blob: 7452a00bbb50c134b529c1d024dfc53fcfca093b [file] [log] [blame]
Christoph Lameterb20a3502006-03-22 00:09:12 -08001/*
2 * Memory Migration functionality - linux/mm/migration.c
3 *
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5 *
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
8 *
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
Christoph Lametercde53532008-07-04 09:59:22 -070012 * Christoph Lameter
Christoph Lameterb20a3502006-03-22 00:09:12 -080013 */
14
15#include <linux/migrate.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040016#include <linux/export.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080017#include <linux/swap.h>
Christoph Lameter06972122006-06-23 02:03:35 -070018#include <linux/swapops.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080019#include <linux/pagemap.h>
Christoph Lametere23ca002006-04-10 22:52:57 -070020#include <linux/buffer_head.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080021#include <linux/mm_inline.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070022#include <linux/nsproxy.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080023#include <linux/pagevec.h>
Hugh Dickinse9995ef2009-12-14 17:59:31 -080024#include <linux/ksm.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080025#include <linux/rmap.h>
26#include <linux/topology.h>
27#include <linux/cpu.h>
28#include <linux/cpuset.h>
Christoph Lameter04e62a22006-06-23 02:03:38 -070029#include <linux/writeback.h>
Christoph Lameter742755a2006-06-23 02:03:55 -070030#include <linux/mempolicy.h>
31#include <linux/vmalloc.h>
David Quigley86c3a762006-06-23 02:04:02 -070032#include <linux/security.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080033#include <linux/memcontrol.h>
Adrian Bunk4f5ca262008-07-23 21:27:02 -070034#include <linux/syscalls.h>
Naoya Horiguchi290408d2010-09-08 10:19:35 +090035#include <linux/hugetlb.h>
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -070036#include <linux/hugetlb_cgroup.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090037#include <linux/gfp.h>
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080038#include <linux/balloon_compaction.h>
Mel Gormanf714f4f2013-12-18 17:08:33 -080039#include <linux/mmu_notifier.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070040#include <linux/page_idle.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080041
Michal Nazarewicz0d1836c2010-12-21 17:24:26 -080042#include <asm/tlbflush.h>
43
Mel Gorman7b2a2d42012-10-19 14:07:31 +010044#define CREATE_TRACE_POINTS
45#include <trace/events/migrate.h>
46
Christoph Lameterb20a3502006-03-22 00:09:12 -080047#include "internal.h"
48
Christoph Lameterb20a3502006-03-22 00:09:12 -080049/*
Christoph Lameter742755a2006-06-23 02:03:55 -070050 * migrate_prep() needs to be called before we start compiling a list of pages
Mel Gorman748446b2010-05-24 14:32:27 -070051 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
52 * undesirable, use migrate_prep_local()
Christoph Lameterb20a3502006-03-22 00:09:12 -080053 */
54int migrate_prep(void)
55{
Christoph Lameterb20a3502006-03-22 00:09:12 -080056 /*
57 * Clear the LRU lists so pages can be isolated.
58 * Note that pages may be moved off the LRU after we have
59 * drained them. Those pages will fail to migrate like other
60 * pages that may be busy.
61 */
62 lru_add_drain_all();
63
64 return 0;
65}
66
Mel Gorman748446b2010-05-24 14:32:27 -070067/* Do the necessary work of migrate_prep but not if it involves other CPUs */
68int migrate_prep_local(void)
69{
70 lru_add_drain();
71
72 return 0;
73}
74
Christoph Lameterb20a3502006-03-22 00:09:12 -080075/*
Rafael Aquini5733c7d2012-12-11 16:02:47 -080076 * Put previously isolated pages back onto the appropriate lists
77 * from where they were once taken off for compaction/migration.
78 *
Joonsoo Kim59c82b72014-01-21 15:51:17 -080079 * This function shall be used whenever the isolated pageset has been
80 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
81 * and isolate_huge_page().
Rafael Aquini5733c7d2012-12-11 16:02:47 -080082 */
83void putback_movable_pages(struct list_head *l)
84{
85 struct page *page;
86 struct page *page2;
87
88 list_for_each_entry_safe(page, page2, l, lru) {
Naoya Horiguchi31caf662013-09-11 14:21:59 -070089 if (unlikely(PageHuge(page))) {
90 putback_active_hugepage(page);
91 continue;
92 }
Rafael Aquini5733c7d2012-12-11 16:02:47 -080093 list_del(&page->lru);
94 dec_zone_page_state(page, NR_ISOLATED_ANON +
95 page_is_file_cache(page));
Rafael Aquini117aad12013-09-30 13:45:16 -070096 if (unlikely(isolated_balloon_page(page)))
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080097 balloon_page_putback(page);
98 else
99 putback_lru_page(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800100 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800101}
102
Christoph Lameter06972122006-06-23 02:03:35 -0700103/*
104 * Restore a potential migration pte to a working pte entry
105 */
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800106static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
107 unsigned long addr, void *old)
Christoph Lameter06972122006-06-23 02:03:35 -0700108{
109 struct mm_struct *mm = vma->vm_mm;
110 swp_entry_t entry;
Christoph Lameter06972122006-06-23 02:03:35 -0700111 pmd_t *pmd;
112 pte_t *ptep, pte;
113 spinlock_t *ptl;
114
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900115 if (unlikely(PageHuge(new))) {
116 ptep = huge_pte_offset(mm, addr);
117 if (!ptep)
118 goto out;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800119 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900120 } else {
Bob Liu62190492012-12-11 16:00:37 -0800121 pmd = mm_find_pmd(mm, addr);
122 if (!pmd)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900123 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700124
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900125 ptep = pte_offset_map(pmd, addr);
Christoph Lameter06972122006-06-23 02:03:35 -0700126
Hugh Dickins486cf462011-10-19 12:50:35 -0700127 /*
128 * Peek to check is_swap_pte() before taking ptlock? No, we
129 * can race mremap's move_ptes(), which skips anon_vma lock.
130 */
Christoph Lameter06972122006-06-23 02:03:35 -0700131
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900132 ptl = pte_lockptr(mm, pmd);
133 }
134
Christoph Lameter06972122006-06-23 02:03:35 -0700135 spin_lock(ptl);
136 pte = *ptep;
137 if (!is_swap_pte(pte))
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800138 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700139
140 entry = pte_to_swp_entry(pte);
141
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800142 if (!is_migration_entry(entry) ||
143 migration_entry_to_page(entry) != old)
144 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700145
Christoph Lameter06972122006-06-23 02:03:35 -0700146 get_page(new);
147 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700148 if (pte_swp_soft_dirty(*ptep))
149 pte = pte_mksoft_dirty(pte);
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100150
151 /* Recheck VMA as permissions can change since migration started */
Christoph Lameter06972122006-06-23 02:03:35 -0700152 if (is_write_migration_entry(entry))
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100153 pte = maybe_mkwrite(pte, vma);
154
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200155#ifdef CONFIG_HUGETLB_PAGE
Tony Lube7517d2013-02-04 14:28:46 -0800156 if (PageHuge(new)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900157 pte = pte_mkhuge(pte);
Tony Lube7517d2013-02-04 14:28:46 -0800158 pte = arch_make_huge_pte(pte, vma, new, 0);
159 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200160#endif
Leonid Yegoshinc2cc4992013-05-24 15:55:18 -0700161 flush_dcache_page(new);
Christoph Lameter06972122006-06-23 02:03:35 -0700162 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700163
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900164 if (PageHuge(new)) {
165 if (PageAnon(new))
166 hugepage_add_anon_rmap(new, vma, addr);
167 else
168 page_dup_rmap(new);
169 } else if (PageAnon(new))
Christoph Lameter04e62a22006-06-23 02:03:38 -0700170 page_add_anon_rmap(new, vma, addr);
171 else
172 page_add_file_rmap(new);
173
174 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000175 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800176unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700177 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800178out:
179 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700180}
181
182/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700183 * Get rid of all migration entries and replace them by
184 * references to the indicated page.
185 */
186static void remove_migration_ptes(struct page *old, struct page *new)
187{
Joonsoo Kim051ac832014-01-21 15:49:48 -0800188 struct rmap_walk_control rwc = {
189 .rmap_one = remove_migration_pte,
190 .arg = old,
191 };
192
193 rmap_walk(new, &rwc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700194}
195
196/*
Christoph Lameter06972122006-06-23 02:03:35 -0700197 * Something used the pte of a page under migration. We need to
198 * get to the page and wait until migration is finished.
199 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700200 */
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800201void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700202 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700203{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700204 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700205 swp_entry_t entry;
206 struct page *page;
207
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700208 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700209 pte = *ptep;
210 if (!is_swap_pte(pte))
211 goto out;
212
213 entry = pte_to_swp_entry(pte);
214 if (!is_migration_entry(entry))
215 goto out;
216
217 page = migration_entry_to_page(entry);
218
Nick Piggine2867812008-07-25 19:45:30 -0700219 /*
220 * Once radix-tree replacement of page migration started, page_count
221 * *must* be zero. And, we don't want to call wait_on_page_locked()
222 * against a page without get_page().
223 * So, we use get_page_unless_zero(), here. Even failed, page fault
224 * will occur again.
225 */
226 if (!get_page_unless_zero(page))
227 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700228 pte_unmap_unlock(ptep, ptl);
229 wait_on_page_locked(page);
230 put_page(page);
231 return;
232out:
233 pte_unmap_unlock(ptep, ptl);
234}
235
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700236void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
237 unsigned long address)
238{
239 spinlock_t *ptl = pte_lockptr(mm, pmd);
240 pte_t *ptep = pte_offset_map(pmd, address);
241 __migration_entry_wait(mm, ptep, ptl);
242}
243
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800244void migration_entry_wait_huge(struct vm_area_struct *vma,
245 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700246{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800247 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700248 __migration_entry_wait(mm, pte, ptl);
249}
250
Mel Gormanb969c4a2012-01-12 17:19:34 -0800251#ifdef CONFIG_BLOCK
252/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800253static bool buffer_migrate_lock_buffers(struct buffer_head *head,
254 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800255{
256 struct buffer_head *bh = head;
257
258 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800259 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800260 do {
261 get_bh(bh);
262 lock_buffer(bh);
263 bh = bh->b_this_page;
264
265 } while (bh != head);
266
267 return true;
268 }
269
270 /* async case, we cannot block on lock_buffer so use trylock_buffer */
271 do {
272 get_bh(bh);
273 if (!trylock_buffer(bh)) {
274 /*
275 * We failed to lock the buffer and cannot stall in
276 * async migration. Release the taken locks
277 */
278 struct buffer_head *failed_bh = bh;
279 put_bh(failed_bh);
280 bh = head;
281 while (bh != failed_bh) {
282 unlock_buffer(bh);
283 put_bh(bh);
284 bh = bh->b_this_page;
285 }
286 return false;
287 }
288
289 bh = bh->b_this_page;
290 } while (bh != head);
291 return true;
292}
293#else
294static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800295 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800296{
297 return true;
298}
299#endif /* CONFIG_BLOCK */
300
Christoph Lameterb20a3502006-03-22 00:09:12 -0800301/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700302 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700303 *
304 * The number of remaining references must be:
305 * 1 for anonymous pages without a mapping
306 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100307 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800308 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800309int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800310 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500311 struct buffer_head *head, enum migrate_mode mode,
312 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800313{
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500314 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800315 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800316
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700317 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700318 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500319 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700320 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800321 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700322 }
323
Nick Piggin19fd6232008-07-25 19:45:32 -0700324 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800325
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800326 pslot = radix_tree_lookup_slot(&mapping->page_tree,
327 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800328
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500329 expected_count += 1 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700330 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800331 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700332 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700333 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800334 }
335
Nick Piggine2867812008-07-25 19:45:30 -0700336 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700337 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700338 return -EAGAIN;
339 }
340
Christoph Lameterb20a3502006-03-22 00:09:12 -0800341 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800342 * In the async migration case of moving a page with buffers, lock the
343 * buffers using trylock before the mapping is moved. If the mapping
344 * was moved, we later failed to lock the buffers and could not move
345 * the mapping back due to an elevated page count, we would have to
346 * block waiting on other references to be dropped.
347 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800348 if (mode == MIGRATE_ASYNC && head &&
349 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800350 page_unfreeze_refs(page, expected_count);
351 spin_unlock_irq(&mapping->tree_lock);
352 return -EAGAIN;
353 }
354
355 /*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800356 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800357 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800358 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800359 if (PageSwapCache(page)) {
360 SetPageSwapCache(newpage);
361 set_page_private(newpage, page_private(page));
362 }
363
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800364 radix_tree_replace_slot(pslot, newpage);
365
366 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800367 * Drop cache reference from old page by unfreezing
368 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800369 * We know this isn't the last reference.
370 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800371 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800372
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700373 /*
374 * If moved to a different zone then also account
375 * the page for that zone. Other VM counters will be
376 * taken care of when we establish references to the
377 * new page and drop references to the old page.
378 *
379 * Note that anonymous pages are accounted for
380 * via NR_FILE_PAGES and NR_ANON_PAGES if they
381 * are mapped to swap space.
382 */
383 __dec_zone_page_state(page, NR_FILE_PAGES);
384 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700385 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700386 __dec_zone_page_state(page, NR_SHMEM);
387 __inc_zone_page_state(newpage, NR_SHMEM);
388 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700389 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800390
Rafael Aquini78bd5202012-12-11 16:02:31 -0800391 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800392}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800393
394/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900395 * The expected number of remaining references is the same as that
396 * of migrate_page_move_mapping().
397 */
398int migrate_huge_page_move_mapping(struct address_space *mapping,
399 struct page *newpage, struct page *page)
400{
401 int expected_count;
402 void **pslot;
403
404 if (!mapping) {
405 if (page_count(page) != 1)
406 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800407 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900408 }
409
410 spin_lock_irq(&mapping->tree_lock);
411
412 pslot = radix_tree_lookup_slot(&mapping->page_tree,
413 page_index(page));
414
415 expected_count = 2 + page_has_private(page);
416 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800417 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900418 spin_unlock_irq(&mapping->tree_lock);
419 return -EAGAIN;
420 }
421
422 if (!page_freeze_refs(page, expected_count)) {
423 spin_unlock_irq(&mapping->tree_lock);
424 return -EAGAIN;
425 }
426
427 get_page(newpage);
428
429 radix_tree_replace_slot(pslot, newpage);
430
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800431 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900432
433 spin_unlock_irq(&mapping->tree_lock);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800434 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900435}
436
437/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800438 * Gigantic pages are so large that we do not guarantee that page++ pointer
439 * arithmetic will work across the entire page. We need something more
440 * specialized.
441 */
442static void __copy_gigantic_page(struct page *dst, struct page *src,
443 int nr_pages)
444{
445 int i;
446 struct page *dst_base = dst;
447 struct page *src_base = src;
448
449 for (i = 0; i < nr_pages; ) {
450 cond_resched();
451 copy_highpage(dst, src);
452
453 i++;
454 dst = mem_map_next(dst, dst_base, i);
455 src = mem_map_next(src, src_base, i);
456 }
457}
458
459static void copy_huge_page(struct page *dst, struct page *src)
460{
461 int i;
462 int nr_pages;
463
464 if (PageHuge(src)) {
465 /* hugetlbfs page */
466 struct hstate *h = page_hstate(src);
467 nr_pages = pages_per_huge_page(h);
468
469 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
470 __copy_gigantic_page(dst, src, nr_pages);
471 return;
472 }
473 } else {
474 /* thp page */
475 BUG_ON(!PageTransHuge(src));
476 nr_pages = hpage_nr_pages(src);
477 }
478
479 for (i = 0; i < nr_pages; i++) {
480 cond_resched();
481 copy_highpage(dst + i, src + i);
482 }
483}
484
485/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800486 * Copy the page to its new location
487 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900488void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800489{
Rik van Riel7851a452013-10-07 11:29:23 +0100490 int cpupid;
491
Mel Gormanb32967f2012-11-19 12:35:47 +0000492 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900493 copy_huge_page(newpage, page);
494 else
495 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800496
497 if (PageError(page))
498 SetPageError(newpage);
499 if (PageReferenced(page))
500 SetPageReferenced(newpage);
501 if (PageUptodate(page))
502 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700503 if (TestClearPageActive(page)) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800504 VM_BUG_ON_PAGE(PageUnevictable(page), page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800505 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800506 } else if (TestClearPageUnevictable(page))
507 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800508 if (PageChecked(page))
509 SetPageChecked(newpage);
510 if (PageMappedToDisk(page))
511 SetPageMappedToDisk(newpage);
512
513 if (PageDirty(page)) {
514 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700515 /*
516 * Want to mark the page and the radix tree as dirty, and
517 * redo the accounting that clear_page_dirty_for_io undid,
518 * but we can't use set_page_dirty because that function
519 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300520 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700521 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700522 if (PageSwapBacked(page))
523 SetPageDirty(newpage);
524 else
525 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800526 }
527
Vladimir Davydov33c3fc72015-09-09 15:35:45 -0700528 if (page_is_young(page))
529 set_page_young(newpage);
530 if (page_is_idle(page))
531 set_page_idle(newpage);
532
Rik van Riel7851a452013-10-07 11:29:23 +0100533 /*
534 * Copy NUMA information to the new page, to prevent over-eager
535 * future migrations of this same page.
536 */
537 cpupid = page_cpupid_xchg_last(page, -1);
538 page_cpupid_xchg_last(newpage, cpupid);
539
Nick Pigginb291f002008-10-18 20:26:44 -0700540 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800541 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800542 /*
543 * Please do not reorder this without considering how mm/ksm.c's
544 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
545 */
Naoya Horiguchib3b3a992015-04-15 16:13:15 -0700546 if (PageSwapCache(page))
547 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800548 ClearPagePrivate(page);
549 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800550
551 /*
552 * If any waiters have accumulated on the new page then
553 * wake them up.
554 */
555 if (PageWriteback(newpage))
556 end_page_writeback(newpage);
557}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800558
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700559/************************************************************
560 * Migration functions
561 ***********************************************************/
562
Christoph Lameterb20a3502006-03-22 00:09:12 -0800563/*
564 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100565 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800566 *
567 * Pages are locked upon entry and exit.
568 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700569int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800570 struct page *newpage, struct page *page,
571 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800572{
573 int rc;
574
575 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
576
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500577 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800578
Rafael Aquini78bd5202012-12-11 16:02:31 -0800579 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800580 return rc;
581
582 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800583 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800584}
585EXPORT_SYMBOL(migrate_page);
586
David Howells93614012006-09-30 20:45:40 +0200587#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800588/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700589 * Migration function for pages with buffers. This function can only be used
590 * if the underlying filesystem guarantees that no other references to "page"
591 * exist.
592 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700593int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800594 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700595{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700596 struct buffer_head *bh, *head;
597 int rc;
598
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700599 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800600 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700601
602 head = page_buffers(page);
603
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500604 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700605
Rafael Aquini78bd5202012-12-11 16:02:31 -0800606 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700607 return rc;
608
Mel Gormanb969c4a2012-01-12 17:19:34 -0800609 /*
610 * In the async case, migrate_page_move_mapping locked the buffers
611 * with an IRQ-safe spinlock held. In the sync case, the buffers
612 * need to be locked now
613 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800614 if (mode != MIGRATE_ASYNC)
615 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700616
617 ClearPagePrivate(page);
618 set_page_private(newpage, page_private(page));
619 set_page_private(page, 0);
620 put_page(page);
621 get_page(newpage);
622
623 bh = head;
624 do {
625 set_bh_page(bh, newpage, bh_offset(bh));
626 bh = bh->b_this_page;
627
628 } while (bh != head);
629
630 SetPagePrivate(newpage);
631
632 migrate_page_copy(newpage, page);
633
634 bh = head;
635 do {
636 unlock_buffer(bh);
637 put_bh(bh);
638 bh = bh->b_this_page;
639
640 } while (bh != head);
641
Rafael Aquini78bd5202012-12-11 16:02:31 -0800642 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700643}
644EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200645#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700646
Christoph Lameter04e62a22006-06-23 02:03:38 -0700647/*
648 * Writeback a page to clean the dirty state
649 */
650static int writeout(struct address_space *mapping, struct page *page)
651{
652 struct writeback_control wbc = {
653 .sync_mode = WB_SYNC_NONE,
654 .nr_to_write = 1,
655 .range_start = 0,
656 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700657 .for_reclaim = 1
658 };
659 int rc;
660
661 if (!mapping->a_ops->writepage)
662 /* No write method for the address space */
663 return -EINVAL;
664
665 if (!clear_page_dirty_for_io(page))
666 /* Someone else already triggered a write */
667 return -EAGAIN;
668
669 /*
670 * A dirty page may imply that the underlying filesystem has
671 * the page on some queue. So the page must be clean for
672 * migration. Writeout may mean we loose the lock and the
673 * page state is no longer what we checked for earlier.
674 * At this point we know that the migration attempt cannot
675 * be successful.
676 */
677 remove_migration_ptes(page, page);
678
679 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700680
681 if (rc != AOP_WRITEPAGE_ACTIVATE)
682 /* unlocked. Relock */
683 lock_page(page);
684
Hugh Dickinsbda85502008-11-19 15:36:36 -0800685 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700686}
687
688/*
689 * Default handling if a filesystem does not provide a migration function.
690 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700691static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800692 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700693{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800694 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800695 /* Only writeback pages in full synchronous migration */
696 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800697 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700698 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800699 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700700
701 /*
702 * Buffers may be managed in a filesystem specific way.
703 * We must have no buffers or drop them.
704 */
David Howells266cf652009-04-03 16:42:36 +0100705 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700706 !try_to_release_page(page, GFP_KERNEL))
707 return -EAGAIN;
708
Mel Gormana6bc32b2012-01-12 17:19:43 -0800709 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700710}
711
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700712/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700713 * Move a page to a newly allocated page
714 * The page is locked and all ptes have been successfully removed.
715 *
716 * The new page will have replaced the old page if this function
717 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700718 *
719 * Return value:
720 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800721 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700722 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700723static int move_to_new_page(struct page *newpage, struct page *page,
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800724 int page_was_mapped, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700725{
726 struct address_space *mapping;
727 int rc;
728
729 /*
730 * Block others from accessing the page when we get around to
731 * establishing additional references. We are the only one
732 * holding a reference to the new page at this point.
733 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200734 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700735 BUG();
736
737 /* Prepare mapping for the new page.*/
738 newpage->index = page->index;
739 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700740 if (PageSwapBacked(page))
741 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700742
743 mapping = page_mapping(page);
744 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800745 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800746 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700747 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800748 * Most pages have a mapping and most filesystems provide a
749 * migratepage callback. Anonymous pages are part of swap
750 * space which also has its own migratepage callback. This
751 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700752 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800753 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800754 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800755 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800756 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700757
Rafael Aquini78bd5202012-12-11 16:02:31 -0800758 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700759 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700760 } else {
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700761 mem_cgroup_migrate(page, newpage, false);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800762 if (page_was_mapped)
Mel Gorman3fe20112010-05-24 14:32:20 -0700763 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800764 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700765 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700766
767 unlock_page(newpage);
768
769 return rc;
770}
771
Minchan Kim0dabec92011-10-31 17:06:57 -0700772static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800773 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700774{
Minchan Kim0dabec92011-10-31 17:06:57 -0700775 int rc = -EAGAIN;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800776 int page_was_mapped = 0;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700777 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700778
Nick Piggin529ae9a2008-08-02 12:01:03 +0200779 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800780 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700781 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800782
783 /*
784 * It's not safe for direct compaction to call lock_page.
785 * For example, during page readahead pages are added locked
786 * to the LRU. Later, when the IO completes the pages are
787 * marked uptodate and unlocked. However, the queueing
788 * could be merging multiple pages for one bio (e.g.
789 * mpage_readpages). If an allocation happens for the
790 * second or third page, the process can end up locking
791 * the same page twice and deadlocking. Rather than
792 * trying to be clever about what pages can be locked,
793 * avoid the use of lock_page for direct compaction
794 * altogether.
795 */
796 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700797 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800798
Christoph Lametere24f0b82006-06-23 02:03:51 -0700799 lock_page(page);
800 }
801
802 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700803 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700804 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800805 * necessary to wait for PageWriteback. In the async case,
806 * the retry loop is too short and in the sync-light case,
807 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700808 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800809 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700810 rc = -EBUSY;
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700811 goto out_unlock;
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700812 }
813 if (!force)
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700814 goto out_unlock;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700815 wait_on_page_writeback(page);
816 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700817 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700818 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
819 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800820 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700821 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700822 * File Caches may use write_page() or lock_page() in migration, then,
823 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700824 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800825 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800826 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000827 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800828 * getting a hold on an anon_vma from outside one of its mms.
829 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700830 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800831 if (anon_vma) {
832 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700833 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800834 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800835 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700836 /*
837 * We cannot be sure that the anon_vma of an unmapped
838 * swapcache page is safe to use because we don't
839 * know in advance if the VMA that this page belonged
840 * to still exists. If the VMA and others sharing the
841 * data have been freed, then the anon_vma could
842 * already be invalid.
843 *
844 * To avoid this possibility, swapcache pages get
845 * migrated but are not remapped when migration
846 * completes
847 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700848 } else {
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700849 goto out_unlock;
Mel Gorman3fe20112010-05-24 14:32:20 -0700850 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700851 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800852
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700853 if (unlikely(isolated_balloon_page(page))) {
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800854 /*
855 * A ballooned page does not need any special attention from
856 * physical to virtual reverse mapping procedures.
857 * Skip any attempt to unmap PTEs or to remap swap cache,
858 * in order to avoid burning cycles at rmap level, and perform
859 * the page migration right away (proteced by page lock).
860 */
861 rc = balloon_page_migrate(newpage, page, mode);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700862 goto out_unlock;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800863 }
864
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700865 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800866 * Corner case handling:
867 * 1. When a new swap-cache page is read into, it is added to the LRU
868 * and treated as swapcache but it has no rmap yet.
869 * Calling try_to_unmap() against a page->mapping==NULL page will
870 * trigger a BUG. So handle it here.
871 * 2. An orphaned page (see truncate_complete_page) might have
872 * fs-private metadata. The page can be picked up due to memory
873 * offlining. Everywhere else except page reclaim, the page is
874 * invisible to the vm, so the page can not be migrated. So try to
875 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700876 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800877 if (!page->mapping) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800878 VM_BUG_ON_PAGE(PageAnon(page), page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800879 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800880 try_to_free_buffers(page);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700881 goto out_unlock;
Shaohua Li62e1c552008-02-04 22:29:33 -0800882 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700883 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800884 }
885
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700886 /* Establish migration ptes or remove ptes */
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800887 if (page_mapped(page)) {
888 try_to_unmap(page,
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700889 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800890 page_was_mapped = 1;
891 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700892
Shaohua Liabfc3482009-09-21 17:01:19 -0700893skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700894 if (!page_mapped(page))
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800895 rc = move_to_new_page(newpage, page, page_was_mapped, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700896
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800897 if (rc && page_was_mapped)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700898 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700899
900 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700901 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700902 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700903
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700904out_unlock:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700905 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700906out:
907 return rc;
908}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700909
Minchan Kim0dabec92011-10-31 17:06:57 -0700910/*
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -0700911 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
912 * around it.
913 */
914#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
915#define ICE_noinline noinline
916#else
917#define ICE_noinline
918#endif
919
920/*
Minchan Kim0dabec92011-10-31 17:06:57 -0700921 * Obtain the lock on page, remove all ptes and migrate the page
922 * to the newly allocated page in newpage.
923 */
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -0700924static ICE_noinline int unmap_and_move(new_page_t get_new_page,
925 free_page_t put_new_page,
926 unsigned long private, struct page *page,
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -0700927 int force, enum migrate_mode mode,
928 enum migrate_reason reason)
Minchan Kim0dabec92011-10-31 17:06:57 -0700929{
930 int rc = 0;
931 int *result = NULL;
932 struct page *newpage = get_new_page(page, private, &result);
933
934 if (!newpage)
935 return -ENOMEM;
936
937 if (page_count(page) == 1) {
938 /* page was freed from under us. So we are done. */
939 goto out;
940 }
941
942 if (unlikely(PageTransHuge(page)))
943 if (unlikely(split_huge_page(page)))
944 goto out;
945
Hugh Dickins9c620e22013-02-22 16:35:14 -0800946 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800947
Minchan Kim0dabec92011-10-31 17:06:57 -0700948out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700949 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700950 /*
951 * A page that has been migrated has all references
952 * removed and will be freed. A page that has not been
953 * migrated will have kepts its references and be
954 * restored.
955 */
956 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700957 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700958 page_is_file_cache(page));
Naoya Horiguchif4c18e62015-08-06 15:47:08 -0700959 /* Soft-offlined page shouldn't go through lru cache list */
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700960 if (reason == MR_MEMORY_FAILURE) {
Naoya Horiguchif4c18e62015-08-06 15:47:08 -0700961 put_page(page);
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700962 if (!test_set_page_hwpoison(page))
963 num_poisoned_pages_inc();
964 } else
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -0700965 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700966 }
David Rientjes68711a72014-06-04 16:08:25 -0700967
Christoph Lameter95a402c2006-06-23 02:03:53 -0700968 /*
David Rientjes68711a72014-06-04 16:08:25 -0700969 * If migration was not successful and there's a freeing callback, use
970 * it. Otherwise, putback_lru_page() will drop the reference grabbed
971 * during isolation.
Christoph Lameter95a402c2006-06-23 02:03:53 -0700972 */
Hugh Dickins8bdd6382014-07-26 12:58:23 -0700973 if (rc != MIGRATEPAGE_SUCCESS && put_new_page) {
974 ClearPageSwapBacked(newpage);
David Rientjes68711a72014-06-04 16:08:25 -0700975 put_new_page(newpage, private);
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700976 } else if (unlikely(__is_movable_balloon_page(newpage))) {
977 /* drop our reference, page already in the balloon */
978 put_page(newpage);
Hugh Dickins8bdd6382014-07-26 12:58:23 -0700979 } else
David Rientjes68711a72014-06-04 16:08:25 -0700980 putback_lru_page(newpage);
981
Christoph Lameter742755a2006-06-23 02:03:55 -0700982 if (result) {
983 if (rc)
984 *result = rc;
985 else
986 *result = page_to_nid(newpage);
987 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700988 return rc;
989}
990
991/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900992 * Counterpart of unmap_and_move_page() for hugepage migration.
993 *
994 * This function doesn't wait the completion of hugepage I/O
995 * because there is no race between I/O and migration for hugepage.
996 * Note that currently hugepage I/O occurs only in direct I/O
997 * where no lock is held and PG_writeback is irrelevant,
998 * and writeback status of all subpages are counted in the reference
999 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1000 * under direct I/O, the reference of the head page is 512 and a bit more.)
1001 * This means that when we try to migrate hugepage whose subpages are
1002 * doing direct I/O, some references remain after try_to_unmap() and
1003 * hugepage migration fails without data corruption.
1004 *
1005 * There is also no race when direct I/O is issued on the page under migration,
1006 * because then pte is replaced with migration swap entry and direct I/O code
1007 * will wait in the page fault for migration to complete.
1008 */
1009static int unmap_and_move_huge_page(new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001010 free_page_t put_new_page, unsigned long private,
1011 struct page *hpage, int force,
1012 enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001013{
1014 int rc = 0;
1015 int *result = NULL;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001016 int page_was_mapped = 0;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001017 struct page *new_hpage;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001018 struct anon_vma *anon_vma = NULL;
1019
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001020 /*
1021 * Movability of hugepages depends on architectures and hugepage size.
1022 * This check is necessary because some callers of hugepage migration
1023 * like soft offline and memory hotremove don't walk through page
1024 * tables or check whether the hugepage is pmd-based or not before
1025 * kicking migration.
1026 */
Naoya Horiguchi100873d2014-06-04 16:10:56 -07001027 if (!hugepage_migration_supported(page_hstate(hpage))) {
Joonsoo Kim32665f22014-01-21 15:51:15 -08001028 putback_active_hugepage(hpage);
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001029 return -ENOSYS;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001030 }
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001031
Joonsoo Kim32665f22014-01-21 15:51:15 -08001032 new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001033 if (!new_hpage)
1034 return -ENOMEM;
1035
1036 rc = -EAGAIN;
1037
1038 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001039 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001040 goto out;
1041 lock_page(hpage);
1042 }
1043
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001044 if (PageAnon(hpage))
1045 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001046
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001047 if (page_mapped(hpage)) {
1048 try_to_unmap(hpage,
1049 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1050 page_was_mapped = 1;
1051 }
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001052
1053 if (!page_mapped(hpage))
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001054 rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001055
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001056 if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001057 remove_migration_ptes(hpage, hpage);
1058
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001059 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001060 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001061
David Rientjes68711a72014-06-04 16:08:25 -07001062 if (rc == MIGRATEPAGE_SUCCESS)
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001063 hugetlb_cgroup_migrate(hpage, new_hpage);
1064
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001065 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001066out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001067 if (rc != -EAGAIN)
1068 putback_active_hugepage(hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001069
1070 /*
1071 * If migration was not successful and there's a freeing callback, use
1072 * it. Otherwise, put_page() will drop the reference grabbed during
1073 * isolation.
1074 */
1075 if (rc != MIGRATEPAGE_SUCCESS && put_new_page)
1076 put_new_page(new_hpage, private);
1077 else
Naoya Horiguchi3aaa76e2015-09-22 14:59:14 -07001078 putback_active_hugepage(new_hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001079
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001080 if (result) {
1081 if (rc)
1082 *result = rc;
1083 else
1084 *result = page_to_nid(new_hpage);
1085 }
1086 return rc;
1087}
1088
1089/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001090 * migrate_pages - migrate the pages specified in a list, to the free pages
1091 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001092 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001093 * @from: The list of pages to be migrated.
1094 * @get_new_page: The function used to allocate free pages to be used
1095 * as the target of the page migration.
David Rientjes68711a72014-06-04 16:08:25 -07001096 * @put_new_page: The function used to free target pages if migration
1097 * fails, or NULL if no special handling is necessary.
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001098 * @private: Private data to be passed on to get_new_page()
1099 * @mode: The migration mode that specifies the constraints for
1100 * page migration, if any.
1101 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001102 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001103 * The function returns after 10 attempts or if no pages are movable any more
1104 * because the list has become empty or no retryable pages exist any more.
1105 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001106 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001107 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001108 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001109 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001110int migrate_pages(struct list_head *from, new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001111 free_page_t put_new_page, unsigned long private,
1112 enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001113{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001114 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001115 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001116 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001117 int pass = 0;
1118 struct page *page;
1119 struct page *page2;
1120 int swapwrite = current->flags & PF_SWAPWRITE;
1121 int rc;
1122
1123 if (!swapwrite)
1124 current->flags |= PF_SWAPWRITE;
1125
Christoph Lametere24f0b82006-06-23 02:03:51 -07001126 for(pass = 0; pass < 10 && retry; pass++) {
1127 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001128
Christoph Lametere24f0b82006-06-23 02:03:51 -07001129 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001130 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001131
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001132 if (PageHuge(page))
1133 rc = unmap_and_move_huge_page(get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001134 put_new_page, private, page,
1135 pass > 2, mode);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001136 else
David Rientjes68711a72014-06-04 16:08:25 -07001137 rc = unmap_and_move(get_new_page, put_new_page,
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -07001138 private, page, pass > 2, mode,
1139 reason);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001140
Christoph Lametere24f0b82006-06-23 02:03:51 -07001141 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001142 case -ENOMEM:
1143 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001144 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001145 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001146 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001147 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001148 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001149 break;
1150 default:
Naoya Horiguchi354a3362014-01-21 15:51:14 -08001151 /*
1152 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1153 * unlike -EAGAIN case, the failed page is
1154 * removed from migration page list and not
1155 * retried in the next outer loop.
1156 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001157 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001158 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001159 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001160 }
1161 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001162 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001163out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001164 if (nr_succeeded)
1165 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1166 if (nr_failed)
1167 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001168 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1169
Christoph Lameterb20a3502006-03-22 00:09:12 -08001170 if (!swapwrite)
1171 current->flags &= ~PF_SWAPWRITE;
1172
Rafael Aquini78bd5202012-12-11 16:02:31 -08001173 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001174}
1175
Christoph Lameter742755a2006-06-23 02:03:55 -07001176#ifdef CONFIG_NUMA
1177/*
1178 * Move a list of individual pages
1179 */
1180struct page_to_node {
1181 unsigned long addr;
1182 struct page *page;
1183 int node;
1184 int status;
1185};
1186
1187static struct page *new_page_node(struct page *p, unsigned long private,
1188 int **result)
1189{
1190 struct page_to_node *pm = (struct page_to_node *)private;
1191
1192 while (pm->node != MAX_NUMNODES && pm->page != p)
1193 pm++;
1194
1195 if (pm->node == MAX_NUMNODES)
1196 return NULL;
1197
1198 *result = &pm->status;
1199
Naoya Horiguchie632a932013-09-11 14:22:04 -07001200 if (PageHuge(p))
1201 return alloc_huge_page_node(page_hstate(compound_head(p)),
1202 pm->node);
1203 else
Vlastimil Babka96db8002015-09-08 15:03:50 -07001204 return __alloc_pages_node(pm->node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001205 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001206}
1207
1208/*
1209 * Move a set of pages as indicated in the pm array. The addr
1210 * field must be set to the virtual address of the page to be moved
1211 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001212 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001213 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001214static int do_move_page_to_node_array(struct mm_struct *mm,
1215 struct page_to_node *pm,
1216 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001217{
1218 int err;
1219 struct page_to_node *pp;
1220 LIST_HEAD(pagelist);
1221
1222 down_read(&mm->mmap_sem);
1223
1224 /*
1225 * Build a list of pages to migrate
1226 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001227 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1228 struct vm_area_struct *vma;
1229 struct page *page;
1230
Christoph Lameter742755a2006-06-23 02:03:55 -07001231 err = -EFAULT;
1232 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001233 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001234 goto set_status;
1235
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001236 /* FOLL_DUMP to ignore special (like zero) pages */
1237 page = follow_page(vma, pp->addr,
1238 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001239
1240 err = PTR_ERR(page);
1241 if (IS_ERR(page))
1242 goto set_status;
1243
Christoph Lameter742755a2006-06-23 02:03:55 -07001244 err = -ENOENT;
1245 if (!page)
1246 goto set_status;
1247
Christoph Lameter742755a2006-06-23 02:03:55 -07001248 pp->page = page;
1249 err = page_to_nid(page);
1250
1251 if (err == pp->node)
1252 /*
1253 * Node already in the right place
1254 */
1255 goto put_and_set;
1256
1257 err = -EACCES;
1258 if (page_mapcount(page) > 1 &&
1259 !migrate_all)
1260 goto put_and_set;
1261
Naoya Horiguchie632a932013-09-11 14:22:04 -07001262 if (PageHuge(page)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08001263 if (PageHead(page))
1264 isolate_huge_page(page, &pagelist);
Naoya Horiguchie632a932013-09-11 14:22:04 -07001265 goto put_and_set;
1266 }
1267
Nick Piggin62695a82008-10-18 20:26:09 -07001268 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001269 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001270 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001271 inc_zone_page_state(page, NR_ISOLATED_ANON +
1272 page_is_file_cache(page));
1273 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001274put_and_set:
1275 /*
1276 * Either remove the duplicate refcount from
1277 * isolate_lru_page() or drop the page ref if it was
1278 * not isolated.
1279 */
1280 put_page(page);
1281set_status:
1282 pp->status = err;
1283 }
1284
Brice Gogline78bbfa2008-10-18 20:27:15 -07001285 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001286 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -07001287 err = migrate_pages(&pagelist, new_page_node, NULL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001288 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001289 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001290 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001291 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001292
1293 up_read(&mm->mmap_sem);
1294 return err;
1295}
1296
1297/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001298 * Migrate an array of page address onto an array of nodes and fill
1299 * the corresponding array of status.
1300 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001301static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001302 unsigned long nr_pages,
1303 const void __user * __user *pages,
1304 const int __user *nodes,
1305 int __user *status, int flags)
1306{
Brice Goglin3140a222009-01-06 14:38:57 -08001307 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001308 unsigned long chunk_nr_pages;
1309 unsigned long chunk_start;
1310 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001311
Brice Goglin3140a222009-01-06 14:38:57 -08001312 err = -ENOMEM;
1313 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1314 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001315 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001316
1317 migrate_prep();
1318
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001319 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001320 * Store a chunk of page_to_node array in a page,
1321 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001322 */
Brice Goglin3140a222009-01-06 14:38:57 -08001323 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001324
Brice Goglin3140a222009-01-06 14:38:57 -08001325 for (chunk_start = 0;
1326 chunk_start < nr_pages;
1327 chunk_start += chunk_nr_pages) {
1328 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001329
Brice Goglin3140a222009-01-06 14:38:57 -08001330 if (chunk_start + chunk_nr_pages > nr_pages)
1331 chunk_nr_pages = nr_pages - chunk_start;
1332
1333 /* fill the chunk pm with addrs and nodes from user-space */
1334 for (j = 0; j < chunk_nr_pages; j++) {
1335 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001336 int node;
1337
Brice Goglin3140a222009-01-06 14:38:57 -08001338 err = -EFAULT;
1339 if (get_user(p, pages + j + chunk_start))
1340 goto out_pm;
1341 pm[j].addr = (unsigned long) p;
1342
1343 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001344 goto out_pm;
1345
1346 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001347 if (node < 0 || node >= MAX_NUMNODES)
1348 goto out_pm;
1349
Lai Jiangshan389162c2012-12-12 13:51:30 -08001350 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001351 goto out_pm;
1352
1353 err = -EACCES;
1354 if (!node_isset(node, task_nodes))
1355 goto out_pm;
1356
Brice Goglin3140a222009-01-06 14:38:57 -08001357 pm[j].node = node;
1358 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001359
Brice Goglin3140a222009-01-06 14:38:57 -08001360 /* End marker for this chunk */
1361 pm[chunk_nr_pages].node = MAX_NUMNODES;
1362
1363 /* Migrate this chunk */
1364 err = do_move_page_to_node_array(mm, pm,
1365 flags & MPOL_MF_MOVE_ALL);
1366 if (err < 0)
1367 goto out_pm;
1368
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001369 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001370 for (j = 0; j < chunk_nr_pages; j++)
1371 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001372 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001373 goto out_pm;
1374 }
1375 }
1376 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001377
1378out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001379 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001380out:
1381 return err;
1382}
1383
1384/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001385 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001386 */
Brice Goglin80bba122008-12-09 13:14:23 -08001387static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1388 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001389{
Brice Goglin2f007e72008-10-18 20:27:16 -07001390 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001391
Christoph Lameter742755a2006-06-23 02:03:55 -07001392 down_read(&mm->mmap_sem);
1393
Brice Goglin2f007e72008-10-18 20:27:16 -07001394 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001395 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001396 struct vm_area_struct *vma;
1397 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001398 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001399
1400 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001401 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001402 goto set_status;
1403
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001404 /* FOLL_DUMP to ignore special (like zero) pages */
1405 page = follow_page(vma, addr, FOLL_DUMP);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001406
1407 err = PTR_ERR(page);
1408 if (IS_ERR(page))
1409 goto set_status;
1410
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001411 err = page ? page_to_nid(page) : -ENOENT;
Christoph Lameter742755a2006-06-23 02:03:55 -07001412set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001413 *status = err;
1414
1415 pages++;
1416 status++;
1417 }
1418
1419 up_read(&mm->mmap_sem);
1420}
1421
1422/*
1423 * Determine the nodes of a user array of pages and store it in
1424 * a user array of status.
1425 */
1426static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1427 const void __user * __user *pages,
1428 int __user *status)
1429{
1430#define DO_PAGES_STAT_CHUNK_NR 16
1431 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1432 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001433
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001434 while (nr_pages) {
1435 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001436
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001437 chunk_nr = nr_pages;
1438 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1439 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1440
1441 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1442 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001443
1444 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1445
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001446 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1447 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001448
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001449 pages += chunk_nr;
1450 status += chunk_nr;
1451 nr_pages -= chunk_nr;
1452 }
1453 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001454}
1455
1456/*
1457 * Move a list of pages in the address space of the currently executing
1458 * process.
1459 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001460SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1461 const void __user * __user *, pages,
1462 const int __user *, nodes,
1463 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001464{
David Howellsc69e8d92008-11-14 10:39:19 +11001465 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001466 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001467 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001468 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001469 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001470
1471 /* Check flags */
1472 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1473 return -EINVAL;
1474
1475 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1476 return -EPERM;
1477
1478 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001479 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001480 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001481 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001482 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001483 return -ESRCH;
1484 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001485 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001486
1487 /*
1488 * Check if this process has the right to modify the specified
1489 * process. The right exists if the process has administrative
1490 * capabilities, superuser privileges or the same
1491 * userid as the target process.
1492 */
David Howellsc69e8d92008-11-14 10:39:19 +11001493 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001494 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1495 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001496 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001497 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001498 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001499 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001500 }
David Howellsc69e8d92008-11-14 10:39:19 +11001501 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001502
David Quigley86c3a762006-06-23 02:04:02 -07001503 err = security_task_movememory(task);
1504 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001505 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001506
Christoph Lameter3268c632012-03-21 16:34:06 -07001507 task_nodes = cpuset_mems_allowed(task);
1508 mm = get_task_mm(task);
1509 put_task_struct(task);
1510
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001511 if (!mm)
1512 return -EINVAL;
1513
1514 if (nodes)
1515 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1516 nodes, status, flags);
1517 else
1518 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001519
1520 mmput(mm);
1521 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001522
Christoph Lameter742755a2006-06-23 02:03:55 -07001523out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001524 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001525 return err;
1526}
Christoph Lameter742755a2006-06-23 02:03:55 -07001527
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001528#ifdef CONFIG_NUMA_BALANCING
1529/*
1530 * Returns true if this is a safe migration target node for misplaced NUMA
1531 * pages. Currently it only checks the watermarks which crude
1532 */
1533static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001534 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001535{
1536 int z;
1537 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1538 struct zone *zone = pgdat->node_zones + z;
1539
1540 if (!populated_zone(zone))
1541 continue;
1542
Lisa Du6e543d52013-09-11 14:22:36 -07001543 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001544 continue;
1545
1546 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1547 if (!zone_watermark_ok(zone, 0,
1548 high_wmark_pages(zone) +
1549 nr_migrate_pages,
1550 0, 0))
1551 continue;
1552 return true;
1553 }
1554 return false;
1555}
1556
1557static struct page *alloc_misplaced_dst_page(struct page *page,
1558 unsigned long data,
1559 int **result)
1560{
1561 int nid = (int) data;
1562 struct page *newpage;
1563
Vlastimil Babka96db8002015-09-08 15:03:50 -07001564 newpage = __alloc_pages_node(nid,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001565 (GFP_HIGHUSER_MOVABLE |
1566 __GFP_THISNODE | __GFP_NOMEMALLOC |
1567 __GFP_NORETRY | __GFP_NOWARN) &
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001568 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001569
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001570 return newpage;
1571}
1572
1573/*
Mel Gormana8f60772012-11-14 21:41:46 +00001574 * page migration rate limiting control.
1575 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1576 * window of time. Default here says do not migrate more than 1280M per second.
1577 */
1578static unsigned int migrate_interval_millisecs __read_mostly = 100;
1579static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1580
Mel Gormanb32967f2012-11-19 12:35:47 +00001581/* Returns true if the node is migrate rate-limited after the update */
Mel Gorman1c30e012014-01-21 15:50:58 -08001582static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1583 unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001584{
Mel Gormanb32967f2012-11-19 12:35:47 +00001585 /*
1586 * Rate-limit the amount of data that is being migrated to a node.
1587 * Optimal placement is no good if the memory bus is saturated and
1588 * all the time is being spent migrating!
1589 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001590 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001591 spin_lock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001592 pgdat->numabalancing_migrate_nr_pages = 0;
1593 pgdat->numabalancing_migrate_next_window = jiffies +
1594 msecs_to_jiffies(migrate_interval_millisecs);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001595 spin_unlock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001596 }
Mel Gormanaf1839d2014-01-21 15:51:01 -08001597 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1598 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1599 nr_pages);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001600 return true;
Mel Gormanaf1839d2014-01-21 15:51:01 -08001601 }
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001602
1603 /*
1604 * This is an unlocked non-atomic update so errors are possible.
1605 * The consequences are failing to migrate when we potentiall should
1606 * have which is not severe enough to warrant locking. If it is ever
1607 * a problem, it can be converted to a per-cpu counter.
1608 */
1609 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1610 return false;
Mel Gormanb32967f2012-11-19 12:35:47 +00001611}
1612
Mel Gorman1c30e012014-01-21 15:50:58 -08001613static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
Mel Gormanb32967f2012-11-19 12:35:47 +00001614{
Hugh Dickins340ef392013-02-22 16:34:33 -08001615 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001616
Sasha Levin309381fea2014-01-23 15:52:54 -08001617 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
Mel Gorman3abef4e2013-02-22 16:34:27 -08001618
Mel Gormanb32967f2012-11-19 12:35:47 +00001619 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001620 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1621 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001622
Hugh Dickins340ef392013-02-22 16:34:33 -08001623 if (isolate_lru_page(page))
1624 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001625
1626 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001627 * migrate_misplaced_transhuge_page() skips page migration's usual
1628 * check on page_count(), so we must do it here, now that the page
1629 * has been isolated: a GUP pin, or any other pin, prevents migration.
1630 * The expected page count is 3: 1 for page's mapcount and 1 for the
1631 * caller's pin and 1 for the reference taken by isolate_lru_page().
1632 */
1633 if (PageTransHuge(page) && page_count(page) != 3) {
1634 putback_lru_page(page);
1635 return 0;
1636 }
1637
1638 page_lru = page_is_file_cache(page);
1639 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1640 hpage_nr_pages(page));
1641
1642 /*
1643 * Isolating the page has taken another reference, so the
1644 * caller's reference can be safely dropped without the page
1645 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001646 */
1647 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001648 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001649}
1650
Mel Gormande466bd2013-12-18 17:08:42 -08001651bool pmd_trans_migrating(pmd_t pmd)
1652{
1653 struct page *page = pmd_page(pmd);
1654 return PageLocked(page);
1655}
1656
Mel Gormana8f60772012-11-14 21:41:46 +00001657/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001658 * Attempt to migrate a misplaced page to the specified destination
1659 * node. Caller is expected to have an elevated reference count on
1660 * the page that will be dropped by this function before returning.
1661 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001662int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1663 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001664{
Mel Gormana8f60772012-11-14 21:41:46 +00001665 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001666 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001667 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001668 LIST_HEAD(migratepages);
1669
1670 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001671 * Don't migrate file pages that are mapped in multiple processes
1672 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001673 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001674 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1675 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001676 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001677
Mel Gormana8f60772012-11-14 21:41:46 +00001678 /*
1679 * Rate-limit the amount of data that is being migrated to a node.
1680 * Optimal placement is no good if the memory bus is saturated and
1681 * all the time is being spent migrating!
1682 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001683 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001684 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001685
Mel Gormanb32967f2012-11-19 12:35:47 +00001686 isolated = numamigrate_isolate_page(pgdat, page);
1687 if (!isolated)
1688 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001689
Mel Gormanb32967f2012-11-19 12:35:47 +00001690 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001691 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
David Rientjes68711a72014-06-04 16:08:25 -07001692 NULL, node, MIGRATE_ASYNC,
1693 MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001694 if (nr_remaining) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001695 if (!list_empty(&migratepages)) {
1696 list_del(&page->lru);
1697 dec_zone_page_state(page, NR_ISOLATED_ANON +
1698 page_is_file_cache(page));
1699 putback_lru_page(page);
1700 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001701 isolated = 0;
1702 } else
1703 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001704 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001705 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001706
1707out:
1708 put_page(page);
1709 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001710}
Mel Gorman220018d2012-12-05 09:32:56 +00001711#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001712
Mel Gorman220018d2012-12-05 09:32:56 +00001713#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001714/*
1715 * Migrates a THP to a given target node. page must be locked and is unlocked
1716 * before returning.
1717 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001718int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1719 struct vm_area_struct *vma,
1720 pmd_t *pmd, pmd_t entry,
1721 unsigned long address,
1722 struct page *page, int node)
1723{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001724 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001725 pg_data_t *pgdat = NODE_DATA(node);
1726 int isolated = 0;
1727 struct page *new_page = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001728 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001729 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1730 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001731 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001732
1733 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001734 * Rate-limit the amount of data that is being migrated to a node.
1735 * Optimal placement is no good if the memory bus is saturated and
1736 * all the time is being spent migrating!
1737 */
Mel Gormand28d43352012-11-29 09:24:36 +00001738 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001739 goto out_dropref;
1740
1741 new_page = alloc_pages_node(node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001742 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1743 HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001744 if (!new_page)
1745 goto out_fail;
1746
Mel Gormanb32967f2012-11-19 12:35:47 +00001747 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001748 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001749 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001750 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001751 }
1752
Mel Gormanb0943d62013-12-18 17:08:46 -08001753 if (mm_tlb_flush_pending(mm))
1754 flush_tlb_range(vma, mmun_start, mmun_end);
1755
Mel Gormanb32967f2012-11-19 12:35:47 +00001756 /* Prepare a page as a migration target */
1757 __set_page_locked(new_page);
1758 SetPageSwapBacked(new_page);
1759
1760 /* anon mapping, we can simply copy page->mapping to the new page: */
1761 new_page->mapping = page->mapping;
1762 new_page->index = page->index;
1763 migrate_page_copy(new_page, page);
1764 WARN_ON(PageLRU(new_page));
1765
1766 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001767 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001768 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001769 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1770fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001771 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001772 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001773
1774 /* Reverse changes made by migrate_page_copy() */
1775 if (TestClearPageActive(new_page))
1776 SetPageActive(page);
1777 if (TestClearPageUnevictable(new_page))
1778 SetPageUnevictable(page);
1779 mlock_migrate_page(page, new_page);
1780
1781 unlock_page(new_page);
1782 put_page(new_page); /* Free it */
1783
Mel Gormana54a4072013-10-07 11:28:46 +01001784 /* Retake the callers reference and putback on LRU */
1785 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001786 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001787 mod_zone_page_state(page_zone(page),
1788 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001789
1790 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001791 }
1792
Mel Gorman2b4847e2013-12-18 17:08:32 -08001793 orig_entry = *pmd;
Mel Gormanb32967f2012-11-19 12:35:47 +00001794 entry = mk_pmd(new_page, vma->vm_page_prot);
Mel Gormanb32967f2012-11-19 12:35:47 +00001795 entry = pmd_mkhuge(entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001796 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001797
Mel Gorman2b4847e2013-12-18 17:08:32 -08001798 /*
1799 * Clear the old entry under pagetable lock and establish the new PTE.
1800 * Any parallel GUP will either observe the old page blocking on the
1801 * page lock, block on the page table lock or observe the new page.
1802 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1803 * guarantee the copy is visible before the pagetable update.
1804 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001805 flush_cache_range(vma, mmun_start, mmun_end);
Mel Gorman11de9922014-06-04 16:07:41 -07001806 page_add_anon_rmap(new_page, vma, mmun_start);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001807 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001808 set_pmd_at(mm, mmun_start, pmd, entry);
1809 flush_tlb_range(vma, mmun_start, mmun_end);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001810 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001811
1812 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08001813 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1814 flush_tlb_range(vma, mmun_start, mmun_end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001815 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001816 update_mmu_cache_pmd(vma, address, &entry);
1817 page_remove_rmap(new_page);
1818 goto fail_putback;
1819 }
1820
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001821 mem_cgroup_migrate(page, new_page, false);
1822
Mel Gormanb32967f2012-11-19 12:35:47 +00001823 page_remove_rmap(page);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001824
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001825 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001826 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001827
Mel Gorman11de9922014-06-04 16:07:41 -07001828 /* Take an "isolate" reference and put new page on the LRU. */
1829 get_page(new_page);
1830 putback_lru_page(new_page);
1831
Mel Gormanb32967f2012-11-19 12:35:47 +00001832 unlock_page(new_page);
1833 unlock_page(page);
1834 put_page(page); /* Drop the rmap reference */
1835 put_page(page); /* Drop the LRU isolation reference */
1836
1837 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1838 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1839
Mel Gormanb32967f2012-11-19 12:35:47 +00001840 mod_zone_page_state(page_zone(page),
1841 NR_ISOLATED_ANON + page_lru,
1842 -HPAGE_PMD_NR);
1843 return isolated;
1844
Hugh Dickins340ef392013-02-22 16:34:33 -08001845out_fail:
1846 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001847out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08001848 ptl = pmd_lock(mm, pmd);
1849 if (pmd_same(*pmd, entry)) {
Mel Gorman4d942462015-02-12 14:58:28 -08001850 entry = pmd_modify(entry, vma->vm_page_prot);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001851 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001852 update_mmu_cache_pmd(vma, address, &entry);
1853 }
1854 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001855
Mel Gormaneb4489f62013-12-18 17:08:39 -08001856out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08001857 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001858 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001859 return 0;
1860}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001861#endif /* CONFIG_NUMA_BALANCING */
1862
1863#endif /* CONFIG_NUMA */