blob: 668a80422630f37c74f5247a3aea4913deea7829 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 */
9#include <linux/module.h>
10#include <linux/mm.h>
11#include <linux/kernel_stat.h>
12#include <linux/swap.h>
Hugh Dickins46017e92008-02-04 22:28:41 -080013#include <linux/swapops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/init.h>
15#include <linux/pagemap.h>
16#include <linux/buffer_head.h>
17#include <linux/backing-dev.h>
Hugh Dickinsc484d412006-01-06 00:10:55 -080018#include <linux/pagevec.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080019#include <linux/migrate.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020
21#include <asm/pgtable.h>
22
23/*
24 * swapper_space is a fiction, retained to simplify the path through
Anderson Briglia2706a1b2007-07-15 23:38:09 -070025 * vmscan's shrink_page_list, to make sync_page look nicer, and to allow
Linus Torvalds1da177e2005-04-16 15:20:36 -070026 * future use of radix_tree tags in the swap cache.
27 */
Christoph Hellwigf5e54d62006-06-28 04:26:44 -070028static const struct address_space_operations swap_aops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -070029 .writepage = swap_writepage,
30 .sync_page = block_sync_page,
31 .set_page_dirty = __set_page_dirty_nobuffers,
Christoph Lametere965f962006-02-01 03:05:41 -080032 .migratepage = migrate_page,
Linus Torvalds1da177e2005-04-16 15:20:36 -070033};
34
35static struct backing_dev_info swap_backing_dev_info = {
36 .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
37 .unplug_io_fn = swap_unplug_io_fn,
38};
39
40struct address_space swapper_space = {
41 .page_tree = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
Ingo Molnare4d91912006-07-03 00:24:34 -070042 .tree_lock = __RW_LOCK_UNLOCKED(swapper_space.tree_lock),
Linus Torvalds1da177e2005-04-16 15:20:36 -070043 .a_ops = &swap_aops,
44 .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear),
45 .backing_dev_info = &swap_backing_dev_info,
46};
Linus Torvalds1da177e2005-04-16 15:20:36 -070047
48#define INC_CACHE_INFO(x) do { swap_cache_info.x++; } while (0)
49
50static struct {
51 unsigned long add_total;
52 unsigned long del_total;
53 unsigned long find_success;
54 unsigned long find_total;
55 unsigned long noent_race;
56 unsigned long exist_race;
57} swap_cache_info;
58
59void show_swap_cache_info(void)
60{
61 printk("Swap cache: add %lu, delete %lu, find %lu/%lu, race %lu+%lu\n",
62 swap_cache_info.add_total, swap_cache_info.del_total,
63 swap_cache_info.find_success, swap_cache_info.find_total,
64 swap_cache_info.noent_race, swap_cache_info.exist_race);
65 printk("Free swap = %lukB\n", nr_swap_pages << (PAGE_SHIFT - 10));
66 printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
67}
68
69/*
70 * __add_to_swap_cache resembles add_to_page_cache on swapper_space,
71 * but sets SwapCache flag and private instead of mapping and index.
72 */
Victor Fusco9de75d12005-09-10 00:26:46 -070073static int __add_to_swap_cache(struct page *page, swp_entry_t entry,
Al Virodd0fc662005-10-07 07:46:04 +010074 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -070075{
76 int error;
77
Nick Pigginb55ed812007-10-16 01:24:42 -070078 BUG_ON(!PageLocked(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 BUG_ON(PageSwapCache(page));
80 BUG_ON(PagePrivate(page));
81 error = radix_tree_preload(gfp_mask);
82 if (!error) {
83 write_lock_irq(&swapper_space.tree_lock);
84 error = radix_tree_insert(&swapper_space.page_tree,
85 entry.val, page);
86 if (!error) {
87 page_cache_get(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 SetPageSwapCache(page);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -070089 set_page_private(page, entry.val);
Linus Torvalds1da177e2005-04-16 15:20:36 -070090 total_swapcache_pages++;
Christoph Lameter347ce432006-06-30 01:55:35 -070091 __inc_zone_page_state(page, NR_FILE_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -070092 }
93 write_unlock_irq(&swapper_space.tree_lock);
94 radix_tree_preload_end();
95 }
96 return error;
97}
98
99static int add_to_swap_cache(struct page *page, swp_entry_t entry)
100{
101 int error;
102
Nick Pigginb55ed812007-10-16 01:24:42 -0700103 BUG_ON(PageLocked(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104 if (!swap_duplicate(entry)) {
105 INC_CACHE_INFO(noent_race);
106 return -ENOENT;
107 }
Nick Pigginb55ed812007-10-16 01:24:42 -0700108 SetPageLocked(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109 error = __add_to_swap_cache(page, entry, GFP_KERNEL);
110 /*
111 * Anon pages are already on the LRU, we don't run lru_cache_add here.
112 */
113 if (error) {
Nick Pigginb55ed812007-10-16 01:24:42 -0700114 ClearPageLocked(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115 swap_free(entry);
116 if (error == -EEXIST)
117 INC_CACHE_INFO(exist_race);
118 return error;
119 }
120 INC_CACHE_INFO(add_total);
121 return 0;
122}
123
124/*
125 * This must be called only on pages that have
126 * been verified to be in the swap cache.
127 */
128void __delete_from_swap_cache(struct page *page)
129{
130 BUG_ON(!PageLocked(page));
131 BUG_ON(!PageSwapCache(page));
132 BUG_ON(PageWriteback(page));
Hugh Dickins3279ffd2005-09-03 15:54:43 -0700133 BUG_ON(PagePrivate(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700135 radix_tree_delete(&swapper_space.page_tree, page_private(page));
136 set_page_private(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 ClearPageSwapCache(page);
138 total_swapcache_pages--;
Christoph Lameter347ce432006-06-30 01:55:35 -0700139 __dec_zone_page_state(page, NR_FILE_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140 INC_CACHE_INFO(del_total);
141}
142
143/**
144 * add_to_swap - allocate swap space for a page
145 * @page: page we want to move to swap
146 *
147 * Allocate swap space for the page and add the page to the
148 * swap cache. Caller needs to hold the page lock.
149 */
Christoph Lameter1480a542006-01-08 01:00:53 -0800150int add_to_swap(struct page * page, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151{
152 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 int err;
154
Eric Sesterhenne74ca2b2006-04-01 01:25:12 +0200155 BUG_ON(!PageLocked(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156
157 for (;;) {
158 entry = get_swap_page();
159 if (!entry.val)
160 return 0;
161
Nick Pigginbd53b712005-05-01 08:58:37 -0700162 /*
163 * Radix-tree node allocations from PF_MEMALLOC contexts could
164 * completely exhaust the page allocator. __GFP_NOMEMALLOC
165 * stops emergency reserves from being allocated.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 *
Nick Pigginbd53b712005-05-01 08:58:37 -0700167 * TODO: this could cause a theoretical memory reclaim
168 * deadlock in the swap out path.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 /*
171 * Add it to the swap cache and mark it dirty
172 */
Nick Pigginbd53b712005-05-01 08:58:37 -0700173 err = __add_to_swap_cache(page, entry,
Christoph Lameter1480a542006-01-08 01:00:53 -0800174 gfp_mask|__GFP_NOMEMALLOC|__GFP_NOWARN);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175
176 switch (err) {
177 case 0: /* Success */
178 SetPageUptodate(page);
179 SetPageDirty(page);
180 INC_CACHE_INFO(add_total);
181 return 1;
182 case -EEXIST:
183 /* Raced with "speculative" read_swap_cache_async */
184 INC_CACHE_INFO(exist_race);
185 swap_free(entry);
186 continue;
187 default:
188 /* -ENOMEM radix-tree allocation failure */
189 swap_free(entry);
190 return 0;
191 }
192 }
193}
194
195/*
196 * This must be called only on pages that have
197 * been verified to be in the swap cache and locked.
198 * It will never put the page into the free list,
199 * the caller has a reference on the page.
200 */
201void delete_from_swap_cache(struct page *page)
202{
203 swp_entry_t entry;
204
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700205 entry.val = page_private(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206
207 write_lock_irq(&swapper_space.tree_lock);
208 __delete_from_swap_cache(page);
209 write_unlock_irq(&swapper_space.tree_lock);
210
211 swap_free(entry);
212 page_cache_release(page);
213}
214
215/*
216 * Strange swizzling function only for use by shmem_writepage
217 */
218int move_to_swap_cache(struct page *page, swp_entry_t entry)
219{
220 int err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
221 if (!err) {
222 remove_from_page_cache(page);
223 page_cache_release(page); /* pagecache ref */
224 if (!swap_duplicate(entry))
225 BUG();
226 SetPageDirty(page);
227 INC_CACHE_INFO(add_total);
228 } else if (err == -EEXIST)
229 INC_CACHE_INFO(exist_race);
230 return err;
231}
232
233/*
234 * Strange swizzling function for shmem_getpage (and shmem_unuse)
235 */
236int move_from_swap_cache(struct page *page, unsigned long index,
237 struct address_space *mapping)
238{
239 int err = add_to_page_cache(page, mapping, index, GFP_ATOMIC);
240 if (!err) {
241 delete_from_swap_cache(page);
242 /* shift page from clean_pages to dirty_pages list */
243 ClearPageDirty(page);
244 set_page_dirty(page);
245 }
246 return err;
247}
248
249/*
250 * If we are the only user, then try to free up the swap cache.
251 *
252 * Its ok to check for PageSwapCache without the page lock
253 * here because we are going to recheck again inside
254 * exclusive_swap_page() _with_ the lock.
255 * - Marcelo
256 */
257static inline void free_swap_cache(struct page *page)
258{
259 if (PageSwapCache(page) && !TestSetPageLocked(page)) {
260 remove_exclusive_swap_page(page);
261 unlock_page(page);
262 }
263}
264
265/*
266 * Perform a free_page(), also freeing any swap cache associated with
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700267 * this page if it is the last user of the page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 */
269void free_page_and_swap_cache(struct page *page)
270{
271 free_swap_cache(page);
272 page_cache_release(page);
273}
274
275/*
276 * Passed an array of pages, drop them all from swapcache and then release
277 * them. They are removed from the LRU and freed if this is their last use.
278 */
279void free_pages_and_swap_cache(struct page **pages, int nr)
280{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 struct page **pagep = pages;
282
283 lru_add_drain();
284 while (nr) {
Hugh Dickinsc484d412006-01-06 00:10:55 -0800285 int todo = min(nr, PAGEVEC_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286 int i;
287
288 for (i = 0; i < todo; i++)
289 free_swap_cache(pagep[i]);
290 release_pages(pagep, todo, 0);
291 pagep += todo;
292 nr -= todo;
293 }
294}
295
296/*
297 * Lookup a swap entry in the swap cache. A found page will be returned
298 * unlocked and with its refcount incremented - we rely on the kernel
299 * lock getting page table operations atomic even if we drop the page
300 * lock before returning.
301 */
302struct page * lookup_swap_cache(swp_entry_t entry)
303{
304 struct page *page;
305
306 page = find_get_page(&swapper_space, entry.val);
307
308 if (page)
309 INC_CACHE_INFO(find_success);
310
311 INC_CACHE_INFO(find_total);
312 return page;
313}
314
315/*
316 * Locate a page of swap in physical memory, reserving swap cache space
317 * and reading the disk if it is not already cached.
318 * A failure return means that either the page allocation failed or that
319 * the swap entry is no longer in use.
320 */
321struct page *read_swap_cache_async(swp_entry_t entry,
322 struct vm_area_struct *vma, unsigned long addr)
323{
324 struct page *found_page, *new_page = NULL;
325 int err;
326
327 do {
328 /*
329 * First check the swap cache. Since this is normally
330 * called after lookup_swap_cache() failed, re-calling
331 * that would confuse statistics.
332 */
333 found_page = find_get_page(&swapper_space, entry.val);
334 if (found_page)
335 break;
336
337 /*
338 * Get a new page to read into from swap.
339 */
340 if (!new_page) {
Mel Gorman769848c2007-07-17 04:03:05 -0700341 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
342 vma, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343 if (!new_page)
344 break; /* Out of memory */
345 }
346
347 /*
348 * Associate the page with swap entry in the swap cache.
349 * May fail (-ENOENT) if swap entry has been freed since
350 * our caller observed it. May fail (-EEXIST) if there
351 * is already a page associated with this entry in the
352 * swap cache: added by a racing read_swap_cache_async,
353 * or by try_to_swap_out (or shmem_writepage) re-using
354 * the just freed swap entry for an existing page.
355 * May fail (-ENOMEM) if radix-tree node allocation failed.
356 */
357 err = add_to_swap_cache(new_page, entry);
358 if (!err) {
359 /*
360 * Initiate read into locked page and return.
361 */
362 lru_cache_add_active(new_page);
363 swap_readpage(NULL, new_page);
364 return new_page;
365 }
366 } while (err != -ENOENT && err != -ENOMEM);
367
368 if (new_page)
369 page_cache_release(new_page);
370 return found_page;
371}
Hugh Dickins46017e92008-02-04 22:28:41 -0800372
373/**
374 * swapin_readahead - swap in pages in hope we need them soon
375 * @entry: swap entry of this memory
376 * @vma: user vma this address belongs to
377 * @addr: target address for mempolicy
378 *
379 * Returns the struct page for entry and addr, after queueing swapin.
380 *
381 * Primitive swap readahead code. We simply read an aligned block of
382 * (1 << page_cluster) entries in the swap area. This method is chosen
383 * because it doesn't cost us any seek time. We also make sure to queue
384 * the 'original' request together with the readahead ones...
385 *
386 * This has been extended to use the NUMA policies from the mm triggering
387 * the readahead.
388 *
389 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
390 */
391struct page *swapin_readahead(swp_entry_t entry,
392 struct vm_area_struct *vma, unsigned long addr)
393{
394 int nr_pages;
395 struct page *page;
396 unsigned long offset;
397 unsigned long end_offset;
398
399 /*
400 * Get starting offset for readaround, and number of pages to read.
401 * Adjust starting address by readbehind (for NUMA interleave case)?
402 * No, it's very unlikely that swap layout would follow vma layout,
403 * more likely that neighbouring swap pages came from the same node:
404 * so use the same "addr" to choose the same node for each swap read.
405 */
406 nr_pages = valid_swaphandles(entry, &offset);
407 for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
408 /* Ok, do the async read-ahead now */
409 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
410 vma, addr);
411 if (!page)
412 break;
413 page_cache_release(page);
414 }
415 lru_add_drain(); /* Push any new pages onto the LRU now */
416 return read_swap_cache_async(entry, vma, addr);
417}