xref: /openbmc/linux/mm/mlock.c (revision 95e9fd10)
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 	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 = SWAP_AGAIN;
114 
115 			/*
116 			 * Optimization: if the page was mapped just once,
117 			 * that's our mapping and we don't need to check all the
118 			 * other vmas.
119 			 */
120 			if (page_mapcount(page) > 1)
121 				ret = try_to_munlock(page);
122 			/*
123 			 * did try_to_unlock() succeed or punt?
124 			 */
125 			if (ret != SWAP_MLOCK)
126 				count_vm_event(UNEVICTABLE_PGMUNLOCKED);
127 
128 			putback_lru_page(page);
129 		} else {
130 			/*
131 			 * Some other task has removed the page from the LRU.
132 			 * putback_lru_page() will take care of removing the
133 			 * page from the unevictable list, if necessary.
134 			 * vmscan [page_referenced()] will move the page back
135 			 * to the unevictable list if some other vma has it
136 			 * mlocked.
137 			 */
138 			if (PageUnevictable(page))
139 				count_vm_event(UNEVICTABLE_PGSTRANDED);
140 			else
141 				count_vm_event(UNEVICTABLE_PGMUNLOCKED);
142 		}
143 	}
144 }
145 
146 /**
147  * __mlock_vma_pages_range() -  mlock a range of pages in the vma.
148  * @vma:   target vma
149  * @start: start address
150  * @end:   end address
151  *
152  * This takes care of making the pages present too.
153  *
154  * return 0 on success, negative error code on error.
155  *
156  * vma->vm_mm->mmap_sem must be held for at least read.
157  */
158 static long __mlock_vma_pages_range(struct vm_area_struct *vma,
159 				    unsigned long start, unsigned long end,
160 				    int *nonblocking)
161 {
162 	struct mm_struct *mm = vma->vm_mm;
163 	unsigned long addr = start;
164 	int nr_pages = (end - start) / PAGE_SIZE;
165 	int gup_flags;
166 
167 	VM_BUG_ON(start & ~PAGE_MASK);
168 	VM_BUG_ON(end   & ~PAGE_MASK);
169 	VM_BUG_ON(start < vma->vm_start);
170 	VM_BUG_ON(end   > vma->vm_end);
171 	VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
172 
173 	gup_flags = FOLL_TOUCH | FOLL_MLOCK;
174 	/*
175 	 * We want to touch writable mappings with a write fault in order
176 	 * to break COW, except for shared mappings because these don't COW
177 	 * and we would not want to dirty them for nothing.
178 	 */
179 	if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
180 		gup_flags |= FOLL_WRITE;
181 
182 	/*
183 	 * We want mlock to succeed for regions that have any permissions
184 	 * other than PROT_NONE.
185 	 */
186 	if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
187 		gup_flags |= FOLL_FORCE;
188 
189 	return __get_user_pages(current, mm, addr, nr_pages, gup_flags,
190 				NULL, NULL, nonblocking);
191 }
192 
193 /*
194  * convert get_user_pages() return value to posix mlock() error
195  */
196 static int __mlock_posix_error_return(long retval)
197 {
198 	if (retval == -EFAULT)
199 		retval = -ENOMEM;
200 	else if (retval == -ENOMEM)
201 		retval = -EAGAIN;
202 	return retval;
203 }
204 
205 /**
206  * mlock_vma_pages_range() - mlock pages in specified vma range.
207  * @vma - the vma containing the specfied address range
208  * @start - starting address in @vma to mlock
209  * @end   - end address [+1] in @vma to mlock
210  *
211  * For mmap()/mremap()/expansion of mlocked vma.
212  *
213  * return 0 on success for "normal" vmas.
214  *
215  * return number of pages [> 0] to be removed from locked_vm on success
216  * of "special" vmas.
217  */
218 long mlock_vma_pages_range(struct vm_area_struct *vma,
219 			unsigned long start, unsigned long end)
220 {
221 	int nr_pages = (end - start) / PAGE_SIZE;
222 	BUG_ON(!(vma->vm_flags & VM_LOCKED));
223 
224 	/*
225 	 * filter unlockable vmas
226 	 */
227 	if (vma->vm_flags & (VM_IO | VM_PFNMAP))
228 		goto no_mlock;
229 
230 	if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
231 			is_vm_hugetlb_page(vma) ||
232 			vma == get_gate_vma(current->mm))) {
233 
234 		__mlock_vma_pages_range(vma, start, end, NULL);
235 
236 		/* Hide errors from mmap() and other callers */
237 		return 0;
238 	}
239 
240 	/*
241 	 * User mapped kernel pages or huge pages:
242 	 * make these pages present to populate the ptes, but
243 	 * fall thru' to reset VM_LOCKED--no need to unlock, and
244 	 * return nr_pages so these don't get counted against task's
245 	 * locked limit.  huge pages are already counted against
246 	 * locked vm limit.
247 	 */
248 	make_pages_present(start, end);
249 
250 no_mlock:
251 	vma->vm_flags &= ~VM_LOCKED;	/* and don't come back! */
252 	return nr_pages;		/* error or pages NOT mlocked */
253 }
254 
255 /*
256  * munlock_vma_pages_range() - munlock all pages in the vma range.'
257  * @vma - vma containing range to be munlock()ed.
258  * @start - start address in @vma of the range
259  * @end - end of range in @vma.
260  *
261  *  For mremap(), munmap() and exit().
262  *
263  * Called with @vma VM_LOCKED.
264  *
265  * Returns with VM_LOCKED cleared.  Callers must be prepared to
266  * deal with this.
267  *
268  * We don't save and restore VM_LOCKED here because pages are
269  * still on lru.  In unmap path, pages might be scanned by reclaim
270  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
271  * free them.  This will result in freeing mlocked pages.
272  */
273 void munlock_vma_pages_range(struct vm_area_struct *vma,
274 			     unsigned long start, unsigned long end)
275 {
276 	unsigned long addr;
277 
278 	lru_add_drain();
279 	vma->vm_flags &= ~VM_LOCKED;
280 
281 	for (addr = start; addr < end; addr += PAGE_SIZE) {
282 		struct page *page;
283 		/*
284 		 * Although FOLL_DUMP is intended for get_dump_page(),
285 		 * it just so happens that its special treatment of the
286 		 * ZERO_PAGE (returning an error instead of doing get_page)
287 		 * suits munlock very well (and if somehow an abnormal page
288 		 * has sneaked into the range, we won't oops here: great).
289 		 */
290 		page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
291 		if (page && !IS_ERR(page)) {
292 			lock_page(page);
293 			/*
294 			 * Like in __mlock_vma_pages_range(),
295 			 * because we lock page here and migration is
296 			 * blocked by the elevated reference, we need
297 			 * only check for file-cache page truncation.
298 			 */
299 			if (page->mapping)
300 				munlock_vma_page(page);
301 			unlock_page(page);
302 			put_page(page);
303 		}
304 		cond_resched();
305 	}
306 }
307 
308 /*
309  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
310  *
311  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
312  * munlock is a no-op.  However, for some special vmas, we go ahead and
313  * populate the ptes via make_pages_present().
314  *
315  * For vmas that pass the filters, merge/split as appropriate.
316  */
317 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
318 	unsigned long start, unsigned long end, vm_flags_t newflags)
319 {
320 	struct mm_struct *mm = vma->vm_mm;
321 	pgoff_t pgoff;
322 	int nr_pages;
323 	int ret = 0;
324 	int lock = !!(newflags & VM_LOCKED);
325 
326 	if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
327 	    is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
328 		goto out;	/* don't set VM_LOCKED,  don't count */
329 
330 	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
331 	*prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
332 			  vma->vm_file, pgoff, vma_policy(vma));
333 	if (*prev) {
334 		vma = *prev;
335 		goto success;
336 	}
337 
338 	if (start != vma->vm_start) {
339 		ret = split_vma(mm, vma, start, 1);
340 		if (ret)
341 			goto out;
342 	}
343 
344 	if (end != vma->vm_end) {
345 		ret = split_vma(mm, vma, end, 0);
346 		if (ret)
347 			goto out;
348 	}
349 
350 success:
351 	/*
352 	 * Keep track of amount of locked VM.
353 	 */
354 	nr_pages = (end - start) >> PAGE_SHIFT;
355 	if (!lock)
356 		nr_pages = -nr_pages;
357 	mm->locked_vm += nr_pages;
358 
359 	/*
360 	 * vm_flags is protected by the mmap_sem held in write mode.
361 	 * It's okay if try_to_unmap_one unmaps a page just after we
362 	 * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
363 	 */
364 
365 	if (lock)
366 		vma->vm_flags = newflags;
367 	else
368 		munlock_vma_pages_range(vma, start, end);
369 
370 out:
371 	*prev = vma;
372 	return ret;
373 }
374 
375 static int do_mlock(unsigned long start, size_t len, int on)
376 {
377 	unsigned long nstart, end, tmp;
378 	struct vm_area_struct * vma, * prev;
379 	int error;
380 
381 	VM_BUG_ON(start & ~PAGE_MASK);
382 	VM_BUG_ON(len != PAGE_ALIGN(len));
383 	end = start + len;
384 	if (end < start)
385 		return -EINVAL;
386 	if (end == start)
387 		return 0;
388 	vma = find_vma(current->mm, start);
389 	if (!vma || vma->vm_start > start)
390 		return -ENOMEM;
391 
392 	prev = vma->vm_prev;
393 	if (start > vma->vm_start)
394 		prev = vma;
395 
396 	for (nstart = start ; ; ) {
397 		vm_flags_t newflags;
398 
399 		/* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
400 
401 		newflags = vma->vm_flags | VM_LOCKED;
402 		if (!on)
403 			newflags &= ~VM_LOCKED;
404 
405 		tmp = vma->vm_end;
406 		if (tmp > end)
407 			tmp = end;
408 		error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
409 		if (error)
410 			break;
411 		nstart = tmp;
412 		if (nstart < prev->vm_end)
413 			nstart = prev->vm_end;
414 		if (nstart >= end)
415 			break;
416 
417 		vma = prev->vm_next;
418 		if (!vma || vma->vm_start != nstart) {
419 			error = -ENOMEM;
420 			break;
421 		}
422 	}
423 	return error;
424 }
425 
426 static int do_mlock_pages(unsigned long start, size_t len, int ignore_errors)
427 {
428 	struct mm_struct *mm = current->mm;
429 	unsigned long end, nstart, nend;
430 	struct vm_area_struct *vma = NULL;
431 	int locked = 0;
432 	int ret = 0;
433 
434 	VM_BUG_ON(start & ~PAGE_MASK);
435 	VM_BUG_ON(len != PAGE_ALIGN(len));
436 	end = start + len;
437 
438 	for (nstart = start; nstart < end; nstart = nend) {
439 		/*
440 		 * We want to fault in pages for [nstart; end) address range.
441 		 * Find first corresponding VMA.
442 		 */
443 		if (!locked) {
444 			locked = 1;
445 			down_read(&mm->mmap_sem);
446 			vma = find_vma(mm, nstart);
447 		} else if (nstart >= vma->vm_end)
448 			vma = vma->vm_next;
449 		if (!vma || vma->vm_start >= end)
450 			break;
451 		/*
452 		 * Set [nstart; nend) to intersection of desired address
453 		 * range with the first VMA. Also, skip undesirable VMA types.
454 		 */
455 		nend = min(end, vma->vm_end);
456 		if (vma->vm_flags & (VM_IO | VM_PFNMAP))
457 			continue;
458 		if (nstart < vma->vm_start)
459 			nstart = vma->vm_start;
460 		/*
461 		 * Now fault in a range of pages. __mlock_vma_pages_range()
462 		 * double checks the vma flags, so that it won't mlock pages
463 		 * if the vma was already munlocked.
464 		 */
465 		ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
466 		if (ret < 0) {
467 			if (ignore_errors) {
468 				ret = 0;
469 				continue;	/* continue at next VMA */
470 			}
471 			ret = __mlock_posix_error_return(ret);
472 			break;
473 		}
474 		nend = nstart + ret * PAGE_SIZE;
475 		ret = 0;
476 	}
477 	if (locked)
478 		up_read(&mm->mmap_sem);
479 	return ret;	/* 0 or negative error code */
480 }
481 
482 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
483 {
484 	unsigned long locked;
485 	unsigned long lock_limit;
486 	int error = -ENOMEM;
487 
488 	if (!can_do_mlock())
489 		return -EPERM;
490 
491 	lru_add_drain_all();	/* flush pagevec */
492 
493 	down_write(&current->mm->mmap_sem);
494 	len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
495 	start &= PAGE_MASK;
496 
497 	locked = len >> PAGE_SHIFT;
498 	locked += current->mm->locked_vm;
499 
500 	lock_limit = rlimit(RLIMIT_MEMLOCK);
501 	lock_limit >>= PAGE_SHIFT;
502 
503 	/* check against resource limits */
504 	if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
505 		error = do_mlock(start, len, 1);
506 	up_write(&current->mm->mmap_sem);
507 	if (!error)
508 		error = do_mlock_pages(start, len, 0);
509 	return error;
510 }
511 
512 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
513 {
514 	int ret;
515 
516 	down_write(&current->mm->mmap_sem);
517 	len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
518 	start &= PAGE_MASK;
519 	ret = do_mlock(start, len, 0);
520 	up_write(&current->mm->mmap_sem);
521 	return ret;
522 }
523 
524 static int do_mlockall(int flags)
525 {
526 	struct vm_area_struct * vma, * prev = NULL;
527 	unsigned int def_flags = 0;
528 
529 	if (flags & MCL_FUTURE)
530 		def_flags = VM_LOCKED;
531 	current->mm->def_flags = def_flags;
532 	if (flags == MCL_FUTURE)
533 		goto out;
534 
535 	for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
536 		vm_flags_t newflags;
537 
538 		newflags = vma->vm_flags | VM_LOCKED;
539 		if (!(flags & MCL_CURRENT))
540 			newflags &= ~VM_LOCKED;
541 
542 		/* Ignore errors */
543 		mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
544 	}
545 out:
546 	return 0;
547 }
548 
549 SYSCALL_DEFINE1(mlockall, int, flags)
550 {
551 	unsigned long lock_limit;
552 	int ret = -EINVAL;
553 
554 	if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
555 		goto out;
556 
557 	ret = -EPERM;
558 	if (!can_do_mlock())
559 		goto out;
560 
561 	if (flags & MCL_CURRENT)
562 		lru_add_drain_all();	/* flush pagevec */
563 
564 	down_write(&current->mm->mmap_sem);
565 
566 	lock_limit = rlimit(RLIMIT_MEMLOCK);
567 	lock_limit >>= PAGE_SHIFT;
568 
569 	ret = -ENOMEM;
570 	if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
571 	    capable(CAP_IPC_LOCK))
572 		ret = do_mlockall(flags);
573 	up_write(&current->mm->mmap_sem);
574 	if (!ret && (flags & MCL_CURRENT)) {
575 		/* Ignore errors */
576 		do_mlock_pages(0, TASK_SIZE, 1);
577 	}
578 out:
579 	return ret;
580 }
581 
582 SYSCALL_DEFINE0(munlockall)
583 {
584 	int ret;
585 
586 	down_write(&current->mm->mmap_sem);
587 	ret = do_mlockall(0);
588 	up_write(&current->mm->mmap_sem);
589 	return ret;
590 }
591 
592 /*
593  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
594  * shm segments) get accounted against the user_struct instead.
595  */
596 static DEFINE_SPINLOCK(shmlock_user_lock);
597 
598 int user_shm_lock(size_t size, struct user_struct *user)
599 {
600 	unsigned long lock_limit, locked;
601 	int allowed = 0;
602 
603 	locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
604 	lock_limit = rlimit(RLIMIT_MEMLOCK);
605 	if (lock_limit == RLIM_INFINITY)
606 		allowed = 1;
607 	lock_limit >>= PAGE_SHIFT;
608 	spin_lock(&shmlock_user_lock);
609 	if (!allowed &&
610 	    locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
611 		goto out;
612 	get_uid(user);
613 	user->locked_shm += locked;
614 	allowed = 1;
615 out:
616 	spin_unlock(&shmlock_user_lock);
617 	return allowed;
618 }
619 
620 void user_shm_unlock(size_t size, struct user_struct *user)
621 {
622 	spin_lock(&shmlock_user_lock);
623 	user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
624 	spin_unlock(&shmlock_user_lock);
625 	free_uid(user);
626 }
627