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/sched.h>
15 #include <linux/ksm.h>
18 * Any behaviour which results in changes to the vma->vm_flags needs to
19 * take mmap_sem for writing. Others, which simply traverse vmas, need
20 * to only take it for reading.
22 static int madvise_need_mmap_write(int behavior)
30 /* be safe, default to 1. list exceptions explicitly */
36 * We can potentially split a vm area into separate
37 * areas, each area with its own behavior.
39 static long madvise_behavior(struct vm_area_struct * vma,
40 struct vm_area_struct **prev,
41 unsigned long start, unsigned long end, int behavior)
43 struct mm_struct * mm = vma->vm_mm;
46 unsigned long new_flags = vma->vm_flags;
50 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
53 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
56 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
59 new_flags |= VM_DONTCOPY;
62 if (vma->vm_flags & VM_IO) {
66 new_flags &= ~VM_DONTCOPY;
69 case MADV_UNMERGEABLE:
70 error = ksm_madvise(vma, start, end, behavior, &new_flags);
76 if (new_flags == vma->vm_flags) {
81 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
82 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
83 vma->vm_file, pgoff, vma_policy(vma));
91 if (start != vma->vm_start) {
92 error = split_vma(mm, vma, start, 1);
97 if (end != vma->vm_end) {
98 error = split_vma(mm, vma, end, 0);
105 * vm_flags is protected by the mmap_sem held in write mode.
107 vma->vm_flags = new_flags;
110 if (error == -ENOMEM)
116 * Schedule all required I/O operations. Do not wait for completion.
118 static long madvise_willneed(struct vm_area_struct * vma,
119 struct vm_area_struct ** prev,
120 unsigned long start, unsigned long end)
122 struct file *file = vma->vm_file;
127 if (file->f_mapping->a_ops->get_xip_mem) {
128 /* no bad return value, but ignore advice */
133 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
134 if (end > vma->vm_end)
136 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
138 force_page_cache_readahead(file->f_mapping, file, start, end - start);
143 * Application no longer needs these pages. If the pages are dirty,
144 * it's OK to just throw them away. The app will be more careful about
145 * data it wants to keep. Be sure to free swap resources too. The
146 * zap_page_range call sets things up for shrink_active_list to actually free
147 * these pages later if no one else has touched them in the meantime,
148 * although we could add these pages to a global reuse list for
149 * shrink_active_list to pick up before reclaiming other pages.
151 * NB: This interface discards data rather than pushes it out to swap,
152 * as some implementations do. This has performance implications for
153 * applications like large transactional databases which want to discard
154 * pages in anonymous maps after committing to backing store the data
155 * that was kept in them. There is no reason to write this data out to
156 * the swap area if the application is discarding it.
158 * An interface that causes the system to free clean pages and flush
159 * dirty pages is already available as msync(MS_INVALIDATE).
161 static long madvise_dontneed(struct vm_area_struct * vma,
162 struct vm_area_struct ** prev,
163 unsigned long start, unsigned long end)
166 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
169 if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
170 struct zap_details details = {
171 .nonlinear_vma = vma,
172 .last_index = ULONG_MAX,
174 zap_page_range(vma, start, end - start, &details);
176 zap_page_range(vma, start, end - start, NULL);
181 * Application wants to free up the pages and associated backing store.
182 * This is effectively punching a hole into the middle of a file.
184 * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
185 * Other filesystems return -ENOSYS.
187 static long madvise_remove(struct vm_area_struct *vma,
188 struct vm_area_struct **prev,
189 unsigned long start, unsigned long end)
191 struct address_space *mapping;
192 loff_t offset, endoff;
195 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
197 if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
200 if (!vma->vm_file || !vma->vm_file->f_mapping
201 || !vma->vm_file->f_mapping->host) {
205 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
208 mapping = vma->vm_file->f_mapping;
210 offset = (loff_t)(start - vma->vm_start)
211 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
212 endoff = (loff_t)(end - vma->vm_start - 1)
213 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
215 /* vmtruncate_range needs to take i_mutex and i_alloc_sem */
216 up_read(¤t->mm->mmap_sem);
217 error = vmtruncate_range(mapping->host, offset, endoff);
218 down_read(¤t->mm->mmap_sem);
222 #ifdef CONFIG_MEMORY_FAILURE
224 * Error injection support for memory error handling.
226 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
230 if (!capable(CAP_SYS_ADMIN))
232 for (; start < end; start += PAGE_SIZE) {
234 int ret = get_user_pages_fast(start, 1, 0, &p);
237 if (bhv == MADV_SOFT_OFFLINE) {
238 printk(KERN_INFO "Soft offlining page %lx at %lx\n",
239 page_to_pfn(p), start);
240 ret = soft_offline_page(p, MF_COUNT_INCREASED);
245 printk(KERN_INFO "Injecting memory failure for page %lx at %lx\n",
246 page_to_pfn(p), start);
247 /* Ignore return value for now */
248 __memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
255 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
256 unsigned long start, unsigned long end, int behavior)
260 return madvise_remove(vma, prev, start, end);
262 return madvise_willneed(vma, prev, start, end);
264 return madvise_dontneed(vma, prev, start, end);
266 return madvise_behavior(vma, prev, start, end, behavior);
271 madvise_behavior_valid(int behavior)
277 case MADV_SEQUENTIAL:
284 case MADV_UNMERGEABLE:
294 * The madvise(2) system call.
296 * Applications can use madvise() to advise the kernel how it should
297 * handle paging I/O in this VM area. The idea is to help the kernel
298 * use appropriate read-ahead and caching techniques. The information
299 * provided is advisory only, and can be safely disregarded by the
300 * kernel without affecting the correct operation of the application.
303 * MADV_NORMAL - the default behavior is to read clusters. This
304 * results in some read-ahead and read-behind.
305 * MADV_RANDOM - the system should read the minimum amount of data
306 * on any access, since it is unlikely that the appli-
307 * cation will need more than what it asks for.
308 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
309 * once, so they can be aggressively read ahead, and
310 * can be freed soon after they are accessed.
311 * MADV_WILLNEED - the application is notifying the system to read
313 * MADV_DONTNEED - the application is finished with the given range,
314 * so the kernel can free resources associated with it.
315 * MADV_REMOVE - the application wants to free up the given range of
316 * pages and associated backing store.
317 * MADV_DONTFORK - omit this area from child's address space when forking:
318 * typically, to avoid COWing pages pinned by get_user_pages().
319 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
320 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
321 * this area with pages of identical content from other such areas.
322 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
326 * -EINVAL - start + len < 0, start is not page-aligned,
327 * "behavior" is not a valid value, or application
328 * is attempting to release locked or shared pages.
329 * -ENOMEM - addresses in the specified range are not currently
330 * mapped, or are outside the AS of the process.
331 * -EIO - an I/O error occurred while paging in data.
332 * -EBADF - map exists, but area maps something that isn't a file.
333 * -EAGAIN - a kernel resource was temporarily unavailable.
335 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
337 unsigned long end, tmp;
338 struct vm_area_struct * vma, *prev;
339 int unmapped_error = 0;
344 #ifdef CONFIG_MEMORY_FAILURE
345 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
346 return madvise_hwpoison(behavior, start, start+len_in);
348 if (!madvise_behavior_valid(behavior))
351 write = madvise_need_mmap_write(behavior);
353 down_write(¤t->mm->mmap_sem);
355 down_read(¤t->mm->mmap_sem);
357 if (start & ~PAGE_MASK)
359 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
361 /* Check to see whether len was rounded up from small -ve to zero */
374 * If the interval [start,end) covers some unmapped address
375 * ranges, just ignore them, but return -ENOMEM at the end.
376 * - different from the way of handling in mlock etc.
378 vma = find_vma_prev(current->mm, start, &prev);
379 if (vma && start > vma->vm_start)
383 /* Still start < end. */
388 /* Here start < (end|vma->vm_end). */
389 if (start < vma->vm_start) {
390 unmapped_error = -ENOMEM;
391 start = vma->vm_start;
396 /* Here vma->vm_start <= start < (end|vma->vm_end) */
401 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
402 error = madvise_vma(vma, &prev, start, tmp, behavior);
406 if (prev && start < prev->vm_end)
407 start = prev->vm_end;
408 error = unmapped_error;
413 else /* madvise_remove dropped mmap_sem */
414 vma = find_vma(current->mm, start);
418 up_write(¤t->mm->mmap_sem);
420 up_read(¤t->mm->mmap_sem);