blob: 9baf41c877ffb98595aac0834b8aa43aedd6dc76 [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
Kirill A. Shutemove3884662016-03-17 14:20:07 -0700175 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
Hugh Dickins51afb122015-11-05 18:49:37 -0800176 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 */
Kirill A. Shutemove3884662016-03-17 14:20:07 -0700190void remove_migration_ptes(struct page *old, struct page *new, bool locked)
Christoph Lameter04e62a22006-06-23 02:03:38 -0700191{
Joonsoo Kim051ac832014-01-21 15:49:48 -0800192 struct rmap_walk_control rwc = {
193 .rmap_one = remove_migration_pte,
194 .arg = old,
195 };
196
Kirill A. Shutemove3884662016-03-17 14:20:07 -0700197 if (locked)
198 rmap_walk_locked(new, &rwc);
199 else
200 rmap_walk(new, &rwc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700201}
202
203/*
Christoph Lameter06972122006-06-23 02:03:35 -0700204 * Something used the pte of a page under migration. We need to
205 * get to the page and wait until migration is finished.
206 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700207 */
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800208void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700209 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700210{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700211 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700212 swp_entry_t entry;
213 struct page *page;
214
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700215 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700216 pte = *ptep;
217 if (!is_swap_pte(pte))
218 goto out;
219
220 entry = pte_to_swp_entry(pte);
221 if (!is_migration_entry(entry))
222 goto out;
223
224 page = migration_entry_to_page(entry);
225
Nick Piggine2867812008-07-25 19:45:30 -0700226 /*
227 * Once radix-tree replacement of page migration started, page_count
228 * *must* be zero. And, we don't want to call wait_on_page_locked()
229 * against a page without get_page().
230 * So, we use get_page_unless_zero(), here. Even failed, page fault
231 * will occur again.
232 */
233 if (!get_page_unless_zero(page))
234 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700235 pte_unmap_unlock(ptep, ptl);
236 wait_on_page_locked(page);
237 put_page(page);
238 return;
239out:
240 pte_unmap_unlock(ptep, ptl);
241}
242
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700243void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
244 unsigned long address)
245{
246 spinlock_t *ptl = pte_lockptr(mm, pmd);
247 pte_t *ptep = pte_offset_map(pmd, address);
248 __migration_entry_wait(mm, ptep, ptl);
249}
250
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800251void migration_entry_wait_huge(struct vm_area_struct *vma,
252 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700253{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800254 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700255 __migration_entry_wait(mm, pte, ptl);
256}
257
Mel Gormanb969c4a2012-01-12 17:19:34 -0800258#ifdef CONFIG_BLOCK
259/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800260static bool buffer_migrate_lock_buffers(struct buffer_head *head,
261 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800262{
263 struct buffer_head *bh = head;
264
265 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800266 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800267 do {
268 get_bh(bh);
269 lock_buffer(bh);
270 bh = bh->b_this_page;
271
272 } while (bh != head);
273
274 return true;
275 }
276
277 /* async case, we cannot block on lock_buffer so use trylock_buffer */
278 do {
279 get_bh(bh);
280 if (!trylock_buffer(bh)) {
281 /*
282 * We failed to lock the buffer and cannot stall in
283 * async migration. Release the taken locks
284 */
285 struct buffer_head *failed_bh = bh;
286 put_bh(failed_bh);
287 bh = head;
288 while (bh != failed_bh) {
289 unlock_buffer(bh);
290 put_bh(bh);
291 bh = bh->b_this_page;
292 }
293 return false;
294 }
295
296 bh = bh->b_this_page;
297 } while (bh != head);
298 return true;
299}
300#else
301static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800302 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800303{
304 return true;
305}
306#endif /* CONFIG_BLOCK */
307
Christoph Lameterb20a3502006-03-22 00:09:12 -0800308/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700309 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700310 *
311 * The number of remaining references must be:
312 * 1 for anonymous pages without a mapping
313 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100314 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800315 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800316int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800317 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500318 struct buffer_head *head, enum migrate_mode mode,
319 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800320{
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800321 struct zone *oldzone, *newzone;
322 int dirty;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500323 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800324 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800325
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700326 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700327 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500328 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700329 return -EAGAIN;
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800330
331 /* No turning back from here */
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800332 newpage->index = page->index;
333 newpage->mapping = page->mapping;
334 if (PageSwapBacked(page))
Hugh Dickinsfa9949d2016-05-19 17:12:41 -0700335 __SetPageSwapBacked(newpage);
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800336
Rafael Aquini78bd5202012-12-11 16:02:31 -0800337 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700338 }
339
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800340 oldzone = page_zone(page);
341 newzone = page_zone(newpage);
342
Nick Piggin19fd6232008-07-25 19:45:32 -0700343 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800344
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800345 pslot = radix_tree_lookup_slot(&mapping->page_tree,
346 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800347
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500348 expected_count += 1 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700349 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800350 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700351 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700352 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800353 }
354
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700355 if (!page_ref_freeze(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700356 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700357 return -EAGAIN;
358 }
359
Christoph Lameterb20a3502006-03-22 00:09:12 -0800360 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800361 * In the async migration case of moving a page with buffers, lock the
362 * buffers using trylock before the mapping is moved. If the mapping
363 * was moved, we later failed to lock the buffers and could not move
364 * the mapping back due to an elevated page count, we would have to
365 * block waiting on other references to be dropped.
366 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800367 if (mode == MIGRATE_ASYNC && head &&
368 !buffer_migrate_lock_buffers(head, mode)) {
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700369 page_ref_unfreeze(page, expected_count);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800370 spin_unlock_irq(&mapping->tree_lock);
371 return -EAGAIN;
372 }
373
374 /*
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800375 * Now we know that no one else is looking at the page:
376 * no turning back from here.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800377 */
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800378 newpage->index = page->index;
379 newpage->mapping = page->mapping;
380 if (PageSwapBacked(page))
Hugh Dickinsfa9949d2016-05-19 17:12:41 -0700381 __SetPageSwapBacked(newpage);
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800382
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800383 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800384 if (PageSwapCache(page)) {
385 SetPageSwapCache(newpage);
386 set_page_private(newpage, page_private(page));
387 }
388
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800389 /* Move dirty while page refs frozen and newpage not yet exposed */
390 dirty = PageDirty(page);
391 if (dirty) {
392 ClearPageDirty(page);
393 SetPageDirty(newpage);
394 }
395
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800396 radix_tree_replace_slot(pslot, newpage);
397
398 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800399 * Drop cache reference from old page by unfreezing
400 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800401 * We know this isn't the last reference.
402 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700403 page_ref_unfreeze(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800404
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800405 spin_unlock(&mapping->tree_lock);
406 /* Leave irq disabled to prevent preemption while updating stats */
407
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700408 /*
409 * If moved to a different zone then also account
410 * the page for that zone. Other VM counters will be
411 * taken care of when we establish references to the
412 * new page and drop references to the old page.
413 *
414 * Note that anonymous pages are accounted for
415 * via NR_FILE_PAGES and NR_ANON_PAGES if they
416 * are mapped to swap space.
417 */
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800418 if (newzone != oldzone) {
419 __dec_zone_state(oldzone, NR_FILE_PAGES);
420 __inc_zone_state(newzone, NR_FILE_PAGES);
421 if (PageSwapBacked(page) && !PageSwapCache(page)) {
422 __dec_zone_state(oldzone, NR_SHMEM);
423 __inc_zone_state(newzone, NR_SHMEM);
424 }
425 if (dirty && mapping_cap_account_dirty(mapping)) {
426 __dec_zone_state(oldzone, NR_FILE_DIRTY);
427 __inc_zone_state(newzone, NR_FILE_DIRTY);
428 }
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700429 }
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800430 local_irq_enable();
Christoph Lameterb20a3502006-03-22 00:09:12 -0800431
Rafael Aquini78bd5202012-12-11 16:02:31 -0800432 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800433}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800434
435/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900436 * The expected number of remaining references is the same as that
437 * of migrate_page_move_mapping().
438 */
439int migrate_huge_page_move_mapping(struct address_space *mapping,
440 struct page *newpage, struct page *page)
441{
442 int expected_count;
443 void **pslot;
444
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900445 spin_lock_irq(&mapping->tree_lock);
446
447 pslot = radix_tree_lookup_slot(&mapping->page_tree,
448 page_index(page));
449
450 expected_count = 2 + page_has_private(page);
451 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800452 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900453 spin_unlock_irq(&mapping->tree_lock);
454 return -EAGAIN;
455 }
456
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700457 if (!page_ref_freeze(page, expected_count)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900458 spin_unlock_irq(&mapping->tree_lock);
459 return -EAGAIN;
460 }
461
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800462 newpage->index = page->index;
463 newpage->mapping = page->mapping;
Johannes Weiner6a93ca82016-03-15 14:57:19 -0700464
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900465 get_page(newpage);
466
467 radix_tree_replace_slot(pslot, newpage);
468
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700469 page_ref_unfreeze(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900470
471 spin_unlock_irq(&mapping->tree_lock);
Johannes Weiner6a93ca82016-03-15 14:57:19 -0700472
Rafael Aquini78bd5202012-12-11 16:02:31 -0800473 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900474}
475
476/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800477 * Gigantic pages are so large that we do not guarantee that page++ pointer
478 * arithmetic will work across the entire page. We need something more
479 * specialized.
480 */
481static void __copy_gigantic_page(struct page *dst, struct page *src,
482 int nr_pages)
483{
484 int i;
485 struct page *dst_base = dst;
486 struct page *src_base = src;
487
488 for (i = 0; i < nr_pages; ) {
489 cond_resched();
490 copy_highpage(dst, src);
491
492 i++;
493 dst = mem_map_next(dst, dst_base, i);
494 src = mem_map_next(src, src_base, i);
495 }
496}
497
498static void copy_huge_page(struct page *dst, struct page *src)
499{
500 int i;
501 int nr_pages;
502
503 if (PageHuge(src)) {
504 /* hugetlbfs page */
505 struct hstate *h = page_hstate(src);
506 nr_pages = pages_per_huge_page(h);
507
508 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
509 __copy_gigantic_page(dst, src, nr_pages);
510 return;
511 }
512 } else {
513 /* thp page */
514 BUG_ON(!PageTransHuge(src));
515 nr_pages = hpage_nr_pages(src);
516 }
517
518 for (i = 0; i < nr_pages; i++) {
519 cond_resched();
520 copy_highpage(dst + i, src + i);
521 }
522}
523
524/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800525 * Copy the page to its new location
526 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900527void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800528{
Rik van Riel7851a452013-10-07 11:29:23 +0100529 int cpupid;
530
Mel Gormanb32967f2012-11-19 12:35:47 +0000531 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900532 copy_huge_page(newpage, page);
533 else
534 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800535
536 if (PageError(page))
537 SetPageError(newpage);
538 if (PageReferenced(page))
539 SetPageReferenced(newpage);
540 if (PageUptodate(page))
541 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700542 if (TestClearPageActive(page)) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800543 VM_BUG_ON_PAGE(PageUnevictable(page), page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800544 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800545 } else if (TestClearPageUnevictable(page))
546 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800547 if (PageChecked(page))
548 SetPageChecked(newpage);
549 if (PageMappedToDisk(page))
550 SetPageMappedToDisk(newpage);
551
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800552 /* Move dirty on pages not done by migrate_page_move_mapping() */
553 if (PageDirty(page))
554 SetPageDirty(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800555
Vladimir Davydov33c3fc72015-09-09 15:35:45 -0700556 if (page_is_young(page))
557 set_page_young(newpage);
558 if (page_is_idle(page))
559 set_page_idle(newpage);
560
Rik van Riel7851a452013-10-07 11:29:23 +0100561 /*
562 * Copy NUMA information to the new page, to prevent over-eager
563 * future migrations of this same page.
564 */
565 cpupid = page_cpupid_xchg_last(page, -1);
566 page_cpupid_xchg_last(newpage, cpupid);
567
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800568 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800569 /*
570 * Please do not reorder this without considering how mm/ksm.c's
571 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
572 */
Naoya Horiguchib3b3a992015-04-15 16:13:15 -0700573 if (PageSwapCache(page))
574 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800575 ClearPagePrivate(page);
576 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800577
578 /*
579 * If any waiters have accumulated on the new page then
580 * wake them up.
581 */
582 if (PageWriteback(newpage))
583 end_page_writeback(newpage);
Vlastimil Babkad435edc2016-03-15 14:56:15 -0700584
585 copy_page_owner(page, newpage);
Johannes Weiner74485cf2016-03-15 14:57:54 -0700586
587 mem_cgroup_migrate(page, newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800588}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800589
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700590/************************************************************
591 * Migration functions
592 ***********************************************************/
593
Christoph Lameterb20a3502006-03-22 00:09:12 -0800594/*
595 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100596 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800597 *
598 * Pages are locked upon entry and exit.
599 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700600int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800601 struct page *newpage, struct page *page,
602 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800603{
604 int rc;
605
606 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
607
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500608 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800609
Rafael Aquini78bd5202012-12-11 16:02:31 -0800610 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800611 return rc;
612
613 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800614 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800615}
616EXPORT_SYMBOL(migrate_page);
617
David Howells93614012006-09-30 20:45:40 +0200618#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800619/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700620 * Migration function for pages with buffers. This function can only be used
621 * if the underlying filesystem guarantees that no other references to "page"
622 * exist.
623 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700624int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800625 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700626{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700627 struct buffer_head *bh, *head;
628 int rc;
629
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700630 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800631 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700632
633 head = page_buffers(page);
634
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500635 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700636
Rafael Aquini78bd5202012-12-11 16:02:31 -0800637 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700638 return rc;
639
Mel Gormanb969c4a2012-01-12 17:19:34 -0800640 /*
641 * In the async case, migrate_page_move_mapping locked the buffers
642 * with an IRQ-safe spinlock held. In the sync case, the buffers
643 * need to be locked now
644 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800645 if (mode != MIGRATE_ASYNC)
646 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700647
648 ClearPagePrivate(page);
649 set_page_private(newpage, page_private(page));
650 set_page_private(page, 0);
651 put_page(page);
652 get_page(newpage);
653
654 bh = head;
655 do {
656 set_bh_page(bh, newpage, bh_offset(bh));
657 bh = bh->b_this_page;
658
659 } while (bh != head);
660
661 SetPagePrivate(newpage);
662
663 migrate_page_copy(newpage, page);
664
665 bh = head;
666 do {
667 unlock_buffer(bh);
668 put_bh(bh);
669 bh = bh->b_this_page;
670
671 } while (bh != head);
672
Rafael Aquini78bd5202012-12-11 16:02:31 -0800673 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700674}
675EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200676#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700677
Christoph Lameter04e62a22006-06-23 02:03:38 -0700678/*
679 * Writeback a page to clean the dirty state
680 */
681static int writeout(struct address_space *mapping, struct page *page)
682{
683 struct writeback_control wbc = {
684 .sync_mode = WB_SYNC_NONE,
685 .nr_to_write = 1,
686 .range_start = 0,
687 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700688 .for_reclaim = 1
689 };
690 int rc;
691
692 if (!mapping->a_ops->writepage)
693 /* No write method for the address space */
694 return -EINVAL;
695
696 if (!clear_page_dirty_for_io(page))
697 /* Someone else already triggered a write */
698 return -EAGAIN;
699
700 /*
701 * A dirty page may imply that the underlying filesystem has
702 * the page on some queue. So the page must be clean for
703 * migration. Writeout may mean we loose the lock and the
704 * page state is no longer what we checked for earlier.
705 * At this point we know that the migration attempt cannot
706 * be successful.
707 */
Kirill A. Shutemove3884662016-03-17 14:20:07 -0700708 remove_migration_ptes(page, page, false);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700709
710 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700711
712 if (rc != AOP_WRITEPAGE_ACTIVATE)
713 /* unlocked. Relock */
714 lock_page(page);
715
Hugh Dickinsbda85502008-11-19 15:36:36 -0800716 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700717}
718
719/*
720 * Default handling if a filesystem does not provide a migration function.
721 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700722static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800723 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700724{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800725 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800726 /* Only writeback pages in full synchronous migration */
727 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800728 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700729 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800730 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700731
732 /*
733 * Buffers may be managed in a filesystem specific way.
734 * We must have no buffers or drop them.
735 */
David Howells266cf652009-04-03 16:42:36 +0100736 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700737 !try_to_release_page(page, GFP_KERNEL))
738 return -EAGAIN;
739
Mel Gormana6bc32b2012-01-12 17:19:43 -0800740 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700741}
742
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700743/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700744 * Move a page to a newly allocated page
745 * The page is locked and all ptes have been successfully removed.
746 *
747 * The new page will have replaced the old page if this function
748 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700749 *
750 * Return value:
751 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800752 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700753 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700754static int move_to_new_page(struct page *newpage, struct page *page,
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800755 enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700756{
757 struct address_space *mapping;
758 int rc;
759
Hugh Dickins7db76712015-11-05 18:49:49 -0800760 VM_BUG_ON_PAGE(!PageLocked(page), page);
761 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700762
Christoph Lametere24f0b82006-06-23 02:03:51 -0700763 mapping = page_mapping(page);
764 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800765 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800766 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700767 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800768 * Most pages have a mapping and most filesystems provide a
769 * migratepage callback. Anonymous pages are part of swap
770 * space which also has its own migratepage callback. This
771 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700772 */
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800773 rc = mapping->a_ops->migratepage(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800774 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800775 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700776
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800777 /*
778 * When successful, old pagecache page->mapping must be cleared before
779 * page is freed; but stats require that PageAnon be left as PageAnon.
780 */
781 if (rc == MIGRATEPAGE_SUCCESS) {
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800782 if (!PageAnon(page))
783 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700784 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700785 return rc;
786}
787
Minchan Kim0dabec92011-10-31 17:06:57 -0700788static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800789 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700790{
Minchan Kim0dabec92011-10-31 17:06:57 -0700791 int rc = -EAGAIN;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800792 int page_was_mapped = 0;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700793 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700794
Nick Piggin529ae9a2008-08-02 12:01:03 +0200795 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800796 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700797 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800798
799 /*
800 * It's not safe for direct compaction to call lock_page.
801 * For example, during page readahead pages are added locked
802 * to the LRU. Later, when the IO completes the pages are
803 * marked uptodate and unlocked. However, the queueing
804 * could be merging multiple pages for one bio (e.g.
805 * mpage_readpages). If an allocation happens for the
806 * second or third page, the process can end up locking
807 * the same page twice and deadlocking. Rather than
808 * trying to be clever about what pages can be locked,
809 * avoid the use of lock_page for direct compaction
810 * altogether.
811 */
812 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700813 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800814
Christoph Lametere24f0b82006-06-23 02:03:51 -0700815 lock_page(page);
816 }
817
818 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700819 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700820 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800821 * necessary to wait for PageWriteback. In the async case,
822 * the retry loop is too short and in the sync-light case,
823 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700824 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800825 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700826 rc = -EBUSY;
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700827 goto out_unlock;
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700828 }
829 if (!force)
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700830 goto out_unlock;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700831 wait_on_page_writeback(page);
832 }
Hugh Dickins03f15c82015-11-05 18:49:56 -0800833
Christoph Lametere24f0b82006-06-23 02:03:51 -0700834 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700835 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
836 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800837 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700838 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700839 * File Caches may use write_page() or lock_page() in migration, then,
840 * just care Anon page here.
Hugh Dickins03f15c82015-11-05 18:49:56 -0800841 *
842 * Only page_get_anon_vma() understands the subtleties of
843 * getting a hold on an anon_vma from outside one of its mms.
844 * But if we cannot get anon_vma, then we won't need it anyway,
845 * because that implies that the anon page is no longer mapped
846 * (and cannot be remapped so long as we hold the page lock).
Christoph Lametere24f0b82006-06-23 02:03:51 -0700847 */
Hugh Dickins03f15c82015-11-05 18:49:56 -0800848 if (PageAnon(page) && !PageKsm(page))
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700849 anon_vma = page_get_anon_vma(page);
Shaohua Li62e1c552008-02-04 22:29:33 -0800850
Hugh Dickins7db76712015-11-05 18:49:49 -0800851 /*
852 * Block others from accessing the new page when we get around to
853 * establishing additional references. We are usually the only one
854 * holding a reference to newpage at this point. We used to have a BUG
855 * here if trylock_page(newpage) fails, but would like to allow for
856 * cases where there might be a race with the previous use of newpage.
857 * This is much like races on refcount of oldpage: just don't BUG().
858 */
859 if (unlikely(!trylock_page(newpage)))
860 goto out_unlock;
861
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700862 if (unlikely(isolated_balloon_page(page))) {
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800863 /*
864 * A ballooned page does not need any special attention from
865 * physical to virtual reverse mapping procedures.
866 * Skip any attempt to unmap PTEs or to remap swap cache,
867 * in order to avoid burning cycles at rmap level, and perform
868 * the page migration right away (proteced by page lock).
869 */
870 rc = balloon_page_migrate(newpage, page, mode);
Hugh Dickins7db76712015-11-05 18:49:49 -0800871 goto out_unlock_both;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800872 }
873
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700874 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800875 * Corner case handling:
876 * 1. When a new swap-cache page is read into, it is added to the LRU
877 * and treated as swapcache but it has no rmap yet.
878 * Calling try_to_unmap() against a page->mapping==NULL page will
879 * trigger a BUG. So handle it here.
880 * 2. An orphaned page (see truncate_complete_page) might have
881 * fs-private metadata. The page can be picked up due to memory
882 * offlining. Everywhere else except page reclaim, the page is
883 * invisible to the vm, so the page can not be migrated. So try to
884 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700885 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800886 if (!page->mapping) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800887 VM_BUG_ON_PAGE(PageAnon(page), page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800888 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800889 try_to_free_buffers(page);
Hugh Dickins7db76712015-11-05 18:49:49 -0800890 goto out_unlock_both;
Shaohua Li62e1c552008-02-04 22:29:33 -0800891 }
Hugh Dickins7db76712015-11-05 18:49:49 -0800892 } else if (page_mapped(page)) {
893 /* Establish migration ptes */
Hugh Dickins03f15c82015-11-05 18:49:56 -0800894 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
895 page);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800896 try_to_unmap(page,
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700897 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800898 page_was_mapped = 1;
899 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700900
Christoph Lametere6a15302006-06-25 05:46:49 -0700901 if (!page_mapped(page))
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800902 rc = move_to_new_page(newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700903
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800904 if (page_was_mapped)
905 remove_migration_ptes(page,
Kirill A. Shutemove3884662016-03-17 14:20:07 -0700906 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700907
Hugh Dickins7db76712015-11-05 18:49:49 -0800908out_unlock_both:
909 unlock_page(newpage);
910out_unlock:
Mel Gorman3f6c8272010-05-24 14:32:17 -0700911 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700912 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700913 put_anon_vma(anon_vma);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700914 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700915out:
916 return rc;
917}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700918
Minchan Kim0dabec92011-10-31 17:06:57 -0700919/*
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -0700920 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
921 * around it.
922 */
923#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
924#define ICE_noinline noinline
925#else
926#define ICE_noinline
927#endif
928
929/*
Minchan Kim0dabec92011-10-31 17:06:57 -0700930 * Obtain the lock on page, remove all ptes and migrate the page
931 * to the newly allocated page in newpage.
932 */
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -0700933static ICE_noinline int unmap_and_move(new_page_t get_new_page,
934 free_page_t put_new_page,
935 unsigned long private, struct page *page,
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -0700936 int force, enum migrate_mode mode,
937 enum migrate_reason reason)
Minchan Kim0dabec92011-10-31 17:06:57 -0700938{
Hugh Dickins2def7422015-11-05 18:49:46 -0800939 int rc = MIGRATEPAGE_SUCCESS;
Minchan Kim0dabec92011-10-31 17:06:57 -0700940 int *result = NULL;
Hugh Dickins2def7422015-11-05 18:49:46 -0800941 struct page *newpage;
Minchan Kim0dabec92011-10-31 17:06:57 -0700942
Hugh Dickins2def7422015-11-05 18:49:46 -0800943 newpage = get_new_page(page, private, &result);
Minchan Kim0dabec92011-10-31 17:06:57 -0700944 if (!newpage)
945 return -ENOMEM;
946
947 if (page_count(page) == 1) {
948 /* page was freed from under us. So we are done. */
949 goto out;
950 }
951
Kirill A. Shutemov4d2fa962016-01-15 16:54:00 -0800952 if (unlikely(PageTransHuge(page))) {
953 lock_page(page);
954 rc = split_huge_page(page);
955 unlock_page(page);
956 if (rc)
Minchan Kim0dabec92011-10-31 17:06:57 -0700957 goto out;
Kirill A. Shutemov4d2fa962016-01-15 16:54:00 -0800958 }
Minchan Kim0dabec92011-10-31 17:06:57 -0700959
Hugh Dickins9c620e22013-02-22 16:35:14 -0800960 rc = __unmap_and_move(page, newpage, force, mode);
Vlastimil Babka7cd12b42016-03-15 14:56:18 -0700961 if (rc == MIGRATEPAGE_SUCCESS) {
Hugh Dickins2def7422015-11-05 18:49:46 -0800962 put_new_page = NULL;
Vlastimil Babka7cd12b42016-03-15 14:56:18 -0700963 set_page_owner_migrate_reason(newpage, reason);
964 }
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800965
Minchan Kim0dabec92011-10-31 17:06:57 -0700966out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700967 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700968 /*
969 * A page that has been migrated has all references
970 * removed and will be freed. A page that has not been
971 * migrated will have kepts its references and be
972 * restored.
973 */
974 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700975 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700976 page_is_file_cache(page));
Naoya Horiguchif4c18e62015-08-06 15:47:08 -0700977 /* Soft-offlined page shouldn't go through lru cache list */
Minchan Kimd7e69482016-04-28 16:18:44 -0700978 if (reason == MR_MEMORY_FAILURE && rc == MIGRATEPAGE_SUCCESS) {
979 /*
980 * With this release, we free successfully migrated
981 * page and set PG_HWPoison on just freed page
982 * intentionally. Although it's rather weird, it's how
983 * HWPoison flag works at the moment.
984 */
Naoya Horiguchif4c18e62015-08-06 15:47:08 -0700985 put_page(page);
Wanpeng Lida1b13c2015-09-08 15:03:27 -0700986 if (!test_set_page_hwpoison(page))
987 num_poisoned_pages_inc();
988 } else
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -0700989 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700990 }
David Rientjes68711a72014-06-04 16:08:25 -0700991
Christoph Lameter95a402c2006-06-23 02:03:53 -0700992 /*
David Rientjes68711a72014-06-04 16:08:25 -0700993 * If migration was not successful and there's a freeing callback, use
994 * it. Otherwise, putback_lru_page() will drop the reference grabbed
995 * during isolation.
Christoph Lameter95a402c2006-06-23 02:03:53 -0700996 */
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800997 if (put_new_page)
David Rientjes68711a72014-06-04 16:08:25 -0700998 put_new_page(newpage, private);
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800999 else if (unlikely(__is_movable_balloon_page(newpage))) {
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -07001000 /* drop our reference, page already in the balloon */
1001 put_page(newpage);
Hugh Dickins8bdd6382014-07-26 12:58:23 -07001002 } else
David Rientjes68711a72014-06-04 16:08:25 -07001003 putback_lru_page(newpage);
1004
Christoph Lameter742755a2006-06-23 02:03:55 -07001005 if (result) {
1006 if (rc)
1007 *result = rc;
1008 else
1009 *result = page_to_nid(newpage);
1010 }
Christoph Lametere24f0b82006-06-23 02:03:51 -07001011 return rc;
1012}
1013
1014/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001015 * Counterpart of unmap_and_move_page() for hugepage migration.
1016 *
1017 * This function doesn't wait the completion of hugepage I/O
1018 * because there is no race between I/O and migration for hugepage.
1019 * Note that currently hugepage I/O occurs only in direct I/O
1020 * where no lock is held and PG_writeback is irrelevant,
1021 * and writeback status of all subpages are counted in the reference
1022 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1023 * under direct I/O, the reference of the head page is 512 and a bit more.)
1024 * This means that when we try to migrate hugepage whose subpages are
1025 * doing direct I/O, some references remain after try_to_unmap() and
1026 * hugepage migration fails without data corruption.
1027 *
1028 * There is also no race when direct I/O is issued on the page under migration,
1029 * because then pte is replaced with migration swap entry and direct I/O code
1030 * will wait in the page fault for migration to complete.
1031 */
1032static int unmap_and_move_huge_page(new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001033 free_page_t put_new_page, unsigned long private,
1034 struct page *hpage, int force,
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001035 enum migrate_mode mode, int reason)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001036{
Hugh Dickins2def7422015-11-05 18:49:46 -08001037 int rc = -EAGAIN;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001038 int *result = NULL;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001039 int page_was_mapped = 0;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001040 struct page *new_hpage;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001041 struct anon_vma *anon_vma = NULL;
1042
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001043 /*
1044 * Movability of hugepages depends on architectures and hugepage size.
1045 * This check is necessary because some callers of hugepage migration
1046 * like soft offline and memory hotremove don't walk through page
1047 * tables or check whether the hugepage is pmd-based or not before
1048 * kicking migration.
1049 */
Naoya Horiguchi100873d2014-06-04 16:10:56 -07001050 if (!hugepage_migration_supported(page_hstate(hpage))) {
Joonsoo Kim32665f22014-01-21 15:51:15 -08001051 putback_active_hugepage(hpage);
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001052 return -ENOSYS;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001053 }
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001054
Joonsoo Kim32665f22014-01-21 15:51:15 -08001055 new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001056 if (!new_hpage)
1057 return -ENOMEM;
1058
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001059 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001060 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001061 goto out;
1062 lock_page(hpage);
1063 }
1064
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001065 if (PageAnon(hpage))
1066 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001067
Hugh Dickins7db76712015-11-05 18:49:49 -08001068 if (unlikely(!trylock_page(new_hpage)))
1069 goto put_anon;
1070
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001071 if (page_mapped(hpage)) {
1072 try_to_unmap(hpage,
1073 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1074 page_was_mapped = 1;
1075 }
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001076
1077 if (!page_mapped(hpage))
Hugh Dickins5c3f9a62015-11-05 18:49:53 -08001078 rc = move_to_new_page(new_hpage, hpage, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001079
Hugh Dickins5c3f9a62015-11-05 18:49:53 -08001080 if (page_was_mapped)
1081 remove_migration_ptes(hpage,
Kirill A. Shutemove3884662016-03-17 14:20:07 -07001082 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001083
Hugh Dickins7db76712015-11-05 18:49:49 -08001084 unlock_page(new_hpage);
1085
1086put_anon:
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001087 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001088 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001089
Hugh Dickins2def7422015-11-05 18:49:46 -08001090 if (rc == MIGRATEPAGE_SUCCESS) {
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001091 hugetlb_cgroup_migrate(hpage, new_hpage);
Hugh Dickins2def7422015-11-05 18:49:46 -08001092 put_new_page = NULL;
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001093 set_page_owner_migrate_reason(new_hpage, reason);
Hugh Dickins2def7422015-11-05 18:49:46 -08001094 }
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001095
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001096 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001097out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001098 if (rc != -EAGAIN)
1099 putback_active_hugepage(hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001100
1101 /*
1102 * If migration was not successful and there's a freeing callback, use
1103 * it. Otherwise, put_page() will drop the reference grabbed during
1104 * isolation.
1105 */
Hugh Dickins2def7422015-11-05 18:49:46 -08001106 if (put_new_page)
David Rientjes68711a72014-06-04 16:08:25 -07001107 put_new_page(new_hpage, private);
1108 else
Naoya Horiguchi3aaa76e2015-09-22 14:59:14 -07001109 putback_active_hugepage(new_hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001110
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001111 if (result) {
1112 if (rc)
1113 *result = rc;
1114 else
1115 *result = page_to_nid(new_hpage);
1116 }
1117 return rc;
1118}
1119
1120/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001121 * migrate_pages - migrate the pages specified in a list, to the free pages
1122 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001123 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001124 * @from: The list of pages to be migrated.
1125 * @get_new_page: The function used to allocate free pages to be used
1126 * as the target of the page migration.
David Rientjes68711a72014-06-04 16:08:25 -07001127 * @put_new_page: The function used to free target pages if migration
1128 * fails, or NULL if no special handling is necessary.
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001129 * @private: Private data to be passed on to get_new_page()
1130 * @mode: The migration mode that specifies the constraints for
1131 * page migration, if any.
1132 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001133 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001134 * The function returns after 10 attempts or if no pages are movable any more
1135 * because the list has become empty or no retryable pages exist any more.
Hugh Dickins14e0f9b2015-11-05 18:49:43 -08001136 * The caller should call putback_movable_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001137 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001138 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001139 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001140 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001141int migrate_pages(struct list_head *from, new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001142 free_page_t put_new_page, unsigned long private,
1143 enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001144{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001145 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001146 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001147 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001148 int pass = 0;
1149 struct page *page;
1150 struct page *page2;
1151 int swapwrite = current->flags & PF_SWAPWRITE;
1152 int rc;
1153
1154 if (!swapwrite)
1155 current->flags |= PF_SWAPWRITE;
1156
Christoph Lametere24f0b82006-06-23 02:03:51 -07001157 for(pass = 0; pass < 10 && retry; pass++) {
1158 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001159
Christoph Lametere24f0b82006-06-23 02:03:51 -07001160 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001161 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001162
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001163 if (PageHuge(page))
1164 rc = unmap_and_move_huge_page(get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001165 put_new_page, private, page,
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001166 pass > 2, mode, reason);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001167 else
David Rientjes68711a72014-06-04 16:08:25 -07001168 rc = unmap_and_move(get_new_page, put_new_page,
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -07001169 private, page, pass > 2, mode,
1170 reason);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001171
Christoph Lametere24f0b82006-06-23 02:03:51 -07001172 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001173 case -ENOMEM:
David Rientjesdfef2ef2016-05-20 16:59:05 -07001174 nr_failed++;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001175 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001176 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001177 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001178 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001179 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001180 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001181 break;
1182 default:
Naoya Horiguchi354a3362014-01-21 15:51:14 -08001183 /*
1184 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1185 * unlike -EAGAIN case, the failed page is
1186 * removed from migration page list and not
1187 * retried in the next outer loop.
1188 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001189 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001190 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001191 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001192 }
1193 }
Vlastimil Babkaf2f81fb2015-11-05 18:47:03 -08001194 nr_failed += retry;
1195 rc = nr_failed;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001196out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001197 if (nr_succeeded)
1198 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1199 if (nr_failed)
1200 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001201 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1202
Christoph Lameterb20a3502006-03-22 00:09:12 -08001203 if (!swapwrite)
1204 current->flags &= ~PF_SWAPWRITE;
1205
Rafael Aquini78bd5202012-12-11 16:02:31 -08001206 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001207}
1208
Christoph Lameter742755a2006-06-23 02:03:55 -07001209#ifdef CONFIG_NUMA
1210/*
1211 * Move a list of individual pages
1212 */
1213struct page_to_node {
1214 unsigned long addr;
1215 struct page *page;
1216 int node;
1217 int status;
1218};
1219
1220static struct page *new_page_node(struct page *p, unsigned long private,
1221 int **result)
1222{
1223 struct page_to_node *pm = (struct page_to_node *)private;
1224
1225 while (pm->node != MAX_NUMNODES && pm->page != p)
1226 pm++;
1227
1228 if (pm->node == MAX_NUMNODES)
1229 return NULL;
1230
1231 *result = &pm->status;
1232
Naoya Horiguchie632a932013-09-11 14:22:04 -07001233 if (PageHuge(p))
1234 return alloc_huge_page_node(page_hstate(compound_head(p)),
1235 pm->node);
1236 else
Vlastimil Babka96db8002015-09-08 15:03:50 -07001237 return __alloc_pages_node(pm->node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001238 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001239}
1240
1241/*
1242 * Move a set of pages as indicated in the pm array. The addr
1243 * field must be set to the virtual address of the page to be moved
1244 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001245 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001246 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001247static int do_move_page_to_node_array(struct mm_struct *mm,
1248 struct page_to_node *pm,
1249 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001250{
1251 int err;
1252 struct page_to_node *pp;
1253 LIST_HEAD(pagelist);
1254
1255 down_read(&mm->mmap_sem);
1256
1257 /*
1258 * Build a list of pages to migrate
1259 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001260 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1261 struct vm_area_struct *vma;
1262 struct page *page;
1263
Christoph Lameter742755a2006-06-23 02:03:55 -07001264 err = -EFAULT;
1265 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001266 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001267 goto set_status;
1268
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001269 /* FOLL_DUMP to ignore special (like zero) pages */
1270 page = follow_page(vma, pp->addr,
1271 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001272
1273 err = PTR_ERR(page);
1274 if (IS_ERR(page))
1275 goto set_status;
1276
Christoph Lameter742755a2006-06-23 02:03:55 -07001277 err = -ENOENT;
1278 if (!page)
1279 goto set_status;
1280
Christoph Lameter742755a2006-06-23 02:03:55 -07001281 pp->page = page;
1282 err = page_to_nid(page);
1283
1284 if (err == pp->node)
1285 /*
1286 * Node already in the right place
1287 */
1288 goto put_and_set;
1289
1290 err = -EACCES;
1291 if (page_mapcount(page) > 1 &&
1292 !migrate_all)
1293 goto put_and_set;
1294
Naoya Horiguchie632a932013-09-11 14:22:04 -07001295 if (PageHuge(page)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08001296 if (PageHead(page))
1297 isolate_huge_page(page, &pagelist);
Naoya Horiguchie632a932013-09-11 14:22:04 -07001298 goto put_and_set;
1299 }
1300
Nick Piggin62695a82008-10-18 20:26:09 -07001301 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001302 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001303 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001304 inc_zone_page_state(page, NR_ISOLATED_ANON +
1305 page_is_file_cache(page));
1306 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001307put_and_set:
1308 /*
1309 * Either remove the duplicate refcount from
1310 * isolate_lru_page() or drop the page ref if it was
1311 * not isolated.
1312 */
1313 put_page(page);
1314set_status:
1315 pp->status = err;
1316 }
1317
Brice Gogline78bbfa2008-10-18 20:27:15 -07001318 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001319 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -07001320 err = migrate_pages(&pagelist, new_page_node, NULL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001321 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001322 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001323 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001324 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001325
1326 up_read(&mm->mmap_sem);
1327 return err;
1328}
1329
1330/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001331 * Migrate an array of page address onto an array of nodes and fill
1332 * the corresponding array of status.
1333 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001334static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001335 unsigned long nr_pages,
1336 const void __user * __user *pages,
1337 const int __user *nodes,
1338 int __user *status, int flags)
1339{
Brice Goglin3140a222009-01-06 14:38:57 -08001340 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001341 unsigned long chunk_nr_pages;
1342 unsigned long chunk_start;
1343 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001344
Brice Goglin3140a222009-01-06 14:38:57 -08001345 err = -ENOMEM;
1346 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1347 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001348 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001349
1350 migrate_prep();
1351
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001352 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001353 * Store a chunk of page_to_node array in a page,
1354 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001355 */
Brice Goglin3140a222009-01-06 14:38:57 -08001356 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001357
Brice Goglin3140a222009-01-06 14:38:57 -08001358 for (chunk_start = 0;
1359 chunk_start < nr_pages;
1360 chunk_start += chunk_nr_pages) {
1361 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001362
Brice Goglin3140a222009-01-06 14:38:57 -08001363 if (chunk_start + chunk_nr_pages > nr_pages)
1364 chunk_nr_pages = nr_pages - chunk_start;
1365
1366 /* fill the chunk pm with addrs and nodes from user-space */
1367 for (j = 0; j < chunk_nr_pages; j++) {
1368 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001369 int node;
1370
Brice Goglin3140a222009-01-06 14:38:57 -08001371 err = -EFAULT;
1372 if (get_user(p, pages + j + chunk_start))
1373 goto out_pm;
1374 pm[j].addr = (unsigned long) p;
1375
1376 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001377 goto out_pm;
1378
1379 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001380 if (node < 0 || node >= MAX_NUMNODES)
1381 goto out_pm;
1382
Lai Jiangshan389162c2012-12-12 13:51:30 -08001383 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001384 goto out_pm;
1385
1386 err = -EACCES;
1387 if (!node_isset(node, task_nodes))
1388 goto out_pm;
1389
Brice Goglin3140a222009-01-06 14:38:57 -08001390 pm[j].node = node;
1391 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001392
Brice Goglin3140a222009-01-06 14:38:57 -08001393 /* End marker for this chunk */
1394 pm[chunk_nr_pages].node = MAX_NUMNODES;
1395
1396 /* Migrate this chunk */
1397 err = do_move_page_to_node_array(mm, pm,
1398 flags & MPOL_MF_MOVE_ALL);
1399 if (err < 0)
1400 goto out_pm;
1401
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001402 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001403 for (j = 0; j < chunk_nr_pages; j++)
1404 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001405 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001406 goto out_pm;
1407 }
1408 }
1409 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001410
1411out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001412 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001413out:
1414 return err;
1415}
1416
1417/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001418 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001419 */
Brice Goglin80bba122008-12-09 13:14:23 -08001420static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1421 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001422{
Brice Goglin2f007e72008-10-18 20:27:16 -07001423 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001424
Christoph Lameter742755a2006-06-23 02:03:55 -07001425 down_read(&mm->mmap_sem);
1426
Brice Goglin2f007e72008-10-18 20:27:16 -07001427 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001428 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001429 struct vm_area_struct *vma;
1430 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001431 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001432
1433 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001434 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001435 goto set_status;
1436
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001437 /* FOLL_DUMP to ignore special (like zero) pages */
1438 page = follow_page(vma, addr, FOLL_DUMP);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001439
1440 err = PTR_ERR(page);
1441 if (IS_ERR(page))
1442 goto set_status;
1443
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001444 err = page ? page_to_nid(page) : -ENOENT;
Christoph Lameter742755a2006-06-23 02:03:55 -07001445set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001446 *status = err;
1447
1448 pages++;
1449 status++;
1450 }
1451
1452 up_read(&mm->mmap_sem);
1453}
1454
1455/*
1456 * Determine the nodes of a user array of pages and store it in
1457 * a user array of status.
1458 */
1459static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1460 const void __user * __user *pages,
1461 int __user *status)
1462{
1463#define DO_PAGES_STAT_CHUNK_NR 16
1464 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1465 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001466
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001467 while (nr_pages) {
1468 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001469
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001470 chunk_nr = nr_pages;
1471 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1472 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1473
1474 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1475 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001476
1477 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1478
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001479 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1480 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001481
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001482 pages += chunk_nr;
1483 status += chunk_nr;
1484 nr_pages -= chunk_nr;
1485 }
1486 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001487}
1488
1489/*
1490 * Move a list of pages in the address space of the currently executing
1491 * process.
1492 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001493SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1494 const void __user * __user *, pages,
1495 const int __user *, nodes,
1496 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001497{
David Howellsc69e8d92008-11-14 10:39:19 +11001498 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001499 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001500 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001501 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001502 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001503
1504 /* Check flags */
1505 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1506 return -EINVAL;
1507
1508 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1509 return -EPERM;
1510
1511 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001512 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001513 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001514 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001515 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001516 return -ESRCH;
1517 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001518 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001519
1520 /*
1521 * Check if this process has the right to modify the specified
1522 * process. The right exists if the process has administrative
1523 * capabilities, superuser privileges or the same
1524 * userid as the target process.
1525 */
David Howellsc69e8d92008-11-14 10:39:19 +11001526 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001527 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1528 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001529 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001530 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001531 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001532 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001533 }
David Howellsc69e8d92008-11-14 10:39:19 +11001534 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001535
David Quigley86c3a762006-06-23 02:04:02 -07001536 err = security_task_movememory(task);
1537 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001538 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001539
Christoph Lameter3268c632012-03-21 16:34:06 -07001540 task_nodes = cpuset_mems_allowed(task);
1541 mm = get_task_mm(task);
1542 put_task_struct(task);
1543
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001544 if (!mm)
1545 return -EINVAL;
1546
1547 if (nodes)
1548 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1549 nodes, status, flags);
1550 else
1551 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001552
1553 mmput(mm);
1554 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001555
Christoph Lameter742755a2006-06-23 02:03:55 -07001556out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001557 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001558 return err;
1559}
Christoph Lameter742755a2006-06-23 02:03:55 -07001560
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001561#ifdef CONFIG_NUMA_BALANCING
1562/*
1563 * Returns true if this is a safe migration target node for misplaced NUMA
1564 * pages. Currently it only checks the watermarks which crude
1565 */
1566static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001567 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001568{
1569 int z;
1570 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1571 struct zone *zone = pgdat->node_zones + z;
1572
1573 if (!populated_zone(zone))
1574 continue;
1575
Lisa Du6e543d52013-09-11 14:22:36 -07001576 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001577 continue;
1578
1579 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1580 if (!zone_watermark_ok(zone, 0,
1581 high_wmark_pages(zone) +
1582 nr_migrate_pages,
1583 0, 0))
1584 continue;
1585 return true;
1586 }
1587 return false;
1588}
1589
1590static struct page *alloc_misplaced_dst_page(struct page *page,
1591 unsigned long data,
1592 int **result)
1593{
1594 int nid = (int) data;
1595 struct page *newpage;
1596
Vlastimil Babka96db8002015-09-08 15:03:50 -07001597 newpage = __alloc_pages_node(nid,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001598 (GFP_HIGHUSER_MOVABLE |
1599 __GFP_THISNODE | __GFP_NOMEMALLOC |
1600 __GFP_NORETRY | __GFP_NOWARN) &
Mel Gorman8479eba2016-02-26 15:19:31 -08001601 ~__GFP_RECLAIM, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001602
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001603 return newpage;
1604}
1605
1606/*
Mel Gormana8f60772012-11-14 21:41:46 +00001607 * page migration rate limiting control.
1608 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1609 * window of time. Default here says do not migrate more than 1280M per second.
1610 */
1611static unsigned int migrate_interval_millisecs __read_mostly = 100;
1612static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1613
Mel Gormanb32967f2012-11-19 12:35:47 +00001614/* Returns true if the node is migrate rate-limited after the update */
Mel Gorman1c30e012014-01-21 15:50:58 -08001615static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1616 unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001617{
Mel Gormanb32967f2012-11-19 12:35:47 +00001618 /*
1619 * Rate-limit the amount of data that is being migrated to a node.
1620 * Optimal placement is no good if the memory bus is saturated and
1621 * all the time is being spent migrating!
1622 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001623 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001624 spin_lock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001625 pgdat->numabalancing_migrate_nr_pages = 0;
1626 pgdat->numabalancing_migrate_next_window = jiffies +
1627 msecs_to_jiffies(migrate_interval_millisecs);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001628 spin_unlock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001629 }
Mel Gormanaf1839d2014-01-21 15:51:01 -08001630 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1631 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1632 nr_pages);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001633 return true;
Mel Gormanaf1839d2014-01-21 15:51:01 -08001634 }
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001635
1636 /*
1637 * This is an unlocked non-atomic update so errors are possible.
1638 * The consequences are failing to migrate when we potentiall should
1639 * have which is not severe enough to warrant locking. If it is ever
1640 * a problem, it can be converted to a per-cpu counter.
1641 */
1642 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1643 return false;
Mel Gormanb32967f2012-11-19 12:35:47 +00001644}
1645
Mel Gorman1c30e012014-01-21 15:50:58 -08001646static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
Mel Gormanb32967f2012-11-19 12:35:47 +00001647{
Hugh Dickins340ef392013-02-22 16:34:33 -08001648 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001649
Sasha Levin309381fea2014-01-23 15:52:54 -08001650 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
Mel Gorman3abef4e2013-02-22 16:34:27 -08001651
Mel Gormanb32967f2012-11-19 12:35:47 +00001652 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001653 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1654 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001655
Hugh Dickins340ef392013-02-22 16:34:33 -08001656 if (isolate_lru_page(page))
1657 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001658
1659 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001660 * migrate_misplaced_transhuge_page() skips page migration's usual
1661 * check on page_count(), so we must do it here, now that the page
1662 * has been isolated: a GUP pin, or any other pin, prevents migration.
1663 * The expected page count is 3: 1 for page's mapcount and 1 for the
1664 * caller's pin and 1 for the reference taken by isolate_lru_page().
1665 */
1666 if (PageTransHuge(page) && page_count(page) != 3) {
1667 putback_lru_page(page);
1668 return 0;
1669 }
1670
1671 page_lru = page_is_file_cache(page);
1672 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1673 hpage_nr_pages(page));
1674
1675 /*
1676 * Isolating the page has taken another reference, so the
1677 * caller's reference can be safely dropped without the page
1678 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001679 */
1680 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001681 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001682}
1683
Mel Gormande466bd2013-12-18 17:08:42 -08001684bool pmd_trans_migrating(pmd_t pmd)
1685{
1686 struct page *page = pmd_page(pmd);
1687 return PageLocked(page);
1688}
1689
Mel Gormana8f60772012-11-14 21:41:46 +00001690/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001691 * Attempt to migrate a misplaced page to the specified destination
1692 * node. Caller is expected to have an elevated reference count on
1693 * the page that will be dropped by this function before returning.
1694 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001695int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1696 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001697{
Mel Gormana8f60772012-11-14 21:41:46 +00001698 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001699 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001700 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001701 LIST_HEAD(migratepages);
1702
1703 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001704 * Don't migrate file pages that are mapped in multiple processes
1705 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001706 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001707 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1708 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001709 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001710
Mel Gormana8f60772012-11-14 21:41:46 +00001711 /*
1712 * Rate-limit the amount of data that is being migrated to a node.
1713 * Optimal placement is no good if the memory bus is saturated and
1714 * all the time is being spent migrating!
1715 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001716 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001717 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001718
Mel Gormanb32967f2012-11-19 12:35:47 +00001719 isolated = numamigrate_isolate_page(pgdat, page);
1720 if (!isolated)
1721 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001722
Mel Gormanb32967f2012-11-19 12:35:47 +00001723 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001724 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
David Rientjes68711a72014-06-04 16:08:25 -07001725 NULL, node, MIGRATE_ASYNC,
1726 MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001727 if (nr_remaining) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001728 if (!list_empty(&migratepages)) {
1729 list_del(&page->lru);
1730 dec_zone_page_state(page, NR_ISOLATED_ANON +
1731 page_is_file_cache(page));
1732 putback_lru_page(page);
1733 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001734 isolated = 0;
1735 } else
1736 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001737 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001738 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001739
1740out:
1741 put_page(page);
1742 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001743}
Mel Gorman220018d2012-12-05 09:32:56 +00001744#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001745
Mel Gorman220018d2012-12-05 09:32:56 +00001746#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001747/*
1748 * Migrates a THP to a given target node. page must be locked and is unlocked
1749 * before returning.
1750 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001751int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1752 struct vm_area_struct *vma,
1753 pmd_t *pmd, pmd_t entry,
1754 unsigned long address,
1755 struct page *page, int node)
1756{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001757 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001758 pg_data_t *pgdat = NODE_DATA(node);
1759 int isolated = 0;
1760 struct page *new_page = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001761 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001762 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1763 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001764 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001765
1766 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001767 * Rate-limit the amount of data that is being migrated to a node.
1768 * Optimal placement is no good if the memory bus is saturated and
1769 * all the time is being spent migrating!
1770 */
Mel Gormand28d43352012-11-29 09:24:36 +00001771 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001772 goto out_dropref;
1773
1774 new_page = alloc_pages_node(node,
Mel Gorman71baba42015-11-06 16:28:28 -08001775 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001776 HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001777 if (!new_page)
1778 goto out_fail;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001779 prep_transhuge_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001780
Mel Gormanb32967f2012-11-19 12:35:47 +00001781 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001782 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001783 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001784 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001785 }
Aneesh Kumar K.V458aa762016-03-17 14:18:56 -07001786 /*
1787 * We are not sure a pending tlb flush here is for a huge page
1788 * mapping or not. Hence use the tlb range variant
1789 */
Mel Gormanb0943d62013-12-18 17:08:46 -08001790 if (mm_tlb_flush_pending(mm))
1791 flush_tlb_range(vma, mmun_start, mmun_end);
1792
Mel Gormanb32967f2012-11-19 12:35:47 +00001793 /* Prepare a page as a migration target */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001794 __SetPageLocked(new_page);
Hugh Dickinsfa9949d2016-05-19 17:12:41 -07001795 __SetPageSwapBacked(new_page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001796
1797 /* anon mapping, we can simply copy page->mapping to the new page: */
1798 new_page->mapping = page->mapping;
1799 new_page->index = page->index;
1800 migrate_page_copy(new_page, page);
1801 WARN_ON(PageLRU(new_page));
1802
1803 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001804 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001805 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001806 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1807fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001808 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001809 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001810
1811 /* Reverse changes made by migrate_page_copy() */
1812 if (TestClearPageActive(new_page))
1813 SetPageActive(page);
1814 if (TestClearPageUnevictable(new_page))
1815 SetPageUnevictable(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001816
1817 unlock_page(new_page);
1818 put_page(new_page); /* Free it */
1819
Mel Gormana54a4072013-10-07 11:28:46 +01001820 /* Retake the callers reference and putback on LRU */
1821 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001822 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001823 mod_zone_page_state(page_zone(page),
1824 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001825
1826 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001827 }
1828
Mel Gorman2b4847e2013-12-18 17:08:32 -08001829 orig_entry = *pmd;
Mel Gormanb32967f2012-11-19 12:35:47 +00001830 entry = mk_pmd(new_page, vma->vm_page_prot);
Mel Gormanb32967f2012-11-19 12:35:47 +00001831 entry = pmd_mkhuge(entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001832 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001833
Mel Gorman2b4847e2013-12-18 17:08:32 -08001834 /*
1835 * Clear the old entry under pagetable lock and establish the new PTE.
1836 * Any parallel GUP will either observe the old page blocking on the
1837 * page lock, block on the page table lock or observe the new page.
1838 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1839 * guarantee the copy is visible before the pagetable update.
1840 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001841 flush_cache_range(vma, mmun_start, mmun_end);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001842 page_add_anon_rmap(new_page, vma, mmun_start, true);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001843 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001844 set_pmd_at(mm, mmun_start, pmd, entry);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001845 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001846
1847 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08001848 set_pmd_at(mm, mmun_start, pmd, orig_entry);
Aneesh Kumar K.V458aa762016-03-17 14:18:56 -07001849 flush_pmd_tlb_range(vma, mmun_start, mmun_end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001850 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001851 update_mmu_cache_pmd(vma, address, &entry);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001852 page_remove_rmap(new_page, true);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001853 goto fail_putback;
1854 }
1855
Hugh Dickins51afb122015-11-05 18:49:37 -08001856 mlock_migrate_page(new_page, page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001857 page_remove_rmap(page, true);
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001858 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001859
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001860 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001861 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001862
Mel Gorman11de9922014-06-04 16:07:41 -07001863 /* Take an "isolate" reference and put new page on the LRU. */
1864 get_page(new_page);
1865 putback_lru_page(new_page);
1866
Mel Gormanb32967f2012-11-19 12:35:47 +00001867 unlock_page(new_page);
1868 unlock_page(page);
1869 put_page(page); /* Drop the rmap reference */
1870 put_page(page); /* Drop the LRU isolation reference */
1871
1872 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1873 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1874
Mel Gormanb32967f2012-11-19 12:35:47 +00001875 mod_zone_page_state(page_zone(page),
1876 NR_ISOLATED_ANON + page_lru,
1877 -HPAGE_PMD_NR);
1878 return isolated;
1879
Hugh Dickins340ef392013-02-22 16:34:33 -08001880out_fail:
1881 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001882out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08001883 ptl = pmd_lock(mm, pmd);
1884 if (pmd_same(*pmd, entry)) {
Mel Gorman4d942462015-02-12 14:58:28 -08001885 entry = pmd_modify(entry, vma->vm_page_prot);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001886 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001887 update_mmu_cache_pmd(vma, address, &entry);
1888 }
1889 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001890
Mel Gormaneb4489f62013-12-18 17:08:39 -08001891out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08001892 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001893 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001894 return 0;
1895}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001896#endif /* CONFIG_NUMA_BALANCING */
1897
1898#endif /* CONFIG_NUMA */