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