1 // SPDX-License-Identifier: GPL-2.0
3 * linux/mm/swap_state.c
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie
8 * Rewritten to use page cache, (C) 1998 Stephen Tweedie
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>
25 #include <asm/pgtable.h>
28 * swapper_space is a fiction, retained to simplify the path through
29 * vmscan's shrink_page_list.
31 static 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,
39 struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly;
40 static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly;
41 static bool enable_vma_readahead __read_mostly = true;
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)
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)
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))
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)
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)
65 unsigned long add_total;
66 unsigned long del_total;
67 unsigned long find_success;
68 unsigned long find_total;
71 unsigned long total_swapcache_pages(void)
73 unsigned int i, j, nr;
74 unsigned long ret = 0;
75 struct address_space *spaces;
76 struct swap_info_struct *si;
78 for (i = 0; i < MAX_SWAPFILES; i++) {
79 swp_entry_t entry = swp_entry(i, 1);
81 /* Avoid get_swap_device() to warn for bad swap entry */
82 if (!swp_swap_info(entry))
84 /* Prevent swapoff to free swapper_spaces */
85 si = get_swap_device(entry);
88 nr = nr_swapper_spaces[i];
89 spaces = swapper_spaces[i];
90 for (j = 0; j < nr; j++)
91 ret += spaces[j].nrpages;
97 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
99 void show_swap_cache_info(void)
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));
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.
114 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp)
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 = hpage_nr_pages(page);
121 VM_BUG_ON_PAGE(!PageLocked(page), page);
122 VM_BUG_ON_PAGE(PageSwapCache(page), page);
123 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
125 page_ref_add(page, nr);
126 SetPageSwapCache(page);
130 xas_create_range(&xas);
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);
139 address_space->nrpages += nr;
140 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
141 ADD_CACHE_INFO(add_total, nr);
143 xas_unlock_irq(&xas);
144 } while (xas_nomem(&xas, gfp));
146 if (!xas_error(&xas))
149 ClearPageSwapCache(page);
150 page_ref_sub(page, nr);
151 return xas_error(&xas);
155 * This must be called only on pages that have
156 * been verified to be in the swap cache.
158 void __delete_from_swap_cache(struct page *page, swp_entry_t entry)
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);
165 VM_BUG_ON_PAGE(!PageLocked(page), page);
166 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
167 VM_BUG_ON_PAGE(PageWriteback(page), page);
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);
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);
182 * add_to_swap - allocate swap space for a page
183 * @page: page we want to move to swap
185 * Allocate swap space for the page and add the page to the
186 * swap cache. Caller needs to hold the page lock.
188 int add_to_swap(struct page *page)
193 VM_BUG_ON_PAGE(!PageLocked(page), page);
194 VM_BUG_ON_PAGE(!PageUptodate(page), page);
196 entry = get_swap_page(page);
201 * XArray node allocations from PF_MEMALLOC contexts could
202 * completely exhaust the page allocator. __GFP_NOMEMALLOC
203 * stops emergency reserves from being allocated.
205 * TODO: this could cause a theoretical memory reclaim
206 * deadlock in the swap out path.
209 * Add it to the swap cache.
211 err = add_to_swap_cache(page, entry,
212 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
215 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
216 * clear SWAP_HAS_CACHE flag.
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
229 set_page_dirty(page);
234 put_swap_page(page, entry);
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.
244 void delete_from_swap_cache(struct page *page)
246 swp_entry_t entry = { .val = page_private(page) };
247 struct address_space *address_space = swap_address_space(entry);
249 xa_lock_irq(&address_space->i_pages);
250 __delete_from_swap_cache(page, entry);
251 xa_unlock_irq(&address_space->i_pages);
253 put_swap_page(page, entry);
254 page_ref_sub(page, hpage_nr_pages(page));
258 * If we are the only user, then try to free up the swap cache.
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.
265 static inline void free_swap_cache(struct page *page)
267 if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
268 try_to_free_swap(page);
274 * Perform a free_page(), also freeing any swap cache associated with
275 * this page if it is the last user of the page.
277 void free_page_and_swap_cache(struct page *page)
279 free_swap_cache(page);
280 if (!is_huge_zero_page(page))
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.
288 void free_pages_and_swap_cache(struct page **pages, int nr)
290 struct page **pagep = pages;
294 for (i = 0; i < nr; i++)
295 free_swap_cache(pagep[i]);
296 release_pages(pagep, nr);
299 static inline bool swap_use_vma_readahead(void)
301 return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
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.
310 struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma,
314 struct swap_info_struct *si;
316 si = get_swap_device(entry);
319 page = find_get_page(swap_address_space(entry), swp_offset(entry));
322 INC_CACHE_INFO(find_total);
324 bool vma_ra = swap_use_vma_readahead();
327 INC_CACHE_INFO(find_success);
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.
332 if (unlikely(PageTransCompound(page)))
335 readahead = TestClearPageReadahead(page);
337 unsigned long ra_val;
340 ra_val = GET_SWAP_RA_VAL(vma);
341 win = SWAP_RA_WIN(ra_val);
342 hits = SWAP_RA_HITS(ra_val);
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));
350 count_vm_event(SWAP_RA_HIT);
352 atomic_inc(&swapin_readahead_hits);
359 struct 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)
363 struct page *found_page = NULL, *new_page = NULL;
364 struct swap_info_struct *si;
366 *new_page_allocated = false;
370 * First check the swap cache. Since this is normally
371 * called after lookup_swap_cache() failed, re-calling
372 * that would confuse statistics.
374 si = get_swap_device(entry);
377 found_page = find_get_page(swap_address_space(entry),
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.
391 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
395 * Get a new page to read into from swap.
398 new_page = alloc_page_vma(gfp_mask, vma, addr);
400 break; /* Out of memory */
404 * Swap entry may have been freed since our caller observed it.
406 err = swapcache_prepare(entry);
407 if (err == -EEXIST) {
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.
415 } else if (err) /* swp entry is obsolete ? */
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);
423 /* Initiate read into locked page */
424 SetPageWorkingset(new_page);
425 lru_cache_add_anon(new_page);
426 *new_page_allocated = true;
429 __ClearPageLocked(new_page);
431 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
432 * clear SWAP_HAS_CACHE flag.
434 put_swap_page(new_page, entry);
435 } while (err != -ENOMEM);
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.
448 struct 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)
451 bool page_was_allocated;
452 struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
453 vma, addr, &page_was_allocated);
455 if (page_was_allocated)
456 swap_readpage(retpage, do_poll);
461 static unsigned int __swapin_nr_pages(unsigned long prev_offset,
462 unsigned long offset,
467 unsigned int pages, last_ra;
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.
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).
481 if (offset != prev_offset + 1 && offset != prev_offset - 1)
484 unsigned int roundup = 4;
485 while (roundup < pages)
490 if (pages > max_pages)
493 /* Don't shrink readahead too fast */
494 last_ra = prev_win / 2;
501 static unsigned long swapin_nr_pages(unsigned long offset)
503 static unsigned long prev_offset;
504 unsigned int hits, pages, max_pages;
505 static atomic_t last_readahead_pages;
507 max_pages = 1 << READ_ONCE(page_cluster);
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));
515 prev_offset = offset;
516 atomic_set(&last_readahead_pages, pages);
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
527 * Returns the struct page for entry and addr, after queueing swapin.
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...
534 * This has been extended to use the NUMA policies from the mm triggering
537 * Caller must hold read mmap_sem if vmf->vma is not NULL.
539 struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
540 struct vm_fault *vmf)
543 unsigned long entry_offset = swp_offset(entry);
544 unsigned long offset = entry_offset;
545 unsigned long start_offset, end_offset;
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;
553 mask = swapin_nr_pages(offset) - 1;
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))
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. */
570 if (end_offset >= si->max)
571 end_offset = si->max - 1;
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);
581 if (page_allocated) {
582 swap_readpage(page, false);
583 if (offset != entry_offset) {
584 SetPageReadahead(page);
585 count_vm_event(SWAP_RA);
590 blk_finish_plug(&plug);
592 lru_add_drain(); /* Push any new pages onto the LRU now */
594 return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll);
597 int init_swap_address_space(unsigned int type, unsigned long nr_pages)
599 struct address_space *spaces, *space;
602 nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
603 spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL);
606 for (i = 0; i < nr; 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);
614 nr_swapper_spaces[type] = nr;
615 swapper_spaces[type] = spaces;
620 void exit_swap_address_space(unsigned int type)
622 kvfree(swapper_spaces[type]);
623 nr_swapper_spaces[type] = 0;
624 swapper_spaces[type] = NULL;
627 static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
631 unsigned long *start,
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));
640 static void swap_ra_info(struct vm_fault *vmf,
641 struct vma_swap_readahead *ra_info)
643 struct vm_area_struct *vma = vmf->vma;
644 unsigned long ra_val;
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;
654 max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
655 SWAP_RA_ORDER_CEILING);
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)))) {
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,
676 atomic_long_set(&vma->swap_readahead_info,
677 SWAP_RA_VAL(faddr, win, 0));
684 /* Copy the PTEs because the page table may be unmapped */
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,
691 left = (win - 1) / 2;
692 swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
695 ra_info->nr_pte = end - start;
696 ra_info->offset = fpfn - start;
697 pte -= ra_info->offset;
701 tpte = ra_info->ptes;
702 for (pfn = start; pfn != end; pfn++)
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
714 * Returns the struct page for entry and addr, after queueing swapin.
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.
719 * Caller must hold read mmap_sem if vmf->vma is not NULL.
722 static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
723 struct vm_fault *vmf)
725 struct blk_plug plug;
726 struct vm_area_struct *vma = vmf->vma;
732 struct vma_swap_readahead ra_info = {0,};
734 swap_ra_info(vmf, &ra_info);
735 if (ra_info.win == 1)
738 blk_start_plug(&plug);
739 for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte;
742 if (pte_none(pentry))
744 if (pte_present(pentry))
746 entry = pte_to_swp_entry(pentry);
747 if (unlikely(non_swap_entry(entry)))
749 page = __read_swap_cache_async(entry, gfp_mask, vma,
750 vmf->address, &page_allocated);
753 if (page_allocated) {
754 swap_readpage(page, false);
755 if (i != ra_info.offset) {
756 SetPageReadahead(page);
757 count_vm_event(SWAP_RA);
762 blk_finish_plug(&plug);
765 return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
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
775 * Returns the struct page for entry and addr, after queueing swapin.
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.
781 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
782 struct vm_fault *vmf)
784 return swap_use_vma_readahead() ?
785 swap_vma_readahead(entry, gfp_mask, vmf) :
786 swap_cluster_readahead(entry, gfp_mask, vmf);
790 static ssize_t vma_ra_enabled_show(struct kobject *kobj,
791 struct kobj_attribute *attr, char *buf)
793 return sprintf(buf, "%s\n", enable_vma_readahead ? "true" : "false");
795 static ssize_t vma_ra_enabled_store(struct kobject *kobj,
796 struct kobj_attribute *attr,
797 const char *buf, size_t count)
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;
808 static struct kobj_attribute vma_ra_enabled_attr =
809 __ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show,
810 vma_ra_enabled_store);
812 static struct attribute *swap_attrs[] = {
813 &vma_ra_enabled_attr.attr,
817 static struct attribute_group swap_attr_group = {
821 static int __init swap_init_sysfs(void)
824 struct kobject *swap_kobj;
826 swap_kobj = kobject_create_and_add("swap", mm_kobj);
828 pr_err("failed to create swap kobject\n");
831 err = sysfs_create_group(swap_kobj, &swap_attr_group);
833 pr_err("failed to register swap group\n");
839 kobject_put(swap_kobj);
842 subsys_initcall(swap_init_sysfs);