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/sched/user.h>
12 #include <linux/swap.h>
13 #include <linux/swapops.h>
14 #include <linux/pagemap.h>
15 #include <linux/pagevec.h>
16 #include <linux/mempolicy.h>
17 #include <linux/syscalls.h>
18 #include <linux/sched.h>
19 #include <linux/export.h>
20 #include <linux/rmap.h>
21 #include <linux/mmzone.h>
22 #include <linux/hugetlb.h>
23 #include <linux/memcontrol.h>
24 #include <linux/mm_inline.h>
28 bool can_do_mlock(void)
30 if (rlimit(RLIMIT_MEMLOCK) != 0)
32 if (capable(CAP_IPC_LOCK))
36 EXPORT_SYMBOL(can_do_mlock);
39 * Mlocked pages are marked with PageMlocked() flag for efficient testing
40 * in vmscan and, possibly, the fault path; and to support semi-accurate
43 * An mlocked page [PageMlocked(page)] is unevictable. As such, it will
44 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
45 * The unevictable list is an LRU sibling list to the [in]active lists.
46 * PageUnevictable is set to indicate the unevictable state.
48 * When lazy mlocking via vmscan, it is important to ensure that the
49 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
50 * may have mlocked a page that is being munlocked. So lazy mlock must take
51 * the mmap_sem for read, and verify that the vma really is locked
56 * LRU accounting for clear_page_mlock()
58 void clear_page_mlock(struct page *page)
60 if (!TestClearPageMlocked(page))
63 mod_zone_page_state(page_zone(page), NR_MLOCK,
64 -hpage_nr_pages(page));
65 count_vm_event(UNEVICTABLE_PGCLEARED);
66 if (!isolate_lru_page(page)) {
67 putback_lru_page(page);
70 * We lost the race. the page already moved to evictable list.
72 if (PageUnevictable(page))
73 count_vm_event(UNEVICTABLE_PGSTRANDED);
78 * Mark page as mlocked if not already.
79 * If page on LRU, isolate and putback to move to unevictable list.
81 void mlock_vma_page(struct page *page)
83 /* Serialize with page migration */
84 BUG_ON(!PageLocked(page));
86 VM_BUG_ON_PAGE(PageTail(page), page);
87 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
89 if (!TestSetPageMlocked(page)) {
90 mod_zone_page_state(page_zone(page), NR_MLOCK,
91 hpage_nr_pages(page));
92 count_vm_event(UNEVICTABLE_PGMLOCKED);
93 if (!isolate_lru_page(page))
94 putback_lru_page(page);
99 * Isolate a page from LRU with optional get_page() pin.
100 * Assumes lru_lock already held and page already pinned.
102 static bool __munlock_isolate_lru_page(struct page *page, bool getpage)
105 struct lruvec *lruvec;
107 lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
111 del_page_from_lru_list(page, lruvec, page_lru(page));
119 * Finish munlock after successful page isolation
121 * Page must be locked. This is a wrapper for try_to_munlock()
122 * and putback_lru_page() with munlock accounting.
124 static void __munlock_isolated_page(struct page *page)
127 * Optimization: if the page was mapped just once, that's our mapping
128 * and we don't need to check all the other vmas.
130 if (page_mapcount(page) > 1)
131 try_to_munlock(page);
133 /* Did try_to_unlock() succeed or punt? */
134 if (!PageMlocked(page))
135 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
137 putback_lru_page(page);
141 * Accounting for page isolation fail during munlock
143 * Performs accounting when page isolation fails in munlock. There is nothing
144 * else to do because it means some other task has already removed the page
145 * from the LRU. putback_lru_page() will take care of removing the page from
146 * the unevictable list, if necessary. vmscan [page_referenced()] will move
147 * the page back to the unevictable list if some other vma has it mlocked.
149 static void __munlock_isolation_failed(struct page *page)
151 if (PageUnevictable(page))
152 __count_vm_event(UNEVICTABLE_PGSTRANDED);
154 __count_vm_event(UNEVICTABLE_PGMUNLOCKED);
158 * munlock_vma_page - munlock a vma page
159 * @page - page to be unlocked, either a normal page or THP page head
161 * returns the size of the page as a page mask (0 for normal page,
162 * HPAGE_PMD_NR - 1 for THP head page)
164 * called from munlock()/munmap() path with page supposedly on the LRU.
165 * When we munlock a page, because the vma where we found the page is being
166 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
167 * page locked so that we can leave it on the unevictable lru list and not
168 * bother vmscan with it. However, to walk the page's rmap list in
169 * try_to_munlock() we must isolate the page from the LRU. If some other
170 * task has removed the page from the LRU, we won't be able to do that.
171 * So we clear the PageMlocked as we might not get another chance. If we
172 * can't isolate the page, we leave it for putback_lru_page() and vmscan
173 * [page_referenced()/try_to_unmap()] to deal with.
175 unsigned int munlock_vma_page(struct page *page)
178 struct zone *zone = page_zone(page);
180 /* For try_to_munlock() and to serialize with page migration */
181 BUG_ON(!PageLocked(page));
183 VM_BUG_ON_PAGE(PageTail(page), page);
186 * Serialize with any parallel __split_huge_page_refcount() which
187 * might otherwise copy PageMlocked to part of the tail pages before
188 * we clear it in the head page. It also stabilizes hpage_nr_pages().
190 spin_lock_irq(zone_lru_lock(zone));
192 if (!TestClearPageMlocked(page)) {
193 /* Potentially, PTE-mapped THP: do not skip the rest PTEs */
198 nr_pages = hpage_nr_pages(page);
199 __mod_zone_page_state(zone, NR_MLOCK, -nr_pages);
201 if (__munlock_isolate_lru_page(page, true)) {
202 spin_unlock_irq(zone_lru_lock(zone));
203 __munlock_isolated_page(page);
206 __munlock_isolation_failed(page);
209 spin_unlock_irq(zone_lru_lock(zone));
216 * convert get_user_pages() return value to posix mlock() error
218 static int __mlock_posix_error_return(long retval)
220 if (retval == -EFAULT)
222 else if (retval == -ENOMEM)
228 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
230 * The fast path is available only for evictable pages with single mapping.
231 * Then we can bypass the per-cpu pvec and get better performance.
232 * when mapcount > 1 we need try_to_munlock() which can fail.
233 * when !page_evictable(), we need the full redo logic of putback_lru_page to
234 * avoid leaving evictable page in unevictable list.
236 * In case of success, @page is added to @pvec and @pgrescued is incremented
237 * in case that the page was previously unevictable. @page is also unlocked.
239 static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
242 VM_BUG_ON_PAGE(PageLRU(page), page);
243 VM_BUG_ON_PAGE(!PageLocked(page), page);
245 if (page_mapcount(page) <= 1 && page_evictable(page)) {
246 pagevec_add(pvec, page);
247 if (TestClearPageUnevictable(page))
257 * Putback multiple evictable pages to the LRU
259 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
260 * the pages might have meanwhile become unevictable but that is OK.
262 static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
264 count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
266 *__pagevec_lru_add() calls release_pages() so we don't call
267 * put_page() explicitly
269 __pagevec_lru_add(pvec);
270 count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
274 * Munlock a batch of pages from the same zone
276 * The work is split to two main phases. First phase clears the Mlocked flag
277 * and attempts to isolate the pages, all under a single zone lru lock.
278 * The second phase finishes the munlock only for pages where isolation
281 * Note that the pagevec may be modified during the process.
283 static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
286 int nr = pagevec_count(pvec);
288 struct pagevec pvec_putback;
291 pagevec_init(&pvec_putback, 0);
293 /* Phase 1: page isolation */
294 spin_lock_irq(zone_lru_lock(zone));
295 for (i = 0; i < nr; i++) {
296 struct page *page = pvec->pages[i];
298 if (TestClearPageMlocked(page)) {
300 * We already have pin from follow_page_mask()
301 * so we can spare the get_page() here.
303 if (__munlock_isolate_lru_page(page, false))
306 __munlock_isolation_failed(page);
310 * We won't be munlocking this page in the next phase
311 * but we still need to release the follow_page_mask()
312 * pin. We cannot do it under lru_lock however. If it's
313 * the last pin, __page_cache_release() would deadlock.
315 pagevec_add(&pvec_putback, pvec->pages[i]);
316 pvec->pages[i] = NULL;
318 delta_munlocked = -nr + pagevec_count(&pvec_putback);
319 __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
320 spin_unlock_irq(zone_lru_lock(zone));
322 /* Now we can release pins of pages that we are not munlocking */
323 pagevec_release(&pvec_putback);
325 /* Phase 2: page munlock */
326 for (i = 0; i < nr; i++) {
327 struct page *page = pvec->pages[i];
331 if (!__putback_lru_fast_prepare(page, &pvec_putback,
334 * Slow path. We don't want to lose the last
335 * pin before unlock_page()
337 get_page(page); /* for putback_lru_page() */
338 __munlock_isolated_page(page);
340 put_page(page); /* from follow_page_mask() */
346 * Phase 3: page putback for pages that qualified for the fast path
347 * This will also call put_page() to return pin from follow_page_mask()
349 if (pagevec_count(&pvec_putback))
350 __putback_lru_fast(&pvec_putback, pgrescued);
354 * Fill up pagevec for __munlock_pagevec using pte walk
356 * The function expects that the struct page corresponding to @start address is
357 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
359 * The rest of @pvec is filled by subsequent pages within the same pmd and same
360 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
361 * pages also get pinned.
363 * Returns the address of the next page that should be scanned. This equals
364 * @start + PAGE_SIZE when no page could be added by the pte walk.
366 static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
367 struct vm_area_struct *vma, int zoneid, unsigned long start,
374 * Initialize pte walk starting at the already pinned page where we
375 * are sure that there is a pte, as it was pinned under the same
378 pte = get_locked_pte(vma->vm_mm, start, &ptl);
379 /* Make sure we do not cross the page table boundary */
380 end = pgd_addr_end(start, end);
381 end = p4d_addr_end(start, end);
382 end = pud_addr_end(start, end);
383 end = pmd_addr_end(start, end);
385 /* The page next to the pinned page is the first we will try to get */
387 while (start < end) {
388 struct page *page = NULL;
390 if (pte_present(*pte))
391 page = vm_normal_page(vma, start, *pte);
393 * Break if page could not be obtained or the page's node+zone does not
396 if (!page || page_zone_id(page) != zoneid)
400 * Do not use pagevec for PTE-mapped THP,
401 * munlock_vma_pages_range() will handle them.
403 if (PageTransCompound(page))
408 * Increase the address that will be returned *before* the
409 * eventual break due to pvec becoming full by adding the page
412 if (pagevec_add(pvec, page) == 0)
415 pte_unmap_unlock(pte, ptl);
420 * munlock_vma_pages_range() - munlock all pages in the vma range.'
421 * @vma - vma containing range to be munlock()ed.
422 * @start - start address in @vma of the range
423 * @end - end of range in @vma.
425 * For mremap(), munmap() and exit().
427 * Called with @vma VM_LOCKED.
429 * Returns with VM_LOCKED cleared. Callers must be prepared to
432 * We don't save and restore VM_LOCKED here because pages are
433 * still on lru. In unmap path, pages might be scanned by reclaim
434 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
435 * free them. This will result in freeing mlocked pages.
437 void munlock_vma_pages_range(struct vm_area_struct *vma,
438 unsigned long start, unsigned long end)
440 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
442 while (start < end) {
444 unsigned int page_mask = 0;
445 unsigned long page_increm;
450 pagevec_init(&pvec, 0);
452 * Although FOLL_DUMP is intended for get_dump_page(),
453 * it just so happens that its special treatment of the
454 * ZERO_PAGE (returning an error instead of doing get_page)
455 * suits munlock very well (and if somehow an abnormal page
456 * has sneaked into the range, we won't oops here: great).
458 page = follow_page(vma, start, FOLL_GET | FOLL_DUMP);
460 if (page && !IS_ERR(page)) {
461 if (PageTransTail(page)) {
462 VM_BUG_ON_PAGE(PageMlocked(page), page);
463 put_page(page); /* follow_page_mask() */
464 } else if (PageTransHuge(page)) {
467 * Any THP page found by follow_page_mask() may
468 * have gotten split before reaching
469 * munlock_vma_page(), so we need to compute
470 * the page_mask here instead.
472 page_mask = munlock_vma_page(page);
474 put_page(page); /* follow_page_mask() */
477 * Non-huge pages are handled in batches via
478 * pagevec. The pin from follow_page_mask()
479 * prevents them from collapsing by THP.
481 pagevec_add(&pvec, page);
482 zone = page_zone(page);
483 zoneid = page_zone_id(page);
486 * Try to fill the rest of pagevec using fast
487 * pte walk. This will also update start to
488 * the next page to process. Then munlock the
491 start = __munlock_pagevec_fill(&pvec, vma,
493 __munlock_pagevec(&pvec, zone);
497 page_increm = 1 + page_mask;
498 start += page_increm * PAGE_SIZE;
505 * mlock_fixup - handle mlock[all]/munlock[all] requests.
507 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
508 * munlock is a no-op. However, for some special vmas, we go ahead and
511 * For vmas that pass the filters, merge/split as appropriate.
513 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
514 unsigned long start, unsigned long end, vm_flags_t newflags)
516 struct mm_struct *mm = vma->vm_mm;
520 int lock = !!(newflags & VM_LOCKED);
521 vm_flags_t old_flags = vma->vm_flags;
523 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
524 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
525 /* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
528 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
529 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
530 vma->vm_file, pgoff, vma_policy(vma),
531 vma->vm_userfaultfd_ctx);
537 if (start != vma->vm_start) {
538 ret = split_vma(mm, vma, start, 1);
543 if (end != vma->vm_end) {
544 ret = split_vma(mm, vma, end, 0);
551 * Keep track of amount of locked VM.
553 nr_pages = (end - start) >> PAGE_SHIFT;
555 nr_pages = -nr_pages;
556 else if (old_flags & VM_LOCKED)
558 mm->locked_vm += nr_pages;
561 * vm_flags is protected by the mmap_sem held in write mode.
562 * It's okay if try_to_unmap_one unmaps a page just after we
563 * set VM_LOCKED, populate_vma_page_range will bring it back.
567 vma->vm_flags = newflags;
569 munlock_vma_pages_range(vma, start, end);
576 static int apply_vma_lock_flags(unsigned long start, size_t len,
579 unsigned long nstart, end, tmp;
580 struct vm_area_struct * vma, * prev;
583 VM_BUG_ON(offset_in_page(start));
584 VM_BUG_ON(len != PAGE_ALIGN(len));
590 vma = find_vma(current->mm, start);
591 if (!vma || vma->vm_start > start)
595 if (start > vma->vm_start)
598 for (nstart = start ; ; ) {
599 vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
603 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
607 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
611 if (nstart < prev->vm_end)
612 nstart = prev->vm_end;
617 if (!vma || vma->vm_start != nstart) {
626 * Go through vma areas and sum size of mlocked
627 * vma pages, as return value.
628 * Note deferred memory locking case(mlock2(,,MLOCK_ONFAULT)
630 * Return value: previously mlocked page counts
632 static int count_mm_mlocked_page_nr(struct mm_struct *mm,
633 unsigned long start, size_t len)
635 struct vm_area_struct *vma;
641 vma = find_vma(mm, start);
645 for (; vma ; vma = vma->vm_next) {
646 if (start >= vma->vm_end)
648 if (start + len <= vma->vm_start)
650 if (vma->vm_flags & VM_LOCKED) {
651 if (start > vma->vm_start)
652 count -= (start - vma->vm_start);
653 if (start + len < vma->vm_end) {
654 count += start + len - vma->vm_start;
657 count += vma->vm_end - vma->vm_start;
661 return count >> PAGE_SHIFT;
664 static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
666 unsigned long locked;
667 unsigned long lock_limit;
673 lru_add_drain_all(); /* flush pagevec */
675 len = PAGE_ALIGN(len + (offset_in_page(start)));
678 lock_limit = rlimit(RLIMIT_MEMLOCK);
679 lock_limit >>= PAGE_SHIFT;
680 locked = len >> PAGE_SHIFT;
682 if (down_write_killable(¤t->mm->mmap_sem))
685 locked += current->mm->locked_vm;
686 if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
688 * It is possible that the regions requested intersect with
689 * previously mlocked areas, that part area in "mm->locked_vm"
690 * should not be counted to new mlock increment count. So check
691 * and adjust locked count if necessary.
693 locked -= count_mm_mlocked_page_nr(current->mm,
697 /* check against resource limits */
698 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
699 error = apply_vma_lock_flags(start, len, flags);
701 up_write(¤t->mm->mmap_sem);
705 error = __mm_populate(start, len, 0);
707 return __mlock_posix_error_return(error);
711 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
713 return do_mlock(start, len, VM_LOCKED);
716 SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
718 vm_flags_t vm_flags = VM_LOCKED;
720 if (flags & ~MLOCK_ONFAULT)
723 if (flags & MLOCK_ONFAULT)
724 vm_flags |= VM_LOCKONFAULT;
726 return do_mlock(start, len, vm_flags);
729 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
733 len = PAGE_ALIGN(len + (offset_in_page(start)));
736 if (down_write_killable(¤t->mm->mmap_sem))
738 ret = apply_vma_lock_flags(start, len, 0);
739 up_write(¤t->mm->mmap_sem);
745 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
746 * and translate into the appropriate modifications to mm->def_flags and/or the
747 * flags for all current VMAs.
749 * There are a couple of subtleties with this. If mlockall() is called multiple
750 * times with different flags, the values do not necessarily stack. If mlockall
751 * is called once including the MCL_FUTURE flag and then a second time without
752 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
754 static int apply_mlockall_flags(int flags)
756 struct vm_area_struct * vma, * prev = NULL;
757 vm_flags_t to_add = 0;
759 current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
760 if (flags & MCL_FUTURE) {
761 current->mm->def_flags |= VM_LOCKED;
763 if (flags & MCL_ONFAULT)
764 current->mm->def_flags |= VM_LOCKONFAULT;
766 if (!(flags & MCL_CURRENT))
770 if (flags & MCL_CURRENT) {
772 if (flags & MCL_ONFAULT)
773 to_add |= VM_LOCKONFAULT;
776 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
779 newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
783 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
784 cond_resched_rcu_qs();
790 SYSCALL_DEFINE1(mlockall, int, flags)
792 unsigned long lock_limit;
795 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)))
801 if (flags & MCL_CURRENT)
802 lru_add_drain_all(); /* flush pagevec */
804 lock_limit = rlimit(RLIMIT_MEMLOCK);
805 lock_limit >>= PAGE_SHIFT;
807 if (down_write_killable(¤t->mm->mmap_sem))
811 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
812 capable(CAP_IPC_LOCK))
813 ret = apply_mlockall_flags(flags);
814 up_write(¤t->mm->mmap_sem);
815 if (!ret && (flags & MCL_CURRENT))
816 mm_populate(0, TASK_SIZE);
821 SYSCALL_DEFINE0(munlockall)
825 if (down_write_killable(¤t->mm->mmap_sem))
827 ret = apply_mlockall_flags(0);
828 up_write(¤t->mm->mmap_sem);
833 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
834 * shm segments) get accounted against the user_struct instead.
836 static DEFINE_SPINLOCK(shmlock_user_lock);
838 int user_shm_lock(size_t size, struct user_struct *user)
840 unsigned long lock_limit, locked;
843 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
844 lock_limit = rlimit(RLIMIT_MEMLOCK);
845 if (lock_limit == RLIM_INFINITY)
847 lock_limit >>= PAGE_SHIFT;
848 spin_lock(&shmlock_user_lock);
850 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
853 user->locked_shm += locked;
856 spin_unlock(&shmlock_user_lock);
860 void user_shm_unlock(size_t size, struct user_struct *user)
862 spin_lock(&shmlock_user_lock);
863 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
864 spin_unlock(&shmlock_user_lock);