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/pagewalk.h>
18 #include <linux/mempolicy.h>
19 #include <linux/syscalls.h>
20 #include <linux/sched.h>
21 #include <linux/export.h>
22 #include <linux/rmap.h>
23 #include <linux/mmzone.h>
24 #include <linux/hugetlb.h>
25 #include <linux/memcontrol.h>
26 #include <linux/mm_inline.h>
27 #include <linux/secretmem.h>
36 static DEFINE_PER_CPU(struct mlock_pvec, mlock_pvec) = {
37 .lock = INIT_LOCAL_LOCK(lock),
40 bool can_do_mlock(void)
42 if (rlimit(RLIMIT_MEMLOCK) != 0)
44 if (capable(CAP_IPC_LOCK))
48 EXPORT_SYMBOL(can_do_mlock);
51 * Mlocked pages are marked with PageMlocked() flag for efficient testing
52 * in vmscan and, possibly, the fault path; and to support semi-accurate
55 * An mlocked page [PageMlocked(page)] is unevictable. As such, it will
56 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
57 * The unevictable list is an LRU sibling list to the [in]active lists.
58 * PageUnevictable is set to indicate the unevictable state.
61 static struct lruvec *__mlock_page(struct page *page, struct lruvec *lruvec)
63 /* There is nothing more we can do while it's off LRU */
64 if (!TestClearPageLRU(page))
67 lruvec = folio_lruvec_relock_irq(page_folio(page), lruvec);
69 if (unlikely(page_evictable(page))) {
71 * This is a little surprising, but quite possible:
72 * PageMlocked must have got cleared already by another CPU.
73 * Could this page be on the Unevictable LRU? I'm not sure,
74 * but move it now if so.
76 if (PageUnevictable(page)) {
77 del_page_from_lru_list(page, lruvec);
78 ClearPageUnevictable(page);
79 add_page_to_lru_list(page, lruvec);
80 __count_vm_events(UNEVICTABLE_PGRESCUED,
86 if (PageUnevictable(page)) {
87 if (PageMlocked(page))
92 del_page_from_lru_list(page, lruvec);
93 ClearPageActive(page);
94 SetPageUnevictable(page);
95 page->mlock_count = !!PageMlocked(page);
96 add_page_to_lru_list(page, lruvec);
97 __count_vm_events(UNEVICTABLE_PGCULLED, thp_nr_pages(page));
103 static struct lruvec *__mlock_new_page(struct page *page, struct lruvec *lruvec)
105 VM_BUG_ON_PAGE(PageLRU(page), page);
107 lruvec = folio_lruvec_relock_irq(page_folio(page), lruvec);
109 /* As above, this is a little surprising, but possible */
110 if (unlikely(page_evictable(page)))
113 SetPageUnevictable(page);
114 page->mlock_count = !!PageMlocked(page);
115 __count_vm_events(UNEVICTABLE_PGCULLED, thp_nr_pages(page));
117 add_page_to_lru_list(page, lruvec);
122 static struct lruvec *__munlock_page(struct page *page, struct lruvec *lruvec)
124 int nr_pages = thp_nr_pages(page);
125 bool isolated = false;
127 if (!TestClearPageLRU(page))
131 lruvec = folio_lruvec_relock_irq(page_folio(page), lruvec);
133 if (PageUnevictable(page)) {
134 /* Then mlock_count is maintained, but might undercount */
135 if (page->mlock_count)
137 if (page->mlock_count)
140 /* else assume that was the last mlock: reclaim will fix it if not */
143 if (TestClearPageMlocked(page)) {
144 __mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
145 if (isolated || !PageUnevictable(page))
146 __count_vm_events(UNEVICTABLE_PGMUNLOCKED, nr_pages);
148 __count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
151 /* page_evictable() has to be checked *after* clearing Mlocked */
152 if (isolated && PageUnevictable(page) && page_evictable(page)) {
153 del_page_from_lru_list(page, lruvec);
154 ClearPageUnevictable(page);
155 add_page_to_lru_list(page, lruvec);
156 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
165 * Flags held in the low bits of a struct page pointer on the mlock_pvec.
169 static inline struct page *mlock_lru(struct page *page)
171 return (struct page *)((unsigned long)page + LRU_PAGE);
174 static inline struct page *mlock_new(struct page *page)
176 return (struct page *)((unsigned long)page + NEW_PAGE);
180 * mlock_pagevec() is derived from pagevec_lru_move_fn():
181 * perhaps that can make use of such page pointer flags in future,
182 * but for now just keep it for mlock. We could use three separate
183 * pagevecs instead, but one feels better (munlocking a full pagevec
184 * does not need to drain mlocking pagevecs first).
186 static void mlock_pagevec(struct pagevec *pvec)
188 struct lruvec *lruvec = NULL;
193 for (i = 0; i < pagevec_count(pvec); i++) {
194 page = pvec->pages[i];
195 mlock = (unsigned long)page & (LRU_PAGE | NEW_PAGE);
196 page = (struct page *)((unsigned long)page - mlock);
197 pvec->pages[i] = page;
199 if (mlock & LRU_PAGE)
200 lruvec = __mlock_page(page, lruvec);
201 else if (mlock & NEW_PAGE)
202 lruvec = __mlock_new_page(page, lruvec);
204 lruvec = __munlock_page(page, lruvec);
208 unlock_page_lruvec_irq(lruvec);
209 release_pages(pvec->pages, pvec->nr);
210 pagevec_reinit(pvec);
213 void mlock_page_drain_local(void)
215 struct pagevec *pvec;
217 local_lock(&mlock_pvec.lock);
218 pvec = this_cpu_ptr(&mlock_pvec.vec);
219 if (pagevec_count(pvec))
221 local_unlock(&mlock_pvec.lock);
224 void mlock_page_drain_remote(int cpu)
226 struct pagevec *pvec;
228 WARN_ON_ONCE(cpu_online(cpu));
229 pvec = &per_cpu(mlock_pvec.vec, cpu);
230 if (pagevec_count(pvec))
234 bool need_mlock_page_drain(int cpu)
236 return pagevec_count(&per_cpu(mlock_pvec.vec, cpu));
240 * mlock_folio - mlock a folio already on (or temporarily off) LRU
241 * @folio: folio to be mlocked.
243 void mlock_folio(struct folio *folio)
245 struct pagevec *pvec;
247 local_lock(&mlock_pvec.lock);
248 pvec = this_cpu_ptr(&mlock_pvec.vec);
250 if (!folio_test_set_mlocked(folio)) {
251 int nr_pages = folio_nr_pages(folio);
253 zone_stat_mod_folio(folio, NR_MLOCK, nr_pages);
254 __count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
258 if (!pagevec_add(pvec, mlock_lru(&folio->page)) ||
259 folio_test_large(folio) || lru_cache_disabled())
261 local_unlock(&mlock_pvec.lock);
265 * mlock_new_page - mlock a newly allocated page not yet on LRU
266 * @page: page to be mlocked, either a normal page or a THP head.
268 void mlock_new_page(struct page *page)
270 struct pagevec *pvec;
271 int nr_pages = thp_nr_pages(page);
273 local_lock(&mlock_pvec.lock);
274 pvec = this_cpu_ptr(&mlock_pvec.vec);
275 SetPageMlocked(page);
276 mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
277 __count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
280 if (!pagevec_add(pvec, mlock_new(page)) ||
281 PageHead(page) || lru_cache_disabled())
283 local_unlock(&mlock_pvec.lock);
287 * munlock_page - munlock a page
288 * @page: page to be munlocked, either a normal page or a THP head.
290 void munlock_page(struct page *page)
292 struct pagevec *pvec;
294 local_lock(&mlock_pvec.lock);
295 pvec = this_cpu_ptr(&mlock_pvec.vec);
297 * TestClearPageMlocked(page) must be left to __munlock_page(),
298 * which will check whether the page is multiply mlocked.
302 if (!pagevec_add(pvec, page) ||
303 PageHead(page) || lru_cache_disabled())
305 local_unlock(&mlock_pvec.lock);
308 static int mlock_pte_range(pmd_t *pmd, unsigned long addr,
309 unsigned long end, struct mm_walk *walk)
312 struct vm_area_struct *vma = walk->vma;
314 pte_t *start_pte, *pte;
317 ptl = pmd_trans_huge_lock(pmd, vma);
319 if (!pmd_present(*pmd))
321 if (is_huge_zero_pmd(*pmd))
323 page = pmd_page(*pmd);
324 if (vma->vm_flags & VM_LOCKED)
325 mlock_folio(page_folio(page));
331 start_pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
332 for (pte = start_pte; addr != end; pte++, addr += PAGE_SIZE) {
333 if (!pte_present(*pte))
335 page = vm_normal_page(vma, addr, *pte);
336 if (!page || is_zone_device_page(page))
338 if (PageTransCompound(page))
340 if (vma->vm_flags & VM_LOCKED)
341 mlock_folio(page_folio(page));
345 pte_unmap(start_pte);
353 * mlock_vma_pages_range() - mlock any pages already in the range,
354 * or munlock all pages in the range.
355 * @vma - vma containing range to be mlock()ed or munlock()ed
356 * @start - start address in @vma of the range
357 * @end - end of range in @vma
358 * @newflags - the new set of flags for @vma.
360 * Called for mlock(), mlock2() and mlockall(), to set @vma VM_LOCKED;
361 * called for munlock() and munlockall(), to clear VM_LOCKED from @vma.
363 static void mlock_vma_pages_range(struct vm_area_struct *vma,
364 unsigned long start, unsigned long end, vm_flags_t newflags)
366 static const struct mm_walk_ops mlock_walk_ops = {
367 .pmd_entry = mlock_pte_range,
371 * There is a slight chance that concurrent page migration,
372 * or page reclaim finding a page of this now-VM_LOCKED vma,
373 * will call mlock_vma_page() and raise page's mlock_count:
374 * double counting, leaving the page unevictable indefinitely.
375 * Communicate this danger to mlock_vma_page() with VM_IO,
376 * which is a VM_SPECIAL flag not allowed on VM_LOCKED vmas.
377 * mmap_lock is held in write mode here, so this weird
378 * combination should not be visible to other mmap_lock users;
379 * but WRITE_ONCE so rmap walkers must see VM_IO if VM_LOCKED.
381 if (newflags & VM_LOCKED)
383 WRITE_ONCE(vma->vm_flags, newflags);
386 walk_page_range(vma->vm_mm, start, end, &mlock_walk_ops, NULL);
389 if (newflags & VM_IO) {
391 WRITE_ONCE(vma->vm_flags, newflags);
396 * mlock_fixup - handle mlock[all]/munlock[all] requests.
398 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
399 * munlock is a no-op. However, for some special vmas, we go ahead and
402 * For vmas that pass the filters, merge/split as appropriate.
404 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
405 unsigned long start, unsigned long end, vm_flags_t newflags)
407 struct mm_struct *mm = vma->vm_mm;
411 vm_flags_t oldflags = vma->vm_flags;
413 if (newflags == oldflags || (oldflags & VM_SPECIAL) ||
414 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm) ||
415 vma_is_dax(vma) || vma_is_secretmem(vma))
416 /* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
419 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
420 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
421 vma->vm_file, pgoff, vma_policy(vma),
422 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
428 if (start != vma->vm_start) {
429 ret = split_vma(mm, vma, start, 1);
434 if (end != vma->vm_end) {
435 ret = split_vma(mm, vma, end, 0);
442 * Keep track of amount of locked VM.
444 nr_pages = (end - start) >> PAGE_SHIFT;
445 if (!(newflags & VM_LOCKED))
446 nr_pages = -nr_pages;
447 else if (oldflags & VM_LOCKED)
449 mm->locked_vm += nr_pages;
452 * vm_flags is protected by the mmap_lock held in write mode.
453 * It's okay if try_to_unmap_one unmaps a page just after we
454 * set VM_LOCKED, populate_vma_page_range will bring it back.
457 if ((newflags & VM_LOCKED) && (oldflags & VM_LOCKED)) {
458 /* No work to do, and mlocking twice would be wrong */
459 vma->vm_flags = newflags;
461 mlock_vma_pages_range(vma, start, end, newflags);
468 static int apply_vma_lock_flags(unsigned long start, size_t len,
471 unsigned long nstart, end, tmp;
472 struct vm_area_struct *vma, *prev;
474 MA_STATE(mas, ¤t->mm->mm_mt, start, start);
476 VM_BUG_ON(offset_in_page(start));
477 VM_BUG_ON(len != PAGE_ALIGN(len));
483 vma = mas_walk(&mas);
487 if (start > vma->vm_start)
490 prev = mas_prev(&mas, 0);
492 for (nstart = start ; ; ) {
493 vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
497 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
501 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
505 if (nstart < prev->vm_end)
506 nstart = prev->vm_end;
510 vma = find_vma(prev->vm_mm, prev->vm_end);
511 if (!vma || vma->vm_start != nstart) {
520 * Go through vma areas and sum size of mlocked
521 * vma pages, as return value.
522 * Note deferred memory locking case(mlock2(,,MLOCK_ONFAULT)
524 * Return value: previously mlocked page counts
526 static unsigned long count_mm_mlocked_page_nr(struct mm_struct *mm,
527 unsigned long start, size_t len)
529 struct vm_area_struct *vma;
530 unsigned long count = 0;
532 VMA_ITERATOR(vmi, mm, start);
534 /* Don't overflow past ULONG_MAX */
535 if (unlikely(ULONG_MAX - len < start))
540 for_each_vma_range(vmi, vma, end) {
541 if (vma->vm_flags & VM_LOCKED) {
542 if (start > vma->vm_start)
543 count -= (start - vma->vm_start);
544 if (end < vma->vm_end) {
545 count += end - vma->vm_start;
548 count += vma->vm_end - vma->vm_start;
552 return count >> PAGE_SHIFT;
556 * convert get_user_pages() return value to posix mlock() error
558 static int __mlock_posix_error_return(long retval)
560 if (retval == -EFAULT)
562 else if (retval == -ENOMEM)
567 static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
569 unsigned long locked;
570 unsigned long lock_limit;
573 start = untagged_addr(start);
578 len = PAGE_ALIGN(len + (offset_in_page(start)));
581 lock_limit = rlimit(RLIMIT_MEMLOCK);
582 lock_limit >>= PAGE_SHIFT;
583 locked = len >> PAGE_SHIFT;
585 if (mmap_write_lock_killable(current->mm))
588 locked += current->mm->locked_vm;
589 if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
591 * It is possible that the regions requested intersect with
592 * previously mlocked areas, that part area in "mm->locked_vm"
593 * should not be counted to new mlock increment count. So check
594 * and adjust locked count if necessary.
596 locked -= count_mm_mlocked_page_nr(current->mm,
600 /* check against resource limits */
601 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
602 error = apply_vma_lock_flags(start, len, flags);
604 mmap_write_unlock(current->mm);
608 error = __mm_populate(start, len, 0);
610 return __mlock_posix_error_return(error);
614 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
616 return do_mlock(start, len, VM_LOCKED);
619 SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
621 vm_flags_t vm_flags = VM_LOCKED;
623 if (flags & ~MLOCK_ONFAULT)
626 if (flags & MLOCK_ONFAULT)
627 vm_flags |= VM_LOCKONFAULT;
629 return do_mlock(start, len, vm_flags);
632 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
636 start = untagged_addr(start);
638 len = PAGE_ALIGN(len + (offset_in_page(start)));
641 if (mmap_write_lock_killable(current->mm))
643 ret = apply_vma_lock_flags(start, len, 0);
644 mmap_write_unlock(current->mm);
650 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
651 * and translate into the appropriate modifications to mm->def_flags and/or the
652 * flags for all current VMAs.
654 * There are a couple of subtleties with this. If mlockall() is called multiple
655 * times with different flags, the values do not necessarily stack. If mlockall
656 * is called once including the MCL_FUTURE flag and then a second time without
657 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
659 static int apply_mlockall_flags(int flags)
661 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
662 struct vm_area_struct *vma, *prev = NULL;
663 vm_flags_t to_add = 0;
665 current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
666 if (flags & MCL_FUTURE) {
667 current->mm->def_flags |= VM_LOCKED;
669 if (flags & MCL_ONFAULT)
670 current->mm->def_flags |= VM_LOCKONFAULT;
672 if (!(flags & MCL_CURRENT))
676 if (flags & MCL_CURRENT) {
678 if (flags & MCL_ONFAULT)
679 to_add |= VM_LOCKONFAULT;
682 mas_for_each(&mas, vma, ULONG_MAX) {
685 newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
689 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
697 SYSCALL_DEFINE1(mlockall, int, flags)
699 unsigned long lock_limit;
702 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)) ||
703 flags == MCL_ONFAULT)
709 lock_limit = rlimit(RLIMIT_MEMLOCK);
710 lock_limit >>= PAGE_SHIFT;
712 if (mmap_write_lock_killable(current->mm))
716 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
717 capable(CAP_IPC_LOCK))
718 ret = apply_mlockall_flags(flags);
719 mmap_write_unlock(current->mm);
720 if (!ret && (flags & MCL_CURRENT))
721 mm_populate(0, TASK_SIZE);
726 SYSCALL_DEFINE0(munlockall)
730 if (mmap_write_lock_killable(current->mm))
732 ret = apply_mlockall_flags(0);
733 mmap_write_unlock(current->mm);
738 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
739 * shm segments) get accounted against the user_struct instead.
741 static DEFINE_SPINLOCK(shmlock_user_lock);
743 int user_shm_lock(size_t size, struct ucounts *ucounts)
745 unsigned long lock_limit, locked;
749 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
750 lock_limit = rlimit(RLIMIT_MEMLOCK);
751 if (lock_limit != RLIM_INFINITY)
752 lock_limit >>= PAGE_SHIFT;
753 spin_lock(&shmlock_user_lock);
754 memlock = inc_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
756 if ((memlock == LONG_MAX || memlock > lock_limit) && !capable(CAP_IPC_LOCK)) {
757 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
760 if (!get_ucounts(ucounts)) {
761 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
767 spin_unlock(&shmlock_user_lock);
771 void user_shm_unlock(size_t size, struct ucounts *ucounts)
773 spin_lock(&shmlock_user_lock);
774 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
775 spin_unlock(&shmlock_user_lock);
776 put_ucounts(ucounts);