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" |
1da177e4 LT |
26 | |
27 | /* | |
28 | * swapper_space is a fiction, retained to simplify the path through | |
7eaceacc | 29 | * vmscan's shrink_page_list. |
1da177e4 | 30 | */ |
f5e54d6e | 31 | static const struct address_space_operations swap_aops = { |
1da177e4 | 32 | .writepage = swap_writepage, |
62c230bc | 33 | .set_page_dirty = swap_set_page_dirty, |
1c93923c | 34 | #ifdef CONFIG_MIGRATION |
e965f963 | 35 | .migratepage = migrate_page, |
1c93923c | 36 | #endif |
1da177e4 LT |
37 | }; |
38 | ||
783cb68e CD |
39 | struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly; |
40 | static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly; | |
f5c754d6 | 41 | static bool enable_vma_readahead __read_mostly = true; |
ec560175 | 42 | |
ec560175 HY |
43 | #define SWAP_RA_WIN_SHIFT (PAGE_SHIFT / 2) |
44 | #define SWAP_RA_HITS_MASK ((1UL << SWAP_RA_WIN_SHIFT) - 1) | |
45 | #define SWAP_RA_HITS_MAX SWAP_RA_HITS_MASK | |
46 | #define SWAP_RA_WIN_MASK (~PAGE_MASK & ~SWAP_RA_HITS_MASK) | |
47 | ||
48 | #define SWAP_RA_HITS(v) ((v) & SWAP_RA_HITS_MASK) | |
49 | #define SWAP_RA_WIN(v) (((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT) | |
50 | #define SWAP_RA_ADDR(v) ((v) & PAGE_MASK) | |
51 | ||
52 | #define SWAP_RA_VAL(addr, win, hits) \ | |
53 | (((addr) & PAGE_MASK) | \ | |
54 | (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) | \ | |
55 | ((hits) & SWAP_RA_HITS_MASK)) | |
56 | ||
57 | /* Initial readahead hits is 4 to start up with a small window */ | |
58 | #define GET_SWAP_RA_VAL(vma) \ | |
59 | (atomic_long_read(&(vma)->swap_readahead_info) ? : 4) | |
1da177e4 | 60 | |
b96a3db2 QC |
61 | #define INC_CACHE_INFO(x) data_race(swap_cache_info.x++) |
62 | #define ADD_CACHE_INFO(x, nr) data_race(swap_cache_info.x += (nr)) | |
1da177e4 LT |
63 | |
64 | static struct { | |
65 | unsigned long add_total; | |
66 | unsigned long del_total; | |
67 | unsigned long find_success; | |
68 | unsigned long find_total; | |
1da177e4 LT |
69 | } swap_cache_info; |
70 | ||
579f8290 SL |
71 | static atomic_t swapin_readahead_hits = ATOMIC_INIT(4); |
72 | ||
1da177e4 LT |
73 | void show_swap_cache_info(void) |
74 | { | |
33806f06 | 75 | printk("%lu pages in swap cache\n", total_swapcache_pages()); |
2c97b7fc | 76 | printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n", |
1da177e4 | 77 | swap_cache_info.add_total, swap_cache_info.del_total, |
bb63be0a | 78 | swap_cache_info.find_success, swap_cache_info.find_total); |
ec8acf20 SL |
79 | printk("Free swap = %ldkB\n", |
80 | get_nr_swap_pages() << (PAGE_SHIFT - 10)); | |
1da177e4 LT |
81 | printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10)); |
82 | } | |
83 | ||
aae466b0 JK |
84 | void *get_shadow_from_swap_cache(swp_entry_t entry) |
85 | { | |
86 | struct address_space *address_space = swap_address_space(entry); | |
87 | pgoff_t idx = swp_offset(entry); | |
88 | struct page *page; | |
89 | ||
8c647dd1 | 90 | page = xa_load(&address_space->i_pages, idx); |
aae466b0 JK |
91 | if (xa_is_value(page)) |
92 | return page; | |
aae466b0 JK |
93 | return NULL; |
94 | } | |
95 | ||
1da177e4 | 96 | /* |
8d93b41c | 97 | * add_to_swap_cache resembles add_to_page_cache_locked on swapper_space, |
1da177e4 LT |
98 | * but sets SwapCache flag and private instead of mapping and index. |
99 | */ | |
3852f676 JK |
100 | int add_to_swap_cache(struct page *page, swp_entry_t entry, |
101 | gfp_t gfp, void **shadowp) | |
1da177e4 | 102 | { |
8d93b41c | 103 | struct address_space *address_space = swap_address_space(entry); |
38d8b4e6 | 104 | pgoff_t idx = swp_offset(entry); |
8d93b41c | 105 | XA_STATE_ORDER(xas, &address_space->i_pages, idx, compound_order(page)); |
6c357848 | 106 | unsigned long i, nr = thp_nr_pages(page); |
3852f676 | 107 | void *old; |
1da177e4 | 108 | |
309381fe SL |
109 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
110 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
111 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
51726b12 | 112 | |
38d8b4e6 | 113 | page_ref_add(page, nr); |
31a56396 | 114 | SetPageSwapCache(page); |
31a56396 | 115 | |
8d93b41c | 116 | do { |
3852f676 JK |
117 | unsigned long nr_shadows = 0; |
118 | ||
8d93b41c MW |
119 | xas_lock_irq(&xas); |
120 | xas_create_range(&xas); | |
121 | if (xas_error(&xas)) | |
122 | goto unlock; | |
123 | for (i = 0; i < nr; i++) { | |
124 | VM_BUG_ON_PAGE(xas.xa_index != idx + i, page); | |
3852f676 JK |
125 | old = xas_load(&xas); |
126 | if (xa_is_value(old)) { | |
127 | nr_shadows++; | |
128 | if (shadowp) | |
129 | *shadowp = old; | |
130 | } | |
8d93b41c | 131 | set_page_private(page + i, entry.val + i); |
4101196b | 132 | xas_store(&xas, page); |
8d93b41c MW |
133 | xas_next(&xas); |
134 | } | |
38d8b4e6 HY |
135 | address_space->nrpages += nr; |
136 | __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr); | |
b6038942 | 137 | __mod_lruvec_page_state(page, NR_SWAPCACHE, nr); |
38d8b4e6 | 138 | ADD_CACHE_INFO(add_total, nr); |
8d93b41c MW |
139 | unlock: |
140 | xas_unlock_irq(&xas); | |
141 | } while (xas_nomem(&xas, gfp)); | |
31a56396 | 142 | |
8d93b41c MW |
143 | if (!xas_error(&xas)) |
144 | return 0; | |
31a56396 | 145 | |
8d93b41c MW |
146 | ClearPageSwapCache(page); |
147 | page_ref_sub(page, nr); | |
148 | return xas_error(&xas); | |
1da177e4 LT |
149 | } |
150 | ||
1da177e4 LT |
151 | /* |
152 | * This must be called only on pages that have | |
153 | * been verified to be in the swap cache. | |
154 | */ | |
3852f676 JK |
155 | void __delete_from_swap_cache(struct page *page, |
156 | swp_entry_t entry, void *shadow) | |
1da177e4 | 157 | { |
4e17ec25 | 158 | struct address_space *address_space = swap_address_space(entry); |
6c357848 | 159 | int i, nr = thp_nr_pages(page); |
4e17ec25 MW |
160 | pgoff_t idx = swp_offset(entry); |
161 | XA_STATE(xas, &address_space->i_pages, idx); | |
33806f06 | 162 | |
309381fe SL |
163 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
164 | VM_BUG_ON_PAGE(!PageSwapCache(page), page); | |
165 | VM_BUG_ON_PAGE(PageWriteback(page), page); | |
1da177e4 | 166 | |
38d8b4e6 | 167 | for (i = 0; i < nr; i++) { |
3852f676 | 168 | void *entry = xas_store(&xas, shadow); |
4101196b | 169 | VM_BUG_ON_PAGE(entry != page, entry); |
38d8b4e6 | 170 | set_page_private(page + i, 0); |
4e17ec25 | 171 | xas_next(&xas); |
38d8b4e6 | 172 | } |
1da177e4 | 173 | ClearPageSwapCache(page); |
38d8b4e6 HY |
174 | address_space->nrpages -= nr; |
175 | __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr); | |
b6038942 | 176 | __mod_lruvec_page_state(page, NR_SWAPCACHE, -nr); |
38d8b4e6 | 177 | ADD_CACHE_INFO(del_total, nr); |
1da177e4 LT |
178 | } |
179 | ||
180 | /** | |
181 | * add_to_swap - allocate swap space for a page | |
182 | * @page: page we want to move to swap | |
183 | * | |
184 | * Allocate swap space for the page and add the page to the | |
185 | * swap cache. Caller needs to hold the page lock. | |
186 | */ | |
0f074658 | 187 | int add_to_swap(struct page *page) |
1da177e4 LT |
188 | { |
189 | swp_entry_t entry; | |
1da177e4 LT |
190 | int err; |
191 | ||
309381fe SL |
192 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
193 | VM_BUG_ON_PAGE(!PageUptodate(page), page); | |
1da177e4 | 194 | |
38d8b4e6 | 195 | entry = get_swap_page(page); |
2ca4532a | 196 | if (!entry.val) |
0f074658 MK |
197 | return 0; |
198 | ||
2ca4532a | 199 | /* |
8d93b41c | 200 | * XArray node allocations from PF_MEMALLOC contexts could |
2ca4532a DN |
201 | * completely exhaust the page allocator. __GFP_NOMEMALLOC |
202 | * stops emergency reserves from being allocated. | |
203 | * | |
204 | * TODO: this could cause a theoretical memory reclaim | |
205 | * deadlock in the swap out path. | |
206 | */ | |
207 | /* | |
854e9ed0 | 208 | * Add it to the swap cache. |
2ca4532a DN |
209 | */ |
210 | err = add_to_swap_cache(page, entry, | |
3852f676 | 211 | __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN, NULL); |
38d8b4e6 | 212 | if (err) |
bd53b714 | 213 | /* |
2ca4532a DN |
214 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely |
215 | * clear SWAP_HAS_CACHE flag. | |
1da177e4 | 216 | */ |
0f074658 | 217 | goto fail; |
9625456c SL |
218 | /* |
219 | * Normally the page will be dirtied in unmap because its pte should be | |
0e9aa675 | 220 | * dirty. A special case is MADV_FREE page. The page's pte could have |
9625456c SL |
221 | * dirty bit cleared but the page's SwapBacked bit is still set because |
222 | * clearing the dirty bit and SwapBacked bit has no lock protected. For | |
223 | * such page, unmap will not set dirty bit for it, so page reclaim will | |
224 | * not write the page out. This can cause data corruption when the page | |
225 | * is swap in later. Always setting the dirty bit for the page solves | |
226 | * the problem. | |
227 | */ | |
228 | set_page_dirty(page); | |
38d8b4e6 HY |
229 | |
230 | return 1; | |
231 | ||
38d8b4e6 | 232 | fail: |
0f074658 | 233 | put_swap_page(page, entry); |
38d8b4e6 | 234 | return 0; |
1da177e4 LT |
235 | } |
236 | ||
237 | /* | |
238 | * This must be called only on pages that have | |
239 | * been verified to be in the swap cache and locked. | |
240 | * It will never put the page into the free list, | |
241 | * the caller has a reference on the page. | |
242 | */ | |
243 | void delete_from_swap_cache(struct page *page) | |
244 | { | |
4e17ec25 MW |
245 | swp_entry_t entry = { .val = page_private(page) }; |
246 | struct address_space *address_space = swap_address_space(entry); | |
1da177e4 | 247 | |
b93b0163 | 248 | xa_lock_irq(&address_space->i_pages); |
3852f676 | 249 | __delete_from_swap_cache(page, entry, NULL); |
b93b0163 | 250 | xa_unlock_irq(&address_space->i_pages); |
1da177e4 | 251 | |
75f6d6d2 | 252 | put_swap_page(page, entry); |
6c357848 | 253 | page_ref_sub(page, thp_nr_pages(page)); |
1da177e4 LT |
254 | } |
255 | ||
3852f676 JK |
256 | void clear_shadow_from_swap_cache(int type, unsigned long begin, |
257 | unsigned long end) | |
258 | { | |
259 | unsigned long curr = begin; | |
260 | void *old; | |
261 | ||
262 | for (;;) { | |
263 | unsigned long nr_shadows = 0; | |
264 | swp_entry_t entry = swp_entry(type, curr); | |
265 | struct address_space *address_space = swap_address_space(entry); | |
266 | XA_STATE(xas, &address_space->i_pages, curr); | |
267 | ||
268 | xa_lock_irq(&address_space->i_pages); | |
269 | xas_for_each(&xas, old, end) { | |
270 | if (!xa_is_value(old)) | |
271 | continue; | |
272 | xas_store(&xas, NULL); | |
273 | nr_shadows++; | |
274 | } | |
3852f676 JK |
275 | xa_unlock_irq(&address_space->i_pages); |
276 | ||
277 | /* search the next swapcache until we meet end */ | |
278 | curr >>= SWAP_ADDRESS_SPACE_SHIFT; | |
279 | curr++; | |
280 | curr <<= SWAP_ADDRESS_SPACE_SHIFT; | |
281 | if (curr > end) | |
282 | break; | |
283 | } | |
284 | } | |
285 | ||
1da177e4 LT |
286 | /* |
287 | * If we are the only user, then try to free up the swap cache. | |
288 | * | |
289 | * Its ok to check for PageSwapCache without the page lock | |
a2c43eed HD |
290 | * here because we are going to recheck again inside |
291 | * try_to_free_swap() _with_ the lock. | |
1da177e4 LT |
292 | * - Marcelo |
293 | */ | |
294 | static inline void free_swap_cache(struct page *page) | |
295 | { | |
a2c43eed HD |
296 | if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) { |
297 | try_to_free_swap(page); | |
1da177e4 LT |
298 | unlock_page(page); |
299 | } | |
300 | } | |
301 | ||
302 | /* | |
303 | * Perform a free_page(), also freeing any swap cache associated with | |
b8072f09 | 304 | * this page if it is the last user of the page. |
1da177e4 LT |
305 | */ |
306 | void free_page_and_swap_cache(struct page *page) | |
307 | { | |
308 | free_swap_cache(page); | |
6fcb52a5 | 309 | if (!is_huge_zero_page(page)) |
770a5370 | 310 | put_page(page); |
1da177e4 LT |
311 | } |
312 | ||
313 | /* | |
314 | * Passed an array of pages, drop them all from swapcache and then release | |
315 | * them. They are removed from the LRU and freed if this is their last use. | |
316 | */ | |
317 | void free_pages_and_swap_cache(struct page **pages, int nr) | |
318 | { | |
1da177e4 | 319 | struct page **pagep = pages; |
aabfb572 | 320 | int i; |
1da177e4 LT |
321 | |
322 | lru_add_drain(); | |
aabfb572 MH |
323 | for (i = 0; i < nr; i++) |
324 | free_swap_cache(pagep[i]); | |
c6f92f9f | 325 | release_pages(pagep, nr); |
1da177e4 LT |
326 | } |
327 | ||
e9e9b7ec MK |
328 | static inline bool swap_use_vma_readahead(void) |
329 | { | |
330 | return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap); | |
331 | } | |
332 | ||
1da177e4 LT |
333 | /* |
334 | * Lookup a swap entry in the swap cache. A found page will be returned | |
335 | * unlocked and with its refcount incremented - we rely on the kernel | |
336 | * lock getting page table operations atomic even if we drop the page | |
337 | * lock before returning. | |
338 | */ | |
ec560175 HY |
339 | struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma, |
340 | unsigned long addr) | |
1da177e4 LT |
341 | { |
342 | struct page *page; | |
eb085574 | 343 | struct swap_info_struct *si; |
1da177e4 | 344 | |
eb085574 HY |
345 | si = get_swap_device(entry); |
346 | if (!si) | |
347 | return NULL; | |
f6ab1f7f | 348 | page = find_get_page(swap_address_space(entry), swp_offset(entry)); |
eb085574 | 349 | put_swap_device(si); |
1da177e4 | 350 | |
ec560175 HY |
351 | INC_CACHE_INFO(find_total); |
352 | if (page) { | |
eaf649eb MK |
353 | bool vma_ra = swap_use_vma_readahead(); |
354 | bool readahead; | |
355 | ||
1da177e4 | 356 | INC_CACHE_INFO(find_success); |
eaf649eb MK |
357 | /* |
358 | * At the moment, we don't support PG_readahead for anon THP | |
359 | * so let's bail out rather than confusing the readahead stat. | |
360 | */ | |
ec560175 HY |
361 | if (unlikely(PageTransCompound(page))) |
362 | return page; | |
eaf649eb | 363 | |
ec560175 | 364 | readahead = TestClearPageReadahead(page); |
eaf649eb MK |
365 | if (vma && vma_ra) { |
366 | unsigned long ra_val; | |
367 | int win, hits; | |
368 | ||
369 | ra_val = GET_SWAP_RA_VAL(vma); | |
370 | win = SWAP_RA_WIN(ra_val); | |
371 | hits = SWAP_RA_HITS(ra_val); | |
ec560175 HY |
372 | if (readahead) |
373 | hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX); | |
374 | atomic_long_set(&vma->swap_readahead_info, | |
375 | SWAP_RA_VAL(addr, win, hits)); | |
376 | } | |
eaf649eb | 377 | |
ec560175 | 378 | if (readahead) { |
cbc65df2 | 379 | count_vm_event(SWAP_RA_HIT); |
eaf649eb | 380 | if (!vma || !vma_ra) |
ec560175 | 381 | atomic_inc(&swapin_readahead_hits); |
cbc65df2 | 382 | } |
579f8290 | 383 | } |
eaf649eb | 384 | |
1da177e4 LT |
385 | return page; |
386 | } | |
387 | ||
61ef1865 MWO |
388 | /** |
389 | * find_get_incore_page - Find and get a page from the page or swap caches. | |
390 | * @mapping: The address_space to search. | |
391 | * @index: The page cache index. | |
392 | * | |
393 | * This differs from find_get_page() in that it will also look for the | |
394 | * page in the swap cache. | |
395 | * | |
396 | * Return: The found page or %NULL. | |
397 | */ | |
398 | struct page *find_get_incore_page(struct address_space *mapping, pgoff_t index) | |
399 | { | |
400 | swp_entry_t swp; | |
401 | struct swap_info_struct *si; | |
44835d20 MWO |
402 | struct page *page = pagecache_get_page(mapping, index, |
403 | FGP_ENTRY | FGP_HEAD, 0); | |
61ef1865 | 404 | |
a6de4b48 | 405 | if (!page) |
61ef1865 | 406 | return page; |
a6de4b48 MWO |
407 | if (!xa_is_value(page)) |
408 | return find_subpage(page, index); | |
61ef1865 MWO |
409 | if (!shmem_mapping(mapping)) |
410 | return NULL; | |
411 | ||
412 | swp = radix_to_swp_entry(page); | |
413 | /* Prevent swapoff from happening to us */ | |
414 | si = get_swap_device(swp); | |
415 | if (!si) | |
416 | return NULL; | |
417 | page = find_get_page(swap_address_space(swp), swp_offset(swp)); | |
418 | put_swap_device(si); | |
419 | return page; | |
420 | } | |
421 | ||
5b999aad DS |
422 | struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, |
423 | struct vm_area_struct *vma, unsigned long addr, | |
424 | bool *new_page_allocated) | |
1da177e4 | 425 | { |
eb085574 | 426 | struct swap_info_struct *si; |
4c6355b2 | 427 | struct page *page; |
aae466b0 | 428 | void *shadow = NULL; |
4c6355b2 | 429 | |
5b999aad | 430 | *new_page_allocated = false; |
1da177e4 | 431 | |
4c6355b2 JW |
432 | for (;;) { |
433 | int err; | |
1da177e4 LT |
434 | /* |
435 | * First check the swap cache. Since this is normally | |
436 | * called after lookup_swap_cache() failed, re-calling | |
437 | * that would confuse statistics. | |
438 | */ | |
eb085574 HY |
439 | si = get_swap_device(entry); |
440 | if (!si) | |
4c6355b2 JW |
441 | return NULL; |
442 | page = find_get_page(swap_address_space(entry), | |
443 | swp_offset(entry)); | |
eb085574 | 444 | put_swap_device(si); |
4c6355b2 JW |
445 | if (page) |
446 | return page; | |
1da177e4 | 447 | |
ba81f838 HY |
448 | /* |
449 | * Just skip read ahead for unused swap slot. | |
450 | * During swap_off when swap_slot_cache is disabled, | |
451 | * we have to handle the race between putting | |
452 | * swap entry in swap cache and marking swap slot | |
453 | * as SWAP_HAS_CACHE. That's done in later part of code or | |
454 | * else swap_off will be aborted if we return NULL. | |
455 | */ | |
456 | if (!__swp_swapcount(entry) && swap_slot_cache_enabled) | |
4c6355b2 | 457 | return NULL; |
e8c26ab6 | 458 | |
1da177e4 | 459 | /* |
4c6355b2 JW |
460 | * Get a new page to read into from swap. Allocate it now, |
461 | * before marking swap_map SWAP_HAS_CACHE, when -EEXIST will | |
462 | * cause any racers to loop around until we add it to cache. | |
1da177e4 | 463 | */ |
4c6355b2 JW |
464 | page = alloc_page_vma(gfp_mask, vma, addr); |
465 | if (!page) | |
466 | return NULL; | |
1da177e4 | 467 | |
f000944d HD |
468 | /* |
469 | * Swap entry may have been freed since our caller observed it. | |
470 | */ | |
355cfa73 | 471 | err = swapcache_prepare(entry); |
4c6355b2 | 472 | if (!err) |
f000944d HD |
473 | break; |
474 | ||
4c6355b2 JW |
475 | put_page(page); |
476 | if (err != -EEXIST) | |
477 | return NULL; | |
478 | ||
2ca4532a | 479 | /* |
4c6355b2 JW |
480 | * We might race against __delete_from_swap_cache(), and |
481 | * stumble across a swap_map entry whose SWAP_HAS_CACHE | |
482 | * has not yet been cleared. Or race against another | |
483 | * __read_swap_cache_async(), which has set SWAP_HAS_CACHE | |
484 | * in swap_map, but not yet added its page to swap cache. | |
2ca4532a | 485 | */ |
4c6355b2 JW |
486 | cond_resched(); |
487 | } | |
488 | ||
489 | /* | |
490 | * The swap entry is ours to swap in. Prepare the new page. | |
491 | */ | |
492 | ||
493 | __SetPageLocked(page); | |
494 | __SetPageSwapBacked(page); | |
495 | ||
0add0c77 | 496 | if (mem_cgroup_swapin_charge_page(page, NULL, gfp_mask, entry)) |
4c6355b2 | 497 | goto fail_unlock; |
4c6355b2 | 498 | |
0add0c77 SB |
499 | /* May fail (-ENOMEM) if XArray node allocation failed. */ |
500 | if (add_to_swap_cache(page, entry, gfp_mask & GFP_RECLAIM_MASK, &shadow)) | |
4c6355b2 | 501 | goto fail_unlock; |
0add0c77 SB |
502 | |
503 | mem_cgroup_swapin_uncharge_swap(entry); | |
4c6355b2 | 504 | |
aae466b0 JK |
505 | if (shadow) |
506 | workingset_refault(page, shadow); | |
314b57fb | 507 | |
4c6355b2 | 508 | /* Caller will initiate read into locked page */ |
6058eaec | 509 | lru_cache_add(page); |
4c6355b2 JW |
510 | *new_page_allocated = true; |
511 | return page; | |
1da177e4 | 512 | |
4c6355b2 | 513 | fail_unlock: |
0add0c77 | 514 | put_swap_page(page, entry); |
4c6355b2 JW |
515 | unlock_page(page); |
516 | put_page(page); | |
517 | return NULL; | |
1da177e4 | 518 | } |
46017e95 | 519 | |
5b999aad DS |
520 | /* |
521 | * Locate a page of swap in physical memory, reserving swap cache space | |
522 | * and reading the disk if it is not already cached. | |
523 | * A failure return means that either the page allocation failed or that | |
524 | * the swap entry is no longer in use. | |
525 | */ | |
526 | struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, | |
23955622 | 527 | struct vm_area_struct *vma, unsigned long addr, bool do_poll) |
5b999aad DS |
528 | { |
529 | bool page_was_allocated; | |
530 | struct page *retpage = __read_swap_cache_async(entry, gfp_mask, | |
531 | vma, addr, &page_was_allocated); | |
532 | ||
533 | if (page_was_allocated) | |
23955622 | 534 | swap_readpage(retpage, do_poll); |
5b999aad DS |
535 | |
536 | return retpage; | |
537 | } | |
538 | ||
ec560175 HY |
539 | static unsigned int __swapin_nr_pages(unsigned long prev_offset, |
540 | unsigned long offset, | |
541 | int hits, | |
542 | int max_pages, | |
543 | int prev_win) | |
579f8290 | 544 | { |
ec560175 | 545 | unsigned int pages, last_ra; |
579f8290 SL |
546 | |
547 | /* | |
548 | * This heuristic has been found to work well on both sequential and | |
549 | * random loads, swapping to hard disk or to SSD: please don't ask | |
550 | * what the "+ 2" means, it just happens to work well, that's all. | |
551 | */ | |
ec560175 | 552 | pages = hits + 2; |
579f8290 SL |
553 | if (pages == 2) { |
554 | /* | |
555 | * We can have no readahead hits to judge by: but must not get | |
556 | * stuck here forever, so check for an adjacent offset instead | |
557 | * (and don't even bother to check whether swap type is same). | |
558 | */ | |
559 | if (offset != prev_offset + 1 && offset != prev_offset - 1) | |
560 | pages = 1; | |
579f8290 SL |
561 | } else { |
562 | unsigned int roundup = 4; | |
563 | while (roundup < pages) | |
564 | roundup <<= 1; | |
565 | pages = roundup; | |
566 | } | |
567 | ||
568 | if (pages > max_pages) | |
569 | pages = max_pages; | |
570 | ||
571 | /* Don't shrink readahead too fast */ | |
ec560175 | 572 | last_ra = prev_win / 2; |
579f8290 SL |
573 | if (pages < last_ra) |
574 | pages = last_ra; | |
ec560175 HY |
575 | |
576 | return pages; | |
577 | } | |
578 | ||
579 | static unsigned long swapin_nr_pages(unsigned long offset) | |
580 | { | |
581 | static unsigned long prev_offset; | |
582 | unsigned int hits, pages, max_pages; | |
583 | static atomic_t last_readahead_pages; | |
584 | ||
585 | max_pages = 1 << READ_ONCE(page_cluster); | |
586 | if (max_pages <= 1) | |
587 | return 1; | |
588 | ||
589 | hits = atomic_xchg(&swapin_readahead_hits, 0); | |
d6c1f098 QC |
590 | pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits, |
591 | max_pages, | |
ec560175 HY |
592 | atomic_read(&last_readahead_pages)); |
593 | if (!hits) | |
d6c1f098 | 594 | WRITE_ONCE(prev_offset, offset); |
579f8290 SL |
595 | atomic_set(&last_readahead_pages, pages); |
596 | ||
597 | return pages; | |
598 | } | |
599 | ||
46017e95 | 600 | /** |
e9e9b7ec | 601 | * swap_cluster_readahead - swap in pages in hope we need them soon |
46017e95 | 602 | * @entry: swap entry of this memory |
7682486b | 603 | * @gfp_mask: memory allocation flags |
e9e9b7ec | 604 | * @vmf: fault information |
46017e95 HD |
605 | * |
606 | * Returns the struct page for entry and addr, after queueing swapin. | |
607 | * | |
608 | * Primitive swap readahead code. We simply read an aligned block of | |
609 | * (1 << page_cluster) entries in the swap area. This method is chosen | |
610 | * because it doesn't cost us any seek time. We also make sure to queue | |
611 | * the 'original' request together with the readahead ones... | |
612 | * | |
613 | * This has been extended to use the NUMA policies from the mm triggering | |
614 | * the readahead. | |
615 | * | |
c1e8d7c6 | 616 | * Caller must hold read mmap_lock if vmf->vma is not NULL. |
46017e95 | 617 | */ |
e9e9b7ec MK |
618 | struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask, |
619 | struct vm_fault *vmf) | |
46017e95 | 620 | { |
46017e95 | 621 | struct page *page; |
579f8290 SL |
622 | unsigned long entry_offset = swp_offset(entry); |
623 | unsigned long offset = entry_offset; | |
67f96aa2 | 624 | unsigned long start_offset, end_offset; |
579f8290 | 625 | unsigned long mask; |
e9a6effa | 626 | struct swap_info_struct *si = swp_swap_info(entry); |
3fb5c298 | 627 | struct blk_plug plug; |
c4fa6309 | 628 | bool do_poll = true, page_allocated; |
e9e9b7ec MK |
629 | struct vm_area_struct *vma = vmf->vma; |
630 | unsigned long addr = vmf->address; | |
46017e95 | 631 | |
579f8290 SL |
632 | mask = swapin_nr_pages(offset) - 1; |
633 | if (!mask) | |
634 | goto skip; | |
635 | ||
8fd2e0b5 | 636 | /* Test swap type to make sure the dereference is safe */ |
32646315 | 637 | if (likely(si->flags & (SWP_BLKDEV | SWP_FS_OPS))) { |
8fd2e0b5 YS |
638 | struct inode *inode = si->swap_file->f_mapping->host; |
639 | if (inode_read_congested(inode)) | |
640 | goto skip; | |
641 | } | |
642 | ||
23955622 | 643 | do_poll = false; |
67f96aa2 RR |
644 | /* Read a page_cluster sized and aligned cluster around offset. */ |
645 | start_offset = offset & ~mask; | |
646 | end_offset = offset | mask; | |
647 | if (!start_offset) /* First page is swap header. */ | |
648 | start_offset++; | |
e9a6effa HY |
649 | if (end_offset >= si->max) |
650 | end_offset = si->max - 1; | |
67f96aa2 | 651 | |
3fb5c298 | 652 | blk_start_plug(&plug); |
67f96aa2 | 653 | for (offset = start_offset; offset <= end_offset ; offset++) { |
46017e95 | 654 | /* Ok, do the async read-ahead now */ |
c4fa6309 HY |
655 | page = __read_swap_cache_async( |
656 | swp_entry(swp_type(entry), offset), | |
657 | gfp_mask, vma, addr, &page_allocated); | |
46017e95 | 658 | if (!page) |
67f96aa2 | 659 | continue; |
c4fa6309 HY |
660 | if (page_allocated) { |
661 | swap_readpage(page, false); | |
eaf649eb | 662 | if (offset != entry_offset) { |
c4fa6309 HY |
663 | SetPageReadahead(page); |
664 | count_vm_event(SWAP_RA); | |
665 | } | |
cbc65df2 | 666 | } |
09cbfeaf | 667 | put_page(page); |
46017e95 | 668 | } |
3fb5c298 CE |
669 | blk_finish_plug(&plug); |
670 | ||
46017e95 | 671 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
579f8290 | 672 | skip: |
23955622 | 673 | return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll); |
46017e95 | 674 | } |
4b3ef9da HY |
675 | |
676 | int init_swap_address_space(unsigned int type, unsigned long nr_pages) | |
677 | { | |
678 | struct address_space *spaces, *space; | |
679 | unsigned int i, nr; | |
680 | ||
681 | nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES); | |
778e1cdd | 682 | spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL); |
4b3ef9da HY |
683 | if (!spaces) |
684 | return -ENOMEM; | |
685 | for (i = 0; i < nr; i++) { | |
686 | space = spaces + i; | |
a2833486 | 687 | xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ); |
4b3ef9da HY |
688 | atomic_set(&space->i_mmap_writable, 0); |
689 | space->a_ops = &swap_aops; | |
690 | /* swap cache doesn't use writeback related tags */ | |
691 | mapping_set_no_writeback_tags(space); | |
4b3ef9da HY |
692 | } |
693 | nr_swapper_spaces[type] = nr; | |
054f1d1f | 694 | swapper_spaces[type] = spaces; |
4b3ef9da HY |
695 | |
696 | return 0; | |
697 | } | |
698 | ||
699 | void exit_swap_address_space(unsigned int type) | |
700 | { | |
054f1d1f | 701 | kvfree(swapper_spaces[type]); |
4b3ef9da | 702 | nr_swapper_spaces[type] = 0; |
054f1d1f | 703 | swapper_spaces[type] = NULL; |
4b3ef9da | 704 | } |
ec560175 HY |
705 | |
706 | static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma, | |
707 | unsigned long faddr, | |
708 | unsigned long lpfn, | |
709 | unsigned long rpfn, | |
710 | unsigned long *start, | |
711 | unsigned long *end) | |
712 | { | |
713 | *start = max3(lpfn, PFN_DOWN(vma->vm_start), | |
714 | PFN_DOWN(faddr & PMD_MASK)); | |
715 | *end = min3(rpfn, PFN_DOWN(vma->vm_end), | |
716 | PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE)); | |
717 | } | |
718 | ||
eaf649eb MK |
719 | static void swap_ra_info(struct vm_fault *vmf, |
720 | struct vma_swap_readahead *ra_info) | |
ec560175 HY |
721 | { |
722 | struct vm_area_struct *vma = vmf->vma; | |
eaf649eb | 723 | unsigned long ra_val; |
ec560175 HY |
724 | swp_entry_t entry; |
725 | unsigned long faddr, pfn, fpfn; | |
726 | unsigned long start, end; | |
eaf649eb | 727 | pte_t *pte, *orig_pte; |
ec560175 HY |
728 | unsigned int max_win, hits, prev_win, win, left; |
729 | #ifndef CONFIG_64BIT | |
730 | pte_t *tpte; | |
731 | #endif | |
732 | ||
61b63972 HY |
733 | max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster), |
734 | SWAP_RA_ORDER_CEILING); | |
735 | if (max_win == 1) { | |
eaf649eb MK |
736 | ra_info->win = 1; |
737 | return; | |
61b63972 HY |
738 | } |
739 | ||
ec560175 | 740 | faddr = vmf->address; |
eaf649eb MK |
741 | orig_pte = pte = pte_offset_map(vmf->pmd, faddr); |
742 | entry = pte_to_swp_entry(*pte); | |
743 | if ((unlikely(non_swap_entry(entry)))) { | |
744 | pte_unmap(orig_pte); | |
745 | return; | |
746 | } | |
ec560175 | 747 | |
ec560175 | 748 | fpfn = PFN_DOWN(faddr); |
eaf649eb MK |
749 | ra_val = GET_SWAP_RA_VAL(vma); |
750 | pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val)); | |
751 | prev_win = SWAP_RA_WIN(ra_val); | |
752 | hits = SWAP_RA_HITS(ra_val); | |
753 | ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits, | |
ec560175 HY |
754 | max_win, prev_win); |
755 | atomic_long_set(&vma->swap_readahead_info, | |
756 | SWAP_RA_VAL(faddr, win, 0)); | |
757 | ||
eaf649eb MK |
758 | if (win == 1) { |
759 | pte_unmap(orig_pte); | |
760 | return; | |
761 | } | |
ec560175 HY |
762 | |
763 | /* Copy the PTEs because the page table may be unmapped */ | |
764 | if (fpfn == pfn + 1) | |
765 | swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end); | |
766 | else if (pfn == fpfn + 1) | |
767 | swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1, | |
768 | &start, &end); | |
769 | else { | |
770 | left = (win - 1) / 2; | |
771 | swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left, | |
772 | &start, &end); | |
773 | } | |
eaf649eb MK |
774 | ra_info->nr_pte = end - start; |
775 | ra_info->offset = fpfn - start; | |
776 | pte -= ra_info->offset; | |
ec560175 | 777 | #ifdef CONFIG_64BIT |
eaf649eb | 778 | ra_info->ptes = pte; |
ec560175 | 779 | #else |
eaf649eb | 780 | tpte = ra_info->ptes; |
ec560175 HY |
781 | for (pfn = start; pfn != end; pfn++) |
782 | *tpte++ = *pte++; | |
783 | #endif | |
eaf649eb | 784 | pte_unmap(orig_pte); |
ec560175 HY |
785 | } |
786 | ||
e9f59873 YS |
787 | /** |
788 | * swap_vma_readahead - swap in pages in hope we need them soon | |
27ec4878 | 789 | * @fentry: swap entry of this memory |
e9f59873 YS |
790 | * @gfp_mask: memory allocation flags |
791 | * @vmf: fault information | |
792 | * | |
793 | * Returns the struct page for entry and addr, after queueing swapin. | |
794 | * | |
cb152a1a | 795 | * Primitive swap readahead code. We simply read in a few pages whose |
e9f59873 YS |
796 | * virtual addresses are around the fault address in the same vma. |
797 | * | |
c1e8d7c6 | 798 | * Caller must hold read mmap_lock if vmf->vma is not NULL. |
e9f59873 YS |
799 | * |
800 | */ | |
f5c754d6 CIK |
801 | static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask, |
802 | struct vm_fault *vmf) | |
ec560175 HY |
803 | { |
804 | struct blk_plug plug; | |
805 | struct vm_area_struct *vma = vmf->vma; | |
806 | struct page *page; | |
807 | pte_t *pte, pentry; | |
808 | swp_entry_t entry; | |
809 | unsigned int i; | |
810 | bool page_allocated; | |
e97af699 ML |
811 | struct vma_swap_readahead ra_info = { |
812 | .win = 1, | |
813 | }; | |
ec560175 | 814 | |
eaf649eb MK |
815 | swap_ra_info(vmf, &ra_info); |
816 | if (ra_info.win == 1) | |
ec560175 HY |
817 | goto skip; |
818 | ||
819 | blk_start_plug(&plug); | |
eaf649eb | 820 | for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte; |
ec560175 HY |
821 | i++, pte++) { |
822 | pentry = *pte; | |
823 | if (pte_none(pentry)) | |
824 | continue; | |
825 | if (pte_present(pentry)) | |
826 | continue; | |
827 | entry = pte_to_swp_entry(pentry); | |
828 | if (unlikely(non_swap_entry(entry))) | |
829 | continue; | |
830 | page = __read_swap_cache_async(entry, gfp_mask, vma, | |
831 | vmf->address, &page_allocated); | |
832 | if (!page) | |
833 | continue; | |
834 | if (page_allocated) { | |
835 | swap_readpage(page, false); | |
eaf649eb | 836 | if (i != ra_info.offset) { |
ec560175 HY |
837 | SetPageReadahead(page); |
838 | count_vm_event(SWAP_RA); | |
839 | } | |
840 | } | |
841 | put_page(page); | |
842 | } | |
843 | blk_finish_plug(&plug); | |
844 | lru_add_drain(); | |
845 | skip: | |
846 | return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, | |
eaf649eb | 847 | ra_info.win == 1); |
ec560175 | 848 | } |
d9bfcfdc | 849 | |
e9e9b7ec MK |
850 | /** |
851 | * swapin_readahead - swap in pages in hope we need them soon | |
852 | * @entry: swap entry of this memory | |
853 | * @gfp_mask: memory allocation flags | |
854 | * @vmf: fault information | |
855 | * | |
856 | * Returns the struct page for entry and addr, after queueing swapin. | |
857 | * | |
858 | * It's a main entry function for swap readahead. By the configuration, | |
859 | * it will read ahead blocks by cluster-based(ie, physical disk based) | |
860 | * or vma-based(ie, virtual address based on faulty address) readahead. | |
861 | */ | |
862 | struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, | |
863 | struct vm_fault *vmf) | |
864 | { | |
865 | return swap_use_vma_readahead() ? | |
866 | swap_vma_readahead(entry, gfp_mask, vmf) : | |
867 | swap_cluster_readahead(entry, gfp_mask, vmf); | |
868 | } | |
869 | ||
d9bfcfdc HY |
870 | #ifdef CONFIG_SYSFS |
871 | static ssize_t vma_ra_enabled_show(struct kobject *kobj, | |
872 | struct kobj_attribute *attr, char *buf) | |
873 | { | |
ae7a927d JP |
874 | return sysfs_emit(buf, "%s\n", |
875 | enable_vma_readahead ? "true" : "false"); | |
d9bfcfdc HY |
876 | } |
877 | static ssize_t vma_ra_enabled_store(struct kobject *kobj, | |
878 | struct kobj_attribute *attr, | |
879 | const char *buf, size_t count) | |
880 | { | |
881 | if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1)) | |
e9e9b7ec | 882 | enable_vma_readahead = true; |
d9bfcfdc | 883 | else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1)) |
e9e9b7ec | 884 | enable_vma_readahead = false; |
d9bfcfdc HY |
885 | else |
886 | return -EINVAL; | |
887 | ||
888 | return count; | |
889 | } | |
890 | static struct kobj_attribute vma_ra_enabled_attr = | |
891 | __ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show, | |
892 | vma_ra_enabled_store); | |
893 | ||
d9bfcfdc HY |
894 | static struct attribute *swap_attrs[] = { |
895 | &vma_ra_enabled_attr.attr, | |
d9bfcfdc HY |
896 | NULL, |
897 | }; | |
898 | ||
e48333b6 | 899 | static const struct attribute_group swap_attr_group = { |
d9bfcfdc HY |
900 | .attrs = swap_attrs, |
901 | }; | |
902 | ||
903 | static int __init swap_init_sysfs(void) | |
904 | { | |
905 | int err; | |
906 | struct kobject *swap_kobj; | |
907 | ||
908 | swap_kobj = kobject_create_and_add("swap", mm_kobj); | |
909 | if (!swap_kobj) { | |
910 | pr_err("failed to create swap kobject\n"); | |
911 | return -ENOMEM; | |
912 | } | |
913 | err = sysfs_create_group(swap_kobj, &swap_attr_group); | |
914 | if (err) { | |
915 | pr_err("failed to register swap group\n"); | |
916 | goto delete_obj; | |
917 | } | |
918 | return 0; | |
919 | ||
920 | delete_obj: | |
921 | kobject_put(swap_kobj); | |
922 | return err; | |
923 | } | |
924 | subsys_initcall(swap_init_sysfs); | |
925 | #endif |