4 * (C) Copyright 1995 Linus Torvalds
5 * (C) Copyright 2002 Christoph Hellwig
8 #include <linux/capability.h>
9 #include <linux/mman.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/syscalls.h>
16 #include <linux/sched.h>
17 #include <linux/module.h>
18 #include <linux/rmap.h>
19 #include <linux/mmzone.h>
20 #include <linux/hugetlb.h>
24 int can_do_mlock(void)
26 if (capable(CAP_IPC_LOCK))
28 if (rlimit(RLIMIT_MEMLOCK) != 0)
32 EXPORT_SYMBOL(can_do_mlock);
35 * Mlocked pages are marked with PageMlocked() flag for efficient testing
36 * in vmscan and, possibly, the fault path; and to support semi-accurate
39 * An mlocked page [PageMlocked(page)] is unevictable. As such, it will
40 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
41 * The unevictable list is an LRU sibling list to the [in]active lists.
42 * PageUnevictable is set to indicate the unevictable state.
44 * When lazy mlocking via vmscan, it is important to ensure that the
45 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
46 * may have mlocked a page that is being munlocked. So lazy mlock must take
47 * the mmap_sem for read, and verify that the vma really is locked
52 * LRU accounting for clear_page_mlock()
54 void __clear_page_mlock(struct page *page)
56 VM_BUG_ON(!PageLocked(page));
58 if (!page->mapping) { /* truncated ? */
62 dec_zone_page_state(page, NR_MLOCK);
63 count_vm_event(UNEVICTABLE_PGCLEARED);
64 if (!isolate_lru_page(page)) {
65 putback_lru_page(page);
68 * We lost the race. the page already moved to evictable list.
70 if (PageUnevictable(page))
71 count_vm_event(UNEVICTABLE_PGSTRANDED);
76 * Mark page as mlocked if not already.
77 * If page on LRU, isolate and putback to move to unevictable list.
79 void mlock_vma_page(struct page *page)
81 BUG_ON(!PageLocked(page));
83 if (!TestSetPageMlocked(page)) {
84 inc_zone_page_state(page, NR_MLOCK);
85 count_vm_event(UNEVICTABLE_PGMLOCKED);
86 if (!isolate_lru_page(page))
87 putback_lru_page(page);
92 * munlock_vma_page - munlock a vma page
93 * @page - page to be unlocked
95 * called from munlock()/munmap() path with page supposedly on the LRU.
96 * When we munlock a page, because the vma where we found the page is being
97 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
98 * page locked so that we can leave it on the unevictable lru list and not
99 * bother vmscan with it. However, to walk the page's rmap list in
100 * try_to_munlock() we must isolate the page from the LRU. If some other
101 * task has removed the page from the LRU, we won't be able to do that.
102 * So we clear the PageMlocked as we might not get another chance. If we
103 * can't isolate the page, we leave it for putback_lru_page() and vmscan
104 * [page_referenced()/try_to_unmap()] to deal with.
106 void munlock_vma_page(struct page *page)
108 BUG_ON(!PageLocked(page));
110 if (TestClearPageMlocked(page)) {
111 dec_zone_page_state(page, NR_MLOCK);
112 if (!isolate_lru_page(page)) {
113 int ret = try_to_munlock(page);
115 * did try_to_unlock() succeed or punt?
117 if (ret != SWAP_MLOCK)
118 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
120 putback_lru_page(page);
123 * Some other task has removed the page from the LRU.
124 * putback_lru_page() will take care of removing the
125 * page from the unevictable list, if necessary.
126 * vmscan [page_referenced()] will move the page back
127 * to the unevictable list if some other vma has it
130 if (PageUnevictable(page))
131 count_vm_event(UNEVICTABLE_PGSTRANDED);
133 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
138 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
140 return (vma->vm_flags & VM_GROWSDOWN) &&
141 (vma->vm_start == addr) &&
142 !vma_stack_continue(vma->vm_prev, addr);
146 * __mlock_vma_pages_range() - mlock a range of pages in the vma.
148 * @start: start address
151 * This takes care of making the pages present too.
153 * return 0 on success, negative error code on error.
155 * vma->vm_mm->mmap_sem must be held for at least read.
157 static long __mlock_vma_pages_range(struct vm_area_struct *vma,
158 unsigned long start, unsigned long end)
160 struct mm_struct *mm = vma->vm_mm;
161 unsigned long addr = start;
162 struct page *pages[16]; /* 16 gives a reasonable batch */
163 int nr_pages = (end - start) / PAGE_SIZE;
167 VM_BUG_ON(start & ~PAGE_MASK);
168 VM_BUG_ON(end & ~PAGE_MASK);
169 VM_BUG_ON(start < vma->vm_start);
170 VM_BUG_ON(end > vma->vm_end);
171 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
173 gup_flags = FOLL_TOUCH | FOLL_GET;
175 * We want to touch writable mappings with a write fault in order
176 * to break COW, except for shared mappings because these don't COW
177 * and we would not want to dirty them for nothing.
179 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
180 gup_flags |= FOLL_WRITE;
182 /* We don't try to access the guard page of a stack vma */
183 if (stack_guard_page(vma, start)) {
188 while (nr_pages > 0) {
194 * get_user_pages makes pages present if we are
195 * setting mlock. and this extra reference count will
196 * disable migration of this page. However, page may
197 * still be truncated out from under us.
199 ret = __get_user_pages(current, mm, addr,
200 min_t(int, nr_pages, ARRAY_SIZE(pages)),
201 gup_flags, pages, NULL);
203 * This can happen for, e.g., VM_NONLINEAR regions before
204 * a page has been allocated and mapped at a given offset,
205 * or for addresses that map beyond end of a file.
206 * We'll mlock the pages if/when they get faulted in.
211 lru_add_drain(); /* push cached pages to LRU */
213 for (i = 0; i < ret; i++) {
214 struct page *page = pages[i];
218 * That preliminary check is mainly to avoid
219 * the pointless overhead of lock_page on the
220 * ZERO_PAGE: which might bounce very badly if
221 * there is contention. However, we're still
222 * dirtying its cacheline with get/put_page:
223 * we'll add another __get_user_pages flag to
224 * avoid it if that case turns out to matter.
228 * Because we lock page here and migration is
229 * blocked by the elevated reference, we need
230 * only check for file-cache page truncation.
233 mlock_vma_page(page);
236 put_page(page); /* ref from get_user_pages() */
239 addr += ret * PAGE_SIZE;
244 return ret; /* 0 or negative error code */
248 * convert get_user_pages() return value to posix mlock() error
250 static int __mlock_posix_error_return(long retval)
252 if (retval == -EFAULT)
254 else if (retval == -ENOMEM)
260 * mlock_vma_pages_range() - mlock pages in specified vma range.
261 * @vma - the vma containing the specfied address range
262 * @start - starting address in @vma to mlock
263 * @end - end address [+1] in @vma to mlock
265 * For mmap()/mremap()/expansion of mlocked vma.
267 * return 0 on success for "normal" vmas.
269 * return number of pages [> 0] to be removed from locked_vm on success
272 long mlock_vma_pages_range(struct vm_area_struct *vma,
273 unsigned long start, unsigned long end)
275 int nr_pages = (end - start) / PAGE_SIZE;
276 BUG_ON(!(vma->vm_flags & VM_LOCKED));
279 * filter unlockable vmas
281 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
284 if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
285 is_vm_hugetlb_page(vma) ||
286 vma == get_gate_vma(current))) {
288 __mlock_vma_pages_range(vma, start, end);
290 /* Hide errors from mmap() and other callers */
295 * User mapped kernel pages or huge pages:
296 * make these pages present to populate the ptes, but
297 * fall thru' to reset VM_LOCKED--no need to unlock, and
298 * return nr_pages so these don't get counted against task's
299 * locked limit. huge pages are already counted against
302 make_pages_present(start, end);
305 vma->vm_flags &= ~VM_LOCKED; /* and don't come back! */
306 return nr_pages; /* error or pages NOT mlocked */
310 * munlock_vma_pages_range() - munlock all pages in the vma range.'
311 * @vma - vma containing range to be munlock()ed.
312 * @start - start address in @vma of the range
313 * @end - end of range in @vma.
315 * For mremap(), munmap() and exit().
317 * Called with @vma VM_LOCKED.
319 * Returns with VM_LOCKED cleared. Callers must be prepared to
322 * We don't save and restore VM_LOCKED here because pages are
323 * still on lru. In unmap path, pages might be scanned by reclaim
324 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
325 * free them. This will result in freeing mlocked pages.
327 void munlock_vma_pages_range(struct vm_area_struct *vma,
328 unsigned long start, unsigned long end)
333 vma->vm_flags &= ~VM_LOCKED;
335 for (addr = start; addr < end; addr += PAGE_SIZE) {
338 * Although FOLL_DUMP is intended for get_dump_page(),
339 * it just so happens that its special treatment of the
340 * ZERO_PAGE (returning an error instead of doing get_page)
341 * suits munlock very well (and if somehow an abnormal page
342 * has sneaked into the range, we won't oops here: great).
344 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
345 if (page && !IS_ERR(page)) {
348 * Like in __mlock_vma_pages_range(),
349 * because we lock page here and migration is
350 * blocked by the elevated reference, we need
351 * only check for file-cache page truncation.
354 munlock_vma_page(page);
363 * mlock_fixup - handle mlock[all]/munlock[all] requests.
365 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
366 * munlock is a no-op. However, for some special vmas, we go ahead and
367 * populate the ptes via make_pages_present().
369 * For vmas that pass the filters, merge/split as appropriate.
371 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
372 unsigned long start, unsigned long end, unsigned int newflags)
374 struct mm_struct *mm = vma->vm_mm;
378 int lock = newflags & VM_LOCKED;
380 if (newflags == vma->vm_flags ||
381 (vma->vm_flags & (VM_IO | VM_PFNMAP)))
382 goto out; /* don't set VM_LOCKED, don't count */
384 if ((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
385 is_vm_hugetlb_page(vma) ||
386 vma == get_gate_vma(current)) {
388 make_pages_present(start, end);
389 goto out; /* don't set VM_LOCKED, don't count */
392 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
393 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
394 vma->vm_file, pgoff, vma_policy(vma));
400 if (start != vma->vm_start) {
401 ret = split_vma(mm, vma, start, 1);
406 if (end != vma->vm_end) {
407 ret = split_vma(mm, vma, end, 0);
414 * Keep track of amount of locked VM.
416 nr_pages = (end - start) >> PAGE_SHIFT;
418 nr_pages = -nr_pages;
419 mm->locked_vm += nr_pages;
422 * vm_flags is protected by the mmap_sem held in write mode.
423 * It's okay if try_to_unmap_one unmaps a page just after we
424 * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
428 vma->vm_flags = newflags;
429 ret = __mlock_vma_pages_range(vma, start, end);
431 ret = __mlock_posix_error_return(ret);
433 munlock_vma_pages_range(vma, start, end);
441 static int do_mlock(unsigned long start, size_t len, int on)
443 unsigned long nstart, end, tmp;
444 struct vm_area_struct * vma, * prev;
447 len = PAGE_ALIGN(len);
453 vma = find_vma_prev(current->mm, start, &prev);
454 if (!vma || vma->vm_start > start)
457 if (start > vma->vm_start)
460 for (nstart = start ; ; ) {
461 unsigned int newflags;
463 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
465 newflags = vma->vm_flags | VM_LOCKED;
467 newflags &= ~VM_LOCKED;
472 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
476 if (nstart < prev->vm_end)
477 nstart = prev->vm_end;
482 if (!vma || vma->vm_start != nstart) {
490 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
492 unsigned long locked;
493 unsigned long lock_limit;
499 lru_add_drain_all(); /* flush pagevec */
501 down_write(¤t->mm->mmap_sem);
502 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
505 locked = len >> PAGE_SHIFT;
506 locked += current->mm->locked_vm;
508 lock_limit = rlimit(RLIMIT_MEMLOCK);
509 lock_limit >>= PAGE_SHIFT;
511 /* check against resource limits */
512 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
513 error = do_mlock(start, len, 1);
514 up_write(¤t->mm->mmap_sem);
518 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
522 down_write(¤t->mm->mmap_sem);
523 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
525 ret = do_mlock(start, len, 0);
526 up_write(¤t->mm->mmap_sem);
530 static int do_mlockall(int flags)
532 struct vm_area_struct * vma, * prev = NULL;
533 unsigned int def_flags = 0;
535 if (flags & MCL_FUTURE)
536 def_flags = VM_LOCKED;
537 current->mm->def_flags = def_flags;
538 if (flags == MCL_FUTURE)
541 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
542 unsigned int newflags;
544 newflags = vma->vm_flags | VM_LOCKED;
545 if (!(flags & MCL_CURRENT))
546 newflags &= ~VM_LOCKED;
549 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
555 SYSCALL_DEFINE1(mlockall, int, flags)
557 unsigned long lock_limit;
560 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
567 lru_add_drain_all(); /* flush pagevec */
569 down_write(¤t->mm->mmap_sem);
571 lock_limit = rlimit(RLIMIT_MEMLOCK);
572 lock_limit >>= PAGE_SHIFT;
575 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
576 capable(CAP_IPC_LOCK))
577 ret = do_mlockall(flags);
578 up_write(¤t->mm->mmap_sem);
583 SYSCALL_DEFINE0(munlockall)
587 down_write(¤t->mm->mmap_sem);
588 ret = do_mlockall(0);
589 up_write(¤t->mm->mmap_sem);
594 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
595 * shm segments) get accounted against the user_struct instead.
597 static DEFINE_SPINLOCK(shmlock_user_lock);
599 int user_shm_lock(size_t size, struct user_struct *user)
601 unsigned long lock_limit, locked;
604 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
605 lock_limit = rlimit(RLIMIT_MEMLOCK);
606 if (lock_limit == RLIM_INFINITY)
608 lock_limit >>= PAGE_SHIFT;
609 spin_lock(&shmlock_user_lock);
611 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
614 user->locked_shm += locked;
617 spin_unlock(&shmlock_user_lock);
621 void user_shm_unlock(size_t size, struct user_struct *user)
623 spin_lock(&shmlock_user_lock);
624 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
625 spin_unlock(&shmlock_user_lock);