]>
Commit | Line | Data |
---|---|---|
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/pagewalk.h> | |
42 | #include <linux/pfn_t.h> | |
43 | #include <linux/memremap.h> | |
44 | #include <linux/userfaultfd_k.h> | |
45 | #include <linux/balloon_compaction.h> | |
46 | #include <linux/mmu_notifier.h> | |
47 | #include <linux/page_idle.h> | |
48 | #include <linux/page_owner.h> | |
49 | #include <linux/sched/mm.h> | |
50 | #include <linux/ptrace.h> | |
51 | #include <linux/oom.h> | |
52 | ||
53 | #include <asm/tlbflush.h> | |
54 | ||
55 | #define CREATE_TRACE_POINTS | |
56 | #include <trace/events/migrate.h> | |
57 | ||
58 | #include "internal.h" | |
59 | ||
60 | /* | |
61 | * migrate_prep() needs to be called before we start compiling a list of pages | |
62 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is | |
63 | * undesirable, use migrate_prep_local() | |
64 | */ | |
65 | void migrate_prep(void) | |
66 | { | |
67 | /* | |
68 | * Clear the LRU lists so pages can be isolated. | |
69 | * Note that pages may be moved off the LRU after we have | |
70 | * drained them. Those pages will fail to migrate like other | |
71 | * pages that may be busy. | |
72 | */ | |
73 | lru_add_drain_all(); | |
74 | } | |
75 | ||
76 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ | |
77 | void migrate_prep_local(void) | |
78 | { | |
79 | lru_add_drain(); | |
80 | } | |
81 | ||
82 | int isolate_movable_page(struct page *page, isolate_mode_t mode) | |
83 | { | |
84 | struct address_space *mapping; | |
85 | ||
86 | /* | |
87 | * Avoid burning cycles with pages that are yet under __free_pages(), | |
88 | * or just got freed under us. | |
89 | * | |
90 | * In case we 'win' a race for a movable page being freed under us and | |
91 | * raise its refcount preventing __free_pages() from doing its job | |
92 | * the put_page() at the end of this block will take care of | |
93 | * release this page, thus avoiding a nasty leakage. | |
94 | */ | |
95 | if (unlikely(!get_page_unless_zero(page))) | |
96 | goto out; | |
97 | ||
98 | /* | |
99 | * Check PageMovable before holding a PG_lock because page's owner | |
100 | * assumes anybody doesn't touch PG_lock of newly allocated page | |
101 | * so unconditionally grabbing the lock ruins page's owner side. | |
102 | */ | |
103 | if (unlikely(!__PageMovable(page))) | |
104 | goto out_putpage; | |
105 | /* | |
106 | * As movable pages are not isolated from LRU lists, concurrent | |
107 | * compaction threads can race against page migration functions | |
108 | * as well as race against the releasing a page. | |
109 | * | |
110 | * In order to avoid having an already isolated movable page | |
111 | * being (wrongly) re-isolated while it is under migration, | |
112 | * or to avoid attempting to isolate pages being released, | |
113 | * lets be sure we have the page lock | |
114 | * before proceeding with the movable page isolation steps. | |
115 | */ | |
116 | if (unlikely(!trylock_page(page))) | |
117 | goto out_putpage; | |
118 | ||
119 | if (!PageMovable(page) || PageIsolated(page)) | |
120 | goto out_no_isolated; | |
121 | ||
122 | mapping = page_mapping(page); | |
123 | VM_BUG_ON_PAGE(!mapping, page); | |
124 | ||
125 | if (!mapping->a_ops->isolate_page(page, mode)) | |
126 | goto out_no_isolated; | |
127 | ||
128 | /* Driver shouldn't use PG_isolated bit of page->flags */ | |
129 | WARN_ON_ONCE(PageIsolated(page)); | |
130 | __SetPageIsolated(page); | |
131 | unlock_page(page); | |
132 | ||
133 | return 0; | |
134 | ||
135 | out_no_isolated: | |
136 | unlock_page(page); | |
137 | out_putpage: | |
138 | put_page(page); | |
139 | out: | |
140 | return -EBUSY; | |
141 | } | |
142 | ||
143 | /* It should be called on page which is PG_movable */ | |
144 | void putback_movable_page(struct page *page) | |
145 | { | |
146 | struct address_space *mapping; | |
147 | ||
148 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
149 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
150 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
151 | ||
152 | mapping = page_mapping(page); | |
153 | mapping->a_ops->putback_page(page); | |
154 | __ClearPageIsolated(page); | |
155 | } | |
156 | ||
157 | /* | |
158 | * Put previously isolated pages back onto the appropriate lists | |
159 | * from where they were once taken off for compaction/migration. | |
160 | * | |
161 | * This function shall be used whenever the isolated pageset has been | |
162 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() | |
163 | * and isolate_huge_page(). | |
164 | */ | |
165 | void putback_movable_pages(struct list_head *l) | |
166 | { | |
167 | struct page *page; | |
168 | struct page *page2; | |
169 | ||
170 | list_for_each_entry_safe(page, page2, l, lru) { | |
171 | if (unlikely(PageHuge(page))) { | |
172 | putback_active_hugepage(page); | |
173 | continue; | |
174 | } | |
175 | list_del(&page->lru); | |
176 | /* | |
177 | * We isolated non-lru movable page so here we can use | |
178 | * __PageMovable because LRU page's mapping cannot have | |
179 | * PAGE_MAPPING_MOVABLE. | |
180 | */ | |
181 | if (unlikely(__PageMovable(page))) { | |
182 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
183 | lock_page(page); | |
184 | if (PageMovable(page)) | |
185 | putback_movable_page(page); | |
186 | else | |
187 | __ClearPageIsolated(page); | |
188 | unlock_page(page); | |
189 | put_page(page); | |
190 | } else { | |
191 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + | |
192 | page_is_file_lru(page), -thp_nr_pages(page)); | |
193 | putback_lru_page(page); | |
194 | } | |
195 | } | |
196 | } | |
197 | ||
198 | /* | |
199 | * Restore a potential migration pte to a working pte entry | |
200 | */ | |
201 | static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, | |
202 | unsigned long addr, void *old) | |
203 | { | |
204 | struct page_vma_mapped_walk pvmw = { | |
205 | .page = old, | |
206 | .vma = vma, | |
207 | .address = addr, | |
208 | .flags = PVMW_SYNC | PVMW_MIGRATION, | |
209 | }; | |
210 | struct page *new; | |
211 | pte_t pte; | |
212 | swp_entry_t entry; | |
213 | ||
214 | VM_BUG_ON_PAGE(PageTail(page), page); | |
215 | while (page_vma_mapped_walk(&pvmw)) { | |
216 | if (PageKsm(page)) | |
217 | new = page; | |
218 | else | |
219 | new = page - pvmw.page->index + | |
220 | linear_page_index(vma, pvmw.address); | |
221 | ||
222 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION | |
223 | /* PMD-mapped THP migration entry */ | |
224 | if (!pvmw.pte) { | |
225 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); | |
226 | remove_migration_pmd(&pvmw, new); | |
227 | continue; | |
228 | } | |
229 | #endif | |
230 | ||
231 | get_page(new); | |
232 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); | |
233 | if (pte_swp_soft_dirty(*pvmw.pte)) | |
234 | pte = pte_mksoft_dirty(pte); | |
235 | ||
236 | /* | |
237 | * Recheck VMA as permissions can change since migration started | |
238 | */ | |
239 | entry = pte_to_swp_entry(*pvmw.pte); | |
240 | if (is_write_migration_entry(entry)) | |
241 | pte = maybe_mkwrite(pte, vma); | |
242 | else if (pte_swp_uffd_wp(*pvmw.pte)) | |
243 | pte = pte_mkuffd_wp(pte); | |
244 | ||
245 | if (unlikely(is_device_private_page(new))) { | |
246 | entry = make_device_private_entry(new, pte_write(pte)); | |
247 | pte = swp_entry_to_pte(entry); | |
248 | if (pte_swp_soft_dirty(*pvmw.pte)) | |
249 | pte = pte_swp_mksoft_dirty(pte); | |
250 | if (pte_swp_uffd_wp(*pvmw.pte)) | |
251 | pte = pte_swp_mkuffd_wp(pte); | |
252 | } | |
253 | ||
254 | #ifdef CONFIG_HUGETLB_PAGE | |
255 | if (PageHuge(new)) { | |
256 | pte = pte_mkhuge(pte); | |
257 | pte = arch_make_huge_pte(pte, vma, new, 0); | |
258 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); | |
259 | if (PageAnon(new)) | |
260 | hugepage_add_anon_rmap(new, vma, pvmw.address); | |
261 | else | |
262 | page_dup_rmap(new, true); | |
263 | } else | |
264 | #endif | |
265 | { | |
266 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); | |
267 | ||
268 | if (PageAnon(new)) | |
269 | page_add_anon_rmap(new, vma, pvmw.address, false); | |
270 | else | |
271 | page_add_file_rmap(new, false); | |
272 | } | |
273 | if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) | |
274 | mlock_vma_page(new); | |
275 | ||
276 | if (PageTransHuge(page) && PageMlocked(page)) | |
277 | clear_page_mlock(page); | |
278 | ||
279 | /* No need to invalidate - it was non-present before */ | |
280 | update_mmu_cache(vma, pvmw.address, pvmw.pte); | |
281 | } | |
282 | ||
283 | return true; | |
284 | } | |
285 | ||
286 | /* | |
287 | * Get rid of all migration entries and replace them by | |
288 | * references to the indicated page. | |
289 | */ | |
290 | void remove_migration_ptes(struct page *old, struct page *new, bool locked) | |
291 | { | |
292 | struct rmap_walk_control rwc = { | |
293 | .rmap_one = remove_migration_pte, | |
294 | .arg = old, | |
295 | }; | |
296 | ||
297 | if (locked) | |
298 | rmap_walk_locked(new, &rwc); | |
299 | else | |
300 | rmap_walk(new, &rwc); | |
301 | } | |
302 | ||
303 | /* | |
304 | * Something used the pte of a page under migration. We need to | |
305 | * get to the page and wait until migration is finished. | |
306 | * When we return from this function the fault will be retried. | |
307 | */ | |
308 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, | |
309 | spinlock_t *ptl) | |
310 | { | |
311 | pte_t pte; | |
312 | swp_entry_t entry; | |
313 | struct page *page; | |
314 | ||
315 | spin_lock(ptl); | |
316 | pte = *ptep; | |
317 | if (!is_swap_pte(pte)) | |
318 | goto out; | |
319 | ||
320 | entry = pte_to_swp_entry(pte); | |
321 | if (!is_migration_entry(entry)) | |
322 | goto out; | |
323 | ||
324 | page = migration_entry_to_page(entry); | |
325 | ||
326 | /* | |
327 | * Once page cache replacement of page migration started, page_count | |
328 | * is zero; but we must not call put_and_wait_on_page_locked() without | |
329 | * a ref. Use get_page_unless_zero(), and just fault again if it fails. | |
330 | */ | |
331 | if (!get_page_unless_zero(page)) | |
332 | goto out; | |
333 | pte_unmap_unlock(ptep, ptl); | |
334 | put_and_wait_on_page_locked(page); | |
335 | return; | |
336 | out: | |
337 | pte_unmap_unlock(ptep, ptl); | |
338 | } | |
339 | ||
340 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, | |
341 | unsigned long address) | |
342 | { | |
343 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
344 | pte_t *ptep = pte_offset_map(pmd, address); | |
345 | __migration_entry_wait(mm, ptep, ptl); | |
346 | } | |
347 | ||
348 | void migration_entry_wait_huge(struct vm_area_struct *vma, | |
349 | struct mm_struct *mm, pte_t *pte) | |
350 | { | |
351 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); | |
352 | __migration_entry_wait(mm, pte, ptl); | |
353 | } | |
354 | ||
355 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION | |
356 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) | |
357 | { | |
358 | spinlock_t *ptl; | |
359 | struct page *page; | |
360 | ||
361 | ptl = pmd_lock(mm, pmd); | |
362 | if (!is_pmd_migration_entry(*pmd)) | |
363 | goto unlock; | |
364 | page = migration_entry_to_page(pmd_to_swp_entry(*pmd)); | |
365 | if (!get_page_unless_zero(page)) | |
366 | goto unlock; | |
367 | spin_unlock(ptl); | |
368 | put_and_wait_on_page_locked(page); | |
369 | return; | |
370 | unlock: | |
371 | spin_unlock(ptl); | |
372 | } | |
373 | #endif | |
374 | ||
375 | static int expected_page_refs(struct address_space *mapping, struct page *page) | |
376 | { | |
377 | int expected_count = 1; | |
378 | ||
379 | /* | |
380 | * Device private pages have an extra refcount as they are | |
381 | * ZONE_DEVICE pages. | |
382 | */ | |
383 | expected_count += is_device_private_page(page); | |
384 | if (mapping) | |
385 | expected_count += thp_nr_pages(page) + page_has_private(page); | |
386 | ||
387 | return expected_count; | |
388 | } | |
389 | ||
390 | /* | |
391 | * Replace the page in the mapping. | |
392 | * | |
393 | * The number of remaining references must be: | |
394 | * 1 for anonymous pages without a mapping | |
395 | * 2 for pages with a mapping | |
396 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. | |
397 | */ | |
398 | int migrate_page_move_mapping(struct address_space *mapping, | |
399 | struct page *newpage, struct page *page, int extra_count) | |
400 | { | |
401 | XA_STATE(xas, &mapping->i_pages, page_index(page)); | |
402 | struct zone *oldzone, *newzone; | |
403 | int dirty; | |
404 | int expected_count = expected_page_refs(mapping, page) + extra_count; | |
405 | ||
406 | if (!mapping) { | |
407 | /* Anonymous page without mapping */ | |
408 | if (page_count(page) != expected_count) | |
409 | return -EAGAIN; | |
410 | ||
411 | /* No turning back from here */ | |
412 | newpage->index = page->index; | |
413 | newpage->mapping = page->mapping; | |
414 | if (PageSwapBacked(page)) | |
415 | __SetPageSwapBacked(newpage); | |
416 | ||
417 | return MIGRATEPAGE_SUCCESS; | |
418 | } | |
419 | ||
420 | oldzone = page_zone(page); | |
421 | newzone = page_zone(newpage); | |
422 | ||
423 | xas_lock_irq(&xas); | |
424 | if (page_count(page) != expected_count || xas_load(&xas) != page) { | |
425 | xas_unlock_irq(&xas); | |
426 | return -EAGAIN; | |
427 | } | |
428 | ||
429 | if (!page_ref_freeze(page, expected_count)) { | |
430 | xas_unlock_irq(&xas); | |
431 | return -EAGAIN; | |
432 | } | |
433 | ||
434 | /* | |
435 | * Now we know that no one else is looking at the page: | |
436 | * no turning back from here. | |
437 | */ | |
438 | newpage->index = page->index; | |
439 | newpage->mapping = page->mapping; | |
440 | page_ref_add(newpage, thp_nr_pages(page)); /* add cache reference */ | |
441 | if (PageSwapBacked(page)) { | |
442 | __SetPageSwapBacked(newpage); | |
443 | if (PageSwapCache(page)) { | |
444 | SetPageSwapCache(newpage); | |
445 | set_page_private(newpage, page_private(page)); | |
446 | } | |
447 | } else { | |
448 | VM_BUG_ON_PAGE(PageSwapCache(page), page); | |
449 | } | |
450 | ||
451 | /* Move dirty while page refs frozen and newpage not yet exposed */ | |
452 | dirty = PageDirty(page); | |
453 | if (dirty) { | |
454 | ClearPageDirty(page); | |
455 | SetPageDirty(newpage); | |
456 | } | |
457 | ||
458 | xas_store(&xas, newpage); | |
459 | if (PageTransHuge(page)) { | |
460 | int i; | |
461 | ||
462 | for (i = 1; i < HPAGE_PMD_NR; i++) { | |
463 | xas_next(&xas); | |
464 | xas_store(&xas, newpage); | |
465 | } | |
466 | } | |
467 | ||
468 | /* | |
469 | * Drop cache reference from old page by unfreezing | |
470 | * to one less reference. | |
471 | * We know this isn't the last reference. | |
472 | */ | |
473 | page_ref_unfreeze(page, expected_count - thp_nr_pages(page)); | |
474 | ||
475 | xas_unlock(&xas); | |
476 | /* Leave irq disabled to prevent preemption while updating stats */ | |
477 | ||
478 | /* | |
479 | * If moved to a different zone then also account | |
480 | * the page for that zone. Other VM counters will be | |
481 | * taken care of when we establish references to the | |
482 | * new page and drop references to the old page. | |
483 | * | |
484 | * Note that anonymous pages are accounted for | |
485 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they | |
486 | * are mapped to swap space. | |
487 | */ | |
488 | if (newzone != oldzone) { | |
489 | struct lruvec *old_lruvec, *new_lruvec; | |
490 | struct mem_cgroup *memcg; | |
491 | ||
492 | memcg = page_memcg(page); | |
493 | old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); | |
494 | new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); | |
495 | ||
496 | __dec_lruvec_state(old_lruvec, NR_FILE_PAGES); | |
497 | __inc_lruvec_state(new_lruvec, NR_FILE_PAGES); | |
498 | if (PageSwapBacked(page) && !PageSwapCache(page)) { | |
499 | __dec_lruvec_state(old_lruvec, NR_SHMEM); | |
500 | __inc_lruvec_state(new_lruvec, NR_SHMEM); | |
501 | } | |
502 | if (dirty && mapping_can_writeback(mapping)) { | |
503 | __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY); | |
504 | __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING); | |
505 | __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY); | |
506 | __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING); | |
507 | } | |
508 | } | |
509 | local_irq_enable(); | |
510 | ||
511 | return MIGRATEPAGE_SUCCESS; | |
512 | } | |
513 | EXPORT_SYMBOL(migrate_page_move_mapping); | |
514 | ||
515 | /* | |
516 | * The expected number of remaining references is the same as that | |
517 | * of migrate_page_move_mapping(). | |
518 | */ | |
519 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
520 | struct page *newpage, struct page *page) | |
521 | { | |
522 | XA_STATE(xas, &mapping->i_pages, page_index(page)); | |
523 | int expected_count; | |
524 | ||
525 | xas_lock_irq(&xas); | |
526 | expected_count = 2 + page_has_private(page); | |
527 | if (page_count(page) != expected_count || xas_load(&xas) != page) { | |
528 | xas_unlock_irq(&xas); | |
529 | return -EAGAIN; | |
530 | } | |
531 | ||
532 | if (!page_ref_freeze(page, expected_count)) { | |
533 | xas_unlock_irq(&xas); | |
534 | return -EAGAIN; | |
535 | } | |
536 | ||
537 | newpage->index = page->index; | |
538 | newpage->mapping = page->mapping; | |
539 | ||
540 | get_page(newpage); | |
541 | ||
542 | xas_store(&xas, newpage); | |
543 | ||
544 | page_ref_unfreeze(page, expected_count - 1); | |
545 | ||
546 | xas_unlock_irq(&xas); | |
547 | ||
548 | return MIGRATEPAGE_SUCCESS; | |
549 | } | |
550 | ||
551 | /* | |
552 | * Gigantic pages are so large that we do not guarantee that page++ pointer | |
553 | * arithmetic will work across the entire page. We need something more | |
554 | * specialized. | |
555 | */ | |
556 | static void __copy_gigantic_page(struct page *dst, struct page *src, | |
557 | int nr_pages) | |
558 | { | |
559 | int i; | |
560 | struct page *dst_base = dst; | |
561 | struct page *src_base = src; | |
562 | ||
563 | for (i = 0; i < nr_pages; ) { | |
564 | cond_resched(); | |
565 | copy_highpage(dst, src); | |
566 | ||
567 | i++; | |
568 | dst = mem_map_next(dst, dst_base, i); | |
569 | src = mem_map_next(src, src_base, i); | |
570 | } | |
571 | } | |
572 | ||
573 | static void copy_huge_page(struct page *dst, struct page *src) | |
574 | { | |
575 | int i; | |
576 | int nr_pages; | |
577 | ||
578 | if (PageHuge(src)) { | |
579 | /* hugetlbfs page */ | |
580 | struct hstate *h = page_hstate(src); | |
581 | nr_pages = pages_per_huge_page(h); | |
582 | ||
583 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { | |
584 | __copy_gigantic_page(dst, src, nr_pages); | |
585 | return; | |
586 | } | |
587 | } else { | |
588 | /* thp page */ | |
589 | BUG_ON(!PageTransHuge(src)); | |
590 | nr_pages = thp_nr_pages(src); | |
591 | } | |
592 | ||
593 | for (i = 0; i < nr_pages; i++) { | |
594 | cond_resched(); | |
595 | copy_highpage(dst + i, src + i); | |
596 | } | |
597 | } | |
598 | ||
599 | /* | |
600 | * Copy the page to its new location | |
601 | */ | |
602 | void migrate_page_states(struct page *newpage, struct page *page) | |
603 | { | |
604 | int cpupid; | |
605 | ||
606 | if (PageError(page)) | |
607 | SetPageError(newpage); | |
608 | if (PageReferenced(page)) | |
609 | SetPageReferenced(newpage); | |
610 | if (PageUptodate(page)) | |
611 | SetPageUptodate(newpage); | |
612 | if (TestClearPageActive(page)) { | |
613 | VM_BUG_ON_PAGE(PageUnevictable(page), page); | |
614 | SetPageActive(newpage); | |
615 | } else if (TestClearPageUnevictable(page)) | |
616 | SetPageUnevictable(newpage); | |
617 | if (PageWorkingset(page)) | |
618 | SetPageWorkingset(newpage); | |
619 | if (PageChecked(page)) | |
620 | SetPageChecked(newpage); | |
621 | if (PageMappedToDisk(page)) | |
622 | SetPageMappedToDisk(newpage); | |
623 | ||
624 | /* Move dirty on pages not done by migrate_page_move_mapping() */ | |
625 | if (PageDirty(page)) | |
626 | SetPageDirty(newpage); | |
627 | ||
628 | if (page_is_young(page)) | |
629 | set_page_young(newpage); | |
630 | if (page_is_idle(page)) | |
631 | set_page_idle(newpage); | |
632 | ||
633 | /* | |
634 | * Copy NUMA information to the new page, to prevent over-eager | |
635 | * future migrations of this same page. | |
636 | */ | |
637 | cpupid = page_cpupid_xchg_last(page, -1); | |
638 | page_cpupid_xchg_last(newpage, cpupid); | |
639 | ||
640 | ksm_migrate_page(newpage, page); | |
641 | /* | |
642 | * Please do not reorder this without considering how mm/ksm.c's | |
643 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
644 | */ | |
645 | if (PageSwapCache(page)) | |
646 | ClearPageSwapCache(page); | |
647 | ClearPagePrivate(page); | |
648 | set_page_private(page, 0); | |
649 | ||
650 | /* | |
651 | * If any waiters have accumulated on the new page then | |
652 | * wake them up. | |
653 | */ | |
654 | if (PageWriteback(newpage)) | |
655 | end_page_writeback(newpage); | |
656 | ||
657 | /* | |
658 | * PG_readahead shares the same bit with PG_reclaim. The above | |
659 | * end_page_writeback() may clear PG_readahead mistakenly, so set the | |
660 | * bit after that. | |
661 | */ | |
662 | if (PageReadahead(page)) | |
663 | SetPageReadahead(newpage); | |
664 | ||
665 | copy_page_owner(page, newpage); | |
666 | ||
667 | if (!PageHuge(page)) | |
668 | mem_cgroup_migrate(page, newpage); | |
669 | } | |
670 | EXPORT_SYMBOL(migrate_page_states); | |
671 | ||
672 | void migrate_page_copy(struct page *newpage, struct page *page) | |
673 | { | |
674 | if (PageHuge(page) || PageTransHuge(page)) | |
675 | copy_huge_page(newpage, page); | |
676 | else | |
677 | copy_highpage(newpage, page); | |
678 | ||
679 | migrate_page_states(newpage, page); | |
680 | } | |
681 | EXPORT_SYMBOL(migrate_page_copy); | |
682 | ||
683 | /************************************************************ | |
684 | * Migration functions | |
685 | ***********************************************************/ | |
686 | ||
687 | /* | |
688 | * Common logic to directly migrate a single LRU page suitable for | |
689 | * pages that do not use PagePrivate/PagePrivate2. | |
690 | * | |
691 | * Pages are locked upon entry and exit. | |
692 | */ | |
693 | int migrate_page(struct address_space *mapping, | |
694 | struct page *newpage, struct page *page, | |
695 | enum migrate_mode mode) | |
696 | { | |
697 | int rc; | |
698 | ||
699 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
700 | ||
701 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); | |
702 | ||
703 | if (rc != MIGRATEPAGE_SUCCESS) | |
704 | return rc; | |
705 | ||
706 | if (mode != MIGRATE_SYNC_NO_COPY) | |
707 | migrate_page_copy(newpage, page); | |
708 | else | |
709 | migrate_page_states(newpage, page); | |
710 | return MIGRATEPAGE_SUCCESS; | |
711 | } | |
712 | EXPORT_SYMBOL(migrate_page); | |
713 | ||
714 | #ifdef CONFIG_BLOCK | |
715 | /* Returns true if all buffers are successfully locked */ | |
716 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
717 | enum migrate_mode mode) | |
718 | { | |
719 | struct buffer_head *bh = head; | |
720 | ||
721 | /* Simple case, sync compaction */ | |
722 | if (mode != MIGRATE_ASYNC) { | |
723 | do { | |
724 | lock_buffer(bh); | |
725 | bh = bh->b_this_page; | |
726 | ||
727 | } while (bh != head); | |
728 | ||
729 | return true; | |
730 | } | |
731 | ||
732 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
733 | do { | |
734 | if (!trylock_buffer(bh)) { | |
735 | /* | |
736 | * We failed to lock the buffer and cannot stall in | |
737 | * async migration. Release the taken locks | |
738 | */ | |
739 | struct buffer_head *failed_bh = bh; | |
740 | bh = head; | |
741 | while (bh != failed_bh) { | |
742 | unlock_buffer(bh); | |
743 | bh = bh->b_this_page; | |
744 | } | |
745 | return false; | |
746 | } | |
747 | ||
748 | bh = bh->b_this_page; | |
749 | } while (bh != head); | |
750 | return true; | |
751 | } | |
752 | ||
753 | static int __buffer_migrate_page(struct address_space *mapping, | |
754 | struct page *newpage, struct page *page, enum migrate_mode mode, | |
755 | bool check_refs) | |
756 | { | |
757 | struct buffer_head *bh, *head; | |
758 | int rc; | |
759 | int expected_count; | |
760 | ||
761 | if (!page_has_buffers(page)) | |
762 | return migrate_page(mapping, newpage, page, mode); | |
763 | ||
764 | /* Check whether page does not have extra refs before we do more work */ | |
765 | expected_count = expected_page_refs(mapping, page); | |
766 | if (page_count(page) != expected_count) | |
767 | return -EAGAIN; | |
768 | ||
769 | head = page_buffers(page); | |
770 | if (!buffer_migrate_lock_buffers(head, mode)) | |
771 | return -EAGAIN; | |
772 | ||
773 | if (check_refs) { | |
774 | bool busy; | |
775 | bool invalidated = false; | |
776 | ||
777 | recheck_buffers: | |
778 | busy = false; | |
779 | spin_lock(&mapping->private_lock); | |
780 | bh = head; | |
781 | do { | |
782 | if (atomic_read(&bh->b_count)) { | |
783 | busy = true; | |
784 | break; | |
785 | } | |
786 | bh = bh->b_this_page; | |
787 | } while (bh != head); | |
788 | if (busy) { | |
789 | if (invalidated) { | |
790 | rc = -EAGAIN; | |
791 | goto unlock_buffers; | |
792 | } | |
793 | spin_unlock(&mapping->private_lock); | |
794 | invalidate_bh_lrus(); | |
795 | invalidated = true; | |
796 | goto recheck_buffers; | |
797 | } | |
798 | } | |
799 | ||
800 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); | |
801 | if (rc != MIGRATEPAGE_SUCCESS) | |
802 | goto unlock_buffers; | |
803 | ||
804 | attach_page_private(newpage, detach_page_private(page)); | |
805 | ||
806 | bh = head; | |
807 | do { | |
808 | set_bh_page(bh, newpage, bh_offset(bh)); | |
809 | bh = bh->b_this_page; | |
810 | ||
811 | } while (bh != head); | |
812 | ||
813 | if (mode != MIGRATE_SYNC_NO_COPY) | |
814 | migrate_page_copy(newpage, page); | |
815 | else | |
816 | migrate_page_states(newpage, page); | |
817 | ||
818 | rc = MIGRATEPAGE_SUCCESS; | |
819 | unlock_buffers: | |
820 | if (check_refs) | |
821 | spin_unlock(&mapping->private_lock); | |
822 | bh = head; | |
823 | do { | |
824 | unlock_buffer(bh); | |
825 | bh = bh->b_this_page; | |
826 | ||
827 | } while (bh != head); | |
828 | ||
829 | return rc; | |
830 | } | |
831 | ||
832 | /* | |
833 | * Migration function for pages with buffers. This function can only be used | |
834 | * if the underlying filesystem guarantees that no other references to "page" | |
835 | * exist. For example attached buffer heads are accessed only under page lock. | |
836 | */ | |
837 | int buffer_migrate_page(struct address_space *mapping, | |
838 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
839 | { | |
840 | return __buffer_migrate_page(mapping, newpage, page, mode, false); | |
841 | } | |
842 | EXPORT_SYMBOL(buffer_migrate_page); | |
843 | ||
844 | /* | |
845 | * Same as above except that this variant is more careful and checks that there | |
846 | * are also no buffer head references. This function is the right one for | |
847 | * mappings where buffer heads are directly looked up and referenced (such as | |
848 | * block device mappings). | |
849 | */ | |
850 | int buffer_migrate_page_norefs(struct address_space *mapping, | |
851 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
852 | { | |
853 | return __buffer_migrate_page(mapping, newpage, page, mode, true); | |
854 | } | |
855 | #endif | |
856 | ||
857 | /* | |
858 | * Writeback a page to clean the dirty state | |
859 | */ | |
860 | static int writeout(struct address_space *mapping, struct page *page) | |
861 | { | |
862 | struct writeback_control wbc = { | |
863 | .sync_mode = WB_SYNC_NONE, | |
864 | .nr_to_write = 1, | |
865 | .range_start = 0, | |
866 | .range_end = LLONG_MAX, | |
867 | .for_reclaim = 1 | |
868 | }; | |
869 | int rc; | |
870 | ||
871 | if (!mapping->a_ops->writepage) | |
872 | /* No write method for the address space */ | |
873 | return -EINVAL; | |
874 | ||
875 | if (!clear_page_dirty_for_io(page)) | |
876 | /* Someone else already triggered a write */ | |
877 | return -EAGAIN; | |
878 | ||
879 | /* | |
880 | * A dirty page may imply that the underlying filesystem has | |
881 | * the page on some queue. So the page must be clean for | |
882 | * migration. Writeout may mean we loose the lock and the | |
883 | * page state is no longer what we checked for earlier. | |
884 | * At this point we know that the migration attempt cannot | |
885 | * be successful. | |
886 | */ | |
887 | remove_migration_ptes(page, page, false); | |
888 | ||
889 | rc = mapping->a_ops->writepage(page, &wbc); | |
890 | ||
891 | if (rc != AOP_WRITEPAGE_ACTIVATE) | |
892 | /* unlocked. Relock */ | |
893 | lock_page(page); | |
894 | ||
895 | return (rc < 0) ? -EIO : -EAGAIN; | |
896 | } | |
897 | ||
898 | /* | |
899 | * Default handling if a filesystem does not provide a migration function. | |
900 | */ | |
901 | static int fallback_migrate_page(struct address_space *mapping, | |
902 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
903 | { | |
904 | if (PageDirty(page)) { | |
905 | /* Only writeback pages in full synchronous migration */ | |
906 | switch (mode) { | |
907 | case MIGRATE_SYNC: | |
908 | case MIGRATE_SYNC_NO_COPY: | |
909 | break; | |
910 | default: | |
911 | return -EBUSY; | |
912 | } | |
913 | return writeout(mapping, page); | |
914 | } | |
915 | ||
916 | /* | |
917 | * Buffers may be managed in a filesystem specific way. | |
918 | * We must have no buffers or drop them. | |
919 | */ | |
920 | if (page_has_private(page) && | |
921 | !try_to_release_page(page, GFP_KERNEL)) | |
922 | return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; | |
923 | ||
924 | return migrate_page(mapping, newpage, page, mode); | |
925 | } | |
926 | ||
927 | /* | |
928 | * Move a page to a newly allocated page | |
929 | * The page is locked and all ptes have been successfully removed. | |
930 | * | |
931 | * The new page will have replaced the old page if this function | |
932 | * is successful. | |
933 | * | |
934 | * Return value: | |
935 | * < 0 - error code | |
936 | * MIGRATEPAGE_SUCCESS - success | |
937 | */ | |
938 | static int move_to_new_page(struct page *newpage, struct page *page, | |
939 | enum migrate_mode mode) | |
940 | { | |
941 | struct address_space *mapping; | |
942 | int rc = -EAGAIN; | |
943 | bool is_lru = !__PageMovable(page); | |
944 | ||
945 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
946 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
947 | ||
948 | mapping = page_mapping(page); | |
949 | ||
950 | if (likely(is_lru)) { | |
951 | if (!mapping) | |
952 | rc = migrate_page(mapping, newpage, page, mode); | |
953 | else if (mapping->a_ops->migratepage) | |
954 | /* | |
955 | * Most pages have a mapping and most filesystems | |
956 | * provide a migratepage callback. Anonymous pages | |
957 | * are part of swap space which also has its own | |
958 | * migratepage callback. This is the most common path | |
959 | * for page migration. | |
960 | */ | |
961 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
962 | page, mode); | |
963 | else | |
964 | rc = fallback_migrate_page(mapping, newpage, | |
965 | page, mode); | |
966 | } else { | |
967 | /* | |
968 | * In case of non-lru page, it could be released after | |
969 | * isolation step. In that case, we shouldn't try migration. | |
970 | */ | |
971 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
972 | if (!PageMovable(page)) { | |
973 | rc = MIGRATEPAGE_SUCCESS; | |
974 | __ClearPageIsolated(page); | |
975 | goto out; | |
976 | } | |
977 | ||
978 | rc = mapping->a_ops->migratepage(mapping, newpage, | |
979 | page, mode); | |
980 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && | |
981 | !PageIsolated(page)); | |
982 | } | |
983 | ||
984 | /* | |
985 | * When successful, old pagecache page->mapping must be cleared before | |
986 | * page is freed; but stats require that PageAnon be left as PageAnon. | |
987 | */ | |
988 | if (rc == MIGRATEPAGE_SUCCESS) { | |
989 | if (__PageMovable(page)) { | |
990 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
991 | ||
992 | /* | |
993 | * We clear PG_movable under page_lock so any compactor | |
994 | * cannot try to migrate this page. | |
995 | */ | |
996 | __ClearPageIsolated(page); | |
997 | } | |
998 | ||
999 | /* | |
1000 | * Anonymous and movable page->mapping will be cleared by | |
1001 | * free_pages_prepare so don't reset it here for keeping | |
1002 | * the type to work PageAnon, for example. | |
1003 | */ | |
1004 | if (!PageMappingFlags(page)) | |
1005 | page->mapping = NULL; | |
1006 | ||
1007 | if (likely(!is_zone_device_page(newpage))) | |
1008 | flush_dcache_page(newpage); | |
1009 | ||
1010 | } | |
1011 | out: | |
1012 | return rc; | |
1013 | } | |
1014 | ||
1015 | static int __unmap_and_move(struct page *page, struct page *newpage, | |
1016 | int force, enum migrate_mode mode) | |
1017 | { | |
1018 | int rc = -EAGAIN; | |
1019 | int page_was_mapped = 0; | |
1020 | struct anon_vma *anon_vma = NULL; | |
1021 | bool is_lru = !__PageMovable(page); | |
1022 | ||
1023 | if (!trylock_page(page)) { | |
1024 | if (!force || mode == MIGRATE_ASYNC) | |
1025 | goto out; | |
1026 | ||
1027 | /* | |
1028 | * It's not safe for direct compaction to call lock_page. | |
1029 | * For example, during page readahead pages are added locked | |
1030 | * to the LRU. Later, when the IO completes the pages are | |
1031 | * marked uptodate and unlocked. However, the queueing | |
1032 | * could be merging multiple pages for one bio (e.g. | |
1033 | * mpage_readahead). If an allocation happens for the | |
1034 | * second or third page, the process can end up locking | |
1035 | * the same page twice and deadlocking. Rather than | |
1036 | * trying to be clever about what pages can be locked, | |
1037 | * avoid the use of lock_page for direct compaction | |
1038 | * altogether. | |
1039 | */ | |
1040 | if (current->flags & PF_MEMALLOC) | |
1041 | goto out; | |
1042 | ||
1043 | lock_page(page); | |
1044 | } | |
1045 | ||
1046 | if (PageWriteback(page)) { | |
1047 | /* | |
1048 | * Only in the case of a full synchronous migration is it | |
1049 | * necessary to wait for PageWriteback. In the async case, | |
1050 | * the retry loop is too short and in the sync-light case, | |
1051 | * the overhead of stalling is too much | |
1052 | */ | |
1053 | switch (mode) { | |
1054 | case MIGRATE_SYNC: | |
1055 | case MIGRATE_SYNC_NO_COPY: | |
1056 | break; | |
1057 | default: | |
1058 | rc = -EBUSY; | |
1059 | goto out_unlock; | |
1060 | } | |
1061 | if (!force) | |
1062 | goto out_unlock; | |
1063 | wait_on_page_writeback(page); | |
1064 | } | |
1065 | ||
1066 | /* | |
1067 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, | |
1068 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1069 | * This get_anon_vma() delays freeing anon_vma pointer until the end | |
1070 | * of migration. File cache pages are no problem because of page_lock() | |
1071 | * File Caches may use write_page() or lock_page() in migration, then, | |
1072 | * just care Anon page here. | |
1073 | * | |
1074 | * Only page_get_anon_vma() understands the subtleties of | |
1075 | * getting a hold on an anon_vma from outside one of its mms. | |
1076 | * But if we cannot get anon_vma, then we won't need it anyway, | |
1077 | * because that implies that the anon page is no longer mapped | |
1078 | * (and cannot be remapped so long as we hold the page lock). | |
1079 | */ | |
1080 | if (PageAnon(page) && !PageKsm(page)) | |
1081 | anon_vma = page_get_anon_vma(page); | |
1082 | ||
1083 | /* | |
1084 | * Block others from accessing the new page when we get around to | |
1085 | * establishing additional references. We are usually the only one | |
1086 | * holding a reference to newpage at this point. We used to have a BUG | |
1087 | * here if trylock_page(newpage) fails, but would like to allow for | |
1088 | * cases where there might be a race with the previous use of newpage. | |
1089 | * This is much like races on refcount of oldpage: just don't BUG(). | |
1090 | */ | |
1091 | if (unlikely(!trylock_page(newpage))) | |
1092 | goto out_unlock; | |
1093 | ||
1094 | if (unlikely(!is_lru)) { | |
1095 | rc = move_to_new_page(newpage, page, mode); | |
1096 | goto out_unlock_both; | |
1097 | } | |
1098 | ||
1099 | /* | |
1100 | * Corner case handling: | |
1101 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
1102 | * and treated as swapcache but it has no rmap yet. | |
1103 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
1104 | * trigger a BUG. So handle it here. | |
1105 | * 2. An orphaned page (see truncate_cleanup_page) might have | |
1106 | * fs-private metadata. The page can be picked up due to memory | |
1107 | * offlining. Everywhere else except page reclaim, the page is | |
1108 | * invisible to the vm, so the page can not be migrated. So try to | |
1109 | * free the metadata, so the page can be freed. | |
1110 | */ | |
1111 | if (!page->mapping) { | |
1112 | VM_BUG_ON_PAGE(PageAnon(page), page); | |
1113 | if (page_has_private(page)) { | |
1114 | try_to_free_buffers(page); | |
1115 | goto out_unlock_both; | |
1116 | } | |
1117 | } else if (page_mapped(page)) { | |
1118 | /* Establish migration ptes */ | |
1119 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, | |
1120 | page); | |
1121 | try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK); | |
1122 | page_was_mapped = 1; | |
1123 | } | |
1124 | ||
1125 | if (!page_mapped(page)) | |
1126 | rc = move_to_new_page(newpage, page, mode); | |
1127 | ||
1128 | if (page_was_mapped) | |
1129 | remove_migration_ptes(page, | |
1130 | rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); | |
1131 | ||
1132 | out_unlock_both: | |
1133 | unlock_page(newpage); | |
1134 | out_unlock: | |
1135 | /* Drop an anon_vma reference if we took one */ | |
1136 | if (anon_vma) | |
1137 | put_anon_vma(anon_vma); | |
1138 | unlock_page(page); | |
1139 | out: | |
1140 | /* | |
1141 | * If migration is successful, decrease refcount of the newpage | |
1142 | * which will not free the page because new page owner increased | |
1143 | * refcounter. As well, if it is LRU page, add the page to LRU | |
1144 | * list in here. Use the old state of the isolated source page to | |
1145 | * determine if we migrated a LRU page. newpage was already unlocked | |
1146 | * and possibly modified by its owner - don't rely on the page | |
1147 | * state. | |
1148 | */ | |
1149 | if (rc == MIGRATEPAGE_SUCCESS) { | |
1150 | if (unlikely(!is_lru)) | |
1151 | put_page(newpage); | |
1152 | else | |
1153 | putback_lru_page(newpage); | |
1154 | } | |
1155 | ||
1156 | return rc; | |
1157 | } | |
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 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 | struct list_head *ret) | |
1169 | { | |
1170 | int rc = MIGRATEPAGE_SUCCESS; | |
1171 | struct page *newpage = NULL; | |
1172 | ||
1173 | if (!thp_migration_supported() && PageTransHuge(page)) | |
1174 | return -ENOSYS; | |
1175 | ||
1176 | if (page_count(page) == 1) { | |
1177 | /* page was freed from under us. So we are done. */ | |
1178 | ClearPageActive(page); | |
1179 | ClearPageUnevictable(page); | |
1180 | if (unlikely(__PageMovable(page))) { | |
1181 | lock_page(page); | |
1182 | if (!PageMovable(page)) | |
1183 | __ClearPageIsolated(page); | |
1184 | unlock_page(page); | |
1185 | } | |
1186 | goto out; | |
1187 | } | |
1188 | ||
1189 | newpage = get_new_page(page, private); | |
1190 | if (!newpage) | |
1191 | return -ENOMEM; | |
1192 | ||
1193 | rc = __unmap_and_move(page, newpage, force, mode); | |
1194 | if (rc == MIGRATEPAGE_SUCCESS) | |
1195 | set_page_owner_migrate_reason(newpage, reason); | |
1196 | ||
1197 | out: | |
1198 | if (rc != -EAGAIN) { | |
1199 | /* | |
1200 | * A page that has been migrated has all references | |
1201 | * removed and will be freed. A page that has not been | |
1202 | * migrated will have kept its references and be restored. | |
1203 | */ | |
1204 | list_del(&page->lru); | |
1205 | } | |
1206 | ||
1207 | /* | |
1208 | * If migration is successful, releases reference grabbed during | |
1209 | * isolation. Otherwise, restore the page to right list unless | |
1210 | * we want to retry. | |
1211 | */ | |
1212 | if (rc == MIGRATEPAGE_SUCCESS) { | |
1213 | /* | |
1214 | * Compaction can migrate also non-LRU pages which are | |
1215 | * not accounted to NR_ISOLATED_*. They can be recognized | |
1216 | * as __PageMovable | |
1217 | */ | |
1218 | if (likely(!__PageMovable(page))) | |
1219 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + | |
1220 | page_is_file_lru(page), -thp_nr_pages(page)); | |
1221 | ||
1222 | if (reason != MR_MEMORY_FAILURE) | |
1223 | /* | |
1224 | * We release the page in page_handle_poison. | |
1225 | */ | |
1226 | put_page(page); | |
1227 | } else { | |
1228 | if (rc != -EAGAIN) | |
1229 | list_add_tail(&page->lru, ret); | |
1230 | ||
1231 | if (put_new_page) | |
1232 | put_new_page(newpage, private); | |
1233 | else | |
1234 | put_page(newpage); | |
1235 | } | |
1236 | ||
1237 | return rc; | |
1238 | } | |
1239 | ||
1240 | /* | |
1241 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
1242 | * | |
1243 | * This function doesn't wait the completion of hugepage I/O | |
1244 | * because there is no race between I/O and migration for hugepage. | |
1245 | * Note that currently hugepage I/O occurs only in direct I/O | |
1246 | * where no lock is held and PG_writeback is irrelevant, | |
1247 | * and writeback status of all subpages are counted in the reference | |
1248 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
1249 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
1250 | * This means that when we try to migrate hugepage whose subpages are | |
1251 | * doing direct I/O, some references remain after try_to_unmap() and | |
1252 | * hugepage migration fails without data corruption. | |
1253 | * | |
1254 | * There is also no race when direct I/O is issued on the page under migration, | |
1255 | * because then pte is replaced with migration swap entry and direct I/O code | |
1256 | * will wait in the page fault for migration to complete. | |
1257 | */ | |
1258 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
1259 | free_page_t put_new_page, unsigned long private, | |
1260 | struct page *hpage, int force, | |
1261 | enum migrate_mode mode, int reason, | |
1262 | struct list_head *ret) | |
1263 | { | |
1264 | int rc = -EAGAIN; | |
1265 | int page_was_mapped = 0; | |
1266 | struct page *new_hpage; | |
1267 | struct anon_vma *anon_vma = NULL; | |
1268 | struct address_space *mapping = NULL; | |
1269 | ||
1270 | /* | |
1271 | * Migratability of hugepages depends on architectures and their size. | |
1272 | * This check is necessary because some callers of hugepage migration | |
1273 | * like soft offline and memory hotremove don't walk through page | |
1274 | * tables or check whether the hugepage is pmd-based or not before | |
1275 | * kicking migration. | |
1276 | */ | |
1277 | if (!hugepage_migration_supported(page_hstate(hpage))) { | |
1278 | list_move_tail(&hpage->lru, ret); | |
1279 | return -ENOSYS; | |
1280 | } | |
1281 | ||
1282 | new_hpage = get_new_page(hpage, private); | |
1283 | if (!new_hpage) | |
1284 | return -ENOMEM; | |
1285 | ||
1286 | if (!trylock_page(hpage)) { | |
1287 | if (!force) | |
1288 | goto out; | |
1289 | switch (mode) { | |
1290 | case MIGRATE_SYNC: | |
1291 | case MIGRATE_SYNC_NO_COPY: | |
1292 | break; | |
1293 | default: | |
1294 | goto out; | |
1295 | } | |
1296 | lock_page(hpage); | |
1297 | } | |
1298 | ||
1299 | /* | |
1300 | * Check for pages which are in the process of being freed. Without | |
1301 | * page_mapping() set, hugetlbfs specific move page routine will not | |
1302 | * be called and we could leak usage counts for subpools. | |
1303 | */ | |
1304 | if (page_private(hpage) && !page_mapping(hpage)) { | |
1305 | rc = -EBUSY; | |
1306 | goto out_unlock; | |
1307 | } | |
1308 | ||
1309 | if (PageAnon(hpage)) | |
1310 | anon_vma = page_get_anon_vma(hpage); | |
1311 | ||
1312 | if (unlikely(!trylock_page(new_hpage))) | |
1313 | goto put_anon; | |
1314 | ||
1315 | if (page_mapped(hpage)) { | |
1316 | bool mapping_locked = false; | |
1317 | enum ttu_flags ttu = TTU_MIGRATION|TTU_IGNORE_MLOCK; | |
1318 | ||
1319 | if (!PageAnon(hpage)) { | |
1320 | /* | |
1321 | * In shared mappings, try_to_unmap could potentially | |
1322 | * call huge_pmd_unshare. Because of this, take | |
1323 | * semaphore in write mode here and set TTU_RMAP_LOCKED | |
1324 | * to let lower levels know we have taken the lock. | |
1325 | */ | |
1326 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1327 | if (unlikely(!mapping)) | |
1328 | goto unlock_put_anon; | |
1329 | ||
1330 | mapping_locked = true; | |
1331 | ttu |= TTU_RMAP_LOCKED; | |
1332 | } | |
1333 | ||
1334 | try_to_unmap(hpage, ttu); | |
1335 | page_was_mapped = 1; | |
1336 | ||
1337 | if (mapping_locked) | |
1338 | i_mmap_unlock_write(mapping); | |
1339 | } | |
1340 | ||
1341 | if (!page_mapped(hpage)) | |
1342 | rc = move_to_new_page(new_hpage, hpage, mode); | |
1343 | ||
1344 | if (page_was_mapped) | |
1345 | remove_migration_ptes(hpage, | |
1346 | rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); | |
1347 | ||
1348 | unlock_put_anon: | |
1349 | unlock_page(new_hpage); | |
1350 | ||
1351 | put_anon: | |
1352 | if (anon_vma) | |
1353 | put_anon_vma(anon_vma); | |
1354 | ||
1355 | if (rc == MIGRATEPAGE_SUCCESS) { | |
1356 | move_hugetlb_state(hpage, new_hpage, reason); | |
1357 | put_new_page = NULL; | |
1358 | } | |
1359 | ||
1360 | out_unlock: | |
1361 | unlock_page(hpage); | |
1362 | out: | |
1363 | if (rc == MIGRATEPAGE_SUCCESS) | |
1364 | putback_active_hugepage(hpage); | |
1365 | else if (rc != -EAGAIN && rc != MIGRATEPAGE_SUCCESS) | |
1366 | list_move_tail(&hpage->lru, ret); | |
1367 | ||
1368 | /* | |
1369 | * If migration was not successful and there's a freeing callback, use | |
1370 | * it. Otherwise, put_page() will drop the reference grabbed during | |
1371 | * isolation. | |
1372 | */ | |
1373 | if (put_new_page) | |
1374 | put_new_page(new_hpage, private); | |
1375 | else | |
1376 | putback_active_hugepage(new_hpage); | |
1377 | ||
1378 | return rc; | |
1379 | } | |
1380 | ||
1381 | static inline int try_split_thp(struct page *page, struct page **page2, | |
1382 | struct list_head *from) | |
1383 | { | |
1384 | int rc = 0; | |
1385 | ||
1386 | lock_page(page); | |
1387 | rc = split_huge_page_to_list(page, from); | |
1388 | unlock_page(page); | |
1389 | if (!rc) | |
1390 | list_safe_reset_next(page, *page2, lru); | |
1391 | ||
1392 | return rc; | |
1393 | } | |
1394 | ||
1395 | /* | |
1396 | * migrate_pages - migrate the pages specified in a list, to the free pages | |
1397 | * supplied as the target for the page migration | |
1398 | * | |
1399 | * @from: The list of pages to be migrated. | |
1400 | * @get_new_page: The function used to allocate free pages to be used | |
1401 | * as the target of the page migration. | |
1402 | * @put_new_page: The function used to free target pages if migration | |
1403 | * fails, or NULL if no special handling is necessary. | |
1404 | * @private: Private data to be passed on to get_new_page() | |
1405 | * @mode: The migration mode that specifies the constraints for | |
1406 | * page migration, if any. | |
1407 | * @reason: The reason for page migration. | |
1408 | * | |
1409 | * The function returns after 10 attempts or if no pages are movable any more | |
1410 | * because the list has become empty or no retryable pages exist any more. | |
1411 | * It is caller's responsibility to call putback_movable_pages() to return pages | |
1412 | * to the LRU or free list only if ret != 0. | |
1413 | * | |
1414 | * Returns the number of pages that were not migrated, or an error code. | |
1415 | */ | |
1416 | int migrate_pages(struct list_head *from, new_page_t get_new_page, | |
1417 | free_page_t put_new_page, unsigned long private, | |
1418 | enum migrate_mode mode, int reason) | |
1419 | { | |
1420 | int retry = 1; | |
1421 | int thp_retry = 1; | |
1422 | int nr_failed = 0; | |
1423 | int nr_succeeded = 0; | |
1424 | int nr_thp_succeeded = 0; | |
1425 | int nr_thp_failed = 0; | |
1426 | int nr_thp_split = 0; | |
1427 | int pass = 0; | |
1428 | bool is_thp = false; | |
1429 | struct page *page; | |
1430 | struct page *page2; | |
1431 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1432 | int rc, nr_subpages; | |
1433 | LIST_HEAD(ret_pages); | |
1434 | ||
1435 | if (!swapwrite) | |
1436 | current->flags |= PF_SWAPWRITE; | |
1437 | ||
1438 | for (pass = 0; pass < 10 && (retry || thp_retry); pass++) { | |
1439 | retry = 0; | |
1440 | thp_retry = 0; | |
1441 | ||
1442 | list_for_each_entry_safe(page, page2, from, lru) { | |
1443 | retry: | |
1444 | /* | |
1445 | * THP statistics is based on the source huge page. | |
1446 | * Capture required information that might get lost | |
1447 | * during migration. | |
1448 | */ | |
1449 | is_thp = PageTransHuge(page) && !PageHuge(page); | |
1450 | nr_subpages = thp_nr_pages(page); | |
1451 | cond_resched(); | |
1452 | ||
1453 | if (PageHuge(page)) | |
1454 | rc = unmap_and_move_huge_page(get_new_page, | |
1455 | put_new_page, private, page, | |
1456 | pass > 2, mode, reason, | |
1457 | &ret_pages); | |
1458 | else | |
1459 | rc = unmap_and_move(get_new_page, put_new_page, | |
1460 | private, page, pass > 2, mode, | |
1461 | reason, &ret_pages); | |
1462 | /* | |
1463 | * The rules are: | |
1464 | * Success: non hugetlb page will be freed, hugetlb | |
1465 | * page will be put back | |
1466 | * -EAGAIN: stay on the from list | |
1467 | * -ENOMEM: stay on the from list | |
1468 | * Other errno: put on ret_pages list then splice to | |
1469 | * from list | |
1470 | */ | |
1471 | switch(rc) { | |
1472 | /* | |
1473 | * THP migration might be unsupported or the | |
1474 | * allocation could've failed so we should | |
1475 | * retry on the same page with the THP split | |
1476 | * to base pages. | |
1477 | * | |
1478 | * Head page is retried immediately and tail | |
1479 | * pages are added to the tail of the list so | |
1480 | * we encounter them after the rest of the list | |
1481 | * is processed. | |
1482 | */ | |
1483 | case -ENOSYS: | |
1484 | /* THP migration is unsupported */ | |
1485 | if (is_thp) { | |
1486 | if (!try_split_thp(page, &page2, from)) { | |
1487 | nr_thp_split++; | |
1488 | goto retry; | |
1489 | } | |
1490 | ||
1491 | nr_thp_failed++; | |
1492 | nr_failed += nr_subpages; | |
1493 | break; | |
1494 | } | |
1495 | ||
1496 | /* Hugetlb migration is unsupported */ | |
1497 | nr_failed++; | |
1498 | break; | |
1499 | case -ENOMEM: | |
1500 | /* | |
1501 | * When memory is low, don't bother to try to migrate | |
1502 | * other pages, just exit. | |
1503 | */ | |
1504 | if (is_thp) { | |
1505 | if (!try_split_thp(page, &page2, from)) { | |
1506 | nr_thp_split++; | |
1507 | goto retry; | |
1508 | } | |
1509 | ||
1510 | nr_thp_failed++; | |
1511 | nr_failed += nr_subpages; | |
1512 | goto out; | |
1513 | } | |
1514 | nr_failed++; | |
1515 | goto out; | |
1516 | case -EAGAIN: | |
1517 | if (is_thp) { | |
1518 | thp_retry++; | |
1519 | break; | |
1520 | } | |
1521 | retry++; | |
1522 | break; | |
1523 | case MIGRATEPAGE_SUCCESS: | |
1524 | if (is_thp) { | |
1525 | nr_thp_succeeded++; | |
1526 | nr_succeeded += nr_subpages; | |
1527 | break; | |
1528 | } | |
1529 | nr_succeeded++; | |
1530 | break; | |
1531 | default: | |
1532 | /* | |
1533 | * Permanent failure (-EBUSY, etc.): | |
1534 | * unlike -EAGAIN case, the failed page is | |
1535 | * removed from migration page list and not | |
1536 | * retried in the next outer loop. | |
1537 | */ | |
1538 | if (is_thp) { | |
1539 | nr_thp_failed++; | |
1540 | nr_failed += nr_subpages; | |
1541 | break; | |
1542 | } | |
1543 | nr_failed++; | |
1544 | break; | |
1545 | } | |
1546 | } | |
1547 | } | |
1548 | nr_failed += retry + thp_retry; | |
1549 | nr_thp_failed += thp_retry; | |
1550 | rc = nr_failed; | |
1551 | out: | |
1552 | /* | |
1553 | * Put the permanent failure page back to migration list, they | |
1554 | * will be put back to the right list by the caller. | |
1555 | */ | |
1556 | list_splice(&ret_pages, from); | |
1557 | ||
1558 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1559 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
1560 | count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded); | |
1561 | count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed); | |
1562 | count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split); | |
1563 | trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded, | |
1564 | nr_thp_failed, nr_thp_split, mode, reason); | |
1565 | ||
1566 | if (!swapwrite) | |
1567 | current->flags &= ~PF_SWAPWRITE; | |
1568 | ||
1569 | return rc; | |
1570 | } | |
1571 | ||
1572 | struct page *alloc_migration_target(struct page *page, unsigned long private) | |
1573 | { | |
1574 | struct migration_target_control *mtc; | |
1575 | gfp_t gfp_mask; | |
1576 | unsigned int order = 0; | |
1577 | struct page *new_page = NULL; | |
1578 | int nid; | |
1579 | int zidx; | |
1580 | ||
1581 | mtc = (struct migration_target_control *)private; | |
1582 | gfp_mask = mtc->gfp_mask; | |
1583 | nid = mtc->nid; | |
1584 | if (nid == NUMA_NO_NODE) | |
1585 | nid = page_to_nid(page); | |
1586 | ||
1587 | if (PageHuge(page)) { | |
1588 | struct hstate *h = page_hstate(compound_head(page)); | |
1589 | ||
1590 | gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); | |
1591 | return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask); | |
1592 | } | |
1593 | ||
1594 | if (PageTransHuge(page)) { | |
1595 | /* | |
1596 | * clear __GFP_RECLAIM to make the migration callback | |
1597 | * consistent with regular THP allocations. | |
1598 | */ | |
1599 | gfp_mask &= ~__GFP_RECLAIM; | |
1600 | gfp_mask |= GFP_TRANSHUGE; | |
1601 | order = HPAGE_PMD_ORDER; | |
1602 | } | |
1603 | zidx = zone_idx(page_zone(page)); | |
1604 | if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) | |
1605 | gfp_mask |= __GFP_HIGHMEM; | |
1606 | ||
1607 | new_page = __alloc_pages_nodemask(gfp_mask, order, nid, mtc->nmask); | |
1608 | ||
1609 | if (new_page && PageTransHuge(new_page)) | |
1610 | prep_transhuge_page(new_page); | |
1611 | ||
1612 | return new_page; | |
1613 | } | |
1614 | ||
1615 | #ifdef CONFIG_NUMA | |
1616 | ||
1617 | static int store_status(int __user *status, int start, int value, int nr) | |
1618 | { | |
1619 | while (nr-- > 0) { | |
1620 | if (put_user(value, status + start)) | |
1621 | return -EFAULT; | |
1622 | start++; | |
1623 | } | |
1624 | ||
1625 | return 0; | |
1626 | } | |
1627 | ||
1628 | static int do_move_pages_to_node(struct mm_struct *mm, | |
1629 | struct list_head *pagelist, int node) | |
1630 | { | |
1631 | int err; | |
1632 | struct migration_target_control mtc = { | |
1633 | .nid = node, | |
1634 | .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, | |
1635 | }; | |
1636 | ||
1637 | err = migrate_pages(pagelist, alloc_migration_target, NULL, | |
1638 | (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL); | |
1639 | if (err) | |
1640 | putback_movable_pages(pagelist); | |
1641 | return err; | |
1642 | } | |
1643 | ||
1644 | /* | |
1645 | * Resolves the given address to a struct page, isolates it from the LRU and | |
1646 | * puts it to the given pagelist. | |
1647 | * Returns: | |
1648 | * errno - if the page cannot be found/isolated | |
1649 | * 0 - when it doesn't have to be migrated because it is already on the | |
1650 | * target node | |
1651 | * 1 - when it has been queued | |
1652 | */ | |
1653 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, | |
1654 | int node, struct list_head *pagelist, bool migrate_all) | |
1655 | { | |
1656 | struct vm_area_struct *vma; | |
1657 | struct page *page; | |
1658 | unsigned int follflags; | |
1659 | int err; | |
1660 | ||
1661 | mmap_read_lock(mm); | |
1662 | err = -EFAULT; | |
1663 | vma = find_vma(mm, addr); | |
1664 | if (!vma || addr < vma->vm_start || !vma_migratable(vma)) | |
1665 | goto out; | |
1666 | ||
1667 | /* FOLL_DUMP to ignore special (like zero) pages */ | |
1668 | follflags = FOLL_GET | FOLL_DUMP; | |
1669 | page = follow_page(vma, addr, follflags); | |
1670 | ||
1671 | err = PTR_ERR(page); | |
1672 | if (IS_ERR(page)) | |
1673 | goto out; | |
1674 | ||
1675 | err = -ENOENT; | |
1676 | if (!page) | |
1677 | goto out; | |
1678 | ||
1679 | err = 0; | |
1680 | if (page_to_nid(page) == node) | |
1681 | goto out_putpage; | |
1682 | ||
1683 | err = -EACCES; | |
1684 | if (page_mapcount(page) > 1 && !migrate_all) | |
1685 | goto out_putpage; | |
1686 | ||
1687 | if (PageHuge(page)) { | |
1688 | if (PageHead(page)) { | |
1689 | isolate_huge_page(page, pagelist); | |
1690 | err = 1; | |
1691 | } | |
1692 | } else { | |
1693 | struct page *head; | |
1694 | ||
1695 | head = compound_head(page); | |
1696 | err = isolate_lru_page(head); | |
1697 | if (err) | |
1698 | goto out_putpage; | |
1699 | ||
1700 | err = 1; | |
1701 | list_add_tail(&head->lru, pagelist); | |
1702 | mod_node_page_state(page_pgdat(head), | |
1703 | NR_ISOLATED_ANON + page_is_file_lru(head), | |
1704 | thp_nr_pages(head)); | |
1705 | } | |
1706 | out_putpage: | |
1707 | /* | |
1708 | * Either remove the duplicate refcount from | |
1709 | * isolate_lru_page() or drop the page ref if it was | |
1710 | * not isolated. | |
1711 | */ | |
1712 | put_page(page); | |
1713 | out: | |
1714 | mmap_read_unlock(mm); | |
1715 | return err; | |
1716 | } | |
1717 | ||
1718 | static int move_pages_and_store_status(struct mm_struct *mm, int node, | |
1719 | struct list_head *pagelist, int __user *status, | |
1720 | int start, int i, unsigned long nr_pages) | |
1721 | { | |
1722 | int err; | |
1723 | ||
1724 | if (list_empty(pagelist)) | |
1725 | return 0; | |
1726 | ||
1727 | err = do_move_pages_to_node(mm, pagelist, node); | |
1728 | if (err) { | |
1729 | /* | |
1730 | * Positive err means the number of failed | |
1731 | * pages to migrate. Since we are going to | |
1732 | * abort and return the number of non-migrated | |
1733 | * pages, so need to include the rest of the | |
1734 | * nr_pages that have not been attempted as | |
1735 | * well. | |
1736 | */ | |
1737 | if (err > 0) | |
1738 | err += nr_pages - i - 1; | |
1739 | return err; | |
1740 | } | |
1741 | return store_status(status, start, node, i - start); | |
1742 | } | |
1743 | ||
1744 | /* | |
1745 | * Migrate an array of page address onto an array of nodes and fill | |
1746 | * the corresponding array of status. | |
1747 | */ | |
1748 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, | |
1749 | unsigned long nr_pages, | |
1750 | const void __user * __user *pages, | |
1751 | const int __user *nodes, | |
1752 | int __user *status, int flags) | |
1753 | { | |
1754 | int current_node = NUMA_NO_NODE; | |
1755 | LIST_HEAD(pagelist); | |
1756 | int start, i; | |
1757 | int err = 0, err1; | |
1758 | ||
1759 | migrate_prep(); | |
1760 | ||
1761 | for (i = start = 0; i < nr_pages; i++) { | |
1762 | const void __user *p; | |
1763 | unsigned long addr; | |
1764 | int node; | |
1765 | ||
1766 | err = -EFAULT; | |
1767 | if (get_user(p, pages + i)) | |
1768 | goto out_flush; | |
1769 | if (get_user(node, nodes + i)) | |
1770 | goto out_flush; | |
1771 | addr = (unsigned long)untagged_addr(p); | |
1772 | ||
1773 | err = -ENODEV; | |
1774 | if (node < 0 || node >= MAX_NUMNODES) | |
1775 | goto out_flush; | |
1776 | if (!node_state(node, N_MEMORY)) | |
1777 | goto out_flush; | |
1778 | ||
1779 | err = -EACCES; | |
1780 | if (!node_isset(node, task_nodes)) | |
1781 | goto out_flush; | |
1782 | ||
1783 | if (current_node == NUMA_NO_NODE) { | |
1784 | current_node = node; | |
1785 | start = i; | |
1786 | } else if (node != current_node) { | |
1787 | err = move_pages_and_store_status(mm, current_node, | |
1788 | &pagelist, status, start, i, nr_pages); | |
1789 | if (err) | |
1790 | goto out; | |
1791 | start = i; | |
1792 | current_node = node; | |
1793 | } | |
1794 | ||
1795 | /* | |
1796 | * Errors in the page lookup or isolation are not fatal and we simply | |
1797 | * report them via status | |
1798 | */ | |
1799 | err = add_page_for_migration(mm, addr, current_node, | |
1800 | &pagelist, flags & MPOL_MF_MOVE_ALL); | |
1801 | ||
1802 | if (err > 0) { | |
1803 | /* The page is successfully queued for migration */ | |
1804 | continue; | |
1805 | } | |
1806 | ||
1807 | /* | |
1808 | * If the page is already on the target node (!err), store the | |
1809 | * node, otherwise, store the err. | |
1810 | */ | |
1811 | err = store_status(status, i, err ? : current_node, 1); | |
1812 | if (err) | |
1813 | goto out_flush; | |
1814 | ||
1815 | err = move_pages_and_store_status(mm, current_node, &pagelist, | |
1816 | status, start, i, nr_pages); | |
1817 | if (err) | |
1818 | goto out; | |
1819 | current_node = NUMA_NO_NODE; | |
1820 | } | |
1821 | out_flush: | |
1822 | /* Make sure we do not overwrite the existing error */ | |
1823 | err1 = move_pages_and_store_status(mm, current_node, &pagelist, | |
1824 | status, start, i, nr_pages); | |
1825 | if (err >= 0) | |
1826 | err = err1; | |
1827 | out: | |
1828 | return err; | |
1829 | } | |
1830 | ||
1831 | /* | |
1832 | * Determine the nodes of an array of pages and store it in an array of status. | |
1833 | */ | |
1834 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, | |
1835 | const void __user **pages, int *status) | |
1836 | { | |
1837 | unsigned long i; | |
1838 | ||
1839 | mmap_read_lock(mm); | |
1840 | ||
1841 | for (i = 0; i < nr_pages; i++) { | |
1842 | unsigned long addr = (unsigned long)(*pages); | |
1843 | struct vm_area_struct *vma; | |
1844 | struct page *page; | |
1845 | int err = -EFAULT; | |
1846 | ||
1847 | vma = find_vma(mm, addr); | |
1848 | if (!vma || addr < vma->vm_start) | |
1849 | goto set_status; | |
1850 | ||
1851 | /* FOLL_DUMP to ignore special (like zero) pages */ | |
1852 | page = follow_page(vma, addr, FOLL_DUMP); | |
1853 | ||
1854 | err = PTR_ERR(page); | |
1855 | if (IS_ERR(page)) | |
1856 | goto set_status; | |
1857 | ||
1858 | err = page ? page_to_nid(page) : -ENOENT; | |
1859 | set_status: | |
1860 | *status = err; | |
1861 | ||
1862 | pages++; | |
1863 | status++; | |
1864 | } | |
1865 | ||
1866 | mmap_read_unlock(mm); | |
1867 | } | |
1868 | ||
1869 | /* | |
1870 | * Determine the nodes of a user array of pages and store it in | |
1871 | * a user array of status. | |
1872 | */ | |
1873 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1874 | const void __user * __user *pages, | |
1875 | int __user *status) | |
1876 | { | |
1877 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1878 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1879 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
1880 | ||
1881 | while (nr_pages) { | |
1882 | unsigned long chunk_nr; | |
1883 | ||
1884 | chunk_nr = nr_pages; | |
1885 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1886 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1887 | ||
1888 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1889 | break; | |
1890 | ||
1891 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1892 | ||
1893 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) | |
1894 | break; | |
1895 | ||
1896 | pages += chunk_nr; | |
1897 | status += chunk_nr; | |
1898 | nr_pages -= chunk_nr; | |
1899 | } | |
1900 | return nr_pages ? -EFAULT : 0; | |
1901 | } | |
1902 | ||
1903 | static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) | |
1904 | { | |
1905 | struct task_struct *task; | |
1906 | struct mm_struct *mm; | |
1907 | ||
1908 | /* | |
1909 | * There is no need to check if current process has the right to modify | |
1910 | * the specified process when they are same. | |
1911 | */ | |
1912 | if (!pid) { | |
1913 | mmget(current->mm); | |
1914 | *mem_nodes = cpuset_mems_allowed(current); | |
1915 | return current->mm; | |
1916 | } | |
1917 | ||
1918 | /* Find the mm_struct */ | |
1919 | rcu_read_lock(); | |
1920 | task = find_task_by_vpid(pid); | |
1921 | if (!task) { | |
1922 | rcu_read_unlock(); | |
1923 | return ERR_PTR(-ESRCH); | |
1924 | } | |
1925 | get_task_struct(task); | |
1926 | ||
1927 | /* | |
1928 | * Check if this process has the right to modify the specified | |
1929 | * process. Use the regular "ptrace_may_access()" checks. | |
1930 | */ | |
1931 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { | |
1932 | rcu_read_unlock(); | |
1933 | mm = ERR_PTR(-EPERM); | |
1934 | goto out; | |
1935 | } | |
1936 | rcu_read_unlock(); | |
1937 | ||
1938 | mm = ERR_PTR(security_task_movememory(task)); | |
1939 | if (IS_ERR(mm)) | |
1940 | goto out; | |
1941 | *mem_nodes = cpuset_mems_allowed(task); | |
1942 | mm = get_task_mm(task); | |
1943 | out: | |
1944 | put_task_struct(task); | |
1945 | if (!mm) | |
1946 | mm = ERR_PTR(-EINVAL); | |
1947 | return mm; | |
1948 | } | |
1949 | ||
1950 | /* | |
1951 | * Move a list of pages in the address space of the currently executing | |
1952 | * process. | |
1953 | */ | |
1954 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, | |
1955 | const void __user * __user *pages, | |
1956 | const int __user *nodes, | |
1957 | int __user *status, int flags) | |
1958 | { | |
1959 | struct mm_struct *mm; | |
1960 | int err; | |
1961 | nodemask_t task_nodes; | |
1962 | ||
1963 | /* Check flags */ | |
1964 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1965 | return -EINVAL; | |
1966 | ||
1967 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1968 | return -EPERM; | |
1969 | ||
1970 | mm = find_mm_struct(pid, &task_nodes); | |
1971 | if (IS_ERR(mm)) | |
1972 | return PTR_ERR(mm); | |
1973 | ||
1974 | if (nodes) | |
1975 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1976 | nodes, status, flags); | |
1977 | else | |
1978 | err = do_pages_stat(mm, nr_pages, pages, status); | |
1979 | ||
1980 | mmput(mm); | |
1981 | return err; | |
1982 | } | |
1983 | ||
1984 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, | |
1985 | const void __user * __user *, pages, | |
1986 | const int __user *, nodes, | |
1987 | int __user *, status, int, flags) | |
1988 | { | |
1989 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
1990 | } | |
1991 | ||
1992 | #ifdef CONFIG_COMPAT | |
1993 | COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages, | |
1994 | compat_uptr_t __user *, pages32, | |
1995 | const int __user *, nodes, | |
1996 | int __user *, status, | |
1997 | int, flags) | |
1998 | { | |
1999 | const void __user * __user *pages; | |
2000 | int i; | |
2001 | ||
2002 | pages = compat_alloc_user_space(nr_pages * sizeof(void *)); | |
2003 | for (i = 0; i < nr_pages; i++) { | |
2004 | compat_uptr_t p; | |
2005 | ||
2006 | if (get_user(p, pages32 + i) || | |
2007 | put_user(compat_ptr(p), pages + i)) | |
2008 | return -EFAULT; | |
2009 | } | |
2010 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); | |
2011 | } | |
2012 | #endif /* CONFIG_COMPAT */ | |
2013 | ||
2014 | #ifdef CONFIG_NUMA_BALANCING | |
2015 | /* | |
2016 | * Returns true if this is a safe migration target node for misplaced NUMA | |
2017 | * pages. Currently it only checks the watermarks which crude | |
2018 | */ | |
2019 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
2020 | unsigned long nr_migrate_pages) | |
2021 | { | |
2022 | int z; | |
2023 | ||
2024 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
2025 | struct zone *zone = pgdat->node_zones + z; | |
2026 | ||
2027 | if (!populated_zone(zone)) | |
2028 | continue; | |
2029 | ||
2030 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
2031 | if (!zone_watermark_ok(zone, 0, | |
2032 | high_wmark_pages(zone) + | |
2033 | nr_migrate_pages, | |
2034 | ZONE_MOVABLE, 0)) | |
2035 | continue; | |
2036 | return true; | |
2037 | } | |
2038 | return false; | |
2039 | } | |
2040 | ||
2041 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
2042 | unsigned long data) | |
2043 | { | |
2044 | int nid = (int) data; | |
2045 | struct page *newpage; | |
2046 | ||
2047 | newpage = __alloc_pages_node(nid, | |
2048 | (GFP_HIGHUSER_MOVABLE | | |
2049 | __GFP_THISNODE | __GFP_NOMEMALLOC | | |
2050 | __GFP_NORETRY | __GFP_NOWARN) & | |
2051 | ~__GFP_RECLAIM, 0); | |
2052 | ||
2053 | return newpage; | |
2054 | } | |
2055 | ||
2056 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) | |
2057 | { | |
2058 | int page_lru; | |
2059 | ||
2060 | VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); | |
2061 | ||
2062 | /* Avoid migrating to a node that is nearly full */ | |
2063 | if (!migrate_balanced_pgdat(pgdat, compound_nr(page))) | |
2064 | return 0; | |
2065 | ||
2066 | if (isolate_lru_page(page)) | |
2067 | return 0; | |
2068 | ||
2069 | /* | |
2070 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
2071 | * check on page_count(), so we must do it here, now that the page | |
2072 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
2073 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
2074 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
2075 | */ | |
2076 | if (PageTransHuge(page) && page_count(page) != 3) { | |
2077 | putback_lru_page(page); | |
2078 | return 0; | |
2079 | } | |
2080 | ||
2081 | page_lru = page_is_file_lru(page); | |
2082 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, | |
2083 | thp_nr_pages(page)); | |
2084 | ||
2085 | /* | |
2086 | * Isolating the page has taken another reference, so the | |
2087 | * caller's reference can be safely dropped without the page | |
2088 | * disappearing underneath us during migration. | |
2089 | */ | |
2090 | put_page(page); | |
2091 | return 1; | |
2092 | } | |
2093 | ||
2094 | bool pmd_trans_migrating(pmd_t pmd) | |
2095 | { | |
2096 | struct page *page = pmd_page(pmd); | |
2097 | return PageLocked(page); | |
2098 | } | |
2099 | ||
2100 | static inline bool is_shared_exec_page(struct vm_area_struct *vma, | |
2101 | struct page *page) | |
2102 | { | |
2103 | if (page_mapcount(page) != 1 && | |
2104 | (page_is_file_lru(page) || vma_is_shmem(vma)) && | |
2105 | (vma->vm_flags & VM_EXEC)) | |
2106 | return true; | |
2107 | ||
2108 | return false; | |
2109 | } | |
2110 | ||
2111 | /* | |
2112 | * Attempt to migrate a misplaced page to the specified destination | |
2113 | * node. Caller is expected to have an elevated reference count on | |
2114 | * the page that will be dropped by this function before returning. | |
2115 | */ | |
2116 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, | |
2117 | int node) | |
2118 | { | |
2119 | pg_data_t *pgdat = NODE_DATA(node); | |
2120 | int isolated; | |
2121 | int nr_remaining; | |
2122 | LIST_HEAD(migratepages); | |
2123 | ||
2124 | /* | |
2125 | * Don't migrate file pages that are mapped in multiple processes | |
2126 | * with execute permissions as they are probably shared libraries. | |
2127 | */ | |
2128 | if (is_shared_exec_page(vma, page)) | |
2129 | goto out; | |
2130 | ||
2131 | /* | |
2132 | * Also do not migrate dirty pages as not all filesystems can move | |
2133 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. | |
2134 | */ | |
2135 | if (page_is_file_lru(page) && PageDirty(page)) | |
2136 | goto out; | |
2137 | ||
2138 | isolated = numamigrate_isolate_page(pgdat, page); | |
2139 | if (!isolated) | |
2140 | goto out; | |
2141 | ||
2142 | list_add(&page->lru, &migratepages); | |
2143 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, | |
2144 | NULL, node, MIGRATE_ASYNC, | |
2145 | MR_NUMA_MISPLACED); | |
2146 | if (nr_remaining) { | |
2147 | if (!list_empty(&migratepages)) { | |
2148 | list_del(&page->lru); | |
2149 | dec_node_page_state(page, NR_ISOLATED_ANON + | |
2150 | page_is_file_lru(page)); | |
2151 | putback_lru_page(page); | |
2152 | } | |
2153 | isolated = 0; | |
2154 | } else | |
2155 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
2156 | BUG_ON(!list_empty(&migratepages)); | |
2157 | return isolated; | |
2158 | ||
2159 | out: | |
2160 | put_page(page); | |
2161 | return 0; | |
2162 | } | |
2163 | #endif /* CONFIG_NUMA_BALANCING */ | |
2164 | ||
2165 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) | |
2166 | /* | |
2167 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
2168 | * before returning. | |
2169 | */ | |
2170 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, | |
2171 | struct vm_area_struct *vma, | |
2172 | pmd_t *pmd, pmd_t entry, | |
2173 | unsigned long address, | |
2174 | struct page *page, int node) | |
2175 | { | |
2176 | spinlock_t *ptl; | |
2177 | pg_data_t *pgdat = NODE_DATA(node); | |
2178 | int isolated = 0; | |
2179 | struct page *new_page = NULL; | |
2180 | int page_lru = page_is_file_lru(page); | |
2181 | unsigned long start = address & HPAGE_PMD_MASK; | |
2182 | ||
2183 | if (is_shared_exec_page(vma, page)) | |
2184 | goto out; | |
2185 | ||
2186 | new_page = alloc_pages_node(node, | |
2187 | (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), | |
2188 | HPAGE_PMD_ORDER); | |
2189 | if (!new_page) | |
2190 | goto out_fail; | |
2191 | prep_transhuge_page(new_page); | |
2192 | ||
2193 | isolated = numamigrate_isolate_page(pgdat, page); | |
2194 | if (!isolated) { | |
2195 | put_page(new_page); | |
2196 | goto out_fail; | |
2197 | } | |
2198 | ||
2199 | /* Prepare a page as a migration target */ | |
2200 | __SetPageLocked(new_page); | |
2201 | if (PageSwapBacked(page)) | |
2202 | __SetPageSwapBacked(new_page); | |
2203 | ||
2204 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
2205 | new_page->mapping = page->mapping; | |
2206 | new_page->index = page->index; | |
2207 | /* flush the cache before copying using the kernel virtual address */ | |
2208 | flush_cache_range(vma, start, start + HPAGE_PMD_SIZE); | |
2209 | migrate_page_copy(new_page, page); | |
2210 | WARN_ON(PageLRU(new_page)); | |
2211 | ||
2212 | /* Recheck the target PMD */ | |
2213 | ptl = pmd_lock(mm, pmd); | |
2214 | if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) { | |
2215 | spin_unlock(ptl); | |
2216 | ||
2217 | /* Reverse changes made by migrate_page_copy() */ | |
2218 | if (TestClearPageActive(new_page)) | |
2219 | SetPageActive(page); | |
2220 | if (TestClearPageUnevictable(new_page)) | |
2221 | SetPageUnevictable(page); | |
2222 | ||
2223 | unlock_page(new_page); | |
2224 | put_page(new_page); /* Free it */ | |
2225 | ||
2226 | /* Retake the callers reference and putback on LRU */ | |
2227 | get_page(page); | |
2228 | putback_lru_page(page); | |
2229 | mod_node_page_state(page_pgdat(page), | |
2230 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
2231 | ||
2232 | goto out_unlock; | |
2233 | } | |
2234 | ||
2235 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); | |
2236 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
2237 | ||
2238 | /* | |
2239 | * Overwrite the old entry under pagetable lock and establish | |
2240 | * the new PTE. Any parallel GUP will either observe the old | |
2241 | * page blocking on the page lock, block on the page table | |
2242 | * lock or observe the new page. The SetPageUptodate on the | |
2243 | * new page and page_add_new_anon_rmap guarantee the copy is | |
2244 | * visible before the pagetable update. | |
2245 | */ | |
2246 | page_add_anon_rmap(new_page, vma, start, true); | |
2247 | /* | |
2248 | * At this point the pmd is numa/protnone (i.e. non present) and the TLB | |
2249 | * has already been flushed globally. So no TLB can be currently | |
2250 | * caching this non present pmd mapping. There's no need to clear the | |
2251 | * pmd before doing set_pmd_at(), nor to flush the TLB after | |
2252 | * set_pmd_at(). Clearing the pmd here would introduce a race | |
2253 | * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the | |
2254 | * mmap_lock for reading. If the pmd is set to NULL at any given time, | |
2255 | * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this | |
2256 | * pmd. | |
2257 | */ | |
2258 | set_pmd_at(mm, start, pmd, entry); | |
2259 | update_mmu_cache_pmd(vma, address, &entry); | |
2260 | ||
2261 | page_ref_unfreeze(page, 2); | |
2262 | mlock_migrate_page(new_page, page); | |
2263 | page_remove_rmap(page, true); | |
2264 | set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); | |
2265 | ||
2266 | spin_unlock(ptl); | |
2267 | ||
2268 | /* Take an "isolate" reference and put new page on the LRU. */ | |
2269 | get_page(new_page); | |
2270 | putback_lru_page(new_page); | |
2271 | ||
2272 | unlock_page(new_page); | |
2273 | unlock_page(page); | |
2274 | put_page(page); /* Drop the rmap reference */ | |
2275 | put_page(page); /* Drop the LRU isolation reference */ | |
2276 | ||
2277 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
2278 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
2279 | ||
2280 | mod_node_page_state(page_pgdat(page), | |
2281 | NR_ISOLATED_ANON + page_lru, | |
2282 | -HPAGE_PMD_NR); | |
2283 | return isolated; | |
2284 | ||
2285 | out_fail: | |
2286 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
2287 | ptl = pmd_lock(mm, pmd); | |
2288 | if (pmd_same(*pmd, entry)) { | |
2289 | entry = pmd_modify(entry, vma->vm_page_prot); | |
2290 | set_pmd_at(mm, start, pmd, entry); | |
2291 | update_mmu_cache_pmd(vma, address, &entry); | |
2292 | } | |
2293 | spin_unlock(ptl); | |
2294 | ||
2295 | out_unlock: | |
2296 | unlock_page(page); | |
2297 | out: | |
2298 | put_page(page); | |
2299 | return 0; | |
2300 | } | |
2301 | #endif /* CONFIG_NUMA_BALANCING */ | |
2302 | ||
2303 | #endif /* CONFIG_NUMA */ | |
2304 | ||
2305 | #ifdef CONFIG_DEVICE_PRIVATE | |
2306 | static int migrate_vma_collect_hole(unsigned long start, | |
2307 | unsigned long end, | |
2308 | __always_unused int depth, | |
2309 | struct mm_walk *walk) | |
2310 | { | |
2311 | struct migrate_vma *migrate = walk->private; | |
2312 | unsigned long addr; | |
2313 | ||
2314 | /* Only allow populating anonymous memory. */ | |
2315 | if (!vma_is_anonymous(walk->vma)) { | |
2316 | for (addr = start; addr < end; addr += PAGE_SIZE) { | |
2317 | migrate->src[migrate->npages] = 0; | |
2318 | migrate->dst[migrate->npages] = 0; | |
2319 | migrate->npages++; | |
2320 | } | |
2321 | return 0; | |
2322 | } | |
2323 | ||
2324 | for (addr = start; addr < end; addr += PAGE_SIZE) { | |
2325 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; | |
2326 | migrate->dst[migrate->npages] = 0; | |
2327 | migrate->npages++; | |
2328 | migrate->cpages++; | |
2329 | } | |
2330 | ||
2331 | return 0; | |
2332 | } | |
2333 | ||
2334 | static int migrate_vma_collect_skip(unsigned long start, | |
2335 | unsigned long end, | |
2336 | struct mm_walk *walk) | |
2337 | { | |
2338 | struct migrate_vma *migrate = walk->private; | |
2339 | unsigned long addr; | |
2340 | ||
2341 | for (addr = start; addr < end; addr += PAGE_SIZE) { | |
2342 | migrate->dst[migrate->npages] = 0; | |
2343 | migrate->src[migrate->npages++] = 0; | |
2344 | } | |
2345 | ||
2346 | return 0; | |
2347 | } | |
2348 | ||
2349 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
2350 | unsigned long start, | |
2351 | unsigned long end, | |
2352 | struct mm_walk *walk) | |
2353 | { | |
2354 | struct migrate_vma *migrate = walk->private; | |
2355 | struct vm_area_struct *vma = walk->vma; | |
2356 | struct mm_struct *mm = vma->vm_mm; | |
2357 | unsigned long addr = start, unmapped = 0; | |
2358 | spinlock_t *ptl; | |
2359 | pte_t *ptep; | |
2360 | ||
2361 | again: | |
2362 | if (pmd_none(*pmdp)) | |
2363 | return migrate_vma_collect_hole(start, end, -1, walk); | |
2364 | ||
2365 | if (pmd_trans_huge(*pmdp)) { | |
2366 | struct page *page; | |
2367 | ||
2368 | ptl = pmd_lock(mm, pmdp); | |
2369 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
2370 | spin_unlock(ptl); | |
2371 | goto again; | |
2372 | } | |
2373 | ||
2374 | page = pmd_page(*pmdp); | |
2375 | if (is_huge_zero_page(page)) { | |
2376 | spin_unlock(ptl); | |
2377 | split_huge_pmd(vma, pmdp, addr); | |
2378 | if (pmd_trans_unstable(pmdp)) | |
2379 | return migrate_vma_collect_skip(start, end, | |
2380 | walk); | |
2381 | } else { | |
2382 | int ret; | |
2383 | ||
2384 | get_page(page); | |
2385 | spin_unlock(ptl); | |
2386 | if (unlikely(!trylock_page(page))) | |
2387 | return migrate_vma_collect_skip(start, end, | |
2388 | walk); | |
2389 | ret = split_huge_page(page); | |
2390 | unlock_page(page); | |
2391 | put_page(page); | |
2392 | if (ret) | |
2393 | return migrate_vma_collect_skip(start, end, | |
2394 | walk); | |
2395 | if (pmd_none(*pmdp)) | |
2396 | return migrate_vma_collect_hole(start, end, -1, | |
2397 | walk); | |
2398 | } | |
2399 | } | |
2400 | ||
2401 | if (unlikely(pmd_bad(*pmdp))) | |
2402 | return migrate_vma_collect_skip(start, end, walk); | |
2403 | ||
2404 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
2405 | arch_enter_lazy_mmu_mode(); | |
2406 | ||
2407 | for (; addr < end; addr += PAGE_SIZE, ptep++) { | |
2408 | unsigned long mpfn = 0, pfn; | |
2409 | struct page *page; | |
2410 | swp_entry_t entry; | |
2411 | pte_t pte; | |
2412 | ||
2413 | pte = *ptep; | |
2414 | ||
2415 | if (pte_none(pte)) { | |
2416 | if (vma_is_anonymous(vma)) { | |
2417 | mpfn = MIGRATE_PFN_MIGRATE; | |
2418 | migrate->cpages++; | |
2419 | } | |
2420 | goto next; | |
2421 | } | |
2422 | ||
2423 | if (!pte_present(pte)) { | |
2424 | /* | |
2425 | * Only care about unaddressable device page special | |
2426 | * page table entry. Other special swap entries are not | |
2427 | * migratable, and we ignore regular swapped page. | |
2428 | */ | |
2429 | entry = pte_to_swp_entry(pte); | |
2430 | if (!is_device_private_entry(entry)) | |
2431 | goto next; | |
2432 | ||
2433 | page = device_private_entry_to_page(entry); | |
2434 | if (!(migrate->flags & | |
2435 | MIGRATE_VMA_SELECT_DEVICE_PRIVATE) || | |
2436 | page->pgmap->owner != migrate->pgmap_owner) | |
2437 | goto next; | |
2438 | ||
2439 | mpfn = migrate_pfn(page_to_pfn(page)) | | |
2440 | MIGRATE_PFN_MIGRATE; | |
2441 | if (is_write_device_private_entry(entry)) | |
2442 | mpfn |= MIGRATE_PFN_WRITE; | |
2443 | } else { | |
2444 | if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) | |
2445 | goto next; | |
2446 | pfn = pte_pfn(pte); | |
2447 | if (is_zero_pfn(pfn)) { | |
2448 | mpfn = MIGRATE_PFN_MIGRATE; | |
2449 | migrate->cpages++; | |
2450 | goto next; | |
2451 | } | |
2452 | page = vm_normal_page(migrate->vma, addr, pte); | |
2453 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; | |
2454 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
2455 | } | |
2456 | ||
2457 | /* FIXME support THP */ | |
2458 | if (!page || !page->mapping || PageTransCompound(page)) { | |
2459 | mpfn = 0; | |
2460 | goto next; | |
2461 | } | |
2462 | ||
2463 | /* | |
2464 | * By getting a reference on the page we pin it and that blocks | |
2465 | * any kind of migration. Side effect is that it "freezes" the | |
2466 | * pte. | |
2467 | * | |
2468 | * We drop this reference after isolating the page from the lru | |
2469 | * for non device page (device page are not on the lru and thus | |
2470 | * can't be dropped from it). | |
2471 | */ | |
2472 | get_page(page); | |
2473 | migrate->cpages++; | |
2474 | ||
2475 | /* | |
2476 | * Optimize for the common case where page is only mapped once | |
2477 | * in one process. If we can lock the page, then we can safely | |
2478 | * set up a special migration page table entry now. | |
2479 | */ | |
2480 | if (trylock_page(page)) { | |
2481 | pte_t swp_pte; | |
2482 | ||
2483 | mpfn |= MIGRATE_PFN_LOCKED; | |
2484 | ptep_get_and_clear(mm, addr, ptep); | |
2485 | ||
2486 | /* Setup special migration page table entry */ | |
2487 | entry = make_migration_entry(page, mpfn & | |
2488 | MIGRATE_PFN_WRITE); | |
2489 | swp_pte = swp_entry_to_pte(entry); | |
2490 | if (pte_present(pte)) { | |
2491 | if (pte_soft_dirty(pte)) | |
2492 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2493 | if (pte_uffd_wp(pte)) | |
2494 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2495 | } else { | |
2496 | if (pte_swp_soft_dirty(pte)) | |
2497 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2498 | if (pte_swp_uffd_wp(pte)) | |
2499 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2500 | } | |
2501 | set_pte_at(mm, addr, ptep, swp_pte); | |
2502 | ||
2503 | /* | |
2504 | * This is like regular unmap: we remove the rmap and | |
2505 | * drop page refcount. Page won't be freed, as we took | |
2506 | * a reference just above. | |
2507 | */ | |
2508 | page_remove_rmap(page, false); | |
2509 | put_page(page); | |
2510 | ||
2511 | if (pte_present(pte)) | |
2512 | unmapped++; | |
2513 | } | |
2514 | ||
2515 | next: | |
2516 | migrate->dst[migrate->npages] = 0; | |
2517 | migrate->src[migrate->npages++] = mpfn; | |
2518 | } | |
2519 | arch_leave_lazy_mmu_mode(); | |
2520 | pte_unmap_unlock(ptep - 1, ptl); | |
2521 | ||
2522 | /* Only flush the TLB if we actually modified any entries */ | |
2523 | if (unmapped) | |
2524 | flush_tlb_range(walk->vma, start, end); | |
2525 | ||
2526 | return 0; | |
2527 | } | |
2528 | ||
2529 | static const struct mm_walk_ops migrate_vma_walk_ops = { | |
2530 | .pmd_entry = migrate_vma_collect_pmd, | |
2531 | .pte_hole = migrate_vma_collect_hole, | |
2532 | }; | |
2533 | ||
2534 | /* | |
2535 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
2536 | * @migrate: migrate struct containing all migration information | |
2537 | * | |
2538 | * This will walk the CPU page table. For each virtual address backed by a | |
2539 | * valid page, it updates the src array and takes a reference on the page, in | |
2540 | * order to pin the page until we lock it and unmap it. | |
2541 | */ | |
2542 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
2543 | { | |
2544 | struct mmu_notifier_range range; | |
2545 | ||
2546 | /* | |
2547 | * Note that the pgmap_owner is passed to the mmu notifier callback so | |
2548 | * that the registered device driver can skip invalidating device | |
2549 | * private page mappings that won't be migrated. | |
2550 | */ | |
2551 | mmu_notifier_range_init_migrate(&range, 0, migrate->vma, | |
2552 | migrate->vma->vm_mm, migrate->start, migrate->end, | |
2553 | migrate->pgmap_owner); | |
2554 | mmu_notifier_invalidate_range_start(&range); | |
2555 | ||
2556 | walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, | |
2557 | &migrate_vma_walk_ops, migrate); | |
2558 | ||
2559 | mmu_notifier_invalidate_range_end(&range); | |
2560 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); | |
2561 | } | |
2562 | ||
2563 | /* | |
2564 | * migrate_vma_check_page() - check if page is pinned or not | |
2565 | * @page: struct page to check | |
2566 | * | |
2567 | * Pinned pages cannot be migrated. This is the same test as in | |
2568 | * migrate_page_move_mapping(), except that here we allow migration of a | |
2569 | * ZONE_DEVICE page. | |
2570 | */ | |
2571 | static bool migrate_vma_check_page(struct page *page) | |
2572 | { | |
2573 | /* | |
2574 | * One extra ref because caller holds an extra reference, either from | |
2575 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
2576 | * a device page. | |
2577 | */ | |
2578 | int extra = 1; | |
2579 | ||
2580 | /* | |
2581 | * FIXME support THP (transparent huge page), it is bit more complex to | |
2582 | * check them than regular pages, because they can be mapped with a pmd | |
2583 | * or with a pte (split pte mapping). | |
2584 | */ | |
2585 | if (PageCompound(page)) | |
2586 | return false; | |
2587 | ||
2588 | /* Page from ZONE_DEVICE have one extra reference */ | |
2589 | if (is_zone_device_page(page)) { | |
2590 | /* | |
2591 | * Private page can never be pin as they have no valid pte and | |
2592 | * GUP will fail for those. Yet if there is a pending migration | |
2593 | * a thread might try to wait on the pte migration entry and | |
2594 | * will bump the page reference count. Sadly there is no way to | |
2595 | * differentiate a regular pin from migration wait. Hence to | |
2596 | * avoid 2 racing thread trying to migrate back to CPU to enter | |
2597 | * infinite loop (one stoping migration because the other is | |
2598 | * waiting on pte migration entry). We always return true here. | |
2599 | * | |
2600 | * FIXME proper solution is to rework migration_entry_wait() so | |
2601 | * it does not need to take a reference on page. | |
2602 | */ | |
2603 | return is_device_private_page(page); | |
2604 | } | |
2605 | ||
2606 | /* For file back page */ | |
2607 | if (page_mapping(page)) | |
2608 | extra += 1 + page_has_private(page); | |
2609 | ||
2610 | if ((page_count(page) - extra) > page_mapcount(page)) | |
2611 | return false; | |
2612 | ||
2613 | return true; | |
2614 | } | |
2615 | ||
2616 | /* | |
2617 | * migrate_vma_prepare() - lock pages and isolate them from the lru | |
2618 | * @migrate: migrate struct containing all migration information | |
2619 | * | |
2620 | * This locks pages that have been collected by migrate_vma_collect(). Once each | |
2621 | * page is locked it is isolated from the lru (for non-device pages). Finally, | |
2622 | * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be | |
2623 | * migrated by concurrent kernel threads. | |
2624 | */ | |
2625 | static void migrate_vma_prepare(struct migrate_vma *migrate) | |
2626 | { | |
2627 | const unsigned long npages = migrate->npages; | |
2628 | const unsigned long start = migrate->start; | |
2629 | unsigned long addr, i, restore = 0; | |
2630 | bool allow_drain = true; | |
2631 | ||
2632 | lru_add_drain(); | |
2633 | ||
2634 | for (i = 0; (i < npages) && migrate->cpages; i++) { | |
2635 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2636 | bool remap = true; | |
2637 | ||
2638 | if (!page) | |
2639 | continue; | |
2640 | ||
2641 | if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { | |
2642 | /* | |
2643 | * Because we are migrating several pages there can be | |
2644 | * a deadlock between 2 concurrent migration where each | |
2645 | * are waiting on each other page lock. | |
2646 | * | |
2647 | * Make migrate_vma() a best effort thing and backoff | |
2648 | * for any page we can not lock right away. | |
2649 | */ | |
2650 | if (!trylock_page(page)) { | |
2651 | migrate->src[i] = 0; | |
2652 | migrate->cpages--; | |
2653 | put_page(page); | |
2654 | continue; | |
2655 | } | |
2656 | remap = false; | |
2657 | migrate->src[i] |= MIGRATE_PFN_LOCKED; | |
2658 | } | |
2659 | ||
2660 | /* ZONE_DEVICE pages are not on LRU */ | |
2661 | if (!is_zone_device_page(page)) { | |
2662 | if (!PageLRU(page) && allow_drain) { | |
2663 | /* Drain CPU's pagevec */ | |
2664 | lru_add_drain_all(); | |
2665 | allow_drain = false; | |
2666 | } | |
2667 | ||
2668 | if (isolate_lru_page(page)) { | |
2669 | if (remap) { | |
2670 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2671 | migrate->cpages--; | |
2672 | restore++; | |
2673 | } else { | |
2674 | migrate->src[i] = 0; | |
2675 | unlock_page(page); | |
2676 | migrate->cpages--; | |
2677 | put_page(page); | |
2678 | } | |
2679 | continue; | |
2680 | } | |
2681 | ||
2682 | /* Drop the reference we took in collect */ | |
2683 | put_page(page); | |
2684 | } | |
2685 | ||
2686 | if (!migrate_vma_check_page(page)) { | |
2687 | if (remap) { | |
2688 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2689 | migrate->cpages--; | |
2690 | restore++; | |
2691 | ||
2692 | if (!is_zone_device_page(page)) { | |
2693 | get_page(page); | |
2694 | putback_lru_page(page); | |
2695 | } | |
2696 | } else { | |
2697 | migrate->src[i] = 0; | |
2698 | unlock_page(page); | |
2699 | migrate->cpages--; | |
2700 | ||
2701 | if (!is_zone_device_page(page)) | |
2702 | putback_lru_page(page); | |
2703 | else | |
2704 | put_page(page); | |
2705 | } | |
2706 | } | |
2707 | } | |
2708 | ||
2709 | for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { | |
2710 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2711 | ||
2712 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2713 | continue; | |
2714 | ||
2715 | remove_migration_pte(page, migrate->vma, addr, page); | |
2716 | ||
2717 | migrate->src[i] = 0; | |
2718 | unlock_page(page); | |
2719 | put_page(page); | |
2720 | restore--; | |
2721 | } | |
2722 | } | |
2723 | ||
2724 | /* | |
2725 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
2726 | * @migrate: migrate struct containing all migration information | |
2727 | * | |
2728 | * Replace page mapping (CPU page table pte) with a special migration pte entry | |
2729 | * and check again if it has been pinned. Pinned pages are restored because we | |
2730 | * cannot migrate them. | |
2731 | * | |
2732 | * This is the last step before we call the device driver callback to allocate | |
2733 | * destination memory and copy contents of original page over to new page. | |
2734 | */ | |
2735 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
2736 | { | |
2737 | int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK; | |
2738 | const unsigned long npages = migrate->npages; | |
2739 | const unsigned long start = migrate->start; | |
2740 | unsigned long addr, i, restore = 0; | |
2741 | ||
2742 | for (i = 0; i < npages; i++) { | |
2743 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2744 | ||
2745 | if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2746 | continue; | |
2747 | ||
2748 | if (page_mapped(page)) { | |
2749 | try_to_unmap(page, flags); | |
2750 | if (page_mapped(page)) | |
2751 | goto restore; | |
2752 | } | |
2753 | ||
2754 | if (migrate_vma_check_page(page)) | |
2755 | continue; | |
2756 | ||
2757 | restore: | |
2758 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
2759 | migrate->cpages--; | |
2760 | restore++; | |
2761 | } | |
2762 | ||
2763 | for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { | |
2764 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
2765 | ||
2766 | if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
2767 | continue; | |
2768 | ||
2769 | remove_migration_ptes(page, page, false); | |
2770 | ||
2771 | migrate->src[i] = 0; | |
2772 | unlock_page(page); | |
2773 | restore--; | |
2774 | ||
2775 | if (is_zone_device_page(page)) | |
2776 | put_page(page); | |
2777 | else | |
2778 | putback_lru_page(page); | |
2779 | } | |
2780 | } | |
2781 | ||
2782 | /** | |
2783 | * migrate_vma_setup() - prepare to migrate a range of memory | |
2784 | * @args: contains the vma, start, and pfns arrays for the migration | |
2785 | * | |
2786 | * Returns: negative errno on failures, 0 when 0 or more pages were migrated | |
2787 | * without an error. | |
2788 | * | |
2789 | * Prepare to migrate a range of memory virtual address range by collecting all | |
2790 | * the pages backing each virtual address in the range, saving them inside the | |
2791 | * src array. Then lock those pages and unmap them. Once the pages are locked | |
2792 | * and unmapped, check whether each page is pinned or not. Pages that aren't | |
2793 | * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the | |
2794 | * corresponding src array entry. Then restores any pages that are pinned, by | |
2795 | * remapping and unlocking those pages. | |
2796 | * | |
2797 | * The caller should then allocate destination memory and copy source memory to | |
2798 | * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE | |
2799 | * flag set). Once these are allocated and copied, the caller must update each | |
2800 | * corresponding entry in the dst array with the pfn value of the destination | |
2801 | * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set | |
2802 | * (destination pages must have their struct pages locked, via lock_page()). | |
2803 | * | |
2804 | * Note that the caller does not have to migrate all the pages that are marked | |
2805 | * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from | |
2806 | * device memory to system memory. If the caller cannot migrate a device page | |
2807 | * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe | |
2808 | * consequences for the userspace process, so it must be avoided if at all | |
2809 | * possible. | |
2810 | * | |
2811 | * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we | |
2812 | * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus | |
2813 | * allowing the caller to allocate device memory for those unback virtual | |
2814 | * address. For this the caller simply has to allocate device memory and | |
2815 | * properly set the destination entry like for regular migration. Note that | |
2816 | * this can still fails and thus inside the device driver must check if the | |
2817 | * migration was successful for those entries after calling migrate_vma_pages() | |
2818 | * just like for regular migration. | |
2819 | * | |
2820 | * After that, the callers must call migrate_vma_pages() to go over each entry | |
2821 | * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
2822 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
2823 | * then migrate_vma_pages() to migrate struct page information from the source | |
2824 | * struct page to the destination struct page. If it fails to migrate the | |
2825 | * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the | |
2826 | * src array. | |
2827 | * | |
2828 | * At this point all successfully migrated pages have an entry in the src | |
2829 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
2830 | * array entry with MIGRATE_PFN_VALID flag set. | |
2831 | * | |
2832 | * Once migrate_vma_pages() returns the caller may inspect which pages were | |
2833 | * successfully migrated, and which were not. Successfully migrated pages will | |
2834 | * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. | |
2835 | * | |
2836 | * It is safe to update device page table after migrate_vma_pages() because | |
2837 | * both destination and source page are still locked, and the mmap_lock is held | |
2838 | * in read mode (hence no one can unmap the range being migrated). | |
2839 | * | |
2840 | * Once the caller is done cleaning up things and updating its page table (if it | |
2841 | * chose to do so, this is not an obligation) it finally calls | |
2842 | * migrate_vma_finalize() to update the CPU page table to point to new pages | |
2843 | * for successfully migrated pages or otherwise restore the CPU page table to | |
2844 | * point to the original source pages. | |
2845 | */ | |
2846 | int migrate_vma_setup(struct migrate_vma *args) | |
2847 | { | |
2848 | long nr_pages = (args->end - args->start) >> PAGE_SHIFT; | |
2849 | ||
2850 | args->start &= PAGE_MASK; | |
2851 | args->end &= PAGE_MASK; | |
2852 | if (!args->vma || is_vm_hugetlb_page(args->vma) || | |
2853 | (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) | |
2854 | return -EINVAL; | |
2855 | if (nr_pages <= 0) | |
2856 | return -EINVAL; | |
2857 | if (args->start < args->vma->vm_start || | |
2858 | args->start >= args->vma->vm_end) | |
2859 | return -EINVAL; | |
2860 | if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) | |
2861 | return -EINVAL; | |
2862 | if (!args->src || !args->dst) | |
2863 | return -EINVAL; | |
2864 | ||
2865 | memset(args->src, 0, sizeof(*args->src) * nr_pages); | |
2866 | args->cpages = 0; | |
2867 | args->npages = 0; | |
2868 | ||
2869 | migrate_vma_collect(args); | |
2870 | ||
2871 | if (args->cpages) | |
2872 | migrate_vma_prepare(args); | |
2873 | if (args->cpages) | |
2874 | migrate_vma_unmap(args); | |
2875 | ||
2876 | /* | |
2877 | * At this point pages are locked and unmapped, and thus they have | |
2878 | * stable content and can safely be copied to destination memory that | |
2879 | * is allocated by the drivers. | |
2880 | */ | |
2881 | return 0; | |
2882 | ||
2883 | } | |
2884 | EXPORT_SYMBOL(migrate_vma_setup); | |
2885 | ||
2886 | /* | |
2887 | * This code closely matches the code in: | |
2888 | * __handle_mm_fault() | |
2889 | * handle_pte_fault() | |
2890 | * do_anonymous_page() | |
2891 | * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE | |
2892 | * private page. | |
2893 | */ | |
2894 | static void migrate_vma_insert_page(struct migrate_vma *migrate, | |
2895 | unsigned long addr, | |
2896 | struct page *page, | |
2897 | unsigned long *src) | |
2898 | { | |
2899 | struct vm_area_struct *vma = migrate->vma; | |
2900 | struct mm_struct *mm = vma->vm_mm; | |
2901 | bool flush = false; | |
2902 | spinlock_t *ptl; | |
2903 | pte_t entry; | |
2904 | pgd_t *pgdp; | |
2905 | p4d_t *p4dp; | |
2906 | pud_t *pudp; | |
2907 | pmd_t *pmdp; | |
2908 | pte_t *ptep; | |
2909 | ||
2910 | /* Only allow populating anonymous memory */ | |
2911 | if (!vma_is_anonymous(vma)) | |
2912 | goto abort; | |
2913 | ||
2914 | pgdp = pgd_offset(mm, addr); | |
2915 | p4dp = p4d_alloc(mm, pgdp, addr); | |
2916 | if (!p4dp) | |
2917 | goto abort; | |
2918 | pudp = pud_alloc(mm, p4dp, addr); | |
2919 | if (!pudp) | |
2920 | goto abort; | |
2921 | pmdp = pmd_alloc(mm, pudp, addr); | |
2922 | if (!pmdp) | |
2923 | goto abort; | |
2924 | ||
2925 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) | |
2926 | goto abort; | |
2927 | ||
2928 | /* | |
2929 | * Use pte_alloc() instead of pte_alloc_map(). We can't run | |
2930 | * pte_offset_map() on pmds where a huge pmd might be created | |
2931 | * from a different thread. | |
2932 | * | |
2933 | * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when | |
2934 | * parallel threads are excluded by other means. | |
2935 | * | |
2936 | * Here we only have mmap_read_lock(mm). | |
2937 | */ | |
2938 | if (pte_alloc(mm, pmdp)) | |
2939 | goto abort; | |
2940 | ||
2941 | /* See the comment in pte_alloc_one_map() */ | |
2942 | if (unlikely(pmd_trans_unstable(pmdp))) | |
2943 | goto abort; | |
2944 | ||
2945 | if (unlikely(anon_vma_prepare(vma))) | |
2946 | goto abort; | |
2947 | if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL)) | |
2948 | goto abort; | |
2949 | ||
2950 | /* | |
2951 | * The memory barrier inside __SetPageUptodate makes sure that | |
2952 | * preceding stores to the page contents become visible before | |
2953 | * the set_pte_at() write. | |
2954 | */ | |
2955 | __SetPageUptodate(page); | |
2956 | ||
2957 | if (is_zone_device_page(page)) { | |
2958 | if (is_device_private_page(page)) { | |
2959 | swp_entry_t swp_entry; | |
2960 | ||
2961 | swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE); | |
2962 | entry = swp_entry_to_pte(swp_entry); | |
2963 | } | |
2964 | } else { | |
2965 | entry = mk_pte(page, vma->vm_page_prot); | |
2966 | if (vma->vm_flags & VM_WRITE) | |
2967 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2968 | } | |
2969 | ||
2970 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
2971 | ||
2972 | if (check_stable_address_space(mm)) | |
2973 | goto unlock_abort; | |
2974 | ||
2975 | if (pte_present(*ptep)) { | |
2976 | unsigned long pfn = pte_pfn(*ptep); | |
2977 | ||
2978 | if (!is_zero_pfn(pfn)) | |
2979 | goto unlock_abort; | |
2980 | flush = true; | |
2981 | } else if (!pte_none(*ptep)) | |
2982 | goto unlock_abort; | |
2983 | ||
2984 | /* | |
2985 | * Check for userfaultfd but do not deliver the fault. Instead, | |
2986 | * just back off. | |
2987 | */ | |
2988 | if (userfaultfd_missing(vma)) | |
2989 | goto unlock_abort; | |
2990 | ||
2991 | inc_mm_counter(mm, MM_ANONPAGES); | |
2992 | page_add_new_anon_rmap(page, vma, addr, false); | |
2993 | if (!is_zone_device_page(page)) | |
2994 | lru_cache_add_inactive_or_unevictable(page, vma); | |
2995 | get_page(page); | |
2996 | ||
2997 | if (flush) { | |
2998 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
2999 | ptep_clear_flush_notify(vma, addr, ptep); | |
3000 | set_pte_at_notify(mm, addr, ptep, entry); | |
3001 | update_mmu_cache(vma, addr, ptep); | |
3002 | } else { | |
3003 | /* No need to invalidate - it was non-present before */ | |
3004 | set_pte_at(mm, addr, ptep, entry); | |
3005 | update_mmu_cache(vma, addr, ptep); | |
3006 | } | |
3007 | ||
3008 | pte_unmap_unlock(ptep, ptl); | |
3009 | *src = MIGRATE_PFN_MIGRATE; | |
3010 | return; | |
3011 | ||
3012 | unlock_abort: | |
3013 | pte_unmap_unlock(ptep, ptl); | |
3014 | abort: | |
3015 | *src &= ~MIGRATE_PFN_MIGRATE; | |
3016 | } | |
3017 | ||
3018 | /** | |
3019 | * migrate_vma_pages() - migrate meta-data from src page to dst page | |
3020 | * @migrate: migrate struct containing all migration information | |
3021 | * | |
3022 | * This migrates struct page meta-data from source struct page to destination | |
3023 | * struct page. This effectively finishes the migration from source page to the | |
3024 | * destination page. | |
3025 | */ | |
3026 | void migrate_vma_pages(struct migrate_vma *migrate) | |
3027 | { | |
3028 | const unsigned long npages = migrate->npages; | |
3029 | const unsigned long start = migrate->start; | |
3030 | struct mmu_notifier_range range; | |
3031 | unsigned long addr, i; | |
3032 | bool notified = false; | |
3033 | ||
3034 | for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { | |
3035 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
3036 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
3037 | struct address_space *mapping; | |
3038 | int r; | |
3039 | ||
3040 | if (!newpage) { | |
3041 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
3042 | continue; | |
3043 | } | |
3044 | ||
3045 | if (!page) { | |
3046 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) | |
3047 | continue; | |
3048 | if (!notified) { | |
3049 | notified = true; | |
3050 | ||
3051 | mmu_notifier_range_init_migrate(&range, 0, | |
3052 | migrate->vma, migrate->vma->vm_mm, | |
3053 | addr, migrate->end, | |
3054 | migrate->pgmap_owner); | |
3055 | mmu_notifier_invalidate_range_start(&range); | |
3056 | } | |
3057 | migrate_vma_insert_page(migrate, addr, newpage, | |
3058 | &migrate->src[i]); | |
3059 | continue; | |
3060 | } | |
3061 | ||
3062 | mapping = page_mapping(page); | |
3063 | ||
3064 | if (is_zone_device_page(newpage)) { | |
3065 | if (is_device_private_page(newpage)) { | |
3066 | /* | |
3067 | * For now only support private anonymous when | |
3068 | * migrating to un-addressable device memory. | |
3069 | */ | |
3070 | if (mapping) { | |
3071 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
3072 | continue; | |
3073 | } | |
3074 | } else { | |
3075 | /* | |
3076 | * Other types of ZONE_DEVICE page are not | |
3077 | * supported. | |
3078 | */ | |
3079 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
3080 | continue; | |
3081 | } | |
3082 | } | |
3083 | ||
3084 | r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); | |
3085 | if (r != MIGRATEPAGE_SUCCESS) | |
3086 | migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; | |
3087 | } | |
3088 | ||
3089 | /* | |
3090 | * No need to double call mmu_notifier->invalidate_range() callback as | |
3091 | * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() | |
3092 | * did already call it. | |
3093 | */ | |
3094 | if (notified) | |
3095 | mmu_notifier_invalidate_range_only_end(&range); | |
3096 | } | |
3097 | EXPORT_SYMBOL(migrate_vma_pages); | |
3098 | ||
3099 | /** | |
3100 | * migrate_vma_finalize() - restore CPU page table entry | |
3101 | * @migrate: migrate struct containing all migration information | |
3102 | * | |
3103 | * This replaces the special migration pte entry with either a mapping to the | |
3104 | * new page if migration was successful for that page, or to the original page | |
3105 | * otherwise. | |
3106 | * | |
3107 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
3108 | * refcount, for device pages. | |
3109 | */ | |
3110 | void migrate_vma_finalize(struct migrate_vma *migrate) | |
3111 | { | |
3112 | const unsigned long npages = migrate->npages; | |
3113 | unsigned long i; | |
3114 | ||
3115 | for (i = 0; i < npages; i++) { | |
3116 | struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); | |
3117 | struct page *page = migrate_pfn_to_page(migrate->src[i]); | |
3118 | ||
3119 | if (!page) { | |
3120 | if (newpage) { | |
3121 | unlock_page(newpage); | |
3122 | put_page(newpage); | |
3123 | } | |
3124 | continue; | |
3125 | } | |
3126 | ||
3127 | if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { | |
3128 | if (newpage) { | |
3129 | unlock_page(newpage); | |
3130 | put_page(newpage); | |
3131 | } | |
3132 | newpage = page; | |
3133 | } | |
3134 | ||
3135 | remove_migration_ptes(page, newpage, false); | |
3136 | unlock_page(page); | |
3137 | ||
3138 | if (is_zone_device_page(page)) | |
3139 | put_page(page); | |
3140 | else | |
3141 | putback_lru_page(page); | |
3142 | ||
3143 | if (newpage != page) { | |
3144 | unlock_page(newpage); | |
3145 | if (is_zone_device_page(newpage)) | |
3146 | put_page(newpage); | |
3147 | else | |
3148 | putback_lru_page(newpage); | |
3149 | } | |
3150 | } | |
3151 | } | |
3152 | EXPORT_SYMBOL(migrate_vma_finalize); | |
3153 | #endif /* CONFIG_DEVICE_PRIVATE */ |