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/swap.h>
21 #include <linux/swapops.h>
24 * Any behaviour which results in changes to the vma->vm_flags needs to
25 * take mmap_sem for writing. Others, which simply traverse vmas, need
26 * to only take it for reading.
28 static int madvise_need_mmap_write(int behavior)
36 /* be safe, default to 1. list exceptions explicitly */
42 * We can potentially split a vm area into separate
43 * areas, each area with its own behavior.
45 static long madvise_behavior(struct vm_area_struct *vma,
46 struct vm_area_struct **prev,
47 unsigned long start, unsigned long end, int behavior)
49 struct mm_struct *mm = vma->vm_mm;
52 unsigned long new_flags = vma->vm_flags;
56 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
59 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
62 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
65 new_flags |= VM_DONTCOPY;
68 if (vma->vm_flags & VM_IO) {
72 new_flags &= ~VM_DONTCOPY;
75 new_flags |= VM_DONTDUMP;
78 if (new_flags & VM_SPECIAL) {
82 new_flags &= ~VM_DONTDUMP;
85 case MADV_UNMERGEABLE:
86 error = ksm_madvise(vma, start, end, behavior, &new_flags);
92 error = hugepage_madvise(vma, &new_flags, behavior);
98 if (new_flags == vma->vm_flags) {
103 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
104 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
105 vma->vm_file, pgoff, vma_policy(vma));
113 if (start != vma->vm_start) {
114 error = split_vma(mm, vma, start, 1);
119 if (end != vma->vm_end) {
120 error = split_vma(mm, vma, end, 0);
127 * vm_flags is protected by the mmap_sem held in write mode.
129 vma->vm_flags = new_flags;
132 if (error == -ENOMEM)
138 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
139 unsigned long end, struct mm_walk *walk)
142 struct vm_area_struct *vma = walk->private;
145 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
148 for (index = start; index != end; index += PAGE_SIZE) {
154 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
155 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
156 pte_unmap_unlock(orig_pte, ptl);
158 if (pte_present(pte) || pte_none(pte))
160 entry = pte_to_swp_entry(pte);
161 if (unlikely(non_swap_entry(entry)))
164 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
167 page_cache_release(page);
173 static void force_swapin_readahead(struct vm_area_struct *vma,
174 unsigned long start, unsigned long end)
176 struct mm_walk walk = {
178 .pmd_entry = swapin_walk_pmd_entry,
182 walk_page_range(start, end, &walk);
184 lru_add_drain(); /* Push any new pages onto the LRU now */
187 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
188 unsigned long start, unsigned long end,
189 struct address_space *mapping)
195 for (; start < end; start += PAGE_SIZE) {
196 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
198 page = find_get_entry(mapping, index);
199 if (!radix_tree_exceptional_entry(page)) {
201 page_cache_release(page);
204 swap = radix_to_swp_entry(page);
205 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
208 page_cache_release(page);
211 lru_add_drain(); /* Push any new pages onto the LRU now */
213 #endif /* CONFIG_SWAP */
216 * Schedule all required I/O operations. Do not wait for completion.
218 static long madvise_willneed(struct vm_area_struct *vma,
219 struct vm_area_struct **prev,
220 unsigned long start, unsigned long end)
222 struct file *file = vma->vm_file;
225 if (!file || mapping_cap_swap_backed(file->f_mapping)) {
228 force_swapin_readahead(vma, start, end);
230 force_shm_swapin_readahead(vma, start, end,
239 if (file->f_mapping->a_ops->get_xip_mem) {
240 /* no bad return value, but ignore advice */
245 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
246 if (end > vma->vm_end)
248 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
250 force_page_cache_readahead(file->f_mapping, file, start, end - start);
255 * Application no longer needs these pages. If the pages are dirty,
256 * it's OK to just throw them away. The app will be more careful about
257 * data it wants to keep. Be sure to free swap resources too. The
258 * zap_page_range call sets things up for shrink_active_list to actually free
259 * these pages later if no one else has touched them in the meantime,
260 * although we could add these pages to a global reuse list for
261 * shrink_active_list to pick up before reclaiming other pages.
263 * NB: This interface discards data rather than pushes it out to swap,
264 * as some implementations do. This has performance implications for
265 * applications like large transactional databases which want to discard
266 * pages in anonymous maps after committing to backing store the data
267 * that was kept in them. There is no reason to write this data out to
268 * the swap area if the application is discarding it.
270 * An interface that causes the system to free clean pages and flush
271 * dirty pages is already available as msync(MS_INVALIDATE).
273 static long madvise_dontneed(struct vm_area_struct *vma,
274 struct vm_area_struct **prev,
275 unsigned long start, unsigned long end)
278 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
281 zap_page_range(vma, start, end - start, NULL);
286 * Application wants to free up the pages and associated backing store.
287 * This is effectively punching a hole into the middle of a file.
289 static long madvise_remove(struct vm_area_struct *vma,
290 struct vm_area_struct **prev,
291 unsigned long start, unsigned long end)
297 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
299 if (vma->vm_flags & (VM_LOCKED | VM_HUGETLB))
304 if (!f || !f->f_mapping || !f->f_mapping->host) {
308 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
311 offset = (loff_t)(start - vma->vm_start)
312 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
315 * Filesystem's fallocate may need to take i_mutex. We need to
316 * explicitly grab a reference because the vma (and hence the
317 * vma's reference to the file) can go away as soon as we drop
321 up_read(¤t->mm->mmap_sem);
322 error = vfs_fallocate(f,
323 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
324 offset, end - start);
326 down_read(¤t->mm->mmap_sem);
330 #ifdef CONFIG_MEMORY_FAILURE
332 * Error injection support for memory error handling.
334 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
337 if (!capable(CAP_SYS_ADMIN))
339 for (; start < end; start += PAGE_SIZE <<
340 compound_order(compound_head(p))) {
343 ret = get_user_pages_fast(start, 1, 0, &p);
347 if (PageHWPoison(p)) {
351 if (bhv == MADV_SOFT_OFFLINE) {
352 pr_info("Soft offlining page %#lx at %#lx\n",
353 page_to_pfn(p), start);
354 ret = soft_offline_page(p, MF_COUNT_INCREASED);
359 pr_info("Injecting memory failure for page %#lx at %#lx\n",
360 page_to_pfn(p), start);
361 /* Ignore return value for now */
362 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
369 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
370 unsigned long start, unsigned long end, int behavior)
374 return madvise_remove(vma, prev, start, end);
376 return madvise_willneed(vma, prev, start, end);
378 return madvise_dontneed(vma, prev, start, end);
380 return madvise_behavior(vma, prev, start, end, behavior);
385 madvise_behavior_valid(int behavior)
391 case MADV_SEQUENTIAL:
398 case MADV_UNMERGEABLE:
400 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
402 case MADV_NOHUGEPAGE:
414 * The madvise(2) system call.
416 * Applications can use madvise() to advise the kernel how it should
417 * handle paging I/O in this VM area. The idea is to help the kernel
418 * use appropriate read-ahead and caching techniques. The information
419 * provided is advisory only, and can be safely disregarded by the
420 * kernel without affecting the correct operation of the application.
423 * MADV_NORMAL - the default behavior is to read clusters. This
424 * results in some read-ahead and read-behind.
425 * MADV_RANDOM - the system should read the minimum amount of data
426 * on any access, since it is unlikely that the appli-
427 * cation will need more than what it asks for.
428 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
429 * once, so they can be aggressively read ahead, and
430 * can be freed soon after they are accessed.
431 * MADV_WILLNEED - the application is notifying the system to read
433 * MADV_DONTNEED - the application is finished with the given range,
434 * so the kernel can free resources associated with it.
435 * MADV_REMOVE - the application wants to free up the given range of
436 * pages and associated backing store.
437 * MADV_DONTFORK - omit this area from child's address space when forking:
438 * typically, to avoid COWing pages pinned by get_user_pages().
439 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
440 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
441 * this area with pages of identical content from other such areas.
442 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
446 * -EINVAL - start + len < 0, start is not page-aligned,
447 * "behavior" is not a valid value, or application
448 * is attempting to release locked or shared pages.
449 * -ENOMEM - addresses in the specified range are not currently
450 * mapped, or are outside the AS of the process.
451 * -EIO - an I/O error occurred while paging in data.
452 * -EBADF - map exists, but area maps something that isn't a file.
453 * -EAGAIN - a kernel resource was temporarily unavailable.
455 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
457 unsigned long end, tmp;
458 struct vm_area_struct *vma, *prev;
459 int unmapped_error = 0;
463 struct blk_plug plug;
465 #ifdef CONFIG_MEMORY_FAILURE
466 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
467 return madvise_hwpoison(behavior, start, start+len_in);
469 if (!madvise_behavior_valid(behavior))
472 if (start & ~PAGE_MASK)
474 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
476 /* Check to see whether len was rounded up from small -ve to zero */
488 write = madvise_need_mmap_write(behavior);
490 down_write(¤t->mm->mmap_sem);
492 down_read(¤t->mm->mmap_sem);
495 * If the interval [start,end) covers some unmapped address
496 * ranges, just ignore them, but return -ENOMEM at the end.
497 * - different from the way of handling in mlock etc.
499 vma = find_vma_prev(current->mm, start, &prev);
500 if (vma && start > vma->vm_start)
503 blk_start_plug(&plug);
505 /* Still start < end. */
510 /* Here start < (end|vma->vm_end). */
511 if (start < vma->vm_start) {
512 unmapped_error = -ENOMEM;
513 start = vma->vm_start;
518 /* Here vma->vm_start <= start < (end|vma->vm_end) */
523 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
524 error = madvise_vma(vma, &prev, start, tmp, behavior);
528 if (prev && start < prev->vm_end)
529 start = prev->vm_end;
530 error = unmapped_error;
535 else /* madvise_remove dropped mmap_sem */
536 vma = find_vma(current->mm, start);
539 blk_finish_plug(&plug);
541 up_write(¤t->mm->mmap_sem);
543 up_read(¤t->mm->mmap_sem);