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>
33 struct folio_batch fbatch;
36 static DEFINE_PER_CPU(struct mlock_fbatch, mlock_fbatch) = {
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 folios are marked with the PG_mlocked flag for efficient testing
52 * in vmscan and, possibly, the fault path; and to support semi-accurate
55 * An mlocked folio [folio_test_mlocked(folio)] is unevictable. As such, it
56 * will be ostensibly placed on the LRU "unevictable" list (actually no such
57 * list exists), rather than the [in]active lists. PG_unevictable is set to
58 * indicate the unevictable state.
61 static struct lruvec *__mlock_folio(struct folio *folio, struct lruvec *lruvec)
63 /* There is nothing more we can do while it's off LRU */
64 if (!folio_test_clear_lru(folio))
67 lruvec = folio_lruvec_relock_irq(folio, lruvec);
69 if (unlikely(folio_evictable(folio))) {
71 * This is a little surprising, but quite possible: PG_mlocked
72 * must have got cleared already by another CPU. Could this
73 * folio be unevictable? I'm not sure, but move it now if so.
75 if (folio_test_unevictable(folio)) {
76 lruvec_del_folio(lruvec, folio);
77 folio_clear_unevictable(folio);
78 lruvec_add_folio(lruvec, folio);
80 __count_vm_events(UNEVICTABLE_PGRESCUED,
81 folio_nr_pages(folio));
86 if (folio_test_unevictable(folio)) {
87 if (folio_test_mlocked(folio))
92 lruvec_del_folio(lruvec, folio);
93 folio_clear_active(folio);
94 folio_set_unevictable(folio);
95 folio->mlock_count = !!folio_test_mlocked(folio);
96 lruvec_add_folio(lruvec, folio);
97 __count_vm_events(UNEVICTABLE_PGCULLED, folio_nr_pages(folio));
103 static struct lruvec *__mlock_new_folio(struct folio *folio, struct lruvec *lruvec)
105 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
107 lruvec = folio_lruvec_relock_irq(folio, lruvec);
109 /* As above, this is a little surprising, but possible */
110 if (unlikely(folio_evictable(folio)))
113 folio_set_unevictable(folio);
114 folio->mlock_count = !!folio_test_mlocked(folio);
115 __count_vm_events(UNEVICTABLE_PGCULLED, folio_nr_pages(folio));
117 lruvec_add_folio(lruvec, folio);
118 folio_set_lru(folio);
122 static struct lruvec *__munlock_folio(struct folio *folio, struct lruvec *lruvec)
124 int nr_pages = folio_nr_pages(folio);
125 bool isolated = false;
127 if (!folio_test_clear_lru(folio))
131 lruvec = folio_lruvec_relock_irq(folio, lruvec);
133 if (folio_test_unevictable(folio)) {
134 /* Then mlock_count is maintained, but might undercount */
135 if (folio->mlock_count)
136 folio->mlock_count--;
137 if (folio->mlock_count)
140 /* else assume that was the last mlock: reclaim will fix it if not */
143 if (folio_test_clear_mlocked(folio)) {
144 __zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
145 if (isolated || !folio_test_unevictable(folio))
146 __count_vm_events(UNEVICTABLE_PGMUNLOCKED, nr_pages);
148 __count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
151 /* folio_evictable() has to be checked *after* clearing Mlocked */
152 if (isolated && folio_test_unevictable(folio) && folio_evictable(folio)) {
153 lruvec_del_folio(lruvec, folio);
154 folio_clear_unevictable(folio);
155 lruvec_add_folio(lruvec, folio);
156 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
160 folio_set_lru(folio);
165 * Flags held in the low bits of a struct folio pointer on the mlock_fbatch.
167 #define LRU_FOLIO 0x1
168 #define NEW_FOLIO 0x2
169 static inline struct folio *mlock_lru(struct folio *folio)
171 return (struct folio *)((unsigned long)folio + LRU_FOLIO);
174 static inline struct folio *mlock_new(struct folio *folio)
176 return (struct folio *)((unsigned long)folio + NEW_FOLIO);
180 * mlock_folio_batch() is derived from folio_batch_move_lru(): perhaps that can
181 * make use of such folio pointer flags in future, but for now just keep it for
182 * mlock. We could use three separate folio batches instead, but one feels
183 * better (munlocking a full folio batch does not need to drain mlocking folio
186 static void mlock_folio_batch(struct folio_batch *fbatch)
188 struct lruvec *lruvec = NULL;
193 for (i = 0; i < folio_batch_count(fbatch); i++) {
194 folio = fbatch->folios[i];
195 mlock = (unsigned long)folio & (LRU_FOLIO | NEW_FOLIO);
196 folio = (struct folio *)((unsigned long)folio - mlock);
197 fbatch->folios[i] = folio;
199 if (mlock & LRU_FOLIO)
200 lruvec = __mlock_folio(folio, lruvec);
201 else if (mlock & NEW_FOLIO)
202 lruvec = __mlock_new_folio(folio, lruvec);
204 lruvec = __munlock_folio(folio, lruvec);
208 unlock_page_lruvec_irq(lruvec);
209 folios_put(fbatch->folios, folio_batch_count(fbatch));
210 folio_batch_reinit(fbatch);
213 void mlock_drain_local(void)
215 struct folio_batch *fbatch;
217 local_lock(&mlock_fbatch.lock);
218 fbatch = this_cpu_ptr(&mlock_fbatch.fbatch);
219 if (folio_batch_count(fbatch))
220 mlock_folio_batch(fbatch);
221 local_unlock(&mlock_fbatch.lock);
224 void mlock_drain_remote(int cpu)
226 struct folio_batch *fbatch;
228 WARN_ON_ONCE(cpu_online(cpu));
229 fbatch = &per_cpu(mlock_fbatch.fbatch, cpu);
230 if (folio_batch_count(fbatch))
231 mlock_folio_batch(fbatch);
234 bool need_mlock_drain(int cpu)
236 return folio_batch_count(&per_cpu(mlock_fbatch.fbatch, 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 folio_batch *fbatch;
247 local_lock(&mlock_fbatch.lock);
248 fbatch = this_cpu_ptr(&mlock_fbatch.fbatch);
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 (!folio_batch_add(fbatch, mlock_lru(folio)) ||
259 folio_test_large(folio) || lru_cache_disabled())
260 mlock_folio_batch(fbatch);
261 local_unlock(&mlock_fbatch.lock);
265 * mlock_new_folio - mlock a newly allocated folio not yet on LRU
266 * @folio: folio to be mlocked, either normal or a THP head.
268 void mlock_new_folio(struct folio *folio)
270 struct folio_batch *fbatch;
271 int nr_pages = folio_nr_pages(folio);
273 local_lock(&mlock_fbatch.lock);
274 fbatch = this_cpu_ptr(&mlock_fbatch.fbatch);
275 folio_set_mlocked(folio);
277 zone_stat_mod_folio(folio, NR_MLOCK, nr_pages);
278 __count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
281 if (!folio_batch_add(fbatch, mlock_new(folio)) ||
282 folio_test_large(folio) || lru_cache_disabled())
283 mlock_folio_batch(fbatch);
284 local_unlock(&mlock_fbatch.lock);
288 * munlock_folio - munlock a folio
289 * @folio: folio to be munlocked, either normal or a THP head.
291 void munlock_folio(struct folio *folio)
293 struct folio_batch *fbatch;
295 local_lock(&mlock_fbatch.lock);
296 fbatch = this_cpu_ptr(&mlock_fbatch.fbatch);
298 * folio_test_clear_mlocked(folio) must be left to __munlock_folio(),
299 * which will check whether the folio is multiply mlocked.
302 if (!folio_batch_add(fbatch, folio) ||
303 folio_test_large(folio) || lru_cache_disabled())
304 mlock_folio_batch(fbatch);
305 local_unlock(&mlock_fbatch.lock);
308 static inline unsigned int folio_mlock_step(struct folio *folio,
309 pte_t *pte, unsigned long addr, unsigned long end)
311 unsigned int count, i, nr = folio_nr_pages(folio);
312 unsigned long pfn = folio_pfn(folio);
313 pte_t ptent = ptep_get(pte);
315 if (!folio_test_large(folio))
318 count = pfn + nr - pte_pfn(ptent);
319 count = min_t(unsigned int, count, (end - addr) >> PAGE_SHIFT);
321 for (i = 0; i < count; i++, pte++) {
322 pte_t entry = ptep_get(pte);
324 if (!pte_present(entry))
326 if (pte_pfn(entry) - pfn >= nr)
333 static inline bool allow_mlock_munlock(struct folio *folio,
334 struct vm_area_struct *vma, unsigned long start,
335 unsigned long end, unsigned int step)
338 * For unlock, allow munlock large folio which is partially
339 * mapped to VMA. As it's possible that large folio is
340 * mlocked and VMA is split later.
342 * During memory pressure, such kind of large folio can
343 * be split. And the pages are not in VM_LOCKed VMA
346 if (!(vma->vm_flags & VM_LOCKED))
349 /* folio_within_range() cannot take KSM, but any small folio is OK */
350 if (!folio_test_large(folio))
353 /* folio not in range [start, end), skip mlock */
354 if (!folio_within_range(folio, vma, start, end))
357 /* folio is not fully mapped, skip mlock */
358 if (step != folio_nr_pages(folio))
364 static int mlock_pte_range(pmd_t *pmd, unsigned long addr,
365 unsigned long end, struct mm_walk *walk)
368 struct vm_area_struct *vma = walk->vma;
370 pte_t *start_pte, *pte;
373 unsigned int step = 1;
374 unsigned long start = addr;
376 ptl = pmd_trans_huge_lock(pmd, vma);
378 if (!pmd_present(*pmd))
380 if (is_huge_zero_pmd(*pmd))
382 folio = page_folio(pmd_page(*pmd));
383 if (vma->vm_flags & VM_LOCKED)
386 munlock_folio(folio);
390 start_pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
392 walk->action = ACTION_AGAIN;
396 for (pte = start_pte; addr != end; pte++, addr += PAGE_SIZE) {
397 ptent = ptep_get(pte);
398 if (!pte_present(ptent))
400 folio = vm_normal_folio(vma, addr, ptent);
401 if (!folio || folio_is_zone_device(folio))
404 step = folio_mlock_step(folio, pte, addr, end);
405 if (!allow_mlock_munlock(folio, vma, start, end, step))
408 if (vma->vm_flags & VM_LOCKED)
411 munlock_folio(folio);
415 addr += (step - 1) << PAGE_SHIFT;
417 pte_unmap(start_pte);
425 * mlock_vma_pages_range() - mlock any pages already in the range,
426 * or munlock all pages in the range.
427 * @vma - vma containing range to be mlock()ed or munlock()ed
428 * @start - start address in @vma of the range
429 * @end - end of range in @vma
430 * @newflags - the new set of flags for @vma.
432 * Called for mlock(), mlock2() and mlockall(), to set @vma VM_LOCKED;
433 * called for munlock() and munlockall(), to clear VM_LOCKED from @vma.
435 static void mlock_vma_pages_range(struct vm_area_struct *vma,
436 unsigned long start, unsigned long end, vm_flags_t newflags)
438 static const struct mm_walk_ops mlock_walk_ops = {
439 .pmd_entry = mlock_pte_range,
440 .walk_lock = PGWALK_WRLOCK_VERIFY,
444 * There is a slight chance that concurrent page migration,
445 * or page reclaim finding a page of this now-VM_LOCKED vma,
446 * will call mlock_vma_folio() and raise page's mlock_count:
447 * double counting, leaving the page unevictable indefinitely.
448 * Communicate this danger to mlock_vma_folio() with VM_IO,
449 * which is a VM_SPECIAL flag not allowed on VM_LOCKED vmas.
450 * mmap_lock is held in write mode here, so this weird
451 * combination should not be visible to other mmap_lock users;
452 * but WRITE_ONCE so rmap walkers must see VM_IO if VM_LOCKED.
454 if (newflags & VM_LOCKED)
456 vma_start_write(vma);
457 vm_flags_reset_once(vma, newflags);
460 walk_page_range(vma->vm_mm, start, end, &mlock_walk_ops, NULL);
463 if (newflags & VM_IO) {
465 vm_flags_reset_once(vma, newflags);
470 * mlock_fixup - handle mlock[all]/munlock[all] requests.
472 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
473 * munlock is a no-op. However, for some special vmas, we go ahead and
476 * For vmas that pass the filters, merge/split as appropriate.
478 static int mlock_fixup(struct vma_iterator *vmi, struct vm_area_struct *vma,
479 struct vm_area_struct **prev, unsigned long start,
480 unsigned long end, vm_flags_t newflags)
482 struct mm_struct *mm = vma->vm_mm;
485 vm_flags_t oldflags = vma->vm_flags;
487 if (newflags == oldflags || (oldflags & VM_SPECIAL) ||
488 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm) ||
489 vma_is_dax(vma) || vma_is_secretmem(vma))
490 /* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
493 vma = vma_modify_flags(vmi, *prev, vma, start, end, newflags);
500 * Keep track of amount of locked VM.
502 nr_pages = (end - start) >> PAGE_SHIFT;
503 if (!(newflags & VM_LOCKED))
504 nr_pages = -nr_pages;
505 else if (oldflags & VM_LOCKED)
507 mm->locked_vm += nr_pages;
510 * vm_flags is protected by the mmap_lock held in write mode.
511 * It's okay if try_to_unmap_one unmaps a page just after we
512 * set VM_LOCKED, populate_vma_page_range will bring it back.
514 if ((newflags & VM_LOCKED) && (oldflags & VM_LOCKED)) {
515 /* No work to do, and mlocking twice would be wrong */
516 vma_start_write(vma);
517 vm_flags_reset(vma, newflags);
519 mlock_vma_pages_range(vma, start, end, newflags);
526 static int apply_vma_lock_flags(unsigned long start, size_t len,
529 unsigned long nstart, end, tmp;
530 struct vm_area_struct *vma, *prev;
531 VMA_ITERATOR(vmi, current->mm, start);
533 VM_BUG_ON(offset_in_page(start));
534 VM_BUG_ON(len != PAGE_ALIGN(len));
540 vma = vma_iter_load(&vmi);
544 prev = vma_prev(&vmi);
545 if (start > vma->vm_start)
550 for_each_vma_range(vmi, vma, end) {
554 if (vma->vm_start != tmp)
557 newflags = vma->vm_flags & ~VM_LOCKED_MASK;
559 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
563 error = mlock_fixup(&vmi, vma, &prev, nstart, tmp, newflags);
566 tmp = vma_iter_end(&vmi);
577 * Go through vma areas and sum size of mlocked
578 * vma pages, as return value.
579 * Note deferred memory locking case(mlock2(,,MLOCK_ONFAULT)
581 * Return value: previously mlocked page counts
583 static unsigned long count_mm_mlocked_page_nr(struct mm_struct *mm,
584 unsigned long start, size_t len)
586 struct vm_area_struct *vma;
587 unsigned long count = 0;
589 VMA_ITERATOR(vmi, mm, start);
591 /* Don't overflow past ULONG_MAX */
592 if (unlikely(ULONG_MAX - len < start))
597 for_each_vma_range(vmi, vma, end) {
598 if (vma->vm_flags & VM_LOCKED) {
599 if (start > vma->vm_start)
600 count -= (start - vma->vm_start);
601 if (end < vma->vm_end) {
602 count += end - vma->vm_start;
605 count += vma->vm_end - vma->vm_start;
609 return count >> PAGE_SHIFT;
613 * convert get_user_pages() return value to posix mlock() error
615 static int __mlock_posix_error_return(long retval)
617 if (retval == -EFAULT)
619 else if (retval == -ENOMEM)
624 static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
626 unsigned long locked;
627 unsigned long lock_limit;
630 start = untagged_addr(start);
635 len = PAGE_ALIGN(len + (offset_in_page(start)));
638 lock_limit = rlimit(RLIMIT_MEMLOCK);
639 lock_limit >>= PAGE_SHIFT;
640 locked = len >> PAGE_SHIFT;
642 if (mmap_write_lock_killable(current->mm))
645 locked += current->mm->locked_vm;
646 if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
648 * It is possible that the regions requested intersect with
649 * previously mlocked areas, that part area in "mm->locked_vm"
650 * should not be counted to new mlock increment count. So check
651 * and adjust locked count if necessary.
653 locked -= count_mm_mlocked_page_nr(current->mm,
657 /* check against resource limits */
658 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
659 error = apply_vma_lock_flags(start, len, flags);
661 mmap_write_unlock(current->mm);
665 error = __mm_populate(start, len, 0);
667 return __mlock_posix_error_return(error);
671 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
673 return do_mlock(start, len, VM_LOCKED);
676 SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
678 vm_flags_t vm_flags = VM_LOCKED;
680 if (flags & ~MLOCK_ONFAULT)
683 if (flags & MLOCK_ONFAULT)
684 vm_flags |= VM_LOCKONFAULT;
686 return do_mlock(start, len, vm_flags);
689 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
693 start = untagged_addr(start);
695 len = PAGE_ALIGN(len + (offset_in_page(start)));
698 if (mmap_write_lock_killable(current->mm))
700 ret = apply_vma_lock_flags(start, len, 0);
701 mmap_write_unlock(current->mm);
707 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
708 * and translate into the appropriate modifications to mm->def_flags and/or the
709 * flags for all current VMAs.
711 * There are a couple of subtleties with this. If mlockall() is called multiple
712 * times with different flags, the values do not necessarily stack. If mlockall
713 * is called once including the MCL_FUTURE flag and then a second time without
714 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
716 static int apply_mlockall_flags(int flags)
718 VMA_ITERATOR(vmi, current->mm, 0);
719 struct vm_area_struct *vma, *prev = NULL;
720 vm_flags_t to_add = 0;
722 current->mm->def_flags &= ~VM_LOCKED_MASK;
723 if (flags & MCL_FUTURE) {
724 current->mm->def_flags |= VM_LOCKED;
726 if (flags & MCL_ONFAULT)
727 current->mm->def_flags |= VM_LOCKONFAULT;
729 if (!(flags & MCL_CURRENT))
733 if (flags & MCL_CURRENT) {
735 if (flags & MCL_ONFAULT)
736 to_add |= VM_LOCKONFAULT;
739 for_each_vma(vmi, vma) {
742 newflags = vma->vm_flags & ~VM_LOCKED_MASK;
746 mlock_fixup(&vmi, vma, &prev, vma->vm_start, vma->vm_end,
754 SYSCALL_DEFINE1(mlockall, int, flags)
756 unsigned long lock_limit;
759 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)) ||
760 flags == MCL_ONFAULT)
766 lock_limit = rlimit(RLIMIT_MEMLOCK);
767 lock_limit >>= PAGE_SHIFT;
769 if (mmap_write_lock_killable(current->mm))
773 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
774 capable(CAP_IPC_LOCK))
775 ret = apply_mlockall_flags(flags);
776 mmap_write_unlock(current->mm);
777 if (!ret && (flags & MCL_CURRENT))
778 mm_populate(0, TASK_SIZE);
783 SYSCALL_DEFINE0(munlockall)
787 if (mmap_write_lock_killable(current->mm))
789 ret = apply_mlockall_flags(0);
790 mmap_write_unlock(current->mm);
795 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
796 * shm segments) get accounted against the user_struct instead.
798 static DEFINE_SPINLOCK(shmlock_user_lock);
800 int user_shm_lock(size_t size, struct ucounts *ucounts)
802 unsigned long lock_limit, locked;
806 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
807 lock_limit = rlimit(RLIMIT_MEMLOCK);
808 if (lock_limit != RLIM_INFINITY)
809 lock_limit >>= PAGE_SHIFT;
810 spin_lock(&shmlock_user_lock);
811 memlock = inc_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
813 if ((memlock == LONG_MAX || memlock > lock_limit) && !capable(CAP_IPC_LOCK)) {
814 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
817 if (!get_ucounts(ucounts)) {
818 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
824 spin_unlock(&shmlock_user_lock);
828 void user_shm_unlock(size_t size, struct ucounts *ucounts)
830 spin_lock(&shmlock_user_lock);
831 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
832 spin_unlock(&shmlock_user_lock);
833 put_ucounts(ucounts);