blob: 9194375b230729fead356e8dd3c8f6bb951415ab [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/*
Lee Schermerhorn894bc312008-10-18 20:26:39 -070075 * Add isolated pages on the list back to the LRU under page lock
76 * to avoid leaking evictable pages back onto unevictable list.
Christoph Lameterb20a3502006-03-22 00:09:12 -080077 */
Minchan Kime13861d2010-05-24 14:31:59 -070078void putback_lru_pages(struct list_head *l)
Christoph Lameterb20a3502006-03-22 00:09:12 -080079{
80 struct page *page;
81 struct page *page2;
Christoph Lameterb20a3502006-03-22 00:09:12 -080082
83 list_for_each_entry_safe(page, page2, l, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -070084 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -070085 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -070086 page_is_file_cache(page));
Rafael Aquini5733c7d2012-12-11 16:02:47 -080087 putback_lru_page(page);
88 }
89}
90
91/*
92 * Put previously isolated pages back onto the appropriate lists
93 * from where they were once taken off for compaction/migration.
94 *
95 * This function shall be used instead of putback_lru_pages(),
96 * whenever the isolated pageset has been built by isolate_migratepages_range()
97 */
98void putback_movable_pages(struct list_head *l)
99{
100 struct page *page;
101 struct page *page2;
102
103 list_for_each_entry_safe(page, page2, l, lru) {
Naoya Horiguchi31caf662013-09-11 14:21:59 -0700104 if (unlikely(PageHuge(page))) {
105 putback_active_hugepage(page);
106 continue;
107 }
Rafael Aquini5733c7d2012-12-11 16:02:47 -0800108 list_del(&page->lru);
109 dec_zone_page_state(page, NR_ISOLATED_ANON +
110 page_is_file_cache(page));
Rafael Aquini117aad12013-09-30 13:45:16 -0700111 if (unlikely(isolated_balloon_page(page)))
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800112 balloon_page_putback(page);
113 else
114 putback_lru_page(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800115 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800116}
117
Christoph Lameter06972122006-06-23 02:03:35 -0700118/*
119 * Restore a potential migration pte to a working pte entry
120 */
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800121static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
122 unsigned long addr, void *old)
Christoph Lameter06972122006-06-23 02:03:35 -0700123{
124 struct mm_struct *mm = vma->vm_mm;
125 swp_entry_t entry;
Christoph Lameter06972122006-06-23 02:03:35 -0700126 pmd_t *pmd;
127 pte_t *ptep, pte;
128 spinlock_t *ptl;
129
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900130 if (unlikely(PageHuge(new))) {
131 ptep = huge_pte_offset(mm, addr);
132 if (!ptep)
133 goto out;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800134 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900135 } else {
Bob Liu62190492012-12-11 16:00:37 -0800136 pmd = mm_find_pmd(mm, addr);
137 if (!pmd)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900138 goto out;
Andrea Arcangeli500d65d2011-01-13 15:46:55 -0800139 if (pmd_trans_huge(*pmd))
140 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700141
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900142 ptep = pte_offset_map(pmd, addr);
Christoph Lameter06972122006-06-23 02:03:35 -0700143
Hugh Dickins486cf462011-10-19 12:50:35 -0700144 /*
145 * Peek to check is_swap_pte() before taking ptlock? No, we
146 * can race mremap's move_ptes(), which skips anon_vma lock.
147 */
Christoph Lameter06972122006-06-23 02:03:35 -0700148
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900149 ptl = pte_lockptr(mm, pmd);
150 }
151
Christoph Lameter06972122006-06-23 02:03:35 -0700152 spin_lock(ptl);
153 pte = *ptep;
154 if (!is_swap_pte(pte))
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800155 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700156
157 entry = pte_to_swp_entry(pte);
158
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800159 if (!is_migration_entry(entry) ||
160 migration_entry_to_page(entry) != old)
161 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700162
Christoph Lameter06972122006-06-23 02:03:35 -0700163 get_page(new);
164 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700165 if (pte_swp_soft_dirty(*ptep))
166 pte = pte_mksoft_dirty(pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700167 if (is_write_migration_entry(entry))
168 pte = pte_mkwrite(pte);
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200169#ifdef CONFIG_HUGETLB_PAGE
Tony Lube7517d2013-02-04 14:28:46 -0800170 if (PageHuge(new)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900171 pte = pte_mkhuge(pte);
Tony Lube7517d2013-02-04 14:28:46 -0800172 pte = arch_make_huge_pte(pte, vma, new, 0);
173 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200174#endif
Leonid Yegoshinc2cc4992013-05-24 15:55:18 -0700175 flush_dcache_page(new);
Christoph Lameter06972122006-06-23 02:03:35 -0700176 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700177
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900178 if (PageHuge(new)) {
179 if (PageAnon(new))
180 hugepage_add_anon_rmap(new, vma, addr);
181 else
182 page_dup_rmap(new);
183 } else if (PageAnon(new))
Christoph Lameter04e62a22006-06-23 02:03:38 -0700184 page_add_anon_rmap(new, vma, addr);
185 else
186 page_add_file_rmap(new);
187
188 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000189 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800190unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700191 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800192out:
193 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700194}
195
196/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700197 * Get rid of all migration entries and replace them by
198 * references to the indicated page.
199 */
200static void remove_migration_ptes(struct page *old, struct page *new)
201{
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800202 rmap_walk(new, remove_migration_pte, old);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700203}
204
205/*
Christoph Lameter06972122006-06-23 02:03:35 -0700206 * Something used the pte of a page under migration. We need to
207 * get to the page and wait until migration is finished.
208 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700209 */
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700210static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
211 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700212{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700213 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700214 swp_entry_t entry;
215 struct page *page;
216
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700217 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700218 pte = *ptep;
219 if (!is_swap_pte(pte))
220 goto out;
221
222 entry = pte_to_swp_entry(pte);
223 if (!is_migration_entry(entry))
224 goto out;
225
226 page = migration_entry_to_page(entry);
227
Nick Piggine2867812008-07-25 19:45:30 -0700228 /*
229 * Once radix-tree replacement of page migration started, page_count
230 * *must* be zero. And, we don't want to call wait_on_page_locked()
231 * against a page without get_page().
232 * So, we use get_page_unless_zero(), here. Even failed, page fault
233 * will occur again.
234 */
235 if (!get_page_unless_zero(page))
236 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700237 pte_unmap_unlock(ptep, ptl);
238 wait_on_page_locked(page);
239 put_page(page);
240 return;
241out:
242 pte_unmap_unlock(ptep, ptl);
243}
244
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700245void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
246 unsigned long address)
247{
248 spinlock_t *ptl = pte_lockptr(mm, pmd);
249 pte_t *ptep = pte_offset_map(pmd, address);
250 __migration_entry_wait(mm, ptep, ptl);
251}
252
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800253void migration_entry_wait_huge(struct vm_area_struct *vma,
254 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700255{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800256 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700257 __migration_entry_wait(mm, pte, ptl);
258}
259
Mel Gormanb969c4a2012-01-12 17:19:34 -0800260#ifdef CONFIG_BLOCK
261/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800262static bool buffer_migrate_lock_buffers(struct buffer_head *head,
263 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800264{
265 struct buffer_head *bh = head;
266
267 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800268 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800269 do {
270 get_bh(bh);
271 lock_buffer(bh);
272 bh = bh->b_this_page;
273
274 } while (bh != head);
275
276 return true;
277 }
278
279 /* async case, we cannot block on lock_buffer so use trylock_buffer */
280 do {
281 get_bh(bh);
282 if (!trylock_buffer(bh)) {
283 /*
284 * We failed to lock the buffer and cannot stall in
285 * async migration. Release the taken locks
286 */
287 struct buffer_head *failed_bh = bh;
288 put_bh(failed_bh);
289 bh = head;
290 while (bh != failed_bh) {
291 unlock_buffer(bh);
292 put_bh(bh);
293 bh = bh->b_this_page;
294 }
295 return false;
296 }
297
298 bh = bh->b_this_page;
299 } while (bh != head);
300 return true;
301}
302#else
303static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800304 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800305{
306 return true;
307}
308#endif /* CONFIG_BLOCK */
309
Christoph Lameterb20a3502006-03-22 00:09:12 -0800310/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700311 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700312 *
313 * The number of remaining references must be:
314 * 1 for anonymous pages without a mapping
315 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100316 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800317 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800318int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800319 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500320 struct buffer_head *head, enum migrate_mode mode,
321 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800322{
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500323 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800324 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800325
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700326 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700327 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500328 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700329 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800330 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700331 }
332
Nick Piggin19fd6232008-07-25 19:45:32 -0700333 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800334
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800335 pslot = radix_tree_lookup_slot(&mapping->page_tree,
336 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800337
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500338 expected_count += 1 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700339 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800340 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700341 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700342 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800343 }
344
Nick Piggine2867812008-07-25 19:45:30 -0700345 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700346 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700347 return -EAGAIN;
348 }
349
Christoph Lameterb20a3502006-03-22 00:09:12 -0800350 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800351 * In the async migration case of moving a page with buffers, lock the
352 * buffers using trylock before the mapping is moved. If the mapping
353 * was moved, we later failed to lock the buffers and could not move
354 * the mapping back due to an elevated page count, we would have to
355 * block waiting on other references to be dropped.
356 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800357 if (mode == MIGRATE_ASYNC && head &&
358 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800359 page_unfreeze_refs(page, expected_count);
360 spin_unlock_irq(&mapping->tree_lock);
361 return -EAGAIN;
362 }
363
364 /*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800365 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800366 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800367 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800368 if (PageSwapCache(page)) {
369 SetPageSwapCache(newpage);
370 set_page_private(newpage, page_private(page));
371 }
372
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800373 radix_tree_replace_slot(pslot, newpage);
374
375 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800376 * Drop cache reference from old page by unfreezing
377 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800378 * We know this isn't the last reference.
379 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800380 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800381
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700382 /*
383 * If moved to a different zone then also account
384 * the page for that zone. Other VM counters will be
385 * taken care of when we establish references to the
386 * new page and drop references to the old page.
387 *
388 * Note that anonymous pages are accounted for
389 * via NR_FILE_PAGES and NR_ANON_PAGES if they
390 * are mapped to swap space.
391 */
392 __dec_zone_page_state(page, NR_FILE_PAGES);
393 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700394 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700395 __dec_zone_page_state(page, NR_SHMEM);
396 __inc_zone_page_state(newpage, NR_SHMEM);
397 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700398 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800399
Rafael Aquini78bd5202012-12-11 16:02:31 -0800400 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800401}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800402
403/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900404 * The expected number of remaining references is the same as that
405 * of migrate_page_move_mapping().
406 */
407int migrate_huge_page_move_mapping(struct address_space *mapping,
408 struct page *newpage, struct page *page)
409{
410 int expected_count;
411 void **pslot;
412
413 if (!mapping) {
414 if (page_count(page) != 1)
415 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800416 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900417 }
418
419 spin_lock_irq(&mapping->tree_lock);
420
421 pslot = radix_tree_lookup_slot(&mapping->page_tree,
422 page_index(page));
423
424 expected_count = 2 + page_has_private(page);
425 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800426 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900427 spin_unlock_irq(&mapping->tree_lock);
428 return -EAGAIN;
429 }
430
431 if (!page_freeze_refs(page, expected_count)) {
432 spin_unlock_irq(&mapping->tree_lock);
433 return -EAGAIN;
434 }
435
436 get_page(newpage);
437
438 radix_tree_replace_slot(pslot, newpage);
439
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800440 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900441
442 spin_unlock_irq(&mapping->tree_lock);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800443 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900444}
445
446/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800447 * Gigantic pages are so large that we do not guarantee that page++ pointer
448 * arithmetic will work across the entire page. We need something more
449 * specialized.
450 */
451static void __copy_gigantic_page(struct page *dst, struct page *src,
452 int nr_pages)
453{
454 int i;
455 struct page *dst_base = dst;
456 struct page *src_base = src;
457
458 for (i = 0; i < nr_pages; ) {
459 cond_resched();
460 copy_highpage(dst, src);
461
462 i++;
463 dst = mem_map_next(dst, dst_base, i);
464 src = mem_map_next(src, src_base, i);
465 }
466}
467
468static void copy_huge_page(struct page *dst, struct page *src)
469{
470 int i;
471 int nr_pages;
472
473 if (PageHuge(src)) {
474 /* hugetlbfs page */
475 struct hstate *h = page_hstate(src);
476 nr_pages = pages_per_huge_page(h);
477
478 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
479 __copy_gigantic_page(dst, src, nr_pages);
480 return;
481 }
482 } else {
483 /* thp page */
484 BUG_ON(!PageTransHuge(src));
485 nr_pages = hpage_nr_pages(src);
486 }
487
488 for (i = 0; i < nr_pages; i++) {
489 cond_resched();
490 copy_highpage(dst + i, src + i);
491 }
492}
493
494/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800495 * Copy the page to its new location
496 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900497void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800498{
Rik van Riel7851a452013-10-07 11:29:23 +0100499 int cpupid;
500
Mel Gormanb32967f2012-11-19 12:35:47 +0000501 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900502 copy_huge_page(newpage, page);
503 else
504 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800505
506 if (PageError(page))
507 SetPageError(newpage);
508 if (PageReferenced(page))
509 SetPageReferenced(newpage);
510 if (PageUptodate(page))
511 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700512 if (TestClearPageActive(page)) {
513 VM_BUG_ON(PageUnevictable(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800514 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800515 } else if (TestClearPageUnevictable(page))
516 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800517 if (PageChecked(page))
518 SetPageChecked(newpage);
519 if (PageMappedToDisk(page))
520 SetPageMappedToDisk(newpage);
521
522 if (PageDirty(page)) {
523 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700524 /*
525 * Want to mark the page and the radix tree as dirty, and
526 * redo the accounting that clear_page_dirty_for_io undid,
527 * but we can't use set_page_dirty because that function
528 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300529 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700530 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700531 if (PageSwapBacked(page))
532 SetPageDirty(newpage);
533 else
534 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800535 }
536
Rik van Riel7851a452013-10-07 11:29:23 +0100537 /*
538 * Copy NUMA information to the new page, to prevent over-eager
539 * future migrations of this same page.
540 */
541 cpupid = page_cpupid_xchg_last(page, -1);
542 page_cpupid_xchg_last(newpage, cpupid);
543
Nick Pigginb291f002008-10-18 20:26:44 -0700544 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800545 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800546 /*
547 * Please do not reorder this without considering how mm/ksm.c's
548 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
549 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800550 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800551 ClearPagePrivate(page);
552 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800553
554 /*
555 * If any waiters have accumulated on the new page then
556 * wake them up.
557 */
558 if (PageWriteback(newpage))
559 end_page_writeback(newpage);
560}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800561
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700562/************************************************************
563 * Migration functions
564 ***********************************************************/
565
566/* Always fail migration. Used for mappings that are not movable */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700567int fail_migrate_page(struct address_space *mapping,
568 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700569{
570 return -EIO;
571}
572EXPORT_SYMBOL(fail_migrate_page);
573
Christoph Lameterb20a3502006-03-22 00:09:12 -0800574/*
575 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100576 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800577 *
578 * Pages are locked upon entry and exit.
579 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700580int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800581 struct page *newpage, struct page *page,
582 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800583{
584 int rc;
585
586 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
587
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500588 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800589
Rafael Aquini78bd5202012-12-11 16:02:31 -0800590 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800591 return rc;
592
593 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800594 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800595}
596EXPORT_SYMBOL(migrate_page);
597
David Howells93614012006-09-30 20:45:40 +0200598#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800599/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700600 * Migration function for pages with buffers. This function can only be used
601 * if the underlying filesystem guarantees that no other references to "page"
602 * exist.
603 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700604int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800605 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700606{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700607 struct buffer_head *bh, *head;
608 int rc;
609
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700610 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800611 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700612
613 head = page_buffers(page);
614
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500615 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700616
Rafael Aquini78bd5202012-12-11 16:02:31 -0800617 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700618 return rc;
619
Mel Gormanb969c4a2012-01-12 17:19:34 -0800620 /*
621 * In the async case, migrate_page_move_mapping locked the buffers
622 * with an IRQ-safe spinlock held. In the sync case, the buffers
623 * need to be locked now
624 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800625 if (mode != MIGRATE_ASYNC)
626 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700627
628 ClearPagePrivate(page);
629 set_page_private(newpage, page_private(page));
630 set_page_private(page, 0);
631 put_page(page);
632 get_page(newpage);
633
634 bh = head;
635 do {
636 set_bh_page(bh, newpage, bh_offset(bh));
637 bh = bh->b_this_page;
638
639 } while (bh != head);
640
641 SetPagePrivate(newpage);
642
643 migrate_page_copy(newpage, page);
644
645 bh = head;
646 do {
647 unlock_buffer(bh);
648 put_bh(bh);
649 bh = bh->b_this_page;
650
651 } while (bh != head);
652
Rafael Aquini78bd5202012-12-11 16:02:31 -0800653 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700654}
655EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200656#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700657
Christoph Lameter04e62a22006-06-23 02:03:38 -0700658/*
659 * Writeback a page to clean the dirty state
660 */
661static int writeout(struct address_space *mapping, struct page *page)
662{
663 struct writeback_control wbc = {
664 .sync_mode = WB_SYNC_NONE,
665 .nr_to_write = 1,
666 .range_start = 0,
667 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700668 .for_reclaim = 1
669 };
670 int rc;
671
672 if (!mapping->a_ops->writepage)
673 /* No write method for the address space */
674 return -EINVAL;
675
676 if (!clear_page_dirty_for_io(page))
677 /* Someone else already triggered a write */
678 return -EAGAIN;
679
680 /*
681 * A dirty page may imply that the underlying filesystem has
682 * the page on some queue. So the page must be clean for
683 * migration. Writeout may mean we loose the lock and the
684 * page state is no longer what we checked for earlier.
685 * At this point we know that the migration attempt cannot
686 * be successful.
687 */
688 remove_migration_ptes(page, page);
689
690 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700691
692 if (rc != AOP_WRITEPAGE_ACTIVATE)
693 /* unlocked. Relock */
694 lock_page(page);
695
Hugh Dickinsbda85502008-11-19 15:36:36 -0800696 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700697}
698
699/*
700 * Default handling if a filesystem does not provide a migration function.
701 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700702static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800703 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700704{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800705 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800706 /* Only writeback pages in full synchronous migration */
707 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800708 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700709 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800710 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700711
712 /*
713 * Buffers may be managed in a filesystem specific way.
714 * We must have no buffers or drop them.
715 */
David Howells266cf652009-04-03 16:42:36 +0100716 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700717 !try_to_release_page(page, GFP_KERNEL))
718 return -EAGAIN;
719
Mel Gormana6bc32b2012-01-12 17:19:43 -0800720 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700721}
722
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700723/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700724 * Move a page to a newly allocated page
725 * The page is locked and all ptes have been successfully removed.
726 *
727 * The new page will have replaced the old page if this function
728 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700729 *
730 * Return value:
731 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800732 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700733 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700734static int move_to_new_page(struct page *newpage, struct page *page,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800735 int remap_swapcache, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700736{
737 struct address_space *mapping;
738 int rc;
739
740 /*
741 * Block others from accessing the page when we get around to
742 * establishing additional references. We are the only one
743 * holding a reference to the new page at this point.
744 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200745 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700746 BUG();
747
748 /* Prepare mapping for the new page.*/
749 newpage->index = page->index;
750 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700751 if (PageSwapBacked(page))
752 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700753
754 mapping = page_mapping(page);
755 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800756 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800757 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700758 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800759 * Most pages have a mapping and most filesystems provide a
760 * migratepage callback. Anonymous pages are part of swap
761 * space which also has its own migratepage callback. This
762 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700763 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800764 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800765 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800766 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800767 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700768
Rafael Aquini78bd5202012-12-11 16:02:31 -0800769 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700770 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700771 } else {
772 if (remap_swapcache)
773 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800774 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700775 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700776
777 unlock_page(newpage);
778
779 return rc;
780}
781
Minchan Kim0dabec92011-10-31 17:06:57 -0700782static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800783 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700784{
Minchan Kim0dabec92011-10-31 17:06:57 -0700785 int rc = -EAGAIN;
Mel Gorman3fe20112010-05-24 14:32:20 -0700786 int remap_swapcache = 1;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -0700787 struct mem_cgroup *mem;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700788 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700789
Nick Piggin529ae9a2008-08-02 12:01:03 +0200790 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800791 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700792 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800793
794 /*
795 * It's not safe for direct compaction to call lock_page.
796 * For example, during page readahead pages are added locked
797 * to the LRU. Later, when the IO completes the pages are
798 * marked uptodate and unlocked. However, the queueing
799 * could be merging multiple pages for one bio (e.g.
800 * mpage_readpages). If an allocation happens for the
801 * second or third page, the process can end up locking
802 * the same page twice and deadlocking. Rather than
803 * trying to be clever about what pages can be locked,
804 * avoid the use of lock_page for direct compaction
805 * altogether.
806 */
807 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700808 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800809
Christoph Lametere24f0b82006-06-23 02:03:51 -0700810 lock_page(page);
811 }
812
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800813 /* charge against new page */
Johannes Weiner0030f532012-07-31 16:45:25 -0700814 mem_cgroup_prepare_migration(page, newpage, &mem);
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800815
Christoph Lametere24f0b82006-06-23 02:03:51 -0700816 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700817 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700818 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800819 * necessary to wait for PageWriteback. In the async case,
820 * the retry loop is too short and in the sync-light case,
821 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700822 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800823 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700824 rc = -EBUSY;
825 goto uncharge;
826 }
827 if (!force)
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800828 goto uncharge;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700829 wait_on_page_writeback(page);
830 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700831 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700832 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
833 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800834 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700835 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700836 * File Caches may use write_page() or lock_page() in migration, then,
837 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700838 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800839 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800840 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000841 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800842 * getting a hold on an anon_vma from outside one of its mms.
843 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700844 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800845 if (anon_vma) {
846 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700847 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800848 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800849 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700850 /*
851 * We cannot be sure that the anon_vma of an unmapped
852 * swapcache page is safe to use because we don't
853 * know in advance if the VMA that this page belonged
854 * to still exists. If the VMA and others sharing the
855 * data have been freed, then the anon_vma could
856 * already be invalid.
857 *
858 * To avoid this possibility, swapcache pages get
859 * migrated but are not remapped when migration
860 * completes
861 */
862 remap_swapcache = 0;
863 } else {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800864 goto uncharge;
Mel Gorman3fe20112010-05-24 14:32:20 -0700865 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700866 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800867
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800868 if (unlikely(balloon_page_movable(page))) {
869 /*
870 * A ballooned page does not need any special attention from
871 * physical to virtual reverse mapping procedures.
872 * Skip any attempt to unmap PTEs or to remap swap cache,
873 * in order to avoid burning cycles at rmap level, and perform
874 * the page migration right away (proteced by page lock).
875 */
876 rc = balloon_page_migrate(newpage, page, mode);
877 goto uncharge;
878 }
879
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700880 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800881 * Corner case handling:
882 * 1. When a new swap-cache page is read into, it is added to the LRU
883 * and treated as swapcache but it has no rmap yet.
884 * Calling try_to_unmap() against a page->mapping==NULL page will
885 * trigger a BUG. So handle it here.
886 * 2. An orphaned page (see truncate_complete_page) might have
887 * fs-private metadata. The page can be picked up due to memory
888 * offlining. Everywhere else except page reclaim, the page is
889 * invisible to the vm, so the page can not be migrated. So try to
890 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700891 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800892 if (!page->mapping) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800893 VM_BUG_ON(PageAnon(page));
894 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800895 try_to_free_buffers(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800896 goto uncharge;
Shaohua Li62e1c552008-02-04 22:29:33 -0800897 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700898 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800899 }
900
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700901 /* Establish migration ptes or remove ptes */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200902 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700903
Shaohua Liabfc3482009-09-21 17:01:19 -0700904skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700905 if (!page_mapped(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800906 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700907
Mel Gorman3fe20112010-05-24 14:32:20 -0700908 if (rc && remap_swapcache)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700909 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700910
911 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700912 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700913 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700914
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800915uncharge:
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800916 mem_cgroup_end_migration(mem, page, newpage,
917 (rc == MIGRATEPAGE_SUCCESS ||
918 rc == MIGRATEPAGE_BALLOON_SUCCESS));
Christoph Lametere24f0b82006-06-23 02:03:51 -0700919 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700920out:
921 return rc;
922}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700923
Minchan Kim0dabec92011-10-31 17:06:57 -0700924/*
925 * Obtain the lock on page, remove all ptes and migrate the page
926 * to the newly allocated page in newpage.
927 */
928static int unmap_and_move(new_page_t get_new_page, unsigned long private,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800929 struct page *page, int force, enum migrate_mode mode)
Minchan Kim0dabec92011-10-31 17:06:57 -0700930{
931 int rc = 0;
932 int *result = NULL;
933 struct page *newpage = get_new_page(page, private, &result);
934
935 if (!newpage)
936 return -ENOMEM;
937
938 if (page_count(page) == 1) {
939 /* page was freed from under us. So we are done. */
940 goto out;
941 }
942
943 if (unlikely(PageTransHuge(page)))
944 if (unlikely(split_huge_page(page)))
945 goto out;
946
Hugh Dickins9c620e22013-02-22 16:35:14 -0800947 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800948
949 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
950 /*
951 * A ballooned page has been migrated already.
952 * Now, it's the time to wrap-up counters,
953 * handle the page back to Buddy and return.
954 */
955 dec_zone_page_state(page, NR_ISOLATED_ANON +
956 page_is_file_cache(page));
957 balloon_page_free(page);
958 return MIGRATEPAGE_SUCCESS;
959 }
Minchan Kim0dabec92011-10-31 17:06:57 -0700960out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700961 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700962 /*
963 * A page that has been migrated has all references
964 * removed and will be freed. A page that has not been
965 * migrated will have kepts its references and be
966 * restored.
967 */
968 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700969 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700970 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700971 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700972 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700973 /*
974 * Move the new page to the LRU. If migration was not successful
975 * then this will free the page.
976 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700977 putback_lru_page(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700978 if (result) {
979 if (rc)
980 *result = rc;
981 else
982 *result = page_to_nid(newpage);
983 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700984 return rc;
985}
986
987/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900988 * Counterpart of unmap_and_move_page() for hugepage migration.
989 *
990 * This function doesn't wait the completion of hugepage I/O
991 * because there is no race between I/O and migration for hugepage.
992 * Note that currently hugepage I/O occurs only in direct I/O
993 * where no lock is held and PG_writeback is irrelevant,
994 * and writeback status of all subpages are counted in the reference
995 * count of the head page (i.e. if all subpages of a 2MB hugepage are
996 * under direct I/O, the reference of the head page is 512 and a bit more.)
997 * This means that when we try to migrate hugepage whose subpages are
998 * doing direct I/O, some references remain after try_to_unmap() and
999 * hugepage migration fails without data corruption.
1000 *
1001 * There is also no race when direct I/O is issued on the page under migration,
1002 * because then pte is replaced with migration swap entry and direct I/O code
1003 * will wait in the page fault for migration to complete.
1004 */
1005static int unmap_and_move_huge_page(new_page_t get_new_page,
1006 unsigned long private, struct page *hpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001007 int force, enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001008{
1009 int rc = 0;
1010 int *result = NULL;
1011 struct page *new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001012 struct anon_vma *anon_vma = NULL;
1013
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001014 /*
1015 * Movability of hugepages depends on architectures and hugepage size.
1016 * This check is necessary because some callers of hugepage migration
1017 * like soft offline and memory hotremove don't walk through page
1018 * tables or check whether the hugepage is pmd-based or not before
1019 * kicking migration.
1020 */
1021 if (!hugepage_migration_support(page_hstate(hpage)))
1022 return -ENOSYS;
1023
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001024 if (!new_hpage)
1025 return -ENOMEM;
1026
1027 rc = -EAGAIN;
1028
1029 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001030 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001031 goto out;
1032 lock_page(hpage);
1033 }
1034
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001035 if (PageAnon(hpage))
1036 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001037
1038 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1039
1040 if (!page_mapped(hpage))
Mel Gormana6bc32b2012-01-12 17:19:43 -08001041 rc = move_to_new_page(new_hpage, hpage, 1, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001042
1043 if (rc)
1044 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
1049 if (!rc)
1050 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);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001056 put_page(new_hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001057 if (result) {
1058 if (rc)
1059 *result = rc;
1060 else
1061 *result = page_to_nid(new_hpage);
1062 }
1063 return rc;
1064}
1065
1066/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001067 * migrate_pages - migrate the pages specified in a list, to the free pages
1068 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001069 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001070 * @from: The list of pages to be migrated.
1071 * @get_new_page: The function used to allocate free pages to be used
1072 * as the target of the page migration.
1073 * @private: Private data to be passed on to get_new_page()
1074 * @mode: The migration mode that specifies the constraints for
1075 * page migration, if any.
1076 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001077 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001078 * The function returns after 10 attempts or if no pages are movable any more
1079 * because the list has become empty or no retryable pages exist any more.
1080 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001081 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001082 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001083 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001084 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001085int migrate_pages(struct list_head *from, new_page_t get_new_page,
1086 unsigned long private, enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001087{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001088 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001089 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001090 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001091 int pass = 0;
1092 struct page *page;
1093 struct page *page2;
1094 int swapwrite = current->flags & PF_SWAPWRITE;
1095 int rc;
1096
1097 if (!swapwrite)
1098 current->flags |= PF_SWAPWRITE;
1099
Christoph Lametere24f0b82006-06-23 02:03:51 -07001100 for(pass = 0; pass < 10 && retry; pass++) {
1101 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001102
Christoph Lametere24f0b82006-06-23 02:03:51 -07001103 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001104 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001105
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001106 if (PageHuge(page))
1107 rc = unmap_and_move_huge_page(get_new_page,
1108 private, page, pass > 2, mode);
1109 else
1110 rc = unmap_and_move(get_new_page, private,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001111 page, pass > 2, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001112
Christoph Lametere24f0b82006-06-23 02:03:51 -07001113 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001114 case -ENOMEM:
1115 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001116 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001117 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001118 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001119 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001120 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001121 break;
1122 default:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001123 /* Permanent failure */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001124 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001125 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001126 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001127 }
1128 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001129 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001130out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001131 if (nr_succeeded)
1132 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1133 if (nr_failed)
1134 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001135 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1136
Christoph Lameterb20a3502006-03-22 00:09:12 -08001137 if (!swapwrite)
1138 current->flags &= ~PF_SWAPWRITE;
1139
Rafael Aquini78bd5202012-12-11 16:02:31 -08001140 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001141}
1142
Christoph Lameter742755a2006-06-23 02:03:55 -07001143#ifdef CONFIG_NUMA
1144/*
1145 * Move a list of individual pages
1146 */
1147struct page_to_node {
1148 unsigned long addr;
1149 struct page *page;
1150 int node;
1151 int status;
1152};
1153
1154static struct page *new_page_node(struct page *p, unsigned long private,
1155 int **result)
1156{
1157 struct page_to_node *pm = (struct page_to_node *)private;
1158
1159 while (pm->node != MAX_NUMNODES && pm->page != p)
1160 pm++;
1161
1162 if (pm->node == MAX_NUMNODES)
1163 return NULL;
1164
1165 *result = &pm->status;
1166
Naoya Horiguchie632a932013-09-11 14:22:04 -07001167 if (PageHuge(p))
1168 return alloc_huge_page_node(page_hstate(compound_head(p)),
1169 pm->node);
1170 else
1171 return alloc_pages_exact_node(pm->node,
Mel Gorman769848c2007-07-17 04:03:05 -07001172 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001173}
1174
1175/*
1176 * Move a set of pages as indicated in the pm array. The addr
1177 * field must be set to the virtual address of the page to be moved
1178 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001179 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001180 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001181static int do_move_page_to_node_array(struct mm_struct *mm,
1182 struct page_to_node *pm,
1183 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001184{
1185 int err;
1186 struct page_to_node *pp;
1187 LIST_HEAD(pagelist);
1188
1189 down_read(&mm->mmap_sem);
1190
1191 /*
1192 * Build a list of pages to migrate
1193 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001194 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1195 struct vm_area_struct *vma;
1196 struct page *page;
1197
Christoph Lameter742755a2006-06-23 02:03:55 -07001198 err = -EFAULT;
1199 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001200 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001201 goto set_status;
1202
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001203 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001204
1205 err = PTR_ERR(page);
1206 if (IS_ERR(page))
1207 goto set_status;
1208
Christoph Lameter742755a2006-06-23 02:03:55 -07001209 err = -ENOENT;
1210 if (!page)
1211 goto set_status;
1212
Hugh Dickins62b61f62009-12-14 17:59:33 -08001213 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001214 if (PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001215 goto put_and_set;
1216
1217 pp->page = page;
1218 err = page_to_nid(page);
1219
1220 if (err == pp->node)
1221 /*
1222 * Node already in the right place
1223 */
1224 goto put_and_set;
1225
1226 err = -EACCES;
1227 if (page_mapcount(page) > 1 &&
1228 !migrate_all)
1229 goto put_and_set;
1230
Naoya Horiguchie632a932013-09-11 14:22:04 -07001231 if (PageHuge(page)) {
1232 isolate_huge_page(page, &pagelist);
1233 goto put_and_set;
1234 }
1235
Nick Piggin62695a82008-10-18 20:26:09 -07001236 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001237 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001238 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001239 inc_zone_page_state(page, NR_ISOLATED_ANON +
1240 page_is_file_cache(page));
1241 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001242put_and_set:
1243 /*
1244 * Either remove the duplicate refcount from
1245 * isolate_lru_page() or drop the page ref if it was
1246 * not isolated.
1247 */
1248 put_page(page);
1249set_status:
1250 pp->status = err;
1251 }
1252
Brice Gogline78bbfa2008-10-18 20:27:15 -07001253 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001254 if (!list_empty(&pagelist)) {
Christoph Lameter742755a2006-06-23 02:03:55 -07001255 err = migrate_pages(&pagelist, new_page_node,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001256 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001257 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001258 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001259 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001260
1261 up_read(&mm->mmap_sem);
1262 return err;
1263}
1264
1265/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001266 * Migrate an array of page address onto an array of nodes and fill
1267 * the corresponding array of status.
1268 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001269static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001270 unsigned long nr_pages,
1271 const void __user * __user *pages,
1272 const int __user *nodes,
1273 int __user *status, int flags)
1274{
Brice Goglin3140a222009-01-06 14:38:57 -08001275 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001276 unsigned long chunk_nr_pages;
1277 unsigned long chunk_start;
1278 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001279
Brice Goglin3140a222009-01-06 14:38:57 -08001280 err = -ENOMEM;
1281 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1282 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001283 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001284
1285 migrate_prep();
1286
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001287 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001288 * Store a chunk of page_to_node array in a page,
1289 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001290 */
Brice Goglin3140a222009-01-06 14:38:57 -08001291 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001292
Brice Goglin3140a222009-01-06 14:38:57 -08001293 for (chunk_start = 0;
1294 chunk_start < nr_pages;
1295 chunk_start += chunk_nr_pages) {
1296 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001297
Brice Goglin3140a222009-01-06 14:38:57 -08001298 if (chunk_start + chunk_nr_pages > nr_pages)
1299 chunk_nr_pages = nr_pages - chunk_start;
1300
1301 /* fill the chunk pm with addrs and nodes from user-space */
1302 for (j = 0; j < chunk_nr_pages; j++) {
1303 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001304 int node;
1305
Brice Goglin3140a222009-01-06 14:38:57 -08001306 err = -EFAULT;
1307 if (get_user(p, pages + j + chunk_start))
1308 goto out_pm;
1309 pm[j].addr = (unsigned long) p;
1310
1311 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001312 goto out_pm;
1313
1314 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001315 if (node < 0 || node >= MAX_NUMNODES)
1316 goto out_pm;
1317
Lai Jiangshan389162c2012-12-12 13:51:30 -08001318 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001319 goto out_pm;
1320
1321 err = -EACCES;
1322 if (!node_isset(node, task_nodes))
1323 goto out_pm;
1324
Brice Goglin3140a222009-01-06 14:38:57 -08001325 pm[j].node = node;
1326 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001327
Brice Goglin3140a222009-01-06 14:38:57 -08001328 /* End marker for this chunk */
1329 pm[chunk_nr_pages].node = MAX_NUMNODES;
1330
1331 /* Migrate this chunk */
1332 err = do_move_page_to_node_array(mm, pm,
1333 flags & MPOL_MF_MOVE_ALL);
1334 if (err < 0)
1335 goto out_pm;
1336
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001337 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001338 for (j = 0; j < chunk_nr_pages; j++)
1339 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001340 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001341 goto out_pm;
1342 }
1343 }
1344 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001345
1346out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001347 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001348out:
1349 return err;
1350}
1351
1352/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001353 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001354 */
Brice Goglin80bba122008-12-09 13:14:23 -08001355static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1356 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001357{
Brice Goglin2f007e72008-10-18 20:27:16 -07001358 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001359
Christoph Lameter742755a2006-06-23 02:03:55 -07001360 down_read(&mm->mmap_sem);
1361
Brice Goglin2f007e72008-10-18 20:27:16 -07001362 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001363 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001364 struct vm_area_struct *vma;
1365 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001366 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001367
1368 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001369 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001370 goto set_status;
1371
Brice Goglin2f007e72008-10-18 20:27:16 -07001372 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001373
1374 err = PTR_ERR(page);
1375 if (IS_ERR(page))
1376 goto set_status;
1377
Christoph Lameter742755a2006-06-23 02:03:55 -07001378 err = -ENOENT;
1379 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001380 if (!page || PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001381 goto set_status;
1382
1383 err = page_to_nid(page);
1384set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001385 *status = err;
1386
1387 pages++;
1388 status++;
1389 }
1390
1391 up_read(&mm->mmap_sem);
1392}
1393
1394/*
1395 * Determine the nodes of a user array of pages and store it in
1396 * a user array of status.
1397 */
1398static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1399 const void __user * __user *pages,
1400 int __user *status)
1401{
1402#define DO_PAGES_STAT_CHUNK_NR 16
1403 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1404 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001405
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001406 while (nr_pages) {
1407 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001408
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001409 chunk_nr = nr_pages;
1410 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1411 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1412
1413 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1414 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001415
1416 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1417
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001418 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1419 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001420
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001421 pages += chunk_nr;
1422 status += chunk_nr;
1423 nr_pages -= chunk_nr;
1424 }
1425 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001426}
1427
1428/*
1429 * Move a list of pages in the address space of the currently executing
1430 * process.
1431 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001432SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1433 const void __user * __user *, pages,
1434 const int __user *, nodes,
1435 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001436{
David Howellsc69e8d92008-11-14 10:39:19 +11001437 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001438 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001439 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001440 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001441 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001442
1443 /* Check flags */
1444 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1445 return -EINVAL;
1446
1447 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1448 return -EPERM;
1449
1450 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001451 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001452 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001453 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001454 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001455 return -ESRCH;
1456 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001457 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001458
1459 /*
1460 * Check if this process has the right to modify the specified
1461 * process. The right exists if the process has administrative
1462 * capabilities, superuser privileges or the same
1463 * userid as the target process.
1464 */
David Howellsc69e8d92008-11-14 10:39:19 +11001465 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001466 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1467 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001468 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001469 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001470 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001471 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001472 }
David Howellsc69e8d92008-11-14 10:39:19 +11001473 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001474
David Quigley86c3a762006-06-23 02:04:02 -07001475 err = security_task_movememory(task);
1476 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001477 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001478
Christoph Lameter3268c632012-03-21 16:34:06 -07001479 task_nodes = cpuset_mems_allowed(task);
1480 mm = get_task_mm(task);
1481 put_task_struct(task);
1482
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001483 if (!mm)
1484 return -EINVAL;
1485
1486 if (nodes)
1487 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1488 nodes, status, flags);
1489 else
1490 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001491
1492 mmput(mm);
1493 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001494
Christoph Lameter742755a2006-06-23 02:03:55 -07001495out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001496 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001497 return err;
1498}
Christoph Lameter742755a2006-06-23 02:03:55 -07001499
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001500/*
1501 * Call migration functions in the vma_ops that may prepare
1502 * memory in a vm for migration. migration functions may perform
1503 * the migration for vmas that do not have an underlying page struct.
1504 */
1505int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1506 const nodemask_t *from, unsigned long flags)
1507{
1508 struct vm_area_struct *vma;
1509 int err = 0;
1510
Daisuke Nishimura1001c9f2009-02-11 13:04:18 -08001511 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001512 if (vma->vm_ops && vma->vm_ops->migrate) {
1513 err = vma->vm_ops->migrate(vma, to, from, flags);
1514 if (err)
1515 break;
1516 }
1517 }
1518 return err;
1519}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001520
1521#ifdef CONFIG_NUMA_BALANCING
1522/*
1523 * Returns true if this is a safe migration target node for misplaced NUMA
1524 * pages. Currently it only checks the watermarks which crude
1525 */
1526static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001527 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001528{
1529 int z;
1530 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1531 struct zone *zone = pgdat->node_zones + z;
1532
1533 if (!populated_zone(zone))
1534 continue;
1535
Lisa Du6e543d52013-09-11 14:22:36 -07001536 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001537 continue;
1538
1539 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1540 if (!zone_watermark_ok(zone, 0,
1541 high_wmark_pages(zone) +
1542 nr_migrate_pages,
1543 0, 0))
1544 continue;
1545 return true;
1546 }
1547 return false;
1548}
1549
1550static struct page *alloc_misplaced_dst_page(struct page *page,
1551 unsigned long data,
1552 int **result)
1553{
1554 int nid = (int) data;
1555 struct page *newpage;
1556
1557 newpage = alloc_pages_exact_node(nid,
1558 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1559 __GFP_NOMEMALLOC | __GFP_NORETRY |
1560 __GFP_NOWARN) &
1561 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001562 if (newpage)
Peter Zijlstra90572892013-10-07 11:29:20 +01001563 page_cpupid_xchg_last(newpage, page_cpupid_last(page));
Hillf Dantonbac03822012-11-27 14:46:24 +00001564
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001565 return newpage;
1566}
1567
1568/*
Mel Gormana8f60772012-11-14 21:41:46 +00001569 * page migration rate limiting control.
1570 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1571 * window of time. Default here says do not migrate more than 1280M per second.
Mel Gormane14808b2012-11-19 10:59:15 +00001572 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1573 * as it is faults that reset the window, pte updates will happen unconditionally
1574 * if there has not been a fault since @pteupdate_interval_millisecs after the
1575 * throttle window closed.
Mel Gormana8f60772012-11-14 21:41:46 +00001576 */
1577static unsigned int migrate_interval_millisecs __read_mostly = 100;
Mel Gormane14808b2012-11-19 10:59:15 +00001578static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
Mel Gormana8f60772012-11-14 21:41:46 +00001579static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1580
Mel Gormane14808b2012-11-19 10:59:15 +00001581/* Returns true if NUMA migration is currently rate limited */
1582bool migrate_ratelimited(int node)
1583{
1584 pg_data_t *pgdat = NODE_DATA(node);
1585
1586 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1587 msecs_to_jiffies(pteupdate_interval_millisecs)))
1588 return false;
1589
1590 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1591 return false;
1592
1593 return true;
1594}
1595
Mel Gormanb32967f2012-11-19 12:35:47 +00001596/* Returns true if the node is migrate rate-limited after the update */
Mel Gormand28d43352012-11-29 09:24:36 +00001597bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001598{
1599 bool rate_limited = false;
1600
1601 /*
1602 * Rate-limit the amount of data that is being migrated to a node.
1603 * Optimal placement is no good if the memory bus is saturated and
1604 * all the time is being spent migrating!
1605 */
1606 spin_lock(&pgdat->numabalancing_migrate_lock);
1607 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1608 pgdat->numabalancing_migrate_nr_pages = 0;
1609 pgdat->numabalancing_migrate_next_window = jiffies +
1610 msecs_to_jiffies(migrate_interval_millisecs);
1611 }
1612 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1613 rate_limited = true;
1614 else
Mel Gormand28d43352012-11-29 09:24:36 +00001615 pgdat->numabalancing_migrate_nr_pages += nr_pages;
Mel Gormanb32967f2012-11-19 12:35:47 +00001616 spin_unlock(&pgdat->numabalancing_migrate_lock);
1617
1618 return rate_limited;
1619}
1620
1621int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1622{
Hugh Dickins340ef392013-02-22 16:34:33 -08001623 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001624
Mel Gorman3abef4e2013-02-22 16:34:27 -08001625 VM_BUG_ON(compound_order(page) && !PageTransHuge(page));
1626
Mel Gormanb32967f2012-11-19 12:35:47 +00001627 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001628 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1629 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001630
Hugh Dickins340ef392013-02-22 16:34:33 -08001631 if (isolate_lru_page(page))
1632 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001633
1634 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001635 * migrate_misplaced_transhuge_page() skips page migration's usual
1636 * check on page_count(), so we must do it here, now that the page
1637 * has been isolated: a GUP pin, or any other pin, prevents migration.
1638 * The expected page count is 3: 1 for page's mapcount and 1 for the
1639 * caller's pin and 1 for the reference taken by isolate_lru_page().
1640 */
1641 if (PageTransHuge(page) && page_count(page) != 3) {
1642 putback_lru_page(page);
1643 return 0;
1644 }
1645
1646 page_lru = page_is_file_cache(page);
1647 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1648 hpage_nr_pages(page));
1649
1650 /*
1651 * Isolating the page has taken another reference, so the
1652 * caller's reference can be safely dropped without the page
1653 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001654 */
1655 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001656 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001657}
1658
Mel Gormande466bd2013-12-18 17:08:42 -08001659bool pmd_trans_migrating(pmd_t pmd)
1660{
1661 struct page *page = pmd_page(pmd);
1662 return PageLocked(page);
1663}
1664
1665void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd)
1666{
1667 struct page *page = pmd_page(*pmd);
1668 wait_on_page_locked(page);
1669}
1670
Mel Gormana8f60772012-11-14 21:41:46 +00001671/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001672 * Attempt to migrate a misplaced page to the specified destination
1673 * node. Caller is expected to have an elevated reference count on
1674 * the page that will be dropped by this function before returning.
1675 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001676int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1677 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001678{
Mel Gormana8f60772012-11-14 21:41:46 +00001679 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001680 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001681 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001682 LIST_HEAD(migratepages);
1683
1684 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001685 * Don't migrate file pages that are mapped in multiple processes
1686 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001687 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001688 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1689 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001690 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001691
Mel Gormana8f60772012-11-14 21:41:46 +00001692 /*
1693 * Rate-limit the amount of data that is being migrated to a node.
1694 * Optimal placement is no good if the memory bus is saturated and
1695 * all the time is being spent migrating!
1696 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001697 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001698 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001699
Mel Gormanb32967f2012-11-19 12:35:47 +00001700 isolated = numamigrate_isolate_page(pgdat, page);
1701 if (!isolated)
1702 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001703
Mel Gormanb32967f2012-11-19 12:35:47 +00001704 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001705 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1706 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001707 if (nr_remaining) {
1708 putback_lru_pages(&migratepages);
1709 isolated = 0;
1710 } else
1711 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001712 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001713 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001714
1715out:
1716 put_page(page);
1717 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001718}
Mel Gorman220018d2012-12-05 09:32:56 +00001719#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001720
Mel Gorman220018d2012-12-05 09:32:56 +00001721#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001722/*
1723 * Migrates a THP to a given target node. page must be locked and is unlocked
1724 * before returning.
1725 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001726int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1727 struct vm_area_struct *vma,
1728 pmd_t *pmd, pmd_t entry,
1729 unsigned long address,
1730 struct page *page, int node)
1731{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001732 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001733 pg_data_t *pgdat = NODE_DATA(node);
1734 int isolated = 0;
1735 struct page *new_page = NULL;
1736 struct mem_cgroup *memcg = NULL;
1737 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001738 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1739 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001740 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001741
1742 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001743 * Rate-limit the amount of data that is being migrated to a node.
1744 * Optimal placement is no good if the memory bus is saturated and
1745 * all the time is being spent migrating!
1746 */
Mel Gormand28d43352012-11-29 09:24:36 +00001747 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001748 goto out_dropref;
1749
1750 new_page = alloc_pages_node(node,
1751 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001752 if (!new_page)
1753 goto out_fail;
1754
Peter Zijlstra90572892013-10-07 11:29:20 +01001755 page_cpupid_xchg_last(new_page, page_cpupid_last(page));
Mel Gormanb32967f2012-11-19 12:35:47 +00001756
1757 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001758 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001759 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001760 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001761 }
1762
Mel Gormanb0943d62013-12-18 17:08:46 -08001763 if (mm_tlb_flush_pending(mm))
1764 flush_tlb_range(vma, mmun_start, mmun_end);
1765
Mel Gormanb32967f2012-11-19 12:35:47 +00001766 /* Prepare a page as a migration target */
1767 __set_page_locked(new_page);
1768 SetPageSwapBacked(new_page);
1769
1770 /* anon mapping, we can simply copy page->mapping to the new page: */
1771 new_page->mapping = page->mapping;
1772 new_page->index = page->index;
1773 migrate_page_copy(new_page, page);
1774 WARN_ON(PageLRU(new_page));
1775
1776 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001777 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001778 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001779 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1780fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001781 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001782 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001783
1784 /* Reverse changes made by migrate_page_copy() */
1785 if (TestClearPageActive(new_page))
1786 SetPageActive(page);
1787 if (TestClearPageUnevictable(new_page))
1788 SetPageUnevictable(page);
1789 mlock_migrate_page(page, new_page);
1790
1791 unlock_page(new_page);
1792 put_page(new_page); /* Free it */
1793
Mel Gormana54a4072013-10-07 11:28:46 +01001794 /* Retake the callers reference and putback on LRU */
1795 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001796 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001797 mod_zone_page_state(page_zone(page),
1798 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001799
1800 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001801 }
1802
1803 /*
1804 * Traditional migration needs to prepare the memcg charge
1805 * transaction early to prevent the old page from being
1806 * uncharged when installing migration entries. Here we can
1807 * save the potential rollback and start the charge transfer
1808 * only when migration is already known to end successfully.
1809 */
1810 mem_cgroup_prepare_migration(page, new_page, &memcg);
1811
Mel Gorman2b4847e2013-12-18 17:08:32 -08001812 orig_entry = *pmd;
Mel Gormanb32967f2012-11-19 12:35:47 +00001813 entry = mk_pmd(new_page, vma->vm_page_prot);
Mel Gormanb32967f2012-11-19 12:35:47 +00001814 entry = pmd_mkhuge(entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001815 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001816
Mel Gorman2b4847e2013-12-18 17:08:32 -08001817 /*
1818 * Clear the old entry under pagetable lock and establish the new PTE.
1819 * Any parallel GUP will either observe the old page blocking on the
1820 * page lock, block on the page table lock or observe the new page.
1821 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1822 * guarantee the copy is visible before the pagetable update.
1823 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001824 flush_cache_range(vma, mmun_start, mmun_end);
1825 page_add_new_anon_rmap(new_page, vma, mmun_start);
1826 pmdp_clear_flush(vma, mmun_start, pmd);
1827 set_pmd_at(mm, mmun_start, pmd, entry);
1828 flush_tlb_range(vma, mmun_start, mmun_end);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001829 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001830
1831 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08001832 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1833 flush_tlb_range(vma, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001834 update_mmu_cache_pmd(vma, address, &entry);
1835 page_remove_rmap(new_page);
1836 goto fail_putback;
1837 }
1838
Mel Gormanb32967f2012-11-19 12:35:47 +00001839 page_remove_rmap(page);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001840
Mel Gormanb32967f2012-11-19 12:35:47 +00001841 /*
1842 * Finish the charge transaction under the page table lock to
1843 * prevent split_huge_page() from dividing up the charge
1844 * before it's fully transferred to the new page.
1845 */
1846 mem_cgroup_end_migration(memcg, page, new_page, true);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001847 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001848 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001849
1850 unlock_page(new_page);
1851 unlock_page(page);
1852 put_page(page); /* Drop the rmap reference */
1853 put_page(page); /* Drop the LRU isolation reference */
1854
1855 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1856 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1857
Mel Gormanb32967f2012-11-19 12:35:47 +00001858 mod_zone_page_state(page_zone(page),
1859 NR_ISOLATED_ANON + page_lru,
1860 -HPAGE_PMD_NR);
1861 return isolated;
1862
Hugh Dickins340ef392013-02-22 16:34:33 -08001863out_fail:
1864 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001865out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08001866 ptl = pmd_lock(mm, pmd);
1867 if (pmd_same(*pmd, entry)) {
1868 entry = pmd_mknonnuma(entry);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001869 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001870 update_mmu_cache_pmd(vma, address, &entry);
1871 }
1872 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001873
Mel Gormaneb4489f62013-12-18 17:08:39 -08001874out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08001875 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001876 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001877 return 0;
1878}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001879#endif /* CONFIG_NUMA_BALANCING */
1880
1881#endif /* CONFIG_NUMA */