]>
Commit | Line | Data |
---|---|---|
1 | // SPDX-License-Identifier: GPL-2.0 | |
2 | /* | |
3 | * linux/mm/swap_state.c | |
4 | * | |
5 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
6 | * Swap reorganised 29.12.95, Stephen Tweedie | |
7 | * | |
8 | * Rewritten to use page cache, (C) 1998 Stephen Tweedie | |
9 | */ | |
10 | #include <linux/mm.h> | |
11 | #include <linux/gfp.h> | |
12 | #include <linux/kernel_stat.h> | |
13 | #include <linux/swap.h> | |
14 | #include <linux/swapops.h> | |
15 | #include <linux/init.h> | |
16 | #include <linux/pagemap.h> | |
17 | #include <linux/backing-dev.h> | |
18 | #include <linux/blkdev.h> | |
19 | #include <linux/migrate.h> | |
20 | #include <linux/vmalloc.h> | |
21 | #include <linux/swap_slots.h> | |
22 | #include <linux/huge_mm.h> | |
23 | #include <linux/shmem_fs.h> | |
24 | #include "internal.h" | |
25 | #include "swap.h" | |
26 | ||
27 | /* | |
28 | * swapper_space is a fiction, retained to simplify the path through | |
29 | * vmscan's shrink_page_list. | |
30 | */ | |
31 | static const struct address_space_operations swap_aops = { | |
32 | .writepage = swap_writepage, | |
33 | .dirty_folio = noop_dirty_folio, | |
34 | #ifdef CONFIG_MIGRATION | |
35 | .migrate_folio = migrate_folio, | |
36 | #endif | |
37 | }; | |
38 | ||
39 | struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly; | |
40 | static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly; | |
41 | static bool enable_vma_readahead __read_mostly = true; | |
42 | ||
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) | |
60 | ||
61 | static atomic_t swapin_readahead_hits = ATOMIC_INIT(4); | |
62 | ||
63 | void show_swap_cache_info(void) | |
64 | { | |
65 | printk("%lu pages in swap cache\n", total_swapcache_pages()); | |
66 | printk("Free swap = %ldkB\n", K(get_nr_swap_pages())); | |
67 | printk("Total swap = %lukB\n", K(total_swap_pages)); | |
68 | } | |
69 | ||
70 | void *get_shadow_from_swap_cache(swp_entry_t entry) | |
71 | { | |
72 | struct address_space *address_space = swap_address_space(entry); | |
73 | pgoff_t idx = swp_offset(entry); | |
74 | struct page *page; | |
75 | ||
76 | page = xa_load(&address_space->i_pages, idx); | |
77 | if (xa_is_value(page)) | |
78 | return page; | |
79 | return NULL; | |
80 | } | |
81 | ||
82 | /* | |
83 | * add_to_swap_cache resembles filemap_add_folio on swapper_space, | |
84 | * but sets SwapCache flag and private instead of mapping and index. | |
85 | */ | |
86 | int add_to_swap_cache(struct folio *folio, swp_entry_t entry, | |
87 | gfp_t gfp, void **shadowp) | |
88 | { | |
89 | struct address_space *address_space = swap_address_space(entry); | |
90 | pgoff_t idx = swp_offset(entry); | |
91 | XA_STATE_ORDER(xas, &address_space->i_pages, idx, folio_order(folio)); | |
92 | unsigned long i, nr = folio_nr_pages(folio); | |
93 | void *old; | |
94 | ||
95 | xas_set_update(&xas, workingset_update_node); | |
96 | ||
97 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); | |
98 | VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); | |
99 | VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio); | |
100 | ||
101 | folio_ref_add(folio, nr); | |
102 | folio_set_swapcache(folio); | |
103 | folio->swap = entry; | |
104 | ||
105 | do { | |
106 | xas_lock_irq(&xas); | |
107 | xas_create_range(&xas); | |
108 | if (xas_error(&xas)) | |
109 | goto unlock; | |
110 | for (i = 0; i < nr; i++) { | |
111 | VM_BUG_ON_FOLIO(xas.xa_index != idx + i, folio); | |
112 | old = xas_load(&xas); | |
113 | if (xa_is_value(old)) { | |
114 | if (shadowp) | |
115 | *shadowp = old; | |
116 | } | |
117 | xas_store(&xas, folio); | |
118 | xas_next(&xas); | |
119 | } | |
120 | address_space->nrpages += nr; | |
121 | __node_stat_mod_folio(folio, NR_FILE_PAGES, nr); | |
122 | __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, nr); | |
123 | unlock: | |
124 | xas_unlock_irq(&xas); | |
125 | } while (xas_nomem(&xas, gfp)); | |
126 | ||
127 | if (!xas_error(&xas)) | |
128 | return 0; | |
129 | ||
130 | folio_clear_swapcache(folio); | |
131 | folio_ref_sub(folio, nr); | |
132 | return xas_error(&xas); | |
133 | } | |
134 | ||
135 | /* | |
136 | * This must be called only on folios that have | |
137 | * been verified to be in the swap cache. | |
138 | */ | |
139 | void __delete_from_swap_cache(struct folio *folio, | |
140 | swp_entry_t entry, void *shadow) | |
141 | { | |
142 | struct address_space *address_space = swap_address_space(entry); | |
143 | int i; | |
144 | long nr = folio_nr_pages(folio); | |
145 | pgoff_t idx = swp_offset(entry); | |
146 | XA_STATE(xas, &address_space->i_pages, idx); | |
147 | ||
148 | xas_set_update(&xas, workingset_update_node); | |
149 | ||
150 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); | |
151 | VM_BUG_ON_FOLIO(!folio_test_swapcache(folio), folio); | |
152 | VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio); | |
153 | ||
154 | for (i = 0; i < nr; i++) { | |
155 | void *entry = xas_store(&xas, shadow); | |
156 | VM_BUG_ON_PAGE(entry != folio, entry); | |
157 | xas_next(&xas); | |
158 | } | |
159 | folio->swap.val = 0; | |
160 | folio_clear_swapcache(folio); | |
161 | address_space->nrpages -= nr; | |
162 | __node_stat_mod_folio(folio, NR_FILE_PAGES, -nr); | |
163 | __lruvec_stat_mod_folio(folio, NR_SWAPCACHE, -nr); | |
164 | } | |
165 | ||
166 | /** | |
167 | * add_to_swap - allocate swap space for a folio | |
168 | * @folio: folio we want to move to swap | |
169 | * | |
170 | * Allocate swap space for the folio and add the folio to the | |
171 | * swap cache. | |
172 | * | |
173 | * Context: Caller needs to hold the folio lock. | |
174 | * Return: Whether the folio was added to the swap cache. | |
175 | */ | |
176 | bool add_to_swap(struct folio *folio) | |
177 | { | |
178 | swp_entry_t entry; | |
179 | int err; | |
180 | ||
181 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); | |
182 | VM_BUG_ON_FOLIO(!folio_test_uptodate(folio), folio); | |
183 | ||
184 | entry = folio_alloc_swap(folio); | |
185 | if (!entry.val) | |
186 | return false; | |
187 | ||
188 | /* | |
189 | * XArray node allocations from PF_MEMALLOC contexts could | |
190 | * completely exhaust the page allocator. __GFP_NOMEMALLOC | |
191 | * stops emergency reserves from being allocated. | |
192 | * | |
193 | * TODO: this could cause a theoretical memory reclaim | |
194 | * deadlock in the swap out path. | |
195 | */ | |
196 | /* | |
197 | * Add it to the swap cache. | |
198 | */ | |
199 | err = add_to_swap_cache(folio, entry, | |
200 | __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN, NULL); | |
201 | if (err) | |
202 | /* | |
203 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely | |
204 | * clear SWAP_HAS_CACHE flag. | |
205 | */ | |
206 | goto fail; | |
207 | /* | |
208 | * Normally the folio will be dirtied in unmap because its | |
209 | * pte should be dirty. A special case is MADV_FREE page. The | |
210 | * page's pte could have dirty bit cleared but the folio's | |
211 | * SwapBacked flag is still set because clearing the dirty bit | |
212 | * and SwapBacked flag has no lock protected. For such folio, | |
213 | * unmap will not set dirty bit for it, so folio reclaim will | |
214 | * not write the folio out. This can cause data corruption when | |
215 | * the folio is swapped in later. Always setting the dirty flag | |
216 | * for the folio solves the problem. | |
217 | */ | |
218 | folio_mark_dirty(folio); | |
219 | ||
220 | return true; | |
221 | ||
222 | fail: | |
223 | put_swap_folio(folio, entry); | |
224 | return false; | |
225 | } | |
226 | ||
227 | /* | |
228 | * This must be called only on folios that have | |
229 | * been verified to be in the swap cache and locked. | |
230 | * It will never put the folio into the free list, | |
231 | * the caller has a reference on the folio. | |
232 | */ | |
233 | void delete_from_swap_cache(struct folio *folio) | |
234 | { | |
235 | swp_entry_t entry = folio->swap; | |
236 | struct address_space *address_space = swap_address_space(entry); | |
237 | ||
238 | xa_lock_irq(&address_space->i_pages); | |
239 | __delete_from_swap_cache(folio, entry, NULL); | |
240 | xa_unlock_irq(&address_space->i_pages); | |
241 | ||
242 | put_swap_folio(folio, entry); | |
243 | folio_ref_sub(folio, folio_nr_pages(folio)); | |
244 | } | |
245 | ||
246 | void clear_shadow_from_swap_cache(int type, unsigned long begin, | |
247 | unsigned long end) | |
248 | { | |
249 | unsigned long curr = begin; | |
250 | void *old; | |
251 | ||
252 | for (;;) { | |
253 | swp_entry_t entry = swp_entry(type, curr); | |
254 | struct address_space *address_space = swap_address_space(entry); | |
255 | XA_STATE(xas, &address_space->i_pages, curr); | |
256 | ||
257 | xas_set_update(&xas, workingset_update_node); | |
258 | ||
259 | xa_lock_irq(&address_space->i_pages); | |
260 | xas_for_each(&xas, old, end) { | |
261 | if (!xa_is_value(old)) | |
262 | continue; | |
263 | xas_store(&xas, NULL); | |
264 | } | |
265 | xa_unlock_irq(&address_space->i_pages); | |
266 | ||
267 | /* search the next swapcache until we meet end */ | |
268 | curr >>= SWAP_ADDRESS_SPACE_SHIFT; | |
269 | curr++; | |
270 | curr <<= SWAP_ADDRESS_SPACE_SHIFT; | |
271 | if (curr > end) | |
272 | break; | |
273 | } | |
274 | } | |
275 | ||
276 | /* | |
277 | * If we are the only user, then try to free up the swap cache. | |
278 | * | |
279 | * Its ok to check the swapcache flag without the folio lock | |
280 | * here because we are going to recheck again inside | |
281 | * folio_free_swap() _with_ the lock. | |
282 | * - Marcelo | |
283 | */ | |
284 | void free_swap_cache(struct page *page) | |
285 | { | |
286 | struct folio *folio = page_folio(page); | |
287 | ||
288 | if (folio_test_swapcache(folio) && !folio_mapped(folio) && | |
289 | folio_trylock(folio)) { | |
290 | folio_free_swap(folio); | |
291 | folio_unlock(folio); | |
292 | } | |
293 | } | |
294 | ||
295 | /* | |
296 | * Perform a free_page(), also freeing any swap cache associated with | |
297 | * this page if it is the last user of the page. | |
298 | */ | |
299 | void free_page_and_swap_cache(struct page *page) | |
300 | { | |
301 | free_swap_cache(page); | |
302 | if (!is_huge_zero_page(page)) | |
303 | put_page(page); | |
304 | } | |
305 | ||
306 | /* | |
307 | * Passed an array of pages, drop them all from swapcache and then release | |
308 | * them. They are removed from the LRU and freed if this is their last use. | |
309 | */ | |
310 | void free_pages_and_swap_cache(struct encoded_page **pages, int nr) | |
311 | { | |
312 | lru_add_drain(); | |
313 | for (int i = 0; i < nr; i++) | |
314 | free_swap_cache(encoded_page_ptr(pages[i])); | |
315 | release_pages(pages, nr); | |
316 | } | |
317 | ||
318 | static inline bool swap_use_vma_readahead(void) | |
319 | { | |
320 | return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap); | |
321 | } | |
322 | ||
323 | /* | |
324 | * Lookup a swap entry in the swap cache. A found folio will be returned | |
325 | * unlocked and with its refcount incremented - we rely on the kernel | |
326 | * lock getting page table operations atomic even if we drop the folio | |
327 | * lock before returning. | |
328 | * | |
329 | * Caller must lock the swap device or hold a reference to keep it valid. | |
330 | */ | |
331 | struct folio *swap_cache_get_folio(swp_entry_t entry, | |
332 | struct vm_area_struct *vma, unsigned long addr) | |
333 | { | |
334 | struct folio *folio; | |
335 | ||
336 | folio = filemap_get_folio(swap_address_space(entry), swp_offset(entry)); | |
337 | if (!IS_ERR(folio)) { | |
338 | bool vma_ra = swap_use_vma_readahead(); | |
339 | bool readahead; | |
340 | ||
341 | /* | |
342 | * At the moment, we don't support PG_readahead for anon THP | |
343 | * so let's bail out rather than confusing the readahead stat. | |
344 | */ | |
345 | if (unlikely(folio_test_large(folio))) | |
346 | return folio; | |
347 | ||
348 | readahead = folio_test_clear_readahead(folio); | |
349 | if (vma && vma_ra) { | |
350 | unsigned long ra_val; | |
351 | int win, hits; | |
352 | ||
353 | ra_val = GET_SWAP_RA_VAL(vma); | |
354 | win = SWAP_RA_WIN(ra_val); | |
355 | hits = SWAP_RA_HITS(ra_val); | |
356 | if (readahead) | |
357 | hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX); | |
358 | atomic_long_set(&vma->swap_readahead_info, | |
359 | SWAP_RA_VAL(addr, win, hits)); | |
360 | } | |
361 | ||
362 | if (readahead) { | |
363 | count_vm_event(SWAP_RA_HIT); | |
364 | if (!vma || !vma_ra) | |
365 | atomic_inc(&swapin_readahead_hits); | |
366 | } | |
367 | } else { | |
368 | folio = NULL; | |
369 | } | |
370 | ||
371 | return folio; | |
372 | } | |
373 | ||
374 | /** | |
375 | * filemap_get_incore_folio - Find and get a folio from the page or swap caches. | |
376 | * @mapping: The address_space to search. | |
377 | * @index: The page cache index. | |
378 | * | |
379 | * This differs from filemap_get_folio() in that it will also look for the | |
380 | * folio in the swap cache. | |
381 | * | |
382 | * Return: The found folio or %NULL. | |
383 | */ | |
384 | struct folio *filemap_get_incore_folio(struct address_space *mapping, | |
385 | pgoff_t index) | |
386 | { | |
387 | swp_entry_t swp; | |
388 | struct swap_info_struct *si; | |
389 | struct folio *folio = filemap_get_entry(mapping, index); | |
390 | ||
391 | if (!folio) | |
392 | return ERR_PTR(-ENOENT); | |
393 | if (!xa_is_value(folio)) | |
394 | return folio; | |
395 | if (!shmem_mapping(mapping)) | |
396 | return ERR_PTR(-ENOENT); | |
397 | ||
398 | swp = radix_to_swp_entry(folio); | |
399 | /* There might be swapin error entries in shmem mapping. */ | |
400 | if (non_swap_entry(swp)) | |
401 | return ERR_PTR(-ENOENT); | |
402 | /* Prevent swapoff from happening to us */ | |
403 | si = get_swap_device(swp); | |
404 | if (!si) | |
405 | return ERR_PTR(-ENOENT); | |
406 | index = swp_offset(swp); | |
407 | folio = filemap_get_folio(swap_address_space(swp), index); | |
408 | put_swap_device(si); | |
409 | return folio; | |
410 | } | |
411 | ||
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) | |
415 | { | |
416 | struct swap_info_struct *si; | |
417 | struct folio *folio; | |
418 | struct page *page; | |
419 | void *shadow = NULL; | |
420 | ||
421 | *new_page_allocated = false; | |
422 | si = get_swap_device(entry); | |
423 | if (!si) | |
424 | return NULL; | |
425 | ||
426 | for (;;) { | |
427 | int err; | |
428 | /* | |
429 | * First check the swap cache. Since this is normally | |
430 | * called after swap_cache_get_folio() failed, re-calling | |
431 | * that would confuse statistics. | |
432 | */ | |
433 | folio = filemap_get_folio(swap_address_space(entry), | |
434 | swp_offset(entry)); | |
435 | if (!IS_ERR(folio)) { | |
436 | page = folio_file_page(folio, swp_offset(entry)); | |
437 | goto got_page; | |
438 | } | |
439 | ||
440 | /* | |
441 | * Just skip read ahead for unused swap slot. | |
442 | * During swap_off when swap_slot_cache is disabled, | |
443 | * we have to handle the race between putting | |
444 | * swap entry in swap cache and marking swap slot | |
445 | * as SWAP_HAS_CACHE. That's done in later part of code or | |
446 | * else swap_off will be aborted if we return NULL. | |
447 | */ | |
448 | if (!swap_swapcount(si, entry) && swap_slot_cache_enabled) | |
449 | goto fail_put_swap; | |
450 | ||
451 | /* | |
452 | * Get a new page to read into from swap. Allocate it now, | |
453 | * before marking swap_map SWAP_HAS_CACHE, when -EEXIST will | |
454 | * cause any racers to loop around until we add it to cache. | |
455 | */ | |
456 | folio = vma_alloc_folio(gfp_mask, 0, vma, addr, false); | |
457 | if (!folio) | |
458 | goto fail_put_swap; | |
459 | ||
460 | /* | |
461 | * Swap entry may have been freed since our caller observed it. | |
462 | */ | |
463 | err = swapcache_prepare(entry); | |
464 | if (!err) | |
465 | break; | |
466 | ||
467 | folio_put(folio); | |
468 | if (err != -EEXIST) | |
469 | goto fail_put_swap; | |
470 | ||
471 | /* | |
472 | * We might race against __delete_from_swap_cache(), and | |
473 | * stumble across a swap_map entry whose SWAP_HAS_CACHE | |
474 | * has not yet been cleared. Or race against another | |
475 | * __read_swap_cache_async(), which has set SWAP_HAS_CACHE | |
476 | * in swap_map, but not yet added its page to swap cache. | |
477 | */ | |
478 | schedule_timeout_uninterruptible(1); | |
479 | } | |
480 | ||
481 | /* | |
482 | * The swap entry is ours to swap in. Prepare the new page. | |
483 | */ | |
484 | ||
485 | __folio_set_locked(folio); | |
486 | __folio_set_swapbacked(folio); | |
487 | ||
488 | if (mem_cgroup_swapin_charge_folio(folio, NULL, gfp_mask, entry)) | |
489 | goto fail_unlock; | |
490 | ||
491 | /* May fail (-ENOMEM) if XArray node allocation failed. */ | |
492 | if (add_to_swap_cache(folio, entry, gfp_mask & GFP_RECLAIM_MASK, &shadow)) | |
493 | goto fail_unlock; | |
494 | ||
495 | mem_cgroup_swapin_uncharge_swap(entry); | |
496 | ||
497 | if (shadow) | |
498 | workingset_refault(folio, shadow); | |
499 | ||
500 | /* Caller will initiate read into locked folio */ | |
501 | folio_add_lru(folio); | |
502 | *new_page_allocated = true; | |
503 | page = &folio->page; | |
504 | got_page: | |
505 | put_swap_device(si); | |
506 | return page; | |
507 | ||
508 | fail_unlock: | |
509 | put_swap_folio(folio, entry); | |
510 | folio_unlock(folio); | |
511 | folio_put(folio); | |
512 | fail_put_swap: | |
513 | put_swap_device(si); | |
514 | return NULL; | |
515 | } | |
516 | ||
517 | /* | |
518 | * Locate a page of swap in physical memory, reserving swap cache space | |
519 | * and reading the disk if it is not already cached. | |
520 | * A failure return means that either the page allocation failed or that | |
521 | * the swap entry is no longer in use. | |
522 | * | |
523 | * get/put_swap_device() aren't needed to call this function, because | |
524 | * __read_swap_cache_async() call them and swap_readpage() holds the | |
525 | * swap cache folio lock. | |
526 | */ | |
527 | struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, | |
528 | struct vm_area_struct *vma, | |
529 | unsigned long addr, struct swap_iocb **plug) | |
530 | { | |
531 | bool page_was_allocated; | |
532 | struct page *retpage = __read_swap_cache_async(entry, gfp_mask, | |
533 | vma, addr, &page_was_allocated); | |
534 | ||
535 | if (page_was_allocated) | |
536 | swap_readpage(retpage, false, plug); | |
537 | ||
538 | return retpage; | |
539 | } | |
540 | ||
541 | static unsigned int __swapin_nr_pages(unsigned long prev_offset, | |
542 | unsigned long offset, | |
543 | int hits, | |
544 | int max_pages, | |
545 | int prev_win) | |
546 | { | |
547 | unsigned int pages, last_ra; | |
548 | ||
549 | /* | |
550 | * This heuristic has been found to work well on both sequential and | |
551 | * random loads, swapping to hard disk or to SSD: please don't ask | |
552 | * what the "+ 2" means, it just happens to work well, that's all. | |
553 | */ | |
554 | pages = hits + 2; | |
555 | if (pages == 2) { | |
556 | /* | |
557 | * We can have no readahead hits to judge by: but must not get | |
558 | * stuck here forever, so check for an adjacent offset instead | |
559 | * (and don't even bother to check whether swap type is same). | |
560 | */ | |
561 | if (offset != prev_offset + 1 && offset != prev_offset - 1) | |
562 | pages = 1; | |
563 | } else { | |
564 | unsigned int roundup = 4; | |
565 | while (roundup < pages) | |
566 | roundup <<= 1; | |
567 | pages = roundup; | |
568 | } | |
569 | ||
570 | if (pages > max_pages) | |
571 | pages = max_pages; | |
572 | ||
573 | /* Don't shrink readahead too fast */ | |
574 | last_ra = prev_win / 2; | |
575 | if (pages < last_ra) | |
576 | pages = last_ra; | |
577 | ||
578 | return pages; | |
579 | } | |
580 | ||
581 | static unsigned long swapin_nr_pages(unsigned long offset) | |
582 | { | |
583 | static unsigned long prev_offset; | |
584 | unsigned int hits, pages, max_pages; | |
585 | static atomic_t last_readahead_pages; | |
586 | ||
587 | max_pages = 1 << READ_ONCE(page_cluster); | |
588 | if (max_pages <= 1) | |
589 | return 1; | |
590 | ||
591 | hits = atomic_xchg(&swapin_readahead_hits, 0); | |
592 | pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits, | |
593 | max_pages, | |
594 | atomic_read(&last_readahead_pages)); | |
595 | if (!hits) | |
596 | WRITE_ONCE(prev_offset, offset); | |
597 | atomic_set(&last_readahead_pages, pages); | |
598 | ||
599 | return pages; | |
600 | } | |
601 | ||
602 | /** | |
603 | * swap_cluster_readahead - swap in pages in hope we need them soon | |
604 | * @entry: swap entry of this memory | |
605 | * @gfp_mask: memory allocation flags | |
606 | * @vmf: fault information | |
607 | * | |
608 | * Returns the struct page for entry and addr, after queueing swapin. | |
609 | * | |
610 | * Primitive swap readahead code. We simply read an aligned block of | |
611 | * (1 << page_cluster) entries in the swap area. This method is chosen | |
612 | * because it doesn't cost us any seek time. We also make sure to queue | |
613 | * the 'original' request together with the readahead ones... | |
614 | * | |
615 | * This has been extended to use the NUMA policies from the mm triggering | |
616 | * the readahead. | |
617 | * | |
618 | * Caller must hold read mmap_lock if vmf->vma is not NULL. | |
619 | */ | |
620 | struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask, | |
621 | struct vm_fault *vmf) | |
622 | { | |
623 | struct page *page; | |
624 | unsigned long entry_offset = swp_offset(entry); | |
625 | unsigned long offset = entry_offset; | |
626 | unsigned long start_offset, end_offset; | |
627 | unsigned long mask; | |
628 | struct swap_info_struct *si = swp_swap_info(entry); | |
629 | struct blk_plug plug; | |
630 | struct swap_iocb *splug = NULL; | |
631 | bool page_allocated; | |
632 | struct vm_area_struct *vma = vmf->vma; | |
633 | unsigned long addr = vmf->address; | |
634 | ||
635 | mask = swapin_nr_pages(offset) - 1; | |
636 | if (!mask) | |
637 | goto skip; | |
638 | ||
639 | /* Read a page_cluster sized and aligned cluster around offset. */ | |
640 | start_offset = offset & ~mask; | |
641 | end_offset = offset | mask; | |
642 | if (!start_offset) /* First page is swap header. */ | |
643 | start_offset++; | |
644 | if (end_offset >= si->max) | |
645 | end_offset = si->max - 1; | |
646 | ||
647 | blk_start_plug(&plug); | |
648 | for (offset = start_offset; offset <= end_offset ; offset++) { | |
649 | /* Ok, do the async read-ahead now */ | |
650 | page = __read_swap_cache_async( | |
651 | swp_entry(swp_type(entry), offset), | |
652 | gfp_mask, vma, addr, &page_allocated); | |
653 | if (!page) | |
654 | continue; | |
655 | if (page_allocated) { | |
656 | swap_readpage(page, false, &splug); | |
657 | if (offset != entry_offset) { | |
658 | SetPageReadahead(page); | |
659 | count_vm_event(SWAP_RA); | |
660 | } | |
661 | } | |
662 | put_page(page); | |
663 | } | |
664 | blk_finish_plug(&plug); | |
665 | swap_read_unplug(splug); | |
666 | ||
667 | lru_add_drain(); /* Push any new pages onto the LRU now */ | |
668 | skip: | |
669 | /* The page was likely read above, so no need for plugging here */ | |
670 | return read_swap_cache_async(entry, gfp_mask, vma, addr, NULL); | |
671 | } | |
672 | ||
673 | int init_swap_address_space(unsigned int type, unsigned long nr_pages) | |
674 | { | |
675 | struct address_space *spaces, *space; | |
676 | unsigned int i, nr; | |
677 | ||
678 | nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES); | |
679 | spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL); | |
680 | if (!spaces) | |
681 | return -ENOMEM; | |
682 | for (i = 0; i < nr; i++) { | |
683 | space = spaces + i; | |
684 | xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ); | |
685 | atomic_set(&space->i_mmap_writable, 0); | |
686 | space->a_ops = &swap_aops; | |
687 | /* swap cache doesn't use writeback related tags */ | |
688 | mapping_set_no_writeback_tags(space); | |
689 | } | |
690 | nr_swapper_spaces[type] = nr; | |
691 | swapper_spaces[type] = spaces; | |
692 | ||
693 | return 0; | |
694 | } | |
695 | ||
696 | void exit_swap_address_space(unsigned int type) | |
697 | { | |
698 | int i; | |
699 | struct address_space *spaces = swapper_spaces[type]; | |
700 | ||
701 | for (i = 0; i < nr_swapper_spaces[type]; i++) | |
702 | VM_WARN_ON_ONCE(!mapping_empty(&spaces[i])); | |
703 | kvfree(spaces); | |
704 | nr_swapper_spaces[type] = 0; | |
705 | swapper_spaces[type] = NULL; | |
706 | } | |
707 | ||
708 | #define SWAP_RA_ORDER_CEILING 5 | |
709 | ||
710 | struct vma_swap_readahead { | |
711 | unsigned short win; | |
712 | unsigned short offset; | |
713 | unsigned short nr_pte; | |
714 | }; | |
715 | ||
716 | static void swap_ra_info(struct vm_fault *vmf, | |
717 | struct vma_swap_readahead *ra_info) | |
718 | { | |
719 | struct vm_area_struct *vma = vmf->vma; | |
720 | unsigned long ra_val; | |
721 | unsigned long faddr, pfn, fpfn, lpfn, rpfn; | |
722 | unsigned long start, end; | |
723 | unsigned int max_win, hits, prev_win, win; | |
724 | ||
725 | max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster), | |
726 | SWAP_RA_ORDER_CEILING); | |
727 | if (max_win == 1) { | |
728 | ra_info->win = 1; | |
729 | return; | |
730 | } | |
731 | ||
732 | faddr = vmf->address; | |
733 | fpfn = PFN_DOWN(faddr); | |
734 | ra_val = GET_SWAP_RA_VAL(vma); | |
735 | pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val)); | |
736 | prev_win = SWAP_RA_WIN(ra_val); | |
737 | hits = SWAP_RA_HITS(ra_val); | |
738 | ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits, | |
739 | max_win, prev_win); | |
740 | atomic_long_set(&vma->swap_readahead_info, | |
741 | SWAP_RA_VAL(faddr, win, 0)); | |
742 | if (win == 1) | |
743 | return; | |
744 | ||
745 | if (fpfn == pfn + 1) { | |
746 | lpfn = fpfn; | |
747 | rpfn = fpfn + win; | |
748 | } else if (pfn == fpfn + 1) { | |
749 | lpfn = fpfn - win + 1; | |
750 | rpfn = fpfn + 1; | |
751 | } else { | |
752 | unsigned int left = (win - 1) / 2; | |
753 | ||
754 | lpfn = fpfn - left; | |
755 | rpfn = fpfn + win - left; | |
756 | } | |
757 | start = max3(lpfn, PFN_DOWN(vma->vm_start), | |
758 | PFN_DOWN(faddr & PMD_MASK)); | |
759 | end = min3(rpfn, PFN_DOWN(vma->vm_end), | |
760 | PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE)); | |
761 | ||
762 | ra_info->nr_pte = end - start; | |
763 | ra_info->offset = fpfn - start; | |
764 | } | |
765 | ||
766 | /** | |
767 | * swap_vma_readahead - swap in pages in hope we need them soon | |
768 | * @fentry: swap entry of this memory | |
769 | * @gfp_mask: memory allocation flags | |
770 | * @vmf: fault information | |
771 | * | |
772 | * Returns the struct page for entry and addr, after queueing swapin. | |
773 | * | |
774 | * Primitive swap readahead code. We simply read in a few pages whose | |
775 | * virtual addresses are around the fault address in the same vma. | |
776 | * | |
777 | * Caller must hold read mmap_lock if vmf->vma is not NULL. | |
778 | * | |
779 | */ | |
780 | static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask, | |
781 | struct vm_fault *vmf) | |
782 | { | |
783 | struct blk_plug plug; | |
784 | struct swap_iocb *splug = NULL; | |
785 | struct vm_area_struct *vma = vmf->vma; | |
786 | struct page *page; | |
787 | pte_t *pte = NULL, pentry; | |
788 | unsigned long addr; | |
789 | swp_entry_t entry; | |
790 | unsigned int i; | |
791 | bool page_allocated; | |
792 | struct vma_swap_readahead ra_info = { | |
793 | .win = 1, | |
794 | }; | |
795 | ||
796 | swap_ra_info(vmf, &ra_info); | |
797 | if (ra_info.win == 1) | |
798 | goto skip; | |
799 | ||
800 | addr = vmf->address - (ra_info.offset * PAGE_SIZE); | |
801 | ||
802 | blk_start_plug(&plug); | |
803 | for (i = 0; i < ra_info.nr_pte; i++, addr += PAGE_SIZE) { | |
804 | if (!pte++) { | |
805 | pte = pte_offset_map(vmf->pmd, addr); | |
806 | if (!pte) | |
807 | break; | |
808 | } | |
809 | pentry = ptep_get_lockless(pte); | |
810 | if (!is_swap_pte(pentry)) | |
811 | continue; | |
812 | entry = pte_to_swp_entry(pentry); | |
813 | if (unlikely(non_swap_entry(entry))) | |
814 | continue; | |
815 | pte_unmap(pte); | |
816 | pte = NULL; | |
817 | page = __read_swap_cache_async(entry, gfp_mask, vma, | |
818 | addr, &page_allocated); | |
819 | if (!page) | |
820 | continue; | |
821 | if (page_allocated) { | |
822 | swap_readpage(page, false, &splug); | |
823 | if (i != ra_info.offset) { | |
824 | SetPageReadahead(page); | |
825 | count_vm_event(SWAP_RA); | |
826 | } | |
827 | } | |
828 | put_page(page); | |
829 | } | |
830 | if (pte) | |
831 | pte_unmap(pte); | |
832 | blk_finish_plug(&plug); | |
833 | swap_read_unplug(splug); | |
834 | lru_add_drain(); | |
835 | skip: | |
836 | /* The page was likely read above, so no need for plugging here */ | |
837 | return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, | |
838 | NULL); | |
839 | } | |
840 | ||
841 | /** | |
842 | * swapin_readahead - swap in pages in hope we need them soon | |
843 | * @entry: swap entry of this memory | |
844 | * @gfp_mask: memory allocation flags | |
845 | * @vmf: fault information | |
846 | * | |
847 | * Returns the struct page for entry and addr, after queueing swapin. | |
848 | * | |
849 | * It's a main entry function for swap readahead. By the configuration, | |
850 | * it will read ahead blocks by cluster-based(ie, physical disk based) | |
851 | * or vma-based(ie, virtual address based on faulty address) readahead. | |
852 | */ | |
853 | struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, | |
854 | struct vm_fault *vmf) | |
855 | { | |
856 | return swap_use_vma_readahead() ? | |
857 | swap_vma_readahead(entry, gfp_mask, vmf) : | |
858 | swap_cluster_readahead(entry, gfp_mask, vmf); | |
859 | } | |
860 | ||
861 | #ifdef CONFIG_SYSFS | |
862 | static ssize_t vma_ra_enabled_show(struct kobject *kobj, | |
863 | struct kobj_attribute *attr, char *buf) | |
864 | { | |
865 | return sysfs_emit(buf, "%s\n", | |
866 | enable_vma_readahead ? "true" : "false"); | |
867 | } | |
868 | static ssize_t vma_ra_enabled_store(struct kobject *kobj, | |
869 | struct kobj_attribute *attr, | |
870 | const char *buf, size_t count) | |
871 | { | |
872 | ssize_t ret; | |
873 | ||
874 | ret = kstrtobool(buf, &enable_vma_readahead); | |
875 | if (ret) | |
876 | return ret; | |
877 | ||
878 | return count; | |
879 | } | |
880 | static struct kobj_attribute vma_ra_enabled_attr = __ATTR_RW(vma_ra_enabled); | |
881 | ||
882 | static struct attribute *swap_attrs[] = { | |
883 | &vma_ra_enabled_attr.attr, | |
884 | NULL, | |
885 | }; | |
886 | ||
887 | static const struct attribute_group swap_attr_group = { | |
888 | .attrs = swap_attrs, | |
889 | }; | |
890 | ||
891 | static int __init swap_init_sysfs(void) | |
892 | { | |
893 | int err; | |
894 | struct kobject *swap_kobj; | |
895 | ||
896 | swap_kobj = kobject_create_and_add("swap", mm_kobj); | |
897 | if (!swap_kobj) { | |
898 | pr_err("failed to create swap kobject\n"); | |
899 | return -ENOMEM; | |
900 | } | |
901 | err = sysfs_create_group(swap_kobj, &swap_attr_group); | |
902 | if (err) { | |
903 | pr_err("failed to register swap group\n"); | |
904 | goto delete_obj; | |
905 | } | |
906 | return 0; | |
907 | ||
908 | delete_obj: | |
909 | kobject_put(swap_kobj); | |
910 | return err; | |
911 | } | |
912 | subsys_initcall(swap_init_sysfs); | |
913 | #endif |