blob: 344cdf692fc8060b20022a559dd222fded2a6e12 [file] [log] [blame]
Christoph Lameterb20a3502006-03-22 00:09:12 -08001/*
2 * Memory Migration functionality - linux/mm/migration.c
3 *
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5 *
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
8 *
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
Christoph Lametercde53532008-07-04 09:59:22 -070012 * Christoph Lameter
Christoph Lameterb20a3502006-03-22 00:09:12 -080013 */
14
15#include <linux/migrate.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040016#include <linux/export.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080017#include <linux/swap.h>
Christoph Lameter06972122006-06-23 02:03:35 -070018#include <linux/swapops.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080019#include <linux/pagemap.h>
Christoph Lametere23ca002006-04-10 22:52:57 -070020#include <linux/buffer_head.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080021#include <linux/mm_inline.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070022#include <linux/nsproxy.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080023#include <linux/pagevec.h>
Hugh Dickinse9995ef2009-12-14 17:59:31 -080024#include <linux/ksm.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080025#include <linux/rmap.h>
26#include <linux/topology.h>
27#include <linux/cpu.h>
28#include <linux/cpuset.h>
Christoph Lameter04e62a22006-06-23 02:03:38 -070029#include <linux/writeback.h>
Christoph Lameter742755a2006-06-23 02:03:55 -070030#include <linux/mempolicy.h>
31#include <linux/vmalloc.h>
David Quigley86c3a762006-06-23 02:04:02 -070032#include <linux/security.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080033#include <linux/memcontrol.h>
Adrian Bunk4f5ca262008-07-23 21:27:02 -070034#include <linux/syscalls.h>
Naoya Horiguchi290408d2010-09-08 10:19:35 +090035#include <linux/hugetlb.h>
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -070036#include <linux/hugetlb_cgroup.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090037#include <linux/gfp.h>
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080038#include <linux/balloon_compaction.h>
Mel Gormanf714f4f2013-12-18 17:08:33 -080039#include <linux/mmu_notifier.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080040
Michal Nazarewicz0d1836c2010-12-21 17:24:26 -080041#include <asm/tlbflush.h>
42
Mel Gorman7b2a2d42012-10-19 14:07:31 +010043#define CREATE_TRACE_POINTS
44#include <trace/events/migrate.h>
45
Christoph Lameterb20a3502006-03-22 00:09:12 -080046#include "internal.h"
47
Christoph Lameterb20a3502006-03-22 00:09:12 -080048/*
Christoph Lameter742755a2006-06-23 02:03:55 -070049 * migrate_prep() needs to be called before we start compiling a list of pages
Mel Gorman748446b2010-05-24 14:32:27 -070050 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
51 * undesirable, use migrate_prep_local()
Christoph Lameterb20a3502006-03-22 00:09:12 -080052 */
53int migrate_prep(void)
54{
Christoph Lameterb20a3502006-03-22 00:09:12 -080055 /*
56 * Clear the LRU lists so pages can be isolated.
57 * Note that pages may be moved off the LRU after we have
58 * drained them. Those pages will fail to migrate like other
59 * pages that may be busy.
60 */
61 lru_add_drain_all();
62
63 return 0;
64}
65
Mel Gorman748446b2010-05-24 14:32:27 -070066/* Do the necessary work of migrate_prep but not if it involves other CPUs */
67int migrate_prep_local(void)
68{
69 lru_add_drain();
70
71 return 0;
72}
73
Christoph Lameterb20a3502006-03-22 00:09:12 -080074/*
Rafael Aquini5733c7d2012-12-11 16:02:47 -080075 * Put previously isolated pages back onto the appropriate lists
76 * from where they were once taken off for compaction/migration.
77 *
Joonsoo Kim59c82b72014-01-21 15:51:17 -080078 * This function shall be used whenever the isolated pageset has been
79 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
80 * and isolate_huge_page().
Rafael Aquini5733c7d2012-12-11 16:02:47 -080081 */
82void putback_movable_pages(struct list_head *l)
83{
84 struct page *page;
85 struct page *page2;
86
87 list_for_each_entry_safe(page, page2, l, lru) {
Naoya Horiguchi31caf662013-09-11 14:21:59 -070088 if (unlikely(PageHuge(page))) {
89 putback_active_hugepage(page);
90 continue;
91 }
Rafael Aquini5733c7d2012-12-11 16:02:47 -080092 list_del(&page->lru);
93 dec_zone_page_state(page, NR_ISOLATED_ANON +
94 page_is_file_cache(page));
Rafael Aquini117aad12013-09-30 13:45:16 -070095 if (unlikely(isolated_balloon_page(page)))
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080096 balloon_page_putback(page);
97 else
98 putback_lru_page(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -080099 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800100}
101
Christoph Lameter06972122006-06-23 02:03:35 -0700102/*
103 * Restore a potential migration pte to a working pte entry
104 */
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800105static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
106 unsigned long addr, void *old)
Christoph Lameter06972122006-06-23 02:03:35 -0700107{
108 struct mm_struct *mm = vma->vm_mm;
109 swp_entry_t entry;
Christoph Lameter06972122006-06-23 02:03:35 -0700110 pmd_t *pmd;
111 pte_t *ptep, pte;
112 spinlock_t *ptl;
113
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900114 if (unlikely(PageHuge(new))) {
115 ptep = huge_pte_offset(mm, addr);
116 if (!ptep)
117 goto out;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800118 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900119 } else {
Bob Liu62190492012-12-11 16:00:37 -0800120 pmd = mm_find_pmd(mm, addr);
121 if (!pmd)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900122 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700123
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900124 ptep = pte_offset_map(pmd, addr);
Christoph Lameter06972122006-06-23 02:03:35 -0700125
Hugh Dickins486cf462011-10-19 12:50:35 -0700126 /*
127 * Peek to check is_swap_pte() before taking ptlock? No, we
128 * can race mremap's move_ptes(), which skips anon_vma lock.
129 */
Christoph Lameter06972122006-06-23 02:03:35 -0700130
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900131 ptl = pte_lockptr(mm, pmd);
132 }
133
Christoph Lameter06972122006-06-23 02:03:35 -0700134 spin_lock(ptl);
135 pte = *ptep;
136 if (!is_swap_pte(pte))
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800137 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700138
139 entry = pte_to_swp_entry(pte);
140
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800141 if (!is_migration_entry(entry) ||
142 migration_entry_to_page(entry) != old)
143 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700144
Christoph Lameter06972122006-06-23 02:03:35 -0700145 get_page(new);
146 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700147 if (pte_swp_soft_dirty(*ptep))
148 pte = pte_mksoft_dirty(pte);
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100149
150 /* Recheck VMA as permissions can change since migration started */
Christoph Lameter06972122006-06-23 02:03:35 -0700151 if (is_write_migration_entry(entry))
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100152 pte = maybe_mkwrite(pte, vma);
153
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200154#ifdef CONFIG_HUGETLB_PAGE
Tony Lube7517d2013-02-04 14:28:46 -0800155 if (PageHuge(new)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900156 pte = pte_mkhuge(pte);
Tony Lube7517d2013-02-04 14:28:46 -0800157 pte = arch_make_huge_pte(pte, vma, new, 0);
158 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200159#endif
Leonid Yegoshinc2cc4992013-05-24 15:55:18 -0700160 flush_dcache_page(new);
Christoph Lameter06972122006-06-23 02:03:35 -0700161 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700162
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900163 if (PageHuge(new)) {
164 if (PageAnon(new))
165 hugepage_add_anon_rmap(new, vma, addr);
166 else
167 page_dup_rmap(new);
168 } else if (PageAnon(new))
Christoph Lameter04e62a22006-06-23 02:03:38 -0700169 page_add_anon_rmap(new, vma, addr);
170 else
171 page_add_file_rmap(new);
172
173 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000174 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800175unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700176 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800177out:
178 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700179}
180
181/*
Hugh Dickins7e09e732014-03-20 21:52:17 -0700182 * Congratulations to trinity for discovering this bug.
183 * mm/fremap.c's remap_file_pages() accepts any range within a single vma to
184 * convert that vma to VM_NONLINEAR; and generic_file_remap_pages() will then
185 * replace the specified range by file ptes throughout (maybe populated after).
186 * If page migration finds a page within that range, while it's still located
187 * by vma_interval_tree rather than lost to i_mmap_nonlinear list, no problem:
188 * zap_pte() clears the temporary migration entry before mmap_sem is dropped.
189 * But if the migrating page is in a part of the vma outside the range to be
190 * remapped, then it will not be cleared, and remove_migration_ptes() needs to
191 * deal with it. Fortunately, this part of the vma is of course still linear,
192 * so we just need to use linear location on the nonlinear list.
193 */
194static int remove_linear_migration_ptes_from_nonlinear(struct page *page,
195 struct address_space *mapping, void *arg)
196{
197 struct vm_area_struct *vma;
198 /* hugetlbfs does not support remap_pages, so no huge pgoff worries */
199 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
200 unsigned long addr;
201
202 list_for_each_entry(vma,
203 &mapping->i_mmap_nonlinear, shared.nonlinear) {
204
205 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
206 if (addr >= vma->vm_start && addr < vma->vm_end)
207 remove_migration_pte(page, vma, addr, arg);
208 }
209 return SWAP_AGAIN;
210}
211
212/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700213 * Get rid of all migration entries and replace them by
214 * references to the indicated page.
215 */
216static void remove_migration_ptes(struct page *old, struct page *new)
217{
Joonsoo Kim051ac832014-01-21 15:49:48 -0800218 struct rmap_walk_control rwc = {
219 .rmap_one = remove_migration_pte,
220 .arg = old,
Hugh Dickins7e09e732014-03-20 21:52:17 -0700221 .file_nonlinear = remove_linear_migration_ptes_from_nonlinear,
Joonsoo Kim051ac832014-01-21 15:49:48 -0800222 };
223
224 rmap_walk(new, &rwc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700225}
226
227/*
Christoph Lameter06972122006-06-23 02:03:35 -0700228 * Something used the pte of a page under migration. We need to
229 * get to the page and wait until migration is finished.
230 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700231 */
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700232static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
233 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700234{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700235 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700236 swp_entry_t entry;
237 struct page *page;
238
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700239 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700240 pte = *ptep;
241 if (!is_swap_pte(pte))
242 goto out;
243
244 entry = pte_to_swp_entry(pte);
245 if (!is_migration_entry(entry))
246 goto out;
247
248 page = migration_entry_to_page(entry);
249
Nick Piggine2867812008-07-25 19:45:30 -0700250 /*
251 * Once radix-tree replacement of page migration started, page_count
252 * *must* be zero. And, we don't want to call wait_on_page_locked()
253 * against a page without get_page().
254 * So, we use get_page_unless_zero(), here. Even failed, page fault
255 * will occur again.
256 */
257 if (!get_page_unless_zero(page))
258 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700259 pte_unmap_unlock(ptep, ptl);
260 wait_on_page_locked(page);
261 put_page(page);
262 return;
263out:
264 pte_unmap_unlock(ptep, ptl);
265}
266
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700267void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
268 unsigned long address)
269{
270 spinlock_t *ptl = pte_lockptr(mm, pmd);
271 pte_t *ptep = pte_offset_map(pmd, address);
272 __migration_entry_wait(mm, ptep, ptl);
273}
274
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800275void migration_entry_wait_huge(struct vm_area_struct *vma,
276 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700277{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800278 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700279 __migration_entry_wait(mm, pte, ptl);
280}
281
Mel Gormanb969c4a2012-01-12 17:19:34 -0800282#ifdef CONFIG_BLOCK
283/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800284static bool buffer_migrate_lock_buffers(struct buffer_head *head,
285 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800286{
287 struct buffer_head *bh = head;
288
289 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800290 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800291 do {
292 get_bh(bh);
293 lock_buffer(bh);
294 bh = bh->b_this_page;
295
296 } while (bh != head);
297
298 return true;
299 }
300
301 /* async case, we cannot block on lock_buffer so use trylock_buffer */
302 do {
303 get_bh(bh);
304 if (!trylock_buffer(bh)) {
305 /*
306 * We failed to lock the buffer and cannot stall in
307 * async migration. Release the taken locks
308 */
309 struct buffer_head *failed_bh = bh;
310 put_bh(failed_bh);
311 bh = head;
312 while (bh != failed_bh) {
313 unlock_buffer(bh);
314 put_bh(bh);
315 bh = bh->b_this_page;
316 }
317 return false;
318 }
319
320 bh = bh->b_this_page;
321 } while (bh != head);
322 return true;
323}
324#else
325static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800326 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800327{
328 return true;
329}
330#endif /* CONFIG_BLOCK */
331
Christoph Lameterb20a3502006-03-22 00:09:12 -0800332/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700333 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700334 *
335 * The number of remaining references must be:
336 * 1 for anonymous pages without a mapping
337 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100338 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800339 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800340int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800341 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500342 struct buffer_head *head, enum migrate_mode mode,
343 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800344{
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500345 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800346 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800347
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700348 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700349 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500350 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700351 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800352 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700353 }
354
Nick Piggin19fd6232008-07-25 19:45:32 -0700355 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800356
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800357 pslot = radix_tree_lookup_slot(&mapping->page_tree,
358 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800359
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500360 expected_count += 1 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700361 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800362 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700363 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700364 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800365 }
366
Nick Piggine2867812008-07-25 19:45:30 -0700367 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700368 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700369 return -EAGAIN;
370 }
371
Christoph Lameterb20a3502006-03-22 00:09:12 -0800372 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800373 * In the async migration case of moving a page with buffers, lock the
374 * buffers using trylock before the mapping is moved. If the mapping
375 * was moved, we later failed to lock the buffers and could not move
376 * the mapping back due to an elevated page count, we would have to
377 * block waiting on other references to be dropped.
378 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800379 if (mode == MIGRATE_ASYNC && head &&
380 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800381 page_unfreeze_refs(page, expected_count);
382 spin_unlock_irq(&mapping->tree_lock);
383 return -EAGAIN;
384 }
385
386 /*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800387 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800388 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800389 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800390 if (PageSwapCache(page)) {
391 SetPageSwapCache(newpage);
392 set_page_private(newpage, page_private(page));
393 }
394
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800395 radix_tree_replace_slot(pslot, newpage);
396
397 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800398 * Drop cache reference from old page by unfreezing
399 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800400 * We know this isn't the last reference.
401 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800402 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800403
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700404 /*
405 * If moved to a different zone then also account
406 * the page for that zone. Other VM counters will be
407 * taken care of when we establish references to the
408 * new page and drop references to the old page.
409 *
410 * Note that anonymous pages are accounted for
411 * via NR_FILE_PAGES and NR_ANON_PAGES if they
412 * are mapped to swap space.
413 */
414 __dec_zone_page_state(page, NR_FILE_PAGES);
415 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700416 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700417 __dec_zone_page_state(page, NR_SHMEM);
418 __inc_zone_page_state(newpage, NR_SHMEM);
419 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700420 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800421
Rafael Aquini78bd5202012-12-11 16:02:31 -0800422 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800423}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800424
425/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900426 * The expected number of remaining references is the same as that
427 * of migrate_page_move_mapping().
428 */
429int migrate_huge_page_move_mapping(struct address_space *mapping,
430 struct page *newpage, struct page *page)
431{
432 int expected_count;
433 void **pslot;
434
435 if (!mapping) {
436 if (page_count(page) != 1)
437 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800438 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900439 }
440
441 spin_lock_irq(&mapping->tree_lock);
442
443 pslot = radix_tree_lookup_slot(&mapping->page_tree,
444 page_index(page));
445
446 expected_count = 2 + page_has_private(page);
447 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800448 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900449 spin_unlock_irq(&mapping->tree_lock);
450 return -EAGAIN;
451 }
452
453 if (!page_freeze_refs(page, expected_count)) {
454 spin_unlock_irq(&mapping->tree_lock);
455 return -EAGAIN;
456 }
457
458 get_page(newpage);
459
460 radix_tree_replace_slot(pslot, newpage);
461
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800462 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900463
464 spin_unlock_irq(&mapping->tree_lock);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800465 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900466}
467
468/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800469 * Gigantic pages are so large that we do not guarantee that page++ pointer
470 * arithmetic will work across the entire page. We need something more
471 * specialized.
472 */
473static void __copy_gigantic_page(struct page *dst, struct page *src,
474 int nr_pages)
475{
476 int i;
477 struct page *dst_base = dst;
478 struct page *src_base = src;
479
480 for (i = 0; i < nr_pages; ) {
481 cond_resched();
482 copy_highpage(dst, src);
483
484 i++;
485 dst = mem_map_next(dst, dst_base, i);
486 src = mem_map_next(src, src_base, i);
487 }
488}
489
490static void copy_huge_page(struct page *dst, struct page *src)
491{
492 int i;
493 int nr_pages;
494
495 if (PageHuge(src)) {
496 /* hugetlbfs page */
497 struct hstate *h = page_hstate(src);
498 nr_pages = pages_per_huge_page(h);
499
500 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
501 __copy_gigantic_page(dst, src, nr_pages);
502 return;
503 }
504 } else {
505 /* thp page */
506 BUG_ON(!PageTransHuge(src));
507 nr_pages = hpage_nr_pages(src);
508 }
509
510 for (i = 0; i < nr_pages; i++) {
511 cond_resched();
512 copy_highpage(dst + i, src + i);
513 }
514}
515
516/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800517 * Copy the page to its new location
518 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900519void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800520{
Rik van Riel7851a452013-10-07 11:29:23 +0100521 int cpupid;
522
Mel Gormanb32967f2012-11-19 12:35:47 +0000523 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900524 copy_huge_page(newpage, page);
525 else
526 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800527
528 if (PageError(page))
529 SetPageError(newpage);
530 if (PageReferenced(page))
531 SetPageReferenced(newpage);
532 if (PageUptodate(page))
533 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700534 if (TestClearPageActive(page)) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800535 VM_BUG_ON_PAGE(PageUnevictable(page), page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800536 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800537 } else if (TestClearPageUnevictable(page))
538 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800539 if (PageChecked(page))
540 SetPageChecked(newpage);
541 if (PageMappedToDisk(page))
542 SetPageMappedToDisk(newpage);
543
544 if (PageDirty(page)) {
545 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700546 /*
547 * Want to mark the page and the radix tree as dirty, and
548 * redo the accounting that clear_page_dirty_for_io undid,
549 * but we can't use set_page_dirty because that function
550 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300551 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700552 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700553 if (PageSwapBacked(page))
554 SetPageDirty(newpage);
555 else
556 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800557 }
558
Rik van Riel7851a452013-10-07 11:29:23 +0100559 /*
560 * Copy NUMA information to the new page, to prevent over-eager
561 * future migrations of this same page.
562 */
563 cpupid = page_cpupid_xchg_last(page, -1);
564 page_cpupid_xchg_last(newpage, cpupid);
565
Nick Pigginb291f002008-10-18 20:26:44 -0700566 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800567 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800568 /*
569 * Please do not reorder this without considering how mm/ksm.c's
570 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
571 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800572 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800573 ClearPagePrivate(page);
574 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800575
576 /*
577 * If any waiters have accumulated on the new page then
578 * wake them up.
579 */
580 if (PageWriteback(newpage))
581 end_page_writeback(newpage);
582}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800583
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700584/************************************************************
585 * Migration functions
586 ***********************************************************/
587
Christoph Lameterb20a3502006-03-22 00:09:12 -0800588/*
589 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100590 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800591 *
592 * Pages are locked upon entry and exit.
593 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700594int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800595 struct page *newpage, struct page *page,
596 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800597{
598 int rc;
599
600 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
601
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500602 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800603
Rafael Aquini78bd5202012-12-11 16:02:31 -0800604 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800605 return rc;
606
607 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800608 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800609}
610EXPORT_SYMBOL(migrate_page);
611
David Howells93614012006-09-30 20:45:40 +0200612#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800613/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700614 * Migration function for pages with buffers. This function can only be used
615 * if the underlying filesystem guarantees that no other references to "page"
616 * exist.
617 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700618int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800619 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700620{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700621 struct buffer_head *bh, *head;
622 int rc;
623
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700624 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800625 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700626
627 head = page_buffers(page);
628
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500629 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700630
Rafael Aquini78bd5202012-12-11 16:02:31 -0800631 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700632 return rc;
633
Mel Gormanb969c4a2012-01-12 17:19:34 -0800634 /*
635 * In the async case, migrate_page_move_mapping locked the buffers
636 * with an IRQ-safe spinlock held. In the sync case, the buffers
637 * need to be locked now
638 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800639 if (mode != MIGRATE_ASYNC)
640 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700641
642 ClearPagePrivate(page);
643 set_page_private(newpage, page_private(page));
644 set_page_private(page, 0);
645 put_page(page);
646 get_page(newpage);
647
648 bh = head;
649 do {
650 set_bh_page(bh, newpage, bh_offset(bh));
651 bh = bh->b_this_page;
652
653 } while (bh != head);
654
655 SetPagePrivate(newpage);
656
657 migrate_page_copy(newpage, page);
658
659 bh = head;
660 do {
661 unlock_buffer(bh);
662 put_bh(bh);
663 bh = bh->b_this_page;
664
665 } while (bh != head);
666
Rafael Aquini78bd5202012-12-11 16:02:31 -0800667 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700668}
669EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200670#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700671
Christoph Lameter04e62a22006-06-23 02:03:38 -0700672/*
673 * Writeback a page to clean the dirty state
674 */
675static int writeout(struct address_space *mapping, struct page *page)
676{
677 struct writeback_control wbc = {
678 .sync_mode = WB_SYNC_NONE,
679 .nr_to_write = 1,
680 .range_start = 0,
681 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700682 .for_reclaim = 1
683 };
684 int rc;
685
686 if (!mapping->a_ops->writepage)
687 /* No write method for the address space */
688 return -EINVAL;
689
690 if (!clear_page_dirty_for_io(page))
691 /* Someone else already triggered a write */
692 return -EAGAIN;
693
694 /*
695 * A dirty page may imply that the underlying filesystem has
696 * the page on some queue. So the page must be clean for
697 * migration. Writeout may mean we loose the lock and the
698 * page state is no longer what we checked for earlier.
699 * At this point we know that the migration attempt cannot
700 * be successful.
701 */
702 remove_migration_ptes(page, page);
703
704 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700705
706 if (rc != AOP_WRITEPAGE_ACTIVATE)
707 /* unlocked. Relock */
708 lock_page(page);
709
Hugh Dickinsbda85502008-11-19 15:36:36 -0800710 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700711}
712
713/*
714 * Default handling if a filesystem does not provide a migration function.
715 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700716static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800717 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700718{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800719 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800720 /* Only writeback pages in full synchronous migration */
721 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800722 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700723 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800724 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700725
726 /*
727 * Buffers may be managed in a filesystem specific way.
728 * We must have no buffers or drop them.
729 */
David Howells266cf652009-04-03 16:42:36 +0100730 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700731 !try_to_release_page(page, GFP_KERNEL))
732 return -EAGAIN;
733
Mel Gormana6bc32b2012-01-12 17:19:43 -0800734 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700735}
736
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700737/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700738 * Move a page to a newly allocated page
739 * The page is locked and all ptes have been successfully removed.
740 *
741 * The new page will have replaced the old page if this function
742 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700743 *
744 * Return value:
745 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800746 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700747 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700748static int move_to_new_page(struct page *newpage, struct page *page,
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800749 int page_was_mapped, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700750{
751 struct address_space *mapping;
752 int rc;
753
754 /*
755 * Block others from accessing the page when we get around to
756 * establishing additional references. We are the only one
757 * holding a reference to the new page at this point.
758 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200759 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700760 BUG();
761
762 /* Prepare mapping for the new page.*/
763 newpage->index = page->index;
764 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700765 if (PageSwapBacked(page))
766 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700767
768 mapping = page_mapping(page);
769 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800770 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800771 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700772 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800773 * Most pages have a mapping and most filesystems provide a
774 * migratepage callback. Anonymous pages are part of swap
775 * space which also has its own migratepage callback. This
776 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700777 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800778 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800779 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800780 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800781 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700782
Rafael Aquini78bd5202012-12-11 16:02:31 -0800783 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700784 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700785 } else {
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700786 mem_cgroup_migrate(page, newpage, false);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800787 if (page_was_mapped)
Mel Gorman3fe20112010-05-24 14:32:20 -0700788 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800789 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700790 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700791
792 unlock_page(newpage);
793
794 return rc;
795}
796
Minchan Kim0dabec92011-10-31 17:06:57 -0700797static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800798 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700799{
Minchan Kim0dabec92011-10-31 17:06:57 -0700800 int rc = -EAGAIN;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800801 int page_was_mapped = 0;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700802 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700803
Nick Piggin529ae9a2008-08-02 12:01:03 +0200804 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800805 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700806 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800807
808 /*
809 * It's not safe for direct compaction to call lock_page.
810 * For example, during page readahead pages are added locked
811 * to the LRU. Later, when the IO completes the pages are
812 * marked uptodate and unlocked. However, the queueing
813 * could be merging multiple pages for one bio (e.g.
814 * mpage_readpages). If an allocation happens for the
815 * second or third page, the process can end up locking
816 * the same page twice and deadlocking. Rather than
817 * trying to be clever about what pages can be locked,
818 * avoid the use of lock_page for direct compaction
819 * altogether.
820 */
821 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700822 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800823
Christoph Lametere24f0b82006-06-23 02:03:51 -0700824 lock_page(page);
825 }
826
827 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700828 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700829 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800830 * necessary to wait for PageWriteback. In the async case,
831 * the retry loop is too short and in the sync-light case,
832 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700833 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800834 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700835 rc = -EBUSY;
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700836 goto out_unlock;
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700837 }
838 if (!force)
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700839 goto out_unlock;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700840 wait_on_page_writeback(page);
841 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700842 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700843 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
844 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800845 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700846 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700847 * File Caches may use write_page() or lock_page() in migration, then,
848 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700849 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800850 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800851 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000852 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800853 * getting a hold on an anon_vma from outside one of its mms.
854 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700855 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800856 if (anon_vma) {
857 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700858 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800859 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800860 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700861 /*
862 * We cannot be sure that the anon_vma of an unmapped
863 * swapcache page is safe to use because we don't
864 * know in advance if the VMA that this page belonged
865 * to still exists. If the VMA and others sharing the
866 * data have been freed, then the anon_vma could
867 * already be invalid.
868 *
869 * To avoid this possibility, swapcache pages get
870 * migrated but are not remapped when migration
871 * completes
872 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700873 } else {
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700874 goto out_unlock;
Mel Gorman3fe20112010-05-24 14:32:20 -0700875 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700876 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800877
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700878 if (unlikely(isolated_balloon_page(page))) {
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800879 /*
880 * A ballooned page does not need any special attention from
881 * physical to virtual reverse mapping procedures.
882 * Skip any attempt to unmap PTEs or to remap swap cache,
883 * in order to avoid burning cycles at rmap level, and perform
884 * the page migration right away (proteced by page lock).
885 */
886 rc = balloon_page_migrate(newpage, page, mode);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700887 goto out_unlock;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800888 }
889
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700890 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800891 * Corner case handling:
892 * 1. When a new swap-cache page is read into, it is added to the LRU
893 * and treated as swapcache but it has no rmap yet.
894 * Calling try_to_unmap() against a page->mapping==NULL page will
895 * trigger a BUG. So handle it here.
896 * 2. An orphaned page (see truncate_complete_page) might have
897 * fs-private metadata. The page can be picked up due to memory
898 * offlining. Everywhere else except page reclaim, the page is
899 * invisible to the vm, so the page can not be migrated. So try to
900 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700901 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800902 if (!page->mapping) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800903 VM_BUG_ON_PAGE(PageAnon(page), page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800904 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800905 try_to_free_buffers(page);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700906 goto out_unlock;
Shaohua Li62e1c552008-02-04 22:29:33 -0800907 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700908 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800909 }
910
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700911 /* Establish migration ptes or remove ptes */
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800912 if (page_mapped(page)) {
913 try_to_unmap(page,
914 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
915 page_was_mapped = 1;
916 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700917
Shaohua Liabfc3482009-09-21 17:01:19 -0700918skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700919 if (!page_mapped(page))
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800920 rc = move_to_new_page(newpage, page, page_was_mapped, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700921
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800922 if (rc && page_was_mapped)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700923 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700924
925 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700926 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700927 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700928
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700929out_unlock:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700930 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700931out:
932 return rc;
933}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700934
Minchan Kim0dabec92011-10-31 17:06:57 -0700935/*
936 * Obtain the lock on page, remove all ptes and migrate the page
937 * to the newly allocated page in newpage.
938 */
David Rientjes68711a72014-06-04 16:08:25 -0700939static int unmap_and_move(new_page_t get_new_page, free_page_t put_new_page,
940 unsigned long private, struct page *page, int force,
941 enum migrate_mode mode)
Minchan Kim0dabec92011-10-31 17:06:57 -0700942{
943 int rc = 0;
944 int *result = NULL;
945 struct page *newpage = get_new_page(page, private, &result);
946
947 if (!newpage)
948 return -ENOMEM;
949
950 if (page_count(page) == 1) {
951 /* page was freed from under us. So we are done. */
952 goto out;
953 }
954
955 if (unlikely(PageTransHuge(page)))
956 if (unlikely(split_huge_page(page)))
957 goto out;
958
Hugh Dickins9c620e22013-02-22 16:35:14 -0800959 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800960
Minchan Kim0dabec92011-10-31 17:06:57 -0700961out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700962 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700963 /*
964 * A page that has been migrated has all references
965 * removed and will be freed. A page that has not been
966 * migrated will have kepts its references and be
967 * restored.
968 */
969 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700970 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700971 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700972 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700973 }
David Rientjes68711a72014-06-04 16:08:25 -0700974
Christoph Lameter95a402c2006-06-23 02:03:53 -0700975 /*
David Rientjes68711a72014-06-04 16:08:25 -0700976 * If migration was not successful and there's a freeing callback, use
977 * it. Otherwise, putback_lru_page() will drop the reference grabbed
978 * during isolation.
Christoph Lameter95a402c2006-06-23 02:03:53 -0700979 */
Hugh Dickins8bdd6382014-07-26 12:58:23 -0700980 if (rc != MIGRATEPAGE_SUCCESS && put_new_page) {
981 ClearPageSwapBacked(newpage);
David Rientjes68711a72014-06-04 16:08:25 -0700982 put_new_page(newpage, private);
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700983 } else if (unlikely(__is_movable_balloon_page(newpage))) {
984 /* drop our reference, page already in the balloon */
985 put_page(newpage);
Hugh Dickins8bdd6382014-07-26 12:58:23 -0700986 } else
David Rientjes68711a72014-06-04 16:08:25 -0700987 putback_lru_page(newpage);
988
Christoph Lameter742755a2006-06-23 02:03:55 -0700989 if (result) {
990 if (rc)
991 *result = rc;
992 else
993 *result = page_to_nid(newpage);
994 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700995 return rc;
996}
997
998/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900999 * Counterpart of unmap_and_move_page() for hugepage migration.
1000 *
1001 * This function doesn't wait the completion of hugepage I/O
1002 * because there is no race between I/O and migration for hugepage.
1003 * Note that currently hugepage I/O occurs only in direct I/O
1004 * where no lock is held and PG_writeback is irrelevant,
1005 * and writeback status of all subpages are counted in the reference
1006 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1007 * under direct I/O, the reference of the head page is 512 and a bit more.)
1008 * This means that when we try to migrate hugepage whose subpages are
1009 * doing direct I/O, some references remain after try_to_unmap() and
1010 * hugepage migration fails without data corruption.
1011 *
1012 * There is also no race when direct I/O is issued on the page under migration,
1013 * because then pte is replaced with migration swap entry and direct I/O code
1014 * will wait in the page fault for migration to complete.
1015 */
1016static int unmap_and_move_huge_page(new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001017 free_page_t put_new_page, unsigned long private,
1018 struct page *hpage, int force,
1019 enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001020{
1021 int rc = 0;
1022 int *result = NULL;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001023 int page_was_mapped = 0;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001024 struct page *new_hpage;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001025 struct anon_vma *anon_vma = NULL;
1026
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001027 /*
1028 * Movability of hugepages depends on architectures and hugepage size.
1029 * This check is necessary because some callers of hugepage migration
1030 * like soft offline and memory hotremove don't walk through page
1031 * tables or check whether the hugepage is pmd-based or not before
1032 * kicking migration.
1033 */
Naoya Horiguchi100873d2014-06-04 16:10:56 -07001034 if (!hugepage_migration_supported(page_hstate(hpage))) {
Joonsoo Kim32665f22014-01-21 15:51:15 -08001035 putback_active_hugepage(hpage);
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001036 return -ENOSYS;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001037 }
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001038
Joonsoo Kim32665f22014-01-21 15:51:15 -08001039 new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001040 if (!new_hpage)
1041 return -ENOMEM;
1042
1043 rc = -EAGAIN;
1044
1045 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001046 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001047 goto out;
1048 lock_page(hpage);
1049 }
1050
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001051 if (PageAnon(hpage))
1052 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001053
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001054 if (page_mapped(hpage)) {
1055 try_to_unmap(hpage,
1056 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1057 page_was_mapped = 1;
1058 }
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001059
1060 if (!page_mapped(hpage))
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001061 rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001062
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001063 if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001064 remove_migration_ptes(hpage, hpage);
1065
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001066 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001067 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001068
David Rientjes68711a72014-06-04 16:08:25 -07001069 if (rc == MIGRATEPAGE_SUCCESS)
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001070 hugetlb_cgroup_migrate(hpage, new_hpage);
1071
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001072 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001073out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001074 if (rc != -EAGAIN)
1075 putback_active_hugepage(hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001076
1077 /*
1078 * If migration was not successful and there's a freeing callback, use
1079 * it. Otherwise, put_page() will drop the reference grabbed during
1080 * isolation.
1081 */
1082 if (rc != MIGRATEPAGE_SUCCESS && put_new_page)
1083 put_new_page(new_hpage, private);
1084 else
1085 put_page(new_hpage);
1086
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001087 if (result) {
1088 if (rc)
1089 *result = rc;
1090 else
1091 *result = page_to_nid(new_hpage);
1092 }
1093 return rc;
1094}
1095
1096/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001097 * migrate_pages - migrate the pages specified in a list, to the free pages
1098 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001099 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001100 * @from: The list of pages to be migrated.
1101 * @get_new_page: The function used to allocate free pages to be used
1102 * as the target of the page migration.
David Rientjes68711a72014-06-04 16:08:25 -07001103 * @put_new_page: The function used to free target pages if migration
1104 * fails, or NULL if no special handling is necessary.
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001105 * @private: Private data to be passed on to get_new_page()
1106 * @mode: The migration mode that specifies the constraints for
1107 * page migration, if any.
1108 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001109 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001110 * The function returns after 10 attempts or if no pages are movable any more
1111 * because the list has become empty or no retryable pages exist any more.
1112 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001113 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001114 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001115 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001116 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001117int migrate_pages(struct list_head *from, new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001118 free_page_t put_new_page, unsigned long private,
1119 enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001120{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001121 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001122 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001123 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001124 int pass = 0;
1125 struct page *page;
1126 struct page *page2;
1127 int swapwrite = current->flags & PF_SWAPWRITE;
1128 int rc;
1129
1130 if (!swapwrite)
1131 current->flags |= PF_SWAPWRITE;
1132
Christoph Lametere24f0b82006-06-23 02:03:51 -07001133 for(pass = 0; pass < 10 && retry; pass++) {
1134 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001135
Christoph Lametere24f0b82006-06-23 02:03:51 -07001136 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001137 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001138
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001139 if (PageHuge(page))
1140 rc = unmap_and_move_huge_page(get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001141 put_new_page, private, page,
1142 pass > 2, mode);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001143 else
David Rientjes68711a72014-06-04 16:08:25 -07001144 rc = unmap_and_move(get_new_page, put_new_page,
1145 private, page, pass > 2, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001146
Christoph Lametere24f0b82006-06-23 02:03:51 -07001147 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001148 case -ENOMEM:
1149 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001150 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001151 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001152 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001153 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001154 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001155 break;
1156 default:
Naoya Horiguchi354a3362014-01-21 15:51:14 -08001157 /*
1158 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1159 * unlike -EAGAIN case, the failed page is
1160 * removed from migration page list and not
1161 * retried in the next outer loop.
1162 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001163 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001164 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001165 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001166 }
1167 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001168 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001169out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001170 if (nr_succeeded)
1171 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1172 if (nr_failed)
1173 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001174 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1175
Christoph Lameterb20a3502006-03-22 00:09:12 -08001176 if (!swapwrite)
1177 current->flags &= ~PF_SWAPWRITE;
1178
Rafael Aquini78bd5202012-12-11 16:02:31 -08001179 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001180}
1181
Christoph Lameter742755a2006-06-23 02:03:55 -07001182#ifdef CONFIG_NUMA
1183/*
1184 * Move a list of individual pages
1185 */
1186struct page_to_node {
1187 unsigned long addr;
1188 struct page *page;
1189 int node;
1190 int status;
1191};
1192
1193static struct page *new_page_node(struct page *p, unsigned long private,
1194 int **result)
1195{
1196 struct page_to_node *pm = (struct page_to_node *)private;
1197
1198 while (pm->node != MAX_NUMNODES && pm->page != p)
1199 pm++;
1200
1201 if (pm->node == MAX_NUMNODES)
1202 return NULL;
1203
1204 *result = &pm->status;
1205
Naoya Horiguchie632a932013-09-11 14:22:04 -07001206 if (PageHuge(p))
1207 return alloc_huge_page_node(page_hstate(compound_head(p)),
1208 pm->node);
1209 else
1210 return alloc_pages_exact_node(pm->node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001211 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001212}
1213
1214/*
1215 * Move a set of pages as indicated in the pm array. The addr
1216 * field must be set to the virtual address of the page to be moved
1217 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001218 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001219 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001220static int do_move_page_to_node_array(struct mm_struct *mm,
1221 struct page_to_node *pm,
1222 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001223{
1224 int err;
1225 struct page_to_node *pp;
1226 LIST_HEAD(pagelist);
1227
1228 down_read(&mm->mmap_sem);
1229
1230 /*
1231 * Build a list of pages to migrate
1232 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001233 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1234 struct vm_area_struct *vma;
1235 struct page *page;
1236
Christoph Lameter742755a2006-06-23 02:03:55 -07001237 err = -EFAULT;
1238 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001239 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001240 goto set_status;
1241
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001242 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001243
1244 err = PTR_ERR(page);
1245 if (IS_ERR(page))
1246 goto set_status;
1247
Christoph Lameter742755a2006-06-23 02:03:55 -07001248 err = -ENOENT;
1249 if (!page)
1250 goto set_status;
1251
Hugh Dickins62b61f62009-12-14 17:59:33 -08001252 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001253 if (PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001254 goto put_and_set;
1255
1256 pp->page = page;
1257 err = page_to_nid(page);
1258
1259 if (err == pp->node)
1260 /*
1261 * Node already in the right place
1262 */
1263 goto put_and_set;
1264
1265 err = -EACCES;
1266 if (page_mapcount(page) > 1 &&
1267 !migrate_all)
1268 goto put_and_set;
1269
Naoya Horiguchie632a932013-09-11 14:22:04 -07001270 if (PageHuge(page)) {
1271 isolate_huge_page(page, &pagelist);
1272 goto put_and_set;
1273 }
1274
Nick Piggin62695a82008-10-18 20:26:09 -07001275 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001276 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001277 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001278 inc_zone_page_state(page, NR_ISOLATED_ANON +
1279 page_is_file_cache(page));
1280 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001281put_and_set:
1282 /*
1283 * Either remove the duplicate refcount from
1284 * isolate_lru_page() or drop the page ref if it was
1285 * not isolated.
1286 */
1287 put_page(page);
1288set_status:
1289 pp->status = err;
1290 }
1291
Brice Gogline78bbfa2008-10-18 20:27:15 -07001292 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001293 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -07001294 err = migrate_pages(&pagelist, new_page_node, NULL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001295 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001296 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001297 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001298 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001299
1300 up_read(&mm->mmap_sem);
1301 return err;
1302}
1303
1304/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001305 * Migrate an array of page address onto an array of nodes and fill
1306 * the corresponding array of status.
1307 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001308static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001309 unsigned long nr_pages,
1310 const void __user * __user *pages,
1311 const int __user *nodes,
1312 int __user *status, int flags)
1313{
Brice Goglin3140a222009-01-06 14:38:57 -08001314 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001315 unsigned long chunk_nr_pages;
1316 unsigned long chunk_start;
1317 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001318
Brice Goglin3140a222009-01-06 14:38:57 -08001319 err = -ENOMEM;
1320 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1321 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001322 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001323
1324 migrate_prep();
1325
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001326 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001327 * Store a chunk of page_to_node array in a page,
1328 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001329 */
Brice Goglin3140a222009-01-06 14:38:57 -08001330 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001331
Brice Goglin3140a222009-01-06 14:38:57 -08001332 for (chunk_start = 0;
1333 chunk_start < nr_pages;
1334 chunk_start += chunk_nr_pages) {
1335 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001336
Brice Goglin3140a222009-01-06 14:38:57 -08001337 if (chunk_start + chunk_nr_pages > nr_pages)
1338 chunk_nr_pages = nr_pages - chunk_start;
1339
1340 /* fill the chunk pm with addrs and nodes from user-space */
1341 for (j = 0; j < chunk_nr_pages; j++) {
1342 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001343 int node;
1344
Brice Goglin3140a222009-01-06 14:38:57 -08001345 err = -EFAULT;
1346 if (get_user(p, pages + j + chunk_start))
1347 goto out_pm;
1348 pm[j].addr = (unsigned long) p;
1349
1350 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001351 goto out_pm;
1352
1353 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001354 if (node < 0 || node >= MAX_NUMNODES)
1355 goto out_pm;
1356
Lai Jiangshan389162c2012-12-12 13:51:30 -08001357 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001358 goto out_pm;
1359
1360 err = -EACCES;
1361 if (!node_isset(node, task_nodes))
1362 goto out_pm;
1363
Brice Goglin3140a222009-01-06 14:38:57 -08001364 pm[j].node = node;
1365 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001366
Brice Goglin3140a222009-01-06 14:38:57 -08001367 /* End marker for this chunk */
1368 pm[chunk_nr_pages].node = MAX_NUMNODES;
1369
1370 /* Migrate this chunk */
1371 err = do_move_page_to_node_array(mm, pm,
1372 flags & MPOL_MF_MOVE_ALL);
1373 if (err < 0)
1374 goto out_pm;
1375
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001376 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001377 for (j = 0; j < chunk_nr_pages; j++)
1378 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001379 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001380 goto out_pm;
1381 }
1382 }
1383 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001384
1385out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001386 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001387out:
1388 return err;
1389}
1390
1391/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001392 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001393 */
Brice Goglin80bba122008-12-09 13:14:23 -08001394static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1395 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001396{
Brice Goglin2f007e72008-10-18 20:27:16 -07001397 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001398
Christoph Lameter742755a2006-06-23 02:03:55 -07001399 down_read(&mm->mmap_sem);
1400
Brice Goglin2f007e72008-10-18 20:27:16 -07001401 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001402 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001403 struct vm_area_struct *vma;
1404 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001405 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001406
1407 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001408 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001409 goto set_status;
1410
Brice Goglin2f007e72008-10-18 20:27:16 -07001411 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001412
1413 err = PTR_ERR(page);
1414 if (IS_ERR(page))
1415 goto set_status;
1416
Christoph Lameter742755a2006-06-23 02:03:55 -07001417 err = -ENOENT;
1418 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001419 if (!page || PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001420 goto set_status;
1421
1422 err = page_to_nid(page);
1423set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001424 *status = err;
1425
1426 pages++;
1427 status++;
1428 }
1429
1430 up_read(&mm->mmap_sem);
1431}
1432
1433/*
1434 * Determine the nodes of a user array of pages and store it in
1435 * a user array of status.
1436 */
1437static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1438 const void __user * __user *pages,
1439 int __user *status)
1440{
1441#define DO_PAGES_STAT_CHUNK_NR 16
1442 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1443 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001444
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001445 while (nr_pages) {
1446 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001447
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001448 chunk_nr = nr_pages;
1449 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1450 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1451
1452 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1453 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001454
1455 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1456
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001457 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1458 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001459
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001460 pages += chunk_nr;
1461 status += chunk_nr;
1462 nr_pages -= chunk_nr;
1463 }
1464 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001465}
1466
1467/*
1468 * Move a list of pages in the address space of the currently executing
1469 * process.
1470 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001471SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1472 const void __user * __user *, pages,
1473 const int __user *, nodes,
1474 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001475{
David Howellsc69e8d92008-11-14 10:39:19 +11001476 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001477 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001478 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001479 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001480 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001481
1482 /* Check flags */
1483 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1484 return -EINVAL;
1485
1486 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1487 return -EPERM;
1488
1489 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001490 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001491 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001492 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001493 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001494 return -ESRCH;
1495 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001496 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001497
1498 /*
1499 * Check if this process has the right to modify the specified
1500 * process. The right exists if the process has administrative
1501 * capabilities, superuser privileges or the same
1502 * userid as the target process.
1503 */
David Howellsc69e8d92008-11-14 10:39:19 +11001504 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001505 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1506 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001507 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001508 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001509 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001510 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001511 }
David Howellsc69e8d92008-11-14 10:39:19 +11001512 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001513
David Quigley86c3a762006-06-23 02:04:02 -07001514 err = security_task_movememory(task);
1515 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001516 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001517
Christoph Lameter3268c632012-03-21 16:34:06 -07001518 task_nodes = cpuset_mems_allowed(task);
1519 mm = get_task_mm(task);
1520 put_task_struct(task);
1521
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001522 if (!mm)
1523 return -EINVAL;
1524
1525 if (nodes)
1526 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1527 nodes, status, flags);
1528 else
1529 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001530
1531 mmput(mm);
1532 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001533
Christoph Lameter742755a2006-06-23 02:03:55 -07001534out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001535 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001536 return err;
1537}
Christoph Lameter742755a2006-06-23 02:03:55 -07001538
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001539#ifdef CONFIG_NUMA_BALANCING
1540/*
1541 * Returns true if this is a safe migration target node for misplaced NUMA
1542 * pages. Currently it only checks the watermarks which crude
1543 */
1544static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001545 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001546{
1547 int z;
1548 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1549 struct zone *zone = pgdat->node_zones + z;
1550
1551 if (!populated_zone(zone))
1552 continue;
1553
Lisa Du6e543d52013-09-11 14:22:36 -07001554 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001555 continue;
1556
1557 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1558 if (!zone_watermark_ok(zone, 0,
1559 high_wmark_pages(zone) +
1560 nr_migrate_pages,
1561 0, 0))
1562 continue;
1563 return true;
1564 }
1565 return false;
1566}
1567
1568static struct page *alloc_misplaced_dst_page(struct page *page,
1569 unsigned long data,
1570 int **result)
1571{
1572 int nid = (int) data;
1573 struct page *newpage;
1574
1575 newpage = alloc_pages_exact_node(nid,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001576 (GFP_HIGHUSER_MOVABLE |
1577 __GFP_THISNODE | __GFP_NOMEMALLOC |
1578 __GFP_NORETRY | __GFP_NOWARN) &
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001579 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001580
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001581 return newpage;
1582}
1583
1584/*
Mel Gormana8f60772012-11-14 21:41:46 +00001585 * page migration rate limiting control.
1586 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1587 * window of time. Default here says do not migrate more than 1280M per second.
Mel Gormane14808b2012-11-19 10:59:15 +00001588 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1589 * as it is faults that reset the window, pte updates will happen unconditionally
1590 * if there has not been a fault since @pteupdate_interval_millisecs after the
1591 * throttle window closed.
Mel Gormana8f60772012-11-14 21:41:46 +00001592 */
1593static unsigned int migrate_interval_millisecs __read_mostly = 100;
Mel Gormane14808b2012-11-19 10:59:15 +00001594static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
Mel Gormana8f60772012-11-14 21:41:46 +00001595static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1596
Mel Gormane14808b2012-11-19 10:59:15 +00001597/* Returns true if NUMA migration is currently rate limited */
1598bool migrate_ratelimited(int node)
1599{
1600 pg_data_t *pgdat = NODE_DATA(node);
1601
1602 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1603 msecs_to_jiffies(pteupdate_interval_millisecs)))
1604 return false;
1605
1606 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1607 return false;
1608
1609 return true;
1610}
1611
Mel Gormanb32967f2012-11-19 12:35:47 +00001612/* Returns true if the node is migrate rate-limited after the update */
Mel Gorman1c30e012014-01-21 15:50:58 -08001613static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1614 unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001615{
Mel Gormanb32967f2012-11-19 12:35:47 +00001616 /*
1617 * Rate-limit the amount of data that is being migrated to a node.
1618 * Optimal placement is no good if the memory bus is saturated and
1619 * all the time is being spent migrating!
1620 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001621 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001622 spin_lock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001623 pgdat->numabalancing_migrate_nr_pages = 0;
1624 pgdat->numabalancing_migrate_next_window = jiffies +
1625 msecs_to_jiffies(migrate_interval_millisecs);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001626 spin_unlock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001627 }
Mel Gormanaf1839d2014-01-21 15:51:01 -08001628 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1629 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1630 nr_pages);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001631 return true;
Mel Gormanaf1839d2014-01-21 15:51:01 -08001632 }
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001633
1634 /*
1635 * This is an unlocked non-atomic update so errors are possible.
1636 * The consequences are failing to migrate when we potentiall should
1637 * have which is not severe enough to warrant locking. If it is ever
1638 * a problem, it can be converted to a per-cpu counter.
1639 */
1640 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1641 return false;
Mel Gormanb32967f2012-11-19 12:35:47 +00001642}
1643
Mel Gorman1c30e012014-01-21 15:50:58 -08001644static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
Mel Gormanb32967f2012-11-19 12:35:47 +00001645{
Hugh Dickins340ef392013-02-22 16:34:33 -08001646 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001647
Sasha Levin309381fea2014-01-23 15:52:54 -08001648 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
Mel Gorman3abef4e2013-02-22 16:34:27 -08001649
Mel Gormanb32967f2012-11-19 12:35:47 +00001650 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001651 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1652 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001653
Hugh Dickins340ef392013-02-22 16:34:33 -08001654 if (isolate_lru_page(page))
1655 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001656
1657 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001658 * migrate_misplaced_transhuge_page() skips page migration's usual
1659 * check on page_count(), so we must do it here, now that the page
1660 * has been isolated: a GUP pin, or any other pin, prevents migration.
1661 * The expected page count is 3: 1 for page's mapcount and 1 for the
1662 * caller's pin and 1 for the reference taken by isolate_lru_page().
1663 */
1664 if (PageTransHuge(page) && page_count(page) != 3) {
1665 putback_lru_page(page);
1666 return 0;
1667 }
1668
1669 page_lru = page_is_file_cache(page);
1670 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1671 hpage_nr_pages(page));
1672
1673 /*
1674 * Isolating the page has taken another reference, so the
1675 * caller's reference can be safely dropped without the page
1676 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001677 */
1678 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001679 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001680}
1681
Mel Gormande466bd2013-12-18 17:08:42 -08001682bool pmd_trans_migrating(pmd_t pmd)
1683{
1684 struct page *page = pmd_page(pmd);
1685 return PageLocked(page);
1686}
1687
1688void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd)
1689{
1690 struct page *page = pmd_page(*pmd);
1691 wait_on_page_locked(page);
1692}
1693
Mel Gormana8f60772012-11-14 21:41:46 +00001694/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001695 * Attempt to migrate a misplaced page to the specified destination
1696 * node. Caller is expected to have an elevated reference count on
1697 * the page that will be dropped by this function before returning.
1698 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001699int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1700 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001701{
Mel Gormana8f60772012-11-14 21:41:46 +00001702 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001703 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001704 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001705 LIST_HEAD(migratepages);
1706
1707 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001708 * Don't migrate file pages that are mapped in multiple processes
1709 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001710 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001711 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1712 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001713 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001714
Mel Gormana8f60772012-11-14 21:41:46 +00001715 /*
1716 * Rate-limit the amount of data that is being migrated to a node.
1717 * Optimal placement is no good if the memory bus is saturated and
1718 * all the time is being spent migrating!
1719 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001720 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001721 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001722
Mel Gormanb32967f2012-11-19 12:35:47 +00001723 isolated = numamigrate_isolate_page(pgdat, page);
1724 if (!isolated)
1725 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001726
Mel Gormanb32967f2012-11-19 12:35:47 +00001727 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001728 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
David Rientjes68711a72014-06-04 16:08:25 -07001729 NULL, node, MIGRATE_ASYNC,
1730 MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001731 if (nr_remaining) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001732 if (!list_empty(&migratepages)) {
1733 list_del(&page->lru);
1734 dec_zone_page_state(page, NR_ISOLATED_ANON +
1735 page_is_file_cache(page));
1736 putback_lru_page(page);
1737 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001738 isolated = 0;
1739 } else
1740 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001741 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001742 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001743
1744out:
1745 put_page(page);
1746 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001747}
Mel Gorman220018d2012-12-05 09:32:56 +00001748#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001749
Mel Gorman220018d2012-12-05 09:32:56 +00001750#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001751/*
1752 * Migrates a THP to a given target node. page must be locked and is unlocked
1753 * before returning.
1754 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001755int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1756 struct vm_area_struct *vma,
1757 pmd_t *pmd, pmd_t entry,
1758 unsigned long address,
1759 struct page *page, int node)
1760{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001761 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001762 pg_data_t *pgdat = NODE_DATA(node);
1763 int isolated = 0;
1764 struct page *new_page = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001765 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001766 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1767 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001768 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001769
1770 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001771 * Rate-limit the amount of data that is being migrated to a node.
1772 * Optimal placement is no good if the memory bus is saturated and
1773 * all the time is being spent migrating!
1774 */
Mel Gormand28d43352012-11-29 09:24:36 +00001775 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001776 goto out_dropref;
1777
1778 new_page = alloc_pages_node(node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001779 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1780 HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001781 if (!new_page)
1782 goto out_fail;
1783
Mel Gormanb32967f2012-11-19 12:35:47 +00001784 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001785 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001786 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001787 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001788 }
1789
Mel Gormanb0943d62013-12-18 17:08:46 -08001790 if (mm_tlb_flush_pending(mm))
1791 flush_tlb_range(vma, mmun_start, mmun_end);
1792
Mel Gormanb32967f2012-11-19 12:35:47 +00001793 /* Prepare a page as a migration target */
1794 __set_page_locked(new_page);
1795 SetPageSwapBacked(new_page);
1796
1797 /* anon mapping, we can simply copy page->mapping to the new page: */
1798 new_page->mapping = page->mapping;
1799 new_page->index = page->index;
1800 migrate_page_copy(new_page, page);
1801 WARN_ON(PageLRU(new_page));
1802
1803 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001804 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001805 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001806 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1807fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001808 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001809 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001810
1811 /* Reverse changes made by migrate_page_copy() */
1812 if (TestClearPageActive(new_page))
1813 SetPageActive(page);
1814 if (TestClearPageUnevictable(new_page))
1815 SetPageUnevictable(page);
1816 mlock_migrate_page(page, new_page);
1817
1818 unlock_page(new_page);
1819 put_page(new_page); /* Free it */
1820
Mel Gormana54a4072013-10-07 11:28:46 +01001821 /* Retake the callers reference and putback on LRU */
1822 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001823 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001824 mod_zone_page_state(page_zone(page),
1825 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001826
1827 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001828 }
1829
Mel Gorman2b4847e2013-12-18 17:08:32 -08001830 orig_entry = *pmd;
Mel Gormanb32967f2012-11-19 12:35:47 +00001831 entry = mk_pmd(new_page, vma->vm_page_prot);
Mel Gormanb32967f2012-11-19 12:35:47 +00001832 entry = pmd_mkhuge(entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001833 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001834
Mel Gorman2b4847e2013-12-18 17:08:32 -08001835 /*
1836 * Clear the old entry under pagetable lock and establish the new PTE.
1837 * Any parallel GUP will either observe the old page blocking on the
1838 * page lock, block on the page table lock or observe the new page.
1839 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1840 * guarantee the copy is visible before the pagetable update.
1841 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001842 flush_cache_range(vma, mmun_start, mmun_end);
Mel Gorman11de9922014-06-04 16:07:41 -07001843 page_add_anon_rmap(new_page, vma, mmun_start);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001844 pmdp_clear_flush_notify(vma, mmun_start, pmd);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001845 set_pmd_at(mm, mmun_start, pmd, entry);
1846 flush_tlb_range(vma, mmun_start, mmun_end);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001847 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001848
1849 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08001850 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1851 flush_tlb_range(vma, mmun_start, mmun_end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001852 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001853 update_mmu_cache_pmd(vma, address, &entry);
1854 page_remove_rmap(new_page);
1855 goto fail_putback;
1856 }
1857
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001858 mem_cgroup_migrate(page, new_page, false);
1859
Mel Gormanb32967f2012-11-19 12:35:47 +00001860 page_remove_rmap(page);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001861
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001862 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001863 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001864
Mel Gorman11de9922014-06-04 16:07:41 -07001865 /* Take an "isolate" reference and put new page on the LRU. */
1866 get_page(new_page);
1867 putback_lru_page(new_page);
1868
Mel Gormanb32967f2012-11-19 12:35:47 +00001869 unlock_page(new_page);
1870 unlock_page(page);
1871 put_page(page); /* Drop the rmap reference */
1872 put_page(page); /* Drop the LRU isolation reference */
1873
1874 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1875 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1876
Mel Gormanb32967f2012-11-19 12:35:47 +00001877 mod_zone_page_state(page_zone(page),
1878 NR_ISOLATED_ANON + page_lru,
1879 -HPAGE_PMD_NR);
1880 return isolated;
1881
Hugh Dickins340ef392013-02-22 16:34:33 -08001882out_fail:
1883 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001884out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08001885 ptl = pmd_lock(mm, pmd);
1886 if (pmd_same(*pmd, entry)) {
1887 entry = pmd_mknonnuma(entry);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001888 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001889 update_mmu_cache_pmd(vma, address, &entry);
1890 }
1891 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001892
Mel Gormaneb4489f62013-12-18 17:08:39 -08001893out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08001894 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001895 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001896 return 0;
1897}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001898#endif /* CONFIG_NUMA_BALANCING */
1899
1900#endif /* CONFIG_NUMA */