[PATCH] mm: split page table 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/init.h>
14 #include <linux/pagemap.h>
15 #include <linux/buffer_head.h>
16 #include <linux/backing-dev.h>
17
18 #include <asm/pgtable.h>
19
20 /*
21  * swapper_space is a fiction, retained to simplify the path through
22  * vmscan's shrink_list, to make sync_page look nicer, and to allow
23  * future use of radix_tree tags in the swap cache.
24  */
25 static struct address_space_operations swap_aops = {
26         .writepage      = swap_writepage,
27         .sync_page      = block_sync_page,
28         .set_page_dirty = __set_page_dirty_nobuffers,
29 };
30
31 static struct backing_dev_info swap_backing_dev_info = {
32         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
33         .unplug_io_fn   = swap_unplug_io_fn,
34 };
35
36 struct address_space swapper_space = {
37         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
38         .tree_lock      = RW_LOCK_UNLOCKED,
39         .a_ops          = &swap_aops,
40         .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear),
41         .backing_dev_info = &swap_backing_dev_info,
42 };
43 EXPORT_SYMBOL(swapper_space);
44
45 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
46
47 static struct {
48         unsigned long add_total;
49         unsigned long del_total;
50         unsigned long find_success;
51         unsigned long find_total;
52         unsigned long noent_race;
53         unsigned long exist_race;
54 } swap_cache_info;
55
56 void show_swap_cache_info(void)
57 {
58         printk("Swap cache: add %lu, delete %lu, find %lu/%lu, race %lu+%lu\n",
59                 swap_cache_info.add_total, swap_cache_info.del_total,
60                 swap_cache_info.find_success, swap_cache_info.find_total,
61                 swap_cache_info.noent_race, swap_cache_info.exist_race);
62         printk("Free swap  = %lukB\n", nr_swap_pages << (PAGE_SHIFT - 10));
63         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
64 }
65
66 /*
67  * __add_to_swap_cache resembles add_to_page_cache on swapper_space,
68  * but sets SwapCache flag and private instead of mapping and index.
69  */
70 static int __add_to_swap_cache(struct page *page, swp_entry_t entry,
71                                gfp_t gfp_mask)
72 {
73         int error;
74
75         BUG_ON(PageSwapCache(page));
76         BUG_ON(PagePrivate(page));
77         error = radix_tree_preload(gfp_mask);
78         if (!error) {
79                 write_lock_irq(&swapper_space.tree_lock);
80                 error = radix_tree_insert(&swapper_space.page_tree,
81                                                 entry.val, page);
82                 if (!error) {
83                         page_cache_get(page);
84                         SetPageLocked(page);
85                         SetPageSwapCache(page);
86                         set_page_private(page, entry.val);
87                         total_swapcache_pages++;
88                         pagecache_acct(1);
89                 }
90                 write_unlock_irq(&swapper_space.tree_lock);
91                 radix_tree_preload_end();
92         }
93         return error;
94 }
95
96 static int add_to_swap_cache(struct page *page, swp_entry_t entry)
97 {
98         int error;
99
100         if (!swap_duplicate(entry)) {
101                 INC_CACHE_INFO(noent_race);
102                 return -ENOENT;
103         }
104         error = __add_to_swap_cache(page, entry, GFP_KERNEL);
105         /*
106          * Anon pages are already on the LRU, we don't run lru_cache_add here.
107          */
108         if (error) {
109                 swap_free(entry);
110                 if (error == -EEXIST)
111                         INC_CACHE_INFO(exist_race);
112                 return error;
113         }
114         INC_CACHE_INFO(add_total);
115         return 0;
116 }
117
118 /*
119  * This must be called only on pages that have
120  * been verified to be in the swap cache.
121  */
122 void __delete_from_swap_cache(struct page *page)
123 {
124         BUG_ON(!PageLocked(page));
125         BUG_ON(!PageSwapCache(page));
126         BUG_ON(PageWriteback(page));
127         BUG_ON(PagePrivate(page));
128
129         radix_tree_delete(&swapper_space.page_tree, page_private(page));
130         set_page_private(page, 0);
131         ClearPageSwapCache(page);
132         total_swapcache_pages--;
133         pagecache_acct(-1);
134         INC_CACHE_INFO(del_total);
135 }
136
137 /**
138  * add_to_swap - allocate swap space for a page
139  * @page: page we want to move to swap
140  *
141  * Allocate swap space for the page and add the page to the
142  * swap cache.  Caller needs to hold the page lock. 
143  */
144 int add_to_swap(struct page * page)
145 {
146         swp_entry_t entry;
147         int err;
148
149         if (!PageLocked(page))
150                 BUG();
151
152         for (;;) {
153                 entry = get_swap_page();
154                 if (!entry.val)
155                         return 0;
156
157                 /*
158                  * Radix-tree node allocations from PF_MEMALLOC contexts could
159                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
160                  * stops emergency reserves from being allocated.
161                  *
162                  * TODO: this could cause a theoretical memory reclaim
163                  * deadlock in the swap out path.
164                  */
165                 /*
166                  * Add it to the swap cache and mark it dirty
167                  */
168                 err = __add_to_swap_cache(page, entry,
169                                 GFP_ATOMIC|__GFP_NOMEMALLOC|__GFP_NOWARN);
170
171                 switch (err) {
172                 case 0:                         /* Success */
173                         SetPageUptodate(page);
174                         SetPageDirty(page);
175                         INC_CACHE_INFO(add_total);
176                         return 1;
177                 case -EEXIST:
178                         /* Raced with "speculative" read_swap_cache_async */
179                         INC_CACHE_INFO(exist_race);
180                         swap_free(entry);
181                         continue;
182                 default:
183                         /* -ENOMEM radix-tree allocation failure */
184                         swap_free(entry);
185                         return 0;
186                 }
187         }
188 }
189
190 /*
191  * This must be called only on pages that have
192  * been verified to be in the swap cache and locked.
193  * It will never put the page into the free list,
194  * the caller has a reference on the page.
195  */
196 void delete_from_swap_cache(struct page *page)
197 {
198         swp_entry_t entry;
199
200         entry.val = page_private(page);
201
202         write_lock_irq(&swapper_space.tree_lock);
203         __delete_from_swap_cache(page);
204         write_unlock_irq(&swapper_space.tree_lock);
205
206         swap_free(entry);
207         page_cache_release(page);
208 }
209
210 /*
211  * Strange swizzling function only for use by shmem_writepage
212  */
213 int move_to_swap_cache(struct page *page, swp_entry_t entry)
214 {
215         int err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
216         if (!err) {
217                 remove_from_page_cache(page);
218                 page_cache_release(page);       /* pagecache ref */
219                 if (!swap_duplicate(entry))
220                         BUG();
221                 SetPageDirty(page);
222                 INC_CACHE_INFO(add_total);
223         } else if (err == -EEXIST)
224                 INC_CACHE_INFO(exist_race);
225         return err;
226 }
227
228 /*
229  * Strange swizzling function for shmem_getpage (and shmem_unuse)
230  */
231 int move_from_swap_cache(struct page *page, unsigned long index,
232                 struct address_space *mapping)
233 {
234         int err = add_to_page_cache(page, mapping, index, GFP_ATOMIC);
235         if (!err) {
236                 delete_from_swap_cache(page);
237                 /* shift page from clean_pages to dirty_pages list */
238                 ClearPageDirty(page);
239                 set_page_dirty(page);
240         }
241         return err;
242 }
243
244 /* 
245  * If we are the only user, then try to free up the swap cache. 
246  * 
247  * Its ok to check for PageSwapCache without the page lock
248  * here because we are going to recheck again inside 
249  * exclusive_swap_page() _with_ the lock. 
250  *                                      - Marcelo
251  */
252 static inline void free_swap_cache(struct page *page)
253 {
254         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
255                 remove_exclusive_swap_page(page);
256                 unlock_page(page);
257         }
258 }
259
260 /* 
261  * Perform a free_page(), also freeing any swap cache associated with
262  * this page if it is the last user of the page. Can not do a lock_page,
263  * as we are holding the page_table_lock spinlock.
264  */
265 void free_page_and_swap_cache(struct page *page)
266 {
267         free_swap_cache(page);
268         page_cache_release(page);
269 }
270
271 /*
272  * Passed an array of pages, drop them all from swapcache and then release
273  * them.  They are removed from the LRU and freed if this is their last use.
274  */
275 void free_pages_and_swap_cache(struct page **pages, int nr)
276 {
277         int chunk = 16;
278         struct page **pagep = pages;
279
280         lru_add_drain();
281         while (nr) {
282                 int todo = min(chunk, nr);
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 }