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