Loading...
1/*
2 * linux/mm/swap_state.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie
6 *
7 * Rewritten to use page cache, (C) 1998 Stephen Tweedie
8 */
9#include <linux/mm.h>
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/page_cgroup.h>
21
22#include <asm/pgtable.h>
23
24/*
25 * swapper_space is a fiction, retained to simplify the path through
26 * vmscan's shrink_page_list.
27 */
28static const struct address_space_operations swap_aops = {
29 .writepage = swap_writepage,
30 .set_page_dirty = swap_set_page_dirty,
31 .migratepage = migrate_page,
32};
33
34static struct backing_dev_info swap_backing_dev_info = {
35 .name = "swap",
36 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
37};
38
39struct address_space swapper_spaces[MAX_SWAPFILES] = {
40 [0 ... MAX_SWAPFILES - 1] = {
41 .page_tree = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
42 .a_ops = &swap_aops,
43 .backing_dev_info = &swap_backing_dev_info,
44 }
45};
46
47#define INC_CACHE_INFO(x) do { swap_cache_info.x++; } while (0)
48
49static struct {
50 unsigned long add_total;
51 unsigned long del_total;
52 unsigned long find_success;
53 unsigned long find_total;
54} swap_cache_info;
55
56unsigned long total_swapcache_pages(void)
57{
58 int i;
59 unsigned long ret = 0;
60
61 for (i = 0; i < MAX_SWAPFILES; i++)
62 ret += swapper_spaces[i].nrpages;
63 return ret;
64}
65
66static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
67
68void show_swap_cache_info(void)
69{
70 printk("%lu pages in swap cache\n", total_swapcache_pages());
71 printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
72 swap_cache_info.add_total, swap_cache_info.del_total,
73 swap_cache_info.find_success, swap_cache_info.find_total);
74 printk("Free swap = %ldkB\n",
75 get_nr_swap_pages() << (PAGE_SHIFT - 10));
76 printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
77}
78
79/*
80 * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
81 * but sets SwapCache flag and private instead of mapping and index.
82 */
83int __add_to_swap_cache(struct page *page, swp_entry_t entry)
84{
85 int error;
86 struct address_space *address_space;
87
88 VM_BUG_ON_PAGE(!PageLocked(page), page);
89 VM_BUG_ON_PAGE(PageSwapCache(page), page);
90 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
91
92 page_cache_get(page);
93 SetPageSwapCache(page);
94 set_page_private(page, entry.val);
95
96 address_space = swap_address_space(entry);
97 spin_lock_irq(&address_space->tree_lock);
98 error = radix_tree_insert(&address_space->page_tree,
99 entry.val, page);
100 if (likely(!error)) {
101 address_space->nrpages++;
102 __inc_zone_page_state(page, NR_FILE_PAGES);
103 INC_CACHE_INFO(add_total);
104 }
105 spin_unlock_irq(&address_space->tree_lock);
106
107 if (unlikely(error)) {
108 /*
109 * Only the context which have set SWAP_HAS_CACHE flag
110 * would call add_to_swap_cache().
111 * So add_to_swap_cache() doesn't returns -EEXIST.
112 */
113 VM_BUG_ON(error == -EEXIST);
114 set_page_private(page, 0UL);
115 ClearPageSwapCache(page);
116 page_cache_release(page);
117 }
118
119 return error;
120}
121
122
123int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
124{
125 int error;
126
127 error = radix_tree_maybe_preload(gfp_mask);
128 if (!error) {
129 error = __add_to_swap_cache(page, entry);
130 radix_tree_preload_end();
131 }
132 return error;
133}
134
135/*
136 * This must be called only on pages that have
137 * been verified to be in the swap cache.
138 */
139void __delete_from_swap_cache(struct page *page)
140{
141 swp_entry_t entry;
142 struct address_space *address_space;
143
144 VM_BUG_ON_PAGE(!PageLocked(page), page);
145 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
146 VM_BUG_ON_PAGE(PageWriteback(page), page);
147
148 entry.val = page_private(page);
149 address_space = swap_address_space(entry);
150 radix_tree_delete(&address_space->page_tree, page_private(page));
151 set_page_private(page, 0);
152 ClearPageSwapCache(page);
153 address_space->nrpages--;
154 __dec_zone_page_state(page, NR_FILE_PAGES);
155 INC_CACHE_INFO(del_total);
156}
157
158/**
159 * add_to_swap - allocate swap space for a page
160 * @page: page we want to move to swap
161 *
162 * Allocate swap space for the page and add the page to the
163 * swap cache. Caller needs to hold the page lock.
164 */
165int add_to_swap(struct page *page, struct list_head *list)
166{
167 swp_entry_t entry;
168 int err;
169
170 VM_BUG_ON_PAGE(!PageLocked(page), page);
171 VM_BUG_ON_PAGE(!PageUptodate(page), page);
172
173 entry = get_swap_page();
174 if (!entry.val)
175 return 0;
176
177 if (unlikely(PageTransHuge(page)))
178 if (unlikely(split_huge_page_to_list(page, list))) {
179 swapcache_free(entry, NULL);
180 return 0;
181 }
182
183 /*
184 * Radix-tree node allocations from PF_MEMALLOC contexts could
185 * completely exhaust the page allocator. __GFP_NOMEMALLOC
186 * stops emergency reserves from being allocated.
187 *
188 * TODO: this could cause a theoretical memory reclaim
189 * deadlock in the swap out path.
190 */
191 /*
192 * Add it to the swap cache and mark it dirty
193 */
194 err = add_to_swap_cache(page, entry,
195 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
196
197 if (!err) { /* Success */
198 SetPageDirty(page);
199 return 1;
200 } else { /* -ENOMEM radix-tree allocation failure */
201 /*
202 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
203 * clear SWAP_HAS_CACHE flag.
204 */
205 swapcache_free(entry, NULL);
206 return 0;
207 }
208}
209
210/*
211 * This must be called only on pages that have
212 * been verified to be in the swap cache and locked.
213 * It will never put the page into the free list,
214 * the caller has a reference on the page.
215 */
216void delete_from_swap_cache(struct page *page)
217{
218 swp_entry_t entry;
219 struct address_space *address_space;
220
221 entry.val = page_private(page);
222
223 address_space = swap_address_space(entry);
224 spin_lock_irq(&address_space->tree_lock);
225 __delete_from_swap_cache(page);
226 spin_unlock_irq(&address_space->tree_lock);
227
228 swapcache_free(entry, page);
229 page_cache_release(page);
230}
231
232/*
233 * If we are the only user, then try to free up the swap cache.
234 *
235 * Its ok to check for PageSwapCache without the page lock
236 * here because we are going to recheck again inside
237 * try_to_free_swap() _with_ the lock.
238 * - Marcelo
239 */
240static inline void free_swap_cache(struct page *page)
241{
242 if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
243 try_to_free_swap(page);
244 unlock_page(page);
245 }
246}
247
248/*
249 * Perform a free_page(), also freeing any swap cache associated with
250 * this page if it is the last user of the page.
251 */
252void free_page_and_swap_cache(struct page *page)
253{
254 free_swap_cache(page);
255 page_cache_release(page);
256}
257
258/*
259 * Passed an array of pages, drop them all from swapcache and then release
260 * them. They are removed from the LRU and freed if this is their last use.
261 */
262void free_pages_and_swap_cache(struct page **pages, int nr)
263{
264 struct page **pagep = pages;
265
266 lru_add_drain();
267 while (nr) {
268 int todo = min(nr, PAGEVEC_SIZE);
269 int i;
270
271 for (i = 0; i < todo; i++)
272 free_swap_cache(pagep[i]);
273 release_pages(pagep, todo, 0);
274 pagep += todo;
275 nr -= todo;
276 }
277}
278
279/*
280 * Lookup a swap entry in the swap cache. A found page will be returned
281 * unlocked and with its refcount incremented - we rely on the kernel
282 * lock getting page table operations atomic even if we drop the page
283 * lock before returning.
284 */
285struct page * lookup_swap_cache(swp_entry_t entry)
286{
287 struct page *page;
288
289 page = find_get_page(swap_address_space(entry), entry.val);
290
291 if (page) {
292 INC_CACHE_INFO(find_success);
293 if (TestClearPageReadahead(page))
294 atomic_inc(&swapin_readahead_hits);
295 }
296
297 INC_CACHE_INFO(find_total);
298 return page;
299}
300
301/*
302 * Locate a page of swap in physical memory, reserving swap cache space
303 * and reading the disk if it is not already cached.
304 * A failure return means that either the page allocation failed or that
305 * the swap entry is no longer in use.
306 */
307struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
308 struct vm_area_struct *vma, unsigned long addr)
309{
310 struct page *found_page, *new_page = NULL;
311 int err;
312
313 do {
314 /*
315 * First check the swap cache. Since this is normally
316 * called after lookup_swap_cache() failed, re-calling
317 * that would confuse statistics.
318 */
319 found_page = find_get_page(swap_address_space(entry),
320 entry.val);
321 if (found_page)
322 break;
323
324 /*
325 * Get a new page to read into from swap.
326 */
327 if (!new_page) {
328 new_page = alloc_page_vma(gfp_mask, vma, addr);
329 if (!new_page)
330 break; /* Out of memory */
331 }
332
333 /*
334 * call radix_tree_preload() while we can wait.
335 */
336 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
337 if (err)
338 break;
339
340 /*
341 * Swap entry may have been freed since our caller observed it.
342 */
343 err = swapcache_prepare(entry);
344 if (err == -EEXIST) {
345 radix_tree_preload_end();
346 /*
347 * We might race against get_swap_page() and stumble
348 * across a SWAP_HAS_CACHE swap_map entry whose page
349 * has not been brought into the swapcache yet, while
350 * the other end is scheduled away waiting on discard
351 * I/O completion at scan_swap_map().
352 *
353 * In order to avoid turning this transitory state
354 * into a permanent loop around this -EEXIST case
355 * if !CONFIG_PREEMPT and the I/O completion happens
356 * to be waiting on the CPU waitqueue where we are now
357 * busy looping, we just conditionally invoke the
358 * scheduler here, if there are some more important
359 * tasks to run.
360 */
361 cond_resched();
362 continue;
363 }
364 if (err) { /* swp entry is obsolete ? */
365 radix_tree_preload_end();
366 break;
367 }
368
369 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
370 __set_page_locked(new_page);
371 SetPageSwapBacked(new_page);
372 err = __add_to_swap_cache(new_page, entry);
373 if (likely(!err)) {
374 radix_tree_preload_end();
375 /*
376 * Initiate read into locked page and return.
377 */
378 lru_cache_add_anon(new_page);
379 swap_readpage(new_page);
380 return new_page;
381 }
382 radix_tree_preload_end();
383 ClearPageSwapBacked(new_page);
384 __clear_page_locked(new_page);
385 /*
386 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
387 * clear SWAP_HAS_CACHE flag.
388 */
389 swapcache_free(entry, NULL);
390 } while (err != -ENOMEM);
391
392 if (new_page)
393 page_cache_release(new_page);
394 return found_page;
395}
396
397static unsigned long swapin_nr_pages(unsigned long offset)
398{
399 static unsigned long prev_offset;
400 unsigned int pages, max_pages, last_ra;
401 static atomic_t last_readahead_pages;
402
403 max_pages = 1 << ACCESS_ONCE(page_cluster);
404 if (max_pages <= 1)
405 return 1;
406
407 /*
408 * This heuristic has been found to work well on both sequential and
409 * random loads, swapping to hard disk or to SSD: please don't ask
410 * what the "+ 2" means, it just happens to work well, that's all.
411 */
412 pages = atomic_xchg(&swapin_readahead_hits, 0) + 2;
413 if (pages == 2) {
414 /*
415 * We can have no readahead hits to judge by: but must not get
416 * stuck here forever, so check for an adjacent offset instead
417 * (and don't even bother to check whether swap type is same).
418 */
419 if (offset != prev_offset + 1 && offset != prev_offset - 1)
420 pages = 1;
421 prev_offset = offset;
422 } else {
423 unsigned int roundup = 4;
424 while (roundup < pages)
425 roundup <<= 1;
426 pages = roundup;
427 }
428
429 if (pages > max_pages)
430 pages = max_pages;
431
432 /* Don't shrink readahead too fast */
433 last_ra = atomic_read(&last_readahead_pages) / 2;
434 if (pages < last_ra)
435 pages = last_ra;
436 atomic_set(&last_readahead_pages, pages);
437
438 return pages;
439}
440
441/**
442 * swapin_readahead - swap in pages in hope we need them soon
443 * @entry: swap entry of this memory
444 * @gfp_mask: memory allocation flags
445 * @vma: user vma this address belongs to
446 * @addr: target address for mempolicy
447 *
448 * Returns the struct page for entry and addr, after queueing swapin.
449 *
450 * Primitive swap readahead code. We simply read an aligned block of
451 * (1 << page_cluster) entries in the swap area. This method is chosen
452 * because it doesn't cost us any seek time. We also make sure to queue
453 * the 'original' request together with the readahead ones...
454 *
455 * This has been extended to use the NUMA policies from the mm triggering
456 * the readahead.
457 *
458 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
459 */
460struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
461 struct vm_area_struct *vma, unsigned long addr)
462{
463 struct page *page;
464 unsigned long entry_offset = swp_offset(entry);
465 unsigned long offset = entry_offset;
466 unsigned long start_offset, end_offset;
467 unsigned long mask;
468 struct blk_plug plug;
469
470 mask = swapin_nr_pages(offset) - 1;
471 if (!mask)
472 goto skip;
473
474 /* Read a page_cluster sized and aligned cluster around offset. */
475 start_offset = offset & ~mask;
476 end_offset = offset | mask;
477 if (!start_offset) /* First page is swap header. */
478 start_offset++;
479
480 blk_start_plug(&plug);
481 for (offset = start_offset; offset <= end_offset ; offset++) {
482 /* Ok, do the async read-ahead now */
483 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
484 gfp_mask, vma, addr);
485 if (!page)
486 continue;
487 if (offset != entry_offset)
488 SetPageReadahead(page);
489 page_cache_release(page);
490 }
491 blk_finish_plug(&plug);
492
493 lru_add_drain(); /* Push any new pages onto the LRU now */
494skip:
495 return read_swap_cache_async(entry, gfp_mask, vma, addr);
496}
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#include <linux/shmem_fs.h>
25#include "internal.h"
26#include "swap.h"
27
28/*
29 * swapper_space is a fiction, retained to simplify the path through
30 * vmscan's shrink_page_list.
31 */
32static const struct address_space_operations swap_aops = {
33 .writepage = swap_writepage,
34 .dirty_folio = noop_dirty_folio,
35#ifdef CONFIG_MIGRATION
36 .migrate_folio = migrate_folio,
37#endif
38};
39
40struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly;
41static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly;
42static bool enable_vma_readahead __read_mostly = true;
43
44#define SWAP_RA_WIN_SHIFT (PAGE_SHIFT / 2)
45#define SWAP_RA_HITS_MASK ((1UL << SWAP_RA_WIN_SHIFT) - 1)
46#define SWAP_RA_HITS_MAX SWAP_RA_HITS_MASK
47#define SWAP_RA_WIN_MASK (~PAGE_MASK & ~SWAP_RA_HITS_MASK)
48
49#define SWAP_RA_HITS(v) ((v) & SWAP_RA_HITS_MASK)
50#define SWAP_RA_WIN(v) (((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
51#define SWAP_RA_ADDR(v) ((v) & PAGE_MASK)
52
53#define SWAP_RA_VAL(addr, win, hits) \
54 (((addr) & PAGE_MASK) | \
55 (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) | \
56 ((hits) & SWAP_RA_HITS_MASK))
57
58/* Initial readahead hits is 4 to start up with a small window */
59#define GET_SWAP_RA_VAL(vma) \
60 (atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
61
62static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
63
64void show_swap_cache_info(void)
65{
66 printk("%lu pages in swap cache\n", total_swapcache_pages());
67 printk("Free swap = %ldkB\n",
68 get_nr_swap_pages() << (PAGE_SHIFT - 10));
69 printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
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 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
98 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
99 VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio);
100
101 folio_ref_add(folio, nr);
102 folio_set_swapcache(folio);
103
104 do {
105 xas_lock_irq(&xas);
106 xas_create_range(&xas);
107 if (xas_error(&xas))
108 goto unlock;
109 for (i = 0; i < nr; i++) {
110 VM_BUG_ON_FOLIO(xas.xa_index != idx + i, folio);
111 old = xas_load(&xas);
112 if (xa_is_value(old)) {
113 if (shadowp)
114 *shadowp = old;
115 }
116 set_page_private(folio_page(folio, i), entry.val + i);
117 xas_store(&xas, folio);
118 xas_next(&xas);
119 }
120 address_space->nrpages += nr;
121 __node_stat_mod_folio(folio, NR_FILE_PAGES, nr);
122 __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, nr);
123unlock:
124 xas_unlock_irq(&xas);
125 } while (xas_nomem(&xas, gfp));
126
127 if (!xas_error(&xas))
128 return 0;
129
130 folio_clear_swapcache(folio);
131 folio_ref_sub(folio, nr);
132 return xas_error(&xas);
133}
134
135/*
136 * This must be called only on folios that have
137 * been verified to be in the swap cache.
138 */
139void __delete_from_swap_cache(struct folio *folio,
140 swp_entry_t entry, void *shadow)
141{
142 struct address_space *address_space = swap_address_space(entry);
143 int i;
144 long nr = folio_nr_pages(folio);
145 pgoff_t idx = swp_offset(entry);
146 XA_STATE(xas, &address_space->i_pages, idx);
147
148 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
149 VM_BUG_ON_FOLIO(!folio_test_swapcache(folio), folio);
150 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
151
152 for (i = 0; i < nr; i++) {
153 void *entry = xas_store(&xas, shadow);
154 VM_BUG_ON_PAGE(entry != folio, entry);
155 set_page_private(folio_page(folio, i), 0);
156 xas_next(&xas);
157 }
158 folio_clear_swapcache(folio);
159 address_space->nrpages -= nr;
160 __node_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
161 __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, -nr);
162}
163
164/**
165 * add_to_swap - allocate swap space for a folio
166 * @folio: folio we want to move to swap
167 *
168 * Allocate swap space for the folio and add the folio to the
169 * swap cache.
170 *
171 * Context: Caller needs to hold the folio lock.
172 * Return: Whether the folio was added to the swap cache.
173 */
174bool add_to_swap(struct folio *folio)
175{
176 swp_entry_t entry;
177 int err;
178
179 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
180 VM_BUG_ON_FOLIO(!folio_test_uptodate(folio), folio);
181
182 entry = folio_alloc_swap(folio);
183 if (!entry.val)
184 return false;
185
186 /*
187 * XArray node allocations from PF_MEMALLOC contexts could
188 * completely exhaust the page allocator. __GFP_NOMEMALLOC
189 * stops emergency reserves from being allocated.
190 *
191 * TODO: this could cause a theoretical memory reclaim
192 * deadlock in the swap out path.
193 */
194 /*
195 * Add it to the swap cache.
196 */
197 err = add_to_swap_cache(folio, entry,
198 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN, NULL);
199 if (err)
200 /*
201 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
202 * clear SWAP_HAS_CACHE flag.
203 */
204 goto fail;
205 /*
206 * Normally the folio will be dirtied in unmap because its
207 * pte should be dirty. A special case is MADV_FREE page. The
208 * page's pte could have dirty bit cleared but the folio's
209 * SwapBacked flag is still set because clearing the dirty bit
210 * and SwapBacked flag has no lock protected. For such folio,
211 * unmap will not set dirty bit for it, so folio reclaim will
212 * not write the folio out. This can cause data corruption when
213 * the folio is swapped in later. Always setting the dirty flag
214 * for the folio solves the problem.
215 */
216 folio_mark_dirty(folio);
217
218 return true;
219
220fail:
221 put_swap_folio(folio, entry);
222 return false;
223}
224
225/*
226 * This must be called only on folios that have
227 * been verified to be in the swap cache and locked.
228 * It will never put the folio into the free list,
229 * the caller has a reference on the folio.
230 */
231void delete_from_swap_cache(struct folio *folio)
232{
233 swp_entry_t entry = folio_swap_entry(folio);
234 struct address_space *address_space = swap_address_space(entry);
235
236 xa_lock_irq(&address_space->i_pages);
237 __delete_from_swap_cache(folio, entry, NULL);
238 xa_unlock_irq(&address_space->i_pages);
239
240 put_swap_folio(folio, entry);
241 folio_ref_sub(folio, folio_nr_pages(folio));
242}
243
244void clear_shadow_from_swap_cache(int type, unsigned long begin,
245 unsigned long end)
246{
247 unsigned long curr = begin;
248 void *old;
249
250 for (;;) {
251 swp_entry_t entry = swp_entry(type, curr);
252 struct address_space *address_space = swap_address_space(entry);
253 XA_STATE(xas, &address_space->i_pages, curr);
254
255 xa_lock_irq(&address_space->i_pages);
256 xas_for_each(&xas, old, end) {
257 if (!xa_is_value(old))
258 continue;
259 xas_store(&xas, NULL);
260 }
261 xa_unlock_irq(&address_space->i_pages);
262
263 /* search the next swapcache until we meet end */
264 curr >>= SWAP_ADDRESS_SPACE_SHIFT;
265 curr++;
266 curr <<= SWAP_ADDRESS_SPACE_SHIFT;
267 if (curr > end)
268 break;
269 }
270}
271
272/*
273 * If we are the only user, then try to free up the swap cache.
274 *
275 * Its ok to check the swapcache flag without the folio lock
276 * here because we are going to recheck again inside
277 * folio_free_swap() _with_ the lock.
278 * - Marcelo
279 */
280void free_swap_cache(struct page *page)
281{
282 struct folio *folio = page_folio(page);
283
284 if (folio_test_swapcache(folio) && !folio_mapped(folio) &&
285 folio_trylock(folio)) {
286 folio_free_swap(folio);
287 folio_unlock(folio);
288 }
289}
290
291/*
292 * Perform a free_page(), also freeing any swap cache associated with
293 * this page if it is the last user of the page.
294 */
295void free_page_and_swap_cache(struct page *page)
296{
297 free_swap_cache(page);
298 if (!is_huge_zero_page(page))
299 put_page(page);
300}
301
302/*
303 * Passed an array of pages, drop them all from swapcache and then release
304 * them. They are removed from the LRU and freed if this is their last use.
305 */
306void free_pages_and_swap_cache(struct encoded_page **pages, int nr)
307{
308 lru_add_drain();
309 for (int i = 0; i < nr; i++)
310 free_swap_cache(encoded_page_ptr(pages[i]));
311 release_pages(pages, nr);
312}
313
314static inline bool swap_use_vma_readahead(void)
315{
316 return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
317}
318
319/*
320 * Lookup a swap entry in the swap cache. A found folio will be returned
321 * unlocked and with its refcount incremented - we rely on the kernel
322 * lock getting page table operations atomic even if we drop the folio
323 * lock before returning.
324 */
325struct folio *swap_cache_get_folio(swp_entry_t entry,
326 struct vm_area_struct *vma, unsigned long addr)
327{
328 struct folio *folio;
329 struct swap_info_struct *si;
330
331 si = get_swap_device(entry);
332 if (!si)
333 return NULL;
334 folio = filemap_get_folio(swap_address_space(entry), swp_offset(entry));
335 put_swap_device(si);
336
337 if (folio) {
338 bool vma_ra = swap_use_vma_readahead();
339 bool readahead;
340
341 /*
342 * At the moment, we don't support PG_readahead for anon THP
343 * so let's bail out rather than confusing the readahead stat.
344 */
345 if (unlikely(folio_test_large(folio)))
346 return folio;
347
348 readahead = folio_test_clear_readahead(folio);
349 if (vma && vma_ra) {
350 unsigned long ra_val;
351 int win, hits;
352
353 ra_val = GET_SWAP_RA_VAL(vma);
354 win = SWAP_RA_WIN(ra_val);
355 hits = SWAP_RA_HITS(ra_val);
356 if (readahead)
357 hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
358 atomic_long_set(&vma->swap_readahead_info,
359 SWAP_RA_VAL(addr, win, hits));
360 }
361
362 if (readahead) {
363 count_vm_event(SWAP_RA_HIT);
364 if (!vma || !vma_ra)
365 atomic_inc(&swapin_readahead_hits);
366 }
367 }
368
369 return folio;
370}
371
372/**
373 * filemap_get_incore_folio - Find and get a folio from the page or swap caches.
374 * @mapping: The address_space to search.
375 * @index: The page cache index.
376 *
377 * This differs from filemap_get_folio() in that it will also look for the
378 * folio in the swap cache.
379 *
380 * Return: The found folio or %NULL.
381 */
382struct folio *filemap_get_incore_folio(struct address_space *mapping,
383 pgoff_t index)
384{
385 swp_entry_t swp;
386 struct swap_info_struct *si;
387 struct folio *folio = __filemap_get_folio(mapping, index, FGP_ENTRY, 0);
388
389 if (!xa_is_value(folio))
390 goto out;
391 if (!shmem_mapping(mapping))
392 return NULL;
393
394 swp = radix_to_swp_entry(folio);
395 /* There might be swapin error entries in shmem mapping. */
396 if (non_swap_entry(swp))
397 return NULL;
398 /* Prevent swapoff from happening to us */
399 si = get_swap_device(swp);
400 if (!si)
401 return NULL;
402 index = swp_offset(swp);
403 folio = filemap_get_folio(swap_address_space(swp), index);
404 put_swap_device(si);
405out:
406 return folio;
407}
408
409struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
410 struct vm_area_struct *vma, unsigned long addr,
411 bool *new_page_allocated)
412{
413 struct swap_info_struct *si;
414 struct folio *folio;
415 void *shadow = NULL;
416
417 *new_page_allocated = false;
418
419 for (;;) {
420 int err;
421 /*
422 * First check the swap cache. Since this is normally
423 * called after swap_cache_get_folio() failed, re-calling
424 * that would confuse statistics.
425 */
426 si = get_swap_device(entry);
427 if (!si)
428 return NULL;
429 folio = filemap_get_folio(swap_address_space(entry),
430 swp_offset(entry));
431 put_swap_device(si);
432 if (folio)
433 return folio_file_page(folio, swp_offset(entry));
434
435 /*
436 * Just skip read ahead for unused swap slot.
437 * During swap_off when swap_slot_cache is disabled,
438 * we have to handle the race between putting
439 * swap entry in swap cache and marking swap slot
440 * as SWAP_HAS_CACHE. That's done in later part of code or
441 * else swap_off will be aborted if we return NULL.
442 */
443 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
444 return NULL;
445
446 /*
447 * Get a new page to read into from swap. Allocate it now,
448 * before marking swap_map SWAP_HAS_CACHE, when -EEXIST will
449 * cause any racers to loop around until we add it to cache.
450 */
451 folio = vma_alloc_folio(gfp_mask, 0, vma, addr, false);
452 if (!folio)
453 return NULL;
454
455 /*
456 * Swap entry may have been freed since our caller observed it.
457 */
458 err = swapcache_prepare(entry);
459 if (!err)
460 break;
461
462 folio_put(folio);
463 if (err != -EEXIST)
464 return NULL;
465
466 /*
467 * We might race against __delete_from_swap_cache(), and
468 * stumble across a swap_map entry whose SWAP_HAS_CACHE
469 * has not yet been cleared. Or race against another
470 * __read_swap_cache_async(), which has set SWAP_HAS_CACHE
471 * in swap_map, but not yet added its page to swap cache.
472 */
473 schedule_timeout_uninterruptible(1);
474 }
475
476 /*
477 * The swap entry is ours to swap in. Prepare the new page.
478 */
479
480 __folio_set_locked(folio);
481 __folio_set_swapbacked(folio);
482
483 if (mem_cgroup_swapin_charge_folio(folio, NULL, gfp_mask, entry))
484 goto fail_unlock;
485
486 /* May fail (-ENOMEM) if XArray node allocation failed. */
487 if (add_to_swap_cache(folio, entry, gfp_mask & GFP_RECLAIM_MASK, &shadow))
488 goto fail_unlock;
489
490 mem_cgroup_swapin_uncharge_swap(entry);
491
492 if (shadow)
493 workingset_refault(folio, shadow);
494
495 /* Caller will initiate read into locked folio */
496 folio_add_lru(folio);
497 *new_page_allocated = true;
498 return &folio->page;
499
500fail_unlock:
501 put_swap_folio(folio, entry);
502 folio_unlock(folio);
503 folio_put(folio);
504 return NULL;
505}
506
507/*
508 * Locate a page of swap in physical memory, reserving swap cache space
509 * and reading the disk if it is not already cached.
510 * A failure return means that either the page allocation failed or that
511 * the swap entry is no longer in use.
512 */
513struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
514 struct vm_area_struct *vma,
515 unsigned long addr, bool do_poll,
516 struct swap_iocb **plug)
517{
518 bool page_was_allocated;
519 struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
520 vma, addr, &page_was_allocated);
521
522 if (page_was_allocated)
523 swap_readpage(retpage, do_poll, plug);
524
525 return retpage;
526}
527
528static unsigned int __swapin_nr_pages(unsigned long prev_offset,
529 unsigned long offset,
530 int hits,
531 int max_pages,
532 int prev_win)
533{
534 unsigned int pages, last_ra;
535
536 /*
537 * This heuristic has been found to work well on both sequential and
538 * random loads, swapping to hard disk or to SSD: please don't ask
539 * what the "+ 2" means, it just happens to work well, that's all.
540 */
541 pages = hits + 2;
542 if (pages == 2) {
543 /*
544 * We can have no readahead hits to judge by: but must not get
545 * stuck here forever, so check for an adjacent offset instead
546 * (and don't even bother to check whether swap type is same).
547 */
548 if (offset != prev_offset + 1 && offset != prev_offset - 1)
549 pages = 1;
550 } else {
551 unsigned int roundup = 4;
552 while (roundup < pages)
553 roundup <<= 1;
554 pages = roundup;
555 }
556
557 if (pages > max_pages)
558 pages = max_pages;
559
560 /* Don't shrink readahead too fast */
561 last_ra = prev_win / 2;
562 if (pages < last_ra)
563 pages = last_ra;
564
565 return pages;
566}
567
568static unsigned long swapin_nr_pages(unsigned long offset)
569{
570 static unsigned long prev_offset;
571 unsigned int hits, pages, max_pages;
572 static atomic_t last_readahead_pages;
573
574 max_pages = 1 << READ_ONCE(page_cluster);
575 if (max_pages <= 1)
576 return 1;
577
578 hits = atomic_xchg(&swapin_readahead_hits, 0);
579 pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits,
580 max_pages,
581 atomic_read(&last_readahead_pages));
582 if (!hits)
583 WRITE_ONCE(prev_offset, offset);
584 atomic_set(&last_readahead_pages, pages);
585
586 return pages;
587}
588
589/**
590 * swap_cluster_readahead - swap in pages in hope we need them soon
591 * @entry: swap entry of this memory
592 * @gfp_mask: memory allocation flags
593 * @vmf: fault information
594 *
595 * Returns the struct page for entry and addr, after queueing swapin.
596 *
597 * Primitive swap readahead code. We simply read an aligned block of
598 * (1 << page_cluster) entries in the swap area. This method is chosen
599 * because it doesn't cost us any seek time. We also make sure to queue
600 * the 'original' request together with the readahead ones...
601 *
602 * This has been extended to use the NUMA policies from the mm triggering
603 * the readahead.
604 *
605 * Caller must hold read mmap_lock if vmf->vma is not NULL.
606 */
607struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
608 struct vm_fault *vmf)
609{
610 struct page *page;
611 unsigned long entry_offset = swp_offset(entry);
612 unsigned long offset = entry_offset;
613 unsigned long start_offset, end_offset;
614 unsigned long mask;
615 struct swap_info_struct *si = swp_swap_info(entry);
616 struct blk_plug plug;
617 struct swap_iocb *splug = NULL;
618 bool do_poll = true, page_allocated;
619 struct vm_area_struct *vma = vmf->vma;
620 unsigned long addr = vmf->address;
621
622 mask = swapin_nr_pages(offset) - 1;
623 if (!mask)
624 goto skip;
625
626 do_poll = false;
627 /* Read a page_cluster sized and aligned cluster around offset. */
628 start_offset = offset & ~mask;
629 end_offset = offset | mask;
630 if (!start_offset) /* First page is swap header. */
631 start_offset++;
632 if (end_offset >= si->max)
633 end_offset = si->max - 1;
634
635 blk_start_plug(&plug);
636 for (offset = start_offset; offset <= end_offset ; offset++) {
637 /* Ok, do the async read-ahead now */
638 page = __read_swap_cache_async(
639 swp_entry(swp_type(entry), offset),
640 gfp_mask, vma, addr, &page_allocated);
641 if (!page)
642 continue;
643 if (page_allocated) {
644 swap_readpage(page, false, &splug);
645 if (offset != entry_offset) {
646 SetPageReadahead(page);
647 count_vm_event(SWAP_RA);
648 }
649 }
650 put_page(page);
651 }
652 blk_finish_plug(&plug);
653 swap_read_unplug(splug);
654
655 lru_add_drain(); /* Push any new pages onto the LRU now */
656skip:
657 /* The page was likely read above, so no need for plugging here */
658 return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll, NULL);
659}
660
661int init_swap_address_space(unsigned int type, unsigned long nr_pages)
662{
663 struct address_space *spaces, *space;
664 unsigned int i, nr;
665
666 nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
667 spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL);
668 if (!spaces)
669 return -ENOMEM;
670 for (i = 0; i < nr; i++) {
671 space = spaces + i;
672 xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ);
673 atomic_set(&space->i_mmap_writable, 0);
674 space->a_ops = &swap_aops;
675 /* swap cache doesn't use writeback related tags */
676 mapping_set_no_writeback_tags(space);
677 }
678 nr_swapper_spaces[type] = nr;
679 swapper_spaces[type] = spaces;
680
681 return 0;
682}
683
684void exit_swap_address_space(unsigned int type)
685{
686 int i;
687 struct address_space *spaces = swapper_spaces[type];
688
689 for (i = 0; i < nr_swapper_spaces[type]; i++)
690 VM_WARN_ON_ONCE(!mapping_empty(&spaces[i]));
691 kvfree(spaces);
692 nr_swapper_spaces[type] = 0;
693 swapper_spaces[type] = NULL;
694}
695
696static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
697 unsigned long faddr,
698 unsigned long lpfn,
699 unsigned long rpfn,
700 unsigned long *start,
701 unsigned long *end)
702{
703 *start = max3(lpfn, PFN_DOWN(vma->vm_start),
704 PFN_DOWN(faddr & PMD_MASK));
705 *end = min3(rpfn, PFN_DOWN(vma->vm_end),
706 PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
707}
708
709static void swap_ra_info(struct vm_fault *vmf,
710 struct vma_swap_readahead *ra_info)
711{
712 struct vm_area_struct *vma = vmf->vma;
713 unsigned long ra_val;
714 unsigned long faddr, pfn, fpfn;
715 unsigned long start, end;
716 pte_t *pte, *orig_pte;
717 unsigned int max_win, hits, prev_win, win, left;
718#ifndef CONFIG_64BIT
719 pte_t *tpte;
720#endif
721
722 max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
723 SWAP_RA_ORDER_CEILING);
724 if (max_win == 1) {
725 ra_info->win = 1;
726 return;
727 }
728
729 faddr = vmf->address;
730 orig_pte = pte = pte_offset_map(vmf->pmd, faddr);
731
732 fpfn = PFN_DOWN(faddr);
733 ra_val = GET_SWAP_RA_VAL(vma);
734 pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val));
735 prev_win = SWAP_RA_WIN(ra_val);
736 hits = SWAP_RA_HITS(ra_val);
737 ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits,
738 max_win, prev_win);
739 atomic_long_set(&vma->swap_readahead_info,
740 SWAP_RA_VAL(faddr, win, 0));
741
742 if (win == 1) {
743 pte_unmap(orig_pte);
744 return;
745 }
746
747 /* Copy the PTEs because the page table may be unmapped */
748 if (fpfn == pfn + 1)
749 swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end);
750 else if (pfn == fpfn + 1)
751 swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1,
752 &start, &end);
753 else {
754 left = (win - 1) / 2;
755 swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
756 &start, &end);
757 }
758 ra_info->nr_pte = end - start;
759 ra_info->offset = fpfn - start;
760 pte -= ra_info->offset;
761#ifdef CONFIG_64BIT
762 ra_info->ptes = pte;
763#else
764 tpte = ra_info->ptes;
765 for (pfn = start; pfn != end; pfn++)
766 *tpte++ = *pte++;
767#endif
768 pte_unmap(orig_pte);
769}
770
771/**
772 * swap_vma_readahead - swap in pages in hope we need them soon
773 * @fentry: swap entry of this memory
774 * @gfp_mask: memory allocation flags
775 * @vmf: fault information
776 *
777 * Returns the struct page for entry and addr, after queueing swapin.
778 *
779 * Primitive swap readahead code. We simply read in a few pages whose
780 * virtual addresses are around the fault address in the same vma.
781 *
782 * Caller must hold read mmap_lock if vmf->vma is not NULL.
783 *
784 */
785static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
786 struct vm_fault *vmf)
787{
788 struct blk_plug plug;
789 struct swap_iocb *splug = NULL;
790 struct vm_area_struct *vma = vmf->vma;
791 struct page *page;
792 pte_t *pte, pentry;
793 swp_entry_t entry;
794 unsigned int i;
795 bool page_allocated;
796 struct vma_swap_readahead ra_info = {
797 .win = 1,
798 };
799
800 swap_ra_info(vmf, &ra_info);
801 if (ra_info.win == 1)
802 goto skip;
803
804 blk_start_plug(&plug);
805 for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte;
806 i++, pte++) {
807 pentry = *pte;
808 if (!is_swap_pte(pentry))
809 continue;
810 entry = pte_to_swp_entry(pentry);
811 if (unlikely(non_swap_entry(entry)))
812 continue;
813 page = __read_swap_cache_async(entry, gfp_mask, vma,
814 vmf->address, &page_allocated);
815 if (!page)
816 continue;
817 if (page_allocated) {
818 swap_readpage(page, false, &splug);
819 if (i != ra_info.offset) {
820 SetPageReadahead(page);
821 count_vm_event(SWAP_RA);
822 }
823 }
824 put_page(page);
825 }
826 blk_finish_plug(&plug);
827 swap_read_unplug(splug);
828 lru_add_drain();
829skip:
830 /* The page was likely read above, so no need for plugging here */
831 return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
832 ra_info.win == 1, NULL);
833}
834
835/**
836 * swapin_readahead - swap in pages in hope we need them soon
837 * @entry: swap entry of this memory
838 * @gfp_mask: memory allocation flags
839 * @vmf: fault information
840 *
841 * Returns the struct page for entry and addr, after queueing swapin.
842 *
843 * It's a main entry function for swap readahead. By the configuration,
844 * it will read ahead blocks by cluster-based(ie, physical disk based)
845 * or vma-based(ie, virtual address based on faulty address) readahead.
846 */
847struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
848 struct vm_fault *vmf)
849{
850 return swap_use_vma_readahead() ?
851 swap_vma_readahead(entry, gfp_mask, vmf) :
852 swap_cluster_readahead(entry, gfp_mask, vmf);
853}
854
855#ifdef CONFIG_SYSFS
856static ssize_t vma_ra_enabled_show(struct kobject *kobj,
857 struct kobj_attribute *attr, char *buf)
858{
859 return sysfs_emit(buf, "%s\n",
860 enable_vma_readahead ? "true" : "false");
861}
862static ssize_t vma_ra_enabled_store(struct kobject *kobj,
863 struct kobj_attribute *attr,
864 const char *buf, size_t count)
865{
866 ssize_t ret;
867
868 ret = kstrtobool(buf, &enable_vma_readahead);
869 if (ret)
870 return ret;
871
872 return count;
873}
874static struct kobj_attribute vma_ra_enabled_attr = __ATTR_RW(vma_ra_enabled);
875
876static struct attribute *swap_attrs[] = {
877 &vma_ra_enabled_attr.attr,
878 NULL,
879};
880
881static const struct attribute_group swap_attr_group = {
882 .attrs = swap_attrs,
883};
884
885static int __init swap_init_sysfs(void)
886{
887 int err;
888 struct kobject *swap_kobj;
889
890 swap_kobj = kobject_create_and_add("swap", mm_kobj);
891 if (!swap_kobj) {
892 pr_err("failed to create swap kobject\n");
893 return -ENOMEM;
894 }
895 err = sysfs_create_group(swap_kobj, &swap_attr_group);
896 if (err) {
897 pr_err("failed to register swap group\n");
898 goto delete_obj;
899 }
900 return 0;
901
902delete_obj:
903 kobject_put(swap_kobj);
904 return err;
905}
906subsys_initcall(swap_init_sysfs);
907#endif