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
25#include <asm/pgtable.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) do { swap_cache_info.x++; } while (0)
62#define ADD_CACHE_INFO(x, nr) do { swap_cache_info.x += (nr); } while (0)
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
71unsigned long total_swapcache_pages(void)
72{
73 unsigned int i, j, nr;
74 unsigned long ret = 0;
75 struct address_space *spaces;
76
77 rcu_read_lock();
78 for (i = 0; i < MAX_SWAPFILES; i++) {
79 /*
80 * The corresponding entries in nr_swapper_spaces and
81 * swapper_spaces will be reused only after at least
82 * one grace period. So it is impossible for them
83 * belongs to different usage.
84 */
85 nr = nr_swapper_spaces[i];
86 spaces = rcu_dereference(swapper_spaces[i]);
87 if (!nr || !spaces)
88 continue;
89 for (j = 0; j < nr; j++)
90 ret += spaces[j].nrpages;
91 }
92 rcu_read_unlock();
93 return ret;
94}
95
96static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
97
98void show_swap_cache_info(void)
99{
100 printk("%lu pages in swap cache\n", total_swapcache_pages());
101 printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
102 swap_cache_info.add_total, swap_cache_info.del_total,
103 swap_cache_info.find_success, swap_cache_info.find_total);
104 printk("Free swap = %ldkB\n",
105 get_nr_swap_pages() << (PAGE_SHIFT - 10));
106 printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
107}
108
109/*
110 * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
111 * but sets SwapCache flag and private instead of mapping and index.
112 */
113int __add_to_swap_cache(struct page *page, swp_entry_t entry)
114{
115 int error, i, nr = hpage_nr_pages(page);
116 struct address_space *address_space;
117 pgoff_t idx = swp_offset(entry);
118
119 VM_BUG_ON_PAGE(!PageLocked(page), page);
120 VM_BUG_ON_PAGE(PageSwapCache(page), page);
121 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
122
123 page_ref_add(page, nr);
124 SetPageSwapCache(page);
125
126 address_space = swap_address_space(entry);
127 xa_lock_irq(&address_space->i_pages);
128 for (i = 0; i < nr; i++) {
129 set_page_private(page + i, entry.val + i);
130 error = radix_tree_insert(&address_space->i_pages,
131 idx + i, page + i);
132 if (unlikely(error))
133 break;
134 }
135 if (likely(!error)) {
136 address_space->nrpages += nr;
137 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
138 ADD_CACHE_INFO(add_total, nr);
139 } else {
140 /*
141 * Only the context which have set SWAP_HAS_CACHE flag
142 * would call add_to_swap_cache().
143 * So add_to_swap_cache() doesn't returns -EEXIST.
144 */
145 VM_BUG_ON(error == -EEXIST);
146 set_page_private(page + i, 0UL);
147 while (i--) {
148 radix_tree_delete(&address_space->i_pages, idx + i);
149 set_page_private(page + i, 0UL);
150 }
151 ClearPageSwapCache(page);
152 page_ref_sub(page, nr);
153 }
154 xa_unlock_irq(&address_space->i_pages);
155
156 return error;
157}
158
159
160int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
161{
162 int error;
163
164 error = radix_tree_maybe_preload_order(gfp_mask, compound_order(page));
165 if (!error) {
166 error = __add_to_swap_cache(page, entry);
167 radix_tree_preload_end();
168 }
169 return error;
170}
171
172/*
173 * This must be called only on pages that have
174 * been verified to be in the swap cache.
175 */
176void __delete_from_swap_cache(struct page *page)
177{
178 struct address_space *address_space;
179 int i, nr = hpage_nr_pages(page);
180 swp_entry_t entry;
181 pgoff_t idx;
182
183 VM_BUG_ON_PAGE(!PageLocked(page), page);
184 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
185 VM_BUG_ON_PAGE(PageWriteback(page), page);
186
187 entry.val = page_private(page);
188 address_space = swap_address_space(entry);
189 idx = swp_offset(entry);
190 for (i = 0; i < nr; i++) {
191 radix_tree_delete(&address_space->i_pages, idx + i);
192 set_page_private(page + i, 0);
193 }
194 ClearPageSwapCache(page);
195 address_space->nrpages -= nr;
196 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
197 ADD_CACHE_INFO(del_total, nr);
198}
199
200/**
201 * add_to_swap - allocate swap space for a page
202 * @page: page we want to move to swap
203 *
204 * Allocate swap space for the page and add the page to the
205 * swap cache. Caller needs to hold the page lock.
206 */
207int add_to_swap(struct page *page)
208{
209 swp_entry_t entry;
210 int err;
211
212 VM_BUG_ON_PAGE(!PageLocked(page), page);
213 VM_BUG_ON_PAGE(!PageUptodate(page), page);
214
215 entry = get_swap_page(page);
216 if (!entry.val)
217 return 0;
218
219 if (mem_cgroup_try_charge_swap(page, entry))
220 goto fail;
221
222 /*
223 * Radix-tree node allocations from PF_MEMALLOC contexts could
224 * completely exhaust the page allocator. __GFP_NOMEMALLOC
225 * stops emergency reserves from being allocated.
226 *
227 * TODO: this could cause a theoretical memory reclaim
228 * deadlock in the swap out path.
229 */
230 /*
231 * Add it to the swap cache.
232 */
233 err = add_to_swap_cache(page, entry,
234 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
235 /* -ENOMEM radix-tree allocation failure */
236 if (err)
237 /*
238 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
239 * clear SWAP_HAS_CACHE flag.
240 */
241 goto fail;
242 /*
243 * Normally the page will be dirtied in unmap because its pte should be
244 * dirty. A special case is MADV_FREE page. The page'e pte could have
245 * dirty bit cleared but the page's SwapBacked bit is still set because
246 * clearing the dirty bit and SwapBacked bit has no lock protected. For
247 * such page, unmap will not set dirty bit for it, so page reclaim will
248 * not write the page out. This can cause data corruption when the page
249 * is swap in later. Always setting the dirty bit for the page solves
250 * the problem.
251 */
252 set_page_dirty(page);
253
254 return 1;
255
256fail:
257 put_swap_page(page, entry);
258 return 0;
259}
260
261/*
262 * This must be called only on pages that have
263 * been verified to be in the swap cache and locked.
264 * It will never put the page into the free list,
265 * the caller has a reference on the page.
266 */
267void delete_from_swap_cache(struct page *page)
268{
269 swp_entry_t entry;
270 struct address_space *address_space;
271
272 entry.val = page_private(page);
273
274 address_space = swap_address_space(entry);
275 xa_lock_irq(&address_space->i_pages);
276 __delete_from_swap_cache(page);
277 xa_unlock_irq(&address_space->i_pages);
278
279 put_swap_page(page, entry);
280 page_ref_sub(page, hpage_nr_pages(page));
281}
282
283/*
284 * If we are the only user, then try to free up the swap cache.
285 *
286 * Its ok to check for PageSwapCache without the page lock
287 * here because we are going to recheck again inside
288 * try_to_free_swap() _with_ the lock.
289 * - Marcelo
290 */
291static inline void free_swap_cache(struct page *page)
292{
293 if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
294 try_to_free_swap(page);
295 unlock_page(page);
296 }
297}
298
299/*
300 * Perform a free_page(), also freeing any swap cache associated with
301 * this page if it is the last user of the page.
302 */
303void free_page_and_swap_cache(struct page *page)
304{
305 free_swap_cache(page);
306 if (!is_huge_zero_page(page))
307 put_page(page);
308}
309
310/*
311 * Passed an array of pages, drop them all from swapcache and then release
312 * them. They are removed from the LRU and freed if this is their last use.
313 */
314void free_pages_and_swap_cache(struct page **pages, int nr)
315{
316 struct page **pagep = pages;
317 int i;
318
319 lru_add_drain();
320 for (i = 0; i < nr; i++)
321 free_swap_cache(pagep[i]);
322 release_pages(pagep, nr);
323}
324
325static inline bool swap_use_vma_readahead(void)
326{
327 return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
328}
329
330/*
331 * Lookup a swap entry in the swap cache. A found page will be returned
332 * unlocked and with its refcount incremented - we rely on the kernel
333 * lock getting page table operations atomic even if we drop the page
334 * lock before returning.
335 */
336struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma,
337 unsigned long addr)
338{
339 struct page *page;
340
341 page = find_get_page(swap_address_space(entry), swp_offset(entry));
342
343 INC_CACHE_INFO(find_total);
344 if (page) {
345 bool vma_ra = swap_use_vma_readahead();
346 bool readahead;
347
348 INC_CACHE_INFO(find_success);
349 /*
350 * At the moment, we don't support PG_readahead for anon THP
351 * so let's bail out rather than confusing the readahead stat.
352 */
353 if (unlikely(PageTransCompound(page)))
354 return page;
355
356 readahead = TestClearPageReadahead(page);
357 if (vma && vma_ra) {
358 unsigned long ra_val;
359 int win, hits;
360
361 ra_val = GET_SWAP_RA_VAL(vma);
362 win = SWAP_RA_WIN(ra_val);
363 hits = SWAP_RA_HITS(ra_val);
364 if (readahead)
365 hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
366 atomic_long_set(&vma->swap_readahead_info,
367 SWAP_RA_VAL(addr, win, hits));
368 }
369
370 if (readahead) {
371 count_vm_event(SWAP_RA_HIT);
372 if (!vma || !vma_ra)
373 atomic_inc(&swapin_readahead_hits);
374 }
375 }
376
377 return page;
378}
379
380struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
381 struct vm_area_struct *vma, unsigned long addr,
382 bool *new_page_allocated)
383{
384 struct page *found_page, *new_page = NULL;
385 struct address_space *swapper_space = swap_address_space(entry);
386 int err;
387 *new_page_allocated = false;
388
389 do {
390 /*
391 * First check the swap cache. Since this is normally
392 * called after lookup_swap_cache() failed, re-calling
393 * that would confuse statistics.
394 */
395 found_page = find_get_page(swapper_space, swp_offset(entry));
396 if (found_page)
397 break;
398
399 /*
400 * Just skip read ahead for unused swap slot.
401 * During swap_off when swap_slot_cache is disabled,
402 * we have to handle the race between putting
403 * swap entry in swap cache and marking swap slot
404 * as SWAP_HAS_CACHE. That's done in later part of code or
405 * else swap_off will be aborted if we return NULL.
406 */
407 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
408 break;
409
410 /*
411 * Get a new page to read into from swap.
412 */
413 if (!new_page) {
414 new_page = alloc_page_vma(gfp_mask, vma, addr);
415 if (!new_page)
416 break; /* Out of memory */
417 }
418
419 /*
420 * call radix_tree_preload() while we can wait.
421 */
422 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
423 if (err)
424 break;
425
426 /*
427 * Swap entry may have been freed since our caller observed it.
428 */
429 err = swapcache_prepare(entry);
430 if (err == -EEXIST) {
431 radix_tree_preload_end();
432 /*
433 * We might race against get_swap_page() and stumble
434 * across a SWAP_HAS_CACHE swap_map entry whose page
435 * has not been brought into the swapcache yet.
436 */
437 cond_resched();
438 continue;
439 }
440 if (err) { /* swp entry is obsolete ? */
441 radix_tree_preload_end();
442 break;
443 }
444
445 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
446 __SetPageLocked(new_page);
447 __SetPageSwapBacked(new_page);
448 err = __add_to_swap_cache(new_page, entry);
449 if (likely(!err)) {
450 radix_tree_preload_end();
451 /*
452 * Initiate read into locked page and return.
453 */
454 lru_cache_add_anon(new_page);
455 *new_page_allocated = true;
456 return new_page;
457 }
458 radix_tree_preload_end();
459 __ClearPageLocked(new_page);
460 /*
461 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
462 * clear SWAP_HAS_CACHE flag.
463 */
464 put_swap_page(new_page, entry);
465 } while (err != -ENOMEM);
466
467 if (new_page)
468 put_page(new_page);
469 return found_page;
470}
471
472/*
473 * Locate a page of swap in physical memory, reserving swap cache space
474 * and reading the disk if it is not already cached.
475 * A failure return means that either the page allocation failed or that
476 * the swap entry is no longer in use.
477 */
478struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
479 struct vm_area_struct *vma, unsigned long addr, bool do_poll)
480{
481 bool page_was_allocated;
482 struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
483 vma, addr, &page_was_allocated);
484
485 if (page_was_allocated)
486 swap_readpage(retpage, do_poll);
487
488 return retpage;
489}
490
491static unsigned int __swapin_nr_pages(unsigned long prev_offset,
492 unsigned long offset,
493 int hits,
494 int max_pages,
495 int prev_win)
496{
497 unsigned int pages, last_ra;
498
499 /*
500 * This heuristic has been found to work well on both sequential and
501 * random loads, swapping to hard disk or to SSD: please don't ask
502 * what the "+ 2" means, it just happens to work well, that's all.
503 */
504 pages = hits + 2;
505 if (pages == 2) {
506 /*
507 * We can have no readahead hits to judge by: but must not get
508 * stuck here forever, so check for an adjacent offset instead
509 * (and don't even bother to check whether swap type is same).
510 */
511 if (offset != prev_offset + 1 && offset != prev_offset - 1)
512 pages = 1;
513 } else {
514 unsigned int roundup = 4;
515 while (roundup < pages)
516 roundup <<= 1;
517 pages = roundup;
518 }
519
520 if (pages > max_pages)
521 pages = max_pages;
522
523 /* Don't shrink readahead too fast */
524 last_ra = prev_win / 2;
525 if (pages < last_ra)
526 pages = last_ra;
527
528 return pages;
529}
530
531static unsigned long swapin_nr_pages(unsigned long offset)
532{
533 static unsigned long prev_offset;
534 unsigned int hits, pages, max_pages;
535 static atomic_t last_readahead_pages;
536
537 max_pages = 1 << READ_ONCE(page_cluster);
538 if (max_pages <= 1)
539 return 1;
540
541 hits = atomic_xchg(&swapin_readahead_hits, 0);
542 pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,
543 atomic_read(&last_readahead_pages));
544 if (!hits)
545 prev_offset = offset;
546 atomic_set(&last_readahead_pages, pages);
547
548 return pages;
549}
550
551/**
552 * swap_cluster_readahead - swap in pages in hope we need them soon
553 * @entry: swap entry of this memory
554 * @gfp_mask: memory allocation flags
555 * @vmf: fault information
556 *
557 * Returns the struct page for entry and addr, after queueing swapin.
558 *
559 * Primitive swap readahead code. We simply read an aligned block of
560 * (1 << page_cluster) entries in the swap area. This method is chosen
561 * because it doesn't cost us any seek time. We also make sure to queue
562 * the 'original' request together with the readahead ones...
563 *
564 * This has been extended to use the NUMA policies from the mm triggering
565 * the readahead.
566 *
567 * Caller must hold down_read on the vma->vm_mm if vmf->vma is not NULL.
568 */
569struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
570 struct vm_fault *vmf)
571{
572 struct page *page;
573 unsigned long entry_offset = swp_offset(entry);
574 unsigned long offset = entry_offset;
575 unsigned long start_offset, end_offset;
576 unsigned long mask;
577 struct swap_info_struct *si = swp_swap_info(entry);
578 struct blk_plug plug;
579 bool do_poll = true, page_allocated;
580 struct vm_area_struct *vma = vmf->vma;
581 unsigned long addr = vmf->address;
582
583 mask = swapin_nr_pages(offset) - 1;
584 if (!mask)
585 goto skip;
586
587 do_poll = false;
588 /* Read a page_cluster sized and aligned cluster around offset. */
589 start_offset = offset & ~mask;
590 end_offset = offset | mask;
591 if (!start_offset) /* First page is swap header. */
592 start_offset++;
593 if (end_offset >= si->max)
594 end_offset = si->max - 1;
595
596 blk_start_plug(&plug);
597 for (offset = start_offset; offset <= end_offset ; offset++) {
598 /* Ok, do the async read-ahead now */
599 page = __read_swap_cache_async(
600 swp_entry(swp_type(entry), offset),
601 gfp_mask, vma, addr, &page_allocated);
602 if (!page)
603 continue;
604 if (page_allocated) {
605 swap_readpage(page, false);
606 if (offset != entry_offset) {
607 SetPageReadahead(page);
608 count_vm_event(SWAP_RA);
609 }
610 }
611 put_page(page);
612 }
613 blk_finish_plug(&plug);
614
615 lru_add_drain(); /* Push any new pages onto the LRU now */
616skip:
617 return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll);
618}
619
620int init_swap_address_space(unsigned int type, unsigned long nr_pages)
621{
622 struct address_space *spaces, *space;
623 unsigned int i, nr;
624
625 nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
626 spaces = kvzalloc(sizeof(struct address_space) * nr, GFP_KERNEL);
627 if (!spaces)
628 return -ENOMEM;
629 for (i = 0; i < nr; i++) {
630 space = spaces + i;
631 INIT_RADIX_TREE(&space->i_pages, GFP_ATOMIC|__GFP_NOWARN);
632 atomic_set(&space->i_mmap_writable, 0);
633 space->a_ops = &swap_aops;
634 /* swap cache doesn't use writeback related tags */
635 mapping_set_no_writeback_tags(space);
636 }
637 nr_swapper_spaces[type] = nr;
638 rcu_assign_pointer(swapper_spaces[type], spaces);
639
640 return 0;
641}
642
643void exit_swap_address_space(unsigned int type)
644{
645 struct address_space *spaces;
646
647 spaces = swapper_spaces[type];
648 nr_swapper_spaces[type] = 0;
649 rcu_assign_pointer(swapper_spaces[type], NULL);
650 synchronize_rcu();
651 kvfree(spaces);
652}
653
654static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
655 unsigned long faddr,
656 unsigned long lpfn,
657 unsigned long rpfn,
658 unsigned long *start,
659 unsigned long *end)
660{
661 *start = max3(lpfn, PFN_DOWN(vma->vm_start),
662 PFN_DOWN(faddr & PMD_MASK));
663 *end = min3(rpfn, PFN_DOWN(vma->vm_end),
664 PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
665}
666
667static void swap_ra_info(struct vm_fault *vmf,
668 struct vma_swap_readahead *ra_info)
669{
670 struct vm_area_struct *vma = vmf->vma;
671 unsigned long ra_val;
672 swp_entry_t entry;
673 unsigned long faddr, pfn, fpfn;
674 unsigned long start, end;
675 pte_t *pte, *orig_pte;
676 unsigned int max_win, hits, prev_win, win, left;
677#ifndef CONFIG_64BIT
678 pte_t *tpte;
679#endif
680
681 max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
682 SWAP_RA_ORDER_CEILING);
683 if (max_win == 1) {
684 ra_info->win = 1;
685 return;
686 }
687
688 faddr = vmf->address;
689 orig_pte = pte = pte_offset_map(vmf->pmd, faddr);
690 entry = pte_to_swp_entry(*pte);
691 if ((unlikely(non_swap_entry(entry)))) {
692 pte_unmap(orig_pte);
693 return;
694 }
695
696 fpfn = PFN_DOWN(faddr);
697 ra_val = GET_SWAP_RA_VAL(vma);
698 pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val));
699 prev_win = SWAP_RA_WIN(ra_val);
700 hits = SWAP_RA_HITS(ra_val);
701 ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits,
702 max_win, prev_win);
703 atomic_long_set(&vma->swap_readahead_info,
704 SWAP_RA_VAL(faddr, win, 0));
705
706 if (win == 1) {
707 pte_unmap(orig_pte);
708 return;
709 }
710
711 /* Copy the PTEs because the page table may be unmapped */
712 if (fpfn == pfn + 1)
713 swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end);
714 else if (pfn == fpfn + 1)
715 swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1,
716 &start, &end);
717 else {
718 left = (win - 1) / 2;
719 swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
720 &start, &end);
721 }
722 ra_info->nr_pte = end - start;
723 ra_info->offset = fpfn - start;
724 pte -= ra_info->offset;
725#ifdef CONFIG_64BIT
726 ra_info->ptes = pte;
727#else
728 tpte = ra_info->ptes;
729 for (pfn = start; pfn != end; pfn++)
730 *tpte++ = *pte++;
731#endif
732 pte_unmap(orig_pte);
733}
734
735static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
736 struct vm_fault *vmf)
737{
738 struct blk_plug plug;
739 struct vm_area_struct *vma = vmf->vma;
740 struct page *page;
741 pte_t *pte, pentry;
742 swp_entry_t entry;
743 unsigned int i;
744 bool page_allocated;
745 struct vma_swap_readahead ra_info = {0,};
746
747 swap_ra_info(vmf, &ra_info);
748 if (ra_info.win == 1)
749 goto skip;
750
751 blk_start_plug(&plug);
752 for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte;
753 i++, pte++) {
754 pentry = *pte;
755 if (pte_none(pentry))
756 continue;
757 if (pte_present(pentry))
758 continue;
759 entry = pte_to_swp_entry(pentry);
760 if (unlikely(non_swap_entry(entry)))
761 continue;
762 page = __read_swap_cache_async(entry, gfp_mask, vma,
763 vmf->address, &page_allocated);
764 if (!page)
765 continue;
766 if (page_allocated) {
767 swap_readpage(page, false);
768 if (i != ra_info.offset) {
769 SetPageReadahead(page);
770 count_vm_event(SWAP_RA);
771 }
772 }
773 put_page(page);
774 }
775 blk_finish_plug(&plug);
776 lru_add_drain();
777skip:
778 return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
779 ra_info.win == 1);
780}
781
782/**
783 * swapin_readahead - swap in pages in hope we need them soon
784 * @entry: swap entry of this memory
785 * @gfp_mask: memory allocation flags
786 * @vmf: fault information
787 *
788 * Returns the struct page for entry and addr, after queueing swapin.
789 *
790 * It's a main entry function for swap readahead. By the configuration,
791 * it will read ahead blocks by cluster-based(ie, physical disk based)
792 * or vma-based(ie, virtual address based on faulty address) readahead.
793 */
794struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
795 struct vm_fault *vmf)
796{
797 return swap_use_vma_readahead() ?
798 swap_vma_readahead(entry, gfp_mask, vmf) :
799 swap_cluster_readahead(entry, gfp_mask, vmf);
800}
801
802#ifdef CONFIG_SYSFS
803static ssize_t vma_ra_enabled_show(struct kobject *kobj,
804 struct kobj_attribute *attr, char *buf)
805{
806 return sprintf(buf, "%s\n", enable_vma_readahead ? "true" : "false");
807}
808static ssize_t vma_ra_enabled_store(struct kobject *kobj,
809 struct kobj_attribute *attr,
810 const char *buf, size_t count)
811{
812 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
813 enable_vma_readahead = true;
814 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
815 enable_vma_readahead = false;
816 else
817 return -EINVAL;
818
819 return count;
820}
821static struct kobj_attribute vma_ra_enabled_attr =
822 __ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show,
823 vma_ra_enabled_store);
824
825static struct attribute *swap_attrs[] = {
826 &vma_ra_enabled_attr.attr,
827 NULL,
828};
829
830static struct attribute_group swap_attr_group = {
831 .attrs = swap_attrs,
832};
833
834static int __init swap_init_sysfs(void)
835{
836 int err;
837 struct kobject *swap_kobj;
838
839 swap_kobj = kobject_create_and_add("swap", mm_kobj);
840 if (!swap_kobj) {
841 pr_err("failed to create swap kobject\n");
842 return -ENOMEM;
843 }
844 err = sysfs_create_group(swap_kobj, &swap_attr_group);
845 if (err) {
846 pr_err("failed to register swap group\n");
847 goto delete_obj;
848 }
849 return 0;
850
851delete_obj:
852 kobject_put(swap_kobj);
853 return err;
854}
855subsys_initcall(swap_init_sysfs);
856#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