4 * Copyright (C) 1999 Linus Torvalds
5 * Copyright (C) 2002 Christoph Hellwig
8 #include <linux/mman.h>
9 #include <linux/pagemap.h>
10 #include <linux/syscalls.h>
11 #include <linux/mempolicy.h>
12 #include <linux/page-isolation.h>
13 #include <linux/hugetlb.h>
14 #include <linux/falloc.h>
15 #include <linux/sched.h>
16 #include <linux/ksm.h>
18 #include <linux/file.h>
19 #include <linux/blkdev.h>
20 #include <linux/backing-dev.h>
21 #include <linux/swap.h>
22 #include <linux/swapops.h>
25 * Any behaviour which results in changes to the vma->vm_flags needs to
26 * take mmap_sem for writing. Others, which simply traverse vmas, need
27 * to only take it for reading.
29 static int madvise_need_mmap_write(int behavior)
37 /* be safe, default to 1. list exceptions explicitly */
43 * We can potentially split a vm area into separate
44 * areas, each area with its own behavior.
46 static long madvise_behavior(struct vm_area_struct *vma,
47 struct vm_area_struct **prev,
48 unsigned long start, unsigned long end, int behavior)
50 struct mm_struct *mm = vma->vm_mm;
53 unsigned long new_flags = vma->vm_flags;
57 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
60 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
63 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
66 new_flags |= VM_DONTCOPY;
69 if (vma->vm_flags & VM_IO) {
73 new_flags &= ~VM_DONTCOPY;
76 new_flags |= VM_DONTDUMP;
79 if (new_flags & VM_SPECIAL) {
83 new_flags &= ~VM_DONTDUMP;
86 case MADV_UNMERGEABLE:
87 error = ksm_madvise(vma, start, end, behavior, &new_flags);
93 error = hugepage_madvise(vma, &new_flags, behavior);
99 if (new_flags == vma->vm_flags) {
104 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
105 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
106 vma->vm_file, pgoff, vma_policy(vma),
107 vma->vm_userfaultfd_ctx);
115 if (start != vma->vm_start) {
116 error = split_vma(mm, vma, start, 1);
121 if (end != vma->vm_end) {
122 error = split_vma(mm, vma, end, 0);
129 * vm_flags is protected by the mmap_sem held in write mode.
131 vma->vm_flags = new_flags;
134 if (error == -ENOMEM)
140 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
141 unsigned long end, struct mm_walk *walk)
144 struct vm_area_struct *vma = walk->private;
147 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
150 for (index = start; index != end; index += PAGE_SIZE) {
156 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
157 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
158 pte_unmap_unlock(orig_pte, ptl);
160 if (pte_present(pte) || pte_none(pte))
162 entry = pte_to_swp_entry(pte);
163 if (unlikely(non_swap_entry(entry)))
166 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
169 page_cache_release(page);
175 static void force_swapin_readahead(struct vm_area_struct *vma,
176 unsigned long start, unsigned long end)
178 struct mm_walk walk = {
180 .pmd_entry = swapin_walk_pmd_entry,
184 walk_page_range(start, end, &walk);
186 lru_add_drain(); /* Push any new pages onto the LRU now */
189 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
190 unsigned long start, unsigned long end,
191 struct address_space *mapping)
197 for (; start < end; start += PAGE_SIZE) {
198 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
200 page = find_get_entry(mapping, index);
201 if (!radix_tree_exceptional_entry(page)) {
203 page_cache_release(page);
206 swap = radix_to_swp_entry(page);
207 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
210 page_cache_release(page);
213 lru_add_drain(); /* Push any new pages onto the LRU now */
215 #endif /* CONFIG_SWAP */
218 * Schedule all required I/O operations. Do not wait for completion.
220 static long madvise_willneed(struct vm_area_struct *vma,
221 struct vm_area_struct **prev,
222 unsigned long start, unsigned long end)
224 struct file *file = vma->vm_file;
229 force_swapin_readahead(vma, start, end);
233 if (shmem_mapping(file->f_mapping)) {
235 force_shm_swapin_readahead(vma, start, end,
244 if (IS_DAX(file_inode(file))) {
245 /* no bad return value, but ignore advice */
250 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
251 if (end > vma->vm_end)
253 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
255 force_page_cache_readahead(file->f_mapping, file, start, end - start);
260 * Application no longer needs these pages. If the pages are dirty,
261 * it's OK to just throw them away. The app will be more careful about
262 * data it wants to keep. Be sure to free swap resources too. The
263 * zap_page_range call sets things up for shrink_active_list to actually free
264 * these pages later if no one else has touched them in the meantime,
265 * although we could add these pages to a global reuse list for
266 * shrink_active_list to pick up before reclaiming other pages.
268 * NB: This interface discards data rather than pushes it out to swap,
269 * as some implementations do. This has performance implications for
270 * applications like large transactional databases which want to discard
271 * pages in anonymous maps after committing to backing store the data
272 * that was kept in them. There is no reason to write this data out to
273 * the swap area if the application is discarding it.
275 * An interface that causes the system to free clean pages and flush
276 * dirty pages is already available as msync(MS_INVALIDATE).
278 static long madvise_dontneed(struct vm_area_struct *vma,
279 struct vm_area_struct **prev,
280 unsigned long start, unsigned long end)
283 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
286 zap_page_range(vma, start, end - start, NULL);
291 * Application wants to free up the pages and associated backing store.
292 * This is effectively punching a hole into the middle of a file.
294 static long madvise_remove(struct vm_area_struct *vma,
295 struct vm_area_struct **prev,
296 unsigned long start, unsigned long end)
302 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
304 if (vma->vm_flags & VM_LOCKED)
309 if (!f || !f->f_mapping || !f->f_mapping->host) {
313 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
316 offset = (loff_t)(start - vma->vm_start)
317 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
320 * Filesystem's fallocate may need to take i_mutex. We need to
321 * explicitly grab a reference because the vma (and hence the
322 * vma's reference to the file) can go away as soon as we drop
326 up_read(¤t->mm->mmap_sem);
327 error = vfs_fallocate(f,
328 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
329 offset, end - start);
331 down_read(¤t->mm->mmap_sem);
335 #ifdef CONFIG_MEMORY_FAILURE
337 * Error injection support for memory error handling.
339 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
342 if (!capable(CAP_SYS_ADMIN))
344 for (; start < end; start += PAGE_SIZE <<
345 compound_order(compound_head(p))) {
348 ret = get_user_pages_fast(start, 1, 0, &p);
352 if (PageHWPoison(p)) {
356 if (bhv == MADV_SOFT_OFFLINE) {
357 pr_info("Soft offlining page %#lx at %#lx\n",
358 page_to_pfn(p), start);
359 ret = soft_offline_page(p, MF_COUNT_INCREASED);
364 pr_info("Injecting memory failure for page %#lx at %#lx\n",
365 page_to_pfn(p), start);
366 /* Ignore return value for now */
367 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
374 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
375 unsigned long start, unsigned long end, int behavior)
379 return madvise_remove(vma, prev, start, end);
381 return madvise_willneed(vma, prev, start, end);
383 return madvise_dontneed(vma, prev, start, end);
385 return madvise_behavior(vma, prev, start, end, behavior);
390 madvise_behavior_valid(int behavior)
396 case MADV_SEQUENTIAL:
403 case MADV_UNMERGEABLE:
405 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
407 case MADV_NOHUGEPAGE:
419 * The madvise(2) system call.
421 * Applications can use madvise() to advise the kernel how it should
422 * handle paging I/O in this VM area. The idea is to help the kernel
423 * use appropriate read-ahead and caching techniques. The information
424 * provided is advisory only, and can be safely disregarded by the
425 * kernel without affecting the correct operation of the application.
428 * MADV_NORMAL - the default behavior is to read clusters. This
429 * results in some read-ahead and read-behind.
430 * MADV_RANDOM - the system should read the minimum amount of data
431 * on any access, since it is unlikely that the appli-
432 * cation will need more than what it asks for.
433 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
434 * once, so they can be aggressively read ahead, and
435 * can be freed soon after they are accessed.
436 * MADV_WILLNEED - the application is notifying the system to read
438 * MADV_DONTNEED - the application is finished with the given range,
439 * so the kernel can free resources associated with it.
440 * MADV_REMOVE - the application wants to free up the given range of
441 * pages and associated backing store.
442 * MADV_DONTFORK - omit this area from child's address space when forking:
443 * typically, to avoid COWing pages pinned by get_user_pages().
444 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
445 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
446 * this area with pages of identical content from other such areas.
447 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
451 * -EINVAL - start + len < 0, start is not page-aligned,
452 * "behavior" is not a valid value, or application
453 * is attempting to release locked or shared pages.
454 * -ENOMEM - addresses in the specified range are not currently
455 * mapped, or are outside the AS of the process.
456 * -EIO - an I/O error occurred while paging in data.
457 * -EBADF - map exists, but area maps something that isn't a file.
458 * -EAGAIN - a kernel resource was temporarily unavailable.
460 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
462 unsigned long end, tmp;
463 struct vm_area_struct *vma, *prev;
464 int unmapped_error = 0;
468 struct blk_plug plug;
470 #ifdef CONFIG_MEMORY_FAILURE
471 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
472 return madvise_hwpoison(behavior, start, start+len_in);
474 if (!madvise_behavior_valid(behavior))
477 if (start & ~PAGE_MASK)
479 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
481 /* Check to see whether len was rounded up from small -ve to zero */
493 write = madvise_need_mmap_write(behavior);
495 down_write(¤t->mm->mmap_sem);
497 down_read(¤t->mm->mmap_sem);
500 * If the interval [start,end) covers some unmapped address
501 * ranges, just ignore them, but return -ENOMEM at the end.
502 * - different from the way of handling in mlock etc.
504 vma = find_vma_prev(current->mm, start, &prev);
505 if (vma && start > vma->vm_start)
508 blk_start_plug(&plug);
510 /* Still start < end. */
515 /* Here start < (end|vma->vm_end). */
516 if (start < vma->vm_start) {
517 unmapped_error = -ENOMEM;
518 start = vma->vm_start;
523 /* Here vma->vm_start <= start < (end|vma->vm_end) */
528 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
529 error = madvise_vma(vma, &prev, start, tmp, behavior);
533 if (prev && start < prev->vm_end)
534 start = prev->vm_end;
535 error = unmapped_error;
540 else /* madvise_remove dropped mmap_sem */
541 vma = find_vma(current->mm, start);
544 blk_finish_plug(&plug);
546 up_write(¤t->mm->mmap_sem);
548 up_read(¤t->mm->mmap_sem);