blob: d1f04ea23c048cb2b03fccab5b10726cd0bc3577
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 | /* swap cache doesn't use writeback related tags */ |
41 | .flags = 1 << AS_NO_WRITEBACK_TAGS, |
42 | } |
43 | }; |
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 | } swap_cache_info; |
53 | |
54 | unsigned long total_swapcache_pages(void) |
55 | { |
56 | int i; |
57 | unsigned long ret = 0; |
58 | |
59 | for (i = 0; i < MAX_SWAPFILES; i++) |
60 | ret += swapper_spaces[i].nrpages; |
61 | return ret; |
62 | } |
63 | |
64 | static atomic_t swapin_readahead_hits = ATOMIC_INIT(4); |
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_PAGE(!PageLocked(page), page); |
87 | VM_BUG_ON_PAGE(PageSwapCache(page), page); |
88 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); |
89 | |
90 | get_page(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 | swp_offset(entry), page); |
98 | if (likely(!error)) { |
99 | address_space->nrpages++; |
100 | __inc_node_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 | put_page(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_maybe_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_PAGE(!PageLocked(page), page); |
143 | VM_BUG_ON_PAGE(!PageSwapCache(page), page); |
144 | VM_BUG_ON_PAGE(PageWriteback(page), page); |
145 | |
146 | entry.val = page_private(page); |
147 | address_space = swap_address_space(entry); |
148 | radix_tree_delete(&address_space->page_tree, swp_offset(entry)); |
149 | set_page_private(page, 0); |
150 | ClearPageSwapCache(page); |
151 | address_space->nrpages--; |
152 | __dec_node_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_PAGE(!PageLocked(page), page); |
169 | VM_BUG_ON_PAGE(!PageUptodate(page), page); |
170 | |
171 | entry = get_swap_page(); |
172 | if (!entry.val) |
173 | return 0; |
174 | |
175 | if (mem_cgroup_try_charge_swap(page, entry)) { |
176 | swapcache_free(entry); |
177 | return 0; |
178 | } |
179 | |
180 | if (unlikely(PageTransHuge(page))) |
181 | if (unlikely(split_huge_page_to_list(page, list))) { |
182 | swapcache_free(entry); |
183 | return 0; |
184 | } |
185 | |
186 | /* |
187 | * Radix-tree node allocations from PF_MEMALLOC contexts could |
188 | * completely exhaust the page allocator. __GFP_NOMEMALLOC |
189 | * stops emergency reserves from being allocated. |
190 | * |
191 | * TODO: this could cause a theoretical memory reclaim |
192 | * deadlock in the swap out path. |
193 | */ |
194 | /* |
195 | * Add it to the swap cache. |
196 | */ |
197 | err = add_to_swap_cache(page, entry, |
198 | __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN); |
199 | |
200 | if (!err) { |
201 | return 1; |
202 | } else { /* -ENOMEM radix-tree allocation failure */ |
203 | /* |
204 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely |
205 | * clear SWAP_HAS_CACHE flag. |
206 | */ |
207 | swapcache_free(entry); |
208 | return 0; |
209 | } |
210 | } |
211 | |
212 | /* |
213 | * This must be called only on pages that have |
214 | * been verified to be in the swap cache and locked. |
215 | * It will never put the page into the free list, |
216 | * the caller has a reference on the page. |
217 | */ |
218 | void delete_from_swap_cache(struct page *page) |
219 | { |
220 | swp_entry_t entry; |
221 | struct address_space *address_space; |
222 | |
223 | entry.val = page_private(page); |
224 | |
225 | address_space = swap_address_space(entry); |
226 | spin_lock_irq(&address_space->tree_lock); |
227 | __delete_from_swap_cache(page); |
228 | spin_unlock_irq(&address_space->tree_lock); |
229 | |
230 | swapcache_free(entry); |
231 | put_page(page); |
232 | } |
233 | |
234 | /* |
235 | * If we are the only user, then try to free up the swap cache. |
236 | * |
237 | * Its ok to check for PageSwapCache without the page lock |
238 | * here because we are going to recheck again inside |
239 | * try_to_free_swap() _with_ the lock. |
240 | * - Marcelo |
241 | */ |
242 | static inline void free_swap_cache(struct page *page) |
243 | { |
244 | if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) { |
245 | try_to_free_swap(page); |
246 | unlock_page(page); |
247 | } |
248 | } |
249 | |
250 | /* |
251 | * Perform a free_page(), also freeing any swap cache associated with |
252 | * this page if it is the last user of the page. |
253 | */ |
254 | void free_page_and_swap_cache(struct page *page) |
255 | { |
256 | free_swap_cache(page); |
257 | if (!is_huge_zero_page(page)) |
258 | put_page(page); |
259 | } |
260 | |
261 | /* |
262 | * Passed an array of pages, drop them all from swapcache and then release |
263 | * them. They are removed from the LRU and freed if this is their last use. |
264 | */ |
265 | void free_pages_and_swap_cache(struct page **pages, int nr) |
266 | { |
267 | struct page **pagep = pages; |
268 | int i; |
269 | |
270 | lru_add_drain(); |
271 | for (i = 0; i < nr; i++) |
272 | free_swap_cache(pagep[i]); |
273 | release_pages(pagep, nr, false); |
274 | } |
275 | |
276 | /* |
277 | * Lookup a swap entry in the swap cache. A found page will be returned |
278 | * unlocked and with its refcount incremented - we rely on the kernel |
279 | * lock getting page table operations atomic even if we drop the page |
280 | * lock before returning. |
281 | */ |
282 | struct page * lookup_swap_cache(swp_entry_t entry) |
283 | { |
284 | struct page *page; |
285 | |
286 | page = find_get_page(swap_address_space(entry), swp_offset(entry)); |
287 | |
288 | if (page) { |
289 | INC_CACHE_INFO(find_success); |
290 | if (TestClearPageReadahead(page)) |
291 | atomic_inc(&swapin_readahead_hits); |
292 | } |
293 | |
294 | INC_CACHE_INFO(find_total); |
295 | return page; |
296 | } |
297 | |
298 | struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, |
299 | struct vm_area_struct *vma, unsigned long addr, |
300 | bool *new_page_allocated) |
301 | { |
302 | struct page *found_page, *new_page = NULL; |
303 | struct address_space *swapper_space = swap_address_space(entry); |
304 | int err; |
305 | *new_page_allocated = false; |
306 | |
307 | do { |
308 | /* |
309 | * First check the swap cache. Since this is normally |
310 | * called after lookup_swap_cache() failed, re-calling |
311 | * that would confuse statistics. |
312 | */ |
313 | found_page = find_get_page(swapper_space, swp_offset(entry)); |
314 | if (found_page) |
315 | break; |
316 | |
317 | /* |
318 | * Get a new page to read into from swap. |
319 | */ |
320 | if (!new_page) { |
321 | #ifdef CONFIG_AMLOGIC_CMA |
322 | new_page = alloc_page_vma(gfp_mask | __GFP_BDEV, |
323 | vma, addr); |
324 | #else |
325 | new_page = alloc_page_vma(gfp_mask, vma, addr); |
326 | #endif |
327 | if (!new_page) |
328 | break; /* Out of memory */ |
329 | } |
330 | |
331 | /* |
332 | * call radix_tree_preload() while we can wait. |
333 | */ |
334 | err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL); |
335 | if (err) |
336 | break; |
337 | |
338 | /* |
339 | * Swap entry may have been freed since our caller observed it. |
340 | */ |
341 | err = swapcache_prepare(entry); |
342 | if (err == -EEXIST) { |
343 | radix_tree_preload_end(); |
344 | /* |
345 | * We might race against get_swap_page() and stumble |
346 | * across a SWAP_HAS_CACHE swap_map entry whose page |
347 | * has not been brought into the swapcache yet, while |
348 | * the other end is scheduled away waiting on discard |
349 | * I/O completion at scan_swap_map(). |
350 | * |
351 | * In order to avoid turning this transitory state |
352 | * into a permanent loop around this -EEXIST case |
353 | * if !CONFIG_PREEMPT and the I/O completion happens |
354 | * to be waiting on the CPU waitqueue where we are now |
355 | * busy looping, we just conditionally invoke the |
356 | * scheduler here, if there are some more important |
357 | * tasks to run. |
358 | */ |
359 | cond_resched(); |
360 | continue; |
361 | } |
362 | if (err) { /* swp entry is obsolete ? */ |
363 | radix_tree_preload_end(); |
364 | break; |
365 | } |
366 | |
367 | /* May fail (-ENOMEM) if radix-tree node allocation failed. */ |
368 | __SetPageLocked(new_page); |
369 | __SetPageSwapBacked(new_page); |
370 | err = __add_to_swap_cache(new_page, entry); |
371 | if (likely(!err)) { |
372 | radix_tree_preload_end(); |
373 | /* |
374 | * Initiate read into locked page and return. |
375 | */ |
376 | SetPageWorkingset(new_page); |
377 | lru_cache_add_anon(new_page); |
378 | *new_page_allocated = true; |
379 | return new_page; |
380 | } |
381 | radix_tree_preload_end(); |
382 | __ClearPageLocked(new_page); |
383 | /* |
384 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely |
385 | * clear SWAP_HAS_CACHE flag. |
386 | */ |
387 | swapcache_free(entry); |
388 | } while (err != -ENOMEM); |
389 | |
390 | if (new_page) |
391 | put_page(new_page); |
392 | return found_page; |
393 | } |
394 | |
395 | /* |
396 | * Locate a page of swap in physical memory, reserving swap cache space |
397 | * and reading the disk if it is not already cached. |
398 | * A failure return means that either the page allocation failed or that |
399 | * the swap entry is no longer in use. |
400 | */ |
401 | struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, |
402 | struct vm_area_struct *vma, unsigned long addr) |
403 | { |
404 | bool page_was_allocated; |
405 | struct page *retpage = __read_swap_cache_async(entry, gfp_mask, |
406 | vma, addr, &page_was_allocated); |
407 | |
408 | if (page_was_allocated) |
409 | swap_readpage(retpage); |
410 | |
411 | return retpage; |
412 | } |
413 | |
414 | static unsigned long swapin_nr_pages(unsigned long offset) |
415 | { |
416 | static unsigned long prev_offset; |
417 | unsigned int pages, max_pages, last_ra; |
418 | static atomic_t last_readahead_pages; |
419 | |
420 | max_pages = 1 << READ_ONCE(page_cluster); |
421 | if (max_pages <= 1) |
422 | return 1; |
423 | |
424 | /* |
425 | * This heuristic has been found to work well on both sequential and |
426 | * random loads, swapping to hard disk or to SSD: please don't ask |
427 | * what the "+ 2" means, it just happens to work well, that's all. |
428 | */ |
429 | pages = atomic_xchg(&swapin_readahead_hits, 0) + 2; |
430 | if (pages == 2) { |
431 | /* |
432 | * We can have no readahead hits to judge by: but must not get |
433 | * stuck here forever, so check for an adjacent offset instead |
434 | * (and don't even bother to check whether swap type is same). |
435 | */ |
436 | if (offset != prev_offset + 1 && offset != prev_offset - 1) |
437 | pages = 1; |
438 | prev_offset = offset; |
439 | } else { |
440 | unsigned int roundup = 4; |
441 | while (roundup < pages) |
442 | roundup <<= 1; |
443 | pages = roundup; |
444 | } |
445 | |
446 | if (pages > max_pages) |
447 | pages = max_pages; |
448 | |
449 | /* Don't shrink readahead too fast */ |
450 | last_ra = atomic_read(&last_readahead_pages) / 2; |
451 | if (pages < last_ra) |
452 | pages = last_ra; |
453 | atomic_set(&last_readahead_pages, pages); |
454 | |
455 | return pages; |
456 | } |
457 | |
458 | /** |
459 | * swapin_readahead - swap in pages in hope we need them soon |
460 | * @entry: swap entry of this memory |
461 | * @gfp_mask: memory allocation flags |
462 | * @vma: user vma this address belongs to |
463 | * @addr: target address for mempolicy |
464 | * |
465 | * Returns the struct page for entry and addr, after queueing swapin. |
466 | * |
467 | * Primitive swap readahead code. We simply read an aligned block of |
468 | * (1 << page_cluster) entries in the swap area. This method is chosen |
469 | * because it doesn't cost us any seek time. We also make sure to queue |
470 | * the 'original' request together with the readahead ones... |
471 | * |
472 | * This has been extended to use the NUMA policies from the mm triggering |
473 | * the readahead. |
474 | * |
475 | * Caller must hold down_read on the vma->vm_mm if vma is not NULL. |
476 | */ |
477 | struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, |
478 | struct vm_area_struct *vma, unsigned long addr) |
479 | { |
480 | struct page *page; |
481 | unsigned long entry_offset = swp_offset(entry); |
482 | unsigned long offset = entry_offset; |
483 | unsigned long start_offset, end_offset; |
484 | unsigned long mask; |
485 | struct blk_plug plug; |
486 | |
487 | mask = swapin_nr_pages(offset) - 1; |
488 | if (!mask) |
489 | goto skip; |
490 | |
491 | /* Read a page_cluster sized and aligned cluster around offset. */ |
492 | start_offset = offset & ~mask; |
493 | end_offset = offset | mask; |
494 | if (!start_offset) /* First page is swap header. */ |
495 | start_offset++; |
496 | |
497 | blk_start_plug(&plug); |
498 | for (offset = start_offset; offset <= end_offset ; offset++) { |
499 | /* Ok, do the async read-ahead now */ |
500 | page = read_swap_cache_async(swp_entry(swp_type(entry), offset), |
501 | gfp_mask, vma, addr); |
502 | if (!page) |
503 | continue; |
504 | if (offset != entry_offset) |
505 | SetPageReadahead(page); |
506 | put_page(page); |
507 | } |
508 | blk_finish_plug(&plug); |
509 | |
510 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
511 | skip: |
512 | return read_swap_cache_async(entry, gfp_mask, vma, addr); |
513 | } |
514 |