]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Memory Migration functionality - linux/mm/migrate.c | |
3 | * | |
4 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter | |
5 | * | |
6 | * Page migration was first developed in the context of the memory hotplug | |
7 | * project. The main authors of the migration code are: | |
8 | * | |
9 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> | |
10 | * Hirokazu Takahashi <taka@valinux.co.jp> | |
11 | * Dave Hansen <haveblue@us.ibm.com> | |
12 | * Christoph Lameter | |
13 | */ | |
14 | ||
15 | #include <linux/migrate.h> | |
16 | #include <linux/export.h> | |
17 | #include <linux/swap.h> | |
18 | #include <linux/swapops.h> | |
19 | #include <linux/pagemap.h> | |
20 | #include <linux/buffer_head.h> | |
21 | #include <linux/mm_inline.h> | |
22 | #include <linux/nsproxy.h> | |
23 | #include <linux/pagevec.h> | |
24 | #include <linux/ksm.h> | |
25 | #include <linux/rmap.h> | |
26 | #include <linux/topology.h> | |
27 | #include <linux/cpu.h> | |
28 | #include <linux/cpuset.h> | |
29 | #include <linux/writeback.h> | |
30 | #include <linux/mempolicy.h> | |
31 | #include <linux/vmalloc.h> | |
32 | #include <linux/security.h> | |
33 | #include <linux/backing-dev.h> | |
34 | #include <linux/compaction.h> | |
35 | #include <linux/syscalls.h> | |
36 | #include <linux/hugetlb.h> | |
37 | #include <linux/hugetlb_cgroup.h> | |
38 | #include <linux/gfp.h> | |
39 | #include <linux/balloon_compaction.h> | |
40 | #include <linux/mmu_notifier.h> | |
41 | #include <linux/page_idle.h> | |
42 | #include <linux/page_owner.h> | |
43 | #include <linux/sched/mm.h> | |
44 | #include <linux/ptrace.h> | |
45 | ||
46 | #include <asm/tlbflush.h> | |
47 | ||
48 | #define CREATE_TRACE_POINTS | |
49 | #include <trace/events/migrate.h> | |
50 | ||
51 | #include "internal.h" | |
52 | ||
53 | /* | |
54 | * migrate_prep() needs to be called before we start compiling a list of pages | |
55 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is | |
56 | * undesirable, use migrate_prep_local() | |
57 | */ | |
58 | int migrate_prep(void) | |
59 | { | |
60 | /* | |
61 | * Clear the LRU lists so pages can be isolated. | |
62 | * Note that pages may be moved off the LRU after we have | |
63 | * drained them. Those pages will fail to migrate like other | |
64 | * pages that may be busy. | |
65 | */ | |
66 | lru_add_drain_all(); | |
67 | ||
68 | return 0; | |
69 | } | |
70 | ||
71 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ | |
72 | int migrate_prep_local(void) | |
73 | { | |
74 | lru_add_drain(); | |
75 | ||
76 | return 0; | |
77 | } | |
78 | ||
79 | int isolate_movable_page(struct page *page, isolate_mode_t mode) | |
80 | { | |
81 | struct address_space *mapping; | |
82 | ||
83 | /* | |
84 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
85 | * or just got freed under us. | |
86 | * | |
87 | * In case we 'win' a race for a movable page being freed under us and | |
88 | * raise its refcount preventing __free_pages() from doing its job | |
89 | * the put_page() at the end of this block will take care of | |
90 | * release this page, thus avoiding a nasty leakage. | |
91 | */ | |
92 | if (unlikely(!get_page_unless_zero(page))) | |
93 | goto out; | |
94 | ||
95 | /* | |
96 | * Check PageMovable before holding a PG_lock because page's owner | |
97 | * assumes anybody doesn't touch PG_lock of newly allocated page | |
98 | * so unconditionally grapping the lock ruins page's owner side. | |
99 | */ | |
100 | if (unlikely(!__PageMovable(page))) | |
101 | goto out_putpage; | |
102 | /* | |
103 | * As movable pages are not isolated from LRU lists, concurrent | |
104 | * compaction threads can race against page migration functions | |
105 | * as well as race against the releasing a page. | |
106 | * | |
107 | * In order to avoid having an already isolated movable page | |
108 | * being (wrongly) re-isolated while it is under migration, | |
109 | * or to avoid attempting to isolate pages being released, | |
110 | * lets be sure we have the page lock | |
111 | * before proceeding with the movable page isolation steps. | |
112 | */ | |
113 | if (unlikely(!trylock_page(page))) | |
114 | goto out_putpage; | |
115 | ||
116 | if (!PageMovable(page) || PageIsolated(page)) | |
117 | goto out_no_isolated; | |
118 | ||
119 | mapping = page_mapping(page); | |
120 | VM_BUG_ON_PAGE(!mapping, page); | |
121 | ||
122 | if (!mapping->a_ops->isolate_page(page, mode)) | |
123 | goto out_no_isolated; | |
124 | ||
125 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
126 | WARN_ON_ONCE(PageIsolated(page)); | |
127 | __SetPageIsolated(page); | |
128 | unlock_page(page); | |
129 | ||
130 | return 0; | |
131 | ||
132 | out_no_isolated: | |
133 | unlock_page(page); | |
134 | out_putpage: | |
135 | put_page(page); | |
136 | out: | |
137 | return -EBUSY; | |
138 | } | |
139 | ||
140 | /* It should be called on page which is PG_movable */ | |
141 | void putback_movable_page(struct page *page) | |
142 | { | |
143 | struct address_space *mapping; | |
144 | ||
145 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
146 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
147 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
148 | ||
149 | mapping = page_mapping(page); | |
150 | mapping->a_ops->putback_page(page); | |
151 | __ClearPageIsolated(page); | |
152 | } | |
153 | ||
154 | /* | |
155 | * Put previously isolated pages back onto the appropriate lists | |
156 | * from where they were once taken off for compaction/migration. | |
157 | * | |
158 | * This function shall be used whenever the isolated pageset has been | |
159 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
160 | * and isolate_huge_page(). | |
161 | */ | |
162 | void putback_movable_pages(struct list_head *l) | |
163 | { | |
164 | struct page *page; | |
165 | struct page *page2; | |
166 | ||
167 | list_for_each_entry_safe(page, page2, l, lru) { | |
168 | if (unlikely(PageHuge(page))) { | |
169 | putback_active_hugepage(page); | |
170 | continue; | |
171 | } | |
172 | list_del(&page->lru); | |
173 | /* | |
174 | * We isolated non-lru movable page so here we can use | |
175 | * __PageMovable because LRU page's mapping cannot have | |
176 | * PAGE_MAPPING_MOVABLE. | |
177 | */ | |
178 | if (unlikely(__PageMovable(page))) { | |
179 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
180 | lock_page(page); | |
181 | if (PageMovable(page)) | |
182 | putback_movable_page(page); | |
183 | else | |
184 | __ClearPageIsolated(page); | |
185 | unlock_page(page); | |
186 | put_page(page); | |
187 | } else { | |
188 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + | |
189 | page_is_file_cache(page), -hpage_nr_pages(page)); | |
190 | putback_lru_page(page); | |
191 | } | |
192 | } | |
193 | } | |
194 | ||
195 | /* | |
196 | * Restore a potential migration pte to a working pte entry | |
197 | */ | |
198 | static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, | |
199 | unsigned long addr, void *old) | |
200 | { | |
201 | struct page_vma_mapped_walk pvmw = { | |
202 | .page = old, | |
203 | .vma = vma, | |
204 | .address = addr, | |
205 | .flags = PVMW_SYNC | PVMW_MIGRATION, | |
206 | }; | |
207 | struct page *new; | |
208 | pte_t pte; | |
209 | swp_entry_t entry; | |
210 | ||
211 | VM_BUG_ON_PAGE(PageTail(page), page); | |
212 | while (page_vma_mapped_walk(&pvmw)) { | |
213 | if (PageKsm(page)) | |
214 | new = page; | |
215 | else | |
216 | new = page - pvmw.page->index + | |
217 | linear_page_index(vma, pvmw.address); | |
218 | ||
219 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION | |
220 | /* PMD-mapped THP migration entry */ | |
221 | if (!pvmw.pte) { | |
222 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); | |
223 | remove_migration_pmd(&pvmw, new); | |
224 | continue; | |
225 | } | |
226 | #endif | |
227 | ||
228 | get_page(new); | |
229 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); | |
230 | if (pte_swp_soft_dirty(*pvmw.pte)) | |
231 | pte = pte_mksoft_dirty(pte); | |
232 | ||
233 | /* | |
234 | * Recheck VMA as permissions can change since migration started | |
235 | */ | |
236 | entry = pte_to_swp_entry(*pvmw.pte); | |
237 | if (is_write_migration_entry(entry)) | |
238 | pte = maybe_mkwrite(pte, vma); | |
239 | ||
240 | flush_dcache_page(new); | |
241 | #ifdef CONFIG_HUGETLB_PAGE | |
242 | if (PageHuge(new)) { | |
243 | pte = pte_mkhuge(pte); | |
244 | pte = arch_make_huge_pte(pte, vma, new, 0); | |
245 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); | |
246 | if (PageAnon(new)) | |
247 | hugepage_add_anon_rmap(new, vma, pvmw.address); | |
248 | else | |
249 | page_dup_rmap(new, true); | |
250 | } else | |
251 | #endif | |
252 | { | |
253 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); | |
254 | ||
255 | if (PageAnon(new)) | |
256 | page_add_anon_rmap(new, vma, pvmw.address, false); | |
257 | else | |
258 | page_add_file_rmap(new, false); | |
259 | } | |
260 | if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) | |
261 | mlock_vma_page(new); | |
262 | ||
263 | /* No need to invalidate - it was non-present before */ | |
264 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
265 | } | |
266 | ||
267 | return true; | |
268 | } | |
269 | ||
270 | /* | |
271 | * Get rid of all migration entries and replace them by | |
272 | * references to the indicated page. | |
273 | */ | |
274 | void remove_migration_ptes(struct page *old, struct page *new, bool locked) | |
275 | { | |
276 | struct rmap_walk_control rwc = { | |
277 | .rmap_one = remove_migration_pte, | |
278 | .arg = old, | |
279 | }; | |
280 | ||
281 | if (locked) | |
282 | rmap_walk_locked(new, &rwc); | |
283 | else | |
284 | rmap_walk(new, &rwc); | |
285 | } | |
286 | ||
287 | /* | |
288 | * Something used the pte of a page under migration. We need to | |
289 | * get to the page and wait until migration is finished. | |
290 | * When we return from this function the fault will be retried. | |
291 | */ | |
292 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, | |
293 | spinlock_t *ptl) | |
294 | { | |
295 | pte_t pte; | |
296 | swp_entry_t entry; | |
297 | struct page *page; | |
298 | ||
299 | spin_lock(ptl); | |
300 | pte = *ptep; | |
301 | if (!is_swap_pte(pte)) | |
302 | goto out; | |
303 | ||
304 | entry = pte_to_swp_entry(pte); | |
305 | if (!is_migration_entry(entry)) | |
306 | goto out; | |
307 | ||
308 | page = migration_entry_to_page(entry); | |
309 | ||
310 | /* | |
311 | * Once radix-tree replacement of page migration started, page_count | |
312 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
313 | * against a page without get_page(). | |
314 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
315 | * will occur again. | |
316 | */ | |
317 | if (!get_page_unless_zero(page)) | |
318 | goto out; | |
319 | pte_unmap_unlock(ptep, ptl); | |
320 | wait_on_page_locked(page); | |
321 | put_page(page); | |
322 | return; | |
323 | out: | |
324 | pte_unmap_unlock(ptep, ptl); | |
325 | } | |
326 | ||
327 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, | |
328 | unsigned long address) | |
329 | { | |
330 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
331 | pte_t *ptep = pte_offset_map(pmd, address); | |
332 | __migration_entry_wait(mm, ptep, ptl); | |
333 | } | |
334 | ||
335 | void migration_entry_wait_huge(struct vm_area_struct *vma, | |
336 | struct mm_struct *mm, pte_t *pte) | |
337 | { | |
338 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); | |
339 | __migration_entry_wait(mm, pte, ptl); | |
340 | } | |
341 | ||
342 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION | |
343 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
344 | { | |
345 | spinlock_t *ptl; | |
346 | struct page *page; | |
347 | ||
348 | ptl = pmd_lock(mm, pmd); | |
349 | if (!is_pmd_migration_entry(*pmd)) | |
350 | goto unlock; | |
351 | page = migration_entry_to_page(pmd_to_swp_entry(*pmd)); | |
352 | if (!get_page_unless_zero(page)) | |
353 | goto unlock; | |
354 | spin_unlock(ptl); | |
355 | wait_on_page_locked(page); | |
356 | put_page(page); | |
357 | return; | |
358 | unlock: | |
359 | spin_unlock(ptl); | |
360 | } | |
361 | #endif | |
362 | ||
363 | #ifdef CONFIG_BLOCK | |
364 | /* Returns true if all buffers are successfully locked */ | |
365 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
366 | enum migrate_mode mode) | |
367 | { | |
368 | struct buffer_head *bh = head; | |
369 | ||
370 | /* Simple case, sync compaction */ | |
371 | if (mode != MIGRATE_ASYNC) { | |
372 | do { | |
373 | get_bh(bh); | |
374 | lock_buffer(bh); | |
375 | bh = bh->b_this_page; | |
376 | ||
377 | } while (bh != head); | |
378 | ||
379 | return true; | |
380 | } | |
381 | ||
382 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
383 | do { | |
384 | get_bh(bh); | |
385 | if (!trylock_buffer(bh)) { | |
386 | /* | |
387 | * We failed to lock the buffer and cannot stall in | |
388 | * async migration. Release the taken locks | |
389 | */ | |
390 | struct buffer_head *failed_bh = bh; | |
391 | put_bh(failed_bh); | |
392 | bh = head; | |
393 | while (bh != failed_bh) { | |
394 | unlock_buffer(bh); | |
395 | put_bh(bh); | |
396 | bh = bh->b_this_page; | |
397 | } | |
398 | return false; | |
399 | } | |
400 | ||
401 | bh = bh->b_this_page; | |
402 | } while (bh != head); | |
403 | return true; | |
404 | } | |
405 | #else | |
406 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
407 | enum migrate_mode mode) | |
408 | { | |
409 | return true; | |
410 | } | |
411 | #endif /* CONFIG_BLOCK */ | |
412 | ||
413 | /* | |
414 | * Replace the page in the mapping. | |
415 | * | |
416 | * The number of remaining references must be: | |
417 | * 1 for anonymous pages without a mapping | |
418 | * 2 for pages with a mapping | |
419 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. | |
420 | */ | |
421 | int migrate_page_move_mapping(struct address_space *mapping, | |
422 | struct page *newpage, struct page *page, | |
423 | struct buffer_head *head, enum migrate_mode mode, | |
424 | int extra_count) | |
425 | { | |
426 | struct zone *oldzone, *newzone; | |
427 | int dirty; | |
428 | int expected_count = 1 + extra_count; | |
429 | void **pslot; | |
430 | ||
431 | /* | |
432 | * ZONE_DEVICE pages have 1 refcount always held by their device | |
433 | * | |
434 | * Note that DAX memory will never reach that point as it does not have | |
435 | * the MEMORY_DEVICE_ALLOW_MIGRATE flag set (see memory_hotplug.h). | |
436 | */ | |
437 | expected_count += is_zone_device_page(page); | |
438 | ||
439 | if (!mapping) { | |
440 | /* Anonymous page without mapping */ | |
441 | if (page_count(page) != expected_count) | |
442 | return -EAGAIN; | |
443 | ||
444 | /* No turning back from here */ | |
445 | newpage->index = page->index; | |
446 | newpage->mapping = page->mapping; | |
447 | if (PageSwapBacked(page)) | |
448 | __SetPageSwapBacked(newpage); | |
449 | ||
450 | return MIGRATEPAGE_SUCCESS; | |
451 | } | |
452 | ||
453 | oldzone = page_zone(page); | |
454 | newzone = page_zone(newpage); | |
455 | ||
456 | spin_lock_irq(&mapping->tree_lock); | |
457 | ||
458 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
459 | page_index(page)); | |
460 | ||
461 | expected_count += 1 + page_has_private(page); | |
462 | if (page_count(page) != expected_count || | |
463 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { | |
464 | spin_unlock_irq(&mapping->tree_lock); | |
465 | return -EAGAIN; | |
466 | } | |
467 | ||
468 | if (!page_ref_freeze(page, expected_count)) { | |
469 | spin_unlock_irq(&mapping->tree_lock); | |
470 | return -EAGAIN; | |
471 | } | |
472 | ||
473 | /* | |
474 | * In the async migration case of moving a page with buffers, lock the | |
475 | * buffers using trylock before the mapping is moved. If the mapping | |
476 | * was moved, we later failed to lock the buffers and could not move | |
477 | * the mapping back due to an elevated page count, we would have to | |
478 | * block waiting on other references to be dropped. | |
479 | */ | |
480 | if (mode == MIGRATE_ASYNC && head && | |
481 | !buffer_migrate_lock_buffers(head, mode)) { | |
482 | page_ref_unfreeze(page, expected_count); | |
483 | spin_unlock_irq(&mapping->tree_lock); | |
484 | return -EAGAIN; | |
485 | } | |
486 | ||
487 | /* | |
488 | * Now we know that no one else is looking at the page: | |
489 | * no turning back from here. | |
490 | */ | |
491 | newpage->index = page->index; | |
492 | newpage->mapping = page->mapping; | |
493 | get_page(newpage); /* add cache reference */ | |
494 | if (PageSwapBacked(page)) { | |
495 | __SetPageSwapBacked(newpage); | |
496 | if (PageSwapCache(page)) { | |
497 | SetPageSwapCache(newpage); | |
498 | set_page_private(newpage, page_private(page)); | |
499 | } | |
500 | } else { | |
501 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
502 | } | |
503 | ||
504 | /* Move dirty while page refs frozen and newpage not yet exposed */ | |
505 | dirty = PageDirty(page); | |
506 | if (dirty) { | |
507 | ClearPageDirty(page); | |
508 | SetPageDirty(newpage); | |
509 | } | |
510 | ||
511 | radix_tree_replace_slot(&mapping->page_tree, pslot, newpage); | |
512 | ||
513 | /* | |
514 | * Drop cache reference from old page by unfreezing | |
515 | * to one less reference. | |
516 | * We know this isn't the last reference. | |
517 | */ | |
518 | page_ref_unfreeze(page, expected_count - 1); | |
519 | ||
520 | spin_unlock(&mapping->tree_lock); | |
521 | /* Leave irq disabled to prevent preemption while updating stats */ | |
522 | ||
523 | /* | |
524 | * If moved to a different zone then also account | |
525 | * the page for that zone. Other VM counters will be | |
526 | * taken care of when we establish references to the | |
527 | * new page and drop references to the old page. | |
528 | * | |
529 | * Note that anonymous pages are accounted for | |
530 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they | |
531 | * are mapped to swap space. | |
532 | */ | |
533 | if (newzone != oldzone) { | |
534 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES); | |
535 | __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES); | |
536 | if (PageSwapBacked(page) && !PageSwapCache(page)) { | |
537 | __dec_node_state(oldzone->zone_pgdat, NR_SHMEM); | |
538 | __inc_node_state(newzone->zone_pgdat, NR_SHMEM); | |
539 | } | |
540 | if (dirty && mapping_cap_account_dirty(mapping)) { | |
541 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY); | |
542 | __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING); | |
543 | __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY); | |
544 | __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING); | |
545 | } | |
546 | } | |
547 | local_irq_enable(); | |
548 | ||
549 | return MIGRATEPAGE_SUCCESS; | |
550 | } | |
551 | EXPORT_SYMBOL(migrate_page_move_mapping); | |
552 | ||
553 | /* | |
554 | * The expected number of remaining references is the same as that | |
555 | * of migrate_page_move_mapping(). | |
556 | */ | |
557 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
558 | struct page *newpage, struct page *page) | |
559 | { | |
560 | int expected_count; | |
561 | void **pslot; | |
562 | ||
563 | spin_lock_irq(&mapping->tree_lock); | |
564 | ||
565 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
566 | page_index(page)); | |
567 | ||
568 | expected_count = 2 + page_has_private(page); | |
569 | if (page_count(page) != expected_count || | |
570 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { | |
571 | spin_unlock_irq(&mapping->tree_lock); | |
572 | return -EAGAIN; | |
573 | } | |
574 | ||
575 | if (!page_ref_freeze(page, expected_count)) { | |
576 | spin_unlock_irq(&mapping->tree_lock); | |
577 | return -EAGAIN; | |
578 | } | |
579 | ||
580 | newpage->index = page->index; | |
581 | newpage->mapping = page->mapping; | |
582 | ||
583 | get_page(newpage); | |
584 | ||
585 | radix_tree_replace_slot(&mapping->page_tree, pslot, newpage); | |
586 | ||
587 | page_ref_unfreeze(page, expected_count - 1); | |
588 | ||
589 | spin_unlock_irq(&mapping->tree_lock); | |
590 | ||
591 | return MIGRATEPAGE_SUCCESS; | |
592 | } | |
593 | ||
594 | /* | |
595 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
596 | * arithmetic will work across the entire page. We need something more | |
597 | * specialized. | |
598 | */ | |
599 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
600 | int nr_pages) | |
601 | { | |
602 | int i; | |
603 | struct page *dst_base = dst; | |
604 | struct page *src_base = src; | |
605 | ||
606 | for (i = 0; i < nr_pages; ) { | |
607 | cond_resched(); | |
608 | copy_highpage(dst, src); | |
609 | ||
610 | i++; | |
611 | dst = mem_map_next(dst, dst_base, i); | |
612 | src = mem_map_next(src, src_base, i); | |
613 | } | |
614 | } | |
615 | ||
616 | static void copy_huge_page(struct page *dst, struct page *src) | |
617 | { | |
618 | int i; | |
619 | int nr_pages; | |
620 | ||
621 | if (PageHuge(src)) { | |
622 | /* hugetlbfs page */ | |
623 | struct hstate *h = page_hstate(src); | |
624 | nr_pages = pages_per_huge_page(h); | |
625 | ||
626 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
627 | __copy_gigantic_page(dst, src, nr_pages); | |
628 | return; | |
629 | } | |
630 | } else { | |
631 | /* thp page */ | |
632 | BUG_ON(!PageTransHuge(src)); | |
633 | nr_pages = hpage_nr_pages(src); | |
634 | } | |
635 | ||
636 | for (i = 0; i < nr_pages; i++) { | |
637 | cond_resched(); | |
638 | copy_highpage(dst + i, src + i); | |
639 | } | |
640 | } | |
641 | ||
642 | /* | |
643 | * Copy the page to its new location | |
644 | */ | |
645 | void migrate_page_states(struct page *newpage, struct page *page) | |
646 | { | |
647 | int cpupid; | |
648 | ||
649 | if (PageError(page)) | |
650 | SetPageError(newpage); | |
651 | if (PageReferenced(page)) | |
652 | SetPageReferenced(newpage); | |
653 | if (PageUptodate(page)) | |
654 | SetPageUptodate(newpage); | |
655 | if (TestClearPageActive(page)) { | |
656 | VM_BUG_ON_PAGE(PageUnevictable(page), page); | |
657 | SetPageActive(newpage); | |
658 | } else if (TestClearPageUnevictable(page)) | |
659 | SetPageUnevictable(newpage); | |
660 | if (PageChecked(page)) | |
661 | SetPageChecked(newpage); | |
662 | if (PageMappedToDisk(page)) | |
663 | SetPageMappedToDisk(newpage); | |
664 | ||
665 | /* Move dirty on pages not done by migrate_page_move_mapping() */ | |
666 | if (PageDirty(page)) | |
667 | SetPageDirty(newpage); | |
668 | ||
669 | if (page_is_young(page)) | |
670 | set_page_young(newpage); | |
671 | if (page_is_idle(page)) | |
672 | set_page_idle(newpage); | |
673 | ||
674 | /* | |
675 | * Copy NUMA information to the new page, to prevent over-eager | |
676 | * future migrations of this same page. | |
677 | */ | |
678 | cpupid = page_cpupid_xchg_last(page, -1); | |
679 | page_cpupid_xchg_last(newpage, cpupid); | |
680 | ||
681 | ksm_migrate_page(newpage, page); | |
682 | /* | |
683 | * Please do not reorder this without considering how mm/ksm.c's | |
684 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
685 | */ | |
686 | if (PageSwapCache(page)) | |
687 | ClearPageSwapCache(page); | |
688 | ClearPagePrivate(page); | |
689 | set_page_private(page, 0); | |
690 | ||
691 | /* | |
692 | * If any waiters have accumulated on the new page then | |
693 | * wake them up. | |
694 | */ | |
695 | if (PageWriteback(newpage)) | |
696 | end_page_writeback(newpage); | |
697 | ||
698 | copy_page_owner(page, newpage); | |
699 | ||
700 | mem_cgroup_migrate(page, newpage); | |
701 | } | |
702 | EXPORT_SYMBOL(migrate_page_states); | |
703 | ||
704 | void migrate_page_copy(struct page *newpage, struct page *page) | |
705 | { | |
706 | if (PageHuge(page) || PageTransHuge(page)) | |
707 | copy_huge_page(newpage, page); | |
708 | else | |
709 | copy_highpage(newpage, page); | |
710 | ||
711 | migrate_page_states(newpage, page); | |
712 | } | |
713 | EXPORT_SYMBOL(migrate_page_copy); | |
714 | ||
715 | /************************************************************ | |
716 | * Migration functions | |
717 | ***********************************************************/ | |
718 | ||
719 | /* | |
720 | * Common logic to directly migrate a single LRU page suitable for | |
721 | * pages that do not use PagePrivate/PagePrivate2. | |
722 | * | |
723 | * Pages are locked upon entry and exit. | |
724 | */ | |
725 | int migrate_page(struct address_space *mapping, | |
726 | struct page *newpage, struct page *page, | |
727 | enum migrate_mode mode) | |
728 | { | |
729 | int rc; | |
730 | ||
731 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
732 | ||
733 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); | |
734 | ||
735 | if (rc != MIGRATEPAGE_SUCCESS) | |
736 | return rc; | |
737 | ||
738 | if (mode != MIGRATE_SYNC_NO_COPY) | |
739 | migrate_page_copy(newpage, page); | |
740 | else | |
741 | migrate_page_states(newpage, page); | |
742 | return MIGRATEPAGE_SUCCESS; | |
743 | } | |
744 | EXPORT_SYMBOL(migrate_page); | |
745 | ||
746 | #ifdef CONFIG_BLOCK | |
747 | /* | |
748 | * Migration function for pages with buffers. This function can only be used | |
749 | * if the underlying filesystem guarantees that no other references to "page" | |
750 | * exist. | |
751 | */ | |
752 | int buffer_migrate_page(struct address_space *mapping, | |
753 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
754 | { | |
755 | struct buffer_head *bh, *head; | |
756 | int rc; | |
757 | ||
758 | if (!page_has_buffers(page)) | |
759 | return migrate_page(mapping, newpage, page, mode); | |
760 | ||
761 | head = page_buffers(page); | |
762 | ||
763 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0); | |
764 | ||
765 | if (rc != MIGRATEPAGE_SUCCESS) | |
766 | return rc; | |
767 | ||
768 | /* | |
769 | * In the async case, migrate_page_move_mapping locked the buffers | |
770 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
771 | * need to be locked now | |
772 | */ | |
773 | if (mode != MIGRATE_ASYNC) | |
774 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
775 | ||
776 | ClearPagePrivate(page); | |
777 | set_page_private(newpage, page_private(page)); | |
778 | set_page_private(page, 0); | |
779 | put_page(page); | |
780 | get_page(newpage); | |
781 | ||
782 | bh = head; | |
783 | do { | |
784 | set_bh_page(bh, newpage, bh_offset(bh)); | |
785 | bh = bh->b_this_page; | |
786 | ||
787 | } while (bh != head); | |
788 | ||
789 | SetPagePrivate(newpage); | |
790 | ||
791 | if (mode != MIGRATE_SYNC_NO_COPY) | |
792 | migrate_page_copy(newpage, page); | |
793 | else | |
794 | migrate_page_states(newpage, page); | |
795 | ||
796 | bh = head; | |
797 | do { | |
798 | unlock_buffer(bh); | |
799 | put_bh(bh); | |
800 | bh = bh->b_this_page; | |
801 | ||
802 | } while (bh != head); | |
803 | ||
804 | return MIGRATEPAGE_SUCCESS; | |
805 | } | |
806 | EXPORT_SYMBOL(buffer_migrate_page); | |
807 | #endif | |
808 | ||
809 | /* | |
810 | * Writeback a page to clean the dirty state | |
811 | */ | |
812 | static int writeout(struct address_space *mapping, struct page *page) | |
813 | { | |
814 | struct writeback_control wbc = { | |
815 | .sync_mode = WB_SYNC_NONE, | |
816 | .nr_to_write = 1, | |
817 | .range_start = 0, | |
818 | .range_end = LLONG_MAX, | |
819 | .for_reclaim = 1 | |
820 | }; | |
821 | int rc; | |
822 | ||
823 | if (!mapping->a_ops->writepage) | |
824 | /* No write method for the address space */ | |
825 | return -EINVAL; | |
826 | ||
827 | if (!clear_page_dirty_for_io(page)) | |
828 | /* Someone else already triggered a write */ | |
829 | return -EAGAIN; | |
830 | ||
831 | /* | |
832 | * A dirty page may imply that the underlying filesystem has | |
833 | * the page on some queue. So the page must be clean for | |
834 | * migration. Writeout may mean we loose the lock and the | |
835 | * page state is no longer what we checked for earlier. | |
836 | * At this point we know that the migration attempt cannot | |
837 | * be successful. | |
838 | */ | |
839 | remove_migration_ptes(page, page, false); | |
840 | ||
841 | rc = mapping->a_ops->writepage(page, &wbc); | |
842 | ||
843 | if (rc != AOP_WRITEPAGE_ACTIVATE) | |
844 | /* unlocked. Relock */ | |
845 | lock_page(page); | |
846 | ||
847 | return (rc < 0) ? -EIO : -EAGAIN; | |
848 | } | |
849 | ||
850 | /* | |
851 | * Default handling if a filesystem does not provide a migration function. | |
852 | */ | |
853 | static int fallback_migrate_page(struct address_space *mapping, | |
854 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
855 | { | |
856 | if (PageDirty(page)) { | |
857 | /* Only writeback pages in full synchronous migration */ | |
858 | switch (mode) { | |
859 | case MIGRATE_SYNC: | |
860 | case MIGRATE_SYNC_NO_COPY: | |
861 | break; | |
862 | default: | |
863 | return -EBUSY; | |
864 | } | |
865 | return writeout(mapping, page); | |
866 | } | |
867 | ||
868 | /* | |
869 | * Buffers may be managed in a filesystem specific way. | |
870 | * We must have no buffers or drop them. | |
871 | */ | |
872 | if (page_has_private(page) && | |
873 | !try_to_release_page(page, GFP_KERNEL)) | |
874 | return -EAGAIN; | |
875 | ||
876 | return migrate_page(mapping, newpage, page, mode); | |
877 | } | |
878 | ||
879 | /* | |
880 | * Move a page to a newly allocated page | |
881 | * The page is locked and all ptes have been successfully removed. | |
882 | * | |
883 | * The new page will have replaced the old page if this function | |
884 | * is successful. | |
885 | * | |
886 | * Return value: | |
887 | * < 0 - error code | |
888 | * MIGRATEPAGE_SUCCESS - success | |
889 | */ | |
890 | static int move_to_new_page(struct page *newpage, struct page *page, | |
891 | enum migrate_mode mode) | |
892 | { | |
893 | struct address_space *mapping; | |
894 | int rc = -EAGAIN; | |
895 | bool is_lru = !__PageMovable(page); | |
896 | ||
897 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
898 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
899 | ||
900 | mapping = page_mapping(page); | |
901 | ||
902 | if (likely(is_lru)) { | |
903 | if (!mapping) | |
904 | rc = migrate_page(mapping, newpage, page, mode); | |
905 | else if (mapping->a_ops->migratepage) | |
906 | /* | |
907 | * Most pages have a mapping and most filesystems | |
908 | * provide a migratepage callback. Anonymous pages | |
909 | * are part of swap space which also has its own | |
910 | * migratepage callback. This is the most common path | |
911 | * for page migration. | |
912 | */ | |
913 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
914 | page, mode); | |
915 | else | |
916 | rc = fallback_migrate_page(mapping, newpage, | |
917 | page, mode); | |
918 | } else { | |
919 | /* | |
920 | * In case of non-lru page, it could be released after | |
921 | * isolation step. In that case, we shouldn't try migration. | |
922 | */ | |
923 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
924 | if (!PageMovable(page)) { | |
925 | rc = MIGRATEPAGE_SUCCESS; | |
926 | __ClearPageIsolated(page); | |
927 | goto out; | |
928 | } | |
929 | ||
930 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
931 | page, mode); | |
932 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && | |
933 | !PageIsolated(page)); | |
934 | } | |
935 | ||
936 | /* | |
937 | * When successful, old pagecache page->mapping must be cleared before | |
938 | * page is freed; but stats require that PageAnon be left as PageAnon. | |
939 | */ | |
940 | if (rc == MIGRATEPAGE_SUCCESS) { | |
941 | if (__PageMovable(page)) { | |
942 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
943 | ||
944 | /* | |
945 | * We clear PG_movable under page_lock so any compactor | |
946 | * cannot try to migrate this page. | |
947 | */ | |
948 | __ClearPageIsolated(page); | |
949 | } | |
950 | ||
951 | /* | |
952 | * Anonymous and movable page->mapping will be cleard by | |
953 | * free_pages_prepare so don't reset it here for keeping | |
954 | * the type to work PageAnon, for example. | |
955 | */ | |
956 | if (!PageMappingFlags(page)) | |
957 | page->mapping = NULL; | |
958 | } | |
959 | out: | |
960 | return rc; | |
961 | } | |
962 | ||
963 | static int __unmap_and_move(struct page *page, struct page *newpage, | |
964 | int force, enum migrate_mode mode) | |
965 | { | |
966 | int rc = -EAGAIN; | |
967 | int page_was_mapped = 0; | |
968 | struct anon_vma *anon_vma = NULL; | |
969 | bool is_lru = !__PageMovable(page); | |
970 | ||
971 | if (!trylock_page(page)) { | |
972 | if (!force || mode == MIGRATE_ASYNC) | |
973 | goto out; | |
974 | ||
975 | /* | |
976 | * It's not safe for direct compaction to call lock_page. | |
977 | * For example, during page readahead pages are added locked | |
978 | * to the LRU. Later, when the IO completes the pages are | |
979 | * marked uptodate and unlocked. However, the queueing | |
980 | * could be merging multiple pages for one bio (e.g. | |
981 | * mpage_readpages). If an allocation happens for the | |
982 | * second or third page, the process can end up locking | |
983 | * the same page twice and deadlocking. Rather than | |
984 | * trying to be clever about what pages can be locked, | |
985 | * avoid the use of lock_page for direct compaction | |
986 | * altogether. | |
987 | */ | |
988 | if (current->flags & PF_MEMALLOC) | |
989 | goto out; | |
990 | ||
991 | lock_page(page); | |
992 | } | |
993 | ||
994 | if (PageWriteback(page)) { | |
995 | /* | |
996 | * Only in the case of a full synchronous migration is it | |
997 | * necessary to wait for PageWriteback. In the async case, | |
998 | * the retry loop is too short and in the sync-light case, | |
999 | * the overhead of stalling is too much | |
1000 | */ | |
1001 | switch (mode) { | |
1002 | case MIGRATE_SYNC: | |
1003 | case MIGRATE_SYNC_NO_COPY: | |
1004 | break; | |
1005 | default: | |
1006 | rc = -EBUSY; | |
1007 | goto out_unlock; | |
1008 | } | |
1009 | if (!force) | |
1010 | goto out_unlock; | |
1011 | wait_on_page_writeback(page); | |
1012 | } | |
1013 | ||
1014 | /* | |
1015 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, | |
1016 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1017 | * This get_anon_vma() delays freeing anon_vma pointer until the end | |
1018 | * of migration. File cache pages are no problem because of page_lock() | |
1019 | * File Caches may use write_page() or lock_page() in migration, then, | |
1020 | * just care Anon page here. | |
1021 | * | |
1022 | * Only page_get_anon_vma() understands the subtleties of | |
1023 | * getting a hold on an anon_vma from outside one of its mms. | |
1024 | * But if we cannot get anon_vma, then we won't need it anyway, | |
1025 | * because that implies that the anon page is no longer mapped | |
1026 | * (and cannot be remapped so long as we hold the page lock). | |
1027 | */ | |
1028 | if (PageAnon(page) && !PageKsm(page)) | |
1029 | anon_vma = page_get_anon_vma(page); | |
1030 | ||
1031 | /* | |
1032 | * Block others from accessing the new page when we get around to | |
1033 | * establishing additional references. We are usually the only one | |
1034 | * holding a reference to newpage at this point. We used to have a BUG | |
1035 | * here if trylock_page(newpage) fails, but would like to allow for | |
1036 | * cases where there might be a race with the previous use of newpage. | |
1037 | * This is much like races on refcount of oldpage: just don't BUG(). | |
1038 | */ | |
1039 | if (unlikely(!trylock_page(newpage))) | |
1040 | goto out_unlock; | |
1041 | ||
1042 | if (unlikely(!is_lru)) { | |
1043 | rc = move_to_new_page(newpage, page, mode); | |
1044 | goto out_unlock_both; | |
1045 | } | |
1046 | ||
1047 | /* | |
1048 | * Corner case handling: | |
1049 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1050 | * and treated as swapcache but it has no rmap yet. | |
1051 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
1052 | * trigger a BUG. So handle it here. | |
1053 | * 2. An orphaned page (see truncate_complete_page) might have | |
1054 | * fs-private metadata. The page can be picked up due to memory | |
1055 | * offlining. Everywhere else except page reclaim, the page is | |
1056 | * invisible to the vm, so the page can not be migrated. So try to | |
1057 | * free the metadata, so the page can be freed. | |
1058 | */ | |
1059 | if (!page->mapping) { | |
1060 | VM_BUG_ON_PAGE(PageAnon(page), page); | |
1061 | if (page_has_private(page)) { | |
1062 | try_to_free_buffers(page); | |
1063 | goto out_unlock_both; | |
1064 | } | |
1065 | } else if (page_mapped(page)) { | |
1066 | /* Establish migration ptes */ | |
1067 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, | |
1068 | page); | |
1069 | try_to_unmap(page, | |
1070 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1071 | page_was_mapped = 1; | |
1072 | } | |
1073 | ||
1074 | if (!page_mapped(page)) | |
1075 | rc = move_to_new_page(newpage, page, mode); | |
1076 | ||
1077 | if (page_was_mapped) | |
1078 | remove_migration_ptes(page, | |
1079 | rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); | |
1080 | ||
1081 | out_unlock_both: | |
1082 | unlock_page(newpage); | |
1083 | out_unlock: | |
1084 | /* Drop an anon_vma reference if we took one */ | |
1085 | if (anon_vma) | |
1086 | put_anon_vma(anon_vma); | |
1087 | unlock_page(page); | |
1088 | out: | |
1089 | /* | |
1090 | * If migration is successful, decrease refcount of the newpage | |
1091 | * which will not free the page because new page owner increased | |
1092 | * refcounter. As well, if it is LRU page, add the page to LRU | |
1093 | * list in here. | |
1094 | */ | |
1095 | if (rc == MIGRATEPAGE_SUCCESS) { | |
1096 | if (unlikely(__PageMovable(newpage))) | |
1097 | put_page(newpage); | |
1098 | else | |
1099 | putback_lru_page(newpage); | |
1100 | } | |
1101 | ||
1102 | return rc; | |
1103 | } | |
1104 | ||
1105 | /* | |
1106 | * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work | |
1107 | * around it. | |
1108 | */ | |
1109 | #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM) | |
1110 | #define ICE_noinline noinline | |
1111 | #else | |
1112 | #define ICE_noinline | |
1113 | #endif | |
1114 | ||
1115 | /* | |
1116 | * Obtain the lock on page, remove all ptes and migrate the page | |
1117 | * to the newly allocated page in newpage. | |
1118 | */ | |
1119 | static ICE_noinline int unmap_and_move(new_page_t get_new_page, | |
1120 | free_page_t put_new_page, | |
1121 | unsigned long private, struct page *page, | |
1122 | int force, enum migrate_mode mode, | |
1123 | enum migrate_reason reason) | |
1124 | { | |
1125 | int rc = MIGRATEPAGE_SUCCESS; | |
1126 | int *result = NULL; | |
1127 | struct page *newpage; | |
1128 | ||
1129 | newpage = get_new_page(page, private, &result); | |
1130 | if (!newpage) | |
1131 | return -ENOMEM; | |
1132 | ||
1133 | if (page_count(page) == 1) { | |
1134 | /* page was freed from under us. So we are done. */ | |
1135 | ClearPageActive(page); | |
1136 | ClearPageUnevictable(page); | |
1137 | if (unlikely(__PageMovable(page))) { | |
1138 | lock_page(page); | |
1139 | if (!PageMovable(page)) | |
1140 | __ClearPageIsolated(page); | |
1141 | unlock_page(page); | |
1142 | } | |
1143 | if (put_new_page) | |
1144 | put_new_page(newpage, private); | |
1145 | else | |
1146 | put_page(newpage); | |
1147 | goto out; | |
1148 | } | |
1149 | ||
1150 | if (unlikely(PageTransHuge(page) && !PageTransHuge(newpage))) { | |
1151 | lock_page(page); | |
1152 | rc = split_huge_page(page); | |
1153 | unlock_page(page); | |
1154 | if (rc) | |
1155 | goto out; | |
1156 | } | |
1157 | ||
1158 | rc = __unmap_and_move(page, newpage, force, mode); | |
1159 | if (rc == MIGRATEPAGE_SUCCESS) | |
1160 | set_page_owner_migrate_reason(newpage, reason); | |
1161 | ||
1162 | out: | |
1163 | if (rc != -EAGAIN) { | |
1164 | /* | |
1165 | * A page that has been migrated has all references | |
1166 | * removed and will be freed. A page that has not been | |
1167 | * migrated will have kepts its references and be | |
1168 | * restored. | |
1169 | */ | |
1170 | list_del(&page->lru); | |
1171 | ||
1172 | /* | |
1173 | * Compaction can migrate also non-LRU pages which are | |
1174 | * not accounted to NR_ISOLATED_*. They can be recognized | |
1175 | * as __PageMovable | |
1176 | */ | |
1177 | if (likely(!__PageMovable(page))) | |
1178 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + | |
1179 | page_is_file_cache(page), -hpage_nr_pages(page)); | |
1180 | } | |
1181 | ||
1182 | /* | |
1183 | * If migration is successful, releases reference grabbed during | |
1184 | * isolation. Otherwise, restore the page to right list unless | |
1185 | * we want to retry. | |
1186 | */ | |
1187 | if (rc == MIGRATEPAGE_SUCCESS) { | |
1188 | put_page(page); | |
1189 | if (reason == MR_MEMORY_FAILURE) { | |
1190 | /* | |
1191 | * Set PG_HWPoison on just freed page | |
1192 | * intentionally. Although it's rather weird, | |
1193 | * it's how HWPoison flag works at the moment. | |
1194 | */ | |
1195 | if (!test_set_page_hwpoison(page)) | |
1196 | num_poisoned_pages_inc(); | |
1197 | } | |
1198 | } else { | |
1199 | if (rc != -EAGAIN) { | |
1200 | if (likely(!__PageMovable(page))) { | |
1201 | putback_lru_page(page); | |
1202 | goto put_new; | |
1203 | } | |
1204 | ||
1205 | lock_page(page); | |
1206 | if (PageMovable(page)) | |
1207 | putback_movable_page(page); | |
1208 | else | |
1209 | __ClearPageIsolated(page); | |
1210 | unlock_page(page); | |
1211 | put_page(page); | |
1212 | } | |
1213 | put_new: | |
1214 | if (put_new_page) | |
1215 | put_new_page(newpage, private); | |
1216 | else | |
1217 | put_page(newpage); | |
1218 | } | |
1219 | ||
1220 | if (result) { | |
1221 | if (rc) | |
1222 | *result = rc; | |
1223 | else | |
1224 | *result = page_to_nid(newpage); | |
1225 | } | |
1226 | return rc; | |
1227 | } | |
1228 | ||
1229 | /* | |
1230 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1231 | * | |
1232 | * This function doesn't wait the completion of hugepage I/O | |
1233 | * because there is no race between I/O and migration for hugepage. | |
1234 | * Note that currently hugepage I/O occurs only in direct I/O | |
1235 | * where no lock is held and PG_writeback is irrelevant, | |
1236 | * and writeback status of all subpages are counted in the reference | |
1237 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1238 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1239 | * This means that when we try to migrate hugepage whose subpages are | |
1240 | * doing direct I/O, some references remain after try_to_unmap() and | |
1241 | * hugepage migration fails without data corruption. | |
1242 | * | |
1243 | * There is also no race when direct I/O is issued on the page under migration, | |
1244 | * because then pte is replaced with migration swap entry and direct I/O code | |
1245 | * will wait in the page fault for migration to complete. | |
1246 | */ | |
1247 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
1248 | free_page_t put_new_page, unsigned long private, | |
1249 | struct page *hpage, int force, | |
1250 | enum migrate_mode mode, int reason) | |
1251 | { | |
1252 | int rc = -EAGAIN; | |
1253 | int *result = NULL; | |
1254 | int page_was_mapped = 0; | |
1255 | struct page *new_hpage; | |
1256 | struct anon_vma *anon_vma = NULL; | |
1257 | ||
1258 | /* | |
1259 | * Movability of hugepages depends on architectures and hugepage size. | |
1260 | * This check is necessary because some callers of hugepage migration | |
1261 | * like soft offline and memory hotremove don't walk through page | |
1262 | * tables or check whether the hugepage is pmd-based or not before | |
1263 | * kicking migration. | |
1264 | */ | |
1265 | if (!hugepage_migration_supported(page_hstate(hpage))) { | |
1266 | putback_active_hugepage(hpage); | |
1267 | return -ENOSYS; | |
1268 | } | |
1269 | ||
1270 | new_hpage = get_new_page(hpage, private, &result); | |
1271 | if (!new_hpage) | |
1272 | return -ENOMEM; | |
1273 | ||
1274 | if (!trylock_page(hpage)) { | |
1275 | if (!force) | |
1276 | goto out; | |
1277 | switch (mode) { | |
1278 | case MIGRATE_SYNC: | |
1279 | case MIGRATE_SYNC_NO_COPY: | |
1280 | break; | |
1281 | default: | |
1282 | goto out; | |
1283 | } | |
1284 | lock_page(hpage); | |
1285 | } | |
1286 | ||
1287 | if (PageAnon(hpage)) | |
1288 | anon_vma = page_get_anon_vma(hpage); | |
1289 | ||
1290 | if (unlikely(!trylock_page(new_hpage))) | |
1291 | goto put_anon; | |
1292 | ||
1293 | if (page_mapped(hpage)) { | |
1294 | try_to_unmap(hpage, | |
1295 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
1296 | page_was_mapped = 1; | |
1297 | } | |
1298 | ||
1299 | if (!page_mapped(hpage)) | |
1300 | rc = move_to_new_page(new_hpage, hpage, mode); | |
1301 | ||
1302 | if (page_was_mapped) | |
1303 | remove_migration_ptes(hpage, | |
1304 | rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); | |
1305 | ||
1306 | unlock_page(new_hpage); | |
1307 | ||
1308 | put_anon: | |
1309 | if (anon_vma) | |
1310 | put_anon_vma(anon_vma); | |
1311 | ||
1312 | if (rc == MIGRATEPAGE_SUCCESS) { | |
1313 | hugetlb_cgroup_migrate(hpage, new_hpage); | |
1314 | put_new_page = NULL; | |
1315 | set_page_owner_migrate_reason(new_hpage, reason); | |
1316 | } | |
1317 | ||
1318 | unlock_page(hpage); | |
1319 | out: | |
1320 | if (rc != -EAGAIN) | |
1321 | putback_active_hugepage(hpage); | |
1322 | if (reason == MR_MEMORY_FAILURE && !test_set_page_hwpoison(hpage)) | |
1323 | num_poisoned_pages_inc(); | |
1324 | ||
1325 | /* | |
1326 | * If migration was not successful and there's a freeing callback, use | |
1327 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1328 | * isolation. | |
1329 | */ | |
1330 | if (put_new_page) | |
1331 | put_new_page(new_hpage, private); | |
1332 | else | |
1333 | putback_active_hugepage(new_hpage); | |
1334 | ||
1335 | if (result) { | |
1336 | if (rc) | |
1337 | *result = rc; | |
1338 | else | |
1339 | *result = page_to_nid(new_hpage); | |
1340 | } | |
1341 | return rc; | |
1342 | } | |
1343 | ||
1344 | /* | |
1345 | * migrate_pages - migrate the pages specified in a list, to the free pages | |
1346 | * supplied as the target for the page migration | |
1347 | * | |
1348 | * @from: The list of pages to be migrated. | |
1349 | * @get_new_page: The function used to allocate free pages to be used | |
1350 | * as the target of the page migration. | |
1351 | * @put_new_page: The function used to free target pages if migration | |
1352 | * fails, or NULL if no special handling is necessary. | |
1353 | * @private: Private data to be passed on to get_new_page() | |
1354 | * @mode: The migration mode that specifies the constraints for | |
1355 | * page migration, if any. | |
1356 | * @reason: The reason for page migration. | |
1357 | * | |
1358 | * The function returns after 10 attempts or if no pages are movable any more | |
1359 | * because the list has become empty or no retryable pages exist any more. | |
1360 | * The caller should call putback_movable_pages() to return pages to the LRU | |
1361 | * or free list only if ret != 0. | |
1362 | * | |
1363 | * Returns the number of pages that were not migrated, or an error code. | |
1364 | */ | |
1365 | int migrate_pages(struct list_head *from, new_page_t get_new_page, | |
1366 | free_page_t put_new_page, unsigned long private, | |
1367 | enum migrate_mode mode, int reason) | |
1368 | { | |
1369 | int retry = 1; | |
1370 | int nr_failed = 0; | |
1371 | int nr_succeeded = 0; | |
1372 | int pass = 0; | |
1373 | struct page *page; | |
1374 | struct page *page2; | |
1375 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1376 | int rc; | |
1377 | ||
1378 | if (!swapwrite) | |
1379 | current->flags |= PF_SWAPWRITE; | |
1380 | ||
1381 | for(pass = 0; pass < 10 && retry; pass++) { | |
1382 | retry = 0; | |
1383 | ||
1384 | list_for_each_entry_safe(page, page2, from, lru) { | |
1385 | cond_resched(); | |
1386 | ||
1387 | if (PageHuge(page)) | |
1388 | rc = unmap_and_move_huge_page(get_new_page, | |
1389 | put_new_page, private, page, | |
1390 | pass > 2, mode, reason); | |
1391 | else | |
1392 | rc = unmap_and_move(get_new_page, put_new_page, | |
1393 | private, page, pass > 2, mode, | |
1394 | reason); | |
1395 | ||
1396 | switch(rc) { | |
1397 | case -ENOMEM: | |
1398 | nr_failed++; | |
1399 | goto out; | |
1400 | case -EAGAIN: | |
1401 | retry++; | |
1402 | break; | |
1403 | case MIGRATEPAGE_SUCCESS: | |
1404 | nr_succeeded++; | |
1405 | break; | |
1406 | default: | |
1407 | /* | |
1408 | * Permanent failure (-EBUSY, -ENOSYS, etc.): | |
1409 | * unlike -EAGAIN case, the failed page is | |
1410 | * removed from migration page list and not | |
1411 | * retried in the next outer loop. | |
1412 | */ | |
1413 | nr_failed++; | |
1414 | break; | |
1415 | } | |
1416 | } | |
1417 | } | |
1418 | nr_failed += retry; | |
1419 | rc = nr_failed; | |
1420 | out: | |
1421 | if (nr_succeeded) | |
1422 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1423 | if (nr_failed) | |
1424 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
1425 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); | |
1426 | ||
1427 | if (!swapwrite) | |
1428 | current->flags &= ~PF_SWAPWRITE; | |
1429 | ||
1430 | return rc; | |
1431 | } | |
1432 | ||
1433 | #ifdef CONFIG_NUMA | |
1434 | /* | |
1435 | * Move a list of individual pages | |
1436 | */ | |
1437 | struct page_to_node { | |
1438 | unsigned long addr; | |
1439 | struct page *page; | |
1440 | int node; | |
1441 | int status; | |
1442 | }; | |
1443 | ||
1444 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1445 | int **result) | |
1446 | { | |
1447 | struct page_to_node *pm = (struct page_to_node *)private; | |
1448 | ||
1449 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1450 | pm++; | |
1451 | ||
1452 | if (pm->node == MAX_NUMNODES) | |
1453 | return NULL; | |
1454 | ||
1455 | *result = &pm->status; | |
1456 | ||
1457 | if (PageHuge(p)) | |
1458 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1459 | pm->node); | |
1460 | else if (thp_migration_supported() && PageTransHuge(p)) { | |
1461 | struct page *thp; | |
1462 | ||
1463 | thp = alloc_pages_node(pm->node, | |
1464 | (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM, | |
1465 | HPAGE_PMD_ORDER); | |
1466 | if (!thp) | |
1467 | return NULL; | |
1468 | prep_transhuge_page(thp); | |
1469 | return thp; | |
1470 | } else | |
1471 | return __alloc_pages_node(pm->node, | |
1472 | GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0); | |
1473 | } | |
1474 | ||
1475 | /* | |
1476 | * Move a set of pages as indicated in the pm array. The addr | |
1477 | * field must be set to the virtual address of the page to be moved | |
1478 | * and the node number must contain a valid target node. | |
1479 | * The pm array ends with node = MAX_NUMNODES. | |
1480 | */ | |
1481 | static int do_move_page_to_node_array(struct mm_struct *mm, | |
1482 | struct page_to_node *pm, | |
1483 | int migrate_all) | |
1484 | { | |
1485 | int err; | |
1486 | struct page_to_node *pp; | |
1487 | LIST_HEAD(pagelist); | |
1488 | ||
1489 | down_read(&mm->mmap_sem); | |
1490 | ||
1491 | /* | |
1492 | * Build a list of pages to migrate | |
1493 | */ | |
1494 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { | |
1495 | struct vm_area_struct *vma; | |
1496 | struct page *page; | |
1497 | struct page *head; | |
1498 | unsigned int follflags; | |
1499 | ||
1500 | err = -EFAULT; | |
1501 | vma = find_vma(mm, pp->addr); | |
1502 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) | |
1503 | goto set_status; | |
1504 | ||
1505 | /* FOLL_DUMP to ignore special (like zero) pages */ | |
1506 | follflags = FOLL_GET | FOLL_DUMP; | |
1507 | if (!thp_migration_supported()) | |
1508 | follflags |= FOLL_SPLIT; | |
1509 | page = follow_page(vma, pp->addr, follflags); | |
1510 | ||
1511 | err = PTR_ERR(page); | |
1512 | if (IS_ERR(page)) | |
1513 | goto set_status; | |
1514 | ||
1515 | err = -ENOENT; | |
1516 | if (!page) | |
1517 | goto set_status; | |
1518 | ||
1519 | err = page_to_nid(page); | |
1520 | ||
1521 | if (err == pp->node) | |
1522 | /* | |
1523 | * Node already in the right place | |
1524 | */ | |
1525 | goto put_and_set; | |
1526 | ||
1527 | err = -EACCES; | |
1528 | if (page_mapcount(page) > 1 && | |
1529 | !migrate_all) | |
1530 | goto put_and_set; | |
1531 | ||
1532 | if (PageHuge(page)) { | |
1533 | if (PageHead(page)) { | |
1534 | isolate_huge_page(page, &pagelist); | |
1535 | err = 0; | |
1536 | pp->page = page; | |
1537 | } | |
1538 | goto put_and_set; | |
1539 | } | |
1540 | ||
1541 | pp->page = compound_head(page); | |
1542 | head = compound_head(page); | |
1543 | err = isolate_lru_page(head); | |
1544 | if (!err) { | |
1545 | list_add_tail(&head->lru, &pagelist); | |
1546 | mod_node_page_state(page_pgdat(head), | |
1547 | NR_ISOLATED_ANON + page_is_file_cache(head), | |
1548 | hpage_nr_pages(head)); | |
1549 | } | |
1550 | put_and_set: | |
1551 | /* | |
1552 | * Either remove the duplicate refcount from | |
1553 | * isolate_lru_page() or drop the page ref if it was | |
1554 | * not isolated. | |
1555 | */ | |
1556 | put_page(page); | |
1557 | set_status: | |
1558 | pp->status = err; | |
1559 | } | |
1560 | ||
1561 | err = 0; | |
1562 | if (!list_empty(&pagelist)) { | |
1563 | err = migrate_pages(&pagelist, new_page_node, NULL, | |
1564 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); | |
1565 | if (err) | |
1566 | putback_movable_pages(&pagelist); | |
1567 | } | |
1568 | ||
1569 | up_read(&mm->mmap_sem); | |
1570 | return err; | |
1571 | } | |
1572 | ||
1573 | /* | |
1574 | * Migrate an array of page address onto an array of nodes and fill | |
1575 | * the corresponding array of status. | |
1576 | */ | |
1577 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, | |
1578 | unsigned long nr_pages, | |
1579 | const void __user * __user *pages, | |
1580 | const int __user *nodes, | |
1581 | int __user *status, int flags) | |
1582 | { | |
1583 | struct page_to_node *pm; | |
1584 | unsigned long chunk_nr_pages; | |
1585 | unsigned long chunk_start; | |
1586 | int err; | |
1587 | ||
1588 | err = -ENOMEM; | |
1589 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1590 | if (!pm) | |
1591 | goto out; | |
1592 | ||
1593 | migrate_prep(); | |
1594 | ||
1595 | /* | |
1596 | * Store a chunk of page_to_node array in a page, | |
1597 | * but keep the last one as a marker | |
1598 | */ | |
1599 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; | |
1600 | ||
1601 | for (chunk_start = 0; | |
1602 | chunk_start < nr_pages; | |
1603 | chunk_start += chunk_nr_pages) { | |
1604 | int j; | |
1605 | ||
1606 | if (chunk_start + chunk_nr_pages > nr_pages) | |
1607 | chunk_nr_pages = nr_pages - chunk_start; | |
1608 | ||
1609 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1610 | for (j = 0; j < chunk_nr_pages; j++) { | |
1611 | const void __user *p; | |
1612 | int node; | |
1613 | ||
1614 | err = -EFAULT; | |
1615 | if (get_user(p, pages + j + chunk_start)) | |
1616 | goto out_pm; | |
1617 | pm[j].addr = (unsigned long) p; | |
1618 | ||
1619 | if (get_user(node, nodes + j + chunk_start)) | |
1620 | goto out_pm; | |
1621 | ||
1622 | err = -ENODEV; | |
1623 | if (node < 0 || node >= MAX_NUMNODES) | |
1624 | goto out_pm; | |
1625 | ||
1626 | if (!node_state(node, N_MEMORY)) | |
1627 | goto out_pm; | |
1628 | ||
1629 | err = -EACCES; | |
1630 | if (!node_isset(node, task_nodes)) | |
1631 | goto out_pm; | |
1632 | ||
1633 | pm[j].node = node; | |
1634 | } | |
1635 | ||
1636 | /* End marker for this chunk */ | |
1637 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1638 | ||
1639 | /* Migrate this chunk */ | |
1640 | err = do_move_page_to_node_array(mm, pm, | |
1641 | flags & MPOL_MF_MOVE_ALL); | |
1642 | if (err < 0) | |
1643 | goto out_pm; | |
1644 | ||
1645 | /* Return status information */ | |
1646 | for (j = 0; j < chunk_nr_pages; j++) | |
1647 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
1648 | err = -EFAULT; | |
1649 | goto out_pm; | |
1650 | } | |
1651 | } | |
1652 | err = 0; | |
1653 | ||
1654 | out_pm: | |
1655 | free_page((unsigned long)pm); | |
1656 | out: | |
1657 | return err; | |
1658 | } | |
1659 | ||
1660 | /* | |
1661 | * Determine the nodes of an array of pages and store it in an array of status. | |
1662 | */ | |
1663 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, | |
1664 | const void __user **pages, int *status) | |
1665 | { | |
1666 | unsigned long i; | |
1667 | ||
1668 | down_read(&mm->mmap_sem); | |
1669 | ||
1670 | for (i = 0; i < nr_pages; i++) { | |
1671 | unsigned long addr = (unsigned long)(*pages); | |
1672 | struct vm_area_struct *vma; | |
1673 | struct page *page; | |
1674 | int err = -EFAULT; | |
1675 | ||
1676 | vma = find_vma(mm, addr); | |
1677 | if (!vma || addr < vma->vm_start) | |
1678 | goto set_status; | |
1679 | ||
1680 | /* FOLL_DUMP to ignore special (like zero) pages */ | |
1681 | page = follow_page(vma, addr, FOLL_DUMP); | |
1682 | ||
1683 | err = PTR_ERR(page); | |
1684 | if (IS_ERR(page)) | |
1685 | goto set_status; | |
1686 | ||
1687 | err = page ? page_to_nid(page) : -ENOENT; | |
1688 | set_status: | |
1689 | *status = err; | |
1690 | ||
1691 | pages++; | |
1692 | status++; | |
1693 | } | |
1694 | ||
1695 | up_read(&mm->mmap_sem); | |
1696 | } | |
1697 | ||
1698 | /* | |
1699 | * Determine the nodes of a user array of pages and store it in | |
1700 | * a user array of status. | |
1701 | */ | |
1702 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1703 | const void __user * __user *pages, | |
1704 | int __user *status) | |
1705 | { | |
1706 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1707 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1708 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
1709 | ||
1710 | while (nr_pages) { | |
1711 | unsigned long chunk_nr; | |
1712 | ||
1713 | chunk_nr = nr_pages; | |
1714 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1715 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1716 | ||
1717 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1718 | break; | |
1719 | ||
1720 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1721 | ||
1722 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) | |
1723 | break; | |
1724 | ||
1725 | pages += chunk_nr; | |
1726 | status += chunk_nr; | |
1727 | nr_pages -= chunk_nr; | |
1728 | } | |
1729 | return nr_pages ? -EFAULT : 0; | |
1730 | } | |
1731 | ||
1732 | /* | |
1733 | * Move a list of pages in the address space of the currently executing | |
1734 | * process. | |
1735 | */ | |
1736 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, | |
1737 | const void __user * __user *, pages, | |
1738 | const int __user *, nodes, | |
1739 | int __user *, status, int, flags) | |
1740 | { | |
1741 | struct task_struct *task; | |
1742 | struct mm_struct *mm; | |
1743 | int err; | |
1744 | nodemask_t task_nodes; | |
1745 | ||
1746 | /* Check flags */ | |
1747 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1748 | return -EINVAL; | |
1749 | ||
1750 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1751 | return -EPERM; | |
1752 | ||
1753 | /* Find the mm_struct */ | |
1754 | rcu_read_lock(); | |
1755 | task = pid ? find_task_by_vpid(pid) : current; | |
1756 | if (!task) { | |
1757 | rcu_read_unlock(); | |
1758 | return -ESRCH; | |
1759 | } | |
1760 | get_task_struct(task); | |
1761 | ||
1762 | /* | |
1763 | * Check if this process has the right to modify the specified | |
1764 | * process. Use the regular "ptrace_may_access()" checks. | |
1765 | */ | |
1766 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { | |
1767 | rcu_read_unlock(); | |
1768 | err = -EPERM; | |
1769 | goto out; | |
1770 | } | |
1771 | rcu_read_unlock(); | |
1772 | ||
1773 | err = security_task_movememory(task); | |
1774 | if (err) | |
1775 | goto out; | |
1776 | ||
1777 | task_nodes = cpuset_mems_allowed(task); | |
1778 | mm = get_task_mm(task); | |
1779 | put_task_struct(task); | |
1780 | ||
1781 | if (!mm) | |
1782 | return -EINVAL; | |
1783 | ||
1784 | if (nodes) | |
1785 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1786 | nodes, status, flags); | |
1787 | else | |
1788 | err = do_pages_stat(mm, nr_pages, pages, status); | |
1789 | ||
1790 | mmput(mm); | |
1791 | return err; | |
1792 | ||
1793 | out: | |
1794 | put_task_struct(task); | |
1795 | return err; | |
1796 | } | |
1797 | ||
1798 | #ifdef CONFIG_NUMA_BALANCING | |
1799 | /* | |
1800 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1801 | * pages. Currently it only checks the watermarks which crude | |
1802 | */ | |
1803 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
1804 | unsigned long nr_migrate_pages) | |
1805 | { | |
1806 | int z; | |
1807 | ||
1808 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1809 | struct zone *zone = pgdat->node_zones + z; | |
1810 | ||
1811 | if (!populated_zone(zone)) | |
1812 | continue; | |
1813 | ||
1814 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1815 | if (!zone_watermark_ok(zone, 0, | |
1816 | high_wmark_pages(zone) + | |
1817 | nr_migrate_pages, | |
1818 | 0, 0)) | |
1819 | continue; | |
1820 | return true; | |
1821 | } | |
1822 | return false; | |
1823 | } | |
1824 | ||
1825 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1826 | unsigned long data, | |
1827 | int **result) | |
1828 | { | |
1829 | int nid = (int) data; | |
1830 | struct page *newpage; | |
1831 | ||
1832 | newpage = __alloc_pages_node(nid, | |
1833 | (GFP_HIGHUSER_MOVABLE | | |
1834 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
1835 | __GFP_NORETRY | __GFP_NOWARN) & | |
1836 | ~__GFP_RECLAIM, 0); | |
1837 | ||
1838 | return newpage; | |
1839 | } | |
1840 | ||
1841 | /* | |
1842 | * page migration rate limiting control. | |
1843 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1844 | * window of time. Default here says do not migrate more than 1280M per second. | |
1845 | */ | |
1846 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
1847 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); | |
1848 | ||
1849 | /* Returns true if the node is migrate rate-limited after the update */ | |
1850 | static bool numamigrate_update_ratelimit(pg_data_t *pgdat, | |
1851 | unsigned long nr_pages) | |
1852 | { | |
1853 | /* | |
1854 | * Rate-limit the amount of data that is being migrated to a node. | |
1855 | * Optimal placement is no good if the memory bus is saturated and | |
1856 | * all the time is being spent migrating! | |
1857 | */ | |
1858 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { | |
1859 | spin_lock(&pgdat->numabalancing_migrate_lock); | |
1860 | pgdat->numabalancing_migrate_nr_pages = 0; | |
1861 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1862 | msecs_to_jiffies(migrate_interval_millisecs); | |
1863 | spin_unlock(&pgdat->numabalancing_migrate_lock); | |
1864 | } | |
1865 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) { | |
1866 | trace_mm_numa_migrate_ratelimit(current, pgdat->node_id, | |
1867 | nr_pages); | |
1868 | return true; | |
1869 | } | |
1870 | ||
1871 | /* | |
1872 | * This is an unlocked non-atomic update so errors are possible. | |
1873 | * The consequences are failing to migrate when we potentiall should | |
1874 | * have which is not severe enough to warrant locking. If it is ever | |
1875 | * a problem, it can be converted to a per-cpu counter. | |
1876 | */ | |
1877 | pgdat->numabalancing_migrate_nr_pages += nr_pages; | |
1878 | return false; | |
1879 | } | |
1880 | ||
1881 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) | |
1882 | { | |
1883 | int page_lru; | |
1884 | ||
1885 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); | |
1886 | ||
1887 | /* Avoid migrating to a node that is nearly full */ | |
1888 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) | |
1889 | return 0; | |
1890 | ||
1891 | if (isolate_lru_page(page)) | |
1892 | return 0; | |
1893 | ||
1894 | /* | |
1895 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1896 | * check on page_count(), so we must do it here, now that the page | |
1897 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1898 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1899 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1900 | */ | |
1901 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1902 | putback_lru_page(page); | |
1903 | return 0; | |
1904 | } | |
1905 | ||
1906 | page_lru = page_is_file_cache(page); | |
1907 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, | |
1908 | hpage_nr_pages(page)); | |
1909 | ||
1910 | /* | |
1911 | * Isolating the page has taken another reference, so the | |
1912 | * caller's reference can be safely dropped without the page | |
1913 | * disappearing underneath us during migration. | |
1914 | */ | |
1915 | put_page(page); | |
1916 | return 1; | |
1917 | } | |
1918 | ||
1919 | bool pmd_trans_migrating(pmd_t pmd) | |
1920 | { | |
1921 | struct page *page = pmd_page(pmd); | |
1922 | return PageLocked(page); | |
1923 | } | |
1924 | ||
1925 | /* | |
1926 | * Attempt to migrate a misplaced page to the specified destination | |
1927 | * node. Caller is expected to have an elevated reference count on | |
1928 | * the page that will be dropped by this function before returning. | |
1929 | */ | |
1930 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, | |
1931 | int node) | |
1932 | { | |
1933 | pg_data_t *pgdat = NODE_DATA(node); | |
1934 | int isolated; | |
1935 | int nr_remaining; | |
1936 | LIST_HEAD(migratepages); | |
1937 | ||
1938 | /* | |
1939 | * Don't migrate file pages that are mapped in multiple processes | |
1940 | * with execute permissions as they are probably shared libraries. | |
1941 | */ | |
1942 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && | |
1943 | (vma->vm_flags & VM_EXEC)) | |
1944 | goto out; | |
1945 | ||
1946 | /* | |
1947 | * Rate-limit the amount of data that is being migrated to a node. | |
1948 | * Optimal placement is no good if the memory bus is saturated and | |
1949 | * all the time is being spent migrating! | |
1950 | */ | |
1951 | if (numamigrate_update_ratelimit(pgdat, 1)) | |
1952 | goto out; | |
1953 | ||
1954 | isolated = numamigrate_isolate_page(pgdat, page); | |
1955 | if (!isolated) | |
1956 | goto out; | |
1957 | ||
1958 | list_add(&page->lru, &migratepages); | |
1959 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, | |
1960 | NULL, node, MIGRATE_ASYNC, | |
1961 | MR_NUMA_MISPLACED); | |
1962 | if (nr_remaining) { | |
1963 | if (!list_empty(&migratepages)) { | |
1964 | list_del(&page->lru); | |
1965 | dec_node_page_state(page, NR_ISOLATED_ANON + | |
1966 | page_is_file_cache(page)); | |
1967 | putback_lru_page(page); | |
1968 | } | |
1969 | isolated = 0; | |
1970 | } else | |
1971 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
1972 | BUG_ON(!list_empty(&migratepages)); | |
1973 | return isolated; | |
1974 | ||
1975 | out: | |
1976 | put_page(page); | |
1977 | return 0; | |
1978 | } | |
1979 | #endif /* CONFIG_NUMA_BALANCING */ | |
1980 | ||
1981 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) | |
1982 | /* | |
1983 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1984 | * before returning. | |
1985 | */ | |
1986 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, | |
1987 | struct vm_area_struct *vma, | |
1988 | pmd_t *pmd, pmd_t entry, | |
1989 | unsigned long address, | |
1990 | struct page *page, int node) | |
1991 | { | |
1992 | spinlock_t *ptl; | |
1993 | pg_data_t *pgdat = NODE_DATA(node); | |
1994 | int isolated = 0; | |
1995 | struct page *new_page = NULL; | |
1996 | int page_lru = page_is_file_cache(page); | |
1997 | unsigned long mmun_start = address & HPAGE_PMD_MASK; | |
1998 | unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE; | |
1999 | ||
2000 | /* | |
2001 | * Rate-limit the amount of data that is being migrated to a node. | |
2002 | * Optimal placement is no good if the memory bus is saturated and | |
2003 | * all the time is being spent migrating! | |
2004 | */ | |
2005 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) | |
2006 | goto out_dropref; | |
2007 | ||
2008 | new_page = alloc_pages_node(node, | |
2009 | (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), | |
2010 | HPAGE_PMD_ORDER); | |
2011 | if (!new_page) | |
2012 | goto out_fail; | |
2013 | prep_transhuge_page(new_page); | |
2014 | ||
2015 | isolated = numamigrate_isolate_page(pgdat, page); | |
2016 | if (!isolated) { | |
2017 | put_page(new_page); | |
2018 | goto out_fail; | |
2019 | } | |
2020 | ||
2021 | /* Prepare a page as a migration target */ | |
2022 | __SetPageLocked(new_page); | |
2023 | if (PageSwapBacked(page)) | |
2024 | __SetPageSwapBacked(new_page); | |
2025 | ||
2026 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
2027 | new_page->mapping = page->mapping; | |
2028 | new_page->index = page->index; | |
2029 | migrate_page_copy(new_page, page); | |
2030 | WARN_ON(PageLRU(new_page)); | |
2031 | ||
2032 | /* Recheck the target PMD */ | |
2033 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
2034 | ptl = pmd_lock(mm, pmd); | |
2035 | if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) { | |
2036 | spin_unlock(ptl); | |
2037 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
2038 | ||
2039 | /* Reverse changes made by migrate_page_copy() */ | |
2040 | if (TestClearPageActive(new_page)) | |
2041 | SetPageActive(page); | |
2042 | if (TestClearPageUnevictable(new_page)) | |
2043 | SetPageUnevictable(page); | |
2044 | ||
2045 | unlock_page(new_page); | |
2046 | put_page(new_page); /* Free it */ | |
2047 | ||
2048 | /* Retake the callers reference and putback on LRU */ | |
2049 | get_page(page); | |
2050 | putback_lru_page(page); | |
2051 | mod_node_page_state(page_pgdat(page), | |
2052 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
2053 | ||
2054 | goto out_unlock; | |
2055 | } | |
2056 | ||
2057 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); | |
2058 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
2059 | ||
2060 | /* | |
2061 | * Clear the old entry under pagetable lock and establish the new PTE. | |
2062 | * Any parallel GUP will either observe the old page blocking on the | |
2063 | * page lock, block on the page table lock or observe the new page. | |
2064 | * The SetPageUptodate on the new page and page_add_new_anon_rmap | |
2065 | * guarantee the copy is visible before the pagetable update. | |
2066 | */ | |
2067 | flush_cache_range(vma, mmun_start, mmun_end); | |
2068 | page_add_anon_rmap(new_page, vma, mmun_start, true); | |
2069 | pmdp_huge_clear_flush_notify(vma, mmun_start, pmd); | |
2070 | set_pmd_at(mm, mmun_start, pmd, entry); | |
2071 | update_mmu_cache_pmd(vma, address, &entry); | |
2072 | ||
2073 | page_ref_unfreeze(page, 2); | |
2074 | mlock_migrate_page(new_page, page); | |
2075 | page_remove_rmap(page, true); | |
2076 | set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); | |
2077 | ||
2078 | spin_unlock(ptl); | |
2079 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
2080 | ||
2081 | /* Take an "isolate" reference and put new page on the LRU. */ | |
2082 | get_page(new_page); | |
2083 | putback_lru_page(new_page); | |
2084 | ||
2085 | unlock_page(new_page); | |
2086 | unlock_page(page); | |
2087 | put_page(page); /* Drop the rmap reference */ | |
2088 | put_page(page); /* Drop the LRU isolation reference */ | |
2089 | ||
2090 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
2091 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
2092 | ||
2093 | mod_node_page_state(page_pgdat(page), | |
2094 | NR_ISOLATED_ANON + page_lru, | |
2095 | -HPAGE_PMD_NR); | |
2096 | return isolated; | |
2097 | ||
2098 | out_fail: | |
2099 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
2100 | out_dropref: | |
2101 | ptl = pmd_lock(mm, pmd); | |
2102 | if (pmd_same(*pmd, entry)) { | |
2103 | entry = pmd_modify(entry, vma->vm_page_prot); | |
2104 | set_pmd_at(mm, mmun_start, pmd, entry); | |
2105 | update_mmu_cache_pmd(vma, address, &entry); | |
2106 | } | |
2107 | spin_unlock(ptl); | |
2108 | ||
2109 | out_unlock: | |
2110 | unlock_page(page); | |
2111 | put_page(page); | |
2112 | return 0; | |
2113 | } | |
2114 | #endif /* CONFIG_NUMA_BALANCING */ | |
2115 | ||
2116 | #endif /* CONFIG_NUMA */ | |
2117 | ||
2118 | ||
2119 | struct migrate_vma { | |
2120 | struct vm_area_struct *vma; | |
2121 | unsigned long *dst; | |
2122 | unsigned long *src; | |
2123 | unsigned long cpages; | |
2124 | unsigned long npages; | |
2125 | unsigned long start; | |
2126 | unsigned long end; | |
2127 | }; | |
2128 | ||
2129 | static int migrate_vma_collect_hole(unsigned long start, | |
2130 | unsigned long end, | |
2131 | struct mm_walk *walk) | |
2132 | { | |
2133 | struct migrate_vma *migrate = walk->private; | |
2134 | unsigned long addr; | |
2135 | ||
2136 | for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) { | |
2137 | migrate->dst[migrate->npages] = 0; | |
2138 | migrate->src[migrate->npages++] = 0; | |
2139 | } | |
2140 | ||
2141 | return 0; | |
2142 | } | |
2143 | ||
2144 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
2145 | unsigned long start, | |
2146 | unsigned long end, | |
2147 | struct mm_walk *walk) | |
2148 | { | |
2149 | struct migrate_vma *migrate = walk->private; | |
2150 | struct vm_area_struct *vma = walk->vma; | |
2151 | struct mm_struct *mm = vma->vm_mm; | |
2152 | unsigned long addr = start, unmapped = 0; | |
2153 | spinlock_t *ptl; | |
2154 | pte_t *ptep; | |
2155 | ||
2156 | again: | |
2157 | if (pmd_none(*pmdp)) | |
2158 | return migrate_vma_collect_hole(start, end, walk); | |
2159 | ||
2160 | if (pmd_trans_huge(*pmdp)) { | |
2161 | struct page *page; | |
2162 | ||
2163 | ptl = pmd_lock(mm, pmdp); | |
2164 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
2165 | spin_unlock(ptl); | |
2166 | goto again; | |
2167 | } | |
2168 | ||
2169 | page = pmd_page(*pmdp); | |
2170 | if (is_huge_zero_page(page)) { | |
2171 | spin_unlock(ptl); | |
2172 | split_huge_pmd(vma, pmdp, addr); | |
2173 | if (pmd_trans_unstable(pmdp)) | |
2174 | return migrate_vma_collect_hole(start, end, | |
2175 | walk); | |
2176 | } else { | |
2177 | int ret; | |
2178 | ||
2179 | get_page(page); | |
2180 | spin_unlock(ptl); | |
2181 | if (unlikely(!trylock_page(page))) | |
2182 | return migrate_vma_collect_hole(start, end, | |
2183 | walk); | |
2184 | ret = split_huge_page(page); | |
2185 | unlock_page(page); | |
2186 | put_page(page); | |
2187 | if (ret || pmd_none(*pmdp)) | |
2188 | return migrate_vma_collect_hole(start, end, | |
2189 | walk); | |
2190 | } | |
2191 | } | |
2192 | ||
2193 | if (unlikely(pmd_bad(*pmdp))) | |
2194 | return migrate_vma_collect_hole(start, end, walk); | |
2195 | ||
2196 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
2197 | arch_enter_lazy_mmu_mode(); | |
2198 | ||
2199 | for (; addr < end; addr += PAGE_SIZE, ptep++) { | |
2200 | unsigned long mpfn, pfn; | |
2201 | struct page *page; | |
2202 | swp_entry_t entry; | |
2203 | pte_t pte; | |
2204 | ||
2205 | pte = *ptep; | |
2206 | pfn = pte_pfn(pte); | |
2207 | ||
2208 | if (!pte_present(pte)) { | |
2209 | mpfn = pfn = 0; | |
2210 | goto next; | |
2211 | } | |
2212 | ||
2213 | /* FIXME support THP */ | |
2214 | page = vm_normal_page(migrate->vma, addr, pte); | |
2215 | if (!page || !page->mapping || PageTransCompound(page)) { | |
2216 | mpfn = pfn = 0; | |
2217 | goto next; | |
2218 | } | |
2219 | ||
2220 | /* | |
2221 | * By getting a reference on the page we pin it and that blocks | |
2222 | * any kind of migration. Side effect is that it "freezes" the | |
2223 | * pte. | |
2224 | * | |
2225 | * We drop this reference after isolating the page from the lru | |
2226 | * for non device page (device page are not on the lru and thus | |
2227 | * can't be dropped from it). | |
2228 | */ | |
2229 | get_page(page); | |
2230 | migrate->cpages++; | |
2231 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; | |
2232 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
2233 | ||
2234 | /* | |
2235 | * Optimize for the common case where page is only mapped once | |
2236 | * in one process. If we can lock the page, then we can safely | |
2237 | * set up a special migration page table entry now. | |
2238 | */ | |
2239 | if (trylock_page(page)) { | |
2240 | pte_t swp_pte; | |
2241 | ||
2242 | mpfn |= MIGRATE_PFN_LOCKED; | |
2243 | ptep_get_and_clear(mm, addr, ptep); | |
2244 | ||
2245 | /* Setup special migration page table entry */ | |
2246 | entry = make_migration_entry(page, pte_write(pte)); | |
2247 | swp_pte = swp_entry_to_pte(entry); | |
2248 | if (pte_soft_dirty(pte)) | |
2249 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2250 | set_pte_at(mm, addr, ptep, swp_pte); | |
2251 | ||
2252 | /* | |
2253 | * This is like regular unmap: we remove the rmap and | |
2254 | * drop page refcount. Page won't be freed, as we took | |
2255 | * a reference just above. | |
2256 | */ | |
2257 | page_remove_rmap(page, false); | |
2258 | put_page(page); | |
2259 | unmapped++; | |
2260 | } | |
2261 | ||
2262 | next: | |
2263 | migrate->src[migrate->npages++] = mpfn; | |
2264 | } | |
2265 | arch_leave_lazy_mmu_mode(); | |
2266 | pte_unmap_unlock(ptep - 1, ptl); | |
2267 | ||
2268 | /* Only flush the TLB if we actually modified any entries */ | |
2269 | if (unmapped) | |
2270 | flush_tlb_range(walk->vma, start, end); | |
2271 | ||
2272 | return 0; | |
2273 | } | |
2274 | ||
2275 | /* | |
2276 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
2277 | * @migrate: migrate struct containing all migration information | |
2278 | * | |
2279 | * This will walk the CPU page table. For each virtual address backed by a | |
2280 | * valid page, it updates the src array and takes a reference on the page, in | |
2281 | * order to pin the page until we lock it and unmap it. | |
2282 | */ | |
2283 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
2284 | { | |
2285 | struct mm_walk mm_walk; | |
2286 | ||
2287 | mm_walk.pmd_entry = migrate_vma_collect_pmd; | |
2288 | mm_walk.pte_entry = NULL; | |
2289 | mm_walk.pte_hole = migrate_vma_collect_hole; | |
2290 | mm_walk.hugetlb_entry = NULL; | |
2291 | mm_walk.test_walk = NULL; | |
2292 | mm_walk.vma = migrate->vma; | |
2293 | mm_walk.mm = migrate->vma->vm_mm; | |
2294 | mm_walk.private = migrate; | |
2295 | ||
2296 | mmu_notifier_invalidate_range_start(mm_walk.mm, | |
2297 | migrate->start, | |
2298 | migrate->end); | |
2299 | walk_page_range(migrate->start, migrate->end, &mm_walk); | |
2300 | mmu_notifier_invalidate_range_end(mm_walk.mm, | |
2301 | migrate->start, | |
2302 | migrate->end); | |
2303 | ||
2304 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); | |
2305 | } | |
2306 | ||
2307 | /* | |
2308 | * migrate_vma_check_page() - check if page is pinned or not | |
2309 | * @page: struct page to check | |
2310 | * | |
2311 | * Pinned pages cannot be migrated. This is the same test as in | |
2312 | * migrate_page_move_mapping(), except that here we allow migration of a | |
2313 | * ZONE_DEVICE page. | |
2314 | */ | |
2315 | static bool migrate_vma_check_page(struct page *page) | |
2316 | { | |
2317 | /* | |
2318 | * One extra ref because caller holds an extra reference, either from | |
2319 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
2320 | * a device page. | |
2321 | */ | |
2322 | int extra = 1; | |
2323 | ||
2324 | /* | |
2325 | * FIXME support THP (transparent huge page), it is bit more complex to | |
2326 | * check them than regular pages, because they can be mapped with a pmd | |
2327 | * or with a pte (split pte mapping). | |
2328 | */ | |
2329 | if (PageCompound(page)) | |
2330 | return false; | |
2331 | ||
2332 | if ((page_count(page) - extra) > page_mapcount(page)) | |
2333 | return false; | |
2334 | ||
2335 | return true; | |
2336 | } | |
2337 | ||
2338 | /* | |
2339 | * migrate_vma_prepare() - lock pages and isolate them from the lru | |
2340 | * @migrate: migrate struct containing all migration information | |
2341 | * | |
2342 | * This locks pages that have been collected by migrate_vma_collect(). Once each | |
2343 | * page is locked it is isolated from the lru (for non-device pages). Finally, | |
2344 | * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be | |
2345 | * migrated by concurrent kernel threads. | |
2346 | */ | |
2347 | static void migrate_vma_prepare(struct migrate_vma *migrate) | |
2348 | { | |
2349 | const unsigned long npages = migrate->npages; | |
2350 | const unsigned long start = migrate->start; | |
2351 | unsigned long addr, i, restore = 0; | |
2352 | bool allow_drain = true; | |
2353 | ||
2354 | lru_add_drain(); | |
2355 | ||
2356 | for (i = 0; (i < npages) && migrate->cpages; i++) { | |
2357 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2358 | bool remap = true; | |
2359 | ||
2360 | if (!page) | |
2361 | continue; | |
2362 | ||
2363 | if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { | |
2364 | /* | |
2365 | * Because we are migrating several pages there can be | |
2366 | * a deadlock between 2 concurrent migration where each | |
2367 | * are waiting on each other page lock. | |
2368 | * | |
2369 | * Make migrate_vma() a best effort thing and backoff | |
2370 | * for any page we can not lock right away. | |
2371 | */ | |
2372 | if (!trylock_page(page)) { | |
2373 | migrate->src[i] = 0; | |
2374 | migrate->cpages--; | |
2375 | put_page(page); | |
2376 | continue; | |
2377 | } | |
2378 | remap = false; | |
2379 | migrate->src[i] |= MIGRATE_PFN_LOCKED; | |
2380 | } | |
2381 | ||
2382 | if (!PageLRU(page) && allow_drain) { | |
2383 | /* Drain CPU's pagevec */ | |
2384 | lru_add_drain_all(); | |
2385 | allow_drain = false; | |
2386 | } | |
2387 | ||
2388 | if (isolate_lru_page(page)) { | |
2389 | if (remap) { | |
2390 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2391 | migrate->cpages--; | |
2392 | restore++; | |
2393 | } else { | |
2394 | migrate->src[i] = 0; | |
2395 | unlock_page(page); | |
2396 | migrate->cpages--; | |
2397 | put_page(page); | |
2398 | } | |
2399 | continue; | |
2400 | } | |
2401 | ||
2402 | if (!migrate_vma_check_page(page)) { | |
2403 | if (remap) { | |
2404 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2405 | migrate->cpages--; | |
2406 | restore++; | |
2407 | ||
2408 | get_page(page); | |
2409 | putback_lru_page(page); | |
2410 | } else { | |
2411 | migrate->src[i] = 0; | |
2412 | unlock_page(page); | |
2413 | migrate->cpages--; | |
2414 | ||
2415 | putback_lru_page(page); | |
2416 | } | |
2417 | } | |
2418 | } | |
2419 | ||
2420 | for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { | |
2421 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2422 | ||
2423 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2424 | continue; | |
2425 | ||
2426 | remove_migration_pte(page, migrate->vma, addr, page); | |
2427 | ||
2428 | migrate->src[i] = 0; | |
2429 | unlock_page(page); | |
2430 | put_page(page); | |
2431 | restore--; | |
2432 | } | |
2433 | } | |
2434 | ||
2435 | /* | |
2436 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
2437 | * @migrate: migrate struct containing all migration information | |
2438 | * | |
2439 | * Replace page mapping (CPU page table pte) with a special migration pte entry | |
2440 | * and check again if it has been pinned. Pinned pages are restored because we | |
2441 | * cannot migrate them. | |
2442 | * | |
2443 | * This is the last step before we call the device driver callback to allocate | |
2444 | * destination memory and copy contents of original page over to new page. | |
2445 | */ | |
2446 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
2447 | { | |
2448 | int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
2449 | const unsigned long npages = migrate->npages; | |
2450 | const unsigned long start = migrate->start; | |
2451 | unsigned long addr, i, restore = 0; | |
2452 | ||
2453 | for (i = 0; i < npages; i++) { | |
2454 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2455 | ||
2456 | if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2457 | continue; | |
2458 | ||
2459 | if (page_mapped(page)) { | |
2460 | try_to_unmap(page, flags); | |
2461 | if (page_mapped(page)) | |
2462 | goto restore; | |
2463 | } | |
2464 | ||
2465 | if (migrate_vma_check_page(page)) | |
2466 | continue; | |
2467 | ||
2468 | restore: | |
2469 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2470 | migrate->cpages--; | |
2471 | restore++; | |
2472 | } | |
2473 | ||
2474 | for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { | |
2475 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2476 | ||
2477 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2478 | continue; | |
2479 | ||
2480 | remove_migration_ptes(page, page, false); | |
2481 | ||
2482 | migrate->src[i] = 0; | |
2483 | unlock_page(page); | |
2484 | restore--; | |
2485 | ||
2486 | putback_lru_page(page); | |
2487 | } | |
2488 | } | |
2489 | ||
2490 | /* | |
2491 | * migrate_vma_pages() - migrate meta-data from src page to dst page | |
2492 | * @migrate: migrate struct containing all migration information | |
2493 | * | |
2494 | * This migrates struct page meta-data from source struct page to destination | |
2495 | * struct page. This effectively finishes the migration from source page to the | |
2496 | * destination page. | |
2497 | */ | |
2498 | static void migrate_vma_pages(struct migrate_vma *migrate) | |
2499 | { | |
2500 | const unsigned long npages = migrate->npages; | |
2501 | const unsigned long start = migrate->start; | |
2502 | unsigned long addr, i; | |
2503 | ||
2504 | for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { | |
2505 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2506 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2507 | struct address_space *mapping; | |
2508 | int r; | |
2509 | ||
2510 | if (!page || !newpage) | |
2511 | continue; | |
2512 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2513 | continue; | |
2514 | ||
2515 | mapping = page_mapping(page); | |
2516 | ||
2517 | r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); | |
2518 | if (r != MIGRATEPAGE_SUCCESS) | |
2519 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2520 | } | |
2521 | } | |
2522 | ||
2523 | /* | |
2524 | * migrate_vma_finalize() - restore CPU page table entry | |
2525 | * @migrate: migrate struct containing all migration information | |
2526 | * | |
2527 | * This replaces the special migration pte entry with either a mapping to the | |
2528 | * new page if migration was successful for that page, or to the original page | |
2529 | * otherwise. | |
2530 | * | |
2531 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
2532 | * refcount, for device pages. | |
2533 | */ | |
2534 | static void migrate_vma_finalize(struct migrate_vma *migrate) | |
2535 | { | |
2536 | const unsigned long npages = migrate->npages; | |
2537 | unsigned long i; | |
2538 | ||
2539 | for (i = 0; i < npages; i++) { | |
2540 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
2541 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2542 | ||
2543 | if (!page) | |
2544 | continue; | |
2545 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { | |
2546 | if (newpage) { | |
2547 | unlock_page(newpage); | |
2548 | put_page(newpage); | |
2549 | } | |
2550 | newpage = page; | |
2551 | } | |
2552 | ||
2553 | remove_migration_ptes(page, newpage, false); | |
2554 | unlock_page(page); | |
2555 | migrate->cpages--; | |
2556 | ||
2557 | putback_lru_page(page); | |
2558 | ||
2559 | if (newpage != page) { | |
2560 | unlock_page(newpage); | |
2561 | putback_lru_page(newpage); | |
2562 | } | |
2563 | } | |
2564 | } | |
2565 | ||
2566 | /* | |
2567 | * migrate_vma() - migrate a range of memory inside vma | |
2568 | * | |
2569 | * @ops: migration callback for allocating destination memory and copying | |
2570 | * @vma: virtual memory area containing the range to be migrated | |
2571 | * @start: start address of the range to migrate (inclusive) | |
2572 | * @end: end address of the range to migrate (exclusive) | |
2573 | * @src: array of hmm_pfn_t containing source pfns | |
2574 | * @dst: array of hmm_pfn_t containing destination pfns | |
2575 | * @private: pointer passed back to each of the callback | |
2576 | * Returns: 0 on success, error code otherwise | |
2577 | * | |
2578 | * This function tries to migrate a range of memory virtual address range, using | |
2579 | * callbacks to allocate and copy memory from source to destination. First it | |
2580 | * collects all the pages backing each virtual address in the range, saving this | |
2581 | * inside the src array. Then it locks those pages and unmaps them. Once the pages | |
2582 | * are locked and unmapped, it checks whether each page is pinned or not. Pages | |
2583 | * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) | |
2584 | * in the corresponding src array entry. It then restores any pages that are | |
2585 | * pinned, by remapping and unlocking those pages. | |
2586 | * | |
2587 | * At this point it calls the alloc_and_copy() callback. For documentation on | |
2588 | * what is expected from that callback, see struct migrate_vma_ops comments in | |
2589 | * include/linux/migrate.h | |
2590 | * | |
2591 | * After the alloc_and_copy() callback, this function goes over each entry in | |
2592 | * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
2593 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
2594 | * then the function tries to migrate struct page information from the source | |
2595 | * struct page to the destination struct page. If it fails to migrate the struct | |
2596 | * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src | |
2597 | * array. | |
2598 | * | |
2599 | * At this point all successfully migrated pages have an entry in the src | |
2600 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
2601 | * array entry with MIGRATE_PFN_VALID flag set. | |
2602 | * | |
2603 | * It then calls the finalize_and_map() callback. See comments for "struct | |
2604 | * migrate_vma_ops", in include/linux/migrate.h for details about | |
2605 | * finalize_and_map() behavior. | |
2606 | * | |
2607 | * After the finalize_and_map() callback, for successfully migrated pages, this | |
2608 | * function updates the CPU page table to point to new pages, otherwise it | |
2609 | * restores the CPU page table to point to the original source pages. | |
2610 | * | |
2611 | * Function returns 0 after the above steps, even if no pages were migrated | |
2612 | * (The function only returns an error if any of the arguments are invalid.) | |
2613 | * | |
2614 | * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT | |
2615 | * unsigned long entries. | |
2616 | */ | |
2617 | int migrate_vma(const struct migrate_vma_ops *ops, | |
2618 | struct vm_area_struct *vma, | |
2619 | unsigned long start, | |
2620 | unsigned long end, | |
2621 | unsigned long *src, | |
2622 | unsigned long *dst, | |
2623 | void *private) | |
2624 | { | |
2625 | struct migrate_vma migrate; | |
2626 | ||
2627 | /* Sanity check the arguments */ | |
2628 | start &= PAGE_MASK; | |
2629 | end &= PAGE_MASK; | |
2630 | if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) | |
2631 | return -EINVAL; | |
2632 | if (start < vma->vm_start || start >= vma->vm_end) | |
2633 | return -EINVAL; | |
2634 | if (end <= vma->vm_start || end > vma->vm_end) | |
2635 | return -EINVAL; | |
2636 | if (!ops || !src || !dst || start >= end) | |
2637 | return -EINVAL; | |
2638 | ||
2639 | memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT)); | |
2640 | migrate.src = src; | |
2641 | migrate.dst = dst; | |
2642 | migrate.start = start; | |
2643 | migrate.npages = 0; | |
2644 | migrate.cpages = 0; | |
2645 | migrate.end = end; | |
2646 | migrate.vma = vma; | |
2647 | ||
2648 | /* Collect, and try to unmap source pages */ | |
2649 | migrate_vma_collect(&migrate); | |
2650 | if (!migrate.cpages) | |
2651 | return 0; | |
2652 | ||
2653 | /* Lock and isolate page */ | |
2654 | migrate_vma_prepare(&migrate); | |
2655 | if (!migrate.cpages) | |
2656 | return 0; | |
2657 | ||
2658 | /* Unmap pages */ | |
2659 | migrate_vma_unmap(&migrate); | |
2660 | if (!migrate.cpages) | |
2661 | return 0; | |
2662 | ||
2663 | /* | |
2664 | * At this point pages are locked and unmapped, and thus they have | |
2665 | * stable content and can safely be copied to destination memory that | |
2666 | * is allocated by the callback. | |
2667 | * | |
2668 | * Note that migration can fail in migrate_vma_struct_page() for each | |
2669 | * individual page. | |
2670 | */ | |
2671 | ops->alloc_and_copy(vma, src, dst, start, end, private); | |
2672 | ||
2673 | /* This does the real migration of struct page */ | |
2674 | migrate_vma_pages(&migrate); | |
2675 | ||
2676 | ops->finalize_and_map(vma, src, dst, start, end, private); | |
2677 | ||
2678 | /* Unlock and remap pages */ | |
2679 | migrate_vma_finalize(&migrate); | |
2680 | ||
2681 | return 0; | |
2682 | } | |
2683 | EXPORT_SYMBOL(migrate_vma); |