]>
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> |
e8edc6e0 | 17 | #include <linux/sched.h> |
f8af4da3 | 18 | #include <linux/ksm.h> |
3f31d075 | 19 | #include <linux/fs.h> |
9ab4233d | 20 | #include <linux/file.h> |
1998cc04 | 21 | #include <linux/blkdev.h> |
66114cad | 22 | #include <linux/backing-dev.h> |
1998cc04 SL |
23 | #include <linux/swap.h> |
24 | #include <linux/swapops.h> | |
3a4f8a0b | 25 | #include <linux/shmem_fs.h> |
854e9ed0 MK |
26 | #include <linux/mmu_notifier.h> |
27 | ||
28 | #include <asm/tlb.h> | |
1da177e4 | 29 | |
23519073 KS |
30 | #include "internal.h" |
31 | ||
0a27a14a NP |
32 | /* |
33 | * Any behaviour which results in changes to the vma->vm_flags needs to | |
34 | * take mmap_sem for writing. Others, which simply traverse vmas, need | |
35 | * to only take it for reading. | |
36 | */ | |
37 | static int madvise_need_mmap_write(int behavior) | |
38 | { | |
39 | switch (behavior) { | |
40 | case MADV_REMOVE: | |
41 | case MADV_WILLNEED: | |
42 | case MADV_DONTNEED: | |
854e9ed0 | 43 | case MADV_FREE: |
0a27a14a NP |
44 | return 0; |
45 | default: | |
46 | /* be safe, default to 1. list exceptions explicitly */ | |
47 | return 1; | |
48 | } | |
49 | } | |
50 | ||
1da177e4 LT |
51 | /* |
52 | * We can potentially split a vm area into separate | |
53 | * areas, each area with its own behavior. | |
54 | */ | |
ec9bed9d | 55 | static long madvise_behavior(struct vm_area_struct *vma, |
05b74384 PM |
56 | struct vm_area_struct **prev, |
57 | unsigned long start, unsigned long end, int behavior) | |
1da177e4 | 58 | { |
ec9bed9d | 59 | struct mm_struct *mm = vma->vm_mm; |
1da177e4 | 60 | int error = 0; |
05b74384 | 61 | pgoff_t pgoff; |
3866ea90 | 62 | unsigned long new_flags = vma->vm_flags; |
e798c6e8 PM |
63 | |
64 | switch (behavior) { | |
f8225661 MT |
65 | case MADV_NORMAL: |
66 | new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ; | |
67 | break; | |
e798c6e8 | 68 | case MADV_SEQUENTIAL: |
f8225661 | 69 | new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ; |
e798c6e8 PM |
70 | break; |
71 | case MADV_RANDOM: | |
f8225661 | 72 | new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ; |
e798c6e8 | 73 | break; |
f8225661 MT |
74 | case MADV_DONTFORK: |
75 | new_flags |= VM_DONTCOPY; | |
76 | break; | |
77 | case MADV_DOFORK: | |
3866ea90 HD |
78 | if (vma->vm_flags & VM_IO) { |
79 | error = -EINVAL; | |
80 | goto out; | |
81 | } | |
f8225661 | 82 | new_flags &= ~VM_DONTCOPY; |
e798c6e8 | 83 | break; |
d2cd9ede RR |
84 | case MADV_WIPEONFORK: |
85 | /* MADV_WIPEONFORK is only supported on anonymous memory. */ | |
86 | if (vma->vm_file || vma->vm_flags & VM_SHARED) { | |
87 | error = -EINVAL; | |
88 | goto out; | |
89 | } | |
90 | new_flags |= VM_WIPEONFORK; | |
91 | break; | |
92 | case MADV_KEEPONFORK: | |
93 | new_flags &= ~VM_WIPEONFORK; | |
94 | break; | |
accb61fe | 95 | case MADV_DONTDUMP: |
0103bd16 | 96 | new_flags |= VM_DONTDUMP; |
accb61fe JB |
97 | break; |
98 | case MADV_DODUMP: | |
0103bd16 KK |
99 | if (new_flags & VM_SPECIAL) { |
100 | error = -EINVAL; | |
101 | goto out; | |
102 | } | |
103 | new_flags &= ~VM_DONTDUMP; | |
accb61fe | 104 | break; |
f8af4da3 HD |
105 | case MADV_MERGEABLE: |
106 | case MADV_UNMERGEABLE: | |
107 | error = ksm_madvise(vma, start, end, behavior, &new_flags); | |
def5efe0 DR |
108 | if (error) { |
109 | /* | |
110 | * madvise() returns EAGAIN if kernel resources, such as | |
111 | * slab, are temporarily unavailable. | |
112 | */ | |
113 | if (error == -ENOMEM) | |
114 | error = -EAGAIN; | |
f8af4da3 | 115 | goto out; |
def5efe0 | 116 | } |
f8af4da3 | 117 | break; |
0af4e98b | 118 | case MADV_HUGEPAGE: |
a664b2d8 | 119 | case MADV_NOHUGEPAGE: |
60ab3244 | 120 | error = hugepage_madvise(vma, &new_flags, behavior); |
def5efe0 DR |
121 | if (error) { |
122 | /* | |
123 | * madvise() returns EAGAIN if kernel resources, such as | |
124 | * slab, are temporarily unavailable. | |
125 | */ | |
126 | if (error == -ENOMEM) | |
127 | error = -EAGAIN; | |
0af4e98b | 128 | goto out; |
def5efe0 | 129 | } |
0af4e98b | 130 | break; |
e798c6e8 PM |
131 | } |
132 | ||
05b74384 PM |
133 | if (new_flags == vma->vm_flags) { |
134 | *prev = vma; | |
836d5ffd | 135 | goto out; |
05b74384 PM |
136 | } |
137 | ||
138 | pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); | |
139 | *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma, | |
19a809af AA |
140 | vma->vm_file, pgoff, vma_policy(vma), |
141 | vma->vm_userfaultfd_ctx); | |
05b74384 PM |
142 | if (*prev) { |
143 | vma = *prev; | |
144 | goto success; | |
145 | } | |
146 | ||
147 | *prev = vma; | |
1da177e4 LT |
148 | |
149 | if (start != vma->vm_start) { | |
def5efe0 DR |
150 | if (unlikely(mm->map_count >= sysctl_max_map_count)) { |
151 | error = -ENOMEM; | |
1da177e4 | 152 | goto out; |
def5efe0 DR |
153 | } |
154 | error = __split_vma(mm, vma, start, 1); | |
155 | if (error) { | |
156 | /* | |
157 | * madvise() returns EAGAIN if kernel resources, such as | |
158 | * slab, are temporarily unavailable. | |
159 | */ | |
160 | if (error == -ENOMEM) | |
161 | error = -EAGAIN; | |
162 | goto out; | |
163 | } | |
1da177e4 LT |
164 | } |
165 | ||
166 | if (end != vma->vm_end) { | |
def5efe0 DR |
167 | if (unlikely(mm->map_count >= sysctl_max_map_count)) { |
168 | error = -ENOMEM; | |
1da177e4 | 169 | goto out; |
def5efe0 DR |
170 | } |
171 | error = __split_vma(mm, vma, end, 0); | |
172 | if (error) { | |
173 | /* | |
174 | * madvise() returns EAGAIN if kernel resources, such as | |
175 | * slab, are temporarily unavailable. | |
176 | */ | |
177 | if (error == -ENOMEM) | |
178 | error = -EAGAIN; | |
179 | goto out; | |
180 | } | |
1da177e4 LT |
181 | } |
182 | ||
836d5ffd | 183 | success: |
1da177e4 LT |
184 | /* |
185 | * vm_flags is protected by the mmap_sem held in write mode. | |
186 | */ | |
e798c6e8 | 187 | vma->vm_flags = new_flags; |
1da177e4 | 188 | out: |
1da177e4 LT |
189 | return error; |
190 | } | |
191 | ||
1998cc04 SL |
192 | #ifdef CONFIG_SWAP |
193 | static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start, | |
194 | unsigned long end, struct mm_walk *walk) | |
195 | { | |
196 | pte_t *orig_pte; | |
197 | struct vm_area_struct *vma = walk->private; | |
198 | unsigned long index; | |
199 | ||
200 | if (pmd_none_or_trans_huge_or_clear_bad(pmd)) | |
201 | return 0; | |
202 | ||
203 | for (index = start; index != end; index += PAGE_SIZE) { | |
204 | pte_t pte; | |
205 | swp_entry_t entry; | |
206 | struct page *page; | |
207 | spinlock_t *ptl; | |
208 | ||
209 | orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl); | |
210 | pte = *(orig_pte + ((index - start) / PAGE_SIZE)); | |
211 | pte_unmap_unlock(orig_pte, ptl); | |
212 | ||
0661a336 | 213 | if (pte_present(pte) || pte_none(pte)) |
1998cc04 SL |
214 | continue; |
215 | entry = pte_to_swp_entry(pte); | |
216 | if (unlikely(non_swap_entry(entry))) | |
217 | continue; | |
218 | ||
219 | page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE, | |
23955622 | 220 | vma, index, false); |
1998cc04 | 221 | if (page) |
09cbfeaf | 222 | put_page(page); |
1998cc04 SL |
223 | } |
224 | ||
225 | return 0; | |
226 | } | |
227 | ||
228 | static void force_swapin_readahead(struct vm_area_struct *vma, | |
229 | unsigned long start, unsigned long end) | |
230 | { | |
231 | struct mm_walk walk = { | |
232 | .mm = vma->vm_mm, | |
233 | .pmd_entry = swapin_walk_pmd_entry, | |
234 | .private = vma, | |
235 | }; | |
236 | ||
237 | walk_page_range(start, end, &walk); | |
238 | ||
239 | lru_add_drain(); /* Push any new pages onto the LRU now */ | |
240 | } | |
241 | ||
242 | static void force_shm_swapin_readahead(struct vm_area_struct *vma, | |
243 | unsigned long start, unsigned long end, | |
244 | struct address_space *mapping) | |
245 | { | |
246 | pgoff_t index; | |
247 | struct page *page; | |
248 | swp_entry_t swap; | |
249 | ||
250 | for (; start < end; start += PAGE_SIZE) { | |
251 | index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
252 | ||
55231e5c | 253 | page = find_get_entry(mapping, index); |
1998cc04 SL |
254 | if (!radix_tree_exceptional_entry(page)) { |
255 | if (page) | |
09cbfeaf | 256 | put_page(page); |
1998cc04 SL |
257 | continue; |
258 | } | |
259 | swap = radix_to_swp_entry(page); | |
260 | page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE, | |
23955622 | 261 | NULL, 0, false); |
1998cc04 | 262 | if (page) |
09cbfeaf | 263 | put_page(page); |
1998cc04 SL |
264 | } |
265 | ||
266 | lru_add_drain(); /* Push any new pages onto the LRU now */ | |
267 | } | |
268 | #endif /* CONFIG_SWAP */ | |
269 | ||
1da177e4 LT |
270 | /* |
271 | * Schedule all required I/O operations. Do not wait for completion. | |
272 | */ | |
ec9bed9d VC |
273 | static long madvise_willneed(struct vm_area_struct *vma, |
274 | struct vm_area_struct **prev, | |
1da177e4 LT |
275 | unsigned long start, unsigned long end) |
276 | { | |
277 | struct file *file = vma->vm_file; | |
278 | ||
6ea8d958 | 279 | *prev = vma; |
1998cc04 | 280 | #ifdef CONFIG_SWAP |
97b713ba | 281 | if (!file) { |
97b713ba | 282 | force_swapin_readahead(vma, start, end); |
1998cc04 SL |
283 | return 0; |
284 | } | |
1998cc04 | 285 | |
97b713ba | 286 | if (shmem_mapping(file->f_mapping)) { |
97b713ba CH |
287 | force_shm_swapin_readahead(vma, start, end, |
288 | file->f_mapping); | |
289 | return 0; | |
290 | } | |
291 | #else | |
1bef4003 S |
292 | if (!file) |
293 | return -EBADF; | |
97b713ba | 294 | #endif |
1bef4003 | 295 | |
e748dcd0 | 296 | if (IS_DAX(file_inode(file))) { |
fe77ba6f CO |
297 | /* no bad return value, but ignore advice */ |
298 | return 0; | |
299 | } | |
300 | ||
1da177e4 LT |
301 | start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; |
302 | if (end > vma->vm_end) | |
303 | end = vma->vm_end; | |
304 | end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
305 | ||
f7e839dd | 306 | force_page_cache_readahead(file->f_mapping, file, start, end - start); |
1da177e4 LT |
307 | return 0; |
308 | } | |
309 | ||
854e9ed0 MK |
310 | static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr, |
311 | unsigned long end, struct mm_walk *walk) | |
312 | ||
313 | { | |
314 | struct mmu_gather *tlb = walk->private; | |
315 | struct mm_struct *mm = tlb->mm; | |
316 | struct vm_area_struct *vma = walk->vma; | |
317 | spinlock_t *ptl; | |
318 | pte_t *orig_pte, *pte, ptent; | |
319 | struct page *page; | |
64b42bc1 | 320 | int nr_swap = 0; |
b8d3c4c3 MK |
321 | unsigned long next; |
322 | ||
323 | next = pmd_addr_end(addr, end); | |
324 | if (pmd_trans_huge(*pmd)) | |
325 | if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next)) | |
326 | goto next; | |
854e9ed0 | 327 | |
854e9ed0 MK |
328 | if (pmd_trans_unstable(pmd)) |
329 | return 0; | |
330 | ||
07e32661 | 331 | tlb_remove_check_page_size_change(tlb, PAGE_SIZE); |
854e9ed0 | 332 | orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl); |
3ea27719 | 333 | flush_tlb_batched_pending(mm); |
854e9ed0 MK |
334 | arch_enter_lazy_mmu_mode(); |
335 | for (; addr != end; pte++, addr += PAGE_SIZE) { | |
336 | ptent = *pte; | |
337 | ||
64b42bc1 | 338 | if (pte_none(ptent)) |
854e9ed0 | 339 | continue; |
64b42bc1 MK |
340 | /* |
341 | * If the pte has swp_entry, just clear page table to | |
342 | * prevent swap-in which is more expensive rather than | |
343 | * (page allocation + zeroing). | |
344 | */ | |
345 | if (!pte_present(ptent)) { | |
346 | swp_entry_t entry; | |
347 | ||
348 | entry = pte_to_swp_entry(ptent); | |
349 | if (non_swap_entry(entry)) | |
350 | continue; | |
351 | nr_swap--; | |
352 | free_swap_and_cache(entry); | |
353 | pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); | |
354 | continue; | |
355 | } | |
854e9ed0 | 356 | |
df6ad698 | 357 | page = _vm_normal_page(vma, addr, ptent, true); |
854e9ed0 MK |
358 | if (!page) |
359 | continue; | |
360 | ||
361 | /* | |
362 | * If pmd isn't transhuge but the page is THP and | |
363 | * is owned by only this process, split it and | |
364 | * deactivate all pages. | |
365 | */ | |
366 | if (PageTransCompound(page)) { | |
367 | if (page_mapcount(page) != 1) | |
368 | goto out; | |
369 | get_page(page); | |
370 | if (!trylock_page(page)) { | |
371 | put_page(page); | |
372 | goto out; | |
373 | } | |
374 | pte_unmap_unlock(orig_pte, ptl); | |
375 | if (split_huge_page(page)) { | |
376 | unlock_page(page); | |
377 | put_page(page); | |
378 | pte_offset_map_lock(mm, pmd, addr, &ptl); | |
379 | goto out; | |
380 | } | |
854e9ed0 | 381 | unlock_page(page); |
263630e8 | 382 | put_page(page); |
854e9ed0 MK |
383 | pte = pte_offset_map_lock(mm, pmd, addr, &ptl); |
384 | pte--; | |
385 | addr -= PAGE_SIZE; | |
386 | continue; | |
387 | } | |
388 | ||
389 | VM_BUG_ON_PAGE(PageTransCompound(page), page); | |
390 | ||
391 | if (PageSwapCache(page) || PageDirty(page)) { | |
392 | if (!trylock_page(page)) | |
393 | continue; | |
394 | /* | |
395 | * If page is shared with others, we couldn't clear | |
396 | * PG_dirty of the page. | |
397 | */ | |
398 | if (page_mapcount(page) != 1) { | |
399 | unlock_page(page); | |
400 | continue; | |
401 | } | |
402 | ||
403 | if (PageSwapCache(page) && !try_to_free_swap(page)) { | |
404 | unlock_page(page); | |
405 | continue; | |
406 | } | |
407 | ||
408 | ClearPageDirty(page); | |
409 | unlock_page(page); | |
410 | } | |
411 | ||
412 | if (pte_young(ptent) || pte_dirty(ptent)) { | |
413 | /* | |
414 | * Some of architecture(ex, PPC) don't update TLB | |
415 | * with set_pte_at and tlb_remove_tlb_entry so for | |
416 | * the portability, remap the pte with old|clean | |
417 | * after pte clearing. | |
418 | */ | |
419 | ptent = ptep_get_and_clear_full(mm, addr, pte, | |
420 | tlb->fullmm); | |
421 | ||
422 | ptent = pte_mkold(ptent); | |
423 | ptent = pte_mkclean(ptent); | |
424 | set_pte_at(mm, addr, pte, ptent); | |
425 | tlb_remove_tlb_entry(tlb, pte, addr); | |
426 | } | |
802a3a92 | 427 | mark_page_lazyfree(page); |
854e9ed0 MK |
428 | } |
429 | out: | |
64b42bc1 MK |
430 | if (nr_swap) { |
431 | if (current->mm == mm) | |
432 | sync_mm_rss(mm); | |
433 | ||
434 | add_mm_counter(mm, MM_SWAPENTS, nr_swap); | |
435 | } | |
854e9ed0 MK |
436 | arch_leave_lazy_mmu_mode(); |
437 | pte_unmap_unlock(orig_pte, ptl); | |
438 | cond_resched(); | |
b8d3c4c3 | 439 | next: |
854e9ed0 MK |
440 | return 0; |
441 | } | |
442 | ||
443 | static void madvise_free_page_range(struct mmu_gather *tlb, | |
444 | struct vm_area_struct *vma, | |
445 | unsigned long addr, unsigned long end) | |
446 | { | |
447 | struct mm_walk free_walk = { | |
448 | .pmd_entry = madvise_free_pte_range, | |
449 | .mm = vma->vm_mm, | |
450 | .private = tlb, | |
451 | }; | |
452 | ||
453 | tlb_start_vma(tlb, vma); | |
454 | walk_page_range(addr, end, &free_walk); | |
455 | tlb_end_vma(tlb, vma); | |
456 | } | |
457 | ||
458 | static int madvise_free_single_vma(struct vm_area_struct *vma, | |
459 | unsigned long start_addr, unsigned long end_addr) | |
460 | { | |
461 | unsigned long start, end; | |
462 | struct mm_struct *mm = vma->vm_mm; | |
463 | struct mmu_gather tlb; | |
464 | ||
854e9ed0 MK |
465 | /* MADV_FREE works for only anon vma at the moment */ |
466 | if (!vma_is_anonymous(vma)) | |
467 | return -EINVAL; | |
468 | ||
469 | start = max(vma->vm_start, start_addr); | |
470 | if (start >= vma->vm_end) | |
471 | return -EINVAL; | |
472 | end = min(vma->vm_end, end_addr); | |
473 | if (end <= vma->vm_start) | |
474 | return -EINVAL; | |
475 | ||
476 | lru_add_drain(); | |
477 | tlb_gather_mmu(&tlb, mm, start, end); | |
478 | update_hiwater_rss(mm); | |
479 | ||
480 | mmu_notifier_invalidate_range_start(mm, start, end); | |
481 | madvise_free_page_range(&tlb, vma, start, end); | |
482 | mmu_notifier_invalidate_range_end(mm, start, end); | |
483 | tlb_finish_mmu(&tlb, start, end); | |
484 | ||
485 | return 0; | |
486 | } | |
487 | ||
1da177e4 LT |
488 | /* |
489 | * Application no longer needs these pages. If the pages are dirty, | |
490 | * it's OK to just throw them away. The app will be more careful about | |
491 | * data it wants to keep. Be sure to free swap resources too. The | |
7e6cbea3 | 492 | * zap_page_range call sets things up for shrink_active_list to actually free |
1da177e4 LT |
493 | * these pages later if no one else has touched them in the meantime, |
494 | * although we could add these pages to a global reuse list for | |
7e6cbea3 | 495 | * shrink_active_list to pick up before reclaiming other pages. |
1da177e4 LT |
496 | * |
497 | * NB: This interface discards data rather than pushes it out to swap, | |
498 | * as some implementations do. This has performance implications for | |
499 | * applications like large transactional databases which want to discard | |
500 | * pages in anonymous maps after committing to backing store the data | |
501 | * that was kept in them. There is no reason to write this data out to | |
502 | * the swap area if the application is discarding it. | |
503 | * | |
504 | * An interface that causes the system to free clean pages and flush | |
505 | * dirty pages is already available as msync(MS_INVALIDATE). | |
506 | */ | |
230ca982 MR |
507 | static long madvise_dontneed_single_vma(struct vm_area_struct *vma, |
508 | unsigned long start, unsigned long end) | |
509 | { | |
510 | zap_page_range(vma, start, end - start); | |
511 | return 0; | |
512 | } | |
513 | ||
514 | static long madvise_dontneed_free(struct vm_area_struct *vma, | |
515 | struct vm_area_struct **prev, | |
516 | unsigned long start, unsigned long end, | |
517 | int behavior) | |
1da177e4 | 518 | { |
05b74384 | 519 | *prev = vma; |
23519073 | 520 | if (!can_madv_dontneed_vma(vma)) |
1da177e4 LT |
521 | return -EINVAL; |
522 | ||
70ccb92f AA |
523 | if (!userfaultfd_remove(vma, start, end)) { |
524 | *prev = NULL; /* mmap_sem has been dropped, prev is stale */ | |
525 | ||
526 | down_read(¤t->mm->mmap_sem); | |
527 | vma = find_vma(current->mm, start); | |
528 | if (!vma) | |
529 | return -ENOMEM; | |
530 | if (start < vma->vm_start) { | |
531 | /* | |
532 | * This "vma" under revalidation is the one | |
533 | * with the lowest vma->vm_start where start | |
534 | * is also < vma->vm_end. If start < | |
535 | * vma->vm_start it means an hole materialized | |
536 | * in the user address space within the | |
230ca982 MR |
537 | * virtual range passed to MADV_DONTNEED |
538 | * or MADV_FREE. | |
70ccb92f AA |
539 | */ |
540 | return -ENOMEM; | |
541 | } | |
542 | if (!can_madv_dontneed_vma(vma)) | |
543 | return -EINVAL; | |
544 | if (end > vma->vm_end) { | |
545 | /* | |
546 | * Don't fail if end > vma->vm_end. If the old | |
547 | * vma was splitted while the mmap_sem was | |
548 | * released the effect of the concurrent | |
230ca982 | 549 | * operation may not cause madvise() to |
70ccb92f AA |
550 | * have an undefined result. There may be an |
551 | * adjacent next vma that we'll walk | |
552 | * next. userfaultfd_remove() will generate an | |
553 | * UFFD_EVENT_REMOVE repetition on the | |
554 | * end-vma->vm_end range, but the manager can | |
555 | * handle a repetition fine. | |
556 | */ | |
557 | end = vma->vm_end; | |
558 | } | |
559 | VM_WARN_ON(start >= end); | |
560 | } | |
230ca982 MR |
561 | |
562 | if (behavior == MADV_DONTNEED) | |
563 | return madvise_dontneed_single_vma(vma, start, end); | |
564 | else if (behavior == MADV_FREE) | |
565 | return madvise_free_single_vma(vma, start, end); | |
566 | else | |
567 | return -EINVAL; | |
1da177e4 LT |
568 | } |
569 | ||
f6b3ec23 BP |
570 | /* |
571 | * Application wants to free up the pages and associated backing store. | |
572 | * This is effectively punching a hole into the middle of a file. | |
f6b3ec23 BP |
573 | */ |
574 | static long madvise_remove(struct vm_area_struct *vma, | |
00e9fa2d | 575 | struct vm_area_struct **prev, |
f6b3ec23 BP |
576 | unsigned long start, unsigned long end) |
577 | { | |
3f31d075 | 578 | loff_t offset; |
90ed52eb | 579 | int error; |
9ab4233d | 580 | struct file *f; |
f6b3ec23 | 581 | |
90ed52eb | 582 | *prev = NULL; /* tell sys_madvise we drop mmap_sem */ |
00e9fa2d | 583 | |
72079ba0 | 584 | if (vma->vm_flags & VM_LOCKED) |
f6b3ec23 BP |
585 | return -EINVAL; |
586 | ||
9ab4233d AL |
587 | f = vma->vm_file; |
588 | ||
589 | if (!f || !f->f_mapping || !f->f_mapping->host) { | |
f6b3ec23 BP |
590 | return -EINVAL; |
591 | } | |
592 | ||
69cf0fac HD |
593 | if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE)) |
594 | return -EACCES; | |
595 | ||
f6b3ec23 BP |
596 | offset = (loff_t)(start - vma->vm_start) |
597 | + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); | |
90ed52eb | 598 | |
9ab4233d AL |
599 | /* |
600 | * Filesystem's fallocate may need to take i_mutex. We need to | |
601 | * explicitly grab a reference because the vma (and hence the | |
602 | * vma's reference to the file) can go away as soon as we drop | |
603 | * mmap_sem. | |
604 | */ | |
605 | get_file(f); | |
70ccb92f AA |
606 | if (userfaultfd_remove(vma, start, end)) { |
607 | /* mmap_sem was not released by userfaultfd_remove() */ | |
608 | up_read(¤t->mm->mmap_sem); | |
609 | } | |
72c72bdf | 610 | error = vfs_fallocate(f, |
3f31d075 HD |
611 | FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, |
612 | offset, end - start); | |
9ab4233d | 613 | fput(f); |
0a27a14a | 614 | down_read(¤t->mm->mmap_sem); |
90ed52eb | 615 | return error; |
f6b3ec23 BP |
616 | } |
617 | ||
9893e49d AK |
618 | #ifdef CONFIG_MEMORY_FAILURE |
619 | /* | |
620 | * Error injection support for memory error handling. | |
621 | */ | |
97167a76 AK |
622 | static int madvise_inject_error(int behavior, |
623 | unsigned long start, unsigned long end) | |
9893e49d | 624 | { |
97167a76 | 625 | struct page *page; |
c461ad6a | 626 | struct zone *zone; |
19bfbe22 | 627 | unsigned int order; |
97167a76 | 628 | |
9893e49d AK |
629 | if (!capable(CAP_SYS_ADMIN)) |
630 | return -EPERM; | |
97167a76 | 631 | |
19bfbe22 AM |
632 | |
633 | for (; start < end; start += PAGE_SIZE << order) { | |
23e7b5c2 | 634 | unsigned long pfn; |
325c4ef5 AM |
635 | int ret; |
636 | ||
97167a76 | 637 | ret = get_user_pages_fast(start, 1, 0, &page); |
9893e49d AK |
638 | if (ret != 1) |
639 | return ret; | |
23e7b5c2 | 640 | pfn = page_to_pfn(page); |
325c4ef5 | 641 | |
19bfbe22 AM |
642 | /* |
643 | * When soft offlining hugepages, after migrating the page | |
644 | * we dissolve it, therefore in the second loop "page" will | |
645 | * no longer be a compound page, and order will be 0. | |
646 | */ | |
647 | order = compound_order(compound_head(page)); | |
648 | ||
97167a76 AK |
649 | if (PageHWPoison(page)) { |
650 | put_page(page); | |
29b4eede WL |
651 | continue; |
652 | } | |
97167a76 AK |
653 | |
654 | if (behavior == MADV_SOFT_OFFLINE) { | |
655 | pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n", | |
23e7b5c2 | 656 | pfn, start); |
97167a76 AK |
657 | |
658 | ret = soft_offline_page(page, MF_COUNT_INCREASED); | |
afcf938e | 659 | if (ret) |
8302423b | 660 | return ret; |
afcf938e AK |
661 | continue; |
662 | } | |
23e7b5c2 | 663 | |
97167a76 | 664 | pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n", |
23e7b5c2 | 665 | pfn, start); |
97167a76 | 666 | |
23e7b5c2 DW |
667 | /* |
668 | * Drop the page reference taken by get_user_pages_fast(). In | |
669 | * the absence of MF_COUNT_INCREASED the memory_failure() | |
670 | * routine is responsible for pinning the page to prevent it | |
671 | * from being released back to the page allocator. | |
672 | */ | |
673 | put_page(page); | |
674 | ret = memory_failure(pfn, 0); | |
23a003bf NH |
675 | if (ret) |
676 | return ret; | |
9893e49d | 677 | } |
c461ad6a MG |
678 | |
679 | /* Ensure that all poisoned pages are removed from per-cpu lists */ | |
680 | for_each_populated_zone(zone) | |
681 | drain_all_pages(zone); | |
682 | ||
325c4ef5 | 683 | return 0; |
9893e49d AK |
684 | } |
685 | #endif | |
686 | ||
165cd402 | 687 | static long |
688 | madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev, | |
689 | unsigned long start, unsigned long end, int behavior) | |
1da177e4 | 690 | { |
1da177e4 | 691 | switch (behavior) { |
f6b3ec23 | 692 | case MADV_REMOVE: |
3866ea90 | 693 | return madvise_remove(vma, prev, start, end); |
1da177e4 | 694 | case MADV_WILLNEED: |
3866ea90 | 695 | return madvise_willneed(vma, prev, start, end); |
854e9ed0 | 696 | case MADV_FREE: |
1da177e4 | 697 | case MADV_DONTNEED: |
230ca982 | 698 | return madvise_dontneed_free(vma, prev, start, end, behavior); |
1da177e4 | 699 | default: |
3866ea90 | 700 | return madvise_behavior(vma, prev, start, end, behavior); |
1da177e4 | 701 | } |
1da177e4 LT |
702 | } |
703 | ||
1ecef9ed | 704 | static bool |
75927af8 NP |
705 | madvise_behavior_valid(int behavior) |
706 | { | |
707 | switch (behavior) { | |
708 | case MADV_DOFORK: | |
709 | case MADV_DONTFORK: | |
710 | case MADV_NORMAL: | |
711 | case MADV_SEQUENTIAL: | |
712 | case MADV_RANDOM: | |
713 | case MADV_REMOVE: | |
714 | case MADV_WILLNEED: | |
715 | case MADV_DONTNEED: | |
854e9ed0 | 716 | case MADV_FREE: |
f8af4da3 HD |
717 | #ifdef CONFIG_KSM |
718 | case MADV_MERGEABLE: | |
719 | case MADV_UNMERGEABLE: | |
0af4e98b AA |
720 | #endif |
721 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
722 | case MADV_HUGEPAGE: | |
a664b2d8 | 723 | case MADV_NOHUGEPAGE: |
f8af4da3 | 724 | #endif |
accb61fe JB |
725 | case MADV_DONTDUMP: |
726 | case MADV_DODUMP: | |
d2cd9ede RR |
727 | case MADV_WIPEONFORK: |
728 | case MADV_KEEPONFORK: | |
5e451be7 AK |
729 | #ifdef CONFIG_MEMORY_FAILURE |
730 | case MADV_SOFT_OFFLINE: | |
731 | case MADV_HWPOISON: | |
732 | #endif | |
1ecef9ed | 733 | return true; |
75927af8 NP |
734 | |
735 | default: | |
1ecef9ed | 736 | return false; |
75927af8 NP |
737 | } |
738 | } | |
3866ea90 | 739 | |
1da177e4 LT |
740 | /* |
741 | * The madvise(2) system call. | |
742 | * | |
743 | * Applications can use madvise() to advise the kernel how it should | |
744 | * handle paging I/O in this VM area. The idea is to help the kernel | |
745 | * use appropriate read-ahead and caching techniques. The information | |
746 | * provided is advisory only, and can be safely disregarded by the | |
747 | * kernel without affecting the correct operation of the application. | |
748 | * | |
749 | * behavior values: | |
750 | * MADV_NORMAL - the default behavior is to read clusters. This | |
751 | * results in some read-ahead and read-behind. | |
752 | * MADV_RANDOM - the system should read the minimum amount of data | |
753 | * on any access, since it is unlikely that the appli- | |
754 | * cation will need more than what it asks for. | |
755 | * MADV_SEQUENTIAL - pages in the given range will probably be accessed | |
756 | * once, so they can be aggressively read ahead, and | |
757 | * can be freed soon after they are accessed. | |
758 | * MADV_WILLNEED - the application is notifying the system to read | |
759 | * some pages ahead. | |
760 | * MADV_DONTNEED - the application is finished with the given range, | |
761 | * so the kernel can free resources associated with it. | |
d7206a70 NH |
762 | * MADV_FREE - the application marks pages in the given range as lazy free, |
763 | * where actual purges are postponed until memory pressure happens. | |
f6b3ec23 BP |
764 | * MADV_REMOVE - the application wants to free up the given range of |
765 | * pages and associated backing store. | |
3866ea90 HD |
766 | * MADV_DONTFORK - omit this area from child's address space when forking: |
767 | * typically, to avoid COWing pages pinned by get_user_pages(). | |
768 | * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking. | |
c02c3009 YS |
769 | * MADV_WIPEONFORK - present the child process with zero-filled memory in this |
770 | * range after a fork. | |
771 | * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK | |
d7206a70 NH |
772 | * MADV_HWPOISON - trigger memory error handler as if the given memory range |
773 | * were corrupted by unrecoverable hardware memory failure. | |
774 | * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory. | |
f8af4da3 HD |
775 | * MADV_MERGEABLE - the application recommends that KSM try to merge pages in |
776 | * this area with pages of identical content from other such areas. | |
777 | * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others. | |
d7206a70 NH |
778 | * MADV_HUGEPAGE - the application wants to back the given range by transparent |
779 | * huge pages in the future. Existing pages might be coalesced and | |
780 | * new pages might be allocated as THP. | |
781 | * MADV_NOHUGEPAGE - mark the given range as not worth being backed by | |
782 | * transparent huge pages so the existing pages will not be | |
783 | * coalesced into THP and new pages will not be allocated as THP. | |
784 | * MADV_DONTDUMP - the application wants to prevent pages in the given range | |
785 | * from being included in its core dump. | |
786 | * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump. | |
1da177e4 LT |
787 | * |
788 | * return values: | |
789 | * zero - success | |
790 | * -EINVAL - start + len < 0, start is not page-aligned, | |
791 | * "behavior" is not a valid value, or application | |
c02c3009 YS |
792 | * is attempting to release locked or shared pages, |
793 | * or the specified address range includes file, Huge TLB, | |
794 | * MAP_SHARED or VMPFNMAP range. | |
1da177e4 LT |
795 | * -ENOMEM - addresses in the specified range are not currently |
796 | * mapped, or are outside the AS of the process. | |
797 | * -EIO - an I/O error occurred while paging in data. | |
798 | * -EBADF - map exists, but area maps something that isn't a file. | |
799 | * -EAGAIN - a kernel resource was temporarily unavailable. | |
800 | */ | |
3480b257 | 801 | SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior) |
1da177e4 | 802 | { |
05b74384 | 803 | unsigned long end, tmp; |
ec9bed9d | 804 | struct vm_area_struct *vma, *prev; |
1da177e4 LT |
805 | int unmapped_error = 0; |
806 | int error = -EINVAL; | |
f7977793 | 807 | int write; |
1da177e4 | 808 | size_t len; |
1998cc04 | 809 | struct blk_plug plug; |
1da177e4 | 810 | |
75927af8 NP |
811 | if (!madvise_behavior_valid(behavior)) |
812 | return error; | |
813 | ||
1da177e4 | 814 | if (start & ~PAGE_MASK) |
84d96d89 | 815 | return error; |
1da177e4 LT |
816 | len = (len_in + ~PAGE_MASK) & PAGE_MASK; |
817 | ||
818 | /* Check to see whether len was rounded up from small -ve to zero */ | |
819 | if (len_in && !len) | |
84d96d89 | 820 | return error; |
1da177e4 LT |
821 | |
822 | end = start + len; | |
823 | if (end < start) | |
84d96d89 | 824 | return error; |
1da177e4 LT |
825 | |
826 | error = 0; | |
827 | if (end == start) | |
84d96d89 RV |
828 | return error; |
829 | ||
5e451be7 AK |
830 | #ifdef CONFIG_MEMORY_FAILURE |
831 | if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE) | |
832 | return madvise_inject_error(behavior, start, start + len_in); | |
833 | #endif | |
834 | ||
84d96d89 | 835 | write = madvise_need_mmap_write(behavior); |
dc0ef0df MH |
836 | if (write) { |
837 | if (down_write_killable(¤t->mm->mmap_sem)) | |
838 | return -EINTR; | |
839 | } else { | |
84d96d89 | 840 | down_read(¤t->mm->mmap_sem); |
dc0ef0df | 841 | } |
1da177e4 LT |
842 | |
843 | /* | |
844 | * If the interval [start,end) covers some unmapped address | |
845 | * ranges, just ignore them, but return -ENOMEM at the end. | |
05b74384 | 846 | * - different from the way of handling in mlock etc. |
1da177e4 | 847 | */ |
05b74384 | 848 | vma = find_vma_prev(current->mm, start, &prev); |
836d5ffd HD |
849 | if (vma && start > vma->vm_start) |
850 | prev = vma; | |
851 | ||
1998cc04 | 852 | blk_start_plug(&plug); |
1da177e4 LT |
853 | for (;;) { |
854 | /* Still start < end. */ | |
855 | error = -ENOMEM; | |
856 | if (!vma) | |
84d96d89 | 857 | goto out; |
1da177e4 | 858 | |
05b74384 | 859 | /* Here start < (end|vma->vm_end). */ |
1da177e4 LT |
860 | if (start < vma->vm_start) { |
861 | unmapped_error = -ENOMEM; | |
862 | start = vma->vm_start; | |
05b74384 | 863 | if (start >= end) |
84d96d89 | 864 | goto out; |
1da177e4 LT |
865 | } |
866 | ||
05b74384 PM |
867 | /* Here vma->vm_start <= start < (end|vma->vm_end) */ |
868 | tmp = vma->vm_end; | |
869 | if (end < tmp) | |
870 | tmp = end; | |
1da177e4 | 871 | |
05b74384 PM |
872 | /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */ |
873 | error = madvise_vma(vma, &prev, start, tmp, behavior); | |
1da177e4 | 874 | if (error) |
84d96d89 | 875 | goto out; |
05b74384 | 876 | start = tmp; |
90ed52eb | 877 | if (prev && start < prev->vm_end) |
05b74384 PM |
878 | start = prev->vm_end; |
879 | error = unmapped_error; | |
880 | if (start >= end) | |
84d96d89 | 881 | goto out; |
90ed52eb HD |
882 | if (prev) |
883 | vma = prev->vm_next; | |
884 | else /* madvise_remove dropped mmap_sem */ | |
885 | vma = find_vma(current->mm, start); | |
1da177e4 | 886 | } |
1da177e4 | 887 | out: |
84d96d89 | 888 | blk_finish_plug(&plug); |
f7977793 | 889 | if (write) |
0a27a14a NP |
890 | up_write(¤t->mm->mmap_sem); |
891 | else | |
892 | up_read(¤t->mm->mmap_sem); | |
893 | ||
1da177e4 LT |
894 | return error; |
895 | } |