mm: get_user_pages: migrate out CMA pages when FOLL_DURABLE flag is set
[linux-3.10.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/export.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 (rlimit(RLIMIT_MEMLOCK) != 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         if (!TestClearPageMlocked(page))
57                 return;
58
59         mod_zone_page_state(page_zone(page), NR_MLOCK,
60                             -hpage_nr_pages(page));
61         count_vm_event(UNEVICTABLE_PGCLEARED);
62         if (!isolate_lru_page(page)) {
63                 putback_lru_page(page);
64         } else {
65                 /*
66                  * We lost the race. the page already moved to evictable list.
67                  */
68                 if (PageUnevictable(page))
69                         count_vm_event(UNEVICTABLE_PGSTRANDED);
70         }
71 }
72
73 /*
74  * Mark page as mlocked if not already.
75  * If page on LRU, isolate and putback to move to unevictable list.
76  */
77 void mlock_vma_page(struct page *page)
78 {
79         /* Serialize with page migration */
80         BUG_ON(!PageLocked(page));
81
82         if (!TestSetPageMlocked(page)) {
83                 mod_zone_page_state(page_zone(page), NR_MLOCK,
84                                     hpage_nr_pages(page));
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 unsigned int munlock_vma_page(struct page *page)
107 {
108         unsigned int page_mask = 0;
109
110         /* For try_to_munlock() and to serialize with page migration */
111         BUG_ON(!PageLocked(page));
112
113         if (TestClearPageMlocked(page)) {
114                 unsigned int nr_pages = hpage_nr_pages(page);
115                 mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
116                 page_mask = nr_pages - 1;
117                 if (!isolate_lru_page(page)) {
118                         int ret = SWAP_AGAIN;
119
120                         /*
121                          * Optimization: if the page was mapped just once,
122                          * that's our mapping and we don't need to check all the
123                          * other vmas.
124                          */
125                         if (page_mapcount(page) > 1)
126                                 ret = try_to_munlock(page);
127                         /*
128                          * did try_to_unlock() succeed or punt?
129                          */
130                         if (ret != SWAP_MLOCK)
131                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
132
133                         putback_lru_page(page);
134                 } else {
135                         /*
136                          * Some other task has removed the page from the LRU.
137                          * putback_lru_page() will take care of removing the
138                          * page from the unevictable list, if necessary.
139                          * vmscan [page_referenced()] will move the page back
140                          * to the unevictable list if some other vma has it
141                          * mlocked.
142                          */
143                         if (PageUnevictable(page))
144                                 count_vm_event(UNEVICTABLE_PGSTRANDED);
145                         else
146                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
147                 }
148         }
149
150         return page_mask;
151 }
152
153 /**
154  * __mlock_vma_pages_range() -  mlock a range of pages in the vma.
155  * @vma:   target vma
156  * @start: start address
157  * @end:   end address
158  *
159  * This takes care of making the pages present too.
160  *
161  * return 0 on success, negative error code on error.
162  *
163  * vma->vm_mm->mmap_sem must be held for at least read.
164  */
165 long __mlock_vma_pages_range(struct vm_area_struct *vma,
166                 unsigned long start, unsigned long end, int *nonblocking)
167 {
168         struct mm_struct *mm = vma->vm_mm;
169         unsigned long nr_pages = (end - start) / PAGE_SIZE;
170         int gup_flags;
171
172         VM_BUG_ON(start & ~PAGE_MASK);
173         VM_BUG_ON(end   & ~PAGE_MASK);
174         VM_BUG_ON(start < vma->vm_start);
175         VM_BUG_ON(end   > vma->vm_end);
176         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
177
178         gup_flags = FOLL_TOUCH | FOLL_MLOCK;
179         /*
180          * We want to touch writable mappings with a write fault in order
181          * to break COW, except for shared mappings because these don't COW
182          * and we would not want to dirty them for nothing.
183          */
184         if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
185                 gup_flags |= FOLL_WRITE;
186
187         /*
188          * We want mlock to succeed for regions that have any permissions
189          * other than PROT_NONE.
190          */
191         if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
192                 gup_flags |= FOLL_FORCE;
193
194         /*
195          * We made sure addr is within a VMA, so the following will
196          * not result in a stack expansion that recurses back here.
197          */
198         return __get_user_pages(current, mm, start, nr_pages, gup_flags,
199                                 NULL, NULL, nonblocking);
200 }
201
202 /*
203  * convert get_user_pages() return value to posix mlock() error
204  */
205 static int __mlock_posix_error_return(long retval)
206 {
207         if (retval == -EFAULT)
208                 retval = -ENOMEM;
209         else if (retval == -ENOMEM)
210                 retval = -EAGAIN;
211         return retval;
212 }
213
214 /*
215  * munlock_vma_pages_range() - munlock all pages in the vma range.'
216  * @vma - vma containing range to be munlock()ed.
217  * @start - start address in @vma of the range
218  * @end - end of range in @vma.
219  *
220  *  For mremap(), munmap() and exit().
221  *
222  * Called with @vma VM_LOCKED.
223  *
224  * Returns with VM_LOCKED cleared.  Callers must be prepared to
225  * deal with this.
226  *
227  * We don't save and restore VM_LOCKED here because pages are
228  * still on lru.  In unmap path, pages might be scanned by reclaim
229  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
230  * free them.  This will result in freeing mlocked pages.
231  */
232 void munlock_vma_pages_range(struct vm_area_struct *vma,
233                              unsigned long start, unsigned long end)
234 {
235         vma->vm_flags &= ~VM_LOCKED;
236
237         while (start < end) {
238                 struct page *page;
239                 unsigned int page_mask, page_increm;
240
241                 /*
242                  * Although FOLL_DUMP is intended for get_dump_page(),
243                  * it just so happens that its special treatment of the
244                  * ZERO_PAGE (returning an error instead of doing get_page)
245                  * suits munlock very well (and if somehow an abnormal page
246                  * has sneaked into the range, we won't oops here: great).
247                  */
248                 page = follow_page_mask(vma, start, FOLL_GET | FOLL_DUMP,
249                                         &page_mask);
250                 if (page && !IS_ERR(page)) {
251                         lock_page(page);
252                         lru_add_drain();
253                         /*
254                          * Any THP page found by follow_page_mask() may have
255                          * gotten split before reaching munlock_vma_page(),
256                          * so we need to recompute the page_mask here.
257                          */
258                         page_mask = munlock_vma_page(page);
259                         unlock_page(page);
260                         put_page(page);
261                 }
262                 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
263                 start += page_increm * PAGE_SIZE;
264                 cond_resched();
265         }
266 }
267
268 /*
269  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
270  *
271  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
272  * munlock is a no-op.  However, for some special vmas, we go ahead and
273  * populate the ptes.
274  *
275  * For vmas that pass the filters, merge/split as appropriate.
276  */
277 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
278         unsigned long start, unsigned long end, vm_flags_t newflags)
279 {
280         struct mm_struct *mm = vma->vm_mm;
281         pgoff_t pgoff;
282         int nr_pages;
283         int ret = 0;
284         int lock = !!(newflags & VM_LOCKED);
285
286         if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
287             is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
288                 goto out;       /* don't set VM_LOCKED,  don't count */
289
290         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
291         *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
292                           vma->vm_file, pgoff, vma_policy(vma),
293                           vma_get_anon_name(vma));
294         if (*prev) {
295                 vma = *prev;
296                 goto success;
297         }
298
299         if (start != vma->vm_start) {
300                 ret = split_vma(mm, vma, start, 1);
301                 if (ret)
302                         goto out;
303         }
304
305         if (end != vma->vm_end) {
306                 ret = split_vma(mm, vma, end, 0);
307                 if (ret)
308                         goto out;
309         }
310
311 success:
312         /*
313          * Keep track of amount of locked VM.
314          */
315         nr_pages = (end - start) >> PAGE_SHIFT;
316         if (!lock)
317                 nr_pages = -nr_pages;
318         mm->locked_vm += nr_pages;
319
320         /*
321          * vm_flags is protected by the mmap_sem held in write mode.
322          * It's okay if try_to_unmap_one unmaps a page just after we
323          * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
324          */
325
326         if (lock)
327                 vma->vm_flags = newflags;
328         else
329                 munlock_vma_pages_range(vma, start, end);
330
331 out:
332         *prev = vma;
333         return ret;
334 }
335
336 static int do_mlock(unsigned long start, size_t len, int on)
337 {
338         unsigned long nstart, end, tmp;
339         struct vm_area_struct * vma, * prev;
340         int error;
341
342         VM_BUG_ON(start & ~PAGE_MASK);
343         VM_BUG_ON(len != PAGE_ALIGN(len));
344         end = start + len;
345         if (end < start)
346                 return -EINVAL;
347         if (end == start)
348                 return 0;
349         vma = find_vma(current->mm, start);
350         if (!vma || vma->vm_start > start)
351                 return -ENOMEM;
352
353         prev = vma->vm_prev;
354         if (start > vma->vm_start)
355                 prev = vma;
356
357         for (nstart = start ; ; ) {
358                 vm_flags_t newflags;
359
360                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
361
362                 newflags = vma->vm_flags & ~VM_LOCKED;
363                 if (on)
364                         newflags |= VM_LOCKED;
365
366                 tmp = vma->vm_end;
367                 if (tmp > end)
368                         tmp = end;
369                 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
370                 if (error)
371                         break;
372                 nstart = tmp;
373                 if (nstart < prev->vm_end)
374                         nstart = prev->vm_end;
375                 if (nstart >= end)
376                         break;
377
378                 vma = prev->vm_next;
379                 if (!vma || vma->vm_start != nstart) {
380                         error = -ENOMEM;
381                         break;
382                 }
383         }
384         return error;
385 }
386
387 /*
388  * __mm_populate - populate and/or mlock pages within a range of address space.
389  *
390  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
391  * flags. VMAs must be already marked with the desired vm_flags, and
392  * mmap_sem must not be held.
393  */
394 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
395 {
396         struct mm_struct *mm = current->mm;
397         unsigned long end, nstart, nend;
398         struct vm_area_struct *vma = NULL;
399         int locked = 0;
400         long ret = 0;
401
402         VM_BUG_ON(start & ~PAGE_MASK);
403         VM_BUG_ON(len != PAGE_ALIGN(len));
404         end = start + len;
405
406         for (nstart = start; nstart < end; nstart = nend) {
407                 /*
408                  * We want to fault in pages for [nstart; end) address range.
409                  * Find first corresponding VMA.
410                  */
411                 if (!locked) {
412                         locked = 1;
413                         down_read(&mm->mmap_sem);
414                         vma = find_vma(mm, nstart);
415                 } else if (nstart >= vma->vm_end)
416                         vma = vma->vm_next;
417                 if (!vma || vma->vm_start >= end)
418                         break;
419                 /*
420                  * Set [nstart; nend) to intersection of desired address
421                  * range with the first VMA. Also, skip undesirable VMA types.
422                  */
423                 nend = min(end, vma->vm_end);
424                 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
425                         continue;
426                 if (nstart < vma->vm_start)
427                         nstart = vma->vm_start;
428                 /*
429                  * Now fault in a range of pages. __mlock_vma_pages_range()
430                  * double checks the vma flags, so that it won't mlock pages
431                  * if the vma was already munlocked.
432                  */
433                 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
434                 if (ret < 0) {
435                         if (ignore_errors) {
436                                 ret = 0;
437                                 continue;       /* continue at next VMA */
438                         }
439                         ret = __mlock_posix_error_return(ret);
440                         break;
441                 }
442                 nend = nstart + ret * PAGE_SIZE;
443                 ret = 0;
444         }
445         if (locked)
446                 up_read(&mm->mmap_sem);
447         return ret;     /* 0 or negative error code */
448 }
449
450 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
451 {
452         unsigned long locked;
453         unsigned long lock_limit;
454         int error = -ENOMEM;
455
456         if (!can_do_mlock())
457                 return -EPERM;
458
459         lru_add_drain_all();    /* flush pagevec */
460
461         down_write(&current->mm->mmap_sem);
462         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
463         start &= PAGE_MASK;
464
465         locked = len >> PAGE_SHIFT;
466         locked += current->mm->locked_vm;
467
468         lock_limit = rlimit(RLIMIT_MEMLOCK);
469         lock_limit >>= PAGE_SHIFT;
470
471         /* check against resource limits */
472         if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
473                 error = do_mlock(start, len, 1);
474         up_write(&current->mm->mmap_sem);
475         if (!error)
476                 error = __mm_populate(start, len, 0);
477         return error;
478 }
479
480 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
481 {
482         int ret;
483
484         down_write(&current->mm->mmap_sem);
485         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
486         start &= PAGE_MASK;
487         ret = do_mlock(start, len, 0);
488         up_write(&current->mm->mmap_sem);
489         return ret;
490 }
491
492 static int do_mlockall(int flags)
493 {
494         struct vm_area_struct * vma, * prev = NULL;
495
496         if (flags & MCL_FUTURE)
497                 current->mm->def_flags |= VM_LOCKED;
498         else
499                 current->mm->def_flags &= ~VM_LOCKED;
500         if (flags == MCL_FUTURE)
501                 goto out;
502
503         for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
504                 vm_flags_t newflags;
505
506                 newflags = vma->vm_flags & ~VM_LOCKED;
507                 if (flags & MCL_CURRENT)
508                         newflags |= VM_LOCKED;
509
510                 /* Ignore errors */
511                 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
512         }
513 out:
514         return 0;
515 }
516
517 SYSCALL_DEFINE1(mlockall, int, flags)
518 {
519         unsigned long lock_limit;
520         int ret = -EINVAL;
521
522         if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
523                 goto out;
524
525         ret = -EPERM;
526         if (!can_do_mlock())
527                 goto out;
528
529         if (flags & MCL_CURRENT)
530                 lru_add_drain_all();    /* flush pagevec */
531
532         down_write(&current->mm->mmap_sem);
533
534         lock_limit = rlimit(RLIMIT_MEMLOCK);
535         lock_limit >>= PAGE_SHIFT;
536
537         ret = -ENOMEM;
538         if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
539             capable(CAP_IPC_LOCK))
540                 ret = do_mlockall(flags);
541         up_write(&current->mm->mmap_sem);
542         if (!ret && (flags & MCL_CURRENT))
543                 mm_populate(0, TASK_SIZE);
544 out:
545         return ret;
546 }
547
548 SYSCALL_DEFINE0(munlockall)
549 {
550         int ret;
551
552         down_write(&current->mm->mmap_sem);
553         ret = do_mlockall(0);
554         up_write(&current->mm->mmap_sem);
555         return ret;
556 }
557
558 /*
559  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
560  * shm segments) get accounted against the user_struct instead.
561  */
562 static DEFINE_SPINLOCK(shmlock_user_lock);
563
564 int user_shm_lock(size_t size, struct user_struct *user)
565 {
566         unsigned long lock_limit, locked;
567         int allowed = 0;
568
569         locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
570         lock_limit = rlimit(RLIMIT_MEMLOCK);
571         if (lock_limit == RLIM_INFINITY)
572                 allowed = 1;
573         lock_limit >>= PAGE_SHIFT;
574         spin_lock(&shmlock_user_lock);
575         if (!allowed &&
576             locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
577                 goto out;
578         get_uid(user);
579         user->locked_shm += locked;
580         allowed = 1;
581 out:
582         spin_unlock(&shmlock_user_lock);
583         return allowed;
584 }
585
586 void user_shm_unlock(size_t size, struct user_struct *user)
587 {
588         spin_lock(&shmlock_user_lock);
589         user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
590         spin_unlock(&shmlock_user_lock);
591         free_uid(user);
592 }