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 /* Serialize with page migration */
83 BUG_ON(!PageLocked(page));
85 if (!TestSetPageMlocked(page)) {
86 mod_zone_page_state(page_zone(page), NR_MLOCK,
87 hpage_nr_pages(page));
88 count_vm_event(UNEVICTABLE_PGMLOCKED);
89 if (!isolate_lru_page(page))
90 putback_lru_page(page);
95 * Isolate a page from LRU with optional get_page() pin.
96 * Assumes lru_lock already held and page already pinned.
98 static bool __munlock_isolate_lru_page(struct page *page, bool getpage)
101 struct lruvec *lruvec;
103 lruvec = mem_cgroup_page_lruvec(page, page_zone(page));
107 del_page_from_lru_list(page, lruvec, page_lru(page));
115 * Finish munlock after successful page isolation
117 * Page must be locked. This is a wrapper for try_to_munlock()
118 * and putback_lru_page() with munlock accounting.
120 static void __munlock_isolated_page(struct page *page)
122 int ret = SWAP_AGAIN;
125 * Optimization: if the page was mapped just once, that's our mapping
126 * and we don't need to check all the other vmas.
128 if (page_mapcount(page) > 1)
129 ret = try_to_munlock(page);
131 /* Did try_to_unlock() succeed or punt? */
132 if (ret != SWAP_MLOCK)
133 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
135 putback_lru_page(page);
139 * Accounting for page isolation fail during munlock
141 * Performs accounting when page isolation fails in munlock. There is nothing
142 * else to do because it means some other task has already removed the page
143 * from the LRU. putback_lru_page() will take care of removing the page from
144 * the unevictable list, if necessary. vmscan [page_referenced()] will move
145 * the page back to the unevictable list if some other vma has it mlocked.
147 static void __munlock_isolation_failed(struct page *page)
149 if (PageUnevictable(page))
150 __count_vm_event(UNEVICTABLE_PGSTRANDED);
152 __count_vm_event(UNEVICTABLE_PGMUNLOCKED);
156 * munlock_vma_page - munlock a vma page
157 * @page - page to be unlocked, either a normal page or THP page head
159 * returns the size of the page as a page mask (0 for normal page,
160 * HPAGE_PMD_NR - 1 for THP head page)
162 * called from munlock()/munmap() path with page supposedly on the LRU.
163 * When we munlock a page, because the vma where we found the page is being
164 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
165 * page locked so that we can leave it on the unevictable lru list and not
166 * bother vmscan with it. However, to walk the page's rmap list in
167 * try_to_munlock() we must isolate the page from the LRU. If some other
168 * task has removed the page from the LRU, we won't be able to do that.
169 * So we clear the PageMlocked as we might not get another chance. If we
170 * can't isolate the page, we leave it for putback_lru_page() and vmscan
171 * [page_referenced()/try_to_unmap()] to deal with.
173 unsigned int munlock_vma_page(struct page *page)
175 unsigned int nr_pages;
176 struct zone *zone = page_zone(page);
178 /* For try_to_munlock() and to serialize with page migration */
179 BUG_ON(!PageLocked(page));
182 * Serialize with any parallel __split_huge_page_refcount() which
183 * might otherwise copy PageMlocked to part of the tail pages before
184 * we clear it in the head page. It also stabilizes hpage_nr_pages().
186 spin_lock_irq(&zone->lru_lock);
188 nr_pages = hpage_nr_pages(page);
189 if (!TestClearPageMlocked(page))
192 __mod_zone_page_state(zone, NR_MLOCK, -nr_pages);
194 if (__munlock_isolate_lru_page(page, true)) {
195 spin_unlock_irq(&zone->lru_lock);
196 __munlock_isolated_page(page);
199 __munlock_isolation_failed(page);
202 spin_unlock_irq(&zone->lru_lock);
209 * __mlock_vma_pages_range() - mlock a range of pages in the vma.
211 * @start: start address
215 * This takes care of making the pages present too.
217 * return 0 on success, negative error code on error.
219 * vma->vm_mm->mmap_sem must be held.
221 * If @nonblocking is NULL, it may be held for read or write and will
224 * If @nonblocking is non-NULL, it must held for read only and may be
225 * released. If it's released, *@nonblocking will be set to 0.
227 long __mlock_vma_pages_range(struct vm_area_struct *vma,
228 unsigned long start, unsigned long end, int *nonblocking)
230 struct mm_struct *mm = vma->vm_mm;
231 unsigned long nr_pages = (end - start) / PAGE_SIZE;
234 VM_BUG_ON(start & ~PAGE_MASK);
235 VM_BUG_ON(end & ~PAGE_MASK);
236 VM_BUG_ON_VMA(start < vma->vm_start, vma);
237 VM_BUG_ON_VMA(end > vma->vm_end, vma);
238 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
240 gup_flags = FOLL_TOUCH | FOLL_MLOCK;
242 * We want to touch writable mappings with a write fault in order
243 * to break COW, except for shared mappings because these don't COW
244 * and we would not want to dirty them for nothing.
246 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
247 gup_flags |= FOLL_WRITE;
250 * We want mlock to succeed for regions that have any permissions
251 * other than PROT_NONE.
253 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
254 gup_flags |= FOLL_FORCE;
257 * We made sure addr is within a VMA, so the following will
258 * not result in a stack expansion that recurses back here.
260 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
261 NULL, NULL, nonblocking);
265 * convert get_user_pages() return value to posix mlock() error
267 static int __mlock_posix_error_return(long retval)
269 if (retval == -EFAULT)
271 else if (retval == -ENOMEM)
277 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
279 * The fast path is available only for evictable pages with single mapping.
280 * Then we can bypass the per-cpu pvec and get better performance.
281 * when mapcount > 1 we need try_to_munlock() which can fail.
282 * when !page_evictable(), we need the full redo logic of putback_lru_page to
283 * avoid leaving evictable page in unevictable list.
285 * In case of success, @page is added to @pvec and @pgrescued is incremented
286 * in case that the page was previously unevictable. @page is also unlocked.
288 static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
291 VM_BUG_ON_PAGE(PageLRU(page), page);
292 VM_BUG_ON_PAGE(!PageLocked(page), page);
294 if (page_mapcount(page) <= 1 && page_evictable(page)) {
295 pagevec_add(pvec, page);
296 if (TestClearPageUnevictable(page))
306 * Putback multiple evictable pages to the LRU
308 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
309 * the pages might have meanwhile become unevictable but that is OK.
311 static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
313 count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
315 *__pagevec_lru_add() calls release_pages() so we don't call
316 * put_page() explicitly
318 __pagevec_lru_add(pvec);
319 count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
323 * Munlock a batch of pages from the same zone
325 * The work is split to two main phases. First phase clears the Mlocked flag
326 * and attempts to isolate the pages, all under a single zone lru lock.
327 * The second phase finishes the munlock only for pages where isolation
330 * Note that the pagevec may be modified during the process.
332 static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
335 int nr = pagevec_count(pvec);
337 struct pagevec pvec_putback;
340 pagevec_init(&pvec_putback, 0);
342 /* Phase 1: page isolation */
343 spin_lock_irq(&zone->lru_lock);
344 for (i = 0; i < nr; i++) {
345 struct page *page = pvec->pages[i];
347 if (TestClearPageMlocked(page)) {
349 * We already have pin from follow_page_mask()
350 * so we can spare the get_page() here.
352 if (__munlock_isolate_lru_page(page, false))
355 __munlock_isolation_failed(page);
359 * We won't be munlocking this page in the next phase
360 * but we still need to release the follow_page_mask()
361 * pin. We cannot do it under lru_lock however. If it's
362 * the last pin, __page_cache_release() would deadlock.
364 pagevec_add(&pvec_putback, pvec->pages[i]);
365 pvec->pages[i] = NULL;
367 delta_munlocked = -nr + pagevec_count(&pvec_putback);
368 __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
369 spin_unlock_irq(&zone->lru_lock);
371 /* Now we can release pins of pages that we are not munlocking */
372 pagevec_release(&pvec_putback);
374 /* Phase 2: page munlock */
375 for (i = 0; i < nr; i++) {
376 struct page *page = pvec->pages[i];
380 if (!__putback_lru_fast_prepare(page, &pvec_putback,
383 * Slow path. We don't want to lose the last
384 * pin before unlock_page()
386 get_page(page); /* for putback_lru_page() */
387 __munlock_isolated_page(page);
389 put_page(page); /* from follow_page_mask() */
395 * Phase 3: page putback for pages that qualified for the fast path
396 * This will also call put_page() to return pin from follow_page_mask()
398 if (pagevec_count(&pvec_putback))
399 __putback_lru_fast(&pvec_putback, pgrescued);
403 * Fill up pagevec for __munlock_pagevec using pte walk
405 * The function expects that the struct page corresponding to @start address is
406 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
408 * The rest of @pvec is filled by subsequent pages within the same pmd and same
409 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
410 * pages also get pinned.
412 * Returns the address of the next page that should be scanned. This equals
413 * @start + PAGE_SIZE when no page could be added by the pte walk.
415 static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
416 struct vm_area_struct *vma, int zoneid, unsigned long start,
423 * Initialize pte walk starting at the already pinned page where we
424 * are sure that there is a pte, as it was pinned under the same
427 pte = get_locked_pte(vma->vm_mm, start, &ptl);
428 /* Make sure we do not cross the page table boundary */
429 end = pgd_addr_end(start, end);
430 end = pud_addr_end(start, end);
431 end = pmd_addr_end(start, end);
433 /* The page next to the pinned page is the first we will try to get */
435 while (start < end) {
436 struct page *page = NULL;
438 if (pte_present(*pte))
439 page = vm_normal_page(vma, start, *pte);
441 * Break if page could not be obtained or the page's node+zone does not
444 if (!page || page_zone_id(page) != zoneid)
449 * Increase the address that will be returned *before* the
450 * eventual break due to pvec becoming full by adding the page
453 if (pagevec_add(pvec, page) == 0)
456 pte_unmap_unlock(pte, ptl);
461 * munlock_vma_pages_range() - munlock all pages in the vma range.'
462 * @vma - vma containing range to be munlock()ed.
463 * @start - start address in @vma of the range
464 * @end - end of range in @vma.
466 * For mremap(), munmap() and exit().
468 * Called with @vma VM_LOCKED.
470 * Returns with VM_LOCKED cleared. Callers must be prepared to
473 * We don't save and restore VM_LOCKED here because pages are
474 * still on lru. In unmap path, pages might be scanned by reclaim
475 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
476 * free them. This will result in freeing mlocked pages.
478 void munlock_vma_pages_range(struct vm_area_struct *vma,
479 unsigned long start, unsigned long end)
481 vma->vm_flags &= ~VM_LOCKED;
483 while (start < end) {
484 struct page *page = NULL;
485 unsigned int page_mask;
486 unsigned long page_increm;
491 pagevec_init(&pvec, 0);
493 * Although FOLL_DUMP is intended for get_dump_page(),
494 * it just so happens that its special treatment of the
495 * ZERO_PAGE (returning an error instead of doing get_page)
496 * suits munlock very well (and if somehow an abnormal page
497 * has sneaked into the range, we won't oops here: great).
499 page = follow_page_mask(vma, start, FOLL_GET | FOLL_DUMP,
502 if (page && !IS_ERR(page)) {
503 if (PageTransHuge(page)) {
506 * Any THP page found by follow_page_mask() may
507 * have gotten split before reaching
508 * munlock_vma_page(), so we need to recompute
509 * the page_mask here.
511 page_mask = munlock_vma_page(page);
513 put_page(page); /* follow_page_mask() */
516 * Non-huge pages are handled in batches via
517 * pagevec. The pin from follow_page_mask()
518 * prevents them from collapsing by THP.
520 pagevec_add(&pvec, page);
521 zone = page_zone(page);
522 zoneid = page_zone_id(page);
525 * Try to fill the rest of pagevec using fast
526 * pte walk. This will also update start to
527 * the next page to process. Then munlock the
530 start = __munlock_pagevec_fill(&pvec, vma,
532 __munlock_pagevec(&pvec, zone);
536 /* It's a bug to munlock in the middle of a THP page */
537 VM_BUG_ON((start >> PAGE_SHIFT) & page_mask);
538 page_increm = 1 + page_mask;
539 start += page_increm * PAGE_SIZE;
546 * mlock_fixup - handle mlock[all]/munlock[all] requests.
548 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
549 * munlock is a no-op. However, for some special vmas, we go ahead and
552 * For vmas that pass the filters, merge/split as appropriate.
554 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
555 unsigned long start, unsigned long end, vm_flags_t newflags)
557 struct mm_struct *mm = vma->vm_mm;
561 int lock = !!(newflags & VM_LOCKED);
563 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
564 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
565 goto out; /* don't set VM_LOCKED, don't count */
567 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
568 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
569 vma->vm_file, pgoff, vma_policy(vma));
575 if (start != vma->vm_start) {
576 ret = split_vma(mm, vma, start, 1);
581 if (end != vma->vm_end) {
582 ret = split_vma(mm, vma, end, 0);
589 * Keep track of amount of locked VM.
591 nr_pages = (end - start) >> PAGE_SHIFT;
593 nr_pages = -nr_pages;
594 mm->locked_vm += nr_pages;
597 * vm_flags is protected by the mmap_sem held in write mode.
598 * It's okay if try_to_unmap_one unmaps a page just after we
599 * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
603 vma->vm_flags = newflags;
605 munlock_vma_pages_range(vma, start, end);
612 static int do_mlock(unsigned long start, size_t len, int on)
614 unsigned long nstart, end, tmp;
615 struct vm_area_struct * vma, * prev;
618 VM_BUG_ON(start & ~PAGE_MASK);
619 VM_BUG_ON(len != PAGE_ALIGN(len));
625 vma = find_vma(current->mm, start);
626 if (!vma || vma->vm_start > start)
630 if (start > vma->vm_start)
633 for (nstart = start ; ; ) {
636 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
638 newflags = vma->vm_flags & ~VM_LOCKED;
640 newflags |= VM_LOCKED;
645 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
649 if (nstart < prev->vm_end)
650 nstart = prev->vm_end;
655 if (!vma || vma->vm_start != nstart) {
664 * __mm_populate - populate and/or mlock pages within a range of address space.
666 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
667 * flags. VMAs must be already marked with the desired vm_flags, and
668 * mmap_sem must not be held.
670 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
672 struct mm_struct *mm = current->mm;
673 unsigned long end, nstart, nend;
674 struct vm_area_struct *vma = NULL;
678 VM_BUG_ON(start & ~PAGE_MASK);
679 VM_BUG_ON(len != PAGE_ALIGN(len));
682 for (nstart = start; nstart < end; nstart = nend) {
684 * We want to fault in pages for [nstart; end) address range.
685 * Find first corresponding VMA.
689 down_read(&mm->mmap_sem);
690 vma = find_vma(mm, nstart);
691 } else if (nstart >= vma->vm_end)
693 if (!vma || vma->vm_start >= end)
696 * Set [nstart; nend) to intersection of desired address
697 * range with the first VMA. Also, skip undesirable VMA types.
699 nend = min(end, vma->vm_end);
700 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
702 if (nstart < vma->vm_start)
703 nstart = vma->vm_start;
705 * Now fault in a range of pages. __mlock_vma_pages_range()
706 * double checks the vma flags, so that it won't mlock pages
707 * if the vma was already munlocked.
709 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
713 continue; /* continue at next VMA */
715 ret = __mlock_posix_error_return(ret);
718 nend = nstart + ret * PAGE_SIZE;
722 up_read(&mm->mmap_sem);
723 return ret; /* 0 or negative error code */
726 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
728 unsigned long locked;
729 unsigned long lock_limit;
735 lru_add_drain_all(); /* flush pagevec */
737 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
740 lock_limit = rlimit(RLIMIT_MEMLOCK);
741 lock_limit >>= PAGE_SHIFT;
742 locked = len >> PAGE_SHIFT;
744 down_write(¤t->mm->mmap_sem);
746 locked += current->mm->locked_vm;
748 /* check against resource limits */
749 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
750 error = do_mlock(start, len, 1);
752 up_write(¤t->mm->mmap_sem);
754 error = __mm_populate(start, len, 0);
758 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
762 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
765 down_write(¤t->mm->mmap_sem);
766 ret = do_mlock(start, len, 0);
767 up_write(¤t->mm->mmap_sem);
772 static int do_mlockall(int flags)
774 struct vm_area_struct * vma, * prev = NULL;
776 if (flags & MCL_FUTURE)
777 current->mm->def_flags |= VM_LOCKED;
779 current->mm->def_flags &= ~VM_LOCKED;
780 if (flags == MCL_FUTURE)
783 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
786 newflags = vma->vm_flags & ~VM_LOCKED;
787 if (flags & MCL_CURRENT)
788 newflags |= VM_LOCKED;
791 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
792 cond_resched_rcu_qs();
798 SYSCALL_DEFINE1(mlockall, int, flags)
800 unsigned long lock_limit;
803 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
810 if (flags & MCL_CURRENT)
811 lru_add_drain_all(); /* flush pagevec */
813 lock_limit = rlimit(RLIMIT_MEMLOCK);
814 lock_limit >>= PAGE_SHIFT;
817 down_write(¤t->mm->mmap_sem);
819 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
820 capable(CAP_IPC_LOCK))
821 ret = do_mlockall(flags);
822 up_write(¤t->mm->mmap_sem);
823 if (!ret && (flags & MCL_CURRENT))
824 mm_populate(0, TASK_SIZE);
829 SYSCALL_DEFINE0(munlockall)
833 down_write(¤t->mm->mmap_sem);
834 ret = do_mlockall(0);
835 up_write(¤t->mm->mmap_sem);
840 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
841 * shm segments) get accounted against the user_struct instead.
843 static DEFINE_SPINLOCK(shmlock_user_lock);
845 int user_shm_lock(size_t size, struct user_struct *user)
847 unsigned long lock_limit, locked;
850 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
851 lock_limit = rlimit(RLIMIT_MEMLOCK);
852 if (lock_limit == RLIM_INFINITY)
854 lock_limit >>= PAGE_SHIFT;
855 spin_lock(&shmlock_user_lock);
857 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
860 user->locked_shm += locked;
863 spin_unlock(&shmlock_user_lock);
867 void user_shm_unlock(size_t size, struct user_struct *user)
869 spin_lock(&shmlock_user_lock);
870 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
871 spin_unlock(&shmlock_user_lock);