Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 LT |
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 | */ | |
1da177e4 | 10 | #include <linux/mm.h> |
5a0e3ad6 | 11 | #include <linux/gfp.h> |
1da177e4 LT |
12 | #include <linux/kernel_stat.h> |
13 | #include <linux/swap.h> | |
46017e95 | 14 | #include <linux/swapops.h> |
1da177e4 LT |
15 | #include <linux/init.h> |
16 | #include <linux/pagemap.h> | |
1da177e4 | 17 | #include <linux/backing-dev.h> |
3fb5c298 | 18 | #include <linux/blkdev.h> |
c484d410 | 19 | #include <linux/pagevec.h> |
b20a3503 | 20 | #include <linux/migrate.h> |
4b3ef9da | 21 | #include <linux/vmalloc.h> |
67afa38e | 22 | #include <linux/swap_slots.h> |
38d8b4e6 | 23 | #include <linux/huge_mm.h> |
61ef1865 | 24 | #include <linux/shmem_fs.h> |
243bce09 | 25 | #include "internal.h" |
014bb1de | 26 | #include "swap.h" |
1da177e4 LT |
27 | |
28 | /* | |
29 | * swapper_space is a fiction, retained to simplify the path through | |
7eaceacc | 30 | * vmscan's shrink_page_list. |
1da177e4 | 31 | */ |
f5e54d6e | 32 | static const struct address_space_operations swap_aops = { |
1da177e4 | 33 | .writepage = swap_writepage, |
4c4a7634 | 34 | .dirty_folio = noop_dirty_folio, |
1c93923c | 35 | #ifdef CONFIG_MIGRATION |
54184650 | 36 | .migrate_folio = migrate_folio, |
1c93923c | 37 | #endif |
1da177e4 LT |
38 | }; |
39 | ||
783cb68e CD |
40 | struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly; |
41 | static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly; | |
f5c754d6 | 42 | static bool enable_vma_readahead __read_mostly = true; |
ec560175 | 43 | |
ec560175 HY |
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) | |
1da177e4 | 61 | |
579f8290 SL |
62 | static atomic_t swapin_readahead_hits = ATOMIC_INIT(4); |
63 | ||
1da177e4 LT |
64 | void show_swap_cache_info(void) |
65 | { | |
33806f06 | 66 | printk("%lu pages in swap cache\n", total_swapcache_pages()); |
ec8acf20 SL |
67 | printk("Free swap = %ldkB\n", |
68 | get_nr_swap_pages() << (PAGE_SHIFT - 10)); | |
1da177e4 LT |
69 | printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10)); |
70 | } | |
71 | ||
aae466b0 JK |
72 | void *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 | ||
8c647dd1 | 78 | page = xa_load(&address_space->i_pages, idx); |
aae466b0 JK |
79 | if (xa_is_value(page)) |
80 | return page; | |
aae466b0 JK |
81 | return NULL; |
82 | } | |
83 | ||
1da177e4 | 84 | /* |
2bb876b5 | 85 | * add_to_swap_cache resembles filemap_add_folio on swapper_space, |
1da177e4 LT |
86 | * but sets SwapCache flag and private instead of mapping and index. |
87 | */ | |
a4c366f0 | 88 | int add_to_swap_cache(struct folio *folio, swp_entry_t entry, |
3852f676 | 89 | gfp_t gfp, void **shadowp) |
1da177e4 | 90 | { |
8d93b41c | 91 | struct address_space *address_space = swap_address_space(entry); |
38d8b4e6 | 92 | pgoff_t idx = swp_offset(entry); |
a4c366f0 MWO |
93 | XA_STATE_ORDER(xas, &address_space->i_pages, idx, folio_order(folio)); |
94 | unsigned long i, nr = folio_nr_pages(folio); | |
3852f676 | 95 | void *old; |
1da177e4 | 96 | |
5649d113 YY |
97 | xas_set_update(&xas, workingset_update_node); |
98 | ||
a4c366f0 MWO |
99 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
100 | VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); | |
101 | VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio); | |
51726b12 | 102 | |
a4c366f0 MWO |
103 | folio_ref_add(folio, nr); |
104 | folio_set_swapcache(folio); | |
31a56396 | 105 | |
8d93b41c MW |
106 | do { |
107 | xas_lock_irq(&xas); | |
108 | xas_create_range(&xas); | |
109 | if (xas_error(&xas)) | |
110 | goto unlock; | |
111 | for (i = 0; i < nr; i++) { | |
a4c366f0 | 112 | VM_BUG_ON_FOLIO(xas.xa_index != idx + i, folio); |
3852f676 JK |
113 | old = xas_load(&xas); |
114 | if (xa_is_value(old)) { | |
3852f676 JK |
115 | if (shadowp) |
116 | *shadowp = old; | |
117 | } | |
a4c366f0 MWO |
118 | set_page_private(folio_page(folio, i), entry.val + i); |
119 | xas_store(&xas, folio); | |
8d93b41c MW |
120 | xas_next(&xas); |
121 | } | |
38d8b4e6 | 122 | address_space->nrpages += nr; |
a4c366f0 MWO |
123 | __node_stat_mod_folio(folio, NR_FILE_PAGES, nr); |
124 | __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, nr); | |
8d93b41c MW |
125 | unlock: |
126 | xas_unlock_irq(&xas); | |
127 | } while (xas_nomem(&xas, gfp)); | |
31a56396 | 128 | |
8d93b41c MW |
129 | if (!xas_error(&xas)) |
130 | return 0; | |
31a56396 | 131 | |
a4c366f0 MWO |
132 | folio_clear_swapcache(folio); |
133 | folio_ref_sub(folio, nr); | |
8d93b41c | 134 | return xas_error(&xas); |
1da177e4 LT |
135 | } |
136 | ||
1da177e4 | 137 | /* |
ceff9d33 | 138 | * This must be called only on folios that have |
1da177e4 LT |
139 | * been verified to be in the swap cache. |
140 | */ | |
ceff9d33 | 141 | void __delete_from_swap_cache(struct folio *folio, |
3852f676 | 142 | swp_entry_t entry, void *shadow) |
1da177e4 | 143 | { |
4e17ec25 | 144 | struct address_space *address_space = swap_address_space(entry); |
ceff9d33 MWO |
145 | int i; |
146 | long nr = folio_nr_pages(folio); | |
4e17ec25 MW |
147 | pgoff_t idx = swp_offset(entry); |
148 | XA_STATE(xas, &address_space->i_pages, idx); | |
33806f06 | 149 | |
5649d113 YY |
150 | xas_set_update(&xas, workingset_update_node); |
151 | ||
ceff9d33 MWO |
152 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
153 | VM_BUG_ON_FOLIO(!folio_test_swapcache(folio), folio); | |
154 | VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio); | |
1da177e4 | 155 | |
38d8b4e6 | 156 | for (i = 0; i < nr; i++) { |
3852f676 | 157 | void *entry = xas_store(&xas, shadow); |
b9eb7776 | 158 | VM_BUG_ON_PAGE(entry != folio, entry); |
ceff9d33 | 159 | set_page_private(folio_page(folio, i), 0); |
4e17ec25 | 160 | xas_next(&xas); |
38d8b4e6 | 161 | } |
ceff9d33 | 162 | folio_clear_swapcache(folio); |
38d8b4e6 | 163 | address_space->nrpages -= nr; |
ceff9d33 MWO |
164 | __node_stat_mod_folio(folio, NR_FILE_PAGES, -nr); |
165 | __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, -nr); | |
1da177e4 LT |
166 | } |
167 | ||
168 | /** | |
09c02e56 MWO |
169 | * add_to_swap - allocate swap space for a folio |
170 | * @folio: folio we want to move to swap | |
1da177e4 | 171 | * |
09c02e56 MWO |
172 | * Allocate swap space for the folio and add the folio to the |
173 | * swap cache. | |
174 | * | |
175 | * Context: Caller needs to hold the folio lock. | |
176 | * Return: Whether the folio was added to the swap cache. | |
1da177e4 | 177 | */ |
09c02e56 | 178 | bool add_to_swap(struct folio *folio) |
1da177e4 LT |
179 | { |
180 | swp_entry_t entry; | |
1da177e4 LT |
181 | int err; |
182 | ||
09c02e56 MWO |
183 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
184 | VM_BUG_ON_FOLIO(!folio_test_uptodate(folio), folio); | |
1da177e4 | 185 | |
e2e3fdc7 | 186 | entry = folio_alloc_swap(folio); |
2ca4532a | 187 | if (!entry.val) |
09c02e56 | 188 | return false; |
0f074658 | 189 | |
2ca4532a | 190 | /* |
8d93b41c | 191 | * XArray node allocations from PF_MEMALLOC contexts could |
2ca4532a DN |
192 | * completely exhaust the page allocator. __GFP_NOMEMALLOC |
193 | * stops emergency reserves from being allocated. | |
194 | * | |
195 | * TODO: this could cause a theoretical memory reclaim | |
196 | * deadlock in the swap out path. | |
197 | */ | |
198 | /* | |
854e9ed0 | 199 | * Add it to the swap cache. |
2ca4532a | 200 | */ |
a4c366f0 | 201 | err = add_to_swap_cache(folio, entry, |
3852f676 | 202 | __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN, NULL); |
38d8b4e6 | 203 | if (err) |
bd53b714 | 204 | /* |
2ca4532a DN |
205 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely |
206 | * clear SWAP_HAS_CACHE flag. | |
1da177e4 | 207 | */ |
0f074658 | 208 | goto fail; |
9625456c | 209 | /* |
09c02e56 MWO |
210 | * Normally the folio will be dirtied in unmap because its |
211 | * pte should be dirty. A special case is MADV_FREE page. The | |
212 | * page's pte could have dirty bit cleared but the folio's | |
213 | * SwapBacked flag is still set because clearing the dirty bit | |
214 | * and SwapBacked flag has no lock protected. For such folio, | |
215 | * unmap will not set dirty bit for it, so folio reclaim will | |
216 | * not write the folio out. This can cause data corruption when | |
217 | * the folio is swapped in later. Always setting the dirty flag | |
218 | * for the folio solves the problem. | |
9625456c | 219 | */ |
09c02e56 | 220 | folio_mark_dirty(folio); |
38d8b4e6 | 221 | |
09c02e56 | 222 | return true; |
38d8b4e6 | 223 | |
38d8b4e6 | 224 | fail: |
4081f744 | 225 | put_swap_folio(folio, entry); |
09c02e56 | 226 | return false; |
1da177e4 LT |
227 | } |
228 | ||
229 | /* | |
75fa68a5 | 230 | * This must be called only on folios that have |
1da177e4 | 231 | * been verified to be in the swap cache and locked. |
75fa68a5 MWO |
232 | * It will never put the folio into the free list, |
233 | * the caller has a reference on the folio. | |
1da177e4 | 234 | */ |
75fa68a5 | 235 | void delete_from_swap_cache(struct folio *folio) |
1da177e4 | 236 | { |
75fa68a5 | 237 | swp_entry_t entry = folio_swap_entry(folio); |
4e17ec25 | 238 | struct address_space *address_space = swap_address_space(entry); |
1da177e4 | 239 | |
b93b0163 | 240 | xa_lock_irq(&address_space->i_pages); |
ceff9d33 | 241 | __delete_from_swap_cache(folio, entry, NULL); |
b93b0163 | 242 | xa_unlock_irq(&address_space->i_pages); |
1da177e4 | 243 | |
4081f744 | 244 | put_swap_folio(folio, entry); |
75fa68a5 | 245 | folio_ref_sub(folio, folio_nr_pages(folio)); |
1da177e4 LT |
246 | } |
247 | ||
3852f676 JK |
248 | void clear_shadow_from_swap_cache(int type, unsigned long begin, |
249 | unsigned long end) | |
250 | { | |
251 | unsigned long curr = begin; | |
252 | void *old; | |
253 | ||
254 | for (;;) { | |
3852f676 JK |
255 | swp_entry_t entry = swp_entry(type, curr); |
256 | struct address_space *address_space = swap_address_space(entry); | |
257 | XA_STATE(xas, &address_space->i_pages, curr); | |
258 | ||
5649d113 YY |
259 | xas_set_update(&xas, workingset_update_node); |
260 | ||
3852f676 JK |
261 | xa_lock_irq(&address_space->i_pages); |
262 | xas_for_each(&xas, old, end) { | |
263 | if (!xa_is_value(old)) | |
264 | continue; | |
265 | xas_store(&xas, NULL); | |
3852f676 | 266 | } |
3852f676 JK |
267 | xa_unlock_irq(&address_space->i_pages); |
268 | ||
269 | /* search the next swapcache until we meet end */ | |
270 | curr >>= SWAP_ADDRESS_SPACE_SHIFT; | |
271 | curr++; | |
272 | curr <<= SWAP_ADDRESS_SPACE_SHIFT; | |
273 | if (curr > end) | |
274 | break; | |
275 | } | |
276 | } | |
277 | ||
1da177e4 LT |
278 | /* |
279 | * If we are the only user, then try to free up the swap cache. | |
280 | * | |
aedd74d4 | 281 | * Its ok to check the swapcache flag without the folio lock |
a2c43eed | 282 | * here because we are going to recheck again inside |
aedd74d4 | 283 | * folio_free_swap() _with_ the lock. |
1da177e4 LT |
284 | * - Marcelo |
285 | */ | |
f4c4a3f4 | 286 | void free_swap_cache(struct page *page) |
1da177e4 | 287 | { |
aedd74d4 MWO |
288 | struct folio *folio = page_folio(page); |
289 | ||
290 | if (folio_test_swapcache(folio) && !folio_mapped(folio) && | |
291 | folio_trylock(folio)) { | |
292 | folio_free_swap(folio); | |
293 | folio_unlock(folio); | |
1da177e4 LT |
294 | } |
295 | } | |
296 | ||
297 | /* | |
298 | * Perform a free_page(), also freeing any swap cache associated with | |
b8072f09 | 299 | * this page if it is the last user of the page. |
1da177e4 LT |
300 | */ |
301 | void free_page_and_swap_cache(struct page *page) | |
302 | { | |
303 | free_swap_cache(page); | |
6fcb52a5 | 304 | if (!is_huge_zero_page(page)) |
770a5370 | 305 | put_page(page); |
1da177e4 LT |
306 | } |
307 | ||
308 | /* | |
309 | * Passed an array of pages, drop them all from swapcache and then release | |
310 | * them. They are removed from the LRU and freed if this is their last use. | |
311 | */ | |
7cc8f9c7 | 312 | void free_pages_and_swap_cache(struct encoded_page **pages, int nr) |
1da177e4 | 313 | { |
1da177e4 | 314 | lru_add_drain(); |
7cc8f9c7 LT |
315 | for (int i = 0; i < nr; i++) |
316 | free_swap_cache(encoded_page_ptr(pages[i])); | |
317 | release_pages(pages, nr); | |
1da177e4 LT |
318 | } |
319 | ||
e9e9b7ec MK |
320 | static inline bool swap_use_vma_readahead(void) |
321 | { | |
322 | return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap); | |
323 | } | |
324 | ||
1da177e4 | 325 | /* |
c9edc242 | 326 | * Lookup a swap entry in the swap cache. A found folio will be returned |
1da177e4 | 327 | * unlocked and with its refcount incremented - we rely on the kernel |
c9edc242 | 328 | * lock getting page table operations atomic even if we drop the folio |
1da177e4 | 329 | * lock before returning. |
cbc2bd98 KS |
330 | * |
331 | * Caller must lock the swap device or hold a reference to keep it valid. | |
1da177e4 | 332 | */ |
c9edc242 MWO |
333 | struct folio *swap_cache_get_folio(swp_entry_t entry, |
334 | struct vm_area_struct *vma, unsigned long addr) | |
1da177e4 | 335 | { |
c9edc242 | 336 | struct folio *folio; |
1da177e4 | 337 | |
c9edc242 | 338 | folio = filemap_get_folio(swap_address_space(entry), swp_offset(entry)); |
c9edc242 | 339 | if (folio) { |
eaf649eb MK |
340 | bool vma_ra = swap_use_vma_readahead(); |
341 | bool readahead; | |
342 | ||
eaf649eb MK |
343 | /* |
344 | * At the moment, we don't support PG_readahead for anon THP | |
345 | * so let's bail out rather than confusing the readahead stat. | |
346 | */ | |
c9edc242 MWO |
347 | if (unlikely(folio_test_large(folio))) |
348 | return folio; | |
eaf649eb | 349 | |
c9edc242 | 350 | readahead = folio_test_clear_readahead(folio); |
eaf649eb MK |
351 | if (vma && vma_ra) { |
352 | unsigned long ra_val; | |
353 | int win, hits; | |
354 | ||
355 | ra_val = GET_SWAP_RA_VAL(vma); | |
356 | win = SWAP_RA_WIN(ra_val); | |
357 | hits = SWAP_RA_HITS(ra_val); | |
ec560175 HY |
358 | if (readahead) |
359 | hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX); | |
360 | atomic_long_set(&vma->swap_readahead_info, | |
361 | SWAP_RA_VAL(addr, win, hits)); | |
362 | } | |
eaf649eb | 363 | |
ec560175 | 364 | if (readahead) { |
cbc65df2 | 365 | count_vm_event(SWAP_RA_HIT); |
eaf649eb | 366 | if (!vma || !vma_ra) |
ec560175 | 367 | atomic_inc(&swapin_readahead_hits); |
cbc65df2 | 368 | } |
579f8290 | 369 | } |
eaf649eb | 370 | |
c9edc242 MWO |
371 | return folio; |
372 | } | |
373 | ||
61ef1865 | 374 | /** |
524984ff | 375 | * filemap_get_incore_folio - Find and get a folio from the page or swap caches. |
61ef1865 MWO |
376 | * @mapping: The address_space to search. |
377 | * @index: The page cache index. | |
378 | * | |
524984ff MWO |
379 | * This differs from filemap_get_folio() in that it will also look for the |
380 | * folio in the swap cache. | |
61ef1865 | 381 | * |
524984ff | 382 | * Return: The found folio or %NULL. |
61ef1865 | 383 | */ |
524984ff MWO |
384 | struct folio *filemap_get_incore_folio(struct address_space *mapping, |
385 | pgoff_t index) | |
61ef1865 MWO |
386 | { |
387 | swp_entry_t swp; | |
388 | struct swap_info_struct *si; | |
dd8095b1 | 389 | struct folio *folio = __filemap_get_folio(mapping, index, FGP_ENTRY, 0); |
61ef1865 | 390 | |
dd8095b1 MWO |
391 | if (!xa_is_value(folio)) |
392 | goto out; | |
61ef1865 MWO |
393 | if (!shmem_mapping(mapping)) |
394 | return NULL; | |
395 | ||
dd8095b1 | 396 | swp = radix_to_swp_entry(folio); |
ba6851b4 ML |
397 | /* There might be swapin error entries in shmem mapping. */ |
398 | if (non_swap_entry(swp)) | |
399 | return NULL; | |
61ef1865 MWO |
400 | /* Prevent swapoff from happening to us */ |
401 | si = get_swap_device(swp); | |
402 | if (!si) | |
403 | return NULL; | |
dd8095b1 MWO |
404 | index = swp_offset(swp); |
405 | folio = filemap_get_folio(swap_address_space(swp), index); | |
61ef1865 | 406 | put_swap_device(si); |
dd8095b1 | 407 | out: |
524984ff | 408 | return folio; |
61ef1865 MWO |
409 | } |
410 | ||
5b999aad DS |
411 | struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, |
412 | struct vm_area_struct *vma, unsigned long addr, | |
413 | bool *new_page_allocated) | |
1da177e4 | 414 | { |
eb085574 | 415 | struct swap_info_struct *si; |
a0d3374b | 416 | struct folio *folio; |
aae466b0 | 417 | void *shadow = NULL; |
4c6355b2 | 418 | |
5b999aad | 419 | *new_page_allocated = false; |
1da177e4 | 420 | |
4c6355b2 JW |
421 | for (;;) { |
422 | int err; | |
1da177e4 LT |
423 | /* |
424 | * First check the swap cache. Since this is normally | |
cb691e2f | 425 | * called after swap_cache_get_folio() failed, re-calling |
1da177e4 LT |
426 | * that would confuse statistics. |
427 | */ | |
eb085574 HY |
428 | si = get_swap_device(entry); |
429 | if (!si) | |
4c6355b2 | 430 | return NULL; |
a0d3374b MWO |
431 | folio = filemap_get_folio(swap_address_space(entry), |
432 | swp_offset(entry)); | |
eb085574 | 433 | put_swap_device(si); |
a0d3374b MWO |
434 | if (folio) |
435 | return folio_file_page(folio, swp_offset(entry)); | |
1da177e4 | 436 | |
ba81f838 HY |
437 | /* |
438 | * Just skip read ahead for unused swap slot. | |
439 | * During swap_off when swap_slot_cache is disabled, | |
440 | * we have to handle the race between putting | |
441 | * swap entry in swap cache and marking swap slot | |
442 | * as SWAP_HAS_CACHE. That's done in later part of code or | |
443 | * else swap_off will be aborted if we return NULL. | |
444 | */ | |
445 | if (!__swp_swapcount(entry) && swap_slot_cache_enabled) | |
4c6355b2 | 446 | return NULL; |
e8c26ab6 | 447 | |
1da177e4 | 448 | /* |
4c6355b2 JW |
449 | * Get a new page to read into from swap. Allocate it now, |
450 | * before marking swap_map SWAP_HAS_CACHE, when -EEXIST will | |
451 | * cause any racers to loop around until we add it to cache. | |
1da177e4 | 452 | */ |
a0d3374b MWO |
453 | folio = vma_alloc_folio(gfp_mask, 0, vma, addr, false); |
454 | if (!folio) | |
4c6355b2 | 455 | return NULL; |
1da177e4 | 456 | |
f000944d HD |
457 | /* |
458 | * Swap entry may have been freed since our caller observed it. | |
459 | */ | |
355cfa73 | 460 | err = swapcache_prepare(entry); |
4c6355b2 | 461 | if (!err) |
f000944d HD |
462 | break; |
463 | ||
a0d3374b | 464 | folio_put(folio); |
4c6355b2 JW |
465 | if (err != -EEXIST) |
466 | return NULL; | |
467 | ||
2ca4532a | 468 | /* |
4c6355b2 JW |
469 | * We might race against __delete_from_swap_cache(), and |
470 | * stumble across a swap_map entry whose SWAP_HAS_CACHE | |
471 | * has not yet been cleared. Or race against another | |
472 | * __read_swap_cache_async(), which has set SWAP_HAS_CACHE | |
473 | * in swap_map, but not yet added its page to swap cache. | |
2ca4532a | 474 | */ |
029c4628 | 475 | schedule_timeout_uninterruptible(1); |
4c6355b2 JW |
476 | } |
477 | ||
478 | /* | |
479 | * The swap entry is ours to swap in. Prepare the new page. | |
480 | */ | |
481 | ||
a0d3374b MWO |
482 | __folio_set_locked(folio); |
483 | __folio_set_swapbacked(folio); | |
4c6355b2 | 484 | |
65995918 | 485 | if (mem_cgroup_swapin_charge_folio(folio, NULL, gfp_mask, entry)) |
4c6355b2 | 486 | goto fail_unlock; |
4c6355b2 | 487 | |
0add0c77 | 488 | /* May fail (-ENOMEM) if XArray node allocation failed. */ |
a4c366f0 | 489 | if (add_to_swap_cache(folio, entry, gfp_mask & GFP_RECLAIM_MASK, &shadow)) |
4c6355b2 | 490 | goto fail_unlock; |
0add0c77 SB |
491 | |
492 | mem_cgroup_swapin_uncharge_swap(entry); | |
4c6355b2 | 493 | |
aae466b0 | 494 | if (shadow) |
a0d3374b | 495 | workingset_refault(folio, shadow); |
314b57fb | 496 | |
a0d3374b MWO |
497 | /* Caller will initiate read into locked folio */ |
498 | folio_add_lru(folio); | |
4c6355b2 | 499 | *new_page_allocated = true; |
a0d3374b | 500 | return &folio->page; |
1da177e4 | 501 | |
4c6355b2 | 502 | fail_unlock: |
4081f744 | 503 | put_swap_folio(folio, entry); |
a0d3374b MWO |
504 | folio_unlock(folio); |
505 | folio_put(folio); | |
4c6355b2 | 506 | return NULL; |
1da177e4 | 507 | } |
46017e95 | 508 | |
5b999aad DS |
509 | /* |
510 | * Locate a page of swap in physical memory, reserving swap cache space | |
511 | * and reading the disk if it is not already cached. | |
512 | * A failure return means that either the page allocation failed or that | |
513 | * the swap entry is no longer in use. | |
514 | */ | |
515 | struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, | |
5169b844 N |
516 | struct vm_area_struct *vma, |
517 | unsigned long addr, bool do_poll, | |
518 | struct swap_iocb **plug) | |
5b999aad DS |
519 | { |
520 | bool page_was_allocated; | |
521 | struct page *retpage = __read_swap_cache_async(entry, gfp_mask, | |
522 | vma, addr, &page_was_allocated); | |
523 | ||
524 | if (page_was_allocated) | |
5169b844 | 525 | swap_readpage(retpage, do_poll, plug); |
5b999aad DS |
526 | |
527 | return retpage; | |
528 | } | |
529 | ||
ec560175 HY |
530 | static unsigned int __swapin_nr_pages(unsigned long prev_offset, |
531 | unsigned long offset, | |
532 | int hits, | |
533 | int max_pages, | |
534 | int prev_win) | |
579f8290 | 535 | { |
ec560175 | 536 | unsigned int pages, last_ra; |
579f8290 SL |
537 | |
538 | /* | |
539 | * This heuristic has been found to work well on both sequential and | |
540 | * random loads, swapping to hard disk or to SSD: please don't ask | |
541 | * what the "+ 2" means, it just happens to work well, that's all. | |
542 | */ | |
ec560175 | 543 | pages = hits + 2; |
579f8290 SL |
544 | if (pages == 2) { |
545 | /* | |
546 | * We can have no readahead hits to judge by: but must not get | |
547 | * stuck here forever, so check for an adjacent offset instead | |
548 | * (and don't even bother to check whether swap type is same). | |
549 | */ | |
550 | if (offset != prev_offset + 1 && offset != prev_offset - 1) | |
551 | pages = 1; | |
579f8290 SL |
552 | } else { |
553 | unsigned int roundup = 4; | |
554 | while (roundup < pages) | |
555 | roundup <<= 1; | |
556 | pages = roundup; | |
557 | } | |
558 | ||
559 | if (pages > max_pages) | |
560 | pages = max_pages; | |
561 | ||
562 | /* Don't shrink readahead too fast */ | |
ec560175 | 563 | last_ra = prev_win / 2; |
579f8290 SL |
564 | if (pages < last_ra) |
565 | pages = last_ra; | |
ec560175 HY |
566 | |
567 | return pages; | |
568 | } | |
569 | ||
570 | static unsigned long swapin_nr_pages(unsigned long offset) | |
571 | { | |
572 | static unsigned long prev_offset; | |
573 | unsigned int hits, pages, max_pages; | |
574 | static atomic_t last_readahead_pages; | |
575 | ||
576 | max_pages = 1 << READ_ONCE(page_cluster); | |
577 | if (max_pages <= 1) | |
578 | return 1; | |
579 | ||
580 | hits = atomic_xchg(&swapin_readahead_hits, 0); | |
d6c1f098 QC |
581 | pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits, |
582 | max_pages, | |
ec560175 HY |
583 | atomic_read(&last_readahead_pages)); |
584 | if (!hits) | |
d6c1f098 | 585 | WRITE_ONCE(prev_offset, offset); |
579f8290 SL |
586 | atomic_set(&last_readahead_pages, pages); |
587 | ||
588 | return pages; | |
589 | } | |
590 | ||
46017e95 | 591 | /** |
e9e9b7ec | 592 | * swap_cluster_readahead - swap in pages in hope we need them soon |
46017e95 | 593 | * @entry: swap entry of this memory |
7682486b | 594 | * @gfp_mask: memory allocation flags |
e9e9b7ec | 595 | * @vmf: fault information |
46017e95 HD |
596 | * |
597 | * Returns the struct page for entry and addr, after queueing swapin. | |
598 | * | |
599 | * Primitive swap readahead code. We simply read an aligned block of | |
600 | * (1 << page_cluster) entries in the swap area. This method is chosen | |
601 | * because it doesn't cost us any seek time. We also make sure to queue | |
602 | * the 'original' request together with the readahead ones... | |
603 | * | |
604 | * This has been extended to use the NUMA policies from the mm triggering | |
605 | * the readahead. | |
606 | * | |
c1e8d7c6 | 607 | * Caller must hold read mmap_lock if vmf->vma is not NULL. |
46017e95 | 608 | */ |
e9e9b7ec MK |
609 | struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask, |
610 | struct vm_fault *vmf) | |
46017e95 | 611 | { |
46017e95 | 612 | struct page *page; |
579f8290 SL |
613 | unsigned long entry_offset = swp_offset(entry); |
614 | unsigned long offset = entry_offset; | |
67f96aa2 | 615 | unsigned long start_offset, end_offset; |
579f8290 | 616 | unsigned long mask; |
e9a6effa | 617 | struct swap_info_struct *si = swp_swap_info(entry); |
3fb5c298 | 618 | struct blk_plug plug; |
5169b844 | 619 | struct swap_iocb *splug = NULL; |
c4fa6309 | 620 | bool do_poll = true, page_allocated; |
e9e9b7ec MK |
621 | struct vm_area_struct *vma = vmf->vma; |
622 | unsigned long addr = vmf->address; | |
46017e95 | 623 | |
579f8290 SL |
624 | mask = swapin_nr_pages(offset) - 1; |
625 | if (!mask) | |
626 | goto skip; | |
627 | ||
23955622 | 628 | do_poll = false; |
67f96aa2 RR |
629 | /* Read a page_cluster sized and aligned cluster around offset. */ |
630 | start_offset = offset & ~mask; | |
631 | end_offset = offset | mask; | |
632 | if (!start_offset) /* First page is swap header. */ | |
633 | start_offset++; | |
e9a6effa HY |
634 | if (end_offset >= si->max) |
635 | end_offset = si->max - 1; | |
67f96aa2 | 636 | |
3fb5c298 | 637 | blk_start_plug(&plug); |
67f96aa2 | 638 | for (offset = start_offset; offset <= end_offset ; offset++) { |
46017e95 | 639 | /* Ok, do the async read-ahead now */ |
c4fa6309 HY |
640 | page = __read_swap_cache_async( |
641 | swp_entry(swp_type(entry), offset), | |
642 | gfp_mask, vma, addr, &page_allocated); | |
46017e95 | 643 | if (!page) |
67f96aa2 | 644 | continue; |
c4fa6309 | 645 | if (page_allocated) { |
5169b844 | 646 | swap_readpage(page, false, &splug); |
eaf649eb | 647 | if (offset != entry_offset) { |
c4fa6309 HY |
648 | SetPageReadahead(page); |
649 | count_vm_event(SWAP_RA); | |
650 | } | |
cbc65df2 | 651 | } |
09cbfeaf | 652 | put_page(page); |
46017e95 | 653 | } |
3fb5c298 | 654 | blk_finish_plug(&plug); |
5169b844 | 655 | swap_read_unplug(splug); |
3fb5c298 | 656 | |
46017e95 | 657 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
579f8290 | 658 | skip: |
5169b844 N |
659 | /* The page was likely read above, so no need for plugging here */ |
660 | return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll, NULL); | |
46017e95 | 661 | } |
4b3ef9da HY |
662 | |
663 | int init_swap_address_space(unsigned int type, unsigned long nr_pages) | |
664 | { | |
665 | struct address_space *spaces, *space; | |
666 | unsigned int i, nr; | |
667 | ||
668 | nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES); | |
778e1cdd | 669 | spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL); |
4b3ef9da HY |
670 | if (!spaces) |
671 | return -ENOMEM; | |
672 | for (i = 0; i < nr; i++) { | |
673 | space = spaces + i; | |
a2833486 | 674 | xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ); |
4b3ef9da HY |
675 | atomic_set(&space->i_mmap_writable, 0); |
676 | space->a_ops = &swap_aops; | |
677 | /* swap cache doesn't use writeback related tags */ | |
678 | mapping_set_no_writeback_tags(space); | |
4b3ef9da HY |
679 | } |
680 | nr_swapper_spaces[type] = nr; | |
054f1d1f | 681 | swapper_spaces[type] = spaces; |
4b3ef9da HY |
682 | |
683 | return 0; | |
684 | } | |
685 | ||
686 | void exit_swap_address_space(unsigned int type) | |
687 | { | |
eea4a501 HY |
688 | int i; |
689 | struct address_space *spaces = swapper_spaces[type]; | |
690 | ||
691 | for (i = 0; i < nr_swapper_spaces[type]; i++) | |
692 | VM_WARN_ON_ONCE(!mapping_empty(&spaces[i])); | |
693 | kvfree(spaces); | |
4b3ef9da | 694 | nr_swapper_spaces[type] = 0; |
054f1d1f | 695 | swapper_spaces[type] = NULL; |
4b3ef9da | 696 | } |
ec560175 | 697 | |
eaf649eb | 698 | static void swap_ra_info(struct vm_fault *vmf, |
16ba391e | 699 | struct vma_swap_readahead *ra_info) |
ec560175 HY |
700 | { |
701 | struct vm_area_struct *vma = vmf->vma; | |
eaf649eb | 702 | unsigned long ra_val; |
16ba391e | 703 | unsigned long faddr, pfn, fpfn, lpfn, rpfn; |
ec560175 | 704 | unsigned long start, end; |
eaf649eb | 705 | pte_t *pte, *orig_pte; |
16ba391e | 706 | unsigned int max_win, hits, prev_win, win; |
ec560175 HY |
707 | #ifndef CONFIG_64BIT |
708 | pte_t *tpte; | |
709 | #endif | |
710 | ||
61b63972 HY |
711 | max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster), |
712 | SWAP_RA_ORDER_CEILING); | |
713 | if (max_win == 1) { | |
eaf649eb MK |
714 | ra_info->win = 1; |
715 | return; | |
61b63972 HY |
716 | } |
717 | ||
ec560175 | 718 | faddr = vmf->address; |
ec560175 | 719 | fpfn = PFN_DOWN(faddr); |
eaf649eb MK |
720 | ra_val = GET_SWAP_RA_VAL(vma); |
721 | pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val)); | |
722 | prev_win = SWAP_RA_WIN(ra_val); | |
723 | hits = SWAP_RA_HITS(ra_val); | |
724 | ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits, | |
ec560175 HY |
725 | max_win, prev_win); |
726 | atomic_long_set(&vma->swap_readahead_info, | |
727 | SWAP_RA_VAL(faddr, win, 0)); | |
728 | ||
18ad72f5 | 729 | if (win == 1) |
eaf649eb | 730 | return; |
ec560175 HY |
731 | |
732 | /* Copy the PTEs because the page table may be unmapped */ | |
18ad72f5 | 733 | orig_pte = pte = pte_offset_map(vmf->pmd, faddr); |
16ba391e KS |
734 | if (fpfn == pfn + 1) { |
735 | lpfn = fpfn; | |
736 | rpfn = fpfn + win; | |
737 | } else if (pfn == fpfn + 1) { | |
738 | lpfn = fpfn - win + 1; | |
739 | rpfn = fpfn + 1; | |
740 | } else { | |
741 | unsigned int left = (win - 1) / 2; | |
742 | ||
743 | lpfn = fpfn - left; | |
744 | rpfn = fpfn + win - left; | |
ec560175 | 745 | } |
16ba391e KS |
746 | start = max3(lpfn, PFN_DOWN(vma->vm_start), |
747 | PFN_DOWN(faddr & PMD_MASK)); | |
748 | end = min3(rpfn, PFN_DOWN(vma->vm_end), | |
749 | PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE)); | |
750 | ||
eaf649eb MK |
751 | ra_info->nr_pte = end - start; |
752 | ra_info->offset = fpfn - start; | |
753 | pte -= ra_info->offset; | |
ec560175 | 754 | #ifdef CONFIG_64BIT |
eaf649eb | 755 | ra_info->ptes = pte; |
ec560175 | 756 | #else |
eaf649eb | 757 | tpte = ra_info->ptes; |
ec560175 HY |
758 | for (pfn = start; pfn != end; pfn++) |
759 | *tpte++ = *pte++; | |
760 | #endif | |
eaf649eb | 761 | pte_unmap(orig_pte); |
ec560175 HY |
762 | } |
763 | ||
e9f59873 YS |
764 | /** |
765 | * swap_vma_readahead - swap in pages in hope we need them soon | |
27ec4878 | 766 | * @fentry: swap entry of this memory |
e9f59873 YS |
767 | * @gfp_mask: memory allocation flags |
768 | * @vmf: fault information | |
769 | * | |
770 | * Returns the struct page for entry and addr, after queueing swapin. | |
771 | * | |
cb152a1a | 772 | * Primitive swap readahead code. We simply read in a few pages whose |
e9f59873 YS |
773 | * virtual addresses are around the fault address in the same vma. |
774 | * | |
c1e8d7c6 | 775 | * Caller must hold read mmap_lock if vmf->vma is not NULL. |
e9f59873 YS |
776 | * |
777 | */ | |
f5c754d6 CIK |
778 | static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask, |
779 | struct vm_fault *vmf) | |
ec560175 HY |
780 | { |
781 | struct blk_plug plug; | |
5169b844 | 782 | struct swap_iocb *splug = NULL; |
ec560175 HY |
783 | struct vm_area_struct *vma = vmf->vma; |
784 | struct page *page; | |
785 | pte_t *pte, pentry; | |
786 | swp_entry_t entry; | |
787 | unsigned int i; | |
788 | bool page_allocated; | |
e97af699 ML |
789 | struct vma_swap_readahead ra_info = { |
790 | .win = 1, | |
791 | }; | |
ec560175 | 792 | |
eaf649eb MK |
793 | swap_ra_info(vmf, &ra_info); |
794 | if (ra_info.win == 1) | |
ec560175 HY |
795 | goto skip; |
796 | ||
797 | blk_start_plug(&plug); | |
eaf649eb | 798 | for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte; |
ec560175 HY |
799 | i++, pte++) { |
800 | pentry = *pte; | |
92bafb20 | 801 | if (!is_swap_pte(pentry)) |
ec560175 HY |
802 | continue; |
803 | entry = pte_to_swp_entry(pentry); | |
804 | if (unlikely(non_swap_entry(entry))) | |
805 | continue; | |
806 | page = __read_swap_cache_async(entry, gfp_mask, vma, | |
807 | vmf->address, &page_allocated); | |
808 | if (!page) | |
809 | continue; | |
810 | if (page_allocated) { | |
5169b844 | 811 | swap_readpage(page, false, &splug); |
eaf649eb | 812 | if (i != ra_info.offset) { |
ec560175 HY |
813 | SetPageReadahead(page); |
814 | count_vm_event(SWAP_RA); | |
815 | } | |
816 | } | |
817 | put_page(page); | |
818 | } | |
819 | blk_finish_plug(&plug); | |
5169b844 | 820 | swap_read_unplug(splug); |
ec560175 HY |
821 | lru_add_drain(); |
822 | skip: | |
5169b844 | 823 | /* The page was likely read above, so no need for plugging here */ |
ec560175 | 824 | return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, |
5169b844 | 825 | ra_info.win == 1, NULL); |
ec560175 | 826 | } |
d9bfcfdc | 827 | |
e9e9b7ec MK |
828 | /** |
829 | * swapin_readahead - swap in pages in hope we need them soon | |
830 | * @entry: swap entry of this memory | |
831 | * @gfp_mask: memory allocation flags | |
832 | * @vmf: fault information | |
833 | * | |
834 | * Returns the struct page for entry and addr, after queueing swapin. | |
835 | * | |
836 | * It's a main entry function for swap readahead. By the configuration, | |
837 | * it will read ahead blocks by cluster-based(ie, physical disk based) | |
838 | * or vma-based(ie, virtual address based on faulty address) readahead. | |
839 | */ | |
840 | struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, | |
841 | struct vm_fault *vmf) | |
842 | { | |
843 | return swap_use_vma_readahead() ? | |
844 | swap_vma_readahead(entry, gfp_mask, vmf) : | |
845 | swap_cluster_readahead(entry, gfp_mask, vmf); | |
846 | } | |
847 | ||
d9bfcfdc HY |
848 | #ifdef CONFIG_SYSFS |
849 | static ssize_t vma_ra_enabled_show(struct kobject *kobj, | |
850 | struct kobj_attribute *attr, char *buf) | |
851 | { | |
ae7a927d JP |
852 | return sysfs_emit(buf, "%s\n", |
853 | enable_vma_readahead ? "true" : "false"); | |
d9bfcfdc HY |
854 | } |
855 | static ssize_t vma_ra_enabled_store(struct kobject *kobj, | |
856 | struct kobj_attribute *attr, | |
857 | const char *buf, size_t count) | |
858 | { | |
717aeab4 JG |
859 | ssize_t ret; |
860 | ||
861 | ret = kstrtobool(buf, &enable_vma_readahead); | |
862 | if (ret) | |
863 | return ret; | |
d9bfcfdc HY |
864 | |
865 | return count; | |
866 | } | |
6106b93e | 867 | static struct kobj_attribute vma_ra_enabled_attr = __ATTR_RW(vma_ra_enabled); |
d9bfcfdc | 868 | |
d9bfcfdc HY |
869 | static struct attribute *swap_attrs[] = { |
870 | &vma_ra_enabled_attr.attr, | |
d9bfcfdc HY |
871 | NULL, |
872 | }; | |
873 | ||
e48333b6 | 874 | static const struct attribute_group swap_attr_group = { |
d9bfcfdc HY |
875 | .attrs = swap_attrs, |
876 | }; | |
877 | ||
878 | static int __init swap_init_sysfs(void) | |
879 | { | |
880 | int err; | |
881 | struct kobject *swap_kobj; | |
882 | ||
883 | swap_kobj = kobject_create_and_add("swap", mm_kobj); | |
884 | if (!swap_kobj) { | |
885 | pr_err("failed to create swap kobject\n"); | |
886 | return -ENOMEM; | |
887 | } | |
888 | err = sysfs_create_group(swap_kobj, &swap_attr_group); | |
889 | if (err) { | |
890 | pr_err("failed to register swap group\n"); | |
891 | goto delete_obj; | |
892 | } | |
893 | return 0; | |
894 | ||
895 | delete_obj: | |
896 | kobject_put(swap_kobj); | |
897 | return err; | |
898 | } | |
899 | subsys_initcall(swap_init_sysfs); | |
900 | #endif |