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
];
41 #define INC_CACHE_INFO(x) do { swap_cache_info.x++; } while (0)
42 #define ADD_CACHE_INFO(x, nr) do { swap_cache_info.x += (nr); } while (0)
45 unsigned long add_total
;
46 unsigned long del_total
;
47 unsigned long find_success
;
48 unsigned long find_total
;
51 unsigned long total_swapcache_pages(void)
53 unsigned int i
, j
, nr
;
54 unsigned long ret
= 0;
55 struct address_space
*spaces
;
58 for (i
= 0; i
< MAX_SWAPFILES
; i
++) {
60 * The corresponding entries in nr_swapper_spaces and
61 * swapper_spaces will be reused only after at least
62 * one grace period. So it is impossible for them
63 * belongs to different usage.
65 nr
= nr_swapper_spaces
[i
];
66 spaces
= rcu_dereference(swapper_spaces
[i
]);
69 for (j
= 0; j
< nr
; j
++)
70 ret
+= spaces
[j
].nrpages
;
76 static atomic_t swapin_readahead_hits
= ATOMIC_INIT(4);
78 void show_swap_cache_info(void)
80 printk("%lu pages in swap cache\n", total_swapcache_pages());
81 printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
82 swap_cache_info
.add_total
, swap_cache_info
.del_total
,
83 swap_cache_info
.find_success
, swap_cache_info
.find_total
);
84 printk("Free swap = %ldkB\n",
85 get_nr_swap_pages() << (PAGE_SHIFT
- 10));
86 printk("Total swap = %lukB\n", total_swap_pages
<< (PAGE_SHIFT
- 10));
90 * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
91 * but sets SwapCache flag and private instead of mapping and index.
93 int __add_to_swap_cache(struct page
*page
, swp_entry_t entry
)
95 int error
, i
, nr
= hpage_nr_pages(page
);
96 struct address_space
*address_space
;
97 pgoff_t idx
= swp_offset(entry
);
99 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
100 VM_BUG_ON_PAGE(PageSwapCache(page
), page
);
101 VM_BUG_ON_PAGE(!PageSwapBacked(page
), page
);
103 page_ref_add(page
, nr
);
104 SetPageSwapCache(page
);
106 address_space
= swap_address_space(entry
);
107 spin_lock_irq(&address_space
->tree_lock
);
108 for (i
= 0; i
< nr
; i
++) {
109 set_page_private(page
+ i
, entry
.val
+ i
);
110 error
= radix_tree_insert(&address_space
->page_tree
,
115 if (likely(!error
)) {
116 address_space
->nrpages
+= nr
;
117 __mod_node_page_state(page_pgdat(page
), NR_FILE_PAGES
, nr
);
118 ADD_CACHE_INFO(add_total
, nr
);
121 * Only the context which have set SWAP_HAS_CACHE flag
122 * would call add_to_swap_cache().
123 * So add_to_swap_cache() doesn't returns -EEXIST.
125 VM_BUG_ON(error
== -EEXIST
);
126 set_page_private(page
+ i
, 0UL);
128 radix_tree_delete(&address_space
->page_tree
, idx
+ i
);
129 set_page_private(page
+ i
, 0UL);
131 ClearPageSwapCache(page
);
132 page_ref_sub(page
, nr
);
134 spin_unlock_irq(&address_space
->tree_lock
);
140 int add_to_swap_cache(struct page
*page
, swp_entry_t entry
, gfp_t gfp_mask
)
144 error
= radix_tree_maybe_preload_order(gfp_mask
, compound_order(page
));
146 error
= __add_to_swap_cache(page
, entry
);
147 radix_tree_preload_end();
153 * This must be called only on pages that have
154 * been verified to be in the swap cache.
156 void __delete_from_swap_cache(struct page
*page
)
158 struct address_space
*address_space
;
159 int i
, nr
= hpage_nr_pages(page
);
163 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
164 VM_BUG_ON_PAGE(!PageSwapCache(page
), page
);
165 VM_BUG_ON_PAGE(PageWriteback(page
), page
);
167 entry
.val
= page_private(page
);
168 address_space
= swap_address_space(entry
);
169 idx
= swp_offset(entry
);
170 for (i
= 0; i
< nr
; i
++) {
171 radix_tree_delete(&address_space
->page_tree
, idx
+ i
);
172 set_page_private(page
+ i
, 0);
174 ClearPageSwapCache(page
);
175 address_space
->nrpages
-= nr
;
176 __mod_node_page_state(page_pgdat(page
), NR_FILE_PAGES
, -nr
);
177 ADD_CACHE_INFO(del_total
, nr
);
181 * add_to_swap - allocate swap space for a page
182 * @page: page we want to move to swap
184 * Allocate swap space for the page and add the page to the
185 * swap cache. Caller needs to hold the page lock.
187 int add_to_swap(struct page
*page
)
192 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
193 VM_BUG_ON_PAGE(!PageUptodate(page
), page
);
195 entry
= get_swap_page(page
);
199 if (mem_cgroup_try_charge_swap(page
, entry
))
203 * Radix-tree node allocations from PF_MEMALLOC contexts could
204 * completely exhaust the page allocator. __GFP_NOMEMALLOC
205 * stops emergency reserves from being allocated.
207 * TODO: this could cause a theoretical memory reclaim
208 * deadlock in the swap out path.
211 * Add it to the swap cache.
213 err
= add_to_swap_cache(page
, entry
,
214 __GFP_HIGH
|__GFP_NOMEMALLOC
|__GFP_NOWARN
);
215 /* -ENOMEM radix-tree allocation failure */
218 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
219 * clear SWAP_HAS_CACHE flag.
223 * Normally the page will be dirtied in unmap because its pte should be
224 * dirty. A special case is MADV_FREE page. The page'e pte could have
225 * dirty bit cleared but the page's SwapBacked bit is still set because
226 * clearing the dirty bit and SwapBacked bit has no lock protected. For
227 * such page, unmap will not set dirty bit for it, so page reclaim will
228 * not write the page out. This can cause data corruption when the page
229 * is swap in later. Always setting the dirty bit for the page solves
232 set_page_dirty(page
);
237 put_swap_page(page
, entry
);
242 * This must be called only on pages that have
243 * been verified to be in the swap cache and locked.
244 * It will never put the page into the free list,
245 * the caller has a reference on the page.
247 void delete_from_swap_cache(struct page
*page
)
250 struct address_space
*address_space
;
252 entry
.val
= page_private(page
);
254 address_space
= swap_address_space(entry
);
255 spin_lock_irq(&address_space
->tree_lock
);
256 __delete_from_swap_cache(page
);
257 spin_unlock_irq(&address_space
->tree_lock
);
259 put_swap_page(page
, entry
);
260 page_ref_sub(page
, hpage_nr_pages(page
));
264 * If we are the only user, then try to free up the swap cache.
266 * Its ok to check for PageSwapCache without the page lock
267 * here because we are going to recheck again inside
268 * try_to_free_swap() _with_ the lock.
271 static inline void free_swap_cache(struct page
*page
)
273 if (PageSwapCache(page
) && !page_mapped(page
) && trylock_page(page
)) {
274 try_to_free_swap(page
);
280 * Perform a free_page(), also freeing any swap cache associated with
281 * this page if it is the last user of the page.
283 void free_page_and_swap_cache(struct page
*page
)
285 free_swap_cache(page
);
286 if (!is_huge_zero_page(page
))
291 * Passed an array of pages, drop them all from swapcache and then release
292 * them. They are removed from the LRU and freed if this is their last use.
294 void free_pages_and_swap_cache(struct page
**pages
, int nr
)
296 struct page
**pagep
= pages
;
300 for (i
= 0; i
< nr
; i
++)
301 free_swap_cache(pagep
[i
]);
302 release_pages(pagep
, nr
, false);
306 * Lookup a swap entry in the swap cache. A found page will be returned
307 * unlocked and with its refcount incremented - we rely on the kernel
308 * lock getting page table operations atomic even if we drop the page
309 * lock before returning.
311 struct page
* lookup_swap_cache(swp_entry_t entry
)
315 page
= find_get_page(swap_address_space(entry
), swp_offset(entry
));
317 if (page
&& likely(!PageTransCompound(page
))) {
318 INC_CACHE_INFO(find_success
);
319 if (TestClearPageReadahead(page
))
320 atomic_inc(&swapin_readahead_hits
);
323 INC_CACHE_INFO(find_total
);
327 struct page
*__read_swap_cache_async(swp_entry_t entry
, gfp_t gfp_mask
,
328 struct vm_area_struct
*vma
, unsigned long addr
,
329 bool *new_page_allocated
)
331 struct page
*found_page
, *new_page
= NULL
;
332 struct address_space
*swapper_space
= swap_address_space(entry
);
334 *new_page_allocated
= false;
338 * First check the swap cache. Since this is normally
339 * called after lookup_swap_cache() failed, re-calling
340 * that would confuse statistics.
342 found_page
= find_get_page(swapper_space
, swp_offset(entry
));
347 * Just skip read ahead for unused swap slot.
348 * During swap_off when swap_slot_cache is disabled,
349 * we have to handle the race between putting
350 * swap entry in swap cache and marking swap slot
351 * as SWAP_HAS_CACHE. That's done in later part of code or
352 * else swap_off will be aborted if we return NULL.
354 if (!__swp_swapcount(entry
) && swap_slot_cache_enabled
)
358 * Get a new page to read into from swap.
361 new_page
= alloc_page_vma(gfp_mask
, vma
, addr
);
363 break; /* Out of memory */
367 * call radix_tree_preload() while we can wait.
369 err
= radix_tree_maybe_preload(gfp_mask
& GFP_KERNEL
);
374 * Swap entry may have been freed since our caller observed it.
376 err
= swapcache_prepare(entry
);
377 if (err
== -EEXIST
) {
378 radix_tree_preload_end();
380 * We might race against get_swap_page() and stumble
381 * across a SWAP_HAS_CACHE swap_map entry whose page
382 * has not been brought into the swapcache yet.
387 if (err
) { /* swp entry is obsolete ? */
388 radix_tree_preload_end();
392 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
393 __SetPageLocked(new_page
);
394 __SetPageSwapBacked(new_page
);
395 err
= __add_to_swap_cache(new_page
, entry
);
397 radix_tree_preload_end();
399 * Initiate read into locked page and return.
401 lru_cache_add_anon(new_page
);
402 *new_page_allocated
= true;
405 radix_tree_preload_end();
406 __ClearPageLocked(new_page
);
408 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
409 * clear SWAP_HAS_CACHE flag.
411 put_swap_page(new_page
, entry
);
412 } while (err
!= -ENOMEM
);
420 * Locate a page of swap in physical memory, reserving swap cache space
421 * and reading the disk if it is not already cached.
422 * A failure return means that either the page allocation failed or that
423 * the swap entry is no longer in use.
425 struct page
*read_swap_cache_async(swp_entry_t entry
, gfp_t gfp_mask
,
426 struct vm_area_struct
*vma
, unsigned long addr
, bool do_poll
)
428 bool page_was_allocated
;
429 struct page
*retpage
= __read_swap_cache_async(entry
, gfp_mask
,
430 vma
, addr
, &page_was_allocated
);
432 if (page_was_allocated
)
433 swap_readpage(retpage
, do_poll
);
438 static unsigned long swapin_nr_pages(unsigned long offset
)
440 static unsigned long prev_offset
;
441 unsigned int pages
, max_pages
, last_ra
;
442 static atomic_t last_readahead_pages
;
444 max_pages
= 1 << READ_ONCE(page_cluster
);
449 * This heuristic has been found to work well on both sequential and
450 * random loads, swapping to hard disk or to SSD: please don't ask
451 * what the "+ 2" means, it just happens to work well, that's all.
453 pages
= atomic_xchg(&swapin_readahead_hits
, 0) + 2;
456 * We can have no readahead hits to judge by: but must not get
457 * stuck here forever, so check for an adjacent offset instead
458 * (and don't even bother to check whether swap type is same).
460 if (offset
!= prev_offset
+ 1 && offset
!= prev_offset
- 1)
462 prev_offset
= offset
;
464 unsigned int roundup
= 4;
465 while (roundup
< pages
)
470 if (pages
> max_pages
)
473 /* Don't shrink readahead too fast */
474 last_ra
= atomic_read(&last_readahead_pages
) / 2;
477 atomic_set(&last_readahead_pages
, pages
);
483 * swapin_readahead - swap in pages in hope we need them soon
484 * @entry: swap entry of this memory
485 * @gfp_mask: memory allocation flags
486 * @vma: user vma this address belongs to
487 * @addr: target address for mempolicy
489 * Returns the struct page for entry and addr, after queueing swapin.
491 * Primitive swap readahead code. We simply read an aligned block of
492 * (1 << page_cluster) entries in the swap area. This method is chosen
493 * because it doesn't cost us any seek time. We also make sure to queue
494 * the 'original' request together with the readahead ones...
496 * This has been extended to use the NUMA policies from the mm triggering
499 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
501 struct page
*swapin_readahead(swp_entry_t entry
, gfp_t gfp_mask
,
502 struct vm_area_struct
*vma
, unsigned long addr
)
505 unsigned long entry_offset
= swp_offset(entry
);
506 unsigned long offset
= entry_offset
;
507 unsigned long start_offset
, end_offset
;
509 struct blk_plug plug
;
512 mask
= swapin_nr_pages(offset
) - 1;
517 /* Read a page_cluster sized and aligned cluster around offset. */
518 start_offset
= offset
& ~mask
;
519 end_offset
= offset
| mask
;
520 if (!start_offset
) /* First page is swap header. */
523 blk_start_plug(&plug
);
524 for (offset
= start_offset
; offset
<= end_offset
; offset
++) {
525 /* Ok, do the async read-ahead now */
526 page
= read_swap_cache_async(swp_entry(swp_type(entry
), offset
),
527 gfp_mask
, vma
, addr
, false);
530 if (offset
!= entry_offset
&& likely(!PageTransCompound(page
)))
531 SetPageReadahead(page
);
534 blk_finish_plug(&plug
);
536 lru_add_drain(); /* Push any new pages onto the LRU now */
538 return read_swap_cache_async(entry
, gfp_mask
, vma
, addr
, do_poll
);
541 int init_swap_address_space(unsigned int type
, unsigned long nr_pages
)
543 struct address_space
*spaces
, *space
;
546 nr
= DIV_ROUND_UP(nr_pages
, SWAP_ADDRESS_SPACE_PAGES
);
547 spaces
= kvzalloc(sizeof(struct address_space
) * nr
, GFP_KERNEL
);
550 for (i
= 0; i
< nr
; i
++) {
552 INIT_RADIX_TREE(&space
->page_tree
, GFP_ATOMIC
|__GFP_NOWARN
);
553 atomic_set(&space
->i_mmap_writable
, 0);
554 space
->a_ops
= &swap_aops
;
555 /* swap cache doesn't use writeback related tags */
556 mapping_set_no_writeback_tags(space
);
557 spin_lock_init(&space
->tree_lock
);
559 nr_swapper_spaces
[type
] = nr
;
560 rcu_assign_pointer(swapper_spaces
[type
], spaces
);
565 void exit_swap_address_space(unsigned int type
)
567 struct address_space
*spaces
;
569 spaces
= swapper_spaces
[type
];
570 nr_swapper_spaces
[type
] = 0;
571 rcu_assign_pointer(swapper_spaces
[type
], NULL
);