Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/swap_state.c | |
3 | * | |
4 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
5 | * Swap reorganised 29.12.95, Stephen Tweedie | |
6 | * | |
7 | * Rewritten to use page cache, (C) 1998 Stephen Tweedie | |
8 | */ | |
1da177e4 | 9 | #include <linux/mm.h> |
5a0e3ad6 | 10 | #include <linux/gfp.h> |
1da177e4 LT |
11 | #include <linux/kernel_stat.h> |
12 | #include <linux/swap.h> | |
46017e95 | 13 | #include <linux/swapops.h> |
1da177e4 LT |
14 | #include <linux/init.h> |
15 | #include <linux/pagemap.h> | |
1da177e4 | 16 | #include <linux/backing-dev.h> |
3fb5c298 | 17 | #include <linux/blkdev.h> |
c484d410 | 18 | #include <linux/pagevec.h> |
b20a3503 | 19 | #include <linux/migrate.h> |
1da177e4 LT |
20 | |
21 | #include <asm/pgtable.h> | |
22 | ||
23 | /* | |
24 | * swapper_space is a fiction, retained to simplify the path through | |
7eaceacc | 25 | * vmscan's shrink_page_list. |
1da177e4 | 26 | */ |
f5e54d6e | 27 | static const struct address_space_operations swap_aops = { |
1da177e4 | 28 | .writepage = swap_writepage, |
62c230bc | 29 | .set_page_dirty = swap_set_page_dirty, |
1c93923c | 30 | #ifdef CONFIG_MIGRATION |
e965f963 | 31 | .migratepage = migrate_page, |
1c93923c | 32 | #endif |
1da177e4 LT |
33 | }; |
34 | ||
35 | static struct backing_dev_info swap_backing_dev_info = { | |
d993831f | 36 | .name = "swap", |
4f98a2fe | 37 | .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED, |
1da177e4 LT |
38 | }; |
39 | ||
33806f06 SL |
40 | struct address_space swapper_spaces[MAX_SWAPFILES] = { |
41 | [0 ... MAX_SWAPFILES - 1] = { | |
42 | .page_tree = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN), | |
4bb5f5d9 | 43 | .i_mmap_writable = ATOMIC_INIT(0), |
33806f06 SL |
44 | .a_ops = &swap_aops, |
45 | .backing_dev_info = &swap_backing_dev_info, | |
46 | } | |
1da177e4 | 47 | }; |
1da177e4 LT |
48 | |
49 | #define INC_CACHE_INFO(x) do { swap_cache_info.x++; } while (0) | |
50 | ||
51 | static struct { | |
52 | unsigned long add_total; | |
53 | unsigned long del_total; | |
54 | unsigned long find_success; | |
55 | unsigned long find_total; | |
1da177e4 LT |
56 | } swap_cache_info; |
57 | ||
33806f06 SL |
58 | unsigned long total_swapcache_pages(void) |
59 | { | |
60 | int i; | |
61 | unsigned long ret = 0; | |
62 | ||
63 | for (i = 0; i < MAX_SWAPFILES; i++) | |
64 | ret += swapper_spaces[i].nrpages; | |
65 | return ret; | |
66 | } | |
67 | ||
579f8290 SL |
68 | static atomic_t swapin_readahead_hits = ATOMIC_INIT(4); |
69 | ||
1da177e4 LT |
70 | void show_swap_cache_info(void) |
71 | { | |
33806f06 | 72 | printk("%lu pages in swap cache\n", total_swapcache_pages()); |
2c97b7fc | 73 | printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n", |
1da177e4 | 74 | swap_cache_info.add_total, swap_cache_info.del_total, |
bb63be0a | 75 | swap_cache_info.find_success, swap_cache_info.find_total); |
ec8acf20 SL |
76 | printk("Free swap = %ldkB\n", |
77 | get_nr_swap_pages() << (PAGE_SHIFT - 10)); | |
1da177e4 LT |
78 | printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10)); |
79 | } | |
80 | ||
81 | /* | |
31a56396 | 82 | * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space, |
1da177e4 LT |
83 | * but sets SwapCache flag and private instead of mapping and index. |
84 | */ | |
2f772e6c | 85 | int __add_to_swap_cache(struct page *page, swp_entry_t entry) |
1da177e4 LT |
86 | { |
87 | int error; | |
33806f06 | 88 | struct address_space *address_space; |
1da177e4 | 89 | |
309381fe SL |
90 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
91 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
92 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
51726b12 | 93 | |
31a56396 DN |
94 | page_cache_get(page); |
95 | SetPageSwapCache(page); | |
96 | set_page_private(page, entry.val); | |
97 | ||
33806f06 SL |
98 | address_space = swap_address_space(entry); |
99 | spin_lock_irq(&address_space->tree_lock); | |
100 | error = radix_tree_insert(&address_space->page_tree, | |
101 | entry.val, page); | |
31a56396 | 102 | if (likely(!error)) { |
33806f06 | 103 | address_space->nrpages++; |
31a56396 DN |
104 | __inc_zone_page_state(page, NR_FILE_PAGES); |
105 | INC_CACHE_INFO(add_total); | |
106 | } | |
33806f06 | 107 | spin_unlock_irq(&address_space->tree_lock); |
31a56396 DN |
108 | |
109 | if (unlikely(error)) { | |
2ca4532a DN |
110 | /* |
111 | * Only the context which have set SWAP_HAS_CACHE flag | |
112 | * would call add_to_swap_cache(). | |
113 | * So add_to_swap_cache() doesn't returns -EEXIST. | |
114 | */ | |
115 | VM_BUG_ON(error == -EEXIST); | |
31a56396 DN |
116 | set_page_private(page, 0UL); |
117 | ClearPageSwapCache(page); | |
118 | page_cache_release(page); | |
119 | } | |
120 | ||
121 | return error; | |
122 | } | |
123 | ||
124 | ||
125 | int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask) | |
126 | { | |
127 | int error; | |
128 | ||
5e4c0d97 | 129 | error = radix_tree_maybe_preload(gfp_mask); |
35c754d7 | 130 | if (!error) { |
31a56396 | 131 | error = __add_to_swap_cache(page, entry); |
1da177e4 | 132 | radix_tree_preload_end(); |
fa1de900 | 133 | } |
1da177e4 LT |
134 | return error; |
135 | } | |
136 | ||
1da177e4 LT |
137 | /* |
138 | * This must be called only on pages that have | |
139 | * been verified to be in the swap cache. | |
140 | */ | |
141 | void __delete_from_swap_cache(struct page *page) | |
142 | { | |
33806f06 SL |
143 | swp_entry_t entry; |
144 | struct address_space *address_space; | |
145 | ||
309381fe SL |
146 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
147 | VM_BUG_ON_PAGE(!PageSwapCache(page), page); | |
148 | VM_BUG_ON_PAGE(PageWriteback(page), page); | |
1da177e4 | 149 | |
33806f06 SL |
150 | entry.val = page_private(page); |
151 | address_space = swap_address_space(entry); | |
152 | radix_tree_delete(&address_space->page_tree, page_private(page)); | |
4c21e2f2 | 153 | set_page_private(page, 0); |
1da177e4 | 154 | ClearPageSwapCache(page); |
33806f06 | 155 | address_space->nrpages--; |
347ce434 | 156 | __dec_zone_page_state(page, NR_FILE_PAGES); |
1da177e4 LT |
157 | INC_CACHE_INFO(del_total); |
158 | } | |
159 | ||
160 | /** | |
161 | * add_to_swap - allocate swap space for a page | |
162 | * @page: page we want to move to swap | |
163 | * | |
164 | * Allocate swap space for the page and add the page to the | |
165 | * swap cache. Caller needs to hold the page lock. | |
166 | */ | |
5bc7b8ac | 167 | int add_to_swap(struct page *page, struct list_head *list) |
1da177e4 LT |
168 | { |
169 | swp_entry_t entry; | |
1da177e4 LT |
170 | int err; |
171 | ||
309381fe SL |
172 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
173 | VM_BUG_ON_PAGE(!PageUptodate(page), page); | |
1da177e4 | 174 | |
2ca4532a DN |
175 | entry = get_swap_page(); |
176 | if (!entry.val) | |
177 | return 0; | |
178 | ||
3f04f62f | 179 | if (unlikely(PageTransHuge(page))) |
5bc7b8ac | 180 | if (unlikely(split_huge_page_to_list(page, list))) { |
0a31bc97 | 181 | swapcache_free(entry); |
3f04f62f AA |
182 | return 0; |
183 | } | |
184 | ||
2ca4532a DN |
185 | /* |
186 | * Radix-tree node allocations from PF_MEMALLOC contexts could | |
187 | * completely exhaust the page allocator. __GFP_NOMEMALLOC | |
188 | * stops emergency reserves from being allocated. | |
189 | * | |
190 | * TODO: this could cause a theoretical memory reclaim | |
191 | * deadlock in the swap out path. | |
192 | */ | |
193 | /* | |
194 | * Add it to the swap cache and mark it dirty | |
195 | */ | |
196 | err = add_to_swap_cache(page, entry, | |
197 | __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN); | |
198 | ||
199 | if (!err) { /* Success */ | |
200 | SetPageDirty(page); | |
201 | return 1; | |
202 | } else { /* -ENOMEM radix-tree allocation failure */ | |
bd53b714 | 203 | /* |
2ca4532a DN |
204 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely |
205 | * clear SWAP_HAS_CACHE flag. | |
1da177e4 | 206 | */ |
0a31bc97 | 207 | swapcache_free(entry); |
2ca4532a | 208 | return 0; |
1da177e4 LT |
209 | } |
210 | } | |
211 | ||
212 | /* | |
213 | * This must be called only on pages that have | |
214 | * been verified to be in the swap cache and locked. | |
215 | * It will never put the page into the free list, | |
216 | * the caller has a reference on the page. | |
217 | */ | |
218 | void delete_from_swap_cache(struct page *page) | |
219 | { | |
220 | swp_entry_t entry; | |
33806f06 | 221 | struct address_space *address_space; |
1da177e4 | 222 | |
4c21e2f2 | 223 | entry.val = page_private(page); |
1da177e4 | 224 | |
33806f06 SL |
225 | address_space = swap_address_space(entry); |
226 | spin_lock_irq(&address_space->tree_lock); | |
1da177e4 | 227 | __delete_from_swap_cache(page); |
33806f06 | 228 | spin_unlock_irq(&address_space->tree_lock); |
1da177e4 | 229 | |
0a31bc97 | 230 | swapcache_free(entry); |
1da177e4 LT |
231 | page_cache_release(page); |
232 | } | |
233 | ||
1da177e4 LT |
234 | /* |
235 | * If we are the only user, then try to free up the swap cache. | |
236 | * | |
237 | * Its ok to check for PageSwapCache without the page lock | |
a2c43eed HD |
238 | * here because we are going to recheck again inside |
239 | * try_to_free_swap() _with_ the lock. | |
1da177e4 LT |
240 | * - Marcelo |
241 | */ | |
242 | static inline void free_swap_cache(struct page *page) | |
243 | { | |
a2c43eed HD |
244 | if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) { |
245 | try_to_free_swap(page); | |
1da177e4 LT |
246 | unlock_page(page); |
247 | } | |
248 | } | |
249 | ||
250 | /* | |
251 | * Perform a free_page(), also freeing any swap cache associated with | |
b8072f09 | 252 | * this page if it is the last user of the page. |
1da177e4 LT |
253 | */ |
254 | void free_page_and_swap_cache(struct page *page) | |
255 | { | |
256 | free_swap_cache(page); | |
257 | page_cache_release(page); | |
258 | } | |
259 | ||
260 | /* | |
261 | * Passed an array of pages, drop them all from swapcache and then release | |
262 | * them. They are removed from the LRU and freed if this is their last use. | |
263 | */ | |
264 | void free_pages_and_swap_cache(struct page **pages, int nr) | |
265 | { | |
1da177e4 | 266 | struct page **pagep = pages; |
aabfb572 | 267 | int i; |
1da177e4 LT |
268 | |
269 | lru_add_drain(); | |
aabfb572 MH |
270 | for (i = 0; i < nr; i++) |
271 | free_swap_cache(pagep[i]); | |
272 | release_pages(pagep, nr, false); | |
1da177e4 LT |
273 | } |
274 | ||
275 | /* | |
276 | * Lookup a swap entry in the swap cache. A found page will be returned | |
277 | * unlocked and with its refcount incremented - we rely on the kernel | |
278 | * lock getting page table operations atomic even if we drop the page | |
279 | * lock before returning. | |
280 | */ | |
281 | struct page * lookup_swap_cache(swp_entry_t entry) | |
282 | { | |
283 | struct page *page; | |
284 | ||
33806f06 | 285 | page = find_get_page(swap_address_space(entry), entry.val); |
1da177e4 | 286 | |
579f8290 | 287 | if (page) { |
1da177e4 | 288 | INC_CACHE_INFO(find_success); |
579f8290 SL |
289 | if (TestClearPageReadahead(page)) |
290 | atomic_inc(&swapin_readahead_hits); | |
291 | } | |
1da177e4 LT |
292 | |
293 | INC_CACHE_INFO(find_total); | |
294 | return page; | |
295 | } | |
296 | ||
297 | /* | |
298 | * Locate a page of swap in physical memory, reserving swap cache space | |
299 | * and reading the disk if it is not already cached. | |
300 | * A failure return means that either the page allocation failed or that | |
301 | * the swap entry is no longer in use. | |
302 | */ | |
02098fea | 303 | struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, |
1da177e4 LT |
304 | struct vm_area_struct *vma, unsigned long addr) |
305 | { | |
306 | struct page *found_page, *new_page = NULL; | |
307 | int err; | |
308 | ||
309 | do { | |
310 | /* | |
311 | * First check the swap cache. Since this is normally | |
312 | * called after lookup_swap_cache() failed, re-calling | |
313 | * that would confuse statistics. | |
314 | */ | |
33806f06 SL |
315 | found_page = find_get_page(swap_address_space(entry), |
316 | entry.val); | |
1da177e4 LT |
317 | if (found_page) |
318 | break; | |
319 | ||
320 | /* | |
321 | * Get a new page to read into from swap. | |
322 | */ | |
323 | if (!new_page) { | |
02098fea | 324 | new_page = alloc_page_vma(gfp_mask, vma, addr); |
1da177e4 LT |
325 | if (!new_page) |
326 | break; /* Out of memory */ | |
327 | } | |
328 | ||
31a56396 DN |
329 | /* |
330 | * call radix_tree_preload() while we can wait. | |
331 | */ | |
5e4c0d97 | 332 | err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL); |
31a56396 DN |
333 | if (err) |
334 | break; | |
335 | ||
f000944d HD |
336 | /* |
337 | * Swap entry may have been freed since our caller observed it. | |
338 | */ | |
355cfa73 | 339 | err = swapcache_prepare(entry); |
cbab0e4e | 340 | if (err == -EEXIST) { |
31a56396 | 341 | radix_tree_preload_end(); |
cbab0e4e RA |
342 | /* |
343 | * We might race against get_swap_page() and stumble | |
344 | * across a SWAP_HAS_CACHE swap_map entry whose page | |
345 | * has not been brought into the swapcache yet, while | |
346 | * the other end is scheduled away waiting on discard | |
347 | * I/O completion at scan_swap_map(). | |
348 | * | |
349 | * In order to avoid turning this transitory state | |
350 | * into a permanent loop around this -EEXIST case | |
351 | * if !CONFIG_PREEMPT and the I/O completion happens | |
352 | * to be waiting on the CPU waitqueue where we are now | |
353 | * busy looping, we just conditionally invoke the | |
354 | * scheduler here, if there are some more important | |
355 | * tasks to run. | |
356 | */ | |
357 | cond_resched(); | |
355cfa73 | 358 | continue; |
31a56396 DN |
359 | } |
360 | if (err) { /* swp entry is obsolete ? */ | |
361 | radix_tree_preload_end(); | |
f000944d | 362 | break; |
31a56396 | 363 | } |
f000944d | 364 | |
2ca4532a | 365 | /* May fail (-ENOMEM) if radix-tree node allocation failed. */ |
f45840b5 | 366 | __set_page_locked(new_page); |
b2e18538 | 367 | SetPageSwapBacked(new_page); |
31a56396 | 368 | err = __add_to_swap_cache(new_page, entry); |
529ae9aa | 369 | if (likely(!err)) { |
31a56396 | 370 | radix_tree_preload_end(); |
1da177e4 LT |
371 | /* |
372 | * Initiate read into locked page and return. | |
373 | */ | |
c5fdae46 | 374 | lru_cache_add_anon(new_page); |
aca8bf32 | 375 | swap_readpage(new_page); |
1da177e4 LT |
376 | return new_page; |
377 | } | |
31a56396 | 378 | radix_tree_preload_end(); |
b2e18538 | 379 | ClearPageSwapBacked(new_page); |
f45840b5 | 380 | __clear_page_locked(new_page); |
2ca4532a DN |
381 | /* |
382 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely | |
383 | * clear SWAP_HAS_CACHE flag. | |
384 | */ | |
0a31bc97 | 385 | swapcache_free(entry); |
f000944d | 386 | } while (err != -ENOMEM); |
1da177e4 LT |
387 | |
388 | if (new_page) | |
389 | page_cache_release(new_page); | |
390 | return found_page; | |
391 | } | |
46017e95 | 392 | |
579f8290 SL |
393 | static unsigned long swapin_nr_pages(unsigned long offset) |
394 | { | |
395 | static unsigned long prev_offset; | |
396 | unsigned int pages, max_pages, last_ra; | |
397 | static atomic_t last_readahead_pages; | |
398 | ||
399 | max_pages = 1 << ACCESS_ONCE(page_cluster); | |
400 | if (max_pages <= 1) | |
401 | return 1; | |
402 | ||
403 | /* | |
404 | * This heuristic has been found to work well on both sequential and | |
405 | * random loads, swapping to hard disk or to SSD: please don't ask | |
406 | * what the "+ 2" means, it just happens to work well, that's all. | |
407 | */ | |
408 | pages = atomic_xchg(&swapin_readahead_hits, 0) + 2; | |
409 | if (pages == 2) { | |
410 | /* | |
411 | * We can have no readahead hits to judge by: but must not get | |
412 | * stuck here forever, so check for an adjacent offset instead | |
413 | * (and don't even bother to check whether swap type is same). | |
414 | */ | |
415 | if (offset != prev_offset + 1 && offset != prev_offset - 1) | |
416 | pages = 1; | |
417 | prev_offset = offset; | |
418 | } else { | |
419 | unsigned int roundup = 4; | |
420 | while (roundup < pages) | |
421 | roundup <<= 1; | |
422 | pages = roundup; | |
423 | } | |
424 | ||
425 | if (pages > max_pages) | |
426 | pages = max_pages; | |
427 | ||
428 | /* Don't shrink readahead too fast */ | |
429 | last_ra = atomic_read(&last_readahead_pages) / 2; | |
430 | if (pages < last_ra) | |
431 | pages = last_ra; | |
432 | atomic_set(&last_readahead_pages, pages); | |
433 | ||
434 | return pages; | |
435 | } | |
436 | ||
46017e95 HD |
437 | /** |
438 | * swapin_readahead - swap in pages in hope we need them soon | |
439 | * @entry: swap entry of this memory | |
7682486b | 440 | * @gfp_mask: memory allocation flags |
46017e95 HD |
441 | * @vma: user vma this address belongs to |
442 | * @addr: target address for mempolicy | |
443 | * | |
444 | * Returns the struct page for entry and addr, after queueing swapin. | |
445 | * | |
446 | * Primitive swap readahead code. We simply read an aligned block of | |
447 | * (1 << page_cluster) entries in the swap area. This method is chosen | |
448 | * because it doesn't cost us any seek time. We also make sure to queue | |
449 | * the 'original' request together with the readahead ones... | |
450 | * | |
451 | * This has been extended to use the NUMA policies from the mm triggering | |
452 | * the readahead. | |
453 | * | |
454 | * Caller must hold down_read on the vma->vm_mm if vma is not NULL. | |
455 | */ | |
02098fea | 456 | struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, |
46017e95 HD |
457 | struct vm_area_struct *vma, unsigned long addr) |
458 | { | |
46017e95 | 459 | struct page *page; |
579f8290 SL |
460 | unsigned long entry_offset = swp_offset(entry); |
461 | unsigned long offset = entry_offset; | |
67f96aa2 | 462 | unsigned long start_offset, end_offset; |
579f8290 | 463 | unsigned long mask; |
3fb5c298 | 464 | struct blk_plug plug; |
46017e95 | 465 | |
579f8290 SL |
466 | mask = swapin_nr_pages(offset) - 1; |
467 | if (!mask) | |
468 | goto skip; | |
469 | ||
67f96aa2 RR |
470 | /* Read a page_cluster sized and aligned cluster around offset. */ |
471 | start_offset = offset & ~mask; | |
472 | end_offset = offset | mask; | |
473 | if (!start_offset) /* First page is swap header. */ | |
474 | start_offset++; | |
475 | ||
3fb5c298 | 476 | blk_start_plug(&plug); |
67f96aa2 | 477 | for (offset = start_offset; offset <= end_offset ; offset++) { |
46017e95 HD |
478 | /* Ok, do the async read-ahead now */ |
479 | page = read_swap_cache_async(swp_entry(swp_type(entry), offset), | |
02098fea | 480 | gfp_mask, vma, addr); |
46017e95 | 481 | if (!page) |
67f96aa2 | 482 | continue; |
579f8290 SL |
483 | if (offset != entry_offset) |
484 | SetPageReadahead(page); | |
46017e95 HD |
485 | page_cache_release(page); |
486 | } | |
3fb5c298 CE |
487 | blk_finish_plug(&plug); |
488 | ||
46017e95 | 489 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
579f8290 | 490 | skip: |
02098fea | 491 | return read_swap_cache_async(entry, gfp_mask, vma, addr); |
46017e95 | 492 | } |