blob: 89a0a1707f4c67deb77dc466cb6f5c2ecfe19051 [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>
Minchan Kimbda807d2016-07-26 15:23:05 -070034#include <linux/compaction.h>
Adrian Bunk4f5ca262008-07-23 21:27:02 -070035#include <linux/syscalls.h>
Naoya Horiguchi290408d2010-09-08 10:19:35 +090036#include <linux/hugetlb.h>
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -070037#include <linux/hugetlb_cgroup.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090038#include <linux/gfp.h>
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080039#include <linux/balloon_compaction.h>
Mel Gormanf714f4f2013-12-18 17:08:33 -080040#include <linux/mmu_notifier.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070041#include <linux/page_idle.h>
Vlastimil Babkad435edc2016-03-15 14:56:15 -070042#include <linux/page_owner.h>
Ingo Molnar6e84f312017-02-08 18:51:29 +010043#include <linux/sched/mm.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080044
Michal Nazarewicz0d1836c2010-12-21 17:24:26 -080045#include <asm/tlbflush.h>
46
Mel Gorman7b2a2d42012-10-19 14:07:31 +010047#define CREATE_TRACE_POINTS
48#include <trace/events/migrate.h>
49
Christoph Lameterb20a3502006-03-22 00:09:12 -080050#include "internal.h"
51
Christoph Lameterb20a3502006-03-22 00:09:12 -080052/*
Christoph Lameter742755a2006-06-23 02:03:55 -070053 * migrate_prep() needs to be called before we start compiling a list of pages
Mel Gorman748446b2010-05-24 14:32:27 -070054 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
55 * undesirable, use migrate_prep_local()
Christoph Lameterb20a3502006-03-22 00:09:12 -080056 */
57int migrate_prep(void)
58{
Christoph Lameterb20a3502006-03-22 00:09:12 -080059 /*
60 * Clear the LRU lists so pages can be isolated.
61 * Note that pages may be moved off the LRU after we have
62 * drained them. Those pages will fail to migrate like other
63 * pages that may be busy.
64 */
65 lru_add_drain_all();
66
67 return 0;
68}
69
Mel Gorman748446b2010-05-24 14:32:27 -070070/* Do the necessary work of migrate_prep but not if it involves other CPUs */
71int migrate_prep_local(void)
72{
73 lru_add_drain();
74
75 return 0;
76}
77
Yisheng Xie9e5bcd62017-02-24 14:57:29 -080078int isolate_movable_page(struct page *page, isolate_mode_t mode)
Minchan Kimbda807d2016-07-26 15:23:05 -070079{
80 struct address_space *mapping;
81
82 /*
83 * Avoid burning cycles with pages that are yet under __free_pages(),
84 * or just got freed under us.
85 *
86 * In case we 'win' a race for a movable page being freed under us and
87 * raise its refcount preventing __free_pages() from doing its job
88 * the put_page() at the end of this block will take care of
89 * release this page, thus avoiding a nasty leakage.
90 */
91 if (unlikely(!get_page_unless_zero(page)))
92 goto out;
93
94 /*
95 * Check PageMovable before holding a PG_lock because page's owner
96 * assumes anybody doesn't touch PG_lock of newly allocated page
97 * so unconditionally grapping the lock ruins page's owner side.
98 */
99 if (unlikely(!__PageMovable(page)))
100 goto out_putpage;
101 /*
102 * As movable pages are not isolated from LRU lists, concurrent
103 * compaction threads can race against page migration functions
104 * as well as race against the releasing a page.
105 *
106 * In order to avoid having an already isolated movable page
107 * being (wrongly) re-isolated while it is under migration,
108 * or to avoid attempting to isolate pages being released,
109 * lets be sure we have the page lock
110 * before proceeding with the movable page isolation steps.
111 */
112 if (unlikely(!trylock_page(page)))
113 goto out_putpage;
114
115 if (!PageMovable(page) || PageIsolated(page))
116 goto out_no_isolated;
117
118 mapping = page_mapping(page);
119 VM_BUG_ON_PAGE(!mapping, page);
120
121 if (!mapping->a_ops->isolate_page(page, mode))
122 goto out_no_isolated;
123
124 /* Driver shouldn't use PG_isolated bit of page->flags */
125 WARN_ON_ONCE(PageIsolated(page));
126 __SetPageIsolated(page);
127 unlock_page(page);
128
Yisheng Xie9e5bcd62017-02-24 14:57:29 -0800129 return 0;
Minchan Kimbda807d2016-07-26 15:23:05 -0700130
131out_no_isolated:
132 unlock_page(page);
133out_putpage:
134 put_page(page);
135out:
Yisheng Xie9e5bcd62017-02-24 14:57:29 -0800136 return -EBUSY;
Minchan Kimbda807d2016-07-26 15:23:05 -0700137}
138
139/* It should be called on page which is PG_movable */
140void putback_movable_page(struct page *page)
141{
142 struct address_space *mapping;
143
144 VM_BUG_ON_PAGE(!PageLocked(page), page);
145 VM_BUG_ON_PAGE(!PageMovable(page), page);
146 VM_BUG_ON_PAGE(!PageIsolated(page), page);
147
148 mapping = page_mapping(page);
149 mapping->a_ops->putback_page(page);
150 __ClearPageIsolated(page);
151}
152
Christoph Lameterb20a3502006-03-22 00:09:12 -0800153/*
Rafael Aquini5733c7d2012-12-11 16:02:47 -0800154 * Put previously isolated pages back onto the appropriate lists
155 * from where they were once taken off for compaction/migration.
156 *
Joonsoo Kim59c82b72014-01-21 15:51:17 -0800157 * This function shall be used whenever the isolated pageset has been
158 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
159 * and isolate_huge_page().
Rafael Aquini5733c7d2012-12-11 16:02:47 -0800160 */
161void putback_movable_pages(struct list_head *l)
162{
163 struct page *page;
164 struct page *page2;
165
166 list_for_each_entry_safe(page, page2, l, lru) {
Naoya Horiguchi31caf662013-09-11 14:21:59 -0700167 if (unlikely(PageHuge(page))) {
168 putback_active_hugepage(page);
169 continue;
170 }
Rafael Aquini5733c7d2012-12-11 16:02:47 -0800171 list_del(&page->lru);
Minchan Kimbda807d2016-07-26 15:23:05 -0700172 /*
173 * We isolated non-lru movable page so here we can use
174 * __PageMovable because LRU page's mapping cannot have
175 * PAGE_MAPPING_MOVABLE.
176 */
Minchan Kimb1123ea62016-07-26 15:23:09 -0700177 if (unlikely(__PageMovable(page))) {
Minchan Kimbda807d2016-07-26 15:23:05 -0700178 VM_BUG_ON_PAGE(!PageIsolated(page), page);
179 lock_page(page);
180 if (PageMovable(page))
181 putback_movable_page(page);
182 else
183 __ClearPageIsolated(page);
184 unlock_page(page);
185 put_page(page);
186 } else {
Ming Ling6afcf8e2016-12-12 16:42:26 -0800187 dec_node_page_state(page, NR_ISOLATED_ANON +
188 page_is_file_cache(page));
Rabin Vincentfc280fe2017-04-20 14:37:46 -0700189 putback_lru_page(page);
Minchan Kimbda807d2016-07-26 15:23:05 -0700190 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800191 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800192}
193
Christoph Lameter06972122006-06-23 02:03:35 -0700194/*
195 * Restore a potential migration pte to a working pte entry
196 */
Minchan Kime4b82222017-05-03 14:54:27 -0700197static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800198 unsigned long addr, void *old)
Christoph Lameter06972122006-06-23 02:03:35 -0700199{
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800200 struct page_vma_mapped_walk pvmw = {
201 .page = old,
202 .vma = vma,
203 .address = addr,
204 .flags = PVMW_SYNC | PVMW_MIGRATION,
205 };
206 struct page *new;
207 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700208 swp_entry_t entry;
Christoph Lameter06972122006-06-23 02:03:35 -0700209
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800210 VM_BUG_ON_PAGE(PageTail(page), page);
211 while (page_vma_mapped_walk(&pvmw)) {
Naoya Horiguchi4b0ece62017-03-31 15:11:44 -0700212 if (PageKsm(page))
213 new = page;
214 else
215 new = page - pvmw.page->index +
216 linear_page_index(vma, pvmw.address);
Christoph Lameter06972122006-06-23 02:03:35 -0700217
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800218 get_page(new);
219 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
220 if (pte_swp_soft_dirty(*pvmw.pte))
221 pte = pte_mksoft_dirty(pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700222
Hugh Dickins486cf462011-10-19 12:50:35 -0700223 /*
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800224 * Recheck VMA as permissions can change since migration started
Hugh Dickins486cf462011-10-19 12:50:35 -0700225 */
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800226 entry = pte_to_swp_entry(*pvmw.pte);
227 if (is_write_migration_entry(entry))
228 pte = maybe_mkwrite(pte, vma);
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100229
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200230#ifdef CONFIG_HUGETLB_PAGE
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800231 if (PageHuge(new)) {
232 pte = pte_mkhuge(pte);
233 pte = arch_make_huge_pte(pte, vma, new, 0);
234 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200235#endif
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800236 flush_dcache_page(new);
237 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700238
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800239 if (PageHuge(new)) {
240 if (PageAnon(new))
241 hugepage_add_anon_rmap(new, vma, pvmw.address);
242 else
243 page_dup_rmap(new, true);
244 } else if (PageAnon(new))
245 page_add_anon_rmap(new, vma, pvmw.address, false);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900246 else
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800247 page_add_file_rmap(new, false);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700248
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800249 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
250 mlock_vma_page(new);
Hugh Dickins51afb122015-11-05 18:49:37 -0800251
Kirill A. Shutemov3fe87962017-02-24 14:58:16 -0800252 /* No need to invalidate - it was non-present before */
253 update_mmu_cache(vma, pvmw.address, pvmw.pte);
254 }
255
Minchan Kime4b82222017-05-03 14:54:27 -0700256 return true;
Christoph Lameter06972122006-06-23 02:03:35 -0700257}
258
259/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700260 * Get rid of all migration entries and replace them by
261 * references to the indicated page.
262 */
Kirill A. Shutemove3884662016-03-17 14:20:07 -0700263void remove_migration_ptes(struct page *old, struct page *new, bool locked)
Christoph Lameter04e62a22006-06-23 02:03:38 -0700264{
Joonsoo Kim051ac832014-01-21 15:49:48 -0800265 struct rmap_walk_control rwc = {
266 .rmap_one = remove_migration_pte,
267 .arg = old,
268 };
269
Kirill A. Shutemove3884662016-03-17 14:20:07 -0700270 if (locked)
271 rmap_walk_locked(new, &rwc);
272 else
273 rmap_walk(new, &rwc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700274}
275
276/*
Christoph Lameter06972122006-06-23 02:03:35 -0700277 * Something used the pte of a page under migration. We need to
278 * get to the page and wait until migration is finished.
279 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700280 */
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800281void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700282 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700283{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700284 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700285 swp_entry_t entry;
286 struct page *page;
287
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700288 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700289 pte = *ptep;
290 if (!is_swap_pte(pte))
291 goto out;
292
293 entry = pte_to_swp_entry(pte);
294 if (!is_migration_entry(entry))
295 goto out;
296
297 page = migration_entry_to_page(entry);
298
Nick Piggine2867812008-07-25 19:45:30 -0700299 /*
300 * Once radix-tree replacement of page migration started, page_count
301 * *must* be zero. And, we don't want to call wait_on_page_locked()
302 * against a page without get_page().
303 * So, we use get_page_unless_zero(), here. Even failed, page fault
304 * will occur again.
305 */
306 if (!get_page_unless_zero(page))
307 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700308 pte_unmap_unlock(ptep, ptl);
309 wait_on_page_locked(page);
310 put_page(page);
311 return;
312out:
313 pte_unmap_unlock(ptep, ptl);
314}
315
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700316void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
317 unsigned long address)
318{
319 spinlock_t *ptl = pte_lockptr(mm, pmd);
320 pte_t *ptep = pte_offset_map(pmd, address);
321 __migration_entry_wait(mm, ptep, ptl);
322}
323
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800324void migration_entry_wait_huge(struct vm_area_struct *vma,
325 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700326{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800327 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700328 __migration_entry_wait(mm, pte, ptl);
329}
330
Mel Gormanb969c4a2012-01-12 17:19:34 -0800331#ifdef CONFIG_BLOCK
332/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800333static bool buffer_migrate_lock_buffers(struct buffer_head *head,
334 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800335{
336 struct buffer_head *bh = head;
337
338 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800339 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800340 do {
341 get_bh(bh);
342 lock_buffer(bh);
343 bh = bh->b_this_page;
344
345 } while (bh != head);
346
347 return true;
348 }
349
350 /* async case, we cannot block on lock_buffer so use trylock_buffer */
351 do {
352 get_bh(bh);
353 if (!trylock_buffer(bh)) {
354 /*
355 * We failed to lock the buffer and cannot stall in
356 * async migration. Release the taken locks
357 */
358 struct buffer_head *failed_bh = bh;
359 put_bh(failed_bh);
360 bh = head;
361 while (bh != failed_bh) {
362 unlock_buffer(bh);
363 put_bh(bh);
364 bh = bh->b_this_page;
365 }
366 return false;
367 }
368
369 bh = bh->b_this_page;
370 } while (bh != head);
371 return true;
372}
373#else
374static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800375 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800376{
377 return true;
378}
379#endif /* CONFIG_BLOCK */
380
Christoph Lameterb20a3502006-03-22 00:09:12 -0800381/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700382 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700383 *
384 * The number of remaining references must be:
385 * 1 for anonymous pages without a mapping
386 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100387 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800388 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800389int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800390 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500391 struct buffer_head *head, enum migrate_mode mode,
392 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800393{
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800394 struct zone *oldzone, *newzone;
395 int dirty;
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500396 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800397 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800398
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700399 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700400 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500401 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700402 return -EAGAIN;
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800403
404 /* No turning back from here */
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800405 newpage->index = page->index;
406 newpage->mapping = page->mapping;
407 if (PageSwapBacked(page))
Hugh Dickinsfa9949d2016-05-19 17:12:41 -0700408 __SetPageSwapBacked(newpage);
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800409
Rafael Aquini78bd5202012-12-11 16:02:31 -0800410 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700411 }
412
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800413 oldzone = page_zone(page);
414 newzone = page_zone(newpage);
415
Nick Piggin19fd6232008-07-25 19:45:32 -0700416 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800417
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800418 pslot = radix_tree_lookup_slot(&mapping->page_tree,
419 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800420
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500421 expected_count += 1 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700422 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800423 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700424 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700425 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800426 }
427
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700428 if (!page_ref_freeze(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700429 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700430 return -EAGAIN;
431 }
432
Christoph Lameterb20a3502006-03-22 00:09:12 -0800433 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800434 * In the async migration case of moving a page with buffers, lock the
435 * buffers using trylock before the mapping is moved. If the mapping
436 * was moved, we later failed to lock the buffers and could not move
437 * the mapping back due to an elevated page count, we would have to
438 * block waiting on other references to be dropped.
439 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800440 if (mode == MIGRATE_ASYNC && head &&
441 !buffer_migrate_lock_buffers(head, mode)) {
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700442 page_ref_unfreeze(page, expected_count);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800443 spin_unlock_irq(&mapping->tree_lock);
444 return -EAGAIN;
445 }
446
447 /*
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800448 * Now we know that no one else is looking at the page:
449 * no turning back from here.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800450 */
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800451 newpage->index = page->index;
452 newpage->mapping = page->mapping;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800453 get_page(newpage); /* add cache reference */
Nicholas Piggin6326fec2016-12-25 13:00:29 +1000454 if (PageSwapBacked(page)) {
455 __SetPageSwapBacked(newpage);
456 if (PageSwapCache(page)) {
457 SetPageSwapCache(newpage);
458 set_page_private(newpage, page_private(page));
459 }
460 } else {
461 VM_BUG_ON_PAGE(PageSwapCache(page), page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800462 }
463
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800464 /* Move dirty while page refs frozen and newpage not yet exposed */
465 dirty = PageDirty(page);
466 if (dirty) {
467 ClearPageDirty(page);
468 SetPageDirty(newpage);
469 }
470
Johannes Weiner6d75f362016-12-12 16:43:43 -0800471 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800472
473 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800474 * Drop cache reference from old page by unfreezing
475 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800476 * We know this isn't the last reference.
477 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700478 page_ref_unfreeze(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800479
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800480 spin_unlock(&mapping->tree_lock);
481 /* Leave irq disabled to prevent preemption while updating stats */
482
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700483 /*
484 * If moved to a different zone then also account
485 * the page for that zone. Other VM counters will be
486 * taken care of when we establish references to the
487 * new page and drop references to the old page.
488 *
489 * Note that anonymous pages are accounted for
Mel Gorman4b9d0fa2016-07-28 15:46:17 -0700490 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700491 * are mapped to swap space.
492 */
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800493 if (newzone != oldzone) {
Mel Gorman11fb9982016-07-28 15:46:20 -0700494 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
495 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800496 if (PageSwapBacked(page) && !PageSwapCache(page)) {
Mel Gorman11fb9982016-07-28 15:46:20 -0700497 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
498 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800499 }
500 if (dirty && mapping_cap_account_dirty(mapping)) {
Mel Gorman11fb9982016-07-28 15:46:20 -0700501 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700502 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
Mel Gorman11fb9982016-07-28 15:46:20 -0700503 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700504 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800505 }
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700506 }
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800507 local_irq_enable();
Christoph Lameterb20a3502006-03-22 00:09:12 -0800508
Rafael Aquini78bd5202012-12-11 16:02:31 -0800509 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800510}
Richard Weinberger1118dce2016-06-16 23:26:14 +0200511EXPORT_SYMBOL(migrate_page_move_mapping);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800512
513/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900514 * The expected number of remaining references is the same as that
515 * of migrate_page_move_mapping().
516 */
517int migrate_huge_page_move_mapping(struct address_space *mapping,
518 struct page *newpage, struct page *page)
519{
520 int expected_count;
521 void **pslot;
522
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900523 spin_lock_irq(&mapping->tree_lock);
524
525 pslot = radix_tree_lookup_slot(&mapping->page_tree,
526 page_index(page));
527
528 expected_count = 2 + page_has_private(page);
529 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800530 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900531 spin_unlock_irq(&mapping->tree_lock);
532 return -EAGAIN;
533 }
534
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700535 if (!page_ref_freeze(page, expected_count)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900536 spin_unlock_irq(&mapping->tree_lock);
537 return -EAGAIN;
538 }
539
Hugh Dickinscf4b7692015-11-05 18:50:02 -0800540 newpage->index = page->index;
541 newpage->mapping = page->mapping;
Johannes Weiner6a93ca82016-03-15 14:57:19 -0700542
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900543 get_page(newpage);
544
Johannes Weiner6d75f362016-12-12 16:43:43 -0800545 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900546
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700547 page_ref_unfreeze(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900548
549 spin_unlock_irq(&mapping->tree_lock);
Johannes Weiner6a93ca82016-03-15 14:57:19 -0700550
Rafael Aquini78bd5202012-12-11 16:02:31 -0800551 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900552}
553
554/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800555 * Gigantic pages are so large that we do not guarantee that page++ pointer
556 * arithmetic will work across the entire page. We need something more
557 * specialized.
558 */
559static void __copy_gigantic_page(struct page *dst, struct page *src,
560 int nr_pages)
561{
562 int i;
563 struct page *dst_base = dst;
564 struct page *src_base = src;
565
566 for (i = 0; i < nr_pages; ) {
567 cond_resched();
568 copy_highpage(dst, src);
569
570 i++;
571 dst = mem_map_next(dst, dst_base, i);
572 src = mem_map_next(src, src_base, i);
573 }
574}
575
576static void copy_huge_page(struct page *dst, struct page *src)
577{
578 int i;
579 int nr_pages;
580
581 if (PageHuge(src)) {
582 /* hugetlbfs page */
583 struct hstate *h = page_hstate(src);
584 nr_pages = pages_per_huge_page(h);
585
586 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
587 __copy_gigantic_page(dst, src, nr_pages);
588 return;
589 }
590 } else {
591 /* thp page */
592 BUG_ON(!PageTransHuge(src));
593 nr_pages = hpage_nr_pages(src);
594 }
595
596 for (i = 0; i < nr_pages; i++) {
597 cond_resched();
598 copy_highpage(dst + i, src + i);
599 }
600}
601
602/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800603 * Copy the page to its new location
604 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900605void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800606{
Rik van Riel7851a452013-10-07 11:29:23 +0100607 int cpupid;
608
Mel Gormanb32967f2012-11-19 12:35:47 +0000609 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900610 copy_huge_page(newpage, page);
611 else
612 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800613
614 if (PageError(page))
615 SetPageError(newpage);
616 if (PageReferenced(page))
617 SetPageReferenced(newpage);
618 if (PageUptodate(page))
619 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700620 if (TestClearPageActive(page)) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800621 VM_BUG_ON_PAGE(PageUnevictable(page), page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800622 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800623 } else if (TestClearPageUnevictable(page))
624 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800625 if (PageChecked(page))
626 SetPageChecked(newpage);
627 if (PageMappedToDisk(page))
628 SetPageMappedToDisk(newpage);
629
Hugh Dickins42cb14b2015-11-05 18:50:05 -0800630 /* Move dirty on pages not done by migrate_page_move_mapping() */
631 if (PageDirty(page))
632 SetPageDirty(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800633
Vladimir Davydov33c3fc72015-09-09 15:35:45 -0700634 if (page_is_young(page))
635 set_page_young(newpage);
636 if (page_is_idle(page))
637 set_page_idle(newpage);
638
Rik van Riel7851a452013-10-07 11:29:23 +0100639 /*
640 * Copy NUMA information to the new page, to prevent over-eager
641 * future migrations of this same page.
642 */
643 cpupid = page_cpupid_xchg_last(page, -1);
644 page_cpupid_xchg_last(newpage, cpupid);
645
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800646 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800647 /*
648 * Please do not reorder this without considering how mm/ksm.c's
649 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
650 */
Naoya Horiguchib3b3a992015-04-15 16:13:15 -0700651 if (PageSwapCache(page))
652 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800653 ClearPagePrivate(page);
654 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800655
656 /*
657 * If any waiters have accumulated on the new page then
658 * wake them up.
659 */
660 if (PageWriteback(newpage))
661 end_page_writeback(newpage);
Vlastimil Babkad435edc2016-03-15 14:56:15 -0700662
663 copy_page_owner(page, newpage);
Johannes Weiner74485cf2016-03-15 14:57:54 -0700664
665 mem_cgroup_migrate(page, newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800666}
Richard Weinberger1118dce2016-06-16 23:26:14 +0200667EXPORT_SYMBOL(migrate_page_copy);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800668
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700669/************************************************************
670 * Migration functions
671 ***********************************************************/
672
Christoph Lameterb20a3502006-03-22 00:09:12 -0800673/*
Minchan Kimbda807d2016-07-26 15:23:05 -0700674 * Common logic to directly migrate a single LRU page suitable for
David Howells266cf652009-04-03 16:42:36 +0100675 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800676 *
677 * Pages are locked upon entry and exit.
678 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700679int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800680 struct page *newpage, struct page *page,
681 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800682{
683 int rc;
684
685 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
686
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500687 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800688
Rafael Aquini78bd5202012-12-11 16:02:31 -0800689 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800690 return rc;
691
692 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800693 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800694}
695EXPORT_SYMBOL(migrate_page);
696
David Howells93614012006-09-30 20:45:40 +0200697#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800698/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700699 * Migration function for pages with buffers. This function can only be used
700 * if the underlying filesystem guarantees that no other references to "page"
701 * exist.
702 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700703int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800704 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700705{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700706 struct buffer_head *bh, *head;
707 int rc;
708
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700709 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800710 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700711
712 head = page_buffers(page);
713
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500714 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700715
Rafael Aquini78bd5202012-12-11 16:02:31 -0800716 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700717 return rc;
718
Mel Gormanb969c4a2012-01-12 17:19:34 -0800719 /*
720 * In the async case, migrate_page_move_mapping locked the buffers
721 * with an IRQ-safe spinlock held. In the sync case, the buffers
722 * need to be locked now
723 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800724 if (mode != MIGRATE_ASYNC)
725 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700726
727 ClearPagePrivate(page);
728 set_page_private(newpage, page_private(page));
729 set_page_private(page, 0);
730 put_page(page);
731 get_page(newpage);
732
733 bh = head;
734 do {
735 set_bh_page(bh, newpage, bh_offset(bh));
736 bh = bh->b_this_page;
737
738 } while (bh != head);
739
740 SetPagePrivate(newpage);
741
742 migrate_page_copy(newpage, page);
743
744 bh = head;
745 do {
746 unlock_buffer(bh);
747 put_bh(bh);
748 bh = bh->b_this_page;
749
750 } while (bh != head);
751
Rafael Aquini78bd5202012-12-11 16:02:31 -0800752 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700753}
754EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200755#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700756
Christoph Lameter04e62a22006-06-23 02:03:38 -0700757/*
758 * Writeback a page to clean the dirty state
759 */
760static int writeout(struct address_space *mapping, struct page *page)
761{
762 struct writeback_control wbc = {
763 .sync_mode = WB_SYNC_NONE,
764 .nr_to_write = 1,
765 .range_start = 0,
766 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700767 .for_reclaim = 1
768 };
769 int rc;
770
771 if (!mapping->a_ops->writepage)
772 /* No write method for the address space */
773 return -EINVAL;
774
775 if (!clear_page_dirty_for_io(page))
776 /* Someone else already triggered a write */
777 return -EAGAIN;
778
779 /*
780 * A dirty page may imply that the underlying filesystem has
781 * the page on some queue. So the page must be clean for
782 * migration. Writeout may mean we loose the lock and the
783 * page state is no longer what we checked for earlier.
784 * At this point we know that the migration attempt cannot
785 * be successful.
786 */
Kirill A. Shutemove3884662016-03-17 14:20:07 -0700787 remove_migration_ptes(page, page, false);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700788
789 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700790
791 if (rc != AOP_WRITEPAGE_ACTIVATE)
792 /* unlocked. Relock */
793 lock_page(page);
794
Hugh Dickinsbda85502008-11-19 15:36:36 -0800795 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700796}
797
798/*
799 * Default handling if a filesystem does not provide a migration function.
800 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700801static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800802 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700803{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800804 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800805 /* Only writeback pages in full synchronous migration */
806 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800807 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700808 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800809 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700810
811 /*
812 * Buffers may be managed in a filesystem specific way.
813 * We must have no buffers or drop them.
814 */
David Howells266cf652009-04-03 16:42:36 +0100815 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700816 !try_to_release_page(page, GFP_KERNEL))
817 return -EAGAIN;
818
Mel Gormana6bc32b2012-01-12 17:19:43 -0800819 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700820}
821
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700822/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700823 * Move a page to a newly allocated page
824 * The page is locked and all ptes have been successfully removed.
825 *
826 * The new page will have replaced the old page if this function
827 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700828 *
829 * Return value:
830 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800831 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700832 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700833static int move_to_new_page(struct page *newpage, struct page *page,
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800834 enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700835{
836 struct address_space *mapping;
Minchan Kimbda807d2016-07-26 15:23:05 -0700837 int rc = -EAGAIN;
838 bool is_lru = !__PageMovable(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700839
Hugh Dickins7db76712015-11-05 18:49:49 -0800840 VM_BUG_ON_PAGE(!PageLocked(page), page);
841 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700842
Christoph Lametere24f0b82006-06-23 02:03:51 -0700843 mapping = page_mapping(page);
Minchan Kimbda807d2016-07-26 15:23:05 -0700844
845 if (likely(is_lru)) {
846 if (!mapping)
847 rc = migrate_page(mapping, newpage, page, mode);
848 else if (mapping->a_ops->migratepage)
849 /*
850 * Most pages have a mapping and most filesystems
851 * provide a migratepage callback. Anonymous pages
852 * are part of swap space which also has its own
853 * migratepage callback. This is the most common path
854 * for page migration.
855 */
856 rc = mapping->a_ops->migratepage(mapping, newpage,
857 page, mode);
858 else
859 rc = fallback_migrate_page(mapping, newpage,
860 page, mode);
861 } else {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700862 /*
Minchan Kimbda807d2016-07-26 15:23:05 -0700863 * In case of non-lru page, it could be released after
864 * isolation step. In that case, we shouldn't try migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700865 */
Minchan Kimbda807d2016-07-26 15:23:05 -0700866 VM_BUG_ON_PAGE(!PageIsolated(page), page);
867 if (!PageMovable(page)) {
868 rc = MIGRATEPAGE_SUCCESS;
869 __ClearPageIsolated(page);
870 goto out;
871 }
872
873 rc = mapping->a_ops->migratepage(mapping, newpage,
874 page, mode);
875 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
876 !PageIsolated(page));
877 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700878
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800879 /*
880 * When successful, old pagecache page->mapping must be cleared before
881 * page is freed; but stats require that PageAnon be left as PageAnon.
882 */
883 if (rc == MIGRATEPAGE_SUCCESS) {
Minchan Kimbda807d2016-07-26 15:23:05 -0700884 if (__PageMovable(page)) {
885 VM_BUG_ON_PAGE(!PageIsolated(page), page);
886
887 /*
888 * We clear PG_movable under page_lock so any compactor
889 * cannot try to migrate this page.
890 */
891 __ClearPageIsolated(page);
892 }
893
894 /*
895 * Anonymous and movable page->mapping will be cleard by
896 * free_pages_prepare so don't reset it here for keeping
897 * the type to work PageAnon, for example.
898 */
899 if (!PageMappingFlags(page))
Hugh Dickins5c3f9a62015-11-05 18:49:53 -0800900 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700901 }
Minchan Kimbda807d2016-07-26 15:23:05 -0700902out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700903 return rc;
904}
905
Minchan Kim0dabec92011-10-31 17:06:57 -0700906static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800907 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700908{
Minchan Kim0dabec92011-10-31 17:06:57 -0700909 int rc = -EAGAIN;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800910 int page_was_mapped = 0;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700911 struct anon_vma *anon_vma = NULL;
Minchan Kimbda807d2016-07-26 15:23:05 -0700912 bool is_lru = !__PageMovable(page);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700913
Nick Piggin529ae9a2008-08-02 12:01:03 +0200914 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800915 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700916 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800917
918 /*
919 * It's not safe for direct compaction to call lock_page.
920 * For example, during page readahead pages are added locked
921 * to the LRU. Later, when the IO completes the pages are
922 * marked uptodate and unlocked. However, the queueing
923 * could be merging multiple pages for one bio (e.g.
924 * mpage_readpages). If an allocation happens for the
925 * second or third page, the process can end up locking
926 * the same page twice and deadlocking. Rather than
927 * trying to be clever about what pages can be locked,
928 * avoid the use of lock_page for direct compaction
929 * altogether.
930 */
931 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700932 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800933
Christoph Lametere24f0b82006-06-23 02:03:51 -0700934 lock_page(page);
935 }
936
937 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700938 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700939 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800940 * necessary to wait for PageWriteback. In the async case,
941 * the retry loop is too short and in the sync-light case,
942 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700943 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800944 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700945 rc = -EBUSY;
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700946 goto out_unlock;
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700947 }
948 if (!force)
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700949 goto out_unlock;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700950 wait_on_page_writeback(page);
951 }
Hugh Dickins03f15c82015-11-05 18:49:56 -0800952
Christoph Lametere24f0b82006-06-23 02:03:51 -0700953 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700954 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
955 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800956 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700957 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700958 * File Caches may use write_page() or lock_page() in migration, then,
959 * just care Anon page here.
Hugh Dickins03f15c82015-11-05 18:49:56 -0800960 *
961 * Only page_get_anon_vma() understands the subtleties of
962 * getting a hold on an anon_vma from outside one of its mms.
963 * But if we cannot get anon_vma, then we won't need it anyway,
964 * because that implies that the anon page is no longer mapped
965 * (and cannot be remapped so long as we hold the page lock).
Christoph Lametere24f0b82006-06-23 02:03:51 -0700966 */
Hugh Dickins03f15c82015-11-05 18:49:56 -0800967 if (PageAnon(page) && !PageKsm(page))
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700968 anon_vma = page_get_anon_vma(page);
Shaohua Li62e1c552008-02-04 22:29:33 -0800969
Hugh Dickins7db76712015-11-05 18:49:49 -0800970 /*
971 * Block others from accessing the new page when we get around to
972 * establishing additional references. We are usually the only one
973 * holding a reference to newpage at this point. We used to have a BUG
974 * here if trylock_page(newpage) fails, but would like to allow for
975 * cases where there might be a race with the previous use of newpage.
976 * This is much like races on refcount of oldpage: just don't BUG().
977 */
978 if (unlikely(!trylock_page(newpage)))
979 goto out_unlock;
980
Minchan Kimbda807d2016-07-26 15:23:05 -0700981 if (unlikely(!is_lru)) {
982 rc = move_to_new_page(newpage, page, mode);
983 goto out_unlock_both;
984 }
985
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700986 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800987 * Corner case handling:
988 * 1. When a new swap-cache page is read into, it is added to the LRU
989 * and treated as swapcache but it has no rmap yet.
990 * Calling try_to_unmap() against a page->mapping==NULL page will
991 * trigger a BUG. So handle it here.
992 * 2. An orphaned page (see truncate_complete_page) might have
993 * fs-private metadata. The page can be picked up due to memory
994 * offlining. Everywhere else except page reclaim, the page is
995 * invisible to the vm, so the page can not be migrated. So try to
996 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700997 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800998 if (!page->mapping) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800999 VM_BUG_ON_PAGE(PageAnon(page), page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -08001000 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -08001001 try_to_free_buffers(page);
Hugh Dickins7db76712015-11-05 18:49:49 -08001002 goto out_unlock_both;
Shaohua Li62e1c552008-02-04 22:29:33 -08001003 }
Hugh Dickins7db76712015-11-05 18:49:49 -08001004 } else if (page_mapped(page)) {
1005 /* Establish migration ptes */
Hugh Dickins03f15c82015-11-05 18:49:56 -08001006 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1007 page);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001008 try_to_unmap(page,
Wanpeng Lida1b13c2015-09-08 15:03:27 -07001009 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001010 page_was_mapped = 1;
1011 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -07001012
Christoph Lametere6a15302006-06-25 05:46:49 -07001013 if (!page_mapped(page))
Hugh Dickins5c3f9a62015-11-05 18:49:53 -08001014 rc = move_to_new_page(newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -07001015
Hugh Dickins5c3f9a62015-11-05 18:49:53 -08001016 if (page_was_mapped)
1017 remove_migration_ptes(page,
Kirill A. Shutemove3884662016-03-17 14:20:07 -07001018 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
Mel Gorman3f6c8272010-05-24 14:32:17 -07001019
Hugh Dickins7db76712015-11-05 18:49:49 -08001020out_unlock_both:
1021 unlock_page(newpage);
1022out_unlock:
Mel Gorman3f6c8272010-05-24 14:32:17 -07001023 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -07001024 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001025 put_anon_vma(anon_vma);
Christoph Lametere24f0b82006-06-23 02:03:51 -07001026 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -07001027out:
Minchan Kimc6c919e2016-07-26 15:23:02 -07001028 /*
1029 * If migration is successful, decrease refcount of the newpage
1030 * which will not free the page because new page owner increased
1031 * refcounter. As well, if it is LRU page, add the page to LRU
1032 * list in here.
1033 */
1034 if (rc == MIGRATEPAGE_SUCCESS) {
Minchan Kimb1123ea62016-07-26 15:23:09 -07001035 if (unlikely(__PageMovable(newpage)))
Minchan Kimc6c919e2016-07-26 15:23:02 -07001036 put_page(newpage);
1037 else
1038 putback_lru_page(newpage);
1039 }
1040
Minchan Kim0dabec92011-10-31 17:06:57 -07001041 return rc;
1042}
Christoph Lameter95a402c2006-06-23 02:03:53 -07001043
Minchan Kim0dabec92011-10-31 17:06:57 -07001044/*
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -07001045 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1046 * around it.
1047 */
1048#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1049#define ICE_noinline noinline
1050#else
1051#define ICE_noinline
1052#endif
1053
1054/*
Minchan Kim0dabec92011-10-31 17:06:57 -07001055 * Obtain the lock on page, remove all ptes and migrate the page
1056 * to the newly allocated page in newpage.
1057 */
Geert Uytterhoevenef2a5152015-04-14 15:44:22 -07001058static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1059 free_page_t put_new_page,
1060 unsigned long private, struct page *page,
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -07001061 int force, enum migrate_mode mode,
1062 enum migrate_reason reason)
Minchan Kim0dabec92011-10-31 17:06:57 -07001063{
Hugh Dickins2def7422015-11-05 18:49:46 -08001064 int rc = MIGRATEPAGE_SUCCESS;
Minchan Kim0dabec92011-10-31 17:06:57 -07001065 int *result = NULL;
Hugh Dickins2def7422015-11-05 18:49:46 -08001066 struct page *newpage;
Minchan Kim0dabec92011-10-31 17:06:57 -07001067
Hugh Dickins2def7422015-11-05 18:49:46 -08001068 newpage = get_new_page(page, private, &result);
Minchan Kim0dabec92011-10-31 17:06:57 -07001069 if (!newpage)
1070 return -ENOMEM;
1071
1072 if (page_count(page) == 1) {
1073 /* page was freed from under us. So we are done. */
Minchan Kimc6c919e2016-07-26 15:23:02 -07001074 ClearPageActive(page);
1075 ClearPageUnevictable(page);
Minchan Kimbda807d2016-07-26 15:23:05 -07001076 if (unlikely(__PageMovable(page))) {
1077 lock_page(page);
1078 if (!PageMovable(page))
1079 __ClearPageIsolated(page);
1080 unlock_page(page);
1081 }
Minchan Kimc6c919e2016-07-26 15:23:02 -07001082 if (put_new_page)
1083 put_new_page(newpage, private);
1084 else
1085 put_page(newpage);
Minchan Kim0dabec92011-10-31 17:06:57 -07001086 goto out;
1087 }
1088
Kirill A. Shutemov4d2fa962016-01-15 16:54:00 -08001089 if (unlikely(PageTransHuge(page))) {
1090 lock_page(page);
1091 rc = split_huge_page(page);
1092 unlock_page(page);
1093 if (rc)
Minchan Kim0dabec92011-10-31 17:06:57 -07001094 goto out;
Kirill A. Shutemov4d2fa962016-01-15 16:54:00 -08001095 }
Minchan Kim0dabec92011-10-31 17:06:57 -07001096
Hugh Dickins9c620e22013-02-22 16:35:14 -08001097 rc = __unmap_and_move(page, newpage, force, mode);
Minchan Kimc6c919e2016-07-26 15:23:02 -07001098 if (rc == MIGRATEPAGE_SUCCESS)
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001099 set_page_owner_migrate_reason(newpage, reason);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -08001100
Minchan Kim0dabec92011-10-31 17:06:57 -07001101out:
Christoph Lametere24f0b82006-06-23 02:03:51 -07001102 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -07001103 /*
1104 * A page that has been migrated has all references
1105 * removed and will be freed. A page that has not been
1106 * migrated will have kepts its references and be
1107 * restored.
1108 */
1109 list_del(&page->lru);
Ming Ling6afcf8e2016-12-12 16:42:26 -08001110
1111 /*
1112 * Compaction can migrate also non-LRU pages which are
1113 * not accounted to NR_ISOLATED_*. They can be recognized
1114 * as __PageMovable
1115 */
1116 if (likely(!__PageMovable(page)))
1117 dec_node_page_state(page, NR_ISOLATED_ANON +
1118 page_is_file_cache(page));
Christoph Lametere24f0b82006-06-23 02:03:51 -07001119 }
David Rientjes68711a72014-06-04 16:08:25 -07001120
Christoph Lameter95a402c2006-06-23 02:03:53 -07001121 /*
Minchan Kimc6c919e2016-07-26 15:23:02 -07001122 * If migration is successful, releases reference grabbed during
1123 * isolation. Otherwise, restore the page to right list unless
1124 * we want to retry.
Christoph Lameter95a402c2006-06-23 02:03:53 -07001125 */
Minchan Kimc6c919e2016-07-26 15:23:02 -07001126 if (rc == MIGRATEPAGE_SUCCESS) {
1127 put_page(page);
1128 if (reason == MR_MEMORY_FAILURE) {
1129 /*
1130 * Set PG_HWPoison on just freed page
1131 * intentionally. Although it's rather weird,
1132 * it's how HWPoison flag works at the moment.
1133 */
1134 if (!test_set_page_hwpoison(page))
1135 num_poisoned_pages_inc();
1136 }
1137 } else {
Minchan Kimbda807d2016-07-26 15:23:05 -07001138 if (rc != -EAGAIN) {
1139 if (likely(!__PageMovable(page))) {
1140 putback_lru_page(page);
1141 goto put_new;
1142 }
1143
1144 lock_page(page);
1145 if (PageMovable(page))
1146 putback_movable_page(page);
1147 else
1148 __ClearPageIsolated(page);
1149 unlock_page(page);
1150 put_page(page);
1151 }
1152put_new:
Minchan Kimc6c919e2016-07-26 15:23:02 -07001153 if (put_new_page)
1154 put_new_page(newpage, private);
1155 else
1156 put_page(newpage);
1157 }
David Rientjes68711a72014-06-04 16:08:25 -07001158
Christoph Lameter742755a2006-06-23 02:03:55 -07001159 if (result) {
1160 if (rc)
1161 *result = rc;
1162 else
1163 *result = page_to_nid(newpage);
1164 }
Christoph Lametere24f0b82006-06-23 02:03:51 -07001165 return rc;
1166}
1167
1168/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001169 * Counterpart of unmap_and_move_page() for hugepage migration.
1170 *
1171 * This function doesn't wait the completion of hugepage I/O
1172 * because there is no race between I/O and migration for hugepage.
1173 * Note that currently hugepage I/O occurs only in direct I/O
1174 * where no lock is held and PG_writeback is irrelevant,
1175 * and writeback status of all subpages are counted in the reference
1176 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1177 * under direct I/O, the reference of the head page is 512 and a bit more.)
1178 * This means that when we try to migrate hugepage whose subpages are
1179 * doing direct I/O, some references remain after try_to_unmap() and
1180 * hugepage migration fails without data corruption.
1181 *
1182 * There is also no race when direct I/O is issued on the page under migration,
1183 * because then pte is replaced with migration swap entry and direct I/O code
1184 * will wait in the page fault for migration to complete.
1185 */
1186static int unmap_and_move_huge_page(new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001187 free_page_t put_new_page, unsigned long private,
1188 struct page *hpage, int force,
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001189 enum migrate_mode mode, int reason)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001190{
Hugh Dickins2def7422015-11-05 18:49:46 -08001191 int rc = -EAGAIN;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001192 int *result = NULL;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001193 int page_was_mapped = 0;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001194 struct page *new_hpage;
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001195 struct anon_vma *anon_vma = NULL;
1196
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001197 /*
1198 * Movability of hugepages depends on architectures and hugepage size.
1199 * This check is necessary because some callers of hugepage migration
1200 * like soft offline and memory hotremove don't walk through page
1201 * tables or check whether the hugepage is pmd-based or not before
1202 * kicking migration.
1203 */
Naoya Horiguchi100873d2014-06-04 16:10:56 -07001204 if (!hugepage_migration_supported(page_hstate(hpage))) {
Joonsoo Kim32665f22014-01-21 15:51:15 -08001205 putback_active_hugepage(hpage);
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001206 return -ENOSYS;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001207 }
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001208
Joonsoo Kim32665f22014-01-21 15:51:15 -08001209 new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001210 if (!new_hpage)
1211 return -ENOMEM;
1212
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001213 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001214 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001215 goto out;
1216 lock_page(hpage);
1217 }
1218
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001219 if (PageAnon(hpage))
1220 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001221
Hugh Dickins7db76712015-11-05 18:49:49 -08001222 if (unlikely(!trylock_page(new_hpage)))
1223 goto put_anon;
1224
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001225 if (page_mapped(hpage)) {
1226 try_to_unmap(hpage,
1227 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1228 page_was_mapped = 1;
1229 }
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001230
1231 if (!page_mapped(hpage))
Hugh Dickins5c3f9a62015-11-05 18:49:53 -08001232 rc = move_to_new_page(new_hpage, hpage, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001233
Hugh Dickins5c3f9a62015-11-05 18:49:53 -08001234 if (page_was_mapped)
1235 remove_migration_ptes(hpage,
Kirill A. Shutemove3884662016-03-17 14:20:07 -07001236 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001237
Hugh Dickins7db76712015-11-05 18:49:49 -08001238 unlock_page(new_hpage);
1239
1240put_anon:
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001241 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001242 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001243
Hugh Dickins2def7422015-11-05 18:49:46 -08001244 if (rc == MIGRATEPAGE_SUCCESS) {
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001245 hugetlb_cgroup_migrate(hpage, new_hpage);
Hugh Dickins2def7422015-11-05 18:49:46 -08001246 put_new_page = NULL;
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001247 set_page_owner_migrate_reason(new_hpage, reason);
Hugh Dickins2def7422015-11-05 18:49:46 -08001248 }
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001249
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001250 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001251out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001252 if (rc != -EAGAIN)
1253 putback_active_hugepage(hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001254
1255 /*
1256 * If migration was not successful and there's a freeing callback, use
1257 * it. Otherwise, put_page() will drop the reference grabbed during
1258 * isolation.
1259 */
Hugh Dickins2def7422015-11-05 18:49:46 -08001260 if (put_new_page)
David Rientjes68711a72014-06-04 16:08:25 -07001261 put_new_page(new_hpage, private);
1262 else
Naoya Horiguchi3aaa76e2015-09-22 14:59:14 -07001263 putback_active_hugepage(new_hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001264
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001265 if (result) {
1266 if (rc)
1267 *result = rc;
1268 else
1269 *result = page_to_nid(new_hpage);
1270 }
1271 return rc;
1272}
1273
1274/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001275 * migrate_pages - migrate the pages specified in a list, to the free pages
1276 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001277 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001278 * @from: The list of pages to be migrated.
1279 * @get_new_page: The function used to allocate free pages to be used
1280 * as the target of the page migration.
David Rientjes68711a72014-06-04 16:08:25 -07001281 * @put_new_page: The function used to free target pages if migration
1282 * fails, or NULL if no special handling is necessary.
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001283 * @private: Private data to be passed on to get_new_page()
1284 * @mode: The migration mode that specifies the constraints for
1285 * page migration, if any.
1286 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001287 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001288 * The function returns after 10 attempts or if no pages are movable any more
1289 * because the list has become empty or no retryable pages exist any more.
Hugh Dickins14e0f9b2015-11-05 18:49:43 -08001290 * The caller should call putback_movable_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001291 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001292 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001293 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001294 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001295int migrate_pages(struct list_head *from, new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001296 free_page_t put_new_page, unsigned long private,
1297 enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001298{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001299 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001300 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001301 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001302 int pass = 0;
1303 struct page *page;
1304 struct page *page2;
1305 int swapwrite = current->flags & PF_SWAPWRITE;
1306 int rc;
1307
1308 if (!swapwrite)
1309 current->flags |= PF_SWAPWRITE;
1310
Christoph Lametere24f0b82006-06-23 02:03:51 -07001311 for(pass = 0; pass < 10 && retry; pass++) {
1312 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001313
Christoph Lametere24f0b82006-06-23 02:03:51 -07001314 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001315 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001316
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001317 if (PageHuge(page))
1318 rc = unmap_and_move_huge_page(get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001319 put_new_page, private, page,
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07001320 pass > 2, mode, reason);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001321 else
David Rientjes68711a72014-06-04 16:08:25 -07001322 rc = unmap_and_move(get_new_page, put_new_page,
Naoya Horiguchiadd05ce2015-06-24 16:56:50 -07001323 private, page, pass > 2, mode,
1324 reason);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001325
Christoph Lametere24f0b82006-06-23 02:03:51 -07001326 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001327 case -ENOMEM:
David Rientjesdfef2ef2016-05-20 16:59:05 -07001328 nr_failed++;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001329 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001330 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001331 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001332 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001333 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001334 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001335 break;
1336 default:
Naoya Horiguchi354a3362014-01-21 15:51:14 -08001337 /*
1338 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1339 * unlike -EAGAIN case, the failed page is
1340 * removed from migration page list and not
1341 * retried in the next outer loop.
1342 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001343 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001344 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001345 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001346 }
1347 }
Vlastimil Babkaf2f81fb2015-11-05 18:47:03 -08001348 nr_failed += retry;
1349 rc = nr_failed;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001350out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001351 if (nr_succeeded)
1352 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1353 if (nr_failed)
1354 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001355 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1356
Christoph Lameterb20a3502006-03-22 00:09:12 -08001357 if (!swapwrite)
1358 current->flags &= ~PF_SWAPWRITE;
1359
Rafael Aquini78bd5202012-12-11 16:02:31 -08001360 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001361}
1362
Christoph Lameter742755a2006-06-23 02:03:55 -07001363#ifdef CONFIG_NUMA
1364/*
1365 * Move a list of individual pages
1366 */
1367struct page_to_node {
1368 unsigned long addr;
1369 struct page *page;
1370 int node;
1371 int status;
1372};
1373
1374static struct page *new_page_node(struct page *p, unsigned long private,
1375 int **result)
1376{
1377 struct page_to_node *pm = (struct page_to_node *)private;
1378
1379 while (pm->node != MAX_NUMNODES && pm->page != p)
1380 pm++;
1381
1382 if (pm->node == MAX_NUMNODES)
1383 return NULL;
1384
1385 *result = &pm->status;
1386
Naoya Horiguchie632a932013-09-11 14:22:04 -07001387 if (PageHuge(p))
1388 return alloc_huge_page_node(page_hstate(compound_head(p)),
1389 pm->node);
1390 else
Vlastimil Babka96db8002015-09-08 15:03:50 -07001391 return __alloc_pages_node(pm->node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001392 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001393}
1394
1395/*
1396 * Move a set of pages as indicated in the pm array. The addr
1397 * field must be set to the virtual address of the page to be moved
1398 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001399 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001400 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001401static int do_move_page_to_node_array(struct mm_struct *mm,
1402 struct page_to_node *pm,
1403 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001404{
1405 int err;
1406 struct page_to_node *pp;
1407 LIST_HEAD(pagelist);
1408
1409 down_read(&mm->mmap_sem);
1410
1411 /*
1412 * Build a list of pages to migrate
1413 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001414 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1415 struct vm_area_struct *vma;
1416 struct page *page;
1417
Christoph Lameter742755a2006-06-23 02:03:55 -07001418 err = -EFAULT;
1419 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001420 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001421 goto set_status;
1422
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001423 /* FOLL_DUMP to ignore special (like zero) pages */
1424 page = follow_page(vma, pp->addr,
1425 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001426
1427 err = PTR_ERR(page);
1428 if (IS_ERR(page))
1429 goto set_status;
1430
Christoph Lameter742755a2006-06-23 02:03:55 -07001431 err = -ENOENT;
1432 if (!page)
1433 goto set_status;
1434
Christoph Lameter742755a2006-06-23 02:03:55 -07001435 pp->page = page;
1436 err = page_to_nid(page);
1437
1438 if (err == pp->node)
1439 /*
1440 * Node already in the right place
1441 */
1442 goto put_and_set;
1443
1444 err = -EACCES;
1445 if (page_mapcount(page) > 1 &&
1446 !migrate_all)
1447 goto put_and_set;
1448
Naoya Horiguchie632a932013-09-11 14:22:04 -07001449 if (PageHuge(page)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08001450 if (PageHead(page))
1451 isolate_huge_page(page, &pagelist);
Naoya Horiguchie632a932013-09-11 14:22:04 -07001452 goto put_and_set;
1453 }
1454
Nick Piggin62695a82008-10-18 20:26:09 -07001455 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001456 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001457 list_add_tail(&page->lru, &pagelist);
Mel Gorman599d0c92016-07-28 15:45:31 -07001458 inc_node_page_state(page, NR_ISOLATED_ANON +
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001459 page_is_file_cache(page));
1460 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001461put_and_set:
1462 /*
1463 * Either remove the duplicate refcount from
1464 * isolate_lru_page() or drop the page ref if it was
1465 * not isolated.
1466 */
1467 put_page(page);
1468set_status:
1469 pp->status = err;
1470 }
1471
Brice Gogline78bbfa2008-10-18 20:27:15 -07001472 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001473 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -07001474 err = migrate_pages(&pagelist, new_page_node, NULL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001475 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001476 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001477 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001478 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001479
1480 up_read(&mm->mmap_sem);
1481 return err;
1482}
1483
1484/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001485 * Migrate an array of page address onto an array of nodes and fill
1486 * the corresponding array of status.
1487 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001488static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001489 unsigned long nr_pages,
1490 const void __user * __user *pages,
1491 const int __user *nodes,
1492 int __user *status, int flags)
1493{
Brice Goglin3140a222009-01-06 14:38:57 -08001494 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001495 unsigned long chunk_nr_pages;
1496 unsigned long chunk_start;
1497 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001498
Brice Goglin3140a222009-01-06 14:38:57 -08001499 err = -ENOMEM;
1500 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1501 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001502 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001503
1504 migrate_prep();
1505
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001506 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001507 * Store a chunk of page_to_node array in a page,
1508 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001509 */
Brice Goglin3140a222009-01-06 14:38:57 -08001510 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001511
Brice Goglin3140a222009-01-06 14:38:57 -08001512 for (chunk_start = 0;
1513 chunk_start < nr_pages;
1514 chunk_start += chunk_nr_pages) {
1515 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001516
Brice Goglin3140a222009-01-06 14:38:57 -08001517 if (chunk_start + chunk_nr_pages > nr_pages)
1518 chunk_nr_pages = nr_pages - chunk_start;
1519
1520 /* fill the chunk pm with addrs and nodes from user-space */
1521 for (j = 0; j < chunk_nr_pages; j++) {
1522 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001523 int node;
1524
Brice Goglin3140a222009-01-06 14:38:57 -08001525 err = -EFAULT;
1526 if (get_user(p, pages + j + chunk_start))
1527 goto out_pm;
1528 pm[j].addr = (unsigned long) p;
1529
1530 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001531 goto out_pm;
1532
1533 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001534 if (node < 0 || node >= MAX_NUMNODES)
1535 goto out_pm;
1536
Lai Jiangshan389162c2012-12-12 13:51:30 -08001537 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001538 goto out_pm;
1539
1540 err = -EACCES;
1541 if (!node_isset(node, task_nodes))
1542 goto out_pm;
1543
Brice Goglin3140a222009-01-06 14:38:57 -08001544 pm[j].node = node;
1545 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001546
Brice Goglin3140a222009-01-06 14:38:57 -08001547 /* End marker for this chunk */
1548 pm[chunk_nr_pages].node = MAX_NUMNODES;
1549
1550 /* Migrate this chunk */
1551 err = do_move_page_to_node_array(mm, pm,
1552 flags & MPOL_MF_MOVE_ALL);
1553 if (err < 0)
1554 goto out_pm;
1555
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001556 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001557 for (j = 0; j < chunk_nr_pages; j++)
1558 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001559 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001560 goto out_pm;
1561 }
1562 }
1563 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001564
1565out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001566 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001567out:
1568 return err;
1569}
1570
1571/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001572 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001573 */
Brice Goglin80bba122008-12-09 13:14:23 -08001574static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1575 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001576{
Brice Goglin2f007e72008-10-18 20:27:16 -07001577 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001578
Christoph Lameter742755a2006-06-23 02:03:55 -07001579 down_read(&mm->mmap_sem);
1580
Brice Goglin2f007e72008-10-18 20:27:16 -07001581 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001582 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001583 struct vm_area_struct *vma;
1584 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001585 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001586
1587 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001588 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001589 goto set_status;
1590
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001591 /* FOLL_DUMP to ignore special (like zero) pages */
1592 page = follow_page(vma, addr, FOLL_DUMP);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001593
1594 err = PTR_ERR(page);
1595 if (IS_ERR(page))
1596 goto set_status;
1597
Kirill A. Shutemovd8998442015-09-04 15:47:53 -07001598 err = page ? page_to_nid(page) : -ENOENT;
Christoph Lameter742755a2006-06-23 02:03:55 -07001599set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001600 *status = err;
1601
1602 pages++;
1603 status++;
1604 }
1605
1606 up_read(&mm->mmap_sem);
1607}
1608
1609/*
1610 * Determine the nodes of a user array of pages and store it in
1611 * a user array of status.
1612 */
1613static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1614 const void __user * __user *pages,
1615 int __user *status)
1616{
1617#define DO_PAGES_STAT_CHUNK_NR 16
1618 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1619 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001620
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001621 while (nr_pages) {
1622 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001623
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001624 chunk_nr = nr_pages;
1625 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1626 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1627
1628 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1629 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001630
1631 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1632
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001633 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1634 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001635
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001636 pages += chunk_nr;
1637 status += chunk_nr;
1638 nr_pages -= chunk_nr;
1639 }
1640 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001641}
1642
1643/*
1644 * Move a list of pages in the address space of the currently executing
1645 * process.
1646 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001647SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1648 const void __user * __user *, pages,
1649 const int __user *, nodes,
1650 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001651{
David Howellsc69e8d92008-11-14 10:39:19 +11001652 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001653 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001654 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001655 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001656 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001657
1658 /* Check flags */
1659 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1660 return -EINVAL;
1661
1662 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1663 return -EPERM;
1664
1665 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001666 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001667 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001668 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001669 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001670 return -ESRCH;
1671 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001672 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001673
1674 /*
1675 * Check if this process has the right to modify the specified
1676 * process. The right exists if the process has administrative
1677 * capabilities, superuser privileges or the same
1678 * userid as the target process.
1679 */
David Howellsc69e8d92008-11-14 10:39:19 +11001680 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001681 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1682 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001683 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001684 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001685 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001686 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001687 }
David Howellsc69e8d92008-11-14 10:39:19 +11001688 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001689
David Quigley86c3a762006-06-23 02:04:02 -07001690 err = security_task_movememory(task);
1691 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001692 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001693
Christoph Lameter3268c632012-03-21 16:34:06 -07001694 task_nodes = cpuset_mems_allowed(task);
1695 mm = get_task_mm(task);
1696 put_task_struct(task);
1697
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001698 if (!mm)
1699 return -EINVAL;
1700
1701 if (nodes)
1702 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1703 nodes, status, flags);
1704 else
1705 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001706
1707 mmput(mm);
1708 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001709
Christoph Lameter742755a2006-06-23 02:03:55 -07001710out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001711 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001712 return err;
1713}
Christoph Lameter742755a2006-06-23 02:03:55 -07001714
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001715#ifdef CONFIG_NUMA_BALANCING
1716/*
1717 * Returns true if this is a safe migration target node for misplaced NUMA
1718 * pages. Currently it only checks the watermarks which crude
1719 */
1720static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001721 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001722{
1723 int z;
Mel Gorman599d0c92016-07-28 15:45:31 -07001724
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001725 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1726 struct zone *zone = pgdat->node_zones + z;
1727
1728 if (!populated_zone(zone))
1729 continue;
1730
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001731 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1732 if (!zone_watermark_ok(zone, 0,
1733 high_wmark_pages(zone) +
1734 nr_migrate_pages,
1735 0, 0))
1736 continue;
1737 return true;
1738 }
1739 return false;
1740}
1741
1742static struct page *alloc_misplaced_dst_page(struct page *page,
1743 unsigned long data,
1744 int **result)
1745{
1746 int nid = (int) data;
1747 struct page *newpage;
1748
Vlastimil Babka96db8002015-09-08 15:03:50 -07001749 newpage = __alloc_pages_node(nid,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001750 (GFP_HIGHUSER_MOVABLE |
1751 __GFP_THISNODE | __GFP_NOMEMALLOC |
1752 __GFP_NORETRY | __GFP_NOWARN) &
Mel Gorman8479eba2016-02-26 15:19:31 -08001753 ~__GFP_RECLAIM, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001754
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001755 return newpage;
1756}
1757
1758/*
Mel Gormana8f60772012-11-14 21:41:46 +00001759 * page migration rate limiting control.
1760 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1761 * window of time. Default here says do not migrate more than 1280M per second.
1762 */
1763static unsigned int migrate_interval_millisecs __read_mostly = 100;
1764static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1765
Mel Gormanb32967f2012-11-19 12:35:47 +00001766/* Returns true if the node is migrate rate-limited after the update */
Mel Gorman1c30e012014-01-21 15:50:58 -08001767static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1768 unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001769{
Mel Gormanb32967f2012-11-19 12:35:47 +00001770 /*
1771 * Rate-limit the amount of data that is being migrated to a node.
1772 * Optimal placement is no good if the memory bus is saturated and
1773 * all the time is being spent migrating!
1774 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001775 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001776 spin_lock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001777 pgdat->numabalancing_migrate_nr_pages = 0;
1778 pgdat->numabalancing_migrate_next_window = jiffies +
1779 msecs_to_jiffies(migrate_interval_millisecs);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001780 spin_unlock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001781 }
Mel Gormanaf1839d2014-01-21 15:51:01 -08001782 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1783 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1784 nr_pages);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001785 return true;
Mel Gormanaf1839d2014-01-21 15:51:01 -08001786 }
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001787
1788 /*
1789 * This is an unlocked non-atomic update so errors are possible.
1790 * The consequences are failing to migrate when we potentiall should
1791 * have which is not severe enough to warrant locking. If it is ever
1792 * a problem, it can be converted to a per-cpu counter.
1793 */
1794 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1795 return false;
Mel Gormanb32967f2012-11-19 12:35:47 +00001796}
1797
Mel Gorman1c30e012014-01-21 15:50:58 -08001798static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
Mel Gormanb32967f2012-11-19 12:35:47 +00001799{
Hugh Dickins340ef392013-02-22 16:34:33 -08001800 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001801
Sasha Levin309381fea2014-01-23 15:52:54 -08001802 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
Mel Gorman3abef4e2013-02-22 16:34:27 -08001803
Mel Gormanb32967f2012-11-19 12:35:47 +00001804 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001805 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1806 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001807
Hugh Dickins340ef392013-02-22 16:34:33 -08001808 if (isolate_lru_page(page))
1809 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001810
1811 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001812 * migrate_misplaced_transhuge_page() skips page migration's usual
1813 * check on page_count(), so we must do it here, now that the page
1814 * has been isolated: a GUP pin, or any other pin, prevents migration.
1815 * The expected page count is 3: 1 for page's mapcount and 1 for the
1816 * caller's pin and 1 for the reference taken by isolate_lru_page().
1817 */
1818 if (PageTransHuge(page) && page_count(page) != 3) {
1819 putback_lru_page(page);
1820 return 0;
1821 }
1822
1823 page_lru = page_is_file_cache(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001824 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
Hugh Dickins340ef392013-02-22 16:34:33 -08001825 hpage_nr_pages(page));
1826
1827 /*
1828 * Isolating the page has taken another reference, so the
1829 * caller's reference can be safely dropped without the page
1830 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001831 */
1832 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001833 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001834}
1835
Mel Gormande466bd2013-12-18 17:08:42 -08001836bool pmd_trans_migrating(pmd_t pmd)
1837{
1838 struct page *page = pmd_page(pmd);
1839 return PageLocked(page);
1840}
1841
Mel Gormana8f60772012-11-14 21:41:46 +00001842/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001843 * Attempt to migrate a misplaced page to the specified destination
1844 * node. Caller is expected to have an elevated reference count on
1845 * the page that will be dropped by this function before returning.
1846 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001847int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1848 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001849{
Mel Gormana8f60772012-11-14 21:41:46 +00001850 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001851 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001852 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001853 LIST_HEAD(migratepages);
1854
1855 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001856 * Don't migrate file pages that are mapped in multiple processes
1857 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001858 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001859 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1860 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001861 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001862
Mel Gormana8f60772012-11-14 21:41:46 +00001863 /*
1864 * Rate-limit the amount of data that is being migrated to a node.
1865 * Optimal placement is no good if the memory bus is saturated and
1866 * all the time is being spent migrating!
1867 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001868 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001869 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001870
Mel Gormanb32967f2012-11-19 12:35:47 +00001871 isolated = numamigrate_isolate_page(pgdat, page);
1872 if (!isolated)
1873 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001874
Mel Gormanb32967f2012-11-19 12:35:47 +00001875 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001876 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
David Rientjes68711a72014-06-04 16:08:25 -07001877 NULL, node, MIGRATE_ASYNC,
1878 MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001879 if (nr_remaining) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001880 if (!list_empty(&migratepages)) {
1881 list_del(&page->lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001882 dec_node_page_state(page, NR_ISOLATED_ANON +
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001883 page_is_file_cache(page));
1884 putback_lru_page(page);
1885 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001886 isolated = 0;
1887 } else
1888 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001889 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001890 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001891
1892out:
1893 put_page(page);
1894 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001895}
Mel Gorman220018d2012-12-05 09:32:56 +00001896#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001897
Mel Gorman220018d2012-12-05 09:32:56 +00001898#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001899/*
1900 * Migrates a THP to a given target node. page must be locked and is unlocked
1901 * before returning.
1902 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001903int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1904 struct vm_area_struct *vma,
1905 pmd_t *pmd, pmd_t entry,
1906 unsigned long address,
1907 struct page *page, int node)
1908{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001909 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001910 pg_data_t *pgdat = NODE_DATA(node);
1911 int isolated = 0;
1912 struct page *new_page = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001913 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001914 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1915 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001916 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001917
1918 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001919 * Rate-limit the amount of data that is being migrated to a node.
1920 * Optimal placement is no good if the memory bus is saturated and
1921 * all the time is being spent migrating!
1922 */
Mel Gormand28d43352012-11-29 09:24:36 +00001923 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001924 goto out_dropref;
1925
1926 new_page = alloc_pages_node(node,
Vlastimil Babka25160352016-07-28 15:49:25 -07001927 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001928 HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001929 if (!new_page)
1930 goto out_fail;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001931 prep_transhuge_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001932
Mel Gormanb32967f2012-11-19 12:35:47 +00001933 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001934 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001935 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001936 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001937 }
Aneesh Kumar K.V458aa762016-03-17 14:18:56 -07001938 /*
1939 * We are not sure a pending tlb flush here is for a huge page
1940 * mapping or not. Hence use the tlb range variant
1941 */
Mel Gormanb0943d62013-12-18 17:08:46 -08001942 if (mm_tlb_flush_pending(mm))
1943 flush_tlb_range(vma, mmun_start, mmun_end);
1944
Mel Gormanb32967f2012-11-19 12:35:47 +00001945 /* Prepare a page as a migration target */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001946 __SetPageLocked(new_page);
Shaohua Lid44d3632017-05-03 14:52:26 -07001947 if (PageSwapBacked(page))
1948 __SetPageSwapBacked(new_page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001949
1950 /* anon mapping, we can simply copy page->mapping to the new page: */
1951 new_page->mapping = page->mapping;
1952 new_page->index = page->index;
1953 migrate_page_copy(new_page, page);
1954 WARN_ON(PageLRU(new_page));
1955
1956 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001957 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001958 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001959 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1960fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001961 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001962 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001963
1964 /* Reverse changes made by migrate_page_copy() */
1965 if (TestClearPageActive(new_page))
1966 SetPageActive(page);
1967 if (TestClearPageUnevictable(new_page))
1968 SetPageUnevictable(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001969
1970 unlock_page(new_page);
1971 put_page(new_page); /* Free it */
1972
Mel Gormana54a4072013-10-07 11:28:46 +01001973 /* Retake the callers reference and putback on LRU */
1974 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001975 putback_lru_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001976 mod_node_page_state(page_pgdat(page),
Mel Gormana54a4072013-10-07 11:28:46 +01001977 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001978
1979 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001980 }
1981
Mel Gorman2b4847e2013-12-18 17:08:32 -08001982 orig_entry = *pmd;
Kirill A. Shutemov10102452016-07-26 15:25:29 -07001983 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001984 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001985
Mel Gorman2b4847e2013-12-18 17:08:32 -08001986 /*
1987 * Clear the old entry under pagetable lock and establish the new PTE.
1988 * Any parallel GUP will either observe the old page blocking on the
1989 * page lock, block on the page table lock or observe the new page.
1990 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1991 * guarantee the copy is visible before the pagetable update.
1992 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001993 flush_cache_range(vma, mmun_start, mmun_end);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001994 page_add_anon_rmap(new_page, vma, mmun_start, true);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001995 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001996 set_pmd_at(mm, mmun_start, pmd, entry);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001997 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001998
1999 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08002000 set_pmd_at(mm, mmun_start, pmd, orig_entry);
Aneesh Kumar K.V458aa762016-03-17 14:18:56 -07002001 flush_pmd_tlb_range(vma, mmun_start, mmun_end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11002002 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08002003 update_mmu_cache_pmd(vma, address, &entry);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002004 page_remove_rmap(new_page, true);
Mel Gorman2b4847e2013-12-18 17:08:32 -08002005 goto fail_putback;
2006 }
2007
Hugh Dickins51afb122015-11-05 18:49:37 -08002008 mlock_migrate_page(new_page, page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002009 page_remove_rmap(page, true);
Vlastimil Babka7cd12b42016-03-15 14:56:18 -07002010 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
Mel Gorman2b4847e2013-12-18 17:08:32 -08002011
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002012 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08002013 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00002014
Mel Gorman11de9922014-06-04 16:07:41 -07002015 /* Take an "isolate" reference and put new page on the LRU. */
2016 get_page(new_page);
2017 putback_lru_page(new_page);
2018
Mel Gormanb32967f2012-11-19 12:35:47 +00002019 unlock_page(new_page);
2020 unlock_page(page);
2021 put_page(page); /* Drop the rmap reference */
2022 put_page(page); /* Drop the LRU isolation reference */
2023
2024 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2025 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2026
Mel Gorman599d0c92016-07-28 15:45:31 -07002027 mod_node_page_state(page_pgdat(page),
Mel Gormanb32967f2012-11-19 12:35:47 +00002028 NR_ISOLATED_ANON + page_lru,
2029 -HPAGE_PMD_NR);
2030 return isolated;
2031
Hugh Dickins340ef392013-02-22 16:34:33 -08002032out_fail:
2033 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00002034out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08002035 ptl = pmd_lock(mm, pmd);
2036 if (pmd_same(*pmd, entry)) {
Mel Gorman4d942462015-02-12 14:58:28 -08002037 entry = pmd_modify(entry, vma->vm_page_prot);
Mel Gormanf714f4f2013-12-18 17:08:33 -08002038 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08002039 update_mmu_cache_pmd(vma, address, &entry);
2040 }
2041 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01002042
Mel Gormaneb4489f62013-12-18 17:08:39 -08002043out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08002044 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00002045 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00002046 return 0;
2047}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02002048#endif /* CONFIG_NUMA_BALANCING */
2049
2050#endif /* CONFIG_NUMA */