2 * linux/mm/swap_state.c
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie
7 * Rewritten to use page cache, (C) 1998 Stephen Tweedie
10 #include <linux/gfp.h>
11 #include <linux/kernel_stat.h>
12 #include <linux/swap.h>
13 #include <linux/swapops.h>
14 #include <linux/init.h>
15 #include <linux/pagemap.h>
16 #include <linux/backing-dev.h>
17 #include <linux/blkdev.h>
18 #include <linux/pagevec.h>
19 #include <linux/migrate.h>
20 #include <linux/vmalloc.h>
21 #include <linux/swap_slots.h>
22 #include <linux/huge_mm.h>
24 #include <asm/pgtable.h>
27 * swapper_space is a fiction, retained to simplify the path through
28 * vmscan's shrink_page_list.
30 static const struct address_space_operations swap_aops
= {
31 .writepage
= swap_writepage
,
32 .set_page_dirty
= swap_set_page_dirty
,
33 #ifdef CONFIG_MIGRATION
34 .migratepage
= migrate_page
,
38 struct address_space
*swapper_spaces
[MAX_SWAPFILES
];
39 static unsigned int nr_swapper_spaces
[MAX_SWAPFILES
];
40 bool swap_vma_readahead
= true;
42 #define SWAP_RA_WIN_SHIFT (PAGE_SHIFT / 2)
43 #define SWAP_RA_HITS_MASK ((1UL << SWAP_RA_WIN_SHIFT) - 1)
44 #define SWAP_RA_HITS_MAX SWAP_RA_HITS_MASK
45 #define SWAP_RA_WIN_MASK (~PAGE_MASK & ~SWAP_RA_HITS_MASK)
47 #define SWAP_RA_HITS(v) ((v) & SWAP_RA_HITS_MASK)
48 #define SWAP_RA_WIN(v) (((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
49 #define SWAP_RA_ADDR(v) ((v) & PAGE_MASK)
51 #define SWAP_RA_VAL(addr, win, hits) \
52 (((addr) & PAGE_MASK) | \
53 (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) | \
54 ((hits) & SWAP_RA_HITS_MASK))
56 /* Initial readahead hits is 4 to start up with a small window */
57 #define GET_SWAP_RA_VAL(vma) \
58 (atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
60 #define INC_CACHE_INFO(x) do { swap_cache_info.x++; } while (0)
61 #define ADD_CACHE_INFO(x, nr) do { swap_cache_info.x += (nr); } while (0)
64 unsigned long add_total
;
65 unsigned long del_total
;
66 unsigned long find_success
;
67 unsigned long find_total
;
70 unsigned long total_swapcache_pages(void)
72 unsigned int i
, j
, nr
;
73 unsigned long ret
= 0;
74 struct address_space
*spaces
;
77 for (i
= 0; i
< MAX_SWAPFILES
; i
++) {
79 * The corresponding entries in nr_swapper_spaces and
80 * swapper_spaces will be reused only after at least
81 * one grace period. So it is impossible for them
82 * belongs to different usage.
84 nr
= nr_swapper_spaces
[i
];
85 spaces
= rcu_dereference(swapper_spaces
[i
]);
88 for (j
= 0; j
< nr
; j
++)
89 ret
+= spaces
[j
].nrpages
;
95 static atomic_t swapin_readahead_hits
= ATOMIC_INIT(4);
97 void show_swap_cache_info(void)
99 printk("%lu pages in swap cache\n", total_swapcache_pages());
100 printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
101 swap_cache_info
.add_total
, swap_cache_info
.del_total
,
102 swap_cache_info
.find_success
, swap_cache_info
.find_total
);
103 printk("Free swap = %ldkB\n",
104 get_nr_swap_pages() << (PAGE_SHIFT
- 10));
105 printk("Total swap = %lukB\n", total_swap_pages
<< (PAGE_SHIFT
- 10));
109 * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
110 * but sets SwapCache flag and private instead of mapping and index.
112 int __add_to_swap_cache(struct page
*page
, swp_entry_t entry
)
114 int error
, i
, nr
= hpage_nr_pages(page
);
115 struct address_space
*address_space
;
116 pgoff_t idx
= swp_offset(entry
);
118 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
119 VM_BUG_ON_PAGE(PageSwapCache(page
), page
);
120 VM_BUG_ON_PAGE(!PageSwapBacked(page
), page
);
122 page_ref_add(page
, nr
);
123 SetPageSwapCache(page
);
125 address_space
= swap_address_space(entry
);
126 spin_lock_irq(&address_space
->tree_lock
);
127 for (i
= 0; i
< nr
; i
++) {
128 set_page_private(page
+ i
, entry
.val
+ i
);
129 error
= radix_tree_insert(&address_space
->page_tree
,
134 if (likely(!error
)) {
135 address_space
->nrpages
+= nr
;
136 __mod_node_page_state(page_pgdat(page
), NR_FILE_PAGES
, nr
);
137 ADD_CACHE_INFO(add_total
, nr
);
140 * Only the context which have set SWAP_HAS_CACHE flag
141 * would call add_to_swap_cache().
142 * So add_to_swap_cache() doesn't returns -EEXIST.
144 VM_BUG_ON(error
== -EEXIST
);
145 set_page_private(page
+ i
, 0UL);
147 radix_tree_delete(&address_space
->page_tree
, idx
+ i
);
148 set_page_private(page
+ i
, 0UL);
150 ClearPageSwapCache(page
);
151 page_ref_sub(page
, nr
);
153 spin_unlock_irq(&address_space
->tree_lock
);
159 int add_to_swap_cache(struct page
*page
, swp_entry_t entry
, gfp_t gfp_mask
)
163 error
= radix_tree_maybe_preload_order(gfp_mask
, compound_order(page
));
165 error
= __add_to_swap_cache(page
, entry
);
166 radix_tree_preload_end();
172 * This must be called only on pages that have
173 * been verified to be in the swap cache.
175 void __delete_from_swap_cache(struct page
*page
)
177 struct address_space
*address_space
;
178 int i
, nr
= hpage_nr_pages(page
);
182 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
183 VM_BUG_ON_PAGE(!PageSwapCache(page
), page
);
184 VM_BUG_ON_PAGE(PageWriteback(page
), page
);
186 entry
.val
= page_private(page
);
187 address_space
= swap_address_space(entry
);
188 idx
= swp_offset(entry
);
189 for (i
= 0; i
< nr
; i
++) {
190 radix_tree_delete(&address_space
->page_tree
, idx
+ i
);
191 set_page_private(page
+ i
, 0);
193 ClearPageSwapCache(page
);
194 address_space
->nrpages
-= nr
;
195 __mod_node_page_state(page_pgdat(page
), NR_FILE_PAGES
, -nr
);
196 ADD_CACHE_INFO(del_total
, nr
);
200 * add_to_swap - allocate swap space for a page
201 * @page: page we want to move to swap
203 * Allocate swap space for the page and add the page to the
204 * swap cache. Caller needs to hold the page lock.
206 int add_to_swap(struct page
*page
)
211 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
212 VM_BUG_ON_PAGE(!PageUptodate(page
), page
);
214 entry
= get_swap_page(page
);
218 if (mem_cgroup_try_charge_swap(page
, entry
))
222 * Radix-tree node allocations from PF_MEMALLOC contexts could
223 * completely exhaust the page allocator. __GFP_NOMEMALLOC
224 * stops emergency reserves from being allocated.
226 * TODO: this could cause a theoretical memory reclaim
227 * deadlock in the swap out path.
230 * Add it to the swap cache.
232 err
= add_to_swap_cache(page
, entry
,
233 __GFP_HIGH
|__GFP_NOMEMALLOC
|__GFP_NOWARN
);
234 /* -ENOMEM radix-tree allocation failure */
237 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
238 * clear SWAP_HAS_CACHE flag.
242 * Normally the page will be dirtied in unmap because its pte should be
243 * dirty. A special case is MADV_FREE page. The page'e pte could have
244 * dirty bit cleared but the page's SwapBacked bit is still set because
245 * clearing the dirty bit and SwapBacked bit has no lock protected. For
246 * such page, unmap will not set dirty bit for it, so page reclaim will
247 * not write the page out. This can cause data corruption when the page
248 * is swap in later. Always setting the dirty bit for the page solves
251 set_page_dirty(page
);
256 put_swap_page(page
, entry
);
261 * This must be called only on pages that have
262 * been verified to be in the swap cache and locked.
263 * It will never put the page into the free list,
264 * the caller has a reference on the page.
266 void delete_from_swap_cache(struct page
*page
)
269 struct address_space
*address_space
;
271 entry
.val
= page_private(page
);
273 address_space
= swap_address_space(entry
);
274 spin_lock_irq(&address_space
->tree_lock
);
275 __delete_from_swap_cache(page
);
276 spin_unlock_irq(&address_space
->tree_lock
);
278 put_swap_page(page
, entry
);
279 page_ref_sub(page
, hpage_nr_pages(page
));
283 * If we are the only user, then try to free up the swap cache.
285 * Its ok to check for PageSwapCache without the page lock
286 * here because we are going to recheck again inside
287 * try_to_free_swap() _with_ the lock.
290 static inline void free_swap_cache(struct page
*page
)
292 if (PageSwapCache(page
) && !page_mapped(page
) && trylock_page(page
)) {
293 try_to_free_swap(page
);
299 * Perform a free_page(), also freeing any swap cache associated with
300 * this page if it is the last user of the page.
302 void free_page_and_swap_cache(struct page
*page
)
304 free_swap_cache(page
);
305 if (!is_huge_zero_page(page
))
310 * Passed an array of pages, drop them all from swapcache and then release
311 * them. They are removed from the LRU and freed if this is their last use.
313 void free_pages_and_swap_cache(struct page
**pages
, int nr
)
315 struct page
**pagep
= pages
;
319 for (i
= 0; i
< nr
; i
++)
320 free_swap_cache(pagep
[i
]);
321 release_pages(pagep
, nr
, false);
325 * Lookup a swap entry in the swap cache. A found page will be returned
326 * unlocked and with its refcount incremented - we rely on the kernel
327 * lock getting page table operations atomic even if we drop the page
328 * lock before returning.
330 struct page
*lookup_swap_cache(swp_entry_t entry
, struct vm_area_struct
*vma
,
334 unsigned long ra_info
;
335 int win
, hits
, readahead
;
337 page
= find_get_page(swap_address_space(entry
), swp_offset(entry
));
339 INC_CACHE_INFO(find_total
);
341 INC_CACHE_INFO(find_success
);
342 if (unlikely(PageTransCompound(page
)))
344 readahead
= TestClearPageReadahead(page
);
346 ra_info
= GET_SWAP_RA_VAL(vma
);
347 win
= SWAP_RA_WIN(ra_info
);
348 hits
= SWAP_RA_HITS(ra_info
);
350 hits
= min_t(int, hits
+ 1, SWAP_RA_HITS_MAX
);
351 atomic_long_set(&vma
->swap_readahead_info
,
352 SWAP_RA_VAL(addr
, win
, hits
));
355 count_vm_event(SWAP_RA_HIT
);
357 atomic_inc(&swapin_readahead_hits
);
363 struct page
*__read_swap_cache_async(swp_entry_t entry
, gfp_t gfp_mask
,
364 struct vm_area_struct
*vma
, unsigned long addr
,
365 bool *new_page_allocated
)
367 struct page
*found_page
, *new_page
= NULL
;
368 struct address_space
*swapper_space
= swap_address_space(entry
);
370 *new_page_allocated
= false;
374 * First check the swap cache. Since this is normally
375 * called after lookup_swap_cache() failed, re-calling
376 * that would confuse statistics.
378 found_page
= find_get_page(swapper_space
, swp_offset(entry
));
383 * Just skip read ahead for unused swap slot.
384 * During swap_off when swap_slot_cache is disabled,
385 * we have to handle the race between putting
386 * swap entry in swap cache and marking swap slot
387 * as SWAP_HAS_CACHE. That's done in later part of code or
388 * else swap_off will be aborted if we return NULL.
390 if (!__swp_swapcount(entry
) && swap_slot_cache_enabled
)
394 * Get a new page to read into from swap.
397 new_page
= alloc_page_vma(gfp_mask
, vma
, addr
);
399 break; /* Out of memory */
403 * call radix_tree_preload() while we can wait.
405 err
= radix_tree_maybe_preload(gfp_mask
& GFP_KERNEL
);
410 * Swap entry may have been freed since our caller observed it.
412 err
= swapcache_prepare(entry
);
413 if (err
== -EEXIST
) {
414 radix_tree_preload_end();
416 * We might race against get_swap_page() and stumble
417 * across a SWAP_HAS_CACHE swap_map entry whose page
418 * has not been brought into the swapcache yet.
423 if (err
) { /* swp entry is obsolete ? */
424 radix_tree_preload_end();
428 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
429 __SetPageLocked(new_page
);
430 __SetPageSwapBacked(new_page
);
431 err
= __add_to_swap_cache(new_page
, entry
);
433 radix_tree_preload_end();
435 * Initiate read into locked page and return.
437 lru_cache_add_anon(new_page
);
438 *new_page_allocated
= true;
441 radix_tree_preload_end();
442 __ClearPageLocked(new_page
);
444 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
445 * clear SWAP_HAS_CACHE flag.
447 put_swap_page(new_page
, entry
);
448 } while (err
!= -ENOMEM
);
456 * Locate a page of swap in physical memory, reserving swap cache space
457 * and reading the disk if it is not already cached.
458 * A failure return means that either the page allocation failed or that
459 * the swap entry is no longer in use.
461 struct page
*read_swap_cache_async(swp_entry_t entry
, gfp_t gfp_mask
,
462 struct vm_area_struct
*vma
, unsigned long addr
, bool do_poll
)
464 bool page_was_allocated
;
465 struct page
*retpage
= __read_swap_cache_async(entry
, gfp_mask
,
466 vma
, addr
, &page_was_allocated
);
468 if (page_was_allocated
)
469 swap_readpage(retpage
, do_poll
);
474 static unsigned int __swapin_nr_pages(unsigned long prev_offset
,
475 unsigned long offset
,
480 unsigned int pages
, last_ra
;
483 * This heuristic has been found to work well on both sequential and
484 * random loads, swapping to hard disk or to SSD: please don't ask
485 * what the "+ 2" means, it just happens to work well, that's all.
490 * We can have no readahead hits to judge by: but must not get
491 * stuck here forever, so check for an adjacent offset instead
492 * (and don't even bother to check whether swap type is same).
494 if (offset
!= prev_offset
+ 1 && offset
!= prev_offset
- 1)
497 unsigned int roundup
= 4;
498 while (roundup
< pages
)
503 if (pages
> max_pages
)
506 /* Don't shrink readahead too fast */
507 last_ra
= prev_win
/ 2;
514 static unsigned long swapin_nr_pages(unsigned long offset
)
516 static unsigned long prev_offset
;
517 unsigned int hits
, pages
, max_pages
;
518 static atomic_t last_readahead_pages
;
520 max_pages
= 1 << READ_ONCE(page_cluster
);
524 hits
= atomic_xchg(&swapin_readahead_hits
, 0);
525 pages
= __swapin_nr_pages(prev_offset
, offset
, hits
, max_pages
,
526 atomic_read(&last_readahead_pages
));
528 prev_offset
= offset
;
529 atomic_set(&last_readahead_pages
, pages
);
535 * swapin_readahead - swap in pages in hope we need them soon
536 * @entry: swap entry of this memory
537 * @gfp_mask: memory allocation flags
538 * @vma: user vma this address belongs to
539 * @addr: target address for mempolicy
541 * Returns the struct page for entry and addr, after queueing swapin.
543 * Primitive swap readahead code. We simply read an aligned block of
544 * (1 << page_cluster) entries in the swap area. This method is chosen
545 * because it doesn't cost us any seek time. We also make sure to queue
546 * the 'original' request together with the readahead ones...
548 * This has been extended to use the NUMA policies from the mm triggering
551 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
553 struct page
*swapin_readahead(swp_entry_t entry
, gfp_t gfp_mask
,
554 struct vm_area_struct
*vma
, unsigned long addr
)
557 unsigned long entry_offset
= swp_offset(entry
);
558 unsigned long offset
= entry_offset
;
559 unsigned long start_offset
, end_offset
;
561 struct blk_plug plug
;
562 bool do_poll
= true, page_allocated
;
564 mask
= swapin_nr_pages(offset
) - 1;
569 /* Read a page_cluster sized and aligned cluster around offset. */
570 start_offset
= offset
& ~mask
;
571 end_offset
= offset
| mask
;
572 if (!start_offset
) /* First page is swap header. */
575 blk_start_plug(&plug
);
576 for (offset
= start_offset
; offset
<= end_offset
; offset
++) {
577 /* Ok, do the async read-ahead now */
578 page
= __read_swap_cache_async(
579 swp_entry(swp_type(entry
), offset
),
580 gfp_mask
, vma
, addr
, &page_allocated
);
583 if (page_allocated
) {
584 swap_readpage(page
, false);
585 if (offset
!= entry_offset
&&
586 likely(!PageTransCompound(page
))) {
587 SetPageReadahead(page
);
588 count_vm_event(SWAP_RA
);
593 blk_finish_plug(&plug
);
595 lru_add_drain(); /* Push any new pages onto the LRU now */
597 return read_swap_cache_async(entry
, gfp_mask
, vma
, addr
, do_poll
);
600 int init_swap_address_space(unsigned int type
, unsigned long nr_pages
)
602 struct address_space
*spaces
, *space
;
605 nr
= DIV_ROUND_UP(nr_pages
, SWAP_ADDRESS_SPACE_PAGES
);
606 spaces
= kvzalloc(sizeof(struct address_space
) * nr
, GFP_KERNEL
);
609 for (i
= 0; i
< nr
; i
++) {
611 INIT_RADIX_TREE(&space
->page_tree
, GFP_ATOMIC
|__GFP_NOWARN
);
612 atomic_set(&space
->i_mmap_writable
, 0);
613 space
->a_ops
= &swap_aops
;
614 /* swap cache doesn't use writeback related tags */
615 mapping_set_no_writeback_tags(space
);
616 spin_lock_init(&space
->tree_lock
);
618 nr_swapper_spaces
[type
] = nr
;
619 rcu_assign_pointer(swapper_spaces
[type
], spaces
);
624 void exit_swap_address_space(unsigned int type
)
626 struct address_space
*spaces
;
628 spaces
= swapper_spaces
[type
];
629 nr_swapper_spaces
[type
] = 0;
630 rcu_assign_pointer(swapper_spaces
[type
], NULL
);
635 static inline void swap_ra_clamp_pfn(struct vm_area_struct
*vma
,
639 unsigned long *start
,
642 *start
= max3(lpfn
, PFN_DOWN(vma
->vm_start
),
643 PFN_DOWN(faddr
& PMD_MASK
));
644 *end
= min3(rpfn
, PFN_DOWN(vma
->vm_end
),
645 PFN_DOWN((faddr
& PMD_MASK
) + PMD_SIZE
));
648 struct page
*swap_readahead_detect(struct vm_fault
*vmf
,
649 struct vma_swap_readahead
*swap_ra
)
651 struct vm_area_struct
*vma
= vmf
->vma
;
652 unsigned long swap_ra_info
;
655 unsigned long faddr
, pfn
, fpfn
;
656 unsigned long start
, end
;
658 unsigned int max_win
, hits
, prev_win
, win
, left
;
663 max_win
= 1 << min_t(unsigned int, READ_ONCE(page_cluster
),
664 SWAP_RA_ORDER_CEILING
);
670 faddr
= vmf
->address
;
671 entry
= pte_to_swp_entry(vmf
->orig_pte
);
672 if ((unlikely(non_swap_entry(entry
))))
674 page
= lookup_swap_cache(entry
, vma
, faddr
);
678 fpfn
= PFN_DOWN(faddr
);
679 swap_ra_info
= GET_SWAP_RA_VAL(vma
);
680 pfn
= PFN_DOWN(SWAP_RA_ADDR(swap_ra_info
));
681 prev_win
= SWAP_RA_WIN(swap_ra_info
);
682 hits
= SWAP_RA_HITS(swap_ra_info
);
683 swap_ra
->win
= win
= __swapin_nr_pages(pfn
, fpfn
, hits
,
685 atomic_long_set(&vma
->swap_readahead_info
,
686 SWAP_RA_VAL(faddr
, win
, 0));
691 /* Copy the PTEs because the page table may be unmapped */
693 swap_ra_clamp_pfn(vma
, faddr
, fpfn
, fpfn
+ win
, &start
, &end
);
694 else if (pfn
== fpfn
+ 1)
695 swap_ra_clamp_pfn(vma
, faddr
, fpfn
- win
+ 1, fpfn
+ 1,
698 left
= (win
- 1) / 2;
699 swap_ra_clamp_pfn(vma
, faddr
, fpfn
- left
, fpfn
+ win
- left
,
702 swap_ra
->nr_pte
= end
- start
;
703 swap_ra
->offset
= fpfn
- start
;
704 pte
= vmf
->pte
- swap_ra
->offset
;
708 tpte
= swap_ra
->ptes
;
709 for (pfn
= start
; pfn
!= end
; pfn
++)
716 struct page
*do_swap_page_readahead(swp_entry_t fentry
, gfp_t gfp_mask
,
717 struct vm_fault
*vmf
,
718 struct vma_swap_readahead
*swap_ra
)
720 struct blk_plug plug
;
721 struct vm_area_struct
*vma
= vmf
->vma
;
728 if (swap_ra
->win
== 1)
731 blk_start_plug(&plug
);
732 for (i
= 0, pte
= swap_ra
->ptes
; i
< swap_ra
->nr_pte
;
735 if (pte_none(pentry
))
737 if (pte_present(pentry
))
739 entry
= pte_to_swp_entry(pentry
);
740 if (unlikely(non_swap_entry(entry
)))
742 page
= __read_swap_cache_async(entry
, gfp_mask
, vma
,
743 vmf
->address
, &page_allocated
);
746 if (page_allocated
) {
747 swap_readpage(page
, false);
748 if (i
!= swap_ra
->offset
&&
749 likely(!PageTransCompound(page
))) {
750 SetPageReadahead(page
);
751 count_vm_event(SWAP_RA
);
756 blk_finish_plug(&plug
);
759 return read_swap_cache_async(fentry
, gfp_mask
, vma
, vmf
->address
,
764 static ssize_t
vma_ra_enabled_show(struct kobject
*kobj
,
765 struct kobj_attribute
*attr
, char *buf
)
767 return sprintf(buf
, "%s\n", swap_vma_readahead
? "true" : "false");
769 static ssize_t
vma_ra_enabled_store(struct kobject
*kobj
,
770 struct kobj_attribute
*attr
,
771 const char *buf
, size_t count
)
773 if (!strncmp(buf
, "true", 4) || !strncmp(buf
, "1", 1))
774 swap_vma_readahead
= true;
775 else if (!strncmp(buf
, "false", 5) || !strncmp(buf
, "0", 1))
776 swap_vma_readahead
= false;
782 static struct kobj_attribute vma_ra_enabled_attr
=
783 __ATTR(vma_ra_enabled
, 0644, vma_ra_enabled_show
,
784 vma_ra_enabled_store
);
786 static struct attribute
*swap_attrs
[] = {
787 &vma_ra_enabled_attr
.attr
,
791 static struct attribute_group swap_attr_group
= {
795 static int __init
swap_init_sysfs(void)
798 struct kobject
*swap_kobj
;
800 swap_kobj
= kobject_create_and_add("swap", mm_kobj
);
802 pr_err("failed to create swap kobject\n");
805 err
= sysfs_create_group(swap_kobj
, &swap_attr_group
);
807 pr_err("failed to register swap group\n");
813 kobject_put(swap_kobj
);
816 subsys_initcall(swap_init_sysfs
);