blob: 568284ec75d419863866f43065003883b8d81939 [file] [log] [blame]
Christoph Lameterb20a3502006-03-22 00:09:12 -08001/*
Hugh Dickins14e0f9b2015-11-05 18:49:43 -08002 * Memory Migration functionality - linux/mm/migrate.c
Christoph Lameterb20a3502006-03-22 00:09:12 -08003 *
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>
Hugh Dickins42cb14b2015-11-05 18:50:05 -080033#include <linux/backing-dev.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>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070040#include <linux/page_idle.h>
Vlastimil Babkad435edc2016-03-15 14:56:15 -070041#include <linux/page_owner.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080042
Michal Nazarewicz0d1836c2010-12-21 17:24:26 -080043#include <asm/tlbflush.h>
44
Mel Gorman7b2a2d42012-10-19 14:07:31 +010045#define CREATE_TRACE_POINTS
46#include <trace/events/migrate.h>
47
Christoph Lameterb20a3502006-03-22 00:09:12 -080048#include "internal.h"
49
Christoph Lameterb20a3502006-03-22 00:09:12 -080050/*
Christoph Lameter742755a2006-06-23 02:03:55 -070051 * migrate_prep() needs to be called before we start compiling a list of pages
Mel Gorman748446b2010-05-24 14:32:27 -070052 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
53 * undesirable, use migrate_prep_local()
Christoph Lameterb20a3502006-03-22 00:09:12 -080054 */
55int migrate_prep(void)
56{
Christoph Lameterb20a3502006-03-22 00:09:12 -080057 /*
58 * Clear the LRU lists so pages can be isolated.
59 * Note that pages may be moved off the LRU after we have
60 * drained them. Those pages will fail to migrate like other
61 * pages that may be busy.
62 */
63 lru_add_drain_all();
64
65 return 0;
66}
67
Mel Gorman748446b2010-05-24 14:32:27 -070068/* Do the necessary work of migrate_prep but not if it involves other CPUs */
69int migrate_prep_local(void)
70{
71 lru_add_drain();
72
73 return 0;
74}
75
Christoph Lameterb20a3502006-03-22 00:09:12 -080076/*
Rafael Aquini5733c7d2012-12-11 16:02:47 -080077 * Put previously isolated pages back onto the appropriate lists
78 * from where they were once taken off for compaction/migration.
79 *
Joonsoo Kim59c82b72014-01-21 15:51:17 -080080 * This function shall be used whenever the isolated pageset has been
81 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
82 * and isolate_huge_page().
Rafael Aquini5733c7d2012-12-11 16:02:47 -080083 */
84void putback_movable_pages(struct list_head *l)
85{
86 struct page *page;
87 struct page *page2;
88
89 list_for_each_entry_safe(page, page2, l, lru) {
Naoya Horiguchi31caf662013-09-11 14:21:59 -070090 if (unlikely(PageHuge(page))) {
91 putback_active_hugepage(page);
92 continue;
93 }
Rafael Aquini5733c7d2012-12-11 16:02:47 -080094 list_del(&page->lru);
95 dec_zone_page_state(page, NR_ISOLATED_ANON +
96 page_is_file_cache(page));
Rafael Aquini117aad12013-09-30 13:45:16 -070097 if (unlikely(isolated_balloon_page(page)))
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080098 balloon_page_putback(page);
99 else
100 putback_lru_page(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800101 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800102}
103
Christoph Lameter06972122006-06-23 02:03:35 -0700104/*
105 * Restore a potential migration pte to a working pte entry
106 */
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800107static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
108 unsigned long addr, void *old)
Christoph Lameter06972122006-06-23 02:03:35 -0700109{
110 struct mm_struct *mm = vma->vm_mm;
111 swp_entry_t entry;
Christoph Lameter06972122006-06-23 02:03:35 -0700112 pmd_t *pmd;
113 pte_t *ptep, pte;
114 spinlock_t *ptl;
115
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900116 if (unlikely(PageHuge(new))) {
117 ptep = huge_pte_offset(mm, addr);
118 if (!ptep)
119 goto out;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800120 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900121 } else {
Bob Liu62190492012-12-11 16:00:37 -0800122 pmd = mm_find_pmd(mm, addr);
123 if (!pmd)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900124 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);
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100151
152 /* Recheck VMA as permissions can change since migration started */
Christoph Lameter06972122006-06-23 02:03:35 -0700153 if (is_write_migration_entry(entry))
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100154 pte = maybe_mkwrite(pte, vma);
155
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200156#ifdef CONFIG_HUGETLB_PAGE
Tony Lube7517d2013-02-04 14:28:46 -0800157 if (PageHuge(new)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900158 pte = pte_mkhuge(pte);
Tony Lube7517d2013-02-04 14:28:46 -0800159 pte = arch_make_huge_pte(pte, vma, new, 0);
160 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200161#endif
Leonid Yegoshinc2cc4992013-05-24 15:55:18 -0700162 flush_dcache_page(new);
Christoph Lameter06972122006-06-23 02:03:35 -0700163 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700164
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900165 if (PageHuge(new)) {
166 if (PageAnon(new))
167 hugepage_add_anon_rmap(new, vma, addr);
168 else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -0800169 page_dup_rmap(new, true);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900170 } else if (PageAnon(new))
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800171 page_add_anon_rmap(new, vma, addr, false);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700172 else
173 page_add_file_rmap(new);
174
Hugh Dickins51afb122015-11-05 18:49:37 -0800175 if (vma->vm_flags & VM_LOCKED)
176 mlock_vma_page(new);
177
Christoph Lameter04e62a22006-06-23 02:03:38 -0700178 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000179 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800180unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700181 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800182out:
183 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700184}
185
186/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700187 * Get rid of all migration entries and replace them by
188 * references to the indicated page.
189 */
190static void remove_migration_ptes(struct page *old, struct page *new)
191{
Joonsoo Kim051ac832014-01-21 15:49:48 -0800192 struct rmap_walk_control rwc = {
193 .rmap_one = remove_migration_pte,
194 .arg = old,
195 };
196
197 rmap_walk(new, &rwc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700198}
199
200/*
Christoph Lameter06972122006-06-23 02:03:35 -0700201 * Something used the pte of a page under migration. We need to
202 * get to the page and wait until migration is finished.
203 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700204 */
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800205void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700206 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700207{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700208 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700209 swp_entry_t entry;
210 struct page *page;
211
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700212 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700213 pte = *ptep;
214 if (!is_swap_pte(pte))
215 goto out;
216
217 entry = pte_to_swp_entry(pte);
218 if (!is_migration_entry(entry))
219 goto out;
220
221 page = migration_entry_to_page(entry);
222
Nick Piggine2867812008-07-25 19:45:30 -0700223 /*
224 * Once radix-tree replacement of page migration started, page_count
225 * *must* be zero. And, we don't want to call wait_on_page_locked()
226 * against a page without get_page().
227 * So, we use get_page_unless_zero(), here. Even failed, page fault
228 * will occur again.
229 */
230 if (!get_page_unless_zero(page))
231 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700232 pte_unmap_unlock(ptep, ptl);
233 wait_on_page_locked(page);
234 put_page(page);
235 return;
236out:
237 pte_unmap_unlock(ptep, ptl);
238}
239
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700240void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
241 unsigned long address)
242{
243 spinlock_t *ptl = pte_lockptr(mm, pmd);
244 pte_t *ptep = pte_offset_map(pmd, address);
245 __migration_entry_wait(mm, ptep, ptl);
246}
247
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800248void migration_entry_wait_huge(struct vm_area_struct *vma,
249 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700250{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800251 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700252 __migration_entry_wait(mm, pte, ptl);
253}
254
Mel Gormanb969c4a2012-01-12 17:19:34 -0800255#ifdef CONFIG_BLOCK
256/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800257static bool buffer_migrate_lock_buffers(struct buffer_head *head,
258 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800259{
260 struct buffer_head *bh = head;
261
262 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800263 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800264 do {
265 get_bh(bh);
266 lock_buffer(bh);
267 bh = bh->b_this_page;
268
269 } while (bh != head);
270
271 return true;
272 }
273
274 /* async case, we cannot block on lock_buffer so use trylock_buffer */
275 do {
276 get_bh(bh);
277 if (!trylock_buffer(bh)) {
278 /*
279 * We failed to lock the buffer and cannot stall in
280 * async migration. Release the taken locks
281 */
282 struct buffer_head *failed_bh = bh;
283 put_bh(failed_bh);
284 bh = head;
285 while (bh != failed_bh) {
286 unlock_buffer(bh);
287 put_bh(bh);
288 bh = bh->b_this_page;
289 }
290 return false;
291 }
292
293 bh = bh->b_this_page;
294 } while (bh != head);
295 return true;
296}
297#else
298static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800299 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800300{
301 return true;
302}
303#endif /* CONFIG_BLOCK */
304
Christoph Lameterb20a3502006-03-22 00:09:12 -0800305/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700306 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700307 *
308 * The number of remaining references must be:
309 * 1 for anonymous pages without a mapping
310 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100311 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800312 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800313int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800314 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500315 struct buffer_head *head, enum migrate_mode mode,
316 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800317{
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800318 struct zone *oldzone, *newzone;
319 int dirty;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500320 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800321 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800322
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700323 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700324 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500325 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700326 return -EAGAIN;
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800327
328 /* No turning back from here */
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800329 newpage->index = page->index;
330 newpage->mapping = page->mapping;
331 if (PageSwapBacked(page))
332 SetPageSwapBacked(newpage);
333
Rafael Aquini78bd5202012-12-11 16:02:31 -0800334 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700335 }
336
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800337 oldzone = page_zone(page);
338 newzone = page_zone(newpage);
339
Nick Piggin19fd6232008-07-25 19:45:32 -0700340 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800341
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800342 pslot = radix_tree_lookup_slot(&mapping->page_tree,
343 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800344
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500345 expected_count += 1 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700346 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800347 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700348 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700349 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800350 }
351
Nick Piggine2867812008-07-25 19:45:30 -0700352 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700353 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700354 return -EAGAIN;
355 }
356
Christoph Lameterb20a3502006-03-22 00:09:12 -0800357 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800358 * In the async migration case of moving a page with buffers, lock the
359 * buffers using trylock before the mapping is moved. If the mapping
360 * was moved, we later failed to lock the buffers and could not move
361 * the mapping back due to an elevated page count, we would have to
362 * block waiting on other references to be dropped.
363 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800364 if (mode == MIGRATE_ASYNC && head &&
365 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800366 page_unfreeze_refs(page, expected_count);
367 spin_unlock_irq(&mapping->tree_lock);
368 return -EAGAIN;
369 }
370
371 /*
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800372 * Now we know that no one else is looking at the page:
373 * no turning back from here.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800374 */
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800375 newpage->index = page->index;
376 newpage->mapping = page->mapping;
377 if (PageSwapBacked(page))
378 SetPageSwapBacked(newpage);
379
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800380 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800381 if (PageSwapCache(page)) {
382 SetPageSwapCache(newpage);
383 set_page_private(newpage, page_private(page));
384 }
385
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800386 /* Move dirty while page refs frozen and newpage not yet exposed */
387 dirty = PageDirty(page);
388 if (dirty) {
389 ClearPageDirty(page);
390 SetPageDirty(newpage);
391 }
392
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800393 radix_tree_replace_slot(pslot, newpage);
394
395 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800396 * Drop cache reference from old page by unfreezing
397 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800398 * We know this isn't the last reference.
399 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800400 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800401
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800402 spin_unlock(&mapping->tree_lock);
403 /* Leave irq disabled to prevent preemption while updating stats */
404
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700405 /*
406 * If moved to a different zone then also account
407 * the page for that zone. Other VM counters will be
408 * taken care of when we establish references to the
409 * new page and drop references to the old page.
410 *
411 * Note that anonymous pages are accounted for
412 * via NR_FILE_PAGES and NR_ANON_PAGES if they
413 * are mapped to swap space.
414 */
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800415 if (newzone != oldzone) {
416 __dec_zone_state(oldzone, NR_FILE_PAGES);
417 __inc_zone_state(newzone, NR_FILE_PAGES);
418 if (PageSwapBacked(page) && !PageSwapCache(page)) {
419 __dec_zone_state(oldzone, NR_SHMEM);
420 __inc_zone_state(newzone, NR_SHMEM);
421 }
422 if (dirty && mapping_cap_account_dirty(mapping)) {
423 __dec_zone_state(oldzone, NR_FILE_DIRTY);
424 __inc_zone_state(newzone, NR_FILE_DIRTY);
425 }
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700426 }
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800427 local_irq_enable();
Christoph Lameterb20a3502006-03-22 00:09:12 -0800428
Rafael Aquini78bd5202012-12-11 16:02:31 -0800429 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800430}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800431
432/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900433 * The expected number of remaining references is the same as that
434 * of migrate_page_move_mapping().
435 */
436int migrate_huge_page_move_mapping(struct address_space *mapping,
437 struct page *newpage, struct page *page)
438{
439 int expected_count;
440 void **pslot;
441
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900442 spin_lock_irq(&mapping->tree_lock);
443
444 pslot = radix_tree_lookup_slot(&mapping->page_tree,
445 page_index(page));
446
447 expected_count = 2 + page_has_private(page);
448 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800449 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900450 spin_unlock_irq(&mapping->tree_lock);
451 return -EAGAIN;
452 }
453
454 if (!page_freeze_refs(page, expected_count)) {
455 spin_unlock_irq(&mapping->tree_lock);
456 return -EAGAIN;
457 }
458
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800459 newpage->index = page->index;
460 newpage->mapping = page->mapping;
Johannes Weiner6a93ca82016-03-15 14:57:19 -0700461
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900462 get_page(newpage);
463
464 radix_tree_replace_slot(pslot, newpage);
465
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800466 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900467
468 spin_unlock_irq(&mapping->tree_lock);
Johannes Weiner6a93ca82016-03-15 14:57:19 -0700469
Rafael Aquini78bd5202012-12-11 16:02:31 -0800470 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900471}
472
473/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800474 * Gigantic pages are so large that we do not guarantee that page++ pointer
475 * arithmetic will work across the entire page. We need something more
476 * specialized.
477 */
478static void __copy_gigantic_page(struct page *dst, struct page *src,
479 int nr_pages)
480{
481 int i;
482 struct page *dst_base = dst;
483 struct page *src_base = src;
484
485 for (i = 0; i < nr_pages; ) {
486 cond_resched();
487 copy_highpage(dst, src);
488
489 i++;
490 dst = mem_map_next(dst, dst_base, i);
491 src = mem_map_next(src, src_base, i);
492 }
493}
494
495static void copy_huge_page(struct page *dst, struct page *src)
496{
497 int i;
498 int nr_pages;
499
500 if (PageHuge(src)) {
501 /* hugetlbfs page */
502 struct hstate *h = page_hstate(src);
503 nr_pages = pages_per_huge_page(h);
504
505 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
506 __copy_gigantic_page(dst, src, nr_pages);
507 return;
508 }
509 } else {
510 /* thp page */
511 BUG_ON(!PageTransHuge(src));
512 nr_pages = hpage_nr_pages(src);
513 }
514
515 for (i = 0; i < nr_pages; i++) {
516 cond_resched();
517 copy_highpage(dst + i, src + i);
518 }
519}
520
521/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800522 * Copy the page to its new location
523 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900524void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800525{
Rik van Riel7851a452013-10-07 11:29:23 +0100526 int cpupid;
527
Mel Gormanb32967f2012-11-19 12:35:47 +0000528 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900529 copy_huge_page(newpage, page);
530 else
531 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800532
533 if (PageError(page))
534 SetPageError(newpage);
535 if (PageReferenced(page))
536 SetPageReferenced(newpage);
537 if (PageUptodate(page))
538 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700539 if (TestClearPageActive(page)) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800540 VM_BUG_ON_PAGE(PageUnevictable(page), page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800541 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800542 } else if (TestClearPageUnevictable(page))
543 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800544 if (PageChecked(page))
545 SetPageChecked(newpage);
546 if (PageMappedToDisk(page))
547 SetPageMappedToDisk(newpage);
548
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800549 /* Move dirty on pages not done by migrate_page_move_mapping() */
550 if (PageDirty(page))
551 SetPageDirty(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800552
Vladimir Davydov33c3fc72015-09-09 15:35:45 -0700553 if (page_is_young(page))
554 set_page_young(newpage);
555 if (page_is_idle(page))
556 set_page_idle(newpage);
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
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800565 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800566 /*
567 * Please do not reorder this without considering how mm/ksm.c's
568 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
569 */
Naoya Horiguchib3b3a992015-04-15 16:13:15 -0700570 if (PageSwapCache(page))
571 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);
Vlastimil Babkad435edc2016-03-15 14:56:15 -0700581
582 copy_page_owner(page, newpage);
Johannes Weiner74485cf2016-03-15 14:57:54 -0700583
584 mem_cgroup_migrate(page, newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800585}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800586
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700587/************************************************************
588 * Migration functions
589 ***********************************************************/
590
Christoph Lameterb20a3502006-03-22 00:09:12 -0800591/*
592 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100593 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800594 *
595 * Pages are locked upon entry and exit.
596 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700597int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800598 struct page *newpage, struct page *page,
599 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800600{
601 int rc;
602
603 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
604
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500605 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800606
Rafael Aquini78bd5202012-12-11 16:02:31 -0800607 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800608 return rc;
609
610 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800611 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800612}
613EXPORT_SYMBOL(migrate_page);
614
David Howells93614012006-09-30 20:45:40 +0200615#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800616/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700617 * Migration function for pages with buffers. This function can only be used
618 * if the underlying filesystem guarantees that no other references to "page"
619 * exist.
620 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700621int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800622 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700623{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700624 struct buffer_head *bh, *head;
625 int rc;
626
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700627 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800628 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700629
630 head = page_buffers(page);
631
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500632 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700633
Rafael Aquini78bd5202012-12-11 16:02:31 -0800634 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700635 return rc;
636
Mel Gormanb969c4a2012-01-12 17:19:34 -0800637 /*
638 * In the async case, migrate_page_move_mapping locked the buffers
639 * with an IRQ-safe spinlock held. In the sync case, the buffers
640 * need to be locked now
641 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800642 if (mode != MIGRATE_ASYNC)
643 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700644
645 ClearPagePrivate(page);
646 set_page_private(newpage, page_private(page));
647 set_page_private(page, 0);
648 put_page(page);
649 get_page(newpage);
650
651 bh = head;
652 do {
653 set_bh_page(bh, newpage, bh_offset(bh));
654 bh = bh->b_this_page;
655
656 } while (bh != head);
657
658 SetPagePrivate(newpage);
659
660 migrate_page_copy(newpage, page);
661
662 bh = head;
663 do {
664 unlock_buffer(bh);
665 put_bh(bh);
666 bh = bh->b_this_page;
667
668 } while (bh != head);
669
Rafael Aquini78bd5202012-12-11 16:02:31 -0800670 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700671}
672EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200673#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700674
Christoph Lameter04e62a22006-06-23 02:03:38 -0700675/*
676 * Writeback a page to clean the dirty state
677 */
678static int writeout(struct address_space *mapping, struct page *page)
679{
680 struct writeback_control wbc = {
681 .sync_mode = WB_SYNC_NONE,
682 .nr_to_write = 1,
683 .range_start = 0,
684 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700685 .for_reclaim = 1
686 };
687 int rc;
688
689 if (!mapping->a_ops->writepage)
690 /* No write method for the address space */
691 return -EINVAL;
692
693 if (!clear_page_dirty_for_io(page))
694 /* Someone else already triggered a write */
695 return -EAGAIN;
696
697 /*
698 * A dirty page may imply that the underlying filesystem has
699 * the page on some queue. So the page must be clean for
700 * migration. Writeout may mean we loose the lock and the
701 * page state is no longer what we checked for earlier.
702 * At this point we know that the migration attempt cannot
703 * be successful.
704 */
705 remove_migration_ptes(page, page);
706
707 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700708
709 if (rc != AOP_WRITEPAGE_ACTIVATE)
710 /* unlocked. Relock */
711 lock_page(page);
712
Hugh Dickinsbda85502008-11-19 15:36:36 -0800713 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700714}
715
716/*
717 * Default handling if a filesystem does not provide a migration function.
718 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700719static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800720 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700721{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800722 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800723 /* Only writeback pages in full synchronous migration */
724 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800725 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700726 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800727 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700728
729 /*
730 * Buffers may be managed in a filesystem specific way.
731 * We must have no buffers or drop them.
732 */
David Howells266cf652009-04-03 16:42:36 +0100733 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700734 !try_to_release_page(page, GFP_KERNEL))
735 return -EAGAIN;
736
Mel Gormana6bc32b2012-01-12 17:19:43 -0800737 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700738}
739
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700740/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700741 * Move a page to a newly allocated page
742 * The page is locked and all ptes have been successfully removed.
743 *
744 * The new page will have replaced the old page if this function
745 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700746 *
747 * Return value:
748 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800749 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700750 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700751static int move_to_new_page(struct page *newpage, struct page *page,
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800752 enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700753{
754 struct address_space *mapping;
755 int rc;
756
Hugh Dickins7db76712015-11-05 18:49:49 -0800757 VM_BUG_ON_PAGE(!PageLocked(page), page);
758 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700759
Christoph Lametere24f0b82006-06-23 02:03:51 -0700760 mapping = page_mapping(page);
761 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800762 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800763 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700764 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800765 * Most pages have a mapping and most filesystems provide a
766 * migratepage callback. Anonymous pages are part of swap
767 * space which also has its own migratepage callback. This
768 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700769 */
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800770 rc = mapping->a_ops->migratepage(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800771 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800772 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700773
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800774 /*
775 * When successful, old pagecache page->mapping must be cleared before
776 * page is freed; but stats require that PageAnon be left as PageAnon.
777 */
778 if (rc == MIGRATEPAGE_SUCCESS) {
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800779 if (!PageAnon(page))
780 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700781 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700782 return rc;
783}
784
Minchan Kim0dabec92011-10-31 17:06:57 -0700785static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800786 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700787{
Minchan Kim0dabec92011-10-31 17:06:57 -0700788 int rc = -EAGAIN;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800789 int page_was_mapped = 0;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700790 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700791
Nick Piggin529ae9a2008-08-02 12:01:03 +0200792 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800793 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700794 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800795
796 /*
797 * It's not safe for direct compaction to call lock_page.
798 * For example, during page readahead pages are added locked
799 * to the LRU. Later, when the IO completes the pages are
800 * marked uptodate and unlocked. However, the queueing
801 * could be merging multiple pages for one bio (e.g.
802 * mpage_readpages). If an allocation happens for the
803 * second or third page, the process can end up locking
804 * the same page twice and deadlocking. Rather than
805 * trying to be clever about what pages can be locked,
806 * avoid the use of lock_page for direct compaction
807 * altogether.
808 */
809 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700810 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800811
Christoph Lametere24f0b82006-06-23 02:03:51 -0700812 lock_page(page);
813 }
814
815 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700816 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700817 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800818 * necessary to wait for PageWriteback. In the async case,
819 * the retry loop is too short and in the sync-light case,
820 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700821 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800822 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700823 rc = -EBUSY;
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700824 goto out_unlock;
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700825 }
826 if (!force)
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700827 goto out_unlock;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700828 wait_on_page_writeback(page);
829 }
Hugh Dickins03f15c82015-11-05 18:49:56 -0800830
Christoph Lametere24f0b82006-06-23 02:03:51 -0700831 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700832 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
833 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800834 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700835 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700836 * File Caches may use write_page() or lock_page() in migration, then,
837 * just care Anon page here.
Hugh Dickins03f15c82015-11-05 18:49:56 -0800838 *
839 * Only page_get_anon_vma() understands the subtleties of
840 * getting a hold on an anon_vma from outside one of its mms.
841 * But if we cannot get anon_vma, then we won't need it anyway,
842 * because that implies that the anon page is no longer mapped
843 * (and cannot be remapped so long as we hold the page lock).
Christoph Lametere24f0b82006-06-23 02:03:51 -0700844 */
Hugh Dickins03f15c82015-11-05 18:49:56 -0800845 if (PageAnon(page) && !PageKsm(page))
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700846 anon_vma = page_get_anon_vma(page);
Shaohua Li62e1c552008-02-04 22:29:33 -0800847
Hugh Dickins7db76712015-11-05 18:49:49 -0800848 /*
849 * Block others from accessing the new page when we get around to
850 * establishing additional references. We are usually the only one
851 * holding a reference to newpage at this point. We used to have a BUG
852 * here if trylock_page(newpage) fails, but would like to allow for
853 * cases where there might be a race with the previous use of newpage.
854 * This is much like races on refcount of oldpage: just don't BUG().
855 */
856 if (unlikely(!trylock_page(newpage)))
857 goto out_unlock;
858
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700859 if (unlikely(isolated_balloon_page(page))) {
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800860 /*
861 * A ballooned page does not need any special attention from
862 * physical to virtual reverse mapping procedures.
863 * Skip any attempt to unmap PTEs or to remap swap cache,
864 * in order to avoid burning cycles at rmap level, and perform
865 * the page migration right away (proteced by page lock).
866 */
867 rc = balloon_page_migrate(newpage, page, mode);
Hugh Dickins7db76712015-11-05 18:49:49 -0800868 goto out_unlock_both;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800869 }
870
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700871 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800872 * Corner case handling:
873 * 1. When a new swap-cache page is read into, it is added to the LRU
874 * and treated as swapcache but it has no rmap yet.
875 * Calling try_to_unmap() against a page->mapping==NULL page will
876 * trigger a BUG. So handle it here.
877 * 2. An orphaned page (see truncate_complete_page) might have
878 * fs-private metadata. The page can be picked up due to memory
879 * offlining. Everywhere else except page reclaim, the page is
880 * invisible to the vm, so the page can not be migrated. So try to
881 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700882 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800883 if (!page->mapping) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800884 VM_BUG_ON_PAGE(PageAnon(page), page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800885 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800886 try_to_free_buffers(page);
Hugh Dickins7db76712015-11-05 18:49:49 -0800887 goto out_unlock_both;
Shaohua Li62e1c552008-02-04 22:29:33 -0800888 }
Hugh Dickins7db76712015-11-05 18:49:49 -0800889 } else if (page_mapped(page)) {
890 /* Establish migration ptes */
Hugh Dickins03f15c82015-11-05 18:49:56 -0800891 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
892 page);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800893 try_to_unmap(page,
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700894 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800895 page_was_mapped = 1;
896 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700897
Christoph Lametere6a15302006-06-25 05:46:49 -0700898 if (!page_mapped(page))
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800899 rc = move_to_new_page(newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700900
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800901 if (page_was_mapped)
902 remove_migration_ptes(page,
903 rc == MIGRATEPAGE_SUCCESS ? newpage : page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700904
Hugh Dickins7db76712015-11-05 18:49:49 -0800905out_unlock_both:
906 unlock_page(newpage);
907out_unlock:
Mel Gorman3f6c8272010-05-24 14:32:17 -0700908 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700909 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700910 put_anon_vma(anon_vma);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700911 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700912out:
913 return rc;
914}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700915
Minchan Kim0dabec92011-10-31 17:06:57 -0700916/*
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -0700917 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
918 * around it.
919 */
920#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
921#define ICE_noinline noinline
922#else
923#define ICE_noinline
924#endif
925
926/*
Minchan Kim0dabec92011-10-31 17:06:57 -0700927 * Obtain the lock on page, remove all ptes and migrate the page
928 * to the newly allocated page in newpage.
929 */
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -0700930static ICE_noinline int unmap_and_move(new_page_t get_new_page,
931 free_page_t put_new_page,
932 unsigned long private, struct page *page,
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -0700933 int force, enum migrate_mode mode,
934 enum migrate_reason reason)
Minchan Kim0dabec92011-10-31 17:06:57 -0700935{
Hugh Dickins2def7422015-11-05 18:49:46 -0800936 int rc = MIGRATEPAGE_SUCCESS;
Minchan Kim0dabec92011-10-31 17:06:57 -0700937 int *result = NULL;
Hugh Dickins2def7422015-11-05 18:49:46 -0800938 struct page *newpage;
Minchan Kim0dabec92011-10-31 17:06:57 -0700939
Hugh Dickins2def7422015-11-05 18:49:46 -0800940 newpage = get_new_page(page, private, &result);
Minchan Kim0dabec92011-10-31 17:06:57 -0700941 if (!newpage)
942 return -ENOMEM;
943
944 if (page_count(page) == 1) {
945 /* page was freed from under us. So we are done. */
946 goto out;
947 }
948
Kirill A. Shutemov4d2fa962016-01-15 16:54:00 -0800949 if (unlikely(PageTransHuge(page))) {
950 lock_page(page);
951 rc = split_huge_page(page);
952 unlock_page(page);
953 if (rc)
Minchan Kim0dabec92011-10-31 17:06:57 -0700954 goto out;
Kirill A. Shutemov4d2fa962016-01-15 16:54:00 -0800955 }
Minchan Kim0dabec92011-10-31 17:06:57 -0700956
Hugh Dickins9c620e22013-02-22 16:35:14 -0800957 rc = __unmap_and_move(page, newpage, force, mode);
Vlastimil Babka7cd12b42016-03-15 14:56:18 -0700958 if (rc == MIGRATEPAGE_SUCCESS) {
Hugh Dickins2def7422015-11-05 18:49:46 -0800959 put_new_page = NULL;
Vlastimil Babka7cd12b42016-03-15 14:56:18 -0700960 set_page_owner_migrate_reason(newpage, reason);
961 }
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800962
Minchan Kim0dabec92011-10-31 17:06:57 -0700963out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700964 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700965 /*
966 * A page that has been migrated has all references
967 * removed and will be freed. A page that has not been
968 * migrated will have kepts its references and be
969 * restored.
970 */
971 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700972 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700973 page_is_file_cache(page));
Naoya Horiguchif4c18e62015-08-06 15:47:08 -0700974 /* Soft-offlined page shouldn't go through lru cache list */
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700975 if (reason == MR_MEMORY_FAILURE) {
Naoya Horiguchif4c18e62015-08-06 15:47:08 -0700976 put_page(page);
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700977 if (!test_set_page_hwpoison(page))
978 num_poisoned_pages_inc();
979 } else
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -0700980 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700981 }
David Rientjes68711a72014-06-04 16:08:25 -0700982
Christoph Lameter95a402c2006-06-23 02:03:53 -0700983 /*
David Rientjes68711a72014-06-04 16:08:25 -0700984 * If migration was not successful and there's a freeing callback, use
985 * it. Otherwise, putback_lru_page() will drop the reference grabbed
986 * during isolation.
Christoph Lameter95a402c2006-06-23 02:03:53 -0700987 */
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800988 if (put_new_page)
David Rientjes68711a72014-06-04 16:08:25 -0700989 put_new_page(newpage, private);
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800990 else if (unlikely(__is_movable_balloon_page(newpage))) {
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700991 /* drop our reference, page already in the balloon */
992 put_page(newpage);
Hugh Dickins8bdd6382014-07-26 12:58:23 -0700993 } else
David Rientjes68711a72014-06-04 16:08:25 -0700994 putback_lru_page(newpage);
995
Christoph Lameter742755a2006-06-23 02:03:55 -0700996 if (result) {
997 if (rc)
998 *result = rc;
999 else
1000 *result = page_to_nid(newpage);
1001 }
Christoph Lametere24f0b82006-06-23 02:03:51 -07001002 return rc;
1003}
1004
1005/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001006 * Counterpart of unmap_and_move_page() for hugepage migration.
1007 *
1008 * This function doesn't wait the completion of hugepage I/O
1009 * because there is no race between I/O and migration for hugepage.
1010 * Note that currently hugepage I/O occurs only in direct I/O
1011 * where no lock is held and PG_writeback is irrelevant,
1012 * and writeback status of all subpages are counted in the reference
1013 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1014 * under direct I/O, the reference of the head page is 512 and a bit more.)
1015 * This means that when we try to migrate hugepage whose subpages are
1016 * doing direct I/O, some references remain after try_to_unmap() and
1017 * hugepage migration fails without data corruption.
1018 *
1019 * There is also no race when direct I/O is issued on the page under migration,
1020 * because then pte is replaced with migration swap entry and direct I/O code
1021 * will wait in the page fault for migration to complete.
1022 */
1023static int unmap_and_move_huge_page(new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001024 free_page_t put_new_page, unsigned long private,
1025 struct page *hpage, int force,
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001026 enum migrate_mode mode, int reason)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001027{
Hugh Dickins2def7422015-11-05 18:49:46 -08001028 int rc = -EAGAIN;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001029 int *result = NULL;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001030 int page_was_mapped = 0;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001031 struct page *new_hpage;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001032 struct anon_vma *anon_vma = NULL;
1033
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001034 /*
1035 * Movability of hugepages depends on architectures and hugepage size.
1036 * This check is necessary because some callers of hugepage migration
1037 * like soft offline and memory hotremove don't walk through page
1038 * tables or check whether the hugepage is pmd-based or not before
1039 * kicking migration.
1040 */
Naoya Horiguchi100873d2014-06-04 16:10:56 -07001041 if (!hugepage_migration_supported(page_hstate(hpage))) {
Joonsoo Kim32665f22014-01-21 15:51:15 -08001042 putback_active_hugepage(hpage);
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001043 return -ENOSYS;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001044 }
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001045
Joonsoo Kim32665f22014-01-21 15:51:15 -08001046 new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001047 if (!new_hpage)
1048 return -ENOMEM;
1049
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001050 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001051 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001052 goto out;
1053 lock_page(hpage);
1054 }
1055
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001056 if (PageAnon(hpage))
1057 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001058
Hugh Dickins7db76712015-11-05 18:49:49 -08001059 if (unlikely(!trylock_page(new_hpage)))
1060 goto put_anon;
1061
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001062 if (page_mapped(hpage)) {
1063 try_to_unmap(hpage,
1064 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1065 page_was_mapped = 1;
1066 }
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001067
1068 if (!page_mapped(hpage))
Hugh Dickins5c3f9a62015-11-05 18:49:53 -08001069 rc = move_to_new_page(new_hpage, hpage, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001070
Hugh Dickins5c3f9a62015-11-05 18:49:53 -08001071 if (page_was_mapped)
1072 remove_migration_ptes(hpage,
1073 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001074
Hugh Dickins7db76712015-11-05 18:49:49 -08001075 unlock_page(new_hpage);
1076
1077put_anon:
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001078 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001079 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001080
Hugh Dickins2def7422015-11-05 18:49:46 -08001081 if (rc == MIGRATEPAGE_SUCCESS) {
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001082 hugetlb_cgroup_migrate(hpage, new_hpage);
Hugh Dickins2def7422015-11-05 18:49:46 -08001083 put_new_page = NULL;
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001084 set_page_owner_migrate_reason(new_hpage, reason);
Hugh Dickins2def7422015-11-05 18:49:46 -08001085 }
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001086
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001087 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001088out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001089 if (rc != -EAGAIN)
1090 putback_active_hugepage(hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001091
1092 /*
1093 * If migration was not successful and there's a freeing callback, use
1094 * it. Otherwise, put_page() will drop the reference grabbed during
1095 * isolation.
1096 */
Hugh Dickins2def7422015-11-05 18:49:46 -08001097 if (put_new_page)
David Rientjes68711a72014-06-04 16:08:25 -07001098 put_new_page(new_hpage, private);
1099 else
Naoya Horiguchi3aaa76e2015-09-22 14:59:14 -07001100 putback_active_hugepage(new_hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001101
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001102 if (result) {
1103 if (rc)
1104 *result = rc;
1105 else
1106 *result = page_to_nid(new_hpage);
1107 }
1108 return rc;
1109}
1110
1111/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001112 * migrate_pages - migrate the pages specified in a list, to the free pages
1113 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001114 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001115 * @from: The list of pages to be migrated.
1116 * @get_new_page: The function used to allocate free pages to be used
1117 * as the target of the page migration.
David Rientjes68711a72014-06-04 16:08:25 -07001118 * @put_new_page: The function used to free target pages if migration
1119 * fails, or NULL if no special handling is necessary.
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001120 * @private: Private data to be passed on to get_new_page()
1121 * @mode: The migration mode that specifies the constraints for
1122 * page migration, if any.
1123 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001124 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001125 * The function returns after 10 attempts or if no pages are movable any more
1126 * because the list has become empty or no retryable pages exist any more.
Hugh Dickins14e0f9b2015-11-05 18:49:43 -08001127 * The caller should call putback_movable_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001128 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001129 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001130 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001131 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001132int migrate_pages(struct list_head *from, new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001133 free_page_t put_new_page, unsigned long private,
1134 enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001135{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001136 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001137 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001138 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001139 int pass = 0;
1140 struct page *page;
1141 struct page *page2;
1142 int swapwrite = current->flags & PF_SWAPWRITE;
1143 int rc;
1144
1145 if (!swapwrite)
1146 current->flags |= PF_SWAPWRITE;
1147
Christoph Lametere24f0b82006-06-23 02:03:51 -07001148 for(pass = 0; pass < 10 && retry; pass++) {
1149 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001150
Christoph Lametere24f0b82006-06-23 02:03:51 -07001151 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001152 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001153
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001154 if (PageHuge(page))
1155 rc = unmap_and_move_huge_page(get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001156 put_new_page, private, page,
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001157 pass > 2, mode, reason);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001158 else
David Rientjes68711a72014-06-04 16:08:25 -07001159 rc = unmap_and_move(get_new_page, put_new_page,
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -07001160 private, page, pass > 2, mode,
1161 reason);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001162
Christoph Lametere24f0b82006-06-23 02:03:51 -07001163 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001164 case -ENOMEM:
1165 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001166 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001167 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001168 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001169 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001170 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001171 break;
1172 default:
Naoya Horiguchi354a3362014-01-21 15:51:14 -08001173 /*
1174 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1175 * unlike -EAGAIN case, the failed page is
1176 * removed from migration page list and not
1177 * retried in the next outer loop.
1178 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001179 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001180 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001181 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001182 }
1183 }
Vlastimil Babkaf2f81fb2015-11-05 18:47:03 -08001184 nr_failed += retry;
1185 rc = nr_failed;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001186out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001187 if (nr_succeeded)
1188 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1189 if (nr_failed)
1190 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001191 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1192
Christoph Lameterb20a3502006-03-22 00:09:12 -08001193 if (!swapwrite)
1194 current->flags &= ~PF_SWAPWRITE;
1195
Rafael Aquini78bd5202012-12-11 16:02:31 -08001196 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001197}
1198
Christoph Lameter742755a2006-06-23 02:03:55 -07001199#ifdef CONFIG_NUMA
1200/*
1201 * Move a list of individual pages
1202 */
1203struct page_to_node {
1204 unsigned long addr;
1205 struct page *page;
1206 int node;
1207 int status;
1208};
1209
1210static struct page *new_page_node(struct page *p, unsigned long private,
1211 int **result)
1212{
1213 struct page_to_node *pm = (struct page_to_node *)private;
1214
1215 while (pm->node != MAX_NUMNODES && pm->page != p)
1216 pm++;
1217
1218 if (pm->node == MAX_NUMNODES)
1219 return NULL;
1220
1221 *result = &pm->status;
1222
Naoya Horiguchie632a932013-09-11 14:22:04 -07001223 if (PageHuge(p))
1224 return alloc_huge_page_node(page_hstate(compound_head(p)),
1225 pm->node);
1226 else
Vlastimil Babka96db8002015-09-08 15:03:50 -07001227 return __alloc_pages_node(pm->node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001228 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001229}
1230
1231/*
1232 * Move a set of pages as indicated in the pm array. The addr
1233 * field must be set to the virtual address of the page to be moved
1234 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001235 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001236 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001237static int do_move_page_to_node_array(struct mm_struct *mm,
1238 struct page_to_node *pm,
1239 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001240{
1241 int err;
1242 struct page_to_node *pp;
1243 LIST_HEAD(pagelist);
1244
1245 down_read(&mm->mmap_sem);
1246
1247 /*
1248 * Build a list of pages to migrate
1249 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001250 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1251 struct vm_area_struct *vma;
1252 struct page *page;
1253
Christoph Lameter742755a2006-06-23 02:03:55 -07001254 err = -EFAULT;
1255 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001256 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001257 goto set_status;
1258
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001259 /* FOLL_DUMP to ignore special (like zero) pages */
1260 page = follow_page(vma, pp->addr,
1261 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001262
1263 err = PTR_ERR(page);
1264 if (IS_ERR(page))
1265 goto set_status;
1266
Christoph Lameter742755a2006-06-23 02:03:55 -07001267 err = -ENOENT;
1268 if (!page)
1269 goto set_status;
1270
Christoph Lameter742755a2006-06-23 02:03:55 -07001271 pp->page = page;
1272 err = page_to_nid(page);
1273
1274 if (err == pp->node)
1275 /*
1276 * Node already in the right place
1277 */
1278 goto put_and_set;
1279
1280 err = -EACCES;
1281 if (page_mapcount(page) > 1 &&
1282 !migrate_all)
1283 goto put_and_set;
1284
Naoya Horiguchie632a932013-09-11 14:22:04 -07001285 if (PageHuge(page)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08001286 if (PageHead(page))
1287 isolate_huge_page(page, &pagelist);
Naoya Horiguchie632a932013-09-11 14:22:04 -07001288 goto put_and_set;
1289 }
1290
Nick Piggin62695a82008-10-18 20:26:09 -07001291 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001292 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001293 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001294 inc_zone_page_state(page, NR_ISOLATED_ANON +
1295 page_is_file_cache(page));
1296 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001297put_and_set:
1298 /*
1299 * Either remove the duplicate refcount from
1300 * isolate_lru_page() or drop the page ref if it was
1301 * not isolated.
1302 */
1303 put_page(page);
1304set_status:
1305 pp->status = err;
1306 }
1307
Brice Gogline78bbfa2008-10-18 20:27:15 -07001308 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001309 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -07001310 err = migrate_pages(&pagelist, new_page_node, NULL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001311 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001312 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001313 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001314 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001315
1316 up_read(&mm->mmap_sem);
1317 return err;
1318}
1319
1320/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001321 * Migrate an array of page address onto an array of nodes and fill
1322 * the corresponding array of status.
1323 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001324static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001325 unsigned long nr_pages,
1326 const void __user * __user *pages,
1327 const int __user *nodes,
1328 int __user *status, int flags)
1329{
Brice Goglin3140a222009-01-06 14:38:57 -08001330 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001331 unsigned long chunk_nr_pages;
1332 unsigned long chunk_start;
1333 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001334
Brice Goglin3140a222009-01-06 14:38:57 -08001335 err = -ENOMEM;
1336 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1337 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001338 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001339
1340 migrate_prep();
1341
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001342 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001343 * Store a chunk of page_to_node array in a page,
1344 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001345 */
Brice Goglin3140a222009-01-06 14:38:57 -08001346 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001347
Brice Goglin3140a222009-01-06 14:38:57 -08001348 for (chunk_start = 0;
1349 chunk_start < nr_pages;
1350 chunk_start += chunk_nr_pages) {
1351 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001352
Brice Goglin3140a222009-01-06 14:38:57 -08001353 if (chunk_start + chunk_nr_pages > nr_pages)
1354 chunk_nr_pages = nr_pages - chunk_start;
1355
1356 /* fill the chunk pm with addrs and nodes from user-space */
1357 for (j = 0; j < chunk_nr_pages; j++) {
1358 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001359 int node;
1360
Brice Goglin3140a222009-01-06 14:38:57 -08001361 err = -EFAULT;
1362 if (get_user(p, pages + j + chunk_start))
1363 goto out_pm;
1364 pm[j].addr = (unsigned long) p;
1365
1366 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001367 goto out_pm;
1368
1369 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001370 if (node < 0 || node >= MAX_NUMNODES)
1371 goto out_pm;
1372
Lai Jiangshan389162c2012-12-12 13:51:30 -08001373 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001374 goto out_pm;
1375
1376 err = -EACCES;
1377 if (!node_isset(node, task_nodes))
1378 goto out_pm;
1379
Brice Goglin3140a222009-01-06 14:38:57 -08001380 pm[j].node = node;
1381 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001382
Brice Goglin3140a222009-01-06 14:38:57 -08001383 /* End marker for this chunk */
1384 pm[chunk_nr_pages].node = MAX_NUMNODES;
1385
1386 /* Migrate this chunk */
1387 err = do_move_page_to_node_array(mm, pm,
1388 flags & MPOL_MF_MOVE_ALL);
1389 if (err < 0)
1390 goto out_pm;
1391
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001392 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001393 for (j = 0; j < chunk_nr_pages; j++)
1394 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001395 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001396 goto out_pm;
1397 }
1398 }
1399 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001400
1401out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001402 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001403out:
1404 return err;
1405}
1406
1407/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001408 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001409 */
Brice Goglin80bba122008-12-09 13:14:23 -08001410static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1411 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001412{
Brice Goglin2f007e72008-10-18 20:27:16 -07001413 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001414
Christoph Lameter742755a2006-06-23 02:03:55 -07001415 down_read(&mm->mmap_sem);
1416
Brice Goglin2f007e72008-10-18 20:27:16 -07001417 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001418 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001419 struct vm_area_struct *vma;
1420 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001421 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001422
1423 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001424 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001425 goto set_status;
1426
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001427 /* FOLL_DUMP to ignore special (like zero) pages */
1428 page = follow_page(vma, addr, FOLL_DUMP);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001429
1430 err = PTR_ERR(page);
1431 if (IS_ERR(page))
1432 goto set_status;
1433
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001434 err = page ? page_to_nid(page) : -ENOENT;
Christoph Lameter742755a2006-06-23 02:03:55 -07001435set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001436 *status = err;
1437
1438 pages++;
1439 status++;
1440 }
1441
1442 up_read(&mm->mmap_sem);
1443}
1444
1445/*
1446 * Determine the nodes of a user array of pages and store it in
1447 * a user array of status.
1448 */
1449static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1450 const void __user * __user *pages,
1451 int __user *status)
1452{
1453#define DO_PAGES_STAT_CHUNK_NR 16
1454 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1455 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001456
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001457 while (nr_pages) {
1458 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001459
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001460 chunk_nr = nr_pages;
1461 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1462 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1463
1464 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1465 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001466
1467 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1468
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001469 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1470 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001471
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001472 pages += chunk_nr;
1473 status += chunk_nr;
1474 nr_pages -= chunk_nr;
1475 }
1476 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001477}
1478
1479/*
1480 * Move a list of pages in the address space of the currently executing
1481 * process.
1482 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001483SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1484 const void __user * __user *, pages,
1485 const int __user *, nodes,
1486 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001487{
David Howellsc69e8d92008-11-14 10:39:19 +11001488 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001489 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001490 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001491 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001492 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001493
1494 /* Check flags */
1495 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1496 return -EINVAL;
1497
1498 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1499 return -EPERM;
1500
1501 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001502 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001503 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001504 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001505 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001506 return -ESRCH;
1507 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001508 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001509
1510 /*
1511 * Check if this process has the right to modify the specified
1512 * process. The right exists if the process has administrative
1513 * capabilities, superuser privileges or the same
1514 * userid as the target process.
1515 */
David Howellsc69e8d92008-11-14 10:39:19 +11001516 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001517 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1518 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001519 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001520 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001521 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001522 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001523 }
David Howellsc69e8d92008-11-14 10:39:19 +11001524 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001525
David Quigley86c3a762006-06-23 02:04:02 -07001526 err = security_task_movememory(task);
1527 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001528 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001529
Christoph Lameter3268c632012-03-21 16:34:06 -07001530 task_nodes = cpuset_mems_allowed(task);
1531 mm = get_task_mm(task);
1532 put_task_struct(task);
1533
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001534 if (!mm)
1535 return -EINVAL;
1536
1537 if (nodes)
1538 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1539 nodes, status, flags);
1540 else
1541 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001542
1543 mmput(mm);
1544 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001545
Christoph Lameter742755a2006-06-23 02:03:55 -07001546out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001547 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001548 return err;
1549}
Christoph Lameter742755a2006-06-23 02:03:55 -07001550
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001551#ifdef CONFIG_NUMA_BALANCING
1552/*
1553 * Returns true if this is a safe migration target node for misplaced NUMA
1554 * pages. Currently it only checks the watermarks which crude
1555 */
1556static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001557 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001558{
1559 int z;
1560 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1561 struct zone *zone = pgdat->node_zones + z;
1562
1563 if (!populated_zone(zone))
1564 continue;
1565
Lisa Du6e543d52013-09-11 14:22:36 -07001566 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001567 continue;
1568
1569 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1570 if (!zone_watermark_ok(zone, 0,
1571 high_wmark_pages(zone) +
1572 nr_migrate_pages,
1573 0, 0))
1574 continue;
1575 return true;
1576 }
1577 return false;
1578}
1579
1580static struct page *alloc_misplaced_dst_page(struct page *page,
1581 unsigned long data,
1582 int **result)
1583{
1584 int nid = (int) data;
1585 struct page *newpage;
1586
Vlastimil Babka96db8002015-09-08 15:03:50 -07001587 newpage = __alloc_pages_node(nid,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001588 (GFP_HIGHUSER_MOVABLE |
1589 __GFP_THISNODE | __GFP_NOMEMALLOC |
1590 __GFP_NORETRY | __GFP_NOWARN) &
Mel Gorman8479eba2016-02-26 15:19:31 -08001591 ~__GFP_RECLAIM, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001592
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001593 return newpage;
1594}
1595
1596/*
Mel Gormana8f60772012-11-14 21:41:46 +00001597 * page migration rate limiting control.
1598 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1599 * window of time. Default here says do not migrate more than 1280M per second.
1600 */
1601static unsigned int migrate_interval_millisecs __read_mostly = 100;
1602static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1603
Mel Gormanb32967f2012-11-19 12:35:47 +00001604/* Returns true if the node is migrate rate-limited after the update */
Mel Gorman1c30e012014-01-21 15:50:58 -08001605static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1606 unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001607{
Mel Gormanb32967f2012-11-19 12:35:47 +00001608 /*
1609 * Rate-limit the amount of data that is being migrated to a node.
1610 * Optimal placement is no good if the memory bus is saturated and
1611 * all the time is being spent migrating!
1612 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001613 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001614 spin_lock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001615 pgdat->numabalancing_migrate_nr_pages = 0;
1616 pgdat->numabalancing_migrate_next_window = jiffies +
1617 msecs_to_jiffies(migrate_interval_millisecs);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001618 spin_unlock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001619 }
Mel Gormanaf1839d2014-01-21 15:51:01 -08001620 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1621 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1622 nr_pages);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001623 return true;
Mel Gormanaf1839d2014-01-21 15:51:01 -08001624 }
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001625
1626 /*
1627 * This is an unlocked non-atomic update so errors are possible.
1628 * The consequences are failing to migrate when we potentiall should
1629 * have which is not severe enough to warrant locking. If it is ever
1630 * a problem, it can be converted to a per-cpu counter.
1631 */
1632 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1633 return false;
Mel Gormanb32967f2012-11-19 12:35:47 +00001634}
1635
Mel Gorman1c30e012014-01-21 15:50:58 -08001636static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
Mel Gormanb32967f2012-11-19 12:35:47 +00001637{
Hugh Dickins340ef392013-02-22 16:34:33 -08001638 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001639
Sasha Levin309381fea2014-01-23 15:52:54 -08001640 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
Mel Gorman3abef4e2013-02-22 16:34:27 -08001641
Mel Gormanb32967f2012-11-19 12:35:47 +00001642 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001643 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1644 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001645
Hugh Dickins340ef392013-02-22 16:34:33 -08001646 if (isolate_lru_page(page))
1647 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001648
1649 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001650 * migrate_misplaced_transhuge_page() skips page migration's usual
1651 * check on page_count(), so we must do it here, now that the page
1652 * has been isolated: a GUP pin, or any other pin, prevents migration.
1653 * The expected page count is 3: 1 for page's mapcount and 1 for the
1654 * caller's pin and 1 for the reference taken by isolate_lru_page().
1655 */
1656 if (PageTransHuge(page) && page_count(page) != 3) {
1657 putback_lru_page(page);
1658 return 0;
1659 }
1660
1661 page_lru = page_is_file_cache(page);
1662 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1663 hpage_nr_pages(page));
1664
1665 /*
1666 * Isolating the page has taken another reference, so the
1667 * caller's reference can be safely dropped without the page
1668 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001669 */
1670 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001671 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001672}
1673
Mel Gormande466bd2013-12-18 17:08:42 -08001674bool pmd_trans_migrating(pmd_t pmd)
1675{
1676 struct page *page = pmd_page(pmd);
1677 return PageLocked(page);
1678}
1679
Mel Gormana8f60772012-11-14 21:41:46 +00001680/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001681 * Attempt to migrate a misplaced page to the specified destination
1682 * node. Caller is expected to have an elevated reference count on
1683 * the page that will be dropped by this function before returning.
1684 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001685int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1686 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001687{
Mel Gormana8f60772012-11-14 21:41:46 +00001688 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001689 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001690 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001691 LIST_HEAD(migratepages);
1692
1693 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001694 * Don't migrate file pages that are mapped in multiple processes
1695 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001696 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001697 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1698 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001699 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001700
Mel Gormana8f60772012-11-14 21:41:46 +00001701 /*
1702 * Rate-limit the amount of data that is being migrated to a node.
1703 * Optimal placement is no good if the memory bus is saturated and
1704 * all the time is being spent migrating!
1705 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001706 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001707 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001708
Mel Gormanb32967f2012-11-19 12:35:47 +00001709 isolated = numamigrate_isolate_page(pgdat, page);
1710 if (!isolated)
1711 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001712
Mel Gormanb32967f2012-11-19 12:35:47 +00001713 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001714 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
David Rientjes68711a72014-06-04 16:08:25 -07001715 NULL, node, MIGRATE_ASYNC,
1716 MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001717 if (nr_remaining) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001718 if (!list_empty(&migratepages)) {
1719 list_del(&page->lru);
1720 dec_zone_page_state(page, NR_ISOLATED_ANON +
1721 page_is_file_cache(page));
1722 putback_lru_page(page);
1723 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001724 isolated = 0;
1725 } else
1726 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001727 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001728 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001729
1730out:
1731 put_page(page);
1732 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001733}
Mel Gorman220018d2012-12-05 09:32:56 +00001734#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001735
Mel Gorman220018d2012-12-05 09:32:56 +00001736#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001737/*
1738 * Migrates a THP to a given target node. page must be locked and is unlocked
1739 * before returning.
1740 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001741int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1742 struct vm_area_struct *vma,
1743 pmd_t *pmd, pmd_t entry,
1744 unsigned long address,
1745 struct page *page, int node)
1746{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001747 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001748 pg_data_t *pgdat = NODE_DATA(node);
1749 int isolated = 0;
1750 struct page *new_page = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001751 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001752 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1753 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001754 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001755
1756 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001757 * Rate-limit the amount of data that is being migrated to a node.
1758 * Optimal placement is no good if the memory bus is saturated and
1759 * all the time is being spent migrating!
1760 */
Mel Gormand28d43352012-11-29 09:24:36 +00001761 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001762 goto out_dropref;
1763
1764 new_page = alloc_pages_node(node,
Mel Gorman71baba42015-11-06 16:28:28 -08001765 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001766 HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001767 if (!new_page)
1768 goto out_fail;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001769 prep_transhuge_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001770
Mel Gormanb32967f2012-11-19 12:35:47 +00001771 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001772 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001773 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001774 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001775 }
1776
Mel Gormanb0943d62013-12-18 17:08:46 -08001777 if (mm_tlb_flush_pending(mm))
1778 flush_tlb_range(vma, mmun_start, mmun_end);
1779
Mel Gormanb32967f2012-11-19 12:35:47 +00001780 /* Prepare a page as a migration target */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001781 __SetPageLocked(new_page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001782 SetPageSwapBacked(new_page);
1783
1784 /* anon mapping, we can simply copy page->mapping to the new page: */
1785 new_page->mapping = page->mapping;
1786 new_page->index = page->index;
1787 migrate_page_copy(new_page, page);
1788 WARN_ON(PageLRU(new_page));
1789
1790 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001791 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001792 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001793 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1794fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001795 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001796 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001797
1798 /* Reverse changes made by migrate_page_copy() */
1799 if (TestClearPageActive(new_page))
1800 SetPageActive(page);
1801 if (TestClearPageUnevictable(new_page))
1802 SetPageUnevictable(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001803
1804 unlock_page(new_page);
1805 put_page(new_page); /* Free it */
1806
Mel Gormana54a4072013-10-07 11:28:46 +01001807 /* Retake the callers reference and putback on LRU */
1808 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001809 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001810 mod_zone_page_state(page_zone(page),
1811 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001812
1813 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001814 }
1815
Mel Gorman2b4847e2013-12-18 17:08:32 -08001816 orig_entry = *pmd;
Mel Gormanb32967f2012-11-19 12:35:47 +00001817 entry = mk_pmd(new_page, vma->vm_page_prot);
Mel Gormanb32967f2012-11-19 12:35:47 +00001818 entry = pmd_mkhuge(entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001819 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001820
Mel Gorman2b4847e2013-12-18 17:08:32 -08001821 /*
1822 * Clear the old entry under pagetable lock and establish the new PTE.
1823 * Any parallel GUP will either observe the old page blocking on the
1824 * page lock, block on the page table lock or observe the new page.
1825 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1826 * guarantee the copy is visible before the pagetable update.
1827 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001828 flush_cache_range(vma, mmun_start, mmun_end);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001829 page_add_anon_rmap(new_page, vma, mmun_start, true);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001830 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001831 set_pmd_at(mm, mmun_start, pmd, entry);
1832 flush_tlb_range(vma, mmun_start, mmun_end);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001833 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001834
1835 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08001836 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1837 flush_tlb_range(vma, mmun_start, mmun_end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001838 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001839 update_mmu_cache_pmd(vma, address, &entry);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001840 page_remove_rmap(new_page, true);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001841 goto fail_putback;
1842 }
1843
Hugh Dickins51afb122015-11-05 18:49:37 -08001844 mlock_migrate_page(new_page, page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001845 page_remove_rmap(page, true);
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001846 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001847
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001848 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001849 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001850
Mel Gorman11de9922014-06-04 16:07:41 -07001851 /* Take an "isolate" reference and put new page on the LRU. */
1852 get_page(new_page);
1853 putback_lru_page(new_page);
1854
Mel Gormanb32967f2012-11-19 12:35:47 +00001855 unlock_page(new_page);
1856 unlock_page(page);
1857 put_page(page); /* Drop the rmap reference */
1858 put_page(page); /* Drop the LRU isolation reference */
1859
1860 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1861 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1862
Mel Gormanb32967f2012-11-19 12:35:47 +00001863 mod_zone_page_state(page_zone(page),
1864 NR_ISOLATED_ANON + page_lru,
1865 -HPAGE_PMD_NR);
1866 return isolated;
1867
Hugh Dickins340ef392013-02-22 16:34:33 -08001868out_fail:
1869 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001870out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08001871 ptl = pmd_lock(mm, pmd);
1872 if (pmd_same(*pmd, entry)) {
Mel Gorman4d942462015-02-12 14:58:28 -08001873 entry = pmd_modify(entry, vma->vm_page_prot);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001874 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001875 update_mmu_cache_pmd(vma, address, &entry);
1876 }
1877 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001878
Mel Gormaneb4489f62013-12-18 17:08:39 -08001879out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08001880 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001881 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001882 return 0;
1883}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001884#endif /* CONFIG_NUMA_BALANCING */
1885
1886#endif /* CONFIG_NUMA */