Line data Source code
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 0 : 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 0 : lru_add_drain_all();
74 0 : }
75 :
76 : /* Do the necessary work of migrate_prep but not if it involves other CPUs */
77 0 : void migrate_prep_local(void)
78 : {
79 0 : lru_add_drain();
80 0 : }
81 :
82 0 : int isolate_movable_page(struct page *page, isolate_mode_t mode)
83 : {
84 0 : 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 0 : if (unlikely(!get_page_unless_zero(page)))
96 0 : 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 0 : if (unlikely(!__PageMovable(page)))
104 0 : 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 0 : if (unlikely(!trylock_page(page)))
117 0 : goto out_putpage;
118 :
119 0 : if (!PageMovable(page) || PageIsolated(page))
120 0 : goto out_no_isolated;
121 :
122 0 : mapping = page_mapping(page);
123 0 : VM_BUG_ON_PAGE(!mapping, page);
124 :
125 0 : if (!mapping->a_ops->isolate_page(page, mode))
126 0 : goto out_no_isolated;
127 :
128 : /* Driver shouldn't use PG_isolated bit of page->flags */
129 0 : WARN_ON_ONCE(PageIsolated(page));
130 0 : __SetPageIsolated(page);
131 0 : unlock_page(page);
132 :
133 0 : return 0;
134 :
135 0 : out_no_isolated:
136 0 : unlock_page(page);
137 0 : out_putpage:
138 0 : put_page(page);
139 : out:
140 : return -EBUSY;
141 : }
142 :
143 : /* It should be called on page which is PG_movable */
144 0 : void putback_movable_page(struct page *page)
145 : {
146 0 : struct address_space *mapping;
147 :
148 0 : VM_BUG_ON_PAGE(!PageLocked(page), page);
149 0 : VM_BUG_ON_PAGE(!PageMovable(page), page);
150 0 : VM_BUG_ON_PAGE(!PageIsolated(page), page);
151 :
152 0 : mapping = page_mapping(page);
153 0 : mapping->a_ops->putback_page(page);
154 0 : __ClearPageIsolated(page);
155 0 : }
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 0 : void putback_movable_pages(struct list_head *l)
166 : {
167 0 : struct page *page;
168 0 : struct page *page2;
169 :
170 0 : list_for_each_entry_safe(page, page2, l, lru) {
171 0 : if (unlikely(PageHuge(page))) {
172 : putback_active_hugepage(page);
173 : continue;
174 : }
175 0 : 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 0 : if (unlikely(__PageMovable(page))) {
182 0 : VM_BUG_ON_PAGE(!PageIsolated(page), page);
183 0 : lock_page(page);
184 0 : if (PageMovable(page))
185 0 : putback_movable_page(page);
186 : else
187 0 : __ClearPageIsolated(page);
188 0 : unlock_page(page);
189 0 : put_page(page);
190 : } else {
191 0 : mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
192 0 : page_is_file_lru(page), -thp_nr_pages(page));
193 0 : putback_lru_page(page);
194 : }
195 : }
196 0 : }
197 :
198 : /*
199 : * Restore a potential migration pte to a working pte entry
200 : */
201 0 : static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
202 : unsigned long addr, void *old)
203 : {
204 0 : struct page_vma_mapped_walk pvmw = {
205 : .page = old,
206 : .vma = vma,
207 : .address = addr,
208 : .flags = PVMW_SYNC | PVMW_MIGRATION,
209 : };
210 0 : struct page *new;
211 0 : pte_t pte;
212 0 : swp_entry_t entry;
213 :
214 0 : VM_BUG_ON_PAGE(PageTail(page), page);
215 0 : while (page_vma_mapped_walk(&pvmw)) {
216 0 : if (PageKsm(page))
217 : new = page;
218 : else
219 0 : new = page - pvmw.page->index +
220 0 : linear_page_index(vma, pvmw.address);
221 :
222 : #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
223 : /* PMD-mapped THP migration entry */
224 0 : if (!pvmw.pte) {
225 0 : VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
226 0 : remove_migration_pmd(&pvmw, new);
227 0 : continue;
228 : }
229 : #endif
230 :
231 0 : get_page(new);
232 0 : pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
233 0 : 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 0 : entry = pte_to_swp_entry(*pvmw.pte);
240 0 : if (is_write_migration_entry(entry))
241 0 : pte = maybe_mkwrite(pte, vma);
242 : else if (pte_swp_uffd_wp(*pvmw.pte))
243 : pte = pte_mkuffd_wp(pte);
244 :
245 0 : 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 0 : set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
267 :
268 0 : if (PageAnon(new))
269 0 : page_add_anon_rmap(new, vma, pvmw.address, false);
270 : else
271 0 : page_add_file_rmap(new, false);
272 : }
273 0 : if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
274 0 : mlock_vma_page(new);
275 :
276 0 : if (PageTransHuge(page) && PageMlocked(page))
277 0 : clear_page_mlock(page);
278 :
279 : /* No need to invalidate - it was non-present before */
280 0 : update_mmu_cache(vma, pvmw.address, pvmw.pte);
281 : }
282 :
283 0 : return true;
284 : }
285 :
286 : /*
287 : * Get rid of all migration entries and replace them by
288 : * references to the indicated page.
289 : */
290 0 : void remove_migration_ptes(struct page *old, struct page *new, bool locked)
291 : {
292 0 : struct rmap_walk_control rwc = {
293 : .rmap_one = remove_migration_pte,
294 : .arg = old,
295 : };
296 :
297 0 : if (locked)
298 0 : rmap_walk_locked(new, &rwc);
299 : else
300 0 : rmap_walk(new, &rwc);
301 0 : }
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 0 : void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
309 : spinlock_t *ptl)
310 : {
311 0 : pte_t pte;
312 0 : swp_entry_t entry;
313 0 : struct page *page;
314 :
315 0 : spin_lock(ptl);
316 0 : pte = *ptep;
317 0 : if (!is_swap_pte(pte))
318 0 : goto out;
319 :
320 0 : entry = pte_to_swp_entry(pte);
321 0 : if (!is_migration_entry(entry))
322 0 : goto out;
323 :
324 0 : 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 0 : if (!get_page_unless_zero(page))
332 0 : goto out;
333 0 : pte_unmap_unlock(ptep, ptl);
334 0 : put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
335 0 : return;
336 0 : out:
337 0 : pte_unmap_unlock(ptep, ptl);
338 : }
339 :
340 0 : void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
341 : unsigned long address)
342 : {
343 0 : spinlock_t *ptl = pte_lockptr(mm, pmd);
344 0 : pte_t *ptep = pte_offset_map(pmd, address);
345 0 : __migration_entry_wait(mm, ptep, ptl);
346 0 : }
347 :
348 0 : void migration_entry_wait_huge(struct vm_area_struct *vma,
349 : struct mm_struct *mm, pte_t *pte)
350 : {
351 0 : spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
352 0 : __migration_entry_wait(mm, pte, ptl);
353 0 : }
354 :
355 : #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
356 0 : void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
357 : {
358 0 : spinlock_t *ptl;
359 0 : struct page *page;
360 :
361 0 : ptl = pmd_lock(mm, pmd);
362 0 : if (!is_pmd_migration_entry(*pmd))
363 0 : goto unlock;
364 0 : page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
365 0 : if (!get_page_unless_zero(page))
366 0 : goto unlock;
367 0 : spin_unlock(ptl);
368 0 : put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
369 0 : return;
370 0 : unlock:
371 0 : spin_unlock(ptl);
372 : }
373 : #endif
374 :
375 0 : static int expected_page_refs(struct address_space *mapping, struct page *page)
376 : {
377 0 : int expected_count = 1;
378 :
379 : /*
380 : * Device private pages have an extra refcount as they are
381 : * ZONE_DEVICE pages.
382 : */
383 0 : expected_count += is_device_private_page(page);
384 0 : if (mapping)
385 0 : expected_count += thp_nr_pages(page) + page_has_private(page);
386 :
387 0 : 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 0 : int migrate_page_move_mapping(struct address_space *mapping,
399 : struct page *newpage, struct page *page, int extra_count)
400 : {
401 0 : XA_STATE(xas, &mapping->i_pages, page_index(page));
402 0 : struct zone *oldzone, *newzone;
403 0 : int dirty;
404 0 : int expected_count = expected_page_refs(mapping, page) + extra_count;
405 0 : int nr = thp_nr_pages(page);
406 :
407 0 : if (!mapping) {
408 : /* Anonymous page without mapping */
409 0 : if (page_count(page) != expected_count)
410 : return -EAGAIN;
411 :
412 : /* No turning back from here */
413 0 : newpage->index = page->index;
414 0 : newpage->mapping = page->mapping;
415 0 : if (PageSwapBacked(page))
416 0 : __SetPageSwapBacked(newpage);
417 :
418 0 : return MIGRATEPAGE_SUCCESS;
419 : }
420 :
421 0 : oldzone = page_zone(page);
422 0 : newzone = page_zone(newpage);
423 :
424 0 : xas_lock_irq(&xas);
425 0 : if (page_count(page) != expected_count || xas_load(&xas) != page) {
426 0 : xas_unlock_irq(&xas);
427 0 : return -EAGAIN;
428 : }
429 :
430 0 : if (!page_ref_freeze(page, expected_count)) {
431 0 : xas_unlock_irq(&xas);
432 0 : return -EAGAIN;
433 : }
434 :
435 : /*
436 : * Now we know that no one else is looking at the page:
437 : * no turning back from here.
438 : */
439 0 : newpage->index = page->index;
440 0 : newpage->mapping = page->mapping;
441 0 : page_ref_add(newpage, nr); /* add cache reference */
442 0 : if (PageSwapBacked(page)) {
443 0 : __SetPageSwapBacked(newpage);
444 0 : if (PageSwapCache(page)) {
445 : SetPageSwapCache(newpage);
446 0 : set_page_private(newpage, page_private(page));
447 : }
448 : } else {
449 0 : VM_BUG_ON_PAGE(PageSwapCache(page), page);
450 : }
451 :
452 : /* Move dirty while page refs frozen and newpage not yet exposed */
453 0 : dirty = PageDirty(page);
454 0 : if (dirty) {
455 0 : ClearPageDirty(page);
456 0 : SetPageDirty(newpage);
457 : }
458 :
459 0 : xas_store(&xas, newpage);
460 0 : if (PageTransHuge(page)) {
461 : int i;
462 :
463 0 : for (i = 1; i < nr; i++) {
464 0 : xas_next(&xas);
465 0 : xas_store(&xas, newpage);
466 : }
467 : }
468 :
469 : /*
470 : * Drop cache reference from old page by unfreezing
471 : * to one less reference.
472 : * We know this isn't the last reference.
473 : */
474 0 : page_ref_unfreeze(page, expected_count - nr);
475 :
476 0 : xas_unlock(&xas);
477 : /* Leave irq disabled to prevent preemption while updating stats */
478 :
479 : /*
480 : * If moved to a different zone then also account
481 : * the page for that zone. Other VM counters will be
482 : * taken care of when we establish references to the
483 : * new page and drop references to the old page.
484 : *
485 : * Note that anonymous pages are accounted for
486 : * via NR_FILE_PAGES and NR_ANON_MAPPED if they
487 : * are mapped to swap space.
488 : */
489 0 : if (newzone != oldzone) {
490 0 : struct lruvec *old_lruvec, *new_lruvec;
491 0 : struct mem_cgroup *memcg;
492 :
493 0 : memcg = page_memcg(page);
494 0 : old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
495 0 : new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
496 :
497 0 : __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
498 0 : __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
499 0 : if (PageSwapBacked(page) && !PageSwapCache(page)) {
500 0 : __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
501 0 : __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
502 : }
503 : #ifdef CONFIG_SWAP
504 : if (PageSwapCache(page)) {
505 : __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
506 : __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
507 : }
508 : #endif
509 0 : if (dirty && mapping_can_writeback(mapping)) {
510 0 : __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
511 0 : __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
512 0 : __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
513 0 : __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
514 : }
515 : }
516 0 : local_irq_enable();
517 :
518 0 : return MIGRATEPAGE_SUCCESS;
519 : }
520 : EXPORT_SYMBOL(migrate_page_move_mapping);
521 :
522 : /*
523 : * The expected number of remaining references is the same as that
524 : * of migrate_page_move_mapping().
525 : */
526 0 : int migrate_huge_page_move_mapping(struct address_space *mapping,
527 : struct page *newpage, struct page *page)
528 : {
529 0 : XA_STATE(xas, &mapping->i_pages, page_index(page));
530 0 : int expected_count;
531 :
532 0 : xas_lock_irq(&xas);
533 0 : expected_count = 2 + page_has_private(page);
534 0 : if (page_count(page) != expected_count || xas_load(&xas) != page) {
535 0 : xas_unlock_irq(&xas);
536 0 : return -EAGAIN;
537 : }
538 :
539 0 : if (!page_ref_freeze(page, expected_count)) {
540 0 : xas_unlock_irq(&xas);
541 0 : return -EAGAIN;
542 : }
543 :
544 0 : newpage->index = page->index;
545 0 : newpage->mapping = page->mapping;
546 :
547 0 : get_page(newpage);
548 :
549 0 : xas_store(&xas, newpage);
550 :
551 0 : page_ref_unfreeze(page, expected_count - 1);
552 :
553 0 : xas_unlock_irq(&xas);
554 :
555 0 : return MIGRATEPAGE_SUCCESS;
556 : }
557 :
558 : /*
559 : * Gigantic pages are so large that we do not guarantee that page++ pointer
560 : * arithmetic will work across the entire page. We need something more
561 : * specialized.
562 : */
563 : static void __copy_gigantic_page(struct page *dst, struct page *src,
564 : int nr_pages)
565 : {
566 : int i;
567 : struct page *dst_base = dst;
568 : struct page *src_base = src;
569 :
570 : for (i = 0; i < nr_pages; ) {
571 : cond_resched();
572 : copy_highpage(dst, src);
573 :
574 : i++;
575 : dst = mem_map_next(dst, dst_base, i);
576 : src = mem_map_next(src, src_base, i);
577 : }
578 : }
579 :
580 0 : static void copy_huge_page(struct page *dst, struct page *src)
581 : {
582 0 : int i;
583 0 : int nr_pages;
584 :
585 0 : if (PageHuge(src)) {
586 : /* hugetlbfs page */
587 : struct hstate *h = page_hstate(src);
588 : nr_pages = pages_per_huge_page(h);
589 :
590 : if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
591 : __copy_gigantic_page(dst, src, nr_pages);
592 : return;
593 : }
594 : } else {
595 : /* thp page */
596 0 : BUG_ON(!PageTransHuge(src));
597 0 : nr_pages = thp_nr_pages(src);
598 : }
599 :
600 0 : for (i = 0; i < nr_pages; i++) {
601 0 : cond_resched();
602 0 : copy_highpage(dst + i, src + i);
603 : }
604 : }
605 :
606 : /*
607 : * Copy the page to its new location
608 : */
609 0 : void migrate_page_states(struct page *newpage, struct page *page)
610 : {
611 0 : int cpupid;
612 :
613 0 : if (PageError(page))
614 0 : SetPageError(newpage);
615 0 : if (PageReferenced(page))
616 0 : SetPageReferenced(newpage);
617 0 : if (PageUptodate(page))
618 0 : SetPageUptodate(newpage);
619 0 : if (TestClearPageActive(page)) {
620 0 : VM_BUG_ON_PAGE(PageUnevictable(page), page);
621 0 : SetPageActive(newpage);
622 0 : } else if (TestClearPageUnevictable(page))
623 0 : SetPageUnevictable(newpage);
624 0 : if (PageWorkingset(page))
625 0 : SetPageWorkingset(newpage);
626 0 : if (PageChecked(page))
627 0 : SetPageChecked(newpage);
628 0 : if (PageMappedToDisk(page))
629 0 : SetPageMappedToDisk(newpage);
630 :
631 : /* Move dirty on pages not done by migrate_page_move_mapping() */
632 0 : if (PageDirty(page))
633 0 : SetPageDirty(newpage);
634 :
635 0 : if (page_is_young(page))
636 0 : set_page_young(newpage);
637 0 : if (page_is_idle(page))
638 0 : set_page_idle(newpage);
639 :
640 : /*
641 : * Copy NUMA information to the new page, to prevent over-eager
642 : * future migrations of this same page.
643 : */
644 0 : cpupid = page_cpupid_xchg_last(page, -1);
645 0 : page_cpupid_xchg_last(newpage, cpupid);
646 :
647 0 : ksm_migrate_page(newpage, page);
648 : /*
649 : * Please do not reorder this without considering how mm/ksm.c's
650 : * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
651 : */
652 0 : if (PageSwapCache(page))
653 0 : ClearPageSwapCache(page);
654 0 : ClearPagePrivate(page);
655 0 : set_page_private(page, 0);
656 :
657 : /*
658 : * If any waiters have accumulated on the new page then
659 : * wake them up.
660 : */
661 0 : if (PageWriteback(newpage))
662 0 : end_page_writeback(newpage);
663 :
664 : /*
665 : * PG_readahead shares the same bit with PG_reclaim. The above
666 : * end_page_writeback() may clear PG_readahead mistakenly, so set the
667 : * bit after that.
668 : */
669 0 : if (PageReadahead(page))
670 0 : SetPageReadahead(newpage);
671 :
672 0 : copy_page_owner(page, newpage);
673 :
674 0 : if (!PageHuge(page))
675 0 : mem_cgroup_migrate(page, newpage);
676 0 : }
677 : EXPORT_SYMBOL(migrate_page_states);
678 :
679 0 : void migrate_page_copy(struct page *newpage, struct page *page)
680 : {
681 0 : if (PageHuge(page) || PageTransHuge(page))
682 0 : copy_huge_page(newpage, page);
683 : else
684 0 : copy_highpage(newpage, page);
685 :
686 0 : migrate_page_states(newpage, page);
687 0 : }
688 : EXPORT_SYMBOL(migrate_page_copy);
689 :
690 : /************************************************************
691 : * Migration functions
692 : ***********************************************************/
693 :
694 : /*
695 : * Common logic to directly migrate a single LRU page suitable for
696 : * pages that do not use PagePrivate/PagePrivate2.
697 : *
698 : * Pages are locked upon entry and exit.
699 : */
700 0 : int migrate_page(struct address_space *mapping,
701 : struct page *newpage, struct page *page,
702 : enum migrate_mode mode)
703 : {
704 0 : int rc;
705 :
706 0 : BUG_ON(PageWriteback(page)); /* Writeback must be complete */
707 :
708 0 : rc = migrate_page_move_mapping(mapping, newpage, page, 0);
709 :
710 0 : if (rc != MIGRATEPAGE_SUCCESS)
711 : return rc;
712 :
713 0 : if (mode != MIGRATE_SYNC_NO_COPY)
714 0 : migrate_page_copy(newpage, page);
715 : else
716 0 : migrate_page_states(newpage, page);
717 : return MIGRATEPAGE_SUCCESS;
718 : }
719 : EXPORT_SYMBOL(migrate_page);
720 :
721 : #ifdef CONFIG_BLOCK
722 : /* Returns true if all buffers are successfully locked */
723 0 : static bool buffer_migrate_lock_buffers(struct buffer_head *head,
724 : enum migrate_mode mode)
725 : {
726 0 : struct buffer_head *bh = head;
727 :
728 : /* Simple case, sync compaction */
729 0 : if (mode != MIGRATE_ASYNC) {
730 0 : do {
731 0 : lock_buffer(bh);
732 0 : bh = bh->b_this_page;
733 :
734 0 : } while (bh != head);
735 :
736 : return true;
737 : }
738 :
739 : /* async case, we cannot block on lock_buffer so use trylock_buffer */
740 0 : do {
741 0 : if (!trylock_buffer(bh)) {
742 : /*
743 : * We failed to lock the buffer and cannot stall in
744 : * async migration. Release the taken locks
745 : */
746 0 : struct buffer_head *failed_bh = bh;
747 : bh = head;
748 0 : while (bh != failed_bh) {
749 0 : unlock_buffer(bh);
750 0 : bh = bh->b_this_page;
751 : }
752 : return false;
753 : }
754 :
755 0 : bh = bh->b_this_page;
756 0 : } while (bh != head);
757 : return true;
758 : }
759 :
760 0 : static int __buffer_migrate_page(struct address_space *mapping,
761 : struct page *newpage, struct page *page, enum migrate_mode mode,
762 : bool check_refs)
763 : {
764 0 : struct buffer_head *bh, *head;
765 0 : int rc;
766 0 : int expected_count;
767 :
768 0 : if (!page_has_buffers(page))
769 0 : return migrate_page(mapping, newpage, page, mode);
770 :
771 : /* Check whether page does not have extra refs before we do more work */
772 0 : expected_count = expected_page_refs(mapping, page);
773 0 : if (page_count(page) != expected_count)
774 : return -EAGAIN;
775 :
776 0 : head = page_buffers(page);
777 0 : if (!buffer_migrate_lock_buffers(head, mode))
778 : return -EAGAIN;
779 :
780 0 : if (check_refs) {
781 : bool busy;
782 : bool invalidated = false;
783 :
784 0 : recheck_buffers:
785 0 : busy = false;
786 0 : spin_lock(&mapping->private_lock);
787 0 : bh = head;
788 0 : do {
789 0 : if (atomic_read(&bh->b_count)) {
790 : busy = true;
791 : break;
792 : }
793 0 : bh = bh->b_this_page;
794 0 : } while (bh != head);
795 0 : if (busy) {
796 0 : if (invalidated) {
797 0 : rc = -EAGAIN;
798 0 : goto unlock_buffers;
799 : }
800 0 : spin_unlock(&mapping->private_lock);
801 0 : invalidate_bh_lrus();
802 0 : invalidated = true;
803 0 : goto recheck_buffers;
804 : }
805 : }
806 :
807 0 : rc = migrate_page_move_mapping(mapping, newpage, page, 0);
808 0 : if (rc != MIGRATEPAGE_SUCCESS)
809 0 : goto unlock_buffers;
810 :
811 0 : attach_page_private(newpage, detach_page_private(page));
812 :
813 0 : bh = head;
814 0 : do {
815 0 : set_bh_page(bh, newpage, bh_offset(bh));
816 0 : bh = bh->b_this_page;
817 :
818 0 : } while (bh != head);
819 :
820 0 : if (mode != MIGRATE_SYNC_NO_COPY)
821 0 : migrate_page_copy(newpage, page);
822 : else
823 0 : migrate_page_states(newpage, page);
824 :
825 : rc = MIGRATEPAGE_SUCCESS;
826 0 : unlock_buffers:
827 0 : if (check_refs)
828 0 : spin_unlock(&mapping->private_lock);
829 : bh = head;
830 0 : do {
831 0 : unlock_buffer(bh);
832 0 : bh = bh->b_this_page;
833 :
834 0 : } while (bh != head);
835 :
836 : return rc;
837 : }
838 :
839 : /*
840 : * Migration function for pages with buffers. This function can only be used
841 : * if the underlying filesystem guarantees that no other references to "page"
842 : * exist. For example attached buffer heads are accessed only under page lock.
843 : */
844 0 : int buffer_migrate_page(struct address_space *mapping,
845 : struct page *newpage, struct page *page, enum migrate_mode mode)
846 : {
847 0 : return __buffer_migrate_page(mapping, newpage, page, mode, false);
848 : }
849 : EXPORT_SYMBOL(buffer_migrate_page);
850 :
851 : /*
852 : * Same as above except that this variant is more careful and checks that there
853 : * are also no buffer head references. This function is the right one for
854 : * mappings where buffer heads are directly looked up and referenced (such as
855 : * block device mappings).
856 : */
857 0 : int buffer_migrate_page_norefs(struct address_space *mapping,
858 : struct page *newpage, struct page *page, enum migrate_mode mode)
859 : {
860 0 : return __buffer_migrate_page(mapping, newpage, page, mode, true);
861 : }
862 : #endif
863 :
864 : /*
865 : * Writeback a page to clean the dirty state
866 : */
867 0 : static int writeout(struct address_space *mapping, struct page *page)
868 : {
869 0 : struct writeback_control wbc = {
870 : .sync_mode = WB_SYNC_NONE,
871 : .nr_to_write = 1,
872 : .range_start = 0,
873 : .range_end = LLONG_MAX,
874 : .for_reclaim = 1
875 : };
876 0 : int rc;
877 :
878 0 : if (!mapping->a_ops->writepage)
879 : /* No write method for the address space */
880 : return -EINVAL;
881 :
882 0 : if (!clear_page_dirty_for_io(page))
883 : /* Someone else already triggered a write */
884 : return -EAGAIN;
885 :
886 : /*
887 : * A dirty page may imply that the underlying filesystem has
888 : * the page on some queue. So the page must be clean for
889 : * migration. Writeout may mean we loose the lock and the
890 : * page state is no longer what we checked for earlier.
891 : * At this point we know that the migration attempt cannot
892 : * be successful.
893 : */
894 0 : remove_migration_ptes(page, page, false);
895 :
896 0 : rc = mapping->a_ops->writepage(page, &wbc);
897 :
898 0 : if (rc != AOP_WRITEPAGE_ACTIVATE)
899 : /* unlocked. Relock */
900 0 : lock_page(page);
901 :
902 0 : return (rc < 0) ? -EIO : -EAGAIN;
903 : }
904 :
905 : /*
906 : * Default handling if a filesystem does not provide a migration function.
907 : */
908 0 : static int fallback_migrate_page(struct address_space *mapping,
909 : struct page *newpage, struct page *page, enum migrate_mode mode)
910 : {
911 0 : if (PageDirty(page)) {
912 : /* Only writeback pages in full synchronous migration */
913 0 : switch (mode) {
914 : case MIGRATE_SYNC:
915 : case MIGRATE_SYNC_NO_COPY:
916 0 : break;
917 : default:
918 : return -EBUSY;
919 : }
920 0 : return writeout(mapping, page);
921 : }
922 :
923 : /*
924 : * Buffers may be managed in a filesystem specific way.
925 : * We must have no buffers or drop them.
926 : */
927 0 : if (page_has_private(page) &&
928 0 : !try_to_release_page(page, GFP_KERNEL))
929 0 : return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
930 :
931 0 : return migrate_page(mapping, newpage, page, mode);
932 : }
933 :
934 : /*
935 : * Move a page to a newly allocated page
936 : * The page is locked and all ptes have been successfully removed.
937 : *
938 : * The new page will have replaced the old page if this function
939 : * is successful.
940 : *
941 : * Return value:
942 : * < 0 - error code
943 : * MIGRATEPAGE_SUCCESS - success
944 : */
945 0 : static int move_to_new_page(struct page *newpage, struct page *page,
946 : enum migrate_mode mode)
947 : {
948 0 : struct address_space *mapping;
949 0 : int rc = -EAGAIN;
950 0 : bool is_lru = !__PageMovable(page);
951 :
952 0 : VM_BUG_ON_PAGE(!PageLocked(page), page);
953 0 : VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
954 :
955 0 : mapping = page_mapping(page);
956 :
957 0 : if (likely(is_lru)) {
958 0 : if (!mapping)
959 0 : rc = migrate_page(mapping, newpage, page, mode);
960 0 : else if (mapping->a_ops->migratepage)
961 : /*
962 : * Most pages have a mapping and most filesystems
963 : * provide a migratepage callback. Anonymous pages
964 : * are part of swap space which also has its own
965 : * migratepage callback. This is the most common path
966 : * for page migration.
967 : */
968 0 : rc = mapping->a_ops->migratepage(mapping, newpage,
969 : page, mode);
970 : else
971 0 : rc = fallback_migrate_page(mapping, newpage,
972 : page, mode);
973 : } else {
974 : /*
975 : * In case of non-lru page, it could be released after
976 : * isolation step. In that case, we shouldn't try migration.
977 : */
978 0 : VM_BUG_ON_PAGE(!PageIsolated(page), page);
979 0 : if (!PageMovable(page)) {
980 0 : rc = MIGRATEPAGE_SUCCESS;
981 0 : __ClearPageIsolated(page);
982 0 : goto out;
983 : }
984 :
985 0 : rc = mapping->a_ops->migratepage(mapping, newpage,
986 : page, mode);
987 0 : WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
988 : !PageIsolated(page));
989 : }
990 :
991 : /*
992 : * When successful, old pagecache page->mapping must be cleared before
993 : * page is freed; but stats require that PageAnon be left as PageAnon.
994 : */
995 0 : if (rc == MIGRATEPAGE_SUCCESS) {
996 0 : if (__PageMovable(page)) {
997 0 : VM_BUG_ON_PAGE(!PageIsolated(page), page);
998 :
999 : /*
1000 : * We clear PG_movable under page_lock so any compactor
1001 : * cannot try to migrate this page.
1002 : */
1003 0 : __ClearPageIsolated(page);
1004 : }
1005 :
1006 : /*
1007 : * Anonymous and movable page->mapping will be cleared by
1008 : * free_pages_prepare so don't reset it here for keeping
1009 : * the type to work PageAnon, for example.
1010 : */
1011 0 : if (!PageMappingFlags(page))
1012 0 : page->mapping = NULL;
1013 :
1014 0 : if (likely(!is_zone_device_page(newpage)))
1015 0 : flush_dcache_page(newpage);
1016 :
1017 : }
1018 0 : out:
1019 0 : return rc;
1020 : }
1021 :
1022 0 : static int __unmap_and_move(struct page *page, struct page *newpage,
1023 : int force, enum migrate_mode mode)
1024 : {
1025 0 : int rc = -EAGAIN;
1026 0 : int page_was_mapped = 0;
1027 0 : struct anon_vma *anon_vma = NULL;
1028 0 : bool is_lru = !__PageMovable(page);
1029 :
1030 0 : if (!trylock_page(page)) {
1031 0 : if (!force || mode == MIGRATE_ASYNC)
1032 0 : goto out;
1033 :
1034 : /*
1035 : * It's not safe for direct compaction to call lock_page.
1036 : * For example, during page readahead pages are added locked
1037 : * to the LRU. Later, when the IO completes the pages are
1038 : * marked uptodate and unlocked. However, the queueing
1039 : * could be merging multiple pages for one bio (e.g.
1040 : * mpage_readahead). If an allocation happens for the
1041 : * second or third page, the process can end up locking
1042 : * the same page twice and deadlocking. Rather than
1043 : * trying to be clever about what pages can be locked,
1044 : * avoid the use of lock_page for direct compaction
1045 : * altogether.
1046 : */
1047 0 : if (current->flags & PF_MEMALLOC)
1048 0 : goto out;
1049 :
1050 0 : lock_page(page);
1051 : }
1052 :
1053 0 : if (PageWriteback(page)) {
1054 : /*
1055 : * Only in the case of a full synchronous migration is it
1056 : * necessary to wait for PageWriteback. In the async case,
1057 : * the retry loop is too short and in the sync-light case,
1058 : * the overhead of stalling is too much
1059 : */
1060 0 : switch (mode) {
1061 : case MIGRATE_SYNC:
1062 : case MIGRATE_SYNC_NO_COPY:
1063 0 : break;
1064 0 : default:
1065 0 : rc = -EBUSY;
1066 0 : goto out_unlock;
1067 : }
1068 0 : if (!force)
1069 0 : goto out_unlock;
1070 0 : wait_on_page_writeback(page);
1071 : }
1072 :
1073 : /*
1074 : * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1075 : * we cannot notice that anon_vma is freed while we migrates a page.
1076 : * This get_anon_vma() delays freeing anon_vma pointer until the end
1077 : * of migration. File cache pages are no problem because of page_lock()
1078 : * File Caches may use write_page() or lock_page() in migration, then,
1079 : * just care Anon page here.
1080 : *
1081 : * Only page_get_anon_vma() understands the subtleties of
1082 : * getting a hold on an anon_vma from outside one of its mms.
1083 : * But if we cannot get anon_vma, then we won't need it anyway,
1084 : * because that implies that the anon page is no longer mapped
1085 : * (and cannot be remapped so long as we hold the page lock).
1086 : */
1087 0 : if (PageAnon(page) && !PageKsm(page))
1088 0 : anon_vma = page_get_anon_vma(page);
1089 :
1090 : /*
1091 : * Block others from accessing the new page when we get around to
1092 : * establishing additional references. We are usually the only one
1093 : * holding a reference to newpage at this point. We used to have a BUG
1094 : * here if trylock_page(newpage) fails, but would like to allow for
1095 : * cases where there might be a race with the previous use of newpage.
1096 : * This is much like races on refcount of oldpage: just don't BUG().
1097 : */
1098 0 : if (unlikely(!trylock_page(newpage)))
1099 0 : goto out_unlock;
1100 :
1101 0 : if (unlikely(!is_lru)) {
1102 0 : rc = move_to_new_page(newpage, page, mode);
1103 0 : goto out_unlock_both;
1104 : }
1105 :
1106 : /*
1107 : * Corner case handling:
1108 : * 1. When a new swap-cache page is read into, it is added to the LRU
1109 : * and treated as swapcache but it has no rmap yet.
1110 : * Calling try_to_unmap() against a page->mapping==NULL page will
1111 : * trigger a BUG. So handle it here.
1112 : * 2. An orphaned page (see truncate_cleanup_page) might have
1113 : * fs-private metadata. The page can be picked up due to memory
1114 : * offlining. Everywhere else except page reclaim, the page is
1115 : * invisible to the vm, so the page can not be migrated. So try to
1116 : * free the metadata, so the page can be freed.
1117 : */
1118 0 : if (!page->mapping) {
1119 0 : VM_BUG_ON_PAGE(PageAnon(page), page);
1120 0 : if (page_has_private(page)) {
1121 0 : try_to_free_buffers(page);
1122 0 : goto out_unlock_both;
1123 : }
1124 0 : } else if (page_mapped(page)) {
1125 : /* Establish migration ptes */
1126 0 : VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1127 : page);
1128 0 : try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK);
1129 0 : page_was_mapped = 1;
1130 : }
1131 :
1132 0 : if (!page_mapped(page))
1133 0 : rc = move_to_new_page(newpage, page, mode);
1134 :
1135 0 : if (page_was_mapped)
1136 0 : remove_migration_ptes(page,
1137 : rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1138 :
1139 0 : out_unlock_both:
1140 0 : unlock_page(newpage);
1141 0 : out_unlock:
1142 : /* Drop an anon_vma reference if we took one */
1143 0 : if (anon_vma)
1144 0 : put_anon_vma(anon_vma);
1145 0 : unlock_page(page);
1146 0 : out:
1147 : /*
1148 : * If migration is successful, decrease refcount of the newpage
1149 : * which will not free the page because new page owner increased
1150 : * refcounter. As well, if it is LRU page, add the page to LRU
1151 : * list in here. Use the old state of the isolated source page to
1152 : * determine if we migrated a LRU page. newpage was already unlocked
1153 : * and possibly modified by its owner - don't rely on the page
1154 : * state.
1155 : */
1156 0 : if (rc == MIGRATEPAGE_SUCCESS) {
1157 0 : if (unlikely(!is_lru))
1158 0 : put_page(newpage);
1159 : else
1160 0 : putback_lru_page(newpage);
1161 : }
1162 :
1163 0 : return rc;
1164 : }
1165 :
1166 : /*
1167 : * Obtain the lock on page, remove all ptes and migrate the page
1168 : * to the newly allocated page in newpage.
1169 : */
1170 0 : static int unmap_and_move(new_page_t get_new_page,
1171 : free_page_t put_new_page,
1172 : unsigned long private, struct page *page,
1173 : int force, enum migrate_mode mode,
1174 : enum migrate_reason reason,
1175 : struct list_head *ret)
1176 : {
1177 0 : int rc = MIGRATEPAGE_SUCCESS;
1178 0 : struct page *newpage = NULL;
1179 :
1180 0 : if (!thp_migration_supported() && PageTransHuge(page))
1181 : return -ENOSYS;
1182 :
1183 0 : if (page_count(page) == 1) {
1184 : /* page was freed from under us. So we are done. */
1185 0 : ClearPageActive(page);
1186 0 : ClearPageUnevictable(page);
1187 0 : if (unlikely(__PageMovable(page))) {
1188 0 : lock_page(page);
1189 0 : if (!PageMovable(page))
1190 0 : __ClearPageIsolated(page);
1191 0 : unlock_page(page);
1192 : }
1193 0 : goto out;
1194 : }
1195 :
1196 0 : newpage = get_new_page(page, private);
1197 0 : if (!newpage)
1198 : return -ENOMEM;
1199 :
1200 0 : rc = __unmap_and_move(page, newpage, force, mode);
1201 0 : if (rc == MIGRATEPAGE_SUCCESS)
1202 0 : set_page_owner_migrate_reason(newpage, reason);
1203 :
1204 0 : out:
1205 0 : if (rc != -EAGAIN) {
1206 : /*
1207 : * A page that has been migrated has all references
1208 : * removed and will be freed. A page that has not been
1209 : * migrated will have kept its references and be restored.
1210 : */
1211 0 : list_del(&page->lru);
1212 : }
1213 :
1214 : /*
1215 : * If migration is successful, releases reference grabbed during
1216 : * isolation. Otherwise, restore the page to right list unless
1217 : * we want to retry.
1218 : */
1219 0 : if (rc == MIGRATEPAGE_SUCCESS) {
1220 : /*
1221 : * Compaction can migrate also non-LRU pages which are
1222 : * not accounted to NR_ISOLATED_*. They can be recognized
1223 : * as __PageMovable
1224 : */
1225 0 : if (likely(!__PageMovable(page)))
1226 0 : mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1227 0 : page_is_file_lru(page), -thp_nr_pages(page));
1228 :
1229 0 : if (reason != MR_MEMORY_FAILURE)
1230 : /*
1231 : * We release the page in page_handle_poison.
1232 : */
1233 0 : put_page(page);
1234 : } else {
1235 0 : if (rc != -EAGAIN)
1236 0 : list_add_tail(&page->lru, ret);
1237 :
1238 0 : if (put_new_page)
1239 0 : put_new_page(newpage, private);
1240 : else
1241 0 : put_page(newpage);
1242 : }
1243 :
1244 : return rc;
1245 : }
1246 :
1247 : /*
1248 : * Counterpart of unmap_and_move_page() for hugepage migration.
1249 : *
1250 : * This function doesn't wait the completion of hugepage I/O
1251 : * because there is no race between I/O and migration for hugepage.
1252 : * Note that currently hugepage I/O occurs only in direct I/O
1253 : * where no lock is held and PG_writeback is irrelevant,
1254 : * and writeback status of all subpages are counted in the reference
1255 : * count of the head page (i.e. if all subpages of a 2MB hugepage are
1256 : * under direct I/O, the reference of the head page is 512 and a bit more.)
1257 : * This means that when we try to migrate hugepage whose subpages are
1258 : * doing direct I/O, some references remain after try_to_unmap() and
1259 : * hugepage migration fails without data corruption.
1260 : *
1261 : * There is also no race when direct I/O is issued on the page under migration,
1262 : * because then pte is replaced with migration swap entry and direct I/O code
1263 : * will wait in the page fault for migration to complete.
1264 : */
1265 : static int unmap_and_move_huge_page(new_page_t get_new_page,
1266 : free_page_t put_new_page, unsigned long private,
1267 : struct page *hpage, int force,
1268 : enum migrate_mode mode, int reason,
1269 : struct list_head *ret)
1270 : {
1271 : int rc = -EAGAIN;
1272 : int page_was_mapped = 0;
1273 : struct page *new_hpage;
1274 : struct anon_vma *anon_vma = NULL;
1275 : struct address_space *mapping = NULL;
1276 :
1277 : /*
1278 : * Migratability of hugepages depends on architectures and their size.
1279 : * This check is necessary because some callers of hugepage migration
1280 : * like soft offline and memory hotremove don't walk through page
1281 : * tables or check whether the hugepage is pmd-based or not before
1282 : * kicking migration.
1283 : */
1284 : if (!hugepage_migration_supported(page_hstate(hpage))) {
1285 : list_move_tail(&hpage->lru, ret);
1286 : return -ENOSYS;
1287 : }
1288 :
1289 : if (page_count(hpage) == 1) {
1290 : /* page was freed from under us. So we are done. */
1291 : putback_active_hugepage(hpage);
1292 : return MIGRATEPAGE_SUCCESS;
1293 : }
1294 :
1295 : new_hpage = get_new_page(hpage, private);
1296 : if (!new_hpage)
1297 : return -ENOMEM;
1298 :
1299 : if (!trylock_page(hpage)) {
1300 : if (!force)
1301 : goto out;
1302 : switch (mode) {
1303 : case MIGRATE_SYNC:
1304 : case MIGRATE_SYNC_NO_COPY:
1305 : break;
1306 : default:
1307 : goto out;
1308 : }
1309 : lock_page(hpage);
1310 : }
1311 :
1312 : /*
1313 : * Check for pages which are in the process of being freed. Without
1314 : * page_mapping() set, hugetlbfs specific move page routine will not
1315 : * be called and we could leak usage counts for subpools.
1316 : */
1317 : if (page_private(hpage) && !page_mapping(hpage)) {
1318 : rc = -EBUSY;
1319 : goto out_unlock;
1320 : }
1321 :
1322 : if (PageAnon(hpage))
1323 : anon_vma = page_get_anon_vma(hpage);
1324 :
1325 : if (unlikely(!trylock_page(new_hpage)))
1326 : goto put_anon;
1327 :
1328 : if (page_mapped(hpage)) {
1329 : bool mapping_locked = false;
1330 : enum ttu_flags ttu = TTU_MIGRATION|TTU_IGNORE_MLOCK;
1331 :
1332 : if (!PageAnon(hpage)) {
1333 : /*
1334 : * In shared mappings, try_to_unmap could potentially
1335 : * call huge_pmd_unshare. Because of this, take
1336 : * semaphore in write mode here and set TTU_RMAP_LOCKED
1337 : * to let lower levels know we have taken the lock.
1338 : */
1339 : mapping = hugetlb_page_mapping_lock_write(hpage);
1340 : if (unlikely(!mapping))
1341 : goto unlock_put_anon;
1342 :
1343 : mapping_locked = true;
1344 : ttu |= TTU_RMAP_LOCKED;
1345 : }
1346 :
1347 : try_to_unmap(hpage, ttu);
1348 : page_was_mapped = 1;
1349 :
1350 : if (mapping_locked)
1351 : i_mmap_unlock_write(mapping);
1352 : }
1353 :
1354 : if (!page_mapped(hpage))
1355 : rc = move_to_new_page(new_hpage, hpage, mode);
1356 :
1357 : if (page_was_mapped)
1358 : remove_migration_ptes(hpage,
1359 : rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1360 :
1361 : unlock_put_anon:
1362 : unlock_page(new_hpage);
1363 :
1364 : put_anon:
1365 : if (anon_vma)
1366 : put_anon_vma(anon_vma);
1367 :
1368 : if (rc == MIGRATEPAGE_SUCCESS) {
1369 : move_hugetlb_state(hpage, new_hpage, reason);
1370 : put_new_page = NULL;
1371 : }
1372 :
1373 : out_unlock:
1374 : unlock_page(hpage);
1375 : out:
1376 : if (rc == MIGRATEPAGE_SUCCESS)
1377 : putback_active_hugepage(hpage);
1378 : else if (rc != -EAGAIN && rc != MIGRATEPAGE_SUCCESS)
1379 : list_move_tail(&hpage->lru, ret);
1380 :
1381 : /*
1382 : * If migration was not successful and there's a freeing callback, use
1383 : * it. Otherwise, put_page() will drop the reference grabbed during
1384 : * isolation.
1385 : */
1386 : if (put_new_page)
1387 : put_new_page(new_hpage, private);
1388 : else
1389 : putback_active_hugepage(new_hpage);
1390 :
1391 : return rc;
1392 : }
1393 :
1394 0 : static inline int try_split_thp(struct page *page, struct page **page2,
1395 : struct list_head *from)
1396 : {
1397 0 : int rc = 0;
1398 :
1399 0 : lock_page(page);
1400 0 : rc = split_huge_page_to_list(page, from);
1401 0 : unlock_page(page);
1402 0 : if (!rc)
1403 0 : list_safe_reset_next(page, *page2, lru);
1404 :
1405 0 : return rc;
1406 : }
1407 :
1408 : /*
1409 : * migrate_pages - migrate the pages specified in a list, to the free pages
1410 : * supplied as the target for the page migration
1411 : *
1412 : * @from: The list of pages to be migrated.
1413 : * @get_new_page: The function used to allocate free pages to be used
1414 : * as the target of the page migration.
1415 : * @put_new_page: The function used to free target pages if migration
1416 : * fails, or NULL if no special handling is necessary.
1417 : * @private: Private data to be passed on to get_new_page()
1418 : * @mode: The migration mode that specifies the constraints for
1419 : * page migration, if any.
1420 : * @reason: The reason for page migration.
1421 : *
1422 : * The function returns after 10 attempts or if no pages are movable any more
1423 : * because the list has become empty or no retryable pages exist any more.
1424 : * It is caller's responsibility to call putback_movable_pages() to return pages
1425 : * to the LRU or free list only if ret != 0.
1426 : *
1427 : * Returns the number of pages that were not migrated, or an error code.
1428 : */
1429 0 : int migrate_pages(struct list_head *from, new_page_t get_new_page,
1430 : free_page_t put_new_page, unsigned long private,
1431 : enum migrate_mode mode, int reason)
1432 : {
1433 0 : int retry = 1;
1434 0 : int thp_retry = 1;
1435 0 : int nr_failed = 0;
1436 0 : int nr_succeeded = 0;
1437 0 : int nr_thp_succeeded = 0;
1438 0 : int nr_thp_failed = 0;
1439 0 : int nr_thp_split = 0;
1440 0 : int pass = 0;
1441 0 : bool is_thp = false;
1442 0 : struct page *page;
1443 0 : struct page *page2;
1444 0 : int swapwrite = current->flags & PF_SWAPWRITE;
1445 0 : int rc, nr_subpages;
1446 0 : LIST_HEAD(ret_pages);
1447 :
1448 0 : if (!swapwrite)
1449 0 : current->flags |= PF_SWAPWRITE;
1450 :
1451 0 : for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1452 0 : retry = 0;
1453 0 : thp_retry = 0;
1454 :
1455 0 : list_for_each_entry_safe(page, page2, from, lru) {
1456 0 : retry:
1457 : /*
1458 : * THP statistics is based on the source huge page.
1459 : * Capture required information that might get lost
1460 : * during migration.
1461 : */
1462 0 : is_thp = PageTransHuge(page) && !PageHuge(page);
1463 0 : nr_subpages = thp_nr_pages(page);
1464 0 : cond_resched();
1465 :
1466 0 : if (PageHuge(page))
1467 : rc = unmap_and_move_huge_page(get_new_page,
1468 : put_new_page, private, page,
1469 : pass > 2, mode, reason,
1470 : &ret_pages);
1471 : else
1472 0 : rc = unmap_and_move(get_new_page, put_new_page,
1473 : private, page, pass > 2, mode,
1474 : reason, &ret_pages);
1475 : /*
1476 : * The rules are:
1477 : * Success: non hugetlb page will be freed, hugetlb
1478 : * page will be put back
1479 : * -EAGAIN: stay on the from list
1480 : * -ENOMEM: stay on the from list
1481 : * Other errno: put on ret_pages list then splice to
1482 : * from list
1483 : */
1484 0 : switch(rc) {
1485 : /*
1486 : * THP migration might be unsupported or the
1487 : * allocation could've failed so we should
1488 : * retry on the same page with the THP split
1489 : * to base pages.
1490 : *
1491 : * Head page is retried immediately and tail
1492 : * pages are added to the tail of the list so
1493 : * we encounter them after the rest of the list
1494 : * is processed.
1495 : */
1496 0 : case -ENOSYS:
1497 : /* THP migration is unsupported */
1498 0 : if (is_thp) {
1499 0 : if (!try_split_thp(page, &page2, from)) {
1500 0 : nr_thp_split++;
1501 0 : goto retry;
1502 : }
1503 :
1504 0 : nr_thp_failed++;
1505 0 : nr_failed += nr_subpages;
1506 0 : break;
1507 : }
1508 :
1509 : /* Hugetlb migration is unsupported */
1510 0 : nr_failed++;
1511 0 : break;
1512 0 : case -ENOMEM:
1513 : /*
1514 : * When memory is low, don't bother to try to migrate
1515 : * other pages, just exit.
1516 : */
1517 0 : if (is_thp) {
1518 0 : if (!try_split_thp(page, &page2, from)) {
1519 0 : nr_thp_split++;
1520 0 : goto retry;
1521 : }
1522 :
1523 0 : nr_thp_failed++;
1524 0 : nr_failed += nr_subpages;
1525 0 : goto out;
1526 : }
1527 0 : nr_failed++;
1528 0 : goto out;
1529 0 : case -EAGAIN:
1530 0 : if (is_thp) {
1531 0 : thp_retry++;
1532 0 : break;
1533 : }
1534 0 : retry++;
1535 0 : break;
1536 0 : case MIGRATEPAGE_SUCCESS:
1537 0 : if (is_thp) {
1538 0 : nr_thp_succeeded++;
1539 0 : nr_succeeded += nr_subpages;
1540 0 : break;
1541 : }
1542 0 : nr_succeeded++;
1543 0 : break;
1544 0 : default:
1545 : /*
1546 : * Permanent failure (-EBUSY, etc.):
1547 : * unlike -EAGAIN case, the failed page is
1548 : * removed from migration page list and not
1549 : * retried in the next outer loop.
1550 : */
1551 0 : if (is_thp) {
1552 0 : nr_thp_failed++;
1553 0 : nr_failed += nr_subpages;
1554 0 : break;
1555 : }
1556 0 : nr_failed++;
1557 0 : break;
1558 : }
1559 : }
1560 : }
1561 0 : nr_failed += retry + thp_retry;
1562 0 : nr_thp_failed += thp_retry;
1563 0 : rc = nr_failed;
1564 0 : out:
1565 : /*
1566 : * Put the permanent failure page back to migration list, they
1567 : * will be put back to the right list by the caller.
1568 : */
1569 0 : list_splice(&ret_pages, from);
1570 :
1571 0 : count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1572 0 : count_vm_events(PGMIGRATE_FAIL, nr_failed);
1573 0 : count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1574 0 : count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1575 0 : count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1576 0 : trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded,
1577 : nr_thp_failed, nr_thp_split, mode, reason);
1578 :
1579 0 : if (!swapwrite)
1580 0 : current->flags &= ~PF_SWAPWRITE;
1581 :
1582 0 : return rc;
1583 : }
1584 :
1585 0 : struct page *alloc_migration_target(struct page *page, unsigned long private)
1586 : {
1587 0 : struct migration_target_control *mtc;
1588 0 : gfp_t gfp_mask;
1589 0 : unsigned int order = 0;
1590 0 : struct page *new_page = NULL;
1591 0 : int nid;
1592 0 : int zidx;
1593 :
1594 0 : mtc = (struct migration_target_control *)private;
1595 0 : gfp_mask = mtc->gfp_mask;
1596 0 : nid = mtc->nid;
1597 0 : if (nid == NUMA_NO_NODE)
1598 0 : nid = page_to_nid(page);
1599 :
1600 0 : if (PageHuge(page)) {
1601 : struct hstate *h = page_hstate(compound_head(page));
1602 :
1603 : gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1604 : return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1605 : }
1606 :
1607 0 : if (PageTransHuge(page)) {
1608 : /*
1609 : * clear __GFP_RECLAIM to make the migration callback
1610 : * consistent with regular THP allocations.
1611 : */
1612 0 : gfp_mask &= ~__GFP_RECLAIM;
1613 0 : gfp_mask |= GFP_TRANSHUGE;
1614 0 : order = HPAGE_PMD_ORDER;
1615 : }
1616 0 : zidx = zone_idx(page_zone(page));
1617 0 : if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1618 0 : gfp_mask |= __GFP_HIGHMEM;
1619 :
1620 0 : new_page = __alloc_pages_nodemask(gfp_mask, order, nid, mtc->nmask);
1621 :
1622 0 : if (new_page && PageTransHuge(new_page))
1623 0 : prep_transhuge_page(new_page);
1624 :
1625 0 : return new_page;
1626 : }
1627 :
1628 : #ifdef CONFIG_NUMA
1629 :
1630 0 : static int store_status(int __user *status, int start, int value, int nr)
1631 : {
1632 0 : while (nr-- > 0) {
1633 0 : if (put_user(value, status + start))
1634 : return -EFAULT;
1635 0 : start++;
1636 : }
1637 :
1638 : return 0;
1639 : }
1640 :
1641 0 : static int do_move_pages_to_node(struct mm_struct *mm,
1642 : struct list_head *pagelist, int node)
1643 : {
1644 0 : int err;
1645 0 : struct migration_target_control mtc = {
1646 : .nid = node,
1647 : .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1648 : };
1649 :
1650 0 : err = migrate_pages(pagelist, alloc_migration_target, NULL,
1651 : (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1652 0 : if (err)
1653 0 : putback_movable_pages(pagelist);
1654 0 : return err;
1655 : }
1656 :
1657 : /*
1658 : * Resolves the given address to a struct page, isolates it from the LRU and
1659 : * puts it to the given pagelist.
1660 : * Returns:
1661 : * errno - if the page cannot be found/isolated
1662 : * 0 - when it doesn't have to be migrated because it is already on the
1663 : * target node
1664 : * 1 - when it has been queued
1665 : */
1666 0 : static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1667 : int node, struct list_head *pagelist, bool migrate_all)
1668 : {
1669 0 : struct vm_area_struct *vma;
1670 0 : struct page *page;
1671 0 : unsigned int follflags;
1672 0 : int err;
1673 :
1674 0 : mmap_read_lock(mm);
1675 0 : err = -EFAULT;
1676 0 : vma = find_vma(mm, addr);
1677 0 : if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1678 0 : goto out;
1679 :
1680 : /* FOLL_DUMP to ignore special (like zero) pages */
1681 0 : follflags = FOLL_GET | FOLL_DUMP;
1682 0 : page = follow_page(vma, addr, follflags);
1683 :
1684 0 : err = PTR_ERR(page);
1685 0 : if (IS_ERR(page))
1686 0 : goto out;
1687 :
1688 0 : err = -ENOENT;
1689 0 : if (!page)
1690 0 : goto out;
1691 :
1692 0 : err = 0;
1693 0 : if (page_to_nid(page) == node)
1694 0 : goto out_putpage;
1695 :
1696 0 : err = -EACCES;
1697 0 : if (page_mapcount(page) > 1 && !migrate_all)
1698 0 : goto out_putpage;
1699 :
1700 0 : if (PageHuge(page)) {
1701 : if (PageHead(page)) {
1702 : isolate_huge_page(page, pagelist);
1703 : err = 1;
1704 : }
1705 : } else {
1706 0 : struct page *head;
1707 :
1708 0 : head = compound_head(page);
1709 0 : err = isolate_lru_page(head);
1710 0 : if (err)
1711 0 : goto out_putpage;
1712 :
1713 0 : err = 1;
1714 0 : list_add_tail(&head->lru, pagelist);
1715 0 : mod_node_page_state(page_pgdat(head),
1716 0 : NR_ISOLATED_ANON + page_is_file_lru(head),
1717 0 : thp_nr_pages(head));
1718 : }
1719 0 : out_putpage:
1720 : /*
1721 : * Either remove the duplicate refcount from
1722 : * isolate_lru_page() or drop the page ref if it was
1723 : * not isolated.
1724 : */
1725 0 : put_page(page);
1726 0 : out:
1727 0 : mmap_read_unlock(mm);
1728 0 : return err;
1729 : }
1730 :
1731 0 : static int move_pages_and_store_status(struct mm_struct *mm, int node,
1732 : struct list_head *pagelist, int __user *status,
1733 : int start, int i, unsigned long nr_pages)
1734 : {
1735 0 : int err;
1736 :
1737 0 : if (list_empty(pagelist))
1738 : return 0;
1739 :
1740 0 : err = do_move_pages_to_node(mm, pagelist, node);
1741 0 : if (err) {
1742 : /*
1743 : * Positive err means the number of failed
1744 : * pages to migrate. Since we are going to
1745 : * abort and return the number of non-migrated
1746 : * pages, so need to include the rest of the
1747 : * nr_pages that have not been attempted as
1748 : * well.
1749 : */
1750 0 : if (err > 0)
1751 0 : err += nr_pages - i - 1;
1752 0 : return err;
1753 : }
1754 0 : return store_status(status, start, node, i - start);
1755 : }
1756 :
1757 : /*
1758 : * Migrate an array of page address onto an array of nodes and fill
1759 : * the corresponding array of status.
1760 : */
1761 0 : static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1762 : unsigned long nr_pages,
1763 : const void __user * __user *pages,
1764 : const int __user *nodes,
1765 : int __user *status, int flags)
1766 : {
1767 0 : int current_node = NUMA_NO_NODE;
1768 0 : LIST_HEAD(pagelist);
1769 0 : int start, i;
1770 0 : int err = 0, err1;
1771 :
1772 0 : migrate_prep();
1773 :
1774 0 : for (i = start = 0; i < nr_pages; i++) {
1775 0 : const void __user *p;
1776 0 : unsigned long addr;
1777 0 : int node;
1778 :
1779 0 : err = -EFAULT;
1780 0 : if (get_user(p, pages + i))
1781 0 : goto out_flush;
1782 0 : if (get_user(node, nodes + i))
1783 0 : goto out_flush;
1784 0 : addr = (unsigned long)untagged_addr(p);
1785 :
1786 0 : err = -ENODEV;
1787 0 : if (node < 0 || node >= MAX_NUMNODES)
1788 0 : goto out_flush;
1789 0 : if (!node_state(node, N_MEMORY))
1790 0 : goto out_flush;
1791 :
1792 0 : err = -EACCES;
1793 0 : if (!node_isset(node, task_nodes))
1794 0 : goto out_flush;
1795 :
1796 0 : if (current_node == NUMA_NO_NODE) {
1797 : current_node = node;
1798 : start = i;
1799 0 : } else if (node != current_node) {
1800 0 : err = move_pages_and_store_status(mm, current_node,
1801 : &pagelist, status, start, i, nr_pages);
1802 0 : if (err)
1803 0 : goto out;
1804 : start = i;
1805 : current_node = node;
1806 : }
1807 :
1808 : /*
1809 : * Errors in the page lookup or isolation are not fatal and we simply
1810 : * report them via status
1811 : */
1812 0 : err = add_page_for_migration(mm, addr, current_node,
1813 0 : &pagelist, flags & MPOL_MF_MOVE_ALL);
1814 :
1815 0 : if (err > 0) {
1816 : /* The page is successfully queued for migration */
1817 0 : continue;
1818 : }
1819 :
1820 : /*
1821 : * If the page is already on the target node (!err), store the
1822 : * node, otherwise, store the err.
1823 : */
1824 0 : err = store_status(status, i, err ? : current_node, 1);
1825 0 : if (err)
1826 0 : goto out_flush;
1827 :
1828 0 : err = move_pages_and_store_status(mm, current_node, &pagelist,
1829 : status, start, i, nr_pages);
1830 0 : if (err)
1831 0 : goto out;
1832 : current_node = NUMA_NO_NODE;
1833 : }
1834 0 : out_flush:
1835 : /* Make sure we do not overwrite the existing error */
1836 0 : err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1837 : status, start, i, nr_pages);
1838 0 : if (err >= 0)
1839 0 : err = err1;
1840 0 : out:
1841 0 : return err;
1842 : }
1843 :
1844 : /*
1845 : * Determine the nodes of an array of pages and store it in an array of status.
1846 : */
1847 0 : static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1848 : const void __user **pages, int *status)
1849 : {
1850 0 : unsigned long i;
1851 :
1852 0 : mmap_read_lock(mm);
1853 :
1854 0 : for (i = 0; i < nr_pages; i++) {
1855 0 : unsigned long addr = (unsigned long)(*pages);
1856 0 : struct vm_area_struct *vma;
1857 0 : struct page *page;
1858 0 : int err = -EFAULT;
1859 :
1860 0 : vma = find_vma(mm, addr);
1861 0 : if (!vma || addr < vma->vm_start)
1862 0 : goto set_status;
1863 :
1864 : /* FOLL_DUMP to ignore special (like zero) pages */
1865 0 : page = follow_page(vma, addr, FOLL_DUMP);
1866 :
1867 0 : err = PTR_ERR(page);
1868 0 : if (IS_ERR(page))
1869 0 : goto set_status;
1870 :
1871 0 : err = page ? page_to_nid(page) : -ENOENT;
1872 0 : set_status:
1873 0 : *status = err;
1874 :
1875 0 : pages++;
1876 0 : status++;
1877 : }
1878 :
1879 0 : mmap_read_unlock(mm);
1880 0 : }
1881 :
1882 : /*
1883 : * Determine the nodes of a user array of pages and store it in
1884 : * a user array of status.
1885 : */
1886 0 : static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1887 : const void __user * __user *pages,
1888 : int __user *status)
1889 : {
1890 : #define DO_PAGES_STAT_CHUNK_NR 16
1891 0 : const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1892 0 : int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1893 :
1894 0 : while (nr_pages) {
1895 0 : unsigned long chunk_nr;
1896 :
1897 0 : chunk_nr = nr_pages;
1898 0 : if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1899 : chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1900 :
1901 0 : if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1902 : break;
1903 :
1904 0 : do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1905 :
1906 0 : if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1907 : break;
1908 :
1909 0 : pages += chunk_nr;
1910 0 : status += chunk_nr;
1911 0 : nr_pages -= chunk_nr;
1912 : }
1913 0 : return nr_pages ? -EFAULT : 0;
1914 : }
1915 :
1916 0 : static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1917 : {
1918 0 : struct task_struct *task;
1919 0 : struct mm_struct *mm;
1920 :
1921 : /*
1922 : * There is no need to check if current process has the right to modify
1923 : * the specified process when they are same.
1924 : */
1925 0 : if (!pid) {
1926 0 : mmget(current->mm);
1927 0 : *mem_nodes = cpuset_mems_allowed(current);
1928 0 : return current->mm;
1929 : }
1930 :
1931 : /* Find the mm_struct */
1932 0 : rcu_read_lock();
1933 0 : task = find_task_by_vpid(pid);
1934 0 : if (!task) {
1935 0 : rcu_read_unlock();
1936 0 : return ERR_PTR(-ESRCH);
1937 : }
1938 0 : get_task_struct(task);
1939 :
1940 : /*
1941 : * Check if this process has the right to modify the specified
1942 : * process. Use the regular "ptrace_may_access()" checks.
1943 : */
1944 0 : if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1945 0 : rcu_read_unlock();
1946 0 : mm = ERR_PTR(-EPERM);
1947 0 : goto out;
1948 : }
1949 0 : rcu_read_unlock();
1950 :
1951 0 : mm = ERR_PTR(security_task_movememory(task));
1952 0 : if (IS_ERR(mm))
1953 0 : goto out;
1954 0 : *mem_nodes = cpuset_mems_allowed(task);
1955 0 : mm = get_task_mm(task);
1956 0 : out:
1957 0 : put_task_struct(task);
1958 0 : if (!mm)
1959 0 : mm = ERR_PTR(-EINVAL);
1960 : return mm;
1961 : }
1962 :
1963 : /*
1964 : * Move a list of pages in the address space of the currently executing
1965 : * process.
1966 : */
1967 0 : static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1968 : const void __user * __user *pages,
1969 : const int __user *nodes,
1970 : int __user *status, int flags)
1971 : {
1972 0 : struct mm_struct *mm;
1973 0 : int err;
1974 0 : nodemask_t task_nodes;
1975 :
1976 : /* Check flags */
1977 0 : if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1978 : return -EINVAL;
1979 :
1980 0 : if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1981 : return -EPERM;
1982 :
1983 0 : mm = find_mm_struct(pid, &task_nodes);
1984 0 : if (IS_ERR(mm))
1985 0 : return PTR_ERR(mm);
1986 :
1987 0 : if (nodes)
1988 0 : err = do_pages_move(mm, task_nodes, nr_pages, pages,
1989 : nodes, status, flags);
1990 : else
1991 0 : err = do_pages_stat(mm, nr_pages, pages, status);
1992 :
1993 0 : mmput(mm);
1994 0 : return err;
1995 : }
1996 :
1997 0 : SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1998 : const void __user * __user *, pages,
1999 : const int __user *, nodes,
2000 : int __user *, status, int, flags)
2001 : {
2002 0 : return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2003 : }
2004 :
2005 : #ifdef CONFIG_COMPAT
2006 0 : COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
2007 : compat_uptr_t __user *, pages32,
2008 : const int __user *, nodes,
2009 : int __user *, status,
2010 : int, flags)
2011 : {
2012 0 : const void __user * __user *pages;
2013 0 : int i;
2014 :
2015 0 : pages = compat_alloc_user_space(nr_pages * sizeof(void *));
2016 0 : for (i = 0; i < nr_pages; i++) {
2017 0 : compat_uptr_t p;
2018 :
2019 0 : if (get_user(p, pages32 + i) ||
2020 0 : put_user(compat_ptr(p), pages + i))
2021 0 : return -EFAULT;
2022 : }
2023 0 : return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2024 : }
2025 : #endif /* CONFIG_COMPAT */
2026 :
2027 : #ifdef CONFIG_NUMA_BALANCING
2028 : /*
2029 : * Returns true if this is a safe migration target node for misplaced NUMA
2030 : * pages. Currently it only checks the watermarks which crude
2031 : */
2032 : static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2033 : unsigned long nr_migrate_pages)
2034 : {
2035 : int z;
2036 :
2037 : for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2038 : struct zone *zone = pgdat->node_zones + z;
2039 :
2040 : if (!populated_zone(zone))
2041 : continue;
2042 :
2043 : /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2044 : if (!zone_watermark_ok(zone, 0,
2045 : high_wmark_pages(zone) +
2046 : nr_migrate_pages,
2047 : ZONE_MOVABLE, 0))
2048 : continue;
2049 : return true;
2050 : }
2051 : return false;
2052 : }
2053 :
2054 : static struct page *alloc_misplaced_dst_page(struct page *page,
2055 : unsigned long data)
2056 : {
2057 : int nid = (int) data;
2058 : struct page *newpage;
2059 :
2060 : newpage = __alloc_pages_node(nid,
2061 : (GFP_HIGHUSER_MOVABLE |
2062 : __GFP_THISNODE | __GFP_NOMEMALLOC |
2063 : __GFP_NORETRY | __GFP_NOWARN) &
2064 : ~__GFP_RECLAIM, 0);
2065 :
2066 : return newpage;
2067 : }
2068 :
2069 : static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2070 : {
2071 : int page_lru;
2072 :
2073 : VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
2074 :
2075 : /* Avoid migrating to a node that is nearly full */
2076 : if (!migrate_balanced_pgdat(pgdat, compound_nr(page)))
2077 : return 0;
2078 :
2079 : if (isolate_lru_page(page))
2080 : return 0;
2081 :
2082 : /*
2083 : * migrate_misplaced_transhuge_page() skips page migration's usual
2084 : * check on page_count(), so we must do it here, now that the page
2085 : * has been isolated: a GUP pin, or any other pin, prevents migration.
2086 : * The expected page count is 3: 1 for page's mapcount and 1 for the
2087 : * caller's pin and 1 for the reference taken by isolate_lru_page().
2088 : */
2089 : if (PageTransHuge(page) && page_count(page) != 3) {
2090 : putback_lru_page(page);
2091 : return 0;
2092 : }
2093 :
2094 : page_lru = page_is_file_lru(page);
2095 : mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2096 : thp_nr_pages(page));
2097 :
2098 : /*
2099 : * Isolating the page has taken another reference, so the
2100 : * caller's reference can be safely dropped without the page
2101 : * disappearing underneath us during migration.
2102 : */
2103 : put_page(page);
2104 : return 1;
2105 : }
2106 :
2107 : bool pmd_trans_migrating(pmd_t pmd)
2108 : {
2109 : struct page *page = pmd_page(pmd);
2110 : return PageLocked(page);
2111 : }
2112 :
2113 : static inline bool is_shared_exec_page(struct vm_area_struct *vma,
2114 : struct page *page)
2115 : {
2116 : if (page_mapcount(page) != 1 &&
2117 : (page_is_file_lru(page) || vma_is_shmem(vma)) &&
2118 : (vma->vm_flags & VM_EXEC))
2119 : return true;
2120 :
2121 : return false;
2122 : }
2123 :
2124 : /*
2125 : * Attempt to migrate a misplaced page to the specified destination
2126 : * node. Caller is expected to have an elevated reference count on
2127 : * the page that will be dropped by this function before returning.
2128 : */
2129 : int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2130 : int node)
2131 : {
2132 : pg_data_t *pgdat = NODE_DATA(node);
2133 : int isolated;
2134 : int nr_remaining;
2135 : LIST_HEAD(migratepages);
2136 :
2137 : /*
2138 : * Don't migrate file pages that are mapped in multiple processes
2139 : * with execute permissions as they are probably shared libraries.
2140 : */
2141 : if (is_shared_exec_page(vma, page))
2142 : goto out;
2143 :
2144 : /*
2145 : * Also do not migrate dirty pages as not all filesystems can move
2146 : * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2147 : */
2148 : if (page_is_file_lru(page) && PageDirty(page))
2149 : goto out;
2150 :
2151 : isolated = numamigrate_isolate_page(pgdat, page);
2152 : if (!isolated)
2153 : goto out;
2154 :
2155 : list_add(&page->lru, &migratepages);
2156 : nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2157 : NULL, node, MIGRATE_ASYNC,
2158 : MR_NUMA_MISPLACED);
2159 : if (nr_remaining) {
2160 : if (!list_empty(&migratepages)) {
2161 : list_del(&page->lru);
2162 : dec_node_page_state(page, NR_ISOLATED_ANON +
2163 : page_is_file_lru(page));
2164 : putback_lru_page(page);
2165 : }
2166 : isolated = 0;
2167 : } else
2168 : count_vm_numa_event(NUMA_PAGE_MIGRATE);
2169 : BUG_ON(!list_empty(&migratepages));
2170 : return isolated;
2171 :
2172 : out:
2173 : put_page(page);
2174 : return 0;
2175 : }
2176 : #endif /* CONFIG_NUMA_BALANCING */
2177 :
2178 : #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2179 : /*
2180 : * Migrates a THP to a given target node. page must be locked and is unlocked
2181 : * before returning.
2182 : */
2183 : int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2184 : struct vm_area_struct *vma,
2185 : pmd_t *pmd, pmd_t entry,
2186 : unsigned long address,
2187 : struct page *page, int node)
2188 : {
2189 : spinlock_t *ptl;
2190 : pg_data_t *pgdat = NODE_DATA(node);
2191 : int isolated = 0;
2192 : struct page *new_page = NULL;
2193 : int page_lru = page_is_file_lru(page);
2194 : unsigned long start = address & HPAGE_PMD_MASK;
2195 :
2196 : if (is_shared_exec_page(vma, page))
2197 : goto out;
2198 :
2199 : new_page = alloc_pages_node(node,
2200 : (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2201 : HPAGE_PMD_ORDER);
2202 : if (!new_page)
2203 : goto out_fail;
2204 : prep_transhuge_page(new_page);
2205 :
2206 : isolated = numamigrate_isolate_page(pgdat, page);
2207 : if (!isolated) {
2208 : put_page(new_page);
2209 : goto out_fail;
2210 : }
2211 :
2212 : /* Prepare a page as a migration target */
2213 : __SetPageLocked(new_page);
2214 : if (PageSwapBacked(page))
2215 : __SetPageSwapBacked(new_page);
2216 :
2217 : /* anon mapping, we can simply copy page->mapping to the new page: */
2218 : new_page->mapping = page->mapping;
2219 : new_page->index = page->index;
2220 : /* flush the cache before copying using the kernel virtual address */
2221 : flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
2222 : migrate_page_copy(new_page, page);
2223 : WARN_ON(PageLRU(new_page));
2224 :
2225 : /* Recheck the target PMD */
2226 : ptl = pmd_lock(mm, pmd);
2227 : if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2228 : spin_unlock(ptl);
2229 :
2230 : /* Reverse changes made by migrate_page_copy() */
2231 : if (TestClearPageActive(new_page))
2232 : SetPageActive(page);
2233 : if (TestClearPageUnevictable(new_page))
2234 : SetPageUnevictable(page);
2235 :
2236 : unlock_page(new_page);
2237 : put_page(new_page); /* Free it */
2238 :
2239 : /* Retake the callers reference and putback on LRU */
2240 : get_page(page);
2241 : putback_lru_page(page);
2242 : mod_node_page_state(page_pgdat(page),
2243 : NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2244 :
2245 : goto out_unlock;
2246 : }
2247 :
2248 : entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2249 : entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2250 :
2251 : /*
2252 : * Overwrite the old entry under pagetable lock and establish
2253 : * the new PTE. Any parallel GUP will either observe the old
2254 : * page blocking on the page lock, block on the page table
2255 : * lock or observe the new page. The SetPageUptodate on the
2256 : * new page and page_add_new_anon_rmap guarantee the copy is
2257 : * visible before the pagetable update.
2258 : */
2259 : page_add_anon_rmap(new_page, vma, start, true);
2260 : /*
2261 : * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2262 : * has already been flushed globally. So no TLB can be currently
2263 : * caching this non present pmd mapping. There's no need to clear the
2264 : * pmd before doing set_pmd_at(), nor to flush the TLB after
2265 : * set_pmd_at(). Clearing the pmd here would introduce a race
2266 : * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2267 : * mmap_lock for reading. If the pmd is set to NULL at any given time,
2268 : * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2269 : * pmd.
2270 : */
2271 : set_pmd_at(mm, start, pmd, entry);
2272 : update_mmu_cache_pmd(vma, address, &entry);
2273 :
2274 : page_ref_unfreeze(page, 2);
2275 : mlock_migrate_page(new_page, page);
2276 : page_remove_rmap(page, true);
2277 : set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2278 :
2279 : spin_unlock(ptl);
2280 :
2281 : /* Take an "isolate" reference and put new page on the LRU. */
2282 : get_page(new_page);
2283 : putback_lru_page(new_page);
2284 :
2285 : unlock_page(new_page);
2286 : unlock_page(page);
2287 : put_page(page); /* Drop the rmap reference */
2288 : put_page(page); /* Drop the LRU isolation reference */
2289 :
2290 : count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2291 : count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2292 :
2293 : mod_node_page_state(page_pgdat(page),
2294 : NR_ISOLATED_ANON + page_lru,
2295 : -HPAGE_PMD_NR);
2296 : return isolated;
2297 :
2298 : out_fail:
2299 : count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2300 : ptl = pmd_lock(mm, pmd);
2301 : if (pmd_same(*pmd, entry)) {
2302 : entry = pmd_modify(entry, vma->vm_page_prot);
2303 : set_pmd_at(mm, start, pmd, entry);
2304 : update_mmu_cache_pmd(vma, address, &entry);
2305 : }
2306 : spin_unlock(ptl);
2307 :
2308 : out_unlock:
2309 : unlock_page(page);
2310 : out:
2311 : put_page(page);
2312 : return 0;
2313 : }
2314 : #endif /* CONFIG_NUMA_BALANCING */
2315 :
2316 : #endif /* CONFIG_NUMA */
2317 :
2318 : #ifdef CONFIG_DEVICE_PRIVATE
2319 : static int migrate_vma_collect_hole(unsigned long start,
2320 : unsigned long end,
2321 : __always_unused int depth,
2322 : struct mm_walk *walk)
2323 : {
2324 : struct migrate_vma *migrate = walk->private;
2325 : unsigned long addr;
2326 :
2327 : /* Only allow populating anonymous memory. */
2328 : if (!vma_is_anonymous(walk->vma)) {
2329 : for (addr = start; addr < end; addr += PAGE_SIZE) {
2330 : migrate->src[migrate->npages] = 0;
2331 : migrate->dst[migrate->npages] = 0;
2332 : migrate->npages++;
2333 : }
2334 : return 0;
2335 : }
2336 :
2337 : for (addr = start; addr < end; addr += PAGE_SIZE) {
2338 : migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2339 : migrate->dst[migrate->npages] = 0;
2340 : migrate->npages++;
2341 : migrate->cpages++;
2342 : }
2343 :
2344 : return 0;
2345 : }
2346 :
2347 : static int migrate_vma_collect_skip(unsigned long start,
2348 : unsigned long end,
2349 : struct mm_walk *walk)
2350 : {
2351 : struct migrate_vma *migrate = walk->private;
2352 : unsigned long addr;
2353 :
2354 : for (addr = start; addr < end; addr += PAGE_SIZE) {
2355 : migrate->dst[migrate->npages] = 0;
2356 : migrate->src[migrate->npages++] = 0;
2357 : }
2358 :
2359 : return 0;
2360 : }
2361 :
2362 : static int migrate_vma_collect_pmd(pmd_t *pmdp,
2363 : unsigned long start,
2364 : unsigned long end,
2365 : struct mm_walk *walk)
2366 : {
2367 : struct migrate_vma *migrate = walk->private;
2368 : struct vm_area_struct *vma = walk->vma;
2369 : struct mm_struct *mm = vma->vm_mm;
2370 : unsigned long addr = start, unmapped = 0;
2371 : spinlock_t *ptl;
2372 : pte_t *ptep;
2373 :
2374 : again:
2375 : if (pmd_none(*pmdp))
2376 : return migrate_vma_collect_hole(start, end, -1, walk);
2377 :
2378 : if (pmd_trans_huge(*pmdp)) {
2379 : struct page *page;
2380 :
2381 : ptl = pmd_lock(mm, pmdp);
2382 : if (unlikely(!pmd_trans_huge(*pmdp))) {
2383 : spin_unlock(ptl);
2384 : goto again;
2385 : }
2386 :
2387 : page = pmd_page(*pmdp);
2388 : if (is_huge_zero_page(page)) {
2389 : spin_unlock(ptl);
2390 : split_huge_pmd(vma, pmdp, addr);
2391 : if (pmd_trans_unstable(pmdp))
2392 : return migrate_vma_collect_skip(start, end,
2393 : walk);
2394 : } else {
2395 : int ret;
2396 :
2397 : get_page(page);
2398 : spin_unlock(ptl);
2399 : if (unlikely(!trylock_page(page)))
2400 : return migrate_vma_collect_skip(start, end,
2401 : walk);
2402 : ret = split_huge_page(page);
2403 : unlock_page(page);
2404 : put_page(page);
2405 : if (ret)
2406 : return migrate_vma_collect_skip(start, end,
2407 : walk);
2408 : if (pmd_none(*pmdp))
2409 : return migrate_vma_collect_hole(start, end, -1,
2410 : walk);
2411 : }
2412 : }
2413 :
2414 : if (unlikely(pmd_bad(*pmdp)))
2415 : return migrate_vma_collect_skip(start, end, walk);
2416 :
2417 : ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2418 : arch_enter_lazy_mmu_mode();
2419 :
2420 : for (; addr < end; addr += PAGE_SIZE, ptep++) {
2421 : unsigned long mpfn = 0, pfn;
2422 : struct page *page;
2423 : swp_entry_t entry;
2424 : pte_t pte;
2425 :
2426 : pte = *ptep;
2427 :
2428 : if (pte_none(pte)) {
2429 : if (vma_is_anonymous(vma)) {
2430 : mpfn = MIGRATE_PFN_MIGRATE;
2431 : migrate->cpages++;
2432 : }
2433 : goto next;
2434 : }
2435 :
2436 : if (!pte_present(pte)) {
2437 : /*
2438 : * Only care about unaddressable device page special
2439 : * page table entry. Other special swap entries are not
2440 : * migratable, and we ignore regular swapped page.
2441 : */
2442 : entry = pte_to_swp_entry(pte);
2443 : if (!is_device_private_entry(entry))
2444 : goto next;
2445 :
2446 : page = device_private_entry_to_page(entry);
2447 : if (!(migrate->flags &
2448 : MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
2449 : page->pgmap->owner != migrate->pgmap_owner)
2450 : goto next;
2451 :
2452 : mpfn = migrate_pfn(page_to_pfn(page)) |
2453 : MIGRATE_PFN_MIGRATE;
2454 : if (is_write_device_private_entry(entry))
2455 : mpfn |= MIGRATE_PFN_WRITE;
2456 : } else {
2457 : if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
2458 : goto next;
2459 : pfn = pte_pfn(pte);
2460 : if (is_zero_pfn(pfn)) {
2461 : mpfn = MIGRATE_PFN_MIGRATE;
2462 : migrate->cpages++;
2463 : goto next;
2464 : }
2465 : page = vm_normal_page(migrate->vma, addr, pte);
2466 : mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2467 : mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2468 : }
2469 :
2470 : /* FIXME support THP */
2471 : if (!page || !page->mapping || PageTransCompound(page)) {
2472 : mpfn = 0;
2473 : goto next;
2474 : }
2475 :
2476 : /*
2477 : * By getting a reference on the page we pin it and that blocks
2478 : * any kind of migration. Side effect is that it "freezes" the
2479 : * pte.
2480 : *
2481 : * We drop this reference after isolating the page from the lru
2482 : * for non device page (device page are not on the lru and thus
2483 : * can't be dropped from it).
2484 : */
2485 : get_page(page);
2486 : migrate->cpages++;
2487 :
2488 : /*
2489 : * Optimize for the common case where page is only mapped once
2490 : * in one process. If we can lock the page, then we can safely
2491 : * set up a special migration page table entry now.
2492 : */
2493 : if (trylock_page(page)) {
2494 : pte_t swp_pte;
2495 :
2496 : mpfn |= MIGRATE_PFN_LOCKED;
2497 : ptep_get_and_clear(mm, addr, ptep);
2498 :
2499 : /* Setup special migration page table entry */
2500 : entry = make_migration_entry(page, mpfn &
2501 : MIGRATE_PFN_WRITE);
2502 : swp_pte = swp_entry_to_pte(entry);
2503 : if (pte_present(pte)) {
2504 : if (pte_soft_dirty(pte))
2505 : swp_pte = pte_swp_mksoft_dirty(swp_pte);
2506 : if (pte_uffd_wp(pte))
2507 : swp_pte = pte_swp_mkuffd_wp(swp_pte);
2508 : } else {
2509 : if (pte_swp_soft_dirty(pte))
2510 : swp_pte = pte_swp_mksoft_dirty(swp_pte);
2511 : if (pte_swp_uffd_wp(pte))
2512 : swp_pte = pte_swp_mkuffd_wp(swp_pte);
2513 : }
2514 : set_pte_at(mm, addr, ptep, swp_pte);
2515 :
2516 : /*
2517 : * This is like regular unmap: we remove the rmap and
2518 : * drop page refcount. Page won't be freed, as we took
2519 : * a reference just above.
2520 : */
2521 : page_remove_rmap(page, false);
2522 : put_page(page);
2523 :
2524 : if (pte_present(pte))
2525 : unmapped++;
2526 : }
2527 :
2528 : next:
2529 : migrate->dst[migrate->npages] = 0;
2530 : migrate->src[migrate->npages++] = mpfn;
2531 : }
2532 : arch_leave_lazy_mmu_mode();
2533 : pte_unmap_unlock(ptep - 1, ptl);
2534 :
2535 : /* Only flush the TLB if we actually modified any entries */
2536 : if (unmapped)
2537 : flush_tlb_range(walk->vma, start, end);
2538 :
2539 : return 0;
2540 : }
2541 :
2542 : static const struct mm_walk_ops migrate_vma_walk_ops = {
2543 : .pmd_entry = migrate_vma_collect_pmd,
2544 : .pte_hole = migrate_vma_collect_hole,
2545 : };
2546 :
2547 : /*
2548 : * migrate_vma_collect() - collect pages over a range of virtual addresses
2549 : * @migrate: migrate struct containing all migration information
2550 : *
2551 : * This will walk the CPU page table. For each virtual address backed by a
2552 : * valid page, it updates the src array and takes a reference on the page, in
2553 : * order to pin the page until we lock it and unmap it.
2554 : */
2555 : static void migrate_vma_collect(struct migrate_vma *migrate)
2556 : {
2557 : struct mmu_notifier_range range;
2558 :
2559 : /*
2560 : * Note that the pgmap_owner is passed to the mmu notifier callback so
2561 : * that the registered device driver can skip invalidating device
2562 : * private page mappings that won't be migrated.
2563 : */
2564 : mmu_notifier_range_init_migrate(&range, 0, migrate->vma,
2565 : migrate->vma->vm_mm, migrate->start, migrate->end,
2566 : migrate->pgmap_owner);
2567 : mmu_notifier_invalidate_range_start(&range);
2568 :
2569 : walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2570 : &migrate_vma_walk_ops, migrate);
2571 :
2572 : mmu_notifier_invalidate_range_end(&range);
2573 : migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2574 : }
2575 :
2576 : /*
2577 : * migrate_vma_check_page() - check if page is pinned or not
2578 : * @page: struct page to check
2579 : *
2580 : * Pinned pages cannot be migrated. This is the same test as in
2581 : * migrate_page_move_mapping(), except that here we allow migration of a
2582 : * ZONE_DEVICE page.
2583 : */
2584 : static bool migrate_vma_check_page(struct page *page)
2585 : {
2586 : /*
2587 : * One extra ref because caller holds an extra reference, either from
2588 : * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2589 : * a device page.
2590 : */
2591 : int extra = 1;
2592 :
2593 : /*
2594 : * FIXME support THP (transparent huge page), it is bit more complex to
2595 : * check them than regular pages, because they can be mapped with a pmd
2596 : * or with a pte (split pte mapping).
2597 : */
2598 : if (PageCompound(page))
2599 : return false;
2600 :
2601 : /* Page from ZONE_DEVICE have one extra reference */
2602 : if (is_zone_device_page(page)) {
2603 : /*
2604 : * Private page can never be pin as they have no valid pte and
2605 : * GUP will fail for those. Yet if there is a pending migration
2606 : * a thread might try to wait on the pte migration entry and
2607 : * will bump the page reference count. Sadly there is no way to
2608 : * differentiate a regular pin from migration wait. Hence to
2609 : * avoid 2 racing thread trying to migrate back to CPU to enter
2610 : * infinite loop (one stopping migration because the other is
2611 : * waiting on pte migration entry). We always return true here.
2612 : *
2613 : * FIXME proper solution is to rework migration_entry_wait() so
2614 : * it does not need to take a reference on page.
2615 : */
2616 : return is_device_private_page(page);
2617 : }
2618 :
2619 : /* For file back page */
2620 : if (page_mapping(page))
2621 : extra += 1 + page_has_private(page);
2622 :
2623 : if ((page_count(page) - extra) > page_mapcount(page))
2624 : return false;
2625 :
2626 : return true;
2627 : }
2628 :
2629 : /*
2630 : * migrate_vma_prepare() - lock pages and isolate them from the lru
2631 : * @migrate: migrate struct containing all migration information
2632 : *
2633 : * This locks pages that have been collected by migrate_vma_collect(). Once each
2634 : * page is locked it is isolated from the lru (for non-device pages). Finally,
2635 : * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2636 : * migrated by concurrent kernel threads.
2637 : */
2638 : static void migrate_vma_prepare(struct migrate_vma *migrate)
2639 : {
2640 : const unsigned long npages = migrate->npages;
2641 : const unsigned long start = migrate->start;
2642 : unsigned long addr, i, restore = 0;
2643 : bool allow_drain = true;
2644 :
2645 : lru_add_drain();
2646 :
2647 : for (i = 0; (i < npages) && migrate->cpages; i++) {
2648 : struct page *page = migrate_pfn_to_page(migrate->src[i]);
2649 : bool remap = true;
2650 :
2651 : if (!page)
2652 : continue;
2653 :
2654 : if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2655 : /*
2656 : * Because we are migrating several pages there can be
2657 : * a deadlock between 2 concurrent migration where each
2658 : * are waiting on each other page lock.
2659 : *
2660 : * Make migrate_vma() a best effort thing and backoff
2661 : * for any page we can not lock right away.
2662 : */
2663 : if (!trylock_page(page)) {
2664 : migrate->src[i] = 0;
2665 : migrate->cpages--;
2666 : put_page(page);
2667 : continue;
2668 : }
2669 : remap = false;
2670 : migrate->src[i] |= MIGRATE_PFN_LOCKED;
2671 : }
2672 :
2673 : /* ZONE_DEVICE pages are not on LRU */
2674 : if (!is_zone_device_page(page)) {
2675 : if (!PageLRU(page) && allow_drain) {
2676 : /* Drain CPU's pagevec */
2677 : lru_add_drain_all();
2678 : allow_drain = false;
2679 : }
2680 :
2681 : if (isolate_lru_page(page)) {
2682 : if (remap) {
2683 : migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2684 : migrate->cpages--;
2685 : restore++;
2686 : } else {
2687 : migrate->src[i] = 0;
2688 : unlock_page(page);
2689 : migrate->cpages--;
2690 : put_page(page);
2691 : }
2692 : continue;
2693 : }
2694 :
2695 : /* Drop the reference we took in collect */
2696 : put_page(page);
2697 : }
2698 :
2699 : if (!migrate_vma_check_page(page)) {
2700 : if (remap) {
2701 : migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2702 : migrate->cpages--;
2703 : restore++;
2704 :
2705 : if (!is_zone_device_page(page)) {
2706 : get_page(page);
2707 : putback_lru_page(page);
2708 : }
2709 : } else {
2710 : migrate->src[i] = 0;
2711 : unlock_page(page);
2712 : migrate->cpages--;
2713 :
2714 : if (!is_zone_device_page(page))
2715 : putback_lru_page(page);
2716 : else
2717 : put_page(page);
2718 : }
2719 : }
2720 : }
2721 :
2722 : for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2723 : struct page *page = migrate_pfn_to_page(migrate->src[i]);
2724 :
2725 : if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2726 : continue;
2727 :
2728 : remove_migration_pte(page, migrate->vma, addr, page);
2729 :
2730 : migrate->src[i] = 0;
2731 : unlock_page(page);
2732 : put_page(page);
2733 : restore--;
2734 : }
2735 : }
2736 :
2737 : /*
2738 : * migrate_vma_unmap() - replace page mapping with special migration pte entry
2739 : * @migrate: migrate struct containing all migration information
2740 : *
2741 : * Replace page mapping (CPU page table pte) with a special migration pte entry
2742 : * and check again if it has been pinned. Pinned pages are restored because we
2743 : * cannot migrate them.
2744 : *
2745 : * This is the last step before we call the device driver callback to allocate
2746 : * destination memory and copy contents of original page over to new page.
2747 : */
2748 : static void migrate_vma_unmap(struct migrate_vma *migrate)
2749 : {
2750 : int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK;
2751 : const unsigned long npages = migrate->npages;
2752 : const unsigned long start = migrate->start;
2753 : unsigned long addr, i, restore = 0;
2754 :
2755 : for (i = 0; i < npages; i++) {
2756 : struct page *page = migrate_pfn_to_page(migrate->src[i]);
2757 :
2758 : if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2759 : continue;
2760 :
2761 : if (page_mapped(page)) {
2762 : try_to_unmap(page, flags);
2763 : if (page_mapped(page))
2764 : goto restore;
2765 : }
2766 :
2767 : if (migrate_vma_check_page(page))
2768 : continue;
2769 :
2770 : restore:
2771 : migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2772 : migrate->cpages--;
2773 : restore++;
2774 : }
2775 :
2776 : for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2777 : struct page *page = migrate_pfn_to_page(migrate->src[i]);
2778 :
2779 : if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2780 : continue;
2781 :
2782 : remove_migration_ptes(page, page, false);
2783 :
2784 : migrate->src[i] = 0;
2785 : unlock_page(page);
2786 : restore--;
2787 :
2788 : if (is_zone_device_page(page))
2789 : put_page(page);
2790 : else
2791 : putback_lru_page(page);
2792 : }
2793 : }
2794 :
2795 : /**
2796 : * migrate_vma_setup() - prepare to migrate a range of memory
2797 : * @args: contains the vma, start, and pfns arrays for the migration
2798 : *
2799 : * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2800 : * without an error.
2801 : *
2802 : * Prepare to migrate a range of memory virtual address range by collecting all
2803 : * the pages backing each virtual address in the range, saving them inside the
2804 : * src array. Then lock those pages and unmap them. Once the pages are locked
2805 : * and unmapped, check whether each page is pinned or not. Pages that aren't
2806 : * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2807 : * corresponding src array entry. Then restores any pages that are pinned, by
2808 : * remapping and unlocking those pages.
2809 : *
2810 : * The caller should then allocate destination memory and copy source memory to
2811 : * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2812 : * flag set). Once these are allocated and copied, the caller must update each
2813 : * corresponding entry in the dst array with the pfn value of the destination
2814 : * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2815 : * (destination pages must have their struct pages locked, via lock_page()).
2816 : *
2817 : * Note that the caller does not have to migrate all the pages that are marked
2818 : * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2819 : * device memory to system memory. If the caller cannot migrate a device page
2820 : * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2821 : * consequences for the userspace process, so it must be avoided if at all
2822 : * possible.
2823 : *
2824 : * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2825 : * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2826 : * allowing the caller to allocate device memory for those unback virtual
2827 : * address. For this the caller simply has to allocate device memory and
2828 : * properly set the destination entry like for regular migration. Note that
2829 : * this can still fails and thus inside the device driver must check if the
2830 : * migration was successful for those entries after calling migrate_vma_pages()
2831 : * just like for regular migration.
2832 : *
2833 : * After that, the callers must call migrate_vma_pages() to go over each entry
2834 : * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2835 : * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2836 : * then migrate_vma_pages() to migrate struct page information from the source
2837 : * struct page to the destination struct page. If it fails to migrate the
2838 : * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2839 : * src array.
2840 : *
2841 : * At this point all successfully migrated pages have an entry in the src
2842 : * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2843 : * array entry with MIGRATE_PFN_VALID flag set.
2844 : *
2845 : * Once migrate_vma_pages() returns the caller may inspect which pages were
2846 : * successfully migrated, and which were not. Successfully migrated pages will
2847 : * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2848 : *
2849 : * It is safe to update device page table after migrate_vma_pages() because
2850 : * both destination and source page are still locked, and the mmap_lock is held
2851 : * in read mode (hence no one can unmap the range being migrated).
2852 : *
2853 : * Once the caller is done cleaning up things and updating its page table (if it
2854 : * chose to do so, this is not an obligation) it finally calls
2855 : * migrate_vma_finalize() to update the CPU page table to point to new pages
2856 : * for successfully migrated pages or otherwise restore the CPU page table to
2857 : * point to the original source pages.
2858 : */
2859 : int migrate_vma_setup(struct migrate_vma *args)
2860 : {
2861 : long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2862 :
2863 : args->start &= PAGE_MASK;
2864 : args->end &= PAGE_MASK;
2865 : if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2866 : (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2867 : return -EINVAL;
2868 : if (nr_pages <= 0)
2869 : return -EINVAL;
2870 : if (args->start < args->vma->vm_start ||
2871 : args->start >= args->vma->vm_end)
2872 : return -EINVAL;
2873 : if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2874 : return -EINVAL;
2875 : if (!args->src || !args->dst)
2876 : return -EINVAL;
2877 :
2878 : memset(args->src, 0, sizeof(*args->src) * nr_pages);
2879 : args->cpages = 0;
2880 : args->npages = 0;
2881 :
2882 : migrate_vma_collect(args);
2883 :
2884 : if (args->cpages)
2885 : migrate_vma_prepare(args);
2886 : if (args->cpages)
2887 : migrate_vma_unmap(args);
2888 :
2889 : /*
2890 : * At this point pages are locked and unmapped, and thus they have
2891 : * stable content and can safely be copied to destination memory that
2892 : * is allocated by the drivers.
2893 : */
2894 : return 0;
2895 :
2896 : }
2897 : EXPORT_SYMBOL(migrate_vma_setup);
2898 :
2899 : /*
2900 : * This code closely matches the code in:
2901 : * __handle_mm_fault()
2902 : * handle_pte_fault()
2903 : * do_anonymous_page()
2904 : * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2905 : * private page.
2906 : */
2907 : static void migrate_vma_insert_page(struct migrate_vma *migrate,
2908 : unsigned long addr,
2909 : struct page *page,
2910 : unsigned long *src)
2911 : {
2912 : struct vm_area_struct *vma = migrate->vma;
2913 : struct mm_struct *mm = vma->vm_mm;
2914 : bool flush = false;
2915 : spinlock_t *ptl;
2916 : pte_t entry;
2917 : pgd_t *pgdp;
2918 : p4d_t *p4dp;
2919 : pud_t *pudp;
2920 : pmd_t *pmdp;
2921 : pte_t *ptep;
2922 :
2923 : /* Only allow populating anonymous memory */
2924 : if (!vma_is_anonymous(vma))
2925 : goto abort;
2926 :
2927 : pgdp = pgd_offset(mm, addr);
2928 : p4dp = p4d_alloc(mm, pgdp, addr);
2929 : if (!p4dp)
2930 : goto abort;
2931 : pudp = pud_alloc(mm, p4dp, addr);
2932 : if (!pudp)
2933 : goto abort;
2934 : pmdp = pmd_alloc(mm, pudp, addr);
2935 : if (!pmdp)
2936 : goto abort;
2937 :
2938 : if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2939 : goto abort;
2940 :
2941 : /*
2942 : * Use pte_alloc() instead of pte_alloc_map(). We can't run
2943 : * pte_offset_map() on pmds where a huge pmd might be created
2944 : * from a different thread.
2945 : *
2946 : * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2947 : * parallel threads are excluded by other means.
2948 : *
2949 : * Here we only have mmap_read_lock(mm).
2950 : */
2951 : if (pte_alloc(mm, pmdp))
2952 : goto abort;
2953 :
2954 : /* See the comment in pte_alloc_one_map() */
2955 : if (unlikely(pmd_trans_unstable(pmdp)))
2956 : goto abort;
2957 :
2958 : if (unlikely(anon_vma_prepare(vma)))
2959 : goto abort;
2960 : if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
2961 : goto abort;
2962 :
2963 : /*
2964 : * The memory barrier inside __SetPageUptodate makes sure that
2965 : * preceding stores to the page contents become visible before
2966 : * the set_pte_at() write.
2967 : */
2968 : __SetPageUptodate(page);
2969 :
2970 : if (is_zone_device_page(page)) {
2971 : if (is_device_private_page(page)) {
2972 : swp_entry_t swp_entry;
2973 :
2974 : swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2975 : entry = swp_entry_to_pte(swp_entry);
2976 : }
2977 : } else {
2978 : entry = mk_pte(page, vma->vm_page_prot);
2979 : if (vma->vm_flags & VM_WRITE)
2980 : entry = pte_mkwrite(pte_mkdirty(entry));
2981 : }
2982 :
2983 : ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2984 :
2985 : if (check_stable_address_space(mm))
2986 : goto unlock_abort;
2987 :
2988 : if (pte_present(*ptep)) {
2989 : unsigned long pfn = pte_pfn(*ptep);
2990 :
2991 : if (!is_zero_pfn(pfn))
2992 : goto unlock_abort;
2993 : flush = true;
2994 : } else if (!pte_none(*ptep))
2995 : goto unlock_abort;
2996 :
2997 : /*
2998 : * Check for userfaultfd but do not deliver the fault. Instead,
2999 : * just back off.
3000 : */
3001 : if (userfaultfd_missing(vma))
3002 : goto unlock_abort;
3003 :
3004 : inc_mm_counter(mm, MM_ANONPAGES);
3005 : page_add_new_anon_rmap(page, vma, addr, false);
3006 : if (!is_zone_device_page(page))
3007 : lru_cache_add_inactive_or_unevictable(page, vma);
3008 : get_page(page);
3009 :
3010 : if (flush) {
3011 : flush_cache_page(vma, addr, pte_pfn(*ptep));
3012 : ptep_clear_flush_notify(vma, addr, ptep);
3013 : set_pte_at_notify(mm, addr, ptep, entry);
3014 : update_mmu_cache(vma, addr, ptep);
3015 : } else {
3016 : /* No need to invalidate - it was non-present before */
3017 : set_pte_at(mm, addr, ptep, entry);
3018 : update_mmu_cache(vma, addr, ptep);
3019 : }
3020 :
3021 : pte_unmap_unlock(ptep, ptl);
3022 : *src = MIGRATE_PFN_MIGRATE;
3023 : return;
3024 :
3025 : unlock_abort:
3026 : pte_unmap_unlock(ptep, ptl);
3027 : abort:
3028 : *src &= ~MIGRATE_PFN_MIGRATE;
3029 : }
3030 :
3031 : /**
3032 : * migrate_vma_pages() - migrate meta-data from src page to dst page
3033 : * @migrate: migrate struct containing all migration information
3034 : *
3035 : * This migrates struct page meta-data from source struct page to destination
3036 : * struct page. This effectively finishes the migration from source page to the
3037 : * destination page.
3038 : */
3039 : void migrate_vma_pages(struct migrate_vma *migrate)
3040 : {
3041 : const unsigned long npages = migrate->npages;
3042 : const unsigned long start = migrate->start;
3043 : struct mmu_notifier_range range;
3044 : unsigned long addr, i;
3045 : bool notified = false;
3046 :
3047 : for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
3048 : struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
3049 : struct page *page = migrate_pfn_to_page(migrate->src[i]);
3050 : struct address_space *mapping;
3051 : int r;
3052 :
3053 : if (!newpage) {
3054 : migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3055 : continue;
3056 : }
3057 :
3058 : if (!page) {
3059 : if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
3060 : continue;
3061 : if (!notified) {
3062 : notified = true;
3063 :
3064 : mmu_notifier_range_init_migrate(&range, 0,
3065 : migrate->vma, migrate->vma->vm_mm,
3066 : addr, migrate->end,
3067 : migrate->pgmap_owner);
3068 : mmu_notifier_invalidate_range_start(&range);
3069 : }
3070 : migrate_vma_insert_page(migrate, addr, newpage,
3071 : &migrate->src[i]);
3072 : continue;
3073 : }
3074 :
3075 : mapping = page_mapping(page);
3076 :
3077 : if (is_zone_device_page(newpage)) {
3078 : if (is_device_private_page(newpage)) {
3079 : /*
3080 : * For now only support private anonymous when
3081 : * migrating to un-addressable device memory.
3082 : */
3083 : if (mapping) {
3084 : migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3085 : continue;
3086 : }
3087 : } else {
3088 : /*
3089 : * Other types of ZONE_DEVICE page are not
3090 : * supported.
3091 : */
3092 : migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3093 : continue;
3094 : }
3095 : }
3096 :
3097 : r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
3098 : if (r != MIGRATEPAGE_SUCCESS)
3099 : migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3100 : }
3101 :
3102 : /*
3103 : * No need to double call mmu_notifier->invalidate_range() callback as
3104 : * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
3105 : * did already call it.
3106 : */
3107 : if (notified)
3108 : mmu_notifier_invalidate_range_only_end(&range);
3109 : }
3110 : EXPORT_SYMBOL(migrate_vma_pages);
3111 :
3112 : /**
3113 : * migrate_vma_finalize() - restore CPU page table entry
3114 : * @migrate: migrate struct containing all migration information
3115 : *
3116 : * This replaces the special migration pte entry with either a mapping to the
3117 : * new page if migration was successful for that page, or to the original page
3118 : * otherwise.
3119 : *
3120 : * This also unlocks the pages and puts them back on the lru, or drops the extra
3121 : * refcount, for device pages.
3122 : */
3123 : void migrate_vma_finalize(struct migrate_vma *migrate)
3124 : {
3125 : const unsigned long npages = migrate->npages;
3126 : unsigned long i;
3127 :
3128 : for (i = 0; i < npages; i++) {
3129 : struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
3130 : struct page *page = migrate_pfn_to_page(migrate->src[i]);
3131 :
3132 : if (!page) {
3133 : if (newpage) {
3134 : unlock_page(newpage);
3135 : put_page(newpage);
3136 : }
3137 : continue;
3138 : }
3139 :
3140 : if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
3141 : if (newpage) {
3142 : unlock_page(newpage);
3143 : put_page(newpage);
3144 : }
3145 : newpage = page;
3146 : }
3147 :
3148 : remove_migration_ptes(page, newpage, false);
3149 : unlock_page(page);
3150 :
3151 : if (is_zone_device_page(page))
3152 : put_page(page);
3153 : else
3154 : putback_lru_page(page);
3155 :
3156 : if (newpage != page) {
3157 : unlock_page(newpage);
3158 : if (is_zone_device_page(newpage))
3159 : put_page(newpage);
3160 : else
3161 : putback_lru_page(newpage);
3162 : }
3163 : }
3164 : }
3165 : EXPORT_SYMBOL(migrate_vma_finalize);
3166 : #endif /* CONFIG_DEVICE_PRIVATE */
|
