Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 LT |
2 | /* |
3 | * linux/mm/madvise.c | |
4 | * | |
5 | * Copyright (C) 1999 Linus Torvalds | |
6 | * Copyright (C) 2002 Christoph Hellwig | |
7 | */ | |
8 | ||
9 | #include <linux/mman.h> | |
10 | #include <linux/pagemap.h> | |
11 | #include <linux/syscalls.h> | |
05b74384 | 12 | #include <linux/mempolicy.h> |
afcf938e | 13 | #include <linux/page-isolation.h> |
05ce7724 | 14 | #include <linux/userfaultfd_k.h> |
1da177e4 | 15 | #include <linux/hugetlb.h> |
3f31d075 | 16 | #include <linux/falloc.h> |
692fe624 | 17 | #include <linux/fadvise.h> |
e8edc6e0 | 18 | #include <linux/sched.h> |
f8af4da3 | 19 | #include <linux/ksm.h> |
3f31d075 | 20 | #include <linux/fs.h> |
9ab4233d | 21 | #include <linux/file.h> |
1998cc04 | 22 | #include <linux/blkdev.h> |
66114cad | 23 | #include <linux/backing-dev.h> |
a520110e | 24 | #include <linux/pagewalk.h> |
1998cc04 SL |
25 | #include <linux/swap.h> |
26 | #include <linux/swapops.h> | |
3a4f8a0b | 27 | #include <linux/shmem_fs.h> |
854e9ed0 MK |
28 | #include <linux/mmu_notifier.h> |
29 | ||
30 | #include <asm/tlb.h> | |
1da177e4 | 31 | |
23519073 KS |
32 | #include "internal.h" |
33 | ||
0a27a14a NP |
34 | /* |
35 | * Any behaviour which results in changes to the vma->vm_flags needs to | |
36 | * take mmap_sem for writing. Others, which simply traverse vmas, need | |
37 | * to only take it for reading. | |
38 | */ | |
39 | static int madvise_need_mmap_write(int behavior) | |
40 | { | |
41 | switch (behavior) { | |
42 | case MADV_REMOVE: | |
43 | case MADV_WILLNEED: | |
44 | case MADV_DONTNEED: | |
854e9ed0 | 45 | case MADV_FREE: |
0a27a14a NP |
46 | return 0; |
47 | default: | |
48 | /* be safe, default to 1. list exceptions explicitly */ | |
49 | return 1; | |
50 | } | |
51 | } | |
52 | ||
1da177e4 LT |
53 | /* |
54 | * We can potentially split a vm area into separate | |
55 | * areas, each area with its own behavior. | |
56 | */ | |
ec9bed9d | 57 | static long madvise_behavior(struct vm_area_struct *vma, |
05b74384 PM |
58 | struct vm_area_struct **prev, |
59 | unsigned long start, unsigned long end, int behavior) | |
1da177e4 | 60 | { |
ec9bed9d | 61 | struct mm_struct *mm = vma->vm_mm; |
1da177e4 | 62 | int error = 0; |
05b74384 | 63 | pgoff_t pgoff; |
3866ea90 | 64 | unsigned long new_flags = vma->vm_flags; |
e798c6e8 PM |
65 | |
66 | switch (behavior) { | |
f8225661 MT |
67 | case MADV_NORMAL: |
68 | new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ; | |
69 | break; | |
e798c6e8 | 70 | case MADV_SEQUENTIAL: |
f8225661 | 71 | new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ; |
e798c6e8 PM |
72 | break; |
73 | case MADV_RANDOM: | |
f8225661 | 74 | new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ; |
e798c6e8 | 75 | break; |
f8225661 MT |
76 | case MADV_DONTFORK: |
77 | new_flags |= VM_DONTCOPY; | |
78 | break; | |
79 | case MADV_DOFORK: | |
3866ea90 HD |
80 | if (vma->vm_flags & VM_IO) { |
81 | error = -EINVAL; | |
82 | goto out; | |
83 | } | |
f8225661 | 84 | new_flags &= ~VM_DONTCOPY; |
e798c6e8 | 85 | break; |
d2cd9ede RR |
86 | case MADV_WIPEONFORK: |
87 | /* MADV_WIPEONFORK is only supported on anonymous memory. */ | |
88 | if (vma->vm_file || vma->vm_flags & VM_SHARED) { | |
89 | error = -EINVAL; | |
90 | goto out; | |
91 | } | |
92 | new_flags |= VM_WIPEONFORK; | |
93 | break; | |
94 | case MADV_KEEPONFORK: | |
95 | new_flags &= ~VM_WIPEONFORK; | |
96 | break; | |
accb61fe | 97 | case MADV_DONTDUMP: |
0103bd16 | 98 | new_flags |= VM_DONTDUMP; |
accb61fe JB |
99 | break; |
100 | case MADV_DODUMP: | |
d41aa525 | 101 | if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) { |
0103bd16 KK |
102 | error = -EINVAL; |
103 | goto out; | |
104 | } | |
105 | new_flags &= ~VM_DONTDUMP; | |
accb61fe | 106 | break; |
f8af4da3 HD |
107 | case MADV_MERGEABLE: |
108 | case MADV_UNMERGEABLE: | |
109 | error = ksm_madvise(vma, start, end, behavior, &new_flags); | |
f3bc0dba MR |
110 | if (error) |
111 | goto out_convert_errno; | |
f8af4da3 | 112 | break; |
0af4e98b | 113 | case MADV_HUGEPAGE: |
a664b2d8 | 114 | case MADV_NOHUGEPAGE: |
60ab3244 | 115 | error = hugepage_madvise(vma, &new_flags, behavior); |
f3bc0dba MR |
116 | if (error) |
117 | goto out_convert_errno; | |
0af4e98b | 118 | break; |
e798c6e8 PM |
119 | } |
120 | ||
05b74384 PM |
121 | if (new_flags == vma->vm_flags) { |
122 | *prev = vma; | |
836d5ffd | 123 | goto out; |
05b74384 PM |
124 | } |
125 | ||
126 | pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); | |
127 | *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma, | |
19a809af AA |
128 | vma->vm_file, pgoff, vma_policy(vma), |
129 | vma->vm_userfaultfd_ctx); | |
05b74384 PM |
130 | if (*prev) { |
131 | vma = *prev; | |
132 | goto success; | |
133 | } | |
134 | ||
135 | *prev = vma; | |
1da177e4 LT |
136 | |
137 | if (start != vma->vm_start) { | |
def5efe0 DR |
138 | if (unlikely(mm->map_count >= sysctl_max_map_count)) { |
139 | error = -ENOMEM; | |
1da177e4 | 140 | goto out; |
def5efe0 DR |
141 | } |
142 | error = __split_vma(mm, vma, start, 1); | |
f3bc0dba MR |
143 | if (error) |
144 | goto out_convert_errno; | |
1da177e4 LT |
145 | } |
146 | ||
147 | if (end != vma->vm_end) { | |
def5efe0 DR |
148 | if (unlikely(mm->map_count >= sysctl_max_map_count)) { |
149 | error = -ENOMEM; | |
1da177e4 | 150 | goto out; |
def5efe0 DR |
151 | } |
152 | error = __split_vma(mm, vma, end, 0); | |
f3bc0dba MR |
153 | if (error) |
154 | goto out_convert_errno; | |
1da177e4 LT |
155 | } |
156 | ||
836d5ffd | 157 | success: |
1da177e4 LT |
158 | /* |
159 | * vm_flags is protected by the mmap_sem held in write mode. | |
160 | */ | |
e798c6e8 | 161 | vma->vm_flags = new_flags; |
f3bc0dba MR |
162 | |
163 | out_convert_errno: | |
164 | /* | |
165 | * madvise() returns EAGAIN if kernel resources, such as | |
166 | * slab, are temporarily unavailable. | |
167 | */ | |
168 | if (error == -ENOMEM) | |
169 | error = -EAGAIN; | |
1da177e4 | 170 | out: |
1da177e4 LT |
171 | return error; |
172 | } | |
173 | ||
1998cc04 SL |
174 | #ifdef CONFIG_SWAP |
175 | static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start, | |
176 | unsigned long end, struct mm_walk *walk) | |
177 | { | |
178 | pte_t *orig_pte; | |
179 | struct vm_area_struct *vma = walk->private; | |
180 | unsigned long index; | |
181 | ||
182 | if (pmd_none_or_trans_huge_or_clear_bad(pmd)) | |
183 | return 0; | |
184 | ||
185 | for (index = start; index != end; index += PAGE_SIZE) { | |
186 | pte_t pte; | |
187 | swp_entry_t entry; | |
188 | struct page *page; | |
189 | spinlock_t *ptl; | |
190 | ||
191 | orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl); | |
192 | pte = *(orig_pte + ((index - start) / PAGE_SIZE)); | |
193 | pte_unmap_unlock(orig_pte, ptl); | |
194 | ||
0661a336 | 195 | if (pte_present(pte) || pte_none(pte)) |
1998cc04 SL |
196 | continue; |
197 | entry = pte_to_swp_entry(pte); | |
198 | if (unlikely(non_swap_entry(entry))) | |
199 | continue; | |
200 | ||
201 | page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE, | |
23955622 | 202 | vma, index, false); |
1998cc04 | 203 | if (page) |
09cbfeaf | 204 | put_page(page); |
1998cc04 SL |
205 | } |
206 | ||
207 | return 0; | |
208 | } | |
209 | ||
7b86ac33 CH |
210 | static const struct mm_walk_ops swapin_walk_ops = { |
211 | .pmd_entry = swapin_walk_pmd_entry, | |
212 | }; | |
1998cc04 SL |
213 | |
214 | static void force_shm_swapin_readahead(struct vm_area_struct *vma, | |
215 | unsigned long start, unsigned long end, | |
216 | struct address_space *mapping) | |
217 | { | |
218 | pgoff_t index; | |
219 | struct page *page; | |
220 | swp_entry_t swap; | |
221 | ||
222 | for (; start < end; start += PAGE_SIZE) { | |
223 | index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
224 | ||
55231e5c | 225 | page = find_get_entry(mapping, index); |
3159f943 | 226 | if (!xa_is_value(page)) { |
1998cc04 | 227 | if (page) |
09cbfeaf | 228 | put_page(page); |
1998cc04 SL |
229 | continue; |
230 | } | |
231 | swap = radix_to_swp_entry(page); | |
232 | page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE, | |
23955622 | 233 | NULL, 0, false); |
1998cc04 | 234 | if (page) |
09cbfeaf | 235 | put_page(page); |
1998cc04 SL |
236 | } |
237 | ||
238 | lru_add_drain(); /* Push any new pages onto the LRU now */ | |
239 | } | |
240 | #endif /* CONFIG_SWAP */ | |
241 | ||
1da177e4 LT |
242 | /* |
243 | * Schedule all required I/O operations. Do not wait for completion. | |
244 | */ | |
ec9bed9d VC |
245 | static long madvise_willneed(struct vm_area_struct *vma, |
246 | struct vm_area_struct **prev, | |
1da177e4 LT |
247 | unsigned long start, unsigned long end) |
248 | { | |
249 | struct file *file = vma->vm_file; | |
692fe624 | 250 | loff_t offset; |
1da177e4 | 251 | |
6ea8d958 | 252 | *prev = vma; |
1998cc04 | 253 | #ifdef CONFIG_SWAP |
97b713ba | 254 | if (!file) { |
7b86ac33 CH |
255 | walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma); |
256 | lru_add_drain(); /* Push any new pages onto the LRU now */ | |
1998cc04 SL |
257 | return 0; |
258 | } | |
1998cc04 | 259 | |
97b713ba | 260 | if (shmem_mapping(file->f_mapping)) { |
97b713ba CH |
261 | force_shm_swapin_readahead(vma, start, end, |
262 | file->f_mapping); | |
263 | return 0; | |
264 | } | |
265 | #else | |
1bef4003 S |
266 | if (!file) |
267 | return -EBADF; | |
97b713ba | 268 | #endif |
1bef4003 | 269 | |
e748dcd0 | 270 | if (IS_DAX(file_inode(file))) { |
fe77ba6f CO |
271 | /* no bad return value, but ignore advice */ |
272 | return 0; | |
273 | } | |
274 | ||
692fe624 JK |
275 | /* |
276 | * Filesystem's fadvise may need to take various locks. We need to | |
277 | * explicitly grab a reference because the vma (and hence the | |
278 | * vma's reference to the file) can go away as soon as we drop | |
279 | * mmap_sem. | |
280 | */ | |
281 | *prev = NULL; /* tell sys_madvise we drop mmap_sem */ | |
282 | get_file(file); | |
283 | up_read(¤t->mm->mmap_sem); | |
284 | offset = (loff_t)(start - vma->vm_start) | |
285 | + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); | |
286 | vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED); | |
287 | fput(file); | |
288 | down_read(¤t->mm->mmap_sem); | |
1da177e4 LT |
289 | return 0; |
290 | } | |
291 | ||
854e9ed0 MK |
292 | static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr, |
293 | unsigned long end, struct mm_walk *walk) | |
294 | ||
295 | { | |
296 | struct mmu_gather *tlb = walk->private; | |
297 | struct mm_struct *mm = tlb->mm; | |
298 | struct vm_area_struct *vma = walk->vma; | |
299 | spinlock_t *ptl; | |
300 | pte_t *orig_pte, *pte, ptent; | |
301 | struct page *page; | |
64b42bc1 | 302 | int nr_swap = 0; |
b8d3c4c3 MK |
303 | unsigned long next; |
304 | ||
305 | next = pmd_addr_end(addr, end); | |
306 | if (pmd_trans_huge(*pmd)) | |
307 | if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next)) | |
308 | goto next; | |
854e9ed0 | 309 | |
854e9ed0 MK |
310 | if (pmd_trans_unstable(pmd)) |
311 | return 0; | |
312 | ||
ed6a7935 | 313 | tlb_change_page_size(tlb, PAGE_SIZE); |
854e9ed0 | 314 | orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl); |
3ea27719 | 315 | flush_tlb_batched_pending(mm); |
854e9ed0 MK |
316 | arch_enter_lazy_mmu_mode(); |
317 | for (; addr != end; pte++, addr += PAGE_SIZE) { | |
318 | ptent = *pte; | |
319 | ||
64b42bc1 | 320 | if (pte_none(ptent)) |
854e9ed0 | 321 | continue; |
64b42bc1 MK |
322 | /* |
323 | * If the pte has swp_entry, just clear page table to | |
324 | * prevent swap-in which is more expensive rather than | |
325 | * (page allocation + zeroing). | |
326 | */ | |
327 | if (!pte_present(ptent)) { | |
328 | swp_entry_t entry; | |
329 | ||
330 | entry = pte_to_swp_entry(ptent); | |
331 | if (non_swap_entry(entry)) | |
332 | continue; | |
333 | nr_swap--; | |
334 | free_swap_and_cache(entry); | |
335 | pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); | |
336 | continue; | |
337 | } | |
854e9ed0 | 338 | |
25b2995a | 339 | page = vm_normal_page(vma, addr, ptent); |
854e9ed0 MK |
340 | if (!page) |
341 | continue; | |
342 | ||
343 | /* | |
344 | * If pmd isn't transhuge but the page is THP and | |
345 | * is owned by only this process, split it and | |
346 | * deactivate all pages. | |
347 | */ | |
348 | if (PageTransCompound(page)) { | |
349 | if (page_mapcount(page) != 1) | |
350 | goto out; | |
351 | get_page(page); | |
352 | if (!trylock_page(page)) { | |
353 | put_page(page); | |
354 | goto out; | |
355 | } | |
356 | pte_unmap_unlock(orig_pte, ptl); | |
357 | if (split_huge_page(page)) { | |
358 | unlock_page(page); | |
359 | put_page(page); | |
360 | pte_offset_map_lock(mm, pmd, addr, &ptl); | |
361 | goto out; | |
362 | } | |
854e9ed0 | 363 | unlock_page(page); |
263630e8 | 364 | put_page(page); |
854e9ed0 MK |
365 | pte = pte_offset_map_lock(mm, pmd, addr, &ptl); |
366 | pte--; | |
367 | addr -= PAGE_SIZE; | |
368 | continue; | |
369 | } | |
370 | ||
371 | VM_BUG_ON_PAGE(PageTransCompound(page), page); | |
372 | ||
373 | if (PageSwapCache(page) || PageDirty(page)) { | |
374 | if (!trylock_page(page)) | |
375 | continue; | |
376 | /* | |
377 | * If page is shared with others, we couldn't clear | |
378 | * PG_dirty of the page. | |
379 | */ | |
380 | if (page_mapcount(page) != 1) { | |
381 | unlock_page(page); | |
382 | continue; | |
383 | } | |
384 | ||
385 | if (PageSwapCache(page) && !try_to_free_swap(page)) { | |
386 | unlock_page(page); | |
387 | continue; | |
388 | } | |
389 | ||
390 | ClearPageDirty(page); | |
391 | unlock_page(page); | |
392 | } | |
393 | ||
394 | if (pte_young(ptent) || pte_dirty(ptent)) { | |
395 | /* | |
396 | * Some of architecture(ex, PPC) don't update TLB | |
397 | * with set_pte_at and tlb_remove_tlb_entry so for | |
398 | * the portability, remap the pte with old|clean | |
399 | * after pte clearing. | |
400 | */ | |
401 | ptent = ptep_get_and_clear_full(mm, addr, pte, | |
402 | tlb->fullmm); | |
403 | ||
404 | ptent = pte_mkold(ptent); | |
405 | ptent = pte_mkclean(ptent); | |
406 | set_pte_at(mm, addr, pte, ptent); | |
407 | tlb_remove_tlb_entry(tlb, pte, addr); | |
408 | } | |
802a3a92 | 409 | mark_page_lazyfree(page); |
854e9ed0 MK |
410 | } |
411 | out: | |
64b42bc1 MK |
412 | if (nr_swap) { |
413 | if (current->mm == mm) | |
414 | sync_mm_rss(mm); | |
415 | ||
416 | add_mm_counter(mm, MM_SWAPENTS, nr_swap); | |
417 | } | |
854e9ed0 MK |
418 | arch_leave_lazy_mmu_mode(); |
419 | pte_unmap_unlock(orig_pte, ptl); | |
420 | cond_resched(); | |
b8d3c4c3 | 421 | next: |
854e9ed0 MK |
422 | return 0; |
423 | } | |
424 | ||
7b86ac33 CH |
425 | static const struct mm_walk_ops madvise_free_walk_ops = { |
426 | .pmd_entry = madvise_free_pte_range, | |
427 | }; | |
854e9ed0 MK |
428 | |
429 | static int madvise_free_single_vma(struct vm_area_struct *vma, | |
430 | unsigned long start_addr, unsigned long end_addr) | |
431 | { | |
854e9ed0 | 432 | struct mm_struct *mm = vma->vm_mm; |
ac46d4f3 | 433 | struct mmu_notifier_range range; |
854e9ed0 MK |
434 | struct mmu_gather tlb; |
435 | ||
854e9ed0 MK |
436 | /* MADV_FREE works for only anon vma at the moment */ |
437 | if (!vma_is_anonymous(vma)) | |
438 | return -EINVAL; | |
439 | ||
ac46d4f3 JG |
440 | range.start = max(vma->vm_start, start_addr); |
441 | if (range.start >= vma->vm_end) | |
854e9ed0 | 442 | return -EINVAL; |
ac46d4f3 JG |
443 | range.end = min(vma->vm_end, end_addr); |
444 | if (range.end <= vma->vm_start) | |
854e9ed0 | 445 | return -EINVAL; |
7269f999 | 446 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, |
6f4f13e8 | 447 | range.start, range.end); |
854e9ed0 MK |
448 | |
449 | lru_add_drain(); | |
ac46d4f3 | 450 | tlb_gather_mmu(&tlb, mm, range.start, range.end); |
854e9ed0 MK |
451 | update_hiwater_rss(mm); |
452 | ||
ac46d4f3 | 453 | mmu_notifier_invalidate_range_start(&range); |
7b86ac33 CH |
454 | tlb_start_vma(&tlb, vma); |
455 | walk_page_range(vma->vm_mm, range.start, range.end, | |
456 | &madvise_free_walk_ops, &tlb); | |
457 | tlb_end_vma(&tlb, vma); | |
ac46d4f3 JG |
458 | mmu_notifier_invalidate_range_end(&range); |
459 | tlb_finish_mmu(&tlb, range.start, range.end); | |
854e9ed0 MK |
460 | |
461 | return 0; | |
462 | } | |
463 | ||
1da177e4 LT |
464 | /* |
465 | * Application no longer needs these pages. If the pages are dirty, | |
466 | * it's OK to just throw them away. The app will be more careful about | |
467 | * data it wants to keep. Be sure to free swap resources too. The | |
7e6cbea3 | 468 | * zap_page_range call sets things up for shrink_active_list to actually free |
1da177e4 LT |
469 | * these pages later if no one else has touched them in the meantime, |
470 | * although we could add these pages to a global reuse list for | |
7e6cbea3 | 471 | * shrink_active_list to pick up before reclaiming other pages. |
1da177e4 LT |
472 | * |
473 | * NB: This interface discards data rather than pushes it out to swap, | |
474 | * as some implementations do. This has performance implications for | |
475 | * applications like large transactional databases which want to discard | |
476 | * pages in anonymous maps after committing to backing store the data | |
477 | * that was kept in them. There is no reason to write this data out to | |
478 | * the swap area if the application is discarding it. | |
479 | * | |
480 | * An interface that causes the system to free clean pages and flush | |
481 | * dirty pages is already available as msync(MS_INVALIDATE). | |
482 | */ | |
230ca982 MR |
483 | static long madvise_dontneed_single_vma(struct vm_area_struct *vma, |
484 | unsigned long start, unsigned long end) | |
485 | { | |
486 | zap_page_range(vma, start, end - start); | |
487 | return 0; | |
488 | } | |
489 | ||
490 | static long madvise_dontneed_free(struct vm_area_struct *vma, | |
491 | struct vm_area_struct **prev, | |
492 | unsigned long start, unsigned long end, | |
493 | int behavior) | |
1da177e4 | 494 | { |
05b74384 | 495 | *prev = vma; |
23519073 | 496 | if (!can_madv_dontneed_vma(vma)) |
1da177e4 LT |
497 | return -EINVAL; |
498 | ||
70ccb92f AA |
499 | if (!userfaultfd_remove(vma, start, end)) { |
500 | *prev = NULL; /* mmap_sem has been dropped, prev is stale */ | |
501 | ||
502 | down_read(¤t->mm->mmap_sem); | |
503 | vma = find_vma(current->mm, start); | |
504 | if (!vma) | |
505 | return -ENOMEM; | |
506 | if (start < vma->vm_start) { | |
507 | /* | |
508 | * This "vma" under revalidation is the one | |
509 | * with the lowest vma->vm_start where start | |
510 | * is also < vma->vm_end. If start < | |
511 | * vma->vm_start it means an hole materialized | |
512 | * in the user address space within the | |
230ca982 MR |
513 | * virtual range passed to MADV_DONTNEED |
514 | * or MADV_FREE. | |
70ccb92f AA |
515 | */ |
516 | return -ENOMEM; | |
517 | } | |
518 | if (!can_madv_dontneed_vma(vma)) | |
519 | return -EINVAL; | |
520 | if (end > vma->vm_end) { | |
521 | /* | |
522 | * Don't fail if end > vma->vm_end. If the old | |
523 | * vma was splitted while the mmap_sem was | |
524 | * released the effect of the concurrent | |
230ca982 | 525 | * operation may not cause madvise() to |
70ccb92f AA |
526 | * have an undefined result. There may be an |
527 | * adjacent next vma that we'll walk | |
528 | * next. userfaultfd_remove() will generate an | |
529 | * UFFD_EVENT_REMOVE repetition on the | |
530 | * end-vma->vm_end range, but the manager can | |
531 | * handle a repetition fine. | |
532 | */ | |
533 | end = vma->vm_end; | |
534 | } | |
535 | VM_WARN_ON(start >= end); | |
536 | } | |
230ca982 MR |
537 | |
538 | if (behavior == MADV_DONTNEED) | |
539 | return madvise_dontneed_single_vma(vma, start, end); | |
540 | else if (behavior == MADV_FREE) | |
541 | return madvise_free_single_vma(vma, start, end); | |
542 | else | |
543 | return -EINVAL; | |
1da177e4 LT |
544 | } |
545 | ||
f6b3ec23 BP |
546 | /* |
547 | * Application wants to free up the pages and associated backing store. | |
548 | * This is effectively punching a hole into the middle of a file. | |
f6b3ec23 BP |
549 | */ |
550 | static long madvise_remove(struct vm_area_struct *vma, | |
00e9fa2d | 551 | struct vm_area_struct **prev, |
f6b3ec23 BP |
552 | unsigned long start, unsigned long end) |
553 | { | |
3f31d075 | 554 | loff_t offset; |
90ed52eb | 555 | int error; |
9ab4233d | 556 | struct file *f; |
f6b3ec23 | 557 | |
90ed52eb | 558 | *prev = NULL; /* tell sys_madvise we drop mmap_sem */ |
00e9fa2d | 559 | |
72079ba0 | 560 | if (vma->vm_flags & VM_LOCKED) |
f6b3ec23 BP |
561 | return -EINVAL; |
562 | ||
9ab4233d AL |
563 | f = vma->vm_file; |
564 | ||
565 | if (!f || !f->f_mapping || !f->f_mapping->host) { | |
f6b3ec23 BP |
566 | return -EINVAL; |
567 | } | |
568 | ||
69cf0fac HD |
569 | if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE)) |
570 | return -EACCES; | |
571 | ||
f6b3ec23 BP |
572 | offset = (loff_t)(start - vma->vm_start) |
573 | + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); | |
90ed52eb | 574 | |
9ab4233d AL |
575 | /* |
576 | * Filesystem's fallocate may need to take i_mutex. We need to | |
577 | * explicitly grab a reference because the vma (and hence the | |
578 | * vma's reference to the file) can go away as soon as we drop | |
579 | * mmap_sem. | |
580 | */ | |
581 | get_file(f); | |
70ccb92f AA |
582 | if (userfaultfd_remove(vma, start, end)) { |
583 | /* mmap_sem was not released by userfaultfd_remove() */ | |
584 | up_read(¤t->mm->mmap_sem); | |
585 | } | |
72c72bdf | 586 | error = vfs_fallocate(f, |
3f31d075 HD |
587 | FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, |
588 | offset, end - start); | |
9ab4233d | 589 | fput(f); |
0a27a14a | 590 | down_read(¤t->mm->mmap_sem); |
90ed52eb | 591 | return error; |
f6b3ec23 BP |
592 | } |
593 | ||
9893e49d AK |
594 | #ifdef CONFIG_MEMORY_FAILURE |
595 | /* | |
596 | * Error injection support for memory error handling. | |
597 | */ | |
97167a76 AK |
598 | static int madvise_inject_error(int behavior, |
599 | unsigned long start, unsigned long end) | |
9893e49d | 600 | { |
97167a76 | 601 | struct page *page; |
c461ad6a | 602 | struct zone *zone; |
19bfbe22 | 603 | unsigned int order; |
97167a76 | 604 | |
9893e49d AK |
605 | if (!capable(CAP_SYS_ADMIN)) |
606 | return -EPERM; | |
97167a76 | 607 | |
19bfbe22 AM |
608 | |
609 | for (; start < end; start += PAGE_SIZE << order) { | |
23e7b5c2 | 610 | unsigned long pfn; |
325c4ef5 AM |
611 | int ret; |
612 | ||
97167a76 | 613 | ret = get_user_pages_fast(start, 1, 0, &page); |
9893e49d AK |
614 | if (ret != 1) |
615 | return ret; | |
23e7b5c2 | 616 | pfn = page_to_pfn(page); |
325c4ef5 | 617 | |
19bfbe22 AM |
618 | /* |
619 | * When soft offlining hugepages, after migrating the page | |
620 | * we dissolve it, therefore in the second loop "page" will | |
621 | * no longer be a compound page, and order will be 0. | |
622 | */ | |
623 | order = compound_order(compound_head(page)); | |
624 | ||
97167a76 AK |
625 | if (PageHWPoison(page)) { |
626 | put_page(page); | |
29b4eede WL |
627 | continue; |
628 | } | |
97167a76 AK |
629 | |
630 | if (behavior == MADV_SOFT_OFFLINE) { | |
631 | pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n", | |
23e7b5c2 | 632 | pfn, start); |
97167a76 AK |
633 | |
634 | ret = soft_offline_page(page, MF_COUNT_INCREASED); | |
afcf938e | 635 | if (ret) |
8302423b | 636 | return ret; |
afcf938e AK |
637 | continue; |
638 | } | |
23e7b5c2 | 639 | |
97167a76 | 640 | pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n", |
23e7b5c2 | 641 | pfn, start); |
97167a76 | 642 | |
23e7b5c2 DW |
643 | /* |
644 | * Drop the page reference taken by get_user_pages_fast(). In | |
645 | * the absence of MF_COUNT_INCREASED the memory_failure() | |
646 | * routine is responsible for pinning the page to prevent it | |
647 | * from being released back to the page allocator. | |
648 | */ | |
649 | put_page(page); | |
650 | ret = memory_failure(pfn, 0); | |
23a003bf NH |
651 | if (ret) |
652 | return ret; | |
9893e49d | 653 | } |
c461ad6a MG |
654 | |
655 | /* Ensure that all poisoned pages are removed from per-cpu lists */ | |
656 | for_each_populated_zone(zone) | |
657 | drain_all_pages(zone); | |
658 | ||
325c4ef5 | 659 | return 0; |
9893e49d AK |
660 | } |
661 | #endif | |
662 | ||
165cd402 | 663 | static long |
664 | madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev, | |
665 | unsigned long start, unsigned long end, int behavior) | |
1da177e4 | 666 | { |
1da177e4 | 667 | switch (behavior) { |
f6b3ec23 | 668 | case MADV_REMOVE: |
3866ea90 | 669 | return madvise_remove(vma, prev, start, end); |
1da177e4 | 670 | case MADV_WILLNEED: |
3866ea90 | 671 | return madvise_willneed(vma, prev, start, end); |
854e9ed0 | 672 | case MADV_FREE: |
1da177e4 | 673 | case MADV_DONTNEED: |
230ca982 | 674 | return madvise_dontneed_free(vma, prev, start, end, behavior); |
1da177e4 | 675 | default: |
3866ea90 | 676 | return madvise_behavior(vma, prev, start, end, behavior); |
1da177e4 | 677 | } |
1da177e4 LT |
678 | } |
679 | ||
1ecef9ed | 680 | static bool |
75927af8 NP |
681 | madvise_behavior_valid(int behavior) |
682 | { | |
683 | switch (behavior) { | |
684 | case MADV_DOFORK: | |
685 | case MADV_DONTFORK: | |
686 | case MADV_NORMAL: | |
687 | case MADV_SEQUENTIAL: | |
688 | case MADV_RANDOM: | |
689 | case MADV_REMOVE: | |
690 | case MADV_WILLNEED: | |
691 | case MADV_DONTNEED: | |
854e9ed0 | 692 | case MADV_FREE: |
f8af4da3 HD |
693 | #ifdef CONFIG_KSM |
694 | case MADV_MERGEABLE: | |
695 | case MADV_UNMERGEABLE: | |
0af4e98b AA |
696 | #endif |
697 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
698 | case MADV_HUGEPAGE: | |
a664b2d8 | 699 | case MADV_NOHUGEPAGE: |
f8af4da3 | 700 | #endif |
accb61fe JB |
701 | case MADV_DONTDUMP: |
702 | case MADV_DODUMP: | |
d2cd9ede RR |
703 | case MADV_WIPEONFORK: |
704 | case MADV_KEEPONFORK: | |
5e451be7 AK |
705 | #ifdef CONFIG_MEMORY_FAILURE |
706 | case MADV_SOFT_OFFLINE: | |
707 | case MADV_HWPOISON: | |
708 | #endif | |
1ecef9ed | 709 | return true; |
75927af8 NP |
710 | |
711 | default: | |
1ecef9ed | 712 | return false; |
75927af8 NP |
713 | } |
714 | } | |
3866ea90 | 715 | |
1da177e4 LT |
716 | /* |
717 | * The madvise(2) system call. | |
718 | * | |
719 | * Applications can use madvise() to advise the kernel how it should | |
720 | * handle paging I/O in this VM area. The idea is to help the kernel | |
721 | * use appropriate read-ahead and caching techniques. The information | |
722 | * provided is advisory only, and can be safely disregarded by the | |
723 | * kernel without affecting the correct operation of the application. | |
724 | * | |
725 | * behavior values: | |
726 | * MADV_NORMAL - the default behavior is to read clusters. This | |
727 | * results in some read-ahead and read-behind. | |
728 | * MADV_RANDOM - the system should read the minimum amount of data | |
729 | * on any access, since it is unlikely that the appli- | |
730 | * cation will need more than what it asks for. | |
731 | * MADV_SEQUENTIAL - pages in the given range will probably be accessed | |
732 | * once, so they can be aggressively read ahead, and | |
733 | * can be freed soon after they are accessed. | |
734 | * MADV_WILLNEED - the application is notifying the system to read | |
735 | * some pages ahead. | |
736 | * MADV_DONTNEED - the application is finished with the given range, | |
737 | * so the kernel can free resources associated with it. | |
d7206a70 NH |
738 | * MADV_FREE - the application marks pages in the given range as lazy free, |
739 | * where actual purges are postponed until memory pressure happens. | |
f6b3ec23 BP |
740 | * MADV_REMOVE - the application wants to free up the given range of |
741 | * pages and associated backing store. | |
3866ea90 HD |
742 | * MADV_DONTFORK - omit this area from child's address space when forking: |
743 | * typically, to avoid COWing pages pinned by get_user_pages(). | |
744 | * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking. | |
c02c3009 YS |
745 | * MADV_WIPEONFORK - present the child process with zero-filled memory in this |
746 | * range after a fork. | |
747 | * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK | |
d7206a70 NH |
748 | * MADV_HWPOISON - trigger memory error handler as if the given memory range |
749 | * were corrupted by unrecoverable hardware memory failure. | |
750 | * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory. | |
f8af4da3 HD |
751 | * MADV_MERGEABLE - the application recommends that KSM try to merge pages in |
752 | * this area with pages of identical content from other such areas. | |
753 | * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others. | |
d7206a70 NH |
754 | * MADV_HUGEPAGE - the application wants to back the given range by transparent |
755 | * huge pages in the future. Existing pages might be coalesced and | |
756 | * new pages might be allocated as THP. | |
757 | * MADV_NOHUGEPAGE - mark the given range as not worth being backed by | |
758 | * transparent huge pages so the existing pages will not be | |
759 | * coalesced into THP and new pages will not be allocated as THP. | |
760 | * MADV_DONTDUMP - the application wants to prevent pages in the given range | |
761 | * from being included in its core dump. | |
762 | * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump. | |
1da177e4 LT |
763 | * |
764 | * return values: | |
765 | * zero - success | |
766 | * -EINVAL - start + len < 0, start is not page-aligned, | |
767 | * "behavior" is not a valid value, or application | |
c02c3009 YS |
768 | * is attempting to release locked or shared pages, |
769 | * or the specified address range includes file, Huge TLB, | |
770 | * MAP_SHARED or VMPFNMAP range. | |
1da177e4 LT |
771 | * -ENOMEM - addresses in the specified range are not currently |
772 | * mapped, or are outside the AS of the process. | |
773 | * -EIO - an I/O error occurred while paging in data. | |
774 | * -EBADF - map exists, but area maps something that isn't a file. | |
775 | * -EAGAIN - a kernel resource was temporarily unavailable. | |
776 | */ | |
3480b257 | 777 | SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior) |
1da177e4 | 778 | { |
05b74384 | 779 | unsigned long end, tmp; |
ec9bed9d | 780 | struct vm_area_struct *vma, *prev; |
1da177e4 LT |
781 | int unmapped_error = 0; |
782 | int error = -EINVAL; | |
f7977793 | 783 | int write; |
1da177e4 | 784 | size_t len; |
1998cc04 | 785 | struct blk_plug plug; |
1da177e4 | 786 | |
75927af8 NP |
787 | if (!madvise_behavior_valid(behavior)) |
788 | return error; | |
789 | ||
1da177e4 | 790 | if (start & ~PAGE_MASK) |
84d96d89 | 791 | return error; |
1da177e4 LT |
792 | len = (len_in + ~PAGE_MASK) & PAGE_MASK; |
793 | ||
794 | /* Check to see whether len was rounded up from small -ve to zero */ | |
795 | if (len_in && !len) | |
84d96d89 | 796 | return error; |
1da177e4 LT |
797 | |
798 | end = start + len; | |
799 | if (end < start) | |
84d96d89 | 800 | return error; |
1da177e4 LT |
801 | |
802 | error = 0; | |
803 | if (end == start) | |
84d96d89 RV |
804 | return error; |
805 | ||
5e451be7 AK |
806 | #ifdef CONFIG_MEMORY_FAILURE |
807 | if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE) | |
808 | return madvise_inject_error(behavior, start, start + len_in); | |
809 | #endif | |
810 | ||
84d96d89 | 811 | write = madvise_need_mmap_write(behavior); |
dc0ef0df MH |
812 | if (write) { |
813 | if (down_write_killable(¤t->mm->mmap_sem)) | |
814 | return -EINTR; | |
815 | } else { | |
84d96d89 | 816 | down_read(¤t->mm->mmap_sem); |
dc0ef0df | 817 | } |
1da177e4 LT |
818 | |
819 | /* | |
820 | * If the interval [start,end) covers some unmapped address | |
821 | * ranges, just ignore them, but return -ENOMEM at the end. | |
05b74384 | 822 | * - different from the way of handling in mlock etc. |
1da177e4 | 823 | */ |
05b74384 | 824 | vma = find_vma_prev(current->mm, start, &prev); |
836d5ffd HD |
825 | if (vma && start > vma->vm_start) |
826 | prev = vma; | |
827 | ||
1998cc04 | 828 | blk_start_plug(&plug); |
1da177e4 LT |
829 | for (;;) { |
830 | /* Still start < end. */ | |
831 | error = -ENOMEM; | |
832 | if (!vma) | |
84d96d89 | 833 | goto out; |
1da177e4 | 834 | |
05b74384 | 835 | /* Here start < (end|vma->vm_end). */ |
1da177e4 LT |
836 | if (start < vma->vm_start) { |
837 | unmapped_error = -ENOMEM; | |
838 | start = vma->vm_start; | |
05b74384 | 839 | if (start >= end) |
84d96d89 | 840 | goto out; |
1da177e4 LT |
841 | } |
842 | ||
05b74384 PM |
843 | /* Here vma->vm_start <= start < (end|vma->vm_end) */ |
844 | tmp = vma->vm_end; | |
845 | if (end < tmp) | |
846 | tmp = end; | |
1da177e4 | 847 | |
05b74384 PM |
848 | /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */ |
849 | error = madvise_vma(vma, &prev, start, tmp, behavior); | |
1da177e4 | 850 | if (error) |
84d96d89 | 851 | goto out; |
05b74384 | 852 | start = tmp; |
90ed52eb | 853 | if (prev && start < prev->vm_end) |
05b74384 PM |
854 | start = prev->vm_end; |
855 | error = unmapped_error; | |
856 | if (start >= end) | |
84d96d89 | 857 | goto out; |
90ed52eb HD |
858 | if (prev) |
859 | vma = prev->vm_next; | |
860 | else /* madvise_remove dropped mmap_sem */ | |
861 | vma = find_vma(current->mm, start); | |
1da177e4 | 862 | } |
1da177e4 | 863 | out: |
84d96d89 | 864 | blk_finish_plug(&plug); |
f7977793 | 865 | if (write) |
0a27a14a NP |
866 | up_write(¤t->mm->mmap_sem); |
867 | else | |
868 | up_read(¤t->mm->mmap_sem); | |
869 | ||
1da177e4 LT |
870 | return error; |
871 | } |