Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/mlock.c
4 *
5 * (C) Copyright 1995 Linus Torvalds
6 * (C) Copyright 2002 Christoph Hellwig
7 */
8
9#include <linux/capability.h>
10#include <linux/mman.h>
11#include <linux/mm.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>
28
29#include "internal.h"
30
31struct mlock_pvec {
32 local_lock_t lock;
33 struct pagevec vec;
34};
35
36static DEFINE_PER_CPU(struct mlock_pvec, mlock_pvec) = {
37 .lock = INIT_LOCAL_LOCK(lock),
38};
39
40bool can_do_mlock(void)
41{
42 if (rlimit(RLIMIT_MEMLOCK) != 0)
43 return true;
44 if (capable(CAP_IPC_LOCK))
45 return true;
46 return false;
47}
48EXPORT_SYMBOL(can_do_mlock);
49
50/*
51 * Mlocked pages are marked with PageMlocked() flag for efficient testing
52 * in vmscan and, possibly, the fault path; and to support semi-accurate
53 * statistics.
54 *
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.
59 */
60
61static struct lruvec *__mlock_page(struct page *page, struct lruvec *lruvec)
62{
63 /* There is nothing more we can do while it's off LRU */
64 if (!TestClearPageLRU(page))
65 return lruvec;
66
67 lruvec = folio_lruvec_relock_irq(page_folio(page), lruvec);
68
69 if (unlikely(page_evictable(page))) {
70 /*
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.
75 */
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,
81 thp_nr_pages(page));
82 }
83 goto out;
84 }
85
86 if (PageUnevictable(page)) {
87 if (PageMlocked(page))
88 page->mlock_count++;
89 goto out;
90 }
91
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));
98out:
99 SetPageLRU(page);
100 return lruvec;
101}
102
103static struct lruvec *__mlock_new_page(struct page *page, struct lruvec *lruvec)
104{
105 VM_BUG_ON_PAGE(PageLRU(page), page);
106
107 lruvec = folio_lruvec_relock_irq(page_folio(page), lruvec);
108
109 /* As above, this is a little surprising, but possible */
110 if (unlikely(page_evictable(page)))
111 goto out;
112
113 SetPageUnevictable(page);
114 page->mlock_count = !!PageMlocked(page);
115 __count_vm_events(UNEVICTABLE_PGCULLED, thp_nr_pages(page));
116out:
117 add_page_to_lru_list(page, lruvec);
118 SetPageLRU(page);
119 return lruvec;
120}
121
122static struct lruvec *__munlock_page(struct page *page, struct lruvec *lruvec)
123{
124 int nr_pages = thp_nr_pages(page);
125 bool isolated = false;
126
127 if (!TestClearPageLRU(page))
128 goto munlock;
129
130 isolated = true;
131 lruvec = folio_lruvec_relock_irq(page_folio(page), lruvec);
132
133 if (PageUnevictable(page)) {
134 /* Then mlock_count is maintained, but might undercount */
135 if (page->mlock_count)
136 page->mlock_count--;
137 if (page->mlock_count)
138 goto out;
139 }
140 /* else assume that was the last mlock: reclaim will fix it if not */
141
142munlock:
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);
147 else
148 __count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
149 }
150
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);
157 }
158out:
159 if (isolated)
160 SetPageLRU(page);
161 return lruvec;
162}
163
164/*
165 * Flags held in the low bits of a struct page pointer on the mlock_pvec.
166 */
167#define LRU_PAGE 0x1
168#define NEW_PAGE 0x2
169static inline struct page *mlock_lru(struct page *page)
170{
171 return (struct page *)((unsigned long)page + LRU_PAGE);
172}
173
174static inline struct page *mlock_new(struct page *page)
175{
176 return (struct page *)((unsigned long)page + NEW_PAGE);
177}
178
179/*
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).
185 */
186static void mlock_pagevec(struct pagevec *pvec)
187{
188 struct lruvec *lruvec = NULL;
189 unsigned long mlock;
190 struct page *page;
191 int i;
192
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;
198
199 if (mlock & LRU_PAGE)
200 lruvec = __mlock_page(page, lruvec);
201 else if (mlock & NEW_PAGE)
202 lruvec = __mlock_new_page(page, lruvec);
203 else
204 lruvec = __munlock_page(page, lruvec);
205 }
206
207 if (lruvec)
208 unlock_page_lruvec_irq(lruvec);
209 release_pages(pvec->pages, pvec->nr);
210 pagevec_reinit(pvec);
211}
212
213void mlock_page_drain_local(void)
214{
215 struct pagevec *pvec;
216
217 local_lock(&mlock_pvec.lock);
218 pvec = this_cpu_ptr(&mlock_pvec.vec);
219 if (pagevec_count(pvec))
220 mlock_pagevec(pvec);
221 local_unlock(&mlock_pvec.lock);
222}
223
224void mlock_page_drain_remote(int cpu)
225{
226 struct pagevec *pvec;
227
228 WARN_ON_ONCE(cpu_online(cpu));
229 pvec = &per_cpu(mlock_pvec.vec, cpu);
230 if (pagevec_count(pvec))
231 mlock_pagevec(pvec);
232}
233
234bool need_mlock_page_drain(int cpu)
235{
236 return pagevec_count(&per_cpu(mlock_pvec.vec, cpu));
237}
238
239/**
240 * mlock_folio - mlock a folio already on (or temporarily off) LRU
241 * @folio: folio to be mlocked.
242 */
243void mlock_folio(struct folio *folio)
244{
245 struct pagevec *pvec;
246
247 local_lock(&mlock_pvec.lock);
248 pvec = this_cpu_ptr(&mlock_pvec.vec);
249
250 if (!folio_test_set_mlocked(folio)) {
251 int nr_pages = folio_nr_pages(folio);
252
253 zone_stat_mod_folio(folio, NR_MLOCK, nr_pages);
254 __count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
255 }
256
257 folio_get(folio);
258 if (!pagevec_add(pvec, mlock_lru(&folio->page)) ||
259 folio_test_large(folio) || lru_cache_disabled())
260 mlock_pagevec(pvec);
261 local_unlock(&mlock_pvec.lock);
262}
263
264/**
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.
267 */
268void mlock_new_page(struct page *page)
269{
270 struct pagevec *pvec;
271 int nr_pages = thp_nr_pages(page);
272
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);
278
279 get_page(page);
280 if (!pagevec_add(pvec, mlock_new(page)) ||
281 PageHead(page) || lru_cache_disabled())
282 mlock_pagevec(pvec);
283 local_unlock(&mlock_pvec.lock);
284}
285
286/**
287 * munlock_page - munlock a page
288 * @page: page to be munlocked, either a normal page or a THP head.
289 */
290void munlock_page(struct page *page)
291{
292 struct pagevec *pvec;
293
294 local_lock(&mlock_pvec.lock);
295 pvec = this_cpu_ptr(&mlock_pvec.vec);
296 /*
297 * TestClearPageMlocked(page) must be left to __munlock_page(),
298 * which will check whether the page is multiply mlocked.
299 */
300
301 get_page(page);
302 if (!pagevec_add(pvec, page) ||
303 PageHead(page) || lru_cache_disabled())
304 mlock_pagevec(pvec);
305 local_unlock(&mlock_pvec.lock);
306}
307
308static int mlock_pte_range(pmd_t *pmd, unsigned long addr,
309 unsigned long end, struct mm_walk *walk)
310
311{
312 struct vm_area_struct *vma = walk->vma;
313 spinlock_t *ptl;
314 pte_t *start_pte, *pte;
315 struct page *page;
316
317 ptl = pmd_trans_huge_lock(pmd, vma);
318 if (ptl) {
319 if (!pmd_present(*pmd))
320 goto out;
321 if (is_huge_zero_pmd(*pmd))
322 goto out;
323 page = pmd_page(*pmd);
324 if (vma->vm_flags & VM_LOCKED)
325 mlock_folio(page_folio(page));
326 else
327 munlock_page(page);
328 goto out;
329 }
330
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))
334 continue;
335 page = vm_normal_page(vma, addr, *pte);
336 if (!page || is_zone_device_page(page))
337 continue;
338 if (PageTransCompound(page))
339 continue;
340 if (vma->vm_flags & VM_LOCKED)
341 mlock_folio(page_folio(page));
342 else
343 munlock_page(page);
344 }
345 pte_unmap(start_pte);
346out:
347 spin_unlock(ptl);
348 cond_resched();
349 return 0;
350}
351
352/*
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.
359 *
360 * Called for mlock(), mlock2() and mlockall(), to set @vma VM_LOCKED;
361 * called for munlock() and munlockall(), to clear VM_LOCKED from @vma.
362 */
363static void mlock_vma_pages_range(struct vm_area_struct *vma,
364 unsigned long start, unsigned long end, vm_flags_t newflags)
365{
366 static const struct mm_walk_ops mlock_walk_ops = {
367 .pmd_entry = mlock_pte_range,
368 };
369
370 /*
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.
380 */
381 if (newflags & VM_LOCKED)
382 newflags |= VM_IO;
383 WRITE_ONCE(vma->vm_flags, newflags);
384
385 lru_add_drain();
386 walk_page_range(vma->vm_mm, start, end, &mlock_walk_ops, NULL);
387 lru_add_drain();
388
389 if (newflags & VM_IO) {
390 newflags &= ~VM_IO;
391 WRITE_ONCE(vma->vm_flags, newflags);
392 }
393}
394
395/*
396 * mlock_fixup - handle mlock[all]/munlock[all] requests.
397 *
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
400 * populate the ptes.
401 *
402 * For vmas that pass the filters, merge/split as appropriate.
403 */
404static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
405 unsigned long start, unsigned long end, vm_flags_t newflags)
406{
407 struct mm_struct *mm = vma->vm_mm;
408 pgoff_t pgoff;
409 int nr_pages;
410 int ret = 0;
411 vm_flags_t oldflags = vma->vm_flags;
412
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 */
417 goto out;
418
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));
423 if (*prev) {
424 vma = *prev;
425 goto success;
426 }
427
428 if (start != vma->vm_start) {
429 ret = split_vma(mm, vma, start, 1);
430 if (ret)
431 goto out;
432 }
433
434 if (end != vma->vm_end) {
435 ret = split_vma(mm, vma, end, 0);
436 if (ret)
437 goto out;
438 }
439
440success:
441 /*
442 * Keep track of amount of locked VM.
443 */
444 nr_pages = (end - start) >> PAGE_SHIFT;
445 if (!(newflags & VM_LOCKED))
446 nr_pages = -nr_pages;
447 else if (oldflags & VM_LOCKED)
448 nr_pages = 0;
449 mm->locked_vm += nr_pages;
450
451 /*
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.
455 */
456
457 if ((newflags & VM_LOCKED) && (oldflags & VM_LOCKED)) {
458 /* No work to do, and mlocking twice would be wrong */
459 vma->vm_flags = newflags;
460 } else {
461 mlock_vma_pages_range(vma, start, end, newflags);
462 }
463out:
464 *prev = vma;
465 return ret;
466}
467
468static int apply_vma_lock_flags(unsigned long start, size_t len,
469 vm_flags_t flags)
470{
471 unsigned long nstart, end, tmp;
472 struct vm_area_struct *vma, *prev;
473 int error;
474 MA_STATE(mas, ¤t->mm->mm_mt, start, start);
475
476 VM_BUG_ON(offset_in_page(start));
477 VM_BUG_ON(len != PAGE_ALIGN(len));
478 end = start + len;
479 if (end < start)
480 return -EINVAL;
481 if (end == start)
482 return 0;
483 vma = mas_walk(&mas);
484 if (!vma)
485 return -ENOMEM;
486
487 if (start > vma->vm_start)
488 prev = vma;
489 else
490 prev = mas_prev(&mas, 0);
491
492 for (nstart = start ; ; ) {
493 vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
494
495 newflags |= flags;
496
497 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
498 tmp = vma->vm_end;
499 if (tmp > end)
500 tmp = end;
501 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
502 if (error)
503 break;
504 nstart = tmp;
505 if (nstart < prev->vm_end)
506 nstart = prev->vm_end;
507 if (nstart >= end)
508 break;
509
510 vma = find_vma(prev->vm_mm, prev->vm_end);
511 if (!vma || vma->vm_start != nstart) {
512 error = -ENOMEM;
513 break;
514 }
515 }
516 return error;
517}
518
519/*
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)
523 * is also counted.
524 * Return value: previously mlocked page counts
525 */
526static unsigned long count_mm_mlocked_page_nr(struct mm_struct *mm,
527 unsigned long start, size_t len)
528{
529 struct vm_area_struct *vma;
530 unsigned long count = 0;
531 unsigned long end;
532 VMA_ITERATOR(vmi, mm, start);
533
534 /* Don't overflow past ULONG_MAX */
535 if (unlikely(ULONG_MAX - len < start))
536 end = ULONG_MAX;
537 else
538 end = start + len;
539
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;
546 break;
547 }
548 count += vma->vm_end - vma->vm_start;
549 }
550 }
551
552 return count >> PAGE_SHIFT;
553}
554
555/*
556 * convert get_user_pages() return value to posix mlock() error
557 */
558static int __mlock_posix_error_return(long retval)
559{
560 if (retval == -EFAULT)
561 retval = -ENOMEM;
562 else if (retval == -ENOMEM)
563 retval = -EAGAIN;
564 return retval;
565}
566
567static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
568{
569 unsigned long locked;
570 unsigned long lock_limit;
571 int error = -ENOMEM;
572
573 start = untagged_addr(start);
574
575 if (!can_do_mlock())
576 return -EPERM;
577
578 len = PAGE_ALIGN(len + (offset_in_page(start)));
579 start &= PAGE_MASK;
580
581 lock_limit = rlimit(RLIMIT_MEMLOCK);
582 lock_limit >>= PAGE_SHIFT;
583 locked = len >> PAGE_SHIFT;
584
585 if (mmap_write_lock_killable(current->mm))
586 return -EINTR;
587
588 locked += current->mm->locked_vm;
589 if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
590 /*
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.
595 */
596 locked -= count_mm_mlocked_page_nr(current->mm,
597 start, len);
598 }
599
600 /* check against resource limits */
601 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
602 error = apply_vma_lock_flags(start, len, flags);
603
604 mmap_write_unlock(current->mm);
605 if (error)
606 return error;
607
608 error = __mm_populate(start, len, 0);
609 if (error)
610 return __mlock_posix_error_return(error);
611 return 0;
612}
613
614SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
615{
616 return do_mlock(start, len, VM_LOCKED);
617}
618
619SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
620{
621 vm_flags_t vm_flags = VM_LOCKED;
622
623 if (flags & ~MLOCK_ONFAULT)
624 return -EINVAL;
625
626 if (flags & MLOCK_ONFAULT)
627 vm_flags |= VM_LOCKONFAULT;
628
629 return do_mlock(start, len, vm_flags);
630}
631
632SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
633{
634 int ret;
635
636 start = untagged_addr(start);
637
638 len = PAGE_ALIGN(len + (offset_in_page(start)));
639 start &= PAGE_MASK;
640
641 if (mmap_write_lock_killable(current->mm))
642 return -EINTR;
643 ret = apply_vma_lock_flags(start, len, 0);
644 mmap_write_unlock(current->mm);
645
646 return ret;
647}
648
649/*
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.
653 *
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.
658 */
659static int apply_mlockall_flags(int flags)
660{
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;
664
665 current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
666 if (flags & MCL_FUTURE) {
667 current->mm->def_flags |= VM_LOCKED;
668
669 if (flags & MCL_ONFAULT)
670 current->mm->def_flags |= VM_LOCKONFAULT;
671
672 if (!(flags & MCL_CURRENT))
673 goto out;
674 }
675
676 if (flags & MCL_CURRENT) {
677 to_add |= VM_LOCKED;
678 if (flags & MCL_ONFAULT)
679 to_add |= VM_LOCKONFAULT;
680 }
681
682 mas_for_each(&mas, vma, ULONG_MAX) {
683 vm_flags_t newflags;
684
685 newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
686 newflags |= to_add;
687
688 /* Ignore errors */
689 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
690 mas_pause(&mas);
691 cond_resched();
692 }
693out:
694 return 0;
695}
696
697SYSCALL_DEFINE1(mlockall, int, flags)
698{
699 unsigned long lock_limit;
700 int ret;
701
702 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)) ||
703 flags == MCL_ONFAULT)
704 return -EINVAL;
705
706 if (!can_do_mlock())
707 return -EPERM;
708
709 lock_limit = rlimit(RLIMIT_MEMLOCK);
710 lock_limit >>= PAGE_SHIFT;
711
712 if (mmap_write_lock_killable(current->mm))
713 return -EINTR;
714
715 ret = -ENOMEM;
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);
722
723 return ret;
724}
725
726SYSCALL_DEFINE0(munlockall)
727{
728 int ret;
729
730 if (mmap_write_lock_killable(current->mm))
731 return -EINTR;
732 ret = apply_mlockall_flags(0);
733 mmap_write_unlock(current->mm);
734 return ret;
735}
736
737/*
738 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
739 * shm segments) get accounted against the user_struct instead.
740 */
741static DEFINE_SPINLOCK(shmlock_user_lock);
742
743int user_shm_lock(size_t size, struct ucounts *ucounts)
744{
745 unsigned long lock_limit, locked;
746 long memlock;
747 int allowed = 0;
748
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);
755
756 if ((memlock == LONG_MAX || memlock > lock_limit) && !capable(CAP_IPC_LOCK)) {
757 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
758 goto out;
759 }
760 if (!get_ucounts(ucounts)) {
761 dec_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_MEMLOCK, locked);
762 allowed = 0;
763 goto out;
764 }
765 allowed = 1;
766out:
767 spin_unlock(&shmlock_user_lock);
768 return allowed;
769}
770
771void user_shm_unlock(size_t size, struct ucounts *ucounts)
772{
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);
777}
1/*
2 * linux/mm/mlock.c
3 *
4 * (C) Copyright 1995 Linus Torvalds
5 * (C) Copyright 2002 Christoph Hellwig
6 */
7
8#include <linux/capability.h>
9#include <linux/mman.h>
10#include <linux/mm.h>
11#include <linux/swap.h>
12#include <linux/swapops.h>
13#include <linux/pagemap.h>
14#include <linux/mempolicy.h>
15#include <linux/syscalls.h>
16#include <linux/sched.h>
17#include <linux/export.h>
18#include <linux/rmap.h>
19#include <linux/mmzone.h>
20#include <linux/hugetlb.h>
21
22#include "internal.h"
23
24int can_do_mlock(void)
25{
26 if (capable(CAP_IPC_LOCK))
27 return 1;
28 if (rlimit(RLIMIT_MEMLOCK) != 0)
29 return 1;
30 return 0;
31}
32EXPORT_SYMBOL(can_do_mlock);
33
34/*
35 * Mlocked pages are marked with PageMlocked() flag for efficient testing
36 * in vmscan and, possibly, the fault path; and to support semi-accurate
37 * statistics.
38 *
39 * An mlocked page [PageMlocked(page)] is unevictable. As such, it will
40 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
41 * The unevictable list is an LRU sibling list to the [in]active lists.
42 * PageUnevictable is set to indicate the unevictable state.
43 *
44 * When lazy mlocking via vmscan, it is important to ensure that the
45 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
46 * may have mlocked a page that is being munlocked. So lazy mlock must take
47 * the mmap_sem for read, and verify that the vma really is locked
48 * (see mm/rmap.c).
49 */
50
51/*
52 * LRU accounting for clear_page_mlock()
53 */
54void __clear_page_mlock(struct page *page)
55{
56 VM_BUG_ON(!PageLocked(page));
57
58 if (!page->mapping) { /* truncated ? */
59 return;
60 }
61
62 dec_zone_page_state(page, NR_MLOCK);
63 count_vm_event(UNEVICTABLE_PGCLEARED);
64 if (!isolate_lru_page(page)) {
65 putback_lru_page(page);
66 } else {
67 /*
68 * We lost the race. the page already moved to evictable list.
69 */
70 if (PageUnevictable(page))
71 count_vm_event(UNEVICTABLE_PGSTRANDED);
72 }
73}
74
75/*
76 * Mark page as mlocked if not already.
77 * If page on LRU, isolate and putback to move to unevictable list.
78 */
79void mlock_vma_page(struct page *page)
80{
81 BUG_ON(!PageLocked(page));
82
83 if (!TestSetPageMlocked(page)) {
84 inc_zone_page_state(page, NR_MLOCK);
85 count_vm_event(UNEVICTABLE_PGMLOCKED);
86 if (!isolate_lru_page(page))
87 putback_lru_page(page);
88 }
89}
90
91/**
92 * munlock_vma_page - munlock a vma page
93 * @page - page to be unlocked
94 *
95 * called from munlock()/munmap() path with page supposedly on the LRU.
96 * When we munlock a page, because the vma where we found the page is being
97 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
98 * page locked so that we can leave it on the unevictable lru list and not
99 * bother vmscan with it. However, to walk the page's rmap list in
100 * try_to_munlock() we must isolate the page from the LRU. If some other
101 * task has removed the page from the LRU, we won't be able to do that.
102 * So we clear the PageMlocked as we might not get another chance. If we
103 * can't isolate the page, we leave it for putback_lru_page() and vmscan
104 * [page_referenced()/try_to_unmap()] to deal with.
105 */
106void munlock_vma_page(struct page *page)
107{
108 BUG_ON(!PageLocked(page));
109
110 if (TestClearPageMlocked(page)) {
111 dec_zone_page_state(page, NR_MLOCK);
112 if (!isolate_lru_page(page)) {
113 int ret = SWAP_AGAIN;
114
115 /*
116 * Optimization: if the page was mapped just once,
117 * that's our mapping and we don't need to check all the
118 * other vmas.
119 */
120 if (page_mapcount(page) > 1)
121 ret = try_to_munlock(page);
122 /*
123 * did try_to_unlock() succeed or punt?
124 */
125 if (ret != SWAP_MLOCK)
126 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
127
128 putback_lru_page(page);
129 } else {
130 /*
131 * Some other task has removed the page from the LRU.
132 * putback_lru_page() will take care of removing the
133 * page from the unevictable list, if necessary.
134 * vmscan [page_referenced()] will move the page back
135 * to the unevictable list if some other vma has it
136 * mlocked.
137 */
138 if (PageUnevictable(page))
139 count_vm_event(UNEVICTABLE_PGSTRANDED);
140 else
141 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
142 }
143 }
144}
145
146/**
147 * __mlock_vma_pages_range() - mlock a range of pages in the vma.
148 * @vma: target vma
149 * @start: start address
150 * @end: end address
151 *
152 * This takes care of making the pages present too.
153 *
154 * return 0 on success, negative error code on error.
155 *
156 * vma->vm_mm->mmap_sem must be held for at least read.
157 */
158static long __mlock_vma_pages_range(struct vm_area_struct *vma,
159 unsigned long start, unsigned long end,
160 int *nonblocking)
161{
162 struct mm_struct *mm = vma->vm_mm;
163 unsigned long addr = start;
164 int nr_pages = (end - start) / PAGE_SIZE;
165 int gup_flags;
166
167 VM_BUG_ON(start & ~PAGE_MASK);
168 VM_BUG_ON(end & ~PAGE_MASK);
169 VM_BUG_ON(start < vma->vm_start);
170 VM_BUG_ON(end > vma->vm_end);
171 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
172
173 gup_flags = FOLL_TOUCH | FOLL_MLOCK;
174 /*
175 * We want to touch writable mappings with a write fault in order
176 * to break COW, except for shared mappings because these don't COW
177 * and we would not want to dirty them for nothing.
178 */
179 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
180 gup_flags |= FOLL_WRITE;
181
182 /*
183 * We want mlock to succeed for regions that have any permissions
184 * other than PROT_NONE.
185 */
186 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
187 gup_flags |= FOLL_FORCE;
188
189 return __get_user_pages(current, mm, addr, nr_pages, gup_flags,
190 NULL, NULL, nonblocking);
191}
192
193/*
194 * convert get_user_pages() return value to posix mlock() error
195 */
196static int __mlock_posix_error_return(long retval)
197{
198 if (retval == -EFAULT)
199 retval = -ENOMEM;
200 else if (retval == -ENOMEM)
201 retval = -EAGAIN;
202 return retval;
203}
204
205/**
206 * mlock_vma_pages_range() - mlock pages in specified vma range.
207 * @vma - the vma containing the specfied address range
208 * @start - starting address in @vma to mlock
209 * @end - end address [+1] in @vma to mlock
210 *
211 * For mmap()/mremap()/expansion of mlocked vma.
212 *
213 * return 0 on success for "normal" vmas.
214 *
215 * return number of pages [> 0] to be removed from locked_vm on success
216 * of "special" vmas.
217 */
218long mlock_vma_pages_range(struct vm_area_struct *vma,
219 unsigned long start, unsigned long end)
220{
221 int nr_pages = (end - start) / PAGE_SIZE;
222 BUG_ON(!(vma->vm_flags & VM_LOCKED));
223
224 /*
225 * filter unlockable vmas
226 */
227 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
228 goto no_mlock;
229
230 if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
231 is_vm_hugetlb_page(vma) ||
232 vma == get_gate_vma(current->mm))) {
233
234 __mlock_vma_pages_range(vma, start, end, NULL);
235
236 /* Hide errors from mmap() and other callers */
237 return 0;
238 }
239
240 /*
241 * User mapped kernel pages or huge pages:
242 * make these pages present to populate the ptes, but
243 * fall thru' to reset VM_LOCKED--no need to unlock, and
244 * return nr_pages so these don't get counted against task's
245 * locked limit. huge pages are already counted against
246 * locked vm limit.
247 */
248 make_pages_present(start, end);
249
250no_mlock:
251 vma->vm_flags &= ~VM_LOCKED; /* and don't come back! */
252 return nr_pages; /* error or pages NOT mlocked */
253}
254
255/*
256 * munlock_vma_pages_range() - munlock all pages in the vma range.'
257 * @vma - vma containing range to be munlock()ed.
258 * @start - start address in @vma of the range
259 * @end - end of range in @vma.
260 *
261 * For mremap(), munmap() and exit().
262 *
263 * Called with @vma VM_LOCKED.
264 *
265 * Returns with VM_LOCKED cleared. Callers must be prepared to
266 * deal with this.
267 *
268 * We don't save and restore VM_LOCKED here because pages are
269 * still on lru. In unmap path, pages might be scanned by reclaim
270 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
271 * free them. This will result in freeing mlocked pages.
272 */
273void munlock_vma_pages_range(struct vm_area_struct *vma,
274 unsigned long start, unsigned long end)
275{
276 unsigned long addr;
277
278 lru_add_drain();
279 vma->vm_flags &= ~VM_LOCKED;
280
281 for (addr = start; addr < end; addr += PAGE_SIZE) {
282 struct page *page;
283 /*
284 * Although FOLL_DUMP is intended for get_dump_page(),
285 * it just so happens that its special treatment of the
286 * ZERO_PAGE (returning an error instead of doing get_page)
287 * suits munlock very well (and if somehow an abnormal page
288 * has sneaked into the range, we won't oops here: great).
289 */
290 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
291 if (page && !IS_ERR(page)) {
292 lock_page(page);
293 /*
294 * Like in __mlock_vma_pages_range(),
295 * because we lock page here and migration is
296 * blocked by the elevated reference, we need
297 * only check for file-cache page truncation.
298 */
299 if (page->mapping)
300 munlock_vma_page(page);
301 unlock_page(page);
302 put_page(page);
303 }
304 cond_resched();
305 }
306}
307
308/*
309 * mlock_fixup - handle mlock[all]/munlock[all] requests.
310 *
311 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
312 * munlock is a no-op. However, for some special vmas, we go ahead and
313 * populate the ptes via make_pages_present().
314 *
315 * For vmas that pass the filters, merge/split as appropriate.
316 */
317static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
318 unsigned long start, unsigned long end, vm_flags_t newflags)
319{
320 struct mm_struct *mm = vma->vm_mm;
321 pgoff_t pgoff;
322 int nr_pages;
323 int ret = 0;
324 int lock = !!(newflags & VM_LOCKED);
325
326 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
327 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
328 goto out; /* don't set VM_LOCKED, don't count */
329
330 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
331 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
332 vma->vm_file, pgoff, vma_policy(vma));
333 if (*prev) {
334 vma = *prev;
335 goto success;
336 }
337
338 if (start != vma->vm_start) {
339 ret = split_vma(mm, vma, start, 1);
340 if (ret)
341 goto out;
342 }
343
344 if (end != vma->vm_end) {
345 ret = split_vma(mm, vma, end, 0);
346 if (ret)
347 goto out;
348 }
349
350success:
351 /*
352 * Keep track of amount of locked VM.
353 */
354 nr_pages = (end - start) >> PAGE_SHIFT;
355 if (!lock)
356 nr_pages = -nr_pages;
357 mm->locked_vm += nr_pages;
358
359 /*
360 * vm_flags is protected by the mmap_sem held in write mode.
361 * It's okay if try_to_unmap_one unmaps a page just after we
362 * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
363 */
364
365 if (lock)
366 vma->vm_flags = newflags;
367 else
368 munlock_vma_pages_range(vma, start, end);
369
370out:
371 *prev = vma;
372 return ret;
373}
374
375static int do_mlock(unsigned long start, size_t len, int on)
376{
377 unsigned long nstart, end, tmp;
378 struct vm_area_struct * vma, * prev;
379 int error;
380
381 VM_BUG_ON(start & ~PAGE_MASK);
382 VM_BUG_ON(len != PAGE_ALIGN(len));
383 end = start + len;
384 if (end < start)
385 return -EINVAL;
386 if (end == start)
387 return 0;
388 vma = find_vma(current->mm, start);
389 if (!vma || vma->vm_start > start)
390 return -ENOMEM;
391
392 prev = vma->vm_prev;
393 if (start > vma->vm_start)
394 prev = vma;
395
396 for (nstart = start ; ; ) {
397 vm_flags_t newflags;
398
399 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
400
401 newflags = vma->vm_flags | VM_LOCKED;
402 if (!on)
403 newflags &= ~VM_LOCKED;
404
405 tmp = vma->vm_end;
406 if (tmp > end)
407 tmp = end;
408 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
409 if (error)
410 break;
411 nstart = tmp;
412 if (nstart < prev->vm_end)
413 nstart = prev->vm_end;
414 if (nstart >= end)
415 break;
416
417 vma = prev->vm_next;
418 if (!vma || vma->vm_start != nstart) {
419 error = -ENOMEM;
420 break;
421 }
422 }
423 return error;
424}
425
426static int do_mlock_pages(unsigned long start, size_t len, int ignore_errors)
427{
428 struct mm_struct *mm = current->mm;
429 unsigned long end, nstart, nend;
430 struct vm_area_struct *vma = NULL;
431 int locked = 0;
432 int ret = 0;
433
434 VM_BUG_ON(start & ~PAGE_MASK);
435 VM_BUG_ON(len != PAGE_ALIGN(len));
436 end = start + len;
437
438 for (nstart = start; nstart < end; nstart = nend) {
439 /*
440 * We want to fault in pages for [nstart; end) address range.
441 * Find first corresponding VMA.
442 */
443 if (!locked) {
444 locked = 1;
445 down_read(&mm->mmap_sem);
446 vma = find_vma(mm, nstart);
447 } else if (nstart >= vma->vm_end)
448 vma = vma->vm_next;
449 if (!vma || vma->vm_start >= end)
450 break;
451 /*
452 * Set [nstart; nend) to intersection of desired address
453 * range with the first VMA. Also, skip undesirable VMA types.
454 */
455 nend = min(end, vma->vm_end);
456 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
457 continue;
458 if (nstart < vma->vm_start)
459 nstart = vma->vm_start;
460 /*
461 * Now fault in a range of pages. __mlock_vma_pages_range()
462 * double checks the vma flags, so that it won't mlock pages
463 * if the vma was already munlocked.
464 */
465 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
466 if (ret < 0) {
467 if (ignore_errors) {
468 ret = 0;
469 continue; /* continue at next VMA */
470 }
471 ret = __mlock_posix_error_return(ret);
472 break;
473 }
474 nend = nstart + ret * PAGE_SIZE;
475 ret = 0;
476 }
477 if (locked)
478 up_read(&mm->mmap_sem);
479 return ret; /* 0 or negative error code */
480}
481
482SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
483{
484 unsigned long locked;
485 unsigned long lock_limit;
486 int error = -ENOMEM;
487
488 if (!can_do_mlock())
489 return -EPERM;
490
491 lru_add_drain_all(); /* flush pagevec */
492
493 down_write(¤t->mm->mmap_sem);
494 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
495 start &= PAGE_MASK;
496
497 locked = len >> PAGE_SHIFT;
498 locked += current->mm->locked_vm;
499
500 lock_limit = rlimit(RLIMIT_MEMLOCK);
501 lock_limit >>= PAGE_SHIFT;
502
503 /* check against resource limits */
504 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
505 error = do_mlock(start, len, 1);
506 up_write(¤t->mm->mmap_sem);
507 if (!error)
508 error = do_mlock_pages(start, len, 0);
509 return error;
510}
511
512SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
513{
514 int ret;
515
516 down_write(¤t->mm->mmap_sem);
517 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
518 start &= PAGE_MASK;
519 ret = do_mlock(start, len, 0);
520 up_write(¤t->mm->mmap_sem);
521 return ret;
522}
523
524static int do_mlockall(int flags)
525{
526 struct vm_area_struct * vma, * prev = NULL;
527 unsigned int def_flags = 0;
528
529 if (flags & MCL_FUTURE)
530 def_flags = VM_LOCKED;
531 current->mm->def_flags = def_flags;
532 if (flags == MCL_FUTURE)
533 goto out;
534
535 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
536 vm_flags_t newflags;
537
538 newflags = vma->vm_flags | VM_LOCKED;
539 if (!(flags & MCL_CURRENT))
540 newflags &= ~VM_LOCKED;
541
542 /* Ignore errors */
543 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
544 }
545out:
546 return 0;
547}
548
549SYSCALL_DEFINE1(mlockall, int, flags)
550{
551 unsigned long lock_limit;
552 int ret = -EINVAL;
553
554 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
555 goto out;
556
557 ret = -EPERM;
558 if (!can_do_mlock())
559 goto out;
560
561 if (flags & MCL_CURRENT)
562 lru_add_drain_all(); /* flush pagevec */
563
564 down_write(¤t->mm->mmap_sem);
565
566 lock_limit = rlimit(RLIMIT_MEMLOCK);
567 lock_limit >>= PAGE_SHIFT;
568
569 ret = -ENOMEM;
570 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
571 capable(CAP_IPC_LOCK))
572 ret = do_mlockall(flags);
573 up_write(¤t->mm->mmap_sem);
574 if (!ret && (flags & MCL_CURRENT)) {
575 /* Ignore errors */
576 do_mlock_pages(0, TASK_SIZE, 1);
577 }
578out:
579 return ret;
580}
581
582SYSCALL_DEFINE0(munlockall)
583{
584 int ret;
585
586 down_write(¤t->mm->mmap_sem);
587 ret = do_mlockall(0);
588 up_write(¤t->mm->mmap_sem);
589 return ret;
590}
591
592/*
593 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
594 * shm segments) get accounted against the user_struct instead.
595 */
596static DEFINE_SPINLOCK(shmlock_user_lock);
597
598int user_shm_lock(size_t size, struct user_struct *user)
599{
600 unsigned long lock_limit, locked;
601 int allowed = 0;
602
603 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
604 lock_limit = rlimit(RLIMIT_MEMLOCK);
605 if (lock_limit == RLIM_INFINITY)
606 allowed = 1;
607 lock_limit >>= PAGE_SHIFT;
608 spin_lock(&shmlock_user_lock);
609 if (!allowed &&
610 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
611 goto out;
612 get_uid(user);
613 user->locked_shm += locked;
614 allowed = 1;
615out:
616 spin_unlock(&shmlock_user_lock);
617 return allowed;
618}
619
620void user_shm_unlock(size_t size, struct user_struct *user)
621{
622 spin_lock(&shmlock_user_lock);
623 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
624 spin_unlock(&shmlock_user_lock);
625 free_uid(user);
626}