info,
743 SWAP_RA_VAL(faddr, win, 0));
744
745 if (win == 1) {
746 pte_unmap(orig_pte);
747 return;
748 }
749
750 /* Copy the PTEs because the page table may be unmapped */
751 if (fpfn == pfn + 1)
752 swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end);
753 else if (pfn == fpfn + 1)
754 swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1,
755 &start, &end);
756 else {
757 left = (win - 1) / 2;
758 swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
759 &start, &end);
760 }
761 ra_info->nr_pte = end - start;
762 ra_info->offset = fpfn - start;
763 pte -= ra_info->offset;
764#ifdef CONFIG_64BIT
765 ra_info->ptes = pte;
766#else
767 tpte = ra_info->ptes;
768 for (pfn = start; pfn != end; pfn++)
769 *tpte++ = *pte++;
770#endif
771 pte_unmap(orig_pte);
772}
773
774/**
775 * swap_vma_readahead - swap in pages in hope we need them soon
776 * @fentry: swap entry of this memory
777 * @gfp_mask: memory allocation flags
778 * @vmf: fault information
779 *
780 * Returns the struct page for entry and addr, after queueing swapin.
781 *
782 * Primitive swap readahead code. We simply read in a few pages whose
783 * virtual addresses are around the fault address in the same vma.
784 *
785 * Caller must hold read mmap_lock if vmf->vma is not NULL.
786 *
787 */
788static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
789 struct vm_fault *vmf)
790{
791 struct blk_plug plug;
792 struct vm_area_struct *vma = vmf->vma;
793 struct page *page;
794 pte_t *pte, pentry;
795 swp_entry_t entry;
796 unsigned int i;
797 bool page_allocated;
798 struct vma_swap_readahead ra_info = {
799 .win = 1,
800 };
801
802 swap_ra_info(vmf, &ra_info);
803 if (ra_info.win == 1)
804 goto skip;
805
806 blk_start_plug(&plug);
807 for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte;
808 i++, pte++) {
809 pentry = *pte;
810 if (pte_none(pentry))
811 continue;
812 if (pte_present(pentry))
813 continue;
814 entry = pte_to_swp_entry(pentry);
815 if (unlikely(non_swap_entry(entry)))
816 continue;
817 page = __read_swap_cache_async(entry, gfp_mask, vma,
818 vmf->address, &page_allocated);
819 if (!page)
820 continue;
821 if (page_allocated) {
822 swap_readpage(page, false);
823 if (i != ra_info.offset) {
824 SetPageReadahead(page);
825 count_vm_event(SWAP_RA);
826 }
827 }
828 put_page(page);
829 }
830 blk_finish_plug(&plug);
831 lru_add_drain();
832skip:
833 return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
834 ra_info.win == 1);
835}
836
837/**
838 * swapin_readahead - swap in pages in hope we need them soon
839 * @entry: swap entry of this memory
840 * @gfp_mask: memory allocation flags
841 * @vmf: fault information
842 *
843 * Returns the struct page for entry and addr, after queueing swapin.
844 *
845 * It's a main entry function for swap readahead. By the configuration,
846 * it will read ahead blocks by cluster-based(ie, physical disk based)
847 * or vma-based(ie, virtual address based on faulty address) readahead.
848 */
849struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
850 struct vm_fault *vmf)
851{
852 return swap_use_vma_readahead() ?
853 swap_vma_readahead(entry, gfp_mask, vmf) :
854 swap_cluster_readahead(entry, gfp_mask, vmf);
855}
856
857#ifdef CONFIG_SYSFS
858static ssize_t vma_ra_enabled_show(struct kobject *kobj,
859 struct kobj_attribute *attr, char *buf)
860{
861 return sysfs_emit(buf, "%s\n",
862 enable_vma_readahead ? "true" : "false");
863}
864static ssize_t vma_ra_enabled_store(struct kobject *kobj,
865 struct kobj_attribute *attr,
866 const char *buf, size_t count)
867{
868 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
869 enable_vma_readahead = true;
870 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
871 enable_vma_readahead = false;
872 else
873 return -EINVAL;
874
875 return count;
876}
877static struct kobj_attribute vma_ra_enabled_attr =
878 __ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show,
879 vma_ra_enabled_store);
880
881static struct attribute *swap_attrs[] = {
882 &vma_ra_enabled_attr.attr,
883 NULL,
884};
885
886static const struct attribute_group swap_attr_group = {
887 .attrs = swap_attrs,
888};
889
890static int __init swap_init_sysfs(void)
891{
892 int err;
893 struct kobject *swap_kobj;
894
895 swap_kobj = kobject_create_and_add("swap", mm_kobj);
896 if (!swap_kobj) {
897 pr_err("failed to create swap kobject\n");
898 return -ENOMEM;
899 }
900 err = sysfs_create_group(swap_kobj, &swap_attr_group);
901 if (err) {
902 pr_err("failed to register swap group\n");
903 goto delete_obj;
904 }
905 return 0;
906
907delete_obj:
908 kobject_put(swap_kobj);
909 return err;
910}
911subsys_initcall(swap_init_sysfs);
912#endif
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/swap_slots.h>
24#include <linux/huge_mm.h>
25#include <linux/shmem_fs.h>
26#include "internal.h"
27#include "swap.h"
28
29/*
30 * swapper_space is a fiction, retained to simplify the path through
31 * vmscan's shrink_folio_list.
32 */
33static const struct address_space_operations swap_aops = {
34 .writepage = swap_writepage,
35 .dirty_folio = noop_dirty_folio,
36#ifdef CONFIG_MIGRATION
37 .migrate_folio = migrate_folio,
38#endif
39};
40
41struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly;
42static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly;
43static bool enable_vma_readahead __read_mostly = true;
44
45#define SWAP_RA_ORDER_CEILING 5
46
47#define SWAP_RA_WIN_SHIFT (PAGE_SHIFT / 2)
48#define SWAP_RA_HITS_MASK ((1UL << SWAP_RA_WIN_SHIFT) - 1)
49#define SWAP_RA_HITS_MAX SWAP_RA_HITS_MASK
50#define SWAP_RA_WIN_MASK (~PAGE_MASK & ~SWAP_RA_HITS_MASK)
51
52#define SWAP_RA_HITS(v) ((v) & SWAP_RA_HITS_MASK)
53#define SWAP_RA_WIN(v) (((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
54#define SWAP_RA_ADDR(v) ((v) & PAGE_MASK)
55
56#define SWAP_RA_VAL(addr, win, hits) \
57 (((addr) & PAGE_MASK) | \
58 (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) | \
59 ((hits) & SWAP_RA_HITS_MASK))
60
61/* Initial readahead hits is 4 to start up with a small window */
62#define GET_SWAP_RA_VAL(vma) \
63 (atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
64
65static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
66
67void show_swap_cache_info(void)
68{
69 printk("%lu pages in swap cache\n", total_swapcache_pages());
70 printk("Free swap = %ldkB\n", K(get_nr_swap_pages()));
71 printk("Total swap = %lukB\n", K(total_swap_pages));
72}
73
74void *get_shadow_from_swap_cache(swp_entry_t entry)
75{
76 struct address_space *address_space = swap_address_space(entry);
77 pgoff_t idx = swap_cache_index(entry);
78 void *shadow;
79
80 shadow = xa_load(&address_space->i_pages, idx);
81 if (xa_is_value(shadow))
82 return shadow;
83 return NULL;
84}
85
86/*
87 * add_to_swap_cache resembles filemap_add_folio on swapper_space,
88 * but sets SwapCache flag and private instead of mapping and index.
89 */
90int add_to_swap_cache(struct folio *folio, swp_entry_t entry,
91 gfp_t gfp, void **shadowp)
92{
93 struct address_space *address_space = swap_address_space(entry);
94 pgoff_t idx = swap_cache_index(entry);
95 XA_STATE_ORDER(xas, &address_space->i_pages, idx, folio_order(folio));
96 unsigned long i, nr = folio_nr_pages(folio);
97 void *old;
98
99 xas_set_update(&xas, workingset_update_node);
100
101 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
102 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
103 VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio);
104
105 folio_ref_add(folio, nr);
106 folio_set_swapcache(folio);
107 folio->swap = entry;
108
109 do {
110 xas_lock_irq(&xas);
111 xas_create_range(&xas);
112 if (xas_error(&xas))
113 goto unlock;
114 for (i = 0; i < nr; i++) {
115 VM_BUG_ON_FOLIO(xas.xa_index != idx + i, folio);
116 if (shadowp) {
117 old = xas_load(&xas);
118 if (xa_is_value(old))
119 *shadowp = old;
120 }
121 xas_store(&xas, folio);
122 xas_next(&xas);
123 }
124 address_space->nrpages += nr;
125 __node_stat_mod_folio(folio, NR_FILE_PAGES, nr);
126 __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, nr);
127unlock:
128 xas_unlock_irq(&xas);
129 } while (xas_nomem(&xas, gfp));
130
131 if (!xas_error(&xas))
132 return 0;
133
134 folio_clear_swapcache(folio);
135 folio_ref_sub(folio, nr);
136 return xas_error(&xas);
137}
138
139/*
140 * This must be called only on folios that have
141 * been verified to be in the swap cache.
142 */
143void __delete_from_swap_cache(struct folio *folio,
144 swp_entry_t entry, void *shadow)
145{
146 struct address_space *address_space = swap_address_space(entry);
147 int i;
148 long nr = folio_nr_pages(folio);
149 pgoff_t idx = swap_cache_index(entry);
150 XA_STATE(xas, &address_space->i_pages, idx);
151
152 xas_set_update(&xas, workingset_update_node);
153
154 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
155 VM_BUG_ON_FOLIO(!folio_test_swapcache(folio), folio);
156 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
157
158 for (i = 0; i < nr; i++) {
159 void *entry = xas_store(&xas, shadow);
160 VM_BUG_ON_PAGE(entry != folio, entry);
161 xas_next(&xas);
162 }
163 folio->swap.val = 0;
164 folio_clear_swapcache(folio);
165 address_space->nrpages -= nr;
166 __node_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
167 __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, -nr);
168}
169
170/**
171 * add_to_swap - allocate swap space for a folio
172 * @folio: folio we want to move to swap
173 *
174 * Allocate swap space for the folio and add the folio to the
175 * swap cache.
176 *
177 * Context: Caller needs to hold the folio lock.
178 * Return: Whether the folio was added to the swap cache.
179 */
180bool add_to_swap(struct folio *folio)
181{
182 swp_entry_t entry;
183 int err;
184
185 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
186 VM_BUG_ON_FOLIO(!folio_test_uptodate(folio), folio);
187
188 entry = folio_alloc_swap(folio);
189 if (!entry.val)
190 return false;
191
192 /*
193 * XArray node allocations from PF_MEMALLOC contexts could
194 * completely exhaust the page allocator. __GFP_NOMEMALLOC
195 * stops emergency reserves from being allocated.
196 *
197 * TODO: this could cause a theoretical memory reclaim
198 * deadlock in the swap out path.
199 */
200 /*
201 * Add it to the swap cache.
202 */
203 err = add_to_swap_cache(folio, entry,
204 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN, NULL);
205 if (err)
206 /*
207 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
208 * clear SWAP_HAS_CACHE flag.
209 */
210 goto fail;
211 /*
212 * Normally the folio will be dirtied in unmap because its
213 * pte should be dirty. A special case is MADV_FREE page. The
214 * page's pte could have dirty bit cleared but the folio's
215 * SwapBacked flag is still set because clearing the dirty bit
216 * and SwapBacked flag has no lock protected. For such folio,
217 * unmap will not set dirty bit for it, so folio reclaim will
218 * not write the folio out. This can cause data corruption when
219 * the folio is swapped in later. Always setting the dirty flag
220 * for the folio solves the problem.
221 */
222 folio_mark_dirty(folio);
223
224 return true;
225
226fail:
227 put_swap_folio(folio, entry);
228 return false;
229}
230
231/*
232 * This must be called only on folios that have
233 * been verified to be in the swap cache and locked.
234 * It will never put the folio into the free list,
235 * the caller has a reference on the folio.
236 */
237void delete_from_swap_cache(struct folio *folio)
238{
239 swp_entry_t entry = folio->swap;
240 struct address_space *address_space = swap_address_space(entry);
241
242 xa_lock_irq(&address_space->i_pages);
243 __delete_from_swap_cache(folio, entry, NULL);
244 xa_unlock_irq(&address_space->i_pages);
245
246 put_swap_folio(folio, entry);
247 folio_ref_sub(folio, folio_nr_pages(folio));
248}
249
250void clear_shadow_from_swap_cache(int type, unsigned long begin,
251 unsigned long end)
252{
253 unsigned long curr = begin;
254 void *old;
255
256 for (;;) {
257 swp_entry_t entry = swp_entry(type, curr);
258 unsigned long index = curr & SWAP_ADDRESS_SPACE_MASK;
259 struct address_space *address_space = swap_address_space(entry);
260 XA_STATE(xas, &address_space->i_pages, index);
261
262 xas_set_update(&xas, workingset_update_node);
263
264 xa_lock_irq(&address_space->i_pages);
265 xas_for_each(&xas, old, min(index + (end - curr), SWAP_ADDRESS_SPACE_PAGES)) {
266 if (!xa_is_value(old))
267 continue;
268 xas_store(&xas, NULL);
269 }
270 xa_unlock_irq(&address_space->i_pages);
271
272 /* search the next swapcache until we meet end */
273 curr >>= SWAP_ADDRESS_SPACE_SHIFT;
274 curr++;
275 curr <<= SWAP_ADDRESS_SPACE_SHIFT;
276 if (curr > end)
277 break;
278 }
279}
280
281/*
282 * If we are the only user, then try to free up the swap cache.
283 *
284 * Its ok to check the swapcache flag without the folio lock
285 * here because we are going to recheck again inside
286 * folio_free_swap() _with_ the lock.
287 * - Marcelo
288 */
289void free_swap_cache(struct folio *folio)
290{
291 if (folio_test_swapcache(folio) && !folio_mapped(folio) &&
292 folio_trylock(folio)) {
293 folio_free_swap(folio);
294 folio_unlock(folio);
295 }
296}
297
298/*
299 * Perform a free_page(), also freeing any swap cache associated with
300 * this page if it is the last user of the page.
301 */
302void free_page_and_swap_cache(struct page *page)
303{
304 struct folio *folio = page_folio(page);
305
306 free_swap_cache(folio);
307 if (!is_huge_zero_folio(folio))
308 folio_put(folio);
309}
310
311/*
312 * Passed an array of pages, drop them all from swapcache and then release
313 * them. They are removed from the LRU and freed if this is their last use.
314 */
315void free_pages_and_swap_cache(struct encoded_page **pages, int nr)
316{
317 struct folio_batch folios;
318 unsigned int refs[PAGEVEC_SIZE];
319
320 lru_add_drain();
321 folio_batch_init(&folios);
322 for (int i = 0; i < nr; i++) {
323 struct folio *folio = page_folio(encoded_page_ptr(pages[i]));
324
325 free_swap_cache(folio);
326 refs[folios.nr] = 1;
327 if (unlikely(encoded_page_flags(pages[i]) &
328 ENCODED_PAGE_BIT_NR_PAGES_NEXT))
329 refs[folios.nr] = encoded_nr_pages(pages[++i]);
330
331 if (folio_batch_add(&folios, folio) == 0)
332 folios_put_refs(&folios, refs);
333 }
334 if (folios.nr)
335 folios_put_refs(&folios, refs);
336}
337
338static inline bool swap_use_vma_readahead(void)
339{
340 return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
341}
342
343/*
344 * Lookup a swap entry in the swap cache. A found folio will be returned
345 * unlocked and with its refcount incremented - we rely on the kernel
346 * lock getting page table operations atomic even if we drop the folio
347 * lock before returning.
348 *
349 * Caller must lock the swap device or hold a reference to keep it valid.
350 */
351struct folio *swap_cache_get_folio(swp_entry_t entry,
352 struct vm_area_struct *vma, unsigned long addr)
353{
354 struct folio *folio;
355
356 folio = filemap_get_folio(swap_address_space(entry), swap_cache_index(entry));
357 if (!IS_ERR(folio)) {
358 bool vma_ra = swap_use_vma_readahead();
359 bool readahead;
360
361 /*
362 * At the moment, we don't support PG_readahead for anon THP
363 * so let's bail out rather than confusing the readahead stat.
364 */
365 if (unlikely(folio_test_large(folio)))
366 return folio;
367
368 readahead = folio_test_clear_readahead(folio);
369 if (vma && vma_ra) {
370 unsigned long ra_val;
371 int win, hits;
372
373 ra_val = GET_SWAP_RA_VAL(vma);
374 win = SWAP_RA_WIN(ra_val);
375 hits = SWAP_RA_HITS(ra_val);
376 if (readahead)
377 hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
378 atomic_long_set(&vma->swap_readahead_info,
379 SWAP_RA_VAL(addr, win, hits));
380 }
381
382 if (readahead) {
383 count_vm_event(SWAP_RA_HIT);
384 if (!vma || !vma_ra)
385 atomic_inc(&swapin_readahead_hits);
386 }
387 } else {
388 folio = NULL;
389 }
390
391 return folio;
392}
393
394/**
395 * filemap_get_incore_folio - Find and get a folio from the page or swap caches.
396 * @mapping: The address_space to search.
397 * @index: The page cache index.
398 *
399 * This differs from filemap_get_folio() in that it will also look for the
400 * folio in the swap cache.
401 *
402 * Return: The found folio or %NULL.
403 */
404struct folio *filemap_get_incore_folio(struct address_space *mapping,
405 pgoff_t index)
406{
407 swp_entry_t swp;
408 struct swap_info_struct *si;
409 struct folio *folio = filemap_get_entry(mapping, index);
410
411 if (!folio)
412 return ERR_PTR(-ENOENT);
413 if (!xa_is_value(folio))
414 return folio;
415 if (!shmem_mapping(mapping))
416 return ERR_PTR(-ENOENT);
417
418 swp = radix_to_swp_entry(folio);
419 /* There might be swapin error entries in shmem mapping. */
420 if (non_swap_entry(swp))
421 return ERR_PTR(-ENOENT);
422 /* Prevent swapoff from happening to us */
423 si = get_swap_device(swp);
424 if (!si)
425 return ERR_PTR(-ENOENT);
426 index = swap_cache_index(swp);
427 folio = filemap_get_folio(swap_address_space(swp), index);
428 put_swap_device(si);
429 return folio;
430}
431
432struct folio *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
433 struct mempolicy *mpol, pgoff_t ilx, bool *new_page_allocated,
434 bool skip_if_exists)
435{
436 struct swap_info_struct *si;
437 struct folio *folio;
438 struct folio *new_folio = NULL;
439 struct folio *result = NULL;
440 void *shadow = NULL;
441
442 *new_page_allocated = false;
443 si = get_swap_device(entry);
444 if (!si)
445 return NULL;
446
447 for (;;) {
448 int err;
449 /*
450 * First check the swap cache. Since this is normally
451 * called after swap_cache_get_folio() failed, re-calling
452 * that would confuse statistics.
453 */
454 folio = filemap_get_folio(swap_address_space(entry),
455 swap_cache_index(entry));
456 if (!IS_ERR(folio))
457 goto got_folio;
458
459 /*
460 * Just skip read ahead for unused swap slot.
461 * During swap_off when swap_slot_cache is disabled,
462 * we have to handle the race between putting
463 * swap entry in swap cache and marking swap slot
464 * as SWAP_HAS_CACHE. That's done in later part of code or
465 * else swap_off will be aborted if we return NULL.
466 */
467 if (!swap_swapcount(si, entry) && swap_slot_cache_enabled)
468 goto put_and_return;
469
470 /*
471 * Get a new folio to read into from swap. Allocate it now if
472 * new_folio not exist, before marking swap_map SWAP_HAS_CACHE,
473 * when -EEXIST will cause any racers to loop around until we
474 * add it to cache.
475 */
476 if (!new_folio) {
477 new_folio = folio_alloc_mpol(gfp_mask, 0, mpol, ilx, numa_node_id());
478 if (!new_folio)
479 goto put_and_return;
480 }
481
482 /*
483 * Swap entry may have been freed since our caller observed it.
484 */
485 err = swapcache_prepare(entry, 1);
486 if (!err)
487 break;
488 else if (err != -EEXIST)
489 goto put_and_return;
490
491 /*
492 * Protect against a recursive call to __read_swap_cache_async()
493 * on the same entry waiting forever here because SWAP_HAS_CACHE
494 * is set but the folio is not the swap cache yet. This can
495 * happen today if mem_cgroup_swapin_charge_folio() below
496 * triggers reclaim through zswap, which may call
497 * __read_swap_cache_async() in the writeback path.
498 */
499 if (skip_if_exists)
500 goto put_and_return;
501
502 /*
503 * We might race against __delete_from_swap_cache(), and
504 * stumble across a swap_map entry whose SWAP_HAS_CACHE
505 * has not yet been cleared. Or race against another
506 * __read_swap_cache_async(), which has set SWAP_HAS_CACHE
507 * in swap_map, but not yet added its folio to swap cache.
508 */
509 schedule_timeout_uninterruptible(1);
510 }
511
512 /*
513 * The swap entry is ours to swap in. Prepare the new folio.
514 */
515 __folio_set_locked(new_folio);
516 __folio_set_swapbacked(new_folio);
517
518 if (mem_cgroup_swapin_charge_folio(new_folio, NULL, gfp_mask, entry))
519 goto fail_unlock;
520
521 /* May fail (-ENOMEM) if XArray node allocation failed. */
522 if (add_to_swap_cache(new_folio, entry, gfp_mask & GFP_RECLAIM_MASK, &shadow))
523 goto fail_unlock;
524
525 mem_cgroup_swapin_uncharge_swap(entry, 1);
526
527 if (shadow)
528 workingset_refault(new_folio, shadow);
529
530 /* Caller will initiate read into locked new_folio */
531 folio_add_lru(new_folio);
532 *new_page_allocated = true;
533 folio = new_folio;
534got_folio:
535 result = folio;
536 goto put_and_return;
537
538fail_unlock:
539 put_swap_folio(new_folio, entry);
540 folio_unlock(new_folio);
541put_and_return:
542 put_swap_device(si);
543 if (!(*new_page_allocated) && new_folio)
544 folio_put(new_folio);
545 return result;
546}
547
548/*
549 * Locate a page of swap in physical memory, reserving swap cache space
550 * and reading the disk if it is not already cached.
551 * A failure return means that either the page allocation failed or that
552 * the swap entry is no longer in use.
553 *
554 * get/put_swap_device() aren't needed to call this function, because
555 * __read_swap_cache_async() call them and swap_read_folio() holds the
556 * swap cache folio lock.
557 */
558struct folio *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
559 struct vm_area_struct *vma, unsigned long addr,
560 struct swap_iocb **plug)
561{
562 bool page_allocated;
563 struct mempolicy *mpol;
564 pgoff_t ilx;
565 struct folio *folio;
566
567 mpol = get_vma_policy(vma, addr, 0, &ilx);
568 folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx,
569 &page_allocated, false);
570 mpol_cond_put(mpol);
571
572 if (page_allocated)
573 swap_read_folio(folio, plug);
574 return folio;
575}
576
577static unsigned int __swapin_nr_pages(unsigned long prev_offset,
578 unsigned long offset,
579 int hits,
580 int max_pages,
581 int prev_win)
582{
583 unsigned int pages, last_ra;
584
585 /*
586 * This heuristic has been found to work well on both sequential and
587 * random loads, swapping to hard disk or to SSD: please don't ask
588 * what the "+ 2" means, it just happens to work well, that's all.
589 */
590 pages = hits + 2;
591 if (pages == 2) {
592 /*
593 * We can have no readahead hits to judge by: but must not get
594 * stuck here forever, so check for an adjacent offset instead
595 * (and don't even bother to check whether swap type is same).
596 */
597 if (offset != prev_offset + 1 && offset != prev_offset - 1)
598 pages = 1;
599 } else {
600 unsigned int roundup = 4;
601 while (roundup < pages)
602 roundup <<= 1;
603 pages = roundup;
604 }
605
606 if (pages > max_pages)
607 pages = max_pages;
608
609 /* Don't shrink readahead too fast */
610 last_ra = prev_win / 2;
611 if (pages < last_ra)
612 pages = last_ra;
613
614 return pages;
615}
616
617static unsigned long swapin_nr_pages(unsigned long offset)
618{
619 static unsigned long prev_offset;
620 unsigned int hits, pages, max_pages;
621 static atomic_t last_readahead_pages;
622
623 max_pages = 1 << READ_ONCE(page_cluster);
624 if (max_pages <= 1)
625 return 1;
626
627 hits = atomic_xchg(&swapin_readahead_hits, 0);
628 pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits,
629 max_pages,
630 atomic_read(&last_readahead_pages));
631 if (!hits)
632 WRITE_ONCE(prev_offset, offset);
633 atomic_set(&last_readahead_pages, pages);
634
635 return pages;
636}
637
638/**
639 * swap_cluster_readahead - swap in pages in hope we need them soon
640 * @entry: swap entry of this memory
641 * @gfp_mask: memory allocation flags
642 * @mpol: NUMA memory allocation policy to be applied
643 * @ilx: NUMA interleave index, for use only when MPOL_INTERLEAVE
644 *
645 * Returns the struct folio for entry and addr, after queueing swapin.
646 *
647 * Primitive swap readahead code. We simply read an aligned block of
648 * (1 << page_cluster) entries in the swap area. This method is chosen
649 * because it doesn't cost us any seek time. We also make sure to queue
650 * the 'original' request together with the readahead ones...
651 *
652 * Note: it is intentional that the same NUMA policy and interleave index
653 * are used for every page of the readahead: neighbouring pages on swap
654 * are fairly likely to have been swapped out from the same node.
655 */
656struct folio *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
657 struct mempolicy *mpol, pgoff_t ilx)
658{
659 struct folio *folio;
660 unsigned long entry_offset = swp_offset(entry);
661 unsigned long offset = entry_offset;
662 unsigned long start_offset, end_offset;
663 unsigned long mask;
664 struct swap_info_struct *si = swp_swap_info(entry);
665 struct blk_plug plug;
666 struct swap_iocb *splug = NULL;
667 bool page_allocated;
668
669 mask = swapin_nr_pages(offset) - 1;
670 if (!mask)
671 goto skip;
672
673 /* Read a page_cluster sized and aligned cluster around offset. */
674 start_offset = offset & ~mask;
675 end_offset = offset | mask;
676 if (!start_offset) /* First page is swap header. */
677 start_offset++;
678 if (end_offset >= si->max)
679 end_offset = si->max - 1;
680
681 blk_start_plug(&plug);
682 for (offset = start_offset; offset <= end_offset ; offset++) {
683 /* Ok, do the async read-ahead now */
684 folio = __read_swap_cache_async(
685 swp_entry(swp_type(entry), offset),
686 gfp_mask, mpol, ilx, &page_allocated, false);
687 if (!folio)
688 continue;
689 if (page_allocated) {
690 swap_read_folio(folio, &splug);
691 if (offset != entry_offset) {
692 folio_set_readahead(folio);
693 count_vm_event(SWAP_RA);
694 }
695 }
696 folio_put(folio);
697 }
698 blk_finish_plug(&plug);
699 swap_read_unplug(splug);
700 lru_add_drain(); /* Push any new pages onto the LRU now */
701skip:
702 /* The page was likely read above, so no need for plugging here */
703 folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx,
704 &page_allocated, false);
705 if (unlikely(page_allocated))
706 swap_read_folio(folio, NULL);
707 return folio;
708}
709
710int init_swap_address_space(unsigned int type, unsigned long nr_pages)
711{
712 struct address_space *spaces, *space;
713 unsigned int i, nr;
714
715 nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
716 spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL);
717 if (!spaces)
718 return -ENOMEM;
719 for (i = 0; i < nr; i++) {
720 space = spaces + i;
721 xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ);
722 atomic_set(&space->i_mmap_writable, 0);
723 space->a_ops = &swap_aops;
724 /* swap cache doesn't use writeback related tags */
725 mapping_set_no_writeback_tags(space);
726 }
727 nr_swapper_spaces[type] = nr;
728 swapper_spaces[type] = spaces;
729
730 return 0;
731}
732
733void exit_swap_address_space(unsigned int type)
734{
735 int i;
736 struct address_space *spaces = swapper_spaces[type];
737
738 for (i = 0; i < nr_swapper_spaces[type]; i++)
739 VM_WARN_ON_ONCE(!mapping_empty(&spaces[i]));
740 kvfree(spaces);
741 nr_swapper_spaces[type] = 0;
742 swapper_spaces[type] = NULL;
743}
744
745static int swap_vma_ra_win(struct vm_fault *vmf, unsigned long *start,
746 unsigned long *end)
747{
748 struct vm_area_struct *vma = vmf->vma;
749 unsigned long ra_val;
750 unsigned long faddr, prev_faddr, left, right;
751 unsigned int max_win, hits, prev_win, win;
752
753 max_win = 1 << min(READ_ONCE(page_cluster), SWAP_RA_ORDER_CEILING);
754 if (max_win == 1)
755 return 1;
756
757 faddr = vmf->address;
758 ra_val = GET_SWAP_RA_VAL(vma);
759 prev_faddr = SWAP_RA_ADDR(ra_val);
760 prev_win = SWAP_RA_WIN(ra_val);
761 hits = SWAP_RA_HITS(ra_val);
762 win = __swapin_nr_pages(PFN_DOWN(prev_faddr), PFN_DOWN(faddr), hits,
763 max_win, prev_win);
764 atomic_long_set(&vma->swap_readahead_info, SWAP_RA_VAL(faddr, win, 0));
765 if (win == 1)
766 return 1;
767
768 if (faddr == prev_faddr + PAGE_SIZE)
769 left = faddr;
770 else if (prev_faddr == faddr + PAGE_SIZE)
771 left = faddr - (win << PAGE_SHIFT) + PAGE_SIZE;
772 else
773 left = faddr - (((win - 1) / 2) << PAGE_SHIFT);
774 right = left + (win << PAGE_SHIFT);
775 if ((long)left < 0)
776 left = 0;
777 *start = max3(left, vma->vm_start, faddr & PMD_MASK);
778 *end = min3(right, vma->vm_end, (faddr & PMD_MASK) + PMD_SIZE);
779
780 return win;
781}
782
783/**
784 * swap_vma_readahead - swap in pages in hope we need them soon
785 * @targ_entry: swap entry of the targeted memory
786 * @gfp_mask: memory allocation flags
787 * @mpol: NUMA memory allocation policy to be applied
788 * @targ_ilx: NUMA interleave index, for use only when MPOL_INTERLEAVE
789 * @vmf: fault information
790 *
791 * Returns the struct folio for entry and addr, after queueing swapin.
792 *
793 * Primitive swap readahead code. We simply read in a few pages whose
794 * virtual addresses are around the fault address in the same vma.
795 *
796 * Caller must hold read mmap_lock if vmf->vma is not NULL.
797 *
798 */
799static struct folio *swap_vma_readahead(swp_entry_t targ_entry, gfp_t gfp_mask,
800 struct mempolicy *mpol, pgoff_t targ_ilx, struct vm_fault *vmf)
801{
802 struct blk_plug plug;
803 struct swap_iocb *splug = NULL;
804 struct folio *folio;
805 pte_t *pte = NULL, pentry;
806 int win;
807 unsigned long start, end, addr;
808 swp_entry_t entry;
809 pgoff_t ilx;
810 bool page_allocated;
811
812 win = swap_vma_ra_win(vmf, &start, &end);
813 if (win == 1)
814 goto skip;
815
816 ilx = targ_ilx - PFN_DOWN(vmf->address - start);
817
818 blk_start_plug(&plug);
819 for (addr = start; addr < end; ilx++, addr += PAGE_SIZE) {
820 if (!pte++) {
821 pte = pte_offset_map(vmf->pmd, addr);
822 if (!pte)
823 break;
824 }
825 pentry = ptep_get_lockless(pte);
826 if (!is_swap_pte(pentry))
827 continue;
828 entry = pte_to_swp_entry(pentry);
829 if (unlikely(non_swap_entry(entry)))
830 continue;
831 pte_unmap(pte);
832 pte = NULL;
833 folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx,
834 &page_allocated, false);
835 if (!folio)
836 continue;
837 if (page_allocated) {
838 swap_read_folio(folio, &splug);
839 if (addr != vmf->address) {
840 folio_set_readahead(folio);
841 count_vm_event(SWAP_RA);
842 }
843 }
844 folio_put(folio);
845 }
846 if (pte)
847 pte_unmap(pte);
848 blk_finish_plug(&plug);
849 swap_read_unplug(splug);
850 lru_add_drain();
851skip:
852 /* The folio was likely read above, so no need for plugging here */
853 folio = __read_swap_cache_async(targ_entry, gfp_mask, mpol, targ_ilx,
854 &page_allocated, false);
855 if (unlikely(page_allocated))
856 swap_read_folio(folio, NULL);
857 return folio;
858}
859
860/**
861 * swapin_readahead - swap in pages in hope we need them soon
862 * @entry: swap entry of this memory
863 * @gfp_mask: memory allocation flags
864 * @vmf: fault information
865 *
866 * Returns the struct folio for entry and addr, after queueing swapin.
867 *
868 * It's a main entry function for swap readahead. By the configuration,
869 * it will read ahead blocks by cluster-based(ie, physical disk based)
870 * or vma-based(ie, virtual address based on faulty address) readahead.
871 */
872struct folio *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
873 struct vm_fault *vmf)
874{
875 struct mempolicy *mpol;
876 pgoff_t ilx;
877 struct folio *folio;
878
879 mpol = get_vma_policy(vmf->vma, vmf->address, 0, &ilx);
880 folio = swap_use_vma_readahead() ?
881 swap_vma_readahead(entry, gfp_mask, mpol, ilx, vmf) :
882 swap_cluster_readahead(entry, gfp_mask, mpol, ilx);
883 mpol_cond_put(mpol);
884
885 return folio;
886}
887
888#ifdef CONFIG_SYSFS
889static ssize_t vma_ra_enabled_show(struct kobject *kobj,
890 struct kobj_attribute *attr, char *buf)
891{
892 return sysfs_emit(buf, "%s\n", str_true_false(enable_vma_readahead));
893}
894static ssize_t vma_ra_enabled_store(struct kobject *kobj,
895 struct kobj_attribute *attr,
896 const char *buf, size_t count)
897{
898 ssize_t ret;
899
900 ret = kstrtobool(buf, &enable_vma_readahead);
901 if (ret)
902 return ret;
903
904 return count;
905}
906static struct kobj_attribute vma_ra_enabled_attr = __ATTR_RW(vma_ra_enabled);
907
908static struct attribute *swap_attrs[] = {
909 &vma_ra_enabled_attr.attr,
910 NULL,
911};
912
913static const struct attribute_group swap_attr_group = {
914 .attrs = swap_attrs,
915};
916
917static int __init swap_init_sysfs(void)
918{
919 int err;
920 struct kobject *swap_kobj;
921
922 swap_kobj = kobject_create_and_add("swap", mm_kobj);
923 if (!swap_kobj) {
924 pr_err("failed to create swap kobject\n");
925 return -ENOMEM;
926 }
927 err = sysfs_create_group(swap_kobj, &swap_attr_group);
928 if (err) {
929 pr_err("failed to register swap group\n");
930 goto delete_obj;
931 }
932 return 0;
933
934delete_obj:
935 kobject_put(swap_kobj);
936 return err;
937}
938subsys_initcall(swap_init_sysfs);
939#endif