blob: 4d3c922ea1a1cb7558e1b69a3e32000833442489 [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
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>
Prasanna Meda05b74382005-06-21 17:14:37 -070012#include <linux/mempolicy.h>
Andi Kleenafcf9382009-12-16 12:20:00 +010013#include <linux/page-isolation.h>
Pavel Emelyanov05ce7722017-02-22 15:42:40 -080014#include <linux/userfaultfd_k.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070015#include <linux/hugetlb.h>
Hugh Dickins3f31d072012-05-29 15:06:40 -070016#include <linux/falloc.h>
Alexey Dobriyane8edc6e2007-05-21 01:22:52 +040017#include <linux/sched.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070018#include <linux/ksm.h>
Hugh Dickins3f31d072012-05-29 15:06:40 -070019#include <linux/fs.h>
Andy Lutomirski9ab42332012-07-05 16:00:11 -070020#include <linux/file.h>
Shaohua Li1998cc02013-02-22 16:32:31 -080021#include <linux/blkdev.h>
Tejun Heo66114ca2015-05-22 17:13:32 -040022#include <linux/backing-dev.h>
Shaohua Li1998cc02013-02-22 16:32:31 -080023#include <linux/swap.h>
24#include <linux/swapops.h>
Hugh Dickins3a4f8a02017-02-24 14:59:36 -080025#include <linux/shmem_fs.h>
Minchan Kim854e9ed2016-01-15 16:54:53 -080026#include <linux/mmu_notifier.h>
27
28#include <asm/tlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029
Kirill A. Shutemov23519072017-02-22 15:46:39 -080030#include "internal.h"
31
Linus Torvalds1da177e2005-04-16 15:20:36 -070032/*
Nick Piggin0a27a142007-05-06 14:49:53 -070033 * 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 */
37static int madvise_need_mmap_write(int behavior)
38{
39 switch (behavior) {
40 case MADV_REMOVE:
41 case MADV_WILLNEED:
42 case MADV_DONTNEED:
Minchan Kim854e9ed2016-01-15 16:54:53 -080043 case MADV_FREE:
Nick Piggin0a27a142007-05-06 14:49:53 -070044 return 0;
45 default:
46 /* be safe, default to 1. list exceptions explicitly */
47 return 1;
48 }
49}
50
51/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 * We can potentially split a vm area into separate
53 * areas, each area with its own behavior.
54 */
Vladimir Cernovec9bed92013-09-11 14:20:15 -070055static long madvise_behavior(struct vm_area_struct *vma,
Prasanna Meda05b74382005-06-21 17:14:37 -070056 struct vm_area_struct **prev,
57 unsigned long start, unsigned long end, int behavior)
Linus Torvalds1da177e2005-04-16 15:20:36 -070058{
Vladimir Cernovec9bed92013-09-11 14:20:15 -070059 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -070060 int error = 0;
Prasanna Meda05b74382005-06-21 17:14:37 -070061 pgoff_t pgoff;
Hugh Dickins3866ea92009-09-21 17:01:52 -070062 unsigned long new_flags = vma->vm_flags;
Prasanna Medae798c6e2005-06-21 17:14:36 -070063
64 switch (behavior) {
Michael S. Tsirkinf8225662006-02-14 13:53:08 -080065 case MADV_NORMAL:
66 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
67 break;
Prasanna Medae798c6e2005-06-21 17:14:36 -070068 case MADV_SEQUENTIAL:
Michael S. Tsirkinf8225662006-02-14 13:53:08 -080069 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
Prasanna Medae798c6e2005-06-21 17:14:36 -070070 break;
71 case MADV_RANDOM:
Michael S. Tsirkinf8225662006-02-14 13:53:08 -080072 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
Prasanna Medae798c6e2005-06-21 17:14:36 -070073 break;
Michael S. Tsirkinf8225662006-02-14 13:53:08 -080074 case MADV_DONTFORK:
75 new_flags |= VM_DONTCOPY;
76 break;
77 case MADV_DOFORK:
Hugh Dickins3866ea92009-09-21 17:01:52 -070078 if (vma->vm_flags & VM_IO) {
79 error = -EINVAL;
80 goto out;
81 }
Michael S. Tsirkinf8225662006-02-14 13:53:08 -080082 new_flags &= ~VM_DONTCOPY;
Prasanna Medae798c6e2005-06-21 17:14:36 -070083 break;
Rik van Rield2cd9ed2017-09-06 16:25:15 -070084 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;
Jason Baronaccb61f2012-03-23 15:02:51 -070095 case MADV_DONTDUMP:
Konstantin Khlebnikov0103bd12012-10-08 16:28:59 -070096 new_flags |= VM_DONTDUMP;
Jason Baronaccb61f2012-03-23 15:02:51 -070097 break;
98 case MADV_DODUMP:
Konstantin Khlebnikov0103bd12012-10-08 16:28:59 -070099 if (new_flags & VM_SPECIAL) {
100 error = -EINVAL;
101 goto out;
102 }
103 new_flags &= ~VM_DONTDUMP;
Jason Baronaccb61f2012-03-23 15:02:51 -0700104 break;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -0700105 case MADV_MERGEABLE:
106 case MADV_UNMERGEABLE:
107 error = ksm_madvise(vma, start, end, behavior, &new_flags);
David Rientjesdef5efe2017-02-24 14:58:47 -0800108 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;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -0700115 goto out;
David Rientjesdef5efe2017-02-24 14:58:47 -0800116 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -0700117 break;
Andrea Arcangeli0af4e982011-01-13 15:46:55 -0800118 case MADV_HUGEPAGE:
Andrea Arcangelia664b2d2011-01-13 15:47:17 -0800119 case MADV_NOHUGEPAGE:
Andrea Arcangeli60ab3242011-01-13 15:47:18 -0800120 error = hugepage_madvise(vma, &new_flags, behavior);
David Rientjesdef5efe2017-02-24 14:58:47 -0800121 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;
Andrea Arcangeli0af4e982011-01-13 15:46:55 -0800128 goto out;
David Rientjesdef5efe2017-02-24 14:58:47 -0800129 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -0800130 break;
Prasanna Medae798c6e2005-06-21 17:14:36 -0700131 }
132
Prasanna Meda05b74382005-06-21 17:14:37 -0700133 if (new_flags == vma->vm_flags) {
134 *prev = vma;
Hugh Dickins836d5ff2005-09-03 15:54:53 -0700135 goto out;
Prasanna Meda05b74382005-06-21 17:14:37 -0700136 }
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,
Andrea Arcangeli19a809a2015-09-04 15:46:24 -0700140 vma->vm_file, pgoff, vma_policy(vma),
141 vma->vm_userfaultfd_ctx);
Prasanna Meda05b74382005-06-21 17:14:37 -0700142 if (*prev) {
143 vma = *prev;
144 goto success;
145 }
146
147 *prev = vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148
149 if (start != vma->vm_start) {
David Rientjesdef5efe2017-02-24 14:58:47 -0800150 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
151 error = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 goto out;
David Rientjesdef5efe2017-02-24 14:58:47 -0800153 }
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 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 }
165
166 if (end != vma->vm_end) {
David Rientjesdef5efe2017-02-24 14:58:47 -0800167 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
168 error = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 goto out;
David Rientjesdef5efe2017-02-24 14:58:47 -0800170 }
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 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 }
182
Hugh Dickins836d5ff2005-09-03 15:54:53 -0700183success:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 /*
185 * vm_flags is protected by the mmap_sem held in write mode.
186 */
Prasanna Medae798c6e2005-06-21 17:14:36 -0700187 vma->vm_flags = new_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 return error;
190}
191
Shaohua Li1998cc02013-02-22 16:32:31 -0800192#ifdef CONFIG_SWAP
193static 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
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800213 if (pte_present(pte) || pte_none(pte))
Shaohua Li1998cc02013-02-22 16:32:31 -0800214 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,
Shaohua Li23955622017-07-10 15:47:11 -0700220 vma, index, false);
Shaohua Li1998cc02013-02-22 16:32:31 -0800221 if (page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300222 put_page(page);
Shaohua Li1998cc02013-02-22 16:32:31 -0800223 }
224
225 return 0;
226}
227
228static 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
242static 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
Johannes Weiner55231e52014-05-22 11:54:17 -0700253 page = find_get_entry(mapping, index);
Shaohua Li1998cc02013-02-22 16:32:31 -0800254 if (!radix_tree_exceptional_entry(page)) {
255 if (page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300256 put_page(page);
Shaohua Li1998cc02013-02-22 16:32:31 -0800257 continue;
258 }
259 swap = radix_to_swp_entry(page);
260 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
Shaohua Li23955622017-07-10 15:47:11 -0700261 NULL, 0, false);
Shaohua Li1998cc02013-02-22 16:32:31 -0800262 if (page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300263 put_page(page);
Shaohua Li1998cc02013-02-22 16:32:31 -0800264 }
265
266 lru_add_drain(); /* Push any new pages onto the LRU now */
267}
268#endif /* CONFIG_SWAP */
269
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270/*
271 * Schedule all required I/O operations. Do not wait for completion.
272 */
Vladimir Cernovec9bed92013-09-11 14:20:15 -0700273static long madvise_willneed(struct vm_area_struct *vma,
274 struct vm_area_struct **prev,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 unsigned long start, unsigned long end)
276{
277 struct file *file = vma->vm_file;
278
chenjie6ea8d952017-11-29 16:10:54 -0800279 *prev = vma;
Shaohua Li1998cc02013-02-22 16:32:31 -0800280#ifdef CONFIG_SWAP
Christoph Hellwig97b713b2015-01-14 10:42:31 +0100281 if (!file) {
Christoph Hellwig97b713b2015-01-14 10:42:31 +0100282 force_swapin_readahead(vma, start, end);
Shaohua Li1998cc02013-02-22 16:32:31 -0800283 return 0;
284 }
Shaohua Li1998cc02013-02-22 16:32:31 -0800285
Christoph Hellwig97b713b2015-01-14 10:42:31 +0100286 if (shmem_mapping(file->f_mapping)) {
Christoph Hellwig97b713b2015-01-14 10:42:31 +0100287 force_shm_swapin_readahead(vma, start, end,
288 file->f_mapping);
289 return 0;
290 }
291#else
Suzuki1bef4002005-10-11 08:29:06 -0700292 if (!file)
293 return -EBADF;
Christoph Hellwig97b713b2015-01-14 10:42:31 +0100294#endif
Suzuki1bef4002005-10-11 08:29:06 -0700295
Matthew Wilcoxe748dcd2015-02-16 15:59:12 -0800296 if (IS_DAX(file_inode(file))) {
Carsten Ottefe77ba62005-06-23 22:05:29 -0700297 /* no bad return value, but ignore advice */
298 return 0;
299 }
300
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301 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
Wu Fengguangf7e839d2009-06-16 15:31:20 -0700306 force_page_cache_readahead(file->f_mapping, file, start, end - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 return 0;
308}
309
Minchan Kim854e9ed2016-01-15 16:54:53 -0800310static 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;
Minchan Kim64b42bc2016-01-15 16:55:06 -0800320 int nr_swap = 0;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -0800321 unsigned long next;
Minchan Kim854e9ed2016-01-15 16:54:53 -0800322
Minchan Kimb8d3c4c2016-01-15 16:55:42 -0800323 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;
327
Minchan Kim854e9ed2016-01-15 16:54:53 -0800328 if (pmd_trans_unstable(pmd))
329 return 0;
330
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -0800331 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
Minchan Kim854e9ed2016-01-15 16:54:53 -0800332 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
Mel Gorman3ea27712017-08-02 13:31:52 -0700333 flush_tlb_batched_pending(mm);
Minchan Kim854e9ed2016-01-15 16:54:53 -0800334 arch_enter_lazy_mmu_mode();
335 for (; addr != end; pte++, addr += PAGE_SIZE) {
336 ptent = *pte;
337
Minchan Kim64b42bc2016-01-15 16:55:06 -0800338 if (pte_none(ptent))
Minchan Kim854e9ed2016-01-15 16:54:53 -0800339 continue;
Minchan Kim64b42bc2016-01-15 16:55:06 -0800340 /*
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 }
Minchan Kim854e9ed2016-01-15 16:54:53 -0800356
Jérôme Glissedf6ad692017-09-08 16:12:24 -0700357 page = _vm_normal_page(vma, addr, ptent, true);
Minchan Kim854e9ed2016-01-15 16:54:53 -0800358 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 }
Minchan Kim854e9ed2016-01-15 16:54:53 -0800381 unlock_page(page);
Eric Biggers263630e2017-08-25 15:55:39 -0700382 put_page(page);
Minchan Kim854e9ed2016-01-15 16:54:53 -0800383 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 }
Shaohua Li802a3a92017-05-03 14:52:32 -0700427 mark_page_lazyfree(page);
Minchan Kim854e9ed2016-01-15 16:54:53 -0800428 }
429out:
Minchan Kim64b42bc2016-01-15 16:55:06 -0800430 if (nr_swap) {
431 if (current->mm == mm)
432 sync_mm_rss(mm);
433
434 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
435 }
Minchan Kim854e9ed2016-01-15 16:54:53 -0800436 arch_leave_lazy_mmu_mode();
437 pte_unmap_unlock(orig_pte, ptl);
438 cond_resched();
Minchan Kimb8d3c4c2016-01-15 16:55:42 -0800439next:
Minchan Kim854e9ed2016-01-15 16:54:53 -0800440 return 0;
441}
442
443static 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
458static 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
Minchan Kim854e9ed2016-01-15 16:54:53 -0800465 /* 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
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488/*
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
Fernando Luis Vazquez Cao7e6cbea2008-07-29 22:33:39 -0700492 * zap_page_range call sets things up for shrink_active_list to actually free
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 * 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
Fernando Luis Vazquez Cao7e6cbea2008-07-29 22:33:39 -0700495 * shrink_active_list to pick up before reclaiming other pages.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 *
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 */
Mike Rapoport230ca982017-07-10 15:49:02 -0700507static 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
514static 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)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518{
Prasanna Meda05b74382005-06-21 17:14:37 -0700519 *prev = vma;
Kirill A. Shutemov23519072017-02-22 15:46:39 -0800520 if (!can_madv_dontneed_vma(vma))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 return -EINVAL;
522
Andrea Arcangeli70ccb922017-03-09 16:17:11 -0800523 if (!userfaultfd_remove(vma, start, end)) {
524 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
525
526 down_read(&current->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
Mike Rapoport230ca982017-07-10 15:49:02 -0700537 * virtual range passed to MADV_DONTNEED
538 * or MADV_FREE.
Andrea Arcangeli70ccb922017-03-09 16:17:11 -0800539 */
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
Mike Rapoport230ca982017-07-10 15:49:02 -0700549 * operation may not cause madvise() to
Andrea Arcangeli70ccb922017-03-09 16:17:11 -0800550 * 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 }
Mike Rapoport230ca982017-07-10 15:49:02 -0700561
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;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568}
569
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800570/*
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.
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800573 */
574static long madvise_remove(struct vm_area_struct *vma,
Nick Piggin00e9fa22007-03-16 13:38:10 -0800575 struct vm_area_struct **prev,
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800576 unsigned long start, unsigned long end)
577{
Hugh Dickins3f31d072012-05-29 15:06:40 -0700578 loff_t offset;
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700579 int error;
Andy Lutomirski9ab42332012-07-05 16:00:11 -0700580 struct file *f;
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800581
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700582 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
Nick Piggin00e9fa22007-03-16 13:38:10 -0800583
Mike Kravetz72079ba2015-09-08 15:01:57 -0700584 if (vma->vm_flags & VM_LOCKED)
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800585 return -EINVAL;
586
Andy Lutomirski9ab42332012-07-05 16:00:11 -0700587 f = vma->vm_file;
588
589 if (!f || !f->f_mapping || !f->f_mapping->host) {
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800590 return -EINVAL;
591 }
592
Hugh Dickins69cf0fa2006-04-17 22:46:32 +0100593 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
594 return -EACCES;
595
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800596 offset = (loff_t)(start - vma->vm_start)
597 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700598
Andy Lutomirski9ab42332012-07-05 16:00:11 -0700599 /*
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);
Andrea Arcangeli70ccb922017-03-09 16:17:11 -0800606 if (userfaultfd_remove(vma, start, end)) {
607 /* mmap_sem was not released by userfaultfd_remove() */
608 up_read(&current->mm->mmap_sem);
609 }
Anna Schumaker72c72bd2014-11-07 14:44:25 -0500610 error = vfs_fallocate(f,
Hugh Dickins3f31d072012-05-29 15:06:40 -0700611 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
612 offset, end - start);
Andy Lutomirski9ab42332012-07-05 16:00:11 -0700613 fput(f);
Nick Piggin0a27a142007-05-06 14:49:53 -0700614 down_read(&current->mm->mmap_sem);
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700615 return error;
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800616}
617
Andi Kleen9893e492009-09-16 11:50:17 +0200618#ifdef CONFIG_MEMORY_FAILURE
619/*
620 * Error injection support for memory error handling.
621 */
Anshuman Khandual97167a72017-05-03 14:55:25 -0700622static int madvise_inject_error(int behavior,
623 unsigned long start, unsigned long end)
Andi Kleen9893e492009-09-16 11:50:17 +0200624{
Anshuman Khandual97167a72017-05-03 14:55:25 -0700625 struct page *page;
Mel Gormanc461ad62017-08-31 16:15:30 -0700626 struct zone *zone;
Alexandru Moise19bfbe22017-10-03 16:14:31 -0700627 unsigned int order;
Anshuman Khandual97167a72017-05-03 14:55:25 -0700628
Andi Kleen9893e492009-09-16 11:50:17 +0200629 if (!capable(CAP_SYS_ADMIN))
630 return -EPERM;
Anshuman Khandual97167a72017-05-03 14:55:25 -0700631
Alexandru Moise19bfbe22017-10-03 16:14:31 -0700632
633 for (; start < end; start += PAGE_SIZE << order) {
Andrew Morton325c4ef2013-09-11 14:23:03 -0700634 int ret;
635
Anshuman Khandual97167a72017-05-03 14:55:25 -0700636 ret = get_user_pages_fast(start, 1, 0, &page);
Andi Kleen9893e492009-09-16 11:50:17 +0200637 if (ret != 1)
638 return ret;
Andrew Morton325c4ef2013-09-11 14:23:03 -0700639
Alexandru Moise19bfbe22017-10-03 16:14:31 -0700640 /*
641 * When soft offlining hugepages, after migrating the page
642 * we dissolve it, therefore in the second loop "page" will
643 * no longer be a compound page, and order will be 0.
644 */
645 order = compound_order(compound_head(page));
646
Anshuman Khandual97167a72017-05-03 14:55:25 -0700647 if (PageHWPoison(page)) {
648 put_page(page);
Wanpeng Li29b4eed2013-09-11 14:22:59 -0700649 continue;
650 }
Anshuman Khandual97167a72017-05-03 14:55:25 -0700651
652 if (behavior == MADV_SOFT_OFFLINE) {
653 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
654 page_to_pfn(page), start);
655
656 ret = soft_offline_page(page, MF_COUNT_INCREASED);
Andi Kleenafcf9382009-12-16 12:20:00 +0100657 if (ret)
Wanpeng Li83024232013-09-11 14:23:02 -0700658 return ret;
Andi Kleenafcf9382009-12-16 12:20:00 +0100659 continue;
660 }
Anshuman Khandual97167a72017-05-03 14:55:25 -0700661 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
662 page_to_pfn(page), start);
663
Eric W. Biederman83b57532017-07-09 18:14:01 -0500664 ret = memory_failure(page_to_pfn(page), MF_COUNT_INCREASED);
Naoya Horiguchi23a003b2016-03-15 14:56:36 -0700665 if (ret)
666 return ret;
Andi Kleen9893e492009-09-16 11:50:17 +0200667 }
Mel Gormanc461ad62017-08-31 16:15:30 -0700668
669 /* Ensure that all poisoned pages are removed from per-cpu lists */
670 for_each_populated_zone(zone)
671 drain_all_pages(zone);
672
Andrew Morton325c4ef2013-09-11 14:23:03 -0700673 return 0;
Andi Kleen9893e492009-09-16 11:50:17 +0200674}
675#endif
676
suzuki165cd402005-07-27 11:43:59 -0700677static long
678madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
679 unsigned long start, unsigned long end, int behavior)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 switch (behavior) {
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800682 case MADV_REMOVE:
Hugh Dickins3866ea92009-09-21 17:01:52 -0700683 return madvise_remove(vma, prev, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 case MADV_WILLNEED:
Hugh Dickins3866ea92009-09-21 17:01:52 -0700685 return madvise_willneed(vma, prev, start, end);
Minchan Kim854e9ed2016-01-15 16:54:53 -0800686 case MADV_FREE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 case MADV_DONTNEED:
Mike Rapoport230ca982017-07-10 15:49:02 -0700688 return madvise_dontneed_free(vma, prev, start, end, behavior);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 default:
Hugh Dickins3866ea92009-09-21 17:01:52 -0700690 return madvise_behavior(vma, prev, start, end, behavior);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692}
693
Nicholas Krause1ecef9e2015-09-04 15:48:24 -0700694static bool
Nick Piggin75927af2009-06-16 15:32:38 -0700695madvise_behavior_valid(int behavior)
696{
697 switch (behavior) {
698 case MADV_DOFORK:
699 case MADV_DONTFORK:
700 case MADV_NORMAL:
701 case MADV_SEQUENTIAL:
702 case MADV_RANDOM:
703 case MADV_REMOVE:
704 case MADV_WILLNEED:
705 case MADV_DONTNEED:
Minchan Kim854e9ed2016-01-15 16:54:53 -0800706 case MADV_FREE:
Hugh Dickinsf8af4da2009-09-21 17:01:57 -0700707#ifdef CONFIG_KSM
708 case MADV_MERGEABLE:
709 case MADV_UNMERGEABLE:
710#endif
Andrea Arcangeli0af4e982011-01-13 15:46:55 -0800711#ifdef CONFIG_TRANSPARENT_HUGEPAGE
712 case MADV_HUGEPAGE:
Andrea Arcangelia664b2d2011-01-13 15:47:17 -0800713 case MADV_NOHUGEPAGE:
Andrea Arcangeli0af4e982011-01-13 15:46:55 -0800714#endif
Jason Baronaccb61f2012-03-23 15:02:51 -0700715 case MADV_DONTDUMP:
716 case MADV_DODUMP:
Rik van Rield2cd9ed2017-09-06 16:25:15 -0700717 case MADV_WIPEONFORK:
718 case MADV_KEEPONFORK:
Anshuman Khandual5e451be2017-05-03 14:55:28 -0700719#ifdef CONFIG_MEMORY_FAILURE
720 case MADV_SOFT_OFFLINE:
721 case MADV_HWPOISON:
722#endif
Nicholas Krause1ecef9e2015-09-04 15:48:24 -0700723 return true;
Nick Piggin75927af2009-06-16 15:32:38 -0700724
725 default:
Nicholas Krause1ecef9e2015-09-04 15:48:24 -0700726 return false;
Nick Piggin75927af2009-06-16 15:32:38 -0700727 }
728}
Hugh Dickins3866ea92009-09-21 17:01:52 -0700729
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730/*
731 * The madvise(2) system call.
732 *
733 * Applications can use madvise() to advise the kernel how it should
734 * handle paging I/O in this VM area. The idea is to help the kernel
735 * use appropriate read-ahead and caching techniques. The information
736 * provided is advisory only, and can be safely disregarded by the
737 * kernel without affecting the correct operation of the application.
738 *
739 * behavior values:
740 * MADV_NORMAL - the default behavior is to read clusters. This
741 * results in some read-ahead and read-behind.
742 * MADV_RANDOM - the system should read the minimum amount of data
743 * on any access, since it is unlikely that the appli-
744 * cation will need more than what it asks for.
745 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
746 * once, so they can be aggressively read ahead, and
747 * can be freed soon after they are accessed.
748 * MADV_WILLNEED - the application is notifying the system to read
749 * some pages ahead.
750 * MADV_DONTNEED - the application is finished with the given range,
751 * so the kernel can free resources associated with it.
Naoya Horiguchid7206a72016-03-15 14:56:58 -0700752 * MADV_FREE - the application marks pages in the given range as lazy free,
753 * where actual purges are postponed until memory pressure happens.
Badari Pulavartyf6b3ec22006-01-06 00:10:38 -0800754 * MADV_REMOVE - the application wants to free up the given range of
755 * pages and associated backing store.
Hugh Dickins3866ea92009-09-21 17:01:52 -0700756 * MADV_DONTFORK - omit this area from child's address space when forking:
757 * typically, to avoid COWing pages pinned by get_user_pages().
758 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
Yang Shic02c3002017-10-13 15:57:37 -0700759 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
760 * range after a fork.
761 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
Naoya Horiguchid7206a72016-03-15 14:56:58 -0700762 * MADV_HWPOISON - trigger memory error handler as if the given memory range
763 * were corrupted by unrecoverable hardware memory failure.
764 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
Hugh Dickinsf8af4da2009-09-21 17:01:57 -0700765 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
766 * this area with pages of identical content from other such areas.
767 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
Naoya Horiguchid7206a72016-03-15 14:56:58 -0700768 * MADV_HUGEPAGE - the application wants to back the given range by transparent
769 * huge pages in the future. Existing pages might be coalesced and
770 * new pages might be allocated as THP.
771 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
772 * transparent huge pages so the existing pages will not be
773 * coalesced into THP and new pages will not be allocated as THP.
774 * MADV_DONTDUMP - the application wants to prevent pages in the given range
775 * from being included in its core dump.
776 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 *
778 * return values:
779 * zero - success
780 * -EINVAL - start + len < 0, start is not page-aligned,
781 * "behavior" is not a valid value, or application
Yang Shic02c3002017-10-13 15:57:37 -0700782 * is attempting to release locked or shared pages,
783 * or the specified address range includes file, Huge TLB,
784 * MAP_SHARED or VMPFNMAP range.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 * -ENOMEM - addresses in the specified range are not currently
786 * mapped, or are outside the AS of the process.
787 * -EIO - an I/O error occurred while paging in data.
788 * -EBADF - map exists, but area maps something that isn't a file.
789 * -EAGAIN - a kernel resource was temporarily unavailable.
790 */
Heiko Carstens3480b252009-01-14 14:14:16 +0100791SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792{
Prasanna Meda05b74382005-06-21 17:14:37 -0700793 unsigned long end, tmp;
Vladimir Cernovec9bed92013-09-11 14:20:15 -0700794 struct vm_area_struct *vma, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 int unmapped_error = 0;
796 int error = -EINVAL;
Jason Baronf7977792007-07-15 23:38:21 -0700797 int write;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 size_t len;
Shaohua Li1998cc02013-02-22 16:32:31 -0800799 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800
Nick Piggin75927af2009-06-16 15:32:38 -0700801 if (!madvise_behavior_valid(behavior))
802 return error;
803
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700804 if (start & ~PAGE_MASK)
805 return error;
806 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
807
808 /* Check to see whether len was rounded up from small -ve to zero */
809 if (len_in && !len)
810 return error;
811
812 end = start + len;
813 if (end < start)
814 return error;
815
816 error = 0;
817 if (end == start)
818 return error;
819
Anshuman Khandual5e451be2017-05-03 14:55:28 -0700820#ifdef CONFIG_MEMORY_FAILURE
821 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
822 return madvise_inject_error(behavior, start, start + len_in);
823#endif
824
Jason Baronf7977792007-07-15 23:38:21 -0700825 write = madvise_need_mmap_write(behavior);
Michal Hockodc0ef0d2016-05-23 16:25:27 -0700826 if (write) {
827 if (down_write_killable(&current->mm->mmap_sem))
828 return -EINTR;
829 } else {
Nick Piggin0a27a142007-05-06 14:49:53 -0700830 down_read(&current->mm->mmap_sem);
Michal Hockodc0ef0d2016-05-23 16:25:27 -0700831 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 /*
834 * If the interval [start,end) covers some unmapped address
835 * ranges, just ignore them, but return -ENOMEM at the end.
Prasanna Meda05b74382005-06-21 17:14:37 -0700836 * - different from the way of handling in mlock etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 */
Prasanna Meda05b74382005-06-21 17:14:37 -0700838 vma = find_vma_prev(current->mm, start, &prev);
Hugh Dickins836d5ff2005-09-03 15:54:53 -0700839 if (vma && start > vma->vm_start)
840 prev = vma;
841
Shaohua Li1998cc02013-02-22 16:32:31 -0800842 blk_start_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 for (;;) {
844 /* Still start < end. */
845 error = -ENOMEM;
846 if (!vma)
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700847 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848
Prasanna Meda05b74382005-06-21 17:14:37 -0700849 /* Here start < (end|vma->vm_end). */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 if (start < vma->vm_start) {
851 unmapped_error = -ENOMEM;
852 start = vma->vm_start;
Prasanna Meda05b74382005-06-21 17:14:37 -0700853 if (start >= end)
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700854 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 }
856
Prasanna Meda05b74382005-06-21 17:14:37 -0700857 /* Here vma->vm_start <= start < (end|vma->vm_end) */
858 tmp = vma->vm_end;
859 if (end < tmp)
860 tmp = end;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
Prasanna Meda05b74382005-06-21 17:14:37 -0700862 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
863 error = madvise_vma(vma, &prev, start, tmp, behavior);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 if (error)
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700865 goto out;
Prasanna Meda05b74382005-06-21 17:14:37 -0700866 start = tmp;
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700867 if (prev && start < prev->vm_end)
Prasanna Meda05b74382005-06-21 17:14:37 -0700868 start = prev->vm_end;
869 error = unmapped_error;
870 if (start >= end)
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700871 goto out;
Hugh Dickins90ed52e2007-03-29 01:20:38 -0700872 if (prev)
873 vma = prev->vm_next;
874 else /* madvise_remove dropped mmap_sem */
875 vma = find_vma(current->mm, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877out:
Rasmus Villemoes84d96d82013-04-29 15:08:23 -0700878 blk_finish_plug(&plug);
Jason Baronf7977792007-07-15 23:38:21 -0700879 if (write)
Nick Piggin0a27a142007-05-06 14:49:53 -0700880 up_write(&current->mm->mmap_sem);
881 else
882 up_read(&current->mm->mmap_sem);
883
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 return error;
885}