blob: bed48809e5d01c14513a1395a12a3c4098341755 [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/*
Hugh Dickins7e09e732014-03-20 21:52:17 -0700181 * Congratulations to trinity for discovering this bug.
182 * mm/fremap.c's remap_file_pages() accepts any range within a single vma to
183 * convert that vma to VM_NONLINEAR; and generic_file_remap_pages() will then
184 * replace the specified range by file ptes throughout (maybe populated after).
185 * If page migration finds a page within that range, while it's still located
186 * by vma_interval_tree rather than lost to i_mmap_nonlinear list, no problem:
187 * zap_pte() clears the temporary migration entry before mmap_sem is dropped.
188 * But if the migrating page is in a part of the vma outside the range to be
189 * remapped, then it will not be cleared, and remove_migration_ptes() needs to
190 * deal with it. Fortunately, this part of the vma is of course still linear,
191 * so we just need to use linear location on the nonlinear list.
192 */
193static int remove_linear_migration_ptes_from_nonlinear(struct page *page,
194 struct address_space *mapping, void *arg)
195{
196 struct vm_area_struct *vma;
197 /* hugetlbfs does not support remap_pages, so no huge pgoff worries */
198 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
199 unsigned long addr;
200
201 list_for_each_entry(vma,
202 &mapping->i_mmap_nonlinear, shared.nonlinear) {
203
204 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
205 if (addr >= vma->vm_start && addr < vma->vm_end)
206 remove_migration_pte(page, vma, addr, arg);
207 }
208 return SWAP_AGAIN;
209}
210
211/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700212 * Get rid of all migration entries and replace them by
213 * references to the indicated page.
214 */
215static void remove_migration_ptes(struct page *old, struct page *new)
216{
Joonsoo Kim051ac832014-01-21 15:49:48 -0800217 struct rmap_walk_control rwc = {
218 .rmap_one = remove_migration_pte,
219 .arg = old,
Hugh Dickins7e09e732014-03-20 21:52:17 -0700220 .file_nonlinear = remove_linear_migration_ptes_from_nonlinear,
Joonsoo Kim051ac832014-01-21 15:49:48 -0800221 };
222
223 rmap_walk(new, &rwc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700224}
225
226/*
Christoph Lameter06972122006-06-23 02:03:35 -0700227 * Something used the pte of a page under migration. We need to
228 * get to the page and wait until migration is finished.
229 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700230 */
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700231static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
232 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700233{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700234 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700235 swp_entry_t entry;
236 struct page *page;
237
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700238 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700239 pte = *ptep;
240 if (!is_swap_pte(pte))
241 goto out;
242
243 entry = pte_to_swp_entry(pte);
244 if (!is_migration_entry(entry))
245 goto out;
246
247 page = migration_entry_to_page(entry);
248
Nick Piggine2867812008-07-25 19:45:30 -0700249 /*
250 * Once radix-tree replacement of page migration started, page_count
251 * *must* be zero. And, we don't want to call wait_on_page_locked()
252 * against a page without get_page().
253 * So, we use get_page_unless_zero(), here. Even failed, page fault
254 * will occur again.
255 */
256 if (!get_page_unless_zero(page))
257 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700258 pte_unmap_unlock(ptep, ptl);
259 wait_on_page_locked(page);
260 put_page(page);
261 return;
262out:
263 pte_unmap_unlock(ptep, ptl);
264}
265
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700266void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
267 unsigned long address)
268{
269 spinlock_t *ptl = pte_lockptr(mm, pmd);
270 pte_t *ptep = pte_offset_map(pmd, address);
271 __migration_entry_wait(mm, ptep, ptl);
272}
273
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800274void migration_entry_wait_huge(struct vm_area_struct *vma,
275 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700276{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800277 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700278 __migration_entry_wait(mm, pte, ptl);
279}
280
Mel Gormanb969c4a2012-01-12 17:19:34 -0800281#ifdef CONFIG_BLOCK
282/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800283static bool buffer_migrate_lock_buffers(struct buffer_head *head,
284 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800285{
286 struct buffer_head *bh = head;
287
288 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800289 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800290 do {
291 get_bh(bh);
292 lock_buffer(bh);
293 bh = bh->b_this_page;
294
295 } while (bh != head);
296
297 return true;
298 }
299
300 /* async case, we cannot block on lock_buffer so use trylock_buffer */
301 do {
302 get_bh(bh);
303 if (!trylock_buffer(bh)) {
304 /*
305 * We failed to lock the buffer and cannot stall in
306 * async migration. Release the taken locks
307 */
308 struct buffer_head *failed_bh = bh;
309 put_bh(failed_bh);
310 bh = head;
311 while (bh != failed_bh) {
312 unlock_buffer(bh);
313 put_bh(bh);
314 bh = bh->b_this_page;
315 }
316 return false;
317 }
318
319 bh = bh->b_this_page;
320 } while (bh != head);
321 return true;
322}
323#else
324static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800325 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800326{
327 return true;
328}
329#endif /* CONFIG_BLOCK */
330
Christoph Lameterb20a3502006-03-22 00:09:12 -0800331/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700332 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700333 *
334 * The number of remaining references must be:
335 * 1 for anonymous pages without a mapping
336 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100337 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800338 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800339int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800340 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500341 struct buffer_head *head, enum migrate_mode mode,
342 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800343{
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500344 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800345 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800346
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700347 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700348 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500349 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700350 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800351 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700352 }
353
Nick Piggin19fd6232008-07-25 19:45:32 -0700354 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800355
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800356 pslot = radix_tree_lookup_slot(&mapping->page_tree,
357 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800358
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500359 expected_count += 1 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700360 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800361 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700362 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700363 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800364 }
365
Nick Piggine2867812008-07-25 19:45:30 -0700366 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700367 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700368 return -EAGAIN;
369 }
370
Christoph Lameterb20a3502006-03-22 00:09:12 -0800371 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800372 * In the async migration case of moving a page with buffers, lock the
373 * buffers using trylock before the mapping is moved. If the mapping
374 * was moved, we later failed to lock the buffers and could not move
375 * the mapping back due to an elevated page count, we would have to
376 * block waiting on other references to be dropped.
377 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800378 if (mode == MIGRATE_ASYNC && head &&
379 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800380 page_unfreeze_refs(page, expected_count);
381 spin_unlock_irq(&mapping->tree_lock);
382 return -EAGAIN;
383 }
384
385 /*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800386 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800387 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800388 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800389 if (PageSwapCache(page)) {
390 SetPageSwapCache(newpage);
391 set_page_private(newpage, page_private(page));
392 }
393
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800394 radix_tree_replace_slot(pslot, newpage);
395
396 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800397 * Drop cache reference from old page by unfreezing
398 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800399 * We know this isn't the last reference.
400 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800401 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800402
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700403 /*
404 * If moved to a different zone then also account
405 * the page for that zone. Other VM counters will be
406 * taken care of when we establish references to the
407 * new page and drop references to the old page.
408 *
409 * Note that anonymous pages are accounted for
410 * via NR_FILE_PAGES and NR_ANON_PAGES if they
411 * are mapped to swap space.
412 */
413 __dec_zone_page_state(page, NR_FILE_PAGES);
414 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700415 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700416 __dec_zone_page_state(page, NR_SHMEM);
417 __inc_zone_page_state(newpage, NR_SHMEM);
418 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700419 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800420
Rafael Aquini78bd5202012-12-11 16:02:31 -0800421 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800422}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800423
424/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900425 * The expected number of remaining references is the same as that
426 * of migrate_page_move_mapping().
427 */
428int migrate_huge_page_move_mapping(struct address_space *mapping,
429 struct page *newpage, struct page *page)
430{
431 int expected_count;
432 void **pslot;
433
434 if (!mapping) {
435 if (page_count(page) != 1)
436 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800437 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900438 }
439
440 spin_lock_irq(&mapping->tree_lock);
441
442 pslot = radix_tree_lookup_slot(&mapping->page_tree,
443 page_index(page));
444
445 expected_count = 2 + page_has_private(page);
446 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800447 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900448 spin_unlock_irq(&mapping->tree_lock);
449 return -EAGAIN;
450 }
451
452 if (!page_freeze_refs(page, expected_count)) {
453 spin_unlock_irq(&mapping->tree_lock);
454 return -EAGAIN;
455 }
456
457 get_page(newpage);
458
459 radix_tree_replace_slot(pslot, newpage);
460
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800461 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900462
463 spin_unlock_irq(&mapping->tree_lock);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800464 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900465}
466
467/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800468 * Gigantic pages are so large that we do not guarantee that page++ pointer
469 * arithmetic will work across the entire page. We need something more
470 * specialized.
471 */
472static void __copy_gigantic_page(struct page *dst, struct page *src,
473 int nr_pages)
474{
475 int i;
476 struct page *dst_base = dst;
477 struct page *src_base = src;
478
479 for (i = 0; i < nr_pages; ) {
480 cond_resched();
481 copy_highpage(dst, src);
482
483 i++;
484 dst = mem_map_next(dst, dst_base, i);
485 src = mem_map_next(src, src_base, i);
486 }
487}
488
489static void copy_huge_page(struct page *dst, struct page *src)
490{
491 int i;
492 int nr_pages;
493
494 if (PageHuge(src)) {
495 /* hugetlbfs page */
496 struct hstate *h = page_hstate(src);
497 nr_pages = pages_per_huge_page(h);
498
499 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
500 __copy_gigantic_page(dst, src, nr_pages);
501 return;
502 }
503 } else {
504 /* thp page */
505 BUG_ON(!PageTransHuge(src));
506 nr_pages = hpage_nr_pages(src);
507 }
508
509 for (i = 0; i < nr_pages; i++) {
510 cond_resched();
511 copy_highpage(dst + i, src + i);
512 }
513}
514
515/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800516 * Copy the page to its new location
517 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900518void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800519{
Rik van Riel7851a452013-10-07 11:29:23 +0100520 int cpupid;
521
Mel Gormanb32967f2012-11-19 12:35:47 +0000522 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900523 copy_huge_page(newpage, page);
524 else
525 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800526
527 if (PageError(page))
528 SetPageError(newpage);
529 if (PageReferenced(page))
530 SetPageReferenced(newpage);
531 if (PageUptodate(page))
532 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700533 if (TestClearPageActive(page)) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800534 VM_BUG_ON_PAGE(PageUnevictable(page), page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800535 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800536 } else if (TestClearPageUnevictable(page))
537 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800538 if (PageChecked(page))
539 SetPageChecked(newpage);
540 if (PageMappedToDisk(page))
541 SetPageMappedToDisk(newpage);
542
543 if (PageDirty(page)) {
544 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700545 /*
546 * Want to mark the page and the radix tree as dirty, and
547 * redo the accounting that clear_page_dirty_for_io undid,
548 * but we can't use set_page_dirty because that function
549 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300550 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700551 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700552 if (PageSwapBacked(page))
553 SetPageDirty(newpage);
554 else
555 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800556 }
557
Rik van Riel7851a452013-10-07 11:29:23 +0100558 /*
559 * Copy NUMA information to the new page, to prevent over-eager
560 * future migrations of this same page.
561 */
562 cpupid = page_cpupid_xchg_last(page, -1);
563 page_cpupid_xchg_last(newpage, cpupid);
564
Nick Pigginb291f002008-10-18 20:26:44 -0700565 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800566 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800567 /*
568 * Please do not reorder this without considering how mm/ksm.c's
569 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
570 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800571 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800572 ClearPagePrivate(page);
573 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800574
575 /*
576 * If any waiters have accumulated on the new page then
577 * wake them up.
578 */
579 if (PageWriteback(newpage))
580 end_page_writeback(newpage);
581}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800582
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700583/************************************************************
584 * Migration functions
585 ***********************************************************/
586
Christoph Lameterb20a3502006-03-22 00:09:12 -0800587/*
588 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100589 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800590 *
591 * Pages are locked upon entry and exit.
592 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700593int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800594 struct page *newpage, struct page *page,
595 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800596{
597 int rc;
598
599 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
600
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500601 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800602
Rafael Aquini78bd5202012-12-11 16:02:31 -0800603 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800604 return rc;
605
606 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800607 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800608}
609EXPORT_SYMBOL(migrate_page);
610
David Howells93614012006-09-30 20:45:40 +0200611#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800612/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700613 * Migration function for pages with buffers. This function can only be used
614 * if the underlying filesystem guarantees that no other references to "page"
615 * exist.
616 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700617int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800618 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700619{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700620 struct buffer_head *bh, *head;
621 int rc;
622
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700623 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800624 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700625
626 head = page_buffers(page);
627
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500628 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700629
Rafael Aquini78bd5202012-12-11 16:02:31 -0800630 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700631 return rc;
632
Mel Gormanb969c4a2012-01-12 17:19:34 -0800633 /*
634 * In the async case, migrate_page_move_mapping locked the buffers
635 * with an IRQ-safe spinlock held. In the sync case, the buffers
636 * need to be locked now
637 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800638 if (mode != MIGRATE_ASYNC)
639 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700640
641 ClearPagePrivate(page);
642 set_page_private(newpage, page_private(page));
643 set_page_private(page, 0);
644 put_page(page);
645 get_page(newpage);
646
647 bh = head;
648 do {
649 set_bh_page(bh, newpage, bh_offset(bh));
650 bh = bh->b_this_page;
651
652 } while (bh != head);
653
654 SetPagePrivate(newpage);
655
656 migrate_page_copy(newpage, page);
657
658 bh = head;
659 do {
660 unlock_buffer(bh);
661 put_bh(bh);
662 bh = bh->b_this_page;
663
664 } while (bh != head);
665
Rafael Aquini78bd5202012-12-11 16:02:31 -0800666 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700667}
668EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200669#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700670
Christoph Lameter04e62a22006-06-23 02:03:38 -0700671/*
672 * Writeback a page to clean the dirty state
673 */
674static int writeout(struct address_space *mapping, struct page *page)
675{
676 struct writeback_control wbc = {
677 .sync_mode = WB_SYNC_NONE,
678 .nr_to_write = 1,
679 .range_start = 0,
680 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700681 .for_reclaim = 1
682 };
683 int rc;
684
685 if (!mapping->a_ops->writepage)
686 /* No write method for the address space */
687 return -EINVAL;
688
689 if (!clear_page_dirty_for_io(page))
690 /* Someone else already triggered a write */
691 return -EAGAIN;
692
693 /*
694 * A dirty page may imply that the underlying filesystem has
695 * the page on some queue. So the page must be clean for
696 * migration. Writeout may mean we loose the lock and the
697 * page state is no longer what we checked for earlier.
698 * At this point we know that the migration attempt cannot
699 * be successful.
700 */
701 remove_migration_ptes(page, page);
702
703 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700704
705 if (rc != AOP_WRITEPAGE_ACTIVATE)
706 /* unlocked. Relock */
707 lock_page(page);
708
Hugh Dickinsbda85502008-11-19 15:36:36 -0800709 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700710}
711
712/*
713 * Default handling if a filesystem does not provide a migration function.
714 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700715static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800716 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700717{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800718 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800719 /* Only writeback pages in full synchronous migration */
720 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800721 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700722 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800723 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700724
725 /*
726 * Buffers may be managed in a filesystem specific way.
727 * We must have no buffers or drop them.
728 */
David Howells266cf652009-04-03 16:42:36 +0100729 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700730 !try_to_release_page(page, GFP_KERNEL))
731 return -EAGAIN;
732
Mel Gormana6bc32b2012-01-12 17:19:43 -0800733 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700734}
735
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700736/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700737 * Move a page to a newly allocated page
738 * The page is locked and all ptes have been successfully removed.
739 *
740 * The new page will have replaced the old page if this function
741 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700742 *
743 * Return value:
744 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800745 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700746 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700747static int move_to_new_page(struct page *newpage, struct page *page,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800748 int remap_swapcache, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700749{
750 struct address_space *mapping;
751 int rc;
752
753 /*
754 * Block others from accessing the page when we get around to
755 * establishing additional references. We are the only one
756 * holding a reference to the new page at this point.
757 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200758 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700759 BUG();
760
761 /* Prepare mapping for the new page.*/
762 newpage->index = page->index;
763 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700764 if (PageSwapBacked(page))
765 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700766
767 mapping = page_mapping(page);
768 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800769 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800770 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700771 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800772 * Most pages have a mapping and most filesystems provide a
773 * migratepage callback. Anonymous pages are part of swap
774 * space which also has its own migratepage callback. This
775 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700776 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800777 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800778 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800779 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800780 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700781
Rafael Aquini78bd5202012-12-11 16:02:31 -0800782 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700783 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700784 } else {
785 if (remap_swapcache)
786 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800787 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700788 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700789
790 unlock_page(newpage);
791
792 return rc;
793}
794
Minchan Kim0dabec92011-10-31 17:06:57 -0700795static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800796 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700797{
Minchan Kim0dabec92011-10-31 17:06:57 -0700798 int rc = -EAGAIN;
Mel Gorman3fe20112010-05-24 14:32:20 -0700799 int remap_swapcache = 1;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -0700800 struct mem_cgroup *mem;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700801 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700802
Nick Piggin529ae9a2008-08-02 12:01:03 +0200803 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800804 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700805 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800806
807 /*
808 * It's not safe for direct compaction to call lock_page.
809 * For example, during page readahead pages are added locked
810 * to the LRU. Later, when the IO completes the pages are
811 * marked uptodate and unlocked. However, the queueing
812 * could be merging multiple pages for one bio (e.g.
813 * mpage_readpages). If an allocation happens for the
814 * second or third page, the process can end up locking
815 * the same page twice and deadlocking. Rather than
816 * trying to be clever about what pages can be locked,
817 * avoid the use of lock_page for direct compaction
818 * altogether.
819 */
820 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700821 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800822
Christoph Lametere24f0b82006-06-23 02:03:51 -0700823 lock_page(page);
824 }
825
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800826 /* charge against new page */
Johannes Weiner0030f532012-07-31 16:45:25 -0700827 mem_cgroup_prepare_migration(page, newpage, &mem);
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800828
Christoph Lametere24f0b82006-06-23 02:03:51 -0700829 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700830 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700831 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800832 * necessary to wait for PageWriteback. In the async case,
833 * the retry loop is too short and in the sync-light case,
834 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700835 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800836 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700837 rc = -EBUSY;
838 goto uncharge;
839 }
840 if (!force)
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800841 goto uncharge;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700842 wait_on_page_writeback(page);
843 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700844 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700845 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
846 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800847 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700848 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700849 * File Caches may use write_page() or lock_page() in migration, then,
850 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700851 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800852 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800853 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000854 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800855 * getting a hold on an anon_vma from outside one of its mms.
856 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700857 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800858 if (anon_vma) {
859 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700860 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800861 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800862 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700863 /*
864 * We cannot be sure that the anon_vma of an unmapped
865 * swapcache page is safe to use because we don't
866 * know in advance if the VMA that this page belonged
867 * to still exists. If the VMA and others sharing the
868 * data have been freed, then the anon_vma could
869 * already be invalid.
870 *
871 * To avoid this possibility, swapcache pages get
872 * migrated but are not remapped when migration
873 * completes
874 */
875 remap_swapcache = 0;
876 } else {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800877 goto uncharge;
Mel Gorman3fe20112010-05-24 14:32:20 -0700878 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700879 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800880
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800881 if (unlikely(balloon_page_movable(page))) {
882 /*
883 * A ballooned page does not need any special attention from
884 * physical to virtual reverse mapping procedures.
885 * Skip any attempt to unmap PTEs or to remap swap cache,
886 * in order to avoid burning cycles at rmap level, and perform
887 * the page migration right away (proteced by page lock).
888 */
889 rc = balloon_page_migrate(newpage, page, mode);
890 goto uncharge;
891 }
892
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700893 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800894 * Corner case handling:
895 * 1. When a new swap-cache page is read into, it is added to the LRU
896 * and treated as swapcache but it has no rmap yet.
897 * Calling try_to_unmap() against a page->mapping==NULL page will
898 * trigger a BUG. So handle it here.
899 * 2. An orphaned page (see truncate_complete_page) might have
900 * fs-private metadata. The page can be picked up due to memory
901 * offlining. Everywhere else except page reclaim, the page is
902 * invisible to the vm, so the page can not be migrated. So try to
903 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700904 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800905 if (!page->mapping) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800906 VM_BUG_ON_PAGE(PageAnon(page), page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800907 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800908 try_to_free_buffers(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800909 goto uncharge;
Shaohua Li62e1c552008-02-04 22:29:33 -0800910 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700911 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800912 }
913
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700914 /* Establish migration ptes or remove ptes */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200915 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700916
Shaohua Liabfc3482009-09-21 17:01:19 -0700917skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700918 if (!page_mapped(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800919 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700920
Mel Gorman3fe20112010-05-24 14:32:20 -0700921 if (rc && remap_swapcache)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700922 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700923
924 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700925 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700926 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700927
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800928uncharge:
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800929 mem_cgroup_end_migration(mem, page, newpage,
930 (rc == MIGRATEPAGE_SUCCESS ||
931 rc == MIGRATEPAGE_BALLOON_SUCCESS));
Christoph Lametere24f0b82006-06-23 02:03:51 -0700932 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700933out:
934 return rc;
935}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700936
Minchan Kim0dabec92011-10-31 17:06:57 -0700937/*
938 * Obtain the lock on page, remove all ptes and migrate the page
939 * to the newly allocated page in newpage.
940 */
941static int unmap_and_move(new_page_t get_new_page, unsigned long private,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800942 struct page *page, int force, enum migrate_mode mode)
Minchan Kim0dabec92011-10-31 17:06:57 -0700943{
944 int rc = 0;
945 int *result = NULL;
946 struct page *newpage = get_new_page(page, private, &result);
947
948 if (!newpage)
949 return -ENOMEM;
950
951 if (page_count(page) == 1) {
952 /* page was freed from under us. So we are done. */
953 goto out;
954 }
955
956 if (unlikely(PageTransHuge(page)))
957 if (unlikely(split_huge_page(page)))
958 goto out;
959
Hugh Dickins9c620e22013-02-22 16:35:14 -0800960 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800961
962 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
963 /*
964 * A ballooned page has been migrated already.
965 * Now, it's the time to wrap-up counters,
966 * handle the page back to Buddy and return.
967 */
968 dec_zone_page_state(page, NR_ISOLATED_ANON +
969 page_is_file_cache(page));
970 balloon_page_free(page);
971 return MIGRATEPAGE_SUCCESS;
972 }
Minchan Kim0dabec92011-10-31 17:06:57 -0700973out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700974 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700975 /*
976 * A page that has been migrated has all references
977 * removed and will be freed. A page that has not been
978 * migrated will have kepts its references and be
979 * restored.
980 */
981 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700982 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700983 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700984 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700985 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700986 /*
987 * Move the new page to the LRU. If migration was not successful
988 * then this will free the page.
989 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700990 putback_lru_page(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700991 if (result) {
992 if (rc)
993 *result = rc;
994 else
995 *result = page_to_nid(newpage);
996 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700997 return rc;
998}
999
1000/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001001 * Counterpart of unmap_and_move_page() for hugepage migration.
1002 *
1003 * This function doesn't wait the completion of hugepage I/O
1004 * because there is no race between I/O and migration for hugepage.
1005 * Note that currently hugepage I/O occurs only in direct I/O
1006 * where no lock is held and PG_writeback is irrelevant,
1007 * and writeback status of all subpages are counted in the reference
1008 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1009 * under direct I/O, the reference of the head page is 512 and a bit more.)
1010 * This means that when we try to migrate hugepage whose subpages are
1011 * doing direct I/O, some references remain after try_to_unmap() and
1012 * hugepage migration fails without data corruption.
1013 *
1014 * There is also no race when direct I/O is issued on the page under migration,
1015 * because then pte is replaced with migration swap entry and direct I/O code
1016 * will wait in the page fault for migration to complete.
1017 */
1018static int unmap_and_move_huge_page(new_page_t get_new_page,
1019 unsigned long private, struct page *hpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001020 int force, enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001021{
1022 int rc = 0;
1023 int *result = NULL;
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 */
Joonsoo Kim32665f22014-01-21 15:51:15 -08001034 if (!hugepage_migration_support(page_hstate(hpage))) {
1035 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
1054 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1055
1056 if (!page_mapped(hpage))
Mel Gormana6bc32b2012-01-12 17:19:43 -08001057 rc = move_to_new_page(new_hpage, hpage, 1, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001058
1059 if (rc)
1060 remove_migration_ptes(hpage, hpage);
1061
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001062 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001063 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001064
1065 if (!rc)
1066 hugetlb_cgroup_migrate(hpage, new_hpage);
1067
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001068 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001069out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001070 if (rc != -EAGAIN)
1071 putback_active_hugepage(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001072 put_page(new_hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001073 if (result) {
1074 if (rc)
1075 *result = rc;
1076 else
1077 *result = page_to_nid(new_hpage);
1078 }
1079 return rc;
1080}
1081
1082/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001083 * migrate_pages - migrate the pages specified in a list, to the free pages
1084 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001085 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001086 * @from: The list of pages to be migrated.
1087 * @get_new_page: The function used to allocate free pages to be used
1088 * as the target of the page migration.
1089 * @private: Private data to be passed on to get_new_page()
1090 * @mode: The migration mode that specifies the constraints for
1091 * page migration, if any.
1092 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001093 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001094 * The function returns after 10 attempts or if no pages are movable any more
1095 * because the list has become empty or no retryable pages exist any more.
1096 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001097 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001098 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001099 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001100 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001101int migrate_pages(struct list_head *from, new_page_t get_new_page,
1102 unsigned long private, enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001103{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001104 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001105 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001106 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001107 int pass = 0;
1108 struct page *page;
1109 struct page *page2;
1110 int swapwrite = current->flags & PF_SWAPWRITE;
1111 int rc;
1112
1113 if (!swapwrite)
1114 current->flags |= PF_SWAPWRITE;
1115
Christoph Lametere24f0b82006-06-23 02:03:51 -07001116 for(pass = 0; pass < 10 && retry; pass++) {
1117 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001118
Christoph Lametere24f0b82006-06-23 02:03:51 -07001119 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001120 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001121
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001122 if (PageHuge(page))
1123 rc = unmap_and_move_huge_page(get_new_page,
1124 private, page, pass > 2, mode);
1125 else
1126 rc = unmap_and_move(get_new_page, private,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001127 page, pass > 2, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001128
Christoph Lametere24f0b82006-06-23 02:03:51 -07001129 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001130 case -ENOMEM:
1131 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001132 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001133 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001134 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001135 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001136 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001137 break;
1138 default:
Naoya Horiguchi354a3362014-01-21 15:51:14 -08001139 /*
1140 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1141 * unlike -EAGAIN case, the failed page is
1142 * removed from migration page list and not
1143 * retried in the next outer loop.
1144 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001145 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001146 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001147 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001148 }
1149 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001150 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001151out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001152 if (nr_succeeded)
1153 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1154 if (nr_failed)
1155 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001156 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1157
Christoph Lameterb20a3502006-03-22 00:09:12 -08001158 if (!swapwrite)
1159 current->flags &= ~PF_SWAPWRITE;
1160
Rafael Aquini78bd5202012-12-11 16:02:31 -08001161 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001162}
1163
Christoph Lameter742755a2006-06-23 02:03:55 -07001164#ifdef CONFIG_NUMA
1165/*
1166 * Move a list of individual pages
1167 */
1168struct page_to_node {
1169 unsigned long addr;
1170 struct page *page;
1171 int node;
1172 int status;
1173};
1174
1175static struct page *new_page_node(struct page *p, unsigned long private,
1176 int **result)
1177{
1178 struct page_to_node *pm = (struct page_to_node *)private;
1179
1180 while (pm->node != MAX_NUMNODES && pm->page != p)
1181 pm++;
1182
1183 if (pm->node == MAX_NUMNODES)
1184 return NULL;
1185
1186 *result = &pm->status;
1187
Naoya Horiguchie632a932013-09-11 14:22:04 -07001188 if (PageHuge(p))
1189 return alloc_huge_page_node(page_hstate(compound_head(p)),
1190 pm->node);
1191 else
1192 return alloc_pages_exact_node(pm->node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001193 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001194}
1195
1196/*
1197 * Move a set of pages as indicated in the pm array. The addr
1198 * field must be set to the virtual address of the page to be moved
1199 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001200 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001201 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001202static int do_move_page_to_node_array(struct mm_struct *mm,
1203 struct page_to_node *pm,
1204 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001205{
1206 int err;
1207 struct page_to_node *pp;
1208 LIST_HEAD(pagelist);
1209
1210 down_read(&mm->mmap_sem);
1211
1212 /*
1213 * Build a list of pages to migrate
1214 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001215 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1216 struct vm_area_struct *vma;
1217 struct page *page;
1218
Christoph Lameter742755a2006-06-23 02:03:55 -07001219 err = -EFAULT;
1220 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001221 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001222 goto set_status;
1223
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001224 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001225
1226 err = PTR_ERR(page);
1227 if (IS_ERR(page))
1228 goto set_status;
1229
Christoph Lameter742755a2006-06-23 02:03:55 -07001230 err = -ENOENT;
1231 if (!page)
1232 goto set_status;
1233
Hugh Dickins62b61f62009-12-14 17:59:33 -08001234 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001235 if (PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001236 goto put_and_set;
1237
1238 pp->page = page;
1239 err = page_to_nid(page);
1240
1241 if (err == pp->node)
1242 /*
1243 * Node already in the right place
1244 */
1245 goto put_and_set;
1246
1247 err = -EACCES;
1248 if (page_mapcount(page) > 1 &&
1249 !migrate_all)
1250 goto put_and_set;
1251
Naoya Horiguchie632a932013-09-11 14:22:04 -07001252 if (PageHuge(page)) {
1253 isolate_huge_page(page, &pagelist);
1254 goto put_and_set;
1255 }
1256
Nick Piggin62695a82008-10-18 20:26:09 -07001257 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001258 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001259 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001260 inc_zone_page_state(page, NR_ISOLATED_ANON +
1261 page_is_file_cache(page));
1262 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001263put_and_set:
1264 /*
1265 * Either remove the duplicate refcount from
1266 * isolate_lru_page() or drop the page ref if it was
1267 * not isolated.
1268 */
1269 put_page(page);
1270set_status:
1271 pp->status = err;
1272 }
1273
Brice Gogline78bbfa2008-10-18 20:27:15 -07001274 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001275 if (!list_empty(&pagelist)) {
Christoph Lameter742755a2006-06-23 02:03:55 -07001276 err = migrate_pages(&pagelist, new_page_node,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001277 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001278 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001279 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001280 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001281
1282 up_read(&mm->mmap_sem);
1283 return err;
1284}
1285
1286/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001287 * Migrate an array of page address onto an array of nodes and fill
1288 * the corresponding array of status.
1289 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001290static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001291 unsigned long nr_pages,
1292 const void __user * __user *pages,
1293 const int __user *nodes,
1294 int __user *status, int flags)
1295{
Brice Goglin3140a222009-01-06 14:38:57 -08001296 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001297 unsigned long chunk_nr_pages;
1298 unsigned long chunk_start;
1299 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001300
Brice Goglin3140a222009-01-06 14:38:57 -08001301 err = -ENOMEM;
1302 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1303 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001304 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001305
1306 migrate_prep();
1307
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001308 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001309 * Store a chunk of page_to_node array in a page,
1310 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001311 */
Brice Goglin3140a222009-01-06 14:38:57 -08001312 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001313
Brice Goglin3140a222009-01-06 14:38:57 -08001314 for (chunk_start = 0;
1315 chunk_start < nr_pages;
1316 chunk_start += chunk_nr_pages) {
1317 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001318
Brice Goglin3140a222009-01-06 14:38:57 -08001319 if (chunk_start + chunk_nr_pages > nr_pages)
1320 chunk_nr_pages = nr_pages - chunk_start;
1321
1322 /* fill the chunk pm with addrs and nodes from user-space */
1323 for (j = 0; j < chunk_nr_pages; j++) {
1324 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001325 int node;
1326
Brice Goglin3140a222009-01-06 14:38:57 -08001327 err = -EFAULT;
1328 if (get_user(p, pages + j + chunk_start))
1329 goto out_pm;
1330 pm[j].addr = (unsigned long) p;
1331
1332 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001333 goto out_pm;
1334
1335 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001336 if (node < 0 || node >= MAX_NUMNODES)
1337 goto out_pm;
1338
Lai Jiangshan389162c2012-12-12 13:51:30 -08001339 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001340 goto out_pm;
1341
1342 err = -EACCES;
1343 if (!node_isset(node, task_nodes))
1344 goto out_pm;
1345
Brice Goglin3140a222009-01-06 14:38:57 -08001346 pm[j].node = node;
1347 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001348
Brice Goglin3140a222009-01-06 14:38:57 -08001349 /* End marker for this chunk */
1350 pm[chunk_nr_pages].node = MAX_NUMNODES;
1351
1352 /* Migrate this chunk */
1353 err = do_move_page_to_node_array(mm, pm,
1354 flags & MPOL_MF_MOVE_ALL);
1355 if (err < 0)
1356 goto out_pm;
1357
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001358 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001359 for (j = 0; j < chunk_nr_pages; j++)
1360 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001361 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001362 goto out_pm;
1363 }
1364 }
1365 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001366
1367out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001368 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001369out:
1370 return err;
1371}
1372
1373/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001374 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001375 */
Brice Goglin80bba122008-12-09 13:14:23 -08001376static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1377 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001378{
Brice Goglin2f007e72008-10-18 20:27:16 -07001379 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001380
Christoph Lameter742755a2006-06-23 02:03:55 -07001381 down_read(&mm->mmap_sem);
1382
Brice Goglin2f007e72008-10-18 20:27:16 -07001383 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001384 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001385 struct vm_area_struct *vma;
1386 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001387 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001388
1389 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001390 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001391 goto set_status;
1392
Brice Goglin2f007e72008-10-18 20:27:16 -07001393 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001394
1395 err = PTR_ERR(page);
1396 if (IS_ERR(page))
1397 goto set_status;
1398
Christoph Lameter742755a2006-06-23 02:03:55 -07001399 err = -ENOENT;
1400 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001401 if (!page || PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001402 goto set_status;
1403
1404 err = page_to_nid(page);
1405set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001406 *status = err;
1407
1408 pages++;
1409 status++;
1410 }
1411
1412 up_read(&mm->mmap_sem);
1413}
1414
1415/*
1416 * Determine the nodes of a user array of pages and store it in
1417 * a user array of status.
1418 */
1419static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1420 const void __user * __user *pages,
1421 int __user *status)
1422{
1423#define DO_PAGES_STAT_CHUNK_NR 16
1424 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1425 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001426
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001427 while (nr_pages) {
1428 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001429
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001430 chunk_nr = nr_pages;
1431 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1432 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1433
1434 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1435 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001436
1437 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1438
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001439 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1440 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001441
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001442 pages += chunk_nr;
1443 status += chunk_nr;
1444 nr_pages -= chunk_nr;
1445 }
1446 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001447}
1448
1449/*
1450 * Move a list of pages in the address space of the currently executing
1451 * process.
1452 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001453SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1454 const void __user * __user *, pages,
1455 const int __user *, nodes,
1456 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001457{
David Howellsc69e8d92008-11-14 10:39:19 +11001458 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001459 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001460 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001461 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001462 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001463
1464 /* Check flags */
1465 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1466 return -EINVAL;
1467
1468 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1469 return -EPERM;
1470
1471 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001472 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001473 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001474 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001475 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001476 return -ESRCH;
1477 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001478 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001479
1480 /*
1481 * Check if this process has the right to modify the specified
1482 * process. The right exists if the process has administrative
1483 * capabilities, superuser privileges or the same
1484 * userid as the target process.
1485 */
David Howellsc69e8d92008-11-14 10:39:19 +11001486 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001487 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1488 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001489 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001490 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001491 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001492 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001493 }
David Howellsc69e8d92008-11-14 10:39:19 +11001494 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001495
David Quigley86c3a762006-06-23 02:04:02 -07001496 err = security_task_movememory(task);
1497 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001498 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001499
Christoph Lameter3268c632012-03-21 16:34:06 -07001500 task_nodes = cpuset_mems_allowed(task);
1501 mm = get_task_mm(task);
1502 put_task_struct(task);
1503
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001504 if (!mm)
1505 return -EINVAL;
1506
1507 if (nodes)
1508 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1509 nodes, status, flags);
1510 else
1511 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001512
1513 mmput(mm);
1514 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001515
Christoph Lameter742755a2006-06-23 02:03:55 -07001516out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001517 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001518 return err;
1519}
Christoph Lameter742755a2006-06-23 02:03:55 -07001520
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001521/*
1522 * Call migration functions in the vma_ops that may prepare
1523 * memory in a vm for migration. migration functions may perform
1524 * the migration for vmas that do not have an underlying page struct.
1525 */
1526int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1527 const nodemask_t *from, unsigned long flags)
1528{
1529 struct vm_area_struct *vma;
1530 int err = 0;
1531
Daisuke Nishimura1001c9f2009-02-11 13:04:18 -08001532 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001533 if (vma->vm_ops && vma->vm_ops->migrate) {
1534 err = vma->vm_ops->migrate(vma, to, from, flags);
1535 if (err)
1536 break;
1537 }
1538 }
1539 return err;
1540}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001541
1542#ifdef CONFIG_NUMA_BALANCING
1543/*
1544 * Returns true if this is a safe migration target node for misplaced NUMA
1545 * pages. Currently it only checks the watermarks which crude
1546 */
1547static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001548 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001549{
1550 int z;
1551 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1552 struct zone *zone = pgdat->node_zones + z;
1553
1554 if (!populated_zone(zone))
1555 continue;
1556
Lisa Du6e543d52013-09-11 14:22:36 -07001557 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001558 continue;
1559
1560 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1561 if (!zone_watermark_ok(zone, 0,
1562 high_wmark_pages(zone) +
1563 nr_migrate_pages,
1564 0, 0))
1565 continue;
1566 return true;
1567 }
1568 return false;
1569}
1570
1571static struct page *alloc_misplaced_dst_page(struct page *page,
1572 unsigned long data,
1573 int **result)
1574{
1575 int nid = (int) data;
1576 struct page *newpage;
1577
1578 newpage = alloc_pages_exact_node(nid,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001579 (GFP_HIGHUSER_MOVABLE |
1580 __GFP_THISNODE | __GFP_NOMEMALLOC |
1581 __GFP_NORETRY | __GFP_NOWARN) &
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001582 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001583
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001584 return newpage;
1585}
1586
1587/*
Mel Gormana8f60772012-11-14 21:41:46 +00001588 * page migration rate limiting control.
1589 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1590 * window of time. Default here says do not migrate more than 1280M per second.
Mel Gormane14808b2012-11-19 10:59:15 +00001591 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1592 * as it is faults that reset the window, pte updates will happen unconditionally
1593 * if there has not been a fault since @pteupdate_interval_millisecs after the
1594 * throttle window closed.
Mel Gormana8f60772012-11-14 21:41:46 +00001595 */
1596static unsigned int migrate_interval_millisecs __read_mostly = 100;
Mel Gormane14808b2012-11-19 10:59:15 +00001597static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
Mel Gormana8f60772012-11-14 21:41:46 +00001598static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1599
Mel Gormane14808b2012-11-19 10:59:15 +00001600/* Returns true if NUMA migration is currently rate limited */
1601bool migrate_ratelimited(int node)
1602{
1603 pg_data_t *pgdat = NODE_DATA(node);
1604
1605 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1606 msecs_to_jiffies(pteupdate_interval_millisecs)))
1607 return false;
1608
1609 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1610 return false;
1611
1612 return true;
1613}
1614
Mel Gormanb32967f2012-11-19 12:35:47 +00001615/* Returns true if the node is migrate rate-limited after the update */
Mel Gorman1c30e012014-01-21 15:50:58 -08001616static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1617 unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001618{
Mel Gormanb32967f2012-11-19 12:35:47 +00001619 /*
1620 * Rate-limit the amount of data that is being migrated to a node.
1621 * Optimal placement is no good if the memory bus is saturated and
1622 * all the time is being spent migrating!
1623 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001624 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001625 spin_lock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001626 pgdat->numabalancing_migrate_nr_pages = 0;
1627 pgdat->numabalancing_migrate_next_window = jiffies +
1628 msecs_to_jiffies(migrate_interval_millisecs);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001629 spin_unlock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001630 }
Mel Gormanaf1839d2014-01-21 15:51:01 -08001631 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1632 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1633 nr_pages);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001634 return true;
Mel Gormanaf1839d2014-01-21 15:51:01 -08001635 }
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001636
1637 /*
1638 * This is an unlocked non-atomic update so errors are possible.
1639 * The consequences are failing to migrate when we potentiall should
1640 * have which is not severe enough to warrant locking. If it is ever
1641 * a problem, it can be converted to a per-cpu counter.
1642 */
1643 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1644 return false;
Mel Gormanb32967f2012-11-19 12:35:47 +00001645}
1646
Mel Gorman1c30e012014-01-21 15:50:58 -08001647static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
Mel Gormanb32967f2012-11-19 12:35:47 +00001648{
Hugh Dickins340ef392013-02-22 16:34:33 -08001649 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001650
Sasha Levin309381fea2014-01-23 15:52:54 -08001651 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
Mel Gorman3abef4e2013-02-22 16:34:27 -08001652
Mel Gormanb32967f2012-11-19 12:35:47 +00001653 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001654 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1655 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001656
Hugh Dickins340ef392013-02-22 16:34:33 -08001657 if (isolate_lru_page(page))
1658 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001659
1660 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001661 * migrate_misplaced_transhuge_page() skips page migration's usual
1662 * check on page_count(), so we must do it here, now that the page
1663 * has been isolated: a GUP pin, or any other pin, prevents migration.
1664 * The expected page count is 3: 1 for page's mapcount and 1 for the
1665 * caller's pin and 1 for the reference taken by isolate_lru_page().
1666 */
1667 if (PageTransHuge(page) && page_count(page) != 3) {
1668 putback_lru_page(page);
1669 return 0;
1670 }
1671
1672 page_lru = page_is_file_cache(page);
1673 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1674 hpage_nr_pages(page));
1675
1676 /*
1677 * Isolating the page has taken another reference, so the
1678 * caller's reference can be safely dropped without the page
1679 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001680 */
1681 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001682 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001683}
1684
Mel Gormande466bd2013-12-18 17:08:42 -08001685bool pmd_trans_migrating(pmd_t pmd)
1686{
1687 struct page *page = pmd_page(pmd);
1688 return PageLocked(page);
1689}
1690
1691void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd)
1692{
1693 struct page *page = pmd_page(*pmd);
1694 wait_on_page_locked(page);
1695}
1696
Mel Gormana8f60772012-11-14 21:41:46 +00001697/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001698 * Attempt to migrate a misplaced page to the specified destination
1699 * node. Caller is expected to have an elevated reference count on
1700 * the page that will be dropped by this function before returning.
1701 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001702int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1703 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001704{
Mel Gormana8f60772012-11-14 21:41:46 +00001705 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001706 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001707 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001708 LIST_HEAD(migratepages);
1709
1710 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001711 * Don't migrate file pages that are mapped in multiple processes
1712 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001713 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001714 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1715 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001716 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001717
Mel Gormana8f60772012-11-14 21:41:46 +00001718 /*
1719 * Rate-limit the amount of data that is being migrated to a node.
1720 * Optimal placement is no good if the memory bus is saturated and
1721 * all the time is being spent migrating!
1722 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001723 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001724 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001725
Mel Gormanb32967f2012-11-19 12:35:47 +00001726 isolated = numamigrate_isolate_page(pgdat, page);
1727 if (!isolated)
1728 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001729
Mel Gormanb32967f2012-11-19 12:35:47 +00001730 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001731 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1732 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001733 if (nr_remaining) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001734 if (!list_empty(&migratepages)) {
1735 list_del(&page->lru);
1736 dec_zone_page_state(page, NR_ISOLATED_ANON +
1737 page_is_file_cache(page));
1738 putback_lru_page(page);
1739 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001740 isolated = 0;
1741 } else
1742 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001743 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001744 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001745
1746out:
1747 put_page(page);
1748 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001749}
Mel Gorman220018d2012-12-05 09:32:56 +00001750#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001751
Mel Gorman220018d2012-12-05 09:32:56 +00001752#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001753/*
1754 * Migrates a THP to a given target node. page must be locked and is unlocked
1755 * before returning.
1756 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001757int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1758 struct vm_area_struct *vma,
1759 pmd_t *pmd, pmd_t entry,
1760 unsigned long address,
1761 struct page *page, int node)
1762{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001763 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001764 pg_data_t *pgdat = NODE_DATA(node);
1765 int isolated = 0;
1766 struct page *new_page = NULL;
1767 struct mem_cgroup *memcg = NULL;
1768 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001769 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1770 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001771 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001772
1773 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001774 * Rate-limit the amount of data that is being migrated to a node.
1775 * Optimal placement is no good if the memory bus is saturated and
1776 * all the time is being spent migrating!
1777 */
Mel Gormand28d43352012-11-29 09:24:36 +00001778 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001779 goto out_dropref;
1780
1781 new_page = alloc_pages_node(node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001782 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1783 HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001784 if (!new_page)
1785 goto out_fail;
1786
Mel Gormanb32967f2012-11-19 12:35:47 +00001787 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001788 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001789 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001790 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001791 }
1792
Mel Gormanb0943d62013-12-18 17:08:46 -08001793 if (mm_tlb_flush_pending(mm))
1794 flush_tlb_range(vma, mmun_start, mmun_end);
1795
Mel Gormanb32967f2012-11-19 12:35:47 +00001796 /* Prepare a page as a migration target */
1797 __set_page_locked(new_page);
1798 SetPageSwapBacked(new_page);
1799
1800 /* anon mapping, we can simply copy page->mapping to the new page: */
1801 new_page->mapping = page->mapping;
1802 new_page->index = page->index;
1803 migrate_page_copy(new_page, page);
1804 WARN_ON(PageLRU(new_page));
1805
1806 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001807 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001808 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001809 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1810fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001811 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001812 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001813
1814 /* Reverse changes made by migrate_page_copy() */
1815 if (TestClearPageActive(new_page))
1816 SetPageActive(page);
1817 if (TestClearPageUnevictable(new_page))
1818 SetPageUnevictable(page);
1819 mlock_migrate_page(page, new_page);
1820
1821 unlock_page(new_page);
1822 put_page(new_page); /* Free it */
1823
Mel Gormana54a4072013-10-07 11:28:46 +01001824 /* Retake the callers reference and putback on LRU */
1825 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001826 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001827 mod_zone_page_state(page_zone(page),
1828 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001829
1830 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001831 }
1832
1833 /*
1834 * Traditional migration needs to prepare the memcg charge
1835 * transaction early to prevent the old page from being
1836 * uncharged when installing migration entries. Here we can
1837 * save the potential rollback and start the charge transfer
1838 * only when migration is already known to end successfully.
1839 */
1840 mem_cgroup_prepare_migration(page, new_page, &memcg);
1841
Mel Gorman2b4847e2013-12-18 17:08:32 -08001842 orig_entry = *pmd;
Mel Gormanb32967f2012-11-19 12:35:47 +00001843 entry = mk_pmd(new_page, vma->vm_page_prot);
Mel Gormanb32967f2012-11-19 12:35:47 +00001844 entry = pmd_mkhuge(entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001845 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001846
Mel Gorman2b4847e2013-12-18 17:08:32 -08001847 /*
1848 * Clear the old entry under pagetable lock and establish the new PTE.
1849 * Any parallel GUP will either observe the old page blocking on the
1850 * page lock, block on the page table lock or observe the new page.
1851 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1852 * guarantee the copy is visible before the pagetable update.
1853 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001854 flush_cache_range(vma, mmun_start, mmun_end);
1855 page_add_new_anon_rmap(new_page, vma, mmun_start);
1856 pmdp_clear_flush(vma, mmun_start, pmd);
1857 set_pmd_at(mm, mmun_start, pmd, entry);
1858 flush_tlb_range(vma, mmun_start, mmun_end);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001859 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001860
1861 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08001862 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1863 flush_tlb_range(vma, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001864 update_mmu_cache_pmd(vma, address, &entry);
1865 page_remove_rmap(new_page);
1866 goto fail_putback;
1867 }
1868
Mel Gormanb32967f2012-11-19 12:35:47 +00001869 page_remove_rmap(page);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001870
Mel Gormanb32967f2012-11-19 12:35:47 +00001871 /*
1872 * Finish the charge transaction under the page table lock to
1873 * prevent split_huge_page() from dividing up the charge
1874 * before it's fully transferred to the new page.
1875 */
1876 mem_cgroup_end_migration(memcg, page, new_page, true);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001877 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001878 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001879
1880 unlock_page(new_page);
1881 unlock_page(page);
1882 put_page(page); /* Drop the rmap reference */
1883 put_page(page); /* Drop the LRU isolation reference */
1884
1885 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1886 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1887
Mel Gormanb32967f2012-11-19 12:35:47 +00001888 mod_zone_page_state(page_zone(page),
1889 NR_ISOLATED_ANON + page_lru,
1890 -HPAGE_PMD_NR);
1891 return isolated;
1892
Hugh Dickins340ef392013-02-22 16:34:33 -08001893out_fail:
1894 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001895out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08001896 ptl = pmd_lock(mm, pmd);
1897 if (pmd_same(*pmd, entry)) {
1898 entry = pmd_mknonnuma(entry);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001899 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001900 update_mmu_cache_pmd(vma, address, &entry);
1901 }
1902 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001903
Mel Gormaneb4489f62013-12-18 17:08:39 -08001904out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08001905 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001906 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001907 return 0;
1908}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001909#endif /* CONFIG_NUMA_BALANCING */
1910
1911#endif /* CONFIG_NUMA */