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 (current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur != 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 * called from munlock()/munmap() path with page supposedly on the LRU.
94 * Note: unlike mlock_vma_page(), we can't just clear the PageMlocked
95 * [in try_to_munlock()] and then attempt to isolate the page. We must
96 * isolate the page to keep others from messing with its unevictable
97 * and mlocked state while trying to munlock. However, we pre-clear the
98 * mlocked state anyway as we might lose the isolation race and we might
99 * not get another chance to clear PageMlocked. If we successfully
100 * isolate the page and try_to_munlock() detects other VM_LOCKED vmas
101 * mapping the page, it will restore the PageMlocked state, unless the page
102 * is mapped in a non-linear vma. So, we go ahead and ClearPageMlocked(),
103 * perhaps redundantly.
104 * If we lose the isolation race, and the page is mapped by other VM_LOCKED
105 * vmas, we'll detect this in vmscan--via try_to_munlock() or try_to_unmap()
106 * either of which will restore the PageMlocked state by calling
107 * mlock_vma_page() above, if it can grab the vma's mmap sem.
109 void munlock_vma_page(struct page *page)
111 BUG_ON(!PageLocked(page));
113 if (TestClearPageMlocked(page)) {
114 dec_zone_page_state(page, NR_MLOCK);
115 if (!isolate_lru_page(page)) {
116 int ret = try_to_munlock(page);
118 * did try_to_unlock() succeed or punt?
120 if (ret == SWAP_SUCCESS || ret == SWAP_AGAIN)
121 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
123 putback_lru_page(page);
126 * We lost the race. let try_to_unmap() deal
127 * with it. At least we get the page state and
128 * mlock stats right. However, page is still on
129 * the noreclaim list. We'll fix that up when
130 * the page is eventually freed or we scan the
133 if (PageUnevictable(page))
134 count_vm_event(UNEVICTABLE_PGSTRANDED);
136 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
141 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
143 return (vma->vm_flags & VM_GROWSDOWN) &&
144 (vma->vm_start == addr) &&
145 !vma_stack_continue(vma->vm_prev, addr);
149 * __mlock_vma_pages_range() - mlock a range of pages in the vma.
151 * @start: start address
154 * This takes care of making the pages present too.
156 * return 0 on success, negative error code on error.
158 * vma->vm_mm->mmap_sem must be held for at least read.
160 static long __mlock_vma_pages_range(struct vm_area_struct *vma,
161 unsigned long start, unsigned long end)
163 struct mm_struct *mm = vma->vm_mm;
164 unsigned long addr = start;
165 struct page *pages[16]; /* 16 gives a reasonable batch */
166 int nr_pages = (end - start) / PAGE_SIZE;
170 VM_BUG_ON(start & ~PAGE_MASK);
171 VM_BUG_ON(end & ~PAGE_MASK);
172 VM_BUG_ON(start < vma->vm_start);
173 VM_BUG_ON(end > vma->vm_end);
174 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
176 gup_flags = FOLL_TOUCH | FOLL_GET;
177 if (vma->vm_flags & VM_WRITE)
178 gup_flags |= FOLL_WRITE;
180 /* We don't try to access the guard page of a stack vma */
181 if (stack_guard_page(vma, start)) {
186 while (nr_pages > 0) {
192 * get_user_pages makes pages present if we are
193 * setting mlock. and this extra reference count will
194 * disable migration of this page. However, page may
195 * still be truncated out from under us.
197 ret = __get_user_pages(current, mm, addr,
198 min_t(int, nr_pages, ARRAY_SIZE(pages)),
199 gup_flags, pages, NULL);
201 * This can happen for, e.g., VM_NONLINEAR regions before
202 * a page has been allocated and mapped at a given offset,
203 * or for addresses that map beyond end of a file.
204 * We'll mlock the pages if/when they get faulted in.
209 lru_add_drain(); /* push cached pages to LRU */
211 for (i = 0; i < ret; i++) {
212 struct page *page = pages[i];
216 * That preliminary check is mainly to avoid
217 * the pointless overhead of lock_page on the
218 * ZERO_PAGE: which might bounce very badly if
219 * there is contention. However, we're still
220 * dirtying its cacheline with get/put_page:
221 * we'll add another __get_user_pages flag to
222 * avoid it if that case turns out to matter.
226 * Because we lock page here and migration is
227 * blocked by the elevated reference, we need
228 * only check for file-cache page truncation.
231 mlock_vma_page(page);
234 put_page(page); /* ref from get_user_pages() */
237 addr += ret * PAGE_SIZE;
242 return ret; /* 0 or negative error code */
246 * convert get_user_pages() return value to posix mlock() error
248 static int __mlock_posix_error_return(long retval)
250 if (retval == -EFAULT)
252 else if (retval == -ENOMEM)
258 * mlock_vma_pages_range() - mlock pages in specified vma range.
259 * @vma - the vma containing the specfied address range
260 * @start - starting address in @vma to mlock
261 * @end - end address [+1] in @vma to mlock
263 * For mmap()/mremap()/expansion of mlocked vma.
265 * return 0 on success for "normal" vmas.
267 * return number of pages [> 0] to be removed from locked_vm on success
270 long mlock_vma_pages_range(struct vm_area_struct *vma,
271 unsigned long start, unsigned long end)
273 int nr_pages = (end - start) / PAGE_SIZE;
274 BUG_ON(!(vma->vm_flags & VM_LOCKED));
277 * filter unlockable vmas
279 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
282 if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
283 is_vm_hugetlb_page(vma) ||
284 vma == get_gate_vma(current))) {
286 __mlock_vma_pages_range(vma, start, end);
288 /* Hide errors from mmap() and other callers */
293 * User mapped kernel pages or huge pages:
294 * make these pages present to populate the ptes, but
295 * fall thru' to reset VM_LOCKED--no need to unlock, and
296 * return nr_pages so these don't get counted against task's
297 * locked limit. huge pages are already counted against
300 make_pages_present(start, end);
303 vma->vm_flags &= ~VM_LOCKED; /* and don't come back! */
304 return nr_pages; /* error or pages NOT mlocked */
308 * munlock_vma_pages_range() - munlock all pages in the vma range.'
309 * @vma - vma containing range to be munlock()ed.
310 * @start - start address in @vma of the range
311 * @end - end of range in @vma.
313 * For mremap(), munmap() and exit().
315 * Called with @vma VM_LOCKED.
317 * Returns with VM_LOCKED cleared. Callers must be prepared to
320 * We don't save and restore VM_LOCKED here because pages are
321 * still on lru. In unmap path, pages might be scanned by reclaim
322 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
323 * free them. This will result in freeing mlocked pages.
325 void munlock_vma_pages_range(struct vm_area_struct *vma,
326 unsigned long start, unsigned long end)
331 vma->vm_flags &= ~VM_LOCKED;
333 for (addr = start; addr < end; addr += PAGE_SIZE) {
336 * Although FOLL_DUMP is intended for get_dump_page(),
337 * it just so happens that its special treatment of the
338 * ZERO_PAGE (returning an error instead of doing get_page)
339 * suits munlock very well (and if somehow an abnormal page
340 * has sneaked into the range, we won't oops here: great).
342 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
343 if (page && !IS_ERR(page)) {
346 * Like in __mlock_vma_pages_range(),
347 * because we lock page here and migration is
348 * blocked by the elevated reference, we need
349 * only check for file-cache page truncation.
352 munlock_vma_page(page);
361 * mlock_fixup - handle mlock[all]/munlock[all] requests.
363 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
364 * munlock is a no-op. However, for some special vmas, we go ahead and
365 * populate the ptes via make_pages_present().
367 * For vmas that pass the filters, merge/split as appropriate.
369 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
370 unsigned long start, unsigned long end, unsigned int newflags)
372 struct mm_struct *mm = vma->vm_mm;
376 int lock = newflags & VM_LOCKED;
378 if (newflags == vma->vm_flags ||
379 (vma->vm_flags & (VM_IO | VM_PFNMAP)))
380 goto out; /* don't set VM_LOCKED, don't count */
382 if ((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
383 is_vm_hugetlb_page(vma) ||
384 vma == get_gate_vma(current)) {
386 make_pages_present(start, end);
387 goto out; /* don't set VM_LOCKED, don't count */
390 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
391 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
392 vma->vm_file, pgoff, vma_policy(vma));
398 if (start != vma->vm_start) {
399 ret = split_vma(mm, vma, start, 1);
404 if (end != vma->vm_end) {
405 ret = split_vma(mm, vma, end, 0);
412 * Keep track of amount of locked VM.
414 nr_pages = (end - start) >> PAGE_SHIFT;
416 nr_pages = -nr_pages;
417 mm->locked_vm += nr_pages;
420 * vm_flags is protected by the mmap_sem held in write mode.
421 * It's okay if try_to_unmap_one unmaps a page just after we
422 * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
426 vma->vm_flags = newflags;
427 ret = __mlock_vma_pages_range(vma, start, end);
429 ret = __mlock_posix_error_return(ret);
431 munlock_vma_pages_range(vma, start, end);
439 static int do_mlock(unsigned long start, size_t len, int on)
441 unsigned long nstart, end, tmp;
442 struct vm_area_struct * vma, * prev;
445 len = PAGE_ALIGN(len);
451 vma = find_vma_prev(current->mm, start, &prev);
452 if (!vma || vma->vm_start > start)
455 if (start > vma->vm_start)
458 for (nstart = start ; ; ) {
459 unsigned int newflags;
461 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
463 newflags = vma->vm_flags | VM_LOCKED;
465 newflags &= ~VM_LOCKED;
470 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
474 if (nstart < prev->vm_end)
475 nstart = prev->vm_end;
480 if (!vma || vma->vm_start != nstart) {
488 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
490 unsigned long locked;
491 unsigned long lock_limit;
497 lru_add_drain_all(); /* flush pagevec */
499 down_write(¤t->mm->mmap_sem);
500 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
503 locked = len >> PAGE_SHIFT;
504 locked += current->mm->locked_vm;
506 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
507 lock_limit >>= PAGE_SHIFT;
509 /* check against resource limits */
510 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
511 error = do_mlock(start, len, 1);
512 up_write(¤t->mm->mmap_sem);
516 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
520 down_write(¤t->mm->mmap_sem);
521 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
523 ret = do_mlock(start, len, 0);
524 up_write(¤t->mm->mmap_sem);
528 static int do_mlockall(int flags)
530 struct vm_area_struct * vma, * prev = NULL;
531 unsigned int def_flags = 0;
533 if (flags & MCL_FUTURE)
534 def_flags = VM_LOCKED;
535 current->mm->def_flags = def_flags;
536 if (flags == MCL_FUTURE)
539 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
540 unsigned int newflags;
542 newflags = vma->vm_flags | VM_LOCKED;
543 if (!(flags & MCL_CURRENT))
544 newflags &= ~VM_LOCKED;
547 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
553 SYSCALL_DEFINE1(mlockall, int, flags)
555 unsigned long lock_limit;
558 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
565 lru_add_drain_all(); /* flush pagevec */
567 down_write(¤t->mm->mmap_sem);
569 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
570 lock_limit >>= PAGE_SHIFT;
573 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
574 capable(CAP_IPC_LOCK))
575 ret = do_mlockall(flags);
576 up_write(¤t->mm->mmap_sem);
581 SYSCALL_DEFINE0(munlockall)
585 down_write(¤t->mm->mmap_sem);
586 ret = do_mlockall(0);
587 up_write(¤t->mm->mmap_sem);
592 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
593 * shm segments) get accounted against the user_struct instead.
595 static DEFINE_SPINLOCK(shmlock_user_lock);
597 int user_shm_lock(size_t size, struct user_struct *user)
599 unsigned long lock_limit, locked;
602 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
603 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
604 if (lock_limit == RLIM_INFINITY)
606 lock_limit >>= PAGE_SHIFT;
607 spin_lock(&shmlock_user_lock);
609 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
612 user->locked_shm += locked;
615 spin_unlock(&shmlock_user_lock);
619 void user_shm_unlock(size_t size, struct user_struct *user)
621 spin_lock(&shmlock_user_lock);
622 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
623 spin_unlock(&shmlock_user_lock);
627 int account_locked_memory(struct mm_struct *mm, struct rlimit *rlim,
630 unsigned long lim, vm, pgsz;
633 pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
635 down_write(&mm->mmap_sem);
637 lim = rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
638 vm = mm->total_vm + pgsz;
642 lim = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
643 vm = mm->locked_vm + pgsz;
647 mm->total_vm += pgsz;
648 mm->locked_vm += pgsz;
652 up_write(&mm->mmap_sem);
656 void refund_locked_memory(struct mm_struct *mm, size_t size)
658 unsigned long pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
660 down_write(&mm->mmap_sem);
662 mm->total_vm -= pgsz;
663 mm->locked_vm -= pgsz;
665 up_write(&mm->mmap_sem);