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>
20 * Any behaviour which results in changes to the vma->vm_flags needs to
21 * take mmap_sem for writing. Others, which simply traverse vmas, need
22 * to only take it for reading.
24 static int madvise_need_mmap_write(int behavior)
32 /* be safe, default to 1. list exceptions explicitly */
38 * We can potentially split a vm area into separate
39 * areas, each area with its own behavior.
41 static long madvise_behavior(struct vm_area_struct * vma,
42 struct vm_area_struct **prev,
43 unsigned long start, unsigned long end, int behavior)
45 struct mm_struct * mm = vma->vm_mm;
48 unsigned long new_flags = vma->vm_flags;
52 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
55 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
58 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
61 new_flags |= VM_DONTCOPY;
64 if (vma->vm_flags & VM_IO) {
68 new_flags &= ~VM_DONTCOPY;
71 new_flags |= VM_NODUMP;
74 new_flags &= ~VM_NODUMP;
77 case MADV_UNMERGEABLE:
78 error = ksm_madvise(vma, start, end, behavior, &new_flags);
84 error = hugepage_madvise(vma, &new_flags, behavior);
90 if (new_flags == vma->vm_flags) {
95 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
96 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
97 vma->vm_file, pgoff, vma_policy(vma));
105 if (start != vma->vm_start) {
106 error = split_vma(mm, vma, start, 1);
111 if (end != vma->vm_end) {
112 error = split_vma(mm, vma, end, 0);
119 * vm_flags is protected by the mmap_sem held in write mode.
121 vma->vm_flags = new_flags;
124 if (error == -ENOMEM)
130 * Schedule all required I/O operations. Do not wait for completion.
132 static long madvise_willneed(struct vm_area_struct * vma,
133 struct vm_area_struct ** prev,
134 unsigned long start, unsigned long end)
136 struct file *file = vma->vm_file;
141 if (file->f_mapping->a_ops->get_xip_mem) {
142 /* no bad return value, but ignore advice */
147 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
148 if (end > vma->vm_end)
150 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
152 force_page_cache_readahead(file->f_mapping, file, start, end - start);
157 * Application no longer needs these pages. If the pages are dirty,
158 * it's OK to just throw them away. The app will be more careful about
159 * data it wants to keep. Be sure to free swap resources too. The
160 * zap_page_range call sets things up for shrink_active_list to actually free
161 * these pages later if no one else has touched them in the meantime,
162 * although we could add these pages to a global reuse list for
163 * shrink_active_list to pick up before reclaiming other pages.
165 * NB: This interface discards data rather than pushes it out to swap,
166 * as some implementations do. This has performance implications for
167 * applications like large transactional databases which want to discard
168 * pages in anonymous maps after committing to backing store the data
169 * that was kept in them. There is no reason to write this data out to
170 * the swap area if the application is discarding it.
172 * An interface that causes the system to free clean pages and flush
173 * dirty pages is already available as msync(MS_INVALIDATE).
175 static long madvise_dontneed(struct vm_area_struct * vma,
176 struct vm_area_struct ** prev,
177 unsigned long start, unsigned long end)
180 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
183 if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
184 struct zap_details details = {
185 .nonlinear_vma = vma,
186 .last_index = ULONG_MAX,
188 zap_page_range(vma, start, end - start, &details);
190 zap_page_range(vma, start, end - start, NULL);
195 * Application wants to free up the pages and associated backing store.
196 * This is effectively punching a hole into the middle of a file.
198 * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
199 * Other filesystems return -ENOSYS.
201 static long madvise_remove(struct vm_area_struct *vma,
202 struct vm_area_struct **prev,
203 unsigned long start, unsigned long end)
208 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
210 if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
213 if (!vma->vm_file || !vma->vm_file->f_mapping
214 || !vma->vm_file->f_mapping->host) {
218 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
221 offset = (loff_t)(start - vma->vm_start)
222 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
224 /* filesystem's fallocate may need to take i_mutex */
225 up_read(¤t->mm->mmap_sem);
226 error = do_fallocate(vma->vm_file,
227 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
228 offset, end - start);
229 down_read(¤t->mm->mmap_sem);
233 #ifdef CONFIG_MEMORY_FAILURE
235 * Error injection support for memory error handling.
237 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
241 if (!capable(CAP_SYS_ADMIN))
243 for (; start < end; start += PAGE_SIZE) {
245 int ret = get_user_pages_fast(start, 1, 0, &p);
248 if (bhv == MADV_SOFT_OFFLINE) {
249 printk(KERN_INFO "Soft offlining page %lx at %lx\n",
250 page_to_pfn(p), start);
251 ret = soft_offline_page(p, MF_COUNT_INCREASED);
256 printk(KERN_INFO "Injecting memory failure for page %lx at %lx\n",
257 page_to_pfn(p), start);
258 /* Ignore return value for now */
259 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
266 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
267 unsigned long start, unsigned long end, int behavior)
271 return madvise_remove(vma, prev, start, end);
273 return madvise_willneed(vma, prev, start, end);
275 return madvise_dontneed(vma, prev, start, end);
277 return madvise_behavior(vma, prev, start, end, behavior);
282 madvise_behavior_valid(int behavior)
288 case MADV_SEQUENTIAL:
295 case MADV_UNMERGEABLE:
297 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
299 case MADV_NOHUGEPAGE:
311 * The madvise(2) system call.
313 * Applications can use madvise() to advise the kernel how it should
314 * handle paging I/O in this VM area. The idea is to help the kernel
315 * use appropriate read-ahead and caching techniques. The information
316 * provided is advisory only, and can be safely disregarded by the
317 * kernel without affecting the correct operation of the application.
320 * MADV_NORMAL - the default behavior is to read clusters. This
321 * results in some read-ahead and read-behind.
322 * MADV_RANDOM - the system should read the minimum amount of data
323 * on any access, since it is unlikely that the appli-
324 * cation will need more than what it asks for.
325 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
326 * once, so they can be aggressively read ahead, and
327 * can be freed soon after they are accessed.
328 * MADV_WILLNEED - the application is notifying the system to read
330 * MADV_DONTNEED - the application is finished with the given range,
331 * so the kernel can free resources associated with it.
332 * MADV_REMOVE - the application wants to free up the given range of
333 * pages and associated backing store.
334 * MADV_DONTFORK - omit this area from child's address space when forking:
335 * typically, to avoid COWing pages pinned by get_user_pages().
336 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
337 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
338 * this area with pages of identical content from other such areas.
339 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
343 * -EINVAL - start + len < 0, start is not page-aligned,
344 * "behavior" is not a valid value, or application
345 * is attempting to release locked or shared pages.
346 * -ENOMEM - addresses in the specified range are not currently
347 * mapped, or are outside the AS of the process.
348 * -EIO - an I/O error occurred while paging in data.
349 * -EBADF - map exists, but area maps something that isn't a file.
350 * -EAGAIN - a kernel resource was temporarily unavailable.
352 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
354 unsigned long end, tmp;
355 struct vm_area_struct * vma, *prev;
356 int unmapped_error = 0;
361 #ifdef CONFIG_MEMORY_FAILURE
362 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
363 return madvise_hwpoison(behavior, start, start+len_in);
365 if (!madvise_behavior_valid(behavior))
368 write = madvise_need_mmap_write(behavior);
370 down_write(¤t->mm->mmap_sem);
372 down_read(¤t->mm->mmap_sem);
374 if (start & ~PAGE_MASK)
376 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
378 /* Check to see whether len was rounded up from small -ve to zero */
391 * If the interval [start,end) covers some unmapped address
392 * ranges, just ignore them, but return -ENOMEM at the end.
393 * - different from the way of handling in mlock etc.
395 vma = find_vma_prev(current->mm, start, &prev);
396 if (vma && start > vma->vm_start)
400 /* Still start < end. */
405 /* Here start < (end|vma->vm_end). */
406 if (start < vma->vm_start) {
407 unmapped_error = -ENOMEM;
408 start = vma->vm_start;
413 /* Here vma->vm_start <= start < (end|vma->vm_end) */
418 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
419 error = madvise_vma(vma, &prev, start, tmp, behavior);
423 if (prev && start < prev->vm_end)
424 start = prev->vm_end;
425 error = unmapped_error;
430 else /* madvise_remove dropped mmap_sem */
431 vma = find_vma(current->mm, start);
435 up_write(¤t->mm->mmap_sem);
437 up_read(¤t->mm->mmap_sem);