]>
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> |
1da177e4 | 24 | |
1da177e4 LT |
25 | |
26 | /* | |
27 | * swapper_space is a fiction, retained to simplify the path through | |
7eaceacc | 28 | * vmscan's shrink_page_list. |
1da177e4 | 29 | */ |
f5e54d6e | 30 | static const struct address_space_operations swap_aops = { |
1da177e4 | 31 | .writepage = swap_writepage, |
62c230bc | 32 | .set_page_dirty = swap_set_page_dirty, |
1c93923c | 33 | #ifdef CONFIG_MIGRATION |
e965f963 | 34 | .migratepage = migrate_page, |
1c93923c | 35 | #endif |
1da177e4 LT |
36 | }; |
37 | ||
783cb68e CD |
38 | struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly; |
39 | static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly; | |
f5c754d6 | 40 | static bool enable_vma_readahead __read_mostly = true; |
ec560175 | 41 | |
ec560175 HY |
42 | #define SWAP_RA_WIN_SHIFT (PAGE_SHIFT / 2) |
43 | #define SWAP_RA_HITS_MASK ((1UL << SWAP_RA_WIN_SHIFT) - 1) | |
44 | #define SWAP_RA_HITS_MAX SWAP_RA_HITS_MASK | |
45 | #define SWAP_RA_WIN_MASK (~PAGE_MASK & ~SWAP_RA_HITS_MASK) | |
46 | ||
47 | #define SWAP_RA_HITS(v) ((v) & SWAP_RA_HITS_MASK) | |
48 | #define SWAP_RA_WIN(v) (((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT) | |
49 | #define SWAP_RA_ADDR(v) ((v) & PAGE_MASK) | |
50 | ||
51 | #define SWAP_RA_VAL(addr, win, hits) \ | |
52 | (((addr) & PAGE_MASK) | \ | |
53 | (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) | \ | |
54 | ((hits) & SWAP_RA_HITS_MASK)) | |
55 | ||
56 | /* Initial readahead hits is 4 to start up with a small window */ | |
57 | #define GET_SWAP_RA_VAL(vma) \ | |
58 | (atomic_long_read(&(vma)->swap_readahead_info) ? : 4) | |
1da177e4 LT |
59 | |
60 | #define INC_CACHE_INFO(x) do { swap_cache_info.x++; } while (0) | |
38d8b4e6 | 61 | #define ADD_CACHE_INFO(x, nr) do { swap_cache_info.x += (nr); } while (0) |
1da177e4 LT |
62 | |
63 | static struct { | |
64 | unsigned long add_total; | |
65 | unsigned long del_total; | |
66 | unsigned long find_success; | |
67 | unsigned long find_total; | |
1da177e4 LT |
68 | } swap_cache_info; |
69 | ||
33806f06 SL |
70 | unsigned long total_swapcache_pages(void) |
71 | { | |
4b3ef9da | 72 | unsigned int i, j, nr; |
33806f06 | 73 | unsigned long ret = 0; |
4b3ef9da | 74 | struct address_space *spaces; |
054f1d1f | 75 | struct swap_info_struct *si; |
33806f06 | 76 | |
4b3ef9da | 77 | for (i = 0; i < MAX_SWAPFILES; i++) { |
054f1d1f HY |
78 | swp_entry_t entry = swp_entry(i, 1); |
79 | ||
80 | /* Avoid get_swap_device() to warn for bad swap entry */ | |
81 | if (!swp_swap_info(entry)) | |
82 | continue; | |
83 | /* Prevent swapoff to free swapper_spaces */ | |
84 | si = get_swap_device(entry); | |
85 | if (!si) | |
4b3ef9da | 86 | continue; |
054f1d1f HY |
87 | nr = nr_swapper_spaces[i]; |
88 | spaces = swapper_spaces[i]; | |
4b3ef9da HY |
89 | for (j = 0; j < nr; j++) |
90 | ret += spaces[j].nrpages; | |
054f1d1f | 91 | put_swap_device(si); |
4b3ef9da | 92 | } |
33806f06 SL |
93 | return ret; |
94 | } | |
95 | ||
579f8290 SL |
96 | static atomic_t swapin_readahead_hits = ATOMIC_INIT(4); |
97 | ||
1da177e4 LT |
98 | void show_swap_cache_info(void) |
99 | { | |
33806f06 | 100 | printk("%lu pages in swap cache\n", total_swapcache_pages()); |
2c97b7fc | 101 | printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n", |
1da177e4 | 102 | swap_cache_info.add_total, swap_cache_info.del_total, |
bb63be0a | 103 | swap_cache_info.find_success, swap_cache_info.find_total); |
ec8acf20 SL |
104 | printk("Free swap = %ldkB\n", |
105 | get_nr_swap_pages() << (PAGE_SHIFT - 10)); | |
1da177e4 LT |
106 | printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10)); |
107 | } | |
108 | ||
109 | /* | |
8d93b41c | 110 | * add_to_swap_cache resembles add_to_page_cache_locked on swapper_space, |
1da177e4 LT |
111 | * but sets SwapCache flag and private instead of mapping and index. |
112 | */ | |
8d93b41c | 113 | int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp) |
1da177e4 | 114 | { |
8d93b41c | 115 | struct address_space *address_space = swap_address_space(entry); |
38d8b4e6 | 116 | pgoff_t idx = swp_offset(entry); |
8d93b41c | 117 | XA_STATE_ORDER(xas, &address_space->i_pages, idx, compound_order(page)); |
cb774451 | 118 | unsigned long i, nr = hpage_nr_pages(page); |
1da177e4 | 119 | |
309381fe SL |
120 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
121 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
122 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
51726b12 | 123 | |
38d8b4e6 | 124 | page_ref_add(page, nr); |
31a56396 | 125 | SetPageSwapCache(page); |
31a56396 | 126 | |
8d93b41c MW |
127 | do { |
128 | xas_lock_irq(&xas); | |
129 | xas_create_range(&xas); | |
130 | if (xas_error(&xas)) | |
131 | goto unlock; | |
132 | for (i = 0; i < nr; i++) { | |
133 | VM_BUG_ON_PAGE(xas.xa_index != idx + i, page); | |
134 | set_page_private(page + i, entry.val + i); | |
4101196b | 135 | xas_store(&xas, page); |
8d93b41c MW |
136 | xas_next(&xas); |
137 | } | |
38d8b4e6 HY |
138 | address_space->nrpages += nr; |
139 | __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr); | |
140 | ADD_CACHE_INFO(add_total, nr); | |
8d93b41c MW |
141 | unlock: |
142 | xas_unlock_irq(&xas); | |
143 | } while (xas_nomem(&xas, gfp)); | |
31a56396 | 144 | |
8d93b41c MW |
145 | if (!xas_error(&xas)) |
146 | return 0; | |
31a56396 | 147 | |
8d93b41c MW |
148 | ClearPageSwapCache(page); |
149 | page_ref_sub(page, nr); | |
150 | return xas_error(&xas); | |
1da177e4 LT |
151 | } |
152 | ||
1da177e4 LT |
153 | /* |
154 | * This must be called only on pages that have | |
155 | * been verified to be in the swap cache. | |
156 | */ | |
4e17ec25 | 157 | void __delete_from_swap_cache(struct page *page, swp_entry_t entry) |
1da177e4 | 158 | { |
4e17ec25 | 159 | struct address_space *address_space = swap_address_space(entry); |
38d8b4e6 | 160 | int i, nr = hpage_nr_pages(page); |
4e17ec25 MW |
161 | pgoff_t idx = swp_offset(entry); |
162 | XA_STATE(xas, &address_space->i_pages, idx); | |
33806f06 | 163 | |
309381fe SL |
164 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
165 | VM_BUG_ON_PAGE(!PageSwapCache(page), page); | |
166 | VM_BUG_ON_PAGE(PageWriteback(page), page); | |
1da177e4 | 167 | |
38d8b4e6 | 168 | for (i = 0; i < nr; i++) { |
4e17ec25 | 169 | void *entry = xas_store(&xas, NULL); |
4101196b | 170 | VM_BUG_ON_PAGE(entry != page, entry); |
38d8b4e6 | 171 | set_page_private(page + i, 0); |
4e17ec25 | 172 | xas_next(&xas); |
38d8b4e6 | 173 | } |
1da177e4 | 174 | ClearPageSwapCache(page); |
38d8b4e6 HY |
175 | address_space->nrpages -= nr; |
176 | __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr); | |
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, | |
211 | __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN); | |
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 | |
220 | * dirty. A special case is MADV_FREE page. The page'e pte could have | |
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); |
4e17ec25 | 249 | __delete_from_swap_cache(page, entry); |
b93b0163 | 250 | xa_unlock_irq(&address_space->i_pages); |
1da177e4 | 251 | |
75f6d6d2 | 252 | put_swap_page(page, entry); |
38d8b4e6 | 253 | page_ref_sub(page, hpage_nr_pages(page)); |
1da177e4 LT |
254 | } |
255 | ||
1da177e4 LT |
256 | /* |
257 | * If we are the only user, then try to free up the swap cache. | |
258 | * | |
259 | * Its ok to check for PageSwapCache without the page lock | |
a2c43eed HD |
260 | * here because we are going to recheck again inside |
261 | * try_to_free_swap() _with_ the lock. | |
1da177e4 LT |
262 | * - Marcelo |
263 | */ | |
264 | static inline void free_swap_cache(struct page *page) | |
265 | { | |
a2c43eed HD |
266 | if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) { |
267 | try_to_free_swap(page); | |
1da177e4 LT |
268 | unlock_page(page); |
269 | } | |
270 | } | |
271 | ||
272 | /* | |
273 | * Perform a free_page(), also freeing any swap cache associated with | |
b8072f09 | 274 | * this page if it is the last user of the page. |
1da177e4 LT |
275 | */ |
276 | void free_page_and_swap_cache(struct page *page) | |
277 | { | |
278 | free_swap_cache(page); | |
6fcb52a5 | 279 | if (!is_huge_zero_page(page)) |
770a5370 | 280 | put_page(page); |
1da177e4 LT |
281 | } |
282 | ||
283 | /* | |
284 | * Passed an array of pages, drop them all from swapcache and then release | |
285 | * them. They are removed from the LRU and freed if this is their last use. | |
286 | */ | |
287 | void free_pages_and_swap_cache(struct page **pages, int nr) | |
288 | { | |
1da177e4 | 289 | struct page **pagep = pages; |
aabfb572 | 290 | int i; |
1da177e4 LT |
291 | |
292 | lru_add_drain(); | |
aabfb572 MH |
293 | for (i = 0; i < nr; i++) |
294 | free_swap_cache(pagep[i]); | |
c6f92f9f | 295 | release_pages(pagep, nr); |
1da177e4 LT |
296 | } |
297 | ||
e9e9b7ec MK |
298 | static inline bool swap_use_vma_readahead(void) |
299 | { | |
300 | return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap); | |
301 | } | |
302 | ||
1da177e4 LT |
303 | /* |
304 | * Lookup a swap entry in the swap cache. A found page will be returned | |
305 | * unlocked and with its refcount incremented - we rely on the kernel | |
306 | * lock getting page table operations atomic even if we drop the page | |
307 | * lock before returning. | |
308 | */ | |
ec560175 HY |
309 | struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma, |
310 | unsigned long addr) | |
1da177e4 LT |
311 | { |
312 | struct page *page; | |
eb085574 | 313 | struct swap_info_struct *si; |
1da177e4 | 314 | |
eb085574 HY |
315 | si = get_swap_device(entry); |
316 | if (!si) | |
317 | return NULL; | |
f6ab1f7f | 318 | page = find_get_page(swap_address_space(entry), swp_offset(entry)); |
eb085574 | 319 | put_swap_device(si); |
1da177e4 | 320 | |
ec560175 HY |
321 | INC_CACHE_INFO(find_total); |
322 | if (page) { | |
eaf649eb MK |
323 | bool vma_ra = swap_use_vma_readahead(); |
324 | bool readahead; | |
325 | ||
1da177e4 | 326 | INC_CACHE_INFO(find_success); |
eaf649eb MK |
327 | /* |
328 | * At the moment, we don't support PG_readahead for anon THP | |
329 | * so let's bail out rather than confusing the readahead stat. | |
330 | */ | |
ec560175 HY |
331 | if (unlikely(PageTransCompound(page))) |
332 | return page; | |
eaf649eb | 333 | |
ec560175 | 334 | readahead = TestClearPageReadahead(page); |
eaf649eb MK |
335 | if (vma && vma_ra) { |
336 | unsigned long ra_val; | |
337 | int win, hits; | |
338 | ||
339 | ra_val = GET_SWAP_RA_VAL(vma); | |
340 | win = SWAP_RA_WIN(ra_val); | |
341 | hits = SWAP_RA_HITS(ra_val); | |
ec560175 HY |
342 | if (readahead) |
343 | hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX); | |
344 | atomic_long_set(&vma->swap_readahead_info, | |
345 | SWAP_RA_VAL(addr, win, hits)); | |
346 | } | |
eaf649eb | 347 | |
ec560175 | 348 | if (readahead) { |
cbc65df2 | 349 | count_vm_event(SWAP_RA_HIT); |
eaf649eb | 350 | if (!vma || !vma_ra) |
ec560175 | 351 | atomic_inc(&swapin_readahead_hits); |
cbc65df2 | 352 | } |
579f8290 | 353 | } |
eaf649eb | 354 | |
1da177e4 LT |
355 | return page; |
356 | } | |
357 | ||
5b999aad DS |
358 | struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, |
359 | struct vm_area_struct *vma, unsigned long addr, | |
360 | bool *new_page_allocated) | |
1da177e4 | 361 | { |
eb085574 | 362 | struct swap_info_struct *si; |
4c6355b2 JW |
363 | struct page *page; |
364 | ||
5b999aad | 365 | *new_page_allocated = false; |
1da177e4 | 366 | |
4c6355b2 JW |
367 | for (;;) { |
368 | int err; | |
1da177e4 LT |
369 | /* |
370 | * First check the swap cache. Since this is normally | |
371 | * called after lookup_swap_cache() failed, re-calling | |
372 | * that would confuse statistics. | |
373 | */ | |
eb085574 HY |
374 | si = get_swap_device(entry); |
375 | if (!si) | |
4c6355b2 JW |
376 | return NULL; |
377 | page = find_get_page(swap_address_space(entry), | |
378 | swp_offset(entry)); | |
eb085574 | 379 | put_swap_device(si); |
4c6355b2 JW |
380 | if (page) |
381 | return page; | |
1da177e4 | 382 | |
ba81f838 HY |
383 | /* |
384 | * Just skip read ahead for unused swap slot. | |
385 | * During swap_off when swap_slot_cache is disabled, | |
386 | * we have to handle the race between putting | |
387 | * swap entry in swap cache and marking swap slot | |
388 | * as SWAP_HAS_CACHE. That's done in later part of code or | |
389 | * else swap_off will be aborted if we return NULL. | |
390 | */ | |
391 | if (!__swp_swapcount(entry) && swap_slot_cache_enabled) | |
4c6355b2 | 392 | return NULL; |
e8c26ab6 | 393 | |
1da177e4 | 394 | /* |
4c6355b2 JW |
395 | * Get a new page to read into from swap. Allocate it now, |
396 | * before marking swap_map SWAP_HAS_CACHE, when -EEXIST will | |
397 | * cause any racers to loop around until we add it to cache. | |
1da177e4 | 398 | */ |
4c6355b2 JW |
399 | page = alloc_page_vma(gfp_mask, vma, addr); |
400 | if (!page) | |
401 | return NULL; | |
1da177e4 | 402 | |
f000944d HD |
403 | /* |
404 | * Swap entry may have been freed since our caller observed it. | |
405 | */ | |
355cfa73 | 406 | err = swapcache_prepare(entry); |
4c6355b2 | 407 | if (!err) |
f000944d HD |
408 | break; |
409 | ||
4c6355b2 JW |
410 | put_page(page); |
411 | if (err != -EEXIST) | |
412 | return NULL; | |
413 | ||
2ca4532a | 414 | /* |
4c6355b2 JW |
415 | * We might race against __delete_from_swap_cache(), and |
416 | * stumble across a swap_map entry whose SWAP_HAS_CACHE | |
417 | * has not yet been cleared. Or race against another | |
418 | * __read_swap_cache_async(), which has set SWAP_HAS_CACHE | |
419 | * in swap_map, but not yet added its page to swap cache. | |
2ca4532a | 420 | */ |
4c6355b2 JW |
421 | cond_resched(); |
422 | } | |
423 | ||
424 | /* | |
425 | * The swap entry is ours to swap in. Prepare the new page. | |
426 | */ | |
427 | ||
428 | __SetPageLocked(page); | |
429 | __SetPageSwapBacked(page); | |
430 | ||
431 | /* May fail (-ENOMEM) if XArray node allocation failed. */ | |
432 | if (add_to_swap_cache(page, entry, gfp_mask & GFP_KERNEL)) { | |
433 | put_swap_page(page, entry); | |
434 | goto fail_unlock; | |
435 | } | |
436 | ||
d9eb1ea2 | 437 | if (mem_cgroup_charge(page, NULL, gfp_mask)) { |
4c6355b2 JW |
438 | delete_from_swap_cache(page); |
439 | goto fail_unlock; | |
440 | } | |
441 | ||
314b57fb JW |
442 | /* XXX: Move to lru_cache_add() when it supports new vs putback */ |
443 | spin_lock_irq(&page_pgdat(page)->lru_lock); | |
96f8bf4f | 444 | lru_note_cost_page(page); |
314b57fb JW |
445 | spin_unlock_irq(&page_pgdat(page)->lru_lock); |
446 | ||
4c6355b2 JW |
447 | /* Caller will initiate read into locked page */ |
448 | SetPageWorkingset(page); | |
6058eaec | 449 | lru_cache_add(page); |
4c6355b2 JW |
450 | *new_page_allocated = true; |
451 | return page; | |
1da177e4 | 452 | |
4c6355b2 JW |
453 | fail_unlock: |
454 | unlock_page(page); | |
455 | put_page(page); | |
456 | return NULL; | |
1da177e4 | 457 | } |
46017e95 | 458 | |
5b999aad DS |
459 | /* |
460 | * Locate a page of swap in physical memory, reserving swap cache space | |
461 | * and reading the disk if it is not already cached. | |
462 | * A failure return means that either the page allocation failed or that | |
463 | * the swap entry is no longer in use. | |
464 | */ | |
465 | struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, | |
23955622 | 466 | struct vm_area_struct *vma, unsigned long addr, bool do_poll) |
5b999aad DS |
467 | { |
468 | bool page_was_allocated; | |
469 | struct page *retpage = __read_swap_cache_async(entry, gfp_mask, | |
470 | vma, addr, &page_was_allocated); | |
471 | ||
472 | if (page_was_allocated) | |
23955622 | 473 | swap_readpage(retpage, do_poll); |
5b999aad DS |
474 | |
475 | return retpage; | |
476 | } | |
477 | ||
ec560175 HY |
478 | static unsigned int __swapin_nr_pages(unsigned long prev_offset, |
479 | unsigned long offset, | |
480 | int hits, | |
481 | int max_pages, | |
482 | int prev_win) | |
579f8290 | 483 | { |
ec560175 | 484 | unsigned int pages, last_ra; |
579f8290 SL |
485 | |
486 | /* | |
487 | * This heuristic has been found to work well on both sequential and | |
488 | * random loads, swapping to hard disk or to SSD: please don't ask | |
489 | * what the "+ 2" means, it just happens to work well, that's all. | |
490 | */ | |
ec560175 | 491 | pages = hits + 2; |
579f8290 SL |
492 | if (pages == 2) { |
493 | /* | |
494 | * We can have no readahead hits to judge by: but must not get | |
495 | * stuck here forever, so check for an adjacent offset instead | |
496 | * (and don't even bother to check whether swap type is same). | |
497 | */ | |
498 | if (offset != prev_offset + 1 && offset != prev_offset - 1) | |
499 | pages = 1; | |
579f8290 SL |
500 | } else { |
501 | unsigned int roundup = 4; | |
502 | while (roundup < pages) | |
503 | roundup <<= 1; | |
504 | pages = roundup; | |
505 | } | |
506 | ||
507 | if (pages > max_pages) | |
508 | pages = max_pages; | |
509 | ||
510 | /* Don't shrink readahead too fast */ | |
ec560175 | 511 | last_ra = prev_win / 2; |
579f8290 SL |
512 | if (pages < last_ra) |
513 | pages = last_ra; | |
ec560175 HY |
514 | |
515 | return pages; | |
516 | } | |
517 | ||
518 | static unsigned long swapin_nr_pages(unsigned long offset) | |
519 | { | |
520 | static unsigned long prev_offset; | |
521 | unsigned int hits, pages, max_pages; | |
522 | static atomic_t last_readahead_pages; | |
523 | ||
524 | max_pages = 1 << READ_ONCE(page_cluster); | |
525 | if (max_pages <= 1) | |
526 | return 1; | |
527 | ||
528 | hits = atomic_xchg(&swapin_readahead_hits, 0); | |
d6c1f098 QC |
529 | pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits, |
530 | max_pages, | |
ec560175 HY |
531 | atomic_read(&last_readahead_pages)); |
532 | if (!hits) | |
d6c1f098 | 533 | WRITE_ONCE(prev_offset, offset); |
579f8290 SL |
534 | atomic_set(&last_readahead_pages, pages); |
535 | ||
536 | return pages; | |
537 | } | |
538 | ||
46017e95 | 539 | /** |
e9e9b7ec | 540 | * swap_cluster_readahead - swap in pages in hope we need them soon |
46017e95 | 541 | * @entry: swap entry of this memory |
7682486b | 542 | * @gfp_mask: memory allocation flags |
e9e9b7ec | 543 | * @vmf: fault information |
46017e95 HD |
544 | * |
545 | * Returns the struct page for entry and addr, after queueing swapin. | |
546 | * | |
547 | * Primitive swap readahead code. We simply read an aligned block of | |
548 | * (1 << page_cluster) entries in the swap area. This method is chosen | |
549 | * because it doesn't cost us any seek time. We also make sure to queue | |
550 | * the 'original' request together with the readahead ones... | |
551 | * | |
552 | * This has been extended to use the NUMA policies from the mm triggering | |
553 | * the readahead. | |
554 | * | |
c1e8d7c6 | 555 | * Caller must hold read mmap_lock if vmf->vma is not NULL. |
46017e95 | 556 | */ |
e9e9b7ec MK |
557 | struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask, |
558 | struct vm_fault *vmf) | |
46017e95 | 559 | { |
46017e95 | 560 | struct page *page; |
579f8290 SL |
561 | unsigned long entry_offset = swp_offset(entry); |
562 | unsigned long offset = entry_offset; | |
67f96aa2 | 563 | unsigned long start_offset, end_offset; |
579f8290 | 564 | unsigned long mask; |
e9a6effa | 565 | struct swap_info_struct *si = swp_swap_info(entry); |
3fb5c298 | 566 | struct blk_plug plug; |
c4fa6309 | 567 | bool do_poll = true, page_allocated; |
e9e9b7ec MK |
568 | struct vm_area_struct *vma = vmf->vma; |
569 | unsigned long addr = vmf->address; | |
46017e95 | 570 | |
579f8290 SL |
571 | mask = swapin_nr_pages(offset) - 1; |
572 | if (!mask) | |
573 | goto skip; | |
574 | ||
8fd2e0b5 YS |
575 | /* Test swap type to make sure the dereference is safe */ |
576 | if (likely(si->flags & (SWP_BLKDEV | SWP_FS))) { | |
577 | struct inode *inode = si->swap_file->f_mapping->host; | |
578 | if (inode_read_congested(inode)) | |
579 | goto skip; | |
580 | } | |
581 | ||
23955622 | 582 | do_poll = false; |
67f96aa2 RR |
583 | /* Read a page_cluster sized and aligned cluster around offset. */ |
584 | start_offset = offset & ~mask; | |
585 | end_offset = offset | mask; | |
586 | if (!start_offset) /* First page is swap header. */ | |
587 | start_offset++; | |
e9a6effa HY |
588 | if (end_offset >= si->max) |
589 | end_offset = si->max - 1; | |
67f96aa2 | 590 | |
3fb5c298 | 591 | blk_start_plug(&plug); |
67f96aa2 | 592 | for (offset = start_offset; offset <= end_offset ; offset++) { |
46017e95 | 593 | /* Ok, do the async read-ahead now */ |
c4fa6309 HY |
594 | page = __read_swap_cache_async( |
595 | swp_entry(swp_type(entry), offset), | |
596 | gfp_mask, vma, addr, &page_allocated); | |
46017e95 | 597 | if (!page) |
67f96aa2 | 598 | continue; |
c4fa6309 HY |
599 | if (page_allocated) { |
600 | swap_readpage(page, false); | |
eaf649eb | 601 | if (offset != entry_offset) { |
c4fa6309 HY |
602 | SetPageReadahead(page); |
603 | count_vm_event(SWAP_RA); | |
604 | } | |
cbc65df2 | 605 | } |
09cbfeaf | 606 | put_page(page); |
46017e95 | 607 | } |
3fb5c298 CE |
608 | blk_finish_plug(&plug); |
609 | ||
46017e95 | 610 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
579f8290 | 611 | skip: |
23955622 | 612 | return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll); |
46017e95 | 613 | } |
4b3ef9da HY |
614 | |
615 | int init_swap_address_space(unsigned int type, unsigned long nr_pages) | |
616 | { | |
617 | struct address_space *spaces, *space; | |
618 | unsigned int i, nr; | |
619 | ||
620 | nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES); | |
778e1cdd | 621 | spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL); |
4b3ef9da HY |
622 | if (!spaces) |
623 | return -ENOMEM; | |
624 | for (i = 0; i < nr; i++) { | |
625 | space = spaces + i; | |
a2833486 | 626 | xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ); |
4b3ef9da HY |
627 | atomic_set(&space->i_mmap_writable, 0); |
628 | space->a_ops = &swap_aops; | |
629 | /* swap cache doesn't use writeback related tags */ | |
630 | mapping_set_no_writeback_tags(space); | |
4b3ef9da HY |
631 | } |
632 | nr_swapper_spaces[type] = nr; | |
054f1d1f | 633 | swapper_spaces[type] = spaces; |
4b3ef9da HY |
634 | |
635 | return 0; | |
636 | } | |
637 | ||
638 | void exit_swap_address_space(unsigned int type) | |
639 | { | |
054f1d1f | 640 | kvfree(swapper_spaces[type]); |
4b3ef9da | 641 | nr_swapper_spaces[type] = 0; |
054f1d1f | 642 | swapper_spaces[type] = NULL; |
4b3ef9da | 643 | } |
ec560175 HY |
644 | |
645 | static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma, | |
646 | unsigned long faddr, | |
647 | unsigned long lpfn, | |
648 | unsigned long rpfn, | |
649 | unsigned long *start, | |
650 | unsigned long *end) | |
651 | { | |
652 | *start = max3(lpfn, PFN_DOWN(vma->vm_start), | |
653 | PFN_DOWN(faddr & PMD_MASK)); | |
654 | *end = min3(rpfn, PFN_DOWN(vma->vm_end), | |
655 | PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE)); | |
656 | } | |
657 | ||
eaf649eb MK |
658 | static void swap_ra_info(struct vm_fault *vmf, |
659 | struct vma_swap_readahead *ra_info) | |
ec560175 HY |
660 | { |
661 | struct vm_area_struct *vma = vmf->vma; | |
eaf649eb | 662 | unsigned long ra_val; |
ec560175 HY |
663 | swp_entry_t entry; |
664 | unsigned long faddr, pfn, fpfn; | |
665 | unsigned long start, end; | |
eaf649eb | 666 | pte_t *pte, *orig_pte; |
ec560175 HY |
667 | unsigned int max_win, hits, prev_win, win, left; |
668 | #ifndef CONFIG_64BIT | |
669 | pte_t *tpte; | |
670 | #endif | |
671 | ||
61b63972 HY |
672 | max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster), |
673 | SWAP_RA_ORDER_CEILING); | |
674 | if (max_win == 1) { | |
eaf649eb MK |
675 | ra_info->win = 1; |
676 | return; | |
61b63972 HY |
677 | } |
678 | ||
ec560175 | 679 | faddr = vmf->address; |
eaf649eb MK |
680 | orig_pte = pte = pte_offset_map(vmf->pmd, faddr); |
681 | entry = pte_to_swp_entry(*pte); | |
682 | if ((unlikely(non_swap_entry(entry)))) { | |
683 | pte_unmap(orig_pte); | |
684 | return; | |
685 | } | |
ec560175 | 686 | |
ec560175 | 687 | fpfn = PFN_DOWN(faddr); |
eaf649eb MK |
688 | ra_val = GET_SWAP_RA_VAL(vma); |
689 | pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val)); | |
690 | prev_win = SWAP_RA_WIN(ra_val); | |
691 | hits = SWAP_RA_HITS(ra_val); | |
692 | ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits, | |
ec560175 HY |
693 | max_win, prev_win); |
694 | atomic_long_set(&vma->swap_readahead_info, | |
695 | SWAP_RA_VAL(faddr, win, 0)); | |
696 | ||
eaf649eb MK |
697 | if (win == 1) { |
698 | pte_unmap(orig_pte); | |
699 | return; | |
700 | } | |
ec560175 HY |
701 | |
702 | /* Copy the PTEs because the page table may be unmapped */ | |
703 | if (fpfn == pfn + 1) | |
704 | swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end); | |
705 | else if (pfn == fpfn + 1) | |
706 | swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1, | |
707 | &start, &end); | |
708 | else { | |
709 | left = (win - 1) / 2; | |
710 | swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left, | |
711 | &start, &end); | |
712 | } | |
eaf649eb MK |
713 | ra_info->nr_pte = end - start; |
714 | ra_info->offset = fpfn - start; | |
715 | pte -= ra_info->offset; | |
ec560175 | 716 | #ifdef CONFIG_64BIT |
eaf649eb | 717 | ra_info->ptes = pte; |
ec560175 | 718 | #else |
eaf649eb | 719 | tpte = ra_info->ptes; |
ec560175 HY |
720 | for (pfn = start; pfn != end; pfn++) |
721 | *tpte++ = *pte++; | |
722 | #endif | |
eaf649eb | 723 | pte_unmap(orig_pte); |
ec560175 HY |
724 | } |
725 | ||
e9f59873 YS |
726 | /** |
727 | * swap_vma_readahead - swap in pages in hope we need them soon | |
728 | * @entry: swap entry of this memory | |
729 | * @gfp_mask: memory allocation flags | |
730 | * @vmf: fault information | |
731 | * | |
732 | * Returns the struct page for entry and addr, after queueing swapin. | |
733 | * | |
734 | * Primitive swap readahead code. We simply read in a few pages whoes | |
735 | * virtual addresses are around the fault address in the same vma. | |
736 | * | |
c1e8d7c6 | 737 | * Caller must hold read mmap_lock if vmf->vma is not NULL. |
e9f59873 YS |
738 | * |
739 | */ | |
f5c754d6 CIK |
740 | static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask, |
741 | struct vm_fault *vmf) | |
ec560175 HY |
742 | { |
743 | struct blk_plug plug; | |
744 | struct vm_area_struct *vma = vmf->vma; | |
745 | struct page *page; | |
746 | pte_t *pte, pentry; | |
747 | swp_entry_t entry; | |
748 | unsigned int i; | |
749 | bool page_allocated; | |
eaf649eb | 750 | struct vma_swap_readahead ra_info = {0,}; |
ec560175 | 751 | |
eaf649eb MK |
752 | swap_ra_info(vmf, &ra_info); |
753 | if (ra_info.win == 1) | |
ec560175 HY |
754 | goto skip; |
755 | ||
756 | blk_start_plug(&plug); | |
eaf649eb | 757 | for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte; |
ec560175 HY |
758 | i++, pte++) { |
759 | pentry = *pte; | |
760 | if (pte_none(pentry)) | |
761 | continue; | |
762 | if (pte_present(pentry)) | |
763 | continue; | |
764 | entry = pte_to_swp_entry(pentry); | |
765 | if (unlikely(non_swap_entry(entry))) | |
766 | continue; | |
767 | page = __read_swap_cache_async(entry, gfp_mask, vma, | |
768 | vmf->address, &page_allocated); | |
769 | if (!page) | |
770 | continue; | |
771 | if (page_allocated) { | |
772 | swap_readpage(page, false); | |
eaf649eb | 773 | if (i != ra_info.offset) { |
ec560175 HY |
774 | SetPageReadahead(page); |
775 | count_vm_event(SWAP_RA); | |
776 | } | |
777 | } | |
778 | put_page(page); | |
779 | } | |
780 | blk_finish_plug(&plug); | |
781 | lru_add_drain(); | |
782 | skip: | |
783 | return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, | |
eaf649eb | 784 | ra_info.win == 1); |
ec560175 | 785 | } |
d9bfcfdc | 786 | |
e9e9b7ec MK |
787 | /** |
788 | * swapin_readahead - swap in pages in hope we need them soon | |
789 | * @entry: swap entry of this memory | |
790 | * @gfp_mask: memory allocation flags | |
791 | * @vmf: fault information | |
792 | * | |
793 | * Returns the struct page for entry and addr, after queueing swapin. | |
794 | * | |
795 | * It's a main entry function for swap readahead. By the configuration, | |
796 | * it will read ahead blocks by cluster-based(ie, physical disk based) | |
797 | * or vma-based(ie, virtual address based on faulty address) readahead. | |
798 | */ | |
799 | struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, | |
800 | struct vm_fault *vmf) | |
801 | { | |
802 | return swap_use_vma_readahead() ? | |
803 | swap_vma_readahead(entry, gfp_mask, vmf) : | |
804 | swap_cluster_readahead(entry, gfp_mask, vmf); | |
805 | } | |
806 | ||
d9bfcfdc HY |
807 | #ifdef CONFIG_SYSFS |
808 | static ssize_t vma_ra_enabled_show(struct kobject *kobj, | |
809 | struct kobj_attribute *attr, char *buf) | |
810 | { | |
e9e9b7ec | 811 | return sprintf(buf, "%s\n", enable_vma_readahead ? "true" : "false"); |
d9bfcfdc HY |
812 | } |
813 | static ssize_t vma_ra_enabled_store(struct kobject *kobj, | |
814 | struct kobj_attribute *attr, | |
815 | const char *buf, size_t count) | |
816 | { | |
817 | if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1)) | |
e9e9b7ec | 818 | enable_vma_readahead = true; |
d9bfcfdc | 819 | else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1)) |
e9e9b7ec | 820 | enable_vma_readahead = false; |
d9bfcfdc HY |
821 | else |
822 | return -EINVAL; | |
823 | ||
824 | return count; | |
825 | } | |
826 | static struct kobj_attribute vma_ra_enabled_attr = | |
827 | __ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show, | |
828 | vma_ra_enabled_store); | |
829 | ||
d9bfcfdc HY |
830 | static struct attribute *swap_attrs[] = { |
831 | &vma_ra_enabled_attr.attr, | |
d9bfcfdc HY |
832 | NULL, |
833 | }; | |
834 | ||
835 | static struct attribute_group swap_attr_group = { | |
836 | .attrs = swap_attrs, | |
837 | }; | |
838 | ||
839 | static int __init swap_init_sysfs(void) | |
840 | { | |
841 | int err; | |
842 | struct kobject *swap_kobj; | |
843 | ||
844 | swap_kobj = kobject_create_and_add("swap", mm_kobj); | |
845 | if (!swap_kobj) { | |
846 | pr_err("failed to create swap kobject\n"); | |
847 | return -ENOMEM; | |
848 | } | |
849 | err = sysfs_create_group(swap_kobj, &swap_attr_group); | |
850 | if (err) { | |
851 | pr_err("failed to register swap group\n"); | |
852 | goto delete_obj; | |
853 | } | |
854 | return 0; | |
855 | ||
856 | delete_obj: | |
857 | kobject_put(swap_kobj); | |
858 | return err; | |
859 | } | |
860 | subsys_initcall(swap_init_sysfs); | |
861 | #endif |