blob: f53838fe3dfe6e84868b86eb98fedf7abd384ffb [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>
Christoph Lameterb20a3502006-03-22 00:09:12 -080040
Michal Nazarewicz0d1836c2010-12-21 17:24:26 -080041#include <asm/tlbflush.h>
42
Mel Gorman7b2a2d42012-10-19 14:07:31 +010043#define CREATE_TRACE_POINTS
44#include <trace/events/migrate.h>
45
Christoph Lameterb20a3502006-03-22 00:09:12 -080046#include "internal.h"
47
Christoph Lameterb20a3502006-03-22 00:09:12 -080048/*
Christoph Lameter742755a2006-06-23 02:03:55 -070049 * migrate_prep() needs to be called before we start compiling a list of pages
Mel Gorman748446b2010-05-24 14:32:27 -070050 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
51 * undesirable, use migrate_prep_local()
Christoph Lameterb20a3502006-03-22 00:09:12 -080052 */
53int migrate_prep(void)
54{
Christoph Lameterb20a3502006-03-22 00:09:12 -080055 /*
56 * Clear the LRU lists so pages can be isolated.
57 * Note that pages may be moved off the LRU after we have
58 * drained them. Those pages will fail to migrate like other
59 * pages that may be busy.
60 */
61 lru_add_drain_all();
62
63 return 0;
64}
65
Mel Gorman748446b2010-05-24 14:32:27 -070066/* Do the necessary work of migrate_prep but not if it involves other CPUs */
67int migrate_prep_local(void)
68{
69 lru_add_drain();
70
71 return 0;
72}
73
Christoph Lameterb20a3502006-03-22 00:09:12 -080074/*
Rafael Aquini5733c7d2012-12-11 16:02:47 -080075 * Put previously isolated pages back onto the appropriate lists
76 * from where they were once taken off for compaction/migration.
77 *
Joonsoo Kim59c82b72014-01-21 15:51:17 -080078 * This function shall be used whenever the isolated pageset has been
79 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
80 * and isolate_huge_page().
Rafael Aquini5733c7d2012-12-11 16:02:47 -080081 */
82void putback_movable_pages(struct list_head *l)
83{
84 struct page *page;
85 struct page *page2;
86
87 list_for_each_entry_safe(page, page2, l, lru) {
Naoya Horiguchi31caf662013-09-11 14:21:59 -070088 if (unlikely(PageHuge(page))) {
89 putback_active_hugepage(page);
90 continue;
91 }
Rafael Aquini5733c7d2012-12-11 16:02:47 -080092 list_del(&page->lru);
93 dec_zone_page_state(page, NR_ISOLATED_ANON +
94 page_is_file_cache(page));
Rafael Aquini117aad12013-09-30 13:45:16 -070095 if (unlikely(isolated_balloon_page(page)))
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080096 balloon_page_putback(page);
97 else
98 putback_lru_page(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -080099 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800100}
101
Christoph Lameter06972122006-06-23 02:03:35 -0700102/*
103 * Restore a potential migration pte to a working pte entry
104 */
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800105static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
106 unsigned long addr, void *old)
Christoph Lameter06972122006-06-23 02:03:35 -0700107{
108 struct mm_struct *mm = vma->vm_mm;
109 swp_entry_t entry;
Christoph Lameter06972122006-06-23 02:03:35 -0700110 pmd_t *pmd;
111 pte_t *ptep, pte;
112 spinlock_t *ptl;
113
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900114 if (unlikely(PageHuge(new))) {
115 ptep = huge_pte_offset(mm, addr);
116 if (!ptep)
117 goto out;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800118 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900119 } else {
Bob Liu62190492012-12-11 16:00:37 -0800120 pmd = mm_find_pmd(mm, addr);
121 if (!pmd)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900122 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700123
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900124 ptep = pte_offset_map(pmd, addr);
Christoph Lameter06972122006-06-23 02:03:35 -0700125
Hugh Dickins486cf462011-10-19 12:50:35 -0700126 /*
127 * Peek to check is_swap_pte() before taking ptlock? No, we
128 * can race mremap's move_ptes(), which skips anon_vma lock.
129 */
Christoph Lameter06972122006-06-23 02:03:35 -0700130
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900131 ptl = pte_lockptr(mm, pmd);
132 }
133
Christoph Lameter06972122006-06-23 02:03:35 -0700134 spin_lock(ptl);
135 pte = *ptep;
136 if (!is_swap_pte(pte))
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800137 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700138
139 entry = pte_to_swp_entry(pte);
140
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800141 if (!is_migration_entry(entry) ||
142 migration_entry_to_page(entry) != old)
143 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700144
Christoph Lameter06972122006-06-23 02:03:35 -0700145 get_page(new);
146 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700147 if (pte_swp_soft_dirty(*ptep))
148 pte = pte_mksoft_dirty(pte);
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100149
150 /* Recheck VMA as permissions can change since migration started */
Christoph Lameter06972122006-06-23 02:03:35 -0700151 if (is_write_migration_entry(entry))
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100152 pte = maybe_mkwrite(pte, vma);
153
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200154#ifdef CONFIG_HUGETLB_PAGE
Tony Lube7517d2013-02-04 14:28:46 -0800155 if (PageHuge(new)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900156 pte = pte_mkhuge(pte);
Tony Lube7517d2013-02-04 14:28:46 -0800157 pte = arch_make_huge_pte(pte, vma, new, 0);
158 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200159#endif
Leonid Yegoshinc2cc4992013-05-24 15:55:18 -0700160 flush_dcache_page(new);
Christoph Lameter06972122006-06-23 02:03:35 -0700161 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700162
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900163 if (PageHuge(new)) {
164 if (PageAnon(new))
165 hugepage_add_anon_rmap(new, vma, addr);
166 else
167 page_dup_rmap(new);
168 } else if (PageAnon(new))
Christoph Lameter04e62a22006-06-23 02:03:38 -0700169 page_add_anon_rmap(new, vma, addr);
170 else
171 page_add_file_rmap(new);
172
173 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000174 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800175unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700176 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800177out:
178 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700179}
180
181/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700182 * Get rid of all migration entries and replace them by
183 * references to the indicated page.
184 */
185static void remove_migration_ptes(struct page *old, struct page *new)
186{
Joonsoo Kim051ac832014-01-21 15:49:48 -0800187 struct rmap_walk_control rwc = {
188 .rmap_one = remove_migration_pte,
189 .arg = old,
190 };
191
192 rmap_walk(new, &rwc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700193}
194
195/*
Christoph Lameter06972122006-06-23 02:03:35 -0700196 * Something used the pte of a page under migration. We need to
197 * get to the page and wait until migration is finished.
198 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700199 */
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800200void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700201 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700202{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700203 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700204 swp_entry_t entry;
205 struct page *page;
206
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700207 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700208 pte = *ptep;
209 if (!is_swap_pte(pte))
210 goto out;
211
212 entry = pte_to_swp_entry(pte);
213 if (!is_migration_entry(entry))
214 goto out;
215
216 page = migration_entry_to_page(entry);
217
Nick Piggine2867812008-07-25 19:45:30 -0700218 /*
219 * Once radix-tree replacement of page migration started, page_count
220 * *must* be zero. And, we don't want to call wait_on_page_locked()
221 * against a page without get_page().
222 * So, we use get_page_unless_zero(), here. Even failed, page fault
223 * will occur again.
224 */
225 if (!get_page_unless_zero(page))
226 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700227 pte_unmap_unlock(ptep, ptl);
228 wait_on_page_locked(page);
229 put_page(page);
230 return;
231out:
232 pte_unmap_unlock(ptep, ptl);
233}
234
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700235void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
236 unsigned long address)
237{
238 spinlock_t *ptl = pte_lockptr(mm, pmd);
239 pte_t *ptep = pte_offset_map(pmd, address);
240 __migration_entry_wait(mm, ptep, ptl);
241}
242
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800243void migration_entry_wait_huge(struct vm_area_struct *vma,
244 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700245{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800246 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700247 __migration_entry_wait(mm, pte, ptl);
248}
249
Mel Gormanb969c4a2012-01-12 17:19:34 -0800250#ifdef CONFIG_BLOCK
251/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800252static bool buffer_migrate_lock_buffers(struct buffer_head *head,
253 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800254{
255 struct buffer_head *bh = head;
256
257 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800258 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800259 do {
260 get_bh(bh);
261 lock_buffer(bh);
262 bh = bh->b_this_page;
263
264 } while (bh != head);
265
266 return true;
267 }
268
269 /* async case, we cannot block on lock_buffer so use trylock_buffer */
270 do {
271 get_bh(bh);
272 if (!trylock_buffer(bh)) {
273 /*
274 * We failed to lock the buffer and cannot stall in
275 * async migration. Release the taken locks
276 */
277 struct buffer_head *failed_bh = bh;
278 put_bh(failed_bh);
279 bh = head;
280 while (bh != failed_bh) {
281 unlock_buffer(bh);
282 put_bh(bh);
283 bh = bh->b_this_page;
284 }
285 return false;
286 }
287
288 bh = bh->b_this_page;
289 } while (bh != head);
290 return true;
291}
292#else
293static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800294 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800295{
296 return true;
297}
298#endif /* CONFIG_BLOCK */
299
Christoph Lameterb20a3502006-03-22 00:09:12 -0800300/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700301 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700302 *
303 * The number of remaining references must be:
304 * 1 for anonymous pages without a mapping
305 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100306 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800307 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800308int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800309 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500310 struct buffer_head *head, enum migrate_mode mode,
311 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800312{
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500313 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800314 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800315
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700316 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700317 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500318 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700319 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800320 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700321 }
322
Nick Piggin19fd6232008-07-25 19:45:32 -0700323 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800324
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800325 pslot = radix_tree_lookup_slot(&mapping->page_tree,
326 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800327
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500328 expected_count += 1 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700329 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800330 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700331 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700332 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800333 }
334
Nick Piggine2867812008-07-25 19:45:30 -0700335 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700336 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700337 return -EAGAIN;
338 }
339
Christoph Lameterb20a3502006-03-22 00:09:12 -0800340 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800341 * In the async migration case of moving a page with buffers, lock the
342 * buffers using trylock before the mapping is moved. If the mapping
343 * was moved, we later failed to lock the buffers and could not move
344 * the mapping back due to an elevated page count, we would have to
345 * block waiting on other references to be dropped.
346 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800347 if (mode == MIGRATE_ASYNC && head &&
348 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800349 page_unfreeze_refs(page, expected_count);
350 spin_unlock_irq(&mapping->tree_lock);
351 return -EAGAIN;
352 }
353
354 /*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800355 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800356 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800357 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800358 if (PageSwapCache(page)) {
359 SetPageSwapCache(newpage);
360 set_page_private(newpage, page_private(page));
361 }
362
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800363 radix_tree_replace_slot(pslot, newpage);
364
365 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800366 * Drop cache reference from old page by unfreezing
367 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800368 * We know this isn't the last reference.
369 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800370 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800371
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700372 /*
373 * If moved to a different zone then also account
374 * the page for that zone. Other VM counters will be
375 * taken care of when we establish references to the
376 * new page and drop references to the old page.
377 *
378 * Note that anonymous pages are accounted for
379 * via NR_FILE_PAGES and NR_ANON_PAGES if they
380 * are mapped to swap space.
381 */
382 __dec_zone_page_state(page, NR_FILE_PAGES);
383 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700384 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700385 __dec_zone_page_state(page, NR_SHMEM);
386 __inc_zone_page_state(newpage, NR_SHMEM);
387 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700388 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800389
Rafael Aquini78bd5202012-12-11 16:02:31 -0800390 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800391}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800392
393/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900394 * The expected number of remaining references is the same as that
395 * of migrate_page_move_mapping().
396 */
397int migrate_huge_page_move_mapping(struct address_space *mapping,
398 struct page *newpage, struct page *page)
399{
400 int expected_count;
401 void **pslot;
402
403 if (!mapping) {
404 if (page_count(page) != 1)
405 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800406 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900407 }
408
409 spin_lock_irq(&mapping->tree_lock);
410
411 pslot = radix_tree_lookup_slot(&mapping->page_tree,
412 page_index(page));
413
414 expected_count = 2 + page_has_private(page);
415 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800416 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900417 spin_unlock_irq(&mapping->tree_lock);
418 return -EAGAIN;
419 }
420
421 if (!page_freeze_refs(page, expected_count)) {
422 spin_unlock_irq(&mapping->tree_lock);
423 return -EAGAIN;
424 }
425
426 get_page(newpage);
427
428 radix_tree_replace_slot(pslot, newpage);
429
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800430 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900431
432 spin_unlock_irq(&mapping->tree_lock);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800433 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900434}
435
436/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800437 * Gigantic pages are so large that we do not guarantee that page++ pointer
438 * arithmetic will work across the entire page. We need something more
439 * specialized.
440 */
441static void __copy_gigantic_page(struct page *dst, struct page *src,
442 int nr_pages)
443{
444 int i;
445 struct page *dst_base = dst;
446 struct page *src_base = src;
447
448 for (i = 0; i < nr_pages; ) {
449 cond_resched();
450 copy_highpage(dst, src);
451
452 i++;
453 dst = mem_map_next(dst, dst_base, i);
454 src = mem_map_next(src, src_base, i);
455 }
456}
457
458static void copy_huge_page(struct page *dst, struct page *src)
459{
460 int i;
461 int nr_pages;
462
463 if (PageHuge(src)) {
464 /* hugetlbfs page */
465 struct hstate *h = page_hstate(src);
466 nr_pages = pages_per_huge_page(h);
467
468 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
469 __copy_gigantic_page(dst, src, nr_pages);
470 return;
471 }
472 } else {
473 /* thp page */
474 BUG_ON(!PageTransHuge(src));
475 nr_pages = hpage_nr_pages(src);
476 }
477
478 for (i = 0; i < nr_pages; i++) {
479 cond_resched();
480 copy_highpage(dst + i, src + i);
481 }
482}
483
484/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800485 * Copy the page to its new location
486 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900487void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800488{
Rik van Riel7851a452013-10-07 11:29:23 +0100489 int cpupid;
490
Mel Gormanb32967f2012-11-19 12:35:47 +0000491 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900492 copy_huge_page(newpage, page);
493 else
494 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800495
496 if (PageError(page))
497 SetPageError(newpage);
498 if (PageReferenced(page))
499 SetPageReferenced(newpage);
500 if (PageUptodate(page))
501 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700502 if (TestClearPageActive(page)) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800503 VM_BUG_ON_PAGE(PageUnevictable(page), page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800504 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800505 } else if (TestClearPageUnevictable(page))
506 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800507 if (PageChecked(page))
508 SetPageChecked(newpage);
509 if (PageMappedToDisk(page))
510 SetPageMappedToDisk(newpage);
511
512 if (PageDirty(page)) {
513 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700514 /*
515 * Want to mark the page and the radix tree as dirty, and
516 * redo the accounting that clear_page_dirty_for_io undid,
517 * but we can't use set_page_dirty because that function
518 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300519 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700520 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700521 if (PageSwapBacked(page))
522 SetPageDirty(newpage);
523 else
524 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800525 }
526
Rik van Riel7851a452013-10-07 11:29:23 +0100527 /*
528 * Copy NUMA information to the new page, to prevent over-eager
529 * future migrations of this same page.
530 */
531 cpupid = page_cpupid_xchg_last(page, -1);
532 page_cpupid_xchg_last(newpage, cpupid);
533
Nick Pigginb291f002008-10-18 20:26:44 -0700534 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800535 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800536 /*
537 * Please do not reorder this without considering how mm/ksm.c's
538 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
539 */
Naoya Horiguchib3b3a992015-04-15 16:13:15 -0700540 if (PageSwapCache(page))
541 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800542 ClearPagePrivate(page);
543 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800544
545 /*
546 * If any waiters have accumulated on the new page then
547 * wake them up.
548 */
549 if (PageWriteback(newpage))
550 end_page_writeback(newpage);
551}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800552
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700553/************************************************************
554 * Migration functions
555 ***********************************************************/
556
Christoph Lameterb20a3502006-03-22 00:09:12 -0800557/*
558 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100559 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800560 *
561 * Pages are locked upon entry and exit.
562 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700563int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800564 struct page *newpage, struct page *page,
565 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800566{
567 int rc;
568
569 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
570
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500571 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800572
Rafael Aquini78bd5202012-12-11 16:02:31 -0800573 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800574 return rc;
575
576 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800577 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800578}
579EXPORT_SYMBOL(migrate_page);
580
David Howells93614012006-09-30 20:45:40 +0200581#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800582/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700583 * Migration function for pages with buffers. This function can only be used
584 * if the underlying filesystem guarantees that no other references to "page"
585 * exist.
586 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700587int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800588 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700589{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700590 struct buffer_head *bh, *head;
591 int rc;
592
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700593 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800594 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700595
596 head = page_buffers(page);
597
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500598 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700599
Rafael Aquini78bd5202012-12-11 16:02:31 -0800600 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700601 return rc;
602
Mel Gormanb969c4a2012-01-12 17:19:34 -0800603 /*
604 * In the async case, migrate_page_move_mapping locked the buffers
605 * with an IRQ-safe spinlock held. In the sync case, the buffers
606 * need to be locked now
607 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800608 if (mode != MIGRATE_ASYNC)
609 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700610
611 ClearPagePrivate(page);
612 set_page_private(newpage, page_private(page));
613 set_page_private(page, 0);
614 put_page(page);
615 get_page(newpage);
616
617 bh = head;
618 do {
619 set_bh_page(bh, newpage, bh_offset(bh));
620 bh = bh->b_this_page;
621
622 } while (bh != head);
623
624 SetPagePrivate(newpage);
625
626 migrate_page_copy(newpage, page);
627
628 bh = head;
629 do {
630 unlock_buffer(bh);
631 put_bh(bh);
632 bh = bh->b_this_page;
633
634 } while (bh != head);
635
Rafael Aquini78bd5202012-12-11 16:02:31 -0800636 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700637}
638EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200639#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700640
Christoph Lameter04e62a22006-06-23 02:03:38 -0700641/*
642 * Writeback a page to clean the dirty state
643 */
644static int writeout(struct address_space *mapping, struct page *page)
645{
646 struct writeback_control wbc = {
647 .sync_mode = WB_SYNC_NONE,
648 .nr_to_write = 1,
649 .range_start = 0,
650 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700651 .for_reclaim = 1
652 };
653 int rc;
654
655 if (!mapping->a_ops->writepage)
656 /* No write method for the address space */
657 return -EINVAL;
658
659 if (!clear_page_dirty_for_io(page))
660 /* Someone else already triggered a write */
661 return -EAGAIN;
662
663 /*
664 * A dirty page may imply that the underlying filesystem has
665 * the page on some queue. So the page must be clean for
666 * migration. Writeout may mean we loose the lock and the
667 * page state is no longer what we checked for earlier.
668 * At this point we know that the migration attempt cannot
669 * be successful.
670 */
671 remove_migration_ptes(page, page);
672
673 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700674
675 if (rc != AOP_WRITEPAGE_ACTIVATE)
676 /* unlocked. Relock */
677 lock_page(page);
678
Hugh Dickinsbda85502008-11-19 15:36:36 -0800679 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700680}
681
682/*
683 * Default handling if a filesystem does not provide a migration function.
684 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700685static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800686 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700687{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800688 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800689 /* Only writeback pages in full synchronous migration */
690 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800691 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700692 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800693 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700694
695 /*
696 * Buffers may be managed in a filesystem specific way.
697 * We must have no buffers or drop them.
698 */
David Howells266cf652009-04-03 16:42:36 +0100699 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700700 !try_to_release_page(page, GFP_KERNEL))
701 return -EAGAIN;
702
Mel Gormana6bc32b2012-01-12 17:19:43 -0800703 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700704}
705
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700706/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700707 * Move a page to a newly allocated page
708 * The page is locked and all ptes have been successfully removed.
709 *
710 * The new page will have replaced the old page if this function
711 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700712 *
713 * Return value:
714 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800715 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700716 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700717static int move_to_new_page(struct page *newpage, struct page *page,
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800718 int page_was_mapped, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700719{
720 struct address_space *mapping;
721 int rc;
722
723 /*
724 * Block others from accessing the page when we get around to
725 * establishing additional references. We are the only one
726 * holding a reference to the new page at this point.
727 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200728 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700729 BUG();
730
731 /* Prepare mapping for the new page.*/
732 newpage->index = page->index;
733 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700734 if (PageSwapBacked(page))
735 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700736
737 mapping = page_mapping(page);
738 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800739 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800740 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700741 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800742 * Most pages have a mapping and most filesystems provide a
743 * migratepage callback. Anonymous pages are part of swap
744 * space which also has its own migratepage callback. This
745 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700746 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800747 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800748 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800749 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800750 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700751
Rafael Aquini78bd5202012-12-11 16:02:31 -0800752 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700753 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700754 } else {
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700755 mem_cgroup_migrate(page, newpage, false);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800756 if (page_was_mapped)
Mel Gorman3fe20112010-05-24 14:32:20 -0700757 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800758 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700759 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700760
761 unlock_page(newpage);
762
763 return rc;
764}
765
Minchan Kim0dabec92011-10-31 17:06:57 -0700766static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800767 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700768{
Minchan Kim0dabec92011-10-31 17:06:57 -0700769 int rc = -EAGAIN;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800770 int page_was_mapped = 0;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700771 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700772
Nick Piggin529ae9a2008-08-02 12:01:03 +0200773 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800774 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700775 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800776
777 /*
778 * It's not safe for direct compaction to call lock_page.
779 * For example, during page readahead pages are added locked
780 * to the LRU. Later, when the IO completes the pages are
781 * marked uptodate and unlocked. However, the queueing
782 * could be merging multiple pages for one bio (e.g.
783 * mpage_readpages). If an allocation happens for the
784 * second or third page, the process can end up locking
785 * the same page twice and deadlocking. Rather than
786 * trying to be clever about what pages can be locked,
787 * avoid the use of lock_page for direct compaction
788 * altogether.
789 */
790 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700791 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800792
Christoph Lametere24f0b82006-06-23 02:03:51 -0700793 lock_page(page);
794 }
795
796 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700797 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700798 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800799 * necessary to wait for PageWriteback. In the async case,
800 * the retry loop is too short and in the sync-light case,
801 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700802 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800803 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700804 rc = -EBUSY;
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700805 goto out_unlock;
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700806 }
807 if (!force)
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700808 goto out_unlock;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700809 wait_on_page_writeback(page);
810 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700811 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700812 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
813 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800814 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700815 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700816 * File Caches may use write_page() or lock_page() in migration, then,
817 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700818 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800819 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800820 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000821 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800822 * getting a hold on an anon_vma from outside one of its mms.
823 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700824 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800825 if (anon_vma) {
826 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700827 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800828 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800829 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700830 /*
831 * We cannot be sure that the anon_vma of an unmapped
832 * swapcache page is safe to use because we don't
833 * know in advance if the VMA that this page belonged
834 * to still exists. If the VMA and others sharing the
835 * data have been freed, then the anon_vma could
836 * already be invalid.
837 *
838 * To avoid this possibility, swapcache pages get
839 * migrated but are not remapped when migration
840 * completes
841 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700842 } else {
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700843 goto out_unlock;
Mel Gorman3fe20112010-05-24 14:32:20 -0700844 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700845 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800846
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700847 if (unlikely(isolated_balloon_page(page))) {
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800848 /*
849 * A ballooned page does not need any special attention from
850 * physical to virtual reverse mapping procedures.
851 * Skip any attempt to unmap PTEs or to remap swap cache,
852 * in order to avoid burning cycles at rmap level, and perform
853 * the page migration right away (proteced by page lock).
854 */
855 rc = balloon_page_migrate(newpage, page, mode);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700856 goto out_unlock;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800857 }
858
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700859 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800860 * Corner case handling:
861 * 1. When a new swap-cache page is read into, it is added to the LRU
862 * and treated as swapcache but it has no rmap yet.
863 * Calling try_to_unmap() against a page->mapping==NULL page will
864 * trigger a BUG. So handle it here.
865 * 2. An orphaned page (see truncate_complete_page) might have
866 * fs-private metadata. The page can be picked up due to memory
867 * offlining. Everywhere else except page reclaim, the page is
868 * invisible to the vm, so the page can not be migrated. So try to
869 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700870 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800871 if (!page->mapping) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800872 VM_BUG_ON_PAGE(PageAnon(page), page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800873 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800874 try_to_free_buffers(page);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700875 goto out_unlock;
Shaohua Li62e1c552008-02-04 22:29:33 -0800876 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700877 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800878 }
879
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700880 /* Establish migration ptes or remove ptes */
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800881 if (page_mapped(page)) {
882 try_to_unmap(page,
883 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
884 page_was_mapped = 1;
885 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700886
Shaohua Liabfc3482009-09-21 17:01:19 -0700887skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700888 if (!page_mapped(page))
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800889 rc = move_to_new_page(newpage, page, page_was_mapped, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700890
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800891 if (rc && page_was_mapped)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700892 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700893
894 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700895 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700896 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700897
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700898out_unlock:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700899 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700900out:
901 return rc;
902}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700903
Minchan Kim0dabec92011-10-31 17:06:57 -0700904/*
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -0700905 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
906 * around it.
907 */
908#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
909#define ICE_noinline noinline
910#else
911#define ICE_noinline
912#endif
913
914/*
Minchan Kim0dabec92011-10-31 17:06:57 -0700915 * Obtain the lock on page, remove all ptes and migrate the page
916 * to the newly allocated page in newpage.
917 */
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -0700918static ICE_noinline int unmap_and_move(new_page_t get_new_page,
919 free_page_t put_new_page,
920 unsigned long private, struct page *page,
921 int force, enum migrate_mode mode)
Minchan Kim0dabec92011-10-31 17:06:57 -0700922{
923 int rc = 0;
924 int *result = NULL;
925 struct page *newpage = get_new_page(page, private, &result);
926
927 if (!newpage)
928 return -ENOMEM;
929
930 if (page_count(page) == 1) {
931 /* page was freed from under us. So we are done. */
932 goto out;
933 }
934
935 if (unlikely(PageTransHuge(page)))
936 if (unlikely(split_huge_page(page)))
937 goto out;
938
Hugh Dickins9c620e22013-02-22 16:35:14 -0800939 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800940
Minchan Kim0dabec92011-10-31 17:06:57 -0700941out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700942 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700943 /*
944 * A page that has been migrated has all references
945 * removed and will be freed. A page that has not been
946 * migrated will have kepts its references and be
947 * restored.
948 */
949 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700950 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700951 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700952 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700953 }
David Rientjes68711a72014-06-04 16:08:25 -0700954
Christoph Lameter95a402c2006-06-23 02:03:53 -0700955 /*
David Rientjes68711a72014-06-04 16:08:25 -0700956 * If migration was not successful and there's a freeing callback, use
957 * it. Otherwise, putback_lru_page() will drop the reference grabbed
958 * during isolation.
Christoph Lameter95a402c2006-06-23 02:03:53 -0700959 */
Hugh Dickins8bdd6382014-07-26 12:58:23 -0700960 if (rc != MIGRATEPAGE_SUCCESS && put_new_page) {
961 ClearPageSwapBacked(newpage);
David Rientjes68711a72014-06-04 16:08:25 -0700962 put_new_page(newpage, private);
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700963 } else if (unlikely(__is_movable_balloon_page(newpage))) {
964 /* drop our reference, page already in the balloon */
965 put_page(newpage);
Hugh Dickins8bdd6382014-07-26 12:58:23 -0700966 } else
David Rientjes68711a72014-06-04 16:08:25 -0700967 putback_lru_page(newpage);
968
Christoph Lameter742755a2006-06-23 02:03:55 -0700969 if (result) {
970 if (rc)
971 *result = rc;
972 else
973 *result = page_to_nid(newpage);
974 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700975 return rc;
976}
977
978/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900979 * Counterpart of unmap_and_move_page() for hugepage migration.
980 *
981 * This function doesn't wait the completion of hugepage I/O
982 * because there is no race between I/O and migration for hugepage.
983 * Note that currently hugepage I/O occurs only in direct I/O
984 * where no lock is held and PG_writeback is irrelevant,
985 * and writeback status of all subpages are counted in the reference
986 * count of the head page (i.e. if all subpages of a 2MB hugepage are
987 * under direct I/O, the reference of the head page is 512 and a bit more.)
988 * This means that when we try to migrate hugepage whose subpages are
989 * doing direct I/O, some references remain after try_to_unmap() and
990 * hugepage migration fails without data corruption.
991 *
992 * There is also no race when direct I/O is issued on the page under migration,
993 * because then pte is replaced with migration swap entry and direct I/O code
994 * will wait in the page fault for migration to complete.
995 */
996static int unmap_and_move_huge_page(new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -0700997 free_page_t put_new_page, unsigned long private,
998 struct page *hpage, int force,
999 enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001000{
1001 int rc = 0;
1002 int *result = NULL;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001003 int page_was_mapped = 0;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001004 struct page *new_hpage;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001005 struct anon_vma *anon_vma = NULL;
1006
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001007 /*
1008 * Movability of hugepages depends on architectures and hugepage size.
1009 * This check is necessary because some callers of hugepage migration
1010 * like soft offline and memory hotremove don't walk through page
1011 * tables or check whether the hugepage is pmd-based or not before
1012 * kicking migration.
1013 */
Naoya Horiguchi100873d2014-06-04 16:10:56 -07001014 if (!hugepage_migration_supported(page_hstate(hpage))) {
Joonsoo Kim32665f22014-01-21 15:51:15 -08001015 putback_active_hugepage(hpage);
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001016 return -ENOSYS;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001017 }
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001018
Joonsoo Kim32665f22014-01-21 15:51:15 -08001019 new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001020 if (!new_hpage)
1021 return -ENOMEM;
1022
1023 rc = -EAGAIN;
1024
1025 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001026 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001027 goto out;
1028 lock_page(hpage);
1029 }
1030
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001031 if (PageAnon(hpage))
1032 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001033
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001034 if (page_mapped(hpage)) {
1035 try_to_unmap(hpage,
1036 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1037 page_was_mapped = 1;
1038 }
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001039
1040 if (!page_mapped(hpage))
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001041 rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001042
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001043 if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001044 remove_migration_ptes(hpage, hpage);
1045
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001046 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001047 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001048
David Rientjes68711a72014-06-04 16:08:25 -07001049 if (rc == MIGRATEPAGE_SUCCESS)
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001050 hugetlb_cgroup_migrate(hpage, new_hpage);
1051
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001052 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001053out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001054 if (rc != -EAGAIN)
1055 putback_active_hugepage(hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001056
1057 /*
1058 * If migration was not successful and there's a freeing callback, use
1059 * it. Otherwise, put_page() will drop the reference grabbed during
1060 * isolation.
1061 */
1062 if (rc != MIGRATEPAGE_SUCCESS && put_new_page)
1063 put_new_page(new_hpage, private);
1064 else
1065 put_page(new_hpage);
1066
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001067 if (result) {
1068 if (rc)
1069 *result = rc;
1070 else
1071 *result = page_to_nid(new_hpage);
1072 }
1073 return rc;
1074}
1075
1076/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001077 * migrate_pages - migrate the pages specified in a list, to the free pages
1078 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001079 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001080 * @from: The list of pages to be migrated.
1081 * @get_new_page: The function used to allocate free pages to be used
1082 * as the target of the page migration.
David Rientjes68711a72014-06-04 16:08:25 -07001083 * @put_new_page: The function used to free target pages if migration
1084 * fails, or NULL if no special handling is necessary.
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001085 * @private: Private data to be passed on to get_new_page()
1086 * @mode: The migration mode that specifies the constraints for
1087 * page migration, if any.
1088 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001089 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001090 * The function returns after 10 attempts or if no pages are movable any more
1091 * because the list has become empty or no retryable pages exist any more.
1092 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001093 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001094 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001095 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001096 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001097int migrate_pages(struct list_head *from, new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001098 free_page_t put_new_page, unsigned long private,
1099 enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001100{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001101 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001102 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001103 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001104 int pass = 0;
1105 struct page *page;
1106 struct page *page2;
1107 int swapwrite = current->flags & PF_SWAPWRITE;
1108 int rc;
1109
1110 if (!swapwrite)
1111 current->flags |= PF_SWAPWRITE;
1112
Christoph Lametere24f0b82006-06-23 02:03:51 -07001113 for(pass = 0; pass < 10 && retry; pass++) {
1114 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001115
Christoph Lametere24f0b82006-06-23 02:03:51 -07001116 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001117 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001118
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001119 if (PageHuge(page))
1120 rc = unmap_and_move_huge_page(get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001121 put_new_page, private, page,
1122 pass > 2, mode);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001123 else
David Rientjes68711a72014-06-04 16:08:25 -07001124 rc = unmap_and_move(get_new_page, put_new_page,
1125 private, page, pass > 2, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001126
Christoph Lametere24f0b82006-06-23 02:03:51 -07001127 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001128 case -ENOMEM:
1129 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001130 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001131 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001132 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001133 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001134 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001135 break;
1136 default:
Naoya Horiguchi354a3362014-01-21 15:51:14 -08001137 /*
1138 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1139 * unlike -EAGAIN case, the failed page is
1140 * removed from migration page list and not
1141 * retried in the next outer loop.
1142 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001143 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001144 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001145 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001146 }
1147 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001148 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001149out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001150 if (nr_succeeded)
1151 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1152 if (nr_failed)
1153 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001154 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1155
Christoph Lameterb20a3502006-03-22 00:09:12 -08001156 if (!swapwrite)
1157 current->flags &= ~PF_SWAPWRITE;
1158
Rafael Aquini78bd5202012-12-11 16:02:31 -08001159 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001160}
1161
Christoph Lameter742755a2006-06-23 02:03:55 -07001162#ifdef CONFIG_NUMA
1163/*
1164 * Move a list of individual pages
1165 */
1166struct page_to_node {
1167 unsigned long addr;
1168 struct page *page;
1169 int node;
1170 int status;
1171};
1172
1173static struct page *new_page_node(struct page *p, unsigned long private,
1174 int **result)
1175{
1176 struct page_to_node *pm = (struct page_to_node *)private;
1177
1178 while (pm->node != MAX_NUMNODES && pm->page != p)
1179 pm++;
1180
1181 if (pm->node == MAX_NUMNODES)
1182 return NULL;
1183
1184 *result = &pm->status;
1185
Naoya Horiguchie632a932013-09-11 14:22:04 -07001186 if (PageHuge(p))
1187 return alloc_huge_page_node(page_hstate(compound_head(p)),
1188 pm->node);
1189 else
1190 return alloc_pages_exact_node(pm->node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001191 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001192}
1193
1194/*
1195 * Move a set of pages as indicated in the pm array. The addr
1196 * field must be set to the virtual address of the page to be moved
1197 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001198 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001199 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001200static int do_move_page_to_node_array(struct mm_struct *mm,
1201 struct page_to_node *pm,
1202 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001203{
1204 int err;
1205 struct page_to_node *pp;
1206 LIST_HEAD(pagelist);
1207
1208 down_read(&mm->mmap_sem);
1209
1210 /*
1211 * Build a list of pages to migrate
1212 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001213 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1214 struct vm_area_struct *vma;
1215 struct page *page;
1216
Christoph Lameter742755a2006-06-23 02:03:55 -07001217 err = -EFAULT;
1218 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001219 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001220 goto set_status;
1221
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001222 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001223
1224 err = PTR_ERR(page);
1225 if (IS_ERR(page))
1226 goto set_status;
1227
Christoph Lameter742755a2006-06-23 02:03:55 -07001228 err = -ENOENT;
1229 if (!page)
1230 goto set_status;
1231
Hugh Dickins62b61f62009-12-14 17:59:33 -08001232 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001233 if (PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001234 goto put_and_set;
1235
1236 pp->page = page;
1237 err = page_to_nid(page);
1238
1239 if (err == pp->node)
1240 /*
1241 * Node already in the right place
1242 */
1243 goto put_and_set;
1244
1245 err = -EACCES;
1246 if (page_mapcount(page) > 1 &&
1247 !migrate_all)
1248 goto put_and_set;
1249
Naoya Horiguchie632a932013-09-11 14:22:04 -07001250 if (PageHuge(page)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08001251 if (PageHead(page))
1252 isolate_huge_page(page, &pagelist);
Naoya Horiguchie632a932013-09-11 14:22:04 -07001253 goto put_and_set;
1254 }
1255
Nick Piggin62695a82008-10-18 20:26:09 -07001256 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001257 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001258 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001259 inc_zone_page_state(page, NR_ISOLATED_ANON +
1260 page_is_file_cache(page));
1261 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001262put_and_set:
1263 /*
1264 * Either remove the duplicate refcount from
1265 * isolate_lru_page() or drop the page ref if it was
1266 * not isolated.
1267 */
1268 put_page(page);
1269set_status:
1270 pp->status = err;
1271 }
1272
Brice Gogline78bbfa2008-10-18 20:27:15 -07001273 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001274 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -07001275 err = migrate_pages(&pagelist, new_page_node, NULL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001276 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001277 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001278 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001279 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001280
1281 up_read(&mm->mmap_sem);
1282 return err;
1283}
1284
1285/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001286 * Migrate an array of page address onto an array of nodes and fill
1287 * the corresponding array of status.
1288 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001289static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001290 unsigned long nr_pages,
1291 const void __user * __user *pages,
1292 const int __user *nodes,
1293 int __user *status, int flags)
1294{
Brice Goglin3140a222009-01-06 14:38:57 -08001295 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001296 unsigned long chunk_nr_pages;
1297 unsigned long chunk_start;
1298 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001299
Brice Goglin3140a222009-01-06 14:38:57 -08001300 err = -ENOMEM;
1301 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1302 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001303 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001304
1305 migrate_prep();
1306
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001307 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001308 * Store a chunk of page_to_node array in a page,
1309 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001310 */
Brice Goglin3140a222009-01-06 14:38:57 -08001311 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001312
Brice Goglin3140a222009-01-06 14:38:57 -08001313 for (chunk_start = 0;
1314 chunk_start < nr_pages;
1315 chunk_start += chunk_nr_pages) {
1316 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001317
Brice Goglin3140a222009-01-06 14:38:57 -08001318 if (chunk_start + chunk_nr_pages > nr_pages)
1319 chunk_nr_pages = nr_pages - chunk_start;
1320
1321 /* fill the chunk pm with addrs and nodes from user-space */
1322 for (j = 0; j < chunk_nr_pages; j++) {
1323 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001324 int node;
1325
Brice Goglin3140a222009-01-06 14:38:57 -08001326 err = -EFAULT;
1327 if (get_user(p, pages + j + chunk_start))
1328 goto out_pm;
1329 pm[j].addr = (unsigned long) p;
1330
1331 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001332 goto out_pm;
1333
1334 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001335 if (node < 0 || node >= MAX_NUMNODES)
1336 goto out_pm;
1337
Lai Jiangshan389162c2012-12-12 13:51:30 -08001338 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001339 goto out_pm;
1340
1341 err = -EACCES;
1342 if (!node_isset(node, task_nodes))
1343 goto out_pm;
1344
Brice Goglin3140a222009-01-06 14:38:57 -08001345 pm[j].node = node;
1346 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001347
Brice Goglin3140a222009-01-06 14:38:57 -08001348 /* End marker for this chunk */
1349 pm[chunk_nr_pages].node = MAX_NUMNODES;
1350
1351 /* Migrate this chunk */
1352 err = do_move_page_to_node_array(mm, pm,
1353 flags & MPOL_MF_MOVE_ALL);
1354 if (err < 0)
1355 goto out_pm;
1356
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001357 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001358 for (j = 0; j < chunk_nr_pages; j++)
1359 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001360 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001361 goto out_pm;
1362 }
1363 }
1364 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001365
1366out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001367 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001368out:
1369 return err;
1370}
1371
1372/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001373 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001374 */
Brice Goglin80bba122008-12-09 13:14:23 -08001375static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1376 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001377{
Brice Goglin2f007e72008-10-18 20:27:16 -07001378 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001379
Christoph Lameter742755a2006-06-23 02:03:55 -07001380 down_read(&mm->mmap_sem);
1381
Brice Goglin2f007e72008-10-18 20:27:16 -07001382 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001383 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001384 struct vm_area_struct *vma;
1385 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001386 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001387
1388 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001389 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001390 goto set_status;
1391
Brice Goglin2f007e72008-10-18 20:27:16 -07001392 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001393
1394 err = PTR_ERR(page);
1395 if (IS_ERR(page))
1396 goto set_status;
1397
Christoph Lameter742755a2006-06-23 02:03:55 -07001398 err = -ENOENT;
1399 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001400 if (!page || PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001401 goto set_status;
1402
1403 err = page_to_nid(page);
1404set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001405 *status = err;
1406
1407 pages++;
1408 status++;
1409 }
1410
1411 up_read(&mm->mmap_sem);
1412}
1413
1414/*
1415 * Determine the nodes of a user array of pages and store it in
1416 * a user array of status.
1417 */
1418static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1419 const void __user * __user *pages,
1420 int __user *status)
1421{
1422#define DO_PAGES_STAT_CHUNK_NR 16
1423 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1424 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001425
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001426 while (nr_pages) {
1427 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001428
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001429 chunk_nr = nr_pages;
1430 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1431 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1432
1433 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1434 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001435
1436 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1437
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001438 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1439 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001440
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001441 pages += chunk_nr;
1442 status += chunk_nr;
1443 nr_pages -= chunk_nr;
1444 }
1445 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001446}
1447
1448/*
1449 * Move a list of pages in the address space of the currently executing
1450 * process.
1451 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001452SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1453 const void __user * __user *, pages,
1454 const int __user *, nodes,
1455 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001456{
David Howellsc69e8d92008-11-14 10:39:19 +11001457 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001458 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001459 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001460 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001461 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001462
1463 /* Check flags */
1464 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1465 return -EINVAL;
1466
1467 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1468 return -EPERM;
1469
1470 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001471 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001472 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001473 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001474 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001475 return -ESRCH;
1476 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001477 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001478
1479 /*
1480 * Check if this process has the right to modify the specified
1481 * process. The right exists if the process has administrative
1482 * capabilities, superuser privileges or the same
1483 * userid as the target process.
1484 */
David Howellsc69e8d92008-11-14 10:39:19 +11001485 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001486 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1487 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001488 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001489 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001490 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001491 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001492 }
David Howellsc69e8d92008-11-14 10:39:19 +11001493 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001494
David Quigley86c3a762006-06-23 02:04:02 -07001495 err = security_task_movememory(task);
1496 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001497 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001498
Christoph Lameter3268c632012-03-21 16:34:06 -07001499 task_nodes = cpuset_mems_allowed(task);
1500 mm = get_task_mm(task);
1501 put_task_struct(task);
1502
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001503 if (!mm)
1504 return -EINVAL;
1505
1506 if (nodes)
1507 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1508 nodes, status, flags);
1509 else
1510 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001511
1512 mmput(mm);
1513 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001514
Christoph Lameter742755a2006-06-23 02:03:55 -07001515out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001516 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001517 return err;
1518}
Christoph Lameter742755a2006-06-23 02:03:55 -07001519
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001520#ifdef CONFIG_NUMA_BALANCING
1521/*
1522 * Returns true if this is a safe migration target node for misplaced NUMA
1523 * pages. Currently it only checks the watermarks which crude
1524 */
1525static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001526 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001527{
1528 int z;
1529 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1530 struct zone *zone = pgdat->node_zones + z;
1531
1532 if (!populated_zone(zone))
1533 continue;
1534
Lisa Du6e543d52013-09-11 14:22:36 -07001535 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001536 continue;
1537
1538 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1539 if (!zone_watermark_ok(zone, 0,
1540 high_wmark_pages(zone) +
1541 nr_migrate_pages,
1542 0, 0))
1543 continue;
1544 return true;
1545 }
1546 return false;
1547}
1548
1549static struct page *alloc_misplaced_dst_page(struct page *page,
1550 unsigned long data,
1551 int **result)
1552{
1553 int nid = (int) data;
1554 struct page *newpage;
1555
1556 newpage = alloc_pages_exact_node(nid,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001557 (GFP_HIGHUSER_MOVABLE |
1558 __GFP_THISNODE | __GFP_NOMEMALLOC |
1559 __GFP_NORETRY | __GFP_NOWARN) &
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001560 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001561
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001562 return newpage;
1563}
1564
1565/*
Mel Gormana8f60772012-11-14 21:41:46 +00001566 * page migration rate limiting control.
1567 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1568 * window of time. Default here says do not migrate more than 1280M per second.
1569 */
1570static unsigned int migrate_interval_millisecs __read_mostly = 100;
1571static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1572
Mel Gormanb32967f2012-11-19 12:35:47 +00001573/* Returns true if the node is migrate rate-limited after the update */
Mel Gorman1c30e012014-01-21 15:50:58 -08001574static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1575 unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001576{
Mel Gormanb32967f2012-11-19 12:35:47 +00001577 /*
1578 * Rate-limit the amount of data that is being migrated to a node.
1579 * Optimal placement is no good if the memory bus is saturated and
1580 * all the time is being spent migrating!
1581 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001582 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001583 spin_lock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001584 pgdat->numabalancing_migrate_nr_pages = 0;
1585 pgdat->numabalancing_migrate_next_window = jiffies +
1586 msecs_to_jiffies(migrate_interval_millisecs);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001587 spin_unlock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001588 }
Mel Gormanaf1839d2014-01-21 15:51:01 -08001589 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1590 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1591 nr_pages);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001592 return true;
Mel Gormanaf1839d2014-01-21 15:51:01 -08001593 }
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001594
1595 /*
1596 * This is an unlocked non-atomic update so errors are possible.
1597 * The consequences are failing to migrate when we potentiall should
1598 * have which is not severe enough to warrant locking. If it is ever
1599 * a problem, it can be converted to a per-cpu counter.
1600 */
1601 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1602 return false;
Mel Gormanb32967f2012-11-19 12:35:47 +00001603}
1604
Mel Gorman1c30e012014-01-21 15:50:58 -08001605static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
Mel Gormanb32967f2012-11-19 12:35:47 +00001606{
Hugh Dickins340ef392013-02-22 16:34:33 -08001607 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001608
Sasha Levin309381fea2014-01-23 15:52:54 -08001609 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
Mel Gorman3abef4e2013-02-22 16:34:27 -08001610
Mel Gormanb32967f2012-11-19 12:35:47 +00001611 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001612 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1613 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001614
Hugh Dickins340ef392013-02-22 16:34:33 -08001615 if (isolate_lru_page(page))
1616 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001617
1618 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001619 * migrate_misplaced_transhuge_page() skips page migration's usual
1620 * check on page_count(), so we must do it here, now that the page
1621 * has been isolated: a GUP pin, or any other pin, prevents migration.
1622 * The expected page count is 3: 1 for page's mapcount and 1 for the
1623 * caller's pin and 1 for the reference taken by isolate_lru_page().
1624 */
1625 if (PageTransHuge(page) && page_count(page) != 3) {
1626 putback_lru_page(page);
1627 return 0;
1628 }
1629
1630 page_lru = page_is_file_cache(page);
1631 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1632 hpage_nr_pages(page));
1633
1634 /*
1635 * Isolating the page has taken another reference, so the
1636 * caller's reference can be safely dropped without the page
1637 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001638 */
1639 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001640 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001641}
1642
Mel Gormande466bd2013-12-18 17:08:42 -08001643bool pmd_trans_migrating(pmd_t pmd)
1644{
1645 struct page *page = pmd_page(pmd);
1646 return PageLocked(page);
1647}
1648
Mel Gormana8f60772012-11-14 21:41:46 +00001649/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001650 * Attempt to migrate a misplaced page to the specified destination
1651 * node. Caller is expected to have an elevated reference count on
1652 * the page that will be dropped by this function before returning.
1653 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001654int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1655 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001656{
Mel Gormana8f60772012-11-14 21:41:46 +00001657 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001658 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001659 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001660 LIST_HEAD(migratepages);
1661
1662 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001663 * Don't migrate file pages that are mapped in multiple processes
1664 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001665 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001666 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1667 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001668 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001669
Mel Gormana8f60772012-11-14 21:41:46 +00001670 /*
1671 * Rate-limit the amount of data that is being migrated to a node.
1672 * Optimal placement is no good if the memory bus is saturated and
1673 * all the time is being spent migrating!
1674 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001675 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001676 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001677
Mel Gormanb32967f2012-11-19 12:35:47 +00001678 isolated = numamigrate_isolate_page(pgdat, page);
1679 if (!isolated)
1680 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001681
Mel Gormanb32967f2012-11-19 12:35:47 +00001682 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001683 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
David Rientjes68711a72014-06-04 16:08:25 -07001684 NULL, node, MIGRATE_ASYNC,
1685 MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001686 if (nr_remaining) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001687 if (!list_empty(&migratepages)) {
1688 list_del(&page->lru);
1689 dec_zone_page_state(page, NR_ISOLATED_ANON +
1690 page_is_file_cache(page));
1691 putback_lru_page(page);
1692 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001693 isolated = 0;
1694 } else
1695 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001696 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001697 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001698
1699out:
1700 put_page(page);
1701 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001702}
Mel Gorman220018d2012-12-05 09:32:56 +00001703#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001704
Mel Gorman220018d2012-12-05 09:32:56 +00001705#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001706/*
1707 * Migrates a THP to a given target node. page must be locked and is unlocked
1708 * before returning.
1709 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001710int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1711 struct vm_area_struct *vma,
1712 pmd_t *pmd, pmd_t entry,
1713 unsigned long address,
1714 struct page *page, int node)
1715{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001716 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001717 pg_data_t *pgdat = NODE_DATA(node);
1718 int isolated = 0;
1719 struct page *new_page = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001720 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001721 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1722 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001723 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001724
1725 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001726 * Rate-limit the amount of data that is being migrated to a node.
1727 * Optimal placement is no good if the memory bus is saturated and
1728 * all the time is being spent migrating!
1729 */
Mel Gormand28d43352012-11-29 09:24:36 +00001730 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001731 goto out_dropref;
1732
1733 new_page = alloc_pages_node(node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001734 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1735 HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001736 if (!new_page)
1737 goto out_fail;
1738
Mel Gormanb32967f2012-11-19 12:35:47 +00001739 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001740 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001741 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001742 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001743 }
1744
Mel Gormanb0943d62013-12-18 17:08:46 -08001745 if (mm_tlb_flush_pending(mm))
1746 flush_tlb_range(vma, mmun_start, mmun_end);
1747
Mel Gormanb32967f2012-11-19 12:35:47 +00001748 /* Prepare a page as a migration target */
1749 __set_page_locked(new_page);
1750 SetPageSwapBacked(new_page);
1751
1752 /* anon mapping, we can simply copy page->mapping to the new page: */
1753 new_page->mapping = page->mapping;
1754 new_page->index = page->index;
1755 migrate_page_copy(new_page, page);
1756 WARN_ON(PageLRU(new_page));
1757
1758 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001759 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001760 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001761 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1762fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001763 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001764 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001765
1766 /* Reverse changes made by migrate_page_copy() */
1767 if (TestClearPageActive(new_page))
1768 SetPageActive(page);
1769 if (TestClearPageUnevictable(new_page))
1770 SetPageUnevictable(page);
1771 mlock_migrate_page(page, new_page);
1772
1773 unlock_page(new_page);
1774 put_page(new_page); /* Free it */
1775
Mel Gormana54a4072013-10-07 11:28:46 +01001776 /* Retake the callers reference and putback on LRU */
1777 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001778 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001779 mod_zone_page_state(page_zone(page),
1780 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001781
1782 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001783 }
1784
Mel Gorman2b4847e2013-12-18 17:08:32 -08001785 orig_entry = *pmd;
Mel Gormanb32967f2012-11-19 12:35:47 +00001786 entry = mk_pmd(new_page, vma->vm_page_prot);
Mel Gormanb32967f2012-11-19 12:35:47 +00001787 entry = pmd_mkhuge(entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001788 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001789
Mel Gorman2b4847e2013-12-18 17:08:32 -08001790 /*
1791 * Clear the old entry under pagetable lock and establish the new PTE.
1792 * Any parallel GUP will either observe the old page blocking on the
1793 * page lock, block on the page table lock or observe the new page.
1794 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1795 * guarantee the copy is visible before the pagetable update.
1796 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001797 flush_cache_range(vma, mmun_start, mmun_end);
Mel Gorman11de9922014-06-04 16:07:41 -07001798 page_add_anon_rmap(new_page, vma, mmun_start);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001799 pmdp_clear_flush_notify(vma, mmun_start, pmd);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001800 set_pmd_at(mm, mmun_start, pmd, entry);
1801 flush_tlb_range(vma, mmun_start, mmun_end);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001802 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001803
1804 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08001805 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1806 flush_tlb_range(vma, mmun_start, mmun_end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001807 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001808 update_mmu_cache_pmd(vma, address, &entry);
1809 page_remove_rmap(new_page);
1810 goto fail_putback;
1811 }
1812
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001813 mem_cgroup_migrate(page, new_page, false);
1814
Mel Gormanb32967f2012-11-19 12:35:47 +00001815 page_remove_rmap(page);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001816
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001817 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001818 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001819
Mel Gorman11de9922014-06-04 16:07:41 -07001820 /* Take an "isolate" reference and put new page on the LRU. */
1821 get_page(new_page);
1822 putback_lru_page(new_page);
1823
Mel Gormanb32967f2012-11-19 12:35:47 +00001824 unlock_page(new_page);
1825 unlock_page(page);
1826 put_page(page); /* Drop the rmap reference */
1827 put_page(page); /* Drop the LRU isolation reference */
1828
1829 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1830 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1831
Mel Gormanb32967f2012-11-19 12:35:47 +00001832 mod_zone_page_state(page_zone(page),
1833 NR_ISOLATED_ANON + page_lru,
1834 -HPAGE_PMD_NR);
1835 return isolated;
1836
Hugh Dickins340ef392013-02-22 16:34:33 -08001837out_fail:
1838 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001839out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08001840 ptl = pmd_lock(mm, pmd);
1841 if (pmd_same(*pmd, entry)) {
Mel Gorman4d942462015-02-12 14:58:28 -08001842 entry = pmd_modify(entry, vma->vm_page_prot);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001843 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001844 update_mmu_cache_pmd(vma, address, &entry);
1845 }
1846 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001847
Mel Gormaneb4489f62013-12-18 17:08:39 -08001848out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08001849 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001850 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001851 return 0;
1852}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001853#endif /* CONFIG_NUMA_BALANCING */
1854
1855#endif /* CONFIG_NUMA */