Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
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 | |
98f32602 | 17 | * Contributions by Hugh Dickins 2003, 2004 |
1da177e4 LT |
18 | */ |
19 | ||
20 | /* | |
21 | * Lock ordering in mm: | |
22 | * | |
1b1dcc1b | 23 | * inode->i_mutex (while writing or truncating, not reading or faulting) |
82591e6e NP |
24 | * 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) | |
6a46079c AK |
39 | * |
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 | |
1da177e4 LT |
44 | */ |
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> | |
5ad64688 | 52 | #include <linux/ksm.h> |
1da177e4 LT |
53 | #include <linux/rmap.h> |
54 | #include <linux/rcupdate.h> | |
a48d07af | 55 | #include <linux/module.h> |
8a9f3ccd | 56 | #include <linux/memcontrol.h> |
cddb8a5c | 57 | #include <linux/mmu_notifier.h> |
64cdd548 | 58 | #include <linux/migrate.h> |
1da177e4 LT |
59 | |
60 | #include <asm/tlbflush.h> | |
61 | ||
b291f000 NP |
62 | #include "internal.h" |
63 | ||
fdd2e5f8 | 64 | static struct kmem_cache *anon_vma_cachep; |
5beb4930 | 65 | static struct kmem_cache *anon_vma_chain_cachep; |
fdd2e5f8 AB |
66 | |
67 | static inline struct anon_vma *anon_vma_alloc(void) | |
68 | { | |
69 | return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); | |
70 | } | |
71 | ||
db114b83 | 72 | void anon_vma_free(struct anon_vma *anon_vma) |
fdd2e5f8 AB |
73 | { |
74 | kmem_cache_free(anon_vma_cachep, anon_vma); | |
75 | } | |
1da177e4 | 76 | |
5beb4930 RR |
77 | static inline struct anon_vma_chain *anon_vma_chain_alloc(void) |
78 | { | |
79 | return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL); | |
80 | } | |
81 | ||
82 | void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) | |
83 | { | |
84 | kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); | |
85 | } | |
86 | ||
d9d332e0 LT |
87 | /** |
88 | * anon_vma_prepare - attach an anon_vma to a memory region | |
89 | * @vma: the memory region in question | |
90 | * | |
91 | * This makes sure the memory mapping described by 'vma' has | |
92 | * an 'anon_vma' attached to it, so that we can associate the | |
93 | * anonymous pages mapped into it with that anon_vma. | |
94 | * | |
95 | * The common case will be that we already have one, but if | |
96 | * if not we either need to find an adjacent mapping that we | |
97 | * can re-use the anon_vma from (very common when the only | |
98 | * reason for splitting a vma has been mprotect()), or we | |
99 | * allocate a new one. | |
100 | * | |
101 | * Anon-vma allocations are very subtle, because we may have | |
102 | * optimistically looked up an anon_vma in page_lock_anon_vma() | |
103 | * and that may actually touch the spinlock even in the newly | |
104 | * allocated vma (it depends on RCU to make sure that the | |
105 | * anon_vma isn't actually destroyed). | |
106 | * | |
107 | * As a result, we need to do proper anon_vma locking even | |
108 | * for the new allocation. At the same time, we do not want | |
109 | * to do any locking for the common case of already having | |
110 | * an anon_vma. | |
111 | * | |
112 | * This must be called with the mmap_sem held for reading. | |
113 | */ | |
1da177e4 LT |
114 | int anon_vma_prepare(struct vm_area_struct *vma) |
115 | { | |
116 | struct anon_vma *anon_vma = vma->anon_vma; | |
5beb4930 | 117 | struct anon_vma_chain *avc; |
1da177e4 LT |
118 | |
119 | might_sleep(); | |
120 | if (unlikely(!anon_vma)) { | |
121 | struct mm_struct *mm = vma->vm_mm; | |
d9d332e0 | 122 | struct anon_vma *allocated; |
1da177e4 | 123 | |
5beb4930 RR |
124 | avc = anon_vma_chain_alloc(); |
125 | if (!avc) | |
126 | goto out_enomem; | |
127 | ||
1da177e4 | 128 | anon_vma = find_mergeable_anon_vma(vma); |
d9d332e0 LT |
129 | allocated = NULL; |
130 | if (!anon_vma) { | |
1da177e4 LT |
131 | anon_vma = anon_vma_alloc(); |
132 | if (unlikely(!anon_vma)) | |
5beb4930 | 133 | goto out_enomem_free_avc; |
1da177e4 | 134 | allocated = anon_vma; |
1da177e4 LT |
135 | } |
136 | ||
31f2b0eb | 137 | spin_lock(&anon_vma->lock); |
1da177e4 LT |
138 | /* page_table_lock to protect against threads */ |
139 | spin_lock(&mm->page_table_lock); | |
140 | if (likely(!vma->anon_vma)) { | |
141 | vma->anon_vma = anon_vma; | |
5beb4930 RR |
142 | avc->anon_vma = anon_vma; |
143 | avc->vma = vma; | |
144 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
145 | list_add(&avc->same_anon_vma, &anon_vma->head); | |
1da177e4 | 146 | allocated = NULL; |
31f2b0eb | 147 | avc = NULL; |
1da177e4 LT |
148 | } |
149 | spin_unlock(&mm->page_table_lock); | |
d9d332e0 | 150 | spin_unlock(&anon_vma->lock); |
31f2b0eb ON |
151 | |
152 | if (unlikely(allocated)) | |
1da177e4 | 153 | anon_vma_free(allocated); |
31f2b0eb | 154 | if (unlikely(avc)) |
5beb4930 | 155 | anon_vma_chain_free(avc); |
1da177e4 LT |
156 | } |
157 | return 0; | |
5beb4930 RR |
158 | |
159 | out_enomem_free_avc: | |
160 | anon_vma_chain_free(avc); | |
161 | out_enomem: | |
162 | return -ENOMEM; | |
1da177e4 LT |
163 | } |
164 | ||
5beb4930 RR |
165 | static void anon_vma_chain_link(struct vm_area_struct *vma, |
166 | struct anon_vma_chain *avc, | |
167 | struct anon_vma *anon_vma) | |
1da177e4 | 168 | { |
5beb4930 RR |
169 | avc->vma = vma; |
170 | avc->anon_vma = anon_vma; | |
171 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
172 | ||
173 | spin_lock(&anon_vma->lock); | |
174 | list_add_tail(&avc->same_anon_vma, &anon_vma->head); | |
175 | spin_unlock(&anon_vma->lock); | |
1da177e4 LT |
176 | } |
177 | ||
5beb4930 RR |
178 | /* |
179 | * Attach the anon_vmas from src to dst. | |
180 | * Returns 0 on success, -ENOMEM on failure. | |
181 | */ | |
182 | int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) | |
1da177e4 | 183 | { |
5beb4930 RR |
184 | struct anon_vma_chain *avc, *pavc; |
185 | ||
646d87b4 | 186 | list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { |
5beb4930 RR |
187 | avc = anon_vma_chain_alloc(); |
188 | if (!avc) | |
189 | goto enomem_failure; | |
190 | anon_vma_chain_link(dst, avc, pavc->anon_vma); | |
191 | } | |
192 | return 0; | |
1da177e4 | 193 | |
5beb4930 RR |
194 | enomem_failure: |
195 | unlink_anon_vmas(dst); | |
196 | return -ENOMEM; | |
1da177e4 LT |
197 | } |
198 | ||
5beb4930 RR |
199 | /* |
200 | * Attach vma to its own anon_vma, as well as to the anon_vmas that | |
201 | * the corresponding VMA in the parent process is attached to. | |
202 | * Returns 0 on success, non-zero on failure. | |
203 | */ | |
204 | int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) | |
1da177e4 | 205 | { |
5beb4930 RR |
206 | struct anon_vma_chain *avc; |
207 | struct anon_vma *anon_vma; | |
1da177e4 | 208 | |
5beb4930 RR |
209 | /* Don't bother if the parent process has no anon_vma here. */ |
210 | if (!pvma->anon_vma) | |
211 | return 0; | |
212 | ||
213 | /* | |
214 | * First, attach the new VMA to the parent VMA's anon_vmas, | |
215 | * so rmap can find non-COWed pages in child processes. | |
216 | */ | |
217 | if (anon_vma_clone(vma, pvma)) | |
218 | return -ENOMEM; | |
219 | ||
220 | /* Then add our own anon_vma. */ | |
221 | anon_vma = anon_vma_alloc(); | |
222 | if (!anon_vma) | |
223 | goto out_error; | |
224 | avc = anon_vma_chain_alloc(); | |
225 | if (!avc) | |
226 | goto out_error_free_anon_vma; | |
227 | anon_vma_chain_link(vma, avc, anon_vma); | |
228 | /* Mark this anon_vma as the one where our new (COWed) pages go. */ | |
229 | vma->anon_vma = anon_vma; | |
230 | ||
231 | return 0; | |
232 | ||
233 | out_error_free_anon_vma: | |
234 | anon_vma_free(anon_vma); | |
235 | out_error: | |
4946d54c | 236 | unlink_anon_vmas(vma); |
5beb4930 | 237 | return -ENOMEM; |
1da177e4 LT |
238 | } |
239 | ||
5beb4930 | 240 | static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain) |
1da177e4 | 241 | { |
5beb4930 | 242 | struct anon_vma *anon_vma = anon_vma_chain->anon_vma; |
1da177e4 LT |
243 | int empty; |
244 | ||
5beb4930 | 245 | /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */ |
1da177e4 LT |
246 | if (!anon_vma) |
247 | return; | |
248 | ||
249 | spin_lock(&anon_vma->lock); | |
5beb4930 | 250 | list_del(&anon_vma_chain->same_anon_vma); |
1da177e4 LT |
251 | |
252 | /* We must garbage collect the anon_vma if it's empty */ | |
db114b83 | 253 | empty = list_empty(&anon_vma->head) && !ksm_refcount(anon_vma); |
1da177e4 LT |
254 | spin_unlock(&anon_vma->lock); |
255 | ||
256 | if (empty) | |
257 | anon_vma_free(anon_vma); | |
258 | } | |
259 | ||
5beb4930 RR |
260 | void unlink_anon_vmas(struct vm_area_struct *vma) |
261 | { | |
262 | struct anon_vma_chain *avc, *next; | |
263 | ||
264 | /* Unlink each anon_vma chained to the VMA. */ | |
265 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { | |
266 | anon_vma_unlink(avc); | |
267 | list_del(&avc->same_vma); | |
268 | anon_vma_chain_free(avc); | |
269 | } | |
270 | } | |
271 | ||
51cc5068 | 272 | static void anon_vma_ctor(void *data) |
1da177e4 | 273 | { |
a35afb83 | 274 | struct anon_vma *anon_vma = data; |
1da177e4 | 275 | |
a35afb83 | 276 | spin_lock_init(&anon_vma->lock); |
db114b83 | 277 | ksm_refcount_init(anon_vma); |
a35afb83 | 278 | INIT_LIST_HEAD(&anon_vma->head); |
1da177e4 LT |
279 | } |
280 | ||
281 | void __init anon_vma_init(void) | |
282 | { | |
283 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
20c2df83 | 284 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor); |
5beb4930 | 285 | anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC); |
1da177e4 LT |
286 | } |
287 | ||
288 | /* | |
289 | * Getting a lock on a stable anon_vma from a page off the LRU is | |
290 | * tricky: page_lock_anon_vma rely on RCU to guard against the races. | |
291 | */ | |
10be22df | 292 | struct anon_vma *page_lock_anon_vma(struct page *page) |
1da177e4 | 293 | { |
34bbd704 | 294 | struct anon_vma *anon_vma; |
1da177e4 LT |
295 | unsigned long anon_mapping; |
296 | ||
297 | rcu_read_lock(); | |
80e14822 | 298 | anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping); |
3ca7b3c5 | 299 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) |
1da177e4 LT |
300 | goto out; |
301 | if (!page_mapped(page)) | |
302 | goto out; | |
303 | ||
304 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
305 | spin_lock(&anon_vma->lock); | |
34bbd704 | 306 | return anon_vma; |
1da177e4 LT |
307 | out: |
308 | rcu_read_unlock(); | |
34bbd704 ON |
309 | return NULL; |
310 | } | |
311 | ||
10be22df | 312 | void page_unlock_anon_vma(struct anon_vma *anon_vma) |
34bbd704 ON |
313 | { |
314 | spin_unlock(&anon_vma->lock); | |
315 | rcu_read_unlock(); | |
1da177e4 LT |
316 | } |
317 | ||
318 | /* | |
3ad33b24 LS |
319 | * At what user virtual address is page expected in @vma? |
320 | * Returns virtual address or -EFAULT if page's index/offset is not | |
321 | * within the range mapped the @vma. | |
1da177e4 LT |
322 | */ |
323 | static inline unsigned long | |
324 | vma_address(struct page *page, struct vm_area_struct *vma) | |
325 | { | |
326 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
327 | unsigned long address; | |
328 | ||
329 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
330 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
3ad33b24 | 331 | /* page should be within @vma mapping range */ |
1da177e4 LT |
332 | return -EFAULT; |
333 | } | |
334 | return address; | |
335 | } | |
336 | ||
337 | /* | |
bf89c8c8 HS |
338 | * At what user virtual address is page expected in vma? |
339 | * checking that the page matches the vma. | |
1da177e4 LT |
340 | */ |
341 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
342 | { | |
343 | if (PageAnon(page)) { | |
3ca7b3c5 | 344 | if (vma->anon_vma != page_anon_vma(page)) |
1da177e4 LT |
345 | return -EFAULT; |
346 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
ee498ed7 HD |
347 | if (!vma->vm_file || |
348 | vma->vm_file->f_mapping != page->mapping) | |
1da177e4 LT |
349 | return -EFAULT; |
350 | } else | |
351 | return -EFAULT; | |
352 | return vma_address(page, vma); | |
353 | } | |
354 | ||
81b4082d ND |
355 | /* |
356 | * Check that @page is mapped at @address into @mm. | |
357 | * | |
479db0bf NP |
358 | * If @sync is false, page_check_address may perform a racy check to avoid |
359 | * the page table lock when the pte is not present (helpful when reclaiming | |
360 | * highly shared pages). | |
361 | * | |
b8072f09 | 362 | * On success returns with pte mapped and locked. |
81b4082d | 363 | */ |
ceffc078 | 364 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
479db0bf | 365 | unsigned long address, spinlock_t **ptlp, int sync) |
81b4082d ND |
366 | { |
367 | pgd_t *pgd; | |
368 | pud_t *pud; | |
369 | pmd_t *pmd; | |
370 | pte_t *pte; | |
c0718806 | 371 | spinlock_t *ptl; |
81b4082d | 372 | |
81b4082d | 373 | pgd = pgd_offset(mm, address); |
c0718806 HD |
374 | if (!pgd_present(*pgd)) |
375 | return NULL; | |
376 | ||
377 | pud = pud_offset(pgd, address); | |
378 | if (!pud_present(*pud)) | |
379 | return NULL; | |
380 | ||
381 | pmd = pmd_offset(pud, address); | |
382 | if (!pmd_present(*pmd)) | |
383 | return NULL; | |
384 | ||
385 | pte = pte_offset_map(pmd, address); | |
386 | /* Make a quick check before getting the lock */ | |
479db0bf | 387 | if (!sync && !pte_present(*pte)) { |
c0718806 HD |
388 | pte_unmap(pte); |
389 | return NULL; | |
390 | } | |
391 | ||
4c21e2f2 | 392 | ptl = pte_lockptr(mm, pmd); |
c0718806 HD |
393 | spin_lock(ptl); |
394 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
395 | *ptlp = ptl; | |
396 | return pte; | |
81b4082d | 397 | } |
c0718806 HD |
398 | pte_unmap_unlock(pte, ptl); |
399 | return NULL; | |
81b4082d ND |
400 | } |
401 | ||
b291f000 NP |
402 | /** |
403 | * page_mapped_in_vma - check whether a page is really mapped in a VMA | |
404 | * @page: the page to test | |
405 | * @vma: the VMA to test | |
406 | * | |
407 | * Returns 1 if the page is mapped into the page tables of the VMA, 0 | |
408 | * if the page is not mapped into the page tables of this VMA. Only | |
409 | * valid for normal file or anonymous VMAs. | |
410 | */ | |
6a46079c | 411 | int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma) |
b291f000 NP |
412 | { |
413 | unsigned long address; | |
414 | pte_t *pte; | |
415 | spinlock_t *ptl; | |
416 | ||
417 | address = vma_address(page, vma); | |
418 | if (address == -EFAULT) /* out of vma range */ | |
419 | return 0; | |
420 | pte = page_check_address(page, vma->vm_mm, address, &ptl, 1); | |
421 | if (!pte) /* the page is not in this mm */ | |
422 | return 0; | |
423 | pte_unmap_unlock(pte, ptl); | |
424 | ||
425 | return 1; | |
426 | } | |
427 | ||
1da177e4 LT |
428 | /* |
429 | * Subfunctions of page_referenced: page_referenced_one called | |
430 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
431 | */ | |
5ad64688 HD |
432 | int page_referenced_one(struct page *page, struct vm_area_struct *vma, |
433 | unsigned long address, unsigned int *mapcount, | |
434 | unsigned long *vm_flags) | |
1da177e4 LT |
435 | { |
436 | struct mm_struct *mm = vma->vm_mm; | |
1da177e4 | 437 | pte_t *pte; |
c0718806 | 438 | spinlock_t *ptl; |
1da177e4 LT |
439 | int referenced = 0; |
440 | ||
479db0bf | 441 | pte = page_check_address(page, mm, address, &ptl, 0); |
c0718806 HD |
442 | if (!pte) |
443 | goto out; | |
1da177e4 | 444 | |
b291f000 NP |
445 | /* |
446 | * Don't want to elevate referenced for mlocked page that gets this far, | |
447 | * in order that it progresses to try_to_unmap and is moved to the | |
448 | * unevictable list. | |
449 | */ | |
5a9bbdcd | 450 | if (vma->vm_flags & VM_LOCKED) { |
5a9bbdcd | 451 | *mapcount = 1; /* break early from loop */ |
03ef83af | 452 | *vm_flags |= VM_LOCKED; |
b291f000 NP |
453 | goto out_unmap; |
454 | } | |
455 | ||
4917e5d0 JW |
456 | if (ptep_clear_flush_young_notify(vma, address, pte)) { |
457 | /* | |
458 | * Don't treat a reference through a sequentially read | |
459 | * mapping as such. If the page has been used in | |
460 | * another mapping, we will catch it; if this other | |
461 | * mapping is already gone, the unmap path will have | |
462 | * set PG_referenced or activated the page. | |
463 | */ | |
464 | if (likely(!VM_SequentialReadHint(vma))) | |
465 | referenced++; | |
466 | } | |
1da177e4 | 467 | |
c0718806 HD |
468 | /* Pretend the page is referenced if the task has the |
469 | swap token and is in the middle of a page fault. */ | |
f7b7fd8f | 470 | if (mm != current->mm && has_swap_token(mm) && |
c0718806 HD |
471 | rwsem_is_locked(&mm->mmap_sem)) |
472 | referenced++; | |
473 | ||
b291f000 | 474 | out_unmap: |
c0718806 HD |
475 | (*mapcount)--; |
476 | pte_unmap_unlock(pte, ptl); | |
273f047e | 477 | |
6fe6b7e3 WF |
478 | if (referenced) |
479 | *vm_flags |= vma->vm_flags; | |
273f047e | 480 | out: |
1da177e4 LT |
481 | return referenced; |
482 | } | |
483 | ||
bed7161a | 484 | static int page_referenced_anon(struct page *page, |
6fe6b7e3 WF |
485 | struct mem_cgroup *mem_cont, |
486 | unsigned long *vm_flags) | |
1da177e4 LT |
487 | { |
488 | unsigned int mapcount; | |
489 | struct anon_vma *anon_vma; | |
5beb4930 | 490 | struct anon_vma_chain *avc; |
1da177e4 LT |
491 | int referenced = 0; |
492 | ||
493 | anon_vma = page_lock_anon_vma(page); | |
494 | if (!anon_vma) | |
495 | return referenced; | |
496 | ||
497 | mapcount = page_mapcount(page); | |
5beb4930 RR |
498 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { |
499 | struct vm_area_struct *vma = avc->vma; | |
1cb1729b HD |
500 | unsigned long address = vma_address(page, vma); |
501 | if (address == -EFAULT) | |
502 | continue; | |
bed7161a BS |
503 | /* |
504 | * If we are reclaiming on behalf of a cgroup, skip | |
505 | * counting on behalf of references from different | |
506 | * cgroups | |
507 | */ | |
bd845e38 | 508 | if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont)) |
bed7161a | 509 | continue; |
1cb1729b | 510 | referenced += page_referenced_one(page, vma, address, |
6fe6b7e3 | 511 | &mapcount, vm_flags); |
1da177e4 LT |
512 | if (!mapcount) |
513 | break; | |
514 | } | |
34bbd704 ON |
515 | |
516 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
517 | return referenced; |
518 | } | |
519 | ||
520 | /** | |
521 | * page_referenced_file - referenced check for object-based rmap | |
522 | * @page: the page we're checking references on. | |
43d8eac4 | 523 | * @mem_cont: target memory controller |
6fe6b7e3 | 524 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page |
1da177e4 LT |
525 | * |
526 | * For an object-based mapped page, find all the places it is mapped and | |
527 | * check/clear the referenced flag. This is done by following the page->mapping | |
528 | * pointer, then walking the chain of vmas it holds. It returns the number | |
529 | * of references it found. | |
530 | * | |
531 | * This function is only called from page_referenced for object-based pages. | |
532 | */ | |
bed7161a | 533 | static int page_referenced_file(struct page *page, |
6fe6b7e3 WF |
534 | struct mem_cgroup *mem_cont, |
535 | unsigned long *vm_flags) | |
1da177e4 LT |
536 | { |
537 | unsigned int mapcount; | |
538 | struct address_space *mapping = page->mapping; | |
539 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
540 | struct vm_area_struct *vma; | |
541 | struct prio_tree_iter iter; | |
542 | int referenced = 0; | |
543 | ||
544 | /* | |
545 | * The caller's checks on page->mapping and !PageAnon have made | |
546 | * sure that this is a file page: the check for page->mapping | |
547 | * excludes the case just before it gets set on an anon page. | |
548 | */ | |
549 | BUG_ON(PageAnon(page)); | |
550 | ||
551 | /* | |
552 | * The page lock not only makes sure that page->mapping cannot | |
553 | * suddenly be NULLified by truncation, it makes sure that the | |
554 | * structure at mapping cannot be freed and reused yet, | |
555 | * so we can safely take mapping->i_mmap_lock. | |
556 | */ | |
557 | BUG_ON(!PageLocked(page)); | |
558 | ||
559 | spin_lock(&mapping->i_mmap_lock); | |
560 | ||
561 | /* | |
562 | * i_mmap_lock does not stabilize mapcount at all, but mapcount | |
563 | * is more likely to be accurate if we note it after spinning. | |
564 | */ | |
565 | mapcount = page_mapcount(page); | |
566 | ||
567 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
1cb1729b HD |
568 | unsigned long address = vma_address(page, vma); |
569 | if (address == -EFAULT) | |
570 | continue; | |
bed7161a BS |
571 | /* |
572 | * If we are reclaiming on behalf of a cgroup, skip | |
573 | * counting on behalf of references from different | |
574 | * cgroups | |
575 | */ | |
bd845e38 | 576 | if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont)) |
bed7161a | 577 | continue; |
1cb1729b | 578 | referenced += page_referenced_one(page, vma, address, |
6fe6b7e3 | 579 | &mapcount, vm_flags); |
1da177e4 LT |
580 | if (!mapcount) |
581 | break; | |
582 | } | |
583 | ||
584 | spin_unlock(&mapping->i_mmap_lock); | |
585 | return referenced; | |
586 | } | |
587 | ||
588 | /** | |
589 | * page_referenced - test if the page was referenced | |
590 | * @page: the page to test | |
591 | * @is_locked: caller holds lock on the page | |
43d8eac4 | 592 | * @mem_cont: target memory controller |
6fe6b7e3 | 593 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page |
1da177e4 LT |
594 | * |
595 | * Quick test_and_clear_referenced for all mappings to a page, | |
596 | * returns the number of ptes which referenced the page. | |
597 | */ | |
6fe6b7e3 WF |
598 | int page_referenced(struct page *page, |
599 | int is_locked, | |
600 | struct mem_cgroup *mem_cont, | |
601 | unsigned long *vm_flags) | |
1da177e4 LT |
602 | { |
603 | int referenced = 0; | |
5ad64688 | 604 | int we_locked = 0; |
1da177e4 | 605 | |
6fe6b7e3 | 606 | *vm_flags = 0; |
3ca7b3c5 | 607 | if (page_mapped(page) && page_rmapping(page)) { |
5ad64688 HD |
608 | if (!is_locked && (!PageAnon(page) || PageKsm(page))) { |
609 | we_locked = trylock_page(page); | |
610 | if (!we_locked) { | |
611 | referenced++; | |
612 | goto out; | |
613 | } | |
614 | } | |
615 | if (unlikely(PageKsm(page))) | |
616 | referenced += page_referenced_ksm(page, mem_cont, | |
617 | vm_flags); | |
618 | else if (PageAnon(page)) | |
6fe6b7e3 WF |
619 | referenced += page_referenced_anon(page, mem_cont, |
620 | vm_flags); | |
5ad64688 | 621 | else if (page->mapping) |
6fe6b7e3 WF |
622 | referenced += page_referenced_file(page, mem_cont, |
623 | vm_flags); | |
5ad64688 | 624 | if (we_locked) |
1da177e4 | 625 | unlock_page(page); |
1da177e4 | 626 | } |
5ad64688 | 627 | out: |
5b7baf05 CB |
628 | if (page_test_and_clear_young(page)) |
629 | referenced++; | |
630 | ||
1da177e4 LT |
631 | return referenced; |
632 | } | |
633 | ||
1cb1729b HD |
634 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma, |
635 | unsigned long address) | |
d08b3851 PZ |
636 | { |
637 | struct mm_struct *mm = vma->vm_mm; | |
c2fda5fe | 638 | pte_t *pte; |
d08b3851 PZ |
639 | spinlock_t *ptl; |
640 | int ret = 0; | |
641 | ||
479db0bf | 642 | pte = page_check_address(page, mm, address, &ptl, 1); |
d08b3851 PZ |
643 | if (!pte) |
644 | goto out; | |
645 | ||
c2fda5fe PZ |
646 | if (pte_dirty(*pte) || pte_write(*pte)) { |
647 | pte_t entry; | |
d08b3851 | 648 | |
c2fda5fe | 649 | flush_cache_page(vma, address, pte_pfn(*pte)); |
cddb8a5c | 650 | entry = ptep_clear_flush_notify(vma, address, pte); |
c2fda5fe PZ |
651 | entry = pte_wrprotect(entry); |
652 | entry = pte_mkclean(entry); | |
d6e88e67 | 653 | set_pte_at(mm, address, pte, entry); |
c2fda5fe PZ |
654 | ret = 1; |
655 | } | |
d08b3851 | 656 | |
d08b3851 PZ |
657 | pte_unmap_unlock(pte, ptl); |
658 | out: | |
659 | return ret; | |
660 | } | |
661 | ||
662 | static int page_mkclean_file(struct address_space *mapping, struct page *page) | |
663 | { | |
664 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
665 | struct vm_area_struct *vma; | |
666 | struct prio_tree_iter iter; | |
667 | int ret = 0; | |
668 | ||
669 | BUG_ON(PageAnon(page)); | |
670 | ||
671 | spin_lock(&mapping->i_mmap_lock); | |
672 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
1cb1729b HD |
673 | if (vma->vm_flags & VM_SHARED) { |
674 | unsigned long address = vma_address(page, vma); | |
675 | if (address == -EFAULT) | |
676 | continue; | |
677 | ret += page_mkclean_one(page, vma, address); | |
678 | } | |
d08b3851 PZ |
679 | } |
680 | spin_unlock(&mapping->i_mmap_lock); | |
681 | return ret; | |
682 | } | |
683 | ||
684 | int page_mkclean(struct page *page) | |
685 | { | |
686 | int ret = 0; | |
687 | ||
688 | BUG_ON(!PageLocked(page)); | |
689 | ||
690 | if (page_mapped(page)) { | |
691 | struct address_space *mapping = page_mapping(page); | |
ce7e9fae | 692 | if (mapping) { |
d08b3851 | 693 | ret = page_mkclean_file(mapping, page); |
ce7e9fae CB |
694 | if (page_test_dirty(page)) { |
695 | page_clear_dirty(page); | |
696 | ret = 1; | |
697 | } | |
6c210482 | 698 | } |
d08b3851 PZ |
699 | } |
700 | ||
701 | return ret; | |
702 | } | |
60b59bea | 703 | EXPORT_SYMBOL_GPL(page_mkclean); |
d08b3851 | 704 | |
c44b6743 RR |
705 | /** |
706 | * page_move_anon_rmap - move a page to our anon_vma | |
707 | * @page: the page to move to our anon_vma | |
708 | * @vma: the vma the page belongs to | |
709 | * @address: the user virtual address mapped | |
710 | * | |
711 | * When a page belongs exclusively to one process after a COW event, | |
712 | * that page can be moved into the anon_vma that belongs to just that | |
713 | * process, so the rmap code will not search the parent or sibling | |
714 | * processes. | |
715 | */ | |
716 | void page_move_anon_rmap(struct page *page, | |
717 | struct vm_area_struct *vma, unsigned long address) | |
718 | { | |
719 | struct anon_vma *anon_vma = vma->anon_vma; | |
720 | ||
721 | VM_BUG_ON(!PageLocked(page)); | |
722 | VM_BUG_ON(!anon_vma); | |
723 | VM_BUG_ON(page->index != linear_page_index(vma, address)); | |
724 | ||
725 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
726 | page->mapping = (struct address_space *) anon_vma; | |
727 | } | |
728 | ||
9617d95e | 729 | /** |
43d8eac4 | 730 | * __page_set_anon_rmap - setup new anonymous rmap |
9617d95e NP |
731 | * @page: the page to add the mapping to |
732 | * @vma: the vm area in which the mapping is added | |
733 | * @address: the user virtual address mapped | |
e8a03feb | 734 | * @exclusive: the page is exclusively owned by the current process |
9617d95e NP |
735 | */ |
736 | static void __page_set_anon_rmap(struct page *page, | |
e8a03feb | 737 | struct vm_area_struct *vma, unsigned long address, int exclusive) |
9617d95e | 738 | { |
e8a03feb | 739 | struct anon_vma *anon_vma = vma->anon_vma; |
ea90002b | 740 | |
e8a03feb | 741 | BUG_ON(!anon_vma); |
ea90002b LT |
742 | |
743 | /* | |
e8a03feb RR |
744 | * If the page isn't exclusively mapped into this vma, |
745 | * we must use the _oldest_ possible anon_vma for the | |
746 | * page mapping! | |
ea90002b | 747 | * |
e8a03feb RR |
748 | * So take the last AVC chain entry in the vma, which is |
749 | * the deepest ancestor, and use the anon_vma from that. | |
ea90002b | 750 | */ |
e8a03feb RR |
751 | if (!exclusive) { |
752 | struct anon_vma_chain *avc; | |
753 | avc = list_entry(vma->anon_vma_chain.prev, struct anon_vma_chain, same_vma); | |
754 | anon_vma = avc->anon_vma; | |
755 | } | |
9617d95e | 756 | |
9617d95e NP |
757 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
758 | page->mapping = (struct address_space *) anon_vma; | |
9617d95e | 759 | page->index = linear_page_index(vma, address); |
9617d95e NP |
760 | } |
761 | ||
c97a9e10 | 762 | /** |
43d8eac4 | 763 | * __page_check_anon_rmap - sanity check anonymous rmap addition |
c97a9e10 NP |
764 | * @page: the page to add the mapping to |
765 | * @vma: the vm area in which the mapping is added | |
766 | * @address: the user virtual address mapped | |
767 | */ | |
768 | static void __page_check_anon_rmap(struct page *page, | |
769 | struct vm_area_struct *vma, unsigned long address) | |
770 | { | |
771 | #ifdef CONFIG_DEBUG_VM | |
772 | /* | |
773 | * The page's anon-rmap details (mapping and index) are guaranteed to | |
774 | * be set up correctly at this point. | |
775 | * | |
776 | * We have exclusion against page_add_anon_rmap because the caller | |
777 | * always holds the page locked, except if called from page_dup_rmap, | |
778 | * in which case the page is already known to be setup. | |
779 | * | |
780 | * We have exclusion against page_add_new_anon_rmap because those pages | |
781 | * are initially only visible via the pagetables, and the pte is locked | |
782 | * over the call to page_add_new_anon_rmap. | |
783 | */ | |
c97a9e10 NP |
784 | BUG_ON(page->index != linear_page_index(vma, address)); |
785 | #endif | |
786 | } | |
787 | ||
1da177e4 LT |
788 | /** |
789 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
790 | * @page: the page to add the mapping to | |
791 | * @vma: the vm area in which the mapping is added | |
792 | * @address: the user virtual address mapped | |
793 | * | |
5ad64688 | 794 | * The caller needs to hold the pte lock, and the page must be locked in |
80e14822 HD |
795 | * the anon_vma case: to serialize mapping,index checking after setting, |
796 | * and to ensure that PageAnon is not being upgraded racily to PageKsm | |
797 | * (but PageKsm is never downgraded to PageAnon). | |
1da177e4 LT |
798 | */ |
799 | void page_add_anon_rmap(struct page *page, | |
800 | struct vm_area_struct *vma, unsigned long address) | |
801 | { | |
5ad64688 HD |
802 | int first = atomic_inc_and_test(&page->_mapcount); |
803 | if (first) | |
804 | __inc_zone_page_state(page, NR_ANON_PAGES); | |
805 | if (unlikely(PageKsm(page))) | |
806 | return; | |
807 | ||
c97a9e10 NP |
808 | VM_BUG_ON(!PageLocked(page)); |
809 | VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); | |
5ad64688 | 810 | if (first) |
e8a03feb | 811 | __page_set_anon_rmap(page, vma, address, 0); |
69029cd5 | 812 | else |
c97a9e10 | 813 | __page_check_anon_rmap(page, vma, address); |
1da177e4 LT |
814 | } |
815 | ||
43d8eac4 | 816 | /** |
9617d95e NP |
817 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page |
818 | * @page: the page to add the mapping to | |
819 | * @vma: the vm area in which the mapping is added | |
820 | * @address: the user virtual address mapped | |
821 | * | |
822 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
823 | * This means the inc-and-test can be bypassed. | |
c97a9e10 | 824 | * Page does not have to be locked. |
9617d95e NP |
825 | */ |
826 | void page_add_new_anon_rmap(struct page *page, | |
827 | struct vm_area_struct *vma, unsigned long address) | |
828 | { | |
b5934c53 | 829 | VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); |
cbf84b7a HD |
830 | SetPageSwapBacked(page); |
831 | atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */ | |
5ad64688 | 832 | __inc_zone_page_state(page, NR_ANON_PAGES); |
e8a03feb | 833 | __page_set_anon_rmap(page, vma, address, 1); |
b5934c53 | 834 | if (page_evictable(page, vma)) |
cbf84b7a | 835 | lru_cache_add_lru(page, LRU_ACTIVE_ANON); |
b5934c53 HD |
836 | else |
837 | add_page_to_unevictable_list(page); | |
9617d95e NP |
838 | } |
839 | ||
1da177e4 LT |
840 | /** |
841 | * page_add_file_rmap - add pte mapping to a file page | |
842 | * @page: the page to add the mapping to | |
843 | * | |
b8072f09 | 844 | * The caller needs to hold the pte lock. |
1da177e4 LT |
845 | */ |
846 | void page_add_file_rmap(struct page *page) | |
847 | { | |
d69b042f | 848 | if (atomic_inc_and_test(&page->_mapcount)) { |
65ba55f5 | 849 | __inc_zone_page_state(page, NR_FILE_MAPPED); |
d8046582 | 850 | mem_cgroup_update_file_mapped(page, 1); |
d69b042f | 851 | } |
1da177e4 LT |
852 | } |
853 | ||
854 | /** | |
855 | * page_remove_rmap - take down pte mapping from a page | |
856 | * @page: page to remove mapping from | |
857 | * | |
b8072f09 | 858 | * The caller needs to hold the pte lock. |
1da177e4 | 859 | */ |
edc315fd | 860 | void page_remove_rmap(struct page *page) |
1da177e4 | 861 | { |
b904dcfe KM |
862 | /* page still mapped by someone else? */ |
863 | if (!atomic_add_negative(-1, &page->_mapcount)) | |
864 | return; | |
865 | ||
866 | /* | |
867 | * Now that the last pte has gone, s390 must transfer dirty | |
868 | * flag from storage key to struct page. We can usually skip | |
869 | * this if the page is anon, so about to be freed; but perhaps | |
870 | * not if it's in swapcache - there might be another pte slot | |
871 | * containing the swap entry, but page not yet written to swap. | |
872 | */ | |
873 | if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) { | |
874 | page_clear_dirty(page); | |
875 | set_page_dirty(page); | |
1da177e4 | 876 | } |
b904dcfe KM |
877 | if (PageAnon(page)) { |
878 | mem_cgroup_uncharge_page(page); | |
879 | __dec_zone_page_state(page, NR_ANON_PAGES); | |
880 | } else { | |
881 | __dec_zone_page_state(page, NR_FILE_MAPPED); | |
d8046582 | 882 | mem_cgroup_update_file_mapped(page, -1); |
b904dcfe | 883 | } |
b904dcfe KM |
884 | /* |
885 | * It would be tidy to reset the PageAnon mapping here, | |
886 | * but that might overwrite a racing page_add_anon_rmap | |
887 | * which increments mapcount after us but sets mapping | |
888 | * before us: so leave the reset to free_hot_cold_page, | |
889 | * and remember that it's only reliable while mapped. | |
890 | * Leaving it set also helps swapoff to reinstate ptes | |
891 | * faster for those pages still in swapcache. | |
892 | */ | |
1da177e4 LT |
893 | } |
894 | ||
895 | /* | |
896 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
897 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
898 | */ | |
5ad64688 HD |
899 | int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
900 | unsigned long address, enum ttu_flags flags) | |
1da177e4 LT |
901 | { |
902 | struct mm_struct *mm = vma->vm_mm; | |
1da177e4 LT |
903 | pte_t *pte; |
904 | pte_t pteval; | |
c0718806 | 905 | spinlock_t *ptl; |
1da177e4 LT |
906 | int ret = SWAP_AGAIN; |
907 | ||
479db0bf | 908 | pte = page_check_address(page, mm, address, &ptl, 0); |
c0718806 | 909 | if (!pte) |
81b4082d | 910 | goto out; |
1da177e4 LT |
911 | |
912 | /* | |
913 | * If the page is mlock()d, we cannot swap it out. | |
914 | * If it's recently referenced (perhaps page_referenced | |
915 | * skipped over this mm) then we should reactivate it. | |
916 | */ | |
14fa31b8 | 917 | if (!(flags & TTU_IGNORE_MLOCK)) { |
caed0f48 KM |
918 | if (vma->vm_flags & VM_LOCKED) |
919 | goto out_mlock; | |
920 | ||
af8e3354 | 921 | if (TTU_ACTION(flags) == TTU_MUNLOCK) |
53f79acb | 922 | goto out_unmap; |
14fa31b8 AK |
923 | } |
924 | if (!(flags & TTU_IGNORE_ACCESS)) { | |
b291f000 NP |
925 | if (ptep_clear_flush_young_notify(vma, address, pte)) { |
926 | ret = SWAP_FAIL; | |
927 | goto out_unmap; | |
928 | } | |
929 | } | |
1da177e4 | 930 | |
1da177e4 LT |
931 | /* Nuke the page table entry. */ |
932 | flush_cache_page(vma, address, page_to_pfn(page)); | |
cddb8a5c | 933 | pteval = ptep_clear_flush_notify(vma, address, pte); |
1da177e4 LT |
934 | |
935 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
936 | if (pte_dirty(pteval)) | |
937 | set_page_dirty(page); | |
938 | ||
365e9c87 HD |
939 | /* Update high watermark before we lower rss */ |
940 | update_hiwater_rss(mm); | |
941 | ||
888b9f7c AK |
942 | if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { |
943 | if (PageAnon(page)) | |
d559db08 | 944 | dec_mm_counter(mm, MM_ANONPAGES); |
888b9f7c | 945 | else |
d559db08 | 946 | dec_mm_counter(mm, MM_FILEPAGES); |
888b9f7c AK |
947 | set_pte_at(mm, address, pte, |
948 | swp_entry_to_pte(make_hwpoison_entry(page))); | |
949 | } else if (PageAnon(page)) { | |
4c21e2f2 | 950 | swp_entry_t entry = { .val = page_private(page) }; |
0697212a CL |
951 | |
952 | if (PageSwapCache(page)) { | |
953 | /* | |
954 | * Store the swap location in the pte. | |
955 | * See handle_pte_fault() ... | |
956 | */ | |
570a335b HD |
957 | if (swap_duplicate(entry) < 0) { |
958 | set_pte_at(mm, address, pte, pteval); | |
959 | ret = SWAP_FAIL; | |
960 | goto out_unmap; | |
961 | } | |
0697212a CL |
962 | if (list_empty(&mm->mmlist)) { |
963 | spin_lock(&mmlist_lock); | |
964 | if (list_empty(&mm->mmlist)) | |
965 | list_add(&mm->mmlist, &init_mm.mmlist); | |
966 | spin_unlock(&mmlist_lock); | |
967 | } | |
d559db08 | 968 | dec_mm_counter(mm, MM_ANONPAGES); |
b084d435 | 969 | inc_mm_counter(mm, MM_SWAPENTS); |
64cdd548 | 970 | } else if (PAGE_MIGRATION) { |
0697212a CL |
971 | /* |
972 | * Store the pfn of the page in a special migration | |
973 | * pte. do_swap_page() will wait until the migration | |
974 | * pte is removed and then restart fault handling. | |
975 | */ | |
14fa31b8 | 976 | BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION); |
0697212a | 977 | entry = make_migration_entry(page, pte_write(pteval)); |
1da177e4 LT |
978 | } |
979 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
980 | BUG_ON(pte_file(*pte)); | |
14fa31b8 | 981 | } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) { |
04e62a29 CL |
982 | /* Establish migration entry for a file page */ |
983 | swp_entry_t entry; | |
984 | entry = make_migration_entry(page, pte_write(pteval)); | |
985 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
986 | } else | |
d559db08 | 987 | dec_mm_counter(mm, MM_FILEPAGES); |
1da177e4 | 988 | |
edc315fd | 989 | page_remove_rmap(page); |
1da177e4 LT |
990 | page_cache_release(page); |
991 | ||
992 | out_unmap: | |
c0718806 | 993 | pte_unmap_unlock(pte, ptl); |
caed0f48 KM |
994 | out: |
995 | return ret; | |
53f79acb | 996 | |
caed0f48 KM |
997 | out_mlock: |
998 | pte_unmap_unlock(pte, ptl); | |
999 | ||
1000 | ||
1001 | /* | |
1002 | * We need mmap_sem locking, Otherwise VM_LOCKED check makes | |
1003 | * unstable result and race. Plus, We can't wait here because | |
1004 | * we now hold anon_vma->lock or mapping->i_mmap_lock. | |
1005 | * if trylock failed, the page remain in evictable lru and later | |
1006 | * vmscan could retry to move the page to unevictable lru if the | |
1007 | * page is actually mlocked. | |
1008 | */ | |
1009 | if (down_read_trylock(&vma->vm_mm->mmap_sem)) { | |
1010 | if (vma->vm_flags & VM_LOCKED) { | |
1011 | mlock_vma_page(page); | |
1012 | ret = SWAP_MLOCK; | |
53f79acb | 1013 | } |
caed0f48 | 1014 | up_read(&vma->vm_mm->mmap_sem); |
53f79acb | 1015 | } |
1da177e4 LT |
1016 | return ret; |
1017 | } | |
1018 | ||
1019 | /* | |
1020 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
1021 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
1022 | * Consequently, given a particular page and its ->index, we cannot locate the | |
1023 | * ptes which are mapping that page without an exhaustive linear search. | |
1024 | * | |
1025 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
1026 | * maps the file to which the target page belongs. The ->vm_private_data field | |
1027 | * holds the current cursor into that scan. Successive searches will circulate | |
1028 | * around the vma's virtual address space. | |
1029 | * | |
1030 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
1031 | * more scanning pressure is placed against them as well. Eventually pages | |
1032 | * will become fully unmapped and are eligible for eviction. | |
1033 | * | |
1034 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
1035 | * there there won't be many ptes located within the scan cluster. In this case | |
1036 | * maybe we could scan further - to the end of the pte page, perhaps. | |
b291f000 NP |
1037 | * |
1038 | * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can | |
1039 | * acquire it without blocking. If vma locked, mlock the pages in the cluster, | |
1040 | * rather than unmapping them. If we encounter the "check_page" that vmscan is | |
1041 | * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN. | |
1da177e4 LT |
1042 | */ |
1043 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
1044 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
1045 | ||
b291f000 NP |
1046 | static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount, |
1047 | struct vm_area_struct *vma, struct page *check_page) | |
1da177e4 LT |
1048 | { |
1049 | struct mm_struct *mm = vma->vm_mm; | |
1050 | pgd_t *pgd; | |
1051 | pud_t *pud; | |
1052 | pmd_t *pmd; | |
c0718806 | 1053 | pte_t *pte; |
1da177e4 | 1054 | pte_t pteval; |
c0718806 | 1055 | spinlock_t *ptl; |
1da177e4 LT |
1056 | struct page *page; |
1057 | unsigned long address; | |
1058 | unsigned long end; | |
b291f000 NP |
1059 | int ret = SWAP_AGAIN; |
1060 | int locked_vma = 0; | |
1da177e4 | 1061 | |
1da177e4 LT |
1062 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
1063 | end = address + CLUSTER_SIZE; | |
1064 | if (address < vma->vm_start) | |
1065 | address = vma->vm_start; | |
1066 | if (end > vma->vm_end) | |
1067 | end = vma->vm_end; | |
1068 | ||
1069 | pgd = pgd_offset(mm, address); | |
1070 | if (!pgd_present(*pgd)) | |
b291f000 | 1071 | return ret; |
1da177e4 LT |
1072 | |
1073 | pud = pud_offset(pgd, address); | |
1074 | if (!pud_present(*pud)) | |
b291f000 | 1075 | return ret; |
1da177e4 LT |
1076 | |
1077 | pmd = pmd_offset(pud, address); | |
1078 | if (!pmd_present(*pmd)) | |
b291f000 NP |
1079 | return ret; |
1080 | ||
1081 | /* | |
af8e3354 | 1082 | * If we can acquire the mmap_sem for read, and vma is VM_LOCKED, |
b291f000 NP |
1083 | * keep the sem while scanning the cluster for mlocking pages. |
1084 | */ | |
af8e3354 | 1085 | if (down_read_trylock(&vma->vm_mm->mmap_sem)) { |
b291f000 NP |
1086 | locked_vma = (vma->vm_flags & VM_LOCKED); |
1087 | if (!locked_vma) | |
1088 | up_read(&vma->vm_mm->mmap_sem); /* don't need it */ | |
1089 | } | |
c0718806 HD |
1090 | |
1091 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1da177e4 | 1092 | |
365e9c87 HD |
1093 | /* Update high watermark before we lower rss */ |
1094 | update_hiwater_rss(mm); | |
1095 | ||
c0718806 | 1096 | for (; address < end; pte++, address += PAGE_SIZE) { |
1da177e4 LT |
1097 | if (!pte_present(*pte)) |
1098 | continue; | |
6aab341e LT |
1099 | page = vm_normal_page(vma, address, *pte); |
1100 | BUG_ON(!page || PageAnon(page)); | |
1da177e4 | 1101 | |
b291f000 NP |
1102 | if (locked_vma) { |
1103 | mlock_vma_page(page); /* no-op if already mlocked */ | |
1104 | if (page == check_page) | |
1105 | ret = SWAP_MLOCK; | |
1106 | continue; /* don't unmap */ | |
1107 | } | |
1108 | ||
cddb8a5c | 1109 | if (ptep_clear_flush_young_notify(vma, address, pte)) |
1da177e4 LT |
1110 | continue; |
1111 | ||
1112 | /* Nuke the page table entry. */ | |
eca35133 | 1113 | flush_cache_page(vma, address, pte_pfn(*pte)); |
cddb8a5c | 1114 | pteval = ptep_clear_flush_notify(vma, address, pte); |
1da177e4 LT |
1115 | |
1116 | /* If nonlinear, store the file page offset in the pte. */ | |
1117 | if (page->index != linear_page_index(vma, address)) | |
1118 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
1119 | ||
1120 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
1121 | if (pte_dirty(pteval)) | |
1122 | set_page_dirty(page); | |
1123 | ||
edc315fd | 1124 | page_remove_rmap(page); |
1da177e4 | 1125 | page_cache_release(page); |
d559db08 | 1126 | dec_mm_counter(mm, MM_FILEPAGES); |
1da177e4 LT |
1127 | (*mapcount)--; |
1128 | } | |
c0718806 | 1129 | pte_unmap_unlock(pte - 1, ptl); |
b291f000 NP |
1130 | if (locked_vma) |
1131 | up_read(&vma->vm_mm->mmap_sem); | |
1132 | return ret; | |
1da177e4 LT |
1133 | } |
1134 | ||
b291f000 NP |
1135 | /** |
1136 | * try_to_unmap_anon - unmap or unlock anonymous page using the object-based | |
1137 | * rmap method | |
1138 | * @page: the page to unmap/unlock | |
8051be5e | 1139 | * @flags: action and flags |
b291f000 NP |
1140 | * |
1141 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
1142 | * contained in the anon_vma struct it points to. | |
1143 | * | |
1144 | * This function is only called from try_to_unmap/try_to_munlock for | |
1145 | * anonymous pages. | |
1146 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma | |
1147 | * where the page was found will be held for write. So, we won't recheck | |
1148 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be | |
1149 | * 'LOCKED. | |
1150 | */ | |
14fa31b8 | 1151 | static int try_to_unmap_anon(struct page *page, enum ttu_flags flags) |
1da177e4 LT |
1152 | { |
1153 | struct anon_vma *anon_vma; | |
5beb4930 | 1154 | struct anon_vma_chain *avc; |
1da177e4 | 1155 | int ret = SWAP_AGAIN; |
b291f000 | 1156 | |
1da177e4 LT |
1157 | anon_vma = page_lock_anon_vma(page); |
1158 | if (!anon_vma) | |
1159 | return ret; | |
1160 | ||
5beb4930 RR |
1161 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { |
1162 | struct vm_area_struct *vma = avc->vma; | |
1cb1729b HD |
1163 | unsigned long address = vma_address(page, vma); |
1164 | if (address == -EFAULT) | |
1165 | continue; | |
1166 | ret = try_to_unmap_one(page, vma, address, flags); | |
53f79acb HD |
1167 | if (ret != SWAP_AGAIN || !page_mapped(page)) |
1168 | break; | |
1da177e4 | 1169 | } |
34bbd704 ON |
1170 | |
1171 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
1172 | return ret; |
1173 | } | |
1174 | ||
1175 | /** | |
b291f000 NP |
1176 | * try_to_unmap_file - unmap/unlock file page using the object-based rmap method |
1177 | * @page: the page to unmap/unlock | |
14fa31b8 | 1178 | * @flags: action and flags |
1da177e4 LT |
1179 | * |
1180 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
1181 | * contained in the address_space struct it points to. | |
1182 | * | |
b291f000 NP |
1183 | * This function is only called from try_to_unmap/try_to_munlock for |
1184 | * object-based pages. | |
1185 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma | |
1186 | * where the page was found will be held for write. So, we won't recheck | |
1187 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be | |
1188 | * 'LOCKED. | |
1da177e4 | 1189 | */ |
14fa31b8 | 1190 | static int try_to_unmap_file(struct page *page, enum ttu_flags flags) |
1da177e4 LT |
1191 | { |
1192 | struct address_space *mapping = page->mapping; | |
1193 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
1194 | struct vm_area_struct *vma; | |
1195 | struct prio_tree_iter iter; | |
1196 | int ret = SWAP_AGAIN; | |
1197 | unsigned long cursor; | |
1198 | unsigned long max_nl_cursor = 0; | |
1199 | unsigned long max_nl_size = 0; | |
1200 | unsigned int mapcount; | |
1201 | ||
1202 | spin_lock(&mapping->i_mmap_lock); | |
1203 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
1cb1729b HD |
1204 | unsigned long address = vma_address(page, vma); |
1205 | if (address == -EFAULT) | |
1206 | continue; | |
1207 | ret = try_to_unmap_one(page, vma, address, flags); | |
53f79acb HD |
1208 | if (ret != SWAP_AGAIN || !page_mapped(page)) |
1209 | goto out; | |
1da177e4 LT |
1210 | } |
1211 | ||
1212 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
1213 | goto out; | |
1214 | ||
53f79acb HD |
1215 | /* |
1216 | * We don't bother to try to find the munlocked page in nonlinears. | |
1217 | * It's costly. Instead, later, page reclaim logic may call | |
1218 | * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily. | |
1219 | */ | |
1220 | if (TTU_ACTION(flags) == TTU_MUNLOCK) | |
1221 | goto out; | |
1222 | ||
1da177e4 LT |
1223 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, |
1224 | shared.vm_set.list) { | |
1da177e4 LT |
1225 | cursor = (unsigned long) vma->vm_private_data; |
1226 | if (cursor > max_nl_cursor) | |
1227 | max_nl_cursor = cursor; | |
1228 | cursor = vma->vm_end - vma->vm_start; | |
1229 | if (cursor > max_nl_size) | |
1230 | max_nl_size = cursor; | |
1231 | } | |
1232 | ||
b291f000 | 1233 | if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */ |
1da177e4 LT |
1234 | ret = SWAP_FAIL; |
1235 | goto out; | |
1236 | } | |
1237 | ||
1238 | /* | |
1239 | * We don't try to search for this page in the nonlinear vmas, | |
1240 | * and page_referenced wouldn't have found it anyway. Instead | |
1241 | * just walk the nonlinear vmas trying to age and unmap some. | |
1242 | * The mapcount of the page we came in with is irrelevant, | |
1243 | * but even so use it as a guide to how hard we should try? | |
1244 | */ | |
1245 | mapcount = page_mapcount(page); | |
1246 | if (!mapcount) | |
1247 | goto out; | |
1248 | cond_resched_lock(&mapping->i_mmap_lock); | |
1249 | ||
1250 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
1251 | if (max_nl_cursor == 0) | |
1252 | max_nl_cursor = CLUSTER_SIZE; | |
1253 | ||
1254 | do { | |
1255 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
1256 | shared.vm_set.list) { | |
1da177e4 | 1257 | cursor = (unsigned long) vma->vm_private_data; |
839b9685 | 1258 | while ( cursor < max_nl_cursor && |
1da177e4 | 1259 | cursor < vma->vm_end - vma->vm_start) { |
53f79acb HD |
1260 | if (try_to_unmap_cluster(cursor, &mapcount, |
1261 | vma, page) == SWAP_MLOCK) | |
1262 | ret = SWAP_MLOCK; | |
1da177e4 LT |
1263 | cursor += CLUSTER_SIZE; |
1264 | vma->vm_private_data = (void *) cursor; | |
1265 | if ((int)mapcount <= 0) | |
1266 | goto out; | |
1267 | } | |
1268 | vma->vm_private_data = (void *) max_nl_cursor; | |
1269 | } | |
1270 | cond_resched_lock(&mapping->i_mmap_lock); | |
1271 | max_nl_cursor += CLUSTER_SIZE; | |
1272 | } while (max_nl_cursor <= max_nl_size); | |
1273 | ||
1274 | /* | |
1275 | * Don't loop forever (perhaps all the remaining pages are | |
1276 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
1277 | * vmas, now forgetting on which ones it had fallen behind. | |
1278 | */ | |
101d2be7 HD |
1279 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
1280 | vma->vm_private_data = NULL; | |
1da177e4 LT |
1281 | out: |
1282 | spin_unlock(&mapping->i_mmap_lock); | |
1283 | return ret; | |
1284 | } | |
1285 | ||
1286 | /** | |
1287 | * try_to_unmap - try to remove all page table mappings to a page | |
1288 | * @page: the page to get unmapped | |
14fa31b8 | 1289 | * @flags: action and flags |
1da177e4 LT |
1290 | * |
1291 | * Tries to remove all the page table entries which are mapping this | |
1292 | * page, used in the pageout path. Caller must hold the page lock. | |
1293 | * Return values are: | |
1294 | * | |
1295 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
1296 | * SWAP_AGAIN - we missed a mapping, try again later | |
1297 | * SWAP_FAIL - the page is unswappable | |
b291f000 | 1298 | * SWAP_MLOCK - page is mlocked. |
1da177e4 | 1299 | */ |
14fa31b8 | 1300 | int try_to_unmap(struct page *page, enum ttu_flags flags) |
1da177e4 LT |
1301 | { |
1302 | int ret; | |
1303 | ||
1da177e4 LT |
1304 | BUG_ON(!PageLocked(page)); |
1305 | ||
5ad64688 HD |
1306 | if (unlikely(PageKsm(page))) |
1307 | ret = try_to_unmap_ksm(page, flags); | |
1308 | else if (PageAnon(page)) | |
14fa31b8 | 1309 | ret = try_to_unmap_anon(page, flags); |
1da177e4 | 1310 | else |
14fa31b8 | 1311 | ret = try_to_unmap_file(page, flags); |
b291f000 | 1312 | if (ret != SWAP_MLOCK && !page_mapped(page)) |
1da177e4 LT |
1313 | ret = SWAP_SUCCESS; |
1314 | return ret; | |
1315 | } | |
81b4082d | 1316 | |
b291f000 NP |
1317 | /** |
1318 | * try_to_munlock - try to munlock a page | |
1319 | * @page: the page to be munlocked | |
1320 | * | |
1321 | * Called from munlock code. Checks all of the VMAs mapping the page | |
1322 | * to make sure nobody else has this page mlocked. The page will be | |
1323 | * returned with PG_mlocked cleared if no other vmas have it mlocked. | |
1324 | * | |
1325 | * Return values are: | |
1326 | * | |
53f79acb | 1327 | * SWAP_AGAIN - no vma is holding page mlocked, or, |
b291f000 | 1328 | * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem |
5ad64688 | 1329 | * SWAP_FAIL - page cannot be located at present |
b291f000 NP |
1330 | * SWAP_MLOCK - page is now mlocked. |
1331 | */ | |
1332 | int try_to_munlock(struct page *page) | |
1333 | { | |
1334 | VM_BUG_ON(!PageLocked(page) || PageLRU(page)); | |
1335 | ||
5ad64688 HD |
1336 | if (unlikely(PageKsm(page))) |
1337 | return try_to_unmap_ksm(page, TTU_MUNLOCK); | |
1338 | else if (PageAnon(page)) | |
14fa31b8 | 1339 | return try_to_unmap_anon(page, TTU_MUNLOCK); |
b291f000 | 1340 | else |
14fa31b8 | 1341 | return try_to_unmap_file(page, TTU_MUNLOCK); |
b291f000 | 1342 | } |
e9995ef9 HD |
1343 | |
1344 | #ifdef CONFIG_MIGRATION | |
1345 | /* | |
1346 | * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file(): | |
1347 | * Called by migrate.c to remove migration ptes, but might be used more later. | |
1348 | */ | |
1349 | static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *, | |
1350 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1351 | { | |
1352 | struct anon_vma *anon_vma; | |
5beb4930 | 1353 | struct anon_vma_chain *avc; |
e9995ef9 HD |
1354 | int ret = SWAP_AGAIN; |
1355 | ||
1356 | /* | |
1357 | * Note: remove_migration_ptes() cannot use page_lock_anon_vma() | |
1358 | * because that depends on page_mapped(); but not all its usages | |
1359 | * are holding mmap_sem, which also gave the necessary guarantee | |
1360 | * (that this anon_vma's slab has not already been destroyed). | |
1361 | * This needs to be reviewed later: avoiding page_lock_anon_vma() | |
1362 | * is risky, and currently limits the usefulness of rmap_walk(). | |
1363 | */ | |
1364 | anon_vma = page_anon_vma(page); | |
1365 | if (!anon_vma) | |
1366 | return ret; | |
1367 | spin_lock(&anon_vma->lock); | |
5beb4930 RR |
1368 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { |
1369 | struct vm_area_struct *vma = avc->vma; | |
e9995ef9 HD |
1370 | unsigned long address = vma_address(page, vma); |
1371 | if (address == -EFAULT) | |
1372 | continue; | |
1373 | ret = rmap_one(page, vma, address, arg); | |
1374 | if (ret != SWAP_AGAIN) | |
1375 | break; | |
1376 | } | |
1377 | spin_unlock(&anon_vma->lock); | |
1378 | return ret; | |
1379 | } | |
1380 | ||
1381 | static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *, | |
1382 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1383 | { | |
1384 | struct address_space *mapping = page->mapping; | |
1385 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
1386 | struct vm_area_struct *vma; | |
1387 | struct prio_tree_iter iter; | |
1388 | int ret = SWAP_AGAIN; | |
1389 | ||
1390 | if (!mapping) | |
1391 | return ret; | |
1392 | spin_lock(&mapping->i_mmap_lock); | |
1393 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
1394 | unsigned long address = vma_address(page, vma); | |
1395 | if (address == -EFAULT) | |
1396 | continue; | |
1397 | ret = rmap_one(page, vma, address, arg); | |
1398 | if (ret != SWAP_AGAIN) | |
1399 | break; | |
1400 | } | |
1401 | /* | |
1402 | * No nonlinear handling: being always shared, nonlinear vmas | |
1403 | * never contain migration ptes. Decide what to do about this | |
1404 | * limitation to linear when we need rmap_walk() on nonlinear. | |
1405 | */ | |
1406 | spin_unlock(&mapping->i_mmap_lock); | |
1407 | return ret; | |
1408 | } | |
1409 | ||
1410 | int rmap_walk(struct page *page, int (*rmap_one)(struct page *, | |
1411 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1412 | { | |
1413 | VM_BUG_ON(!PageLocked(page)); | |
1414 | ||
1415 | if (unlikely(PageKsm(page))) | |
1416 | return rmap_walk_ksm(page, rmap_one, arg); | |
1417 | else if (PageAnon(page)) | |
1418 | return rmap_walk_anon(page, rmap_one, arg); | |
1419 | else | |
1420 | return rmap_walk_file(page, rmap_one, arg); | |
1421 | } | |
1422 | #endif /* CONFIG_MIGRATION */ |