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