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 (rlimit(RLIMIT_MEMLOCK) != 0)
31 if (capable(CAP_IPC_LOCK))
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 * convert get_user_pages() return value to posix mlock() error
211 static int __mlock_posix_error_return(long retval)
213 if (retval == -EFAULT)
215 else if (retval == -ENOMEM)
221 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
223 * The fast path is available only for evictable pages with single mapping.
224 * Then we can bypass the per-cpu pvec and get better performance.
225 * when mapcount > 1 we need try_to_munlock() which can fail.
226 * when !page_evictable(), we need the full redo logic of putback_lru_page to
227 * avoid leaving evictable page in unevictable list.
229 * In case of success, @page is added to @pvec and @pgrescued is incremented
230 * in case that the page was previously unevictable. @page is also unlocked.
232 static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
235 VM_BUG_ON_PAGE(PageLRU(page), page);
236 VM_BUG_ON_PAGE(!PageLocked(page), page);
238 if (page_mapcount(page) <= 1 && page_evictable(page)) {
239 pagevec_add(pvec, page);
240 if (TestClearPageUnevictable(page))
250 * Putback multiple evictable pages to the LRU
252 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
253 * the pages might have meanwhile become unevictable but that is OK.
255 static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
257 count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
259 *__pagevec_lru_add() calls release_pages() so we don't call
260 * put_page() explicitly
262 __pagevec_lru_add(pvec);
263 count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
267 * Munlock a batch of pages from the same zone
269 * The work is split to two main phases. First phase clears the Mlocked flag
270 * and attempts to isolate the pages, all under a single zone lru lock.
271 * The second phase finishes the munlock only for pages where isolation
274 * Note that the pagevec may be modified during the process.
276 static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
279 int nr = pagevec_count(pvec);
281 struct pagevec pvec_putback;
284 pagevec_init(&pvec_putback, 0);
286 /* Phase 1: page isolation */
287 spin_lock_irq(&zone->lru_lock);
288 for (i = 0; i < nr; i++) {
289 struct page *page = pvec->pages[i];
291 if (TestClearPageMlocked(page)) {
293 * We already have pin from follow_page_mask()
294 * so we can spare the get_page() here.
296 if (__munlock_isolate_lru_page(page, false))
299 __munlock_isolation_failed(page);
303 * We won't be munlocking this page in the next phase
304 * but we still need to release the follow_page_mask()
305 * pin. We cannot do it under lru_lock however. If it's
306 * the last pin, __page_cache_release() would deadlock.
308 pagevec_add(&pvec_putback, pvec->pages[i]);
309 pvec->pages[i] = NULL;
311 delta_munlocked = -nr + pagevec_count(&pvec_putback);
312 __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
313 spin_unlock_irq(&zone->lru_lock);
315 /* Now we can release pins of pages that we are not munlocking */
316 pagevec_release(&pvec_putback);
318 /* Phase 2: page munlock */
319 for (i = 0; i < nr; i++) {
320 struct page *page = pvec->pages[i];
324 if (!__putback_lru_fast_prepare(page, &pvec_putback,
327 * Slow path. We don't want to lose the last
328 * pin before unlock_page()
330 get_page(page); /* for putback_lru_page() */
331 __munlock_isolated_page(page);
333 put_page(page); /* from follow_page_mask() */
339 * Phase 3: page putback for pages that qualified for the fast path
340 * This will also call put_page() to return pin from follow_page_mask()
342 if (pagevec_count(&pvec_putback))
343 __putback_lru_fast(&pvec_putback, pgrescued);
347 * Fill up pagevec for __munlock_pagevec using pte walk
349 * The function expects that the struct page corresponding to @start address is
350 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
352 * The rest of @pvec is filled by subsequent pages within the same pmd and same
353 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
354 * pages also get pinned.
356 * Returns the address of the next page that should be scanned. This equals
357 * @start + PAGE_SIZE when no page could be added by the pte walk.
359 static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
360 struct vm_area_struct *vma, int zoneid, unsigned long start,
367 * Initialize pte walk starting at the already pinned page where we
368 * are sure that there is a pte, as it was pinned under the same
371 pte = get_locked_pte(vma->vm_mm, start, &ptl);
372 /* Make sure we do not cross the page table boundary */
373 end = pgd_addr_end(start, end);
374 end = pud_addr_end(start, end);
375 end = pmd_addr_end(start, end);
377 /* The page next to the pinned page is the first we will try to get */
379 while (start < end) {
380 struct page *page = NULL;
382 if (pte_present(*pte))
383 page = vm_normal_page(vma, start, *pte);
385 * Break if page could not be obtained or the page's node+zone does not
388 if (!page || page_zone_id(page) != zoneid)
393 * Increase the address that will be returned *before* the
394 * eventual break due to pvec becoming full by adding the page
397 if (pagevec_add(pvec, page) == 0)
400 pte_unmap_unlock(pte, ptl);
405 * munlock_vma_pages_range() - munlock all pages in the vma range.'
406 * @vma - vma containing range to be munlock()ed.
407 * @start - start address in @vma of the range
408 * @end - end of range in @vma.
410 * For mremap(), munmap() and exit().
412 * Called with @vma VM_LOCKED.
414 * Returns with VM_LOCKED cleared. Callers must be prepared to
417 * We don't save and restore VM_LOCKED here because pages are
418 * still on lru. In unmap path, pages might be scanned by reclaim
419 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
420 * free them. This will result in freeing mlocked pages.
422 void munlock_vma_pages_range(struct vm_area_struct *vma,
423 unsigned long start, unsigned long end)
425 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
427 while (start < end) {
428 struct page *page = NULL;
429 unsigned int page_mask;
430 unsigned long page_increm;
435 pagevec_init(&pvec, 0);
437 * Although FOLL_DUMP is intended for get_dump_page(),
438 * it just so happens that its special treatment of the
439 * ZERO_PAGE (returning an error instead of doing get_page)
440 * suits munlock very well (and if somehow an abnormal page
441 * has sneaked into the range, we won't oops here: great).
443 page = follow_page_mask(vma, start, FOLL_GET | FOLL_DUMP,
446 if (page && !IS_ERR(page)) {
447 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 recompute
453 * the page_mask here.
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);
466 zoneid = page_zone_id(page);
469 * Try to fill the rest of pagevec using fast
470 * pte walk. This will also update start to
471 * the next page to process. Then munlock the
474 start = __munlock_pagevec_fill(&pvec, vma,
476 __munlock_pagevec(&pvec, zone);
480 /* It's a bug to munlock in the middle of a THP page */
481 VM_BUG_ON((start >> PAGE_SHIFT) & page_mask);
482 page_increm = 1 + page_mask;
483 start += page_increm * PAGE_SIZE;
490 * mlock_fixup - handle mlock[all]/munlock[all] requests.
492 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
493 * munlock is a no-op. However, for some special vmas, we go ahead and
496 * For vmas that pass the filters, merge/split as appropriate.
498 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
499 unsigned long start, unsigned long end, vm_flags_t newflags)
501 struct mm_struct *mm = vma->vm_mm;
505 int lock = !!(newflags & VM_LOCKED);
507 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
508 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
509 /* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
512 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
513 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
514 vma->vm_file, pgoff, vma_policy(vma),
515 vma->vm_userfaultfd_ctx);
521 if (start != vma->vm_start) {
522 ret = split_vma(mm, vma, start, 1);
527 if (end != vma->vm_end) {
528 ret = split_vma(mm, vma, end, 0);
535 * Keep track of amount of locked VM.
537 nr_pages = (end - start) >> PAGE_SHIFT;
539 nr_pages = -nr_pages;
540 mm->locked_vm += nr_pages;
543 * vm_flags is protected by the mmap_sem 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) {
607 static int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
609 unsigned long locked;
610 unsigned long lock_limit;
616 lru_add_drain_all(); /* flush pagevec */
618 len = PAGE_ALIGN(len + (offset_in_page(start)));
621 lock_limit = rlimit(RLIMIT_MEMLOCK);
622 lock_limit >>= PAGE_SHIFT;
623 locked = len >> PAGE_SHIFT;
625 down_write(¤t->mm->mmap_sem);
627 locked += current->mm->locked_vm;
629 /* check against resource limits */
630 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
631 error = apply_vma_lock_flags(start, len, flags);
633 up_write(¤t->mm->mmap_sem);
637 error = __mm_populate(start, len, 0);
639 return __mlock_posix_error_return(error);
643 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
645 return do_mlock(start, len, VM_LOCKED);
648 SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
650 vm_flags_t vm_flags = VM_LOCKED;
652 if (flags & ~MLOCK_ONFAULT)
655 if (flags & MLOCK_ONFAULT)
656 vm_flags |= VM_LOCKONFAULT;
658 return do_mlock(start, len, vm_flags);
661 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
665 len = PAGE_ALIGN(len + (offset_in_page(start)));
668 down_write(¤t->mm->mmap_sem);
669 ret = apply_vma_lock_flags(start, len, 0);
670 up_write(¤t->mm->mmap_sem);
676 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
677 * and translate into the appropriate modifications to mm->def_flags and/or the
678 * flags for all current VMAs.
680 * There are a couple of subtleties with this. If mlockall() is called multiple
681 * times with different flags, the values do not necessarily stack. If mlockall
682 * is called once including the MCL_FUTURE flag and then a second time without
683 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
685 static int apply_mlockall_flags(int flags)
687 struct vm_area_struct * vma, * prev = NULL;
688 vm_flags_t to_add = 0;
690 current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
691 if (flags & MCL_FUTURE) {
692 current->mm->def_flags |= VM_LOCKED;
694 if (flags & MCL_ONFAULT)
695 current->mm->def_flags |= VM_LOCKONFAULT;
697 if (!(flags & MCL_CURRENT))
701 if (flags & MCL_CURRENT) {
703 if (flags & MCL_ONFAULT)
704 to_add |= VM_LOCKONFAULT;
707 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
710 newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
714 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
715 cond_resched_rcu_qs();
721 SYSCALL_DEFINE1(mlockall, int, flags)
723 unsigned long lock_limit;
726 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)))
732 if (flags & MCL_CURRENT)
733 lru_add_drain_all(); /* flush pagevec */
735 lock_limit = rlimit(RLIMIT_MEMLOCK);
736 lock_limit >>= PAGE_SHIFT;
739 down_write(¤t->mm->mmap_sem);
741 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
742 capable(CAP_IPC_LOCK))
743 ret = apply_mlockall_flags(flags);
744 up_write(¤t->mm->mmap_sem);
745 if (!ret && (flags & MCL_CURRENT))
746 mm_populate(0, TASK_SIZE);
751 SYSCALL_DEFINE0(munlockall)
755 down_write(¤t->mm->mmap_sem);
756 ret = apply_mlockall_flags(0);
757 up_write(¤t->mm->mmap_sem);
762 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
763 * shm segments) get accounted against the user_struct instead.
765 static DEFINE_SPINLOCK(shmlock_user_lock);
767 int user_shm_lock(size_t size, struct user_struct *user)
769 unsigned long lock_limit, locked;
772 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
773 lock_limit = rlimit(RLIMIT_MEMLOCK);
774 if (lock_limit == RLIM_INFINITY)
776 lock_limit >>= PAGE_SHIFT;
777 spin_lock(&shmlock_user_lock);
779 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
782 user->locked_shm += locked;
785 spin_unlock(&shmlock_user_lock);
789 void user_shm_unlock(size_t size, struct user_struct *user)
791 spin_lock(&shmlock_user_lock);
792 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
793 spin_unlock(&shmlock_user_lock);