Loading...
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/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>
24
25#include "internal.h"
26
27bool can_do_mlock(void)
28{
29 if (rlimit(RLIMIT_MEMLOCK) != 0)
30 return true;
31 if (capable(CAP_IPC_LOCK))
32 return true;
33 return false;
34}
35EXPORT_SYMBOL(can_do_mlock);
36
37/*
38 * Mlocked pages are marked with PageMlocked() flag for efficient testing
39 * in vmscan and, possibly, the fault path; and to support semi-accurate
40 * statistics.
41 *
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.
46 *
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
51 * (see mm/rmap.c).
52 */
53
54/*
55 * LRU accounting for clear_page_mlock()
56 */
57void clear_page_mlock(struct page *page)
58{
59 if (!TestClearPageMlocked(page))
60 return;
61
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);
67 } else {
68 /*
69 * We lost the race. the page already moved to evictable list.
70 */
71 if (PageUnevictable(page))
72 count_vm_event(UNEVICTABLE_PGSTRANDED);
73 }
74}
75
76/*
77 * Mark page as mlocked if not already.
78 * If page on LRU, isolate and putback to move to unevictable list.
79 */
80void mlock_vma_page(struct page *page)
81{
82 /* Serialize with page migration */
83 BUG_ON(!PageLocked(page));
84
85 VM_BUG_ON_PAGE(PageTail(page), page);
86 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
87
88 if (!TestSetPageMlocked(page)) {
89 mod_zone_page_state(page_zone(page), NR_MLOCK,
90 hpage_nr_pages(page));
91 count_vm_event(UNEVICTABLE_PGMLOCKED);
92 if (!isolate_lru_page(page))
93 putback_lru_page(page);
94 }
95}
96
97/*
98 * Isolate a page from LRU with optional get_page() pin.
99 * Assumes lru_lock already held and page already pinned.
100 */
101static bool __munlock_isolate_lru_page(struct page *page, bool getpage)
102{
103 if (PageLRU(page)) {
104 struct lruvec *lruvec;
105
106 lruvec = mem_cgroup_page_lruvec(page, page_zone(page));
107 if (getpage)
108 get_page(page);
109 ClearPageLRU(page);
110 del_page_from_lru_list(page, lruvec, page_lru(page));
111 return true;
112 }
113
114 return false;
115}
116
117/*
118 * Finish munlock after successful page isolation
119 *
120 * Page must be locked. This is a wrapper for try_to_munlock()
121 * and putback_lru_page() with munlock accounting.
122 */
123static void __munlock_isolated_page(struct page *page)
124{
125 int ret = SWAP_AGAIN;
126
127 /*
128 * Optimization: if the page was mapped just once, that's our mapping
129 * and we don't need to check all the other vmas.
130 */
131 if (page_mapcount(page) > 1)
132 ret = try_to_munlock(page);
133
134 /* Did try_to_unlock() succeed or punt? */
135 if (ret != SWAP_MLOCK)
136 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
137
138 putback_lru_page(page);
139}
140
141/*
142 * Accounting for page isolation fail during munlock
143 *
144 * Performs accounting when page isolation fails in munlock. There is nothing
145 * else to do because it means some other task has already removed the page
146 * from the LRU. putback_lru_page() will take care of removing the page from
147 * the unevictable list, if necessary. vmscan [page_referenced()] will move
148 * the page back to the unevictable list if some other vma has it mlocked.
149 */
150static void __munlock_isolation_failed(struct page *page)
151{
152 if (PageUnevictable(page))
153 __count_vm_event(UNEVICTABLE_PGSTRANDED);
154 else
155 __count_vm_event(UNEVICTABLE_PGMUNLOCKED);
156}
157
158/**
159 * munlock_vma_page - munlock a vma page
160 * @page - page to be unlocked, either a normal page or THP page head
161 *
162 * returns the size of the page as a page mask (0 for normal page,
163 * HPAGE_PMD_NR - 1 for THP head page)
164 *
165 * called from munlock()/munmap() path with page supposedly on the LRU.
166 * When we munlock a page, because the vma where we found the page is being
167 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
168 * page locked so that we can leave it on the unevictable lru list and not
169 * bother vmscan with it. However, to walk the page's rmap list in
170 * try_to_munlock() we must isolate the page from the LRU. If some other
171 * task has removed the page from the LRU, we won't be able to do that.
172 * So we clear the PageMlocked as we might not get another chance. If we
173 * can't isolate the page, we leave it for putback_lru_page() and vmscan
174 * [page_referenced()/try_to_unmap()] to deal with.
175 */
176unsigned int munlock_vma_page(struct page *page)
177{
178 int nr_pages;
179 struct zone *zone = page_zone(page);
180
181 /* For try_to_munlock() and to serialize with page migration */
182 BUG_ON(!PageLocked(page));
183
184 VM_BUG_ON_PAGE(PageTail(page), page);
185
186 /*
187 * Serialize with any parallel __split_huge_page_refcount() which
188 * might otherwise copy PageMlocked to part of the tail pages before
189 * we clear it in the head page. It also stabilizes hpage_nr_pages().
190 */
191 spin_lock_irq(&zone->lru_lock);
192
193 nr_pages = hpage_nr_pages(page);
194 if (!TestClearPageMlocked(page))
195 goto unlock_out;
196
197 __mod_zone_page_state(zone, NR_MLOCK, -nr_pages);
198
199 if (__munlock_isolate_lru_page(page, true)) {
200 spin_unlock_irq(&zone->lru_lock);
201 __munlock_isolated_page(page);
202 goto out;
203 }
204 __munlock_isolation_failed(page);
205
206unlock_out:
207 spin_unlock_irq(&zone->lru_lock);
208
209out:
210 return nr_pages - 1;
211}
212
213/*
214 * convert get_user_pages() return value to posix mlock() error
215 */
216static int __mlock_posix_error_return(long retval)
217{
218 if (retval == -EFAULT)
219 retval = -ENOMEM;
220 else if (retval == -ENOMEM)
221 retval = -EAGAIN;
222 return retval;
223}
224
225/*
226 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
227 *
228 * The fast path is available only for evictable pages with single mapping.
229 * Then we can bypass the per-cpu pvec and get better performance.
230 * when mapcount > 1 we need try_to_munlock() which can fail.
231 * when !page_evictable(), we need the full redo logic of putback_lru_page to
232 * avoid leaving evictable page in unevictable list.
233 *
234 * In case of success, @page is added to @pvec and @pgrescued is incremented
235 * in case that the page was previously unevictable. @page is also unlocked.
236 */
237static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
238 int *pgrescued)
239{
240 VM_BUG_ON_PAGE(PageLRU(page), page);
241 VM_BUG_ON_PAGE(!PageLocked(page), page);
242
243 if (page_mapcount(page) <= 1 && page_evictable(page)) {
244 pagevec_add(pvec, page);
245 if (TestClearPageUnevictable(page))
246 (*pgrescued)++;
247 unlock_page(page);
248 return true;
249 }
250
251 return false;
252}
253
254/*
255 * Putback multiple evictable pages to the LRU
256 *
257 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
258 * the pages might have meanwhile become unevictable but that is OK.
259 */
260static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
261{
262 count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
263 /*
264 *__pagevec_lru_add() calls release_pages() so we don't call
265 * put_page() explicitly
266 */
267 __pagevec_lru_add(pvec);
268 count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
269}
270
271/*
272 * Munlock a batch of pages from the same zone
273 *
274 * The work is split to two main phases. First phase clears the Mlocked flag
275 * and attempts to isolate the pages, all under a single zone lru lock.
276 * The second phase finishes the munlock only for pages where isolation
277 * succeeded.
278 *
279 * Note that the pagevec may be modified during the process.
280 */
281static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
282{
283 int i;
284 int nr = pagevec_count(pvec);
285 int delta_munlocked;
286 struct pagevec pvec_putback;
287 int pgrescued = 0;
288
289 pagevec_init(&pvec_putback, 0);
290
291 /* Phase 1: page isolation */
292 spin_lock_irq(&zone->lru_lock);
293 for (i = 0; i < nr; i++) {
294 struct page *page = pvec->pages[i];
295
296 if (TestClearPageMlocked(page)) {
297 /*
298 * We already have pin from follow_page_mask()
299 * so we can spare the get_page() here.
300 */
301 if (__munlock_isolate_lru_page(page, false))
302 continue;
303 else
304 __munlock_isolation_failed(page);
305 }
306
307 /*
308 * We won't be munlocking this page in the next phase
309 * but we still need to release the follow_page_mask()
310 * pin. We cannot do it under lru_lock however. If it's
311 * the last pin, __page_cache_release() would deadlock.
312 */
313 pagevec_add(&pvec_putback, pvec->pages[i]);
314 pvec->pages[i] = NULL;
315 }
316 delta_munlocked = -nr + pagevec_count(&pvec_putback);
317 __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
318 spin_unlock_irq(&zone->lru_lock);
319
320 /* Now we can release pins of pages that we are not munlocking */
321 pagevec_release(&pvec_putback);
322
323 /* Phase 2: page munlock */
324 for (i = 0; i < nr; i++) {
325 struct page *page = pvec->pages[i];
326
327 if (page) {
328 lock_page(page);
329 if (!__putback_lru_fast_prepare(page, &pvec_putback,
330 &pgrescued)) {
331 /*
332 * Slow path. We don't want to lose the last
333 * pin before unlock_page()
334 */
335 get_page(page); /* for putback_lru_page() */
336 __munlock_isolated_page(page);
337 unlock_page(page);
338 put_page(page); /* from follow_page_mask() */
339 }
340 }
341 }
342
343 /*
344 * Phase 3: page putback for pages that qualified for the fast path
345 * This will also call put_page() to return pin from follow_page_mask()
346 */
347 if (pagevec_count(&pvec_putback))
348 __putback_lru_fast(&pvec_putback, pgrescued);
349}
350
351/*
352 * Fill up pagevec for __munlock_pagevec using pte walk
353 *
354 * The function expects that the struct page corresponding to @start address is
355 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
356 *
357 * The rest of @pvec is filled by subsequent pages within the same pmd and same
358 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
359 * pages also get pinned.
360 *
361 * Returns the address of the next page that should be scanned. This equals
362 * @start + PAGE_SIZE when no page could be added by the pte walk.
363 */
364static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
365 struct vm_area_struct *vma, int zoneid, unsigned long start,
366 unsigned long end)
367{
368 pte_t *pte;
369 spinlock_t *ptl;
370
371 /*
372 * Initialize pte walk starting at the already pinned page where we
373 * are sure that there is a pte, as it was pinned under the same
374 * mmap_sem write op.
375 */
376 pte = get_locked_pte(vma->vm_mm, start, &ptl);
377 /* Make sure we do not cross the page table boundary */
378 end = pgd_addr_end(start, end);
379 end = pud_addr_end(start, end);
380 end = pmd_addr_end(start, end);
381
382 /* The page next to the pinned page is the first we will try to get */
383 start += PAGE_SIZE;
384 while (start < end) {
385 struct page *page = NULL;
386 pte++;
387 if (pte_present(*pte))
388 page = vm_normal_page(vma, start, *pte);
389 /*
390 * Break if page could not be obtained or the page's node+zone does not
391 * match
392 */
393 if (!page || page_zone_id(page) != zoneid)
394 break;
395
396 /*
397 * Do not use pagevec for PTE-mapped THP,
398 * munlock_vma_pages_range() will handle them.
399 */
400 if (PageTransCompound(page))
401 break;
402
403 get_page(page);
404 /*
405 * Increase the address that will be returned *before* the
406 * eventual break due to pvec becoming full by adding the page
407 */
408 start += PAGE_SIZE;
409 if (pagevec_add(pvec, page) == 0)
410 break;
411 }
412 pte_unmap_unlock(pte, ptl);
413 return start;
414}
415
416/*
417 * munlock_vma_pages_range() - munlock all pages in the vma range.'
418 * @vma - vma containing range to be munlock()ed.
419 * @start - start address in @vma of the range
420 * @end - end of range in @vma.
421 *
422 * For mremap(), munmap() and exit().
423 *
424 * Called with @vma VM_LOCKED.
425 *
426 * Returns with VM_LOCKED cleared. Callers must be prepared to
427 * deal with this.
428 *
429 * We don't save and restore VM_LOCKED here because pages are
430 * still on lru. In unmap path, pages might be scanned by reclaim
431 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
432 * free them. This will result in freeing mlocked pages.
433 */
434void munlock_vma_pages_range(struct vm_area_struct *vma,
435 unsigned long start, unsigned long end)
436{
437 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
438
439 while (start < end) {
440 struct page *page;
441 unsigned int page_mask;
442 unsigned long page_increm;
443 struct pagevec pvec;
444 struct zone *zone;
445 int zoneid;
446
447 pagevec_init(&pvec, 0);
448 /*
449 * Although FOLL_DUMP is intended for get_dump_page(),
450 * it just so happens that its special treatment of the
451 * ZERO_PAGE (returning an error instead of doing get_page)
452 * suits munlock very well (and if somehow an abnormal page
453 * has sneaked into the range, we won't oops here: great).
454 */
455 page = follow_page_mask(vma, start, FOLL_GET | FOLL_DUMP,
456 &page_mask);
457
458 if (page && !IS_ERR(page)) {
459 if (PageTransTail(page)) {
460 VM_BUG_ON_PAGE(PageMlocked(page), page);
461 put_page(page); /* follow_page_mask() */
462 } else if (PageTransHuge(page)) {
463 lock_page(page);
464 /*
465 * Any THP page found by follow_page_mask() may
466 * have gotten split before reaching
467 * munlock_vma_page(), so we need to recompute
468 * the page_mask here.
469 */
470 page_mask = munlock_vma_page(page);
471 unlock_page(page);
472 put_page(page); /* follow_page_mask() */
473 } else {
474 /*
475 * Non-huge pages are handled in batches via
476 * pagevec. The pin from follow_page_mask()
477 * prevents them from collapsing by THP.
478 */
479 pagevec_add(&pvec, page);
480 zone = page_zone(page);
481 zoneid = page_zone_id(page);
482
483 /*
484 * Try to fill the rest of pagevec using fast
485 * pte walk. This will also update start to
486 * the next page to process. Then munlock the
487 * pagevec.
488 */
489 start = __munlock_pagevec_fill(&pvec, vma,
490 zoneid, start, end);
491 __munlock_pagevec(&pvec, zone);
492 goto next;
493 }
494 }
495 page_increm = 1 + page_mask;
496 start += page_increm * PAGE_SIZE;
497next:
498 cond_resched();
499 }
500}
501
502/*
503 * mlock_fixup - handle mlock[all]/munlock[all] requests.
504 *
505 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
506 * munlock is a no-op. However, for some special vmas, we go ahead and
507 * populate the ptes.
508 *
509 * For vmas that pass the filters, merge/split as appropriate.
510 */
511static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
512 unsigned long start, unsigned long end, vm_flags_t newflags)
513{
514 struct mm_struct *mm = vma->vm_mm;
515 pgoff_t pgoff;
516 int nr_pages;
517 int ret = 0;
518 int lock = !!(newflags & VM_LOCKED);
519
520 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
521 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
522 /* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
523 goto out;
524
525 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
526 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
527 vma->vm_file, pgoff, vma_policy(vma),
528 vma->vm_userfaultfd_ctx);
529 if (*prev) {
530 vma = *prev;
531 goto success;
532 }
533
534 if (start != vma->vm_start) {
535 ret = split_vma(mm, vma, start, 1);
536 if (ret)
537 goto out;
538 }
539
540 if (end != vma->vm_end) {
541 ret = split_vma(mm, vma, end, 0);
542 if (ret)
543 goto out;
544 }
545
546success:
547 /*
548 * Keep track of amount of locked VM.
549 */
550 nr_pages = (end - start) >> PAGE_SHIFT;
551 if (!lock)
552 nr_pages = -nr_pages;
553 mm->locked_vm += nr_pages;
554
555 /*
556 * vm_flags is protected by the mmap_sem held in write mode.
557 * It's okay if try_to_unmap_one unmaps a page just after we
558 * set VM_LOCKED, populate_vma_page_range will bring it back.
559 */
560
561 if (lock)
562 vma->vm_flags = newflags;
563 else
564 munlock_vma_pages_range(vma, start, end);
565
566out:
567 *prev = vma;
568 return ret;
569}
570
571static int apply_vma_lock_flags(unsigned long start, size_t len,
572 vm_flags_t flags)
573{
574 unsigned long nstart, end, tmp;
575 struct vm_area_struct * vma, * prev;
576 int error;
577
578 VM_BUG_ON(offset_in_page(start));
579 VM_BUG_ON(len != PAGE_ALIGN(len));
580 end = start + len;
581 if (end < start)
582 return -EINVAL;
583 if (end == start)
584 return 0;
585 vma = find_vma(current->mm, start);
586 if (!vma || vma->vm_start > start)
587 return -ENOMEM;
588
589 prev = vma->vm_prev;
590 if (start > vma->vm_start)
591 prev = vma;
592
593 for (nstart = start ; ; ) {
594 vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
595
596 newflags |= flags;
597
598 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
599 tmp = vma->vm_end;
600 if (tmp > end)
601 tmp = end;
602 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
603 if (error)
604 break;
605 nstart = tmp;
606 if (nstart < prev->vm_end)
607 nstart = prev->vm_end;
608 if (nstart >= end)
609 break;
610
611 vma = prev->vm_next;
612 if (!vma || vma->vm_start != nstart) {
613 error = -ENOMEM;
614 break;
615 }
616 }
617 return error;
618}
619
620static int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
621{
622 unsigned long locked;
623 unsigned long lock_limit;
624 int error = -ENOMEM;
625
626 if (!can_do_mlock())
627 return -EPERM;
628
629 lru_add_drain_all(); /* flush pagevec */
630
631 len = PAGE_ALIGN(len + (offset_in_page(start)));
632 start &= PAGE_MASK;
633
634 lock_limit = rlimit(RLIMIT_MEMLOCK);
635 lock_limit >>= PAGE_SHIFT;
636 locked = len >> PAGE_SHIFT;
637
638 down_write(¤t->mm->mmap_sem);
639
640 locked += current->mm->locked_vm;
641
642 /* check against resource limits */
643 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
644 error = apply_vma_lock_flags(start, len, flags);
645
646 up_write(¤t->mm->mmap_sem);
647 if (error)
648 return error;
649
650 error = __mm_populate(start, len, 0);
651 if (error)
652 return __mlock_posix_error_return(error);
653 return 0;
654}
655
656SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
657{
658 return do_mlock(start, len, VM_LOCKED);
659}
660
661SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
662{
663 vm_flags_t vm_flags = VM_LOCKED;
664
665 if (flags & ~MLOCK_ONFAULT)
666 return -EINVAL;
667
668 if (flags & MLOCK_ONFAULT)
669 vm_flags |= VM_LOCKONFAULT;
670
671 return do_mlock(start, len, vm_flags);
672}
673
674SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
675{
676 int ret;
677
678 len = PAGE_ALIGN(len + (offset_in_page(start)));
679 start &= PAGE_MASK;
680
681 down_write(¤t->mm->mmap_sem);
682 ret = apply_vma_lock_flags(start, len, 0);
683 up_write(¤t->mm->mmap_sem);
684
685 return ret;
686}
687
688/*
689 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
690 * and translate into the appropriate modifications to mm->def_flags and/or the
691 * flags for all current VMAs.
692 *
693 * There are a couple of subtleties with this. If mlockall() is called multiple
694 * times with different flags, the values do not necessarily stack. If mlockall
695 * is called once including the MCL_FUTURE flag and then a second time without
696 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
697 */
698static int apply_mlockall_flags(int flags)
699{
700 struct vm_area_struct * vma, * prev = NULL;
701 vm_flags_t to_add = 0;
702
703 current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
704 if (flags & MCL_FUTURE) {
705 current->mm->def_flags |= VM_LOCKED;
706
707 if (flags & MCL_ONFAULT)
708 current->mm->def_flags |= VM_LOCKONFAULT;
709
710 if (!(flags & MCL_CURRENT))
711 goto out;
712 }
713
714 if (flags & MCL_CURRENT) {
715 to_add |= VM_LOCKED;
716 if (flags & MCL_ONFAULT)
717 to_add |= VM_LOCKONFAULT;
718 }
719
720 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
721 vm_flags_t newflags;
722
723 newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
724 newflags |= to_add;
725
726 /* Ignore errors */
727 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
728 cond_resched_rcu_qs();
729 }
730out:
731 return 0;
732}
733
734SYSCALL_DEFINE1(mlockall, int, flags)
735{
736 unsigned long lock_limit;
737 int ret;
738
739 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)))
740 return -EINVAL;
741
742 if (!can_do_mlock())
743 return -EPERM;
744
745 if (flags & MCL_CURRENT)
746 lru_add_drain_all(); /* flush pagevec */
747
748 lock_limit = rlimit(RLIMIT_MEMLOCK);
749 lock_limit >>= PAGE_SHIFT;
750
751 ret = -ENOMEM;
752 down_write(¤t->mm->mmap_sem);
753
754 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
755 capable(CAP_IPC_LOCK))
756 ret = apply_mlockall_flags(flags);
757 up_write(¤t->mm->mmap_sem);
758 if (!ret && (flags & MCL_CURRENT))
759 mm_populate(0, TASK_SIZE);
760
761 return ret;
762}
763
764SYSCALL_DEFINE0(munlockall)
765{
766 int ret;
767
768 down_write(¤t->mm->mmap_sem);
769 ret = apply_mlockall_flags(0);
770 up_write(¤t->mm->mmap_sem);
771 return ret;
772}
773
774/*
775 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
776 * shm segments) get accounted against the user_struct instead.
777 */
778static DEFINE_SPINLOCK(shmlock_user_lock);
779
780int user_shm_lock(size_t size, struct user_struct *user)
781{
782 unsigned long lock_limit, locked;
783 int allowed = 0;
784
785 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
786 lock_limit = rlimit(RLIMIT_MEMLOCK);
787 if (lock_limit == RLIM_INFINITY)
788 allowed = 1;
789 lock_limit >>= PAGE_SHIFT;
790 spin_lock(&shmlock_user_lock);
791 if (!allowed &&
792 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
793 goto out;
794 get_uid(user);
795 user->locked_shm += locked;
796 allowed = 1;
797out:
798 spin_unlock(&shmlock_user_lock);
799 return allowed;
800}
801
802void user_shm_unlock(size_t size, struct user_struct *user)
803{
804 spin_lock(&shmlock_user_lock);
805 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
806 spin_unlock(&shmlock_user_lock);
807 free_uid(user);
808}
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/module.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 = try_to_munlock(page);
114 /*
115 * did try_to_unlock() succeed or punt?
116 */
117 if (ret != SWAP_MLOCK)
118 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
119
120 putback_lru_page(page);
121 } else {
122 /*
123 * Some other task has removed the page from the LRU.
124 * putback_lru_page() will take care of removing the
125 * page from the unevictable list, if necessary.
126 * vmscan [page_referenced()] will move the page back
127 * to the unevictable list if some other vma has it
128 * mlocked.
129 */
130 if (PageUnevictable(page))
131 count_vm_event(UNEVICTABLE_PGSTRANDED);
132 else
133 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
134 }
135 }
136}
137
138/**
139 * __mlock_vma_pages_range() - mlock a range of pages in the vma.
140 * @vma: target vma
141 * @start: start address
142 * @end: end address
143 *
144 * This takes care of making the pages present too.
145 *
146 * return 0 on success, negative error code on error.
147 *
148 * vma->vm_mm->mmap_sem must be held for at least read.
149 */
150static long __mlock_vma_pages_range(struct vm_area_struct *vma,
151 unsigned long start, unsigned long end,
152 int *nonblocking)
153{
154 struct mm_struct *mm = vma->vm_mm;
155 unsigned long addr = start;
156 int nr_pages = (end - start) / PAGE_SIZE;
157 int gup_flags;
158
159 VM_BUG_ON(start & ~PAGE_MASK);
160 VM_BUG_ON(end & ~PAGE_MASK);
161 VM_BUG_ON(start < vma->vm_start);
162 VM_BUG_ON(end > vma->vm_end);
163 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
164
165 gup_flags = FOLL_TOUCH | FOLL_MLOCK;
166 /*
167 * We want to touch writable mappings with a write fault in order
168 * to break COW, except for shared mappings because these don't COW
169 * and we would not want to dirty them for nothing.
170 */
171 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
172 gup_flags |= FOLL_WRITE;
173
174 /*
175 * We want mlock to succeed for regions that have any permissions
176 * other than PROT_NONE.
177 */
178 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
179 gup_flags |= FOLL_FORCE;
180
181 return __get_user_pages(current, mm, addr, nr_pages, gup_flags,
182 NULL, NULL, nonblocking);
183}
184
185/*
186 * convert get_user_pages() return value to posix mlock() error
187 */
188static int __mlock_posix_error_return(long retval)
189{
190 if (retval == -EFAULT)
191 retval = -ENOMEM;
192 else if (retval == -ENOMEM)
193 retval = -EAGAIN;
194 return retval;
195}
196
197/**
198 * mlock_vma_pages_range() - mlock pages in specified vma range.
199 * @vma - the vma containing the specfied address range
200 * @start - starting address in @vma to mlock
201 * @end - end address [+1] in @vma to mlock
202 *
203 * For mmap()/mremap()/expansion of mlocked vma.
204 *
205 * return 0 on success for "normal" vmas.
206 *
207 * return number of pages [> 0] to be removed from locked_vm on success
208 * of "special" vmas.
209 */
210long mlock_vma_pages_range(struct vm_area_struct *vma,
211 unsigned long start, unsigned long end)
212{
213 int nr_pages = (end - start) / PAGE_SIZE;
214 BUG_ON(!(vma->vm_flags & VM_LOCKED));
215
216 /*
217 * filter unlockable vmas
218 */
219 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
220 goto no_mlock;
221
222 if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
223 is_vm_hugetlb_page(vma) ||
224 vma == get_gate_vma(current->mm))) {
225
226 __mlock_vma_pages_range(vma, start, end, NULL);
227
228 /* Hide errors from mmap() and other callers */
229 return 0;
230 }
231
232 /*
233 * User mapped kernel pages or huge pages:
234 * make these pages present to populate the ptes, but
235 * fall thru' to reset VM_LOCKED--no need to unlock, and
236 * return nr_pages so these don't get counted against task's
237 * locked limit. huge pages are already counted against
238 * locked vm limit.
239 */
240 make_pages_present(start, end);
241
242no_mlock:
243 vma->vm_flags &= ~VM_LOCKED; /* and don't come back! */
244 return nr_pages; /* error or pages NOT mlocked */
245}
246
247/*
248 * munlock_vma_pages_range() - munlock all pages in the vma range.'
249 * @vma - vma containing range to be munlock()ed.
250 * @start - start address in @vma of the range
251 * @end - end of range in @vma.
252 *
253 * For mremap(), munmap() and exit().
254 *
255 * Called with @vma VM_LOCKED.
256 *
257 * Returns with VM_LOCKED cleared. Callers must be prepared to
258 * deal with this.
259 *
260 * We don't save and restore VM_LOCKED here because pages are
261 * still on lru. In unmap path, pages might be scanned by reclaim
262 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
263 * free them. This will result in freeing mlocked pages.
264 */
265void munlock_vma_pages_range(struct vm_area_struct *vma,
266 unsigned long start, unsigned long end)
267{
268 unsigned long addr;
269
270 lru_add_drain();
271 vma->vm_flags &= ~VM_LOCKED;
272
273 for (addr = start; addr < end; addr += PAGE_SIZE) {
274 struct page *page;
275 /*
276 * Although FOLL_DUMP is intended for get_dump_page(),
277 * it just so happens that its special treatment of the
278 * ZERO_PAGE (returning an error instead of doing get_page)
279 * suits munlock very well (and if somehow an abnormal page
280 * has sneaked into the range, we won't oops here: great).
281 */
282 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
283 if (page && !IS_ERR(page)) {
284 lock_page(page);
285 /*
286 * Like in __mlock_vma_pages_range(),
287 * because we lock page here and migration is
288 * blocked by the elevated reference, we need
289 * only check for file-cache page truncation.
290 */
291 if (page->mapping)
292 munlock_vma_page(page);
293 unlock_page(page);
294 put_page(page);
295 }
296 cond_resched();
297 }
298}
299
300/*
301 * mlock_fixup - handle mlock[all]/munlock[all] requests.
302 *
303 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
304 * munlock is a no-op. However, for some special vmas, we go ahead and
305 * populate the ptes via make_pages_present().
306 *
307 * For vmas that pass the filters, merge/split as appropriate.
308 */
309static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
310 unsigned long start, unsigned long end, vm_flags_t newflags)
311{
312 struct mm_struct *mm = vma->vm_mm;
313 pgoff_t pgoff;
314 int nr_pages;
315 int ret = 0;
316 int lock = !!(newflags & VM_LOCKED);
317
318 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
319 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
320 goto out; /* don't set VM_LOCKED, don't count */
321
322 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
323 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
324 vma->vm_file, pgoff, vma_policy(vma));
325 if (*prev) {
326 vma = *prev;
327 goto success;
328 }
329
330 if (start != vma->vm_start) {
331 ret = split_vma(mm, vma, start, 1);
332 if (ret)
333 goto out;
334 }
335
336 if (end != vma->vm_end) {
337 ret = split_vma(mm, vma, end, 0);
338 if (ret)
339 goto out;
340 }
341
342success:
343 /*
344 * Keep track of amount of locked VM.
345 */
346 nr_pages = (end - start) >> PAGE_SHIFT;
347 if (!lock)
348 nr_pages = -nr_pages;
349 mm->locked_vm += nr_pages;
350
351 /*
352 * vm_flags is protected by the mmap_sem held in write mode.
353 * It's okay if try_to_unmap_one unmaps a page just after we
354 * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
355 */
356
357 if (lock)
358 vma->vm_flags = newflags;
359 else
360 munlock_vma_pages_range(vma, start, end);
361
362out:
363 *prev = vma;
364 return ret;
365}
366
367static int do_mlock(unsigned long start, size_t len, int on)
368{
369 unsigned long nstart, end, tmp;
370 struct vm_area_struct * vma, * prev;
371 int error;
372
373 VM_BUG_ON(start & ~PAGE_MASK);
374 VM_BUG_ON(len != PAGE_ALIGN(len));
375 end = start + len;
376 if (end < start)
377 return -EINVAL;
378 if (end == start)
379 return 0;
380 vma = find_vma_prev(current->mm, start, &prev);
381 if (!vma || vma->vm_start > start)
382 return -ENOMEM;
383
384 if (start > vma->vm_start)
385 prev = vma;
386
387 for (nstart = start ; ; ) {
388 vm_flags_t newflags;
389
390 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
391
392 newflags = vma->vm_flags | VM_LOCKED;
393 if (!on)
394 newflags &= ~VM_LOCKED;
395
396 tmp = vma->vm_end;
397 if (tmp > end)
398 tmp = end;
399 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
400 if (error)
401 break;
402 nstart = tmp;
403 if (nstart < prev->vm_end)
404 nstart = prev->vm_end;
405 if (nstart >= end)
406 break;
407
408 vma = prev->vm_next;
409 if (!vma || vma->vm_start != nstart) {
410 error = -ENOMEM;
411 break;
412 }
413 }
414 return error;
415}
416
417static int do_mlock_pages(unsigned long start, size_t len, int ignore_errors)
418{
419 struct mm_struct *mm = current->mm;
420 unsigned long end, nstart, nend;
421 struct vm_area_struct *vma = NULL;
422 int locked = 0;
423 int ret = 0;
424
425 VM_BUG_ON(start & ~PAGE_MASK);
426 VM_BUG_ON(len != PAGE_ALIGN(len));
427 end = start + len;
428
429 for (nstart = start; nstart < end; nstart = nend) {
430 /*
431 * We want to fault in pages for [nstart; end) address range.
432 * Find first corresponding VMA.
433 */
434 if (!locked) {
435 locked = 1;
436 down_read(&mm->mmap_sem);
437 vma = find_vma(mm, nstart);
438 } else if (nstart >= vma->vm_end)
439 vma = vma->vm_next;
440 if (!vma || vma->vm_start >= end)
441 break;
442 /*
443 * Set [nstart; nend) to intersection of desired address
444 * range with the first VMA. Also, skip undesirable VMA types.
445 */
446 nend = min(end, vma->vm_end);
447 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
448 continue;
449 if (nstart < vma->vm_start)
450 nstart = vma->vm_start;
451 /*
452 * Now fault in a range of pages. __mlock_vma_pages_range()
453 * double checks the vma flags, so that it won't mlock pages
454 * if the vma was already munlocked.
455 */
456 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
457 if (ret < 0) {
458 if (ignore_errors) {
459 ret = 0;
460 continue; /* continue at next VMA */
461 }
462 ret = __mlock_posix_error_return(ret);
463 break;
464 }
465 nend = nstart + ret * PAGE_SIZE;
466 ret = 0;
467 }
468 if (locked)
469 up_read(&mm->mmap_sem);
470 return ret; /* 0 or negative error code */
471}
472
473SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
474{
475 unsigned long locked;
476 unsigned long lock_limit;
477 int error = -ENOMEM;
478
479 if (!can_do_mlock())
480 return -EPERM;
481
482 lru_add_drain_all(); /* flush pagevec */
483
484 down_write(¤t->mm->mmap_sem);
485 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
486 start &= PAGE_MASK;
487
488 locked = len >> PAGE_SHIFT;
489 locked += current->mm->locked_vm;
490
491 lock_limit = rlimit(RLIMIT_MEMLOCK);
492 lock_limit >>= PAGE_SHIFT;
493
494 /* check against resource limits */
495 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
496 error = do_mlock(start, len, 1);
497 up_write(¤t->mm->mmap_sem);
498 if (!error)
499 error = do_mlock_pages(start, len, 0);
500 return error;
501}
502
503SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
504{
505 int ret;
506
507 down_write(¤t->mm->mmap_sem);
508 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
509 start &= PAGE_MASK;
510 ret = do_mlock(start, len, 0);
511 up_write(¤t->mm->mmap_sem);
512 return ret;
513}
514
515static int do_mlockall(int flags)
516{
517 struct vm_area_struct * vma, * prev = NULL;
518 unsigned int def_flags = 0;
519
520 if (flags & MCL_FUTURE)
521 def_flags = VM_LOCKED;
522 current->mm->def_flags = def_flags;
523 if (flags == MCL_FUTURE)
524 goto out;
525
526 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
527 vm_flags_t newflags;
528
529 newflags = vma->vm_flags | VM_LOCKED;
530 if (!(flags & MCL_CURRENT))
531 newflags &= ~VM_LOCKED;
532
533 /* Ignore errors */
534 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
535 }
536out:
537 return 0;
538}
539
540SYSCALL_DEFINE1(mlockall, int, flags)
541{
542 unsigned long lock_limit;
543 int ret = -EINVAL;
544
545 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
546 goto out;
547
548 ret = -EPERM;
549 if (!can_do_mlock())
550 goto out;
551
552 lru_add_drain_all(); /* flush pagevec */
553
554 down_write(¤t->mm->mmap_sem);
555
556 lock_limit = rlimit(RLIMIT_MEMLOCK);
557 lock_limit >>= PAGE_SHIFT;
558
559 ret = -ENOMEM;
560 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
561 capable(CAP_IPC_LOCK))
562 ret = do_mlockall(flags);
563 up_write(¤t->mm->mmap_sem);
564 if (!ret && (flags & MCL_CURRENT)) {
565 /* Ignore errors */
566 do_mlock_pages(0, TASK_SIZE, 1);
567 }
568out:
569 return ret;
570}
571
572SYSCALL_DEFINE0(munlockall)
573{
574 int ret;
575
576 down_write(¤t->mm->mmap_sem);
577 ret = do_mlockall(0);
578 up_write(¤t->mm->mmap_sem);
579 return ret;
580}
581
582/*
583 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
584 * shm segments) get accounted against the user_struct instead.
585 */
586static DEFINE_SPINLOCK(shmlock_user_lock);
587
588int user_shm_lock(size_t size, struct user_struct *user)
589{
590 unsigned long lock_limit, locked;
591 int allowed = 0;
592
593 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
594 lock_limit = rlimit(RLIMIT_MEMLOCK);
595 if (lock_limit == RLIM_INFINITY)
596 allowed = 1;
597 lock_limit >>= PAGE_SHIFT;
598 spin_lock(&shmlock_user_lock);
599 if (!allowed &&
600 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
601 goto out;
602 get_uid(user);
603 user->locked_shm += locked;
604 allowed = 1;
605out:
606 spin_unlock(&shmlock_user_lock);
607 return allowed;
608}
609
610void user_shm_unlock(size_t size, struct user_struct *user)
611{
612 spin_lock(&shmlock_user_lock);
613 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
614 spin_unlock(&shmlock_user_lock);
615 free_uid(user);
616}