mm: clarify __add_to_swap_cache locking
[linux-2.6.git] / mm / swap_state.c
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  */
9 #include <linux/module.h>
10 #include <linux/mm.h>
11 #include <linux/kernel_stat.h>
12 #include <linux/swap.h>
13 #include <linux/init.h>
14 #include <linux/pagemap.h>
15 #include <linux/buffer_head.h>
16 #include <linux/backing-dev.h>
17 #include <linux/pagevec.h>
18 #include <linux/migrate.h>
19
20 #include <asm/pgtable.h>
21
22 /*
23  * swapper_space is a fiction, retained to simplify the path through
24  * vmscan's shrink_page_list, to make sync_page look nicer, and to allow
25  * future use of radix_tree tags in the swap cache.
26  */
27 static const struct address_space_operations swap_aops = {
28         .writepage      = swap_writepage,
29         .sync_page      = block_sync_page,
30         .set_page_dirty = __set_page_dirty_nobuffers,
31         .migratepage    = migrate_page,
32 };
33
34 static struct backing_dev_info swap_backing_dev_info = {
35         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
36         .unplug_io_fn   = swap_unplug_io_fn,
37 };
38
39 struct address_space swapper_space = {
40         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
41         .tree_lock      = __RW_LOCK_UNLOCKED(swapper_space.tree_lock),
42         .a_ops          = &swap_aops,
43         .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear),
44         .backing_dev_info = &swap_backing_dev_info,
45 };
46
47 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
48
49 static struct {
50         unsigned long add_total;
51         unsigned long del_total;
52         unsigned long find_success;
53         unsigned long find_total;
54         unsigned long noent_race;
55         unsigned long exist_race;
56 } swap_cache_info;
57
58 void show_swap_cache_info(void)
59 {
60         printk("Swap cache: add %lu, delete %lu, find %lu/%lu, race %lu+%lu\n",
61                 swap_cache_info.add_total, swap_cache_info.del_total,
62                 swap_cache_info.find_success, swap_cache_info.find_total,
63                 swap_cache_info.noent_race, swap_cache_info.exist_race);
64         printk("Free swap  = %lukB\n", nr_swap_pages << (PAGE_SHIFT - 10));
65         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
66 }
67
68 /*
69  * __add_to_swap_cache resembles add_to_page_cache on swapper_space,
70  * but sets SwapCache flag and private instead of mapping and index.
71  */
72 static int __add_to_swap_cache(struct page *page, swp_entry_t entry,
73                                gfp_t gfp_mask)
74 {
75         int error;
76
77         BUG_ON(!PageLocked(page));
78         BUG_ON(PageSwapCache(page));
79         BUG_ON(PagePrivate(page));
80         error = radix_tree_preload(gfp_mask);
81         if (!error) {
82                 write_lock_irq(&swapper_space.tree_lock);
83                 error = radix_tree_insert(&swapper_space.page_tree,
84                                                 entry.val, page);
85                 if (!error) {
86                         page_cache_get(page);
87                         SetPageSwapCache(page);
88                         set_page_private(page, entry.val);
89                         total_swapcache_pages++;
90                         __inc_zone_page_state(page, NR_FILE_PAGES);
91                 }
92                 write_unlock_irq(&swapper_space.tree_lock);
93                 radix_tree_preload_end();
94         }
95         return error;
96 }
97
98 static int add_to_swap_cache(struct page *page, swp_entry_t entry)
99 {
100         int error;
101
102         BUG_ON(PageLocked(page));
103         if (!swap_duplicate(entry)) {
104                 INC_CACHE_INFO(noent_race);
105                 return -ENOENT;
106         }
107         SetPageLocked(page);
108         error = __add_to_swap_cache(page, entry, GFP_KERNEL);
109         /*
110          * Anon pages are already on the LRU, we don't run lru_cache_add here.
111          */
112         if (error) {
113                 ClearPageLocked(page);
114                 swap_free(entry);
115                 if (error == -EEXIST)
116                         INC_CACHE_INFO(exist_race);
117                 return error;
118         }
119         INC_CACHE_INFO(add_total);
120         return 0;
121 }
122
123 /*
124  * This must be called only on pages that have
125  * been verified to be in the swap cache.
126  */
127 void __delete_from_swap_cache(struct page *page)
128 {
129         BUG_ON(!PageLocked(page));
130         BUG_ON(!PageSwapCache(page));
131         BUG_ON(PageWriteback(page));
132         BUG_ON(PagePrivate(page));
133
134         radix_tree_delete(&swapper_space.page_tree, page_private(page));
135         set_page_private(page, 0);
136         ClearPageSwapCache(page);
137         total_swapcache_pages--;
138         __dec_zone_page_state(page, NR_FILE_PAGES);
139         INC_CACHE_INFO(del_total);
140 }
141
142 /**
143  * add_to_swap - allocate swap space for a page
144  * @page: page we want to move to swap
145  *
146  * Allocate swap space for the page and add the page to the
147  * swap cache.  Caller needs to hold the page lock. 
148  */
149 int add_to_swap(struct page * page, gfp_t gfp_mask)
150 {
151         swp_entry_t entry;
152         int err;
153
154         BUG_ON(!PageLocked(page));
155
156         for (;;) {
157                 entry = get_swap_page();
158                 if (!entry.val)
159                         return 0;
160
161                 /*
162                  * Radix-tree node allocations from PF_MEMALLOC contexts could
163                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
164                  * stops emergency reserves from being allocated.
165                  *
166                  * TODO: this could cause a theoretical memory reclaim
167                  * deadlock in the swap out path.
168                  */
169                 /*
170                  * Add it to the swap cache and mark it dirty
171                  */
172                 err = __add_to_swap_cache(page, entry,
173                                 gfp_mask|__GFP_NOMEMALLOC|__GFP_NOWARN);
174
175                 switch (err) {
176                 case 0:                         /* Success */
177                         SetPageUptodate(page);
178                         SetPageDirty(page);
179                         INC_CACHE_INFO(add_total);
180                         return 1;
181                 case -EEXIST:
182                         /* Raced with "speculative" read_swap_cache_async */
183                         INC_CACHE_INFO(exist_race);
184                         swap_free(entry);
185                         continue;
186                 default:
187                         /* -ENOMEM radix-tree allocation failure */
188                         swap_free(entry);
189                         return 0;
190                 }
191         }
192 }
193
194 /*
195  * This must be called only on pages that have
196  * been verified to be in the swap cache and locked.
197  * It will never put the page into the free list,
198  * the caller has a reference on the page.
199  */
200 void delete_from_swap_cache(struct page *page)
201 {
202         swp_entry_t entry;
203
204         entry.val = page_private(page);
205
206         write_lock_irq(&swapper_space.tree_lock);
207         __delete_from_swap_cache(page);
208         write_unlock_irq(&swapper_space.tree_lock);
209
210         swap_free(entry);
211         page_cache_release(page);
212 }
213
214 /*
215  * Strange swizzling function only for use by shmem_writepage
216  */
217 int move_to_swap_cache(struct page *page, swp_entry_t entry)
218 {
219         int err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
220         if (!err) {
221                 remove_from_page_cache(page);
222                 page_cache_release(page);       /* pagecache ref */
223                 if (!swap_duplicate(entry))
224                         BUG();
225                 SetPageDirty(page);
226                 INC_CACHE_INFO(add_total);
227         } else if (err == -EEXIST)
228                 INC_CACHE_INFO(exist_race);
229         return err;
230 }
231
232 /*
233  * Strange swizzling function for shmem_getpage (and shmem_unuse)
234  */
235 int move_from_swap_cache(struct page *page, unsigned long index,
236                 struct address_space *mapping)
237 {
238         int err = add_to_page_cache(page, mapping, index, GFP_ATOMIC);
239         if (!err) {
240                 delete_from_swap_cache(page);
241                 /* shift page from clean_pages to dirty_pages list */
242                 ClearPageDirty(page);
243                 set_page_dirty(page);
244         }
245         return err;
246 }
247
248 /* 
249  * If we are the only user, then try to free up the swap cache. 
250  * 
251  * Its ok to check for PageSwapCache without the page lock
252  * here because we are going to recheck again inside 
253  * exclusive_swap_page() _with_ the lock. 
254  *                                      - Marcelo
255  */
256 static inline void free_swap_cache(struct page *page)
257 {
258         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
259                 remove_exclusive_swap_page(page);
260                 unlock_page(page);
261         }
262 }
263
264 /* 
265  * Perform a free_page(), also freeing any swap cache associated with
266  * this page if it is the last user of the page.
267  */
268 void free_page_and_swap_cache(struct page *page)
269 {
270         free_swap_cache(page);
271         page_cache_release(page);
272 }
273
274 /*
275  * Passed an array of pages, drop them all from swapcache and then release
276  * them.  They are removed from the LRU and freed if this is their last use.
277  */
278 void free_pages_and_swap_cache(struct page **pages, int nr)
279 {
280         struct page **pagep = pages;
281
282         lru_add_drain();
283         while (nr) {
284                 int todo = min(nr, PAGEVEC_SIZE);
285                 int i;
286
287                 for (i = 0; i < todo; i++)
288                         free_swap_cache(pagep[i]);
289                 release_pages(pagep, todo, 0);
290                 pagep += todo;
291                 nr -= todo;
292         }
293 }
294
295 /*
296  * Lookup a swap entry in the swap cache. A found page will be returned
297  * unlocked and with its refcount incremented - we rely on the kernel
298  * lock getting page table operations atomic even if we drop the page
299  * lock before returning.
300  */
301 struct page * lookup_swap_cache(swp_entry_t entry)
302 {
303         struct page *page;
304
305         page = find_get_page(&swapper_space, entry.val);
306
307         if (page)
308                 INC_CACHE_INFO(find_success);
309
310         INC_CACHE_INFO(find_total);
311         return page;
312 }
313
314 /* 
315  * Locate a page of swap in physical memory, reserving swap cache space
316  * and reading the disk if it is not already cached.
317  * A failure return means that either the page allocation failed or that
318  * the swap entry is no longer in use.
319  */
320 struct page *read_swap_cache_async(swp_entry_t entry,
321                         struct vm_area_struct *vma, unsigned long addr)
322 {
323         struct page *found_page, *new_page = NULL;
324         int err;
325
326         do {
327                 /*
328                  * First check the swap cache.  Since this is normally
329                  * called after lookup_swap_cache() failed, re-calling
330                  * that would confuse statistics.
331                  */
332                 found_page = find_get_page(&swapper_space, entry.val);
333                 if (found_page)
334                         break;
335
336                 /*
337                  * Get a new page to read into from swap.
338                  */
339                 if (!new_page) {
340                         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
341                                                                 vma, addr);
342                         if (!new_page)
343                                 break;          /* Out of memory */
344                 }
345
346                 /*
347                  * Associate the page with swap entry in the swap cache.
348                  * May fail (-ENOENT) if swap entry has been freed since
349                  * our caller observed it.  May fail (-EEXIST) if there
350                  * is already a page associated with this entry in the
351                  * swap cache: added by a racing read_swap_cache_async,
352                  * or by try_to_swap_out (or shmem_writepage) re-using
353                  * the just freed swap entry for an existing page.
354                  * May fail (-ENOMEM) if radix-tree node allocation failed.
355                  */
356                 err = add_to_swap_cache(new_page, entry);
357                 if (!err) {
358                         /*
359                          * Initiate read into locked page and return.
360                          */
361                         lru_cache_add_active(new_page);
362                         swap_readpage(NULL, new_page);
363                         return new_page;
364                 }
365         } while (err != -ENOENT && err != -ENOMEM);
366
367         if (new_page)
368                 page_cache_release(new_page);
369         return found_page;
370 }