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/hugetlb.h>
13 #include <linux/sched.h>
14 #include <linux/ksm.h>
17 * Any behaviour which results in changes to the vma->vm_flags needs to
18 * take mmap_sem for writing. Others, which simply traverse vmas, need
19 * to only take it for reading.
21 static int madvise_need_mmap_write(int behavior)
29 /* be safe, default to 1. list exceptions explicitly */
35 * We can potentially split a vm area into separate
36 * areas, each area with its own behavior.
38 static long madvise_behavior(struct vm_area_struct * vma,
39 struct vm_area_struct **prev,
40 unsigned long start, unsigned long end, int behavior)
42 struct mm_struct * mm = vma->vm_mm;
45 unsigned long new_flags = vma->vm_flags;
49 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
52 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
55 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
58 new_flags |= VM_DONTCOPY;
61 if (vma->vm_flags & VM_IO) {
65 new_flags &= ~VM_DONTCOPY;
68 case MADV_UNMERGEABLE:
69 error = ksm_madvise(vma, start, end, behavior, &new_flags);
75 if (new_flags == vma->vm_flags) {
80 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
81 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
82 vma->vm_file, pgoff, vma_policy(vma));
90 if (start != vma->vm_start) {
91 error = split_vma(mm, vma, start, 1);
96 if (end != vma->vm_end) {
97 error = split_vma(mm, vma, end, 0);
104 * vm_flags is protected by the mmap_sem held in write mode.
106 vma->vm_flags = new_flags;
109 if (error == -ENOMEM)
115 * Schedule all required I/O operations. Do not wait for completion.
117 static long madvise_willneed(struct vm_area_struct * vma,
118 struct vm_area_struct ** prev,
119 unsigned long start, unsigned long end)
121 struct file *file = vma->vm_file;
126 if (file->f_mapping->a_ops->get_xip_mem) {
127 /* no bad return value, but ignore advice */
132 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
133 if (end > vma->vm_end)
135 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
137 force_page_cache_readahead(file->f_mapping, file, start, end - start);
142 * Application no longer needs these pages. If the pages are dirty,
143 * it's OK to just throw them away. The app will be more careful about
144 * data it wants to keep. Be sure to free swap resources too. The
145 * zap_page_range call sets things up for shrink_active_list to actually free
146 * these pages later if no one else has touched them in the meantime,
147 * although we could add these pages to a global reuse list for
148 * shrink_active_list to pick up before reclaiming other pages.
150 * NB: This interface discards data rather than pushes it out to swap,
151 * as some implementations do. This has performance implications for
152 * applications like large transactional databases which want to discard
153 * pages in anonymous maps after committing to backing store the data
154 * that was kept in them. There is no reason to write this data out to
155 * the swap area if the application is discarding it.
157 * An interface that causes the system to free clean pages and flush
158 * dirty pages is already available as msync(MS_INVALIDATE).
160 static long madvise_dontneed(struct vm_area_struct * vma,
161 struct vm_area_struct ** prev,
162 unsigned long start, unsigned long end)
165 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
168 if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
169 struct zap_details details = {
170 .nonlinear_vma = vma,
171 .last_index = ULONG_MAX,
173 zap_page_range(vma, start, end - start, &details);
175 zap_page_range(vma, start, end - start, NULL);
180 * Application wants to free up the pages and associated backing store.
181 * This is effectively punching a hole into the middle of a file.
183 * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
184 * Other filesystems return -ENOSYS.
186 static long madvise_remove(struct vm_area_struct *vma,
187 struct vm_area_struct **prev,
188 unsigned long start, unsigned long end)
190 struct address_space *mapping;
191 loff_t offset, endoff;
194 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
196 if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
199 if (!vma->vm_file || !vma->vm_file->f_mapping
200 || !vma->vm_file->f_mapping->host) {
204 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
207 mapping = vma->vm_file->f_mapping;
209 offset = (loff_t)(start - vma->vm_start)
210 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
211 endoff = (loff_t)(end - vma->vm_start - 1)
212 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
214 /* vmtruncate_range needs to take i_mutex and i_alloc_sem */
215 up_read(¤t->mm->mmap_sem);
216 error = vmtruncate_range(mapping->host, offset, endoff);
217 down_read(¤t->mm->mmap_sem);
221 #ifdef CONFIG_MEMORY_FAILURE
223 * Error injection support for memory error handling.
225 static int madvise_hwpoison(unsigned long start, unsigned long end)
229 if (!capable(CAP_SYS_ADMIN))
231 for (; start < end; start += PAGE_SIZE) {
233 int ret = get_user_pages(current, current->mm, start, 1,
237 printk(KERN_INFO "Injecting memory failure for page %lx at %lx\n",
238 page_to_pfn(p), start);
239 /* Ignore return value for now */
240 __memory_failure(page_to_pfn(p), 0, 1);
248 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
249 unsigned long start, unsigned long end, int behavior)
253 return madvise_remove(vma, prev, start, end);
255 return madvise_willneed(vma, prev, start, end);
257 return madvise_dontneed(vma, prev, start, end);
259 return madvise_behavior(vma, prev, start, end, behavior);
264 madvise_behavior_valid(int behavior)
270 case MADV_SEQUENTIAL:
277 case MADV_UNMERGEABLE:
287 * The madvise(2) system call.
289 * Applications can use madvise() to advise the kernel how it should
290 * handle paging I/O in this VM area. The idea is to help the kernel
291 * use appropriate read-ahead and caching techniques. The information
292 * provided is advisory only, and can be safely disregarded by the
293 * kernel without affecting the correct operation of the application.
296 * MADV_NORMAL - the default behavior is to read clusters. This
297 * results in some read-ahead and read-behind.
298 * MADV_RANDOM - the system should read the minimum amount of data
299 * on any access, since it is unlikely that the appli-
300 * cation will need more than what it asks for.
301 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
302 * once, so they can be aggressively read ahead, and
303 * can be freed soon after they are accessed.
304 * MADV_WILLNEED - the application is notifying the system to read
306 * MADV_DONTNEED - the application is finished with the given range,
307 * so the kernel can free resources associated with it.
308 * MADV_REMOVE - the application wants to free up the given range of
309 * pages and associated backing store.
310 * MADV_DONTFORK - omit this area from child's address space when forking:
311 * typically, to avoid COWing pages pinned by get_user_pages().
312 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
313 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
314 * this area with pages of identical content from other such areas.
315 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
319 * -EINVAL - start + len < 0, start is not page-aligned,
320 * "behavior" is not a valid value, or application
321 * is attempting to release locked or shared pages.
322 * -ENOMEM - addresses in the specified range are not currently
323 * mapped, or are outside the AS of the process.
324 * -EIO - an I/O error occurred while paging in data.
325 * -EBADF - map exists, but area maps something that isn't a file.
326 * -EAGAIN - a kernel resource was temporarily unavailable.
328 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
330 unsigned long end, tmp;
331 struct vm_area_struct * vma, *prev;
332 int unmapped_error = 0;
337 #ifdef CONFIG_MEMORY_FAILURE
338 if (behavior == MADV_HWPOISON)
339 return madvise_hwpoison(start, start+len_in);
341 if (!madvise_behavior_valid(behavior))
344 write = madvise_need_mmap_write(behavior);
346 down_write(¤t->mm->mmap_sem);
348 down_read(¤t->mm->mmap_sem);
350 if (start & ~PAGE_MASK)
352 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
354 /* Check to see whether len was rounded up from small -ve to zero */
367 * If the interval [start,end) covers some unmapped address
368 * ranges, just ignore them, but return -ENOMEM at the end.
369 * - different from the way of handling in mlock etc.
371 vma = find_vma_prev(current->mm, start, &prev);
372 if (vma && start > vma->vm_start)
376 /* Still start < end. */
381 /* Here start < (end|vma->vm_end). */
382 if (start < vma->vm_start) {
383 unmapped_error = -ENOMEM;
384 start = vma->vm_start;
389 /* Here vma->vm_start <= start < (end|vma->vm_end) */
394 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
395 error = madvise_vma(vma, &prev, start, tmp, behavior);
399 if (prev && start < prev->vm_end)
400 start = prev->vm_end;
401 error = unmapped_error;
406 else /* madvise_remove dropped mmap_sem */
407 vma = find_vma(current->mm, start);
411 up_write(¤t->mm->mmap_sem);
413 up_read(¤t->mm->mmap_sem);