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_page_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_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
63static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
64
65void 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
72void *get_shadow_from_swap_cache(swp_entry_t entry)
73{
74 struct address_space *address_space = swap_address_space(entry);
75 pgoff_t idx = swp_offset(entry);
76 struct page *page;
77
78 page = xa_load(&address_space->i_pages, idx);
79 if (xa_is_value(page))
80 return page;
81 return NULL;
82}
83
84/*
85 * add_to_swap_cache resembles filemap_add_folio on swapper_space,
86 * but sets SwapCache flag and private instead of mapping and index.
87 */
88int 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 = swp_offset(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);
125unlock:
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 */
141void __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 = swp_offset(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 * add_to_swap - allocate swap space for a folio
170 * @folio: folio we want to move to swap
171 *
172 * Allocate swap space for the folio and add the folio to the
173 * swap cache.
174 *
175 * Context: Caller needs to hold the folio lock.
176 * Return: Whether the folio was added to the swap cache.
177 */
178bool add_to_swap(struct folio *folio)
179{
180 swp_entry_t entry;
181 int err;
182
183 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
184 VM_BUG_ON_FOLIO(!folio_test_uptodate(folio), folio);
185
186 entry = folio_alloc_swap(folio);
187 if (!entry.val)
188 return false;
189
190 /*
191 * XArray node allocations from PF_MEMALLOC contexts could
192 * completely exhaust the page allocator. __GFP_NOMEMALLOC
193 * stops emergency reserves from being allocated.
194 *
195 * TODO: this could cause a theoretical memory reclaim
196 * deadlock in the swap out path.
197 */
198 /*
199 * Add it to the swap cache.
200 */
201 err = add_to_swap_cache(folio, entry,
202 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN, NULL);
203 if (err)
204 /*
205 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
206 * clear SWAP_HAS_CACHE flag.
207 */
208 goto fail;
209 /*
210 * Normally the folio will be dirtied in unmap because its
211 * pte should be dirty. A special case is MADV_FREE page. The
212 * page's pte could have dirty bit cleared but the folio's
213 * SwapBacked flag is still set because clearing the dirty bit
214 * and SwapBacked flag has no lock protected. For such folio,
215 * unmap will not set dirty bit for it, so folio reclaim will
216 * not write the folio out. This can cause data corruption when
217 * the folio is swapped in later. Always setting the dirty flag
218 * for the folio solves the problem.
219 */
220 folio_mark_dirty(folio);
221
222 return true;
223
224fail:
225 put_swap_folio(folio, entry);
226 return false;
227}
228
229/*
230 * This must be called only on folios that have
231 * been verified to be in the swap cache and locked.
232 * It will never put the folio into the free list,
233 * the caller has a reference on the folio.
234 */
235void delete_from_swap_cache(struct folio *folio)
236{
237 swp_entry_t entry = folio->swap;
238 struct address_space *address_space = swap_address_space(entry);
239
240 xa_lock_irq(&address_space->i_pages);
241 __delete_from_swap_cache(folio, entry, NULL);
242 xa_unlock_irq(&address_space->i_pages);
243
244 put_swap_folio(folio, entry);
245 folio_ref_sub(folio, folio_nr_pages(folio));
246}
247
248void clear_shadow_from_swap_cache(int type, unsigned long begin,
249 unsigned long end)
250{
251 unsigned long curr = begin;
252 void *old;
253
254 for (;;) {
255 swp_entry_t entry = swp_entry(type, curr);
256 struct address_space *address_space = swap_address_space(entry);
257 XA_STATE(xas, &address_space->i_pages, curr);
258
259 xas_set_update(&xas, workingset_update_node);
260
261 xa_lock_irq(&address_space->i_pages);
262 xas_for_each(&xas, old, end) {
263 if (!xa_is_value(old))
264 continue;
265 xas_store(&xas, NULL);
266 }
267 xa_unlock_irq(&address_space->i_pages);
268
269 /* search the next swapcache until we meet end */
270 curr >>= SWAP_ADDRESS_SPACE_SHIFT;
271 curr++;
272 curr <<= SWAP_ADDRESS_SPACE_SHIFT;
273 if (curr > end)
274 break;
275 }
276}
277
278/*
279 * If we are the only user, then try to free up the swap cache.
280 *
281 * Its ok to check the swapcache flag without the folio lock
282 * here because we are going to recheck again inside
283 * folio_free_swap() _with_ the lock.
284 * - Marcelo
285 */
286void free_swap_cache(struct folio *folio)
287{
288 if (folio_test_swapcache(folio) && !folio_mapped(folio) &&
289 folio_trylock(folio)) {
290 folio_free_swap(folio);
291 folio_unlock(folio);
292 }
293}
294
295/*
296 * Perform a free_page(), also freeing any swap cache associated with
297 * this page if it is the last user of the page.
298 */
299void free_page_and_swap_cache(struct page *page)
300{
301 struct folio *folio = page_folio(page);
302
303 free_swap_cache(folio);
304 if (!is_huge_zero_page(page))
305 folio_put(folio);
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 encoded_page **pages, int nr)
313{
314 struct folio_batch folios;
315 unsigned int refs[PAGEVEC_SIZE];
316
317 lru_add_drain();
318 folio_batch_init(&folios);
319 for (int i = 0; i < nr; i++) {
320 struct folio *folio = page_folio(encoded_page_ptr(pages[i]));
321
322 free_swap_cache(folio);
323 refs[folios.nr] = 1;
324 if (unlikely(encoded_page_flags(pages[i]) &
325 ENCODED_PAGE_BIT_NR_PAGES_NEXT))
326 refs[folios.nr] = encoded_nr_pages(pages[++i]);
327
328 if (folio_batch_add(&folios, folio) == 0)
329 folios_put_refs(&folios, refs);
330 }
331 if (folios.nr)
332 folios_put_refs(&folios, refs);
333}
334
335static inline bool swap_use_vma_readahead(void)
336{
337 return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
338}
339
340/*
341 * Lookup a swap entry in the swap cache. A found folio will be returned
342 * unlocked and with its refcount incremented - we rely on the kernel
343 * lock getting page table operations atomic even if we drop the folio
344 * lock before returning.
345 *
346 * Caller must lock the swap device or hold a reference to keep it valid.
347 */
348struct folio *swap_cache_get_folio(swp_entry_t entry,
349 struct vm_area_struct *vma, unsigned long addr)
350{
351 struct folio *folio;
352
353 folio = filemap_get_folio(swap_address_space(entry), swp_offset(entry));
354 if (!IS_ERR(folio)) {
355 bool vma_ra = swap_use_vma_readahead();
356 bool readahead;
357
358 /*
359 * At the moment, we don't support PG_readahead for anon THP
360 * so let's bail out rather than confusing the readahead stat.
361 */
362 if (unlikely(folio_test_large(folio)))
363 return folio;
364
365 readahead = folio_test_clear_readahead(folio);
366 if (vma && vma_ra) {
367 unsigned long ra_val;
368 int win, hits;
369
370 ra_val = GET_SWAP_RA_VAL(vma);
371 win = SWAP_RA_WIN(ra_val);
372 hits = SWAP_RA_HITS(ra_val);
373 if (readahead)
374 hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
375 atomic_long_set(&vma->swap_readahead_info,
376 SWAP_RA_VAL(addr, win, hits));
377 }
378
379 if (readahead) {
380 count_vm_event(SWAP_RA_HIT);
381 if (!vma || !vma_ra)
382 atomic_inc(&swapin_readahead_hits);
383 }
384 } else {
385 folio = NULL;
386 }
387
388 return folio;
389}
390
391/**
392 * filemap_get_incore_folio - Find and get a folio from the page or swap caches.
393 * @mapping: The address_space to search.
394 * @index: The page cache index.
395 *
396 * This differs from filemap_get_folio() in that it will also look for the
397 * folio in the swap cache.
398 *
399 * Return: The found folio or %NULL.
400 */
401struct folio *filemap_get_incore_folio(struct address_space *mapping,
402 pgoff_t index)
403{
404 swp_entry_t swp;
405 struct swap_info_struct *si;
406 struct folio *folio = filemap_get_entry(mapping, index);
407
408 if (!folio)
409 return ERR_PTR(-ENOENT);
410 if (!xa_is_value(folio))
411 return folio;
412 if (!shmem_mapping(mapping))
413 return ERR_PTR(-ENOENT);
414
415 swp = radix_to_swp_entry(folio);
416 /* There might be swapin error entries in shmem mapping. */
417 if (non_swap_entry(swp))
418 return ERR_PTR(-ENOENT);
419 /* Prevent swapoff from happening to us */
420 si = get_swap_device(swp);
421 if (!si)
422 return ERR_PTR(-ENOENT);
423 index = swp_offset(swp);
424 folio = filemap_get_folio(swap_address_space(swp), index);
425 put_swap_device(si);
426 return folio;
427}
428
429struct folio *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
430 struct mempolicy *mpol, pgoff_t ilx, bool *new_page_allocated,
431 bool skip_if_exists)
432{
433 struct swap_info_struct *si;
434 struct folio *folio;
435 void *shadow = NULL;
436
437 *new_page_allocated = false;
438 si = get_swap_device(entry);
439 if (!si)
440 return NULL;
441
442 for (;;) {
443 int err;
444 /*
445 * First check the swap cache. Since this is normally
446 * called after swap_cache_get_folio() failed, re-calling
447 * that would confuse statistics.
448 */
449 folio = filemap_get_folio(swap_address_space(entry),
450 swp_offset(entry));
451 if (!IS_ERR(folio))
452 goto got_folio;
453
454 /*
455 * Just skip read ahead for unused swap slot.
456 * During swap_off when swap_slot_cache is disabled,
457 * we have to handle the race between putting
458 * swap entry in swap cache and marking swap slot
459 * as SWAP_HAS_CACHE. That's done in later part of code or
460 * else swap_off will be aborted if we return NULL.
461 */
462 if (!swap_swapcount(si, entry) && swap_slot_cache_enabled)
463 goto fail_put_swap;
464
465 /*
466 * Get a new folio to read into from swap. Allocate it now,
467 * before marking swap_map SWAP_HAS_CACHE, when -EEXIST will
468 * cause any racers to loop around until we add it to cache.
469 */
470 folio = (struct folio *)alloc_pages_mpol(gfp_mask, 0,
471 mpol, ilx, numa_node_id());
472 if (!folio)
473 goto fail_put_swap;
474
475 /*
476 * Swap entry may have been freed since our caller observed it.
477 */
478 err = swapcache_prepare(entry);
479 if (!err)
480 break;
481
482 folio_put(folio);
483 if (err != -EEXIST)
484 goto fail_put_swap;
485
486 /*
487 * Protect against a recursive call to __read_swap_cache_async()
488 * on the same entry waiting forever here because SWAP_HAS_CACHE
489 * is set but the folio is not the swap cache yet. This can
490 * happen today if mem_cgroup_swapin_charge_folio() below
491 * triggers reclaim through zswap, which may call
492 * __read_swap_cache_async() in the writeback path.
493 */
494 if (skip_if_exists)
495 goto fail_put_swap;
496
497 /*
498 * We might race against __delete_from_swap_cache(), and
499 * stumble across a swap_map entry whose SWAP_HAS_CACHE
500 * has not yet been cleared. Or race against another
501 * __read_swap_cache_async(), which has set SWAP_HAS_CACHE
502 * in swap_map, but not yet added its folio to swap cache.
503 */
504 schedule_timeout_uninterruptible(1);
505 }
506
507 /*
508 * The swap entry is ours to swap in. Prepare the new folio.
509 */
510
511 __folio_set_locked(folio);
512 __folio_set_swapbacked(folio);
513
514 if (mem_cgroup_swapin_charge_folio(folio, NULL, gfp_mask, entry))
515 goto fail_unlock;
516
517 /* May fail (-ENOMEM) if XArray node allocation failed. */
518 if (add_to_swap_cache(folio, entry, gfp_mask & GFP_RECLAIM_MASK, &shadow))
519 goto fail_unlock;
520
521 mem_cgroup_swapin_uncharge_swap(entry);
522
523 if (shadow)
524 workingset_refault(folio, shadow);
525
526 /* Caller will initiate read into locked folio */
527 folio_add_lru(folio);
528 *new_page_allocated = true;
529got_folio:
530 put_swap_device(si);
531 return folio;
532
533fail_unlock:
534 put_swap_folio(folio, entry);
535 folio_unlock(folio);
536 folio_put(folio);
537fail_put_swap:
538 put_swap_device(si);
539 return NULL;
540}
541
542/*
543 * Locate a page of swap in physical memory, reserving swap cache space
544 * and reading the disk if it is not already cached.
545 * A failure return means that either the page allocation failed or that
546 * the swap entry is no longer in use.
547 *
548 * get/put_swap_device() aren't needed to call this function, because
549 * __read_swap_cache_async() call them and swap_read_folio() holds the
550 * swap cache folio lock.
551 */
552struct folio *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
553 struct vm_area_struct *vma, unsigned long addr,
554 struct swap_iocb **plug)
555{
556 bool page_allocated;
557 struct mempolicy *mpol;
558 pgoff_t ilx;
559 struct folio *folio;
560
561 mpol = get_vma_policy(vma, addr, 0, &ilx);
562 folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx,
563 &page_allocated, false);
564 mpol_cond_put(mpol);
565
566 if (page_allocated)
567 swap_read_folio(folio, false, plug);
568 return folio;
569}
570
571static unsigned int __swapin_nr_pages(unsigned long prev_offset,
572 unsigned long offset,
573 int hits,
574 int max_pages,
575 int prev_win)
576{
577 unsigned int pages, last_ra;
578
579 /*
580 * This heuristic has been found to work well on both sequential and
581 * random loads, swapping to hard disk or to SSD: please don't ask
582 * what the "+ 2" means, it just happens to work well, that's all.
583 */
584 pages = hits + 2;
585 if (pages == 2) {
586 /*
587 * We can have no readahead hits to judge by: but must not get
588 * stuck here forever, so check for an adjacent offset instead
589 * (and don't even bother to check whether swap type is same).
590 */
591 if (offset != prev_offset + 1 && offset != prev_offset - 1)
592 pages = 1;
593 } else {
594 unsigned int roundup = 4;
595 while (roundup < pages)
596 roundup <<= 1;
597 pages = roundup;
598 }
599
600 if (pages > max_pages)
601 pages = max_pages;
602
603 /* Don't shrink readahead too fast */
604 last_ra = prev_win / 2;
605 if (pages < last_ra)
606 pages = last_ra;
607
608 return pages;
609}
610
611static unsigned long swapin_nr_pages(unsigned long offset)
612{
613 static unsigned long prev_offset;
614 unsigned int hits, pages, max_pages;
615 static atomic_t last_readahead_pages;
616
617 max_pages = 1 << READ_ONCE(page_cluster);
618 if (max_pages <= 1)
619 return 1;
620
621 hits = atomic_xchg(&swapin_readahead_hits, 0);
622 pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits,
623 max_pages,
624 atomic_read(&last_readahead_pages));
625 if (!hits)
626 WRITE_ONCE(prev_offset, offset);
627 atomic_set(&last_readahead_pages, pages);
628
629 return pages;
630}
631
632/**
633 * swap_cluster_readahead - swap in pages in hope we need them soon
634 * @entry: swap entry of this memory
635 * @gfp_mask: memory allocation flags
636 * @mpol: NUMA memory allocation policy to be applied
637 * @ilx: NUMA interleave index, for use only when MPOL_INTERLEAVE
638 *
639 * Returns the struct folio for entry and addr, after queueing swapin.
640 *
641 * Primitive swap readahead code. We simply read an aligned block of
642 * (1 << page_cluster) entries in the swap area. This method is chosen
643 * because it doesn't cost us any seek time. We also make sure to queue
644 * the 'original' request together with the readahead ones...
645 *
646 * Note: it is intentional that the same NUMA policy and interleave index
647 * are used for every page of the readahead: neighbouring pages on swap
648 * are fairly likely to have been swapped out from the same node.
649 */
650struct folio *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
651 struct mempolicy *mpol, pgoff_t ilx)
652{
653 struct folio *folio;
654 unsigned long entry_offset = swp_offset(entry);
655 unsigned long offset = entry_offset;
656 unsigned long start_offset, end_offset;
657 unsigned long mask;
658 struct swap_info_struct *si = swp_swap_info(entry);
659 struct blk_plug plug;
660 struct swap_iocb *splug = NULL;
661 bool page_allocated;
662
663 mask = swapin_nr_pages(offset) - 1;
664 if (!mask)
665 goto skip;
666
667 /* Read a page_cluster sized and aligned cluster around offset. */
668 start_offset = offset & ~mask;
669 end_offset = offset | mask;
670 if (!start_offset) /* First page is swap header. */
671 start_offset++;
672 if (end_offset >= si->max)
673 end_offset = si->max - 1;
674
675 blk_start_plug(&plug);
676 for (offset = start_offset; offset <= end_offset ; offset++) {
677 /* Ok, do the async read-ahead now */
678 folio = __read_swap_cache_async(
679 swp_entry(swp_type(entry), offset),
680 gfp_mask, mpol, ilx, &page_allocated, false);
681 if (!folio)
682 continue;
683 if (page_allocated) {
684 swap_read_folio(folio, false, &splug);
685 if (offset != entry_offset) {
686 folio_set_readahead(folio);
687 count_vm_event(SWAP_RA);
688 }
689 }
690 folio_put(folio);
691 }
692 blk_finish_plug(&plug);
693 swap_read_unplug(splug);
694 lru_add_drain(); /* Push any new pages onto the LRU now */
695skip:
696 /* The page was likely read above, so no need for plugging here */
697 folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx,
698 &page_allocated, false);
699 if (unlikely(page_allocated)) {
700 zswap_folio_swapin(folio);
701 swap_read_folio(folio, false, NULL);
702 }
703 return folio;
704}
705
706int init_swap_address_space(unsigned int type, unsigned long nr_pages)
707{
708 struct address_space *spaces, *space;
709 unsigned int i, nr;
710
711 nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
712 spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL);
713 if (!spaces)
714 return -ENOMEM;
715 for (i = 0; i < nr; i++) {
716 space = spaces + i;
717 xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ);
718 atomic_set(&space->i_mmap_writable, 0);
719 space->a_ops = &swap_aops;
720 /* swap cache doesn't use writeback related tags */
721 mapping_set_no_writeback_tags(space);
722 }
723 nr_swapper_spaces[type] = nr;
724 swapper_spaces[type] = spaces;
725
726 return 0;
727}
728
729void exit_swap_address_space(unsigned int type)
730{
731 int i;
732 struct address_space *spaces = swapper_spaces[type];
733
734 for (i = 0; i < nr_swapper_spaces[type]; i++)
735 VM_WARN_ON_ONCE(!mapping_empty(&spaces[i]));
736 kvfree(spaces);
737 nr_swapper_spaces[type] = 0;
738 swapper_spaces[type] = NULL;
739}
740
741#define SWAP_RA_ORDER_CEILING 5
742
743struct vma_swap_readahead {
744 unsigned short win;
745 unsigned short offset;
746 unsigned short nr_pte;
747};
748
749static void swap_ra_info(struct vm_fault *vmf,
750 struct vma_swap_readahead *ra_info)
751{
752 struct vm_area_struct *vma = vmf->vma;
753 unsigned long ra_val;
754 unsigned long faddr, pfn, fpfn, lpfn, rpfn;
755 unsigned long start, end;
756 unsigned int max_win, hits, prev_win, win;
757
758 max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
759 SWAP_RA_ORDER_CEILING);
760 if (max_win == 1) {
761 ra_info->win = 1;
762 return;
763 }
764
765 faddr = vmf->address;
766 fpfn = PFN_DOWN(faddr);
767 ra_val = GET_SWAP_RA_VAL(vma);
768 pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val));
769 prev_win = SWAP_RA_WIN(ra_val);
770 hits = SWAP_RA_HITS(ra_val);
771 ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits,
772 max_win, prev_win);
773 atomic_long_set(&vma->swap_readahead_info,
774 SWAP_RA_VAL(faddr, win, 0));
775 if (win == 1)
776 return;
777
778 if (fpfn == pfn + 1) {
779 lpfn = fpfn;
780 rpfn = fpfn + win;
781 } else if (pfn == fpfn + 1) {
782 lpfn = fpfn - win + 1;
783 rpfn = fpfn + 1;
784 } else {
785 unsigned int left = (win - 1) / 2;
786
787 lpfn = fpfn - left;
788 rpfn = fpfn + win - left;
789 }
790 start = max3(lpfn, PFN_DOWN(vma->vm_start),
791 PFN_DOWN(faddr & PMD_MASK));
792 end = min3(rpfn, PFN_DOWN(vma->vm_end),
793 PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
794
795 ra_info->nr_pte = end - start;
796 ra_info->offset = fpfn - start;
797}
798
799/**
800 * swap_vma_readahead - swap in pages in hope we need them soon
801 * @targ_entry: swap entry of the targeted memory
802 * @gfp_mask: memory allocation flags
803 * @mpol: NUMA memory allocation policy to be applied
804 * @targ_ilx: NUMA interleave index, for use only when MPOL_INTERLEAVE
805 * @vmf: fault information
806 *
807 * Returns the struct folio for entry and addr, after queueing swapin.
808 *
809 * Primitive swap readahead code. We simply read in a few pages whose
810 * virtual addresses are around the fault address in the same vma.
811 *
812 * Caller must hold read mmap_lock if vmf->vma is not NULL.
813 *
814 */
815static struct folio *swap_vma_readahead(swp_entry_t targ_entry, gfp_t gfp_mask,
816 struct mempolicy *mpol, pgoff_t targ_ilx, struct vm_fault *vmf)
817{
818 struct blk_plug plug;
819 struct swap_iocb *splug = NULL;
820 struct folio *folio;
821 pte_t *pte = NULL, pentry;
822 unsigned long addr;
823 swp_entry_t entry;
824 pgoff_t ilx;
825 unsigned int i;
826 bool page_allocated;
827 struct vma_swap_readahead ra_info = {
828 .win = 1,
829 };
830
831 swap_ra_info(vmf, &ra_info);
832 if (ra_info.win == 1)
833 goto skip;
834
835 addr = vmf->address - (ra_info.offset * PAGE_SIZE);
836 ilx = targ_ilx - ra_info.offset;
837
838 blk_start_plug(&plug);
839 for (i = 0; i < ra_info.nr_pte; i++, ilx++, addr += PAGE_SIZE) {
840 if (!pte++) {
841 pte = pte_offset_map(vmf->pmd, addr);
842 if (!pte)
843 break;
844 }
845 pentry = ptep_get_lockless(pte);
846 if (!is_swap_pte(pentry))
847 continue;
848 entry = pte_to_swp_entry(pentry);
849 if (unlikely(non_swap_entry(entry)))
850 continue;
851 pte_unmap(pte);
852 pte = NULL;
853 folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx,
854 &page_allocated, false);
855 if (!folio)
856 continue;
857 if (page_allocated) {
858 swap_read_folio(folio, false, &splug);
859 if (i != ra_info.offset) {
860 folio_set_readahead(folio);
861 count_vm_event(SWAP_RA);
862 }
863 }
864 folio_put(folio);
865 }
866 if (pte)
867 pte_unmap(pte);
868 blk_finish_plug(&plug);
869 swap_read_unplug(splug);
870 lru_add_drain();
871skip:
872 /* The folio was likely read above, so no need for plugging here */
873 folio = __read_swap_cache_async(targ_entry, gfp_mask, mpol, targ_ilx,
874 &page_allocated, false);
875 if (unlikely(page_allocated)) {
876 zswap_folio_swapin(folio);
877 swap_read_folio(folio, false, NULL);
878 }
879 return folio;
880}
881
882/**
883 * swapin_readahead - swap in pages in hope we need them soon
884 * @entry: swap entry of this memory
885 * @gfp_mask: memory allocation flags
886 * @vmf: fault information
887 *
888 * Returns the struct page for entry and addr, after queueing swapin.
889 *
890 * It's a main entry function for swap readahead. By the configuration,
891 * it will read ahead blocks by cluster-based(ie, physical disk based)
892 * or vma-based(ie, virtual address based on faulty address) readahead.
893 */
894struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
895 struct vm_fault *vmf)
896{
897 struct mempolicy *mpol;
898 pgoff_t ilx;
899 struct folio *folio;
900
901 mpol = get_vma_policy(vmf->vma, vmf->address, 0, &ilx);
902 folio = swap_use_vma_readahead() ?
903 swap_vma_readahead(entry, gfp_mask, mpol, ilx, vmf) :
904 swap_cluster_readahead(entry, gfp_mask, mpol, ilx);
905 mpol_cond_put(mpol);
906
907 if (!folio)
908 return NULL;
909 return folio_file_page(folio, swp_offset(entry));
910}
911
912#ifdef CONFIG_SYSFS
913static ssize_t vma_ra_enabled_show(struct kobject *kobj,
914 struct kobj_attribute *attr, char *buf)
915{
916 return sysfs_emit(buf, "%s\n",
917 enable_vma_readahead ? "true" : "false");
918}
919static ssize_t vma_ra_enabled_store(struct kobject *kobj,
920 struct kobj_attribute *attr,
921 const char *buf, size_t count)
922{
923 ssize_t ret;
924
925 ret = kstrtobool(buf, &enable_vma_readahead);
926 if (ret)
927 return ret;
928
929 return count;
930}
931static struct kobj_attribute vma_ra_enabled_attr = __ATTR_RW(vma_ra_enabled);
932
933static struct attribute *swap_attrs[] = {
934 &vma_ra_enabled_attr.attr,
935 NULL,
936};
937
938static const struct attribute_group swap_attr_group = {
939 .attrs = swap_attrs,
940};
941
942static int __init swap_init_sysfs(void)
943{
944 int err;
945 struct kobject *swap_kobj;
946
947 swap_kobj = kobject_create_and_add("swap", mm_kobj);
948 if (!swap_kobj) {
949 pr_err("failed to create swap kobject\n");
950 return -ENOMEM;
951 }
952 err = sysfs_create_group(swap_kobj, &swap_attr_group);
953 if (err) {
954 pr_err("failed to register swap group\n");
955 goto delete_obj;
956 }
957 return 0;
958
959delete_obj:
960 kobject_put(swap_kobj);
961 return err;
962}
963subsys_initcall(swap_init_sysfs);
964#endif