1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * Memory Migration functionality - linux/mm/migrate.c |
4 | * |
5 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter |
6 | * |
7 | * Page migration was first developed in the context of the memory hotplug |
8 | * project. The main authors of the migration code are: |
9 | * |
10 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> |
11 | * Hirokazu Takahashi <taka@valinux.co.jp> |
12 | * Dave Hansen <haveblue@us.ibm.com> |
13 | * Christoph Lameter |
14 | */ |
15 | |
16 | #include <linux/migrate.h> |
17 | #include <linux/export.h> |
18 | #include <linux/swap.h> |
19 | #include <linux/swapops.h> |
20 | #include <linux/pagemap.h> |
21 | #include <linux/buffer_head.h> |
22 | #include <linux/mm_inline.h> |
23 | #include <linux/nsproxy.h> |
24 | #include <linux/pagevec.h> |
25 | #include <linux/ksm.h> |
26 | #include <linux/rmap.h> |
27 | #include <linux/topology.h> |
28 | #include <linux/cpu.h> |
29 | #include <linux/cpuset.h> |
30 | #include <linux/writeback.h> |
31 | #include <linux/mempolicy.h> |
32 | #include <linux/vmalloc.h> |
33 | #include <linux/security.h> |
34 | #include <linux/backing-dev.h> |
35 | #include <linux/compaction.h> |
36 | #include <linux/syscalls.h> |
37 | #include <linux/compat.h> |
38 | #include <linux/hugetlb.h> |
39 | #include <linux/hugetlb_cgroup.h> |
40 | #include <linux/gfp.h> |
41 | #include <linux/pfn_t.h> |
42 | #include <linux/memremap.h> |
43 | #include <linux/userfaultfd_k.h> |
44 | #include <linux/balloon_compaction.h> |
45 | #include <linux/mmu_notifier.h> |
46 | #include <linux/page_idle.h> |
47 | #include <linux/page_owner.h> |
48 | #include <linux/sched/mm.h> |
49 | #include <linux/ptrace.h> |
50 | |
51 | #include <asm/tlbflush.h> |
52 | |
53 | #define CREATE_TRACE_POINTS |
54 | #include <trace/events/migrate.h> |
55 | |
56 | #include "internal.h" |
57 | |
58 | /* |
59 | * migrate_prep() needs to be called before we start compiling a list of pages |
60 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is |
61 | * undesirable, use migrate_prep_local() |
62 | */ |
63 | int migrate_prep(void) |
64 | { |
65 | /* |
66 | * Clear the LRU lists so pages can be isolated. |
67 | * Note that pages may be moved off the LRU after we have |
68 | * drained them. Those pages will fail to migrate like other |
69 | * pages that may be busy. |
70 | */ |
71 | lru_add_drain_all(); |
72 | |
73 | return 0; |
74 | } |
75 | |
76 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ |
77 | int migrate_prep_local(void) |
78 | { |
79 | lru_add_drain(); |
80 | |
81 | return 0; |
82 | } |
83 | |
84 | int isolate_movable_page(struct page *page, isolate_mode_t mode) |
85 | { |
86 | struct address_space *mapping; |
87 | |
88 | /* |
89 | * Avoid burning cycles with pages that are yet under __free_pages(), |
90 | * or just got freed under us. |
91 | * |
92 | * In case we 'win' a race for a movable page being freed under us and |
93 | * raise its refcount preventing __free_pages() from doing its job |
94 | * the put_page() at the end of this block will take care of |
95 | * release this page, thus avoiding a nasty leakage. |
96 | */ |
97 | if (unlikely(!get_page_unless_zero(page))) |
98 | goto out; |
99 | |
100 | /* |
101 | * Check PageMovable before holding a PG_lock because page's owner |
102 | * assumes anybody doesn't touch PG_lock of newly allocated page |
103 | * so unconditionally grabbing the lock ruins page's owner side. |
104 | */ |
105 | if (unlikely(!__PageMovable(page))) |
106 | goto out_putpage; |
107 | /* |
108 | * As movable pages are not isolated from LRU lists, concurrent |
109 | * compaction threads can race against page migration functions |
110 | * as well as race against the releasing a page. |
111 | * |
112 | * In order to avoid having an already isolated movable page |
113 | * being (wrongly) re-isolated while it is under migration, |
114 | * or to avoid attempting to isolate pages being released, |
115 | * lets be sure we have the page lock |
116 | * before proceeding with the movable page isolation steps. |
117 | */ |
118 | if (unlikely(!trylock_page(page))) |
119 | goto out_putpage; |
120 | |
121 | if (!PageMovable(page) || PageIsolated(page)) |
122 | goto out_no_isolated; |
123 | |
124 | mapping = page_mapping(page); |
125 | VM_BUG_ON_PAGE(!mapping, page); |
126 | |
127 | if (!mapping->a_ops->isolate_page(page, mode)) |
128 | goto out_no_isolated; |
129 | |
130 | /* Driver shouldn't use PG_isolated bit of page->flags */ |
131 | WARN_ON_ONCE(PageIsolated(page)); |
132 | __SetPageIsolated(page); |
133 | unlock_page(page); |
134 | |
135 | return 0; |
136 | |
137 | out_no_isolated: |
138 | unlock_page(page); |
139 | out_putpage: |
140 | put_page(page); |
141 | out: |
142 | return -EBUSY; |
143 | } |
144 | |
145 | /* It should be called on page which is PG_movable */ |
146 | void putback_movable_page(struct page *page) |
147 | { |
148 | struct address_space *mapping; |
149 | |
150 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
151 | VM_BUG_ON_PAGE(!PageMovable(page), page); |
152 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
153 | |
154 | mapping = page_mapping(page); |
155 | mapping->a_ops->putback_page(page); |
156 | __ClearPageIsolated(page); |
157 | } |
158 | |
159 | /* |
160 | * Put previously isolated pages back onto the appropriate lists |
161 | * from where they were once taken off for compaction/migration. |
162 | * |
163 | * This function shall be used whenever the isolated pageset has been |
164 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() |
165 | * and isolate_huge_page(). |
166 | */ |
167 | void putback_movable_pages(struct list_head *l) |
168 | { |
169 | struct page *page; |
170 | struct page *page2; |
171 | |
172 | list_for_each_entry_safe(page, page2, l, lru) { |
173 | if (unlikely(PageHuge(page))) { |
174 | putback_active_hugepage(page); |
175 | continue; |
176 | } |
177 | list_del(&page->lru); |
178 | /* |
179 | * We isolated non-lru movable page so here we can use |
180 | * __PageMovable because LRU page's mapping cannot have |
181 | * PAGE_MAPPING_MOVABLE. |
182 | */ |
183 | if (unlikely(__PageMovable(page))) { |
184 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
185 | lock_page(page); |
186 | if (PageMovable(page)) |
187 | putback_movable_page(page); |
188 | else |
189 | __ClearPageIsolated(page); |
190 | unlock_page(page); |
191 | put_page(page); |
192 | } else { |
193 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
194 | page_is_file_cache(page), -hpage_nr_pages(page)); |
195 | putback_lru_page(page); |
196 | } |
197 | } |
198 | } |
199 | |
200 | /* |
201 | * Restore a potential migration pte to a working pte entry |
202 | */ |
203 | static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, |
204 | unsigned long addr, void *old) |
205 | { |
206 | struct page_vma_mapped_walk pvmw = { |
207 | .page = old, |
208 | .vma = vma, |
209 | .address = addr, |
210 | .flags = PVMW_SYNC | PVMW_MIGRATION, |
211 | }; |
212 | struct page *new; |
213 | pte_t pte; |
214 | swp_entry_t entry; |
215 | |
216 | VM_BUG_ON_PAGE(PageTail(page), page); |
217 | while (page_vma_mapped_walk(&pvmw)) { |
218 | if (PageKsm(page)) |
219 | new = page; |
220 | else |
221 | new = page - pvmw.page->index + |
222 | linear_page_index(vma, pvmw.address); |
223 | |
224 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
225 | /* PMD-mapped THP migration entry */ |
226 | if (!pvmw.pte) { |
227 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); |
228 | remove_migration_pmd(&pvmw, new); |
229 | continue; |
230 | } |
231 | #endif |
232 | |
233 | get_page(new); |
234 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); |
235 | if (pte_swp_soft_dirty(*pvmw.pte)) |
236 | pte = pte_mksoft_dirty(pte); |
237 | |
238 | /* |
239 | * Recheck VMA as permissions can change since migration started |
240 | */ |
241 | entry = pte_to_swp_entry(*pvmw.pte); |
242 | if (is_write_migration_entry(entry)) |
243 | pte = maybe_mkwrite(pte, vma); |
244 | |
245 | if (unlikely(is_zone_device_page(new))) { |
246 | if (is_device_private_page(new)) { |
247 | entry = make_device_private_entry(new, pte_write(pte)); |
248 | pte = swp_entry_to_pte(entry); |
249 | } else if (is_device_public_page(new)) { |
250 | pte = pte_mkdevmap(pte); |
251 | flush_dcache_page(new); |
252 | } |
253 | } else |
254 | flush_dcache_page(new); |
255 | |
256 | #ifdef CONFIG_HUGETLB_PAGE |
257 | if (PageHuge(new)) { |
258 | pte = pte_mkhuge(pte); |
259 | pte = arch_make_huge_pte(pte, vma, new, 0); |
260 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
261 | if (PageAnon(new)) |
262 | hugepage_add_anon_rmap(new, vma, pvmw.address); |
263 | else |
264 | page_dup_rmap(new, true); |
265 | } else |
266 | #endif |
267 | { |
268 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
269 | |
270 | if (PageAnon(new)) |
271 | page_add_anon_rmap(new, vma, pvmw.address, false); |
272 | else |
273 | page_add_file_rmap(new, false); |
274 | } |
275 | if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) |
276 | mlock_vma_page(new); |
277 | |
278 | if (PageTransHuge(page) && PageMlocked(page)) |
279 | clear_page_mlock(page); |
280 | |
281 | /* No need to invalidate - it was non-present before */ |
282 | update_mmu_cache(vma, pvmw.address, pvmw.pte); |
283 | } |
284 | |
285 | return true; |
286 | } |
287 | |
288 | /* |
289 | * Get rid of all migration entries and replace them by |
290 | * references to the indicated page. |
291 | */ |
292 | void remove_migration_ptes(struct page *old, struct page *new, bool locked) |
293 | { |
294 | struct rmap_walk_control rwc = { |
295 | .rmap_one = remove_migration_pte, |
296 | .arg = old, |
297 | }; |
298 | |
299 | if (locked) |
300 | rmap_walk_locked(new, &rwc); |
301 | else |
302 | rmap_walk(new, &rwc); |
303 | } |
304 | |
305 | /* |
306 | * Something used the pte of a page under migration. We need to |
307 | * get to the page and wait until migration is finished. |
308 | * When we return from this function the fault will be retried. |
309 | */ |
310 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
311 | spinlock_t *ptl) |
312 | { |
313 | pte_t pte; |
314 | swp_entry_t entry; |
315 | struct page *page; |
316 | |
317 | spin_lock(ptl); |
318 | pte = *ptep; |
319 | if (!is_swap_pte(pte)) |
320 | goto out; |
321 | |
322 | entry = pte_to_swp_entry(pte); |
323 | if (!is_migration_entry(entry)) |
324 | goto out; |
325 | |
326 | page = migration_entry_to_page(entry); |
327 | |
328 | /* |
329 | * Once page cache replacement of page migration started, page_count |
330 | * is zero; but we must not call put_and_wait_on_page_locked() without |
331 | * a ref. Use get_page_unless_zero(), and just fault again if it fails. |
332 | */ |
333 | if (!get_page_unless_zero(page)) |
334 | goto out; |
335 | pte_unmap_unlock(ptep, ptl); |
336 | put_and_wait_on_page_locked(page); |
337 | return; |
338 | out: |
339 | pte_unmap_unlock(ptep, ptl); |
340 | } |
341 | |
342 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
343 | unsigned long address) |
344 | { |
345 | spinlock_t *ptl = pte_lockptr(mm, pmd); |
346 | pte_t *ptep = pte_offset_map(pmd, address); |
347 | __migration_entry_wait(mm, ptep, ptl); |
348 | } |
349 | |
350 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
351 | struct mm_struct *mm, pte_t *pte) |
352 | { |
353 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
354 | __migration_entry_wait(mm, pte, ptl); |
355 | } |
356 | |
357 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
358 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) |
359 | { |
360 | spinlock_t *ptl; |
361 | struct page *page; |
362 | |
363 | ptl = pmd_lock(mm, pmd); |
364 | if (!is_pmd_migration_entry(*pmd)) |
365 | goto unlock; |
366 | page = migration_entry_to_page(pmd_to_swp_entry(*pmd)); |
367 | if (!get_page_unless_zero(page)) |
368 | goto unlock; |
369 | spin_unlock(ptl); |
370 | put_and_wait_on_page_locked(page); |
371 | return; |
372 | unlock: |
373 | spin_unlock(ptl); |
374 | } |
375 | #endif |
376 | |
377 | static int expected_page_refs(struct address_space *mapping, struct page *page) |
378 | { |
379 | int expected_count = 1; |
380 | |
381 | /* |
382 | * Device public or private pages have an extra refcount as they are |
383 | * ZONE_DEVICE pages. |
384 | */ |
385 | expected_count += is_device_private_page(page); |
386 | expected_count += is_device_public_page(page); |
387 | if (mapping) |
388 | expected_count += hpage_nr_pages(page) + page_has_private(page); |
389 | |
390 | return expected_count; |
391 | } |
392 | |
393 | /* |
394 | * Replace the page in the mapping. |
395 | * |
396 | * The number of remaining references must be: |
397 | * 1 for anonymous pages without a mapping |
398 | * 2 for pages with a mapping |
399 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
400 | */ |
401 | int migrate_page_move_mapping(struct address_space *mapping, |
402 | struct page *newpage, struct page *page, enum migrate_mode mode, |
403 | int ) |
404 | { |
405 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
406 | struct zone *oldzone, *newzone; |
407 | int dirty; |
408 | int expected_count = expected_page_refs(mapping, page) + extra_count; |
409 | |
410 | if (!mapping) { |
411 | /* Anonymous page without mapping */ |
412 | if (page_count(page) != expected_count) |
413 | return -EAGAIN; |
414 | |
415 | /* No turning back from here */ |
416 | newpage->index = page->index; |
417 | newpage->mapping = page->mapping; |
418 | if (PageSwapBacked(page)) |
419 | __SetPageSwapBacked(newpage); |
420 | |
421 | return MIGRATEPAGE_SUCCESS; |
422 | } |
423 | |
424 | oldzone = page_zone(page); |
425 | newzone = page_zone(newpage); |
426 | |
427 | xas_lock_irq(&xas); |
428 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
429 | xas_unlock_irq(&xas); |
430 | return -EAGAIN; |
431 | } |
432 | |
433 | if (!page_ref_freeze(page, expected_count)) { |
434 | xas_unlock_irq(&xas); |
435 | return -EAGAIN; |
436 | } |
437 | |
438 | /* |
439 | * Now we know that no one else is looking at the page: |
440 | * no turning back from here. |
441 | */ |
442 | newpage->index = page->index; |
443 | newpage->mapping = page->mapping; |
444 | page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */ |
445 | if (PageSwapBacked(page)) { |
446 | __SetPageSwapBacked(newpage); |
447 | if (PageSwapCache(page)) { |
448 | SetPageSwapCache(newpage); |
449 | set_page_private(newpage, page_private(page)); |
450 | } |
451 | } else { |
452 | VM_BUG_ON_PAGE(PageSwapCache(page), page); |
453 | } |
454 | |
455 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
456 | dirty = PageDirty(page); |
457 | if (dirty) { |
458 | ClearPageDirty(page); |
459 | SetPageDirty(newpage); |
460 | } |
461 | |
462 | xas_store(&xas, newpage); |
463 | if (PageTransHuge(page)) { |
464 | int i; |
465 | |
466 | for (i = 1; i < HPAGE_PMD_NR; i++) { |
467 | xas_next(&xas); |
468 | xas_store(&xas, newpage + i); |
469 | } |
470 | } |
471 | |
472 | /* |
473 | * Drop cache reference from old page by unfreezing |
474 | * to one less reference. |
475 | * We know this isn't the last reference. |
476 | */ |
477 | page_ref_unfreeze(page, expected_count - hpage_nr_pages(page)); |
478 | |
479 | xas_unlock(&xas); |
480 | /* Leave irq disabled to prevent preemption while updating stats */ |
481 | |
482 | /* |
483 | * If moved to a different zone then also account |
484 | * the page for that zone. Other VM counters will be |
485 | * taken care of when we establish references to the |
486 | * new page and drop references to the old page. |
487 | * |
488 | * Note that anonymous pages are accounted for |
489 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
490 | * are mapped to swap space. |
491 | */ |
492 | if (newzone != oldzone) { |
493 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES); |
494 | __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES); |
495 | if (PageSwapBacked(page) && !PageSwapCache(page)) { |
496 | __dec_node_state(oldzone->zone_pgdat, NR_SHMEM); |
497 | __inc_node_state(newzone->zone_pgdat, NR_SHMEM); |
498 | } |
499 | if (dirty && mapping_cap_account_dirty(mapping)) { |
500 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY); |
501 | __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING); |
502 | __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY); |
503 | __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING); |
504 | } |
505 | } |
506 | local_irq_enable(); |
507 | |
508 | return MIGRATEPAGE_SUCCESS; |
509 | } |
510 | EXPORT_SYMBOL(migrate_page_move_mapping); |
511 | |
512 | /* |
513 | * The expected number of remaining references is the same as that |
514 | * of migrate_page_move_mapping(). |
515 | */ |
516 | int migrate_huge_page_move_mapping(struct address_space *mapping, |
517 | struct page *newpage, struct page *page) |
518 | { |
519 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
520 | int expected_count; |
521 | |
522 | xas_lock_irq(&xas); |
523 | expected_count = 2 + page_has_private(page); |
524 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
525 | xas_unlock_irq(&xas); |
526 | return -EAGAIN; |
527 | } |
528 | |
529 | if (!page_ref_freeze(page, expected_count)) { |
530 | xas_unlock_irq(&xas); |
531 | return -EAGAIN; |
532 | } |
533 | |
534 | newpage->index = page->index; |
535 | newpage->mapping = page->mapping; |
536 | |
537 | get_page(newpage); |
538 | |
539 | xas_store(&xas, newpage); |
540 | |
541 | page_ref_unfreeze(page, expected_count - 1); |
542 | |
543 | xas_unlock_irq(&xas); |
544 | |
545 | return MIGRATEPAGE_SUCCESS; |
546 | } |
547 | |
548 | /* |
549 | * Gigantic pages are so large that we do not guarantee that page++ pointer |
550 | * arithmetic will work across the entire page. We need something more |
551 | * specialized. |
552 | */ |
553 | static void __copy_gigantic_page(struct page *dst, struct page *src, |
554 | int nr_pages) |
555 | { |
556 | int i; |
557 | struct page *dst_base = dst; |
558 | struct page *src_base = src; |
559 | |
560 | for (i = 0; i < nr_pages; ) { |
561 | cond_resched(); |
562 | copy_highpage(dst, src); |
563 | |
564 | i++; |
565 | dst = mem_map_next(dst, dst_base, i); |
566 | src = mem_map_next(src, src_base, i); |
567 | } |
568 | } |
569 | |
570 | static void copy_huge_page(struct page *dst, struct page *src) |
571 | { |
572 | int i; |
573 | int nr_pages; |
574 | |
575 | if (PageHuge(src)) { |
576 | /* hugetlbfs page */ |
577 | struct hstate *h = page_hstate(src); |
578 | nr_pages = pages_per_huge_page(h); |
579 | |
580 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { |
581 | __copy_gigantic_page(dst, src, nr_pages); |
582 | return; |
583 | } |
584 | } else { |
585 | /* thp page */ |
586 | BUG_ON(!PageTransHuge(src)); |
587 | nr_pages = hpage_nr_pages(src); |
588 | } |
589 | |
590 | for (i = 0; i < nr_pages; i++) { |
591 | cond_resched(); |
592 | copy_highpage(dst + i, src + i); |
593 | } |
594 | } |
595 | |
596 | /* |
597 | * Copy the page to its new location |
598 | */ |
599 | void migrate_page_states(struct page *newpage, struct page *page) |
600 | { |
601 | int cpupid; |
602 | |
603 | if (PageError(page)) |
604 | SetPageError(newpage); |
605 | if (PageReferenced(page)) |
606 | SetPageReferenced(newpage); |
607 | if (PageUptodate(page)) |
608 | SetPageUptodate(newpage); |
609 | if (TestClearPageActive(page)) { |
610 | VM_BUG_ON_PAGE(PageUnevictable(page), page); |
611 | SetPageActive(newpage); |
612 | } else if (TestClearPageUnevictable(page)) |
613 | SetPageUnevictable(newpage); |
614 | if (PageWorkingset(page)) |
615 | SetPageWorkingset(newpage); |
616 | if (PageChecked(page)) |
617 | SetPageChecked(newpage); |
618 | if (PageMappedToDisk(page)) |
619 | SetPageMappedToDisk(newpage); |
620 | |
621 | /* Move dirty on pages not done by migrate_page_move_mapping() */ |
622 | if (PageDirty(page)) |
623 | SetPageDirty(newpage); |
624 | |
625 | if (page_is_young(page)) |
626 | set_page_young(newpage); |
627 | if (page_is_idle(page)) |
628 | set_page_idle(newpage); |
629 | |
630 | /* |
631 | * Copy NUMA information to the new page, to prevent over-eager |
632 | * future migrations of this same page. |
633 | */ |
634 | cpupid = page_cpupid_xchg_last(page, -1); |
635 | page_cpupid_xchg_last(newpage, cpupid); |
636 | |
637 | ksm_migrate_page(newpage, page); |
638 | /* |
639 | * Please do not reorder this without considering how mm/ksm.c's |
640 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). |
641 | */ |
642 | if (PageSwapCache(page)) |
643 | ClearPageSwapCache(page); |
644 | ClearPagePrivate(page); |
645 | set_page_private(page, 0); |
646 | |
647 | /* |
648 | * If any waiters have accumulated on the new page then |
649 | * wake them up. |
650 | */ |
651 | if (PageWriteback(newpage)) |
652 | end_page_writeback(newpage); |
653 | |
654 | copy_page_owner(page, newpage); |
655 | |
656 | mem_cgroup_migrate(page, newpage); |
657 | } |
658 | EXPORT_SYMBOL(migrate_page_states); |
659 | |
660 | void migrate_page_copy(struct page *newpage, struct page *page) |
661 | { |
662 | if (PageHuge(page) || PageTransHuge(page)) |
663 | copy_huge_page(newpage, page); |
664 | else |
665 | copy_highpage(newpage, page); |
666 | |
667 | migrate_page_states(newpage, page); |
668 | } |
669 | EXPORT_SYMBOL(migrate_page_copy); |
670 | |
671 | /************************************************************ |
672 | * Migration functions |
673 | ***********************************************************/ |
674 | |
675 | /* |
676 | * Common logic to directly migrate a single LRU page suitable for |
677 | * pages that do not use PagePrivate/PagePrivate2. |
678 | * |
679 | * Pages are locked upon entry and exit. |
680 | */ |
681 | int migrate_page(struct address_space *mapping, |
682 | struct page *newpage, struct page *page, |
683 | enum migrate_mode mode) |
684 | { |
685 | int rc; |
686 | |
687 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ |
688 | |
689 | rc = migrate_page_move_mapping(mapping, newpage, page, mode, 0); |
690 | |
691 | if (rc != MIGRATEPAGE_SUCCESS) |
692 | return rc; |
693 | |
694 | if (mode != MIGRATE_SYNC_NO_COPY) |
695 | migrate_page_copy(newpage, page); |
696 | else |
697 | migrate_page_states(newpage, page); |
698 | return MIGRATEPAGE_SUCCESS; |
699 | } |
700 | EXPORT_SYMBOL(migrate_page); |
701 | |
702 | #ifdef CONFIG_BLOCK |
703 | /* Returns true if all buffers are successfully locked */ |
704 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
705 | enum migrate_mode mode) |
706 | { |
707 | struct buffer_head *bh = head; |
708 | |
709 | /* Simple case, sync compaction */ |
710 | if (mode != MIGRATE_ASYNC) { |
711 | do { |
712 | lock_buffer(bh); |
713 | bh = bh->b_this_page; |
714 | |
715 | } while (bh != head); |
716 | |
717 | return true; |
718 | } |
719 | |
720 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ |
721 | do { |
722 | if (!trylock_buffer(bh)) { |
723 | /* |
724 | * We failed to lock the buffer and cannot stall in |
725 | * async migration. Release the taken locks |
726 | */ |
727 | struct buffer_head *failed_bh = bh; |
728 | bh = head; |
729 | while (bh != failed_bh) { |
730 | unlock_buffer(bh); |
731 | bh = bh->b_this_page; |
732 | } |
733 | return false; |
734 | } |
735 | |
736 | bh = bh->b_this_page; |
737 | } while (bh != head); |
738 | return true; |
739 | } |
740 | |
741 | static int __buffer_migrate_page(struct address_space *mapping, |
742 | struct page *newpage, struct page *page, enum migrate_mode mode, |
743 | bool check_refs) |
744 | { |
745 | struct buffer_head *bh, *head; |
746 | int rc; |
747 | int expected_count; |
748 | |
749 | if (!page_has_buffers(page)) |
750 | return migrate_page(mapping, newpage, page, mode); |
751 | |
752 | /* Check whether page does not have extra refs before we do more work */ |
753 | expected_count = expected_page_refs(mapping, page); |
754 | if (page_count(page) != expected_count) |
755 | return -EAGAIN; |
756 | |
757 | head = page_buffers(page); |
758 | if (!buffer_migrate_lock_buffers(head, mode)) |
759 | return -EAGAIN; |
760 | |
761 | if (check_refs) { |
762 | bool busy; |
763 | bool invalidated = false; |
764 | |
765 | recheck_buffers: |
766 | busy = false; |
767 | spin_lock(&mapping->private_lock); |
768 | bh = head; |
769 | do { |
770 | if (atomic_read(&bh->b_count)) { |
771 | busy = true; |
772 | break; |
773 | } |
774 | bh = bh->b_this_page; |
775 | } while (bh != head); |
776 | spin_unlock(&mapping->private_lock); |
777 | if (busy) { |
778 | if (invalidated) { |
779 | rc = -EAGAIN; |
780 | goto unlock_buffers; |
781 | } |
782 | invalidate_bh_lrus(); |
783 | invalidated = true; |
784 | goto recheck_buffers; |
785 | } |
786 | } |
787 | |
788 | rc = migrate_page_move_mapping(mapping, newpage, page, mode, 0); |
789 | if (rc != MIGRATEPAGE_SUCCESS) |
790 | goto unlock_buffers; |
791 | |
792 | ClearPagePrivate(page); |
793 | set_page_private(newpage, page_private(page)); |
794 | set_page_private(page, 0); |
795 | put_page(page); |
796 | get_page(newpage); |
797 | |
798 | bh = head; |
799 | do { |
800 | set_bh_page(bh, newpage, bh_offset(bh)); |
801 | bh = bh->b_this_page; |
802 | |
803 | } while (bh != head); |
804 | |
805 | SetPagePrivate(newpage); |
806 | |
807 | if (mode != MIGRATE_SYNC_NO_COPY) |
808 | migrate_page_copy(newpage, page); |
809 | else |
810 | migrate_page_states(newpage, page); |
811 | |
812 | rc = MIGRATEPAGE_SUCCESS; |
813 | unlock_buffers: |
814 | bh = head; |
815 | do { |
816 | unlock_buffer(bh); |
817 | bh = bh->b_this_page; |
818 | |
819 | } while (bh != head); |
820 | |
821 | return rc; |
822 | } |
823 | |
824 | /* |
825 | * Migration function for pages with buffers. This function can only be used |
826 | * if the underlying filesystem guarantees that no other references to "page" |
827 | * exist. For example attached buffer heads are accessed only under page lock. |
828 | */ |
829 | int buffer_migrate_page(struct address_space *mapping, |
830 | struct page *newpage, struct page *page, enum migrate_mode mode) |
831 | { |
832 | return __buffer_migrate_page(mapping, newpage, page, mode, false); |
833 | } |
834 | EXPORT_SYMBOL(buffer_migrate_page); |
835 | |
836 | /* |
837 | * Same as above except that this variant is more careful and checks that there |
838 | * are also no buffer head references. This function is the right one for |
839 | * mappings where buffer heads are directly looked up and referenced (such as |
840 | * block device mappings). |
841 | */ |
842 | int buffer_migrate_page_norefs(struct address_space *mapping, |
843 | struct page *newpage, struct page *page, enum migrate_mode mode) |
844 | { |
845 | return __buffer_migrate_page(mapping, newpage, page, mode, true); |
846 | } |
847 | #endif |
848 | |
849 | /* |
850 | * Writeback a page to clean the dirty state |
851 | */ |
852 | static int writeout(struct address_space *mapping, struct page *page) |
853 | { |
854 | struct writeback_control wbc = { |
855 | .sync_mode = WB_SYNC_NONE, |
856 | .nr_to_write = 1, |
857 | .range_start = 0, |
858 | .range_end = LLONG_MAX, |
859 | .for_reclaim = 1 |
860 | }; |
861 | int rc; |
862 | |
863 | if (!mapping->a_ops->writepage) |
864 | /* No write method for the address space */ |
865 | return -EINVAL; |
866 | |
867 | if (!clear_page_dirty_for_io(page)) |
868 | /* Someone else already triggered a write */ |
869 | return -EAGAIN; |
870 | |
871 | /* |
872 | * A dirty page may imply that the underlying filesystem has |
873 | * the page on some queue. So the page must be clean for |
874 | * migration. Writeout may mean we loose the lock and the |
875 | * page state is no longer what we checked for earlier. |
876 | * At this point we know that the migration attempt cannot |
877 | * be successful. |
878 | */ |
879 | remove_migration_ptes(page, page, false); |
880 | |
881 | rc = mapping->a_ops->writepage(page, &wbc); |
882 | |
883 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
884 | /* unlocked. Relock */ |
885 | lock_page(page); |
886 | |
887 | return (rc < 0) ? -EIO : -EAGAIN; |
888 | } |
889 | |
890 | /* |
891 | * Default handling if a filesystem does not provide a migration function. |
892 | */ |
893 | static int fallback_migrate_page(struct address_space *mapping, |
894 | struct page *newpage, struct page *page, enum migrate_mode mode) |
895 | { |
896 | if (PageDirty(page)) { |
897 | /* Only writeback pages in full synchronous migration */ |
898 | switch (mode) { |
899 | case MIGRATE_SYNC: |
900 | case MIGRATE_SYNC_NO_COPY: |
901 | break; |
902 | default: |
903 | return -EBUSY; |
904 | } |
905 | return writeout(mapping, page); |
906 | } |
907 | |
908 | /* |
909 | * Buffers may be managed in a filesystem specific way. |
910 | * We must have no buffers or drop them. |
911 | */ |
912 | if (page_has_private(page) && |
913 | !try_to_release_page(page, GFP_KERNEL)) |
914 | return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; |
915 | |
916 | return migrate_page(mapping, newpage, page, mode); |
917 | } |
918 | |
919 | /* |
920 | * Move a page to a newly allocated page |
921 | * The page is locked and all ptes have been successfully removed. |
922 | * |
923 | * The new page will have replaced the old page if this function |
924 | * is successful. |
925 | * |
926 | * Return value: |
927 | * < 0 - error code |
928 | * MIGRATEPAGE_SUCCESS - success |
929 | */ |
930 | static int move_to_new_page(struct page *newpage, struct page *page, |
931 | enum migrate_mode mode) |
932 | { |
933 | struct address_space *mapping; |
934 | int rc = -EAGAIN; |
935 | bool is_lru = !__PageMovable(page); |
936 | |
937 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
938 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); |
939 | |
940 | mapping = page_mapping(page); |
941 | |
942 | if (likely(is_lru)) { |
943 | if (!mapping) |
944 | rc = migrate_page(mapping, newpage, page, mode); |
945 | else if (mapping->a_ops->migratepage) |
946 | /* |
947 | * Most pages have a mapping and most filesystems |
948 | * provide a migratepage callback. Anonymous pages |
949 | * are part of swap space which also has its own |
950 | * migratepage callback. This is the most common path |
951 | * for page migration. |
952 | */ |
953 | rc = mapping->a_ops->migratepage(mapping, newpage, |
954 | page, mode); |
955 | else |
956 | rc = fallback_migrate_page(mapping, newpage, |
957 | page, mode); |
958 | } else { |
959 | /* |
960 | * In case of non-lru page, it could be released after |
961 | * isolation step. In that case, we shouldn't try migration. |
962 | */ |
963 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
964 | if (!PageMovable(page)) { |
965 | rc = MIGRATEPAGE_SUCCESS; |
966 | __ClearPageIsolated(page); |
967 | goto out; |
968 | } |
969 | |
970 | rc = mapping->a_ops->migratepage(mapping, newpage, |
971 | page, mode); |
972 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && |
973 | !PageIsolated(page)); |
974 | } |
975 | |
976 | /* |
977 | * When successful, old pagecache page->mapping must be cleared before |
978 | * page is freed; but stats require that PageAnon be left as PageAnon. |
979 | */ |
980 | if (rc == MIGRATEPAGE_SUCCESS) { |
981 | if (__PageMovable(page)) { |
982 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
983 | |
984 | /* |
985 | * We clear PG_movable under page_lock so any compactor |
986 | * cannot try to migrate this page. |
987 | */ |
988 | __ClearPageIsolated(page); |
989 | } |
990 | |
991 | /* |
992 | * Anonymous and movable page->mapping will be cleard by |
993 | * free_pages_prepare so don't reset it here for keeping |
994 | * the type to work PageAnon, for example. |
995 | */ |
996 | if (!PageMappingFlags(page)) |
997 | page->mapping = NULL; |
998 | } |
999 | out: |
1000 | return rc; |
1001 | } |
1002 | |
1003 | static int __unmap_and_move(struct page *page, struct page *newpage, |
1004 | int force, enum migrate_mode mode) |
1005 | { |
1006 | int rc = -EAGAIN; |
1007 | int page_was_mapped = 0; |
1008 | struct anon_vma *anon_vma = NULL; |
1009 | bool is_lru = !__PageMovable(page); |
1010 | |
1011 | if (!trylock_page(page)) { |
1012 | if (!force || mode == MIGRATE_ASYNC) |
1013 | goto out; |
1014 | |
1015 | /* |
1016 | * It's not safe for direct compaction to call lock_page. |
1017 | * For example, during page readahead pages are added locked |
1018 | * to the LRU. Later, when the IO completes the pages are |
1019 | * marked uptodate and unlocked. However, the queueing |
1020 | * could be merging multiple pages for one bio (e.g. |
1021 | * mpage_readpages). If an allocation happens for the |
1022 | * second or third page, the process can end up locking |
1023 | * the same page twice and deadlocking. Rather than |
1024 | * trying to be clever about what pages can be locked, |
1025 | * avoid the use of lock_page for direct compaction |
1026 | * altogether. |
1027 | */ |
1028 | if (current->flags & PF_MEMALLOC) |
1029 | goto out; |
1030 | |
1031 | lock_page(page); |
1032 | } |
1033 | |
1034 | if (PageWriteback(page)) { |
1035 | /* |
1036 | * Only in the case of a full synchronous migration is it |
1037 | * necessary to wait for PageWriteback. In the async case, |
1038 | * the retry loop is too short and in the sync-light case, |
1039 | * the overhead of stalling is too much |
1040 | */ |
1041 | switch (mode) { |
1042 | case MIGRATE_SYNC: |
1043 | case MIGRATE_SYNC_NO_COPY: |
1044 | break; |
1045 | default: |
1046 | rc = -EBUSY; |
1047 | goto out_unlock; |
1048 | } |
1049 | if (!force) |
1050 | goto out_unlock; |
1051 | wait_on_page_writeback(page); |
1052 | } |
1053 | |
1054 | /* |
1055 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
1056 | * we cannot notice that anon_vma is freed while we migrates a page. |
1057 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
1058 | * of migration. File cache pages are no problem because of page_lock() |
1059 | * File Caches may use write_page() or lock_page() in migration, then, |
1060 | * just care Anon page here. |
1061 | * |
1062 | * Only page_get_anon_vma() understands the subtleties of |
1063 | * getting a hold on an anon_vma from outside one of its mms. |
1064 | * But if we cannot get anon_vma, then we won't need it anyway, |
1065 | * because that implies that the anon page is no longer mapped |
1066 | * (and cannot be remapped so long as we hold the page lock). |
1067 | */ |
1068 | if (PageAnon(page) && !PageKsm(page)) |
1069 | anon_vma = page_get_anon_vma(page); |
1070 | |
1071 | /* |
1072 | * Block others from accessing the new page when we get around to |
1073 | * establishing additional references. We are usually the only one |
1074 | * holding a reference to newpage at this point. We used to have a BUG |
1075 | * here if trylock_page(newpage) fails, but would like to allow for |
1076 | * cases where there might be a race with the previous use of newpage. |
1077 | * This is much like races on refcount of oldpage: just don't BUG(). |
1078 | */ |
1079 | if (unlikely(!trylock_page(newpage))) |
1080 | goto out_unlock; |
1081 | |
1082 | if (unlikely(!is_lru)) { |
1083 | rc = move_to_new_page(newpage, page, mode); |
1084 | goto out_unlock_both; |
1085 | } |
1086 | |
1087 | /* |
1088 | * Corner case handling: |
1089 | * 1. When a new swap-cache page is read into, it is added to the LRU |
1090 | * and treated as swapcache but it has no rmap yet. |
1091 | * Calling try_to_unmap() against a page->mapping==NULL page will |
1092 | * trigger a BUG. So handle it here. |
1093 | * 2. An orphaned page (see truncate_complete_page) might have |
1094 | * fs-private metadata. The page can be picked up due to memory |
1095 | * offlining. Everywhere else except page reclaim, the page is |
1096 | * invisible to the vm, so the page can not be migrated. So try to |
1097 | * free the metadata, so the page can be freed. |
1098 | */ |
1099 | if (!page->mapping) { |
1100 | VM_BUG_ON_PAGE(PageAnon(page), page); |
1101 | if (page_has_private(page)) { |
1102 | try_to_free_buffers(page); |
1103 | goto out_unlock_both; |
1104 | } |
1105 | } else if (page_mapped(page)) { |
1106 | /* Establish migration ptes */ |
1107 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, |
1108 | page); |
1109 | try_to_unmap(page, |
1110 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
1111 | page_was_mapped = 1; |
1112 | } |
1113 | |
1114 | if (!page_mapped(page)) |
1115 | rc = move_to_new_page(newpage, page, mode); |
1116 | |
1117 | if (page_was_mapped) |
1118 | remove_migration_ptes(page, |
1119 | rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); |
1120 | |
1121 | out_unlock_both: |
1122 | unlock_page(newpage); |
1123 | out_unlock: |
1124 | /* Drop an anon_vma reference if we took one */ |
1125 | if (anon_vma) |
1126 | put_anon_vma(anon_vma); |
1127 | unlock_page(page); |
1128 | out: |
1129 | /* |
1130 | * If migration is successful, decrease refcount of the newpage |
1131 | * which will not free the page because new page owner increased |
1132 | * refcounter. As well, if it is LRU page, add the page to LRU |
1133 | * list in here. Use the old state of the isolated source page to |
1134 | * determine if we migrated a LRU page. newpage was already unlocked |
1135 | * and possibly modified by its owner - don't rely on the page |
1136 | * state. |
1137 | */ |
1138 | if (rc == MIGRATEPAGE_SUCCESS) { |
1139 | if (unlikely(!is_lru)) |
1140 | put_page(newpage); |
1141 | else |
1142 | putback_lru_page(newpage); |
1143 | } |
1144 | |
1145 | return rc; |
1146 | } |
1147 | |
1148 | /* |
1149 | * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work |
1150 | * around it. |
1151 | */ |
1152 | #if defined(CONFIG_ARM) && \ |
1153 | defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700 |
1154 | #define ICE_noinline noinline |
1155 | #else |
1156 | #define ICE_noinline |
1157 | #endif |
1158 | |
1159 | /* |
1160 | * Obtain the lock on page, remove all ptes and migrate the page |
1161 | * to the newly allocated page in newpage. |
1162 | */ |
1163 | static ICE_noinline int unmap_and_move(new_page_t get_new_page, |
1164 | free_page_t put_new_page, |
1165 | unsigned long private, struct page *page, |
1166 | int force, enum migrate_mode mode, |
1167 | enum migrate_reason reason) |
1168 | { |
1169 | int rc = MIGRATEPAGE_SUCCESS; |
1170 | struct page *newpage; |
1171 | |
1172 | if (!thp_migration_supported() && PageTransHuge(page)) |
1173 | return -ENOMEM; |
1174 | |
1175 | newpage = get_new_page(page, private); |
1176 | if (!newpage) |
1177 | return -ENOMEM; |
1178 | |
1179 | if (page_count(page) == 1) { |
1180 | /* page was freed from under us. So we are done. */ |
1181 | ClearPageActive(page); |
1182 | ClearPageUnevictable(page); |
1183 | if (unlikely(__PageMovable(page))) { |
1184 | lock_page(page); |
1185 | if (!PageMovable(page)) |
1186 | __ClearPageIsolated(page); |
1187 | unlock_page(page); |
1188 | } |
1189 | if (put_new_page) |
1190 | put_new_page(newpage, private); |
1191 | else |
1192 | put_page(newpage); |
1193 | goto out; |
1194 | } |
1195 | |
1196 | rc = __unmap_and_move(page, newpage, force, mode); |
1197 | if (rc == MIGRATEPAGE_SUCCESS) |
1198 | set_page_owner_migrate_reason(newpage, reason); |
1199 | |
1200 | out: |
1201 | if (rc != -EAGAIN) { |
1202 | /* |
1203 | * A page that has been migrated has all references |
1204 | * removed and will be freed. A page that has not been |
1205 | * migrated will have kepts its references and be |
1206 | * restored. |
1207 | */ |
1208 | list_del(&page->lru); |
1209 | |
1210 | /* |
1211 | * Compaction can migrate also non-LRU pages which are |
1212 | * not accounted to NR_ISOLATED_*. They can be recognized |
1213 | * as __PageMovable |
1214 | */ |
1215 | if (likely(!__PageMovable(page))) |
1216 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
1217 | page_is_file_cache(page), -hpage_nr_pages(page)); |
1218 | } |
1219 | |
1220 | /* |
1221 | * If migration is successful, releases reference grabbed during |
1222 | * isolation. Otherwise, restore the page to right list unless |
1223 | * we want to retry. |
1224 | */ |
1225 | if (rc == MIGRATEPAGE_SUCCESS) { |
1226 | put_page(page); |
1227 | if (reason == MR_MEMORY_FAILURE) { |
1228 | /* |
1229 | * Set PG_HWPoison on just freed page |
1230 | * intentionally. Although it's rather weird, |
1231 | * it's how HWPoison flag works at the moment. |
1232 | */ |
1233 | if (set_hwpoison_free_buddy_page(page)) |
1234 | num_poisoned_pages_inc(); |
1235 | } |
1236 | } else { |
1237 | if (rc != -EAGAIN) { |
1238 | if (likely(!__PageMovable(page))) { |
1239 | putback_lru_page(page); |
1240 | goto put_new; |
1241 | } |
1242 | |
1243 | lock_page(page); |
1244 | if (PageMovable(page)) |
1245 | putback_movable_page(page); |
1246 | else |
1247 | __ClearPageIsolated(page); |
1248 | unlock_page(page); |
1249 | put_page(page); |
1250 | } |
1251 | put_new: |
1252 | if (put_new_page) |
1253 | put_new_page(newpage, private); |
1254 | else |
1255 | put_page(newpage); |
1256 | } |
1257 | |
1258 | return rc; |
1259 | } |
1260 | |
1261 | /* |
1262 | * Counterpart of unmap_and_move_page() for hugepage migration. |
1263 | * |
1264 | * This function doesn't wait the completion of hugepage I/O |
1265 | * because there is no race between I/O and migration for hugepage. |
1266 | * Note that currently hugepage I/O occurs only in direct I/O |
1267 | * where no lock is held and PG_writeback is irrelevant, |
1268 | * and writeback status of all subpages are counted in the reference |
1269 | * count of the head page (i.e. if all subpages of a 2MB hugepage are |
1270 | * under direct I/O, the reference of the head page is 512 and a bit more.) |
1271 | * This means that when we try to migrate hugepage whose subpages are |
1272 | * doing direct I/O, some references remain after try_to_unmap() and |
1273 | * hugepage migration fails without data corruption. |
1274 | * |
1275 | * There is also no race when direct I/O is issued on the page under migration, |
1276 | * because then pte is replaced with migration swap entry and direct I/O code |
1277 | * will wait in the page fault for migration to complete. |
1278 | */ |
1279 | static int unmap_and_move_huge_page(new_page_t get_new_page, |
1280 | free_page_t put_new_page, unsigned long private, |
1281 | struct page *hpage, int force, |
1282 | enum migrate_mode mode, int reason) |
1283 | { |
1284 | int rc = -EAGAIN; |
1285 | int page_was_mapped = 0; |
1286 | struct page *new_hpage; |
1287 | struct anon_vma *anon_vma = NULL; |
1288 | |
1289 | /* |
1290 | * Migratability of hugepages depends on architectures and their size. |
1291 | * This check is necessary because some callers of hugepage migration |
1292 | * like soft offline and memory hotremove don't walk through page |
1293 | * tables or check whether the hugepage is pmd-based or not before |
1294 | * kicking migration. |
1295 | */ |
1296 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
1297 | putback_active_hugepage(hpage); |
1298 | return -ENOSYS; |
1299 | } |
1300 | |
1301 | new_hpage = get_new_page(hpage, private); |
1302 | if (!new_hpage) |
1303 | return -ENOMEM; |
1304 | |
1305 | if (!trylock_page(hpage)) { |
1306 | if (!force) |
1307 | goto out; |
1308 | switch (mode) { |
1309 | case MIGRATE_SYNC: |
1310 | case MIGRATE_SYNC_NO_COPY: |
1311 | break; |
1312 | default: |
1313 | goto out; |
1314 | } |
1315 | lock_page(hpage); |
1316 | } |
1317 | |
1318 | /* |
1319 | * Check for pages which are in the process of being freed. Without |
1320 | * page_mapping() set, hugetlbfs specific move page routine will not |
1321 | * be called and we could leak usage counts for subpools. |
1322 | */ |
1323 | if (page_private(hpage) && !page_mapping(hpage)) { |
1324 | rc = -EBUSY; |
1325 | goto out_unlock; |
1326 | } |
1327 | |
1328 | if (PageAnon(hpage)) |
1329 | anon_vma = page_get_anon_vma(hpage); |
1330 | |
1331 | if (unlikely(!trylock_page(new_hpage))) |
1332 | goto put_anon; |
1333 | |
1334 | if (page_mapped(hpage)) { |
1335 | try_to_unmap(hpage, |
1336 | TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
1337 | page_was_mapped = 1; |
1338 | } |
1339 | |
1340 | if (!page_mapped(hpage)) |
1341 | rc = move_to_new_page(new_hpage, hpage, mode); |
1342 | |
1343 | if (page_was_mapped) |
1344 | remove_migration_ptes(hpage, |
1345 | rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); |
1346 | |
1347 | unlock_page(new_hpage); |
1348 | |
1349 | put_anon: |
1350 | if (anon_vma) |
1351 | put_anon_vma(anon_vma); |
1352 | |
1353 | if (rc == MIGRATEPAGE_SUCCESS) { |
1354 | move_hugetlb_state(hpage, new_hpage, reason); |
1355 | put_new_page = NULL; |
1356 | } |
1357 | |
1358 | out_unlock: |
1359 | unlock_page(hpage); |
1360 | out: |
1361 | if (rc != -EAGAIN) |
1362 | putback_active_hugepage(hpage); |
1363 | |
1364 | /* |
1365 | * If migration was not successful and there's a freeing callback, use |
1366 | * it. Otherwise, put_page() will drop the reference grabbed during |
1367 | * isolation. |
1368 | */ |
1369 | if (put_new_page) |
1370 | put_new_page(new_hpage, private); |
1371 | else |
1372 | putback_active_hugepage(new_hpage); |
1373 | |
1374 | return rc; |
1375 | } |
1376 | |
1377 | /* |
1378 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1379 | * supplied as the target for the page migration |
1380 | * |
1381 | * @from: The list of pages to be migrated. |
1382 | * @get_new_page: The function used to allocate free pages to be used |
1383 | * as the target of the page migration. |
1384 | * @put_new_page: The function used to free target pages if migration |
1385 | * fails, or NULL if no special handling is necessary. |
1386 | * @private: Private data to be passed on to get_new_page() |
1387 | * @mode: The migration mode that specifies the constraints for |
1388 | * page migration, if any. |
1389 | * @reason: The reason for page migration. |
1390 | * |
1391 | * The function returns after 10 attempts or if no pages are movable any more |
1392 | * because the list has become empty or no retryable pages exist any more. |
1393 | * The caller should call putback_movable_pages() to return pages to the LRU |
1394 | * or free list only if ret != 0. |
1395 | * |
1396 | * Returns the number of pages that were not migrated, or an error code. |
1397 | */ |
1398 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
1399 | free_page_t put_new_page, unsigned long private, |
1400 | enum migrate_mode mode, int reason) |
1401 | { |
1402 | int retry = 1; |
1403 | int nr_failed = 0; |
1404 | int nr_succeeded = 0; |
1405 | int pass = 0; |
1406 | struct page *page; |
1407 | struct page *page2; |
1408 | int swapwrite = current->flags & PF_SWAPWRITE; |
1409 | int rc; |
1410 | |
1411 | if (!swapwrite) |
1412 | current->flags |= PF_SWAPWRITE; |
1413 | |
1414 | for(pass = 0; pass < 10 && retry; pass++) { |
1415 | retry = 0; |
1416 | |
1417 | list_for_each_entry_safe(page, page2, from, lru) { |
1418 | retry: |
1419 | cond_resched(); |
1420 | |
1421 | if (PageHuge(page)) |
1422 | rc = unmap_and_move_huge_page(get_new_page, |
1423 | put_new_page, private, page, |
1424 | pass > 2, mode, reason); |
1425 | else |
1426 | rc = unmap_and_move(get_new_page, put_new_page, |
1427 | private, page, pass > 2, mode, |
1428 | reason); |
1429 | |
1430 | switch(rc) { |
1431 | case -ENOMEM: |
1432 | /* |
1433 | * THP migration might be unsupported or the |
1434 | * allocation could've failed so we should |
1435 | * retry on the same page with the THP split |
1436 | * to base pages. |
1437 | * |
1438 | * Head page is retried immediately and tail |
1439 | * pages are added to the tail of the list so |
1440 | * we encounter them after the rest of the list |
1441 | * is processed. |
1442 | */ |
1443 | if (PageTransHuge(page) && !PageHuge(page)) { |
1444 | lock_page(page); |
1445 | rc = split_huge_page_to_list(page, from); |
1446 | unlock_page(page); |
1447 | if (!rc) { |
1448 | list_safe_reset_next(page, page2, lru); |
1449 | goto retry; |
1450 | } |
1451 | } |
1452 | nr_failed++; |
1453 | goto out; |
1454 | case -EAGAIN: |
1455 | retry++; |
1456 | break; |
1457 | case MIGRATEPAGE_SUCCESS: |
1458 | nr_succeeded++; |
1459 | break; |
1460 | default: |
1461 | /* |
1462 | * Permanent failure (-EBUSY, -ENOSYS, etc.): |
1463 | * unlike -EAGAIN case, the failed page is |
1464 | * removed from migration page list and not |
1465 | * retried in the next outer loop. |
1466 | */ |
1467 | nr_failed++; |
1468 | break; |
1469 | } |
1470 | } |
1471 | } |
1472 | nr_failed += retry; |
1473 | rc = nr_failed; |
1474 | out: |
1475 | if (nr_succeeded) |
1476 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); |
1477 | if (nr_failed) |
1478 | count_vm_events(PGMIGRATE_FAIL, nr_failed); |
1479 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1480 | |
1481 | if (!swapwrite) |
1482 | current->flags &= ~PF_SWAPWRITE; |
1483 | |
1484 | return rc; |
1485 | } |
1486 | |
1487 | #ifdef CONFIG_NUMA |
1488 | |
1489 | static int store_status(int __user *status, int start, int value, int nr) |
1490 | { |
1491 | while (nr-- > 0) { |
1492 | if (put_user(value, status + start)) |
1493 | return -EFAULT; |
1494 | start++; |
1495 | } |
1496 | |
1497 | return 0; |
1498 | } |
1499 | |
1500 | static int do_move_pages_to_node(struct mm_struct *mm, |
1501 | struct list_head *pagelist, int node) |
1502 | { |
1503 | int err; |
1504 | |
1505 | if (list_empty(pagelist)) |
1506 | return 0; |
1507 | |
1508 | err = migrate_pages(pagelist, alloc_new_node_page, NULL, node, |
1509 | MIGRATE_SYNC, MR_SYSCALL); |
1510 | if (err) |
1511 | putback_movable_pages(pagelist); |
1512 | return err; |
1513 | } |
1514 | |
1515 | /* |
1516 | * Resolves the given address to a struct page, isolates it from the LRU and |
1517 | * puts it to the given pagelist. |
1518 | * Returns -errno if the page cannot be found/isolated or 0 when it has been |
1519 | * queued or the page doesn't need to be migrated because it is already on |
1520 | * the target node |
1521 | */ |
1522 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, |
1523 | int node, struct list_head *pagelist, bool migrate_all) |
1524 | { |
1525 | struct vm_area_struct *vma; |
1526 | struct page *page; |
1527 | unsigned int follflags; |
1528 | int err; |
1529 | |
1530 | down_read(&mm->mmap_sem); |
1531 | err = -EFAULT; |
1532 | vma = find_vma(mm, addr); |
1533 | if (!vma || addr < vma->vm_start || !vma_migratable(vma)) |
1534 | goto out; |
1535 | |
1536 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1537 | follflags = FOLL_GET | FOLL_DUMP; |
1538 | page = follow_page(vma, addr, follflags); |
1539 | |
1540 | err = PTR_ERR(page); |
1541 | if (IS_ERR(page)) |
1542 | goto out; |
1543 | |
1544 | err = -ENOENT; |
1545 | if (!page) |
1546 | goto out; |
1547 | |
1548 | err = 0; |
1549 | if (page_to_nid(page) == node) |
1550 | goto out_putpage; |
1551 | |
1552 | err = -EACCES; |
1553 | if (page_mapcount(page) > 1 && !migrate_all) |
1554 | goto out_putpage; |
1555 | |
1556 | if (PageHuge(page)) { |
1557 | if (PageHead(page)) { |
1558 | isolate_huge_page(page, pagelist); |
1559 | err = 0; |
1560 | } |
1561 | } else { |
1562 | struct page *head; |
1563 | |
1564 | head = compound_head(page); |
1565 | err = isolate_lru_page(head); |
1566 | if (err) |
1567 | goto out_putpage; |
1568 | |
1569 | err = 0; |
1570 | list_add_tail(&head->lru, pagelist); |
1571 | mod_node_page_state(page_pgdat(head), |
1572 | NR_ISOLATED_ANON + page_is_file_cache(head), |
1573 | hpage_nr_pages(head)); |
1574 | } |
1575 | out_putpage: |
1576 | /* |
1577 | * Either remove the duplicate refcount from |
1578 | * isolate_lru_page() or drop the page ref if it was |
1579 | * not isolated. |
1580 | */ |
1581 | put_page(page); |
1582 | out: |
1583 | up_read(&mm->mmap_sem); |
1584 | return err; |
1585 | } |
1586 | |
1587 | /* |
1588 | * Migrate an array of page address onto an array of nodes and fill |
1589 | * the corresponding array of status. |
1590 | */ |
1591 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
1592 | unsigned long nr_pages, |
1593 | const void __user * __user *pages, |
1594 | const int __user *nodes, |
1595 | int __user *status, int flags) |
1596 | { |
1597 | int current_node = NUMA_NO_NODE; |
1598 | LIST_HEAD(pagelist); |
1599 | int start, i; |
1600 | int err = 0, err1; |
1601 | |
1602 | migrate_prep(); |
1603 | |
1604 | for (i = start = 0; i < nr_pages; i++) { |
1605 | const void __user *p; |
1606 | unsigned long addr; |
1607 | int node; |
1608 | |
1609 | err = -EFAULT; |
1610 | if (get_user(p, pages + i)) |
1611 | goto out_flush; |
1612 | if (get_user(node, nodes + i)) |
1613 | goto out_flush; |
1614 | addr = (unsigned long)p; |
1615 | |
1616 | err = -ENODEV; |
1617 | if (node < 0 || node >= MAX_NUMNODES) |
1618 | goto out_flush; |
1619 | if (!node_state(node, N_MEMORY)) |
1620 | goto out_flush; |
1621 | |
1622 | err = -EACCES; |
1623 | if (!node_isset(node, task_nodes)) |
1624 | goto out_flush; |
1625 | |
1626 | if (current_node == NUMA_NO_NODE) { |
1627 | current_node = node; |
1628 | start = i; |
1629 | } else if (node != current_node) { |
1630 | err = do_move_pages_to_node(mm, &pagelist, current_node); |
1631 | if (err) |
1632 | goto out; |
1633 | err = store_status(status, start, current_node, i - start); |
1634 | if (err) |
1635 | goto out; |
1636 | start = i; |
1637 | current_node = node; |
1638 | } |
1639 | |
1640 | /* |
1641 | * Errors in the page lookup or isolation are not fatal and we simply |
1642 | * report them via status |
1643 | */ |
1644 | err = add_page_for_migration(mm, addr, current_node, |
1645 | &pagelist, flags & MPOL_MF_MOVE_ALL); |
1646 | if (!err) |
1647 | continue; |
1648 | |
1649 | err = store_status(status, i, err, 1); |
1650 | if (err) |
1651 | goto out_flush; |
1652 | |
1653 | err = do_move_pages_to_node(mm, &pagelist, current_node); |
1654 | if (err) |
1655 | goto out; |
1656 | if (i > start) { |
1657 | err = store_status(status, start, current_node, i - start); |
1658 | if (err) |
1659 | goto out; |
1660 | } |
1661 | current_node = NUMA_NO_NODE; |
1662 | } |
1663 | out_flush: |
1664 | if (list_empty(&pagelist)) |
1665 | return err; |
1666 | |
1667 | /* Make sure we do not overwrite the existing error */ |
1668 | err1 = do_move_pages_to_node(mm, &pagelist, current_node); |
1669 | if (!err1) |
1670 | err1 = store_status(status, start, current_node, i - start); |
1671 | if (!err) |
1672 | err = err1; |
1673 | out: |
1674 | return err; |
1675 | } |
1676 | |
1677 | /* |
1678 | * Determine the nodes of an array of pages and store it in an array of status. |
1679 | */ |
1680 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1681 | const void __user **pages, int *status) |
1682 | { |
1683 | unsigned long i; |
1684 | |
1685 | down_read(&mm->mmap_sem); |
1686 | |
1687 | for (i = 0; i < nr_pages; i++) { |
1688 | unsigned long addr = (unsigned long)(*pages); |
1689 | struct vm_area_struct *vma; |
1690 | struct page *page; |
1691 | int err = -EFAULT; |
1692 | |
1693 | vma = find_vma(mm, addr); |
1694 | if (!vma || addr < vma->vm_start) |
1695 | goto set_status; |
1696 | |
1697 | /* FOLL_DUMP to ignore special (like zero) pages */ |
1698 | page = follow_page(vma, addr, FOLL_DUMP); |
1699 | |
1700 | err = PTR_ERR(page); |
1701 | if (IS_ERR(page)) |
1702 | goto set_status; |
1703 | |
1704 | err = page ? page_to_nid(page) : -ENOENT; |
1705 | set_status: |
1706 | *status = err; |
1707 | |
1708 | pages++; |
1709 | status++; |
1710 | } |
1711 | |
1712 | up_read(&mm->mmap_sem); |
1713 | } |
1714 | |
1715 | /* |
1716 | * Determine the nodes of a user array of pages and store it in |
1717 | * a user array of status. |
1718 | */ |
1719 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, |
1720 | const void __user * __user *pages, |
1721 | int __user *status) |
1722 | { |
1723 | #define DO_PAGES_STAT_CHUNK_NR 16 |
1724 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; |
1725 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; |
1726 | |
1727 | while (nr_pages) { |
1728 | unsigned long chunk_nr; |
1729 | |
1730 | chunk_nr = nr_pages; |
1731 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) |
1732 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; |
1733 | |
1734 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) |
1735 | break; |
1736 | |
1737 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); |
1738 | |
1739 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1740 | break; |
1741 | |
1742 | pages += chunk_nr; |
1743 | status += chunk_nr; |
1744 | nr_pages -= chunk_nr; |
1745 | } |
1746 | return nr_pages ? -EFAULT : 0; |
1747 | } |
1748 | |
1749 | /* |
1750 | * Move a list of pages in the address space of the currently executing |
1751 | * process. |
1752 | */ |
1753 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, |
1754 | const void __user * __user *pages, |
1755 | const int __user *nodes, |
1756 | int __user *status, int flags) |
1757 | { |
1758 | struct task_struct *task; |
1759 | struct mm_struct *mm; |
1760 | int err; |
1761 | nodemask_t task_nodes; |
1762 | |
1763 | /* Check flags */ |
1764 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) |
1765 | return -EINVAL; |
1766 | |
1767 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
1768 | return -EPERM; |
1769 | |
1770 | /* Find the mm_struct */ |
1771 | rcu_read_lock(); |
1772 | task = pid ? find_task_by_vpid(pid) : current; |
1773 | if (!task) { |
1774 | rcu_read_unlock(); |
1775 | return -ESRCH; |
1776 | } |
1777 | get_task_struct(task); |
1778 | |
1779 | /* |
1780 | * Check if this process has the right to modify the specified |
1781 | * process. Use the regular "ptrace_may_access()" checks. |
1782 | */ |
1783 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
1784 | rcu_read_unlock(); |
1785 | err = -EPERM; |
1786 | goto out; |
1787 | } |
1788 | rcu_read_unlock(); |
1789 | |
1790 | err = security_task_movememory(task); |
1791 | if (err) |
1792 | goto out; |
1793 | |
1794 | task_nodes = cpuset_mems_allowed(task); |
1795 | mm = get_task_mm(task); |
1796 | put_task_struct(task); |
1797 | |
1798 | if (!mm) |
1799 | return -EINVAL; |
1800 | |
1801 | if (nodes) |
1802 | err = do_pages_move(mm, task_nodes, nr_pages, pages, |
1803 | nodes, status, flags); |
1804 | else |
1805 | err = do_pages_stat(mm, nr_pages, pages, status); |
1806 | |
1807 | mmput(mm); |
1808 | return err; |
1809 | |
1810 | out: |
1811 | put_task_struct(task); |
1812 | return err; |
1813 | } |
1814 | |
1815 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1816 | const void __user * __user *, pages, |
1817 | const int __user *, nodes, |
1818 | int __user *, status, int, flags) |
1819 | { |
1820 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); |
1821 | } |
1822 | |
1823 | #ifdef CONFIG_COMPAT |
1824 | COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages, |
1825 | compat_uptr_t __user *, pages32, |
1826 | const int __user *, nodes, |
1827 | int __user *, status, |
1828 | int, flags) |
1829 | { |
1830 | const void __user * __user *pages; |
1831 | int i; |
1832 | |
1833 | pages = compat_alloc_user_space(nr_pages * sizeof(void *)); |
1834 | for (i = 0; i < nr_pages; i++) { |
1835 | compat_uptr_t p; |
1836 | |
1837 | if (get_user(p, pages32 + i) || |
1838 | put_user(compat_ptr(p), pages + i)) |
1839 | return -EFAULT; |
1840 | } |
1841 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); |
1842 | } |
1843 | #endif /* CONFIG_COMPAT */ |
1844 | |
1845 | #ifdef CONFIG_NUMA_BALANCING |
1846 | /* |
1847 | * Returns true if this is a safe migration target node for misplaced NUMA |
1848 | * pages. Currently it only checks the watermarks which crude |
1849 | */ |
1850 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, |
1851 | unsigned long nr_migrate_pages) |
1852 | { |
1853 | int z; |
1854 | |
1855 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
1856 | struct zone *zone = pgdat->node_zones + z; |
1857 | |
1858 | if (!populated_zone(zone)) |
1859 | continue; |
1860 | |
1861 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
1862 | if (!zone_watermark_ok(zone, 0, |
1863 | high_wmark_pages(zone) + |
1864 | nr_migrate_pages, |
1865 | 0, 0)) |
1866 | continue; |
1867 | return true; |
1868 | } |
1869 | return false; |
1870 | } |
1871 | |
1872 | static struct page *alloc_misplaced_dst_page(struct page *page, |
1873 | unsigned long data) |
1874 | { |
1875 | int nid = (int) data; |
1876 | struct page *newpage; |
1877 | |
1878 | newpage = __alloc_pages_node(nid, |
1879 | (GFP_HIGHUSER_MOVABLE | |
1880 | __GFP_THISNODE | __GFP_NOMEMALLOC | |
1881 | __GFP_NORETRY | __GFP_NOWARN) & |
1882 | ~__GFP_RECLAIM, 0); |
1883 | |
1884 | return newpage; |
1885 | } |
1886 | |
1887 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
1888 | { |
1889 | int page_lru; |
1890 | |
1891 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); |
1892 | |
1893 | /* Avoid migrating to a node that is nearly full */ |
1894 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1895 | return 0; |
1896 | |
1897 | if (isolate_lru_page(page)) |
1898 | return 0; |
1899 | |
1900 | /* |
1901 | * migrate_misplaced_transhuge_page() skips page migration's usual |
1902 | * check on page_count(), so we must do it here, now that the page |
1903 | * has been isolated: a GUP pin, or any other pin, prevents migration. |
1904 | * The expected page count is 3: 1 for page's mapcount and 1 for the |
1905 | * caller's pin and 1 for the reference taken by isolate_lru_page(). |
1906 | */ |
1907 | if (PageTransHuge(page) && page_count(page) != 3) { |
1908 | putback_lru_page(page); |
1909 | return 0; |
1910 | } |
1911 | |
1912 | page_lru = page_is_file_cache(page); |
1913 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, |
1914 | hpage_nr_pages(page)); |
1915 | |
1916 | /* |
1917 | * Isolating the page has taken another reference, so the |
1918 | * caller's reference can be safely dropped without the page |
1919 | * disappearing underneath us during migration. |
1920 | */ |
1921 | put_page(page); |
1922 | return 1; |
1923 | } |
1924 | |
1925 | bool pmd_trans_migrating(pmd_t pmd) |
1926 | { |
1927 | struct page *page = pmd_page(pmd); |
1928 | return PageLocked(page); |
1929 | } |
1930 | |
1931 | /* |
1932 | * Attempt to migrate a misplaced page to the specified destination |
1933 | * node. Caller is expected to have an elevated reference count on |
1934 | * the page that will be dropped by this function before returning. |
1935 | */ |
1936 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1937 | int node) |
1938 | { |
1939 | pg_data_t *pgdat = NODE_DATA(node); |
1940 | int isolated; |
1941 | int nr_remaining; |
1942 | LIST_HEAD(migratepages); |
1943 | |
1944 | /* |
1945 | * Don't migrate file pages that are mapped in multiple processes |
1946 | * with execute permissions as they are probably shared libraries. |
1947 | */ |
1948 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1949 | (vma->vm_flags & VM_EXEC)) |
1950 | goto out; |
1951 | |
1952 | /* |
1953 | * Also do not migrate dirty pages as not all filesystems can move |
1954 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. |
1955 | */ |
1956 | if (page_is_file_cache(page) && PageDirty(page)) |
1957 | goto out; |
1958 | |
1959 | isolated = numamigrate_isolate_page(pgdat, page); |
1960 | if (!isolated) |
1961 | goto out; |
1962 | |
1963 | list_add(&page->lru, &migratepages); |
1964 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
1965 | NULL, node, MIGRATE_ASYNC, |
1966 | MR_NUMA_MISPLACED); |
1967 | if (nr_remaining) { |
1968 | if (!list_empty(&migratepages)) { |
1969 | list_del(&page->lru); |
1970 | dec_node_page_state(page, NR_ISOLATED_ANON + |
1971 | page_is_file_cache(page)); |
1972 | putback_lru_page(page); |
1973 | } |
1974 | isolated = 0; |
1975 | } else |
1976 | count_vm_numa_event(NUMA_PAGE_MIGRATE); |
1977 | BUG_ON(!list_empty(&migratepages)); |
1978 | return isolated; |
1979 | |
1980 | out: |
1981 | put_page(page); |
1982 | return 0; |
1983 | } |
1984 | #endif /* CONFIG_NUMA_BALANCING */ |
1985 | |
1986 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
1987 | /* |
1988 | * Migrates a THP to a given target node. page must be locked and is unlocked |
1989 | * before returning. |
1990 | */ |
1991 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1992 | struct vm_area_struct *vma, |
1993 | pmd_t *pmd, pmd_t entry, |
1994 | unsigned long address, |
1995 | struct page *page, int node) |
1996 | { |
1997 | spinlock_t *ptl; |
1998 | pg_data_t *pgdat = NODE_DATA(node); |
1999 | int isolated = 0; |
2000 | struct page *new_page = NULL; |
2001 | int page_lru = page_is_file_cache(page); |
2002 | unsigned long start = address & HPAGE_PMD_MASK; |
2003 | |
2004 | new_page = alloc_pages_node(node, |
2005 | (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), |
2006 | HPAGE_PMD_ORDER); |
2007 | if (!new_page) |
2008 | goto out_fail; |
2009 | prep_transhuge_page(new_page); |
2010 | |
2011 | isolated = numamigrate_isolate_page(pgdat, page); |
2012 | if (!isolated) { |
2013 | put_page(new_page); |
2014 | goto out_fail; |
2015 | } |
2016 | |
2017 | /* Prepare a page as a migration target */ |
2018 | __SetPageLocked(new_page); |
2019 | if (PageSwapBacked(page)) |
2020 | __SetPageSwapBacked(new_page); |
2021 | |
2022 | /* anon mapping, we can simply copy page->mapping to the new page: */ |
2023 | new_page->mapping = page->mapping; |
2024 | new_page->index = page->index; |
2025 | /* flush the cache before copying using the kernel virtual address */ |
2026 | flush_cache_range(vma, start, start + HPAGE_PMD_SIZE); |
2027 | migrate_page_copy(new_page, page); |
2028 | WARN_ON(PageLRU(new_page)); |
2029 | |
2030 | /* Recheck the target PMD */ |
2031 | ptl = pmd_lock(mm, pmd); |
2032 | if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) { |
2033 | spin_unlock(ptl); |
2034 | |
2035 | /* Reverse changes made by migrate_page_copy() */ |
2036 | if (TestClearPageActive(new_page)) |
2037 | SetPageActive(page); |
2038 | if (TestClearPageUnevictable(new_page)) |
2039 | SetPageUnevictable(page); |
2040 | |
2041 | unlock_page(new_page); |
2042 | put_page(new_page); /* Free it */ |
2043 | |
2044 | /* Retake the callers reference and putback on LRU */ |
2045 | get_page(page); |
2046 | putback_lru_page(page); |
2047 | mod_node_page_state(page_pgdat(page), |
2048 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); |
2049 | |
2050 | goto out_unlock; |
2051 | } |
2052 | |
2053 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
2054 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
2055 | |
2056 | /* |
2057 | * Overwrite the old entry under pagetable lock and establish |
2058 | * the new PTE. Any parallel GUP will either observe the old |
2059 | * page blocking on the page lock, block on the page table |
2060 | * lock or observe the new page. The SetPageUptodate on the |
2061 | * new page and page_add_new_anon_rmap guarantee the copy is |
2062 | * visible before the pagetable update. |
2063 | */ |
2064 | page_add_anon_rmap(new_page, vma, start, true); |
2065 | /* |
2066 | * At this point the pmd is numa/protnone (i.e. non present) and the TLB |
2067 | * has already been flushed globally. So no TLB can be currently |
2068 | * caching this non present pmd mapping. There's no need to clear the |
2069 | * pmd before doing set_pmd_at(), nor to flush the TLB after |
2070 | * set_pmd_at(). Clearing the pmd here would introduce a race |
2071 | * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the |
2072 | * mmap_sem for reading. If the pmd is set to NULL at any given time, |
2073 | * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this |
2074 | * pmd. |
2075 | */ |
2076 | set_pmd_at(mm, start, pmd, entry); |
2077 | update_mmu_cache_pmd(vma, address, &entry); |
2078 | |
2079 | page_ref_unfreeze(page, 2); |
2080 | mlock_migrate_page(new_page, page); |
2081 | page_remove_rmap(page, true); |
2082 | set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); |
2083 | |
2084 | spin_unlock(ptl); |
2085 | |
2086 | /* Take an "isolate" reference and put new page on the LRU. */ |
2087 | get_page(new_page); |
2088 | putback_lru_page(new_page); |
2089 | |
2090 | unlock_page(new_page); |
2091 | unlock_page(page); |
2092 | put_page(page); /* Drop the rmap reference */ |
2093 | put_page(page); /* Drop the LRU isolation reference */ |
2094 | |
2095 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); |
2096 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); |
2097 | |
2098 | mod_node_page_state(page_pgdat(page), |
2099 | NR_ISOLATED_ANON + page_lru, |
2100 | -HPAGE_PMD_NR); |
2101 | return isolated; |
2102 | |
2103 | out_fail: |
2104 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); |
2105 | ptl = pmd_lock(mm, pmd); |
2106 | if (pmd_same(*pmd, entry)) { |
2107 | entry = pmd_modify(entry, vma->vm_page_prot); |
2108 | set_pmd_at(mm, start, pmd, entry); |
2109 | update_mmu_cache_pmd(vma, address, &entry); |
2110 | } |
2111 | spin_unlock(ptl); |
2112 | |
2113 | out_unlock: |
2114 | unlock_page(page); |
2115 | put_page(page); |
2116 | return 0; |
2117 | } |
2118 | #endif /* CONFIG_NUMA_BALANCING */ |
2119 | |
2120 | #endif /* CONFIG_NUMA */ |
2121 | |
2122 | #if defined(CONFIG_MIGRATE_VMA_HELPER) |
2123 | struct migrate_vma { |
2124 | struct vm_area_struct *vma; |
2125 | unsigned long *dst; |
2126 | unsigned long *src; |
2127 | unsigned long cpages; |
2128 | unsigned long npages; |
2129 | unsigned long start; |
2130 | unsigned long end; |
2131 | }; |
2132 | |
2133 | static int migrate_vma_collect_hole(unsigned long start, |
2134 | unsigned long end, |
2135 | struct mm_walk *walk) |
2136 | { |
2137 | struct migrate_vma *migrate = walk->private; |
2138 | unsigned long addr; |
2139 | |
2140 | for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) { |
2141 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; |
2142 | migrate->dst[migrate->npages] = 0; |
2143 | migrate->npages++; |
2144 | migrate->cpages++; |
2145 | } |
2146 | |
2147 | return 0; |
2148 | } |
2149 | |
2150 | static int migrate_vma_collect_skip(unsigned long start, |
2151 | unsigned long end, |
2152 | struct mm_walk *walk) |
2153 | { |
2154 | struct migrate_vma *migrate = walk->private; |
2155 | unsigned long addr; |
2156 | |
2157 | for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) { |
2158 | migrate->dst[migrate->npages] = 0; |
2159 | migrate->src[migrate->npages++] = 0; |
2160 | } |
2161 | |
2162 | return 0; |
2163 | } |
2164 | |
2165 | static int migrate_vma_collect_pmd(pmd_t *pmdp, |
2166 | unsigned long start, |
2167 | unsigned long end, |
2168 | struct mm_walk *walk) |
2169 | { |
2170 | struct migrate_vma *migrate = walk->private; |
2171 | struct vm_area_struct *vma = walk->vma; |
2172 | struct mm_struct *mm = vma->vm_mm; |
2173 | unsigned long addr = start, unmapped = 0; |
2174 | spinlock_t *ptl; |
2175 | pte_t *ptep; |
2176 | |
2177 | again: |
2178 | if (pmd_none(*pmdp)) |
2179 | return migrate_vma_collect_hole(start, end, walk); |
2180 | |
2181 | if (pmd_trans_huge(*pmdp)) { |
2182 | struct page *page; |
2183 | |
2184 | ptl = pmd_lock(mm, pmdp); |
2185 | if (unlikely(!pmd_trans_huge(*pmdp))) { |
2186 | spin_unlock(ptl); |
2187 | goto again; |
2188 | } |
2189 | |
2190 | page = pmd_page(*pmdp); |
2191 | if (is_huge_zero_page(page)) { |
2192 | spin_unlock(ptl); |
2193 | split_huge_pmd(vma, pmdp, addr); |
2194 | if (pmd_trans_unstable(pmdp)) |
2195 | return migrate_vma_collect_skip(start, end, |
2196 | walk); |
2197 | } else { |
2198 | int ret; |
2199 | |
2200 | get_page(page); |
2201 | spin_unlock(ptl); |
2202 | if (unlikely(!trylock_page(page))) |
2203 | return migrate_vma_collect_skip(start, end, |
2204 | walk); |
2205 | ret = split_huge_page(page); |
2206 | unlock_page(page); |
2207 | put_page(page); |
2208 | if (ret) |
2209 | return migrate_vma_collect_skip(start, end, |
2210 | walk); |
2211 | if (pmd_none(*pmdp)) |
2212 | return migrate_vma_collect_hole(start, end, |
2213 | walk); |
2214 | } |
2215 | } |
2216 | |
2217 | if (unlikely(pmd_bad(*pmdp))) |
2218 | return migrate_vma_collect_skip(start, end, walk); |
2219 | |
2220 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); |
2221 | arch_enter_lazy_mmu_mode(); |
2222 | |
2223 | for (; addr < end; addr += PAGE_SIZE, ptep++) { |
2224 | unsigned long mpfn, pfn; |
2225 | struct page *page; |
2226 | swp_entry_t entry; |
2227 | pte_t pte; |
2228 | |
2229 | pte = *ptep; |
2230 | pfn = pte_pfn(pte); |
2231 | |
2232 | if (pte_none(pte)) { |
2233 | mpfn = MIGRATE_PFN_MIGRATE; |
2234 | migrate->cpages++; |
2235 | pfn = 0; |
2236 | goto next; |
2237 | } |
2238 | |
2239 | if (!pte_present(pte)) { |
2240 | mpfn = pfn = 0; |
2241 | |
2242 | /* |
2243 | * Only care about unaddressable device page special |
2244 | * page table entry. Other special swap entries are not |
2245 | * migratable, and we ignore regular swapped page. |
2246 | */ |
2247 | entry = pte_to_swp_entry(pte); |
2248 | if (!is_device_private_entry(entry)) |
2249 | goto next; |
2250 | |
2251 | page = device_private_entry_to_page(entry); |
2252 | mpfn = migrate_pfn(page_to_pfn(page))| |
2253 | MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE; |
2254 | if (is_write_device_private_entry(entry)) |
2255 | mpfn |= MIGRATE_PFN_WRITE; |
2256 | } else { |
2257 | if (is_zero_pfn(pfn)) { |
2258 | mpfn = MIGRATE_PFN_MIGRATE; |
2259 | migrate->cpages++; |
2260 | pfn = 0; |
2261 | goto next; |
2262 | } |
2263 | page = _vm_normal_page(migrate->vma, addr, pte, true); |
2264 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
2265 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; |
2266 | } |
2267 | |
2268 | /* FIXME support THP */ |
2269 | if (!page || !page->mapping || PageTransCompound(page)) { |
2270 | mpfn = pfn = 0; |
2271 | goto next; |
2272 | } |
2273 | pfn = page_to_pfn(page); |
2274 | |
2275 | /* |
2276 | * By getting a reference on the page we pin it and that blocks |
2277 | * any kind of migration. Side effect is that it "freezes" the |
2278 | * pte. |
2279 | * |
2280 | * We drop this reference after isolating the page from the lru |
2281 | * for non device page (device page are not on the lru and thus |
2282 | * can't be dropped from it). |
2283 | */ |
2284 | get_page(page); |
2285 | migrate->cpages++; |
2286 | |
2287 | /* |
2288 | * Optimize for the common case where page is only mapped once |
2289 | * in one process. If we can lock the page, then we can safely |
2290 | * set up a special migration page table entry now. |
2291 | */ |
2292 | if (trylock_page(page)) { |
2293 | pte_t swp_pte; |
2294 | |
2295 | mpfn |= MIGRATE_PFN_LOCKED; |
2296 | ptep_get_and_clear(mm, addr, ptep); |
2297 | |
2298 | /* Setup special migration page table entry */ |
2299 | entry = make_migration_entry(page, mpfn & |
2300 | MIGRATE_PFN_WRITE); |
2301 | swp_pte = swp_entry_to_pte(entry); |
2302 | if (pte_soft_dirty(pte)) |
2303 | swp_pte = pte_swp_mksoft_dirty(swp_pte); |
2304 | set_pte_at(mm, addr, ptep, swp_pte); |
2305 | |
2306 | /* |
2307 | * This is like regular unmap: we remove the rmap and |
2308 | * drop page refcount. Page won't be freed, as we took |
2309 | * a reference just above. |
2310 | */ |
2311 | page_remove_rmap(page, false); |
2312 | put_page(page); |
2313 | |
2314 | if (pte_present(pte)) |
2315 | unmapped++; |
2316 | } |
2317 | |
2318 | next: |
2319 | migrate->dst[migrate->npages] = 0; |
2320 | migrate->src[migrate->npages++] = mpfn; |
2321 | } |
2322 | arch_leave_lazy_mmu_mode(); |
2323 | pte_unmap_unlock(ptep - 1, ptl); |
2324 | |
2325 | /* Only flush the TLB if we actually modified any entries */ |
2326 | if (unmapped) |
2327 | flush_tlb_range(walk->vma, start, end); |
2328 | |
2329 | return 0; |
2330 | } |
2331 | |
2332 | /* |
2333 | * migrate_vma_collect() - collect pages over a range of virtual addresses |
2334 | * @migrate: migrate struct containing all migration information |
2335 | * |
2336 | * This will walk the CPU page table. For each virtual address backed by a |
2337 | * valid page, it updates the src array and takes a reference on the page, in |
2338 | * order to pin the page until we lock it and unmap it. |
2339 | */ |
2340 | static void migrate_vma_collect(struct migrate_vma *migrate) |
2341 | { |
2342 | struct mmu_notifier_range range; |
2343 | struct mm_walk mm_walk; |
2344 | |
2345 | mm_walk.pmd_entry = migrate_vma_collect_pmd; |
2346 | mm_walk.pte_entry = NULL; |
2347 | mm_walk.pte_hole = migrate_vma_collect_hole; |
2348 | mm_walk.hugetlb_entry = NULL; |
2349 | mm_walk.test_walk = NULL; |
2350 | mm_walk.vma = migrate->vma; |
2351 | mm_walk.mm = migrate->vma->vm_mm; |
2352 | mm_walk.private = migrate; |
2353 | |
2354 | mmu_notifier_range_init(&range, mm_walk.mm, migrate->start, |
2355 | migrate->end); |
2356 | mmu_notifier_invalidate_range_start(&range); |
2357 | walk_page_range(migrate->start, migrate->end, &mm_walk); |
2358 | mmu_notifier_invalidate_range_end(&range); |
2359 | |
2360 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); |
2361 | } |
2362 | |
2363 | /* |
2364 | * migrate_vma_check_page() - check if page is pinned or not |
2365 | * @page: struct page to check |
2366 | * |
2367 | * Pinned pages cannot be migrated. This is the same test as in |
2368 | * migrate_page_move_mapping(), except that here we allow migration of a |
2369 | * ZONE_DEVICE page. |
2370 | */ |
2371 | static bool migrate_vma_check_page(struct page *page) |
2372 | { |
2373 | /* |
2374 | * One extra ref because caller holds an extra reference, either from |
2375 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for |
2376 | * a device page. |
2377 | */ |
2378 | int = 1; |
2379 | |
2380 | /* |
2381 | * FIXME support THP (transparent huge page), it is bit more complex to |
2382 | * check them than regular pages, because they can be mapped with a pmd |
2383 | * or with a pte (split pte mapping). |
2384 | */ |
2385 | if (PageCompound(page)) |
2386 | return false; |
2387 | |
2388 | /* Page from ZONE_DEVICE have one extra reference */ |
2389 | if (is_zone_device_page(page)) { |
2390 | /* |
2391 | * Private page can never be pin as they have no valid pte and |
2392 | * GUP will fail for those. Yet if there is a pending migration |
2393 | * a thread might try to wait on the pte migration entry and |
2394 | * will bump the page reference count. Sadly there is no way to |
2395 | * differentiate a regular pin from migration wait. Hence to |
2396 | * avoid 2 racing thread trying to migrate back to CPU to enter |
2397 | * infinite loop (one stoping migration because the other is |
2398 | * waiting on pte migration entry). We always return true here. |
2399 | * |
2400 | * FIXME proper solution is to rework migration_entry_wait() so |
2401 | * it does not need to take a reference on page. |
2402 | */ |
2403 | if (is_device_private_page(page)) |
2404 | return true; |
2405 | |
2406 | /* |
2407 | * Only allow device public page to be migrated and account for |
2408 | * the extra reference count imply by ZONE_DEVICE pages. |
2409 | */ |
2410 | if (!is_device_public_page(page)) |
2411 | return false; |
2412 | extra++; |
2413 | } |
2414 | |
2415 | /* For file back page */ |
2416 | if (page_mapping(page)) |
2417 | extra += 1 + page_has_private(page); |
2418 | |
2419 | if ((page_count(page) - extra) > page_mapcount(page)) |
2420 | return false; |
2421 | |
2422 | return true; |
2423 | } |
2424 | |
2425 | /* |
2426 | * migrate_vma_prepare() - lock pages and isolate them from the lru |
2427 | * @migrate: migrate struct containing all migration information |
2428 | * |
2429 | * This locks pages that have been collected by migrate_vma_collect(). Once each |
2430 | * page is locked it is isolated from the lru (for non-device pages). Finally, |
2431 | * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be |
2432 | * migrated by concurrent kernel threads. |
2433 | */ |
2434 | static void migrate_vma_prepare(struct migrate_vma *migrate) |
2435 | { |
2436 | const unsigned long npages = migrate->npages; |
2437 | const unsigned long start = migrate->start; |
2438 | unsigned long addr, i, restore = 0; |
2439 | bool allow_drain = true; |
2440 | |
2441 | lru_add_drain(); |
2442 | |
2443 | for (i = 0; (i < npages) && migrate->cpages; i++) { |
2444 | struct page *page = migrate_pfn_to_page(migrate->src[i]); |
2445 | bool remap = true; |
2446 | |
2447 | if (!page) |
2448 | continue; |
2449 | |
2450 | if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { |
2451 | /* |
2452 | * Because we are migrating several pages there can be |
2453 | * a deadlock between 2 concurrent migration where each |
2454 | * are waiting on each other page lock. |
2455 | * |
2456 | * Make migrate_vma() a best effort thing and backoff |
2457 | * for any page we can not lock right away. |
2458 | */ |
2459 | if (!trylock_page(page)) { |
2460 | migrate->src[i] = 0; |
2461 | migrate->cpages--; |
2462 | put_page(page); |
2463 | continue; |
2464 | } |
2465 | remap = false; |
2466 | migrate->src[i] |= MIGRATE_PFN_LOCKED; |
2467 | } |
2468 | |
2469 | /* ZONE_DEVICE pages are not on LRU */ |
2470 | if (!is_zone_device_page(page)) { |
2471 | if (!PageLRU(page) && allow_drain) { |
2472 | /* Drain CPU's pagevec */ |
2473 | lru_add_drain_all(); |
2474 | allow_drain = false; |
2475 | } |
2476 | |
2477 | if (isolate_lru_page(page)) { |
2478 | if (remap) { |
2479 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; |
2480 | migrate->cpages--; |
2481 | restore++; |
2482 | } else { |
2483 | migrate->src[i] = 0; |
2484 | unlock_page(page); |
2485 | migrate->cpages--; |
2486 | put_page(page); |
2487 | } |
2488 | continue; |
2489 | } |
2490 | |
2491 | /* Drop the reference we took in collect */ |
2492 | put_page(page); |
2493 | } |
2494 | |
2495 | if (!migrate_vma_check_page(page)) { |
2496 | if (remap) { |
2497 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; |
2498 | migrate->cpages--; |
2499 | restore++; |
2500 | |
2501 | if (!is_zone_device_page(page)) { |
2502 | get_page(page); |
2503 | putback_lru_page(page); |
2504 | } |
2505 | } else { |
2506 | migrate->src[i] = 0; |
2507 | unlock_page(page); |
2508 | migrate->cpages--; |
2509 | |
2510 | if (!is_zone_device_page(page)) |
2511 | putback_lru_page(page); |
2512 | else |
2513 | put_page(page); |
2514 | } |
2515 | } |
2516 | } |
2517 | |
2518 | for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { |
2519 | struct page *page = migrate_pfn_to_page(migrate->src[i]); |
2520 | |
2521 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) |
2522 | continue; |
2523 | |
2524 | remove_migration_pte(page, migrate->vma, addr, page); |
2525 | |
2526 | migrate->src[i] = 0; |
2527 | unlock_page(page); |
2528 | put_page(page); |
2529 | restore--; |
2530 | } |
2531 | } |
2532 | |
2533 | /* |
2534 | * migrate_vma_unmap() - replace page mapping with special migration pte entry |
2535 | * @migrate: migrate struct containing all migration information |
2536 | * |
2537 | * Replace page mapping (CPU page table pte) with a special migration pte entry |
2538 | * and check again if it has been pinned. Pinned pages are restored because we |
2539 | * cannot migrate them. |
2540 | * |
2541 | * This is the last step before we call the device driver callback to allocate |
2542 | * destination memory and copy contents of original page over to new page. |
2543 | */ |
2544 | static void migrate_vma_unmap(struct migrate_vma *migrate) |
2545 | { |
2546 | int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; |
2547 | const unsigned long npages = migrate->npages; |
2548 | const unsigned long start = migrate->start; |
2549 | unsigned long addr, i, restore = 0; |
2550 | |
2551 | for (i = 0; i < npages; i++) { |
2552 | struct page *page = migrate_pfn_to_page(migrate->src[i]); |
2553 | |
2554 | if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) |
2555 | continue; |
2556 | |
2557 | if (page_mapped(page)) { |
2558 | try_to_unmap(page, flags); |
2559 | if (page_mapped(page)) |
2560 | goto restore; |
2561 | } |
2562 | |
2563 | if (migrate_vma_check_page(page)) |
2564 | continue; |
2565 | |
2566 | restore: |
2567 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; |
2568 | migrate->cpages--; |
2569 | restore++; |
2570 | } |
2571 | |
2572 | for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { |
2573 | struct page *page = migrate_pfn_to_page(migrate->src[i]); |
2574 | |
2575 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) |
2576 | continue; |
2577 | |
2578 | remove_migration_ptes(page, page, false); |
2579 | |
2580 | migrate->src[i] = 0; |
2581 | unlock_page(page); |
2582 | restore--; |
2583 | |
2584 | if (is_zone_device_page(page)) |
2585 | put_page(page); |
2586 | else |
2587 | putback_lru_page(page); |
2588 | } |
2589 | } |
2590 | |
2591 | static void migrate_vma_insert_page(struct migrate_vma *migrate, |
2592 | unsigned long addr, |
2593 | struct page *page, |
2594 | unsigned long *src, |
2595 | unsigned long *dst) |
2596 | { |
2597 | struct vm_area_struct *vma = migrate->vma; |
2598 | struct mm_struct *mm = vma->vm_mm; |
2599 | struct mem_cgroup *memcg; |
2600 | bool flush = false; |
2601 | spinlock_t *ptl; |
2602 | pte_t entry; |
2603 | pgd_t *pgdp; |
2604 | p4d_t *p4dp; |
2605 | pud_t *pudp; |
2606 | pmd_t *pmdp; |
2607 | pte_t *ptep; |
2608 | |
2609 | /* Only allow populating anonymous memory */ |
2610 | if (!vma_is_anonymous(vma)) |
2611 | goto abort; |
2612 | |
2613 | pgdp = pgd_offset(mm, addr); |
2614 | p4dp = p4d_alloc(mm, pgdp, addr); |
2615 | if (!p4dp) |
2616 | goto abort; |
2617 | pudp = pud_alloc(mm, p4dp, addr); |
2618 | if (!pudp) |
2619 | goto abort; |
2620 | pmdp = pmd_alloc(mm, pudp, addr); |
2621 | if (!pmdp) |
2622 | goto abort; |
2623 | |
2624 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) |
2625 | goto abort; |
2626 | |
2627 | /* |
2628 | * Use pte_alloc() instead of pte_alloc_map(). We can't run |
2629 | * pte_offset_map() on pmds where a huge pmd might be created |
2630 | * from a different thread. |
2631 | * |
2632 | * pte_alloc_map() is safe to use under down_write(mmap_sem) or when |
2633 | * parallel threads are excluded by other means. |
2634 | * |
2635 | * Here we only have down_read(mmap_sem). |
2636 | */ |
2637 | if (pte_alloc(mm, pmdp)) |
2638 | goto abort; |
2639 | |
2640 | /* See the comment in pte_alloc_one_map() */ |
2641 | if (unlikely(pmd_trans_unstable(pmdp))) |
2642 | goto abort; |
2643 | |
2644 | if (unlikely(anon_vma_prepare(vma))) |
2645 | goto abort; |
2646 | if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false)) |
2647 | goto abort; |
2648 | |
2649 | /* |
2650 | * The memory barrier inside __SetPageUptodate makes sure that |
2651 | * preceding stores to the page contents become visible before |
2652 | * the set_pte_at() write. |
2653 | */ |
2654 | __SetPageUptodate(page); |
2655 | |
2656 | if (is_zone_device_page(page)) { |
2657 | if (is_device_private_page(page)) { |
2658 | swp_entry_t swp_entry; |
2659 | |
2660 | swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE); |
2661 | entry = swp_entry_to_pte(swp_entry); |
2662 | } else if (is_device_public_page(page)) { |
2663 | entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot))); |
2664 | if (vma->vm_flags & VM_WRITE) |
2665 | entry = pte_mkwrite(pte_mkdirty(entry)); |
2666 | entry = pte_mkdevmap(entry); |
2667 | } |
2668 | } else { |
2669 | entry = mk_pte(page, vma->vm_page_prot); |
2670 | if (vma->vm_flags & VM_WRITE) |
2671 | entry = pte_mkwrite(pte_mkdirty(entry)); |
2672 | } |
2673 | |
2674 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); |
2675 | |
2676 | if (pte_present(*ptep)) { |
2677 | unsigned long pfn = pte_pfn(*ptep); |
2678 | |
2679 | if (!is_zero_pfn(pfn)) { |
2680 | pte_unmap_unlock(ptep, ptl); |
2681 | mem_cgroup_cancel_charge(page, memcg, false); |
2682 | goto abort; |
2683 | } |
2684 | flush = true; |
2685 | } else if (!pte_none(*ptep)) { |
2686 | pte_unmap_unlock(ptep, ptl); |
2687 | mem_cgroup_cancel_charge(page, memcg, false); |
2688 | goto abort; |
2689 | } |
2690 | |
2691 | /* |
2692 | * Check for usefaultfd but do not deliver the fault. Instead, |
2693 | * just back off. |
2694 | */ |
2695 | if (userfaultfd_missing(vma)) { |
2696 | pte_unmap_unlock(ptep, ptl); |
2697 | mem_cgroup_cancel_charge(page, memcg, false); |
2698 | goto abort; |
2699 | } |
2700 | |
2701 | inc_mm_counter(mm, MM_ANONPAGES); |
2702 | page_add_new_anon_rmap(page, vma, addr, false); |
2703 | mem_cgroup_commit_charge(page, memcg, false, false); |
2704 | if (!is_zone_device_page(page)) |
2705 | lru_cache_add_active_or_unevictable(page, vma); |
2706 | get_page(page); |
2707 | |
2708 | if (flush) { |
2709 | flush_cache_page(vma, addr, pte_pfn(*ptep)); |
2710 | ptep_clear_flush_notify(vma, addr, ptep); |
2711 | set_pte_at_notify(mm, addr, ptep, entry); |
2712 | update_mmu_cache(vma, addr, ptep); |
2713 | } else { |
2714 | /* No need to invalidate - it was non-present before */ |
2715 | set_pte_at(mm, addr, ptep, entry); |
2716 | update_mmu_cache(vma, addr, ptep); |
2717 | } |
2718 | |
2719 | pte_unmap_unlock(ptep, ptl); |
2720 | *src = MIGRATE_PFN_MIGRATE; |
2721 | return; |
2722 | |
2723 | abort: |
2724 | *src &= ~MIGRATE_PFN_MIGRATE; |
2725 | } |
2726 | |
2727 | /* |
2728 | * migrate_vma_pages() - migrate meta-data from src page to dst page |
2729 | * @migrate: migrate struct containing all migration information |
2730 | * |
2731 | * This migrates struct page meta-data from source struct page to destination |
2732 | * struct page. This effectively finishes the migration from source page to the |
2733 | * destination page. |
2734 | */ |
2735 | static void migrate_vma_pages(struct migrate_vma *migrate) |
2736 | { |
2737 | const unsigned long npages = migrate->npages; |
2738 | const unsigned long start = migrate->start; |
2739 | struct mmu_notifier_range range; |
2740 | unsigned long addr, i; |
2741 | bool notified = false; |
2742 | |
2743 | for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { |
2744 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); |
2745 | struct page *page = migrate_pfn_to_page(migrate->src[i]); |
2746 | struct address_space *mapping; |
2747 | int r; |
2748 | |
2749 | if (!newpage) { |
2750 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; |
2751 | continue; |
2752 | } |
2753 | |
2754 | if (!page) { |
2755 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) { |
2756 | continue; |
2757 | } |
2758 | if (!notified) { |
2759 | notified = true; |
2760 | |
2761 | mmu_notifier_range_init(&range, |
2762 | migrate->vma->vm_mm, |
2763 | addr, migrate->end); |
2764 | mmu_notifier_invalidate_range_start(&range); |
2765 | } |
2766 | migrate_vma_insert_page(migrate, addr, newpage, |
2767 | &migrate->src[i], |
2768 | &migrate->dst[i]); |
2769 | continue; |
2770 | } |
2771 | |
2772 | mapping = page_mapping(page); |
2773 | |
2774 | if (is_zone_device_page(newpage)) { |
2775 | if (is_device_private_page(newpage)) { |
2776 | /* |
2777 | * For now only support private anonymous when |
2778 | * migrating to un-addressable device memory. |
2779 | */ |
2780 | if (mapping) { |
2781 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; |
2782 | continue; |
2783 | } |
2784 | } else if (!is_device_public_page(newpage)) { |
2785 | /* |
2786 | * Other types of ZONE_DEVICE page are not |
2787 | * supported. |
2788 | */ |
2789 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; |
2790 | continue; |
2791 | } |
2792 | } |
2793 | |
2794 | r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); |
2795 | if (r != MIGRATEPAGE_SUCCESS) |
2796 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; |
2797 | } |
2798 | |
2799 | /* |
2800 | * No need to double call mmu_notifier->invalidate_range() callback as |
2801 | * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() |
2802 | * did already call it. |
2803 | */ |
2804 | if (notified) |
2805 | mmu_notifier_invalidate_range_only_end(&range); |
2806 | } |
2807 | |
2808 | /* |
2809 | * migrate_vma_finalize() - restore CPU page table entry |
2810 | * @migrate: migrate struct containing all migration information |
2811 | * |
2812 | * This replaces the special migration pte entry with either a mapping to the |
2813 | * new page if migration was successful for that page, or to the original page |
2814 | * otherwise. |
2815 | * |
2816 | * This also unlocks the pages and puts them back on the lru, or drops the extra |
2817 | * refcount, for device pages. |
2818 | */ |
2819 | static void migrate_vma_finalize(struct migrate_vma *migrate) |
2820 | { |
2821 | const unsigned long npages = migrate->npages; |
2822 | unsigned long i; |
2823 | |
2824 | for (i = 0; i < npages; i++) { |
2825 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); |
2826 | struct page *page = migrate_pfn_to_page(migrate->src[i]); |
2827 | |
2828 | if (!page) { |
2829 | if (newpage) { |
2830 | unlock_page(newpage); |
2831 | put_page(newpage); |
2832 | } |
2833 | continue; |
2834 | } |
2835 | |
2836 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
2837 | if (newpage) { |
2838 | unlock_page(newpage); |
2839 | put_page(newpage); |
2840 | } |
2841 | newpage = page; |
2842 | } |
2843 | |
2844 | remove_migration_ptes(page, newpage, false); |
2845 | unlock_page(page); |
2846 | migrate->cpages--; |
2847 | |
2848 | if (is_zone_device_page(page)) |
2849 | put_page(page); |
2850 | else |
2851 | putback_lru_page(page); |
2852 | |
2853 | if (newpage != page) { |
2854 | unlock_page(newpage); |
2855 | if (is_zone_device_page(newpage)) |
2856 | put_page(newpage); |
2857 | else |
2858 | putback_lru_page(newpage); |
2859 | } |
2860 | } |
2861 | } |
2862 | |
2863 | /* |
2864 | * migrate_vma() - migrate a range of memory inside vma |
2865 | * |
2866 | * @ops: migration callback for allocating destination memory and copying |
2867 | * @vma: virtual memory area containing the range to be migrated |
2868 | * @start: start address of the range to migrate (inclusive) |
2869 | * @end: end address of the range to migrate (exclusive) |
2870 | * @src: array of hmm_pfn_t containing source pfns |
2871 | * @dst: array of hmm_pfn_t containing destination pfns |
2872 | * @private: pointer passed back to each of the callback |
2873 | * Returns: 0 on success, error code otherwise |
2874 | * |
2875 | * This function tries to migrate a range of memory virtual address range, using |
2876 | * callbacks to allocate and copy memory from source to destination. First it |
2877 | * collects all the pages backing each virtual address in the range, saving this |
2878 | * inside the src array. Then it locks those pages and unmaps them. Once the pages |
2879 | * are locked and unmapped, it checks whether each page is pinned or not. Pages |
2880 | * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) |
2881 | * in the corresponding src array entry. It then restores any pages that are |
2882 | * pinned, by remapping and unlocking those pages. |
2883 | * |
2884 | * At this point it calls the alloc_and_copy() callback. For documentation on |
2885 | * what is expected from that callback, see struct migrate_vma_ops comments in |
2886 | * include/linux/migrate.h |
2887 | * |
2888 | * After the alloc_and_copy() callback, this function goes over each entry in |
2889 | * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag |
2890 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, |
2891 | * then the function tries to migrate struct page information from the source |
2892 | * struct page to the destination struct page. If it fails to migrate the struct |
2893 | * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src |
2894 | * array. |
2895 | * |
2896 | * At this point all successfully migrated pages have an entry in the src |
2897 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst |
2898 | * array entry with MIGRATE_PFN_VALID flag set. |
2899 | * |
2900 | * It then calls the finalize_and_map() callback. See comments for "struct |
2901 | * migrate_vma_ops", in include/linux/migrate.h for details about |
2902 | * finalize_and_map() behavior. |
2903 | * |
2904 | * After the finalize_and_map() callback, for successfully migrated pages, this |
2905 | * function updates the CPU page table to point to new pages, otherwise it |
2906 | * restores the CPU page table to point to the original source pages. |
2907 | * |
2908 | * Function returns 0 after the above steps, even if no pages were migrated |
2909 | * (The function only returns an error if any of the arguments are invalid.) |
2910 | * |
2911 | * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT |
2912 | * unsigned long entries. |
2913 | */ |
2914 | int migrate_vma(const struct migrate_vma_ops *ops, |
2915 | struct vm_area_struct *vma, |
2916 | unsigned long start, |
2917 | unsigned long end, |
2918 | unsigned long *src, |
2919 | unsigned long *dst, |
2920 | void *private) |
2921 | { |
2922 | struct migrate_vma migrate; |
2923 | |
2924 | /* Sanity check the arguments */ |
2925 | start &= PAGE_MASK; |
2926 | end &= PAGE_MASK; |
2927 | if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) || |
2928 | vma_is_dax(vma)) |
2929 | return -EINVAL; |
2930 | if (start < vma->vm_start || start >= vma->vm_end) |
2931 | return -EINVAL; |
2932 | if (end <= vma->vm_start || end > vma->vm_end) |
2933 | return -EINVAL; |
2934 | if (!ops || !src || !dst || start >= end) |
2935 | return -EINVAL; |
2936 | |
2937 | memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT)); |
2938 | migrate.src = src; |
2939 | migrate.dst = dst; |
2940 | migrate.start = start; |
2941 | migrate.npages = 0; |
2942 | migrate.cpages = 0; |
2943 | migrate.end = end; |
2944 | migrate.vma = vma; |
2945 | |
2946 | /* Collect, and try to unmap source pages */ |
2947 | migrate_vma_collect(&migrate); |
2948 | if (!migrate.cpages) |
2949 | return 0; |
2950 | |
2951 | /* Lock and isolate page */ |
2952 | migrate_vma_prepare(&migrate); |
2953 | if (!migrate.cpages) |
2954 | return 0; |
2955 | |
2956 | /* Unmap pages */ |
2957 | migrate_vma_unmap(&migrate); |
2958 | if (!migrate.cpages) |
2959 | return 0; |
2960 | |
2961 | /* |
2962 | * At this point pages are locked and unmapped, and thus they have |
2963 | * stable content and can safely be copied to destination memory that |
2964 | * is allocated by the callback. |
2965 | * |
2966 | * Note that migration can fail in migrate_vma_struct_page() for each |
2967 | * individual page. |
2968 | */ |
2969 | ops->alloc_and_copy(vma, src, dst, start, end, private); |
2970 | |
2971 | /* This does the real migration of struct page */ |
2972 | migrate_vma_pages(&migrate); |
2973 | |
2974 | ops->finalize_and_map(vma, src, dst, start, end, private); |
2975 | |
2976 | /* Unlock and remap pages */ |
2977 | migrate_vma_finalize(&migrate); |
2978 | |
2979 | return 0; |
2980 | } |
2981 | EXPORT_SYMBOL(migrate_vma); |
2982 | #endif /* defined(MIGRATE_VMA_HELPER) */ |
2983 | |