1 // SPDX-License-Identifier: GPL-2.0
5 * (C) Copyright 1995 Linus Torvalds
6 * (C) Copyright 2002 Christoph Hellwig
9 #include <linux/capability.h>
10 #include <linux/mman.h>
12 #include <linux/sched/user.h>
13 #include <linux/swap.h>
14 #include <linux/swapops.h>
15 #include <linux/pagemap.h>
16 #include <linux/pagevec.h>
17 #include <linux/mempolicy.h>
18 #include <linux/syscalls.h>
19 #include <linux/sched.h>
20 #include <linux/export.h>
21 #include <linux/rmap.h>
22 #include <linux/mmzone.h>
23 #include <linux/hugetlb.h>
24 #include <linux/memcontrol.h>
25 #include <linux/mm_inline.h>
29 bool can_do_mlock(void)
31 if (rlimit(RLIMIT_MEMLOCK) != 0)
33 if (capable(CAP_IPC_LOCK))
37 EXPORT_SYMBOL(can_do_mlock);
40 * Mlocked pages are marked with PageMlocked() flag for efficient testing
41 * in vmscan and, possibly, the fault path; and to support semi-accurate
44 * An mlocked page [PageMlocked(page)] is unevictable. As such, it will
45 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
46 * The unevictable list is an LRU sibling list to the [in]active lists.
47 * PageUnevictable is set to indicate the unevictable state.
49 * When lazy mlocking via vmscan, it is important to ensure that the
50 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
51 * may have mlocked a page that is being munlocked. So lazy mlock must take
52 * the mmap_lock for read, and verify that the vma really is locked
57 * LRU accounting for clear_page_mlock()
59 void clear_page_mlock(struct page *page)
63 if (!TestClearPageMlocked(page))
66 nr_pages = thp_nr_pages(page);
67 mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
68 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
70 * The previous TestClearPageMlocked() corresponds to the smp_mb()
71 * in __pagevec_lru_add_fn().
73 * See __pagevec_lru_add_fn for more explanation.
75 if (!isolate_lru_page(page)) {
76 putback_lru_page(page);
79 * We lost the race. the page already moved to evictable list.
81 if (PageUnevictable(page))
82 count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
87 * Mark page as mlocked if not already.
88 * If page on LRU, isolate and putback to move to unevictable list.
90 void mlock_vma_page(struct page *page)
92 /* Serialize with page migration */
93 BUG_ON(!PageLocked(page));
95 VM_BUG_ON_PAGE(PageTail(page), page);
96 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
98 if (!TestSetPageMlocked(page)) {
99 int nr_pages = thp_nr_pages(page);
101 mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
102 count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
103 if (!isolate_lru_page(page))
104 putback_lru_page(page);
109 * Finish munlock after successful page isolation
111 * Page must be locked. This is a wrapper for try_to_munlock()
112 * and putback_lru_page() with munlock accounting.
114 static void __munlock_isolated_page(struct page *page)
117 * Optimization: if the page was mapped just once, that's our mapping
118 * and we don't need to check all the other vmas.
120 if (page_mapcount(page) > 1)
121 try_to_munlock(page);
123 /* Did try_to_unlock() succeed or punt? */
124 if (!PageMlocked(page))
125 count_vm_events(UNEVICTABLE_PGMUNLOCKED, thp_nr_pages(page));
127 putback_lru_page(page);
131 * Accounting for page isolation fail during munlock
133 * Performs accounting when page isolation fails in munlock. There is nothing
134 * else to do because it means some other task has already removed the page
135 * from the LRU. putback_lru_page() will take care of removing the page from
136 * the unevictable list, if necessary. vmscan [page_referenced()] will move
137 * the page back to the unevictable list if some other vma has it mlocked.
139 static void __munlock_isolation_failed(struct page *page)
141 int nr_pages = thp_nr_pages(page);
143 if (PageUnevictable(page))
144 __count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
146 __count_vm_events(UNEVICTABLE_PGMUNLOCKED, nr_pages);
150 * munlock_vma_page - munlock a vma page
151 * @page: page to be unlocked, either a normal page or THP page head
153 * returns the size of the page as a page mask (0 for normal page,
154 * HPAGE_PMD_NR - 1 for THP head page)
156 * called from munlock()/munmap() path with page supposedly on the LRU.
157 * When we munlock a page, because the vma where we found the page is being
158 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
159 * page locked so that we can leave it on the unevictable lru list and not
160 * bother vmscan with it. However, to walk the page's rmap list in
161 * try_to_munlock() we must isolate the page from the LRU. If some other
162 * task has removed the page from the LRU, we won't be able to do that.
163 * So we clear the PageMlocked as we might not get another chance. If we
164 * can't isolate the page, we leave it for putback_lru_page() and vmscan
165 * [page_referenced()/try_to_unmap()] to deal with.
167 unsigned int munlock_vma_page(struct page *page)
171 /* For try_to_munlock() and to serialize with page migration */
172 BUG_ON(!PageLocked(page));
173 VM_BUG_ON_PAGE(PageTail(page), page);
175 if (!TestClearPageMlocked(page)) {
176 /* Potentially, PTE-mapped THP: do not skip the rest PTEs */
180 nr_pages = thp_nr_pages(page);
181 mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
183 if (!isolate_lru_page(page))
184 __munlock_isolated_page(page);
186 __munlock_isolation_failed(page);
192 * convert get_user_pages() return value to posix mlock() error
194 static int __mlock_posix_error_return(long retval)
196 if (retval == -EFAULT)
198 else if (retval == -ENOMEM)
204 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
206 * The fast path is available only for evictable pages with single mapping.
207 * Then we can bypass the per-cpu pvec and get better performance.
208 * when mapcount > 1 we need try_to_munlock() which can fail.
209 * when !page_evictable(), we need the full redo logic of putback_lru_page to
210 * avoid leaving evictable page in unevictable list.
212 * In case of success, @page is added to @pvec and @pgrescued is incremented
213 * in case that the page was previously unevictable. @page is also unlocked.
215 static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
218 VM_BUG_ON_PAGE(PageLRU(page), page);
219 VM_BUG_ON_PAGE(!PageLocked(page), page);
221 if (page_mapcount(page) <= 1 && page_evictable(page)) {
222 pagevec_add(pvec, page);
223 if (TestClearPageUnevictable(page))
233 * Putback multiple evictable pages to the LRU
235 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
236 * the pages might have meanwhile become unevictable but that is OK.
238 static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
240 count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
242 *__pagevec_lru_add() calls release_pages() so we don't call
243 * put_page() explicitly
245 __pagevec_lru_add(pvec);
246 count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
250 * Munlock a batch of pages from the same zone
252 * The work is split to two main phases. First phase clears the Mlocked flag
253 * and attempts to isolate the pages, all under a single zone lru lock.
254 * The second phase finishes the munlock only for pages where isolation
257 * Note that the pagevec may be modified during the process.
259 static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
262 int nr = pagevec_count(pvec);
263 int delta_munlocked = -nr;
264 struct pagevec pvec_putback;
265 struct lruvec *lruvec = NULL;
268 pagevec_init(&pvec_putback);
270 /* Phase 1: page isolation */
271 for (i = 0; i < nr; i++) {
272 struct page *page = pvec->pages[i];
274 if (TestClearPageMlocked(page)) {
276 * We already have pin from follow_page_mask()
277 * so we can spare the get_page() here.
279 if (TestClearPageLRU(page)) {
280 lruvec = relock_page_lruvec_irq(page, lruvec);
281 del_page_from_lru_list(page, lruvec);
284 __munlock_isolation_failed(page);
290 * We won't be munlocking this page in the next phase
291 * but we still need to release the follow_page_mask()
292 * pin. We cannot do it under lru_lock however. If it's
293 * the last pin, __page_cache_release() would deadlock.
295 pagevec_add(&pvec_putback, pvec->pages[i]);
296 pvec->pages[i] = NULL;
299 __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
300 unlock_page_lruvec_irq(lruvec);
301 } else if (delta_munlocked) {
302 mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
305 /* Now we can release pins of pages that we are not munlocking */
306 pagevec_release(&pvec_putback);
308 /* Phase 2: page munlock */
309 for (i = 0; i < nr; i++) {
310 struct page *page = pvec->pages[i];
314 if (!__putback_lru_fast_prepare(page, &pvec_putback,
317 * Slow path. We don't want to lose the last
318 * pin before unlock_page()
320 get_page(page); /* for putback_lru_page() */
321 __munlock_isolated_page(page);
323 put_page(page); /* from follow_page_mask() */
329 * Phase 3: page putback for pages that qualified for the fast path
330 * This will also call put_page() to return pin from follow_page_mask()
332 if (pagevec_count(&pvec_putback))
333 __putback_lru_fast(&pvec_putback, pgrescued);
337 * Fill up pagevec for __munlock_pagevec using pte walk
339 * The function expects that the struct page corresponding to @start address is
340 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
342 * The rest of @pvec is filled by subsequent pages within the same pmd and same
343 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
344 * pages also get pinned.
346 * Returns the address of the next page that should be scanned. This equals
347 * @start + PAGE_SIZE when no page could be added by the pte walk.
349 static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
350 struct vm_area_struct *vma, struct zone *zone,
351 unsigned long start, unsigned long end)
357 * Initialize pte walk starting at the already pinned page where we
358 * are sure that there is a pte, as it was pinned under the same
359 * mmap_lock write op.
361 pte = get_locked_pte(vma->vm_mm, start, &ptl);
362 /* Make sure we do not cross the page table boundary */
363 end = pgd_addr_end(start, end);
364 end = p4d_addr_end(start, end);
365 end = pud_addr_end(start, end);
366 end = pmd_addr_end(start, end);
368 /* The page next to the pinned page is the first we will try to get */
370 while (start < end) {
371 struct page *page = NULL;
373 if (pte_present(*pte))
374 page = vm_normal_page(vma, start, *pte);
376 * Break if page could not be obtained or the page's node+zone does not
379 if (!page || page_zone(page) != zone)
383 * Do not use pagevec for PTE-mapped THP,
384 * munlock_vma_pages_range() will handle them.
386 if (PageTransCompound(page))
391 * Increase the address that will be returned *before* the
392 * eventual break due to pvec becoming full by adding the page
395 if (pagevec_add(pvec, page) == 0)
398 pte_unmap_unlock(pte, ptl);
403 * munlock_vma_pages_range() - munlock all pages in the vma range.'
404 * @vma - vma containing range to be munlock()ed.
405 * @start - start address in @vma of the range
406 * @end - end of range in @vma.
408 * For mremap(), munmap() and exit().
410 * Called with @vma VM_LOCKED.
412 * Returns with VM_LOCKED cleared. Callers must be prepared to
415 * We don't save and restore VM_LOCKED here because pages are
416 * still on lru. In unmap path, pages might be scanned by reclaim
417 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
418 * free them. This will result in freeing mlocked pages.
420 void munlock_vma_pages_range(struct vm_area_struct *vma,
421 unsigned long start, unsigned long end)
423 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
425 while (start < end) {
427 unsigned int page_mask = 0;
428 unsigned long page_increm;
434 * Although FOLL_DUMP is intended for get_dump_page(),
435 * it just so happens that its special treatment of the
436 * ZERO_PAGE (returning an error instead of doing get_page)
437 * suits munlock very well (and if somehow an abnormal page
438 * has sneaked into the range, we won't oops here: great).
440 page = follow_page(vma, start, FOLL_GET | FOLL_DUMP);
442 if (page && !IS_ERR(page)) {
443 if (PageTransTail(page)) {
444 VM_BUG_ON_PAGE(PageMlocked(page), page);
445 put_page(page); /* follow_page_mask() */
446 } else if (PageTransHuge(page)) {
449 * Any THP page found by follow_page_mask() may
450 * have gotten split before reaching
451 * munlock_vma_page(), so we need to compute
452 * the page_mask here instead.
454 page_mask = munlock_vma_page(page);
456 put_page(page); /* follow_page_mask() */
459 * Non-huge pages are handled in batches via
460 * pagevec. The pin from follow_page_mask()
461 * prevents them from collapsing by THP.
463 pagevec_add(&pvec, page);
464 zone = page_zone(page);
467 * Try to fill the rest of pagevec using fast
468 * pte walk. This will also update start to
469 * the next page to process. Then munlock the
472 start = __munlock_pagevec_fill(&pvec, vma,
474 __munlock_pagevec(&pvec, zone);
478 page_increm = 1 + page_mask;
479 start += page_increm * PAGE_SIZE;
486 * mlock_fixup - handle mlock[all]/munlock[all] requests.
488 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
489 * munlock is a no-op. However, for some special vmas, we go ahead and
492 * For vmas that pass the filters, merge/split as appropriate.
494 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
495 unsigned long start, unsigned long end, vm_flags_t newflags)
497 struct mm_struct *mm = vma->vm_mm;
501 int lock = !!(newflags & VM_LOCKED);
502 vm_flags_t old_flags = vma->vm_flags;
504 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
505 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm) ||
507 /* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
510 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
511 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
512 vma->vm_file, pgoff, vma_policy(vma),
513 vma->vm_userfaultfd_ctx);
519 if (start != vma->vm_start) {
520 ret = split_vma(mm, vma, start, 1);
525 if (end != vma->vm_end) {
526 ret = split_vma(mm, vma, end, 0);
533 * Keep track of amount of locked VM.
535 nr_pages = (end - start) >> PAGE_SHIFT;
537 nr_pages = -nr_pages;
538 else if (old_flags & VM_LOCKED)
540 mm->locked_vm += nr_pages;
543 * vm_flags is protected by the mmap_lock held in write mode.
544 * It's okay if try_to_unmap_one unmaps a page just after we
545 * set VM_LOCKED, populate_vma_page_range will bring it back.
549 vma->vm_flags = newflags;
551 munlock_vma_pages_range(vma, start, end);
558 static int apply_vma_lock_flags(unsigned long start, size_t len,
561 unsigned long nstart, end, tmp;
562 struct vm_area_struct *vma, *prev;
565 VM_BUG_ON(offset_in_page(start));
566 VM_BUG_ON(len != PAGE_ALIGN(len));
572 vma = find_vma(current->mm, start);
573 if (!vma || vma->vm_start > start)
577 if (start > vma->vm_start)
580 for (nstart = start ; ; ) {
581 vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
585 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
589 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
593 if (nstart < prev->vm_end)
594 nstart = prev->vm_end;
599 if (!vma || vma->vm_start != nstart) {
608 * Go through vma areas and sum size of mlocked
609 * vma pages, as return value.
610 * Note deferred memory locking case(mlock2(,,MLOCK_ONFAULT)
612 * Return value: previously mlocked page counts
614 static unsigned long count_mm_mlocked_page_nr(struct mm_struct *mm,
615 unsigned long start, size_t len)
617 struct vm_area_struct *vma;
618 unsigned long count = 0;
623 vma = find_vma(mm, start);
627 for (; vma ; vma = vma->vm_next) {
628 if (start >= vma->vm_end)
630 if (start + len <= vma->vm_start)
632 if (vma->vm_flags & VM_LOCKED) {
633 if (start > vma->vm_start)
634 count -= (start - vma->vm_start);
635 if (start + len < vma->vm_end) {
636 count += start + len - vma->vm_start;
639 count += vma->vm_end - vma->vm_start;
643 return count >> PAGE_SHIFT;
646 static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
648 unsigned long locked;
649 unsigned long lock_limit;
652 start = untagged_addr(start);
657 len = PAGE_ALIGN(len + (offset_in_page(start)));
660 lock_limit = rlimit(RLIMIT_MEMLOCK);
661 lock_limit >>= PAGE_SHIFT;
662 locked = len >> PAGE_SHIFT;
664 if (mmap_write_lock_killable(current->mm))
667 locked += current->mm->locked_vm;
668 if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
670 * It is possible that the regions requested intersect with
671 * previously mlocked areas, that part area in "mm->locked_vm"
672 * should not be counted to new mlock increment count. So check
673 * and adjust locked count if necessary.
675 locked -= count_mm_mlocked_page_nr(current->mm,
679 /* check against resource limits */
680 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
681 error = apply_vma_lock_flags(start, len, flags);
683 mmap_write_unlock(current->mm);
687 error = __mm_populate(start, len, 0);
689 return __mlock_posix_error_return(error);
693 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
695 return do_mlock(start, len, VM_LOCKED);
698 SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
700 vm_flags_t vm_flags = VM_LOCKED;
702 if (flags & ~MLOCK_ONFAULT)
705 if (flags & MLOCK_ONFAULT)
706 vm_flags |= VM_LOCKONFAULT;
708 return do_mlock(start, len, vm_flags);
711 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
715 start = untagged_addr(start);
717 len = PAGE_ALIGN(len + (offset_in_page(start)));
720 if (mmap_write_lock_killable(current->mm))
722 ret = apply_vma_lock_flags(start, len, 0);
723 mmap_write_unlock(current->mm);
729 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
730 * and translate into the appropriate modifications to mm->def_flags and/or the
731 * flags for all current VMAs.
733 * There are a couple of subtleties with this. If mlockall() is called multiple
734 * times with different flags, the values do not necessarily stack. If mlockall
735 * is called once including the MCL_FUTURE flag and then a second time without
736 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
738 static int apply_mlockall_flags(int flags)
740 struct vm_area_struct *vma, *prev = NULL;
741 vm_flags_t to_add = 0;
743 current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
744 if (flags & MCL_FUTURE) {
745 current->mm->def_flags |= VM_LOCKED;
747 if (flags & MCL_ONFAULT)
748 current->mm->def_flags |= VM_LOCKONFAULT;
750 if (!(flags & MCL_CURRENT))
754 if (flags & MCL_CURRENT) {
756 if (flags & MCL_ONFAULT)
757 to_add |= VM_LOCKONFAULT;
760 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
763 newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
767 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
774 SYSCALL_DEFINE1(mlockall, int, flags)
776 unsigned long lock_limit;
779 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)) ||
780 flags == MCL_ONFAULT)
786 lock_limit = rlimit(RLIMIT_MEMLOCK);
787 lock_limit >>= PAGE_SHIFT;
789 if (mmap_write_lock_killable(current->mm))
793 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
794 capable(CAP_IPC_LOCK))
795 ret = apply_mlockall_flags(flags);
796 mmap_write_unlock(current->mm);
797 if (!ret && (flags & MCL_CURRENT))
798 mm_populate(0, TASK_SIZE);
803 SYSCALL_DEFINE0(munlockall)
807 if (mmap_write_lock_killable(current->mm))
809 ret = apply_mlockall_flags(0);
810 mmap_write_unlock(current->mm);
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);