blob: dd43373a483fa764ce35a3bdc541ea4148653af4 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * mm/rmap.c - physical to virtual reverse mappings
3 *
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
6 *
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
9 *
10 * Provides methods for unmapping each kind of mapped page:
11 * the anon methods track anonymous pages, and
12 * the file methods track pages belonging to an inode.
13 *
14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
Hugh Dickins98f32602009-05-21 20:33:58 +010017 * Contributions by Hugh Dickins 2003, 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -070018 */
19
20/*
21 * Lock ordering in mm:
22 *
Jes Sorensen1b1dcc12006-01-09 15:59:24 -080023 * inode->i_mutex (while writing or truncating, not reading or faulting)
Nick Piggin82591e62006-10-19 23:29:10 -070024 * inode->i_alloc_sem (vmtruncate_range)
25 * mm->mmap_sem
26 * page->flags PG_locked (lock_page)
27 * mapping->i_mmap_lock
28 * anon_vma->lock
29 * mm->page_table_lock or pte_lock
30 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * inode_lock (in set_page_dirty's __mark_inode_dirty)
35 * sb_lock (within inode_lock in fs/fs-writeback.c)
36 * mapping->tree_lock (widely used, in set_page_dirty,
37 * in arch-dependent flush_dcache_mmap_lock,
38 * within inode_lock in __sync_single_inode)
Andi Kleen6a460792009-09-16 11:50:15 +020039 *
40 * (code doesn't rely on that order so it could be switched around)
41 * ->tasklist_lock
42 * anon_vma->lock (memory_failure, collect_procs_anon)
43 * pte map lock
Linus Torvalds1da177e2005-04-16 15:20:36 -070044 */
45
46#include <linux/mm.h>
47#include <linux/pagemap.h>
48#include <linux/swap.h>
49#include <linux/swapops.h>
50#include <linux/slab.h>
51#include <linux/init.h>
52#include <linux/rmap.h>
53#include <linux/rcupdate.h>
Christoph Lametera48d07a2006-02-01 03:05:38 -080054#include <linux/module.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080055#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070056#include <linux/mmu_notifier.h>
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -080057#include <linux/migrate.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070058
59#include <asm/tlbflush.h>
60
Nick Pigginb291f002008-10-18 20:26:44 -070061#include "internal.h"
62
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070063static struct kmem_cache *anon_vma_cachep;
64
65static inline struct anon_vma *anon_vma_alloc(void)
66{
67 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
68}
69
70static inline void anon_vma_free(struct anon_vma *anon_vma)
71{
72 kmem_cache_free(anon_vma_cachep, anon_vma);
73}
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
Linus Torvaldsd9d332e2008-10-19 10:32:20 -070075/**
76 * anon_vma_prepare - attach an anon_vma to a memory region
77 * @vma: the memory region in question
78 *
79 * This makes sure the memory mapping described by 'vma' has
80 * an 'anon_vma' attached to it, so that we can associate the
81 * anonymous pages mapped into it with that anon_vma.
82 *
83 * The common case will be that we already have one, but if
84 * if not we either need to find an adjacent mapping that we
85 * can re-use the anon_vma from (very common when the only
86 * reason for splitting a vma has been mprotect()), or we
87 * allocate a new one.
88 *
89 * Anon-vma allocations are very subtle, because we may have
90 * optimistically looked up an anon_vma in page_lock_anon_vma()
91 * and that may actually touch the spinlock even in the newly
92 * allocated vma (it depends on RCU to make sure that the
93 * anon_vma isn't actually destroyed).
94 *
95 * As a result, we need to do proper anon_vma locking even
96 * for the new allocation. At the same time, we do not want
97 * to do any locking for the common case of already having
98 * an anon_vma.
99 *
100 * This must be called with the mmap_sem held for reading.
101 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102int anon_vma_prepare(struct vm_area_struct *vma)
103{
104 struct anon_vma *anon_vma = vma->anon_vma;
105
106 might_sleep();
107 if (unlikely(!anon_vma)) {
108 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700109 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
111 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700112 allocated = NULL;
113 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114 anon_vma = anon_vma_alloc();
115 if (unlikely(!anon_vma))
116 return -ENOMEM;
117 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118 }
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700119 spin_lock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120
121 /* page_table_lock to protect against threads */
122 spin_lock(&mm->page_table_lock);
123 if (likely(!vma->anon_vma)) {
124 vma->anon_vma = anon_vma;
Christoph Lameter06972122006-06-23 02:03:35 -0700125 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 allocated = NULL;
127 }
128 spin_unlock(&mm->page_table_lock);
129
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700130 spin_unlock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 if (unlikely(allocated))
132 anon_vma_free(allocated);
133 }
134 return 0;
135}
136
137void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
138{
139 BUG_ON(vma->anon_vma != next->anon_vma);
140 list_del(&next->anon_vma_node);
141}
142
143void __anon_vma_link(struct vm_area_struct *vma)
144{
145 struct anon_vma *anon_vma = vma->anon_vma;
146
Hugh Dickins30acbab2007-06-27 14:09:53 -0700147 if (anon_vma)
Christoph Lameter06972122006-06-23 02:03:35 -0700148 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149}
150
151void anon_vma_link(struct vm_area_struct *vma)
152{
153 struct anon_vma *anon_vma = vma->anon_vma;
154
155 if (anon_vma) {
156 spin_lock(&anon_vma->lock);
Christoph Lameter06972122006-06-23 02:03:35 -0700157 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158 spin_unlock(&anon_vma->lock);
159 }
160}
161
162void anon_vma_unlink(struct vm_area_struct *vma)
163{
164 struct anon_vma *anon_vma = vma->anon_vma;
165 int empty;
166
167 if (!anon_vma)
168 return;
169
170 spin_lock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 list_del(&vma->anon_vma_node);
172
173 /* We must garbage collect the anon_vma if it's empty */
174 empty = list_empty(&anon_vma->head);
175 spin_unlock(&anon_vma->lock);
176
177 if (empty)
178 anon_vma_free(anon_vma);
179}
180
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700181static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182{
Christoph Lametera35afb82007-05-16 22:10:57 -0700183 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184
Christoph Lametera35afb82007-05-16 22:10:57 -0700185 spin_lock_init(&anon_vma->lock);
186 INIT_LIST_HEAD(&anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187}
188
189void __init anon_vma_init(void)
190{
191 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900192 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193}
194
195/*
196 * Getting a lock on a stable anon_vma from a page off the LRU is
197 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
198 */
Andi Kleen10be22d2009-09-16 11:50:04 +0200199struct anon_vma *page_lock_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200{
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800201 struct anon_vma *anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 unsigned long anon_mapping;
203
204 rcu_read_lock();
205 anon_mapping = (unsigned long) page->mapping;
206 if (!(anon_mapping & PAGE_MAPPING_ANON))
207 goto out;
208 if (!page_mapped(page))
209 goto out;
210
211 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
212 spin_lock(&anon_vma->lock);
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800213 return anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214out:
215 rcu_read_unlock();
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800216 return NULL;
217}
218
Andi Kleen10be22d2009-09-16 11:50:04 +0200219void page_unlock_anon_vma(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800220{
221 spin_unlock(&anon_vma->lock);
222 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223}
224
225/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800226 * At what user virtual address is page expected in @vma?
227 * Returns virtual address or -EFAULT if page's index/offset is not
228 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 */
230static inline unsigned long
231vma_address(struct page *page, struct vm_area_struct *vma)
232{
233 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
234 unsigned long address;
235
236 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
237 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800238 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239 return -EFAULT;
240 }
241 return address;
242}
243
244/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700245 * At what user virtual address is page expected in vma?
246 * checking that the page matches the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 */
248unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
249{
250 if (PageAnon(page)) {
251 if ((void *)vma->anon_vma !=
252 (void *)page->mapping - PAGE_MAPPING_ANON)
253 return -EFAULT;
254 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800255 if (!vma->vm_file ||
256 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 return -EFAULT;
258 } else
259 return -EFAULT;
260 return vma_address(page, vma);
261}
262
263/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700264 * Check that @page is mapped at @address into @mm.
265 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700266 * If @sync is false, page_check_address may perform a racy check to avoid
267 * the page table lock when the pte is not present (helpful when reclaiming
268 * highly shared pages).
269 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700270 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700271 */
Carsten Otteceffc072005-06-23 22:05:25 -0700272pte_t *page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700273 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700274{
275 pgd_t *pgd;
276 pud_t *pud;
277 pmd_t *pmd;
278 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700279 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700280
Nikita Danilov81b40822005-05-01 08:58:36 -0700281 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700282 if (!pgd_present(*pgd))
283 return NULL;
284
285 pud = pud_offset(pgd, address);
286 if (!pud_present(*pud))
287 return NULL;
288
289 pmd = pmd_offset(pud, address);
290 if (!pmd_present(*pmd))
291 return NULL;
292
293 pte = pte_offset_map(pmd, address);
294 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700295 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700296 pte_unmap(pte);
297 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700298 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700299
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700300 ptl = pte_lockptr(mm, pmd);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700301 spin_lock(ptl);
302 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
303 *ptlp = ptl;
304 return pte;
305 }
306 pte_unmap_unlock(pte, ptl);
307 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700308}
309
Nick Pigginb291f002008-10-18 20:26:44 -0700310/**
311 * page_mapped_in_vma - check whether a page is really mapped in a VMA
312 * @page: the page to test
313 * @vma: the VMA to test
314 *
315 * Returns 1 if the page is mapped into the page tables of the VMA, 0
316 * if the page is not mapped into the page tables of this VMA. Only
317 * valid for normal file or anonymous VMAs.
318 */
Andi Kleen6a460792009-09-16 11:50:15 +0200319int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700320{
321 unsigned long address;
322 pte_t *pte;
323 spinlock_t *ptl;
324
325 address = vma_address(page, vma);
326 if (address == -EFAULT) /* out of vma range */
327 return 0;
328 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
329 if (!pte) /* the page is not in this mm */
330 return 0;
331 pte_unmap_unlock(pte, ptl);
332
333 return 1;
334}
335
Nikita Danilov81b40822005-05-01 08:58:36 -0700336/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 * Subfunctions of page_referenced: page_referenced_one called
338 * repeatedly from either page_referenced_anon or page_referenced_file.
339 */
340static int page_referenced_one(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700341 struct vm_area_struct *vma,
342 unsigned int *mapcount,
343 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344{
345 struct mm_struct *mm = vma->vm_mm;
346 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700348 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 int referenced = 0;
350
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351 address = vma_address(page, vma);
352 if (address == -EFAULT)
353 goto out;
354
Nick Piggin479db0b2008-08-20 14:09:18 -0700355 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700356 if (!pte)
357 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
Nick Pigginb291f002008-10-18 20:26:44 -0700359 /*
360 * Don't want to elevate referenced for mlocked page that gets this far,
361 * in order that it progresses to try_to_unmap and is moved to the
362 * unevictable list.
363 */
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800364 if (vma->vm_flags & VM_LOCKED) {
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800365 *mapcount = 1; /* break early from loop */
Minchan Kim03ef83a2009-08-26 14:29:23 -0700366 *vm_flags |= VM_LOCKED;
Nick Pigginb291f002008-10-18 20:26:44 -0700367 goto out_unmap;
368 }
369
Johannes Weiner4917e5d2009-01-06 14:39:17 -0800370 if (ptep_clear_flush_young_notify(vma, address, pte)) {
371 /*
372 * Don't treat a reference through a sequentially read
373 * mapping as such. If the page has been used in
374 * another mapping, we will catch it; if this other
375 * mapping is already gone, the unmap path will have
376 * set PG_referenced or activated the page.
377 */
378 if (likely(!VM_SequentialReadHint(vma)))
379 referenced++;
380 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381
Hugh Dickinsc0718802005-10-29 18:16:31 -0700382 /* Pretend the page is referenced if the task has the
383 swap token and is in the middle of a page fault. */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800384 if (mm != current->mm && has_swap_token(mm) &&
Hugh Dickinsc0718802005-10-29 18:16:31 -0700385 rwsem_is_locked(&mm->mmap_sem))
386 referenced++;
387
Nick Pigginb291f002008-10-18 20:26:44 -0700388out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700389 (*mapcount)--;
390 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391out:
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700392 if (referenced)
393 *vm_flags |= vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 return referenced;
395}
396
Balbir Singhbed71612008-02-07 00:14:01 -0800397static int page_referenced_anon(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700398 struct mem_cgroup *mem_cont,
399 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400{
401 unsigned int mapcount;
402 struct anon_vma *anon_vma;
403 struct vm_area_struct *vma;
404 int referenced = 0;
405
406 anon_vma = page_lock_anon_vma(page);
407 if (!anon_vma)
408 return referenced;
409
410 mapcount = page_mapcount(page);
411 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Balbir Singhbed71612008-02-07 00:14:01 -0800412 /*
413 * If we are reclaiming on behalf of a cgroup, skip
414 * counting on behalf of references from different
415 * cgroups
416 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800417 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800418 continue;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700419 referenced += page_referenced_one(page, vma,
420 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 if (!mapcount)
422 break;
423 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800424
425 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426 return referenced;
427}
428
429/**
430 * page_referenced_file - referenced check for object-based rmap
431 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700432 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700433 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 *
435 * For an object-based mapped page, find all the places it is mapped and
436 * check/clear the referenced flag. This is done by following the page->mapping
437 * pointer, then walking the chain of vmas it holds. It returns the number
438 * of references it found.
439 *
440 * This function is only called from page_referenced for object-based pages.
441 */
Balbir Singhbed71612008-02-07 00:14:01 -0800442static int page_referenced_file(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700443 struct mem_cgroup *mem_cont,
444 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445{
446 unsigned int mapcount;
447 struct address_space *mapping = page->mapping;
448 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
449 struct vm_area_struct *vma;
450 struct prio_tree_iter iter;
451 int referenced = 0;
452
453 /*
454 * The caller's checks on page->mapping and !PageAnon have made
455 * sure that this is a file page: the check for page->mapping
456 * excludes the case just before it gets set on an anon page.
457 */
458 BUG_ON(PageAnon(page));
459
460 /*
461 * The page lock not only makes sure that page->mapping cannot
462 * suddenly be NULLified by truncation, it makes sure that the
463 * structure at mapping cannot be freed and reused yet,
464 * so we can safely take mapping->i_mmap_lock.
465 */
466 BUG_ON(!PageLocked(page));
467
468 spin_lock(&mapping->i_mmap_lock);
469
470 /*
471 * i_mmap_lock does not stabilize mapcount at all, but mapcount
472 * is more likely to be accurate if we note it after spinning.
473 */
474 mapcount = page_mapcount(page);
475
476 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Balbir Singhbed71612008-02-07 00:14:01 -0800477 /*
478 * If we are reclaiming on behalf of a cgroup, skip
479 * counting on behalf of references from different
480 * cgroups
481 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800482 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800483 continue;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700484 referenced += page_referenced_one(page, vma,
485 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 if (!mapcount)
487 break;
488 }
489
490 spin_unlock(&mapping->i_mmap_lock);
491 return referenced;
492}
493
494/**
495 * page_referenced - test if the page was referenced
496 * @page: the page to test
497 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700498 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700499 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 *
501 * Quick test_and_clear_referenced for all mappings to a page,
502 * returns the number of ptes which referenced the page.
503 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700504int page_referenced(struct page *page,
505 int is_locked,
506 struct mem_cgroup *mem_cont,
507 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508{
509 int referenced = 0;
510
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 if (TestClearPageReferenced(page))
512 referenced++;
513
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700514 *vm_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 if (page_mapped(page) && page->mapping) {
516 if (PageAnon(page))
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700517 referenced += page_referenced_anon(page, mem_cont,
518 vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 else if (is_locked)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700520 referenced += page_referenced_file(page, mem_cont,
521 vm_flags);
Nick Piggin529ae9a2008-08-02 12:01:03 +0200522 else if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 referenced++;
524 else {
525 if (page->mapping)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700526 referenced += page_referenced_file(page,
527 mem_cont, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 unlock_page(page);
529 }
530 }
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100531
532 if (page_test_and_clear_young(page))
533 referenced++;
534
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 return referenced;
536}
537
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700538static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
539{
540 struct mm_struct *mm = vma->vm_mm;
541 unsigned long address;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100542 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700543 spinlock_t *ptl;
544 int ret = 0;
545
546 address = vma_address(page, vma);
547 if (address == -EFAULT)
548 goto out;
549
Nick Piggin479db0b2008-08-20 14:09:18 -0700550 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700551 if (!pte)
552 goto out;
553
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100554 if (pte_dirty(*pte) || pte_write(*pte)) {
555 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700556
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100557 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700558 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100559 entry = pte_wrprotect(entry);
560 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800561 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100562 ret = 1;
563 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700564
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700565 pte_unmap_unlock(pte, ptl);
566out:
567 return ret;
568}
569
570static int page_mkclean_file(struct address_space *mapping, struct page *page)
571{
572 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
573 struct vm_area_struct *vma;
574 struct prio_tree_iter iter;
575 int ret = 0;
576
577 BUG_ON(PageAnon(page));
578
579 spin_lock(&mapping->i_mmap_lock);
580 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
581 if (vma->vm_flags & VM_SHARED)
582 ret += page_mkclean_one(page, vma);
583 }
584 spin_unlock(&mapping->i_mmap_lock);
585 return ret;
586}
587
588int page_mkclean(struct page *page)
589{
590 int ret = 0;
591
592 BUG_ON(!PageLocked(page));
593
594 if (page_mapped(page)) {
595 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100596 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700597 ret = page_mkclean_file(mapping, page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100598 if (page_test_dirty(page)) {
599 page_clear_dirty(page);
600 ret = 1;
601 }
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200602 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700603 }
604
605 return ret;
606}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700607EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700608
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700610 * __page_set_anon_rmap - setup new anonymous rmap
Nick Piggin9617d952006-01-06 00:11:12 -0800611 * @page: the page to add the mapping to
612 * @vma: the vm area in which the mapping is added
613 * @address: the user virtual address mapped
614 */
615static void __page_set_anon_rmap(struct page *page,
616 struct vm_area_struct *vma, unsigned long address)
617{
618 struct anon_vma *anon_vma = vma->anon_vma;
619
620 BUG_ON(!anon_vma);
621 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
622 page->mapping = (struct address_space *) anon_vma;
623
624 page->index = linear_page_index(vma, address);
625
Nick Piggina74609f2006-01-06 00:11:20 -0800626 /*
627 * nr_mapped state can be updated without turning off
628 * interrupts because it is not modified via interrupt.
629 */
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700630 __inc_zone_page_state(page, NR_ANON_PAGES);
Nick Piggin9617d952006-01-06 00:11:12 -0800631}
632
633/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700634 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700635 * @page: the page to add the mapping to
636 * @vma: the vm area in which the mapping is added
637 * @address: the user virtual address mapped
638 */
639static void __page_check_anon_rmap(struct page *page,
640 struct vm_area_struct *vma, unsigned long address)
641{
642#ifdef CONFIG_DEBUG_VM
643 /*
644 * The page's anon-rmap details (mapping and index) are guaranteed to
645 * be set up correctly at this point.
646 *
647 * We have exclusion against page_add_anon_rmap because the caller
648 * always holds the page locked, except if called from page_dup_rmap,
649 * in which case the page is already known to be setup.
650 *
651 * We have exclusion against page_add_new_anon_rmap because those pages
652 * are initially only visible via the pagetables, and the pte is locked
653 * over the call to page_add_new_anon_rmap.
654 */
655 struct anon_vma *anon_vma = vma->anon_vma;
656 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
657 BUG_ON(page->mapping != (struct address_space *)anon_vma);
658 BUG_ON(page->index != linear_page_index(vma, address));
659#endif
660}
661
662/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 * page_add_anon_rmap - add pte mapping to an anonymous page
664 * @page: the page to add the mapping to
665 * @vma: the vm area in which the mapping is added
666 * @address: the user virtual address mapped
667 *
Nick Pigginc97a9e12007-05-16 22:11:21 -0700668 * The caller needs to hold the pte lock and the page must be locked.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 */
670void page_add_anon_rmap(struct page *page,
671 struct vm_area_struct *vma, unsigned long address)
672{
Nick Pigginc97a9e12007-05-16 22:11:21 -0700673 VM_BUG_ON(!PageLocked(page));
674 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Nick Piggin9617d952006-01-06 00:11:12 -0800675 if (atomic_inc_and_test(&page->_mapcount))
676 __page_set_anon_rmap(page, vma, address);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -0700677 else
Nick Pigginc97a9e12007-05-16 22:11:21 -0700678 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679}
680
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700681/**
Nick Piggin9617d952006-01-06 00:11:12 -0800682 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
683 * @page: the page to add the mapping to
684 * @vma: the vm area in which the mapping is added
685 * @address: the user virtual address mapped
686 *
687 * Same as page_add_anon_rmap but must only be called on *new* pages.
688 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -0700689 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -0800690 */
691void page_add_new_anon_rmap(struct page *page,
692 struct vm_area_struct *vma, unsigned long address)
693{
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800694 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800695 SetPageSwapBacked(page);
696 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Nick Piggin9617d952006-01-06 00:11:12 -0800697 __page_set_anon_rmap(page, vma, address);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800698 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800699 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800700 else
701 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -0800702}
703
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704/**
705 * page_add_file_rmap - add pte mapping to a file page
706 * @page: the page to add the mapping to
707 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700708 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709 */
710void page_add_file_rmap(struct page *page)
711{
Balbir Singhd69b0422009-06-17 16:26:34 -0700712 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700713 __inc_zone_page_state(page, NR_FILE_MAPPED);
Balbir Singhd69b0422009-06-17 16:26:34 -0700714 mem_cgroup_update_mapped_file_stat(page, 1);
715 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716}
717
718/**
719 * page_remove_rmap - take down pte mapping from a page
720 * @page: page to remove mapping from
721 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700722 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800724void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725{
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700726 /* page still mapped by someone else? */
727 if (!atomic_add_negative(-1, &page->_mapcount))
728 return;
729
730 /*
731 * Now that the last pte has gone, s390 must transfer dirty
732 * flag from storage key to struct page. We can usually skip
733 * this if the page is anon, so about to be freed; but perhaps
734 * not if it's in swapcache - there might be another pte slot
735 * containing the swap entry, but page not yet written to swap.
736 */
737 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
738 page_clear_dirty(page);
739 set_page_dirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700741 if (PageAnon(page)) {
742 mem_cgroup_uncharge_page(page);
743 __dec_zone_page_state(page, NR_ANON_PAGES);
744 } else {
745 __dec_zone_page_state(page, NR_FILE_MAPPED);
746 }
747 mem_cgroup_update_mapped_file_stat(page, -1);
748 /*
749 * It would be tidy to reset the PageAnon mapping here,
750 * but that might overwrite a racing page_add_anon_rmap
751 * which increments mapcount after us but sets mapping
752 * before us: so leave the reset to free_hot_cold_page,
753 * and remember that it's only reliable while mapped.
754 * Leaving it set also helps swapoff to reinstate ptes
755 * faster for those pages still in swapcache.
756 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757}
758
759/*
760 * Subfunctions of try_to_unmap: try_to_unmap_one called
761 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
762 */
Christoph Lametera48d07a2006-02-01 03:05:38 -0800763static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
Andi Kleen14fa31b2009-09-16 11:50:10 +0200764 enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765{
766 struct mm_struct *mm = vma->vm_mm;
767 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 pte_t *pte;
769 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700770 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771 int ret = SWAP_AGAIN;
772
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 address = vma_address(page, vma);
774 if (address == -EFAULT)
775 goto out;
776
Nick Piggin479db0b2008-08-20 14:09:18 -0700777 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700778 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700779 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780
781 /*
782 * If the page is mlock()d, we cannot swap it out.
783 * If it's recently referenced (perhaps page_referenced
784 * skipped over this mm) then we should reactivate it.
785 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200786 if (!(flags & TTU_IGNORE_MLOCK)) {
Nick Pigginb291f002008-10-18 20:26:44 -0700787 if (vma->vm_flags & VM_LOCKED) {
788 ret = SWAP_MLOCK;
789 goto out_unmap;
790 }
Andi Kleen14fa31b2009-09-16 11:50:10 +0200791 }
792 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -0700793 if (ptep_clear_flush_young_notify(vma, address, pte)) {
794 ret = SWAP_FAIL;
795 goto out_unmap;
796 }
797 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 /* Nuke the page table entry. */
800 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700801 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802
803 /* Move the dirty bit to the physical page now the pte is gone. */
804 if (pte_dirty(pteval))
805 set_page_dirty(page);
806
Hugh Dickins365e9c872005-10-29 18:16:18 -0700807 /* Update high watermark before we lower rss */
808 update_hiwater_rss(mm);
809
Andi Kleen888b9f72009-09-16 11:50:11 +0200810 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
811 if (PageAnon(page))
812 dec_mm_counter(mm, anon_rss);
813 else
814 dec_mm_counter(mm, file_rss);
815 set_pte_at(mm, address, pte,
816 swp_entry_to_pte(make_hwpoison_entry(page)));
817 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700818 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -0700819
820 if (PageSwapCache(page)) {
821 /*
822 * Store the swap location in the pte.
823 * See handle_pte_fault() ...
824 */
825 swap_duplicate(entry);
826 if (list_empty(&mm->mmlist)) {
827 spin_lock(&mmlist_lock);
828 if (list_empty(&mm->mmlist))
829 list_add(&mm->mmlist, &init_mm.mmlist);
830 spin_unlock(&mmlist_lock);
831 }
Christoph Lameter442c9132006-06-23 02:03:38 -0700832 dec_mm_counter(mm, anon_rss);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -0800833 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -0700834 /*
835 * Store the pfn of the page in a special migration
836 * pte. do_swap_page() will wait until the migration
837 * pte is removed and then restart fault handling.
838 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200839 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -0700840 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 }
842 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
843 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +0200844 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -0700845 /* Establish migration entry for a file page */
846 swp_entry_t entry;
847 entry = make_migration_entry(page, pte_write(pteval));
848 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
849 } else
Hugh Dickins42946212005-10-29 18:16:05 -0700850 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851
Christoph Lameter04e62a22006-06-23 02:03:38 -0700852
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800853 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 page_cache_release(page);
855
856out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700857 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858out:
859 return ret;
860}
861
862/*
863 * objrmap doesn't work for nonlinear VMAs because the assumption that
864 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
865 * Consequently, given a particular page and its ->index, we cannot locate the
866 * ptes which are mapping that page without an exhaustive linear search.
867 *
868 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
869 * maps the file to which the target page belongs. The ->vm_private_data field
870 * holds the current cursor into that scan. Successive searches will circulate
871 * around the vma's virtual address space.
872 *
873 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
874 * more scanning pressure is placed against them as well. Eventually pages
875 * will become fully unmapped and are eligible for eviction.
876 *
877 * For very sparsely populated VMAs this is a little inefficient - chances are
878 * there there won't be many ptes located within the scan cluster. In this case
879 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -0700880 *
881 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
882 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
883 * rather than unmapping them. If we encounter the "check_page" that vmscan is
884 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 */
886#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
887#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
888
Nick Pigginb291f002008-10-18 20:26:44 -0700889static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
890 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891{
892 struct mm_struct *mm = vma->vm_mm;
893 pgd_t *pgd;
894 pud_t *pud;
895 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700896 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700898 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 struct page *page;
900 unsigned long address;
901 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -0700902 int ret = SWAP_AGAIN;
903 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 address = (vma->vm_start + cursor) & CLUSTER_MASK;
906 end = address + CLUSTER_SIZE;
907 if (address < vma->vm_start)
908 address = vma->vm_start;
909 if (end > vma->vm_end)
910 end = vma->vm_end;
911
912 pgd = pgd_offset(mm, address);
913 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -0700914 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915
916 pud = pud_offset(pgd, address);
917 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -0700918 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919
920 pmd = pmd_offset(pud, address);
921 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -0700922 return ret;
923
924 /*
925 * MLOCK_PAGES => feature is configured.
926 * if we can acquire the mmap_sem for read, and vma is VM_LOCKED,
927 * keep the sem while scanning the cluster for mlocking pages.
928 */
929 if (MLOCK_PAGES && down_read_trylock(&vma->vm_mm->mmap_sem)) {
930 locked_vma = (vma->vm_flags & VM_LOCKED);
931 if (!locked_vma)
932 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
933 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700934
935 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936
Hugh Dickins365e9c872005-10-29 18:16:18 -0700937 /* Update high watermark before we lower rss */
938 update_hiwater_rss(mm);
939
Hugh Dickinsc0718802005-10-29 18:16:31 -0700940 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 if (!pte_present(*pte))
942 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800943 page = vm_normal_page(vma, address, *pte);
944 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945
Nick Pigginb291f002008-10-18 20:26:44 -0700946 if (locked_vma) {
947 mlock_vma_page(page); /* no-op if already mlocked */
948 if (page == check_page)
949 ret = SWAP_MLOCK;
950 continue; /* don't unmap */
951 }
952
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700953 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 continue;
955
956 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -0800957 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700958 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959
960 /* If nonlinear, store the file page offset in the pte. */
961 if (page->index != linear_page_index(vma, address))
962 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
963
964 /* Move the dirty bit to the physical page now the pte is gone. */
965 if (pte_dirty(pteval))
966 set_page_dirty(page);
967
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800968 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 page_cache_release(page);
Hugh Dickins42946212005-10-29 18:16:05 -0700970 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 (*mapcount)--;
972 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700973 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -0700974 if (locked_vma)
975 up_read(&vma->vm_mm->mmap_sem);
976 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977}
978
Nick Pigginb291f002008-10-18 20:26:44 -0700979/*
980 * common handling for pages mapped in VM_LOCKED vmas
981 */
982static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma)
983{
984 int mlocked = 0;
985
986 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
987 if (vma->vm_flags & VM_LOCKED) {
988 mlock_vma_page(page);
989 mlocked++; /* really mlocked the page */
990 }
991 up_read(&vma->vm_mm->mmap_sem);
992 }
993 return mlocked;
994}
995
996/**
997 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
998 * rmap method
999 * @page: the page to unmap/unlock
1000 * @unlock: request for unlock rather than unmap [unlikely]
1001 * @migration: unmapping for migration - ignored if @unlock
1002 *
1003 * Find all the mappings of a page using the mapping pointer and the vma chains
1004 * contained in the anon_vma struct it points to.
1005 *
1006 * This function is only called from try_to_unmap/try_to_munlock for
1007 * anonymous pages.
1008 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1009 * where the page was found will be held for write. So, we won't recheck
1010 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1011 * 'LOCKED.
1012 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001013static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014{
1015 struct anon_vma *anon_vma;
1016 struct vm_area_struct *vma;
Nick Pigginb291f002008-10-18 20:26:44 -07001017 unsigned int mlocked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018 int ret = SWAP_AGAIN;
Andi Kleen14fa31b2009-09-16 11:50:10 +02001019 int unlock = TTU_ACTION(flags) == TTU_MUNLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020
Nick Pigginb291f002008-10-18 20:26:44 -07001021 if (MLOCK_PAGES && unlikely(unlock))
1022 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
1023
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024 anon_vma = page_lock_anon_vma(page);
1025 if (!anon_vma)
1026 return ret;
1027
1028 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Nick Pigginb291f002008-10-18 20:26:44 -07001029 if (MLOCK_PAGES && unlikely(unlock)) {
1030 if (!((vma->vm_flags & VM_LOCKED) &&
1031 page_mapped_in_vma(page, vma)))
1032 continue; /* must visit all unlocked vmas */
1033 ret = SWAP_MLOCK; /* saw at least one mlocked vma */
1034 } else {
Andi Kleen14fa31b2009-09-16 11:50:10 +02001035 ret = try_to_unmap_one(page, vma, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001036 if (ret == SWAP_FAIL || !page_mapped(page))
1037 break;
1038 }
1039 if (ret == SWAP_MLOCK) {
1040 mlocked = try_to_mlock_page(page, vma);
1041 if (mlocked)
1042 break; /* stop if actually mlocked page */
1043 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001045
1046 page_unlock_anon_vma(anon_vma);
Nick Pigginb291f002008-10-18 20:26:44 -07001047
1048 if (mlocked)
1049 ret = SWAP_MLOCK; /* actually mlocked the page */
1050 else if (ret == SWAP_MLOCK)
1051 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
1052
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053 return ret;
1054}
1055
1056/**
Nick Pigginb291f002008-10-18 20:26:44 -07001057 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1058 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001059 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060 *
1061 * Find all the mappings of a page using the mapping pointer and the vma chains
1062 * contained in the address_space struct it points to.
1063 *
Nick Pigginb291f002008-10-18 20:26:44 -07001064 * This function is only called from try_to_unmap/try_to_munlock for
1065 * object-based pages.
1066 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1067 * where the page was found will be held for write. So, we won't recheck
1068 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1069 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001071static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072{
1073 struct address_space *mapping = page->mapping;
1074 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1075 struct vm_area_struct *vma;
1076 struct prio_tree_iter iter;
1077 int ret = SWAP_AGAIN;
1078 unsigned long cursor;
1079 unsigned long max_nl_cursor = 0;
1080 unsigned long max_nl_size = 0;
1081 unsigned int mapcount;
Nick Pigginb291f002008-10-18 20:26:44 -07001082 unsigned int mlocked = 0;
Andi Kleen14fa31b2009-09-16 11:50:10 +02001083 int unlock = TTU_ACTION(flags) == TTU_MUNLOCK;
Nick Pigginb291f002008-10-18 20:26:44 -07001084
1085 if (MLOCK_PAGES && unlikely(unlock))
1086 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087
1088 spin_lock(&mapping->i_mmap_lock);
1089 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Nick Pigginb291f002008-10-18 20:26:44 -07001090 if (MLOCK_PAGES && unlikely(unlock)) {
MinChan Kim508b9f82009-02-11 13:04:27 -08001091 if (!((vma->vm_flags & VM_LOCKED) &&
1092 page_mapped_in_vma(page, vma)))
Nick Pigginb291f002008-10-18 20:26:44 -07001093 continue; /* must visit all vmas */
1094 ret = SWAP_MLOCK;
1095 } else {
Andi Kleen14fa31b2009-09-16 11:50:10 +02001096 ret = try_to_unmap_one(page, vma, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001097 if (ret == SWAP_FAIL || !page_mapped(page))
1098 goto out;
1099 }
1100 if (ret == SWAP_MLOCK) {
1101 mlocked = try_to_mlock_page(page, vma);
1102 if (mlocked)
1103 break; /* stop if actually mlocked page */
1104 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105 }
1106
Nick Pigginb291f002008-10-18 20:26:44 -07001107 if (mlocked)
1108 goto out;
1109
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 if (list_empty(&mapping->i_mmap_nonlinear))
1111 goto out;
1112
1113 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1114 shared.vm_set.list) {
Nick Pigginb291f002008-10-18 20:26:44 -07001115 if (MLOCK_PAGES && unlikely(unlock)) {
1116 if (!(vma->vm_flags & VM_LOCKED))
1117 continue; /* must visit all vmas */
1118 ret = SWAP_MLOCK; /* leave mlocked == 0 */
1119 goto out; /* no need to look further */
1120 }
Andi Kleen14fa31b2009-09-16 11:50:10 +02001121 if (!MLOCK_PAGES && !(flags & TTU_IGNORE_MLOCK) &&
1122 (vma->vm_flags & VM_LOCKED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 continue;
1124 cursor = (unsigned long) vma->vm_private_data;
1125 if (cursor > max_nl_cursor)
1126 max_nl_cursor = cursor;
1127 cursor = vma->vm_end - vma->vm_start;
1128 if (cursor > max_nl_size)
1129 max_nl_size = cursor;
1130 }
1131
Nick Pigginb291f002008-10-18 20:26:44 -07001132 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 ret = SWAP_FAIL;
1134 goto out;
1135 }
1136
1137 /*
1138 * We don't try to search for this page in the nonlinear vmas,
1139 * and page_referenced wouldn't have found it anyway. Instead
1140 * just walk the nonlinear vmas trying to age and unmap some.
1141 * The mapcount of the page we came in with is irrelevant,
1142 * but even so use it as a guide to how hard we should try?
1143 */
1144 mapcount = page_mapcount(page);
1145 if (!mapcount)
1146 goto out;
1147 cond_resched_lock(&mapping->i_mmap_lock);
1148
1149 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1150 if (max_nl_cursor == 0)
1151 max_nl_cursor = CLUSTER_SIZE;
1152
1153 do {
1154 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1155 shared.vm_set.list) {
Andi Kleen14fa31b2009-09-16 11:50:10 +02001156 if (!MLOCK_PAGES && !(flags & TTU_IGNORE_MLOCK) &&
Nick Pigginb291f002008-10-18 20:26:44 -07001157 (vma->vm_flags & VM_LOCKED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 continue;
1159 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001160 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 cursor < vma->vm_end - vma->vm_start) {
Nick Pigginb291f002008-10-18 20:26:44 -07001162 ret = try_to_unmap_cluster(cursor, &mapcount,
1163 vma, page);
1164 if (ret == SWAP_MLOCK)
1165 mlocked = 2; /* to return below */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 cursor += CLUSTER_SIZE;
1167 vma->vm_private_data = (void *) cursor;
1168 if ((int)mapcount <= 0)
1169 goto out;
1170 }
1171 vma->vm_private_data = (void *) max_nl_cursor;
1172 }
1173 cond_resched_lock(&mapping->i_mmap_lock);
1174 max_nl_cursor += CLUSTER_SIZE;
1175 } while (max_nl_cursor <= max_nl_size);
1176
1177 /*
1178 * Don't loop forever (perhaps all the remaining pages are
1179 * in locked vmas). Reset cursor on all unreserved nonlinear
1180 * vmas, now forgetting on which ones it had fallen behind.
1181 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001182 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1183 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184out:
1185 spin_unlock(&mapping->i_mmap_lock);
Nick Pigginb291f002008-10-18 20:26:44 -07001186 if (mlocked)
1187 ret = SWAP_MLOCK; /* actually mlocked the page */
1188 else if (ret == SWAP_MLOCK)
1189 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190 return ret;
1191}
1192
1193/**
1194 * try_to_unmap - try to remove all page table mappings to a page
1195 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001196 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 *
1198 * Tries to remove all the page table entries which are mapping this
1199 * page, used in the pageout path. Caller must hold the page lock.
1200 * Return values are:
1201 *
1202 * SWAP_SUCCESS - we succeeded in removing all mappings
1203 * SWAP_AGAIN - we missed a mapping, try again later
1204 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001205 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001207int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208{
1209 int ret;
1210
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 BUG_ON(!PageLocked(page));
1212
1213 if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001214 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001216 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001217 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 ret = SWAP_SUCCESS;
1219 return ret;
1220}
Nikita Danilov81b40822005-05-01 08:58:36 -07001221
Nick Pigginb291f002008-10-18 20:26:44 -07001222/**
1223 * try_to_munlock - try to munlock a page
1224 * @page: the page to be munlocked
1225 *
1226 * Called from munlock code. Checks all of the VMAs mapping the page
1227 * to make sure nobody else has this page mlocked. The page will be
1228 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1229 *
1230 * Return values are:
1231 *
1232 * SWAP_SUCCESS - no vma's holding page mlocked.
1233 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1234 * SWAP_MLOCK - page is now mlocked.
1235 */
1236int try_to_munlock(struct page *page)
1237{
1238 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1239
1240 if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001241 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001242 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001243 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001244}
KOSAKI Motohiro68377652009-06-16 15:32:51 -07001245