memcg: fix deadlock between lock_page_cgroup and mapping tree_lock
[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/swapops.h>
14 #include <linux/init.h>
15 #include <linux/pagemap.h>
16 #include <linux/buffer_head.h>
17 #include <linux/backing-dev.h>
18 #include <linux/pagevec.h>
19 #include <linux/migrate.h>
20 #include <linux/page_cgroup.h>
21
22 #include <asm/pgtable.h>
23
24 /*
25  * swapper_space is a fiction, retained to simplify the path through
26  * vmscan's shrink_page_list, to make sync_page look nicer, and to allow
27  * future use of radix_tree tags in the swap cache.
28  */
29 static const struct address_space_operations swap_aops = {
30         .writepage      = swap_writepage,
31         .sync_page      = block_sync_page,
32         .set_page_dirty = __set_page_dirty_nobuffers,
33         .migratepage    = migrate_page,
34 };
35
36 static struct backing_dev_info swap_backing_dev_info = {
37         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
38         .unplug_io_fn   = swap_unplug_io_fn,
39 };
40
41 struct address_space swapper_space = {
42         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
43         .tree_lock      = __SPIN_LOCK_UNLOCKED(swapper_space.tree_lock),
44         .a_ops          = &swap_aops,
45         .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear),
46         .backing_dev_info = &swap_backing_dev_info,
47 };
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;
56 } swap_cache_info;
57
58 void show_swap_cache_info(void)
59 {
60         printk("%lu pages in swap cache\n", total_swapcache_pages);
61         printk("Swap cache stats: add %lu, delete %lu, find %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         printk("Free swap  = %ldkB\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_locked on swapper_space,
70  * but sets SwapCache flag and private instead of mapping and index.
71  */
72 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
73 {
74         int error;
75
76         VM_BUG_ON(!PageLocked(page));
77         VM_BUG_ON(PageSwapCache(page));
78         VM_BUG_ON(!PageSwapBacked(page));
79
80         error = radix_tree_preload(gfp_mask);
81         if (!error) {
82                 page_cache_get(page);
83                 SetPageSwapCache(page);
84                 set_page_private(page, entry.val);
85
86                 spin_lock_irq(&swapper_space.tree_lock);
87                 error = radix_tree_insert(&swapper_space.page_tree,
88                                                 entry.val, page);
89                 if (likely(!error)) {
90                         total_swapcache_pages++;
91                         __inc_zone_page_state(page, NR_FILE_PAGES);
92                         INC_CACHE_INFO(add_total);
93                 }
94                 spin_unlock_irq(&swapper_space.tree_lock);
95                 radix_tree_preload_end();
96
97                 if (unlikely(error)) {
98                         set_page_private(page, 0UL);
99                         ClearPageSwapCache(page);
100                         page_cache_release(page);
101                 }
102         }
103         return error;
104 }
105
106 /*
107  * This must be called only on pages that have
108  * been verified to be in the swap cache.
109  */
110 void __delete_from_swap_cache(struct page *page)
111 {
112         VM_BUG_ON(!PageLocked(page));
113         VM_BUG_ON(!PageSwapCache(page));
114         VM_BUG_ON(PageWriteback(page));
115
116         radix_tree_delete(&swapper_space.page_tree, page_private(page));
117         set_page_private(page, 0);
118         ClearPageSwapCache(page);
119         total_swapcache_pages--;
120         __dec_zone_page_state(page, NR_FILE_PAGES);
121         INC_CACHE_INFO(del_total);
122 }
123
124 /**
125  * add_to_swap - allocate swap space for a page
126  * @page: page we want to move to swap
127  * @gfp_mask: memory allocation flags
128  *
129  * Allocate swap space for the page and add the page to the
130  * swap cache.  Caller needs to hold the page lock. 
131  */
132 int add_to_swap(struct page *page)
133 {
134         swp_entry_t entry;
135         int err;
136
137         VM_BUG_ON(!PageLocked(page));
138         VM_BUG_ON(!PageUptodate(page));
139
140         for (;;) {
141                 entry = get_swap_page();
142                 if (!entry.val)
143                         return 0;
144
145                 /*
146                  * Radix-tree node allocations from PF_MEMALLOC contexts could
147                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
148                  * stops emergency reserves from being allocated.
149                  *
150                  * TODO: this could cause a theoretical memory reclaim
151                  * deadlock in the swap out path.
152                  */
153                 /*
154                  * Add it to the swap cache and mark it dirty
155                  */
156                 err = add_to_swap_cache(page, entry,
157                                 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
158
159                 switch (err) {
160                 case 0:                         /* Success */
161                         SetPageDirty(page);
162                         return 1;
163                 case -EEXIST:
164                         /* Raced with "speculative" read_swap_cache_async */
165                         swap_free(entry);
166                         continue;
167                 default:
168                         /* -ENOMEM radix-tree allocation failure */
169                         swap_free(entry);
170                         return 0;
171                 }
172         }
173 }
174
175 /*
176  * This must be called only on pages that have
177  * been verified to be in the swap cache and locked.
178  * It will never put the page into the free list,
179  * the caller has a reference on the page.
180  */
181 void delete_from_swap_cache(struct page *page)
182 {
183         swp_entry_t entry;
184
185         entry.val = page_private(page);
186
187         spin_lock_irq(&swapper_space.tree_lock);
188         __delete_from_swap_cache(page);
189         spin_unlock_irq(&swapper_space.tree_lock);
190
191         mem_cgroup_uncharge_swapcache(page, entry);
192         swap_free(entry);
193         page_cache_release(page);
194 }
195
196 /* 
197  * If we are the only user, then try to free up the swap cache. 
198  * 
199  * Its ok to check for PageSwapCache without the page lock
200  * here because we are going to recheck again inside
201  * try_to_free_swap() _with_ the lock.
202  *                                      - Marcelo
203  */
204 static inline void free_swap_cache(struct page *page)
205 {
206         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
207                 try_to_free_swap(page);
208                 unlock_page(page);
209         }
210 }
211
212 /* 
213  * Perform a free_page(), also freeing any swap cache associated with
214  * this page if it is the last user of the page.
215  */
216 void free_page_and_swap_cache(struct page *page)
217 {
218         free_swap_cache(page);
219         page_cache_release(page);
220 }
221
222 /*
223  * Passed an array of pages, drop them all from swapcache and then release
224  * them.  They are removed from the LRU and freed if this is their last use.
225  */
226 void free_pages_and_swap_cache(struct page **pages, int nr)
227 {
228         struct page **pagep = pages;
229
230         lru_add_drain();
231         while (nr) {
232                 int todo = min(nr, PAGEVEC_SIZE);
233                 int i;
234
235                 for (i = 0; i < todo; i++)
236                         free_swap_cache(pagep[i]);
237                 release_pages(pagep, todo, 0);
238                 pagep += todo;
239                 nr -= todo;
240         }
241 }
242
243 /*
244  * Lookup a swap entry in the swap cache. A found page will be returned
245  * unlocked and with its refcount incremented - we rely on the kernel
246  * lock getting page table operations atomic even if we drop the page
247  * lock before returning.
248  */
249 struct page * lookup_swap_cache(swp_entry_t entry)
250 {
251         struct page *page;
252
253         page = find_get_page(&swapper_space, entry.val);
254
255         if (page)
256                 INC_CACHE_INFO(find_success);
257
258         INC_CACHE_INFO(find_total);
259         return page;
260 }
261
262 /* 
263  * Locate a page of swap in physical memory, reserving swap cache space
264  * and reading the disk if it is not already cached.
265  * A failure return means that either the page allocation failed or that
266  * the swap entry is no longer in use.
267  */
268 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
269                         struct vm_area_struct *vma, unsigned long addr)
270 {
271         struct page *found_page, *new_page = NULL;
272         int err;
273
274         do {
275                 /*
276                  * First check the swap cache.  Since this is normally
277                  * called after lookup_swap_cache() failed, re-calling
278                  * that would confuse statistics.
279                  */
280                 found_page = find_get_page(&swapper_space, entry.val);
281                 if (found_page)
282                         break;
283
284                 /*
285                  * Get a new page to read into from swap.
286                  */
287                 if (!new_page) {
288                         new_page = alloc_page_vma(gfp_mask, vma, addr);
289                         if (!new_page)
290                                 break;          /* Out of memory */
291                 }
292
293                 /*
294                  * Swap entry may have been freed since our caller observed it.
295                  */
296                 if (!swap_duplicate(entry))
297                         break;
298
299                 /*
300                  * Associate the page with swap entry in the swap cache.
301                  * May fail (-EEXIST) if there is already a page associated
302                  * with this entry in the swap cache: added by a racing
303                  * read_swap_cache_async, or add_to_swap or shmem_writepage
304                  * re-using the just freed swap entry for an existing page.
305                  * May fail (-ENOMEM) if radix-tree node allocation failed.
306                  */
307                 __set_page_locked(new_page);
308                 SetPageSwapBacked(new_page);
309                 err = add_to_swap_cache(new_page, entry, gfp_mask & GFP_KERNEL);
310                 if (likely(!err)) {
311                         /*
312                          * Initiate read into locked page and return.
313                          */
314                         lru_cache_add_anon(new_page);
315                         swap_readpage(NULL, new_page);
316                         return new_page;
317                 }
318                 ClearPageSwapBacked(new_page);
319                 __clear_page_locked(new_page);
320                 swap_free(entry);
321         } while (err != -ENOMEM);
322
323         if (new_page)
324                 page_cache_release(new_page);
325         return found_page;
326 }
327
328 /**
329  * swapin_readahead - swap in pages in hope we need them soon
330  * @entry: swap entry of this memory
331  * @gfp_mask: memory allocation flags
332  * @vma: user vma this address belongs to
333  * @addr: target address for mempolicy
334  *
335  * Returns the struct page for entry and addr, after queueing swapin.
336  *
337  * Primitive swap readahead code. We simply read an aligned block of
338  * (1 << page_cluster) entries in the swap area. This method is chosen
339  * because it doesn't cost us any seek time.  We also make sure to queue
340  * the 'original' request together with the readahead ones...
341  *
342  * This has been extended to use the NUMA policies from the mm triggering
343  * the readahead.
344  *
345  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
346  */
347 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
348                         struct vm_area_struct *vma, unsigned long addr)
349 {
350         int nr_pages;
351         struct page *page;
352         unsigned long offset;
353         unsigned long end_offset;
354
355         /*
356          * Get starting offset for readaround, and number of pages to read.
357          * Adjust starting address by readbehind (for NUMA interleave case)?
358          * No, it's very unlikely that swap layout would follow vma layout,
359          * more likely that neighbouring swap pages came from the same node:
360          * so use the same "addr" to choose the same node for each swap read.
361          */
362         nr_pages = valid_swaphandles(entry, &offset);
363         for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
364                 /* Ok, do the async read-ahead now */
365                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
366                                                 gfp_mask, vma, addr);
367                 if (!page)
368                         break;
369                 page_cache_release(page);
370         }
371         lru_add_drain();        /* Push any new pages onto the LRU now */
372         return read_swap_cache_async(entry, gfp_mask, vma, addr);
373 }