blob: b494fdb9a6363e058f802069ba5de8e677005b45 [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;
Andrea Arcangeli500d65d2011-01-13 15:46:55 -0800123 if (pmd_trans_huge(*pmd))
124 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700125
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900126 ptep = pte_offset_map(pmd, addr);
Christoph Lameter06972122006-06-23 02:03:35 -0700127
Hugh Dickins486cf462011-10-19 12:50:35 -0700128 /*
129 * Peek to check is_swap_pte() before taking ptlock? No, we
130 * can race mremap's move_ptes(), which skips anon_vma lock.
131 */
Christoph Lameter06972122006-06-23 02:03:35 -0700132
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900133 ptl = pte_lockptr(mm, pmd);
134 }
135
Christoph Lameter06972122006-06-23 02:03:35 -0700136 spin_lock(ptl);
137 pte = *ptep;
138 if (!is_swap_pte(pte))
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800139 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700140
141 entry = pte_to_swp_entry(pte);
142
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800143 if (!is_migration_entry(entry) ||
144 migration_entry_to_page(entry) != old)
145 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700146
Christoph Lameter06972122006-06-23 02:03:35 -0700147 get_page(new);
148 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700149 if (pte_swp_soft_dirty(*ptep))
150 pte = pte_mksoft_dirty(pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700151 if (is_write_migration_entry(entry))
152 pte = pte_mkwrite(pte);
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200153#ifdef CONFIG_HUGETLB_PAGE
Tony Lube7517d2013-02-04 14:28:46 -0800154 if (PageHuge(new)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900155 pte = pte_mkhuge(pte);
Tony Lube7517d2013-02-04 14:28:46 -0800156 pte = arch_make_huge_pte(pte, vma, new, 0);
157 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200158#endif
Leonid Yegoshinc2cc4992013-05-24 15:55:18 -0700159 flush_dcache_page(new);
Christoph Lameter06972122006-06-23 02:03:35 -0700160 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700161
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900162 if (PageHuge(new)) {
163 if (PageAnon(new))
164 hugepage_add_anon_rmap(new, vma, addr);
165 else
166 page_dup_rmap(new);
167 } else if (PageAnon(new))
Christoph Lameter04e62a22006-06-23 02:03:38 -0700168 page_add_anon_rmap(new, vma, addr);
169 else
170 page_add_file_rmap(new);
171
172 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000173 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800174unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700175 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800176out:
177 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700178}
179
180/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700181 * Get rid of all migration entries and replace them by
182 * references to the indicated page.
183 */
184static void remove_migration_ptes(struct page *old, struct page *new)
185{
Joonsoo Kim051ac832014-01-21 15:49:48 -0800186 struct rmap_walk_control rwc = {
187 .rmap_one = remove_migration_pte,
188 .arg = old,
189 };
190
191 rmap_walk(new, &rwc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700192}
193
194/*
Christoph Lameter06972122006-06-23 02:03:35 -0700195 * Something used the pte of a page under migration. We need to
196 * get to the page and wait until migration is finished.
197 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700198 */
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700199static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
200 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700201{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700202 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700203 swp_entry_t entry;
204 struct page *page;
205
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700206 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700207 pte = *ptep;
208 if (!is_swap_pte(pte))
209 goto out;
210
211 entry = pte_to_swp_entry(pte);
212 if (!is_migration_entry(entry))
213 goto out;
214
215 page = migration_entry_to_page(entry);
216
Nick Piggine2867812008-07-25 19:45:30 -0700217 /*
218 * Once radix-tree replacement of page migration started, page_count
219 * *must* be zero. And, we don't want to call wait_on_page_locked()
220 * against a page without get_page().
221 * So, we use get_page_unless_zero(), here. Even failed, page fault
222 * will occur again.
223 */
224 if (!get_page_unless_zero(page))
225 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700226 pte_unmap_unlock(ptep, ptl);
227 wait_on_page_locked(page);
228 put_page(page);
229 return;
230out:
231 pte_unmap_unlock(ptep, ptl);
232}
233
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700234void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
235 unsigned long address)
236{
237 spinlock_t *ptl = pte_lockptr(mm, pmd);
238 pte_t *ptep = pte_offset_map(pmd, address);
239 __migration_entry_wait(mm, ptep, ptl);
240}
241
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800242void migration_entry_wait_huge(struct vm_area_struct *vma,
243 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700244{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800245 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700246 __migration_entry_wait(mm, pte, ptl);
247}
248
Mel Gormanb969c4a2012-01-12 17:19:34 -0800249#ifdef CONFIG_BLOCK
250/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800251static bool buffer_migrate_lock_buffers(struct buffer_head *head,
252 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800253{
254 struct buffer_head *bh = head;
255
256 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800257 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800258 do {
259 get_bh(bh);
260 lock_buffer(bh);
261 bh = bh->b_this_page;
262
263 } while (bh != head);
264
265 return true;
266 }
267
268 /* async case, we cannot block on lock_buffer so use trylock_buffer */
269 do {
270 get_bh(bh);
271 if (!trylock_buffer(bh)) {
272 /*
273 * We failed to lock the buffer and cannot stall in
274 * async migration. Release the taken locks
275 */
276 struct buffer_head *failed_bh = bh;
277 put_bh(failed_bh);
278 bh = head;
279 while (bh != failed_bh) {
280 unlock_buffer(bh);
281 put_bh(bh);
282 bh = bh->b_this_page;
283 }
284 return false;
285 }
286
287 bh = bh->b_this_page;
288 } while (bh != head);
289 return true;
290}
291#else
292static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800293 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800294{
295 return true;
296}
297#endif /* CONFIG_BLOCK */
298
Christoph Lameterb20a3502006-03-22 00:09:12 -0800299/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700300 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700301 *
302 * The number of remaining references must be:
303 * 1 for anonymous pages without a mapping
304 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100305 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800306 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800307int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800308 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500309 struct buffer_head *head, enum migrate_mode mode,
310 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800311{
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500312 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800313 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800314
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700315 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700316 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500317 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700318 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800319 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700320 }
321
Nick Piggin19fd6232008-07-25 19:45:32 -0700322 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800323
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800324 pslot = radix_tree_lookup_slot(&mapping->page_tree,
325 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800326
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500327 expected_count += 1 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700328 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800329 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700330 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700331 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800332 }
333
Nick Piggine2867812008-07-25 19:45:30 -0700334 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700335 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700336 return -EAGAIN;
337 }
338
Christoph Lameterb20a3502006-03-22 00:09:12 -0800339 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800340 * In the async migration case of moving a page with buffers, lock the
341 * buffers using trylock before the mapping is moved. If the mapping
342 * was moved, we later failed to lock the buffers and could not move
343 * the mapping back due to an elevated page count, we would have to
344 * block waiting on other references to be dropped.
345 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800346 if (mode == MIGRATE_ASYNC && head &&
347 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800348 page_unfreeze_refs(page, expected_count);
349 spin_unlock_irq(&mapping->tree_lock);
350 return -EAGAIN;
351 }
352
353 /*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800354 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800355 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800356 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800357 if (PageSwapCache(page)) {
358 SetPageSwapCache(newpage);
359 set_page_private(newpage, page_private(page));
360 }
361
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800362 radix_tree_replace_slot(pslot, newpage);
363
364 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800365 * Drop cache reference from old page by unfreezing
366 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800367 * We know this isn't the last reference.
368 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800369 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800370
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700371 /*
372 * If moved to a different zone then also account
373 * the page for that zone. Other VM counters will be
374 * taken care of when we establish references to the
375 * new page and drop references to the old page.
376 *
377 * Note that anonymous pages are accounted for
378 * via NR_FILE_PAGES and NR_ANON_PAGES if they
379 * are mapped to swap space.
380 */
381 __dec_zone_page_state(page, NR_FILE_PAGES);
382 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700383 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700384 __dec_zone_page_state(page, NR_SHMEM);
385 __inc_zone_page_state(newpage, NR_SHMEM);
386 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700387 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800388
Rafael Aquini78bd5202012-12-11 16:02:31 -0800389 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800390}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800391
392/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900393 * The expected number of remaining references is the same as that
394 * of migrate_page_move_mapping().
395 */
396int migrate_huge_page_move_mapping(struct address_space *mapping,
397 struct page *newpage, struct page *page)
398{
399 int expected_count;
400 void **pslot;
401
402 if (!mapping) {
403 if (page_count(page) != 1)
404 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800405 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900406 }
407
408 spin_lock_irq(&mapping->tree_lock);
409
410 pslot = radix_tree_lookup_slot(&mapping->page_tree,
411 page_index(page));
412
413 expected_count = 2 + page_has_private(page);
414 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800415 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900416 spin_unlock_irq(&mapping->tree_lock);
417 return -EAGAIN;
418 }
419
420 if (!page_freeze_refs(page, expected_count)) {
421 spin_unlock_irq(&mapping->tree_lock);
422 return -EAGAIN;
423 }
424
425 get_page(newpage);
426
427 radix_tree_replace_slot(pslot, newpage);
428
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800429 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900430
431 spin_unlock_irq(&mapping->tree_lock);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800432 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900433}
434
435/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800436 * Gigantic pages are so large that we do not guarantee that page++ pointer
437 * arithmetic will work across the entire page. We need something more
438 * specialized.
439 */
440static void __copy_gigantic_page(struct page *dst, struct page *src,
441 int nr_pages)
442{
443 int i;
444 struct page *dst_base = dst;
445 struct page *src_base = src;
446
447 for (i = 0; i < nr_pages; ) {
448 cond_resched();
449 copy_highpage(dst, src);
450
451 i++;
452 dst = mem_map_next(dst, dst_base, i);
453 src = mem_map_next(src, src_base, i);
454 }
455}
456
457static void copy_huge_page(struct page *dst, struct page *src)
458{
459 int i;
460 int nr_pages;
461
462 if (PageHuge(src)) {
463 /* hugetlbfs page */
464 struct hstate *h = page_hstate(src);
465 nr_pages = pages_per_huge_page(h);
466
467 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
468 __copy_gigantic_page(dst, src, nr_pages);
469 return;
470 }
471 } else {
472 /* thp page */
473 BUG_ON(!PageTransHuge(src));
474 nr_pages = hpage_nr_pages(src);
475 }
476
477 for (i = 0; i < nr_pages; i++) {
478 cond_resched();
479 copy_highpage(dst + i, src + i);
480 }
481}
482
483/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800484 * Copy the page to its new location
485 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900486void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800487{
Rik van Riel7851a452013-10-07 11:29:23 +0100488 int cpupid;
489
Mel Gormanb32967f2012-11-19 12:35:47 +0000490 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900491 copy_huge_page(newpage, page);
492 else
493 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800494
495 if (PageError(page))
496 SetPageError(newpage);
497 if (PageReferenced(page))
498 SetPageReferenced(newpage);
499 if (PageUptodate(page))
500 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700501 if (TestClearPageActive(page)) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800502 VM_BUG_ON_PAGE(PageUnevictable(page), page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800503 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800504 } else if (TestClearPageUnevictable(page))
505 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800506 if (PageChecked(page))
507 SetPageChecked(newpage);
508 if (PageMappedToDisk(page))
509 SetPageMappedToDisk(newpage);
510
511 if (PageDirty(page)) {
512 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700513 /*
514 * Want to mark the page and the radix tree as dirty, and
515 * redo the accounting that clear_page_dirty_for_io undid,
516 * but we can't use set_page_dirty because that function
517 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300518 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700519 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700520 if (PageSwapBacked(page))
521 SetPageDirty(newpage);
522 else
523 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800524 }
525
Rik van Riel7851a452013-10-07 11:29:23 +0100526 /*
527 * Copy NUMA information to the new page, to prevent over-eager
528 * future migrations of this same page.
529 */
530 cpupid = page_cpupid_xchg_last(page, -1);
531 page_cpupid_xchg_last(newpage, cpupid);
532
Nick Pigginb291f002008-10-18 20:26:44 -0700533 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800534 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800535 /*
536 * Please do not reorder this without considering how mm/ksm.c's
537 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
538 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800539 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800540 ClearPagePrivate(page);
541 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800542
543 /*
544 * If any waiters have accumulated on the new page then
545 * wake them up.
546 */
547 if (PageWriteback(newpage))
548 end_page_writeback(newpage);
549}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800550
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700551/************************************************************
552 * Migration functions
553 ***********************************************************/
554
Christoph Lameterb20a3502006-03-22 00:09:12 -0800555/*
556 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100557 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800558 *
559 * Pages are locked upon entry and exit.
560 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700561int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800562 struct page *newpage, struct page *page,
563 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800564{
565 int rc;
566
567 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
568
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500569 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800570
Rafael Aquini78bd5202012-12-11 16:02:31 -0800571 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800572 return rc;
573
574 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800575 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800576}
577EXPORT_SYMBOL(migrate_page);
578
David Howells93614012006-09-30 20:45:40 +0200579#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800580/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700581 * Migration function for pages with buffers. This function can only be used
582 * if the underlying filesystem guarantees that no other references to "page"
583 * exist.
584 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700585int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800586 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700587{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700588 struct buffer_head *bh, *head;
589 int rc;
590
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700591 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800592 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700593
594 head = page_buffers(page);
595
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500596 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700597
Rafael Aquini78bd5202012-12-11 16:02:31 -0800598 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700599 return rc;
600
Mel Gormanb969c4a2012-01-12 17:19:34 -0800601 /*
602 * In the async case, migrate_page_move_mapping locked the buffers
603 * with an IRQ-safe spinlock held. In the sync case, the buffers
604 * need to be locked now
605 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800606 if (mode != MIGRATE_ASYNC)
607 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700608
609 ClearPagePrivate(page);
610 set_page_private(newpage, page_private(page));
611 set_page_private(page, 0);
612 put_page(page);
613 get_page(newpage);
614
615 bh = head;
616 do {
617 set_bh_page(bh, newpage, bh_offset(bh));
618 bh = bh->b_this_page;
619
620 } while (bh != head);
621
622 SetPagePrivate(newpage);
623
624 migrate_page_copy(newpage, page);
625
626 bh = head;
627 do {
628 unlock_buffer(bh);
629 put_bh(bh);
630 bh = bh->b_this_page;
631
632 } while (bh != head);
633
Rafael Aquini78bd5202012-12-11 16:02:31 -0800634 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700635}
636EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200637#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700638
Christoph Lameter04e62a22006-06-23 02:03:38 -0700639/*
640 * Writeback a page to clean the dirty state
641 */
642static int writeout(struct address_space *mapping, struct page *page)
643{
644 struct writeback_control wbc = {
645 .sync_mode = WB_SYNC_NONE,
646 .nr_to_write = 1,
647 .range_start = 0,
648 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700649 .for_reclaim = 1
650 };
651 int rc;
652
653 if (!mapping->a_ops->writepage)
654 /* No write method for the address space */
655 return -EINVAL;
656
657 if (!clear_page_dirty_for_io(page))
658 /* Someone else already triggered a write */
659 return -EAGAIN;
660
661 /*
662 * A dirty page may imply that the underlying filesystem has
663 * the page on some queue. So the page must be clean for
664 * migration. Writeout may mean we loose the lock and the
665 * page state is no longer what we checked for earlier.
666 * At this point we know that the migration attempt cannot
667 * be successful.
668 */
669 remove_migration_ptes(page, page);
670
671 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700672
673 if (rc != AOP_WRITEPAGE_ACTIVATE)
674 /* unlocked. Relock */
675 lock_page(page);
676
Hugh Dickinsbda85502008-11-19 15:36:36 -0800677 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700678}
679
680/*
681 * Default handling if a filesystem does not provide a migration function.
682 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700683static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800684 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700685{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800686 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800687 /* Only writeback pages in full synchronous migration */
688 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800689 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700690 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800691 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700692
693 /*
694 * Buffers may be managed in a filesystem specific way.
695 * We must have no buffers or drop them.
696 */
David Howells266cf652009-04-03 16:42:36 +0100697 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700698 !try_to_release_page(page, GFP_KERNEL))
699 return -EAGAIN;
700
Mel Gormana6bc32b2012-01-12 17:19:43 -0800701 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700702}
703
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700704/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700705 * Move a page to a newly allocated page
706 * The page is locked and all ptes have been successfully removed.
707 *
708 * The new page will have replaced the old page if this function
709 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700710 *
711 * Return value:
712 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800713 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700714 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700715static int move_to_new_page(struct page *newpage, struct page *page,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800716 int remap_swapcache, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700717{
718 struct address_space *mapping;
719 int rc;
720
721 /*
722 * Block others from accessing the page when we get around to
723 * establishing additional references. We are the only one
724 * holding a reference to the new page at this point.
725 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200726 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700727 BUG();
728
729 /* Prepare mapping for the new page.*/
730 newpage->index = page->index;
731 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700732 if (PageSwapBacked(page))
733 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700734
735 mapping = page_mapping(page);
736 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800737 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800738 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700739 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800740 * Most pages have a mapping and most filesystems provide a
741 * migratepage callback. Anonymous pages are part of swap
742 * space which also has its own migratepage callback. This
743 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700744 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800745 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800746 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800747 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800748 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700749
Rafael Aquini78bd5202012-12-11 16:02:31 -0800750 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700751 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700752 } else {
753 if (remap_swapcache)
754 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800755 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700756 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700757
758 unlock_page(newpage);
759
760 return rc;
761}
762
Minchan Kim0dabec92011-10-31 17:06:57 -0700763static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800764 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700765{
Minchan Kim0dabec92011-10-31 17:06:57 -0700766 int rc = -EAGAIN;
Mel Gorman3fe20112010-05-24 14:32:20 -0700767 int remap_swapcache = 1;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -0700768 struct mem_cgroup *mem;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700769 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700770
Nick Piggin529ae9a2008-08-02 12:01:03 +0200771 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800772 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700773 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800774
775 /*
776 * It's not safe for direct compaction to call lock_page.
777 * For example, during page readahead pages are added locked
778 * to the LRU. Later, when the IO completes the pages are
779 * marked uptodate and unlocked. However, the queueing
780 * could be merging multiple pages for one bio (e.g.
781 * mpage_readpages). If an allocation happens for the
782 * second or third page, the process can end up locking
783 * the same page twice and deadlocking. Rather than
784 * trying to be clever about what pages can be locked,
785 * avoid the use of lock_page for direct compaction
786 * altogether.
787 */
788 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700789 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800790
Christoph Lametere24f0b82006-06-23 02:03:51 -0700791 lock_page(page);
792 }
793
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800794 /* charge against new page */
Johannes Weiner0030f532012-07-31 16:45:25 -0700795 mem_cgroup_prepare_migration(page, newpage, &mem);
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800796
Christoph Lametere24f0b82006-06-23 02:03:51 -0700797 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700798 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700799 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800800 * necessary to wait for PageWriteback. In the async case,
801 * the retry loop is too short and in the sync-light case,
802 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700803 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800804 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700805 rc = -EBUSY;
806 goto uncharge;
807 }
808 if (!force)
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800809 goto uncharge;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700810 wait_on_page_writeback(page);
811 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700812 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700813 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
814 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800815 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700816 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700817 * File Caches may use write_page() or lock_page() in migration, then,
818 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700819 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800820 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800821 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000822 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800823 * getting a hold on an anon_vma from outside one of its mms.
824 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700825 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800826 if (anon_vma) {
827 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700828 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800829 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800830 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700831 /*
832 * We cannot be sure that the anon_vma of an unmapped
833 * swapcache page is safe to use because we don't
834 * know in advance if the VMA that this page belonged
835 * to still exists. If the VMA and others sharing the
836 * data have been freed, then the anon_vma could
837 * already be invalid.
838 *
839 * To avoid this possibility, swapcache pages get
840 * migrated but are not remapped when migration
841 * completes
842 */
843 remap_swapcache = 0;
844 } else {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800845 goto uncharge;
Mel Gorman3fe20112010-05-24 14:32:20 -0700846 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700847 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800848
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800849 if (unlikely(balloon_page_movable(page))) {
850 /*
851 * A ballooned page does not need any special attention from
852 * physical to virtual reverse mapping procedures.
853 * Skip any attempt to unmap PTEs or to remap swap cache,
854 * in order to avoid burning cycles at rmap level, and perform
855 * the page migration right away (proteced by page lock).
856 */
857 rc = balloon_page_migrate(newpage, page, mode);
858 goto uncharge;
859 }
860
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700861 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800862 * Corner case handling:
863 * 1. When a new swap-cache page is read into, it is added to the LRU
864 * and treated as swapcache but it has no rmap yet.
865 * Calling try_to_unmap() against a page->mapping==NULL page will
866 * trigger a BUG. So handle it here.
867 * 2. An orphaned page (see truncate_complete_page) might have
868 * fs-private metadata. The page can be picked up due to memory
869 * offlining. Everywhere else except page reclaim, the page is
870 * invisible to the vm, so the page can not be migrated. So try to
871 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700872 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800873 if (!page->mapping) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800874 VM_BUG_ON_PAGE(PageAnon(page), page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800875 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800876 try_to_free_buffers(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800877 goto uncharge;
Shaohua Li62e1c552008-02-04 22:29:33 -0800878 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700879 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800880 }
881
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700882 /* Establish migration ptes or remove ptes */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200883 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700884
Shaohua Liabfc3482009-09-21 17:01:19 -0700885skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700886 if (!page_mapped(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800887 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700888
Mel Gorman3fe20112010-05-24 14:32:20 -0700889 if (rc && remap_swapcache)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700890 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700891
892 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700893 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700894 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700895
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800896uncharge:
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800897 mem_cgroup_end_migration(mem, page, newpage,
898 (rc == MIGRATEPAGE_SUCCESS ||
899 rc == MIGRATEPAGE_BALLOON_SUCCESS));
Christoph Lametere24f0b82006-06-23 02:03:51 -0700900 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700901out:
902 return rc;
903}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700904
Minchan Kim0dabec92011-10-31 17:06:57 -0700905/*
906 * Obtain the lock on page, remove all ptes and migrate the page
907 * to the newly allocated page in newpage.
908 */
909static int unmap_and_move(new_page_t get_new_page, unsigned long private,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800910 struct page *page, int force, enum migrate_mode mode)
Minchan Kim0dabec92011-10-31 17:06:57 -0700911{
912 int rc = 0;
913 int *result = NULL;
914 struct page *newpage = get_new_page(page, private, &result);
915
916 if (!newpage)
917 return -ENOMEM;
918
919 if (page_count(page) == 1) {
920 /* page was freed from under us. So we are done. */
921 goto out;
922 }
923
924 if (unlikely(PageTransHuge(page)))
925 if (unlikely(split_huge_page(page)))
926 goto out;
927
Hugh Dickins9c620e22013-02-22 16:35:14 -0800928 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800929
930 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
931 /*
932 * A ballooned page has been migrated already.
933 * Now, it's the time to wrap-up counters,
934 * handle the page back to Buddy and return.
935 */
936 dec_zone_page_state(page, NR_ISOLATED_ANON +
937 page_is_file_cache(page));
938 balloon_page_free(page);
939 return MIGRATEPAGE_SUCCESS;
940 }
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 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700954 /*
955 * Move the new page to the LRU. If migration was not successful
956 * then this will free the page.
957 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700958 putback_lru_page(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700959 if (result) {
960 if (rc)
961 *result = rc;
962 else
963 *result = page_to_nid(newpage);
964 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700965 return rc;
966}
967
968/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900969 * Counterpart of unmap_and_move_page() for hugepage migration.
970 *
971 * This function doesn't wait the completion of hugepage I/O
972 * because there is no race between I/O and migration for hugepage.
973 * Note that currently hugepage I/O occurs only in direct I/O
974 * where no lock is held and PG_writeback is irrelevant,
975 * and writeback status of all subpages are counted in the reference
976 * count of the head page (i.e. if all subpages of a 2MB hugepage are
977 * under direct I/O, the reference of the head page is 512 and a bit more.)
978 * This means that when we try to migrate hugepage whose subpages are
979 * doing direct I/O, some references remain after try_to_unmap() and
980 * hugepage migration fails without data corruption.
981 *
982 * There is also no race when direct I/O is issued on the page under migration,
983 * because then pte is replaced with migration swap entry and direct I/O code
984 * will wait in the page fault for migration to complete.
985 */
986static int unmap_and_move_huge_page(new_page_t get_new_page,
987 unsigned long private, struct page *hpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800988 int force, enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900989{
990 int rc = 0;
991 int *result = NULL;
Joonsoo Kim32665f22014-01-21 15:51:15 -0800992 struct page *new_hpage;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900993 struct anon_vma *anon_vma = NULL;
994
Naoya Horiguchi83467ef2013-09-11 14:22:11 -0700995 /*
996 * Movability of hugepages depends on architectures and hugepage size.
997 * This check is necessary because some callers of hugepage migration
998 * like soft offline and memory hotremove don't walk through page
999 * tables or check whether the hugepage is pmd-based or not before
1000 * kicking migration.
1001 */
Joonsoo Kim32665f22014-01-21 15:51:15 -08001002 if (!hugepage_migration_support(page_hstate(hpage))) {
1003 putback_active_hugepage(hpage);
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001004 return -ENOSYS;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001005 }
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001006
Joonsoo Kim32665f22014-01-21 15:51:15 -08001007 new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001008 if (!new_hpage)
1009 return -ENOMEM;
1010
1011 rc = -EAGAIN;
1012
1013 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001014 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001015 goto out;
1016 lock_page(hpage);
1017 }
1018
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001019 if (PageAnon(hpage))
1020 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001021
1022 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1023
1024 if (!page_mapped(hpage))
Mel Gormana6bc32b2012-01-12 17:19:43 -08001025 rc = move_to_new_page(new_hpage, hpage, 1, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001026
1027 if (rc)
1028 remove_migration_ptes(hpage, hpage);
1029
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001030 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001031 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001032
1033 if (!rc)
1034 hugetlb_cgroup_migrate(hpage, new_hpage);
1035
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001036 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001037out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001038 if (rc != -EAGAIN)
1039 putback_active_hugepage(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001040 put_page(new_hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001041 if (result) {
1042 if (rc)
1043 *result = rc;
1044 else
1045 *result = page_to_nid(new_hpage);
1046 }
1047 return rc;
1048}
1049
1050/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001051 * migrate_pages - migrate the pages specified in a list, to the free pages
1052 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001053 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001054 * @from: The list of pages to be migrated.
1055 * @get_new_page: The function used to allocate free pages to be used
1056 * as the target of the page migration.
1057 * @private: Private data to be passed on to get_new_page()
1058 * @mode: The migration mode that specifies the constraints for
1059 * page migration, if any.
1060 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001061 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001062 * The function returns after 10 attempts or if no pages are movable any more
1063 * because the list has become empty or no retryable pages exist any more.
1064 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001065 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001066 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001067 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001068 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001069int migrate_pages(struct list_head *from, new_page_t get_new_page,
1070 unsigned long private, enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001071{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001072 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001073 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001074 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001075 int pass = 0;
1076 struct page *page;
1077 struct page *page2;
1078 int swapwrite = current->flags & PF_SWAPWRITE;
1079 int rc;
1080
1081 if (!swapwrite)
1082 current->flags |= PF_SWAPWRITE;
1083
Christoph Lametere24f0b82006-06-23 02:03:51 -07001084 for(pass = 0; pass < 10 && retry; pass++) {
1085 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001086
Christoph Lametere24f0b82006-06-23 02:03:51 -07001087 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001088 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001089
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001090 if (PageHuge(page))
1091 rc = unmap_and_move_huge_page(get_new_page,
1092 private, page, pass > 2, mode);
1093 else
1094 rc = unmap_and_move(get_new_page, private,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001095 page, pass > 2, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001096
Christoph Lametere24f0b82006-06-23 02:03:51 -07001097 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001098 case -ENOMEM:
1099 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001100 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001101 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001102 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001103 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001104 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001105 break;
1106 default:
Naoya Horiguchi354a3362014-01-21 15:51:14 -08001107 /*
1108 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1109 * unlike -EAGAIN case, the failed page is
1110 * removed from migration page list and not
1111 * retried in the next outer loop.
1112 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001113 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001114 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001115 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001116 }
1117 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001118 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001119out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001120 if (nr_succeeded)
1121 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1122 if (nr_failed)
1123 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001124 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1125
Christoph Lameterb20a3502006-03-22 00:09:12 -08001126 if (!swapwrite)
1127 current->flags &= ~PF_SWAPWRITE;
1128
Rafael Aquini78bd5202012-12-11 16:02:31 -08001129 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001130}
1131
Christoph Lameter742755a2006-06-23 02:03:55 -07001132#ifdef CONFIG_NUMA
1133/*
1134 * Move a list of individual pages
1135 */
1136struct page_to_node {
1137 unsigned long addr;
1138 struct page *page;
1139 int node;
1140 int status;
1141};
1142
1143static struct page *new_page_node(struct page *p, unsigned long private,
1144 int **result)
1145{
1146 struct page_to_node *pm = (struct page_to_node *)private;
1147
1148 while (pm->node != MAX_NUMNODES && pm->page != p)
1149 pm++;
1150
1151 if (pm->node == MAX_NUMNODES)
1152 return NULL;
1153
1154 *result = &pm->status;
1155
Naoya Horiguchie632a932013-09-11 14:22:04 -07001156 if (PageHuge(p))
1157 return alloc_huge_page_node(page_hstate(compound_head(p)),
1158 pm->node);
1159 else
1160 return alloc_pages_exact_node(pm->node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001161 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001162}
1163
1164/*
1165 * Move a set of pages as indicated in the pm array. The addr
1166 * field must be set to the virtual address of the page to be moved
1167 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001168 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001169 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001170static int do_move_page_to_node_array(struct mm_struct *mm,
1171 struct page_to_node *pm,
1172 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001173{
1174 int err;
1175 struct page_to_node *pp;
1176 LIST_HEAD(pagelist);
1177
1178 down_read(&mm->mmap_sem);
1179
1180 /*
1181 * Build a list of pages to migrate
1182 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001183 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1184 struct vm_area_struct *vma;
1185 struct page *page;
1186
Christoph Lameter742755a2006-06-23 02:03:55 -07001187 err = -EFAULT;
1188 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001189 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001190 goto set_status;
1191
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001192 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001193
1194 err = PTR_ERR(page);
1195 if (IS_ERR(page))
1196 goto set_status;
1197
Christoph Lameter742755a2006-06-23 02:03:55 -07001198 err = -ENOENT;
1199 if (!page)
1200 goto set_status;
1201
Hugh Dickins62b61f62009-12-14 17:59:33 -08001202 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001203 if (PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001204 goto put_and_set;
1205
1206 pp->page = page;
1207 err = page_to_nid(page);
1208
1209 if (err == pp->node)
1210 /*
1211 * Node already in the right place
1212 */
1213 goto put_and_set;
1214
1215 err = -EACCES;
1216 if (page_mapcount(page) > 1 &&
1217 !migrate_all)
1218 goto put_and_set;
1219
Naoya Horiguchie632a932013-09-11 14:22:04 -07001220 if (PageHuge(page)) {
1221 isolate_huge_page(page, &pagelist);
1222 goto put_and_set;
1223 }
1224
Nick Piggin62695a82008-10-18 20:26:09 -07001225 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001226 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001227 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001228 inc_zone_page_state(page, NR_ISOLATED_ANON +
1229 page_is_file_cache(page));
1230 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001231put_and_set:
1232 /*
1233 * Either remove the duplicate refcount from
1234 * isolate_lru_page() or drop the page ref if it was
1235 * not isolated.
1236 */
1237 put_page(page);
1238set_status:
1239 pp->status = err;
1240 }
1241
Brice Gogline78bbfa2008-10-18 20:27:15 -07001242 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001243 if (!list_empty(&pagelist)) {
Christoph Lameter742755a2006-06-23 02:03:55 -07001244 err = migrate_pages(&pagelist, new_page_node,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001245 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001246 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001247 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001248 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001249
1250 up_read(&mm->mmap_sem);
1251 return err;
1252}
1253
1254/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001255 * Migrate an array of page address onto an array of nodes and fill
1256 * the corresponding array of status.
1257 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001258static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001259 unsigned long nr_pages,
1260 const void __user * __user *pages,
1261 const int __user *nodes,
1262 int __user *status, int flags)
1263{
Brice Goglin3140a222009-01-06 14:38:57 -08001264 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001265 unsigned long chunk_nr_pages;
1266 unsigned long chunk_start;
1267 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001268
Brice Goglin3140a222009-01-06 14:38:57 -08001269 err = -ENOMEM;
1270 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1271 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001272 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001273
1274 migrate_prep();
1275
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001276 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001277 * Store a chunk of page_to_node array in a page,
1278 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001279 */
Brice Goglin3140a222009-01-06 14:38:57 -08001280 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001281
Brice Goglin3140a222009-01-06 14:38:57 -08001282 for (chunk_start = 0;
1283 chunk_start < nr_pages;
1284 chunk_start += chunk_nr_pages) {
1285 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001286
Brice Goglin3140a222009-01-06 14:38:57 -08001287 if (chunk_start + chunk_nr_pages > nr_pages)
1288 chunk_nr_pages = nr_pages - chunk_start;
1289
1290 /* fill the chunk pm with addrs and nodes from user-space */
1291 for (j = 0; j < chunk_nr_pages; j++) {
1292 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001293 int node;
1294
Brice Goglin3140a222009-01-06 14:38:57 -08001295 err = -EFAULT;
1296 if (get_user(p, pages + j + chunk_start))
1297 goto out_pm;
1298 pm[j].addr = (unsigned long) p;
1299
1300 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001301 goto out_pm;
1302
1303 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001304 if (node < 0 || node >= MAX_NUMNODES)
1305 goto out_pm;
1306
Lai Jiangshan389162c2012-12-12 13:51:30 -08001307 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001308 goto out_pm;
1309
1310 err = -EACCES;
1311 if (!node_isset(node, task_nodes))
1312 goto out_pm;
1313
Brice Goglin3140a222009-01-06 14:38:57 -08001314 pm[j].node = node;
1315 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001316
Brice Goglin3140a222009-01-06 14:38:57 -08001317 /* End marker for this chunk */
1318 pm[chunk_nr_pages].node = MAX_NUMNODES;
1319
1320 /* Migrate this chunk */
1321 err = do_move_page_to_node_array(mm, pm,
1322 flags & MPOL_MF_MOVE_ALL);
1323 if (err < 0)
1324 goto out_pm;
1325
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001326 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001327 for (j = 0; j < chunk_nr_pages; j++)
1328 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001329 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001330 goto out_pm;
1331 }
1332 }
1333 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001334
1335out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001336 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001337out:
1338 return err;
1339}
1340
1341/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001342 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001343 */
Brice Goglin80bba122008-12-09 13:14:23 -08001344static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1345 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001346{
Brice Goglin2f007e72008-10-18 20:27:16 -07001347 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001348
Christoph Lameter742755a2006-06-23 02:03:55 -07001349 down_read(&mm->mmap_sem);
1350
Brice Goglin2f007e72008-10-18 20:27:16 -07001351 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001352 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001353 struct vm_area_struct *vma;
1354 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001355 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001356
1357 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001358 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001359 goto set_status;
1360
Brice Goglin2f007e72008-10-18 20:27:16 -07001361 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001362
1363 err = PTR_ERR(page);
1364 if (IS_ERR(page))
1365 goto set_status;
1366
Christoph Lameter742755a2006-06-23 02:03:55 -07001367 err = -ENOENT;
1368 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001369 if (!page || PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001370 goto set_status;
1371
1372 err = page_to_nid(page);
1373set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001374 *status = err;
1375
1376 pages++;
1377 status++;
1378 }
1379
1380 up_read(&mm->mmap_sem);
1381}
1382
1383/*
1384 * Determine the nodes of a user array of pages and store it in
1385 * a user array of status.
1386 */
1387static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1388 const void __user * __user *pages,
1389 int __user *status)
1390{
1391#define DO_PAGES_STAT_CHUNK_NR 16
1392 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1393 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001394
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001395 while (nr_pages) {
1396 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001397
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001398 chunk_nr = nr_pages;
1399 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1400 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1401
1402 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1403 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001404
1405 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1406
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001407 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1408 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001409
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001410 pages += chunk_nr;
1411 status += chunk_nr;
1412 nr_pages -= chunk_nr;
1413 }
1414 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001415}
1416
1417/*
1418 * Move a list of pages in the address space of the currently executing
1419 * process.
1420 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001421SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1422 const void __user * __user *, pages,
1423 const int __user *, nodes,
1424 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001425{
David Howellsc69e8d92008-11-14 10:39:19 +11001426 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001427 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001428 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001429 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001430 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001431
1432 /* Check flags */
1433 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1434 return -EINVAL;
1435
1436 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1437 return -EPERM;
1438
1439 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001440 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001441 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001442 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001443 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001444 return -ESRCH;
1445 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001446 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001447
1448 /*
1449 * Check if this process has the right to modify the specified
1450 * process. The right exists if the process has administrative
1451 * capabilities, superuser privileges or the same
1452 * userid as the target process.
1453 */
David Howellsc69e8d92008-11-14 10:39:19 +11001454 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001455 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1456 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001457 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001458 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001459 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001460 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001461 }
David Howellsc69e8d92008-11-14 10:39:19 +11001462 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001463
David Quigley86c3a762006-06-23 02:04:02 -07001464 err = security_task_movememory(task);
1465 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001466 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001467
Christoph Lameter3268c632012-03-21 16:34:06 -07001468 task_nodes = cpuset_mems_allowed(task);
1469 mm = get_task_mm(task);
1470 put_task_struct(task);
1471
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001472 if (!mm)
1473 return -EINVAL;
1474
1475 if (nodes)
1476 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1477 nodes, status, flags);
1478 else
1479 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001480
1481 mmput(mm);
1482 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001483
Christoph Lameter742755a2006-06-23 02:03:55 -07001484out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001485 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001486 return err;
1487}
Christoph Lameter742755a2006-06-23 02:03:55 -07001488
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001489/*
1490 * Call migration functions in the vma_ops that may prepare
1491 * memory in a vm for migration. migration functions may perform
1492 * the migration for vmas that do not have an underlying page struct.
1493 */
1494int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1495 const nodemask_t *from, unsigned long flags)
1496{
1497 struct vm_area_struct *vma;
1498 int err = 0;
1499
Daisuke Nishimura1001c9f2009-02-11 13:04:18 -08001500 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001501 if (vma->vm_ops && vma->vm_ops->migrate) {
1502 err = vma->vm_ops->migrate(vma, to, from, flags);
1503 if (err)
1504 break;
1505 }
1506 }
1507 return err;
1508}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001509
1510#ifdef CONFIG_NUMA_BALANCING
1511/*
1512 * Returns true if this is a safe migration target node for misplaced NUMA
1513 * pages. Currently it only checks the watermarks which crude
1514 */
1515static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001516 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001517{
1518 int z;
1519 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1520 struct zone *zone = pgdat->node_zones + z;
1521
1522 if (!populated_zone(zone))
1523 continue;
1524
Lisa Du6e543d52013-09-11 14:22:36 -07001525 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001526 continue;
1527
1528 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1529 if (!zone_watermark_ok(zone, 0,
1530 high_wmark_pages(zone) +
1531 nr_migrate_pages,
1532 0, 0))
1533 continue;
1534 return true;
1535 }
1536 return false;
1537}
1538
1539static struct page *alloc_misplaced_dst_page(struct page *page,
1540 unsigned long data,
1541 int **result)
1542{
1543 int nid = (int) data;
1544 struct page *newpage;
1545
1546 newpage = alloc_pages_exact_node(nid,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001547 (GFP_HIGHUSER_MOVABLE |
1548 __GFP_THISNODE | __GFP_NOMEMALLOC |
1549 __GFP_NORETRY | __GFP_NOWARN) &
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001550 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001551
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001552 return newpage;
1553}
1554
1555/*
Mel Gormana8f60772012-11-14 21:41:46 +00001556 * page migration rate limiting control.
1557 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1558 * window of time. Default here says do not migrate more than 1280M per second.
Mel Gormane14808b2012-11-19 10:59:15 +00001559 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1560 * as it is faults that reset the window, pte updates will happen unconditionally
1561 * if there has not been a fault since @pteupdate_interval_millisecs after the
1562 * throttle window closed.
Mel Gormana8f60772012-11-14 21:41:46 +00001563 */
1564static unsigned int migrate_interval_millisecs __read_mostly = 100;
Mel Gormane14808b2012-11-19 10:59:15 +00001565static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
Mel Gormana8f60772012-11-14 21:41:46 +00001566static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1567
Mel Gormane14808b2012-11-19 10:59:15 +00001568/* Returns true if NUMA migration is currently rate limited */
1569bool migrate_ratelimited(int node)
1570{
1571 pg_data_t *pgdat = NODE_DATA(node);
1572
1573 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1574 msecs_to_jiffies(pteupdate_interval_millisecs)))
1575 return false;
1576
1577 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1578 return false;
1579
1580 return true;
1581}
1582
Mel Gormanb32967f2012-11-19 12:35:47 +00001583/* Returns true if the node is migrate rate-limited after the update */
Mel Gorman1c30e012014-01-21 15:50:58 -08001584static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1585 unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001586{
Mel Gormanb32967f2012-11-19 12:35:47 +00001587 /*
1588 * Rate-limit the amount of data that is being migrated to a node.
1589 * Optimal placement is no good if the memory bus is saturated and
1590 * all the time is being spent migrating!
1591 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001592 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001593 spin_lock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001594 pgdat->numabalancing_migrate_nr_pages = 0;
1595 pgdat->numabalancing_migrate_next_window = jiffies +
1596 msecs_to_jiffies(migrate_interval_millisecs);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001597 spin_unlock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001598 }
Mel Gormanaf1839d2014-01-21 15:51:01 -08001599 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1600 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1601 nr_pages);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001602 return true;
Mel Gormanaf1839d2014-01-21 15:51:01 -08001603 }
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001604
1605 /*
1606 * This is an unlocked non-atomic update so errors are possible.
1607 * The consequences are failing to migrate when we potentiall should
1608 * have which is not severe enough to warrant locking. If it is ever
1609 * a problem, it can be converted to a per-cpu counter.
1610 */
1611 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1612 return false;
Mel Gormanb32967f2012-11-19 12:35:47 +00001613}
1614
Mel Gorman1c30e012014-01-21 15:50:58 -08001615static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
Mel Gormanb32967f2012-11-19 12:35:47 +00001616{
Hugh Dickins340ef392013-02-22 16:34:33 -08001617 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001618
Sasha Levin309381fea2014-01-23 15:52:54 -08001619 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
Mel Gorman3abef4e2013-02-22 16:34:27 -08001620
Mel Gormanb32967f2012-11-19 12:35:47 +00001621 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001622 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1623 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001624
Hugh Dickins340ef392013-02-22 16:34:33 -08001625 if (isolate_lru_page(page))
1626 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001627
1628 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001629 * migrate_misplaced_transhuge_page() skips page migration's usual
1630 * check on page_count(), so we must do it here, now that the page
1631 * has been isolated: a GUP pin, or any other pin, prevents migration.
1632 * The expected page count is 3: 1 for page's mapcount and 1 for the
1633 * caller's pin and 1 for the reference taken by isolate_lru_page().
1634 */
1635 if (PageTransHuge(page) && page_count(page) != 3) {
1636 putback_lru_page(page);
1637 return 0;
1638 }
1639
1640 page_lru = page_is_file_cache(page);
1641 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1642 hpage_nr_pages(page));
1643
1644 /*
1645 * Isolating the page has taken another reference, so the
1646 * caller's reference can be safely dropped without the page
1647 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001648 */
1649 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001650 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001651}
1652
Mel Gormande466bd2013-12-18 17:08:42 -08001653bool pmd_trans_migrating(pmd_t pmd)
1654{
1655 struct page *page = pmd_page(pmd);
1656 return PageLocked(page);
1657}
1658
1659void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd)
1660{
1661 struct page *page = pmd_page(*pmd);
1662 wait_on_page_locked(page);
1663}
1664
Mel Gormana8f60772012-11-14 21:41:46 +00001665/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001666 * Attempt to migrate a misplaced page to the specified destination
1667 * node. Caller is expected to have an elevated reference count on
1668 * the page that will be dropped by this function before returning.
1669 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001670int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1671 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001672{
Mel Gormana8f60772012-11-14 21:41:46 +00001673 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001674 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001675 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001676 LIST_HEAD(migratepages);
1677
1678 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001679 * Don't migrate file pages that are mapped in multiple processes
1680 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001681 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001682 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1683 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001684 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001685
Mel Gormana8f60772012-11-14 21:41:46 +00001686 /*
1687 * Rate-limit the amount of data that is being migrated to a node.
1688 * Optimal placement is no good if the memory bus is saturated and
1689 * all the time is being spent migrating!
1690 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001691 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001692 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001693
Mel Gormanb32967f2012-11-19 12:35:47 +00001694 isolated = numamigrate_isolate_page(pgdat, page);
1695 if (!isolated)
1696 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001697
Mel Gormanb32967f2012-11-19 12:35:47 +00001698 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001699 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1700 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001701 if (nr_remaining) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001702 if (!list_empty(&migratepages)) {
1703 list_del(&page->lru);
1704 dec_zone_page_state(page, NR_ISOLATED_ANON +
1705 page_is_file_cache(page));
1706 putback_lru_page(page);
1707 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001708 isolated = 0;
1709 } else
1710 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001711 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001712 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001713
1714out:
1715 put_page(page);
1716 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001717}
Mel Gorman220018d2012-12-05 09:32:56 +00001718#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001719
Mel Gorman220018d2012-12-05 09:32:56 +00001720#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001721/*
1722 * Migrates a THP to a given target node. page must be locked and is unlocked
1723 * before returning.
1724 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001725int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1726 struct vm_area_struct *vma,
1727 pmd_t *pmd, pmd_t entry,
1728 unsigned long address,
1729 struct page *page, int node)
1730{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001731 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001732 pg_data_t *pgdat = NODE_DATA(node);
1733 int isolated = 0;
1734 struct page *new_page = NULL;
1735 struct mem_cgroup *memcg = NULL;
1736 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001737 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1738 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001739 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001740
1741 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001742 * Rate-limit the amount of data that is being migrated to a node.
1743 * Optimal placement is no good if the memory bus is saturated and
1744 * all the time is being spent migrating!
1745 */
Mel Gormand28d43352012-11-29 09:24:36 +00001746 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001747 goto out_dropref;
1748
1749 new_page = alloc_pages_node(node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001750 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1751 HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001752 if (!new_page)
1753 goto out_fail;
1754
Mel Gormanb32967f2012-11-19 12:35:47 +00001755 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001756 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001757 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001758 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001759 }
1760
Mel Gormanb0943d62013-12-18 17:08:46 -08001761 if (mm_tlb_flush_pending(mm))
1762 flush_tlb_range(vma, mmun_start, mmun_end);
1763
Mel Gormanb32967f2012-11-19 12:35:47 +00001764 /* Prepare a page as a migration target */
1765 __set_page_locked(new_page);
1766 SetPageSwapBacked(new_page);
1767
1768 /* anon mapping, we can simply copy page->mapping to the new page: */
1769 new_page->mapping = page->mapping;
1770 new_page->index = page->index;
1771 migrate_page_copy(new_page, page);
1772 WARN_ON(PageLRU(new_page));
1773
1774 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001775 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001776 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001777 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1778fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001779 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001780 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001781
1782 /* Reverse changes made by migrate_page_copy() */
1783 if (TestClearPageActive(new_page))
1784 SetPageActive(page);
1785 if (TestClearPageUnevictable(new_page))
1786 SetPageUnevictable(page);
1787 mlock_migrate_page(page, new_page);
1788
1789 unlock_page(new_page);
1790 put_page(new_page); /* Free it */
1791
Mel Gormana54a4072013-10-07 11:28:46 +01001792 /* Retake the callers reference and putback on LRU */
1793 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001794 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001795 mod_zone_page_state(page_zone(page),
1796 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001797
1798 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001799 }
1800
1801 /*
1802 * Traditional migration needs to prepare the memcg charge
1803 * transaction early to prevent the old page from being
1804 * uncharged when installing migration entries. Here we can
1805 * save the potential rollback and start the charge transfer
1806 * only when migration is already known to end successfully.
1807 */
1808 mem_cgroup_prepare_migration(page, new_page, &memcg);
1809
Mel Gorman2b4847e2013-12-18 17:08:32 -08001810 orig_entry = *pmd;
Mel Gormanb32967f2012-11-19 12:35:47 +00001811 entry = mk_pmd(new_page, vma->vm_page_prot);
Mel Gormanb32967f2012-11-19 12:35:47 +00001812 entry = pmd_mkhuge(entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001813 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001814
Mel Gorman2b4847e2013-12-18 17:08:32 -08001815 /*
1816 * Clear the old entry under pagetable lock and establish the new PTE.
1817 * Any parallel GUP will either observe the old page blocking on the
1818 * page lock, block on the page table lock or observe the new page.
1819 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1820 * guarantee the copy is visible before the pagetable update.
1821 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001822 flush_cache_range(vma, mmun_start, mmun_end);
1823 page_add_new_anon_rmap(new_page, vma, mmun_start);
1824 pmdp_clear_flush(vma, mmun_start, pmd);
1825 set_pmd_at(mm, mmun_start, pmd, entry);
1826 flush_tlb_range(vma, mmun_start, mmun_end);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001827 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001828
1829 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08001830 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1831 flush_tlb_range(vma, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001832 update_mmu_cache_pmd(vma, address, &entry);
1833 page_remove_rmap(new_page);
1834 goto fail_putback;
1835 }
1836
Mel Gormanb32967f2012-11-19 12:35:47 +00001837 page_remove_rmap(page);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001838
Mel Gormanb32967f2012-11-19 12:35:47 +00001839 /*
1840 * Finish the charge transaction under the page table lock to
1841 * prevent split_huge_page() from dividing up the charge
1842 * before it's fully transferred to the new page.
1843 */
1844 mem_cgroup_end_migration(memcg, page, new_page, true);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001845 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001846 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001847
1848 unlock_page(new_page);
1849 unlock_page(page);
1850 put_page(page); /* Drop the rmap reference */
1851 put_page(page); /* Drop the LRU isolation reference */
1852
1853 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1854 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1855
Mel Gormanb32967f2012-11-19 12:35:47 +00001856 mod_zone_page_state(page_zone(page),
1857 NR_ISOLATED_ANON + page_lru,
1858 -HPAGE_PMD_NR);
1859 return isolated;
1860
Hugh Dickins340ef392013-02-22 16:34:33 -08001861out_fail:
1862 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001863out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08001864 ptl = pmd_lock(mm, pmd);
1865 if (pmd_same(*pmd, entry)) {
1866 entry = pmd_mknonnuma(entry);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001867 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001868 update_mmu_cache_pmd(vma, address, &entry);
1869 }
1870 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001871
Mel Gormaneb4489f62013-12-18 17:08:39 -08001872out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08001873 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001874 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001875 return 0;
1876}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001877#endif /* CONFIG_NUMA_BALANCING */
1878
1879#endif /* CONFIG_NUMA */