1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 *	linux/mm/madvise.c
  4 *
  5 * Copyright (C) 1999  Linus Torvalds
  6 * Copyright (C) 2002  Christoph Hellwig
  7 */
  8
  9#include <linux/mman.h>
 10#include <linux/pagemap.h>
 11#include <linux/syscalls.h>
 12#include <linux/mempolicy.h>
 13#include <linux/page-isolation.h>
 
 14#include <linux/userfaultfd_k.h>
 15#include <linux/hugetlb.h>
 16#include <linux/falloc.h>
 
 17#include <linux/sched.h>
 
 
 
 
 18#include <linux/ksm.h>
 19#include <linux/fs.h>
 20#include <linux/file.h>
 21#include <linux/blkdev.h>
 22#include <linux/backing-dev.h>
 
 23#include <linux/swap.h>
 24#include <linux/swapops.h>
 25#include <linux/shmem_fs.h>
 26#include <linux/mmu_notifier.h>
 27
 28#include <asm/tlb.h>
 29
 30#include "internal.h"
 
 
 
 
 
 
 
 
 
 
 
 
 31
 32/*
 33 * Any behaviour which results in changes to the vma->vm_flags needs to
 34 * take mmap_sem for writing. Others, which simply traverse vmas, need
 35 * to only take it for reading.
 36 */
 37static int madvise_need_mmap_write(int behavior)
 38{
 39	switch (behavior) {
 40	case MADV_REMOVE:
 41	case MADV_WILLNEED:
 42	case MADV_DONTNEED:
 
 
 
 43	case MADV_FREE:
 
 
 
 
 
 44		return 0;
 45	default:
 46		/* be safe, default to 1. list exceptions explicitly */
 47		return 1;
 48	}
 49}
 50
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 51/*
 52 * We can potentially split a vm area into separate
 53 * areas, each area with its own behavior.
 
 
 54 */
 55static long madvise_behavior(struct vm_area_struct *vma,
 56		     struct vm_area_struct **prev,
 57		     unsigned long start, unsigned long end, int behavior)
 
 58{
 59	struct mm_struct *mm = vma->vm_mm;
 60	int error = 0;
 61	pgoff_t pgoff;
 62	unsigned long new_flags = vma->vm_flags;
 63
 64	switch (behavior) {
 65	case MADV_NORMAL:
 66		new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
 67		break;
 68	case MADV_SEQUENTIAL:
 69		new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
 70		break;
 71	case MADV_RANDOM:
 72		new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
 73		break;
 74	case MADV_DONTFORK:
 75		new_flags |= VM_DONTCOPY;
 76		break;
 77	case MADV_DOFORK:
 78		if (vma->vm_flags & VM_IO) {
 79			error = -EINVAL;
 80			goto out;
 81		}
 82		new_flags &= ~VM_DONTCOPY;
 83		break;
 84	case MADV_WIPEONFORK:
 85		/* MADV_WIPEONFORK is only supported on anonymous memory. */
 86		if (vma->vm_file || vma->vm_flags & VM_SHARED) {
 87			error = -EINVAL;
 88			goto out;
 89		}
 90		new_flags |= VM_WIPEONFORK;
 91		break;
 92	case MADV_KEEPONFORK:
 93		new_flags &= ~VM_WIPEONFORK;
 94		break;
 95	case MADV_DONTDUMP:
 96		new_flags |= VM_DONTDUMP;
 97		break;
 98	case MADV_DODUMP:
 99		if (new_flags & VM_SPECIAL) {
100			error = -EINVAL;
101			goto out;
102		}
103		new_flags &= ~VM_DONTDUMP;
104		break;
105	case MADV_MERGEABLE:
106	case MADV_UNMERGEABLE:
107		error = ksm_madvise(vma, start, end, behavior, &new_flags);
108		if (error) {
109			/*
110			 * madvise() returns EAGAIN if kernel resources, such as
111			 * slab, are temporarily unavailable.
112			 */
113			if (error == -ENOMEM)
114				error = -EAGAIN;
115			goto out;
116		}
117		break;
118	case MADV_HUGEPAGE:
119	case MADV_NOHUGEPAGE:
120		error = hugepage_madvise(vma, &new_flags, behavior);
121		if (error) {
122			/*
123			 * madvise() returns EAGAIN if kernel resources, such as
124			 * slab, are temporarily unavailable.
125			 */
126			if (error == -ENOMEM)
127				error = -EAGAIN;
128			goto out;
129		}
130		break;
131	}
132
133	if (new_flags == vma->vm_flags) {
134		*prev = vma;
135		goto out;
136	}
137
138	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
139	*prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
140			  vma->vm_file, pgoff, vma_policy(vma),
141			  vma->vm_userfaultfd_ctx);
142	if (*prev) {
143		vma = *prev;
144		goto success;
145	}
146
147	*prev = vma;
148
149	if (start != vma->vm_start) {
150		if (unlikely(mm->map_count >= sysctl_max_map_count)) {
151			error = -ENOMEM;
152			goto out;
153		}
154		error = __split_vma(mm, vma, start, 1);
155		if (error) {
156			/*
157			 * madvise() returns EAGAIN if kernel resources, such as
158			 * slab, are temporarily unavailable.
159			 */
160			if (error == -ENOMEM)
161				error = -EAGAIN;
162			goto out;
163		}
164	}
165
166	if (end != vma->vm_end) {
167		if (unlikely(mm->map_count >= sysctl_max_map_count)) {
168			error = -ENOMEM;
169			goto out;
170		}
171		error = __split_vma(mm, vma, end, 0);
172		if (error) {
173			/*
174			 * madvise() returns EAGAIN if kernel resources, such as
175			 * slab, are temporarily unavailable.
176			 */
177			if (error == -ENOMEM)
178				error = -EAGAIN;
179			goto out;
180		}
181	}
182
183success:
184	/*
185	 * vm_flags is protected by the mmap_sem held in write mode.
186	 */
187	vma->vm_flags = new_flags;
188out:
189	return error;
190}
191
192#ifdef CONFIG_SWAP
193static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
194	unsigned long end, struct mm_walk *walk)
195{
196	pte_t *orig_pte;
197	struct vm_area_struct *vma = walk->private;
198	unsigned long index;
199
200	if (pmd_none_or_trans_huge_or_clear_bad(pmd))
201		return 0;
202
203	for (index = start; index != end; index += PAGE_SIZE) {
204		pte_t pte;
205		swp_entry_t entry;
206		struct page *page;
207		spinlock_t *ptl;
208
209		orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
210		pte = *(orig_pte + ((index - start) / PAGE_SIZE));
211		pte_unmap_unlock(orig_pte, ptl);
 
 
212
213		if (pte_present(pte) || pte_none(pte))
 
214			continue;
215		entry = pte_to_swp_entry(pte);
216		if (unlikely(non_swap_entry(entry)))
217			continue;
218
219		page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
220							vma, index, false);
221		if (page)
222			put_page(page);
 
 
 
223	}
224
 
 
 
 
 
225	return 0;
226}
227
228static void force_swapin_readahead(struct vm_area_struct *vma,
229		unsigned long start, unsigned long end)
230{
231	struct mm_walk walk = {
232		.mm = vma->vm_mm,
233		.pmd_entry = swapin_walk_pmd_entry,
234		.private = vma,
235	};
236
237	walk_page_range(start, end, &walk);
238
239	lru_add_drain();	/* Push any new pages onto the LRU now */
240}
241
242static void force_shm_swapin_readahead(struct vm_area_struct *vma,
243		unsigned long start, unsigned long end,
244		struct address_space *mapping)
245{
246	pgoff_t index;
247	struct page *page;
248	swp_entry_t swap;
249
250	for (; start < end; start += PAGE_SIZE) {
251		index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 
 
 
252
253		page = find_get_entry(mapping, index);
254		if (!radix_tree_exceptional_entry(page)) {
255			if (page)
256				put_page(page);
 
257			continue;
258		}
259		swap = radix_to_swp_entry(page);
260		page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
261							NULL, 0, false);
262		if (page)
263			put_page(page);
264	}
265
266	lru_add_drain();	/* Push any new pages onto the LRU now */
 
 
 
 
 
 
 
 
 
 
 
 
 
267}
268#endif		/* CONFIG_SWAP */
269
270/*
271 * Schedule all required I/O operations.  Do not wait for completion.
272 */
273static long madvise_willneed(struct vm_area_struct *vma,
274			     struct vm_area_struct **prev,
275			     unsigned long start, unsigned long end)
276{
 
277	struct file *file = vma->vm_file;
 
278
279	*prev = vma;
280#ifdef CONFIG_SWAP
281	if (!file) {
282		force_swapin_readahead(vma, start, end);
 
283		return 0;
284	}
285
286	if (shmem_mapping(file->f_mapping)) {
287		force_shm_swapin_readahead(vma, start, end,
288					file->f_mapping);
289		return 0;
290	}
291#else
292	if (!file)
293		return -EBADF;
294#endif
295
296	if (IS_DAX(file_inode(file))) {
297		/* no bad return value, but ignore advice */
298		return 0;
299	}
300
301	start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
302	if (end > vma->vm_end)
303		end = vma->vm_end;
304	end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
305
306	force_page_cache_readahead(file->f_mapping, file, start, end - start);
307	return 0;
308}
309
310static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
311				unsigned long end, struct mm_walk *walk)
312
313{
 
314	struct mmu_gather *tlb = walk->private;
315	struct mm_struct *mm = tlb->mm;
316	struct vm_area_struct *vma = walk->vma;
317	spinlock_t *ptl;
318	pte_t *orig_pte, *pte, ptent;
319	struct page *page;
320	int nr_swap = 0;
321	unsigned long next;
 
322
323	next = pmd_addr_end(addr, end);
324	if (pmd_trans_huge(*pmd))
325		if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
326			goto next;
327
328	if (pmd_trans_unstable(pmd))
 
 
329		return 0;
330
331	tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
332	orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
333	flush_tlb_batched_pending(mm);
334	arch_enter_lazy_mmu_mode();
335	for (; addr != end; pte++, addr += PAGE_SIZE) {
336		ptent = *pte;
 
337
338		if (pte_none(ptent))
339			continue;
340		/*
341		 * If the pte has swp_entry, just clear page table to
342		 * prevent swap-in which is more expensive rather than
343		 * (page allocation + zeroing).
344		 */
345		if (!pte_present(ptent)) {
346			swp_entry_t entry;
347
348			entry = pte_to_swp_entry(ptent);
349			if (non_swap_entry(entry))
350				continue;
351			nr_swap--;
352			free_swap_and_cache(entry);
353			pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 
 
 
 
 
354			continue;
355		}
356
357		page = _vm_normal_page(vma, addr, ptent, true);
358		if (!page)
359			continue;
360
361		/*
362		 * If pmd isn't transhuge but the page is THP and
363		 * is owned by only this process, split it and
364		 * deactivate all pages.
 
 
365		 */
366		if (PageTransCompound(page)) {
367			if (page_mapcount(page) != 1)
368				goto out;
369			get_page(page);
370			if (!trylock_page(page)) {
371				put_page(page);
372				goto out;
373			}
374			pte_unmap_unlock(orig_pte, ptl);
375			if (split_huge_page(page)) {
376				unlock_page(page);
377				put_page(page);
378				pte_offset_map_lock(mm, pmd, addr, &ptl);
379				goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
380			}
381			unlock_page(page);
382			put_page(page);
383			pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
384			pte--;
385			addr -= PAGE_SIZE;
386			continue;
387		}
388
389		VM_BUG_ON_PAGE(PageTransCompound(page), page);
 
 
 
 
390
391		if (PageSwapCache(page) || PageDirty(page)) {
392			if (!trylock_page(page))
393				continue;
394			/*
395			 * If page is shared with others, we couldn't clear
396			 * PG_dirty of the page.
 
397			 */
398			if (page_mapcount(page) != 1) {
399				unlock_page(page);
400				continue;
401			}
402
403			if (PageSwapCache(page) && !try_to_free_swap(page)) {
404				unlock_page(page);
 
405				continue;
406			}
407
408			ClearPageDirty(page);
409			unlock_page(page);
410		}
411
412		if (pte_young(ptent) || pte_dirty(ptent)) {
413			/*
414			 * Some of architecture(ex, PPC) don't update TLB
415			 * with set_pte_at and tlb_remove_tlb_entry so for
416			 * the portability, remap the pte with old|clean
417			 * after pte clearing.
418			 */
419			ptent = ptep_get_and_clear_full(mm, addr, pte,
420							tlb->fullmm);
421
422			ptent = pte_mkold(ptent);
423			ptent = pte_mkclean(ptent);
424			set_pte_at(mm, addr, pte, ptent);
425			tlb_remove_tlb_entry(tlb, pte, addr);
426		}
427		mark_page_lazyfree(page);
428	}
429out:
430	if (nr_swap) {
431		if (current->mm == mm)
432			sync_mm_rss(mm);
433
 
434		add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 
 
 
435	}
436	arch_leave_lazy_mmu_mode();
437	pte_unmap_unlock(orig_pte, ptl);
438	cond_resched();
439next:
440	return 0;
441}
442
443static void madvise_free_page_range(struct mmu_gather *tlb,
444			     struct vm_area_struct *vma,
445			     unsigned long addr, unsigned long end)
446{
447	struct mm_walk free_walk = {
448		.pmd_entry = madvise_free_pte_range,
449		.mm = vma->vm_mm,
450		.private = tlb,
451	};
452
453	tlb_start_vma(tlb, vma);
454	walk_page_range(addr, end, &free_walk);
455	tlb_end_vma(tlb, vma);
456}
457
458static int madvise_free_single_vma(struct vm_area_struct *vma,
459			unsigned long start_addr, unsigned long end_addr)
460{
461	unsigned long start, end;
462	struct mm_struct *mm = vma->vm_mm;
 
463	struct mmu_gather tlb;
464
465	/* MADV_FREE works for only anon vma at the moment */
466	if (!vma_is_anonymous(vma))
467		return -EINVAL;
468
469	start = max(vma->vm_start, start_addr);
470	if (start >= vma->vm_end)
471		return -EINVAL;
472	end = min(vma->vm_end, end_addr);
473	if (end <= vma->vm_start)
474		return -EINVAL;
 
 
475
476	lru_add_drain();
477	tlb_gather_mmu(&tlb, mm, start, end);
478	update_hiwater_rss(mm);
479
480	mmu_notifier_invalidate_range_start(mm, start, end);
481	madvise_free_page_range(&tlb, vma, start, end);
482	mmu_notifier_invalidate_range_end(mm, start, end);
483	tlb_finish_mmu(&tlb, start, end);
 
 
 
484
485	return 0;
486}
487
488/*
489 * Application no longer needs these pages.  If the pages are dirty,
490 * it's OK to just throw them away.  The app will be more careful about
491 * data it wants to keep.  Be sure to free swap resources too.  The
492 * zap_page_range call sets things up for shrink_active_list to actually free
493 * these pages later if no one else has touched them in the meantime,
494 * although we could add these pages to a global reuse list for
495 * shrink_active_list to pick up before reclaiming other pages.
496 *
497 * NB: This interface discards data rather than pushes it out to swap,
498 * as some implementations do.  This has performance implications for
499 * applications like large transactional databases which want to discard
500 * pages in anonymous maps after committing to backing store the data
501 * that was kept in them.  There is no reason to write this data out to
502 * the swap area if the application is discarding it.
503 *
504 * An interface that causes the system to free clean pages and flush
505 * dirty pages is already available as msync(MS_INVALIDATE).
506 */
507static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
508					unsigned long start, unsigned long end)
509{
510	zap_page_range(vma, start, end - start);
511	return 0;
512}
513
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
514static long madvise_dontneed_free(struct vm_area_struct *vma,
515				  struct vm_area_struct **prev,
516				  unsigned long start, unsigned long end,
517				  int behavior)
518{
 
 
519	*prev = vma;
520	if (!can_madv_dontneed_vma(vma))
521		return -EINVAL;
522
 
 
 
523	if (!userfaultfd_remove(vma, start, end)) {
524		*prev = NULL; /* mmap_sem has been dropped, prev is stale */
525
526		down_read(&current->mm->mmap_sem);
527		vma = find_vma(current->mm, start);
528		if (!vma)
529			return -ENOMEM;
530		if (start < vma->vm_start) {
531			/*
532			 * This "vma" under revalidation is the one
533			 * with the lowest vma->vm_start where start
534			 * is also < vma->vm_end. If start <
535			 * vma->vm_start it means an hole materialized
536			 * in the user address space within the
537			 * virtual range passed to MADV_DONTNEED
538			 * or MADV_FREE.
539			 */
540			return -ENOMEM;
541		}
542		if (!can_madv_dontneed_vma(vma))
543			return -EINVAL;
544		if (end > vma->vm_end) {
545			/*
546			 * Don't fail if end > vma->vm_end. If the old
547			 * vma was splitted while the mmap_sem was
548			 * released the effect of the concurrent
549			 * operation may not cause madvise() to
550			 * have an undefined result. There may be an
551			 * adjacent next vma that we'll walk
552			 * next. userfaultfd_remove() will generate an
553			 * UFFD_EVENT_REMOVE repetition on the
554			 * end-vma->vm_end range, but the manager can
555			 * handle a repetition fine.
556			 */
557			end = vma->vm_end;
558		}
559		VM_WARN_ON(start >= end);
 
 
 
 
 
 
 
 
 
560	}
561
562	if (behavior == MADV_DONTNEED)
563		return madvise_dontneed_single_vma(vma, start, end);
564	else if (behavior == MADV_FREE)
565		return madvise_free_single_vma(vma, start, end);
566	else
567		return -EINVAL;
568}
569
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
570/*
571 * Application wants to free up the pages and associated backing store.
572 * This is effectively punching a hole into the middle of a file.
573 */
574static long madvise_remove(struct vm_area_struct *vma,
575				struct vm_area_struct **prev,
576				unsigned long start, unsigned long end)
577{
578	loff_t offset;
579	int error;
580	struct file *f;
 
581
582	*prev = NULL;	/* tell sys_madvise we drop mmap_sem */
583
584	if (vma->vm_flags & VM_LOCKED)
585		return -EINVAL;
586
587	f = vma->vm_file;
588
589	if (!f || !f->f_mapping || !f->f_mapping->host) {
590			return -EINVAL;
591	}
592
593	if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
594		return -EACCES;
595
596	offset = (loff_t)(start - vma->vm_start)
597			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
598
599	/*
600	 * Filesystem's fallocate may need to take i_mutex.  We need to
601	 * explicitly grab a reference because the vma (and hence the
602	 * vma's reference to the file) can go away as soon as we drop
603	 * mmap_sem.
604	 */
605	get_file(f);
606	if (userfaultfd_remove(vma, start, end)) {
607		/* mmap_sem was not released by userfaultfd_remove() */
608		up_read(&current->mm->mmap_sem);
609	}
610	error = vfs_fallocate(f,
611				FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
612				offset, end - start);
613	fput(f);
614	down_read(&current->mm->mmap_sem);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
615	return error;
616}
617
618#ifdef CONFIG_MEMORY_FAILURE
619/*
620 * Error injection support for memory error handling.
621 */
622static int madvise_inject_error(int behavior,
623		unsigned long start, unsigned long end)
624{
625	struct page *page;
626	struct zone *zone;
627	unsigned int order;
628
629	if (!capable(CAP_SYS_ADMIN))
630		return -EPERM;
631
632
633	for (; start < end; start += PAGE_SIZE << order) {
 
 
634		int ret;
635
636		ret = get_user_pages_fast(start, 1, 0, &page);
637		if (ret != 1)
638			return ret;
 
639
640		/*
641		 * When soft offlining hugepages, after migrating the page
642		 * we dissolve it, therefore in the second loop "page" will
643		 * no longer be a compound page, and order will be 0.
644		 */
645		order = compound_order(compound_head(page));
646
647		if (PageHWPoison(page)) {
648			put_page(page);
649			continue;
650		}
651
652		if (behavior == MADV_SOFT_OFFLINE) {
653			pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
654						page_to_pfn(page), start);
655
656			ret = soft_offline_page(page, MF_COUNT_INCREASED);
657			if (ret)
658				return ret;
659			continue;
 
 
660		}
661		pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
662						page_to_pfn(page), start);
663
664		ret = memory_failure(page_to_pfn(page), MF_COUNT_INCREASED);
665		if (ret)
666			return ret;
667	}
668
669	/* Ensure that all poisoned pages are removed from per-cpu lists */
670	for_each_populated_zone(zone)
671		drain_all_pages(zone);
672
673	return 0;
674}
675#endif
676
677static long
678madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
679		unsigned long start, unsigned long end, int behavior)
680{
681	switch (behavior) {
682	case MADV_REMOVE:
683		return madvise_remove(vma, prev, start, end);
684	case MADV_WILLNEED:
685		return madvise_willneed(vma, prev, start, end);
686	case MADV_FREE:
687	case MADV_DONTNEED:
688		return madvise_dontneed_free(vma, prev, start, end, behavior);
689	default:
690		return madvise_behavior(vma, prev, start, end, behavior);
691	}
692}
693
694static bool
695madvise_behavior_valid(int behavior)
696{
697	switch (behavior) {
698	case MADV_DOFORK:
699	case MADV_DONTFORK:
700	case MADV_NORMAL:
701	case MADV_SEQUENTIAL:
702	case MADV_RANDOM:
703	case MADV_REMOVE:
704	case MADV_WILLNEED:
705	case MADV_DONTNEED:
 
706	case MADV_FREE:
 
 
 
 
707#ifdef CONFIG_KSM
708	case MADV_MERGEABLE:
709	case MADV_UNMERGEABLE:
710#endif
711#ifdef CONFIG_TRANSPARENT_HUGEPAGE
712	case MADV_HUGEPAGE:
713	case MADV_NOHUGEPAGE:
 
714#endif
715	case MADV_DONTDUMP:
716	case MADV_DODUMP:
717	case MADV_WIPEONFORK:
718	case MADV_KEEPONFORK:
 
 
719#ifdef CONFIG_MEMORY_FAILURE
720	case MADV_SOFT_OFFLINE:
721	case MADV_HWPOISON:
722#endif
723		return true;
724
725	default:
726		return false;
727	}
728}
729
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
730/*
731 * The madvise(2) system call.
732 *
733 * Applications can use madvise() to advise the kernel how it should
734 * handle paging I/O in this VM area.  The idea is to help the kernel
735 * use appropriate read-ahead and caching techniques.  The information
736 * provided is advisory only, and can be safely disregarded by the
737 * kernel without affecting the correct operation of the application.
738 *
739 * behavior values:
740 *  MADV_NORMAL - the default behavior is to read clusters.  This
741 *		results in some read-ahead and read-behind.
742 *  MADV_RANDOM - the system should read the minimum amount of data
743 *		on any access, since it is unlikely that the appli-
744 *		cation will need more than what it asks for.
745 *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
746 *		once, so they can be aggressively read ahead, and
747 *		can be freed soon after they are accessed.
748 *  MADV_WILLNEED - the application is notifying the system to read
749 *		some pages ahead.
750 *  MADV_DONTNEED - the application is finished with the given range,
751 *		so the kernel can free resources associated with it.
752 *  MADV_FREE - the application marks pages in the given range as lazy free,
753 *		where actual purges are postponed until memory pressure happens.
754 *  MADV_REMOVE - the application wants to free up the given range of
755 *		pages and associated backing store.
756 *  MADV_DONTFORK - omit this area from child's address space when forking:
757 *		typically, to avoid COWing pages pinned by get_user_pages().
758 *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
759 *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
760 *              range after a fork.
761 *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
762 *  MADV_HWPOISON - trigger memory error handler as if the given memory range
763 *		were corrupted by unrecoverable hardware memory failure.
764 *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
765 *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
766 *		this area with pages of identical content from other such areas.
767 *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
768 *  MADV_HUGEPAGE - the application wants to back the given range by transparent
769 *		huge pages in the future. Existing pages might be coalesced and
770 *		new pages might be allocated as THP.
771 *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
772 *		transparent huge pages so the existing pages will not be
773 *		coalesced into THP and new pages will not be allocated as THP.
 
774 *  MADV_DONTDUMP - the application wants to prevent pages in the given range
775 *		from being included in its core dump.
776 *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
 
 
 
 
 
 
 
 
 
777 *
778 * return values:
779 *  zero    - success
780 *  -EINVAL - start + len < 0, start is not page-aligned,
781 *		"behavior" is not a valid value, or application
782 *		is attempting to release locked or shared pages,
783 *		or the specified address range includes file, Huge TLB,
784 *		MAP_SHARED or VMPFNMAP range.
785 *  -ENOMEM - addresses in the specified range are not currently
786 *		mapped, or are outside the AS of the process.
787 *  -EIO    - an I/O error occurred while paging in data.
788 *  -EBADF  - map exists, but area maps something that isn't a file.
789 *  -EAGAIN - a kernel resource was temporarily unavailable.
 
790 */
791SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
792{
793	unsigned long end, tmp;
794	struct vm_area_struct *vma, *prev;
795	int unmapped_error = 0;
796	int error = -EINVAL;
797	int write;
798	size_t len;
799	struct blk_plug plug;
800
801	if (!madvise_behavior_valid(behavior))
802		return error;
803
804	if (start & ~PAGE_MASK)
805		return error;
806	len = (len_in + ~PAGE_MASK) & PAGE_MASK;
807
808	/* Check to see whether len was rounded up from small -ve to zero */
809	if (len_in && !len)
810		return error;
811
812	end = start + len;
813	if (end < start)
814		return error;
815
816	error = 0;
817	if (end == start)
818		return error;
819
820#ifdef CONFIG_MEMORY_FAILURE
821	if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
822		return madvise_inject_error(behavior, start, start + len_in);
823#endif
824
825	write = madvise_need_mmap_write(behavior);
826	if (write) {
827		if (down_write_killable(&current->mm->mmap_sem))
828			return -EINTR;
829	} else {
830		down_read(&current->mm->mmap_sem);
831	}
832
833	/*
834	 * If the interval [start,end) covers some unmapped address
835	 * ranges, just ignore them, but return -ENOMEM at the end.
836	 * - different from the way of handling in mlock etc.
837	 */
838	vma = find_vma_prev(current->mm, start, &prev);
839	if (vma && start > vma->vm_start)
840		prev = vma;
841
842	blk_start_plug(&plug);
843	for (;;) {
844		/* Still start < end. */
845		error = -ENOMEM;
846		if (!vma)
847			goto out;
848
849		/* Here start < (end|vma->vm_end). */
850		if (start < vma->vm_start) {
851			unmapped_error = -ENOMEM;
852			start = vma->vm_start;
853			if (start >= end)
854				goto out;
855		}
856
857		/* Here vma->vm_start <= start < (end|vma->vm_end) */
858		tmp = vma->vm_end;
859		if (end < tmp)
860			tmp = end;
861
862		/* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
863		error = madvise_vma(vma, &prev, start, tmp, behavior);
864		if (error)
865			goto out;
866		start = tmp;
867		if (prev && start < prev->vm_end)
868			start = prev->vm_end;
869		error = unmapped_error;
870		if (start >= end)
871			goto out;
872		if (prev)
873			vma = prev->vm_next;
874		else	/* madvise_remove dropped mmap_sem */
875			vma = find_vma(current->mm, start);
876	}
877out:
878	blk_finish_plug(&plug);
 
879	if (write)
880		up_write(&current->mm->mmap_sem);
881	else
882		up_read(&current->mm->mmap_sem);
883
884	return error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
885}