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 | |
a4c366f0 MWO |
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); | |
51726b12 | 100 | |
a4c366f0 MWO |
101 | folio_ref_add(folio, nr); |
102 | folio_set_swapcache(folio); | |
31a56396 | 103 | |
8d93b41c MW |
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++) { | |
a4c366f0 | 110 | VM_BUG_ON_FOLIO(xas.xa_index != idx + i, folio); |
3852f676 JK |
111 | old = xas_load(&xas); |
112 | if (xa_is_value(old)) { | |
3852f676 JK |
113 | if (shadowp) |
114 | *shadowp = old; | |
115 | } | |
a4c366f0 MWO |
116 | set_page_private(folio_page(folio, i), entry.val + i); |
117 | xas_store(&xas, folio); | |
8d93b41c MW |
118 | xas_next(&xas); |
119 | } | |
38d8b4e6 | 120 | address_space->nrpages += nr; |
a4c366f0 MWO |
121 | __node_stat_mod_folio(folio, NR_FILE_PAGES, nr); |
122 | __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, nr); | |
8d93b41c MW |
123 | unlock: |
124 | xas_unlock_irq(&xas); | |
125 | } while (xas_nomem(&xas, gfp)); | |
31a56396 | 126 | |
8d93b41c MW |
127 | if (!xas_error(&xas)) |
128 | return 0; | |
31a56396 | 129 | |
a4c366f0 MWO |
130 | folio_clear_swapcache(folio); |
131 | folio_ref_sub(folio, nr); | |
8d93b41c | 132 | return xas_error(&xas); |
1da177e4 LT |
133 | } |
134 | ||
1da177e4 | 135 | /* |
ceff9d33 | 136 | * This must be called only on folios that have |
1da177e4 LT |
137 | * been verified to be in the swap cache. |
138 | */ | |
ceff9d33 | 139 | void __delete_from_swap_cache(struct folio *folio, |
3852f676 | 140 | swp_entry_t entry, void *shadow) |
1da177e4 | 141 | { |
4e17ec25 | 142 | struct address_space *address_space = swap_address_space(entry); |
ceff9d33 MWO |
143 | int i; |
144 | long nr = folio_nr_pages(folio); | |
4e17ec25 MW |
145 | pgoff_t idx = swp_offset(entry); |
146 | XA_STATE(xas, &address_space->i_pages, idx); | |
33806f06 | 147 | |
ceff9d33 MWO |
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); | |
1da177e4 | 151 | |
38d8b4e6 | 152 | for (i = 0; i < nr; i++) { |
3852f676 | 153 | void *entry = xas_store(&xas, shadow); |
b9eb7776 | 154 | VM_BUG_ON_PAGE(entry != folio, entry); |
ceff9d33 | 155 | set_page_private(folio_page(folio, i), 0); |
4e17ec25 | 156 | xas_next(&xas); |
38d8b4e6 | 157 | } |
ceff9d33 | 158 | folio_clear_swapcache(folio); |
38d8b4e6 | 159 | address_space->nrpages -= nr; |
ceff9d33 MWO |
160 | __node_stat_mod_folio(folio, NR_FILE_PAGES, -nr); |
161 | __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, -nr); | |
1da177e4 LT |
162 | } |
163 | ||
164 | /** | |
09c02e56 MWO |
165 | * add_to_swap - allocate swap space for a folio |
166 | * @folio: folio we want to move to swap | |
1da177e4 | 167 | * |
09c02e56 MWO |
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. | |
1da177e4 | 173 | */ |
09c02e56 | 174 | bool add_to_swap(struct folio *folio) |
1da177e4 LT |
175 | { |
176 | swp_entry_t entry; | |
1da177e4 LT |
177 | int err; |
178 | ||
09c02e56 MWO |
179 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
180 | VM_BUG_ON_FOLIO(!folio_test_uptodate(folio), folio); | |
1da177e4 | 181 | |
e2e3fdc7 | 182 | entry = folio_alloc_swap(folio); |
2ca4532a | 183 | if (!entry.val) |
09c02e56 | 184 | return false; |
0f074658 | 185 | |
2ca4532a | 186 | /* |
8d93b41c | 187 | * XArray node allocations from PF_MEMALLOC contexts could |
2ca4532a DN |
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 | /* | |
854e9ed0 | 195 | * Add it to the swap cache. |
2ca4532a | 196 | */ |
a4c366f0 | 197 | err = add_to_swap_cache(folio, entry, |
3852f676 | 198 | __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN, NULL); |
38d8b4e6 | 199 | if (err) |
bd53b714 | 200 | /* |
2ca4532a DN |
201 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely |
202 | * clear SWAP_HAS_CACHE flag. | |
1da177e4 | 203 | */ |
0f074658 | 204 | goto fail; |
9625456c | 205 | /* |
09c02e56 MWO |
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. | |
9625456c | 215 | */ |
09c02e56 | 216 | folio_mark_dirty(folio); |
38d8b4e6 | 217 | |
09c02e56 | 218 | return true; |
38d8b4e6 | 219 | |
38d8b4e6 | 220 | fail: |
4081f744 | 221 | put_swap_folio(folio, entry); |
09c02e56 | 222 | return false; |
1da177e4 LT |
223 | } |
224 | ||
225 | /* | |
75fa68a5 | 226 | * This must be called only on folios that have |
1da177e4 | 227 | * been verified to be in the swap cache and locked. |
75fa68a5 MWO |
228 | * It will never put the folio into the free list, |
229 | * the caller has a reference on the folio. | |
1da177e4 | 230 | */ |
75fa68a5 | 231 | void delete_from_swap_cache(struct folio *folio) |
1da177e4 | 232 | { |
75fa68a5 | 233 | swp_entry_t entry = folio_swap_entry(folio); |
4e17ec25 | 234 | struct address_space *address_space = swap_address_space(entry); |
1da177e4 | 235 | |
b93b0163 | 236 | xa_lock_irq(&address_space->i_pages); |
ceff9d33 | 237 | __delete_from_swap_cache(folio, entry, NULL); |
b93b0163 | 238 | xa_unlock_irq(&address_space->i_pages); |
1da177e4 | 239 | |
4081f744 | 240 | put_swap_folio(folio, entry); |
75fa68a5 | 241 | folio_ref_sub(folio, folio_nr_pages(folio)); |
1da177e4 LT |
242 | } |
243 | ||
3852f676 JK |
244 | void 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 (;;) { | |
3852f676 JK |
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); | |
3852f676 | 260 | } |
3852f676 JK |
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 | ||
1da177e4 LT |
272 | /* |
273 | * If we are the only user, then try to free up the swap cache. | |
274 | * | |
aedd74d4 | 275 | * Its ok to check the swapcache flag without the folio lock |
a2c43eed | 276 | * here because we are going to recheck again inside |
aedd74d4 | 277 | * folio_free_swap() _with_ the lock. |
1da177e4 LT |
278 | * - Marcelo |
279 | */ | |
f4c4a3f4 | 280 | void free_swap_cache(struct page *page) |
1da177e4 | 281 | { |
aedd74d4 MWO |
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); | |
1da177e4 LT |
288 | } |
289 | } | |
290 | ||
291 | /* | |
292 | * Perform a free_page(), also freeing any swap cache associated with | |
b8072f09 | 293 | * this page if it is the last user of the page. |
1da177e4 LT |
294 | */ |
295 | void free_page_and_swap_cache(struct page *page) | |
296 | { | |
297 | free_swap_cache(page); | |
6fcb52a5 | 298 | if (!is_huge_zero_page(page)) |
770a5370 | 299 | put_page(page); |
1da177e4 LT |
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 | */ | |
306 | void free_pages_and_swap_cache(struct page **pages, int nr) | |
307 | { | |
1da177e4 | 308 | struct page **pagep = pages; |
aabfb572 | 309 | int i; |
1da177e4 LT |
310 | |
311 | lru_add_drain(); | |
aabfb572 MH |
312 | for (i = 0; i < nr; i++) |
313 | free_swap_cache(pagep[i]); | |
c6f92f9f | 314 | release_pages(pagep, nr); |
1da177e4 LT |
315 | } |
316 | ||
e9e9b7ec MK |
317 | static inline bool swap_use_vma_readahead(void) |
318 | { | |
319 | return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap); | |
320 | } | |
321 | ||
1da177e4 | 322 | /* |
c9edc242 | 323 | * Lookup a swap entry in the swap cache. A found folio will be returned |
1da177e4 | 324 | * unlocked and with its refcount incremented - we rely on the kernel |
c9edc242 | 325 | * lock getting page table operations atomic even if we drop the folio |
1da177e4 LT |
326 | * lock before returning. |
327 | */ | |
c9edc242 MWO |
328 | struct folio *swap_cache_get_folio(swp_entry_t entry, |
329 | struct vm_area_struct *vma, unsigned long addr) | |
1da177e4 | 330 | { |
c9edc242 | 331 | struct folio *folio; |
eb085574 | 332 | struct swap_info_struct *si; |
1da177e4 | 333 | |
eb085574 HY |
334 | si = get_swap_device(entry); |
335 | if (!si) | |
336 | return NULL; | |
c9edc242 | 337 | folio = filemap_get_folio(swap_address_space(entry), swp_offset(entry)); |
eb085574 | 338 | put_swap_device(si); |
1da177e4 | 339 | |
c9edc242 | 340 | if (folio) { |
eaf649eb MK |
341 | bool vma_ra = swap_use_vma_readahead(); |
342 | bool readahead; | |
343 | ||
eaf649eb MK |
344 | /* |
345 | * At the moment, we don't support PG_readahead for anon THP | |
346 | * so let's bail out rather than confusing the readahead stat. | |
347 | */ | |
c9edc242 MWO |
348 | if (unlikely(folio_test_large(folio))) |
349 | return folio; | |
eaf649eb | 350 | |
c9edc242 | 351 | readahead = folio_test_clear_readahead(folio); |
eaf649eb MK |
352 | if (vma && vma_ra) { |
353 | unsigned long ra_val; | |
354 | int win, hits; | |
355 | ||
356 | ra_val = GET_SWAP_RA_VAL(vma); | |
357 | win = SWAP_RA_WIN(ra_val); | |
358 | hits = SWAP_RA_HITS(ra_val); | |
ec560175 HY |
359 | if (readahead) |
360 | hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX); | |
361 | atomic_long_set(&vma->swap_readahead_info, | |
362 | SWAP_RA_VAL(addr, win, hits)); | |
363 | } | |
eaf649eb | 364 | |
ec560175 | 365 | if (readahead) { |
cbc65df2 | 366 | count_vm_event(SWAP_RA_HIT); |
eaf649eb | 367 | if (!vma || !vma_ra) |
ec560175 | 368 | atomic_inc(&swapin_readahead_hits); |
cbc65df2 | 369 | } |
579f8290 | 370 | } |
eaf649eb | 371 | |
c9edc242 MWO |
372 | return folio; |
373 | } | |
374 | ||
61ef1865 MWO |
375 | /** |
376 | * find_get_incore_page - Find and get a page from the page or swap caches. | |
377 | * @mapping: The address_space to search. | |
378 | * @index: The page cache index. | |
379 | * | |
380 | * This differs from find_get_page() in that it will also look for the | |
381 | * page in the swap cache. | |
382 | * | |
383 | * Return: The found page or %NULL. | |
384 | */ | |
385 | struct page *find_get_incore_page(struct address_space *mapping, pgoff_t index) | |
386 | { | |
387 | swp_entry_t swp; | |
388 | struct swap_info_struct *si; | |
44835d20 MWO |
389 | struct page *page = pagecache_get_page(mapping, index, |
390 | FGP_ENTRY | FGP_HEAD, 0); | |
61ef1865 | 391 | |
a6de4b48 | 392 | if (!page) |
61ef1865 | 393 | return page; |
a6de4b48 MWO |
394 | if (!xa_is_value(page)) |
395 | return find_subpage(page, index); | |
61ef1865 MWO |
396 | if (!shmem_mapping(mapping)) |
397 | return NULL; | |
398 | ||
399 | swp = radix_to_swp_entry(page); | |
ba6851b4 ML |
400 | /* There might be swapin error entries in shmem mapping. */ |
401 | if (non_swap_entry(swp)) | |
402 | return NULL; | |
61ef1865 MWO |
403 | /* Prevent swapoff from happening to us */ |
404 | si = get_swap_device(swp); | |
405 | if (!si) | |
406 | return NULL; | |
407 | page = find_get_page(swap_address_space(swp), swp_offset(swp)); | |
408 | put_swap_device(si); | |
409 | return page; | |
410 | } | |
411 | ||
5b999aad DS |
412 | struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, |
413 | struct vm_area_struct *vma, unsigned long addr, | |
414 | bool *new_page_allocated) | |
1da177e4 | 415 | { |
eb085574 | 416 | struct swap_info_struct *si; |
a0d3374b | 417 | struct folio *folio; |
aae466b0 | 418 | void *shadow = NULL; |
4c6355b2 | 419 | |
5b999aad | 420 | *new_page_allocated = false; |
1da177e4 | 421 | |
4c6355b2 JW |
422 | for (;;) { |
423 | int err; | |
1da177e4 LT |
424 | /* |
425 | * First check the swap cache. Since this is normally | |
cb691e2f | 426 | * called after swap_cache_get_folio() failed, re-calling |
1da177e4 LT |
427 | * that would confuse statistics. |
428 | */ | |
eb085574 HY |
429 | si = get_swap_device(entry); |
430 | if (!si) | |
4c6355b2 | 431 | return NULL; |
a0d3374b MWO |
432 | folio = filemap_get_folio(swap_address_space(entry), |
433 | swp_offset(entry)); | |
eb085574 | 434 | put_swap_device(si); |
a0d3374b MWO |
435 | if (folio) |
436 | return folio_file_page(folio, swp_offset(entry)); | |
1da177e4 | 437 | |
ba81f838 HY |
438 | /* |
439 | * Just skip read ahead for unused swap slot. | |
440 | * During swap_off when swap_slot_cache is disabled, | |
441 | * we have to handle the race between putting | |
442 | * swap entry in swap cache and marking swap slot | |
443 | * as SWAP_HAS_CACHE. That's done in later part of code or | |
444 | * else swap_off will be aborted if we return NULL. | |
445 | */ | |
446 | if (!__swp_swapcount(entry) && swap_slot_cache_enabled) | |
4c6355b2 | 447 | return NULL; |
e8c26ab6 | 448 | |
1da177e4 | 449 | /* |
4c6355b2 JW |
450 | * Get a new page to read into from swap. Allocate it now, |
451 | * before marking swap_map SWAP_HAS_CACHE, when -EEXIST will | |
452 | * cause any racers to loop around until we add it to cache. | |
1da177e4 | 453 | */ |
a0d3374b MWO |
454 | folio = vma_alloc_folio(gfp_mask, 0, vma, addr, false); |
455 | if (!folio) | |
4c6355b2 | 456 | return NULL; |
1da177e4 | 457 | |
f000944d HD |
458 | /* |
459 | * Swap entry may have been freed since our caller observed it. | |
460 | */ | |
355cfa73 | 461 | err = swapcache_prepare(entry); |
4c6355b2 | 462 | if (!err) |
f000944d HD |
463 | break; |
464 | ||
a0d3374b | 465 | folio_put(folio); |
4c6355b2 JW |
466 | if (err != -EEXIST) |
467 | return NULL; | |
468 | ||
2ca4532a | 469 | /* |
4c6355b2 JW |
470 | * We might race against __delete_from_swap_cache(), and |
471 | * stumble across a swap_map entry whose SWAP_HAS_CACHE | |
472 | * has not yet been cleared. Or race against another | |
473 | * __read_swap_cache_async(), which has set SWAP_HAS_CACHE | |
474 | * in swap_map, but not yet added its page to swap cache. | |
2ca4532a | 475 | */ |
029c4628 | 476 | schedule_timeout_uninterruptible(1); |
4c6355b2 JW |
477 | } |
478 | ||
479 | /* | |
480 | * The swap entry is ours to swap in. Prepare the new page. | |
481 | */ | |
482 | ||
a0d3374b MWO |
483 | __folio_set_locked(folio); |
484 | __folio_set_swapbacked(folio); | |
4c6355b2 | 485 | |
65995918 | 486 | if (mem_cgroup_swapin_charge_folio(folio, NULL, gfp_mask, entry)) |
4c6355b2 | 487 | goto fail_unlock; |
4c6355b2 | 488 | |
0add0c77 | 489 | /* May fail (-ENOMEM) if XArray node allocation failed. */ |
a4c366f0 | 490 | if (add_to_swap_cache(folio, entry, gfp_mask & GFP_RECLAIM_MASK, &shadow)) |
4c6355b2 | 491 | goto fail_unlock; |
0add0c77 SB |
492 | |
493 | mem_cgroup_swapin_uncharge_swap(entry); | |
4c6355b2 | 494 | |
aae466b0 | 495 | if (shadow) |
a0d3374b | 496 | workingset_refault(folio, shadow); |
314b57fb | 497 | |
a0d3374b MWO |
498 | /* Caller will initiate read into locked folio */ |
499 | folio_add_lru(folio); | |
4c6355b2 | 500 | *new_page_allocated = true; |
a0d3374b | 501 | return &folio->page; |
1da177e4 | 502 | |
4c6355b2 | 503 | fail_unlock: |
4081f744 | 504 | put_swap_folio(folio, entry); |
a0d3374b MWO |
505 | folio_unlock(folio); |
506 | folio_put(folio); | |
4c6355b2 | 507 | return NULL; |
1da177e4 | 508 | } |
46017e95 | 509 | |
5b999aad DS |
510 | /* |
511 | * Locate a page of swap in physical memory, reserving swap cache space | |
512 | * and reading the disk if it is not already cached. | |
513 | * A failure return means that either the page allocation failed or that | |
514 | * the swap entry is no longer in use. | |
515 | */ | |
516 | struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, | |
5169b844 N |
517 | struct vm_area_struct *vma, |
518 | unsigned long addr, bool do_poll, | |
519 | struct swap_iocb **plug) | |
5b999aad DS |
520 | { |
521 | bool page_was_allocated; | |
522 | struct page *retpage = __read_swap_cache_async(entry, gfp_mask, | |
523 | vma, addr, &page_was_allocated); | |
524 | ||
525 | if (page_was_allocated) | |
5169b844 | 526 | swap_readpage(retpage, do_poll, plug); |
5b999aad DS |
527 | |
528 | return retpage; | |
529 | } | |
530 | ||
ec560175 HY |
531 | static unsigned int __swapin_nr_pages(unsigned long prev_offset, |
532 | unsigned long offset, | |
533 | int hits, | |
534 | int max_pages, | |
535 | int prev_win) | |
579f8290 | 536 | { |
ec560175 | 537 | unsigned int pages, last_ra; |
579f8290 SL |
538 | |
539 | /* | |
540 | * This heuristic has been found to work well on both sequential and | |
541 | * random loads, swapping to hard disk or to SSD: please don't ask | |
542 | * what the "+ 2" means, it just happens to work well, that's all. | |
543 | */ | |
ec560175 | 544 | pages = hits + 2; |
579f8290 SL |
545 | if (pages == 2) { |
546 | /* | |
547 | * We can have no readahead hits to judge by: but must not get | |
548 | * stuck here forever, so check for an adjacent offset instead | |
549 | * (and don't even bother to check whether swap type is same). | |
550 | */ | |
551 | if (offset != prev_offset + 1 && offset != prev_offset - 1) | |
552 | pages = 1; | |
579f8290 SL |
553 | } else { |
554 | unsigned int roundup = 4; | |
555 | while (roundup < pages) | |
556 | roundup <<= 1; | |
557 | pages = roundup; | |
558 | } | |
559 | ||
560 | if (pages > max_pages) | |
561 | pages = max_pages; | |
562 | ||
563 | /* Don't shrink readahead too fast */ | |
ec560175 | 564 | last_ra = prev_win / 2; |
579f8290 SL |
565 | if (pages < last_ra) |
566 | pages = last_ra; | |
ec560175 HY |
567 | |
568 | return pages; | |
569 | } | |
570 | ||
571 | static unsigned long swapin_nr_pages(unsigned long offset) | |
572 | { | |
573 | static unsigned long prev_offset; | |
574 | unsigned int hits, pages, max_pages; | |
575 | static atomic_t last_readahead_pages; | |
576 | ||
577 | max_pages = 1 << READ_ONCE(page_cluster); | |
578 | if (max_pages <= 1) | |
579 | return 1; | |
580 | ||
581 | hits = atomic_xchg(&swapin_readahead_hits, 0); | |
d6c1f098 QC |
582 | pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits, |
583 | max_pages, | |
ec560175 HY |
584 | atomic_read(&last_readahead_pages)); |
585 | if (!hits) | |
d6c1f098 | 586 | WRITE_ONCE(prev_offset, offset); |
579f8290 SL |
587 | atomic_set(&last_readahead_pages, pages); |
588 | ||
589 | return pages; | |
590 | } | |
591 | ||
46017e95 | 592 | /** |
e9e9b7ec | 593 | * swap_cluster_readahead - swap in pages in hope we need them soon |
46017e95 | 594 | * @entry: swap entry of this memory |
7682486b | 595 | * @gfp_mask: memory allocation flags |
e9e9b7ec | 596 | * @vmf: fault information |
46017e95 HD |
597 | * |
598 | * Returns the struct page for entry and addr, after queueing swapin. | |
599 | * | |
600 | * Primitive swap readahead code. We simply read an aligned block of | |
601 | * (1 << page_cluster) entries in the swap area. This method is chosen | |
602 | * because it doesn't cost us any seek time. We also make sure to queue | |
603 | * the 'original' request together with the readahead ones... | |
604 | * | |
605 | * This has been extended to use the NUMA policies from the mm triggering | |
606 | * the readahead. | |
607 | * | |
c1e8d7c6 | 608 | * Caller must hold read mmap_lock if vmf->vma is not NULL. |
46017e95 | 609 | */ |
e9e9b7ec MK |
610 | struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask, |
611 | struct vm_fault *vmf) | |
46017e95 | 612 | { |
46017e95 | 613 | struct page *page; |
579f8290 SL |
614 | unsigned long entry_offset = swp_offset(entry); |
615 | unsigned long offset = entry_offset; | |
67f96aa2 | 616 | unsigned long start_offset, end_offset; |
579f8290 | 617 | unsigned long mask; |
e9a6effa | 618 | struct swap_info_struct *si = swp_swap_info(entry); |
3fb5c298 | 619 | struct blk_plug plug; |
5169b844 | 620 | struct swap_iocb *splug = NULL; |
c4fa6309 | 621 | bool do_poll = true, page_allocated; |
e9e9b7ec MK |
622 | struct vm_area_struct *vma = vmf->vma; |
623 | unsigned long addr = vmf->address; | |
46017e95 | 624 | |
579f8290 SL |
625 | mask = swapin_nr_pages(offset) - 1; |
626 | if (!mask) | |
627 | goto skip; | |
628 | ||
23955622 | 629 | do_poll = false; |
67f96aa2 RR |
630 | /* Read a page_cluster sized and aligned cluster around offset. */ |
631 | start_offset = offset & ~mask; | |
632 | end_offset = offset | mask; | |
633 | if (!start_offset) /* First page is swap header. */ | |
634 | start_offset++; | |
e9a6effa HY |
635 | if (end_offset >= si->max) |
636 | end_offset = si->max - 1; | |
67f96aa2 | 637 | |
3fb5c298 | 638 | blk_start_plug(&plug); |
67f96aa2 | 639 | for (offset = start_offset; offset <= end_offset ; offset++) { |
46017e95 | 640 | /* Ok, do the async read-ahead now */ |
c4fa6309 HY |
641 | page = __read_swap_cache_async( |
642 | swp_entry(swp_type(entry), offset), | |
643 | gfp_mask, vma, addr, &page_allocated); | |
46017e95 | 644 | if (!page) |
67f96aa2 | 645 | continue; |
c4fa6309 | 646 | if (page_allocated) { |
5169b844 | 647 | swap_readpage(page, false, &splug); |
eaf649eb | 648 | if (offset != entry_offset) { |
c4fa6309 HY |
649 | SetPageReadahead(page); |
650 | count_vm_event(SWAP_RA); | |
651 | } | |
cbc65df2 | 652 | } |
09cbfeaf | 653 | put_page(page); |
46017e95 | 654 | } |
3fb5c298 | 655 | blk_finish_plug(&plug); |
5169b844 | 656 | swap_read_unplug(splug); |
3fb5c298 | 657 | |
46017e95 | 658 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
579f8290 | 659 | skip: |
5169b844 N |
660 | /* The page was likely read above, so no need for plugging here */ |
661 | return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll, NULL); | |
46017e95 | 662 | } |
4b3ef9da HY |
663 | |
664 | int init_swap_address_space(unsigned int type, unsigned long nr_pages) | |
665 | { | |
666 | struct address_space *spaces, *space; | |
667 | unsigned int i, nr; | |
668 | ||
669 | nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES); | |
778e1cdd | 670 | spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL); |
4b3ef9da HY |
671 | if (!spaces) |
672 | return -ENOMEM; | |
673 | for (i = 0; i < nr; i++) { | |
674 | space = spaces + i; | |
a2833486 | 675 | xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ); |
4b3ef9da HY |
676 | atomic_set(&space->i_mmap_writable, 0); |
677 | space->a_ops = &swap_aops; | |
678 | /* swap cache doesn't use writeback related tags */ | |
679 | mapping_set_no_writeback_tags(space); | |
4b3ef9da HY |
680 | } |
681 | nr_swapper_spaces[type] = nr; | |
054f1d1f | 682 | swapper_spaces[type] = spaces; |
4b3ef9da HY |
683 | |
684 | return 0; | |
685 | } | |
686 | ||
687 | void exit_swap_address_space(unsigned int type) | |
688 | { | |
eea4a501 HY |
689 | int i; |
690 | struct address_space *spaces = swapper_spaces[type]; | |
691 | ||
692 | for (i = 0; i < nr_swapper_spaces[type]; i++) | |
693 | VM_WARN_ON_ONCE(!mapping_empty(&spaces[i])); | |
694 | kvfree(spaces); | |
4b3ef9da | 695 | nr_swapper_spaces[type] = 0; |
054f1d1f | 696 | swapper_spaces[type] = NULL; |
4b3ef9da | 697 | } |
ec560175 HY |
698 | |
699 | static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma, | |
700 | unsigned long faddr, | |
701 | unsigned long lpfn, | |
702 | unsigned long rpfn, | |
703 | unsigned long *start, | |
704 | unsigned long *end) | |
705 | { | |
706 | *start = max3(lpfn, PFN_DOWN(vma->vm_start), | |
707 | PFN_DOWN(faddr & PMD_MASK)); | |
708 | *end = min3(rpfn, PFN_DOWN(vma->vm_end), | |
709 | PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE)); | |
710 | } | |
711 | ||
eaf649eb MK |
712 | static void swap_ra_info(struct vm_fault *vmf, |
713 | struct vma_swap_readahead *ra_info) | |
ec560175 HY |
714 | { |
715 | struct vm_area_struct *vma = vmf->vma; | |
eaf649eb | 716 | unsigned long ra_val; |
ec560175 HY |
717 | unsigned long faddr, pfn, fpfn; |
718 | unsigned long start, end; | |
eaf649eb | 719 | pte_t *pte, *orig_pte; |
ec560175 HY |
720 | unsigned int max_win, hits, prev_win, win, left; |
721 | #ifndef CONFIG_64BIT | |
722 | pte_t *tpte; | |
723 | #endif | |
724 | ||
61b63972 HY |
725 | max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster), |
726 | SWAP_RA_ORDER_CEILING); | |
727 | if (max_win == 1) { | |
eaf649eb MK |
728 | ra_info->win = 1; |
729 | return; | |
61b63972 HY |
730 | } |
731 | ||
ec560175 | 732 | faddr = vmf->address; |
eaf649eb | 733 | orig_pte = pte = pte_offset_map(vmf->pmd, faddr); |
ec560175 | 734 | |
ec560175 | 735 | fpfn = PFN_DOWN(faddr); |
eaf649eb MK |
736 | ra_val = GET_SWAP_RA_VAL(vma); |
737 | pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val)); | |
738 | prev_win = SWAP_RA_WIN(ra_val); | |
739 | hits = SWAP_RA_HITS(ra_val); | |
740 | ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits, | |
ec560175 HY |
741 | max_win, prev_win); |
742 | atomic_long_set(&vma->swap_readahead_info, | |
743 | SWAP_RA_VAL(faddr, win, 0)); | |
744 | ||
eaf649eb MK |
745 | if (win == 1) { |
746 | pte_unmap(orig_pte); | |
747 | return; | |
748 | } | |
ec560175 HY |
749 | |
750 | /* Copy the PTEs because the page table may be unmapped */ | |
751 | if (fpfn == pfn + 1) | |
752 | swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end); | |
753 | else if (pfn == fpfn + 1) | |
754 | swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1, | |
755 | &start, &end); | |
756 | else { | |
757 | left = (win - 1) / 2; | |
758 | swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left, | |
759 | &start, &end); | |
760 | } | |
eaf649eb MK |
761 | ra_info->nr_pte = end - start; |
762 | ra_info->offset = fpfn - start; | |
763 | pte -= ra_info->offset; | |
ec560175 | 764 | #ifdef CONFIG_64BIT |
eaf649eb | 765 | ra_info->ptes = pte; |
ec560175 | 766 | #else |
eaf649eb | 767 | tpte = ra_info->ptes; |
ec560175 HY |
768 | for (pfn = start; pfn != end; pfn++) |
769 | *tpte++ = *pte++; | |
770 | #endif | |
eaf649eb | 771 | pte_unmap(orig_pte); |
ec560175 HY |
772 | } |
773 | ||
e9f59873 YS |
774 | /** |
775 | * swap_vma_readahead - swap in pages in hope we need them soon | |
27ec4878 | 776 | * @fentry: swap entry of this memory |
e9f59873 YS |
777 | * @gfp_mask: memory allocation flags |
778 | * @vmf: fault information | |
779 | * | |
780 | * Returns the struct page for entry and addr, after queueing swapin. | |
781 | * | |
cb152a1a | 782 | * Primitive swap readahead code. We simply read in a few pages whose |
e9f59873 YS |
783 | * virtual addresses are around the fault address in the same vma. |
784 | * | |
c1e8d7c6 | 785 | * Caller must hold read mmap_lock if vmf->vma is not NULL. |
e9f59873 YS |
786 | * |
787 | */ | |
f5c754d6 CIK |
788 | static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask, |
789 | struct vm_fault *vmf) | |
ec560175 HY |
790 | { |
791 | struct blk_plug plug; | |
5169b844 | 792 | struct swap_iocb *splug = NULL; |
ec560175 HY |
793 | struct vm_area_struct *vma = vmf->vma; |
794 | struct page *page; | |
795 | pte_t *pte, pentry; | |
796 | swp_entry_t entry; | |
797 | unsigned int i; | |
798 | bool page_allocated; | |
e97af699 ML |
799 | struct vma_swap_readahead ra_info = { |
800 | .win = 1, | |
801 | }; | |
ec560175 | 802 | |
eaf649eb MK |
803 | swap_ra_info(vmf, &ra_info); |
804 | if (ra_info.win == 1) | |
ec560175 HY |
805 | goto skip; |
806 | ||
807 | blk_start_plug(&plug); | |
eaf649eb | 808 | for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte; |
ec560175 HY |
809 | i++, pte++) { |
810 | pentry = *pte; | |
92bafb20 | 811 | if (!is_swap_pte(pentry)) |
ec560175 HY |
812 | continue; |
813 | entry = pte_to_swp_entry(pentry); | |
814 | if (unlikely(non_swap_entry(entry))) | |
815 | continue; | |
816 | page = __read_swap_cache_async(entry, gfp_mask, vma, | |
817 | vmf->address, &page_allocated); | |
818 | if (!page) | |
819 | continue; | |
820 | if (page_allocated) { | |
5169b844 | 821 | swap_readpage(page, false, &splug); |
eaf649eb | 822 | if (i != ra_info.offset) { |
ec560175 HY |
823 | SetPageReadahead(page); |
824 | count_vm_event(SWAP_RA); | |
825 | } | |
826 | } | |
827 | put_page(page); | |
828 | } | |
829 | blk_finish_plug(&plug); | |
5169b844 | 830 | swap_read_unplug(splug); |
ec560175 HY |
831 | lru_add_drain(); |
832 | skip: | |
5169b844 | 833 | /* The page was likely read above, so no need for plugging here */ |
ec560175 | 834 | return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, |
5169b844 | 835 | ra_info.win == 1, NULL); |
ec560175 | 836 | } |
d9bfcfdc | 837 | |
e9e9b7ec MK |
838 | /** |
839 | * swapin_readahead - swap in pages in hope we need them soon | |
840 | * @entry: swap entry of this memory | |
841 | * @gfp_mask: memory allocation flags | |
842 | * @vmf: fault information | |
843 | * | |
844 | * Returns the struct page for entry and addr, after queueing swapin. | |
845 | * | |
846 | * It's a main entry function for swap readahead. By the configuration, | |
847 | * it will read ahead blocks by cluster-based(ie, physical disk based) | |
848 | * or vma-based(ie, virtual address based on faulty address) readahead. | |
849 | */ | |
850 | struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, | |
851 | struct vm_fault *vmf) | |
852 | { | |
853 | return swap_use_vma_readahead() ? | |
854 | swap_vma_readahead(entry, gfp_mask, vmf) : | |
855 | swap_cluster_readahead(entry, gfp_mask, vmf); | |
856 | } | |
857 | ||
d9bfcfdc HY |
858 | #ifdef CONFIG_SYSFS |
859 | static ssize_t vma_ra_enabled_show(struct kobject *kobj, | |
860 | struct kobj_attribute *attr, char *buf) | |
861 | { | |
ae7a927d JP |
862 | return sysfs_emit(buf, "%s\n", |
863 | enable_vma_readahead ? "true" : "false"); | |
d9bfcfdc HY |
864 | } |
865 | static ssize_t vma_ra_enabled_store(struct kobject *kobj, | |
866 | struct kobj_attribute *attr, | |
867 | const char *buf, size_t count) | |
868 | { | |
717aeab4 JG |
869 | ssize_t ret; |
870 | ||
871 | ret = kstrtobool(buf, &enable_vma_readahead); | |
872 | if (ret) | |
873 | return ret; | |
d9bfcfdc HY |
874 | |
875 | return count; | |
876 | } | |
6106b93e | 877 | static struct kobj_attribute vma_ra_enabled_attr = __ATTR_RW(vma_ra_enabled); |
d9bfcfdc | 878 | |
d9bfcfdc HY |
879 | static struct attribute *swap_attrs[] = { |
880 | &vma_ra_enabled_attr.attr, | |
d9bfcfdc HY |
881 | NULL, |
882 | }; | |
883 | ||
e48333b6 | 884 | static const struct attribute_group swap_attr_group = { |
d9bfcfdc HY |
885 | .attrs = swap_attrs, |
886 | }; | |
887 | ||
888 | static int __init swap_init_sysfs(void) | |
889 | { | |
890 | int err; | |
891 | struct kobject *swap_kobj; | |
892 | ||
893 | swap_kobj = kobject_create_and_add("swap", mm_kobj); | |
894 | if (!swap_kobj) { | |
895 | pr_err("failed to create swap kobject\n"); | |
896 | return -ENOMEM; | |
897 | } | |
898 | err = sysfs_create_group(swap_kobj, &swap_attr_group); | |
899 | if (err) { | |
900 | pr_err("failed to register swap group\n"); | |
901 | goto delete_obj; | |
902 | } | |
903 | return 0; | |
904 | ||
905 | delete_obj: | |
906 | kobject_put(swap_kobj); | |
907 | return err; | |
908 | } | |
909 | subsys_initcall(swap_init_sysfs); | |
910 | #endif |