blob: 7890d0bb5e23c3db75cc682878a2c5ec89b0e513 [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>
Christoph Lameterb20a3502006-03-22 00:09:12 -080041
Michal Nazarewicz0d1836c2010-12-21 17:24:26 -080042#include <asm/tlbflush.h>
43
Mel Gorman7b2a2d42012-10-19 14:07:31 +010044#define CREATE_TRACE_POINTS
45#include <trace/events/migrate.h>
46
Christoph Lameterb20a3502006-03-22 00:09:12 -080047#include "internal.h"
48
Christoph Lameterb20a3502006-03-22 00:09:12 -080049/*
Christoph Lameter742755a2006-06-23 02:03:55 -070050 * migrate_prep() needs to be called before we start compiling a list of pages
Mel Gorman748446b2010-05-24 14:32:27 -070051 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
52 * undesirable, use migrate_prep_local()
Christoph Lameterb20a3502006-03-22 00:09:12 -080053 */
54int migrate_prep(void)
55{
Christoph Lameterb20a3502006-03-22 00:09:12 -080056 /*
57 * Clear the LRU lists so pages can be isolated.
58 * Note that pages may be moved off the LRU after we have
59 * drained them. Those pages will fail to migrate like other
60 * pages that may be busy.
61 */
62 lru_add_drain_all();
63
64 return 0;
65}
66
Mel Gorman748446b2010-05-24 14:32:27 -070067/* Do the necessary work of migrate_prep but not if it involves other CPUs */
68int migrate_prep_local(void)
69{
70 lru_add_drain();
71
72 return 0;
73}
74
Christoph Lameterb20a3502006-03-22 00:09:12 -080075/*
Rafael Aquini5733c7d2012-12-11 16:02:47 -080076 * Put previously isolated pages back onto the appropriate lists
77 * from where they were once taken off for compaction/migration.
78 *
Joonsoo Kim59c82b72014-01-21 15:51:17 -080079 * This function shall be used whenever the isolated pageset has been
80 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
81 * and isolate_huge_page().
Rafael Aquini5733c7d2012-12-11 16:02:47 -080082 */
83void putback_movable_pages(struct list_head *l)
84{
85 struct page *page;
86 struct page *page2;
87
88 list_for_each_entry_safe(page, page2, l, lru) {
Naoya Horiguchi31caf662013-09-11 14:21:59 -070089 if (unlikely(PageHuge(page))) {
90 putback_active_hugepage(page);
91 continue;
92 }
Rafael Aquini5733c7d2012-12-11 16:02:47 -080093 list_del(&page->lru);
94 dec_zone_page_state(page, NR_ISOLATED_ANON +
95 page_is_file_cache(page));
Rafael Aquini117aad12013-09-30 13:45:16 -070096 if (unlikely(isolated_balloon_page(page)))
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080097 balloon_page_putback(page);
98 else
99 putback_lru_page(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800100 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800101}
102
Christoph Lameter06972122006-06-23 02:03:35 -0700103/*
104 * Restore a potential migration pte to a working pte entry
105 */
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800106static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
107 unsigned long addr, void *old)
Christoph Lameter06972122006-06-23 02:03:35 -0700108{
109 struct mm_struct *mm = vma->vm_mm;
110 swp_entry_t entry;
Christoph Lameter06972122006-06-23 02:03:35 -0700111 pmd_t *pmd;
112 pte_t *ptep, pte;
113 spinlock_t *ptl;
114
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900115 if (unlikely(PageHuge(new))) {
116 ptep = huge_pte_offset(mm, addr);
117 if (!ptep)
118 goto out;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800119 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900120 } else {
Bob Liu62190492012-12-11 16:00:37 -0800121 pmd = mm_find_pmd(mm, addr);
122 if (!pmd)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900123 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700124
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900125 ptep = pte_offset_map(pmd, addr);
Christoph Lameter06972122006-06-23 02:03:35 -0700126
Hugh Dickins486cf462011-10-19 12:50:35 -0700127 /*
128 * Peek to check is_swap_pte() before taking ptlock? No, we
129 * can race mremap's move_ptes(), which skips anon_vma lock.
130 */
Christoph Lameter06972122006-06-23 02:03:35 -0700131
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900132 ptl = pte_lockptr(mm, pmd);
133 }
134
Christoph Lameter06972122006-06-23 02:03:35 -0700135 spin_lock(ptl);
136 pte = *ptep;
137 if (!is_swap_pte(pte))
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800138 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700139
140 entry = pte_to_swp_entry(pte);
141
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800142 if (!is_migration_entry(entry) ||
143 migration_entry_to_page(entry) != old)
144 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700145
Christoph Lameter06972122006-06-23 02:03:35 -0700146 get_page(new);
147 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700148 if (pte_swp_soft_dirty(*ptep))
149 pte = pte_mksoft_dirty(pte);
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100150
151 /* Recheck VMA as permissions can change since migration started */
Christoph Lameter06972122006-06-23 02:03:35 -0700152 if (is_write_migration_entry(entry))
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100153 pte = maybe_mkwrite(pte, vma);
154
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200155#ifdef CONFIG_HUGETLB_PAGE
Tony Lube7517d2013-02-04 14:28:46 -0800156 if (PageHuge(new)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900157 pte = pte_mkhuge(pte);
Tony Lube7517d2013-02-04 14:28:46 -0800158 pte = arch_make_huge_pte(pte, vma, new, 0);
159 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200160#endif
Leonid Yegoshinc2cc4992013-05-24 15:55:18 -0700161 flush_dcache_page(new);
Christoph Lameter06972122006-06-23 02:03:35 -0700162 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700163
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900164 if (PageHuge(new)) {
165 if (PageAnon(new))
166 hugepage_add_anon_rmap(new, vma, addr);
167 else
168 page_dup_rmap(new);
169 } else if (PageAnon(new))
Christoph Lameter04e62a22006-06-23 02:03:38 -0700170 page_add_anon_rmap(new, vma, addr);
171 else
172 page_add_file_rmap(new);
173
Hugh Dickins51afb122015-11-05 18:49:37 -0800174 if (vma->vm_flags & VM_LOCKED)
175 mlock_vma_page(new);
176
Christoph Lameter04e62a22006-06-23 02:03:38 -0700177 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000178 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800179unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700180 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800181out:
182 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700183}
184
185/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700186 * Get rid of all migration entries and replace them by
187 * references to the indicated page.
188 */
189static void remove_migration_ptes(struct page *old, struct page *new)
190{
Joonsoo Kim051ac832014-01-21 15:49:48 -0800191 struct rmap_walk_control rwc = {
192 .rmap_one = remove_migration_pte,
193 .arg = old,
194 };
195
196 rmap_walk(new, &rwc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700197}
198
199/*
Christoph Lameter06972122006-06-23 02:03:35 -0700200 * Something used the pte of a page under migration. We need to
201 * get to the page and wait until migration is finished.
202 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700203 */
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800204void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700205 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700206{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700207 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700208 swp_entry_t entry;
209 struct page *page;
210
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700211 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700212 pte = *ptep;
213 if (!is_swap_pte(pte))
214 goto out;
215
216 entry = pte_to_swp_entry(pte);
217 if (!is_migration_entry(entry))
218 goto out;
219
220 page = migration_entry_to_page(entry);
221
Nick Piggine2867812008-07-25 19:45:30 -0700222 /*
223 * Once radix-tree replacement of page migration started, page_count
224 * *must* be zero. And, we don't want to call wait_on_page_locked()
225 * against a page without get_page().
226 * So, we use get_page_unless_zero(), here. Even failed, page fault
227 * will occur again.
228 */
229 if (!get_page_unless_zero(page))
230 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700231 pte_unmap_unlock(ptep, ptl);
232 wait_on_page_locked(page);
233 put_page(page);
234 return;
235out:
236 pte_unmap_unlock(ptep, ptl);
237}
238
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700239void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
240 unsigned long address)
241{
242 spinlock_t *ptl = pte_lockptr(mm, pmd);
243 pte_t *ptep = pte_offset_map(pmd, address);
244 __migration_entry_wait(mm, ptep, ptl);
245}
246
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800247void migration_entry_wait_huge(struct vm_area_struct *vma,
248 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700249{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800250 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700251 __migration_entry_wait(mm, pte, ptl);
252}
253
Mel Gormanb969c4a2012-01-12 17:19:34 -0800254#ifdef CONFIG_BLOCK
255/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800256static bool buffer_migrate_lock_buffers(struct buffer_head *head,
257 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800258{
259 struct buffer_head *bh = head;
260
261 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800262 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800263 do {
264 get_bh(bh);
265 lock_buffer(bh);
266 bh = bh->b_this_page;
267
268 } while (bh != head);
269
270 return true;
271 }
272
273 /* async case, we cannot block on lock_buffer so use trylock_buffer */
274 do {
275 get_bh(bh);
276 if (!trylock_buffer(bh)) {
277 /*
278 * We failed to lock the buffer and cannot stall in
279 * async migration. Release the taken locks
280 */
281 struct buffer_head *failed_bh = bh;
282 put_bh(failed_bh);
283 bh = head;
284 while (bh != failed_bh) {
285 unlock_buffer(bh);
286 put_bh(bh);
287 bh = bh->b_this_page;
288 }
289 return false;
290 }
291
292 bh = bh->b_this_page;
293 } while (bh != head);
294 return true;
295}
296#else
297static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800298 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800299{
300 return true;
301}
302#endif /* CONFIG_BLOCK */
303
Christoph Lameterb20a3502006-03-22 00:09:12 -0800304/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700305 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700306 *
307 * The number of remaining references must be:
308 * 1 for anonymous pages without a mapping
309 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100310 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800311 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800312int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800313 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500314 struct buffer_head *head, enum migrate_mode mode,
315 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800316{
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800317 struct zone *oldzone, *newzone;
318 int dirty;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500319 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800320 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800321
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700322 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700323 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500324 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700325 return -EAGAIN;
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800326
327 /* No turning back from here */
328 set_page_memcg(newpage, page_memcg(page));
329 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 set_page_memcg(newpage, page_memcg(page));
376 newpage->index = page->index;
377 newpage->mapping = page->mapping;
378 if (PageSwapBacked(page))
379 SetPageSwapBacked(newpage);
380
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800381 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800382 if (PageSwapCache(page)) {
383 SetPageSwapCache(newpage);
384 set_page_private(newpage, page_private(page));
385 }
386
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800387 /* Move dirty while page refs frozen and newpage not yet exposed */
388 dirty = PageDirty(page);
389 if (dirty) {
390 ClearPageDirty(page);
391 SetPageDirty(newpage);
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
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800403 spin_unlock(&mapping->tree_lock);
404 /* Leave irq disabled to prevent preemption while updating stats */
405
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700406 /*
407 * If moved to a different zone then also account
408 * the page for that zone. Other VM counters will be
409 * taken care of when we establish references to the
410 * new page and drop references to the old page.
411 *
412 * Note that anonymous pages are accounted for
413 * via NR_FILE_PAGES and NR_ANON_PAGES if they
414 * are mapped to swap space.
415 */
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800416 if (newzone != oldzone) {
417 __dec_zone_state(oldzone, NR_FILE_PAGES);
418 __inc_zone_state(newzone, NR_FILE_PAGES);
419 if (PageSwapBacked(page) && !PageSwapCache(page)) {
420 __dec_zone_state(oldzone, NR_SHMEM);
421 __inc_zone_state(newzone, NR_SHMEM);
422 }
423 if (dirty && mapping_cap_account_dirty(mapping)) {
424 __dec_zone_state(oldzone, NR_FILE_DIRTY);
425 __inc_zone_state(newzone, NR_FILE_DIRTY);
426 }
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700427 }
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800428 local_irq_enable();
Christoph Lameterb20a3502006-03-22 00:09:12 -0800429
Rafael Aquini78bd5202012-12-11 16:02:31 -0800430 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800431}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800432
433/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900434 * The expected number of remaining references is the same as that
435 * of migrate_page_move_mapping().
436 */
437int migrate_huge_page_move_mapping(struct address_space *mapping,
438 struct page *newpage, struct page *page)
439{
440 int expected_count;
441 void **pslot;
442
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900443 spin_lock_irq(&mapping->tree_lock);
444
445 pslot = radix_tree_lookup_slot(&mapping->page_tree,
446 page_index(page));
447
448 expected_count = 2 + page_has_private(page);
449 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800450 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900451 spin_unlock_irq(&mapping->tree_lock);
452 return -EAGAIN;
453 }
454
455 if (!page_freeze_refs(page, expected_count)) {
456 spin_unlock_irq(&mapping->tree_lock);
457 return -EAGAIN;
458 }
459
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800460 set_page_memcg(newpage, page_memcg(page));
461 newpage->index = page->index;
462 newpage->mapping = page->mapping;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900463 get_page(newpage);
464
465 radix_tree_replace_slot(pslot, newpage);
466
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800467 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900468
469 spin_unlock_irq(&mapping->tree_lock);
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);
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,
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800748 enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700749{
750 struct address_space *mapping;
751 int rc;
752
Hugh Dickins7db76712015-11-05 18:49:49 -0800753 VM_BUG_ON_PAGE(!PageLocked(page), page);
754 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700755
Christoph Lametere24f0b82006-06-23 02:03:51 -0700756 mapping = page_mapping(page);
757 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800758 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800759 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700760 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800761 * Most pages have a mapping and most filesystems provide a
762 * migratepage callback. Anonymous pages are part of swap
763 * space which also has its own migratepage callback. This
764 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700765 */
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800766 rc = mapping->a_ops->migratepage(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800767 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800768 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700769
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800770 /*
771 * When successful, old pagecache page->mapping must be cleared before
772 * page is freed; but stats require that PageAnon be left as PageAnon.
773 */
774 if (rc == MIGRATEPAGE_SUCCESS) {
775 set_page_memcg(page, NULL);
776 if (!PageAnon(page))
777 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700778 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700779 return rc;
780}
781
Minchan Kim0dabec92011-10-31 17:06:57 -0700782static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800783 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700784{
Minchan Kim0dabec92011-10-31 17:06:57 -0700785 int rc = -EAGAIN;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800786 int page_was_mapped = 0;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700787 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700788
Nick Piggin529ae9a2008-08-02 12:01:03 +0200789 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800790 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700791 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800792
793 /*
794 * It's not safe for direct compaction to call lock_page.
795 * For example, during page readahead pages are added locked
796 * to the LRU. Later, when the IO completes the pages are
797 * marked uptodate and unlocked. However, the queueing
798 * could be merging multiple pages for one bio (e.g.
799 * mpage_readpages). If an allocation happens for the
800 * second or third page, the process can end up locking
801 * the same page twice and deadlocking. Rather than
802 * trying to be clever about what pages can be locked,
803 * avoid the use of lock_page for direct compaction
804 * altogether.
805 */
806 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700807 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800808
Christoph Lametere24f0b82006-06-23 02:03:51 -0700809 lock_page(page);
810 }
811
812 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700813 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700814 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800815 * necessary to wait for PageWriteback. In the async case,
816 * the retry loop is too short and in the sync-light case,
817 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700818 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800819 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700820 rc = -EBUSY;
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700821 goto out_unlock;
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700822 }
823 if (!force)
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700824 goto out_unlock;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700825 wait_on_page_writeback(page);
826 }
Hugh Dickins03f15c82015-11-05 18:49:56 -0800827
Christoph Lametere24f0b82006-06-23 02:03:51 -0700828 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700829 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
830 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800831 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700832 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700833 * File Caches may use write_page() or lock_page() in migration, then,
834 * just care Anon page here.
Hugh Dickins03f15c82015-11-05 18:49:56 -0800835 *
836 * Only page_get_anon_vma() understands the subtleties of
837 * getting a hold on an anon_vma from outside one of its mms.
838 * But if we cannot get anon_vma, then we won't need it anyway,
839 * because that implies that the anon page is no longer mapped
840 * (and cannot be remapped so long as we hold the page lock).
Christoph Lametere24f0b82006-06-23 02:03:51 -0700841 */
Hugh Dickins03f15c82015-11-05 18:49:56 -0800842 if (PageAnon(page) && !PageKsm(page))
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700843 anon_vma = page_get_anon_vma(page);
Shaohua Li62e1c552008-02-04 22:29:33 -0800844
Hugh Dickins7db76712015-11-05 18:49:49 -0800845 /*
846 * Block others from accessing the new page when we get around to
847 * establishing additional references. We are usually the only one
848 * holding a reference to newpage at this point. We used to have a BUG
849 * here if trylock_page(newpage) fails, but would like to allow for
850 * cases where there might be a race with the previous use of newpage.
851 * This is much like races on refcount of oldpage: just don't BUG().
852 */
853 if (unlikely(!trylock_page(newpage)))
854 goto out_unlock;
855
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700856 if (unlikely(isolated_balloon_page(page))) {
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800857 /*
858 * A ballooned page does not need any special attention from
859 * physical to virtual reverse mapping procedures.
860 * Skip any attempt to unmap PTEs or to remap swap cache,
861 * in order to avoid burning cycles at rmap level, and perform
862 * the page migration right away (proteced by page lock).
863 */
864 rc = balloon_page_migrate(newpage, page, mode);
Hugh Dickins7db76712015-11-05 18:49:49 -0800865 goto out_unlock_both;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800866 }
867
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700868 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800869 * Corner case handling:
870 * 1. When a new swap-cache page is read into, it is added to the LRU
871 * and treated as swapcache but it has no rmap yet.
872 * Calling try_to_unmap() against a page->mapping==NULL page will
873 * trigger a BUG. So handle it here.
874 * 2. An orphaned page (see truncate_complete_page) might have
875 * fs-private metadata. The page can be picked up due to memory
876 * offlining. Everywhere else except page reclaim, the page is
877 * invisible to the vm, so the page can not be migrated. So try to
878 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700879 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800880 if (!page->mapping) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800881 VM_BUG_ON_PAGE(PageAnon(page), page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800882 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800883 try_to_free_buffers(page);
Hugh Dickins7db76712015-11-05 18:49:49 -0800884 goto out_unlock_both;
Shaohua Li62e1c552008-02-04 22:29:33 -0800885 }
Hugh Dickins7db76712015-11-05 18:49:49 -0800886 } else if (page_mapped(page)) {
887 /* Establish migration ptes */
Hugh Dickins03f15c82015-11-05 18:49:56 -0800888 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
889 page);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800890 try_to_unmap(page,
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700891 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800892 page_was_mapped = 1;
893 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700894
Christoph Lametere6a15302006-06-25 05:46:49 -0700895 if (!page_mapped(page))
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800896 rc = move_to_new_page(newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700897
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800898 if (page_was_mapped)
899 remove_migration_ptes(page,
900 rc == MIGRATEPAGE_SUCCESS ? newpage : page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700901
Hugh Dickins7db76712015-11-05 18:49:49 -0800902out_unlock_both:
903 unlock_page(newpage);
904out_unlock:
Mel Gorman3f6c8272010-05-24 14:32:17 -0700905 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700906 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700907 put_anon_vma(anon_vma);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700908 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700909out:
910 return rc;
911}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700912
Minchan Kim0dabec92011-10-31 17:06:57 -0700913/*
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -0700914 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
915 * around it.
916 */
917#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
918#define ICE_noinline noinline
919#else
920#define ICE_noinline
921#endif
922
923/*
Minchan Kim0dabec92011-10-31 17:06:57 -0700924 * Obtain the lock on page, remove all ptes and migrate the page
925 * to the newly allocated page in newpage.
926 */
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -0700927static ICE_noinline int unmap_and_move(new_page_t get_new_page,
928 free_page_t put_new_page,
929 unsigned long private, struct page *page,
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -0700930 int force, enum migrate_mode mode,
931 enum migrate_reason reason)
Minchan Kim0dabec92011-10-31 17:06:57 -0700932{
Hugh Dickins2def7422015-11-05 18:49:46 -0800933 int rc = MIGRATEPAGE_SUCCESS;
Minchan Kim0dabec92011-10-31 17:06:57 -0700934 int *result = NULL;
Hugh Dickins2def7422015-11-05 18:49:46 -0800935 struct page *newpage;
Minchan Kim0dabec92011-10-31 17:06:57 -0700936
Hugh Dickins2def7422015-11-05 18:49:46 -0800937 newpage = get_new_page(page, private, &result);
Minchan Kim0dabec92011-10-31 17:06:57 -0700938 if (!newpage)
939 return -ENOMEM;
940
941 if (page_count(page) == 1) {
942 /* page was freed from under us. So we are done. */
943 goto out;
944 }
945
946 if (unlikely(PageTransHuge(page)))
947 if (unlikely(split_huge_page(page)))
948 goto out;
949
Hugh Dickins9c620e22013-02-22 16:35:14 -0800950 rc = __unmap_and_move(page, newpage, force, mode);
Hugh Dickins2def7422015-11-05 18:49:46 -0800951 if (rc == MIGRATEPAGE_SUCCESS)
952 put_new_page = NULL;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800953
Minchan Kim0dabec92011-10-31 17:06:57 -0700954out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700955 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700956 /*
957 * A page that has been migrated has all references
958 * removed and will be freed. A page that has not been
959 * migrated will have kepts its references and be
960 * restored.
961 */
962 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700963 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700964 page_is_file_cache(page));
Naoya Horiguchif4c18e62015-08-06 15:47:08 -0700965 /* Soft-offlined page shouldn't go through lru cache list */
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700966 if (reason == MR_MEMORY_FAILURE) {
Naoya Horiguchif4c18e62015-08-06 15:47:08 -0700967 put_page(page);
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700968 if (!test_set_page_hwpoison(page))
969 num_poisoned_pages_inc();
970 } else
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -0700971 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700972 }
David Rientjes68711a72014-06-04 16:08:25 -0700973
Christoph Lameter95a402c2006-06-23 02:03:53 -0700974 /*
David Rientjes68711a72014-06-04 16:08:25 -0700975 * If migration was not successful and there's a freeing callback, use
976 * it. Otherwise, putback_lru_page() will drop the reference grabbed
977 * during isolation.
Christoph Lameter95a402c2006-06-23 02:03:53 -0700978 */
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800979 if (put_new_page)
David Rientjes68711a72014-06-04 16:08:25 -0700980 put_new_page(newpage, private);
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800981 else if (unlikely(__is_movable_balloon_page(newpage))) {
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700982 /* drop our reference, page already in the balloon */
983 put_page(newpage);
Hugh Dickins8bdd6382014-07-26 12:58:23 -0700984 } else
David Rientjes68711a72014-06-04 16:08:25 -0700985 putback_lru_page(newpage);
986
Christoph Lameter742755a2006-06-23 02:03:55 -0700987 if (result) {
988 if (rc)
989 *result = rc;
990 else
991 *result = page_to_nid(newpage);
992 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700993 return rc;
994}
995
996/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900997 * Counterpart of unmap_and_move_page() for hugepage migration.
998 *
999 * This function doesn't wait the completion of hugepage I/O
1000 * because there is no race between I/O and migration for hugepage.
1001 * Note that currently hugepage I/O occurs only in direct I/O
1002 * where no lock is held and PG_writeback is irrelevant,
1003 * and writeback status of all subpages are counted in the reference
1004 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1005 * under direct I/O, the reference of the head page is 512 and a bit more.)
1006 * This means that when we try to migrate hugepage whose subpages are
1007 * doing direct I/O, some references remain after try_to_unmap() and
1008 * hugepage migration fails without data corruption.
1009 *
1010 * There is also no race when direct I/O is issued on the page under migration,
1011 * because then pte is replaced with migration swap entry and direct I/O code
1012 * will wait in the page fault for migration to complete.
1013 */
1014static int unmap_and_move_huge_page(new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001015 free_page_t put_new_page, unsigned long private,
1016 struct page *hpage, int force,
1017 enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001018{
Hugh Dickins2def7422015-11-05 18:49:46 -08001019 int rc = -EAGAIN;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001020 int *result = NULL;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001021 int page_was_mapped = 0;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001022 struct page *new_hpage;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001023 struct anon_vma *anon_vma = NULL;
1024
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001025 /*
1026 * Movability of hugepages depends on architectures and hugepage size.
1027 * This check is necessary because some callers of hugepage migration
1028 * like soft offline and memory hotremove don't walk through page
1029 * tables or check whether the hugepage is pmd-based or not before
1030 * kicking migration.
1031 */
Naoya Horiguchi100873d2014-06-04 16:10:56 -07001032 if (!hugepage_migration_supported(page_hstate(hpage))) {
Joonsoo Kim32665f22014-01-21 15:51:15 -08001033 putback_active_hugepage(hpage);
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001034 return -ENOSYS;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001035 }
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001036
Joonsoo Kim32665f22014-01-21 15:51:15 -08001037 new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001038 if (!new_hpage)
1039 return -ENOMEM;
1040
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001041 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001042 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001043 goto out;
1044 lock_page(hpage);
1045 }
1046
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001047 if (PageAnon(hpage))
1048 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001049
Hugh Dickins7db76712015-11-05 18:49:49 -08001050 if (unlikely(!trylock_page(new_hpage)))
1051 goto put_anon;
1052
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001053 if (page_mapped(hpage)) {
1054 try_to_unmap(hpage,
1055 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1056 page_was_mapped = 1;
1057 }
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001058
1059 if (!page_mapped(hpage))
Hugh Dickins5c3f9a62015-11-05 18:49:53 -08001060 rc = move_to_new_page(new_hpage, hpage, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001061
Hugh Dickins5c3f9a62015-11-05 18:49:53 -08001062 if (page_was_mapped)
1063 remove_migration_ptes(hpage,
1064 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001065
Hugh Dickins7db76712015-11-05 18:49:49 -08001066 unlock_page(new_hpage);
1067
1068put_anon:
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001069 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001070 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001071
Hugh Dickins2def7422015-11-05 18:49:46 -08001072 if (rc == MIGRATEPAGE_SUCCESS) {
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001073 hugetlb_cgroup_migrate(hpage, new_hpage);
Hugh Dickins2def7422015-11-05 18:49:46 -08001074 put_new_page = NULL;
1075 }
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001076
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001077 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001078out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001079 if (rc != -EAGAIN)
1080 putback_active_hugepage(hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001081
1082 /*
1083 * If migration was not successful and there's a freeing callback, use
1084 * it. Otherwise, put_page() will drop the reference grabbed during
1085 * isolation.
1086 */
Hugh Dickins2def7422015-11-05 18:49:46 -08001087 if (put_new_page)
David Rientjes68711a72014-06-04 16:08:25 -07001088 put_new_page(new_hpage, private);
1089 else
Naoya Horiguchi3aaa76e2015-09-22 14:59:14 -07001090 putback_active_hugepage(new_hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001091
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001092 if (result) {
1093 if (rc)
1094 *result = rc;
1095 else
1096 *result = page_to_nid(new_hpage);
1097 }
1098 return rc;
1099}
1100
1101/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001102 * migrate_pages - migrate the pages specified in a list, to the free pages
1103 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001104 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001105 * @from: The list of pages to be migrated.
1106 * @get_new_page: The function used to allocate free pages to be used
1107 * as the target of the page migration.
David Rientjes68711a72014-06-04 16:08:25 -07001108 * @put_new_page: The function used to free target pages if migration
1109 * fails, or NULL if no special handling is necessary.
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001110 * @private: Private data to be passed on to get_new_page()
1111 * @mode: The migration mode that specifies the constraints for
1112 * page migration, if any.
1113 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001114 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001115 * The function returns after 10 attempts or if no pages are movable any more
1116 * because the list has become empty or no retryable pages exist any more.
Hugh Dickins14e0f9b2015-11-05 18:49:43 -08001117 * The caller should call putback_movable_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001118 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001119 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001120 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001121 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001122int migrate_pages(struct list_head *from, new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001123 free_page_t put_new_page, unsigned long private,
1124 enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001125{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001126 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001127 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001128 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001129 int pass = 0;
1130 struct page *page;
1131 struct page *page2;
1132 int swapwrite = current->flags & PF_SWAPWRITE;
1133 int rc;
1134
1135 if (!swapwrite)
1136 current->flags |= PF_SWAPWRITE;
1137
Christoph Lametere24f0b82006-06-23 02:03:51 -07001138 for(pass = 0; pass < 10 && retry; pass++) {
1139 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001140
Christoph Lametere24f0b82006-06-23 02:03:51 -07001141 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001142 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001143
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001144 if (PageHuge(page))
1145 rc = unmap_and_move_huge_page(get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001146 put_new_page, private, page,
1147 pass > 2, mode);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001148 else
David Rientjes68711a72014-06-04 16:08:25 -07001149 rc = unmap_and_move(get_new_page, put_new_page,
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -07001150 private, page, pass > 2, mode,
1151 reason);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001152
Christoph Lametere24f0b82006-06-23 02:03:51 -07001153 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001154 case -ENOMEM:
1155 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001156 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001157 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001158 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001159 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001160 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001161 break;
1162 default:
Naoya Horiguchi354a3362014-01-21 15:51:14 -08001163 /*
1164 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1165 * unlike -EAGAIN case, the failed page is
1166 * removed from migration page list and not
1167 * retried in the next outer loop.
1168 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001169 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001170 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001171 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001172 }
1173 }
Vlastimil Babkaf2f81fb2015-11-05 18:47:03 -08001174 nr_failed += retry;
1175 rc = nr_failed;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001176out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001177 if (nr_succeeded)
1178 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1179 if (nr_failed)
1180 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001181 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1182
Christoph Lameterb20a3502006-03-22 00:09:12 -08001183 if (!swapwrite)
1184 current->flags &= ~PF_SWAPWRITE;
1185
Rafael Aquini78bd5202012-12-11 16:02:31 -08001186 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001187}
1188
Christoph Lameter742755a2006-06-23 02:03:55 -07001189#ifdef CONFIG_NUMA
1190/*
1191 * Move a list of individual pages
1192 */
1193struct page_to_node {
1194 unsigned long addr;
1195 struct page *page;
1196 int node;
1197 int status;
1198};
1199
1200static struct page *new_page_node(struct page *p, unsigned long private,
1201 int **result)
1202{
1203 struct page_to_node *pm = (struct page_to_node *)private;
1204
1205 while (pm->node != MAX_NUMNODES && pm->page != p)
1206 pm++;
1207
1208 if (pm->node == MAX_NUMNODES)
1209 return NULL;
1210
1211 *result = &pm->status;
1212
Naoya Horiguchie632a932013-09-11 14:22:04 -07001213 if (PageHuge(p))
1214 return alloc_huge_page_node(page_hstate(compound_head(p)),
1215 pm->node);
1216 else
Vlastimil Babka96db8002015-09-08 15:03:50 -07001217 return __alloc_pages_node(pm->node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001218 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001219}
1220
1221/*
1222 * Move a set of pages as indicated in the pm array. The addr
1223 * field must be set to the virtual address of the page to be moved
1224 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001225 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001226 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001227static int do_move_page_to_node_array(struct mm_struct *mm,
1228 struct page_to_node *pm,
1229 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001230{
1231 int err;
1232 struct page_to_node *pp;
1233 LIST_HEAD(pagelist);
1234
1235 down_read(&mm->mmap_sem);
1236
1237 /*
1238 * Build a list of pages to migrate
1239 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001240 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1241 struct vm_area_struct *vma;
1242 struct page *page;
1243
Christoph Lameter742755a2006-06-23 02:03:55 -07001244 err = -EFAULT;
1245 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001246 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001247 goto set_status;
1248
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001249 /* FOLL_DUMP to ignore special (like zero) pages */
1250 page = follow_page(vma, pp->addr,
1251 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001252
1253 err = PTR_ERR(page);
1254 if (IS_ERR(page))
1255 goto set_status;
1256
Christoph Lameter742755a2006-06-23 02:03:55 -07001257 err = -ENOENT;
1258 if (!page)
1259 goto set_status;
1260
Christoph Lameter742755a2006-06-23 02:03:55 -07001261 pp->page = page;
1262 err = page_to_nid(page);
1263
1264 if (err == pp->node)
1265 /*
1266 * Node already in the right place
1267 */
1268 goto put_and_set;
1269
1270 err = -EACCES;
1271 if (page_mapcount(page) > 1 &&
1272 !migrate_all)
1273 goto put_and_set;
1274
Naoya Horiguchie632a932013-09-11 14:22:04 -07001275 if (PageHuge(page)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08001276 if (PageHead(page))
1277 isolate_huge_page(page, &pagelist);
Naoya Horiguchie632a932013-09-11 14:22:04 -07001278 goto put_and_set;
1279 }
1280
Nick Piggin62695a82008-10-18 20:26:09 -07001281 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001282 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001283 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001284 inc_zone_page_state(page, NR_ISOLATED_ANON +
1285 page_is_file_cache(page));
1286 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001287put_and_set:
1288 /*
1289 * Either remove the duplicate refcount from
1290 * isolate_lru_page() or drop the page ref if it was
1291 * not isolated.
1292 */
1293 put_page(page);
1294set_status:
1295 pp->status = err;
1296 }
1297
Brice Gogline78bbfa2008-10-18 20:27:15 -07001298 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001299 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -07001300 err = migrate_pages(&pagelist, new_page_node, NULL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001301 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001302 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001303 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001304 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001305
1306 up_read(&mm->mmap_sem);
1307 return err;
1308}
1309
1310/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001311 * Migrate an array of page address onto an array of nodes and fill
1312 * the corresponding array of status.
1313 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001314static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001315 unsigned long nr_pages,
1316 const void __user * __user *pages,
1317 const int __user *nodes,
1318 int __user *status, int flags)
1319{
Brice Goglin3140a222009-01-06 14:38:57 -08001320 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001321 unsigned long chunk_nr_pages;
1322 unsigned long chunk_start;
1323 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001324
Brice Goglin3140a222009-01-06 14:38:57 -08001325 err = -ENOMEM;
1326 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1327 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001328 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001329
1330 migrate_prep();
1331
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001332 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001333 * Store a chunk of page_to_node array in a page,
1334 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001335 */
Brice Goglin3140a222009-01-06 14:38:57 -08001336 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001337
Brice Goglin3140a222009-01-06 14:38:57 -08001338 for (chunk_start = 0;
1339 chunk_start < nr_pages;
1340 chunk_start += chunk_nr_pages) {
1341 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001342
Brice Goglin3140a222009-01-06 14:38:57 -08001343 if (chunk_start + chunk_nr_pages > nr_pages)
1344 chunk_nr_pages = nr_pages - chunk_start;
1345
1346 /* fill the chunk pm with addrs and nodes from user-space */
1347 for (j = 0; j < chunk_nr_pages; j++) {
1348 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001349 int node;
1350
Brice Goglin3140a222009-01-06 14:38:57 -08001351 err = -EFAULT;
1352 if (get_user(p, pages + j + chunk_start))
1353 goto out_pm;
1354 pm[j].addr = (unsigned long) p;
1355
1356 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001357 goto out_pm;
1358
1359 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001360 if (node < 0 || node >= MAX_NUMNODES)
1361 goto out_pm;
1362
Lai Jiangshan389162c2012-12-12 13:51:30 -08001363 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001364 goto out_pm;
1365
1366 err = -EACCES;
1367 if (!node_isset(node, task_nodes))
1368 goto out_pm;
1369
Brice Goglin3140a222009-01-06 14:38:57 -08001370 pm[j].node = node;
1371 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001372
Brice Goglin3140a222009-01-06 14:38:57 -08001373 /* End marker for this chunk */
1374 pm[chunk_nr_pages].node = MAX_NUMNODES;
1375
1376 /* Migrate this chunk */
1377 err = do_move_page_to_node_array(mm, pm,
1378 flags & MPOL_MF_MOVE_ALL);
1379 if (err < 0)
1380 goto out_pm;
1381
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001382 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001383 for (j = 0; j < chunk_nr_pages; j++)
1384 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001385 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001386 goto out_pm;
1387 }
1388 }
1389 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001390
1391out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001392 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001393out:
1394 return err;
1395}
1396
1397/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001398 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001399 */
Brice Goglin80bba122008-12-09 13:14:23 -08001400static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1401 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001402{
Brice Goglin2f007e72008-10-18 20:27:16 -07001403 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001404
Christoph Lameter742755a2006-06-23 02:03:55 -07001405 down_read(&mm->mmap_sem);
1406
Brice Goglin2f007e72008-10-18 20:27:16 -07001407 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001408 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001409 struct vm_area_struct *vma;
1410 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001411 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001412
1413 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001414 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001415 goto set_status;
1416
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001417 /* FOLL_DUMP to ignore special (like zero) pages */
1418 page = follow_page(vma, addr, FOLL_DUMP);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001419
1420 err = PTR_ERR(page);
1421 if (IS_ERR(page))
1422 goto set_status;
1423
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001424 err = page ? page_to_nid(page) : -ENOENT;
Christoph Lameter742755a2006-06-23 02:03:55 -07001425set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001426 *status = err;
1427
1428 pages++;
1429 status++;
1430 }
1431
1432 up_read(&mm->mmap_sem);
1433}
1434
1435/*
1436 * Determine the nodes of a user array of pages and store it in
1437 * a user array of status.
1438 */
1439static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1440 const void __user * __user *pages,
1441 int __user *status)
1442{
1443#define DO_PAGES_STAT_CHUNK_NR 16
1444 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1445 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001446
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001447 while (nr_pages) {
1448 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001449
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001450 chunk_nr = nr_pages;
1451 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1452 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1453
1454 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1455 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001456
1457 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1458
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001459 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1460 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001461
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001462 pages += chunk_nr;
1463 status += chunk_nr;
1464 nr_pages -= chunk_nr;
1465 }
1466 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001467}
1468
1469/*
1470 * Move a list of pages in the address space of the currently executing
1471 * process.
1472 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001473SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1474 const void __user * __user *, pages,
1475 const int __user *, nodes,
1476 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001477{
David Howellsc69e8d92008-11-14 10:39:19 +11001478 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001479 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001480 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001481 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001482 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001483
1484 /* Check flags */
1485 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1486 return -EINVAL;
1487
1488 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1489 return -EPERM;
1490
1491 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001492 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001493 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001494 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001495 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001496 return -ESRCH;
1497 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001498 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001499
1500 /*
1501 * Check if this process has the right to modify the specified
1502 * process. The right exists if the process has administrative
1503 * capabilities, superuser privileges or the same
1504 * userid as the target process.
1505 */
David Howellsc69e8d92008-11-14 10:39:19 +11001506 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001507 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1508 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001509 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001510 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001511 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001512 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001513 }
David Howellsc69e8d92008-11-14 10:39:19 +11001514 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001515
David Quigley86c3a762006-06-23 02:04:02 -07001516 err = security_task_movememory(task);
1517 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001518 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001519
Christoph Lameter3268c632012-03-21 16:34:06 -07001520 task_nodes = cpuset_mems_allowed(task);
1521 mm = get_task_mm(task);
1522 put_task_struct(task);
1523
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001524 if (!mm)
1525 return -EINVAL;
1526
1527 if (nodes)
1528 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1529 nodes, status, flags);
1530 else
1531 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001532
1533 mmput(mm);
1534 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001535
Christoph Lameter742755a2006-06-23 02:03:55 -07001536out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001537 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001538 return err;
1539}
Christoph Lameter742755a2006-06-23 02:03:55 -07001540
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001541#ifdef CONFIG_NUMA_BALANCING
1542/*
1543 * Returns true if this is a safe migration target node for misplaced NUMA
1544 * pages. Currently it only checks the watermarks which crude
1545 */
1546static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001547 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001548{
1549 int z;
1550 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1551 struct zone *zone = pgdat->node_zones + z;
1552
1553 if (!populated_zone(zone))
1554 continue;
1555
Lisa Du6e543d52013-09-11 14:22:36 -07001556 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001557 continue;
1558
1559 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1560 if (!zone_watermark_ok(zone, 0,
1561 high_wmark_pages(zone) +
1562 nr_migrate_pages,
1563 0, 0))
1564 continue;
1565 return true;
1566 }
1567 return false;
1568}
1569
1570static struct page *alloc_misplaced_dst_page(struct page *page,
1571 unsigned long data,
1572 int **result)
1573{
1574 int nid = (int) data;
1575 struct page *newpage;
1576
Vlastimil Babka96db8002015-09-08 15:03:50 -07001577 newpage = __alloc_pages_node(nid,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001578 (GFP_HIGHUSER_MOVABLE |
1579 __GFP_THISNODE | __GFP_NOMEMALLOC |
1580 __GFP_NORETRY | __GFP_NOWARN) &
Mel Gormand0164ad2015-11-06 16:28:21 -08001581 ~(__GFP_IO | __GFP_FS), 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001582
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001583 return newpage;
1584}
1585
1586/*
Mel Gormana8f60772012-11-14 21:41:46 +00001587 * page migration rate limiting control.
1588 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1589 * window of time. Default here says do not migrate more than 1280M per second.
1590 */
1591static unsigned int migrate_interval_millisecs __read_mostly = 100;
1592static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1593
Mel Gormanb32967f2012-11-19 12:35:47 +00001594/* Returns true if the node is migrate rate-limited after the update */
Mel Gorman1c30e012014-01-21 15:50:58 -08001595static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1596 unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001597{
Mel Gormanb32967f2012-11-19 12:35:47 +00001598 /*
1599 * Rate-limit the amount of data that is being migrated to a node.
1600 * Optimal placement is no good if the memory bus is saturated and
1601 * all the time is being spent migrating!
1602 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001603 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001604 spin_lock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001605 pgdat->numabalancing_migrate_nr_pages = 0;
1606 pgdat->numabalancing_migrate_next_window = jiffies +
1607 msecs_to_jiffies(migrate_interval_millisecs);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001608 spin_unlock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001609 }
Mel Gormanaf1839d2014-01-21 15:51:01 -08001610 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1611 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1612 nr_pages);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001613 return true;
Mel Gormanaf1839d2014-01-21 15:51:01 -08001614 }
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001615
1616 /*
1617 * This is an unlocked non-atomic update so errors are possible.
1618 * The consequences are failing to migrate when we potentiall should
1619 * have which is not severe enough to warrant locking. If it is ever
1620 * a problem, it can be converted to a per-cpu counter.
1621 */
1622 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1623 return false;
Mel Gormanb32967f2012-11-19 12:35:47 +00001624}
1625
Mel Gorman1c30e012014-01-21 15:50:58 -08001626static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
Mel Gormanb32967f2012-11-19 12:35:47 +00001627{
Hugh Dickins340ef392013-02-22 16:34:33 -08001628 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001629
Sasha Levin309381fea2014-01-23 15:52:54 -08001630 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
Mel Gorman3abef4e2013-02-22 16:34:27 -08001631
Mel Gormanb32967f2012-11-19 12:35:47 +00001632 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001633 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1634 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001635
Hugh Dickins340ef392013-02-22 16:34:33 -08001636 if (isolate_lru_page(page))
1637 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001638
1639 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001640 * migrate_misplaced_transhuge_page() skips page migration's usual
1641 * check on page_count(), so we must do it here, now that the page
1642 * has been isolated: a GUP pin, or any other pin, prevents migration.
1643 * The expected page count is 3: 1 for page's mapcount and 1 for the
1644 * caller's pin and 1 for the reference taken by isolate_lru_page().
1645 */
1646 if (PageTransHuge(page) && page_count(page) != 3) {
1647 putback_lru_page(page);
1648 return 0;
1649 }
1650
1651 page_lru = page_is_file_cache(page);
1652 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1653 hpage_nr_pages(page));
1654
1655 /*
1656 * Isolating the page has taken another reference, so the
1657 * caller's reference can be safely dropped without the page
1658 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001659 */
1660 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001661 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001662}
1663
Mel Gormande466bd2013-12-18 17:08:42 -08001664bool pmd_trans_migrating(pmd_t pmd)
1665{
1666 struct page *page = pmd_page(pmd);
1667 return PageLocked(page);
1668}
1669
Mel Gormana8f60772012-11-14 21:41:46 +00001670/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001671 * Attempt to migrate a misplaced page to the specified destination
1672 * node. Caller is expected to have an elevated reference count on
1673 * the page that will be dropped by this function before returning.
1674 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001675int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1676 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001677{
Mel Gormana8f60772012-11-14 21:41:46 +00001678 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001679 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001680 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001681 LIST_HEAD(migratepages);
1682
1683 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001684 * Don't migrate file pages that are mapped in multiple processes
1685 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001686 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001687 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1688 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001689 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001690
Mel Gormana8f60772012-11-14 21:41:46 +00001691 /*
1692 * Rate-limit the amount of data that is being migrated to a node.
1693 * Optimal placement is no good if the memory bus is saturated and
1694 * all the time is being spent migrating!
1695 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001696 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001697 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001698
Mel Gormanb32967f2012-11-19 12:35:47 +00001699 isolated = numamigrate_isolate_page(pgdat, page);
1700 if (!isolated)
1701 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001702
Mel Gormanb32967f2012-11-19 12:35:47 +00001703 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001704 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
David Rientjes68711a72014-06-04 16:08:25 -07001705 NULL, node, MIGRATE_ASYNC,
1706 MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001707 if (nr_remaining) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001708 if (!list_empty(&migratepages)) {
1709 list_del(&page->lru);
1710 dec_zone_page_state(page, NR_ISOLATED_ANON +
1711 page_is_file_cache(page));
1712 putback_lru_page(page);
1713 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001714 isolated = 0;
1715 } else
1716 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001717 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001718 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001719
1720out:
1721 put_page(page);
1722 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001723}
Mel Gorman220018d2012-12-05 09:32:56 +00001724#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001725
Mel Gorman220018d2012-12-05 09:32:56 +00001726#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001727/*
1728 * Migrates a THP to a given target node. page must be locked and is unlocked
1729 * before returning.
1730 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001731int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1732 struct vm_area_struct *vma,
1733 pmd_t *pmd, pmd_t entry,
1734 unsigned long address,
1735 struct page *page, int node)
1736{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001737 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001738 pg_data_t *pgdat = NODE_DATA(node);
1739 int isolated = 0;
1740 struct page *new_page = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001741 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001742 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1743 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001744 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001745
1746 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001747 * Rate-limit the amount of data that is being migrated to a node.
1748 * Optimal placement is no good if the memory bus is saturated and
1749 * all the time is being spent migrating!
1750 */
Mel Gormand28d43352012-11-29 09:24:36 +00001751 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001752 goto out_dropref;
1753
1754 new_page = alloc_pages_node(node,
Mel Gorman71baba42015-11-06 16:28:28 -08001755 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001756 HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001757 if (!new_page)
1758 goto out_fail;
1759
Mel Gormanb32967f2012-11-19 12:35:47 +00001760 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001761 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001762 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001763 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001764 }
1765
Mel Gormanb0943d62013-12-18 17:08:46 -08001766 if (mm_tlb_flush_pending(mm))
1767 flush_tlb_range(vma, mmun_start, mmun_end);
1768
Mel Gormanb32967f2012-11-19 12:35:47 +00001769 /* Prepare a page as a migration target */
1770 __set_page_locked(new_page);
1771 SetPageSwapBacked(new_page);
1772
1773 /* anon mapping, we can simply copy page->mapping to the new page: */
1774 new_page->mapping = page->mapping;
1775 new_page->index = page->index;
1776 migrate_page_copy(new_page, page);
1777 WARN_ON(PageLRU(new_page));
1778
1779 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001780 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001781 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001782 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1783fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001784 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001785 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001786
1787 /* Reverse changes made by migrate_page_copy() */
1788 if (TestClearPageActive(new_page))
1789 SetPageActive(page);
1790 if (TestClearPageUnevictable(new_page))
1791 SetPageUnevictable(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001792
1793 unlock_page(new_page);
1794 put_page(new_page); /* Free it */
1795
Mel Gormana54a4072013-10-07 11:28:46 +01001796 /* Retake the callers reference and putback on LRU */
1797 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001798 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001799 mod_zone_page_state(page_zone(page),
1800 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001801
1802 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001803 }
1804
Mel Gorman2b4847e2013-12-18 17:08:32 -08001805 orig_entry = *pmd;
Mel Gormanb32967f2012-11-19 12:35:47 +00001806 entry = mk_pmd(new_page, vma->vm_page_prot);
Mel Gormanb32967f2012-11-19 12:35:47 +00001807 entry = pmd_mkhuge(entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001808 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001809
Mel Gorman2b4847e2013-12-18 17:08:32 -08001810 /*
1811 * Clear the old entry under pagetable lock and establish the new PTE.
1812 * Any parallel GUP will either observe the old page blocking on the
1813 * page lock, block on the page table lock or observe the new page.
1814 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1815 * guarantee the copy is visible before the pagetable update.
1816 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001817 flush_cache_range(vma, mmun_start, mmun_end);
Mel Gorman11de9922014-06-04 16:07:41 -07001818 page_add_anon_rmap(new_page, vma, mmun_start);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001819 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001820 set_pmd_at(mm, mmun_start, pmd, entry);
1821 flush_tlb_range(vma, mmun_start, mmun_end);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001822 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001823
1824 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08001825 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1826 flush_tlb_range(vma, mmun_start, mmun_end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001827 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001828 update_mmu_cache_pmd(vma, address, &entry);
1829 page_remove_rmap(new_page);
1830 goto fail_putback;
1831 }
1832
Hugh Dickins51afb122015-11-05 18:49:37 -08001833 mlock_migrate_page(new_page, page);
Hugh Dickins45637ba2015-11-05 18:49:40 -08001834 set_page_memcg(new_page, page_memcg(page));
1835 set_page_memcg(page, NULL);
Mel Gormanb32967f2012-11-19 12:35:47 +00001836 page_remove_rmap(page);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001837
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001838 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001839 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001840
Mel Gorman11de9922014-06-04 16:07:41 -07001841 /* Take an "isolate" reference and put new page on the LRU. */
1842 get_page(new_page);
1843 putback_lru_page(new_page);
1844
Mel Gormanb32967f2012-11-19 12:35:47 +00001845 unlock_page(new_page);
1846 unlock_page(page);
1847 put_page(page); /* Drop the rmap reference */
1848 put_page(page); /* Drop the LRU isolation reference */
1849
1850 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1851 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1852
Mel Gormanb32967f2012-11-19 12:35:47 +00001853 mod_zone_page_state(page_zone(page),
1854 NR_ISOLATED_ANON + page_lru,
1855 -HPAGE_PMD_NR);
1856 return isolated;
1857
Hugh Dickins340ef392013-02-22 16:34:33 -08001858out_fail:
1859 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001860out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08001861 ptl = pmd_lock(mm, pmd);
1862 if (pmd_same(*pmd, entry)) {
Mel Gorman4d942462015-02-12 14:58:28 -08001863 entry = pmd_modify(entry, vma->vm_page_prot);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001864 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001865 update_mmu_cache_pmd(vma, address, &entry);
1866 }
1867 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001868
Mel Gormaneb4489f62013-12-18 17:08:39 -08001869out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08001870 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001871 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001872 return 0;
1873}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001874#endif /* CONFIG_NUMA_BALANCING */
1875
1876#endif /* CONFIG_NUMA */