Linux Audio

Check our new training course

Loading...
v5.9
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 *	linux/mm/mlock.c
  4 *
  5 *  (C) Copyright 1995 Linus Torvalds
  6 *  (C) Copyright 2002 Christoph Hellwig
  7 */
  8
  9#include <linux/capability.h>
 10#include <linux/mman.h>
 11#include <linux/mm.h>
 12#include <linux/sched/user.h>
 13#include <linux/swap.h>
 14#include <linux/swapops.h>
 15#include <linux/pagemap.h>
 16#include <linux/pagevec.h>
 17#include <linux/mempolicy.h>
 18#include <linux/syscalls.h>
 19#include <linux/sched.h>
 20#include <linux/export.h>
 21#include <linux/rmap.h>
 22#include <linux/mmzone.h>
 23#include <linux/hugetlb.h>
 24#include <linux/memcontrol.h>
 25#include <linux/mm_inline.h>
 26
 27#include "internal.h"
 28
 29bool can_do_mlock(void)
 30{
 31	if (rlimit(RLIMIT_MEMLOCK) != 0)
 32		return true;
 33	if (capable(CAP_IPC_LOCK))
 34		return true;
 35	return false;
 36}
 37EXPORT_SYMBOL(can_do_mlock);
 38
 39/*
 40 * Mlocked pages are marked with PageMlocked() flag for efficient testing
 41 * in vmscan and, possibly, the fault path; and to support semi-accurate
 42 * statistics.
 43 *
 44 * An mlocked page [PageMlocked(page)] is unevictable.  As such, it will
 45 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
 46 * The unevictable list is an LRU sibling list to the [in]active lists.
 47 * PageUnevictable is set to indicate the unevictable state.
 48 *
 49 * When lazy mlocking via vmscan, it is important to ensure that the
 50 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
 51 * may have mlocked a page that is being munlocked. So lazy mlock must take
 52 * the mmap_lock for read, and verify that the vma really is locked
 53 * (see mm/rmap.c).
 54 */
 55
 56/*
 57 *  LRU accounting for clear_page_mlock()
 58 */
 59void clear_page_mlock(struct page *page)
 60{
 61	int nr_pages;
 62
 63	if (!TestClearPageMlocked(page))
 64		return;
 65
 66	nr_pages = thp_nr_pages(page);
 67	mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
 68	count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
 69	/*
 70	 * The previous TestClearPageMlocked() corresponds to the smp_mb()
 71	 * in __pagevec_lru_add_fn().
 72	 *
 73	 * See __pagevec_lru_add_fn for more explanation.
 74	 */
 75	if (!isolate_lru_page(page)) {
 76		putback_lru_page(page);
 77	} else {
 78		/*
 79		 * We lost the race. the page already moved to evictable list.
 80		 */
 81		if (PageUnevictable(page))
 82			count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
 83	}
 84}
 85
 86/*
 87 * Mark page as mlocked if not already.
 88 * If page on LRU, isolate and putback to move to unevictable list.
 89 */
 90void mlock_vma_page(struct page *page)
 91{
 92	/* Serialize with page migration */
 93	BUG_ON(!PageLocked(page));
 94
 95	VM_BUG_ON_PAGE(PageTail(page), page);
 96	VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
 97
 98	if (!TestSetPageMlocked(page)) {
 99		int nr_pages = thp_nr_pages(page);
100
101		mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
102		count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
103		if (!isolate_lru_page(page))
104			putback_lru_page(page);
105	}
106}
107
108/*
109 * Isolate a page from LRU with optional get_page() pin.
110 * Assumes lru_lock already held and page already pinned.
111 */
112static bool __munlock_isolate_lru_page(struct page *page, bool getpage)
113{
114	if (PageLRU(page)) {
115		struct lruvec *lruvec;
116
117		lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
118		if (getpage)
119			get_page(page);
120		ClearPageLRU(page);
121		del_page_from_lru_list(page, lruvec, page_lru(page));
122		return true;
123	}
124
125	return false;
126}
127
128/*
129 * Finish munlock after successful page isolation
130 *
131 * Page must be locked. This is a wrapper for try_to_munlock()
132 * and putback_lru_page() with munlock accounting.
133 */
134static void __munlock_isolated_page(struct page *page)
135{
 
 
136	/*
137	 * Optimization: if the page was mapped just once, that's our mapping
138	 * and we don't need to check all the other vmas.
139	 */
140	if (page_mapcount(page) > 1)
141		try_to_munlock(page);
142
143	/* Did try_to_unlock() succeed or punt? */
144	if (!PageMlocked(page))
145		count_vm_events(UNEVICTABLE_PGMUNLOCKED, thp_nr_pages(page));
146
147	putback_lru_page(page);
148}
149
150/*
151 * Accounting for page isolation fail during munlock
152 *
153 * Performs accounting when page isolation fails in munlock. There is nothing
154 * else to do because it means some other task has already removed the page
155 * from the LRU. putback_lru_page() will take care of removing the page from
156 * the unevictable list, if necessary. vmscan [page_referenced()] will move
157 * the page back to the unevictable list if some other vma has it mlocked.
158 */
159static void __munlock_isolation_failed(struct page *page)
160{
161	int nr_pages = thp_nr_pages(page);
162
163	if (PageUnevictable(page))
164		__count_vm_events(UNEVICTABLE_PGSTRANDED, nr_pages);
165	else
166		__count_vm_events(UNEVICTABLE_PGMUNLOCKED, nr_pages);
167}
168
169/**
170 * munlock_vma_page - munlock a vma page
171 * @page: page to be unlocked, either a normal page or THP page head
172 *
173 * returns the size of the page as a page mask (0 for normal page,
174 *         HPAGE_PMD_NR - 1 for THP head page)
175 *
176 * called from munlock()/munmap() path with page supposedly on the LRU.
177 * When we munlock a page, because the vma where we found the page is being
178 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
179 * page locked so that we can leave it on the unevictable lru list and not
180 * bother vmscan with it.  However, to walk the page's rmap list in
181 * try_to_munlock() we must isolate the page from the LRU.  If some other
182 * task has removed the page from the LRU, we won't be able to do that.
183 * So we clear the PageMlocked as we might not get another chance.  If we
184 * can't isolate the page, we leave it for putback_lru_page() and vmscan
185 * [page_referenced()/try_to_unmap()] to deal with.
186 */
187unsigned int munlock_vma_page(struct page *page)
188{
189	int nr_pages;
190	pg_data_t *pgdat = page_pgdat(page);
191
192	/* For try_to_munlock() and to serialize with page migration */
193	BUG_ON(!PageLocked(page));
194
195	VM_BUG_ON_PAGE(PageTail(page), page);
196
197	/*
198	 * Serialize with any parallel __split_huge_page_refcount() which
199	 * might otherwise copy PageMlocked to part of the tail pages before
200	 * we clear it in the head page. It also stabilizes thp_nr_pages().
201	 */
202	spin_lock_irq(&pgdat->lru_lock);
203
204	if (!TestClearPageMlocked(page)) {
205		/* Potentially, PTE-mapped THP: do not skip the rest PTEs */
206		nr_pages = 1;
207		goto unlock_out;
208	}
209
210	nr_pages = thp_nr_pages(page);
211	__mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
212
213	if (__munlock_isolate_lru_page(page, true)) {
214		spin_unlock_irq(&pgdat->lru_lock);
215		__munlock_isolated_page(page);
216		goto out;
217	}
218	__munlock_isolation_failed(page);
219
220unlock_out:
221	spin_unlock_irq(&pgdat->lru_lock);
222
223out:
224	return nr_pages - 1;
225}
226
227/*
228 * convert get_user_pages() return value to posix mlock() error
229 */
230static int __mlock_posix_error_return(long retval)
231{
232	if (retval == -EFAULT)
233		retval = -ENOMEM;
234	else if (retval == -ENOMEM)
235		retval = -EAGAIN;
236	return retval;
237}
238
239/*
240 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
241 *
242 * The fast path is available only for evictable pages with single mapping.
243 * Then we can bypass the per-cpu pvec and get better performance.
244 * when mapcount > 1 we need try_to_munlock() which can fail.
245 * when !page_evictable(), we need the full redo logic of putback_lru_page to
246 * avoid leaving evictable page in unevictable list.
247 *
248 * In case of success, @page is added to @pvec and @pgrescued is incremented
249 * in case that the page was previously unevictable. @page is also unlocked.
250 */
251static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
252		int *pgrescued)
253{
254	VM_BUG_ON_PAGE(PageLRU(page), page);
255	VM_BUG_ON_PAGE(!PageLocked(page), page);
256
257	if (page_mapcount(page) <= 1 && page_evictable(page)) {
258		pagevec_add(pvec, page);
259		if (TestClearPageUnevictable(page))
260			(*pgrescued)++;
261		unlock_page(page);
262		return true;
263	}
264
265	return false;
266}
267
268/*
269 * Putback multiple evictable pages to the LRU
270 *
271 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
272 * the pages might have meanwhile become unevictable but that is OK.
273 */
274static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
275{
276	count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
277	/*
278	 *__pagevec_lru_add() calls release_pages() so we don't call
279	 * put_page() explicitly
280	 */
281	__pagevec_lru_add(pvec);
282	count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
283}
284
285/*
286 * Munlock a batch of pages from the same zone
287 *
288 * The work is split to two main phases. First phase clears the Mlocked flag
289 * and attempts to isolate the pages, all under a single zone lru lock.
290 * The second phase finishes the munlock only for pages where isolation
291 * succeeded.
292 *
293 * Note that the pagevec may be modified during the process.
294 */
295static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
296{
297	int i;
298	int nr = pagevec_count(pvec);
299	int delta_munlocked = -nr;
300	struct pagevec pvec_putback;
301	int pgrescued = 0;
302
303	pagevec_init(&pvec_putback);
304
305	/* Phase 1: page isolation */
306	spin_lock_irq(&zone->zone_pgdat->lru_lock);
307	for (i = 0; i < nr; i++) {
308		struct page *page = pvec->pages[i];
309
310		if (TestClearPageMlocked(page)) {
311			/*
312			 * We already have pin from follow_page_mask()
313			 * so we can spare the get_page() here.
314			 */
315			if (__munlock_isolate_lru_page(page, false))
316				continue;
317			else
318				__munlock_isolation_failed(page);
319		} else {
320			delta_munlocked++;
321		}
322
323		/*
324		 * We won't be munlocking this page in the next phase
325		 * but we still need to release the follow_page_mask()
326		 * pin. We cannot do it under lru_lock however. If it's
327		 * the last pin, __page_cache_release() would deadlock.
328		 */
329		pagevec_add(&pvec_putback, pvec->pages[i]);
330		pvec->pages[i] = NULL;
331	}
 
332	__mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
333	spin_unlock_irq(&zone->zone_pgdat->lru_lock);
334
335	/* Now we can release pins of pages that we are not munlocking */
336	pagevec_release(&pvec_putback);
337
338	/* Phase 2: page munlock */
339	for (i = 0; i < nr; i++) {
340		struct page *page = pvec->pages[i];
341
342		if (page) {
343			lock_page(page);
344			if (!__putback_lru_fast_prepare(page, &pvec_putback,
345					&pgrescued)) {
346				/*
347				 * Slow path. We don't want to lose the last
348				 * pin before unlock_page()
349				 */
350				get_page(page); /* for putback_lru_page() */
351				__munlock_isolated_page(page);
352				unlock_page(page);
353				put_page(page); /* from follow_page_mask() */
354			}
355		}
356	}
357
358	/*
359	 * Phase 3: page putback for pages that qualified for the fast path
360	 * This will also call put_page() to return pin from follow_page_mask()
361	 */
362	if (pagevec_count(&pvec_putback))
363		__putback_lru_fast(&pvec_putback, pgrescued);
364}
365
366/*
367 * Fill up pagevec for __munlock_pagevec using pte walk
368 *
369 * The function expects that the struct page corresponding to @start address is
370 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
371 *
372 * The rest of @pvec is filled by subsequent pages within the same pmd and same
373 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
374 * pages also get pinned.
375 *
376 * Returns the address of the next page that should be scanned. This equals
377 * @start + PAGE_SIZE when no page could be added by the pte walk.
378 */
379static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
380			struct vm_area_struct *vma, struct zone *zone,
381			unsigned long start, unsigned long end)
382{
383	pte_t *pte;
384	spinlock_t *ptl;
385
386	/*
387	 * Initialize pte walk starting at the already pinned page where we
388	 * are sure that there is a pte, as it was pinned under the same
389	 * mmap_lock write op.
390	 */
391	pte = get_locked_pte(vma->vm_mm, start,	&ptl);
392	/* Make sure we do not cross the page table boundary */
393	end = pgd_addr_end(start, end);
394	end = p4d_addr_end(start, end);
395	end = pud_addr_end(start, end);
396	end = pmd_addr_end(start, end);
397
398	/* The page next to the pinned page is the first we will try to get */
399	start += PAGE_SIZE;
400	while (start < end) {
401		struct page *page = NULL;
402		pte++;
403		if (pte_present(*pte))
404			page = vm_normal_page(vma, start, *pte);
405		/*
406		 * Break if page could not be obtained or the page's node+zone does not
407		 * match
408		 */
409		if (!page || page_zone(page) != zone)
410			break;
411
412		/*
413		 * Do not use pagevec for PTE-mapped THP,
414		 * munlock_vma_pages_range() will handle them.
415		 */
416		if (PageTransCompound(page))
417			break;
418
419		get_page(page);
420		/*
421		 * Increase the address that will be returned *before* the
422		 * eventual break due to pvec becoming full by adding the page
423		 */
424		start += PAGE_SIZE;
425		if (pagevec_add(pvec, page) == 0)
426			break;
427	}
428	pte_unmap_unlock(pte, ptl);
429	return start;
430}
431
432/*
433 * munlock_vma_pages_range() - munlock all pages in the vma range.'
434 * @vma - vma containing range to be munlock()ed.
435 * @start - start address in @vma of the range
436 * @end - end of range in @vma.
437 *
438 *  For mremap(), munmap() and exit().
439 *
440 * Called with @vma VM_LOCKED.
441 *
442 * Returns with VM_LOCKED cleared.  Callers must be prepared to
443 * deal with this.
444 *
445 * We don't save and restore VM_LOCKED here because pages are
446 * still on lru.  In unmap path, pages might be scanned by reclaim
447 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
448 * free them.  This will result in freeing mlocked pages.
449 */
450void munlock_vma_pages_range(struct vm_area_struct *vma,
451			     unsigned long start, unsigned long end)
452{
453	vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
454
455	while (start < end) {
456		struct page *page;
457		unsigned int page_mask = 0;
458		unsigned long page_increm;
459		struct pagevec pvec;
460		struct zone *zone;
 
461
462		pagevec_init(&pvec);
463		/*
464		 * Although FOLL_DUMP is intended for get_dump_page(),
465		 * it just so happens that its special treatment of the
466		 * ZERO_PAGE (returning an error instead of doing get_page)
467		 * suits munlock very well (and if somehow an abnormal page
468		 * has sneaked into the range, we won't oops here: great).
469		 */
470		page = follow_page(vma, start, FOLL_GET | FOLL_DUMP);
471
472		if (page && !IS_ERR(page)) {
473			if (PageTransTail(page)) {
474				VM_BUG_ON_PAGE(PageMlocked(page), page);
475				put_page(page); /* follow_page_mask() */
476			} else if (PageTransHuge(page)) {
477				lock_page(page);
478				/*
479				 * Any THP page found by follow_page_mask() may
480				 * have gotten split before reaching
481				 * munlock_vma_page(), so we need to compute
482				 * the page_mask here instead.
483				 */
484				page_mask = munlock_vma_page(page);
485				unlock_page(page);
486				put_page(page); /* follow_page_mask() */
487			} else {
488				/*
489				 * Non-huge pages are handled in batches via
490				 * pagevec. The pin from follow_page_mask()
491				 * prevents them from collapsing by THP.
492				 */
493				pagevec_add(&pvec, page);
494				zone = page_zone(page);
 
495
496				/*
497				 * Try to fill the rest of pagevec using fast
498				 * pte walk. This will also update start to
499				 * the next page to process. Then munlock the
500				 * pagevec.
501				 */
502				start = __munlock_pagevec_fill(&pvec, vma,
503						zone, start, end);
504				__munlock_pagevec(&pvec, zone);
505				goto next;
506			}
507		}
508		page_increm = 1 + page_mask;
509		start += page_increm * PAGE_SIZE;
510next:
511		cond_resched();
512	}
513}
514
515/*
516 * mlock_fixup  - handle mlock[all]/munlock[all] requests.
517 *
518 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
519 * munlock is a no-op.  However, for some special vmas, we go ahead and
520 * populate the ptes.
521 *
522 * For vmas that pass the filters, merge/split as appropriate.
523 */
524static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
525	unsigned long start, unsigned long end, vm_flags_t newflags)
526{
527	struct mm_struct *mm = vma->vm_mm;
528	pgoff_t pgoff;
529	int nr_pages;
530	int ret = 0;
531	int lock = !!(newflags & VM_LOCKED);
532	vm_flags_t old_flags = vma->vm_flags;
533
534	if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
535	    is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm) ||
536	    vma_is_dax(vma))
537		/* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
538		goto out;
539
540	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
541	*prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
542			  vma->vm_file, pgoff, vma_policy(vma),
543			  vma->vm_userfaultfd_ctx);
544	if (*prev) {
545		vma = *prev;
546		goto success;
547	}
548
549	if (start != vma->vm_start) {
550		ret = split_vma(mm, vma, start, 1);
551		if (ret)
552			goto out;
553	}
554
555	if (end != vma->vm_end) {
556		ret = split_vma(mm, vma, end, 0);
557		if (ret)
558			goto out;
559	}
560
561success:
562	/*
563	 * Keep track of amount of locked VM.
564	 */
565	nr_pages = (end - start) >> PAGE_SHIFT;
566	if (!lock)
567		nr_pages = -nr_pages;
568	else if (old_flags & VM_LOCKED)
569		nr_pages = 0;
570	mm->locked_vm += nr_pages;
571
572	/*
573	 * vm_flags is protected by the mmap_lock held in write mode.
574	 * It's okay if try_to_unmap_one unmaps a page just after we
575	 * set VM_LOCKED, populate_vma_page_range will bring it back.
576	 */
577
578	if (lock)
579		vma->vm_flags = newflags;
580	else
581		munlock_vma_pages_range(vma, start, end);
582
583out:
584	*prev = vma;
585	return ret;
586}
587
588static int apply_vma_lock_flags(unsigned long start, size_t len,
589				vm_flags_t flags)
590{
591	unsigned long nstart, end, tmp;
592	struct vm_area_struct * vma, * prev;
593	int error;
594
595	VM_BUG_ON(offset_in_page(start));
596	VM_BUG_ON(len != PAGE_ALIGN(len));
597	end = start + len;
598	if (end < start)
599		return -EINVAL;
600	if (end == start)
601		return 0;
602	vma = find_vma(current->mm, start);
603	if (!vma || vma->vm_start > start)
604		return -ENOMEM;
605
606	prev = vma->vm_prev;
607	if (start > vma->vm_start)
608		prev = vma;
609
610	for (nstart = start ; ; ) {
611		vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
612
613		newflags |= flags;
614
615		/* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
616		tmp = vma->vm_end;
617		if (tmp > end)
618			tmp = end;
619		error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
620		if (error)
621			break;
622		nstart = tmp;
623		if (nstart < prev->vm_end)
624			nstart = prev->vm_end;
625		if (nstart >= end)
626			break;
627
628		vma = prev->vm_next;
629		if (!vma || vma->vm_start != nstart) {
630			error = -ENOMEM;
631			break;
632		}
633	}
634	return error;
635}
636
637/*
638 * Go through vma areas and sum size of mlocked
639 * vma pages, as return value.
640 * Note deferred memory locking case(mlock2(,,MLOCK_ONFAULT)
641 * is also counted.
642 * Return value: previously mlocked page counts
643 */
644static unsigned long count_mm_mlocked_page_nr(struct mm_struct *mm,
645		unsigned long start, size_t len)
646{
647	struct vm_area_struct *vma;
648	unsigned long count = 0;
649
650	if (mm == NULL)
651		mm = current->mm;
652
653	vma = find_vma(mm, start);
654	if (vma == NULL)
655		vma = mm->mmap;
656
657	for (; vma ; vma = vma->vm_next) {
658		if (start >= vma->vm_end)
659			continue;
660		if (start + len <=  vma->vm_start)
661			break;
662		if (vma->vm_flags & VM_LOCKED) {
663			if (start > vma->vm_start)
664				count -= (start - vma->vm_start);
665			if (start + len < vma->vm_end) {
666				count += start + len - vma->vm_start;
667				break;
668			}
669			count += vma->vm_end - vma->vm_start;
670		}
671	}
672
673	return count >> PAGE_SHIFT;
674}
675
676static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
677{
678	unsigned long locked;
679	unsigned long lock_limit;
680	int error = -ENOMEM;
681
682	start = untagged_addr(start);
683
684	if (!can_do_mlock())
685		return -EPERM;
686
 
 
687	len = PAGE_ALIGN(len + (offset_in_page(start)));
688	start &= PAGE_MASK;
689
690	lock_limit = rlimit(RLIMIT_MEMLOCK);
691	lock_limit >>= PAGE_SHIFT;
692	locked = len >> PAGE_SHIFT;
693
694	if (mmap_write_lock_killable(current->mm))
695		return -EINTR;
696
697	locked += current->mm->locked_vm;
698	if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
699		/*
700		 * It is possible that the regions requested intersect with
701		 * previously mlocked areas, that part area in "mm->locked_vm"
702		 * should not be counted to new mlock increment count. So check
703		 * and adjust locked count if necessary.
704		 */
705		locked -= count_mm_mlocked_page_nr(current->mm,
706				start, len);
707	}
708
709	/* check against resource limits */
710	if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
711		error = apply_vma_lock_flags(start, len, flags);
712
713	mmap_write_unlock(current->mm);
714	if (error)
715		return error;
716
717	error = __mm_populate(start, len, 0);
718	if (error)
719		return __mlock_posix_error_return(error);
720	return 0;
721}
722
723SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
724{
725	return do_mlock(start, len, VM_LOCKED);
726}
727
728SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
729{
730	vm_flags_t vm_flags = VM_LOCKED;
731
732	if (flags & ~MLOCK_ONFAULT)
733		return -EINVAL;
734
735	if (flags & MLOCK_ONFAULT)
736		vm_flags |= VM_LOCKONFAULT;
737
738	return do_mlock(start, len, vm_flags);
739}
740
741SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
742{
743	int ret;
744
745	start = untagged_addr(start);
746
747	len = PAGE_ALIGN(len + (offset_in_page(start)));
748	start &= PAGE_MASK;
749
750	if (mmap_write_lock_killable(current->mm))
751		return -EINTR;
752	ret = apply_vma_lock_flags(start, len, 0);
753	mmap_write_unlock(current->mm);
754
755	return ret;
756}
757
758/*
759 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
760 * and translate into the appropriate modifications to mm->def_flags and/or the
761 * flags for all current VMAs.
762 *
763 * There are a couple of subtleties with this.  If mlockall() is called multiple
764 * times with different flags, the values do not necessarily stack.  If mlockall
765 * is called once including the MCL_FUTURE flag and then a second time without
766 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
767 */
768static int apply_mlockall_flags(int flags)
769{
770	struct vm_area_struct * vma, * prev = NULL;
771	vm_flags_t to_add = 0;
772
773	current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
774	if (flags & MCL_FUTURE) {
775		current->mm->def_flags |= VM_LOCKED;
776
777		if (flags & MCL_ONFAULT)
778			current->mm->def_flags |= VM_LOCKONFAULT;
779
780		if (!(flags & MCL_CURRENT))
781			goto out;
782	}
783
784	if (flags & MCL_CURRENT) {
785		to_add |= VM_LOCKED;
786		if (flags & MCL_ONFAULT)
787			to_add |= VM_LOCKONFAULT;
788	}
789
790	for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
791		vm_flags_t newflags;
792
793		newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
794		newflags |= to_add;
795
796		/* Ignore errors */
797		mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
798		cond_resched();
799	}
800out:
801	return 0;
802}
803
804SYSCALL_DEFINE1(mlockall, int, flags)
805{
806	unsigned long lock_limit;
807	int ret;
808
809	if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)) ||
810	    flags == MCL_ONFAULT)
811		return -EINVAL;
812
813	if (!can_do_mlock())
814		return -EPERM;
815
 
 
 
816	lock_limit = rlimit(RLIMIT_MEMLOCK);
817	lock_limit >>= PAGE_SHIFT;
818
819	if (mmap_write_lock_killable(current->mm))
820		return -EINTR;
821
822	ret = -ENOMEM;
823	if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
824	    capable(CAP_IPC_LOCK))
825		ret = apply_mlockall_flags(flags);
826	mmap_write_unlock(current->mm);
827	if (!ret && (flags & MCL_CURRENT))
828		mm_populate(0, TASK_SIZE);
829
830	return ret;
831}
832
833SYSCALL_DEFINE0(munlockall)
834{
835	int ret;
836
837	if (mmap_write_lock_killable(current->mm))
838		return -EINTR;
839	ret = apply_mlockall_flags(0);
840	mmap_write_unlock(current->mm);
841	return ret;
842}
843
844/*
845 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
846 * shm segments) get accounted against the user_struct instead.
847 */
848static DEFINE_SPINLOCK(shmlock_user_lock);
849
850int user_shm_lock(size_t size, struct user_struct *user)
851{
852	unsigned long lock_limit, locked;
853	int allowed = 0;
854
855	locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
856	lock_limit = rlimit(RLIMIT_MEMLOCK);
857	if (lock_limit == RLIM_INFINITY)
858		allowed = 1;
859	lock_limit >>= PAGE_SHIFT;
860	spin_lock(&shmlock_user_lock);
861	if (!allowed &&
862	    locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
863		goto out;
864	get_uid(user);
865	user->locked_shm += locked;
866	allowed = 1;
867out:
868	spin_unlock(&shmlock_user_lock);
869	return allowed;
870}
871
872void user_shm_unlock(size_t size, struct user_struct *user)
873{
874	spin_lock(&shmlock_user_lock);
875	user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
876	spin_unlock(&shmlock_user_lock);
877	free_uid(user);
878}
v4.10.11
 
  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/pagevec.h>
 15#include <linux/mempolicy.h>
 16#include <linux/syscalls.h>
 17#include <linux/sched.h>
 18#include <linux/export.h>
 19#include <linux/rmap.h>
 20#include <linux/mmzone.h>
 21#include <linux/hugetlb.h>
 22#include <linux/memcontrol.h>
 23#include <linux/mm_inline.h>
 24
 25#include "internal.h"
 26
 27bool can_do_mlock(void)
 28{
 29	if (rlimit(RLIMIT_MEMLOCK) != 0)
 30		return true;
 31	if (capable(CAP_IPC_LOCK))
 32		return true;
 33	return false;
 34}
 35EXPORT_SYMBOL(can_do_mlock);
 36
 37/*
 38 * Mlocked pages are marked with PageMlocked() flag for efficient testing
 39 * in vmscan and, possibly, the fault path; and to support semi-accurate
 40 * statistics.
 41 *
 42 * An mlocked page [PageMlocked(page)] is unevictable.  As such, it will
 43 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
 44 * The unevictable list is an LRU sibling list to the [in]active lists.
 45 * PageUnevictable is set to indicate the unevictable state.
 46 *
 47 * When lazy mlocking via vmscan, it is important to ensure that the
 48 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
 49 * may have mlocked a page that is being munlocked. So lazy mlock must take
 50 * the mmap_sem for read, and verify that the vma really is locked
 51 * (see mm/rmap.c).
 52 */
 53
 54/*
 55 *  LRU accounting for clear_page_mlock()
 56 */
 57void clear_page_mlock(struct page *page)
 58{
 
 
 59	if (!TestClearPageMlocked(page))
 60		return;
 61
 62	mod_zone_page_state(page_zone(page), NR_MLOCK,
 63			    -hpage_nr_pages(page));
 64	count_vm_event(UNEVICTABLE_PGCLEARED);
 
 
 
 
 
 
 65	if (!isolate_lru_page(page)) {
 66		putback_lru_page(page);
 67	} else {
 68		/*
 69		 * We lost the race. the page already moved to evictable list.
 70		 */
 71		if (PageUnevictable(page))
 72			count_vm_event(UNEVICTABLE_PGSTRANDED);
 73	}
 74}
 75
 76/*
 77 * Mark page as mlocked if not already.
 78 * If page on LRU, isolate and putback to move to unevictable list.
 79 */
 80void mlock_vma_page(struct page *page)
 81{
 82	/* Serialize with page migration */
 83	BUG_ON(!PageLocked(page));
 84
 85	VM_BUG_ON_PAGE(PageTail(page), page);
 86	VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
 87
 88	if (!TestSetPageMlocked(page)) {
 89		mod_zone_page_state(page_zone(page), NR_MLOCK,
 90				    hpage_nr_pages(page));
 91		count_vm_event(UNEVICTABLE_PGMLOCKED);
 
 92		if (!isolate_lru_page(page))
 93			putback_lru_page(page);
 94	}
 95}
 96
 97/*
 98 * Isolate a page from LRU with optional get_page() pin.
 99 * Assumes lru_lock already held and page already pinned.
100 */
101static bool __munlock_isolate_lru_page(struct page *page, bool getpage)
102{
103	if (PageLRU(page)) {
104		struct lruvec *lruvec;
105
106		lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
107		if (getpage)
108			get_page(page);
109		ClearPageLRU(page);
110		del_page_from_lru_list(page, lruvec, page_lru(page));
111		return true;
112	}
113
114	return false;
115}
116
117/*
118 * Finish munlock after successful page isolation
119 *
120 * Page must be locked. This is a wrapper for try_to_munlock()
121 * and putback_lru_page() with munlock accounting.
122 */
123static void __munlock_isolated_page(struct page *page)
124{
125	int ret = SWAP_AGAIN;
126
127	/*
128	 * Optimization: if the page was mapped just once, that's our mapping
129	 * and we don't need to check all the other vmas.
130	 */
131	if (page_mapcount(page) > 1)
132		ret = try_to_munlock(page);
133
134	/* Did try_to_unlock() succeed or punt? */
135	if (ret != SWAP_MLOCK)
136		count_vm_event(UNEVICTABLE_PGMUNLOCKED);
137
138	putback_lru_page(page);
139}
140
141/*
142 * Accounting for page isolation fail during munlock
143 *
144 * Performs accounting when page isolation fails in munlock. There is nothing
145 * else to do because it means some other task has already removed the page
146 * from the LRU. putback_lru_page() will take care of removing the page from
147 * the unevictable list, if necessary. vmscan [page_referenced()] will move
148 * the page back to the unevictable list if some other vma has it mlocked.
149 */
150static void __munlock_isolation_failed(struct page *page)
151{
 
 
152	if (PageUnevictable(page))
153		__count_vm_event(UNEVICTABLE_PGSTRANDED);
154	else
155		__count_vm_event(UNEVICTABLE_PGMUNLOCKED);
156}
157
158/**
159 * munlock_vma_page - munlock a vma page
160 * @page - page to be unlocked, either a normal page or THP page head
161 *
162 * returns the size of the page as a page mask (0 for normal page,
163 *         HPAGE_PMD_NR - 1 for THP head page)
164 *
165 * called from munlock()/munmap() path with page supposedly on the LRU.
166 * When we munlock a page, because the vma where we found the page is being
167 * munlock()ed or munmap()ed, we want to check whether other vmas hold the
168 * page locked so that we can leave it on the unevictable lru list and not
169 * bother vmscan with it.  However, to walk the page's rmap list in
170 * try_to_munlock() we must isolate the page from the LRU.  If some other
171 * task has removed the page from the LRU, we won't be able to do that.
172 * So we clear the PageMlocked as we might not get another chance.  If we
173 * can't isolate the page, we leave it for putback_lru_page() and vmscan
174 * [page_referenced()/try_to_unmap()] to deal with.
175 */
176unsigned int munlock_vma_page(struct page *page)
177{
178	int nr_pages;
179	struct zone *zone = page_zone(page);
180
181	/* For try_to_munlock() and to serialize with page migration */
182	BUG_ON(!PageLocked(page));
183
184	VM_BUG_ON_PAGE(PageTail(page), page);
185
186	/*
187	 * Serialize with any parallel __split_huge_page_refcount() which
188	 * might otherwise copy PageMlocked to part of the tail pages before
189	 * we clear it in the head page. It also stabilizes hpage_nr_pages().
190	 */
191	spin_lock_irq(zone_lru_lock(zone));
192
193	if (!TestClearPageMlocked(page)) {
194		/* Potentially, PTE-mapped THP: do not skip the rest PTEs */
195		nr_pages = 1;
196		goto unlock_out;
197	}
198
199	nr_pages = hpage_nr_pages(page);
200	__mod_zone_page_state(zone, NR_MLOCK, -nr_pages);
201
202	if (__munlock_isolate_lru_page(page, true)) {
203		spin_unlock_irq(zone_lru_lock(zone));
204		__munlock_isolated_page(page);
205		goto out;
206	}
207	__munlock_isolation_failed(page);
208
209unlock_out:
210	spin_unlock_irq(zone_lru_lock(zone));
211
212out:
213	return nr_pages - 1;
214}
215
216/*
217 * convert get_user_pages() return value to posix mlock() error
218 */
219static int __mlock_posix_error_return(long retval)
220{
221	if (retval == -EFAULT)
222		retval = -ENOMEM;
223	else if (retval == -ENOMEM)
224		retval = -EAGAIN;
225	return retval;
226}
227
228/*
229 * Prepare page for fast batched LRU putback via putback_lru_evictable_pagevec()
230 *
231 * The fast path is available only for evictable pages with single mapping.
232 * Then we can bypass the per-cpu pvec and get better performance.
233 * when mapcount > 1 we need try_to_munlock() which can fail.
234 * when !page_evictable(), we need the full redo logic of putback_lru_page to
235 * avoid leaving evictable page in unevictable list.
236 *
237 * In case of success, @page is added to @pvec and @pgrescued is incremented
238 * in case that the page was previously unevictable. @page is also unlocked.
239 */
240static bool __putback_lru_fast_prepare(struct page *page, struct pagevec *pvec,
241		int *pgrescued)
242{
243	VM_BUG_ON_PAGE(PageLRU(page), page);
244	VM_BUG_ON_PAGE(!PageLocked(page), page);
245
246	if (page_mapcount(page) <= 1 && page_evictable(page)) {
247		pagevec_add(pvec, page);
248		if (TestClearPageUnevictable(page))
249			(*pgrescued)++;
250		unlock_page(page);
251		return true;
252	}
253
254	return false;
255}
256
257/*
258 * Putback multiple evictable pages to the LRU
259 *
260 * Batched putback of evictable pages that bypasses the per-cpu pvec. Some of
261 * the pages might have meanwhile become unevictable but that is OK.
262 */
263static void __putback_lru_fast(struct pagevec *pvec, int pgrescued)
264{
265	count_vm_events(UNEVICTABLE_PGMUNLOCKED, pagevec_count(pvec));
266	/*
267	 *__pagevec_lru_add() calls release_pages() so we don't call
268	 * put_page() explicitly
269	 */
270	__pagevec_lru_add(pvec);
271	count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
272}
273
274/*
275 * Munlock a batch of pages from the same zone
276 *
277 * The work is split to two main phases. First phase clears the Mlocked flag
278 * and attempts to isolate the pages, all under a single zone lru lock.
279 * The second phase finishes the munlock only for pages where isolation
280 * succeeded.
281 *
282 * Note that the pagevec may be modified during the process.
283 */
284static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone)
285{
286	int i;
287	int nr = pagevec_count(pvec);
288	int delta_munlocked;
289	struct pagevec pvec_putback;
290	int pgrescued = 0;
291
292	pagevec_init(&pvec_putback, 0);
293
294	/* Phase 1: page isolation */
295	spin_lock_irq(zone_lru_lock(zone));
296	for (i = 0; i < nr; i++) {
297		struct page *page = pvec->pages[i];
298
299		if (TestClearPageMlocked(page)) {
300			/*
301			 * We already have pin from follow_page_mask()
302			 * so we can spare the get_page() here.
303			 */
304			if (__munlock_isolate_lru_page(page, false))
305				continue;
306			else
307				__munlock_isolation_failed(page);
 
 
308		}
309
310		/*
311		 * We won't be munlocking this page in the next phase
312		 * but we still need to release the follow_page_mask()
313		 * pin. We cannot do it under lru_lock however. If it's
314		 * the last pin, __page_cache_release() would deadlock.
315		 */
316		pagevec_add(&pvec_putback, pvec->pages[i]);
317		pvec->pages[i] = NULL;
318	}
319	delta_munlocked = -nr + pagevec_count(&pvec_putback);
320	__mod_zone_page_state(zone, NR_MLOCK, delta_munlocked);
321	spin_unlock_irq(zone_lru_lock(zone));
322
323	/* Now we can release pins of pages that we are not munlocking */
324	pagevec_release(&pvec_putback);
325
326	/* Phase 2: page munlock */
327	for (i = 0; i < nr; i++) {
328		struct page *page = pvec->pages[i];
329
330		if (page) {
331			lock_page(page);
332			if (!__putback_lru_fast_prepare(page, &pvec_putback,
333					&pgrescued)) {
334				/*
335				 * Slow path. We don't want to lose the last
336				 * pin before unlock_page()
337				 */
338				get_page(page); /* for putback_lru_page() */
339				__munlock_isolated_page(page);
340				unlock_page(page);
341				put_page(page); /* from follow_page_mask() */
342			}
343		}
344	}
345
346	/*
347	 * Phase 3: page putback for pages that qualified for the fast path
348	 * This will also call put_page() to return pin from follow_page_mask()
349	 */
350	if (pagevec_count(&pvec_putback))
351		__putback_lru_fast(&pvec_putback, pgrescued);
352}
353
354/*
355 * Fill up pagevec for __munlock_pagevec using pte walk
356 *
357 * The function expects that the struct page corresponding to @start address is
358 * a non-TPH page already pinned and in the @pvec, and that it belongs to @zone.
359 *
360 * The rest of @pvec is filled by subsequent pages within the same pmd and same
361 * zone, as long as the pte's are present and vm_normal_page() succeeds. These
362 * pages also get pinned.
363 *
364 * Returns the address of the next page that should be scanned. This equals
365 * @start + PAGE_SIZE when no page could be added by the pte walk.
366 */
367static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
368		struct vm_area_struct *vma, int zoneid,	unsigned long start,
369		unsigned long end)
370{
371	pte_t *pte;
372	spinlock_t *ptl;
373
374	/*
375	 * Initialize pte walk starting at the already pinned page where we
376	 * are sure that there is a pte, as it was pinned under the same
377	 * mmap_sem write op.
378	 */
379	pte = get_locked_pte(vma->vm_mm, start,	&ptl);
380	/* Make sure we do not cross the page table boundary */
381	end = pgd_addr_end(start, end);
 
382	end = pud_addr_end(start, end);
383	end = pmd_addr_end(start, end);
384
385	/* The page next to the pinned page is the first we will try to get */
386	start += PAGE_SIZE;
387	while (start < end) {
388		struct page *page = NULL;
389		pte++;
390		if (pte_present(*pte))
391			page = vm_normal_page(vma, start, *pte);
392		/*
393		 * Break if page could not be obtained or the page's node+zone does not
394		 * match
395		 */
396		if (!page || page_zone_id(page) != zoneid)
397			break;
398
399		/*
400		 * Do not use pagevec for PTE-mapped THP,
401		 * munlock_vma_pages_range() will handle them.
402		 */
403		if (PageTransCompound(page))
404			break;
405
406		get_page(page);
407		/*
408		 * Increase the address that will be returned *before* the
409		 * eventual break due to pvec becoming full by adding the page
410		 */
411		start += PAGE_SIZE;
412		if (pagevec_add(pvec, page) == 0)
413			break;
414	}
415	pte_unmap_unlock(pte, ptl);
416	return start;
417}
418
419/*
420 * munlock_vma_pages_range() - munlock all pages in the vma range.'
421 * @vma - vma containing range to be munlock()ed.
422 * @start - start address in @vma of the range
423 * @end - end of range in @vma.
424 *
425 *  For mremap(), munmap() and exit().
426 *
427 * Called with @vma VM_LOCKED.
428 *
429 * Returns with VM_LOCKED cleared.  Callers must be prepared to
430 * deal with this.
431 *
432 * We don't save and restore VM_LOCKED here because pages are
433 * still on lru.  In unmap path, pages might be scanned by reclaim
434 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
435 * free them.  This will result in freeing mlocked pages.
436 */
437void munlock_vma_pages_range(struct vm_area_struct *vma,
438			     unsigned long start, unsigned long end)
439{
440	vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
441
442	while (start < end) {
443		struct page *page;
444		unsigned int page_mask = 0;
445		unsigned long page_increm;
446		struct pagevec pvec;
447		struct zone *zone;
448		int zoneid;
449
450		pagevec_init(&pvec, 0);
451		/*
452		 * Although FOLL_DUMP is intended for get_dump_page(),
453		 * it just so happens that its special treatment of the
454		 * ZERO_PAGE (returning an error instead of doing get_page)
455		 * suits munlock very well (and if somehow an abnormal page
456		 * has sneaked into the range, we won't oops here: great).
457		 */
458		page = follow_page(vma, start, FOLL_GET | FOLL_DUMP);
459
460		if (page && !IS_ERR(page)) {
461			if (PageTransTail(page)) {
462				VM_BUG_ON_PAGE(PageMlocked(page), page);
463				put_page(page); /* follow_page_mask() */
464			} else if (PageTransHuge(page)) {
465				lock_page(page);
466				/*
467				 * Any THP page found by follow_page_mask() may
468				 * have gotten split before reaching
469				 * munlock_vma_page(), so we need to compute
470				 * the page_mask here instead.
471				 */
472				page_mask = munlock_vma_page(page);
473				unlock_page(page);
474				put_page(page); /* follow_page_mask() */
475			} else {
476				/*
477				 * Non-huge pages are handled in batches via
478				 * pagevec. The pin from follow_page_mask()
479				 * prevents them from collapsing by THP.
480				 */
481				pagevec_add(&pvec, page);
482				zone = page_zone(page);
483				zoneid = page_zone_id(page);
484
485				/*
486				 * Try to fill the rest of pagevec using fast
487				 * pte walk. This will also update start to
488				 * the next page to process. Then munlock the
489				 * pagevec.
490				 */
491				start = __munlock_pagevec_fill(&pvec, vma,
492						zoneid, start, end);
493				__munlock_pagevec(&pvec, zone);
494				goto next;
495			}
496		}
497		page_increm = 1 + page_mask;
498		start += page_increm * PAGE_SIZE;
499next:
500		cond_resched();
501	}
502}
503
504/*
505 * mlock_fixup  - handle mlock[all]/munlock[all] requests.
506 *
507 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
508 * munlock is a no-op.  However, for some special vmas, we go ahead and
509 * populate the ptes.
510 *
511 * For vmas that pass the filters, merge/split as appropriate.
512 */
513static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
514	unsigned long start, unsigned long end, vm_flags_t newflags)
515{
516	struct mm_struct *mm = vma->vm_mm;
517	pgoff_t pgoff;
518	int nr_pages;
519	int ret = 0;
520	int lock = !!(newflags & VM_LOCKED);
521	vm_flags_t old_flags = vma->vm_flags;
522
523	if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
524	    is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
 
525		/* don't set VM_LOCKED or VM_LOCKONFAULT and don't count */
526		goto out;
527
528	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
529	*prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
530			  vma->vm_file, pgoff, vma_policy(vma),
531			  vma->vm_userfaultfd_ctx);
532	if (*prev) {
533		vma = *prev;
534		goto success;
535	}
536
537	if (start != vma->vm_start) {
538		ret = split_vma(mm, vma, start, 1);
539		if (ret)
540			goto out;
541	}
542
543	if (end != vma->vm_end) {
544		ret = split_vma(mm, vma, end, 0);
545		if (ret)
546			goto out;
547	}
548
549success:
550	/*
551	 * Keep track of amount of locked VM.
552	 */
553	nr_pages = (end - start) >> PAGE_SHIFT;
554	if (!lock)
555		nr_pages = -nr_pages;
556	else if (old_flags & VM_LOCKED)
557		nr_pages = 0;
558	mm->locked_vm += nr_pages;
559
560	/*
561	 * vm_flags is protected by the mmap_sem held in write mode.
562	 * It's okay if try_to_unmap_one unmaps a page just after we
563	 * set VM_LOCKED, populate_vma_page_range will bring it back.
564	 */
565
566	if (lock)
567		vma->vm_flags = newflags;
568	else
569		munlock_vma_pages_range(vma, start, end);
570
571out:
572	*prev = vma;
573	return ret;
574}
575
576static int apply_vma_lock_flags(unsigned long start, size_t len,
577				vm_flags_t flags)
578{
579	unsigned long nstart, end, tmp;
580	struct vm_area_struct * vma, * prev;
581	int error;
582
583	VM_BUG_ON(offset_in_page(start));
584	VM_BUG_ON(len != PAGE_ALIGN(len));
585	end = start + len;
586	if (end < start)
587		return -EINVAL;
588	if (end == start)
589		return 0;
590	vma = find_vma(current->mm, start);
591	if (!vma || vma->vm_start > start)
592		return -ENOMEM;
593
594	prev = vma->vm_prev;
595	if (start > vma->vm_start)
596		prev = vma;
597
598	for (nstart = start ; ; ) {
599		vm_flags_t newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
600
601		newflags |= flags;
602
603		/* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
604		tmp = vma->vm_end;
605		if (tmp > end)
606			tmp = end;
607		error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
608		if (error)
609			break;
610		nstart = tmp;
611		if (nstart < prev->vm_end)
612			nstart = prev->vm_end;
613		if (nstart >= end)
614			break;
615
616		vma = prev->vm_next;
617		if (!vma || vma->vm_start != nstart) {
618			error = -ENOMEM;
619			break;
620		}
621	}
622	return error;
623}
624
625/*
626 * Go through vma areas and sum size of mlocked
627 * vma pages, as return value.
628 * Note deferred memory locking case(mlock2(,,MLOCK_ONFAULT)
629 * is also counted.
630 * Return value: previously mlocked page counts
631 */
632static int count_mm_mlocked_page_nr(struct mm_struct *mm,
633		unsigned long start, size_t len)
634{
635	struct vm_area_struct *vma;
636	int count = 0;
637
638	if (mm == NULL)
639		mm = current->mm;
640
641	vma = find_vma(mm, start);
642	if (vma == NULL)
643		vma = mm->mmap;
644
645	for (; vma ; vma = vma->vm_next) {
646		if (start >= vma->vm_end)
647			continue;
648		if (start + len <=  vma->vm_start)
649			break;
650		if (vma->vm_flags & VM_LOCKED) {
651			if (start > vma->vm_start)
652				count -= (start - vma->vm_start);
653			if (start + len < vma->vm_end) {
654				count += start + len - vma->vm_start;
655				break;
656			}
657			count += vma->vm_end - vma->vm_start;
658		}
659	}
660
661	return count >> PAGE_SHIFT;
662}
663
664static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t flags)
665{
666	unsigned long locked;
667	unsigned long lock_limit;
668	int error = -ENOMEM;
669
 
 
670	if (!can_do_mlock())
671		return -EPERM;
672
673	lru_add_drain_all();	/* flush pagevec */
674
675	len = PAGE_ALIGN(len + (offset_in_page(start)));
676	start &= PAGE_MASK;
677
678	lock_limit = rlimit(RLIMIT_MEMLOCK);
679	lock_limit >>= PAGE_SHIFT;
680	locked = len >> PAGE_SHIFT;
681
682	if (down_write_killable(&current->mm->mmap_sem))
683		return -EINTR;
684
685	locked += current->mm->locked_vm;
686	if ((locked > lock_limit) && (!capable(CAP_IPC_LOCK))) {
687		/*
688		 * It is possible that the regions requested intersect with
689		 * previously mlocked areas, that part area in "mm->locked_vm"
690		 * should not be counted to new mlock increment count. So check
691		 * and adjust locked count if necessary.
692		 */
693		locked -= count_mm_mlocked_page_nr(current->mm,
694				start, len);
695	}
696
697	/* check against resource limits */
698	if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
699		error = apply_vma_lock_flags(start, len, flags);
700
701	up_write(&current->mm->mmap_sem);
702	if (error)
703		return error;
704
705	error = __mm_populate(start, len, 0);
706	if (error)
707		return __mlock_posix_error_return(error);
708	return 0;
709}
710
711SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
712{
713	return do_mlock(start, len, VM_LOCKED);
714}
715
716SYSCALL_DEFINE3(mlock2, unsigned long, start, size_t, len, int, flags)
717{
718	vm_flags_t vm_flags = VM_LOCKED;
719
720	if (flags & ~MLOCK_ONFAULT)
721		return -EINVAL;
722
723	if (flags & MLOCK_ONFAULT)
724		vm_flags |= VM_LOCKONFAULT;
725
726	return do_mlock(start, len, vm_flags);
727}
728
729SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
730{
731	int ret;
732
 
 
733	len = PAGE_ALIGN(len + (offset_in_page(start)));
734	start &= PAGE_MASK;
735
736	if (down_write_killable(&current->mm->mmap_sem))
737		return -EINTR;
738	ret = apply_vma_lock_flags(start, len, 0);
739	up_write(&current->mm->mmap_sem);
740
741	return ret;
742}
743
744/*
745 * Take the MCL_* flags passed into mlockall (or 0 if called from munlockall)
746 * and translate into the appropriate modifications to mm->def_flags and/or the
747 * flags for all current VMAs.
748 *
749 * There are a couple of subtleties with this.  If mlockall() is called multiple
750 * times with different flags, the values do not necessarily stack.  If mlockall
751 * is called once including the MCL_FUTURE flag and then a second time without
752 * it, VM_LOCKED and VM_LOCKONFAULT will be cleared from mm->def_flags.
753 */
754static int apply_mlockall_flags(int flags)
755{
756	struct vm_area_struct * vma, * prev = NULL;
757	vm_flags_t to_add = 0;
758
759	current->mm->def_flags &= VM_LOCKED_CLEAR_MASK;
760	if (flags & MCL_FUTURE) {
761		current->mm->def_flags |= VM_LOCKED;
762
763		if (flags & MCL_ONFAULT)
764			current->mm->def_flags |= VM_LOCKONFAULT;
765
766		if (!(flags & MCL_CURRENT))
767			goto out;
768	}
769
770	if (flags & MCL_CURRENT) {
771		to_add |= VM_LOCKED;
772		if (flags & MCL_ONFAULT)
773			to_add |= VM_LOCKONFAULT;
774	}
775
776	for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
777		vm_flags_t newflags;
778
779		newflags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
780		newflags |= to_add;
781
782		/* Ignore errors */
783		mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
784		cond_resched_rcu_qs();
785	}
786out:
787	return 0;
788}
789
790SYSCALL_DEFINE1(mlockall, int, flags)
791{
792	unsigned long lock_limit;
793	int ret;
794
795	if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT)))
 
796		return -EINVAL;
797
798	if (!can_do_mlock())
799		return -EPERM;
800
801	if (flags & MCL_CURRENT)
802		lru_add_drain_all();	/* flush pagevec */
803
804	lock_limit = rlimit(RLIMIT_MEMLOCK);
805	lock_limit >>= PAGE_SHIFT;
806
807	if (down_write_killable(&current->mm->mmap_sem))
808		return -EINTR;
809
810	ret = -ENOMEM;
811	if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
812	    capable(CAP_IPC_LOCK))
813		ret = apply_mlockall_flags(flags);
814	up_write(&current->mm->mmap_sem);
815	if (!ret && (flags & MCL_CURRENT))
816		mm_populate(0, TASK_SIZE);
817
818	return ret;
819}
820
821SYSCALL_DEFINE0(munlockall)
822{
823	int ret;
824
825	if (down_write_killable(&current->mm->mmap_sem))
826		return -EINTR;
827	ret = apply_mlockall_flags(0);
828	up_write(&current->mm->mmap_sem);
829	return ret;
830}
831
832/*
833 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
834 * shm segments) get accounted against the user_struct instead.
835 */
836static DEFINE_SPINLOCK(shmlock_user_lock);
837
838int user_shm_lock(size_t size, struct user_struct *user)
839{
840	unsigned long lock_limit, locked;
841	int allowed = 0;
842
843	locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
844	lock_limit = rlimit(RLIMIT_MEMLOCK);
845	if (lock_limit == RLIM_INFINITY)
846		allowed = 1;
847	lock_limit >>= PAGE_SHIFT;
848	spin_lock(&shmlock_user_lock);
849	if (!allowed &&
850	    locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
851		goto out;
852	get_uid(user);
853	user->locked_shm += locked;
854	allowed = 1;
855out:
856	spin_unlock(&shmlock_user_lock);
857	return allowed;
858}
859
860void user_shm_unlock(size_t size, struct user_struct *user)
861{
862	spin_lock(&shmlock_user_lock);
863	user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
864	spin_unlock(&shmlock_user_lock);
865	free_uid(user);
866}