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>
26 #include <linux/secretmem.h>
30 bool can_do_mlock(void)
32 if (rlimit(RLIMIT_MEMLOCK) != 0)
34 if (capable(CAP_IPC_LOCK))
38 EXPORT_SYMBOL(can_do_mlock);
41 * Mlocked pages are marked with PageMlocked() flag for efficient testing
42 * in vmscan and, possibly, the fault path; and to support semi-accurate
45 * An mlocked page [PageMlocked(page)] is unevictable. As such, it will
46 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
47 * The unevictable list is an LRU sibling list to the [in]active lists.
48 * PageUnevictable is set to indicate the unevictable state.
50 * When lazy mlocking via vmscan, it is important to ensure that the
51 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
52 * may have mlocked a page that is being munlocked. So lazy mlock must take
53 * the mmap_lock for read, and verify that the vma really is locked
58 * LRU accounting for clear_page_mlock()
60 void clear_page_mlock(struct page *page)
64 if (!TestClearPageMlocked(page))
67 nr_pages = thp_nr_pages(page);
68 mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
69 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
71 * The previous TestClearPageMlocked() corresponds to the smp_mb()
72 * in __pagevec_lru_add_fn().
74 * See __pagevec_lru_add_fn for more explanation.
76 if (!isolate_lru_page(page)) {
77 putback_lru_page(page);
80 * We lost the race. the page already moved to evictable list.
82 if (PageUnevictable(page))
83 count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
88 * Mark page as mlocked if not already.
89 * If page on LRU, isolate and putback to move to unevictable list.
91 void mlock_vma_page(struct page *page)
93 /* Serialize with page migration */
94 BUG_ON(!PageLocked(page));
96 VM_BUG_ON_PAGE(PageTail(page), page);
97 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
99 if (!TestSetPageMlocked(page)) {
100 int nr_pages = thp_nr_pages(page);
102 mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
103 count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
104 if (!isolate_lru_page(page))
105 putback_lru_page(page);
110 * Finish munlock after successful page isolation
112 * Page must be locked. This is a wrapper for page_mlock()
113 * and putback_lru_page() with munlock accounting.
115 static void __munlock_isolated_page(struct page *page)
118 * Optimization: if the page was mapped just once, that's our mapping
119 * and we don't need to check all the other vmas.
121 if (page_mapcount(page) > 1)
124 /* Did try_to_unlock() succeed or punt? */
125 if (!PageMlocked(page))
126 count_vm_events(UNEVICTABLE_PGMUNLOCKED, thp_nr_pages(page));
128 putback_lru_page(page);
132 * Accounting for page isolation fail during munlock
134 * Performs accounting when page isolation fails in munlock. There is nothing
135 * else to do because it means some other task has already removed the page
136 * from the LRU. putback_lru_page() will take care of removing the page from
137 * the unevictable list, if necessary. vmscan [page_referenced()] will move
138 * the page back to the unevictable list if some other vma has it mlocked.
140 static void __munlock_isolation_failed(struct page *page)
142 int nr_pages = thp_nr_pages(page);
144 if (PageUnevictable(page))
145 __count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
147 __count_vm_events(UNEVICTABLE_PGMUNLOCKED, nr_pages);
151 * munlock_vma_page - munlock a vma page
152 * @page: page to be unlocked, either a normal page or THP page head
154 * returns the size of the page as a page mask (0 for normal page,
155 * HPAGE_PMD_NR - 1 for THP head page)
157 * called from munlock()/munmap() path with page supposedly on the LRU.
158 * When we munlock a page, because the vma where we found the page is being
159 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
160 * page locked so that we can leave it on the unevictable lru list and not
161 * bother vmscan with it. However, to walk the page's rmap list in
162 * page_mlock() we must isolate the page from the LRU. If some other
163 * task has removed the page from the LRU, we won't be able to do that.
164 * So we clear the PageMlocked as we might not get another chance. If we
165 * can't isolate the page, we leave it for putback_lru_page() and vmscan
166 * [page_referenced()/try_to_unmap()] to deal with.
168 unsigned int munlock_vma_page(struct page *page)
172 /* For page_mlock() and to serialize with page migration */
173 BUG_ON(!PageLocked(page));
174 VM_BUG_ON_PAGE(PageTail(page), page);
176 if (!TestClearPageMlocked(page)) {
177 /* Potentially, PTE-mapped THP: do not skip the rest PTEs */
181 nr_pages = thp_nr_pages(page);
182 mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
184 if (!isolate_lru_page(page))
185 __munlock_isolated_page(page);
187 __munlock_isolation_failed(page);
193 * convert get_user_pages() return value to posix mlock() error
195 static int __mlock_posix_error_return(long retval)
197 if (retval == -EFAULT)
199 else if (retval == -ENOMEM)
205 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
207 * The fast path is available only for evictable pages with single mapping.
208 * Then we can bypass the per-cpu pvec and get better performance.
209 * when mapcount > 1 we need page_mlock() which can fail.
210 * when !page_evictable(), we need the full redo logic of putback_lru_page to
211 * avoid leaving evictable page in unevictable list.
213 * In case of success, @page is added to @pvec and @pgrescued is incremented
214 * in case that the page was previously unevictable. @page is also unlocked.
216 static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
219 VM_BUG_ON_PAGE(PageLRU(page), page);
220 VM_BUG_ON_PAGE(!PageLocked(page), page);
222 if (page_mapcount(page) <= 1 && page_evictable(page)) {
223 pagevec_add(pvec, page);
224 if (TestClearPageUnevictable(page))
234 * Putback multiple evictable pages to the LRU
236 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
237 * the pages might have meanwhile become unevictable but that is OK.
239 static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
241 count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
243 *__pagevec_lru_add() calls release_pages() so we don't call
244 * put_page() explicitly
246 __pagevec_lru_add(pvec);
247 count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
251 * Munlock a batch of pages from the same zone
253 * The work is split to two main phases. First phase clears the Mlocked flag
254 * and attempts to isolate the pages, all under a single zone lru lock.
255 * The second phase finishes the munlock only for pages where isolation
258 * Note that the pagevec may be modified during the process.
260 static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
263 int nr = pagevec_count(pvec);
264 int delta_munlocked = -nr;
265 struct pagevec pvec_putback;
266 struct lruvec *lruvec = NULL;
269 pagevec_init(&pvec_putback);
271 /* Phase 1: page isolation */
272 for (i = 0; i < nr; i++) {
273 struct page *page = pvec->pages[i];
275 if (TestClearPageMlocked(page)) {
277 * We already have pin from follow_page_mask()
278 * so we can spare the get_page() here.
280 if (TestClearPageLRU(page)) {
281 lruvec = relock_page_lruvec_irq(page, lruvec);
282 del_page_from_lru_list(page, lruvec);
285 __munlock_isolation_failed(page);
291 * We won't be munlocking this page in the next phase
292 * but we still need to release the follow_page_mask()
293 * pin. We cannot do it under lru_lock however. If it's
294 * the last pin, __page_cache_release() would deadlock.
296 pagevec_add(&pvec_putback, pvec->pages[i]);
297 pvec->pages[i] = NULL;
300 __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
301 unlock_page_lruvec_irq(lruvec);
302 } else if (delta_munlocked) {
303 mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
306 /* Now we can release pins of pages that we are not munlocking */
307 pagevec_release(&pvec_putback);
309 /* Phase 2: page munlock */
310 for (i = 0; i < nr; i++) {
311 struct page *page = pvec->pages[i];
315 if (!__putback_lru_fast_prepare(page, &pvec_putback,
318 * Slow path. We don't want to lose the last
319 * pin before unlock_page()
321 get_page(page); /* for putback_lru_page() */
322 __munlock_isolated_page(page);
324 put_page(page); /* from follow_page_mask() */
330 * Phase 3: page putback for pages that qualified for the fast path
331 * This will also call put_page() to return pin from follow_page_mask()
333 if (pagevec_count(&pvec_putback))
334 __putback_lru_fast(&pvec_putback, pgrescued);
338 * Fill up pagevec for __munlock_pagevec using pte walk
340 * The function expects that the struct page corresponding to @start address is
341 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
343 * The rest of @pvec is filled by subsequent pages within the same pmd and same
344 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
345 * pages also get pinned.
347 * Returns the address of the next page that should be scanned. This equals
348 * @start + PAGE_SIZE when no page could be added by the pte walk.
350 static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
351 struct vm_area_struct *vma, struct zone *zone,
352 unsigned long start, unsigned long end)
358 * Initialize pte walk starting at the already pinned page where we
359 * are sure that there is a pte, as it was pinned under the same
360 * mmap_lock write op.
362 pte = get_locked_pte(vma->vm_mm, start, &ptl);
363 /* Make sure we do not cross the page table boundary */
364 end = pgd_addr_end(start, end);
365 end = p4d_addr_end(start, end);
366 end = pud_addr_end(start, end);
367 end = pmd_addr_end(start, end);
369 /* The page next to the pinned page is the first we will try to get */
371 while (start < end) {
372 struct page *page = NULL;
374 if (pte_present(*pte))
375 page = vm_normal_page(vma, start, *pte);
377 * Break if page could not be obtained or the page's node+zone does not
380 if (!page || page_zone(page) != zone)
384 * Do not use pagevec for PTE-mapped THP,
385 * munlock_vma_pages_range() will handle them.
387 if (PageTransCompound(page))
392 * Increase the address that will be returned *before* the
393 * eventual break due to pvec becoming full by adding the page
396 if (pagevec_add(pvec, page) == 0)
399 pte_unmap_unlock(pte, ptl);
404 * munlock_vma_pages_range() - munlock all pages in the vma range.'
405 * @vma - vma containing range to be munlock()ed.
406 * @start - start address in @vma of the range
407 * @end - end of range in @vma.
409 * For mremap(), munmap() and exit().
411 * Called with @vma VM_LOCKED.
413 * Returns with VM_LOCKED cleared. Callers must be prepared to
416 * We don't save and restore VM_LOCKED here because pages are
417 * still on lru. In unmap path, pages might be scanned by reclaim
418 * and re-mlocked by page_mlock/try_to_unmap before we unmap and
419 * free them. This will result in freeing mlocked pages.
421 void munlock_vma_pages_range(struct vm_area_struct *vma,
422 unsigned long start, unsigned long end)
424 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
426 while (start < end) {
428 unsigned int page_mask = 0;
429 unsigned long page_increm;
435 * Although FOLL_DUMP is intended for get_dump_page(),
436 * it just so happens that its special treatment of the
437 * ZERO_PAGE (returning an error instead of doing get_page)
438 * suits munlock very well (and if somehow an abnormal page
439 * has sneaked into the range, we won't oops here: great).
441 page = follow_page(vma, start, FOLL_GET | FOLL_DUMP);
443 if (page && !IS_ERR(page)) {
444 if (PageTransTail(page)) {
445 VM_BUG_ON_PAGE(PageMlocked(page), page);
446 put_page(page); /* follow_page_mask() */
447 } else if (PageTransHuge(page)) {
450 * Any THP page found by follow_page_mask() may
451 * have gotten split before reaching
452 * munlock_vma_page(), so we need to compute
453 * the page_mask here instead.
455 page_mask = munlock_vma_page(page);
457 put_page(page); /* follow_page_mask() */
460 * Non-huge pages are handled in batches via
461 * pagevec. The pin from follow_page_mask()
462 * prevents them from collapsing by THP.
464 pagevec_add(&pvec, page);
465 zone = page_zone(page);
468 * Try to fill the rest of pagevec using fast
469 * pte walk. This will also update start to
470 * the next page to process. Then munlock the
473 start = __munlock_pagevec_fill(&pvec, vma,
475 __munlock_pagevec(&pvec, zone);
479 page_increm = 1 + page_mask;
480 start += page_increm * PAGE_SIZE;
487 * mlock_fixup - handle mlock[all]/munlock[all] requests.
489 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
490 * munlock is a no-op. However, for some special vmas, we go ahead and
493 * For vmas that pass the filters, merge/split as appropriate.
495 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
496 unsigned long start, unsigned long end, vm_flags_t newflags)
498 struct mm_struct *mm = vma->vm_mm;
502 int lock = !!(newflags & VM_LOCKED);
503 vm_flags_t old_flags = vma->vm_flags;
505 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
506 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm) ||
507 vma_is_dax(vma) || vma_is_secretmem(vma))
508 /* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
511 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
512 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
513 vma->vm_file, pgoff, vma_policy(vma),
514 vma->vm_userfaultfd_ctx);
520 if (start != vma->vm_start) {
521 ret = split_vma(mm, vma, start, 1);
526 if (end != vma->vm_end) {
527 ret = split_vma(mm, vma, end, 0);
534 * Keep track of amount of locked VM.
536 nr_pages = (end - start) >> PAGE_SHIFT;
538 nr_pages = -nr_pages;
539 else if (old_flags & VM_LOCKED)
541 mm->locked_vm += nr_pages;
544 * vm_flags is protected by the mmap_lock held in write mode.
545 * It's okay if try_to_unmap_one unmaps a page just after we
546 * set VM_LOCKED, populate_vma_page_range will bring it back.
550 vma->vm_flags = newflags;
552 munlock_vma_pages_range(vma, start, end);
559 static int apply_vma_lock_flags(unsigned long start, size_t len,
562 unsigned long nstart, end, tmp;
563 struct vm_area_struct *vma, *prev;
566 VM_BUG_ON(offset_in_page(start));
567 VM_BUG_ON(len != PAGE_ALIGN(len));
573 vma = find_vma(current->mm, start);
574 if (!vma || vma->vm_start > start)
578 if (start > vma->vm_start)
581 for (nstart = start ; ; ) {
582 vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
586 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
590 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
594 if (nstart < prev->vm_end)
595 nstart = prev->vm_end;
600 if (!vma || vma->vm_start != nstart) {
609 * Go through vma areas and sum size of mlocked
610 * vma pages, as return value.
611 * Note deferred memory locking case(mlock2(,,MLOCK_ONFAULT)
613 * Return value: previously mlocked page counts
615 static unsigned long count_mm_mlocked_page_nr(struct mm_struct *mm,
616 unsigned long start, size_t len)
618 struct vm_area_struct *vma;
619 unsigned long count = 0;
624 vma = find_vma(mm, start);
628 for (; vma ; vma = vma->vm_next) {
629 if (start >= vma->vm_end)
631 if (start + len <= vma->vm_start)
633 if (vma->vm_flags & VM_LOCKED) {
634 if (start > vma->vm_start)
635 count -= (start - vma->vm_start);
636 if (start + len < vma->vm_end) {
637 count += start + len - vma->vm_start;
640 count += vma->vm_end - vma->vm_start;
644 return count >> PAGE_SHIFT;
647 static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
649 unsigned long locked;
650 unsigned long lock_limit;
653 start = untagged_addr(start);
658 len = PAGE_ALIGN(len + (offset_in_page(start)));
661 lock_limit = rlimit(RLIMIT_MEMLOCK);
662 lock_limit >>= PAGE_SHIFT;
663 locked = len >> PAGE_SHIFT;
665 if (mmap_write_lock_killable(current->mm))
668 locked += current->mm->locked_vm;
669 if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
671 * It is possible that the regions requested intersect with
672 * previously mlocked areas, that part area in "mm->locked_vm"
673 * should not be counted to new mlock increment count. So check
674 * and adjust locked count if necessary.
676 locked -= count_mm_mlocked_page_nr(current->mm,
680 /* check against resource limits */
681 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
682 error = apply_vma_lock_flags(start, len, flags);
684 mmap_write_unlock(current->mm);
688 error = __mm_populate(start, len, 0);
690 return __mlock_posix_error_return(error);
694 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
696 return do_mlock(start, len, VM_LOCKED);
699 SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
701 vm_flags_t vm_flags = VM_LOCKED;
703 if (flags & ~MLOCK_ONFAULT)
706 if (flags & MLOCK_ONFAULT)
707 vm_flags |= VM_LOCKONFAULT;
709 return do_mlock(start, len, vm_flags);
712 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
716 start = untagged_addr(start);
718 len = PAGE_ALIGN(len + (offset_in_page(start)));
721 if (mmap_write_lock_killable(current->mm))
723 ret = apply_vma_lock_flags(start, len, 0);
724 mmap_write_unlock(current->mm);
730 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
731 * and translate into the appropriate modifications to mm->def_flags and/or the
732 * flags for all current VMAs.
734 * There are a couple of subtleties with this. If mlockall() is called multiple
735 * times with different flags, the values do not necessarily stack. If mlockall
736 * is called once including the MCL_FUTURE flag and then a second time without
737 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
739 static int apply_mlockall_flags(int flags)
741 struct vm_area_struct *vma, *prev = NULL;
742 vm_flags_t to_add = 0;
744 current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
745 if (flags & MCL_FUTURE) {
746 current->mm->def_flags |= VM_LOCKED;
748 if (flags & MCL_ONFAULT)
749 current->mm->def_flags |= VM_LOCKONFAULT;
751 if (!(flags & MCL_CURRENT))
755 if (flags & MCL_CURRENT) {
757 if (flags & MCL_ONFAULT)
758 to_add |= VM_LOCKONFAULT;
761 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
764 newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
768 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
775 SYSCALL_DEFINE1(mlockall, int, flags)
777 unsigned long lock_limit;
780 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)) ||
781 flags == MCL_ONFAULT)
787 lock_limit = rlimit(RLIMIT_MEMLOCK);
788 lock_limit >>= PAGE_SHIFT;
790 if (mmap_write_lock_killable(current->mm))
794 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
795 capable(CAP_IPC_LOCK))
796 ret = apply_mlockall_flags(flags);
797 mmap_write_unlock(current->mm);
798 if (!ret && (flags & MCL_CURRENT))
799 mm_populate(0, TASK_SIZE);
804 SYSCALL_DEFINE0(munlockall)
808 if (mmap_write_lock_killable(current->mm))
810 ret = apply_mlockall_flags(0);
811 mmap_write_unlock(current->mm);
816 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
817 * shm segments) get accounted against the user_struct instead.
819 static DEFINE_SPINLOCK(shmlock_user_lock);
821 int user_shm_lock(size_t size, struct ucounts *ucounts)
823 unsigned long lock_limit, locked;
827 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
828 lock_limit = rlimit(RLIMIT_MEMLOCK);
829 if (lock_limit != RLIM_INFINITY)
830 lock_limit >>= PAGE_SHIFT;
831 spin_lock(&shmlock_user_lock);
832 memlock = inc_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
834 if ((memlock == LONG_MAX || memlock > lock_limit) && !capable(CAP_IPC_LOCK)) {
835 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
838 if (!get_ucounts(ucounts)) {
839 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
845 spin_unlock(&shmlock_user_lock);
849 void user_shm_unlock(size_t size, struct ucounts *ucounts)
851 spin_lock(&shmlock_user_lock);
852 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
853 spin_unlock(&shmlock_user_lock);
854 put_ucounts(ucounts);