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/pagevec.h>
15 #include <linux/mempolicy.h>
16 #include <linux/syscalls.h>
17 #include <linux/sched.h>
18 #include <linux/export.h>
19 #include <linux/rmap.h>
20 #include <linux/mmzone.h>
21 #include <linux/hugetlb.h>
22 #include <linux/memcontrol.h>
23 #include <linux/mm_inline.h>
27 int can_do_mlock(void)
29 if (capable(CAP_IPC_LOCK))
31 if (rlimit(RLIMIT_MEMLOCK) != 0)
35 EXPORT_SYMBOL(can_do_mlock);
38 * Mlocked pages are marked with PageMlocked() flag for efficient testing
39 * in vmscan and, possibly, the fault path; and to support semi-accurate
42 * An mlocked page [PageMlocked(page)] is unevictable. As such, it will
43 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
44 * The unevictable list is an LRU sibling list to the [in]active lists.
45 * PageUnevictable is set to indicate the unevictable state.
47 * When lazy mlocking via vmscan, it is important to ensure that the
48 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
49 * may have mlocked a page that is being munlocked. So lazy mlock must take
50 * the mmap_sem for read, and verify that the vma really is locked
55 * LRU accounting for clear_page_mlock()
57 void clear_page_mlock(struct page *page)
59 if (!TestClearPageMlocked(page))
62 mod_zone_page_state(page_zone(page), NR_MLOCK,
63 -hpage_nr_pages(page));
64 count_vm_event(UNEVICTABLE_PGCLEARED);
65 if (!isolate_lru_page(page)) {
66 putback_lru_page(page);
69 * We lost the race. the page already moved to evictable list.
71 if (PageUnevictable(page))
72 count_vm_event(UNEVICTABLE_PGSTRANDED);
77 * Mark page as mlocked if not already.
78 * If page on LRU, isolate and putback to move to unevictable list.
80 void mlock_vma_page(struct page *page)
82 BUG_ON(!PageLocked(page));
84 if (!TestSetPageMlocked(page)) {
85 mod_zone_page_state(page_zone(page), NR_MLOCK,
86 hpage_nr_pages(page));
87 count_vm_event(UNEVICTABLE_PGMLOCKED);
88 if (!isolate_lru_page(page))
89 putback_lru_page(page);
94 * Finish munlock after successful page isolation
96 * Page must be locked. This is a wrapper for try_to_munlock()
97 * and putback_lru_page() with munlock accounting.
99 static void __munlock_isolated_page(struct page *page)
101 int ret = SWAP_AGAIN;
104 * Optimization: if the page was mapped just once, that's our mapping
105 * and we don't need to check all the other vmas.
107 if (page_mapcount(page) > 1)
108 ret = try_to_munlock(page);
110 /* Did try_to_unlock() succeed or punt? */
111 if (ret != SWAP_MLOCK)
112 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
114 putback_lru_page(page);
118 * Accounting for page isolation fail during munlock
120 * Performs accounting when page isolation fails in munlock. There is nothing
121 * else to do because it means some other task has already removed the page
122 * from the LRU. putback_lru_page() will take care of removing the page from
123 * the unevictable list, if necessary. vmscan [page_referenced()] will move
124 * the page back to the unevictable list if some other vma has it mlocked.
126 static void __munlock_isolation_failed(struct page *page)
128 if (PageUnevictable(page))
129 count_vm_event(UNEVICTABLE_PGSTRANDED);
131 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
135 * munlock_vma_page - munlock a vma page
136 * @page - page to be unlocked, either a normal page or THP page head
138 * returns the size of the page as a page mask (0 for normal page,
139 * HPAGE_PMD_NR - 1 for THP head page)
141 * called from munlock()/munmap() path with page supposedly on the LRU.
142 * When we munlock a page, because the vma where we found the page is being
143 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
144 * page locked so that we can leave it on the unevictable lru list and not
145 * bother vmscan with it. However, to walk the page's rmap list in
146 * try_to_munlock() we must isolate the page from the LRU. If some other
147 * task has removed the page from the LRU, we won't be able to do that.
148 * So we clear the PageMlocked as we might not get another chance. If we
149 * can't isolate the page, we leave it for putback_lru_page() and vmscan
150 * [page_referenced()/try_to_unmap()] to deal with.
152 unsigned int munlock_vma_page(struct page *page)
154 unsigned int nr_pages;
156 BUG_ON(!PageLocked(page));
158 if (TestClearPageMlocked(page)) {
159 nr_pages = hpage_nr_pages(page);
160 mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
161 if (!isolate_lru_page(page))
162 __munlock_isolated_page(page);
164 __munlock_isolation_failed(page);
166 nr_pages = hpage_nr_pages(page);
170 * Regardless of the original PageMlocked flag, we determine nr_pages
171 * after touching the flag. This leaves a possible race with a THP page
172 * split, such that a whole THP page was munlocked, but nr_pages == 1.
173 * Returning a smaller mask due to that is OK, the worst that can
174 * happen is subsequent useless scanning of the former tail pages.
175 * The NR_MLOCK accounting can however become broken.
181 * __mlock_vma_pages_range() - mlock a range of pages in the vma.
183 * @start: start address
186 * This takes care of making the pages present too.
188 * return 0 on success, negative error code on error.
190 * vma->vm_mm->mmap_sem must be held for at least read.
192 long __mlock_vma_pages_range(struct vm_area_struct *vma,
193 unsigned long start, unsigned long end, int *nonblocking)
195 struct mm_struct *mm = vma->vm_mm;
196 unsigned long nr_pages = (end - start) / PAGE_SIZE;
199 VM_BUG_ON(start & ~PAGE_MASK);
200 VM_BUG_ON(end & ~PAGE_MASK);
201 VM_BUG_ON(start < vma->vm_start);
202 VM_BUG_ON(end > vma->vm_end);
203 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
205 gup_flags = FOLL_TOUCH | FOLL_MLOCK;
207 * We want to touch writable mappings with a write fault in order
208 * to break COW, except for shared mappings because these don't COW
209 * and we would not want to dirty them for nothing.
211 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
212 gup_flags |= FOLL_WRITE;
215 * We want mlock to succeed for regions that have any permissions
216 * other than PROT_NONE.
218 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
219 gup_flags |= FOLL_FORCE;
222 * We made sure addr is within a VMA, so the following will
223 * not result in a stack expansion that recurses back here.
225 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
226 NULL, NULL, nonblocking);
230 * convert get_user_pages() return value to posix mlock() error
232 static int __mlock_posix_error_return(long retval)
234 if (retval == -EFAULT)
236 else if (retval == -ENOMEM)
242 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
244 * The fast path is available only for evictable pages with single mapping.
245 * Then we can bypass the per-cpu pvec and get better performance.
246 * when mapcount > 1 we need try_to_munlock() which can fail.
247 * when !page_evictable(), we need the full redo logic of putback_lru_page to
248 * avoid leaving evictable page in unevictable list.
250 * In case of success, @page is added to @pvec and @pgrescued is incremented
251 * in case that the page was previously unevictable. @page is also unlocked.
253 static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
256 VM_BUG_ON(PageLRU(page));
257 VM_BUG_ON(!PageLocked(page));
259 if (page_mapcount(page) <= 1 && page_evictable(page)) {
260 pagevec_add(pvec, page);
261 if (TestClearPageUnevictable(page))
271 * Putback multiple evictable pages to the LRU
273 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
274 * the pages might have meanwhile become unevictable but that is OK.
276 static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
278 count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
280 *__pagevec_lru_add() calls release_pages() so we don't call
281 * put_page() explicitly
283 __pagevec_lru_add(pvec);
284 count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
288 * Munlock a batch of pages from the same zone
290 * The work is split to two main phases. First phase clears the Mlocked flag
291 * and attempts to isolate the pages, all under a single zone lru lock.
292 * The second phase finishes the munlock only for pages where isolation
295 * Note that the pagevec may be modified during the process.
297 static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
300 int nr = pagevec_count(pvec);
302 struct pagevec pvec_putback;
305 pagevec_init(&pvec_putback, 0);
307 /* Phase 1: page isolation */
308 spin_lock_irq(&zone->lru_lock);
309 for (i = 0; i < nr; i++) {
310 struct page *page = pvec->pages[i];
312 if (TestClearPageMlocked(page)) {
313 struct lruvec *lruvec;
317 lruvec = mem_cgroup_page_lruvec(page, zone);
318 lru = page_lru(page);
320 * We already have pin from follow_page_mask()
321 * so we can spare the get_page() here.
324 del_page_from_lru_list(page, lruvec, lru);
326 __munlock_isolation_failed(page);
333 * We won't be munlocking this page in the next phase
334 * but we still need to release the follow_page_mask()
335 * pin. We cannot do it under lru_lock however. If it's
336 * the last pin, __page_cache_release would deadlock.
338 pagevec_add(&pvec_putback, pvec->pages[i]);
339 pvec->pages[i] = NULL;
342 delta_munlocked = -nr + pagevec_count(&pvec_putback);
343 __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
344 spin_unlock_irq(&zone->lru_lock);
346 /* Now we can release pins of pages that we are not munlocking */
347 pagevec_release(&pvec_putback);
349 /* Phase 2: page munlock */
350 for (i = 0; i < nr; i++) {
351 struct page *page = pvec->pages[i];
355 if (!__putback_lru_fast_prepare(page, &pvec_putback,
358 * Slow path. We don't want to lose the last
359 * pin before unlock_page()
361 get_page(page); /* for putback_lru_page() */
362 __munlock_isolated_page(page);
364 put_page(page); /* from follow_page_mask() */
370 * Phase 3: page putback for pages that qualified for the fast path
371 * This will also call put_page() to return pin from follow_page_mask()
373 if (pagevec_count(&pvec_putback))
374 __putback_lru_fast(&pvec_putback, pgrescued);
378 * Fill up pagevec for __munlock_pagevec using pte walk
380 * The function expects that the struct page corresponding to @start address is
381 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
383 * The rest of @pvec is filled by subsequent pages within the same pmd and same
384 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
385 * pages also get pinned.
387 * Returns the address of the next page that should be scanned. This equals
388 * @start + PAGE_SIZE when no page could be added by the pte walk.
390 static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
391 struct vm_area_struct *vma, int zoneid, unsigned long start,
398 * Initialize pte walk starting at the already pinned page where we
399 * are sure that there is a pte, as it was pinned under the same
402 pte = get_locked_pte(vma->vm_mm, start, &ptl);
403 /* Make sure we do not cross the page table boundary */
404 end = pgd_addr_end(start, end);
405 end = pud_addr_end(start, end);
406 end = pmd_addr_end(start, end);
408 /* The page next to the pinned page is the first we will try to get */
410 while (start < end) {
411 struct page *page = NULL;
413 if (pte_present(*pte))
414 page = vm_normal_page(vma, start, *pte);
416 * Break if page could not be obtained or the page's node+zone does not
419 if (!page || page_zone_id(page) != zoneid)
424 * Increase the address that will be returned *before* the
425 * eventual break due to pvec becoming full by adding the page
428 if (pagevec_add(pvec, page) == 0)
431 pte_unmap_unlock(pte, ptl);
436 * munlock_vma_pages_range() - munlock all pages in the vma range.'
437 * @vma - vma containing range to be munlock()ed.
438 * @start - start address in @vma of the range
439 * @end - end of range in @vma.
441 * For mremap(), munmap() and exit().
443 * Called with @vma VM_LOCKED.
445 * Returns with VM_LOCKED cleared. Callers must be prepared to
448 * We don't save and restore VM_LOCKED here because pages are
449 * still on lru. In unmap path, pages might be scanned by reclaim
450 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
451 * free them. This will result in freeing mlocked pages.
453 void munlock_vma_pages_range(struct vm_area_struct *vma,
454 unsigned long start, unsigned long end)
456 vma->vm_flags &= ~VM_LOCKED;
458 while (start < end) {
459 struct page *page = NULL;
460 unsigned int page_mask;
461 unsigned long page_increm;
466 pagevec_init(&pvec, 0);
468 * Although FOLL_DUMP is intended for get_dump_page(),
469 * it just so happens that its special treatment of the
470 * ZERO_PAGE (returning an error instead of doing get_page)
471 * suits munlock very well (and if somehow an abnormal page
472 * has sneaked into the range, we won't oops here: great).
474 page = follow_page_mask(vma, start, FOLL_GET | FOLL_DUMP,
477 if (page && !IS_ERR(page)) {
478 if (PageTransHuge(page)) {
481 * Any THP page found by follow_page_mask() may
482 * have gotten split before reaching
483 * munlock_vma_page(), so we need to recompute
484 * the page_mask here.
486 page_mask = munlock_vma_page(page);
488 put_page(page); /* follow_page_mask() */
491 * Non-huge pages are handled in batches via
492 * pagevec. The pin from follow_page_mask()
493 * prevents them from collapsing by THP.
495 pagevec_add(&pvec, page);
496 zone = page_zone(page);
497 zoneid = page_zone_id(page);
500 * Try to fill the rest of pagevec using fast
501 * pte walk. This will also update start to
502 * the next page to process. Then munlock the
505 start = __munlock_pagevec_fill(&pvec, vma,
507 __munlock_pagevec(&pvec, zone);
511 /* It's a bug to munlock in the middle of a THP page */
512 VM_BUG_ON((start >> PAGE_SHIFT) & page_mask);
513 page_increm = 1 + page_mask;
514 start += page_increm * PAGE_SIZE;
521 * mlock_fixup - handle mlock[all]/munlock[all] requests.
523 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
524 * munlock is a no-op. However, for some special vmas, we go ahead and
527 * For vmas that pass the filters, merge/split as appropriate.
529 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
530 unsigned long start, unsigned long end, vm_flags_t newflags)
532 struct mm_struct *mm = vma->vm_mm;
536 int lock = !!(newflags & VM_LOCKED);
538 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
539 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
540 goto out; /* don't set VM_LOCKED, don't count */
542 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
543 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
544 vma->vm_file, pgoff, vma_policy(vma));
550 if (start != vma->vm_start) {
551 ret = split_vma(mm, vma, start, 1);
556 if (end != vma->vm_end) {
557 ret = split_vma(mm, vma, end, 0);
564 * Keep track of amount of locked VM.
566 nr_pages = (end - start) >> PAGE_SHIFT;
568 nr_pages = -nr_pages;
569 mm->locked_vm += nr_pages;
572 * vm_flags is protected by the mmap_sem held in write mode.
573 * It's okay if try_to_unmap_one unmaps a page just after we
574 * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
578 vma->vm_flags = newflags;
580 munlock_vma_pages_range(vma, start, end);
587 static int do_mlock(unsigned long start, size_t len, int on)
589 unsigned long nstart, end, tmp;
590 struct vm_area_struct * vma, * prev;
593 VM_BUG_ON(start & ~PAGE_MASK);
594 VM_BUG_ON(len != PAGE_ALIGN(len));
600 vma = find_vma(current->mm, start);
601 if (!vma || vma->vm_start > start)
605 if (start > vma->vm_start)
608 for (nstart = start ; ; ) {
611 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
613 newflags = vma->vm_flags & ~VM_LOCKED;
615 newflags |= VM_LOCKED;
620 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
624 if (nstart < prev->vm_end)
625 nstart = prev->vm_end;
630 if (!vma || vma->vm_start != nstart) {
639 * __mm_populate - populate and/or mlock pages within a range of address space.
641 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
642 * flags. VMAs must be already marked with the desired vm_flags, and
643 * mmap_sem must not be held.
645 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
647 struct mm_struct *mm = current->mm;
648 unsigned long end, nstart, nend;
649 struct vm_area_struct *vma = NULL;
653 VM_BUG_ON(start & ~PAGE_MASK);
654 VM_BUG_ON(len != PAGE_ALIGN(len));
657 for (nstart = start; nstart < end; nstart = nend) {
659 * We want to fault in pages for [nstart; end) address range.
660 * Find first corresponding VMA.
664 down_read(&mm->mmap_sem);
665 vma = find_vma(mm, nstart);
666 } else if (nstart >= vma->vm_end)
668 if (!vma || vma->vm_start >= end)
671 * Set [nstart; nend) to intersection of desired address
672 * range with the first VMA. Also, skip undesirable VMA types.
674 nend = min(end, vma->vm_end);
675 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
677 if (nstart < vma->vm_start)
678 nstart = vma->vm_start;
680 * Now fault in a range of pages. __mlock_vma_pages_range()
681 * double checks the vma flags, so that it won't mlock pages
682 * if the vma was already munlocked.
684 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
688 continue; /* continue at next VMA */
690 ret = __mlock_posix_error_return(ret);
693 nend = nstart + ret * PAGE_SIZE;
697 up_read(&mm->mmap_sem);
698 return ret; /* 0 or negative error code */
701 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
703 unsigned long locked;
704 unsigned long lock_limit;
710 lru_add_drain_all(); /* flush pagevec */
712 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
715 lock_limit = rlimit(RLIMIT_MEMLOCK);
716 lock_limit >>= PAGE_SHIFT;
717 locked = len >> PAGE_SHIFT;
719 down_write(¤t->mm->mmap_sem);
721 locked += current->mm->locked_vm;
723 /* check against resource limits */
724 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
725 error = do_mlock(start, len, 1);
727 up_write(¤t->mm->mmap_sem);
729 error = __mm_populate(start, len, 0);
733 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
737 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
740 down_write(¤t->mm->mmap_sem);
741 ret = do_mlock(start, len, 0);
742 up_write(¤t->mm->mmap_sem);
747 static int do_mlockall(int flags)
749 struct vm_area_struct * vma, * prev = NULL;
751 if (flags & MCL_FUTURE)
752 current->mm->def_flags |= VM_LOCKED;
754 current->mm->def_flags &= ~VM_LOCKED;
755 if (flags == MCL_FUTURE)
758 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
761 newflags = vma->vm_flags & ~VM_LOCKED;
762 if (flags & MCL_CURRENT)
763 newflags |= VM_LOCKED;
766 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
773 SYSCALL_DEFINE1(mlockall, int, flags)
775 unsigned long lock_limit;
778 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
785 if (flags & MCL_CURRENT)
786 lru_add_drain_all(); /* flush pagevec */
788 lock_limit = rlimit(RLIMIT_MEMLOCK);
789 lock_limit >>= PAGE_SHIFT;
792 down_write(¤t->mm->mmap_sem);
794 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
795 capable(CAP_IPC_LOCK))
796 ret = do_mlockall(flags);
797 up_write(¤t->mm->mmap_sem);
798 if (!ret && (flags & MCL_CURRENT))
799 mm_populate(0, TASK_SIZE);
804 SYSCALL_DEFINE0(munlockall)
808 down_write(¤t->mm->mmap_sem);
809 ret = do_mlockall(0);
810 up_write(¤t->mm->mmap_sem);
815 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
816 * shm segments) get accounted against the user_struct instead.
818 static DEFINE_SPINLOCK(shmlock_user_lock);
820 int user_shm_lock(size_t size, struct user_struct *user)
822 unsigned long lock_limit, locked;
825 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
826 lock_limit = rlimit(RLIMIT_MEMLOCK);
827 if (lock_limit == RLIM_INFINITY)
829 lock_limit >>= PAGE_SHIFT;
830 spin_lock(&shmlock_user_lock);
832 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
835 user->locked_shm += locked;
838 spin_unlock(&shmlock_user_lock);
842 void user_shm_unlock(size_t size, struct user_struct *user)
844 spin_lock(&shmlock_user_lock);
845 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
846 spin_unlock(&shmlock_user_lock);