xref: /openbmc/linux/mm/mlock.c (revision 18693050)
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 	BUG_ON(!PageLocked(page));
80 
81 	if (!TestSetPageMlocked(page)) {
82 		mod_zone_page_state(page_zone(page), NR_MLOCK,
83 				    hpage_nr_pages(page));
84 		count_vm_event(UNEVICTABLE_PGMLOCKED);
85 		if (!isolate_lru_page(page))
86 			putback_lru_page(page);
87 	}
88 }
89 
90 /**
91  * munlock_vma_page - munlock a vma page
92  * @page - page to be unlocked
93  *
94  * called from munlock()/munmap() path with page supposedly on the LRU.
95  * When we munlock a page, because the vma where we found the page is being
96  * munlock()ed or munmap()ed, we want to check whether other vmas hold the
97  * page locked so that we can leave it on the unevictable lru list and not
98  * bother vmscan with it.  However, to walk the page's rmap list in
99  * try_to_munlock() we must isolate the page from the LRU.  If some other
100  * task has removed the page from the LRU, we won't be able to do that.
101  * So we clear the PageMlocked as we might not get another chance.  If we
102  * can't isolate the page, we leave it for putback_lru_page() and vmscan
103  * [page_referenced()/try_to_unmap()] to deal with.
104  */
105 void munlock_vma_page(struct page *page)
106 {
107 	BUG_ON(!PageLocked(page));
108 
109 	if (TestClearPageMlocked(page)) {
110 		mod_zone_page_state(page_zone(page), NR_MLOCK,
111 				    -hpage_nr_pages(page));
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 long __mlock_vma_pages_range(struct vm_area_struct *vma,
159 		unsigned long start, unsigned long end, int *nonblocking)
160 {
161 	struct mm_struct *mm = vma->vm_mm;
162 	unsigned long addr = start;
163 	int nr_pages = (end - start) / PAGE_SIZE;
164 	int gup_flags;
165 
166 	VM_BUG_ON(start & ~PAGE_MASK);
167 	VM_BUG_ON(end   & ~PAGE_MASK);
168 	VM_BUG_ON(start < vma->vm_start);
169 	VM_BUG_ON(end   > vma->vm_end);
170 	VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
171 
172 	gup_flags = FOLL_TOUCH | FOLL_MLOCK;
173 	/*
174 	 * We want to touch writable mappings with a write fault in order
175 	 * to break COW, except for shared mappings because these don't COW
176 	 * and we would not want to dirty them for nothing.
177 	 */
178 	if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
179 		gup_flags |= FOLL_WRITE;
180 
181 	/*
182 	 * We want mlock to succeed for regions that have any permissions
183 	 * other than PROT_NONE.
184 	 */
185 	if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
186 		gup_flags |= FOLL_FORCE;
187 
188 	return __get_user_pages(current, mm, addr, nr_pages, gup_flags,
189 				NULL, NULL, nonblocking);
190 }
191 
192 /*
193  * convert get_user_pages() return value to posix mlock() error
194  */
195 static 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  * munlock_vma_pages_range() - munlock all pages in the vma range.'
206  * @vma - vma containing range to be munlock()ed.
207  * @start - start address in @vma of the range
208  * @end - end of range in @vma.
209  *
210  *  For mremap(), munmap() and exit().
211  *
212  * Called with @vma VM_LOCKED.
213  *
214  * Returns with VM_LOCKED cleared.  Callers must be prepared to
215  * deal with this.
216  *
217  * We don't save and restore VM_LOCKED here because pages are
218  * still on lru.  In unmap path, pages might be scanned by reclaim
219  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
220  * free them.  This will result in freeing mlocked pages.
221  */
222 void munlock_vma_pages_range(struct vm_area_struct *vma,
223 			     unsigned long start, unsigned long end)
224 {
225 	unsigned long addr;
226 
227 	lru_add_drain();
228 	vma->vm_flags &= ~VM_LOCKED;
229 
230 	for (addr = start; addr < end; addr += PAGE_SIZE) {
231 		struct page *page;
232 		/*
233 		 * Although FOLL_DUMP is intended for get_dump_page(),
234 		 * it just so happens that its special treatment of the
235 		 * ZERO_PAGE (returning an error instead of doing get_page)
236 		 * suits munlock very well (and if somehow an abnormal page
237 		 * has sneaked into the range, we won't oops here: great).
238 		 */
239 		page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
240 		if (page && !IS_ERR(page)) {
241 			lock_page(page);
242 			munlock_vma_page(page);
243 			unlock_page(page);
244 			put_page(page);
245 		}
246 		cond_resched();
247 	}
248 }
249 
250 /*
251  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
252  *
253  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
254  * munlock is a no-op.  However, for some special vmas, we go ahead and
255  * populate the ptes.
256  *
257  * For vmas that pass the filters, merge/split as appropriate.
258  */
259 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
260 	unsigned long start, unsigned long end, vm_flags_t newflags)
261 {
262 	struct mm_struct *mm = vma->vm_mm;
263 	pgoff_t pgoff;
264 	int nr_pages;
265 	int ret = 0;
266 	int lock = !!(newflags & VM_LOCKED);
267 
268 	if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
269 	    is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
270 		goto out;	/* don't set VM_LOCKED,  don't count */
271 
272 	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
273 	*prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
274 			  vma->vm_file, pgoff, vma_policy(vma));
275 	if (*prev) {
276 		vma = *prev;
277 		goto success;
278 	}
279 
280 	if (start != vma->vm_start) {
281 		ret = split_vma(mm, vma, start, 1);
282 		if (ret)
283 			goto out;
284 	}
285 
286 	if (end != vma->vm_end) {
287 		ret = split_vma(mm, vma, end, 0);
288 		if (ret)
289 			goto out;
290 	}
291 
292 success:
293 	/*
294 	 * Keep track of amount of locked VM.
295 	 */
296 	nr_pages = (end - start) >> PAGE_SHIFT;
297 	if (!lock)
298 		nr_pages = -nr_pages;
299 	mm->locked_vm += nr_pages;
300 
301 	/*
302 	 * vm_flags is protected by the mmap_sem held in write mode.
303 	 * It's okay if try_to_unmap_one unmaps a page just after we
304 	 * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
305 	 */
306 
307 	if (lock)
308 		vma->vm_flags = newflags;
309 	else
310 		munlock_vma_pages_range(vma, start, end);
311 
312 out:
313 	*prev = vma;
314 	return ret;
315 }
316 
317 static int do_mlock(unsigned long start, size_t len, int on)
318 {
319 	unsigned long nstart, end, tmp;
320 	struct vm_area_struct * vma, * prev;
321 	int error;
322 
323 	VM_BUG_ON(start & ~PAGE_MASK);
324 	VM_BUG_ON(len != PAGE_ALIGN(len));
325 	end = start + len;
326 	if (end < start)
327 		return -EINVAL;
328 	if (end == start)
329 		return 0;
330 	vma = find_vma(current->mm, start);
331 	if (!vma || vma->vm_start > start)
332 		return -ENOMEM;
333 
334 	prev = vma->vm_prev;
335 	if (start > vma->vm_start)
336 		prev = vma;
337 
338 	for (nstart = start ; ; ) {
339 		vm_flags_t newflags;
340 
341 		/* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
342 
343 		newflags = vma->vm_flags & ~VM_LOCKED;
344 		if (on)
345 			newflags |= VM_LOCKED | VM_POPULATE;
346 
347 		tmp = vma->vm_end;
348 		if (tmp > end)
349 			tmp = end;
350 		error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
351 		if (error)
352 			break;
353 		nstart = tmp;
354 		if (nstart < prev->vm_end)
355 			nstart = prev->vm_end;
356 		if (nstart >= end)
357 			break;
358 
359 		vma = prev->vm_next;
360 		if (!vma || vma->vm_start != nstart) {
361 			error = -ENOMEM;
362 			break;
363 		}
364 	}
365 	return error;
366 }
367 
368 /*
369  * __mm_populate - populate and/or mlock pages within a range of address space.
370  *
371  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
372  * flags. VMAs must be already marked with the desired vm_flags, and
373  * mmap_sem must not be held.
374  */
375 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
376 {
377 	struct mm_struct *mm = current->mm;
378 	unsigned long end, nstart, nend;
379 	struct vm_area_struct *vma = NULL;
380 	int locked = 0;
381 	int ret = 0;
382 
383 	VM_BUG_ON(start & ~PAGE_MASK);
384 	VM_BUG_ON(len != PAGE_ALIGN(len));
385 	end = start + len;
386 
387 	for (nstart = start; nstart < end; nstart = nend) {
388 		/*
389 		 * We want to fault in pages for [nstart; end) address range.
390 		 * Find first corresponding VMA.
391 		 */
392 		if (!locked) {
393 			locked = 1;
394 			down_read(&mm->mmap_sem);
395 			vma = find_vma(mm, nstart);
396 		} else if (nstart >= vma->vm_end)
397 			vma = vma->vm_next;
398 		if (!vma || vma->vm_start >= end)
399 			break;
400 		/*
401 		 * Set [nstart; nend) to intersection of desired address
402 		 * range with the first VMA. Also, skip undesirable VMA types.
403 		 */
404 		nend = min(end, vma->vm_end);
405 		if ((vma->vm_flags & (VM_IO | VM_PFNMAP | VM_POPULATE)) !=
406 		    VM_POPULATE)
407 			continue;
408 		if (nstart < vma->vm_start)
409 			nstart = vma->vm_start;
410 		/*
411 		 * Now fault in a range of pages. __mlock_vma_pages_range()
412 		 * double checks the vma flags, so that it won't mlock pages
413 		 * if the vma was already munlocked.
414 		 */
415 		ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
416 		if (ret < 0) {
417 			if (ignore_errors) {
418 				ret = 0;
419 				continue;	/* continue at next VMA */
420 			}
421 			ret = __mlock_posix_error_return(ret);
422 			break;
423 		}
424 		nend = nstart + ret * PAGE_SIZE;
425 		ret = 0;
426 	}
427 	if (locked)
428 		up_read(&mm->mmap_sem);
429 	return ret;	/* 0 or negative error code */
430 }
431 
432 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
433 {
434 	unsigned long locked;
435 	unsigned long lock_limit;
436 	int error = -ENOMEM;
437 
438 	if (!can_do_mlock())
439 		return -EPERM;
440 
441 	lru_add_drain_all();	/* flush pagevec */
442 
443 	down_write(&current->mm->mmap_sem);
444 	len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
445 	start &= PAGE_MASK;
446 
447 	locked = len >> PAGE_SHIFT;
448 	locked += current->mm->locked_vm;
449 
450 	lock_limit = rlimit(RLIMIT_MEMLOCK);
451 	lock_limit >>= PAGE_SHIFT;
452 
453 	/* check against resource limits */
454 	if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
455 		error = do_mlock(start, len, 1);
456 	up_write(&current->mm->mmap_sem);
457 	if (!error)
458 		error = __mm_populate(start, len, 0);
459 	return error;
460 }
461 
462 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
463 {
464 	int ret;
465 
466 	down_write(&current->mm->mmap_sem);
467 	len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
468 	start &= PAGE_MASK;
469 	ret = do_mlock(start, len, 0);
470 	up_write(&current->mm->mmap_sem);
471 	return ret;
472 }
473 
474 static int do_mlockall(int flags)
475 {
476 	struct vm_area_struct * vma, * prev = NULL;
477 
478 	if (flags & MCL_FUTURE)
479 		current->mm->def_flags |= VM_LOCKED | VM_POPULATE;
480 	else
481 		current->mm->def_flags &= ~(VM_LOCKED | VM_POPULATE);
482 	if (flags == MCL_FUTURE)
483 		goto out;
484 
485 	for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
486 		vm_flags_t newflags;
487 
488 		newflags = vma->vm_flags & ~VM_LOCKED;
489 		if (flags & MCL_CURRENT)
490 			newflags |= VM_LOCKED | VM_POPULATE;
491 
492 		/* Ignore errors */
493 		mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
494 	}
495 out:
496 	return 0;
497 }
498 
499 SYSCALL_DEFINE1(mlockall, int, flags)
500 {
501 	unsigned long lock_limit;
502 	int ret = -EINVAL;
503 
504 	if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
505 		goto out;
506 
507 	ret = -EPERM;
508 	if (!can_do_mlock())
509 		goto out;
510 
511 	if (flags & MCL_CURRENT)
512 		lru_add_drain_all();	/* flush pagevec */
513 
514 	down_write(&current->mm->mmap_sem);
515 
516 	lock_limit = rlimit(RLIMIT_MEMLOCK);
517 	lock_limit >>= PAGE_SHIFT;
518 
519 	ret = -ENOMEM;
520 	if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
521 	    capable(CAP_IPC_LOCK))
522 		ret = do_mlockall(flags);
523 	up_write(&current->mm->mmap_sem);
524 	if (!ret && (flags & MCL_CURRENT))
525 		mm_populate(0, TASK_SIZE);
526 out:
527 	return ret;
528 }
529 
530 SYSCALL_DEFINE0(munlockall)
531 {
532 	int ret;
533 
534 	down_write(&current->mm->mmap_sem);
535 	ret = do_mlockall(0);
536 	up_write(&current->mm->mmap_sem);
537 	return ret;
538 }
539 
540 /*
541  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
542  * shm segments) get accounted against the user_struct instead.
543  */
544 static DEFINE_SPINLOCK(shmlock_user_lock);
545 
546 int user_shm_lock(size_t size, struct user_struct *user)
547 {
548 	unsigned long lock_limit, locked;
549 	int allowed = 0;
550 
551 	locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
552 	lock_limit = rlimit(RLIMIT_MEMLOCK);
553 	if (lock_limit == RLIM_INFINITY)
554 		allowed = 1;
555 	lock_limit >>= PAGE_SHIFT;
556 	spin_lock(&shmlock_user_lock);
557 	if (!allowed &&
558 	    locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
559 		goto out;
560 	get_uid(user);
561 	user->locked_shm += locked;
562 	allowed = 1;
563 out:
564 	spin_unlock(&shmlock_user_lock);
565 	return allowed;
566 }
567 
568 void user_shm_unlock(size_t size, struct user_struct *user)
569 {
570 	spin_lock(&shmlock_user_lock);
571 	user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
572 	spin_unlock(&shmlock_user_lock);
573 	free_uid(user);
574 }
575