blob: c83896a725042d45694777305107fd33ba9c9c28 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/mlock.c
3 *
4 * (C) Copyright 1995 Linus Torvalds
5 * (C) Copyright 2002 Christoph Hellwig
6 */
7
Randy.Dunlapc59ede72006-01-11 12:17:46 -08008#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/mman.h>
10#include <linux/mm.h>
Nick Pigginb291f002008-10-18 20:26:44 -070011#include <linux/swap.h>
12#include <linux/swapops.h>
13#include <linux/pagemap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/mempolicy.h>
15#include <linux/syscalls.h>
Alexey Dobriyane8edc6e2007-05-21 01:22:52 +040016#include <linux/sched.h>
17#include <linux/module.h>
Nick Pigginb291f002008-10-18 20:26:44 -070018#include <linux/rmap.h>
19#include <linux/mmzone.h>
20#include <linux/hugetlb.h>
21
22#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070023
Alexey Dobriyane8edc6e2007-05-21 01:22:52 +040024int can_do_mlock(void)
25{
26 if (capable(CAP_IPC_LOCK))
27 return 1;
28 if (current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur != 0)
29 return 1;
30 return 0;
31}
32EXPORT_SYMBOL(can_do_mlock);
Linus Torvalds1da177e2005-04-16 15:20:36 -070033
Nick Pigginb291f002008-10-18 20:26:44 -070034#ifdef CONFIG_UNEVICTABLE_LRU
35/*
36 * Mlocked pages are marked with PageMlocked() flag for efficient testing
37 * in vmscan and, possibly, the fault path; and to support semi-accurate
38 * statistics.
39 *
40 * An mlocked page [PageMlocked(page)] is unevictable. As such, it will
41 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
42 * The unevictable list is an LRU sibling list to the [in]active lists.
43 * PageUnevictable is set to indicate the unevictable state.
44 *
45 * When lazy mlocking via vmscan, it is important to ensure that the
46 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
47 * may have mlocked a page that is being munlocked. So lazy mlock must take
48 * the mmap_sem for read, and verify that the vma really is locked
49 * (see mm/rmap.c).
50 */
51
52/*
53 * LRU accounting for clear_page_mlock()
54 */
55void __clear_page_mlock(struct page *page)
56{
57 VM_BUG_ON(!PageLocked(page));
58
59 if (!page->mapping) { /* truncated ? */
60 return;
61 }
62
63 if (!isolate_lru_page(page)) {
64 putback_lru_page(page);
65 } else {
66 /*
67 * Page not on the LRU yet. Flush all pagevecs and retry.
68 */
69 lru_add_drain_all();
70 if (!isolate_lru_page(page))
71 putback_lru_page(page);
72 }
73}
74
75/*
76 * Mark page as mlocked if not already.
77 * If page on LRU, isolate and putback to move to unevictable list.
78 */
79void mlock_vma_page(struct page *page)
80{
81 BUG_ON(!PageLocked(page));
82
83 if (!TestSetPageMlocked(page) && !isolate_lru_page(page))
84 putback_lru_page(page);
85}
86
87/*
88 * called from munlock()/munmap() path with page supposedly on the LRU.
89 *
90 * Note: unlike mlock_vma_page(), we can't just clear the PageMlocked
91 * [in try_to_munlock()] and then attempt to isolate the page. We must
92 * isolate the page to keep others from messing with its unevictable
93 * and mlocked state while trying to munlock. However, we pre-clear the
94 * mlocked state anyway as we might lose the isolation race and we might
95 * not get another chance to clear PageMlocked. If we successfully
96 * isolate the page and try_to_munlock() detects other VM_LOCKED vmas
97 * mapping the page, it will restore the PageMlocked state, unless the page
98 * is mapped in a non-linear vma. So, we go ahead and SetPageMlocked(),
99 * perhaps redundantly.
100 * If we lose the isolation race, and the page is mapped by other VM_LOCKED
101 * vmas, we'll detect this in vmscan--via try_to_munlock() or try_to_unmap()
102 * either of which will restore the PageMlocked state by calling
103 * mlock_vma_page() above, if it can grab the vma's mmap sem.
104 */
105static void munlock_vma_page(struct page *page)
106{
107 BUG_ON(!PageLocked(page));
108
109 if (TestClearPageMlocked(page) && !isolate_lru_page(page)) {
110 try_to_munlock(page);
111 putback_lru_page(page);
112 }
113}
114
115/*
116 * mlock a range of pages in the vma.
117 *
118 * This takes care of making the pages present too.
119 *
120 * vma->vm_mm->mmap_sem must be held for write.
121 */
122static int __mlock_vma_pages_range(struct vm_area_struct *vma,
123 unsigned long start, unsigned long end)
124{
125 struct mm_struct *mm = vma->vm_mm;
126 unsigned long addr = start;
127 struct page *pages[16]; /* 16 gives a reasonable batch */
128 int write = !!(vma->vm_flags & VM_WRITE);
129 int nr_pages = (end - start) / PAGE_SIZE;
130 int ret;
131
132 VM_BUG_ON(start & ~PAGE_MASK || end & ~PAGE_MASK);
133 VM_BUG_ON(start < vma->vm_start || end > vma->vm_end);
134 VM_BUG_ON(!rwsem_is_locked(&vma->vm_mm->mmap_sem));
135
136 lru_add_drain_all(); /* push cached pages to LRU */
137
138 while (nr_pages > 0) {
139 int i;
140
141 cond_resched();
142
143 /*
144 * get_user_pages makes pages present if we are
145 * setting mlock. and this extra reference count will
146 * disable migration of this page. However, page may
147 * still be truncated out from under us.
148 */
149 ret = get_user_pages(current, mm, addr,
150 min_t(int, nr_pages, ARRAY_SIZE(pages)),
151 write, 0, pages, NULL);
152 /*
153 * This can happen for, e.g., VM_NONLINEAR regions before
154 * a page has been allocated and mapped at a given offset,
155 * or for addresses that map beyond end of a file.
156 * We'll mlock the the pages if/when they get faulted in.
157 */
158 if (ret < 0)
159 break;
160 if (ret == 0) {
161 /*
162 * We know the vma is there, so the only time
163 * we cannot get a single page should be an
164 * error (ret < 0) case.
165 */
166 WARN_ON(1);
167 break;
168 }
169
170 lru_add_drain(); /* push cached pages to LRU */
171
172 for (i = 0; i < ret; i++) {
173 struct page *page = pages[i];
174
175 lock_page(page);
176 /*
177 * Because we lock page here and migration is blocked
178 * by the elevated reference, we need only check for
179 * page truncation (file-cache only).
180 */
181 if (page->mapping)
182 mlock_vma_page(page);
183 unlock_page(page);
184 put_page(page); /* ref from get_user_pages() */
185
186 /*
187 * here we assume that get_user_pages() has given us
188 * a list of virtually contiguous pages.
189 */
190 addr += PAGE_SIZE; /* for next get_user_pages() */
191 nr_pages--;
192 }
193 }
194
195 lru_add_drain_all(); /* to update stats */
196
197 return 0; /* count entire vma as locked_vm */
198}
199
200/*
201 * private structure for munlock page table walk
202 */
203struct munlock_page_walk {
204 struct vm_area_struct *vma;
205 pmd_t *pmd; /* for migration_entry_wait() */
206};
207
208/*
209 * munlock normal pages for present ptes
210 */
211static int __munlock_pte_handler(pte_t *ptep, unsigned long addr,
212 unsigned long end, struct mm_walk *walk)
213{
214 struct munlock_page_walk *mpw = walk->private;
215 swp_entry_t entry;
216 struct page *page;
217 pte_t pte;
218
219retry:
220 pte = *ptep;
221 /*
222 * If it's a swap pte, we might be racing with page migration.
223 */
224 if (unlikely(!pte_present(pte))) {
225 if (!is_swap_pte(pte))
226 goto out;
227 entry = pte_to_swp_entry(pte);
228 if (is_migration_entry(entry)) {
229 migration_entry_wait(mpw->vma->vm_mm, mpw->pmd, addr);
230 goto retry;
231 }
232 goto out;
233 }
234
235 page = vm_normal_page(mpw->vma, addr, pte);
236 if (!page)
237 goto out;
238
239 lock_page(page);
240 if (!page->mapping) {
241 unlock_page(page);
242 goto retry;
243 }
244 munlock_vma_page(page);
245 unlock_page(page);
246
247out:
248 return 0;
249}
250
251/*
252 * Save pmd for pte handler for waiting on migration entries
253 */
254static int __munlock_pmd_handler(pmd_t *pmd, unsigned long addr,
255 unsigned long end, struct mm_walk *walk)
256{
257 struct munlock_page_walk *mpw = walk->private;
258
259 mpw->pmd = pmd;
260 return 0;
261}
262
263
264/*
265 * munlock a range of pages in the vma using standard page table walk.
266 *
267 * vma->vm_mm->mmap_sem must be held for write.
268 */
269static void __munlock_vma_pages_range(struct vm_area_struct *vma,
270 unsigned long start, unsigned long end)
271{
272 struct mm_struct *mm = vma->vm_mm;
273 struct munlock_page_walk mpw = {
274 .vma = vma,
275 };
276 struct mm_walk munlock_page_walk = {
277 .pmd_entry = __munlock_pmd_handler,
278 .pte_entry = __munlock_pte_handler,
279 .private = &mpw,
280 .mm = mm,
281 };
282
283 VM_BUG_ON(start & ~PAGE_MASK || end & ~PAGE_MASK);
284 VM_BUG_ON(!rwsem_is_locked(&vma->vm_mm->mmap_sem));
285 VM_BUG_ON(start < vma->vm_start);
286 VM_BUG_ON(end > vma->vm_end);
287
288 lru_add_drain_all(); /* push cached pages to LRU */
289 walk_page_range(start, end, &munlock_page_walk);
290 lru_add_drain_all(); /* to update stats */
291}
292
293#else /* CONFIG_UNEVICTABLE_LRU */
294
295/*
296 * Just make pages present if VM_LOCKED. No-op if unlocking.
297 */
298static int __mlock_vma_pages_range(struct vm_area_struct *vma,
299 unsigned long start, unsigned long end)
300{
301 if (vma->vm_flags & VM_LOCKED)
302 make_pages_present(start, end);
303 return 0;
304}
305
306/*
307 * munlock a range of pages in the vma -- no-op.
308 */
309static void __munlock_vma_pages_range(struct vm_area_struct *vma,
310 unsigned long start, unsigned long end)
311{
312}
313#endif /* CONFIG_UNEVICTABLE_LRU */
314
315/*
316 * mlock all pages in this vma range. For mmap()/mremap()/...
317 */
318int mlock_vma_pages_range(struct vm_area_struct *vma,
319 unsigned long start, unsigned long end)
320{
Lee Schermerhorn8edb08c2008-10-18 20:26:49 -0700321 struct mm_struct *mm = vma->vm_mm;
Nick Pigginb291f002008-10-18 20:26:44 -0700322 int nr_pages = (end - start) / PAGE_SIZE;
323 BUG_ON(!(vma->vm_flags & VM_LOCKED));
324
325 /*
326 * filter unlockable vmas
327 */
328 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
329 goto no_mlock;
330
331 if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
332 is_vm_hugetlb_page(vma) ||
Lee Schermerhorn8edb08c2008-10-18 20:26:49 -0700333 vma == get_gate_vma(current))) {
334 downgrade_write(&mm->mmap_sem);
335 nr_pages = __mlock_vma_pages_range(vma, start, end);
336
337 up_read(&mm->mmap_sem);
338 /* vma can change or disappear */
339 down_write(&mm->mmap_sem);
340 vma = find_vma(mm, start);
341 /* non-NULL vma must contain @start, but need to check @end */
342 if (!vma || end > vma->vm_end)
343 return -EAGAIN;
344 return nr_pages;
345 }
Nick Pigginb291f002008-10-18 20:26:44 -0700346
347 /*
348 * User mapped kernel pages or huge pages:
349 * make these pages present to populate the ptes, but
350 * fall thru' to reset VM_LOCKED--no need to unlock, and
351 * return nr_pages so these don't get counted against task's
352 * locked limit. huge pages are already counted against
353 * locked vm limit.
354 */
355 make_pages_present(start, end);
356
357no_mlock:
358 vma->vm_flags &= ~VM_LOCKED; /* and don't come back! */
359 return nr_pages; /* pages NOT mlocked */
360}
361
362
363/*
364 * munlock all pages in vma. For munmap() and exit().
365 */
366void munlock_vma_pages_all(struct vm_area_struct *vma)
367{
368 vma->vm_flags &= ~VM_LOCKED;
369 __munlock_vma_pages_range(vma, vma->vm_start, vma->vm_end);
370}
371
372/*
373 * mlock_fixup - handle mlock[all]/munlock[all] requests.
374 *
375 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
376 * munlock is a no-op. However, for some special vmas, we go ahead and
377 * populate the ptes via make_pages_present().
378 *
379 * For vmas that pass the filters, merge/split as appropriate.
380 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
382 unsigned long start, unsigned long end, unsigned int newflags)
383{
Nick Pigginb291f002008-10-18 20:26:44 -0700384 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 pgoff_t pgoff;
Nick Pigginb291f002008-10-18 20:26:44 -0700386 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 int ret = 0;
Nick Pigginb291f002008-10-18 20:26:44 -0700388 int lock = newflags & VM_LOCKED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
Nick Pigginb291f002008-10-18 20:26:44 -0700390 if (newflags == vma->vm_flags ||
391 (vma->vm_flags & (VM_IO | VM_PFNMAP)))
392 goto out; /* don't set VM_LOCKED, don't count */
393
394 if ((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
395 is_vm_hugetlb_page(vma) ||
396 vma == get_gate_vma(current)) {
397 if (lock)
398 make_pages_present(start, end);
399 goto out; /* don't set VM_LOCKED, don't count */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 }
401
402 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
403 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
404 vma->vm_file, pgoff, vma_policy(vma));
405 if (*prev) {
406 vma = *prev;
407 goto success;
408 }
409
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 if (start != vma->vm_start) {
411 ret = split_vma(mm, vma, start, 1);
412 if (ret)
413 goto out;
414 }
415
416 if (end != vma->vm_end) {
417 ret = split_vma(mm, vma, end, 0);
418 if (ret)
419 goto out;
420 }
421
422success:
423 /*
Nick Pigginb291f002008-10-18 20:26:44 -0700424 * Keep track of amount of locked VM.
425 */
426 nr_pages = (end - start) >> PAGE_SHIFT;
427 if (!lock)
428 nr_pages = -nr_pages;
429 mm->locked_vm += nr_pages;
430
431 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 * vm_flags is protected by the mmap_sem held in write mode.
433 * It's okay if try_to_unmap_one unmaps a page just after we
Nick Pigginb291f002008-10-18 20:26:44 -0700434 * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 */
436 vma->vm_flags = newflags;
437
Nick Pigginb291f002008-10-18 20:26:44 -0700438 if (lock) {
Lee Schermerhorn8edb08c2008-10-18 20:26:49 -0700439 /*
440 * mmap_sem is currently held for write. Downgrade the write
441 * lock to a read lock so that other faults, mmap scans, ...
442 * while we fault in all pages.
443 */
444 downgrade_write(&mm->mmap_sem);
445
Nick Pigginb291f002008-10-18 20:26:44 -0700446 ret = __mlock_vma_pages_range(vma, start, end);
447 if (ret > 0) {
448 mm->locked_vm -= ret;
449 ret = 0;
450 }
Lee Schermerhorn8edb08c2008-10-18 20:26:49 -0700451 /*
452 * Need to reacquire mmap sem in write mode, as our callers
453 * expect this. We have no support for atomically upgrading
454 * a sem to write, so we need to check for ranges while sem
455 * is unlocked.
456 */
457 up_read(&mm->mmap_sem);
458 /* vma can change or disappear */
459 down_write(&mm->mmap_sem);
460 *prev = find_vma(mm, start);
461 /* non-NULL *prev must contain @start, but need to check @end */
462 if (!(*prev) || end > (*prev)->vm_end)
463 ret = -EAGAIN;
464 } else {
465 /*
466 * TODO: for unlocking, pages will already be resident, so
467 * we don't need to wait for allocations/reclaim/pagein, ...
468 * However, unlocking a very large region can still take a
469 * while. Should we downgrade the semaphore for both lock
470 * AND unlock ?
471 */
Nick Pigginb291f002008-10-18 20:26:44 -0700472 __munlock_vma_pages_range(vma, start, end);
Lee Schermerhorn8edb08c2008-10-18 20:26:49 -0700473 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475out:
Nick Pigginb291f002008-10-18 20:26:44 -0700476 *prev = vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 return ret;
478}
479
480static int do_mlock(unsigned long start, size_t len, int on)
481{
482 unsigned long nstart, end, tmp;
483 struct vm_area_struct * vma, * prev;
484 int error;
485
486 len = PAGE_ALIGN(len);
487 end = start + len;
488 if (end < start)
489 return -EINVAL;
490 if (end == start)
491 return 0;
492 vma = find_vma_prev(current->mm, start, &prev);
493 if (!vma || vma->vm_start > start)
494 return -ENOMEM;
495
496 if (start > vma->vm_start)
497 prev = vma;
498
499 for (nstart = start ; ; ) {
500 unsigned int newflags;
501
502 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
503
504 newflags = vma->vm_flags | VM_LOCKED;
505 if (!on)
506 newflags &= ~VM_LOCKED;
507
508 tmp = vma->vm_end;
509 if (tmp > end)
510 tmp = end;
511 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
512 if (error)
513 break;
514 nstart = tmp;
515 if (nstart < prev->vm_end)
516 nstart = prev->vm_end;
517 if (nstart >= end)
518 break;
519
520 vma = prev->vm_next;
521 if (!vma || vma->vm_start != nstart) {
522 error = -ENOMEM;
523 break;
524 }
525 }
526 return error;
527}
528
529asmlinkage long sys_mlock(unsigned long start, size_t len)
530{
531 unsigned long locked;
532 unsigned long lock_limit;
533 int error = -ENOMEM;
534
535 if (!can_do_mlock())
536 return -EPERM;
537
538 down_write(&current->mm->mmap_sem);
539 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
540 start &= PAGE_MASK;
541
542 locked = len >> PAGE_SHIFT;
543 locked += current->mm->locked_vm;
544
545 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
546 lock_limit >>= PAGE_SHIFT;
547
548 /* check against resource limits */
549 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
550 error = do_mlock(start, len, 1);
551 up_write(&current->mm->mmap_sem);
552 return error;
553}
554
555asmlinkage long sys_munlock(unsigned long start, size_t len)
556{
557 int ret;
558
559 down_write(&current->mm->mmap_sem);
560 len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
561 start &= PAGE_MASK;
562 ret = do_mlock(start, len, 0);
563 up_write(&current->mm->mmap_sem);
564 return ret;
565}
566
567static int do_mlockall(int flags)
568{
569 struct vm_area_struct * vma, * prev = NULL;
570 unsigned int def_flags = 0;
571
572 if (flags & MCL_FUTURE)
573 def_flags = VM_LOCKED;
574 current->mm->def_flags = def_flags;
575 if (flags == MCL_FUTURE)
576 goto out;
577
578 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
579 unsigned int newflags;
580
581 newflags = vma->vm_flags | VM_LOCKED;
582 if (!(flags & MCL_CURRENT))
583 newflags &= ~VM_LOCKED;
584
585 /* Ignore errors */
586 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
587 }
588out:
589 return 0;
590}
591
592asmlinkage long sys_mlockall(int flags)
593{
594 unsigned long lock_limit;
595 int ret = -EINVAL;
596
597 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
598 goto out;
599
600 ret = -EPERM;
601 if (!can_do_mlock())
602 goto out;
603
604 down_write(&current->mm->mmap_sem);
605
606 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
607 lock_limit >>= PAGE_SHIFT;
608
609 ret = -ENOMEM;
610 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
611 capable(CAP_IPC_LOCK))
612 ret = do_mlockall(flags);
613 up_write(&current->mm->mmap_sem);
614out:
615 return ret;
616}
617
618asmlinkage long sys_munlockall(void)
619{
620 int ret;
621
622 down_write(&current->mm->mmap_sem);
623 ret = do_mlockall(0);
624 up_write(&current->mm->mmap_sem);
625 return ret;
626}
627
628/*
629 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
630 * shm segments) get accounted against the user_struct instead.
631 */
632static DEFINE_SPINLOCK(shmlock_user_lock);
633
634int user_shm_lock(size_t size, struct user_struct *user)
635{
636 unsigned long lock_limit, locked;
637 int allowed = 0;
638
639 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
640 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
Herbert van den Bergh5ed44a42007-07-15 23:38:25 -0700641 if (lock_limit == RLIM_INFINITY)
642 allowed = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 lock_limit >>= PAGE_SHIFT;
644 spin_lock(&shmlock_user_lock);
Herbert van den Bergh5ed44a42007-07-15 23:38:25 -0700645 if (!allowed &&
646 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 goto out;
648 get_uid(user);
649 user->locked_shm += locked;
650 allowed = 1;
651out:
652 spin_unlock(&shmlock_user_lock);
653 return allowed;
654}
655
656void user_shm_unlock(size_t size, struct user_struct *user)
657{
658 spin_lock(&shmlock_user_lock);
659 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
660 spin_unlock(&shmlock_user_lock);
661 free_uid(user);
662}