16ff89d058f4e6663eb4e84214bfdc16e36864f8
[muen/linux.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/vmalloc.h>
21 #include <linux/swap_slots.h>
22 #include <linux/huge_mm.h>
23
24 #include <asm/pgtable.h>
25
26 /*
27  * swapper_space is a fiction, retained to simplify the path through
28  * vmscan's shrink_page_list.
29  */
30 static const struct address_space_operations swap_aops = {
31         .writepage      = swap_writepage,
32         .set_page_dirty = swap_set_page_dirty,
33 #ifdef CONFIG_MIGRATION
34         .migratepage    = migrate_page,
35 #endif
36 };
37
38 struct address_space *swapper_spaces[MAX_SWAPFILES];
39 static unsigned int nr_swapper_spaces[MAX_SWAPFILES];
40
41 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
42 #define ADD_CACHE_INFO(x, nr)   do { swap_cache_info.x += (nr); } while (0)
43
44 static struct {
45         unsigned long add_total;
46         unsigned long del_total;
47         unsigned long find_success;
48         unsigned long find_total;
49 } swap_cache_info;
50
51 unsigned long total_swapcache_pages(void)
52 {
53         unsigned int i, j, nr;
54         unsigned long ret = 0;
55         struct address_space *spaces;
56
57         rcu_read_lock();
58         for (i = 0; i < MAX_SWAPFILES; i++) {
59                 /*
60                  * The corresponding entries in nr_swapper_spaces and
61                  * swapper_spaces will be reused only after at least
62                  * one grace period.  So it is impossible for them
63                  * belongs to different usage.
64                  */
65                 nr = nr_swapper_spaces[i];
66                 spaces = rcu_dereference(swapper_spaces[i]);
67                 if (!nr || !spaces)
68                         continue;
69                 for (j = 0; j < nr; j++)
70                         ret += spaces[j].nrpages;
71         }
72         rcu_read_unlock();
73         return ret;
74 }
75
76 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
77
78 void show_swap_cache_info(void)
79 {
80         printk("%lu pages in swap cache\n", total_swapcache_pages());
81         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
82                 swap_cache_info.add_total, swap_cache_info.del_total,
83                 swap_cache_info.find_success, swap_cache_info.find_total);
84         printk("Free swap  = %ldkB\n",
85                 get_nr_swap_pages() << (PAGE_SHIFT - 10));
86         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
87 }
88
89 /*
90  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
91  * but sets SwapCache flag and private instead of mapping and index.
92  */
93 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
94 {
95         int error, i, nr = hpage_nr_pages(page);
96         struct address_space *address_space;
97         pgoff_t idx = swp_offset(entry);
98
99         VM_BUG_ON_PAGE(!PageLocked(page), page);
100         VM_BUG_ON_PAGE(PageSwapCache(page), page);
101         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
102
103         page_ref_add(page, nr);
104         SetPageSwapCache(page);
105
106         address_space = swap_address_space(entry);
107         spin_lock_irq(&address_space->tree_lock);
108         for (i = 0; i < nr; i++) {
109                 set_page_private(page + i, entry.val + i);
110                 error = radix_tree_insert(&address_space->page_tree,
111                                           idx + i, page + i);
112                 if (unlikely(error))
113                         break;
114         }
115         if (likely(!error)) {
116                 address_space->nrpages += nr;
117                 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
118                 ADD_CACHE_INFO(add_total, nr);
119         } else {
120                 /*
121                  * Only the context which have set SWAP_HAS_CACHE flag
122                  * would call add_to_swap_cache().
123                  * So add_to_swap_cache() doesn't returns -EEXIST.
124                  */
125                 VM_BUG_ON(error == -EEXIST);
126                 set_page_private(page + i, 0UL);
127                 while (i--) {
128                         radix_tree_delete(&address_space->page_tree, idx + i);
129                         set_page_private(page + i, 0UL);
130                 }
131                 ClearPageSwapCache(page);
132                 page_ref_sub(page, nr);
133         }
134         spin_unlock_irq(&address_space->tree_lock);
135
136         return error;
137 }
138
139
140 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
141 {
142         int error;
143
144         error = radix_tree_maybe_preload_order(gfp_mask, compound_order(page));
145         if (!error) {
146                 error = __add_to_swap_cache(page, entry);
147                 radix_tree_preload_end();
148         }
149         return error;
150 }
151
152 /*
153  * This must be called only on pages that have
154  * been verified to be in the swap cache.
155  */
156 void __delete_from_swap_cache(struct page *page)
157 {
158         struct address_space *address_space;
159         int i, nr = hpage_nr_pages(page);
160         swp_entry_t entry;
161         pgoff_t idx;
162
163         VM_BUG_ON_PAGE(!PageLocked(page), page);
164         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
165         VM_BUG_ON_PAGE(PageWriteback(page), page);
166
167         entry.val = page_private(page);
168         address_space = swap_address_space(entry);
169         idx = swp_offset(entry);
170         for (i = 0; i < nr; i++) {
171                 radix_tree_delete(&address_space->page_tree, idx + i);
172                 set_page_private(page + i, 0);
173         }
174         ClearPageSwapCache(page);
175         address_space->nrpages -= nr;
176         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
177         ADD_CACHE_INFO(del_total, nr);
178 }
179
180 /**
181  * add_to_swap - allocate swap space for a page
182  * @page: page we want to move to swap
183  *
184  * Allocate swap space for the page and add the page to the
185  * swap cache.  Caller needs to hold the page lock. 
186  */
187 int add_to_swap(struct page *page, struct list_head *list)
188 {
189         swp_entry_t entry;
190         int err;
191
192         VM_BUG_ON_PAGE(!PageLocked(page), page);
193         VM_BUG_ON_PAGE(!PageUptodate(page), page);
194
195 retry:
196         entry = get_swap_page(page);
197         if (!entry.val)
198                 goto fail;
199         if (mem_cgroup_try_charge_swap(page, entry))
200                 goto fail_free;
201
202         /*
203          * Radix-tree node allocations from PF_MEMALLOC contexts could
204          * completely exhaust the page allocator. __GFP_NOMEMALLOC
205          * stops emergency reserves from being allocated.
206          *
207          * TODO: this could cause a theoretical memory reclaim
208          * deadlock in the swap out path.
209          */
210         /*
211          * Add it to the swap cache.
212          */
213         err = add_to_swap_cache(page, entry,
214                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
215         /* -ENOMEM radix-tree allocation failure */
216         if (err)
217                 /*
218                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
219                  * clear SWAP_HAS_CACHE flag.
220                  */
221                 goto fail_free;
222
223         if (PageTransHuge(page)) {
224                 err = split_huge_page_to_list(page, list);
225                 if (err) {
226                         delete_from_swap_cache(page);
227                         return 0;
228                 }
229         }
230
231         return 1;
232
233 fail_free:
234         if (PageTransHuge(page))
235                 swapcache_free_cluster(entry);
236         else
237                 swapcache_free(entry);
238 fail:
239         if (PageTransHuge(page) && !split_huge_page_to_list(page, list))
240                 goto retry;
241         return 0;
242 }
243
244 /*
245  * This must be called only on pages that have
246  * been verified to be in the swap cache and locked.
247  * It will never put the page into the free list,
248  * the caller has a reference on the page.
249  */
250 void delete_from_swap_cache(struct page *page)
251 {
252         swp_entry_t entry;
253         struct address_space *address_space;
254
255         entry.val = page_private(page);
256
257         address_space = swap_address_space(entry);
258         spin_lock_irq(&address_space->tree_lock);
259         __delete_from_swap_cache(page);
260         spin_unlock_irq(&address_space->tree_lock);
261
262         if (PageTransHuge(page))
263                 swapcache_free_cluster(entry);
264         else
265                 swapcache_free(entry);
266
267         page_ref_sub(page, hpage_nr_pages(page));
268 }
269
270 /* 
271  * If we are the only user, then try to free up the swap cache. 
272  * 
273  * Its ok to check for PageSwapCache without the page lock
274  * here because we are going to recheck again inside
275  * try_to_free_swap() _with_ the lock.
276  *                                      - Marcelo
277  */
278 static inline void free_swap_cache(struct page *page)
279 {
280         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
281                 try_to_free_swap(page);
282                 unlock_page(page);
283         }
284 }
285
286 /* 
287  * Perform a free_page(), also freeing any swap cache associated with
288  * this page if it is the last user of the page.
289  */
290 void free_page_and_swap_cache(struct page *page)
291 {
292         free_swap_cache(page);
293         if (!is_huge_zero_page(page))
294                 put_page(page);
295 }
296
297 /*
298  * Passed an array of pages, drop them all from swapcache and then release
299  * them.  They are removed from the LRU and freed if this is their last use.
300  */
301 void free_pages_and_swap_cache(struct page **pages, int nr)
302 {
303         struct page **pagep = pages;
304         int i;
305
306         lru_add_drain();
307         for (i = 0; i < nr; i++)
308                 free_swap_cache(pagep[i]);
309         release_pages(pagep, nr, false);
310 }
311
312 /*
313  * Lookup a swap entry in the swap cache. A found page will be returned
314  * unlocked and with its refcount incremented - we rely on the kernel
315  * lock getting page table operations atomic even if we drop the page
316  * lock before returning.
317  */
318 struct page * lookup_swap_cache(swp_entry_t entry)
319 {
320         struct page *page;
321
322         page = find_get_page(swap_address_space(entry), swp_offset(entry));
323
324         if (page && likely(!PageTransCompound(page))) {
325                 INC_CACHE_INFO(find_success);
326                 if (TestClearPageReadahead(page))
327                         atomic_inc(&swapin_readahead_hits);
328         }
329
330         INC_CACHE_INFO(find_total);
331         return page;
332 }
333
334 struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
335                         struct vm_area_struct *vma, unsigned long addr,
336                         bool *new_page_allocated)
337 {
338         struct page *found_page, *new_page = NULL;
339         struct address_space *swapper_space = swap_address_space(entry);
340         int err;
341         *new_page_allocated = false;
342
343         do {
344                 /*
345                  * First check the swap cache.  Since this is normally
346                  * called after lookup_swap_cache() failed, re-calling
347                  * that would confuse statistics.
348                  */
349                 found_page = find_get_page(swapper_space, swp_offset(entry));
350                 if (found_page)
351                         break;
352
353                 /*
354                  * Just skip read ahead for unused swap slot.
355                  * During swap_off when swap_slot_cache is disabled,
356                  * we have to handle the race between putting
357                  * swap entry in swap cache and marking swap slot
358                  * as SWAP_HAS_CACHE.  That's done in later part of code or
359                  * else swap_off will be aborted if we return NULL.
360                  */
361                 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
362                         break;
363
364                 /*
365                  * Get a new page to read into from swap.
366                  */
367                 if (!new_page) {
368                         new_page = alloc_page_vma(gfp_mask, vma, addr);
369                         if (!new_page)
370                                 break;          /* Out of memory */
371                 }
372
373                 /*
374                  * call radix_tree_preload() while we can wait.
375                  */
376                 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
377                 if (err)
378                         break;
379
380                 /*
381                  * Swap entry may have been freed since our caller observed it.
382                  */
383                 err = swapcache_prepare(entry);
384                 if (err == -EEXIST) {
385                         radix_tree_preload_end();
386                         /*
387                          * We might race against get_swap_page() and stumble
388                          * across a SWAP_HAS_CACHE swap_map entry whose page
389                          * has not been brought into the swapcache yet.
390                          */
391                         cond_resched();
392                         continue;
393                 }
394                 if (err) {              /* swp entry is obsolete ? */
395                         radix_tree_preload_end();
396                         break;
397                 }
398
399                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
400                 __SetPageLocked(new_page);
401                 __SetPageSwapBacked(new_page);
402                 err = __add_to_swap_cache(new_page, entry);
403                 if (likely(!err)) {
404                         radix_tree_preload_end();
405                         /*
406                          * Initiate read into locked page and return.
407                          */
408                         lru_cache_add_anon(new_page);
409                         *new_page_allocated = true;
410                         return new_page;
411                 }
412                 radix_tree_preload_end();
413                 __ClearPageLocked(new_page);
414                 /*
415                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
416                  * clear SWAP_HAS_CACHE flag.
417                  */
418                 swapcache_free(entry);
419         } while (err != -ENOMEM);
420
421         if (new_page)
422                 put_page(new_page);
423         return found_page;
424 }
425
426 /*
427  * Locate a page of swap in physical memory, reserving swap cache space
428  * and reading the disk if it is not already cached.
429  * A failure return means that either the page allocation failed or that
430  * the swap entry is no longer in use.
431  */
432 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
433                         struct vm_area_struct *vma, unsigned long addr)
434 {
435         bool page_was_allocated;
436         struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
437                         vma, addr, &page_was_allocated);
438
439         if (page_was_allocated)
440                 swap_readpage(retpage);
441
442         return retpage;
443 }
444
445 static unsigned long swapin_nr_pages(unsigned long offset)
446 {
447         static unsigned long prev_offset;
448         unsigned int pages, max_pages, last_ra;
449         static atomic_t last_readahead_pages;
450
451         max_pages = 1 << READ_ONCE(page_cluster);
452         if (max_pages <= 1)
453                 return 1;
454
455         /*
456          * This heuristic has been found to work well on both sequential and
457          * random loads, swapping to hard disk or to SSD: please don't ask
458          * what the "+ 2" means, it just happens to work well, that's all.
459          */
460         pages = atomic_xchg(&swapin_readahead_hits, 0) + 2;
461         if (pages == 2) {
462                 /*
463                  * We can have no readahead hits to judge by: but must not get
464                  * stuck here forever, so check for an adjacent offset instead
465                  * (and don't even bother to check whether swap type is same).
466                  */
467                 if (offset != prev_offset + 1 && offset != prev_offset - 1)
468                         pages = 1;
469                 prev_offset = offset;
470         } else {
471                 unsigned int roundup = 4;
472                 while (roundup < pages)
473                         roundup <<= 1;
474                 pages = roundup;
475         }
476
477         if (pages > max_pages)
478                 pages = max_pages;
479
480         /* Don't shrink readahead too fast */
481         last_ra = atomic_read(&last_readahead_pages) / 2;
482         if (pages < last_ra)
483                 pages = last_ra;
484         atomic_set(&last_readahead_pages, pages);
485
486         return pages;
487 }
488
489 /**
490  * swapin_readahead - swap in pages in hope we need them soon
491  * @entry: swap entry of this memory
492  * @gfp_mask: memory allocation flags
493  * @vma: user vma this address belongs to
494  * @addr: target address for mempolicy
495  *
496  * Returns the struct page for entry and addr, after queueing swapin.
497  *
498  * Primitive swap readahead code. We simply read an aligned block of
499  * (1 << page_cluster) entries in the swap area. This method is chosen
500  * because it doesn't cost us any seek time.  We also make sure to queue
501  * the 'original' request together with the readahead ones...
502  *
503  * This has been extended to use the NUMA policies from the mm triggering
504  * the readahead.
505  *
506  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
507  */
508 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
509                         struct vm_area_struct *vma, unsigned long addr)
510 {
511         struct page *page;
512         unsigned long entry_offset = swp_offset(entry);
513         unsigned long offset = entry_offset;
514         unsigned long start_offset, end_offset;
515         unsigned long mask;
516         struct blk_plug plug;
517
518         mask = swapin_nr_pages(offset) - 1;
519         if (!mask)
520                 goto skip;
521
522         /* Read a page_cluster sized and aligned cluster around offset. */
523         start_offset = offset & ~mask;
524         end_offset = offset | mask;
525         if (!start_offset)      /* First page is swap header. */
526                 start_offset++;
527
528         blk_start_plug(&plug);
529         for (offset = start_offset; offset <= end_offset ; offset++) {
530                 /* Ok, do the async read-ahead now */
531                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
532                                                 gfp_mask, vma, addr);
533                 if (!page)
534                         continue;
535                 if (offset != entry_offset && likely(!PageTransCompound(page)))
536                         SetPageReadahead(page);
537                 put_page(page);
538         }
539         blk_finish_plug(&plug);
540
541         lru_add_drain();        /* Push any new pages onto the LRU now */
542 skip:
543         return read_swap_cache_async(entry, gfp_mask, vma, addr);
544 }
545
546 int init_swap_address_space(unsigned int type, unsigned long nr_pages)
547 {
548         struct address_space *spaces, *space;
549         unsigned int i, nr;
550
551         nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
552         spaces = kvzalloc(sizeof(struct address_space) * nr, GFP_KERNEL);
553         if (!spaces)
554                 return -ENOMEM;
555         for (i = 0; i < nr; i++) {
556                 space = spaces + i;
557                 INIT_RADIX_TREE(&space->page_tree, GFP_ATOMIC|__GFP_NOWARN);
558                 atomic_set(&space->i_mmap_writable, 0);
559                 space->a_ops = &swap_aops;
560                 /* swap cache doesn't use writeback related tags */
561                 mapping_set_no_writeback_tags(space);
562                 spin_lock_init(&space->tree_lock);
563         }
564         nr_swapper_spaces[type] = nr;
565         rcu_assign_pointer(swapper_spaces[type], spaces);
566
567         return 0;
568 }
569
570 void exit_swap_address_space(unsigned int type)
571 {
572         struct address_space *spaces;
573
574         spaces = swapper_spaces[type];
575         nr_swapper_spaces[type] = 0;
576         rcu_assign_pointer(swapper_spaces[type], NULL);
577         synchronize_rcu();
578         kvfree(spaces);
579 }