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/mempolicy.h>
18#include <linux/syscalls.h>
19#include <linux/sched.h>
20#include <linux/export.h>
21#include <linux/rmap.h>
22#include <linux/mmzone.h>
23#include <linux/hugetlb.h>
24#include <linux/memcontrol.h>
25#include <linux/mm_inline.h>
26
27#include "internal.h"
28
29bool can_do_mlock(void)
30{
31 if (rlimit(RLIMIT_MEMLOCK) != 0)
32 return true;
33 if (capable(CAP_IPC_LOCK))
34 return true;
35 return false;
36}
37EXPORT_SYMBOL(can_do_mlock);
38
39/*
40 * Mlocked pages are marked with PageMlocked() flag for efficient testing
41 * in vmscan and, possibly, the fault path; and to support semi-accurate
42 * statistics.
43 *
44 * An mlocked page [PageMlocked(page)] is unevictable. As such, it will
45 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
46 * The unevictable list is an LRU sibling list to the [in]active lists.
47 * PageUnevictable is set to indicate the unevictable state.
48 *
49 * When lazy mlocking via vmscan, it is important to ensure that the
50 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
51 * may have mlocked a page that is being munlocked. So lazy mlock must take
52 * the mmap_lock for read, and verify that the vma really is locked
53 * (see mm/rmap.c).
54 */
55
56/*
57 * LRU accounting for clear_page_mlock()
58 */
59void clear_page_mlock(struct page *page)
60{
61 int nr_pages;
62
63 if (!TestClearPageMlocked(page))
64 return;
65
66 nr_pages = thp_nr_pages(page);
67 mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
68 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
69 /*
70 * The previous TestClearPageMlocked() corresponds to the smp_mb()
71 * in __pagevec_lru_add_fn().
72 *
73 * See __pagevec_lru_add_fn for more explanation.
74 */
75 if (!isolate_lru_page(page)) {
76 putback_lru_page(page);
77 } else {
78 /*
79 * We lost the race. the page already moved to evictable list.
80 */
81 if (PageUnevictable(page))
82 count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
83 }
84}
85
86/*
87 * Mark page as mlocked if not already.
88 * If page on LRU, isolate and putback to move to unevictable list.
89 */
90void mlock_vma_page(struct page *page)
91{
92 /* Serialize with page migration */
93 BUG_ON(!PageLocked(page));
94
95 VM_BUG_ON_PAGE(PageTail(page), page);
96 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
97
98 if (!TestSetPageMlocked(page)) {
99 int nr_pages = thp_nr_pages(page);
100
101 mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
102 count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
103 if (!isolate_lru_page(page))
104 putback_lru_page(page);
105 }
106}
107
108/*
109 * Isolate a page from LRU with optional get_page() pin.
110 * Assumes lru_lock already held and page already pinned.
111 */
112static bool __munlock_isolate_lru_page(struct page *page, bool getpage)
113{
114 if (PageLRU(page)) {
115 struct lruvec *lruvec;
116
117 lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
118 if (getpage)
119 get_page(page);
120 ClearPageLRU(page);
121 del_page_from_lru_list(page, lruvec, page_lru(page));
122 return true;
123 }
124
125 return false;
126}
127
128/*
129 * Finish munlock after successful page isolation
130 *
131 * Page must be locked. This is a wrapper for try_to_munlock()
132 * and putback_lru_page() with munlock accounting.
133 */
134static void __munlock_isolated_page(struct page *page)
135{
136 /*
137 * Optimization: if the page was mapped just once, that's our mapping
138 * and we don't need to check all the other vmas.
139 */
140 if (page_mapcount(page) > 1)
141 try_to_munlock(page);
142
143 /* Did try_to_unlock() succeed or punt? */
144 if (!PageMlocked(page))
145 count_vm_events(UNEVICTABLE_PGMUNLOCKED, thp_nr_pages(page));
146
147 putback_lru_page(page);
148}
149
150/*
151 * Accounting for page isolation fail during munlock
152 *
153 * Performs accounting when page isolation fails in munlock. There is nothing
154 * else to do because it means some other task has already removed the page
155 * from the LRU. putback_lru_page() will take care of removing the page from
156 * the unevictable list, if necessary. vmscan [page_referenced()] will move
157 * the page back to the unevictable list if some other vma has it mlocked.
158 */
159static void __munlock_isolation_failed(struct page *page)
160{
161 int nr_pages = thp_nr_pages(page);
162
163 if (PageUnevictable(page))
164 __count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
165 else
166 __count_vm_events(UNEVICTABLE_PGMUNLOCKED, nr_pages);
167}
168
169/**
170 * munlock_vma_page - munlock a vma page
171 * @page: page to be unlocked, either a normal page or THP page head
172 *
173 * returns the size of the page as a page mask (0 for normal page,
174 * HPAGE_PMD_NR - 1 for THP head page)
175 *
176 * called from munlock()/munmap() path with page supposedly on the LRU.
177 * When we munlock a page, because the vma where we found the page is being
178 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
179 * page locked so that we can leave it on the unevictable lru list and not
180 * bother vmscan with it. However, to walk the page's rmap list in
181 * try_to_munlock() we must isolate the page from the LRU. If some other
182 * task has removed the page from the LRU, we won't be able to do that.
183 * So we clear the PageMlocked as we might not get another chance. If we
184 * can't isolate the page, we leave it for putback_lru_page() and vmscan
185 * [page_referenced()/try_to_unmap()] to deal with.
186 */
187unsigned int munlock_vma_page(struct page *page)
188{
189 int nr_pages;
190 pg_data_t *pgdat = page_pgdat(page);
191
192 /* For try_to_munlock() and to serialize with page migration */
193 BUG_ON(!PageLocked(page));
194
195 VM_BUG_ON_PAGE(PageTail(page), page);
196
197 /*
198 * Serialize with any parallel __split_huge_page_refcount() which
199 * might otherwise copy PageMlocked to part of the tail pages before
200 * we clear it in the head page. It also stabilizes thp_nr_pages().
201 */
202 spin_lock_irq(&pgdat->lru_lock);
203
204 if (!TestClearPageMlocked(page)) {
205 /* Potentially, PTE-mapped THP: do not skip the rest PTEs */
206 nr_pages = 1;
207 goto unlock_out;
208 }
209
210 nr_pages = thp_nr_pages(page);
211 __mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
212
213 if (__munlock_isolate_lru_page(page, true)) {
214 spin_unlock_irq(&pgdat->lru_lock);
215 __munlock_isolated_page(page);
216 goto out;
217 }
218 __munlock_isolation_failed(page);
219
220unlock_out:
221 spin_unlock_irq(&pgdat->lru_lock);
222
223out:
224 return nr_pages - 1;
225}
226
227/*
228 * convert get_user_pages() return value to posix mlock() error
229 */
230static int __mlock_posix_error_return(long retval)
231{
232 if (retval == -EFAULT)
233 retval = -ENOMEM;
234 else if (retval == -ENOMEM)
235 retval = -EAGAIN;
236 return retval;
237}
238
239/*
240 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
241 *
242 * The fast path is available only for evictable pages with single mapping.
243 * Then we can bypass the per-cpu pvec and get better performance.
244 * when mapcount > 1 we need try_to_munlock() which can fail.
245 * when !page_evictable(), we need the full redo logic of putback_lru_page to
246 * avoid leaving evictable page in unevictable list.
247 *
248 * In case of success, @page is added to @pvec and @pgrescued is incremented
249 * in case that the page was previously unevictable. @page is also unlocked.
250 */
251static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
252 int *pgrescued)
253{
254 VM_BUG_ON_PAGE(PageLRU(page), page);
255 VM_BUG_ON_PAGE(!PageLocked(page), page);
256
257 if (page_mapcount(page) <= 1 && page_evictable(page)) {
258 pagevec_add(pvec, page);
259 if (TestClearPageUnevictable(page))
260 (*pgrescued)++;
261 unlock_page(page);
262 return true;
263 }
264
265 return false;
266}
267
268/*
269 * Putback multiple evictable pages to the LRU
270 *
271 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
272 * the pages might have meanwhile become unevictable but that is OK.
273 */
274static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
275{
276 count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
277 /*
278 *__pagevec_lru_add() calls release_pages() so we don't call
279 * put_page() explicitly
280 */
281 __pagevec_lru_add(pvec);
282 count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
283}
284
285/*
286 * Munlock a batch of pages from the same zone
287 *
288 * The work is split to two main phases. First phase clears the Mlocked flag
289 * and attempts to isolate the pages, all under a single zone lru lock.
290 * The second phase finishes the munlock only for pages where isolation
291 * succeeded.
292 *
293 * Note that the pagevec may be modified during the process.
294 */
295static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
296{
297 int i;
298 int nr = pagevec_count(pvec);
299 int delta_munlocked = -nr;
300 struct pagevec pvec_putback;
301 int pgrescued = 0;
302
303 pagevec_init(&pvec_putback);
304
305 /* Phase 1: page isolation */
306 spin_lock_irq(&zone->zone_pgdat->lru_lock);
307 for (i = 0; i < nr; i++) {
308 struct page *page = pvec->pages[i];
309
310 if (TestClearPageMlocked(page)) {
311 /*
312 * We already have pin from follow_page_mask()
313 * so we can spare the get_page() here.
314 */
315 if (__munlock_isolate_lru_page(page, false))
316 continue;
317 else
318 __munlock_isolation_failed(page);
319 } else {
320 delta_munlocked++;
321 }
322
323 /*
324 * We won't be munlocking this page in the next phase
325 * but we still need to release the follow_page_mask()
326 * pin. We cannot do it under lru_lock however. If it's
327 * the last pin, __page_cache_release() would deadlock.
328 */
329 pagevec_add(&pvec_putback, pvec->pages[i]);
330 pvec->pages[i] = NULL;
331 }
332 __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
333 spin_unlock_irq(&zone->zone_pgdat->lru_lock);
334
335 /* Now we can release pins of pages that we are not munlocking */
336 pagevec_release(&pvec_putback);
337
338 /* Phase 2: page munlock */
339 for (i = 0; i < nr; i++) {
340 struct page *page = pvec->pages[i];
341
342 if (page) {
343 lock_page(page);
344 if (!__putback_lru_fast_prepare(page, &pvec_putback,
345 &pgrescued)) {
346 /*
347 * Slow path. We don't want to lose the last
348 * pin before unlock_page()
349 */
350 get_page(page); /* for putback_lru_page() */
351 __munlock_isolated_page(page);
352 unlock_page(page);
353 put_page(page); /* from follow_page_mask() */
354 }
355 }
356 }
357
358 /*
359 * Phase 3: page putback for pages that qualified for the fast path
360 * This will also call put_page() to return pin from follow_page_mask()
361 */
362 if (pagevec_count(&pvec_putback))
363 __putback_lru_fast(&pvec_putback, pgrescued);
364}
365
366/*
367 * Fill up pagevec for __munlock_pagevec using pte walk
368 *
369 * The function expects that the struct page corresponding to @start address is
370 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
371 *
372 * The rest of @pvec is filled by subsequent pages within the same pmd and same
373 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
374 * pages also get pinned.
375 *
376 * Returns the address of the next page that should be scanned. This equals
377 * @start + PAGE_SIZE when no page could be added by the pte walk.
378 */
379static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
380 struct vm_area_struct *vma, struct zone *zone,
381 unsigned long start, unsigned long end)
382{
383 pte_t *pte;
384 spinlock_t *ptl;
385
386 /*
387 * Initialize pte walk starting at the already pinned page where we
388 * are sure that there is a pte, as it was pinned under the same
389 * mmap_lock write op.
390 */
391 pte = get_locked_pte(vma->vm_mm, start, &ptl);
392 /* Make sure we do not cross the page table boundary */
393 end = pgd_addr_end(start, end);
394 end = p4d_addr_end(start, end);
395 end = pud_addr_end(start, end);
396 end = pmd_addr_end(start, end);
397
398 /* The page next to the pinned page is the first we will try to get */
399 start += PAGE_SIZE;
400 while (start < end) {
401 struct page *page = NULL;
402 pte++;
403 if (pte_present(*pte))
404 page = vm_normal_page(vma, start, *pte);
405 /*
406 * Break if page could not be obtained or the page's node+zone does not
407 * match
408 */
409 if (!page || page_zone(page) != zone)
410 break;
411
412 /*
413 * Do not use pagevec for PTE-mapped THP,
414 * munlock_vma_pages_range() will handle them.
415 */
416 if (PageTransCompound(page))
417 break;
418
419 get_page(page);
420 /*
421 * Increase the address that will be returned *before* the
422 * eventual break due to pvec becoming full by adding the page
423 */
424 start += PAGE_SIZE;
425 if (pagevec_add(pvec, page) == 0)
426 break;
427 }
428 pte_unmap_unlock(pte, ptl);
429 return start;
430}
431
432/*
433 * munlock_vma_pages_range() - munlock all pages in the vma range.'
434 * @vma - vma containing range to be munlock()ed.
435 * @start - start address in @vma of the range
436 * @end - end of range in @vma.
437 *
438 * For mremap(), munmap() and exit().
439 *
440 * Called with @vma VM_LOCKED.
441 *
442 * Returns with VM_LOCKED cleared. Callers must be prepared to
443 * deal with this.
444 *
445 * We don't save and restore VM_LOCKED here because pages are
446 * still on lru. In unmap path, pages might be scanned by reclaim
447 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
448 * free them. This will result in freeing mlocked pages.
449 */
450void munlock_vma_pages_range(struct vm_area_struct *vma,
451 unsigned long start, unsigned long end)
452{
453 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
454
455 while (start < end) {
456 struct page *page;
457 unsigned int page_mask = 0;
458 unsigned long page_increm;
459 struct pagevec pvec;
460 struct zone *zone;
461
462 pagevec_init(&pvec);
463 /*
464 * Although FOLL_DUMP is intended for get_dump_page(),
465 * it just so happens that its special treatment of the
466 * ZERO_PAGE (returning an error instead of doing get_page)
467 * suits munlock very well (and if somehow an abnormal page
468 * has sneaked into the range, we won't oops here: great).
469 */
470 page = follow_page(vma, start, FOLL_GET | FOLL_DUMP);
471
472 if (page && !IS_ERR(page)) {
473 if (PageTransTail(page)) {
474 VM_BUG_ON_PAGE(PageMlocked(page), page);
475 put_page(page); /* follow_page_mask() */
476 } else if (PageTransHuge(page)) {
477 lock_page(page);
478 /*
479 * Any THP page found by follow_page_mask() may
480 * have gotten split before reaching
481 * munlock_vma_page(), so we need to compute
482 * the page_mask here instead.
483 */
484 page_mask = munlock_vma_page(page);
485 unlock_page(page);
486 put_page(page); /* follow_page_mask() */
487 } else {
488 /*
489 * Non-huge pages are handled in batches via
490 * pagevec. The pin from follow_page_mask()
491 * prevents them from collapsing by THP.
492 */
493 pagevec_add(&pvec, page);
494 zone = page_zone(page);
495
496 /*
497 * Try to fill the rest of pagevec using fast
498 * pte walk. This will also update start to
499 * the next page to process. Then munlock the
500 * pagevec.
501 */
502 start = __munlock_pagevec_fill(&pvec, vma,
503 zone, start, end);
504 __munlock_pagevec(&pvec, zone);
505 goto next;
506 }
507 }
508 page_increm = 1 + page_mask;
509 start += page_increm * PAGE_SIZE;
510next:
511 cond_resched();
512 }
513}
514
515/*
516 * mlock_fixup - handle mlock[all]/munlock[all] requests.
517 *
518 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
519 * munlock is a no-op. However, for some special vmas, we go ahead and
520 * populate the ptes.
521 *
522 * For vmas that pass the filters, merge/split as appropriate.
523 */
524static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
525 unsigned long start, unsigned long end, vm_flags_t newflags)
526{
527 struct mm_struct *mm = vma->vm_mm;
528 pgoff_t pgoff;
529 int nr_pages;
530 int ret = 0;
531 int lock = !!(newflags & VM_LOCKED);
532 vm_flags_t old_flags = vma->vm_flags;
533
534 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
535 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm) ||
536 vma_is_dax(vma))
537 /* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
538 goto out;
539
540 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
541 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
542 vma->vm_file, pgoff, vma_policy(vma),
543 vma->vm_userfaultfd_ctx);
544 if (*prev) {
545 vma = *prev;
546 goto success;
547 }
548
549 if (start != vma->vm_start) {
550 ret = split_vma(mm, vma, start, 1);
551 if (ret)
552 goto out;
553 }
554
555 if (end != vma->vm_end) {
556 ret = split_vma(mm, vma, end, 0);
557 if (ret)
558 goto out;
559 }
560
561success:
562 /*
563 * Keep track of amount of locked VM.
564 */
565 nr_pages = (end - start) >> PAGE_SHIFT;
566 if (!lock)
567 nr_pages = -nr_pages;
568 else if (old_flags & VM_LOCKED)
569 nr_pages = 0;
570 mm->locked_vm += nr_pages;
571
572 /*
573 * vm_flags is protected by the mmap_lock held in write mode.
574 * It's okay if try_to_unmap_one unmaps a page just after we
575 * set VM_LOCKED, populate_vma_page_range will bring it back.
576 */
577
578 if (lock)
579 vma->vm_flags = newflags;
580 else
581 munlock_vma_pages_range(vma, start, end);
582
583out:
584 *prev = vma;
585 return ret;
586}
587
588static int apply_vma_lock_flags(unsigned long start, size_t len,
589 vm_flags_t flags)
590{
591 unsigned long nstart, end, tmp;
592 struct vm_area_struct * vma, * prev;
593 int error;
594
595 VM_BUG_ON(offset_in_page(start));
596 VM_BUG_ON(len != PAGE_ALIGN(len));
597 end = start + len;
598 if (end < start)
599 return -EINVAL;
600 if (end == start)
601 return 0;
602 vma = find_vma(current->mm, start);
603 if (!vma || vma->vm_start > start)
604 return -ENOMEM;
605
606 prev = vma->vm_prev;
607 if (start > vma->vm_start)
608 prev = vma;
609
610 for (nstart = start ; ; ) {
611 vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
612
613 newflags |= flags;
614
615 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
616 tmp = vma->vm_end;
617 if (tmp > end)
618 tmp = end;
619 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
620 if (error)
621 break;
622 nstart = tmp;
623 if (nstart < prev->vm_end)
624 nstart = prev->vm_end;
625 if (nstart >= end)
626 break;
627
628 vma = prev->vm_next;
629 if (!vma || vma->vm_start != nstart) {
630 error = -ENOMEM;
631 break;
632 }
633 }
634 return error;
635}
636
637/*
638 * Go through vma areas and sum size of mlocked
639 * vma pages, as return value.
640 * Note deferred memory locking case(mlock2(,,MLOCK_ONFAULT)
641 * is also counted.
642 * Return value: previously mlocked page counts
643 */
644static unsigned long count_mm_mlocked_page_nr(struct mm_struct *mm,
645 unsigned long start, size_t len)
646{
647 struct vm_area_struct *vma;
648 unsigned long count = 0;
649
650 if (mm == NULL)
651 mm = current->mm;
652
653 vma = find_vma(mm, start);
654 if (vma == NULL)
655 vma = mm->mmap;
656
657 for (; vma ; vma = vma->vm_next) {
658 if (start >= vma->vm_end)
659 continue;
660 if (start + len <= vma->vm_start)
661 break;
662 if (vma->vm_flags & VM_LOCKED) {
663 if (start > vma->vm_start)
664 count -= (start - vma->vm_start);
665 if (start + len < vma->vm_end) {
666 count += start + len - vma->vm_start;
667 break;
668 }
669 count += vma->vm_end - vma->vm_start;
670 }
671 }
672
673 return count >> PAGE_SHIFT;
674}
675
676static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
677{
678 unsigned long locked;
679 unsigned long lock_limit;
680 int error = -ENOMEM;
681
682 start = untagged_addr(start);
683
684 if (!can_do_mlock())
685 return -EPERM;
686
687 len = PAGE_ALIGN(len + (offset_in_page(start)));
688 start &= PAGE_MASK;
689
690 lock_limit = rlimit(RLIMIT_MEMLOCK);
691 lock_limit >>= PAGE_SHIFT;
692 locked = len >> PAGE_SHIFT;
693
694 if (mmap_write_lock_killable(current->mm))
695 return -EINTR;
696
697 locked += current->mm->locked_vm;
698 if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
699 /*
700 * It is possible that the regions requested intersect with
701 * previously mlocked areas, that part area in "mm->locked_vm"
702 * should not be counted to new mlock increment count. So check
703 * and adjust locked count if necessary.
704 */
705 locked -= count_mm_mlocked_page_nr(current->mm,
706 start, len);
707 }
708
709 /* check against resource limits */
710 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
711 error = apply_vma_lock_flags(start, len, flags);
712
713 mmap_write_unlock(current->mm);
714 if (error)
715 return error;
716
717 error = __mm_populate(start, len, 0);
718 if (error)
719 return __mlock_posix_error_return(error);
720 return 0;
721}
722
723SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
724{
725 return do_mlock(start, len, VM_LOCKED);
726}
727
728SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
729{
730 vm_flags_t vm_flags = VM_LOCKED;
731
732 if (flags & ~MLOCK_ONFAULT)
733 return -EINVAL;
734
735 if (flags & MLOCK_ONFAULT)
736 vm_flags |= VM_LOCKONFAULT;
737
738 return do_mlock(start, len, vm_flags);
739}
740
741SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
742{
743 int ret;
744
745 start = untagged_addr(start);
746
747 len = PAGE_ALIGN(len + (offset_in_page(start)));
748 start &= PAGE_MASK;
749
750 if (mmap_write_lock_killable(current->mm))
751 return -EINTR;
752 ret = apply_vma_lock_flags(start, len, 0);
753 mmap_write_unlock(current->mm);
754
755 return ret;
756}
757
758/*
759 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
760 * and translate into the appropriate modifications to mm->def_flags and/or the
761 * flags for all current VMAs.
762 *
763 * There are a couple of subtleties with this. If mlockall() is called multiple
764 * times with different flags, the values do not necessarily stack. If mlockall
765 * is called once including the MCL_FUTURE flag and then a second time without
766 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
767 */
768static int apply_mlockall_flags(int flags)
769{
770 struct vm_area_struct * vma, * prev = NULL;
771 vm_flags_t to_add = 0;
772
773 current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
774 if (flags & MCL_FUTURE) {
775 current->mm->def_flags |= VM_LOCKED;
776
777 if (flags & MCL_ONFAULT)
778 current->mm->def_flags |= VM_LOCKONFAULT;
779
780 if (!(flags & MCL_CURRENT))
781 goto out;
782 }
783
784 if (flags & MCL_CURRENT) {
785 to_add |= VM_LOCKED;
786 if (flags & MCL_ONFAULT)
787 to_add |= VM_LOCKONFAULT;
788 }
789
790 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
791 vm_flags_t newflags;
792
793 newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
794 newflags |= to_add;
795
796 /* Ignore errors */
797 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
798 cond_resched();
799 }
800out:
801 return 0;
802}
803
804SYSCALL_DEFINE1(mlockall, int, flags)
805{
806 unsigned long lock_limit;
807 int ret;
808
809 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)) ||
810 flags == MCL_ONFAULT)
811 return -EINVAL;
812
813 if (!can_do_mlock())
814 return -EPERM;
815
816 lock_limit = rlimit(RLIMIT_MEMLOCK);
817 lock_limit >>= PAGE_SHIFT;
818
819 if (mmap_write_lock_killable(current->mm))
820 return -EINTR;
821
822 ret = -ENOMEM;
823 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
824 capable(CAP_IPC_LOCK))
825 ret = apply_mlockall_flags(flags);
826 mmap_write_unlock(current->mm);
827 if (!ret && (flags & MCL_CURRENT))
828 mm_populate(0, TASK_SIZE);
829
830 return ret;
831}
832
833SYSCALL_DEFINE0(munlockall)
834{
835 int ret;
836
837 if (mmap_write_lock_killable(current->mm))
838 return -EINTR;
839 ret = apply_mlockall_flags(0);
840 mmap_write_unlock(current->mm);
841 return ret;
842}
843
844/*
845 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
846 * shm segments) get accounted against the user_struct instead.
847 */
848static DEFINE_SPINLOCK(shmlock_user_lock);
849
850int user_shm_lock(size_t size, struct user_struct *user)
851{
852 unsigned long lock_limit, locked;
853 int allowed = 0;
854
855 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
856 lock_limit = rlimit(RLIMIT_MEMLOCK);
857 if (lock_limit == RLIM_INFINITY)
858 allowed = 1;
859 lock_limit >>= PAGE_SHIFT;
860 spin_lock(&shmlock_user_lock);
861 if (!allowed &&
862 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
863 goto out;
864 get_uid(user);
865 user->locked_shm += locked;
866 allowed = 1;
867out:
868 spin_unlock(&shmlock_user_lock);
869 return allowed;
870}
871
872void user_shm_unlock(size_t size, struct user_struct *user)
873{
874 spin_lock(&shmlock_user_lock);
875 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
876 spin_unlock(&shmlock_user_lock);
877 free_uid(user);
878}
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/mempolicy.h>
18#include <linux/syscalls.h>
19#include <linux/sched.h>
20#include <linux/export.h>
21#include <linux/rmap.h>
22#include <linux/mmzone.h>
23#include <linux/hugetlb.h>
24#include <linux/memcontrol.h>
25#include <linux/mm_inline.h>
26
27#include "internal.h"
28
29bool can_do_mlock(void)
30{
31 if (rlimit(RLIMIT_MEMLOCK) != 0)
32 return true;
33 if (capable(CAP_IPC_LOCK))
34 return true;
35 return false;
36}
37EXPORT_SYMBOL(can_do_mlock);
38
39/*
40 * Mlocked pages are marked with PageMlocked() flag for efficient testing
41 * in vmscan and, possibly, the fault path; and to support semi-accurate
42 * statistics.
43 *
44 * An mlocked page [PageMlocked(page)] is unevictable. As such, it will
45 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
46 * The unevictable list is an LRU sibling list to the [in]active lists.
47 * PageUnevictable is set to indicate the unevictable state.
48 *
49 * When lazy mlocking via vmscan, it is important to ensure that the
50 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
51 * may have mlocked a page that is being munlocked. So lazy mlock must take
52 * the mmap_sem for read, and verify that the vma really is locked
53 * (see mm/rmap.c).
54 */
55
56/*
57 * LRU accounting for clear_page_mlock()
58 */
59void clear_page_mlock(struct page *page)
60{
61 if (!TestClearPageMlocked(page))
62 return;
63
64 mod_zone_page_state(page_zone(page), NR_MLOCK,
65 -hpage_nr_pages(page));
66 count_vm_event(UNEVICTABLE_PGCLEARED);
67 /*
68 * The previous TestClearPageMlocked() corresponds to the smp_mb()
69 * in __pagevec_lru_add_fn().
70 *
71 * See __pagevec_lru_add_fn for more explanation.
72 */
73 if (!isolate_lru_page(page)) {
74 putback_lru_page(page);
75 } else {
76 /*
77 * We lost the race. the page already moved to evictable list.
78 */
79 if (PageUnevictable(page))
80 count_vm_event(UNEVICTABLE_PGSTRANDED);
81 }
82}
83
84/*
85 * Mark page as mlocked if not already.
86 * If page on LRU, isolate and putback to move to unevictable list.
87 */
88void mlock_vma_page(struct page *page)
89{
90 /* Serialize with page migration */
91 BUG_ON(!PageLocked(page));
92
93 VM_BUG_ON_PAGE(PageTail(page), page);
94 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
95
96 if (!TestSetPageMlocked(page)) {
97 mod_zone_page_state(page_zone(page), NR_MLOCK,
98 hpage_nr_pages(page));
99 count_vm_event(UNEVICTABLE_PGMLOCKED);
100 if (!isolate_lru_page(page))
101 putback_lru_page(page);
102 }
103}
104
105/*
106 * Isolate a page from LRU with optional get_page() pin.
107 * Assumes lru_lock already held and page already pinned.
108 */
109static bool __munlock_isolate_lru_page(struct page *page, bool getpage)
110{
111 if (PageLRU(page)) {
112 struct lruvec *lruvec;
113
114 lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
115 if (getpage)
116 get_page(page);
117 ClearPageLRU(page);
118 del_page_from_lru_list(page, lruvec, page_lru(page));
119 return true;
120 }
121
122 return false;
123}
124
125/*
126 * Finish munlock after successful page isolation
127 *
128 * Page must be locked. This is a wrapper for try_to_munlock()
129 * and putback_lru_page() with munlock accounting.
130 */
131static void __munlock_isolated_page(struct page *page)
132{
133 /*
134 * Optimization: if the page was mapped just once, that's our mapping
135 * and we don't need to check all the other vmas.
136 */
137 if (page_mapcount(page) > 1)
138 try_to_munlock(page);
139
140 /* Did try_to_unlock() succeed or punt? */
141 if (!PageMlocked(page))
142 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
143
144 putback_lru_page(page);
145}
146
147/*
148 * Accounting for page isolation fail during munlock
149 *
150 * Performs accounting when page isolation fails in munlock. There is nothing
151 * else to do because it means some other task has already removed the page
152 * from the LRU. putback_lru_page() will take care of removing the page from
153 * the unevictable list, if necessary. vmscan [page_referenced()] will move
154 * the page back to the unevictable list if some other vma has it mlocked.
155 */
156static void __munlock_isolation_failed(struct page *page)
157{
158 if (PageUnevictable(page))
159 __count_vm_event(UNEVICTABLE_PGSTRANDED);
160 else
161 __count_vm_event(UNEVICTABLE_PGMUNLOCKED);
162}
163
164/**
165 * munlock_vma_page - munlock a vma page
166 * @page: page to be unlocked, either a normal page or THP page head
167 *
168 * returns the size of the page as a page mask (0 for normal page,
169 * HPAGE_PMD_NR - 1 for THP head page)
170 *
171 * called from munlock()/munmap() path with page supposedly on the LRU.
172 * When we munlock a page, because the vma where we found the page is being
173 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
174 * page locked so that we can leave it on the unevictable lru list and not
175 * bother vmscan with it. However, to walk the page's rmap list in
176 * try_to_munlock() we must isolate the page from the LRU. If some other
177 * task has removed the page from the LRU, we won't be able to do that.
178 * So we clear the PageMlocked as we might not get another chance. If we
179 * can't isolate the page, we leave it for putback_lru_page() and vmscan
180 * [page_referenced()/try_to_unmap()] to deal with.
181 */
182unsigned int munlock_vma_page(struct page *page)
183{
184 int nr_pages;
185 pg_data_t *pgdat = page_pgdat(page);
186
187 /* For try_to_munlock() and to serialize with page migration */
188 BUG_ON(!PageLocked(page));
189
190 VM_BUG_ON_PAGE(PageTail(page), page);
191
192 /*
193 * Serialize with any parallel __split_huge_page_refcount() which
194 * might otherwise copy PageMlocked to part of the tail pages before
195 * we clear it in the head page. It also stabilizes hpage_nr_pages().
196 */
197 spin_lock_irq(&pgdat->lru_lock);
198
199 if (!TestClearPageMlocked(page)) {
200 /* Potentially, PTE-mapped THP: do not skip the rest PTEs */
201 nr_pages = 1;
202 goto unlock_out;
203 }
204
205 nr_pages = hpage_nr_pages(page);
206 __mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
207
208 if (__munlock_isolate_lru_page(page, true)) {
209 spin_unlock_irq(&pgdat->lru_lock);
210 __munlock_isolated_page(page);
211 goto out;
212 }
213 __munlock_isolation_failed(page);
214
215unlock_out:
216 spin_unlock_irq(&pgdat->lru_lock);
217
218out:
219 return nr_pages - 1;
220}
221
222/*
223 * convert get_user_pages() return value to posix mlock() error
224 */
225static int __mlock_posix_error_return(long retval)
226{
227 if (retval == -EFAULT)
228 retval = -ENOMEM;
229 else if (retval == -ENOMEM)
230 retval = -EAGAIN;
231 return retval;
232}
233
234/*
235 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
236 *
237 * The fast path is available only for evictable pages with single mapping.
238 * Then we can bypass the per-cpu pvec and get better performance.
239 * when mapcount > 1 we need try_to_munlock() which can fail.
240 * when !page_evictable(), we need the full redo logic of putback_lru_page to
241 * avoid leaving evictable page in unevictable list.
242 *
243 * In case of success, @page is added to @pvec and @pgrescued is incremented
244 * in case that the page was previously unevictable. @page is also unlocked.
245 */
246static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
247 int *pgrescued)
248{
249 VM_BUG_ON_PAGE(PageLRU(page), page);
250 VM_BUG_ON_PAGE(!PageLocked(page), page);
251
252 if (page_mapcount(page) <= 1 && page_evictable(page)) {
253 pagevec_add(pvec, page);
254 if (TestClearPageUnevictable(page))
255 (*pgrescued)++;
256 unlock_page(page);
257 return true;
258 }
259
260 return false;
261}
262
263/*
264 * Putback multiple evictable pages to the LRU
265 *
266 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
267 * the pages might have meanwhile become unevictable but that is OK.
268 */
269static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
270{
271 count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
272 /*
273 *__pagevec_lru_add() calls release_pages() so we don't call
274 * put_page() explicitly
275 */
276 __pagevec_lru_add(pvec);
277 count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
278}
279
280/*
281 * Munlock a batch of pages from the same zone
282 *
283 * The work is split to two main phases. First phase clears the Mlocked flag
284 * and attempts to isolate the pages, all under a single zone lru lock.
285 * The second phase finishes the munlock only for pages where isolation
286 * succeeded.
287 *
288 * Note that the pagevec may be modified during the process.
289 */
290static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
291{
292 int i;
293 int nr = pagevec_count(pvec);
294 int delta_munlocked = -nr;
295 struct pagevec pvec_putback;
296 int pgrescued = 0;
297
298 pagevec_init(&pvec_putback);
299
300 /* Phase 1: page isolation */
301 spin_lock_irq(&zone->zone_pgdat->lru_lock);
302 for (i = 0; i < nr; i++) {
303 struct page *page = pvec->pages[i];
304
305 if (TestClearPageMlocked(page)) {
306 /*
307 * We already have pin from follow_page_mask()
308 * so we can spare the get_page() here.
309 */
310 if (__munlock_isolate_lru_page(page, false))
311 continue;
312 else
313 __munlock_isolation_failed(page);
314 } else {
315 delta_munlocked++;
316 }
317
318 /*
319 * We won't be munlocking this page in the next phase
320 * but we still need to release the follow_page_mask()
321 * pin. We cannot do it under lru_lock however. If it's
322 * the last pin, __page_cache_release() would deadlock.
323 */
324 pagevec_add(&pvec_putback, pvec->pages[i]);
325 pvec->pages[i] = NULL;
326 }
327 __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
328 spin_unlock_irq(&zone->zone_pgdat->lru_lock);
329
330 /* Now we can release pins of pages that we are not munlocking */
331 pagevec_release(&pvec_putback);
332
333 /* Phase 2: page munlock */
334 for (i = 0; i < nr; i++) {
335 struct page *page = pvec->pages[i];
336
337 if (page) {
338 lock_page(page);
339 if (!__putback_lru_fast_prepare(page, &pvec_putback,
340 &pgrescued)) {
341 /*
342 * Slow path. We don't want to lose the last
343 * pin before unlock_page()
344 */
345 get_page(page); /* for putback_lru_page() */
346 __munlock_isolated_page(page);
347 unlock_page(page);
348 put_page(page); /* from follow_page_mask() */
349 }
350 }
351 }
352
353 /*
354 * Phase 3: page putback for pages that qualified for the fast path
355 * This will also call put_page() to return pin from follow_page_mask()
356 */
357 if (pagevec_count(&pvec_putback))
358 __putback_lru_fast(&pvec_putback, pgrescued);
359}
360
361/*
362 * Fill up pagevec for __munlock_pagevec using pte walk
363 *
364 * The function expects that the struct page corresponding to @start address is
365 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
366 *
367 * The rest of @pvec is filled by subsequent pages within the same pmd and same
368 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
369 * pages also get pinned.
370 *
371 * Returns the address of the next page that should be scanned. This equals
372 * @start + PAGE_SIZE when no page could be added by the pte walk.
373 */
374static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
375 struct vm_area_struct *vma, struct zone *zone,
376 unsigned long start, unsigned long end)
377{
378 pte_t *pte;
379 spinlock_t *ptl;
380
381 /*
382 * Initialize pte walk starting at the already pinned page where we
383 * are sure that there is a pte, as it was pinned under the same
384 * mmap_sem write op.
385 */
386 pte = get_locked_pte(vma->vm_mm, start, &ptl);
387 /* Make sure we do not cross the page table boundary */
388 end = pgd_addr_end(start, end);
389 end = p4d_addr_end(start, end);
390 end = pud_addr_end(start, end);
391 end = pmd_addr_end(start, end);
392
393 /* The page next to the pinned page is the first we will try to get */
394 start += PAGE_SIZE;
395 while (start < end) {
396 struct page *page = NULL;
397 pte++;
398 if (pte_present(*pte))
399 page = vm_normal_page(vma, start, *pte);
400 /*
401 * Break if page could not be obtained or the page's node+zone does not
402 * match
403 */
404 if (!page || page_zone(page) != zone)
405 break;
406
407 /*
408 * Do not use pagevec for PTE-mapped THP,
409 * munlock_vma_pages_range() will handle them.
410 */
411 if (PageTransCompound(page))
412 break;
413
414 get_page(page);
415 /*
416 * Increase the address that will be returned *before* the
417 * eventual break due to pvec becoming full by adding the page
418 */
419 start += PAGE_SIZE;
420 if (pagevec_add(pvec, page) == 0)
421 break;
422 }
423 pte_unmap_unlock(pte, ptl);
424 return start;
425}
426
427/*
428 * munlock_vma_pages_range() - munlock all pages in the vma range.'
429 * @vma - vma containing range to be munlock()ed.
430 * @start - start address in @vma of the range
431 * @end - end of range in @vma.
432 *
433 * For mremap(), munmap() and exit().
434 *
435 * Called with @vma VM_LOCKED.
436 *
437 * Returns with VM_LOCKED cleared. Callers must be prepared to
438 * deal with this.
439 *
440 * We don't save and restore VM_LOCKED here because pages are
441 * still on lru. In unmap path, pages might be scanned by reclaim
442 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
443 * free them. This will result in freeing mlocked pages.
444 */
445void munlock_vma_pages_range(struct vm_area_struct *vma,
446 unsigned long start, unsigned long end)
447{
448 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
449
450 while (start < end) {
451 struct page *page;
452 unsigned int page_mask = 0;
453 unsigned long page_increm;
454 struct pagevec pvec;
455 struct zone *zone;
456
457 pagevec_init(&pvec);
458 /*
459 * Although FOLL_DUMP is intended for get_dump_page(),
460 * it just so happens that its special treatment of the
461 * ZERO_PAGE (returning an error instead of doing get_page)
462 * suits munlock very well (and if somehow an abnormal page
463 * has sneaked into the range, we won't oops here: great).
464 */
465 page = follow_page(vma, start, FOLL_GET | FOLL_DUMP);
466
467 if (page && !IS_ERR(page)) {
468 if (PageTransTail(page)) {
469 VM_BUG_ON_PAGE(PageMlocked(page), page);
470 put_page(page); /* follow_page_mask() */
471 } else if (PageTransHuge(page)) {
472 lock_page(page);
473 /*
474 * Any THP page found by follow_page_mask() may
475 * have gotten split before reaching
476 * munlock_vma_page(), so we need to compute
477 * the page_mask here instead.
478 */
479 page_mask = munlock_vma_page(page);
480 unlock_page(page);
481 put_page(page); /* follow_page_mask() */
482 } else {
483 /*
484 * Non-huge pages are handled in batches via
485 * pagevec. The pin from follow_page_mask()
486 * prevents them from collapsing by THP.
487 */
488 pagevec_add(&pvec, page);
489 zone = page_zone(page);
490
491 /*
492 * Try to fill the rest of pagevec using fast
493 * pte walk. This will also update start to
494 * the next page to process. Then munlock the
495 * pagevec.
496 */
497 start = __munlock_pagevec_fill(&pvec, vma,
498 zone, start, end);
499 __munlock_pagevec(&pvec, zone);
500 goto next;
501 }
502 }
503 page_increm = 1 + page_mask;
504 start += page_increm * PAGE_SIZE;
505next:
506 cond_resched();
507 }
508}
509
510/*
511 * mlock_fixup - handle mlock[all]/munlock[all] requests.
512 *
513 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
514 * munlock is a no-op. However, for some special vmas, we go ahead and
515 * populate the ptes.
516 *
517 * For vmas that pass the filters, merge/split as appropriate.
518 */
519static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
520 unsigned long start, unsigned long end, vm_flags_t newflags)
521{
522 struct mm_struct *mm = vma->vm_mm;
523 pgoff_t pgoff;
524 int nr_pages;
525 int ret = 0;
526 int lock = !!(newflags & VM_LOCKED);
527 vm_flags_t old_flags = vma->vm_flags;
528
529 if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
530 is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm) ||
531 vma_is_dax(vma))
532 /* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
533 goto out;
534
535 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
536 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
537 vma->vm_file, pgoff, vma_policy(vma),
538 vma->vm_userfaultfd_ctx);
539 if (*prev) {
540 vma = *prev;
541 goto success;
542 }
543
544 if (start != vma->vm_start) {
545 ret = split_vma(mm, vma, start, 1);
546 if (ret)
547 goto out;
548 }
549
550 if (end != vma->vm_end) {
551 ret = split_vma(mm, vma, end, 0);
552 if (ret)
553 goto out;
554 }
555
556success:
557 /*
558 * Keep track of amount of locked VM.
559 */
560 nr_pages = (end - start) >> PAGE_SHIFT;
561 if (!lock)
562 nr_pages = -nr_pages;
563 else if (old_flags & VM_LOCKED)
564 nr_pages = 0;
565 mm->locked_vm += nr_pages;
566
567 /*
568 * vm_flags is protected by the mmap_sem held in write mode.
569 * It's okay if try_to_unmap_one unmaps a page just after we
570 * set VM_LOCKED, populate_vma_page_range will bring it back.
571 */
572
573 if (lock)
574 vma->vm_flags = newflags;
575 else
576 munlock_vma_pages_range(vma, start, end);
577
578out:
579 *prev = vma;
580 return ret;
581}
582
583static int apply_vma_lock_flags(unsigned long start, size_t len,
584 vm_flags_t flags)
585{
586 unsigned long nstart, end, tmp;
587 struct vm_area_struct * vma, * prev;
588 int error;
589
590 VM_BUG_ON(offset_in_page(start));
591 VM_BUG_ON(len != PAGE_ALIGN(len));
592 end = start + len;
593 if (end < start)
594 return -EINVAL;
595 if (end == start)
596 return 0;
597 vma = find_vma(current->mm, start);
598 if (!vma || vma->vm_start > start)
599 return -ENOMEM;
600
601 prev = vma->vm_prev;
602 if (start > vma->vm_start)
603 prev = vma;
604
605 for (nstart = start ; ; ) {
606 vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
607
608 newflags |= flags;
609
610 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
611 tmp = vma->vm_end;
612 if (tmp > end)
613 tmp = end;
614 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
615 if (error)
616 break;
617 nstart = tmp;
618 if (nstart < prev->vm_end)
619 nstart = prev->vm_end;
620 if (nstart >= end)
621 break;
622
623 vma = prev->vm_next;
624 if (!vma || vma->vm_start != nstart) {
625 error = -ENOMEM;
626 break;
627 }
628 }
629 return error;
630}
631
632/*
633 * Go through vma areas and sum size of mlocked
634 * vma pages, as return value.
635 * Note deferred memory locking case(mlock2(,,MLOCK_ONFAULT)
636 * is also counted.
637 * Return value: previously mlocked page counts
638 */
639static unsigned long count_mm_mlocked_page_nr(struct mm_struct *mm,
640 unsigned long start, size_t len)
641{
642 struct vm_area_struct *vma;
643 unsigned long count = 0;
644
645 if (mm == NULL)
646 mm = current->mm;
647
648 vma = find_vma(mm, start);
649 if (vma == NULL)
650 vma = mm->mmap;
651
652 for (; vma ; vma = vma->vm_next) {
653 if (start >= vma->vm_end)
654 continue;
655 if (start + len <= vma->vm_start)
656 break;
657 if (vma->vm_flags & VM_LOCKED) {
658 if (start > vma->vm_start)
659 count -= (start - vma->vm_start);
660 if (start + len < vma->vm_end) {
661 count += start + len - vma->vm_start;
662 break;
663 }
664 count += vma->vm_end - vma->vm_start;
665 }
666 }
667
668 return count >> PAGE_SHIFT;
669}
670
671static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
672{
673 unsigned long locked;
674 unsigned long lock_limit;
675 int error = -ENOMEM;
676
677 start = untagged_addr(start);
678
679 if (!can_do_mlock())
680 return -EPERM;
681
682 len = PAGE_ALIGN(len + (offset_in_page(start)));
683 start &= PAGE_MASK;
684
685 lock_limit = rlimit(RLIMIT_MEMLOCK);
686 lock_limit >>= PAGE_SHIFT;
687 locked = len >> PAGE_SHIFT;
688
689 if (down_write_killable(¤t->mm->mmap_sem))
690 return -EINTR;
691
692 locked += current->mm->locked_vm;
693 if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
694 /*
695 * It is possible that the regions requested intersect with
696 * previously mlocked areas, that part area in "mm->locked_vm"
697 * should not be counted to new mlock increment count. So check
698 * and adjust locked count if necessary.
699 */
700 locked -= count_mm_mlocked_page_nr(current->mm,
701 start, len);
702 }
703
704 /* check against resource limits */
705 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
706 error = apply_vma_lock_flags(start, len, flags);
707
708 up_write(¤t->mm->mmap_sem);
709 if (error)
710 return error;
711
712 error = __mm_populate(start, len, 0);
713 if (error)
714 return __mlock_posix_error_return(error);
715 return 0;
716}
717
718SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
719{
720 return do_mlock(start, len, VM_LOCKED);
721}
722
723SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
724{
725 vm_flags_t vm_flags = VM_LOCKED;
726
727 if (flags & ~MLOCK_ONFAULT)
728 return -EINVAL;
729
730 if (flags & MLOCK_ONFAULT)
731 vm_flags |= VM_LOCKONFAULT;
732
733 return do_mlock(start, len, vm_flags);
734}
735
736SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
737{
738 int ret;
739
740 start = untagged_addr(start);
741
742 len = PAGE_ALIGN(len + (offset_in_page(start)));
743 start &= PAGE_MASK;
744
745 if (down_write_killable(¤t->mm->mmap_sem))
746 return -EINTR;
747 ret = apply_vma_lock_flags(start, len, 0);
748 up_write(¤t->mm->mmap_sem);
749
750 return ret;
751}
752
753/*
754 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
755 * and translate into the appropriate modifications to mm->def_flags and/or the
756 * flags for all current VMAs.
757 *
758 * There are a couple of subtleties with this. If mlockall() is called multiple
759 * times with different flags, the values do not necessarily stack. If mlockall
760 * is called once including the MCL_FUTURE flag and then a second time without
761 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
762 */
763static int apply_mlockall_flags(int flags)
764{
765 struct vm_area_struct * vma, * prev = NULL;
766 vm_flags_t to_add = 0;
767
768 current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
769 if (flags & MCL_FUTURE) {
770 current->mm->def_flags |= VM_LOCKED;
771
772 if (flags & MCL_ONFAULT)
773 current->mm->def_flags |= VM_LOCKONFAULT;
774
775 if (!(flags & MCL_CURRENT))
776 goto out;
777 }
778
779 if (flags & MCL_CURRENT) {
780 to_add |= VM_LOCKED;
781 if (flags & MCL_ONFAULT)
782 to_add |= VM_LOCKONFAULT;
783 }
784
785 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
786 vm_flags_t newflags;
787
788 newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
789 newflags |= to_add;
790
791 /* Ignore errors */
792 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
793 cond_resched();
794 }
795out:
796 return 0;
797}
798
799SYSCALL_DEFINE1(mlockall, int, flags)
800{
801 unsigned long lock_limit;
802 int ret;
803
804 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)) ||
805 flags == MCL_ONFAULT)
806 return -EINVAL;
807
808 if (!can_do_mlock())
809 return -EPERM;
810
811 lock_limit = rlimit(RLIMIT_MEMLOCK);
812 lock_limit >>= PAGE_SHIFT;
813
814 if (down_write_killable(¤t->mm->mmap_sem))
815 return -EINTR;
816
817 ret = -ENOMEM;
818 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
819 capable(CAP_IPC_LOCK))
820 ret = apply_mlockall_flags(flags);
821 up_write(¤t->mm->mmap_sem);
822 if (!ret && (flags & MCL_CURRENT))
823 mm_populate(0, TASK_SIZE);
824
825 return ret;
826}
827
828SYSCALL_DEFINE0(munlockall)
829{
830 int ret;
831
832 if (down_write_killable(¤t->mm->mmap_sem))
833 return -EINTR;
834 ret = apply_mlockall_flags(0);
835 up_write(¤t->mm->mmap_sem);
836 return ret;
837}
838
839/*
840 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
841 * shm segments) get accounted against the user_struct instead.
842 */
843static DEFINE_SPINLOCK(shmlock_user_lock);
844
845int user_shm_lock(size_t size, struct user_struct *user)
846{
847 unsigned long lock_limit, locked;
848 int allowed = 0;
849
850 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
851 lock_limit = rlimit(RLIMIT_MEMLOCK);
852 if (lock_limit == RLIM_INFINITY)
853 allowed = 1;
854 lock_limit >>= PAGE_SHIFT;
855 spin_lock(&shmlock_user_lock);
856 if (!allowed &&
857 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
858 goto out;
859 get_uid(user);
860 user->locked_shm += locked;
861 allowed = 1;
862out:
863 spin_unlock(&shmlock_user_lock);
864 return allowed;
865}
866
867void user_shm_unlock(size_t size, struct user_struct *user)
868{
869 spin_lock(&shmlock_user_lock);
870 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
871 spin_unlock(&shmlock_user_lock);
872 free_uid(user);
873}