blob: a01147359f3bb71c9c34b9c14f97ad0a730d33e5 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/madvise.c
3 *
4 * Copyright (C) 1999 Linus Torvalds
5 * Copyright (C) 2002 Christoph Hellwig
6 */
7
8#include <linux/mman.h>
9#include <linux/pagemap.h>
10#include <linux/syscalls.h>
Prasanna Meda05b74382005-06-21 17:14:37 -070011#include <linux/mempolicy.h>
Andi Kleenafcf9382009-12-16 12:20:00 +010012#include <linux/page-isolation.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/hugetlb.h>
Hugh Dickins3f31d072012-05-29 15:06:40 -070014#include <linux/falloc.h>
Alexey Dobriyane8edc6e2007-05-21 01:22:52 +040015#include <linux/sched.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070016#include <linux/ksm.h>
Hugh Dickins3f31d072012-05-29 15:06:40 -070017#include <linux/fs.h>
Andy Lutomirski9ab42332012-07-05 16:00:11 -070018#include <linux/file.h>
Shaohua Li1998cc02013-02-22 16:32:31 -080019#include <linux/blkdev.h>
Tejun Heo66114ca2015-05-22 17:13:32 -040020#include <linux/backing-dev.h>
Shaohua Li1998cc02013-02-22 16:32:31 -080021#include <linux/swap.h>
22#include <linux/swapops.h>
Minchan Kim854e9ed2016-01-15 16:54:53 -080023#include <linux/mmu_notifier.h>
24
25#include <asm/tlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026
27/*
Nick Piggin0a27a142007-05-06 14:49:53 -070028 * Any behaviour which results in changes to the vma->vm_flags needs to
29 * take mmap_sem for writing. Others, which simply traverse vmas, need
30 * to only take it for reading.
31 */
32static int madvise_need_mmap_write(int behavior)
33{
34 switch (behavior) {
35 case MADV_REMOVE:
36 case MADV_WILLNEED:
37 case MADV_DONTNEED:
Minchan Kim854e9ed2016-01-15 16:54:53 -080038 case MADV_FREE:
Nick Piggin0a27a142007-05-06 14:49:53 -070039 return 0;
40 default:
41 /* be safe, default to 1. list exceptions explicitly */
42 return 1;
43 }
44}
45
46/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070047 * We can potentially split a vm area into separate
48 * areas, each area with its own behavior.
49 */
Vladimir Cernovec9bed92013-09-11 14:20:15 -070050static long madvise_behavior(struct vm_area_struct *vma,
Prasanna Meda05b74382005-06-21 17:14:37 -070051 struct vm_area_struct **prev,
52 unsigned long start, unsigned long end, int behavior)
Linus Torvalds1da177e2005-04-16 15:20:36 -070053{
Vladimir Cernovec9bed92013-09-11 14:20:15 -070054 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -070055 int error = 0;
Prasanna Meda05b74382005-06-21 17:14:37 -070056 pgoff_t pgoff;
Hugh Dickins3866ea92009-09-21 17:01:52 -070057 unsigned long new_flags = vma->vm_flags;
Prasanna Medae798c6e2005-06-21 17:14:36 -070058
59 switch (behavior) {
Michael S. Tsirkinf8225662006-02-14 13:53:08 -080060 case MADV_NORMAL:
61 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
62 break;
Prasanna Medae798c6e2005-06-21 17:14:36 -070063 case MADV_SEQUENTIAL:
Michael S. Tsirkinf8225662006-02-14 13:53:08 -080064 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
Prasanna Medae798c6e2005-06-21 17:14:36 -070065 break;
66 case MADV_RANDOM:
Michael S. Tsirkinf8225662006-02-14 13:53:08 -080067 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
Prasanna Medae798c6e2005-06-21 17:14:36 -070068 break;
Michael S. Tsirkinf8225662006-02-14 13:53:08 -080069 case MADV_DONTFORK:
70 new_flags |= VM_DONTCOPY;
71 break;
72 case MADV_DOFORK:
Hugh Dickins3866ea92009-09-21 17:01:52 -070073 if (vma->vm_flags & VM_IO) {
74 error = -EINVAL;
75 goto out;
76 }
Michael S. Tsirkinf8225662006-02-14 13:53:08 -080077 new_flags &= ~VM_DONTCOPY;
Prasanna Medae798c6e2005-06-21 17:14:36 -070078 break;
Jason Baronaccb61f2012-03-23 15:02:51 -070079 case MADV_DONTDUMP:
Konstantin Khlebnikov0103bd12012-10-08 16:28:59 -070080 new_flags |= VM_DONTDUMP;
Jason Baronaccb61f2012-03-23 15:02:51 -070081 break;
82 case MADV_DODUMP:
Konstantin Khlebnikov0103bd12012-10-08 16:28:59 -070083 if (new_flags & VM_SPECIAL) {
84 error = -EINVAL;
85 goto out;
86 }
87 new_flags &= ~VM_DONTDUMP;
Jason Baronaccb61f2012-03-23 15:02:51 -070088 break;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070089 case MADV_MERGEABLE:
90 case MADV_UNMERGEABLE:
91 error = ksm_madvise(vma, start, end, behavior, &new_flags);
92 if (error)
93 goto out;
94 break;
Andrea Arcangeli0af4e982011-01-13 15:46:55 -080095 case MADV_HUGEPAGE:
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080096 case MADV_NOHUGEPAGE:
Andrea Arcangeli60ab3242011-01-13 15:47:18 -080097 error = hugepage_madvise(vma, &new_flags, behavior);
Andrea Arcangeli0af4e982011-01-13 15:46:55 -080098 if (error)
99 goto out;
100 break;
Prasanna Medae798c6e2005-06-21 17:14:36 -0700101 }
102
Prasanna Meda05b74382005-06-21 17:14:37 -0700103 if (new_flags == vma->vm_flags) {
104 *prev = vma;
Hugh Dickins836d5ff2005-09-03 15:54:53 -0700105 goto out;
Prasanna Meda05b74382005-06-21 17:14:37 -0700106 }
107
108 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
109 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
Andrea Arcangeli19a809a2015-09-04 15:46:24 -0700110 vma->vm_file, pgoff, vma_policy(vma),
111 vma->vm_userfaultfd_ctx);
Prasanna Meda05b74382005-06-21 17:14:37 -0700112 if (*prev) {
113 vma = *prev;
114 goto success;
115 }
116
117 *prev = vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118
119 if (start != vma->vm_start) {
120 error = split_vma(mm, vma, start, 1);
121 if (error)
122 goto out;
123 }
124
125 if (end != vma->vm_end) {
126 error = split_vma(mm, vma, end, 0);
127 if (error)
128 goto out;
129 }
130
Hugh Dickins836d5ff2005-09-03 15:54:53 -0700131success:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 /*
133 * vm_flags is protected by the mmap_sem held in write mode.
134 */
Prasanna Medae798c6e2005-06-21 17:14:36 -0700135 vma->vm_flags = new_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136
137out:
138 if (error == -ENOMEM)
139 error = -EAGAIN;
140 return error;
141}
142
Shaohua Li1998cc02013-02-22 16:32:31 -0800143#ifdef CONFIG_SWAP
144static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
145 unsigned long end, struct mm_walk *walk)
146{
147 pte_t *orig_pte;
148 struct vm_area_struct *vma = walk->private;
149 unsigned long index;
150
151 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
152 return 0;
153
154 for (index = start; index != end; index += PAGE_SIZE) {
155 pte_t pte;
156 swp_entry_t entry;
157 struct page *page;
158 spinlock_t *ptl;
159
160 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
161 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
162 pte_unmap_unlock(orig_pte, ptl);
163
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800164 if (pte_present(pte) || pte_none(pte))
Shaohua Li1998cc02013-02-22 16:32:31 -0800165 continue;
166 entry = pte_to_swp_entry(pte);
167 if (unlikely(non_swap_entry(entry)))
168 continue;
169
170 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
171 vma, index);
172 if (page)
173 page_cache_release(page);
174 }
175
176 return 0;
177}
178
179static void force_swapin_readahead(struct vm_area_struct *vma,
180 unsigned long start, unsigned long end)
181{
182 struct mm_walk walk = {
183 .mm = vma->vm_mm,
184 .pmd_entry = swapin_walk_pmd_entry,
185 .private = vma,
186 };
187
188 walk_page_range(start, end, &walk);
189
190 lru_add_drain(); /* Push any new pages onto the LRU now */
191}
192
193static void force_shm_swapin_readahead(struct vm_area_struct *vma,
194 unsigned long start, unsigned long end,
195 struct address_space *mapping)
196{
197 pgoff_t index;
198 struct page *page;
199 swp_entry_t swap;
200
201 for (; start < end; start += PAGE_SIZE) {
202 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
203
Johannes Weiner55231e52014-05-22 11:54:17 -0700204 page = find_get_entry(mapping, index);
Shaohua Li1998cc02013-02-22 16:32:31 -0800205 if (!radix_tree_exceptional_entry(page)) {
206 if (page)
207 page_cache_release(page);
208 continue;
209 }
210 swap = radix_to_swp_entry(page);
211 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
212 NULL, 0);
213 if (page)
214 page_cache_release(page);
215 }
216
217 lru_add_drain(); /* Push any new pages onto the LRU now */
218}
219#endif /* CONFIG_SWAP */
220
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221/*
222 * Schedule all required I/O operations. Do not wait for completion.
223 */
Vladimir Cernovec9bed92013-09-11 14:20:15 -0700224static long madvise_willneed(struct vm_area_struct *vma,
225 struct vm_area_struct **prev,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 unsigned long start, unsigned long end)
227{
228 struct file *file = vma->vm_file;
229
Shaohua Li1998cc02013-02-22 16:32:31 -0800230#ifdef CONFIG_SWAP
Christoph Hellwig97b713b2015-01-14 10:42:31 +0100231 if (!file) {
Shaohua Li1998cc02013-02-22 16:32:31 -0800232 *prev = vma;
Christoph Hellwig97b713b2015-01-14 10:42:31 +0100233 force_swapin_readahead(vma, start, end);
Shaohua Li1998cc02013-02-22 16:32:31 -0800234 return 0;
235 }
Shaohua Li1998cc02013-02-22 16:32:31 -0800236
Christoph Hellwig97b713b2015-01-14 10:42:31 +0100237 if (shmem_mapping(file->f_mapping)) {
238 *prev = vma;
239 force_shm_swapin_readahead(vma, start, end,
240 file->f_mapping);
241 return 0;
242 }
243#else
Suzuki1bef4002005-10-11 08:29:06 -0700244 if (!file)
245 return -EBADF;
Christoph Hellwig97b713b2015-01-14 10:42:31 +0100246#endif
Suzuki1bef4002005-10-11 08:29:06 -0700247
Matthew Wilcoxe748dcd2015-02-16 15:59:12 -0800248 if (IS_DAX(file_inode(file))) {
Carsten Ottefe77ba62005-06-23 22:05:29 -0700249 /* no bad return value, but ignore advice */
250 return 0;
251 }
252
Prasanna Meda05b74382005-06-21 17:14:37 -0700253 *prev = vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
255 if (end > vma->vm_end)
256 end = vma->vm_end;
257 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
258
Wu Fengguangf7e839d2009-06-16 15:31:20 -0700259 force_page_cache_readahead(file->f_mapping, file, start, end - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 return 0;
261}
262
Minchan Kim854e9ed2016-01-15 16:54:53 -0800263static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
264 unsigned long end, struct mm_walk *walk)
265
266{
267 struct mmu_gather *tlb = walk->private;
268 struct mm_struct *mm = tlb->mm;
269 struct vm_area_struct *vma = walk->vma;
270 spinlock_t *ptl;
271 pte_t *orig_pte, *pte, ptent;
272 struct page *page;
Minchan Kim64b42bc2016-01-15 16:55:06 -0800273 int nr_swap = 0;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -0800274 unsigned long next;
Minchan Kim854e9ed2016-01-15 16:54:53 -0800275
Minchan Kimb8d3c4c2016-01-15 16:55:42 -0800276 next = pmd_addr_end(addr, end);
277 if (pmd_trans_huge(*pmd))
278 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
279 goto next;
280
Minchan Kim854e9ed2016-01-15 16:54:53 -0800281 if (pmd_trans_unstable(pmd))
282 return 0;
283
284 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
285 arch_enter_lazy_mmu_mode();
286 for (; addr != end; pte++, addr += PAGE_SIZE) {
287 ptent = *pte;
288
Minchan Kim64b42bc2016-01-15 16:55:06 -0800289 if (pte_none(ptent))
Minchan Kim854e9ed2016-01-15 16:54:53 -0800290 continue;
Minchan Kim64b42bc2016-01-15 16:55:06 -0800291 /*
292 * If the pte has swp_entry, just clear page table to
293 * prevent swap-in which is more expensive rather than
294 * (page allocation + zeroing).
295 */
296 if (!pte_present(ptent)) {
297 swp_entry_t entry;
298
299 entry = pte_to_swp_entry(ptent);
300 if (non_swap_entry(entry))
301 continue;
302 nr_swap--;
303 free_swap_and_cache(entry);
304 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
305 continue;
306 }
Minchan Kim854e9ed2016-01-15 16:54:53 -0800307
308 page = vm_normal_page(vma, addr, ptent);
309 if (!page)
310 continue;
311
312 /*
313 * If pmd isn't transhuge but the page is THP and
314 * is owned by only this process, split it and
315 * deactivate all pages.
316 */
317 if (PageTransCompound(page)) {
318 if (page_mapcount(page) != 1)
319 goto out;
320 get_page(page);
321 if (!trylock_page(page)) {
322 put_page(page);
323 goto out;
324 }
325 pte_unmap_unlock(orig_pte, ptl);
326 if (split_huge_page(page)) {
327 unlock_page(page);
328 put_page(page);
329 pte_offset_map_lock(mm, pmd, addr, &ptl);
330 goto out;
331 }
332 put_page(page);
333 unlock_page(page);
334 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
335 pte--;
336 addr -= PAGE_SIZE;
337 continue;
338 }
339
340 VM_BUG_ON_PAGE(PageTransCompound(page), page);
341
342 if (PageSwapCache(page) || PageDirty(page)) {
343 if (!trylock_page(page))
344 continue;
345 /*
346 * If page is shared with others, we couldn't clear
347 * PG_dirty of the page.
348 */
349 if (page_mapcount(page) != 1) {
350 unlock_page(page);
351 continue;
352 }
353
354 if (PageSwapCache(page) && !try_to_free_swap(page)) {
355 unlock_page(page);
356 continue;
357 }
358
359 ClearPageDirty(page);
360 unlock_page(page);
361 }
362
363 if (pte_young(ptent) || pte_dirty(ptent)) {
364 /*
365 * Some of architecture(ex, PPC) don't update TLB
366 * with set_pte_at and tlb_remove_tlb_entry so for
367 * the portability, remap the pte with old|clean
368 * after pte clearing.
369 */
370 ptent = ptep_get_and_clear_full(mm, addr, pte,
371 tlb->fullmm);
372
373 ptent = pte_mkold(ptent);
374 ptent = pte_mkclean(ptent);
375 set_pte_at(mm, addr, pte, ptent);
Minchan Kim10853a02016-01-15 16:55:11 -0800376 if (PageActive(page))
377 deactivate_page(page);
Minchan Kim854e9ed2016-01-15 16:54:53 -0800378 tlb_remove_tlb_entry(tlb, pte, addr);
379 }
380 }
381out:
Minchan Kim64b42bc2016-01-15 16:55:06 -0800382 if (nr_swap) {
383 if (current->mm == mm)
384 sync_mm_rss(mm);
385
386 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
387 }
Minchan Kim854e9ed2016-01-15 16:54:53 -0800388 arch_leave_lazy_mmu_mode();
389 pte_unmap_unlock(orig_pte, ptl);
390 cond_resched();
Minchan Kimb8d3c4c2016-01-15 16:55:42 -0800391next:
Minchan Kim854e9ed2016-01-15 16:54:53 -0800392 return 0;
393}
394
395static void madvise_free_page_range(struct mmu_gather *tlb,
396 struct vm_area_struct *vma,
397 unsigned long addr, unsigned long end)
398{
399 struct mm_walk free_walk = {
400 .pmd_entry = madvise_free_pte_range,
401 .mm = vma->vm_mm,
402 .private = tlb,
403 };
404
405 tlb_start_vma(tlb, vma);
406 walk_page_range(addr, end, &free_walk);
407 tlb_end_vma(tlb, vma);
408}
409
410static int madvise_free_single_vma(struct vm_area_struct *vma,
411 unsigned long start_addr, unsigned long end_addr)
412{
413 unsigned long start, end;
414 struct mm_struct *mm = vma->vm_mm;
415 struct mmu_gather tlb;
416
417 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
418 return -EINVAL;
419
420 /* MADV_FREE works for only anon vma at the moment */
421 if (!vma_is_anonymous(vma))
422 return -EINVAL;
423
424 start = max(vma->vm_start, start_addr);
425 if (start >= vma->vm_end)
426 return -EINVAL;
427 end = min(vma->vm_end, end_addr);
428 if (end <= vma->vm_start)
429 return -EINVAL;
430
431 lru_add_drain();
432 tlb_gather_mmu(&tlb, mm, start, end);
433 update_hiwater_rss(mm);
434
435 mmu_notifier_invalidate_range_start(mm, start, end);
436 madvise_free_page_range(&tlb, vma, start, end);
437 mmu_notifier_invalidate_range_end(mm, start, end);
438 tlb_finish_mmu(&tlb, start, end);
439
440 return 0;
441}
442
443static long madvise_free(struct vm_area_struct *vma,
444 struct vm_area_struct **prev,
445 unsigned long start, unsigned long end)
446{
447 *prev = vma;
448 return madvise_free_single_vma(vma, start, end);
449}
450
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451/*
452 * Application no longer needs these pages. If the pages are dirty,
453 * it's OK to just throw them away. The app will be more careful about
454 * data it wants to keep. Be sure to free swap resources too. The
Fernando Luis Vazquez Cao7e6cbea2008-07-29 22:33:39 -0700455 * zap_page_range call sets things up for shrink_active_list to actually free
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 * these pages later if no one else has touched them in the meantime,
457 * although we could add these pages to a global reuse list for
Fernando Luis Vazquez Cao7e6cbea2008-07-29 22:33:39 -0700458 * shrink_active_list to pick up before reclaiming other pages.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 *
460 * NB: This interface discards data rather than pushes it out to swap,
461 * as some implementations do. This has performance implications for
462 * applications like large transactional databases which want to discard
463 * pages in anonymous maps after committing to backing store the data
464 * that was kept in them. There is no reason to write this data out to
465 * the swap area if the application is discarding it.
466 *
467 * An interface that causes the system to free clean pages and flush
468 * dirty pages is already available as msync(MS_INVALIDATE).
469 */
Vladimir Cernovec9bed92013-09-11 14:20:15 -0700470static long madvise_dontneed(struct vm_area_struct *vma,
471 struct vm_area_struct **prev,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 unsigned long start, unsigned long end)
473{
Prasanna Meda05b74382005-06-21 17:14:37 -0700474 *prev = vma;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800475 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476 return -EINVAL;
477
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -0800478 zap_page_range(vma, start, end - start, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479 return 0;
480}
481
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800482/*
483 * Application wants to free up the pages and associated backing store.
484 * This is effectively punching a hole into the middle of a file.
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800485 */
486static long madvise_remove(struct vm_area_struct *vma,
Nick Piggin00e9fa22007-03-16 13:38:10 -0800487 struct vm_area_struct **prev,
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800488 unsigned long start, unsigned long end)
489{
Hugh Dickins3f31d072012-05-29 15:06:40 -0700490 loff_t offset;
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700491 int error;
Andy Lutomirski9ab42332012-07-05 16:00:11 -0700492 struct file *f;
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800493
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700494 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
Nick Piggin00e9fa22007-03-16 13:38:10 -0800495
Mike Kravetz72079ba2015-09-08 15:01:57 -0700496 if (vma->vm_flags & VM_LOCKED)
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800497 return -EINVAL;
498
Andy Lutomirski9ab42332012-07-05 16:00:11 -0700499 f = vma->vm_file;
500
501 if (!f || !f->f_mapping || !f->f_mapping->host) {
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800502 return -EINVAL;
503 }
504
Hugh Dickins69cf0fa2006-04-17 22:46:32 +0100505 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
506 return -EACCES;
507
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800508 offset = (loff_t)(start - vma->vm_start)
509 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700510
Andy Lutomirski9ab42332012-07-05 16:00:11 -0700511 /*
512 * Filesystem's fallocate may need to take i_mutex. We need to
513 * explicitly grab a reference because the vma (and hence the
514 * vma's reference to the file) can go away as soon as we drop
515 * mmap_sem.
516 */
517 get_file(f);
Nick Piggin0a27a142007-05-06 14:49:53 -0700518 up_read(&current->mm->mmap_sem);
Anna Schumaker72c72bd2014-11-07 14:44:25 -0500519 error = vfs_fallocate(f,
Hugh Dickins3f31d072012-05-29 15:06:40 -0700520 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
521 offset, end - start);
Andy Lutomirski9ab42332012-07-05 16:00:11 -0700522 fput(f);
Nick Piggin0a27a142007-05-06 14:49:53 -0700523 down_read(&current->mm->mmap_sem);
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700524 return error;
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800525}
526
Andi Kleen9893e492009-09-16 11:50:17 +0200527#ifdef CONFIG_MEMORY_FAILURE
528/*
529 * Error injection support for memory error handling.
530 */
Andi Kleenafcf9382009-12-16 12:20:00 +0100531static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
Andi Kleen9893e492009-09-16 11:50:17 +0200532{
Wanpeng Li20cb6ca2013-09-30 13:45:21 -0700533 struct page *p;
Andi Kleen9893e492009-09-16 11:50:17 +0200534 if (!capable(CAP_SYS_ADMIN))
535 return -EPERM;
Wanpeng Li20cb6ca2013-09-30 13:45:21 -0700536 for (; start < end; start += PAGE_SIZE <<
537 compound_order(compound_head(p))) {
Andrew Morton325c4ef2013-09-11 14:23:03 -0700538 int ret;
539
540 ret = get_user_pages_fast(start, 1, 0, &p);
Andi Kleen9893e492009-09-16 11:50:17 +0200541 if (ret != 1)
542 return ret;
Andrew Morton325c4ef2013-09-11 14:23:03 -0700543
Wanpeng Li29b4eed2013-09-11 14:22:59 -0700544 if (PageHWPoison(p)) {
545 put_page(p);
546 continue;
547 }
Andi Kleenafcf9382009-12-16 12:20:00 +0100548 if (bhv == MADV_SOFT_OFFLINE) {
Wanpeng Lib194b8c2013-09-11 14:22:57 -0700549 pr_info("Soft offlining page %#lx at %#lx\n",
Andi Kleenafcf9382009-12-16 12:20:00 +0100550 page_to_pfn(p), start);
551 ret = soft_offline_page(p, MF_COUNT_INCREASED);
552 if (ret)
Wanpeng Li83024232013-09-11 14:23:02 -0700553 return ret;
Andi Kleenafcf9382009-12-16 12:20:00 +0100554 continue;
555 }
Wanpeng Lib194b8c2013-09-11 14:22:57 -0700556 pr_info("Injecting memory failure for page %#lx at %#lx\n",
Andi Kleen9893e492009-09-16 11:50:17 +0200557 page_to_pfn(p), start);
Naoya Horiguchi23a003b2016-03-15 14:56:36 -0700558 ret = memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
559 if (ret)
560 return ret;
Andi Kleen9893e492009-09-16 11:50:17 +0200561 }
Andrew Morton325c4ef2013-09-11 14:23:03 -0700562 return 0;
Andi Kleen9893e492009-09-16 11:50:17 +0200563}
564#endif
565
suzuki165cd402005-07-27 11:43:59 -0700566static long
567madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
568 unsigned long start, unsigned long end, int behavior)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 switch (behavior) {
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800571 case MADV_REMOVE:
Hugh Dickins3866ea92009-09-21 17:01:52 -0700572 return madvise_remove(vma, prev, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 case MADV_WILLNEED:
Hugh Dickins3866ea92009-09-21 17:01:52 -0700574 return madvise_willneed(vma, prev, start, end);
Minchan Kim854e9ed2016-01-15 16:54:53 -0800575 case MADV_FREE:
576 /*
577 * XXX: In this implementation, MADV_FREE works like
578 * MADV_DONTNEED on swapless system or full swap.
579 */
580 if (get_nr_swap_pages() > 0)
581 return madvise_free(vma, prev, start, end);
582 /* passthrough */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 case MADV_DONTNEED:
Hugh Dickins3866ea92009-09-21 17:01:52 -0700584 return madvise_dontneed(vma, prev, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 default:
Hugh Dickins3866ea92009-09-21 17:01:52 -0700586 return madvise_behavior(vma, prev, start, end, behavior);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588}
589
Nicholas Krause1ecef9e2015-09-04 15:48:24 -0700590static bool
Nick Piggin75927af2009-06-16 15:32:38 -0700591madvise_behavior_valid(int behavior)
592{
593 switch (behavior) {
594 case MADV_DOFORK:
595 case MADV_DONTFORK:
596 case MADV_NORMAL:
597 case MADV_SEQUENTIAL:
598 case MADV_RANDOM:
599 case MADV_REMOVE:
600 case MADV_WILLNEED:
601 case MADV_DONTNEED:
Minchan Kim854e9ed2016-01-15 16:54:53 -0800602 case MADV_FREE:
Hugh Dickinsf8af4da2009-09-21 17:01:57 -0700603#ifdef CONFIG_KSM
604 case MADV_MERGEABLE:
605 case MADV_UNMERGEABLE:
606#endif
Andrea Arcangeli0af4e982011-01-13 15:46:55 -0800607#ifdef CONFIG_TRANSPARENT_HUGEPAGE
608 case MADV_HUGEPAGE:
Andrea Arcangelia664b2d2011-01-13 15:47:17 -0800609 case MADV_NOHUGEPAGE:
Andrea Arcangeli0af4e982011-01-13 15:46:55 -0800610#endif
Jason Baronaccb61f2012-03-23 15:02:51 -0700611 case MADV_DONTDUMP:
612 case MADV_DODUMP:
Nicholas Krause1ecef9e2015-09-04 15:48:24 -0700613 return true;
Nick Piggin75927af2009-06-16 15:32:38 -0700614
615 default:
Nicholas Krause1ecef9e2015-09-04 15:48:24 -0700616 return false;
Nick Piggin75927af2009-06-16 15:32:38 -0700617 }
618}
Hugh Dickins3866ea92009-09-21 17:01:52 -0700619
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620/*
621 * The madvise(2) system call.
622 *
623 * Applications can use madvise() to advise the kernel how it should
624 * handle paging I/O in this VM area. The idea is to help the kernel
625 * use appropriate read-ahead and caching techniques. The information
626 * provided is advisory only, and can be safely disregarded by the
627 * kernel without affecting the correct operation of the application.
628 *
629 * behavior values:
630 * MADV_NORMAL - the default behavior is to read clusters. This
631 * results in some read-ahead and read-behind.
632 * MADV_RANDOM - the system should read the minimum amount of data
633 * on any access, since it is unlikely that the appli-
634 * cation will need more than what it asks for.
635 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
636 * once, so they can be aggressively read ahead, and
637 * can be freed soon after they are accessed.
638 * MADV_WILLNEED - the application is notifying the system to read
639 * some pages ahead.
640 * MADV_DONTNEED - the application is finished with the given range,
641 * so the kernel can free resources associated with it.
Naoya Horiguchid7206a72016-03-15 14:56:58 -0700642 * MADV_FREE - the application marks pages in the given range as lazy free,
643 * where actual purges are postponed until memory pressure happens.
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800644 * MADV_REMOVE - the application wants to free up the given range of
645 * pages and associated backing store.
Hugh Dickins3866ea92009-09-21 17:01:52 -0700646 * MADV_DONTFORK - omit this area from child's address space when forking:
647 * typically, to avoid COWing pages pinned by get_user_pages().
648 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
Naoya Horiguchid7206a72016-03-15 14:56:58 -0700649 * MADV_HWPOISON - trigger memory error handler as if the given memory range
650 * were corrupted by unrecoverable hardware memory failure.
651 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
Hugh Dickinsf8af4da2009-09-21 17:01:57 -0700652 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
653 * this area with pages of identical content from other such areas.
654 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
Naoya Horiguchid7206a72016-03-15 14:56:58 -0700655 * MADV_HUGEPAGE - the application wants to back the given range by transparent
656 * huge pages in the future. Existing pages might be coalesced and
657 * new pages might be allocated as THP.
658 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
659 * transparent huge pages so the existing pages will not be
660 * coalesced into THP and new pages will not be allocated as THP.
661 * MADV_DONTDUMP - the application wants to prevent pages in the given range
662 * from being included in its core dump.
663 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 *
665 * return values:
666 * zero - success
667 * -EINVAL - start + len < 0, start is not page-aligned,
668 * "behavior" is not a valid value, or application
669 * is attempting to release locked or shared pages.
670 * -ENOMEM - addresses in the specified range are not currently
671 * mapped, or are outside the AS of the process.
672 * -EIO - an I/O error occurred while paging in data.
673 * -EBADF - map exists, but area maps something that isn't a file.
674 * -EAGAIN - a kernel resource was temporarily unavailable.
675 */
Heiko Carstens3480b252009-01-14 14:14:16 +0100676SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677{
Prasanna Meda05b74382005-06-21 17:14:37 -0700678 unsigned long end, tmp;
Vladimir Cernovec9bed92013-09-11 14:20:15 -0700679 struct vm_area_struct *vma, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 int unmapped_error = 0;
681 int error = -EINVAL;
Jason Baronf7977792007-07-15 23:38:21 -0700682 int write;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 size_t len;
Shaohua Li1998cc02013-02-22 16:32:31 -0800684 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685
Andi Kleen9893e492009-09-16 11:50:17 +0200686#ifdef CONFIG_MEMORY_FAILURE
Andi Kleenafcf9382009-12-16 12:20:00 +0100687 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
688 return madvise_hwpoison(behavior, start, start+len_in);
Andi Kleen9893e492009-09-16 11:50:17 +0200689#endif
Nick Piggin75927af2009-06-16 15:32:38 -0700690 if (!madvise_behavior_valid(behavior))
691 return error;
692
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700693 if (start & ~PAGE_MASK)
694 return error;
695 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
696
697 /* Check to see whether len was rounded up from small -ve to zero */
698 if (len_in && !len)
699 return error;
700
701 end = start + len;
702 if (end < start)
703 return error;
704
705 error = 0;
706 if (end == start)
707 return error;
708
Jason Baronf7977792007-07-15 23:38:21 -0700709 write = madvise_need_mmap_write(behavior);
710 if (write)
Nick Piggin0a27a142007-05-06 14:49:53 -0700711 down_write(&current->mm->mmap_sem);
712 else
713 down_read(&current->mm->mmap_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 /*
716 * If the interval [start,end) covers some unmapped address
717 * ranges, just ignore them, but return -ENOMEM at the end.
Prasanna Meda05b74382005-06-21 17:14:37 -0700718 * - different from the way of handling in mlock etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 */
Prasanna Meda05b74382005-06-21 17:14:37 -0700720 vma = find_vma_prev(current->mm, start, &prev);
Hugh Dickins836d5ff2005-09-03 15:54:53 -0700721 if (vma && start > vma->vm_start)
722 prev = vma;
723
Shaohua Li1998cc02013-02-22 16:32:31 -0800724 blk_start_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 for (;;) {
726 /* Still start < end. */
727 error = -ENOMEM;
728 if (!vma)
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700729 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730
Prasanna Meda05b74382005-06-21 17:14:37 -0700731 /* Here start < (end|vma->vm_end). */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 if (start < vma->vm_start) {
733 unmapped_error = -ENOMEM;
734 start = vma->vm_start;
Prasanna Meda05b74382005-06-21 17:14:37 -0700735 if (start >= end)
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700736 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 }
738
Prasanna Meda05b74382005-06-21 17:14:37 -0700739 /* Here vma->vm_start <= start < (end|vma->vm_end) */
740 tmp = vma->vm_end;
741 if (end < tmp)
742 tmp = end;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743
Prasanna Meda05b74382005-06-21 17:14:37 -0700744 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
745 error = madvise_vma(vma, &prev, start, tmp, behavior);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 if (error)
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700747 goto out;
Prasanna Meda05b74382005-06-21 17:14:37 -0700748 start = tmp;
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700749 if (prev && start < prev->vm_end)
Prasanna Meda05b74382005-06-21 17:14:37 -0700750 start = prev->vm_end;
751 error = unmapped_error;
752 if (start >= end)
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700753 goto out;
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700754 if (prev)
755 vma = prev->vm_next;
756 else /* madvise_remove dropped mmap_sem */
757 vma = find_vma(current->mm, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759out:
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700760 blk_finish_plug(&plug);
Jason Baronf7977792007-07-15 23:38:21 -0700761 if (write)
Nick Piggin0a27a142007-05-06 14:49:53 -0700762 up_write(&current->mm->mmap_sem);
763 else
764 up_read(&current->mm->mmap_sem);
765
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766 return error;
767}