mm: migration: add migrate_entry_wait_huge()
[linux-3.10.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/mm.h>
10 #include <linux/gfp.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/backing-dev.h>
17 #include <linux/blkdev.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.
27  */
28 static const struct address_space_operations swap_aops = {
29         .writepage      = swap_writepage,
30         .set_page_dirty = swap_set_page_dirty,
31         .migratepage    = migrate_page,
32 };
33
34 static struct backing_dev_info swap_backing_dev_info = {
35         .name           = "swap",
36         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
37 };
38
39 struct address_space swapper_spaces[MAX_SWAPFILES] = {
40         [0 ... MAX_SWAPFILES - 1] = {
41                 .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
42                 .a_ops          = &swap_aops,
43                 .backing_dev_info = &swap_backing_dev_info,
44         }
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 } swap_cache_info;
55
56 unsigned long total_swapcache_pages(void)
57 {
58         int i;
59         unsigned long ret = 0;
60
61         for (i = 0; i < MAX_SWAPFILES; i++)
62                 ret += swapper_spaces[i].nrpages;
63         return ret;
64 }
65
66 void show_swap_cache_info(void)
67 {
68         printk("%lu pages in swap cache\n", total_swapcache_pages());
69         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
70                 swap_cache_info.add_total, swap_cache_info.del_total,
71                 swap_cache_info.find_success, swap_cache_info.find_total);
72         printk("Free swap  = %ldkB\n",
73                 get_nr_swap_pages() << (PAGE_SHIFT - 10));
74         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
75 }
76
77 /*
78  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
79  * but sets SwapCache flag and private instead of mapping and index.
80  */
81 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
82 {
83         int error;
84         struct address_space *address_space;
85
86         VM_BUG_ON(!PageLocked(page));
87         VM_BUG_ON(PageSwapCache(page));
88         VM_BUG_ON(!PageSwapBacked(page));
89
90         page_cache_get(page);
91         SetPageSwapCache(page);
92         set_page_private(page, entry.val);
93
94         address_space = swap_address_space(entry);
95         spin_lock_irq(&address_space->tree_lock);
96         error = radix_tree_insert(&address_space->page_tree,
97                                         entry.val, page);
98         if (likely(!error)) {
99                 address_space->nrpages++;
100                 __inc_zone_page_state(page, NR_FILE_PAGES);
101                 INC_CACHE_INFO(add_total);
102         }
103         spin_unlock_irq(&address_space->tree_lock);
104
105         if (unlikely(error)) {
106                 /*
107                  * Only the context which have set SWAP_HAS_CACHE flag
108                  * would call add_to_swap_cache().
109                  * So add_to_swap_cache() doesn't returns -EEXIST.
110                  */
111                 VM_BUG_ON(error == -EEXIST);
112                 set_page_private(page, 0UL);
113                 ClearPageSwapCache(page);
114                 page_cache_release(page);
115         }
116
117         return error;
118 }
119
120
121 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
122 {
123         int error;
124
125         error = radix_tree_preload(gfp_mask);
126         if (!error) {
127                 error = __add_to_swap_cache(page, entry);
128                 radix_tree_preload_end();
129         }
130         return error;
131 }
132
133 /*
134  * This must be called only on pages that have
135  * been verified to be in the swap cache.
136  */
137 void __delete_from_swap_cache(struct page *page)
138 {
139         swp_entry_t entry;
140         struct address_space *address_space;
141
142         VM_BUG_ON(!PageLocked(page));
143         VM_BUG_ON(!PageSwapCache(page));
144         VM_BUG_ON(PageWriteback(page));
145
146         entry.val = page_private(page);
147         address_space = swap_address_space(entry);
148         radix_tree_delete(&address_space->page_tree, page_private(page));
149         set_page_private(page, 0);
150         ClearPageSwapCache(page);
151         address_space->nrpages--;
152         __dec_zone_page_state(page, NR_FILE_PAGES);
153         INC_CACHE_INFO(del_total);
154 }
155
156 /**
157  * add_to_swap - allocate swap space for a page
158  * @page: page we want to move to swap
159  *
160  * Allocate swap space for the page and add the page to the
161  * swap cache.  Caller needs to hold the page lock. 
162  */
163 int add_to_swap(struct page *page, struct list_head *list)
164 {
165         swp_entry_t entry;
166         int err;
167
168         VM_BUG_ON(!PageLocked(page));
169         VM_BUG_ON(!PageUptodate(page));
170
171         entry = get_swap_page();
172         if (!entry.val)
173                 return 0;
174
175         if (unlikely(PageTransHuge(page)))
176                 if (unlikely(split_huge_page_to_list(page, list))) {
177                         swapcache_free(entry, NULL);
178                         return 0;
179                 }
180
181         /*
182          * Radix-tree node allocations from PF_MEMALLOC contexts could
183          * completely exhaust the page allocator. __GFP_NOMEMALLOC
184          * stops emergency reserves from being allocated.
185          *
186          * TODO: this could cause a theoretical memory reclaim
187          * deadlock in the swap out path.
188          */
189         /*
190          * Add it to the swap cache and mark it dirty
191          */
192         err = add_to_swap_cache(page, entry,
193                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
194
195         if (!err) {     /* Success */
196                 SetPageDirty(page);
197                 return 1;
198         } else {        /* -ENOMEM radix-tree allocation failure */
199                 /*
200                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
201                  * clear SWAP_HAS_CACHE flag.
202                  */
203                 swapcache_free(entry, NULL);
204                 return 0;
205         }
206 }
207
208 /*
209  * This must be called only on pages that have
210  * been verified to be in the swap cache and locked.
211  * It will never put the page into the free list,
212  * the caller has a reference on the page.
213  */
214 void delete_from_swap_cache(struct page *page)
215 {
216         swp_entry_t entry;
217         struct address_space *address_space;
218
219         entry.val = page_private(page);
220
221         address_space = swap_address_space(entry);
222         spin_lock_irq(&address_space->tree_lock);
223         __delete_from_swap_cache(page);
224         spin_unlock_irq(&address_space->tree_lock);
225
226         swapcache_free(entry, page);
227         page_cache_release(page);
228 }
229
230 /* 
231  * If we are the only user, then try to free up the swap cache. 
232  * 
233  * Its ok to check for PageSwapCache without the page lock
234  * here because we are going to recheck again inside
235  * try_to_free_swap() _with_ the lock.
236  *                                      - Marcelo
237  */
238 static inline void free_swap_cache(struct page *page)
239 {
240         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
241                 try_to_free_swap(page);
242                 unlock_page(page);
243         }
244 }
245
246 /* 
247  * Perform a free_page(), also freeing any swap cache associated with
248  * this page if it is the last user of the page.
249  */
250 void free_page_and_swap_cache(struct page *page)
251 {
252         free_swap_cache(page);
253         page_cache_release(page);
254 }
255
256 /*
257  * Passed an array of pages, drop them all from swapcache and then release
258  * them.  They are removed from the LRU and freed if this is their last use.
259  */
260 void free_pages_and_swap_cache(struct page **pages, int nr)
261 {
262         struct page **pagep = pages;
263
264         lru_add_drain();
265         while (nr) {
266                 int todo = min(nr, PAGEVEC_SIZE);
267                 int i;
268
269                 for (i = 0; i < todo; i++)
270                         free_swap_cache(pagep[i]);
271                 release_pages(pagep, todo, 0);
272                 pagep += todo;
273                 nr -= todo;
274         }
275 }
276
277 /*
278  * Lookup a swap entry in the swap cache. A found page will be returned
279  * unlocked and with its refcount incremented - we rely on the kernel
280  * lock getting page table operations atomic even if we drop the page
281  * lock before returning.
282  */
283 struct page * lookup_swap_cache(swp_entry_t entry)
284 {
285         struct page *page;
286
287         page = find_get_page(swap_address_space(entry), entry.val);
288
289         if (page)
290                 INC_CACHE_INFO(find_success);
291
292         INC_CACHE_INFO(find_total);
293         return page;
294 }
295
296 /* 
297  * Locate a page of swap in physical memory, reserving swap cache space
298  * and reading the disk if it is not already cached.
299  * A failure return means that either the page allocation failed or that
300  * the swap entry is no longer in use.
301  */
302 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
303                         struct vm_area_struct *vma, unsigned long addr)
304 {
305         struct page *found_page, *new_page = NULL;
306         int err;
307
308         do {
309                 /*
310                  * First check the swap cache.  Since this is normally
311                  * called after lookup_swap_cache() failed, re-calling
312                  * that would confuse statistics.
313                  */
314                 found_page = find_get_page(swap_address_space(entry),
315                                         entry.val);
316                 if (found_page)
317                         break;
318
319                 /*
320                  * Get a new page to read into from swap.
321                  */
322                 if (!new_page) {
323                         new_page = alloc_page_vma(gfp_mask, vma, addr);
324                         if (!new_page)
325                                 break;          /* Out of memory */
326                 }
327
328                 /*
329                  * call radix_tree_preload() while we can wait.
330                  */
331                 err = radix_tree_preload(gfp_mask & GFP_KERNEL);
332                 if (err)
333                         break;
334
335                 /*
336                  * Swap entry may have been freed since our caller observed it.
337                  */
338                 err = swapcache_prepare(entry);
339                 if (err == -EEXIST) {
340                         radix_tree_preload_end();
341                         /*
342                          * We might race against get_swap_page() and stumble
343                          * across a SWAP_HAS_CACHE swap_map entry whose page
344                          * has not been brought into the swapcache yet, while
345                          * the other end is scheduled away waiting on discard
346                          * I/O completion at scan_swap_map().
347                          *
348                          * In order to avoid turning this transitory state
349                          * into a permanent loop around this -EEXIST case
350                          * if !CONFIG_PREEMPT and the I/O completion happens
351                          * to be waiting on the CPU waitqueue where we are now
352                          * busy looping, we just conditionally invoke the
353                          * scheduler here, if there are some more important
354                          * tasks to run.
355                          */
356                         cond_resched();
357                         continue;
358                 }
359                 if (err) {              /* swp entry is obsolete ? */
360                         radix_tree_preload_end();
361                         break;
362                 }
363
364                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
365                 __set_page_locked(new_page);
366                 SetPageSwapBacked(new_page);
367                 err = __add_to_swap_cache(new_page, entry);
368                 if (likely(!err)) {
369                         radix_tree_preload_end();
370                         /*
371                          * Initiate read into locked page and return.
372                          */
373                         lru_cache_add_anon(new_page);
374                         swap_readpage(new_page);
375                         return new_page;
376                 }
377                 radix_tree_preload_end();
378                 ClearPageSwapBacked(new_page);
379                 __clear_page_locked(new_page);
380                 /*
381                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
382                  * clear SWAP_HAS_CACHE flag.
383                  */
384                 swapcache_free(entry, NULL);
385         } while (err != -ENOMEM);
386
387         if (new_page)
388                 page_cache_release(new_page);
389         return found_page;
390 }
391
392 /**
393  * swapin_readahead - swap in pages in hope we need them soon
394  * @entry: swap entry of this memory
395  * @gfp_mask: memory allocation flags
396  * @vma: user vma this address belongs to
397  * @addr: target address for mempolicy
398  *
399  * Returns the struct page for entry and addr, after queueing swapin.
400  *
401  * Primitive swap readahead code. We simply read an aligned block of
402  * (1 << page_cluster) entries in the swap area. This method is chosen
403  * because it doesn't cost us any seek time.  We also make sure to queue
404  * the 'original' request together with the readahead ones...
405  *
406  * This has been extended to use the NUMA policies from the mm triggering
407  * the readahead.
408  *
409  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
410  */
411 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
412                         struct vm_area_struct *vma, unsigned long addr)
413 {
414         struct page *page;
415         unsigned long offset = swp_offset(entry);
416         unsigned long start_offset, end_offset;
417         unsigned long mask = (1UL << page_cluster) - 1;
418         struct blk_plug plug;
419
420         /* Read a page_cluster sized and aligned cluster around offset. */
421         start_offset = offset & ~mask;
422         end_offset = offset | mask;
423         if (!start_offset)      /* First page is swap header. */
424                 start_offset++;
425
426         blk_start_plug(&plug);
427         for (offset = start_offset; offset <= end_offset ; offset++) {
428                 /* Ok, do the async read-ahead now */
429                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
430                                                 gfp_mask, vma, addr);
431                 if (!page)
432                         continue;
433                 page_cache_release(page);
434         }
435         blk_finish_plug(&plug);
436
437         lru_add_drain();        /* Push any new pages onto the LRU now */
438         return read_swap_cache_async(entry, gfp_mask, vma, addr);
439 }