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