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_sem 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)
61 if (!TestClearPageMlocked(page))
64 mod_zone_page_state(page_zone(page), NR_MLOCK,
65 -hpage_nr_pages(page));
66 count_vm_event(UNEVICTABLE_PGCLEARED);
68 * The previous TestClearPageMlocked() corresponds to the smp_mb()
69 * in __pagevec_lru_add_fn().
71 * See __pagevec_lru_add_fn for more explanation.
73 if (!isolate_lru_page(page)) {
74 putback_lru_page(page);
77 * We lost the race. the page already moved to evictable list.
79 if (PageUnevictable(page))
80 count_vm_event(UNEVICTABLE_PGSTRANDED);
85 * Mark page as mlocked if not already.
86 * If page on LRU, isolate and putback to move to unevictable list.
88 void mlock_vma_page(struct page *page)
90 /* Serialize with page migration */
91 BUG_ON(!PageLocked(page));
93 VM_BUG_ON_PAGE(PageTail(page), page);
94 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
96 if (!TestSetPageMlocked(page)) {
97 mod_zone_page_state(page_zone(page), NR_MLOCK,
98 hpage_nr_pages(page));
99 count_vm_event(UNEVICTABLE_PGMLOCKED);
100 if (!isolate_lru_page(page))
101 putback_lru_page(page);
106 * Isolate a page from LRU with optional get_page() pin.
107 * Assumes lru_lock already held and page already pinned.
109 static bool __munlock_isolate_lru_page(struct page *page, bool getpage)
112 struct lruvec *lruvec;
114 lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
118 del_page_from_lru_list(page, lruvec, page_lru(page));
126 * Finish munlock after successful page isolation
128 * Page must be locked. This is a wrapper for try_to_munlock()
129 * and putback_lru_page() with munlock accounting.
131 static void __munlock_isolated_page(struct page *page)
134 * Optimization: if the page was mapped just once, that's our mapping
135 * and we don't need to check all the other vmas.
137 if (page_mapcount(page) > 1)
138 try_to_munlock(page);
140 /* Did try_to_unlock() succeed or punt? */
141 if (!PageMlocked(page))
142 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
144 putback_lru_page(page);
148 * Accounting for page isolation fail during munlock
150 * Performs accounting when page isolation fails in munlock. There is nothing
151 * else to do because it means some other task has already removed the page
152 * from the LRU. putback_lru_page() will take care of removing the page from
153 * the unevictable list, if necessary. vmscan [page_referenced()] will move
154 * the page back to the unevictable list if some other vma has it mlocked.
156 static void __munlock_isolation_failed(struct page *page)
158 if (PageUnevictable(page))
159 __count_vm_event(UNEVICTABLE_PGSTRANDED);
161 __count_vm_event(UNEVICTABLE_PGMUNLOCKED);
165 * munlock_vma_page - munlock a vma page
166 * @page: page to be unlocked, either a normal page or THP page head
168 * returns the size of the page as a page mask (0 for normal page,
169 * HPAGE_PMD_NR - 1 for THP head page)
171 * called from munlock()/munmap() path with page supposedly on the LRU.
172 * When we munlock a page, because the vma where we found the page is being
173 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
174 * page locked so that we can leave it on the unevictable lru list and not
175 * bother vmscan with it. However, to walk the page's rmap list in
176 * try_to_munlock() we must isolate the page from the LRU. If some other
177 * task has removed the page from the LRU, we won't be able to do that.
178 * So we clear the PageMlocked as we might not get another chance. If we
179 * can't isolate the page, we leave it for putback_lru_page() and vmscan
180 * [page_referenced()/try_to_unmap()] to deal with.
182 unsigned int munlock_vma_page(struct page *page)
185 pg_data_t *pgdat = page_pgdat(page);
187 /* For try_to_munlock() and to serialize with page migration */
188 BUG_ON(!PageLocked(page));
190 VM_BUG_ON_PAGE(PageTail(page), page);
193 * Serialize with any parallel __split_huge_page_refcount() which
194 * might otherwise copy PageMlocked to part of the tail pages before
195 * we clear it in the head page. It also stabilizes hpage_nr_pages().
197 spin_lock_irq(&pgdat->lru_lock);
199 if (!TestClearPageMlocked(page)) {
200 /* Potentially, PTE-mapped THP: do not skip the rest PTEs */
205 nr_pages = hpage_nr_pages(page);
206 __mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
208 if (__munlock_isolate_lru_page(page, true)) {
209 spin_unlock_irq(&pgdat->lru_lock);
210 __munlock_isolated_page(page);
213 __munlock_isolation_failed(page);
216 spin_unlock_irq(&pgdat->lru_lock);
223 * convert get_user_pages() return value to posix mlock() error
225 static int __mlock_posix_error_return(long retval)
227 if (retval == -EFAULT)
229 else if (retval == -ENOMEM)
235 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
237 * The fast path is available only for evictable pages with single mapping.
238 * Then we can bypass the per-cpu pvec and get better performance.
239 * when mapcount > 1 we need try_to_munlock() which can fail.
240 * when !page_evictable(), we need the full redo logic of putback_lru_page to
241 * avoid leaving evictable page in unevictable list.
243 * In case of success, @page is added to @pvec and @pgrescued is incremented
244 * in case that the page was previously unevictable. @page is also unlocked.
246 static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
249 VM_BUG_ON_PAGE(PageLRU(page), page);
250 VM_BUG_ON_PAGE(!PageLocked(page), page);
252 if (page_mapcount(page) <= 1 && page_evictable(page)) {
253 pagevec_add(pvec, page);
254 if (TestClearPageUnevictable(page))
264 * Putback multiple evictable pages to the LRU
266 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
267 * the pages might have meanwhile become unevictable but that is OK.
269 static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
271 count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
273 *__pagevec_lru_add() calls release_pages() so we don't call
274 * put_page() explicitly
276 __pagevec_lru_add(pvec);
277 count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
281 * Munlock a batch of pages from the same zone
283 * The work is split to two main phases. First phase clears the Mlocked flag
284 * and attempts to isolate the pages, all under a single zone lru lock.
285 * The second phase finishes the munlock only for pages where isolation
288 * Note that the pagevec may be modified during the process.
290 static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
293 int nr = pagevec_count(pvec);
294 int delta_munlocked = -nr;
295 struct pagevec pvec_putback;
298 pagevec_init(&pvec_putback);
300 /* Phase 1: page isolation */
301 spin_lock_irq(&zone->zone_pgdat->lru_lock);
302 for (i = 0; i < nr; i++) {
303 struct page *page = pvec->pages[i];
305 if (TestClearPageMlocked(page)) {
307 * We already have pin from follow_page_mask()
308 * so we can spare the get_page() here.
310 if (__munlock_isolate_lru_page(page, false))
313 __munlock_isolation_failed(page);
319 * We won't be munlocking this page in the next phase
320 * but we still need to release the follow_page_mask()
321 * pin. We cannot do it under lru_lock however. If it's
322 * the last pin, __page_cache_release() would deadlock.
324 pagevec_add(&pvec_putback, pvec->pages[i]);
325 pvec->pages[i] = NULL;
327 __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
328 spin_unlock_irq(&zone->zone_pgdat->lru_lock);
330 /* Now we can release pins of pages that we are not munlocking */
331 pagevec_release(&pvec_putback);
333 /* Phase 2: page munlock */
334 for (i = 0; i < nr; i++) {
335 struct page *page = pvec->pages[i];
339 if (!__putback_lru_fast_prepare(page, &pvec_putback,
342 * Slow path. We don't want to lose the last
343 * pin before unlock_page()
345 get_page(page); /* for putback_lru_page() */
346 __munlock_isolated_page(page);
348 put_page(page); /* from follow_page_mask() */
354 * Phase 3: page putback for pages that qualified for the fast path
355 * This will also call put_page() to return pin from follow_page_mask()
357 if (pagevec_count(&pvec_putback))
358 __putback_lru_fast(&pvec_putback, pgrescued);
362 * Fill up pagevec for __munlock_pagevec using pte walk
364 * The function expects that the struct page corresponding to @start address is
365 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
367 * The rest of @pvec is filled by subsequent pages within the same pmd and same
368 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
369 * pages also get pinned.
371 * Returns the address of the next page that should be scanned. This equals
372 * @start + PAGE_SIZE when no page could be added by the pte walk.
374 static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
375 struct vm_area_struct *vma, struct zone *zone,
376 unsigned long start, unsigned long end)
382 * Initialize pte walk starting at the already pinned page where we
383 * are sure that there is a pte, as it was pinned under the same
386 pte = get_locked_pte(vma->vm_mm, start, &ptl);
387 /* Make sure we do not cross the page table boundary */
388 end = pgd_addr_end(start, end);
389 end = p4d_addr_end(start, end);
390 end = pud_addr_end(start, end);
391 end = pmd_addr_end(start, end);
393 /* The page next to the pinned page is the first we will try to get */
395 while (start < end) {
396 struct page *page = NULL;
398 if (pte_present(*pte))
399 page = vm_normal_page(vma, start, *pte);
401 * Break if page could not be obtained or the page's node+zone does not
404 if (!page || page_zone(page) != zone)
408 * Do not use pagevec for PTE-mapped THP,
409 * munlock_vma_pages_range() will handle them.
411 if (PageTransCompound(page))
416 * Increase the address that will be returned *before* the
417 * eventual break due to pvec becoming full by adding the page
420 if (pagevec_add(pvec, page) == 0)
423 pte_unmap_unlock(pte, ptl);
428 * munlock_vma_pages_range() - munlock all pages in the vma range.'
429 * @vma - vma containing range to be munlock()ed.
430 * @start - start address in @vma of the range
431 * @end - end of range in @vma.
433 * For mremap(), munmap() and exit().
435 * Called with @vma VM_LOCKED.
437 * Returns with VM_LOCKED cleared. Callers must be prepared to
440 * We don't save and restore VM_LOCKED here because pages are
441 * still on lru. In unmap path, pages might be scanned by reclaim
442 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
443 * free them. This will result in freeing mlocked pages.
445 void munlock_vma_pages_range(struct vm_area_struct *vma,
446 unsigned long start, unsigned long end)
448 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
450 while (start < end) {
452 unsigned int page_mask = 0;
453 unsigned long page_increm;
459 * Although FOLL_DUMP is intended for get_dump_page(),
460 * it just so happens that its special treatment of the
461 * ZERO_PAGE (returning an error instead of doing get_page)
462 * suits munlock very well (and if somehow an abnormal page
463 * has sneaked into the range, we won't oops here: great).
465 page = follow_page(vma, start, FOLL_GET | FOLL_DUMP);
467 if (page && !IS_ERR(page)) {
468 if (PageTransTail(page)) {
469 VM_BUG_ON_PAGE(PageMlocked(page), page);
470 put_page(page); /* follow_page_mask() */
471 } else if (PageTransHuge(page)) {
474 * Any THP page found by follow_page_mask() may
475 * have gotten split before reaching
476 * munlock_vma_page(), so we need to compute
477 * the page_mask here instead.
479 page_mask = munlock_vma_page(page);
481 put_page(page); /* follow_page_mask() */
484 * Non-huge pages are handled in batches via
485 * pagevec. The pin from follow_page_mask()
486 * prevents them from collapsing by THP.
488 pagevec_add(&pvec, page);
489 zone = page_zone(page);
492 * Try to fill the rest of pagevec using fast
493 * pte walk. This will also update start to
494 * the next page to process. Then munlock the
497 start = __munlock_pagevec_fill(&pvec, vma,
499 __munlock_pagevec(&pvec, zone);
503 page_increm = 1 + page_mask;
504 start += page_increm * PAGE_SIZE;
511 * mlock_fixup - handle mlock[all]/munlock[all] requests.
513 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
514 * munlock is a no-op. However, for some special vmas, we go ahead and
517 * For vmas that pass the filters, merge/split as appropriate.
519 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
520 unsigned long start, unsigned long end, vm_flags_t newflags)
522 struct mm_struct *mm = vma->vm_mm;
526 int lock = !!(newflags & VM_LOCKED);
527 vm_flags_t old_flags = vma->vm_flags;
529 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
530 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm) ||
532 /* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
535 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
536 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
537 vma->vm_file, pgoff, vma_policy(vma),
538 vma->vm_userfaultfd_ctx);
544 if (start != vma->vm_start) {
545 ret = split_vma(mm, vma, start, 1);
550 if (end != vma->vm_end) {
551 ret = split_vma(mm, vma, end, 0);
558 * Keep track of amount of locked VM.
560 nr_pages = (end - start) >> PAGE_SHIFT;
562 nr_pages = -nr_pages;
563 else if (old_flags & VM_LOCKED)
565 mm->locked_vm += nr_pages;
568 * vm_flags is protected by the mmap_sem held in write mode.
569 * It's okay if try_to_unmap_one unmaps a page just after we
570 * set VM_LOCKED, populate_vma_page_range will bring it back.
574 vma->vm_flags = newflags;
576 munlock_vma_pages_range(vma, start, end);
583 static int apply_vma_lock_flags(unsigned long start, size_t len,
586 unsigned long nstart, end, tmp;
587 struct vm_area_struct * vma, * prev;
590 VM_BUG_ON(offset_in_page(start));
591 VM_BUG_ON(len != PAGE_ALIGN(len));
597 vma = find_vma(current->mm, start);
598 if (!vma || vma->vm_start > start)
602 if (start > vma->vm_start)
605 for (nstart = start ; ; ) {
606 vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
610 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
614 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
618 if (nstart < prev->vm_end)
619 nstart = prev->vm_end;
624 if (!vma || vma->vm_start != nstart) {
633 * Go through vma areas and sum size of mlocked
634 * vma pages, as return value.
635 * Note deferred memory locking case(mlock2(,,MLOCK_ONFAULT)
637 * Return value: previously mlocked page counts
639 static unsigned long count_mm_mlocked_page_nr(struct mm_struct *mm,
640 unsigned long start, size_t len)
642 struct vm_area_struct *vma;
643 unsigned long count = 0;
648 vma = find_vma(mm, start);
652 for (; vma ; vma = vma->vm_next) {
653 if (start >= vma->vm_end)
655 if (start + len <= vma->vm_start)
657 if (vma->vm_flags & VM_LOCKED) {
658 if (start > vma->vm_start)
659 count -= (start - vma->vm_start);
660 if (start + len < vma->vm_end) {
661 count += start + len - vma->vm_start;
664 count += vma->vm_end - vma->vm_start;
668 return count >> PAGE_SHIFT;
671 static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
673 unsigned long locked;
674 unsigned long lock_limit;
677 start = untagged_addr(start);
682 len = PAGE_ALIGN(len + (offset_in_page(start)));
685 lock_limit = rlimit(RLIMIT_MEMLOCK);
686 lock_limit >>= PAGE_SHIFT;
687 locked = len >> PAGE_SHIFT;
689 if (down_write_killable(¤t->mm->mmap_sem))
692 locked += current->mm->locked_vm;
693 if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
695 * It is possible that the regions requested intersect with
696 * previously mlocked areas, that part area in "mm->locked_vm"
697 * should not be counted to new mlock increment count. So check
698 * and adjust locked count if necessary.
700 locked -= count_mm_mlocked_page_nr(current->mm,
704 /* check against resource limits */
705 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
706 error = apply_vma_lock_flags(start, len, flags);
708 up_write(¤t->mm->mmap_sem);
712 error = __mm_populate(start, len, 0);
714 return __mlock_posix_error_return(error);
718 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
720 return do_mlock(start, len, VM_LOCKED);
723 SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
725 vm_flags_t vm_flags = VM_LOCKED;
727 if (flags & ~MLOCK_ONFAULT)
730 if (flags & MLOCK_ONFAULT)
731 vm_flags |= VM_LOCKONFAULT;
733 return do_mlock(start, len, vm_flags);
736 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
740 start = untagged_addr(start);
742 len = PAGE_ALIGN(len + (offset_in_page(start)));
745 if (down_write_killable(¤t->mm->mmap_sem))
747 ret = apply_vma_lock_flags(start, len, 0);
748 up_write(¤t->mm->mmap_sem);
754 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
755 * and translate into the appropriate modifications to mm->def_flags and/or the
756 * flags for all current VMAs.
758 * There are a couple of subtleties with this. If mlockall() is called multiple
759 * times with different flags, the values do not necessarily stack. If mlockall
760 * is called once including the MCL_FUTURE flag and then a second time without
761 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
763 static int apply_mlockall_flags(int flags)
765 struct vm_area_struct * vma, * prev = NULL;
766 vm_flags_t to_add = 0;
768 current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
769 if (flags & MCL_FUTURE) {
770 current->mm->def_flags |= VM_LOCKED;
772 if (flags & MCL_ONFAULT)
773 current->mm->def_flags |= VM_LOCKONFAULT;
775 if (!(flags & MCL_CURRENT))
779 if (flags & MCL_CURRENT) {
781 if (flags & MCL_ONFAULT)
782 to_add |= VM_LOCKONFAULT;
785 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
788 newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
792 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
799 SYSCALL_DEFINE1(mlockall, int, flags)
801 unsigned long lock_limit;
804 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)) ||
805 flags == MCL_ONFAULT)
811 lock_limit = rlimit(RLIMIT_MEMLOCK);
812 lock_limit >>= PAGE_SHIFT;
814 if (down_write_killable(¤t->mm->mmap_sem))
818 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
819 capable(CAP_IPC_LOCK))
820 ret = apply_mlockall_flags(flags);
821 up_write(¤t->mm->mmap_sem);
822 if (!ret && (flags & MCL_CURRENT))
823 mm_populate(0, TASK_SIZE);
828 SYSCALL_DEFINE0(munlockall)
832 if (down_write_killable(¤t->mm->mmap_sem))
834 ret = apply_mlockall_flags(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);