mlock: replace stale comments in munlock_vma_page()
[linux-2.6.git] / mm / mlock.c
1 /*
2  *      linux/mm/mlock.c
3  *
4  *  (C) Copyright 1995 Linus Torvalds
5  *  (C) Copyright 2002 Christoph Hellwig
6  */
7
8 #include <linux/capability.h>
9 #include <linux/mman.h>
10 #include <linux/mm.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/syscalls.h>
16 #include <linux/sched.h>
17 #include <linux/module.h>
18 #include <linux/rmap.h>
19 #include <linux/mmzone.h>
20 #include <linux/hugetlb.h>
21
22 #include "internal.h"
23
24 int 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 }
32 EXPORT_SYMBOL(can_do_mlock);
33
34 /*
35  * Mlocked pages are marked with PageMlocked() flag for efficient testing
36  * in vmscan and, possibly, the fault path; and to support semi-accurate
37  * statistics.
38  *
39  * An mlocked page [PageMlocked(page)] is unevictable.  As such, it will
40  * be placed on the LRU "unevictable" list, rather than the [in]active lists.
41  * The unevictable list is an LRU sibling list to the [in]active lists.
42  * PageUnevictable is set to indicate the unevictable state.
43  *
44  * When lazy mlocking via vmscan, it is important to ensure that the
45  * vma's VM_LOCKED status is not concurrently being modified, otherwise we
46  * may have mlocked a page that is being munlocked. So lazy mlock must take
47  * the mmap_sem for read, and verify that the vma really is locked
48  * (see mm/rmap.c).
49  */
50
51 /*
52  *  LRU accounting for clear_page_mlock()
53  */
54 void __clear_page_mlock(struct page *page)
55 {
56         VM_BUG_ON(!PageLocked(page));
57
58         if (!page->mapping) {   /* truncated ? */
59                 return;
60         }
61
62         dec_zone_page_state(page, NR_MLOCK);
63         count_vm_event(UNEVICTABLE_PGCLEARED);
64         if (!isolate_lru_page(page)) {
65                 putback_lru_page(page);
66         } else {
67                 /*
68                  * We lost the race. the page already moved to evictable list.
69                  */
70                 if (PageUnevictable(page))
71                         count_vm_event(UNEVICTABLE_PGSTRANDED);
72         }
73 }
74
75 /*
76  * Mark page as mlocked if not already.
77  * If page on LRU, isolate and putback to move to unevictable list.
78  */
79 void mlock_vma_page(struct page *page)
80 {
81         BUG_ON(!PageLocked(page));
82
83         if (!TestSetPageMlocked(page)) {
84                 inc_zone_page_state(page, NR_MLOCK);
85                 count_vm_event(UNEVICTABLE_PGMLOCKED);
86                 if (!isolate_lru_page(page))
87                         putback_lru_page(page);
88         }
89 }
90
91 /**
92  * munlock_vma_page - munlock a vma page
93  * @page - page to be unlocked
94  *
95  * called from munlock()/munmap() path with page supposedly on the LRU.
96  * When we munlock a page, because the vma where we found the page is being
97  * munlock()ed or munmap()ed, we want to check whether other vmas hold the
98  * page locked so that we can leave it on the unevictable lru list and not
99  * bother vmscan with it.  However, to walk the page's rmap list in
100  * try_to_munlock() we must isolate the page from the LRU.  If some other
101  * task has removed the page from the LRU, we won't be able to do that.
102  * So we clear the PageMlocked as we might not get another chance.  If we
103  * can't isolate the page, we leave it for putback_lru_page() and vmscan
104  * [page_referenced()/try_to_unmap()] to deal with.
105  */
106 void munlock_vma_page(struct page *page)
107 {
108         BUG_ON(!PageLocked(page));
109
110         if (TestClearPageMlocked(page)) {
111                 dec_zone_page_state(page, NR_MLOCK);
112                 if (!isolate_lru_page(page)) {
113                         int ret = try_to_munlock(page);
114                         /*
115                          * did try_to_unlock() succeed or punt?
116                          */
117                         if (ret != SWAP_MLOCK)
118                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
119
120                         putback_lru_page(page);
121                 } else {
122                         /*
123                          * Some other task has removed the page from the LRU.
124                          * putback_lru_page() will take care of removing the
125                          * page from the unevictable list, if necessary.
126                          * vmscan [page_referenced()] will move the page back
127                          * to the unevictable list if some other vma has it
128                          * mlocked.
129                          */
130                         if (PageUnevictable(page))
131                                 count_vm_event(UNEVICTABLE_PGSTRANDED);
132                         else
133                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
134                 }
135         }
136 }
137
138 /**
139  * __mlock_vma_pages_range() -  mlock a range of pages in the vma.
140  * @vma:   target vma
141  * @start: start address
142  * @end:   end address
143  *
144  * This takes care of making the pages present too.
145  *
146  * return 0 on success, negative error code on error.
147  *
148  * vma->vm_mm->mmap_sem must be held for at least read.
149  */
150 static long __mlock_vma_pages_range(struct vm_area_struct *vma,
151                                     unsigned long start, unsigned long end)
152 {
153         struct mm_struct *mm = vma->vm_mm;
154         unsigned long addr = start;
155         struct page *pages[16]; /* 16 gives a reasonable batch */
156         int nr_pages = (end - start) / PAGE_SIZE;
157         int ret = 0;
158         int gup_flags;
159
160         VM_BUG_ON(start & ~PAGE_MASK);
161         VM_BUG_ON(end   & ~PAGE_MASK);
162         VM_BUG_ON(start < vma->vm_start);
163         VM_BUG_ON(end   > vma->vm_end);
164         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
165
166         gup_flags = FOLL_TOUCH | FOLL_GET;
167         if (vma->vm_flags & VM_WRITE)
168                 gup_flags |= FOLL_WRITE;
169
170         while (nr_pages > 0) {
171                 int i;
172
173                 cond_resched();
174
175                 /*
176                  * get_user_pages makes pages present if we are
177                  * setting mlock. and this extra reference count will
178                  * disable migration of this page.  However, page may
179                  * still be truncated out from under us.
180                  */
181                 ret = __get_user_pages(current, mm, addr,
182                                 min_t(int, nr_pages, ARRAY_SIZE(pages)),
183                                 gup_flags, pages, NULL);
184                 /*
185                  * This can happen for, e.g., VM_NONLINEAR regions before
186                  * a page has been allocated and mapped at a given offset,
187                  * or for addresses that map beyond end of a file.
188                  * We'll mlock the pages if/when they get faulted in.
189                  */
190                 if (ret < 0)
191                         break;
192
193                 lru_add_drain();        /* push cached pages to LRU */
194
195                 for (i = 0; i < ret; i++) {
196                         struct page *page = pages[i];
197
198                         if (page->mapping) {
199                                 /*
200                                  * That preliminary check is mainly to avoid
201                                  * the pointless overhead of lock_page on the
202                                  * ZERO_PAGE: which might bounce very badly if
203                                  * there is contention.  However, we're still
204                                  * dirtying its cacheline with get/put_page:
205                                  * we'll add another __get_user_pages flag to
206                                  * avoid it if that case turns out to matter.
207                                  */
208                                 lock_page(page);
209                                 /*
210                                  * Because we lock page here and migration is
211                                  * blocked by the elevated reference, we need
212                                  * only check for file-cache page truncation.
213                                  */
214                                 if (page->mapping)
215                                         mlock_vma_page(page);
216                                 unlock_page(page);
217                         }
218                         put_page(page); /* ref from get_user_pages() */
219                 }
220
221                 addr += ret * PAGE_SIZE;
222                 nr_pages -= ret;
223                 ret = 0;
224         }
225
226         return ret;     /* 0 or negative error code */
227 }
228
229 /*
230  * convert get_user_pages() return value to posix mlock() error
231  */
232 static int __mlock_posix_error_return(long retval)
233 {
234         if (retval == -EFAULT)
235                 retval = -ENOMEM;
236         else if (retval == -ENOMEM)
237                 retval = -EAGAIN;
238         return retval;
239 }
240
241 /**
242  * mlock_vma_pages_range() - mlock pages in specified vma range.
243  * @vma - the vma containing the specfied address range
244  * @start - starting address in @vma to mlock
245  * @end   - end address [+1] in @vma to mlock
246  *
247  * For mmap()/mremap()/expansion of mlocked vma.
248  *
249  * return 0 on success for "normal" vmas.
250  *
251  * return number of pages [> 0] to be removed from locked_vm on success
252  * of "special" vmas.
253  */
254 long mlock_vma_pages_range(struct vm_area_struct *vma,
255                         unsigned long start, unsigned long end)
256 {
257         int nr_pages = (end - start) / PAGE_SIZE;
258         BUG_ON(!(vma->vm_flags & VM_LOCKED));
259
260         /*
261          * filter unlockable vmas
262          */
263         if (vma->vm_flags & (VM_IO | VM_PFNMAP))
264                 goto no_mlock;
265
266         if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
267                         is_vm_hugetlb_page(vma) ||
268                         vma == get_gate_vma(current))) {
269
270                 __mlock_vma_pages_range(vma, start, end);
271
272                 /* Hide errors from mmap() and other callers */
273                 return 0;
274         }
275
276         /*
277          * User mapped kernel pages or huge pages:
278          * make these pages present to populate the ptes, but
279          * fall thru' to reset VM_LOCKED--no need to unlock, and
280          * return nr_pages so these don't get counted against task's
281          * locked limit.  huge pages are already counted against
282          * locked vm limit.
283          */
284         make_pages_present(start, end);
285
286 no_mlock:
287         vma->vm_flags &= ~VM_LOCKED;    /* and don't come back! */
288         return nr_pages;                /* error or pages NOT mlocked */
289 }
290
291 /*
292  * munlock_vma_pages_range() - munlock all pages in the vma range.'
293  * @vma - vma containing range to be munlock()ed.
294  * @start - start address in @vma of the range
295  * @end - end of range in @vma.
296  *
297  *  For mremap(), munmap() and exit().
298  *
299  * Called with @vma VM_LOCKED.
300  *
301  * Returns with VM_LOCKED cleared.  Callers must be prepared to
302  * deal with this.
303  *
304  * We don't save and restore VM_LOCKED here because pages are
305  * still on lru.  In unmap path, pages might be scanned by reclaim
306  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
307  * free them.  This will result in freeing mlocked pages.
308  */
309 void munlock_vma_pages_range(struct vm_area_struct *vma,
310                              unsigned long start, unsigned long end)
311 {
312         unsigned long addr;
313
314         lru_add_drain();
315         vma->vm_flags &= ~VM_LOCKED;
316
317         for (addr = start; addr < end; addr += PAGE_SIZE) {
318                 struct page *page;
319                 /*
320                  * Although FOLL_DUMP is intended for get_dump_page(),
321                  * it just so happens that its special treatment of the
322                  * ZERO_PAGE (returning an error instead of doing get_page)
323                  * suits munlock very well (and if somehow an abnormal page
324                  * has sneaked into the range, we won't oops here: great).
325                  */
326                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
327                 if (page && !IS_ERR(page)) {
328                         lock_page(page);
329                         /*
330                          * Like in __mlock_vma_pages_range(),
331                          * because we lock page here and migration is
332                          * blocked by the elevated reference, we need
333                          * only check for file-cache page truncation.
334                          */
335                         if (page->mapping)
336                                 munlock_vma_page(page);
337                         unlock_page(page);
338                         put_page(page);
339                 }
340                 cond_resched();
341         }
342 }
343
344 /*
345  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
346  *
347  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
348  * munlock is a no-op.  However, for some special vmas, we go ahead and
349  * populate the ptes via make_pages_present().
350  *
351  * For vmas that pass the filters, merge/split as appropriate.
352  */
353 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
354         unsigned long start, unsigned long end, unsigned int newflags)
355 {
356         struct mm_struct *mm = vma->vm_mm;
357         pgoff_t pgoff;
358         int nr_pages;
359         int ret = 0;
360         int lock = newflags & VM_LOCKED;
361
362         if (newflags == vma->vm_flags ||
363                         (vma->vm_flags & (VM_IO | VM_PFNMAP)))
364                 goto out;       /* don't set VM_LOCKED,  don't count */
365
366         if ((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
367                         is_vm_hugetlb_page(vma) ||
368                         vma == get_gate_vma(current)) {
369                 if (lock)
370                         make_pages_present(start, end);
371                 goto out;       /* don't set VM_LOCKED,  don't count */
372         }
373
374         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
375         *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
376                           vma->vm_file, pgoff, vma_policy(vma));
377         if (*prev) {
378                 vma = *prev;
379                 goto success;
380         }
381
382         if (start != vma->vm_start) {
383                 ret = split_vma(mm, vma, start, 1);
384                 if (ret)
385                         goto out;
386         }
387
388         if (end != vma->vm_end) {
389                 ret = split_vma(mm, vma, end, 0);
390                 if (ret)
391                         goto out;
392         }
393
394 success:
395         /*
396          * Keep track of amount of locked VM.
397          */
398         nr_pages = (end - start) >> PAGE_SHIFT;
399         if (!lock)
400                 nr_pages = -nr_pages;
401         mm->locked_vm += nr_pages;
402
403         /*
404          * vm_flags is protected by the mmap_sem held in write mode.
405          * It's okay if try_to_unmap_one unmaps a page just after we
406          * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
407          */
408
409         if (lock) {
410                 vma->vm_flags = newflags;
411                 ret = __mlock_vma_pages_range(vma, start, end);
412                 if (ret < 0)
413                         ret = __mlock_posix_error_return(ret);
414         } else {
415                 munlock_vma_pages_range(vma, start, end);
416         }
417
418 out:
419         *prev = vma;
420         return ret;
421 }
422
423 static int do_mlock(unsigned long start, size_t len, int on)
424 {
425         unsigned long nstart, end, tmp;
426         struct vm_area_struct * vma, * prev;
427         int error;
428
429         len = PAGE_ALIGN(len);
430         end = start + len;
431         if (end < start)
432                 return -EINVAL;
433         if (end == start)
434                 return 0;
435         vma = find_vma_prev(current->mm, start, &prev);
436         if (!vma || vma->vm_start > start)
437                 return -ENOMEM;
438
439         if (start > vma->vm_start)
440                 prev = vma;
441
442         for (nstart = start ; ; ) {
443                 unsigned int newflags;
444
445                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
446
447                 newflags = vma->vm_flags | VM_LOCKED;
448                 if (!on)
449                         newflags &= ~VM_LOCKED;
450
451                 tmp = vma->vm_end;
452                 if (tmp > end)
453                         tmp = end;
454                 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
455                 if (error)
456                         break;
457                 nstart = tmp;
458                 if (nstart < prev->vm_end)
459                         nstart = prev->vm_end;
460                 if (nstart >= end)
461                         break;
462
463                 vma = prev->vm_next;
464                 if (!vma || vma->vm_start != nstart) {
465                         error = -ENOMEM;
466                         break;
467                 }
468         }
469         return error;
470 }
471
472 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
473 {
474         unsigned long locked;
475         unsigned long lock_limit;
476         int error = -ENOMEM;
477
478         if (!can_do_mlock())
479                 return -EPERM;
480
481         lru_add_drain_all();    /* flush pagevec */
482
483         down_write(&current->mm->mmap_sem);
484         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
485         start &= PAGE_MASK;
486
487         locked = len >> PAGE_SHIFT;
488         locked += current->mm->locked_vm;
489
490         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
491         lock_limit >>= PAGE_SHIFT;
492
493         /* check against resource limits */
494         if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
495                 error = do_mlock(start, len, 1);
496         up_write(&current->mm->mmap_sem);
497         return error;
498 }
499
500 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
501 {
502         int ret;
503
504         down_write(&current->mm->mmap_sem);
505         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
506         start &= PAGE_MASK;
507         ret = do_mlock(start, len, 0);
508         up_write(&current->mm->mmap_sem);
509         return ret;
510 }
511
512 static int do_mlockall(int flags)
513 {
514         struct vm_area_struct * vma, * prev = NULL;
515         unsigned int def_flags = 0;
516
517         if (flags & MCL_FUTURE)
518                 def_flags = VM_LOCKED;
519         current->mm->def_flags = def_flags;
520         if (flags == MCL_FUTURE)
521                 goto out;
522
523         for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
524                 unsigned int newflags;
525
526                 newflags = vma->vm_flags | VM_LOCKED;
527                 if (!(flags & MCL_CURRENT))
528                         newflags &= ~VM_LOCKED;
529
530                 /* Ignore errors */
531                 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
532         }
533 out:
534         return 0;
535 }
536
537 SYSCALL_DEFINE1(mlockall, int, flags)
538 {
539         unsigned long lock_limit;
540         int ret = -EINVAL;
541
542         if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
543                 goto out;
544
545         ret = -EPERM;
546         if (!can_do_mlock())
547                 goto out;
548
549         lru_add_drain_all();    /* flush pagevec */
550
551         down_write(&current->mm->mmap_sem);
552
553         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
554         lock_limit >>= PAGE_SHIFT;
555
556         ret = -ENOMEM;
557         if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
558             capable(CAP_IPC_LOCK))
559                 ret = do_mlockall(flags);
560         up_write(&current->mm->mmap_sem);
561 out:
562         return ret;
563 }
564
565 SYSCALL_DEFINE0(munlockall)
566 {
567         int ret;
568
569         down_write(&current->mm->mmap_sem);
570         ret = do_mlockall(0);
571         up_write(&current->mm->mmap_sem);
572         return ret;
573 }
574
575 /*
576  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
577  * shm segments) get accounted against the user_struct instead.
578  */
579 static DEFINE_SPINLOCK(shmlock_user_lock);
580
581 int user_shm_lock(size_t size, struct user_struct *user)
582 {
583         unsigned long lock_limit, locked;
584         int allowed = 0;
585
586         locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
587         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
588         if (lock_limit == RLIM_INFINITY)
589                 allowed = 1;
590         lock_limit >>= PAGE_SHIFT;
591         spin_lock(&shmlock_user_lock);
592         if (!allowed &&
593             locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
594                 goto out;
595         get_uid(user);
596         user->locked_shm += locked;
597         allowed = 1;
598 out:
599         spin_unlock(&shmlock_user_lock);
600         return allowed;
601 }
602
603 void user_shm_unlock(size_t size, struct user_struct *user)
604 {
605         spin_lock(&shmlock_user_lock);
606         user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
607         spin_unlock(&shmlock_user_lock);
608         free_uid(user);
609 }
610
611 int account_locked_memory(struct mm_struct *mm, struct rlimit *rlim,
612                           size_t size)
613 {
614         unsigned long lim, vm, pgsz;
615         int error = -ENOMEM;
616
617         pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
618
619         down_write(&mm->mmap_sem);
620
621         lim = rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
622         vm   = mm->total_vm + pgsz;
623         if (lim < vm)
624                 goto out;
625
626         lim = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
627         vm   = mm->locked_vm + pgsz;
628         if (lim < vm)
629                 goto out;
630
631         mm->total_vm  += pgsz;
632         mm->locked_vm += pgsz;
633
634         error = 0;
635  out:
636         up_write(&mm->mmap_sem);
637         return error;
638 }
639
640 void refund_locked_memory(struct mm_struct *mm, size_t size)
641 {
642         unsigned long pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
643
644         down_write(&mm->mmap_sem);
645
646         mm->total_vm  -= pgsz;
647         mm->locked_vm -= pgsz;
648
649         up_write(&mm->mmap_sem);
650 }