[PATCH] page migration reorg
[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_list, to make sync_page look nicer, and to allow
25  * future use of radix_tree tags in the swap cache.
26  */
27 static 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,
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(PageSwapCache(page));
78         BUG_ON(PagePrivate(page));
79         error = radix_tree_preload(gfp_mask);
80         if (!error) {
81                 write_lock_irq(&swapper_space.tree_lock);
82                 error = radix_tree_insert(&swapper_space.page_tree,
83                                                 entry.val, page);
84                 if (!error) {
85                         page_cache_get(page);
86                         SetPageLocked(page);
87                         SetPageSwapCache(page);
88                         set_page_private(page, entry.val);
89                         total_swapcache_pages++;
90                         pagecache_acct(1);
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         if (!swap_duplicate(entry)) {
103                 INC_CACHE_INFO(noent_race);
104                 return -ENOENT;
105         }
106         error = __add_to_swap_cache(page, entry, GFP_KERNEL);
107         /*
108          * Anon pages are already on the LRU, we don't run lru_cache_add here.
109          */
110         if (error) {
111                 swap_free(entry);
112                 if (error == -EEXIST)
113                         INC_CACHE_INFO(exist_race);
114                 return error;
115         }
116         INC_CACHE_INFO(add_total);
117         return 0;
118 }
119
120 /*
121  * This must be called only on pages that have
122  * been verified to be in the swap cache.
123  */
124 void __delete_from_swap_cache(struct page *page)
125 {
126         BUG_ON(!PageLocked(page));
127         BUG_ON(!PageSwapCache(page));
128         BUG_ON(PageWriteback(page));
129         BUG_ON(PagePrivate(page));
130
131         radix_tree_delete(&swapper_space.page_tree, page_private(page));
132         set_page_private(page, 0);
133         ClearPageSwapCache(page);
134         total_swapcache_pages--;
135         pagecache_acct(-1);
136         INC_CACHE_INFO(del_total);
137 }
138
139 /**
140  * add_to_swap - allocate swap space for a page
141  * @page: page we want to move to swap
142  *
143  * Allocate swap space for the page and add the page to the
144  * swap cache.  Caller needs to hold the page lock. 
145  */
146 int add_to_swap(struct page * page, gfp_t gfp_mask)
147 {
148         swp_entry_t entry;
149         int err;
150
151         if (!PageLocked(page))
152                 BUG();
153
154         for (;;) {
155                 entry = get_swap_page();
156                 if (!entry.val)
157                         return 0;
158
159                 /*
160                  * Radix-tree node allocations from PF_MEMALLOC contexts could
161                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
162                  * stops emergency reserves from being allocated.
163                  *
164                  * TODO: this could cause a theoretical memory reclaim
165                  * deadlock in the swap out path.
166                  */
167                 /*
168                  * Add it to the swap cache and mark it dirty
169                  */
170                 err = __add_to_swap_cache(page, entry,
171                                 gfp_mask|__GFP_NOMEMALLOC|__GFP_NOWARN);
172
173                 switch (err) {
174                 case 0:                         /* Success */
175                         SetPageUptodate(page);
176                         SetPageDirty(page);
177                         INC_CACHE_INFO(add_total);
178                         return 1;
179                 case -EEXIST:
180                         /* Raced with "speculative" read_swap_cache_async */
181                         INC_CACHE_INFO(exist_race);
182                         swap_free(entry);
183                         continue;
184                 default:
185                         /* -ENOMEM radix-tree allocation failure */
186                         swap_free(entry);
187                         return 0;
188                 }
189         }
190 }
191
192 /*
193  * This must be called only on pages that have
194  * been verified to be in the swap cache and locked.
195  * It will never put the page into the free list,
196  * the caller has a reference on the page.
197  */
198 void delete_from_swap_cache(struct page *page)
199 {
200         swp_entry_t entry;
201
202         entry.val = page_private(page);
203
204         write_lock_irq(&swapper_space.tree_lock);
205         __delete_from_swap_cache(page);
206         write_unlock_irq(&swapper_space.tree_lock);
207
208         swap_free(entry);
209         page_cache_release(page);
210 }
211
212 /*
213  * Strange swizzling function only for use by shmem_writepage
214  */
215 int move_to_swap_cache(struct page *page, swp_entry_t entry)
216 {
217         int err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
218         if (!err) {
219                 remove_from_page_cache(page);
220                 page_cache_release(page);       /* pagecache ref */
221                 if (!swap_duplicate(entry))
222                         BUG();
223                 SetPageDirty(page);
224                 INC_CACHE_INFO(add_total);
225         } else if (err == -EEXIST)
226                 INC_CACHE_INFO(exist_race);
227         return err;
228 }
229
230 /*
231  * Strange swizzling function for shmem_getpage (and shmem_unuse)
232  */
233 int move_from_swap_cache(struct page *page, unsigned long index,
234                 struct address_space *mapping)
235 {
236         int err = add_to_page_cache(page, mapping, index, GFP_ATOMIC);
237         if (!err) {
238                 delete_from_swap_cache(page);
239                 /* shift page from clean_pages to dirty_pages list */
240                 ClearPageDirty(page);
241                 set_page_dirty(page);
242         }
243         return err;
244 }
245
246 /* 
247  * If we are the only user, then try to free up the swap cache. 
248  * 
249  * Its ok to check for PageSwapCache without the page lock
250  * here because we are going to recheck again inside 
251  * exclusive_swap_page() _with_ the lock. 
252  *                                      - Marcelo
253  */
254 static inline void free_swap_cache(struct page *page)
255 {
256         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
257                 remove_exclusive_swap_page(page);
258                 unlock_page(page);
259         }
260 }
261
262 /* 
263  * Perform a free_page(), also freeing any swap cache associated with
264  * this page if it is the last user of the page.
265  */
266 void free_page_and_swap_cache(struct page *page)
267 {
268         free_swap_cache(page);
269         page_cache_release(page);
270 }
271
272 /*
273  * Passed an array of pages, drop them all from swapcache and then release
274  * them.  They are removed from the LRU and freed if this is their last use.
275  */
276 void free_pages_and_swap_cache(struct page **pages, int nr)
277 {
278         struct page **pagep = pages;
279
280         lru_add_drain();
281         while (nr) {
282                 int todo = min(nr, PAGEVEC_SIZE);
283                 int i;
284
285                 for (i = 0; i < todo; i++)
286                         free_swap_cache(pagep[i]);
287                 release_pages(pagep, todo, 0);
288                 pagep += todo;
289                 nr -= todo;
290         }
291 }
292
293 /*
294  * Lookup a swap entry in the swap cache. A found page will be returned
295  * unlocked and with its refcount incremented - we rely on the kernel
296  * lock getting page table operations atomic even if we drop the page
297  * lock before returning.
298  */
299 struct page * lookup_swap_cache(swp_entry_t entry)
300 {
301         struct page *page;
302
303         page = find_get_page(&swapper_space, entry.val);
304
305         if (page)
306                 INC_CACHE_INFO(find_success);
307
308         INC_CACHE_INFO(find_total);
309         return page;
310 }
311
312 /* 
313  * Locate a page of swap in physical memory, reserving swap cache space
314  * and reading the disk if it is not already cached.
315  * A failure return means that either the page allocation failed or that
316  * the swap entry is no longer in use.
317  */
318 struct page *read_swap_cache_async(swp_entry_t entry,
319                         struct vm_area_struct *vma, unsigned long addr)
320 {
321         struct page *found_page, *new_page = NULL;
322         int err;
323
324         do {
325                 /*
326                  * First check the swap cache.  Since this is normally
327                  * called after lookup_swap_cache() failed, re-calling
328                  * that would confuse statistics.
329                  */
330                 found_page = find_get_page(&swapper_space, entry.val);
331                 if (found_page)
332                         break;
333
334                 /*
335                  * Get a new page to read into from swap.
336                  */
337                 if (!new_page) {
338                         new_page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
339                         if (!new_page)
340                                 break;          /* Out of memory */
341                 }
342
343                 /*
344                  * Associate the page with swap entry in the swap cache.
345                  * May fail (-ENOENT) if swap entry has been freed since
346                  * our caller observed it.  May fail (-EEXIST) if there
347                  * is already a page associated with this entry in the
348                  * swap cache: added by a racing read_swap_cache_async,
349                  * or by try_to_swap_out (or shmem_writepage) re-using
350                  * the just freed swap entry for an existing page.
351                  * May fail (-ENOMEM) if radix-tree node allocation failed.
352                  */
353                 err = add_to_swap_cache(new_page, entry);
354                 if (!err) {
355                         /*
356                          * Initiate read into locked page and return.
357                          */
358                         lru_cache_add_active(new_page);
359                         swap_readpage(NULL, new_page);
360                         return new_page;
361                 }
362         } while (err != -ENOENT && err != -ENOMEM);
363
364         if (new_page)
365                 page_cache_release(new_page);
366         return found_page;
367 }