projects/linux·blob v3.13 · refs

4290 lines · 115.7 KiB

1/*
2 * linux/mm/memory.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
10 */
11
12/*
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
15 *
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
18 * far as I could see.
19 *
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21 */
22
23/*
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
29 */
30
31/*
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh ([email protected])
34 *
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * ([email protected])
37 *
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39 */
40
41#include <linux/kernel_stat.h>
42#include <linux/mm.h>
43#include <linux/hugetlb.h>
44#include <linux/mman.h>
45#include <linux/swap.h>
46#include <linux/highmem.h>
47#include <linux/pagemap.h>
48#include <linux/ksm.h>
49#include <linux/rmap.h>
50#include <linux/export.h>
51#include <linux/delayacct.h>
52#include <linux/init.h>
53#include <linux/writeback.h>
54#include <linux/memcontrol.h>
55#include <linux/mmu_notifier.h>
56#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
59#include <linux/gfp.h>
60#include <linux/migrate.h>
61#include <linux/string.h>
62
63#include <asm/io.h>
64#include <asm/pgalloc.h>
65#include <asm/uaccess.h>
66#include <asm/tlb.h>
67#include <asm/tlbflush.h>
68#include <asm/pgtable.h>
69
70#include "internal.h"
71
72#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
73#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
74#endif
75
76#ifndef CONFIG_NEED_MULTIPLE_NODES
77/* use the per-pgdat data instead for discontigmem - mbligh */
78unsigned long max_mapnr;
79struct page *mem_map;
80
81EXPORT_SYMBOL(max_mapnr);
82EXPORT_SYMBOL(mem_map);
83#endif
84
85/*
86 * A number of key systems in x86 including ioremap() rely on the assumption
87 * that high_memory defines the upper bound on direct map memory, then end
88 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
89 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
90 * and ZONE_HIGHMEM.
91 */
92void * high_memory;
93
94EXPORT_SYMBOL(high_memory);
95
96/*
97 * Randomize the address space (stacks, mmaps, brk, etc.).
98 *
99 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
100 * as ancient (libc5 based) binaries can segfault. )
101 */
102int randomize_va_space __read_mostly =
103#ifdef CONFIG_COMPAT_BRK
104 1;
105#else
106 2;
107#endif
108
109static int __init disable_randmaps(char *s)
110{
111 randomize_va_space = 0;
112 return 1;
113}
114__setup("norandmaps", disable_randmaps);
115
116unsigned long zero_pfn __read_mostly;
117unsigned long highest_memmap_pfn __read_mostly;
118
119/*
120 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
121 */
122static int __init init_zero_pfn(void)
123{
124 zero_pfn = page_to_pfn(ZERO_PAGE(0));
125 return 0;
126}
127core_initcall(init_zero_pfn);
128
129
130#if defined(SPLIT_RSS_COUNTING)
131
132void sync_mm_rss(struct mm_struct *mm)
133{
134 int i;
135
136 for (i = 0; i < NR_MM_COUNTERS; i++) {
137 if (current->rss_stat.count[i]) {
138 add_mm_counter(mm, i, current->rss_stat.count[i]);
139 current->rss_stat.count[i] = 0;
140 }
141 }
142 current->rss_stat.events = 0;
143}
144
145static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
146{
147 struct task_struct *task = current;
148
149 if (likely(task->mm == mm))
150 task->rss_stat.count[member] += val;
151 else
152 add_mm_counter(mm, member, val);
153}
154#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
155#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
156
157/* sync counter once per 64 page faults */
158#define TASK_RSS_EVENTS_THRESH (64)
159static void check_sync_rss_stat(struct task_struct *task)
160{
161 if (unlikely(task != current))
162 return;
163 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
164 sync_mm_rss(task->mm);
165}
166#else /* SPLIT_RSS_COUNTING */
167
168#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
169#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
170
171static void check_sync_rss_stat(struct task_struct *task)
172{
173}
174
175#endif /* SPLIT_RSS_COUNTING */
176
177#ifdef HAVE_GENERIC_MMU_GATHER
178
179static int tlb_next_batch(struct mmu_gather *tlb)
180{
181 struct mmu_gather_batch *batch;
182
183 batch = tlb->active;
184 if (batch->next) {
185 tlb->active = batch->next;
186 return 1;
187 }
188
189 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
190 return 0;
191
192 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
193 if (!batch)
194 return 0;
195
196 tlb->batch_count++;
197 batch->next = NULL;
198 batch->nr = 0;
199 batch->max = MAX_GATHER_BATCH;
200
201 tlb->active->next = batch;
202 tlb->active = batch;
203
204 return 1;
205}
206
207/* tlb_gather_mmu
208 * Called to initialize an (on-stack) mmu_gather structure for page-table
209 * tear-down from @mm. The @fullmm argument is used when @mm is without
210 * users and we're going to destroy the full address space (exit/execve).
211 */
212void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
213{
214 tlb->mm = mm;
215
216 /* Is it from 0 to ~0? */
217 tlb->fullmm = !(start | (end+1));
218 tlb->need_flush_all = 0;
219 tlb->start = start;
220 tlb->end = end;
221 tlb->need_flush = 0;
222 tlb->local.next = NULL;
223 tlb->local.nr = 0;
224 tlb->local.max = ARRAY_SIZE(tlb->__pages);
225 tlb->active = &tlb->local;
226 tlb->batch_count = 0;
227
228#ifdef CONFIG_HAVE_RCU_TABLE_FREE
229 tlb->batch = NULL;
230#endif
231}
232
233void tlb_flush_mmu(struct mmu_gather *tlb)
234{
235 struct mmu_gather_batch *batch;
236
237 if (!tlb->need_flush)
238 return;
239 tlb->need_flush = 0;
240 tlb_flush(tlb);
241#ifdef CONFIG_HAVE_RCU_TABLE_FREE
242 tlb_table_flush(tlb);
243#endif
244
245 for (batch = &tlb->local; batch; batch = batch->next) {
246 free_pages_and_swap_cache(batch->pages, batch->nr);
247 batch->nr = 0;
248 }
249 tlb->active = &tlb->local;
250}
251
252/* tlb_finish_mmu
253 * Called at the end of the shootdown operation to free up any resources
254 * that were required.
255 */
256void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
257{
258 struct mmu_gather_batch *batch, *next;
259
260 tlb_flush_mmu(tlb);
261
262 /* keep the page table cache within bounds */
263 check_pgt_cache();
264
265 for (batch = tlb->local.next; batch; batch = next) {
266 next = batch->next;
267 free_pages((unsigned long)batch, 0);
268 }
269 tlb->local.next = NULL;
270}
271
272/* __tlb_remove_page
273 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
274 * handling the additional races in SMP caused by other CPUs caching valid
275 * mappings in their TLBs. Returns the number of free page slots left.
276 * When out of page slots we must call tlb_flush_mmu().
277 */
278int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
279{
280 struct mmu_gather_batch *batch;
281
282 VM_BUG_ON(!tlb->need_flush);
283
284 batch = tlb->active;
285 batch->pages[batch->nr++] = page;
286 if (batch->nr == batch->max) {
287 if (!tlb_next_batch(tlb))
288 return 0;
289 batch = tlb->active;
290 }
291 VM_BUG_ON(batch->nr > batch->max);
292
293 return batch->max - batch->nr;
294}
295
296#endif /* HAVE_GENERIC_MMU_GATHER */
297
298#ifdef CONFIG_HAVE_RCU_TABLE_FREE
299
300/*
301 * See the comment near struct mmu_table_batch.
302 */
303
304static void tlb_remove_table_smp_sync(void *arg)
305{
306 /* Simply deliver the interrupt */
307}
308
309static void tlb_remove_table_one(void *table)
310{
311 /*
312 * This isn't an RCU grace period and hence the page-tables cannot be
313 * assumed to be actually RCU-freed.
314 *
315 * It is however sufficient for software page-table walkers that rely on
316 * IRQ disabling. See the comment near struct mmu_table_batch.
317 */
318 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
319 __tlb_remove_table(table);
320}
321
322static void tlb_remove_table_rcu(struct rcu_head *head)
323{
324 struct mmu_table_batch *batch;
325 int i;
326
327 batch = container_of(head, struct mmu_table_batch, rcu);
328
329 for (i = 0; i < batch->nr; i++)
330 __tlb_remove_table(batch->tables[i]);
331
332 free_page((unsigned long)batch);
333}
334
335void tlb_table_flush(struct mmu_gather *tlb)
336{
337 struct mmu_table_batch **batch = &tlb->batch;
338
339 if (*batch) {
340 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
341 *batch = NULL;
342 }
343}
344
345void tlb_remove_table(struct mmu_gather *tlb, void *table)
346{
347 struct mmu_table_batch **batch = &tlb->batch;
348
349 tlb->need_flush = 1;
350
351 /*
352 * When there's less then two users of this mm there cannot be a
353 * concurrent page-table walk.
354 */
355 if (atomic_read(&tlb->mm->mm_users) < 2) {
356 __tlb_remove_table(table);
357 return;
358 }
359
360 if (*batch == NULL) {
361 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
362 if (*batch == NULL) {
363 tlb_remove_table_one(table);
364 return;
365 }
366 (*batch)->nr = 0;
367 }
368 (*batch)->tables[(*batch)->nr++] = table;
369 if ((*batch)->nr == MAX_TABLE_BATCH)
370 tlb_table_flush(tlb);
371}
372
373#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
374
375/*
376 * Note: this doesn't free the actual pages themselves. That
377 * has been handled earlier when unmapping all the memory regions.
378 */
379static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
380 unsigned long addr)
381{
382 pgtable_t token = pmd_pgtable(*pmd);
383 pmd_clear(pmd);
384 pte_free_tlb(tlb, token, addr);
385 atomic_long_dec(&tlb->mm->nr_ptes);
386}
387
388static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
389 unsigned long addr, unsigned long end,
390 unsigned long floor, unsigned long ceiling)
391{
392 pmd_t *pmd;
393 unsigned long next;
394 unsigned long start;
395
396 start = addr;
397 pmd = pmd_offset(pud, addr);
398 do {
399 next = pmd_addr_end(addr, end);
400 if (pmd_none_or_clear_bad(pmd))
401 continue;
402 free_pte_range(tlb, pmd, addr);
403 } while (pmd++, addr = next, addr != end);
404
405 start &= PUD_MASK;
406 if (start < floor)
407 return;
408 if (ceiling) {
409 ceiling &= PUD_MASK;
410 if (!ceiling)
411 return;
412 }
413 if (end - 1 > ceiling - 1)
414 return;
415
416 pmd = pmd_offset(pud, start);
417 pud_clear(pud);
418 pmd_free_tlb(tlb, pmd, start);
419}
420
421static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
422 unsigned long addr, unsigned long end,
423 unsigned long floor, unsigned long ceiling)
424{
425 pud_t *pud;
426 unsigned long next;
427 unsigned long start;
428
429 start = addr;
430 pud = pud_offset(pgd, addr);
431 do {
432 next = pud_addr_end(addr, end);
433 if (pud_none_or_clear_bad(pud))
434 continue;
435 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
436 } while (pud++, addr = next, addr != end);
437
438 start &= PGDIR_MASK;
439 if (start < floor)
440 return;
441 if (ceiling) {
442 ceiling &= PGDIR_MASK;
443 if (!ceiling)
444 return;
445 }
446 if (end - 1 > ceiling - 1)
447 return;
448
449 pud = pud_offset(pgd, start);
450 pgd_clear(pgd);
451 pud_free_tlb(tlb, pud, start);
452}
453
454/*
455 * This function frees user-level page tables of a process.
456 */
457void free_pgd_range(struct mmu_gather *tlb,
458 unsigned long addr, unsigned long end,
459 unsigned long floor, unsigned long ceiling)
460{
461 pgd_t *pgd;
462 unsigned long next;
463
464 /*
465 * The next few lines have given us lots of grief...
466 *
467 * Why are we testing PMD* at this top level? Because often
468 * there will be no work to do at all, and we'd prefer not to
469 * go all the way down to the bottom just to discover that.
470 *
471 * Why all these "- 1"s? Because 0 represents both the bottom
472 * of the address space and the top of it (using -1 for the
473 * top wouldn't help much: the masks would do the wrong thing).
474 * The rule is that addr 0 and floor 0 refer to the bottom of
475 * the address space, but end 0 and ceiling 0 refer to the top
476 * Comparisons need to use "end - 1" and "ceiling - 1" (though
477 * that end 0 case should be mythical).
478 *
479 * Wherever addr is brought up or ceiling brought down, we must
480 * be careful to reject "the opposite 0" before it confuses the
481 * subsequent tests. But what about where end is brought down
482 * by PMD_SIZE below? no, end can't go down to 0 there.
483 *
484 * Whereas we round start (addr) and ceiling down, by different
485 * masks at different levels, in order to test whether a table
486 * now has no other vmas using it, so can be freed, we don't
487 * bother to round floor or end up - the tests don't need that.
488 */
489
490 addr &= PMD_MASK;
491 if (addr < floor) {
492 addr += PMD_SIZE;
493 if (!addr)
494 return;
495 }
496 if (ceiling) {
497 ceiling &= PMD_MASK;
498 if (!ceiling)
499 return;
500 }
501 if (end - 1 > ceiling - 1)
502 end -= PMD_SIZE;
503 if (addr > end - 1)
504 return;
505
506 pgd = pgd_offset(tlb->mm, addr);
507 do {
508 next = pgd_addr_end(addr, end);
509 if (pgd_none_or_clear_bad(pgd))
510 continue;
511 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
512 } while (pgd++, addr = next, addr != end);
513}
514
515void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
516 unsigned long floor, unsigned long ceiling)
517{
518 while (vma) {
519 struct vm_area_struct *next = vma->vm_next;
520 unsigned long addr = vma->vm_start;
521
522 /*
523 * Hide vma from rmap and truncate_pagecache before freeing
524 * pgtables
525 */
526 unlink_anon_vmas(vma);
527 unlink_file_vma(vma);
528
529 if (is_vm_hugetlb_page(vma)) {
530 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
531 floor, next? next->vm_start: ceiling);
532 } else {
533 /*
534 * Optimization: gather nearby vmas into one call down
535 */
536 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
537 && !is_vm_hugetlb_page(next)) {
538 vma = next;
539 next = vma->vm_next;
540 unlink_anon_vmas(vma);
541 unlink_file_vma(vma);
542 }
543 free_pgd_range(tlb, addr, vma->vm_end,
544 floor, next? next->vm_start: ceiling);
545 }
546 vma = next;
547 }
548}
549
550int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
551 pmd_t *pmd, unsigned long address)
552{
553 spinlock_t *ptl;
554 pgtable_t new = pte_alloc_one(mm, address);
555 int wait_split_huge_page;
556 if (!new)
557 return -ENOMEM;
558
559 /*
560 * Ensure all pte setup (eg. pte page lock and page clearing) are
561 * visible before the pte is made visible to other CPUs by being
562 * put into page tables.
563 *
564 * The other side of the story is the pointer chasing in the page
565 * table walking code (when walking the page table without locking;
566 * ie. most of the time). Fortunately, these data accesses consist
567 * of a chain of data-dependent loads, meaning most CPUs (alpha
568 * being the notable exception) will already guarantee loads are
569 * seen in-order. See the alpha page table accessors for the
570 * smp_read_barrier_depends() barriers in page table walking code.
571 */
572 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
573
574 ptl = pmd_lock(mm, pmd);
575 wait_split_huge_page = 0;
576 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
577 atomic_long_inc(&mm->nr_ptes);
578 pmd_populate(mm, pmd, new);
579 new = NULL;
580 } else if (unlikely(pmd_trans_splitting(*pmd)))
581 wait_split_huge_page = 1;
582 spin_unlock(ptl);
583 if (new)
584 pte_free(mm, new);
585 if (wait_split_huge_page)
586 wait_split_huge_page(vma->anon_vma, pmd);
587 return 0;
588}
589
590int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
591{
592 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
593 if (!new)
594 return -ENOMEM;
595
596 smp_wmb(); /* See comment in __pte_alloc */
597
598 spin_lock(&init_mm.page_table_lock);
599 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
600 pmd_populate_kernel(&init_mm, pmd, new);
601 new = NULL;
602 } else
603 VM_BUG_ON(pmd_trans_splitting(*pmd));
604 spin_unlock(&init_mm.page_table_lock);
605 if (new)
606 pte_free_kernel(&init_mm, new);
607 return 0;
608}
609
610static inline void init_rss_vec(int *rss)
611{
612 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
613}
614
615static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
616{
617 int i;
618
619 if (current->mm == mm)
620 sync_mm_rss(mm);
621 for (i = 0; i < NR_MM_COUNTERS; i++)
622 if (rss[i])
623 add_mm_counter(mm, i, rss[i]);
624}
625
626/*
627 * This function is called to print an error when a bad pte
628 * is found. For example, we might have a PFN-mapped pte in
629 * a region that doesn't allow it.
630 *
631 * The calling function must still handle the error.
632 */
633static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
634 pte_t pte, struct page *page)
635{
636 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
637 pud_t *pud = pud_offset(pgd, addr);
638 pmd_t *pmd = pmd_offset(pud, addr);
639 struct address_space *mapping;
640 pgoff_t index;
641 static unsigned long resume;
642 static unsigned long nr_shown;
643 static unsigned long nr_unshown;
644
645 /*
646 * Allow a burst of 60 reports, then keep quiet for that minute;
647 * or allow a steady drip of one report per second.
648 */
649 if (nr_shown == 60) {
650 if (time_before(jiffies, resume)) {
651 nr_unshown++;
652 return;
653 }
654 if (nr_unshown) {
655 printk(KERN_ALERT
656 "BUG: Bad page map: %lu messages suppressed\n",
657 nr_unshown);
658 nr_unshown = 0;
659 }
660 nr_shown = 0;
661 }
662 if (nr_shown++ == 0)
663 resume = jiffies + 60 * HZ;
664
665 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
666 index = linear_page_index(vma, addr);
667
668 printk(KERN_ALERT
669 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
670 current->comm,
671 (long long)pte_val(pte), (long long)pmd_val(*pmd));
672 if (page)
673 dump_page(page);
674 printk(KERN_ALERT
675 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
676 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
677 /*
678 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
679 */
680 if (vma->vm_ops)
681 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
682 vma->vm_ops->fault);
683 if (vma->vm_file)
684 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
685 vma->vm_file->f_op->mmap);
686 dump_stack();
687 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
688}
689
690static inline bool is_cow_mapping(vm_flags_t flags)
691{
692 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
693}
694
695/*
696 * vm_normal_page -- This function gets the "struct page" associated with a pte.
697 *
698 * "Special" mappings do not wish to be associated with a "struct page" (either
699 * it doesn't exist, or it exists but they don't want to touch it). In this
700 * case, NULL is returned here. "Normal" mappings do have a struct page.
701 *
702 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
703 * pte bit, in which case this function is trivial. Secondly, an architecture
704 * may not have a spare pte bit, which requires a more complicated scheme,
705 * described below.
706 *
707 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
708 * special mapping (even if there are underlying and valid "struct pages").
709 * COWed pages of a VM_PFNMAP are always normal.
710 *
711 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
712 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
713 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
714 * mapping will always honor the rule
715 *
716 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
717 *
718 * And for normal mappings this is false.
719 *
720 * This restricts such mappings to be a linear translation from virtual address
721 * to pfn. To get around this restriction, we allow arbitrary mappings so long
722 * as the vma is not a COW mapping; in that case, we know that all ptes are
723 * special (because none can have been COWed).
724 *
725 *
726 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
727 *
728 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
729 * page" backing, however the difference is that _all_ pages with a struct
730 * page (that is, those where pfn_valid is true) are refcounted and considered
731 * normal pages by the VM. The disadvantage is that pages are refcounted
732 * (which can be slower and simply not an option for some PFNMAP users). The
733 * advantage is that we don't have to follow the strict linearity rule of
734 * PFNMAP mappings in order to support COWable mappings.
735 *
736 */
737#ifdef __HAVE_ARCH_PTE_SPECIAL
738# define HAVE_PTE_SPECIAL 1
739#else
740# define HAVE_PTE_SPECIAL 0
741#endif
742struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
743 pte_t pte)
744{
745 unsigned long pfn = pte_pfn(pte);
746
747 if (HAVE_PTE_SPECIAL) {
748 if (likely(!pte_special(pte)))
749 goto check_pfn;
750 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
751 return NULL;
752 if (!is_zero_pfn(pfn))
753 print_bad_pte(vma, addr, pte, NULL);
754 return NULL;
755 }
756
757 /* !HAVE_PTE_SPECIAL case follows: */
758
759 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
760 if (vma->vm_flags & VM_MIXEDMAP) {
761 if (!pfn_valid(pfn))
762 return NULL;
763 goto out;
764 } else {
765 unsigned long off;
766 off = (addr - vma->vm_start) >> PAGE_SHIFT;
767 if (pfn == vma->vm_pgoff + off)
768 return NULL;
769 if (!is_cow_mapping(vma->vm_flags))
770 return NULL;
771 }
772 }
773
774 if (is_zero_pfn(pfn))
775 return NULL;
776check_pfn:
777 if (unlikely(pfn > highest_memmap_pfn)) {
778 print_bad_pte(vma, addr, pte, NULL);
779 return NULL;
780 }
781
782 /*
783 * NOTE! We still have PageReserved() pages in the page tables.
784 * eg. VDSO mappings can cause them to exist.
785 */
786out:
787 return pfn_to_page(pfn);
788}
789
790/*
791 * copy one vm_area from one task to the other. Assumes the page tables
792 * already present in the new task to be cleared in the whole range
793 * covered by this vma.
794 */
795
796static inline unsigned long
797copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
798 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
799 unsigned long addr, int *rss)
800{
801 unsigned long vm_flags = vma->vm_flags;
802 pte_t pte = *src_pte;
803 struct page *page;
804
805 /* pte contains position in swap or file, so copy. */
806 if (unlikely(!pte_present(pte))) {
807 if (!pte_file(pte)) {
808 swp_entry_t entry = pte_to_swp_entry(pte);
809
810 if (swap_duplicate(entry) < 0)
811 return entry.val;
812
813 /* make sure dst_mm is on swapoff's mmlist. */
814 if (unlikely(list_empty(&dst_mm->mmlist))) {
815 spin_lock(&mmlist_lock);
816 if (list_empty(&dst_mm->mmlist))
817 list_add(&dst_mm->mmlist,
818 &src_mm->mmlist);
819 spin_unlock(&mmlist_lock);
820 }
821 if (likely(!non_swap_entry(entry)))
822 rss[MM_SWAPENTS]++;
823 else if (is_migration_entry(entry)) {
824 page = migration_entry_to_page(entry);
825
826 if (PageAnon(page))
827 rss[MM_ANONPAGES]++;
828 else
829 rss[MM_FILEPAGES]++;
830
831 if (is_write_migration_entry(entry) &&
832 is_cow_mapping(vm_flags)) {
833 /*
834 * COW mappings require pages in both
835 * parent and child to be set to read.
836 */
837 make_migration_entry_read(&entry);
838 pte = swp_entry_to_pte(entry);
839 if (pte_swp_soft_dirty(*src_pte))
840 pte = pte_swp_mksoft_dirty(pte);
841 set_pte_at(src_mm, addr, src_pte, pte);
842 }
843 }
844 }
845 goto out_set_pte;
846 }
847
848 /*
849 * If it's a COW mapping, write protect it both
850 * in the parent and the child
851 */
852 if (is_cow_mapping(vm_flags)) {
853 ptep_set_wrprotect(src_mm, addr, src_pte);
854 pte = pte_wrprotect(pte);
855 }
856
857 /*
858 * If it's a shared mapping, mark it clean in
859 * the child
860 */
861 if (vm_flags & VM_SHARED)
862 pte = pte_mkclean(pte);
863 pte = pte_mkold(pte);
864
865 page = vm_normal_page(vma, addr, pte);
866 if (page) {
867 get_page(page);
868 page_dup_rmap(page);
869 if (PageAnon(page))
870 rss[MM_ANONPAGES]++;
871 else
872 rss[MM_FILEPAGES]++;
873 }
874
875out_set_pte:
876 set_pte_at(dst_mm, addr, dst_pte, pte);
877 return 0;
878}
879
880int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
881 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
882 unsigned long addr, unsigned long end)
883{
884 pte_t *orig_src_pte, *orig_dst_pte;
885 pte_t *src_pte, *dst_pte;
886 spinlock_t *src_ptl, *dst_ptl;
887 int progress = 0;
888 int rss[NR_MM_COUNTERS];
889 swp_entry_t entry = (swp_entry_t){0};
890
891again:
892 init_rss_vec(rss);
893
894 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
895 if (!dst_pte)
896 return -ENOMEM;
897 src_pte = pte_offset_map(src_pmd, addr);
898 src_ptl = pte_lockptr(src_mm, src_pmd);
899 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
900 orig_src_pte = src_pte;
901 orig_dst_pte = dst_pte;
902 arch_enter_lazy_mmu_mode();
903
904 do {
905 /*
906 * We are holding two locks at this point - either of them
907 * could generate latencies in another task on another CPU.
908 */
909 if (progress >= 32) {
910 progress = 0;
911 if (need_resched() ||
912 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
913 break;
914 }
915 if (pte_none(*src_pte)) {
916 progress++;
917 continue;
918 }
919 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
920 vma, addr, rss);
921 if (entry.val)
922 break;
923 progress += 8;
924 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
925
926 arch_leave_lazy_mmu_mode();
927 spin_unlock(src_ptl);
928 pte_unmap(orig_src_pte);
929 add_mm_rss_vec(dst_mm, rss);
930 pte_unmap_unlock(orig_dst_pte, dst_ptl);
931 cond_resched();
932
933 if (entry.val) {
934 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
935 return -ENOMEM;
936 progress = 0;
937 }
938 if (addr != end)
939 goto again;
940 return 0;
941}
942
943static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
944 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
945 unsigned long addr, unsigned long end)
946{
947 pmd_t *src_pmd, *dst_pmd;
948 unsigned long next;
949
950 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
951 if (!dst_pmd)
952 return -ENOMEM;
953 src_pmd = pmd_offset(src_pud, addr);
954 do {
955 next = pmd_addr_end(addr, end);
956 if (pmd_trans_huge(*src_pmd)) {
957 int err;
958 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
959 err = copy_huge_pmd(dst_mm, src_mm,
960 dst_pmd, src_pmd, addr, vma);
961 if (err == -ENOMEM)
962 return -ENOMEM;
963 if (!err)
964 continue;
965 /* fall through */
966 }
967 if (pmd_none_or_clear_bad(src_pmd))
968 continue;
969 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
970 vma, addr, next))
971 return -ENOMEM;
972 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
973 return 0;
974}
975
976static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
977 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
978 unsigned long addr, unsigned long end)
979{
980 pud_t *src_pud, *dst_pud;
981 unsigned long next;
982
983 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
984 if (!dst_pud)
985 return -ENOMEM;
986 src_pud = pud_offset(src_pgd, addr);
987 do {
988 next = pud_addr_end(addr, end);
989 if (pud_none_or_clear_bad(src_pud))
990 continue;
991 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
992 vma, addr, next))
993 return -ENOMEM;
994 } while (dst_pud++, src_pud++, addr = next, addr != end);
995 return 0;
996}
997
998int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
999 struct vm_area_struct *vma)
1000{
1001 pgd_t *src_pgd, *dst_pgd;
1002 unsigned long next;
1003 unsigned long addr = vma->vm_start;
1004 unsigned long end = vma->vm_end;
1005 unsigned long mmun_start; /* For mmu_notifiers */
1006 unsigned long mmun_end; /* For mmu_notifiers */
1007 bool is_cow;
1008 int ret;
1009
1010 /*
1011 * Don't copy ptes where a page fault will fill them correctly.
1012 * Fork becomes much lighter when there are big shared or private
1013 * readonly mappings. The tradeoff is that copy_page_range is more
1014 * efficient than faulting.
1015 */
1016 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1017 VM_PFNMAP | VM_MIXEDMAP))) {
1018 if (!vma->anon_vma)
1019 return 0;
1020 }
1021
1022 if (is_vm_hugetlb_page(vma))
1023 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1024
1025 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1026 /*
1027 * We do not free on error cases below as remove_vma
1028 * gets called on error from higher level routine
1029 */
1030 ret = track_pfn_copy(vma);
1031 if (ret)
1032 return ret;
1033 }
1034
1035 /*
1036 * We need to invalidate the secondary MMU mappings only when
1037 * there could be a permission downgrade on the ptes of the
1038 * parent mm. And a permission downgrade will only happen if
1039 * is_cow_mapping() returns true.
1040 */
1041 is_cow = is_cow_mapping(vma->vm_flags);
1042 mmun_start = addr;
1043 mmun_end = end;
1044 if (is_cow)
1045 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1046 mmun_end);
1047
1048 ret = 0;
1049 dst_pgd = pgd_offset(dst_mm, addr);
1050 src_pgd = pgd_offset(src_mm, addr);
1051 do {
1052 next = pgd_addr_end(addr, end);
1053 if (pgd_none_or_clear_bad(src_pgd))
1054 continue;
1055 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1056 vma, addr, next))) {
1057 ret = -ENOMEM;
1058 break;
1059 }
1060 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1061
1062 if (is_cow)
1063 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1064 return ret;
1065}
1066
1067static unsigned long zap_pte_range(struct mmu_gather *tlb,
1068 struct vm_area_struct *vma, pmd_t *pmd,
1069 unsigned long addr, unsigned long end,
1070 struct zap_details *details)
1071{
1072 struct mm_struct *mm = tlb->mm;
1073 int force_flush = 0;
1074 int rss[NR_MM_COUNTERS];
1075 spinlock_t *ptl;
1076 pte_t *start_pte;
1077 pte_t *pte;
1078
1079again:
1080 init_rss_vec(rss);
1081 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1082 pte = start_pte;
1083 arch_enter_lazy_mmu_mode();
1084 do {
1085 pte_t ptent = *pte;
1086 if (pte_none(ptent)) {
1087 continue;
1088 }
1089
1090 if (pte_present(ptent)) {
1091 struct page *page;
1092
1093 page = vm_normal_page(vma, addr, ptent);
1094 if (unlikely(details) && page) {
1095 /*
1096 * unmap_shared_mapping_pages() wants to
1097 * invalidate cache without truncating:
1098 * unmap shared but keep private pages.
1099 */
1100 if (details->check_mapping &&
1101 details->check_mapping != page->mapping)
1102 continue;
1103 /*
1104 * Each page->index must be checked when
1105 * invalidating or truncating nonlinear.
1106 */
1107 if (details->nonlinear_vma &&
1108 (page->index < details->first_index ||
1109 page->index > details->last_index))
1110 continue;
1111 }
1112 ptent = ptep_get_and_clear_full(mm, addr, pte,
1113 tlb->fullmm);
1114 tlb_remove_tlb_entry(tlb, pte, addr);
1115 if (unlikely(!page))
1116 continue;
1117 if (unlikely(details) && details->nonlinear_vma
1118 && linear_page_index(details->nonlinear_vma,
1119 addr) != page->index) {
1120 pte_t ptfile = pgoff_to_pte(page->index);
1121 if (pte_soft_dirty(ptent))
1122 pte_file_mksoft_dirty(ptfile);
1123 set_pte_at(mm, addr, pte, ptfile);
1124 }
1125 if (PageAnon(page))
1126 rss[MM_ANONPAGES]--;
1127 else {
1128 if (pte_dirty(ptent))
1129 set_page_dirty(page);
1130 if (pte_young(ptent) &&
1131 likely(!(vma->vm_flags & VM_SEQ_READ)))
1132 mark_page_accessed(page);
1133 rss[MM_FILEPAGES]--;
1134 }
1135 page_remove_rmap(page);
1136 if (unlikely(page_mapcount(page) < 0))
1137 print_bad_pte(vma, addr, ptent, page);
1138 force_flush = !__tlb_remove_page(tlb, page);
1139 if (force_flush)
1140 break;
1141 continue;
1142 }
1143 /*
1144 * If details->check_mapping, we leave swap entries;
1145 * if details->nonlinear_vma, we leave file entries.
1146 */
1147 if (unlikely(details))
1148 continue;
1149 if (pte_file(ptent)) {
1150 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1151 print_bad_pte(vma, addr, ptent, NULL);
1152 } else {
1153 swp_entry_t entry = pte_to_swp_entry(ptent);
1154
1155 if (!non_swap_entry(entry))
1156 rss[MM_SWAPENTS]--;
1157 else if (is_migration_entry(entry)) {
1158 struct page *page;
1159
1160 page = migration_entry_to_page(entry);
1161
1162 if (PageAnon(page))
1163 rss[MM_ANONPAGES]--;
1164 else
1165 rss[MM_FILEPAGES]--;
1166 }
1167 if (unlikely(!free_swap_and_cache(entry)))
1168 print_bad_pte(vma, addr, ptent, NULL);
1169 }
1170 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1171 } while (pte++, addr += PAGE_SIZE, addr != end);
1172
1173 add_mm_rss_vec(mm, rss);
1174 arch_leave_lazy_mmu_mode();
1175 pte_unmap_unlock(start_pte, ptl);
1176
1177 /*
1178 * mmu_gather ran out of room to batch pages, we break out of
1179 * the PTE lock to avoid doing the potential expensive TLB invalidate
1180 * and page-free while holding it.
1181 */
1182 if (force_flush) {
1183 unsigned long old_end;
1184
1185 force_flush = 0;
1186
1187 /*
1188 * Flush the TLB just for the previous segment,
1189 * then update the range to be the remaining
1190 * TLB range.
1191 */
1192 old_end = tlb->end;
1193 tlb->end = addr;
1194
1195 tlb_flush_mmu(tlb);
1196
1197 tlb->start = addr;
1198 tlb->end = old_end;
1199
1200 if (addr != end)
1201 goto again;
1202 }
1203
1204 return addr;
1205}
1206
1207static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1208 struct vm_area_struct *vma, pud_t *pud,
1209 unsigned long addr, unsigned long end,
1210 struct zap_details *details)
1211{
1212 pmd_t *pmd;
1213 unsigned long next;
1214
1215 pmd = pmd_offset(pud, addr);
1216 do {
1217 next = pmd_addr_end(addr, end);
1218 if (pmd_trans_huge(*pmd)) {
1219 if (next - addr != HPAGE_PMD_SIZE) {
1220#ifdef CONFIG_DEBUG_VM
1221 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1222 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1223 __func__, addr, end,
1224 vma->vm_start,
1225 vma->vm_end);
1226 BUG();
1227 }
1228#endif
1229 split_huge_page_pmd(vma, addr, pmd);
1230 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1231 goto next;
1232 /* fall through */
1233 }
1234 /*
1235 * Here there can be other concurrent MADV_DONTNEED or
1236 * trans huge page faults running, and if the pmd is
1237 * none or trans huge it can change under us. This is
1238 * because MADV_DONTNEED holds the mmap_sem in read
1239 * mode.
1240 */
1241 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1242 goto next;
1243 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1244next:
1245 cond_resched();
1246 } while (pmd++, addr = next, addr != end);
1247
1248 return addr;
1249}
1250
1251static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1252 struct vm_area_struct *vma, pgd_t *pgd,
1253 unsigned long addr, unsigned long end,
1254 struct zap_details *details)
1255{
1256 pud_t *pud;
1257 unsigned long next;
1258
1259 pud = pud_offset(pgd, addr);
1260 do {
1261 next = pud_addr_end(addr, end);
1262 if (pud_none_or_clear_bad(pud))
1263 continue;
1264 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1265 } while (pud++, addr = next, addr != end);
1266
1267 return addr;
1268}
1269
1270static void unmap_page_range(struct mmu_gather *tlb,
1271 struct vm_area_struct *vma,
1272 unsigned long addr, unsigned long end,
1273 struct zap_details *details)
1274{
1275 pgd_t *pgd;
1276 unsigned long next;
1277
1278 if (details && !details->check_mapping && !details->nonlinear_vma)
1279 details = NULL;
1280
1281 BUG_ON(addr >= end);
1282 mem_cgroup_uncharge_start();
1283 tlb_start_vma(tlb, vma);
1284 pgd = pgd_offset(vma->vm_mm, addr);
1285 do {
1286 next = pgd_addr_end(addr, end);
1287 if (pgd_none_or_clear_bad(pgd))
1288 continue;
1289 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1290 } while (pgd++, addr = next, addr != end);
1291 tlb_end_vma(tlb, vma);
1292 mem_cgroup_uncharge_end();
1293}
1294
1295
1296static void unmap_single_vma(struct mmu_gather *tlb,
1297 struct vm_area_struct *vma, unsigned long start_addr,
1298 unsigned long end_addr,
1299 struct zap_details *details)
1300{
1301 unsigned long start = max(vma->vm_start, start_addr);
1302 unsigned long end;
1303
1304 if (start >= vma->vm_end)
1305 return;
1306 end = min(vma->vm_end, end_addr);
1307 if (end <= vma->vm_start)
1308 return;
1309
1310 if (vma->vm_file)
1311 uprobe_munmap(vma, start, end);
1312
1313 if (unlikely(vma->vm_flags & VM_PFNMAP))
1314 untrack_pfn(vma, 0, 0);
1315
1316 if (start != end) {
1317 if (unlikely(is_vm_hugetlb_page(vma))) {
1318 /*
1319 * It is undesirable to test vma->vm_file as it
1320 * should be non-null for valid hugetlb area.
1321 * However, vm_file will be NULL in the error
1322 * cleanup path of do_mmap_pgoff. When
1323 * hugetlbfs ->mmap method fails,
1324 * do_mmap_pgoff() nullifies vma->vm_file
1325 * before calling this function to clean up.
1326 * Since no pte has actually been setup, it is
1327 * safe to do nothing in this case.
1328 */
1329 if (vma->vm_file) {
1330 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1331 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1332 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1333 }
1334 } else
1335 unmap_page_range(tlb, vma, start, end, details);
1336 }
1337}
1338
1339/**
1340 * unmap_vmas - unmap a range of memory covered by a list of vma's
1341 * @tlb: address of the caller's struct mmu_gather
1342 * @vma: the starting vma
1343 * @start_addr: virtual address at which to start unmapping
1344 * @end_addr: virtual address at which to end unmapping
1345 *
1346 * Unmap all pages in the vma list.
1347 *
1348 * Only addresses between `start' and `end' will be unmapped.
1349 *
1350 * The VMA list must be sorted in ascending virtual address order.
1351 *
1352 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1353 * range after unmap_vmas() returns. So the only responsibility here is to
1354 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1355 * drops the lock and schedules.
1356 */
1357void unmap_vmas(struct mmu_gather *tlb,
1358 struct vm_area_struct *vma, unsigned long start_addr,
1359 unsigned long end_addr)
1360{
1361 struct mm_struct *mm = vma->vm_mm;
1362
1363 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1364 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1365 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1366 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1367}
1368
1369/**
1370 * zap_page_range - remove user pages in a given range
1371 * @vma: vm_area_struct holding the applicable pages
1372 * @start: starting address of pages to zap
1373 * @size: number of bytes to zap
1374 * @details: details of nonlinear truncation or shared cache invalidation
1375 *
1376 * Caller must protect the VMA list
1377 */
1378void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1379 unsigned long size, struct zap_details *details)
1380{
1381 struct mm_struct *mm = vma->vm_mm;
1382 struct mmu_gather tlb;
1383 unsigned long end = start + size;
1384
1385 lru_add_drain();
1386 tlb_gather_mmu(&tlb, mm, start, end);
1387 update_hiwater_rss(mm);
1388 mmu_notifier_invalidate_range_start(mm, start, end);
1389 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1390 unmap_single_vma(&tlb, vma, start, end, details);
1391 mmu_notifier_invalidate_range_end(mm, start, end);
1392 tlb_finish_mmu(&tlb, start, end);
1393}
1394
1395/**
1396 * zap_page_range_single - remove user pages in a given range
1397 * @vma: vm_area_struct holding the applicable pages
1398 * @address: starting address of pages to zap
1399 * @size: number of bytes to zap
1400 * @details: details of nonlinear truncation or shared cache invalidation
1401 *
1402 * The range must fit into one VMA.
1403 */
1404static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1405 unsigned long size, struct zap_details *details)
1406{
1407 struct mm_struct *mm = vma->vm_mm;
1408 struct mmu_gather tlb;
1409 unsigned long end = address + size;
1410
1411 lru_add_drain();
1412 tlb_gather_mmu(&tlb, mm, address, end);
1413 update_hiwater_rss(mm);
1414 mmu_notifier_invalidate_range_start(mm, address, end);
1415 unmap_single_vma(&tlb, vma, address, end, details);
1416 mmu_notifier_invalidate_range_end(mm, address, end);
1417 tlb_finish_mmu(&tlb, address, end);
1418}
1419
1420/**
1421 * zap_vma_ptes - remove ptes mapping the vma
1422 * @vma: vm_area_struct holding ptes to be zapped
1423 * @address: starting address of pages to zap
1424 * @size: number of bytes to zap
1425 *
1426 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1427 *
1428 * The entire address range must be fully contained within the vma.
1429 *
1430 * Returns 0 if successful.
1431 */
1432int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1433 unsigned long size)
1434{
1435 if (address < vma->vm_start || address + size > vma->vm_end ||
1436 !(vma->vm_flags & VM_PFNMAP))
1437 return -1;
1438 zap_page_range_single(vma, address, size, NULL);
1439 return 0;
1440}
1441EXPORT_SYMBOL_GPL(zap_vma_ptes);
1442
1443/**
1444 * follow_page_mask - look up a page descriptor from a user-virtual address
1445 * @vma: vm_area_struct mapping @address
1446 * @address: virtual address to look up
1447 * @flags: flags modifying lookup behaviour
1448 * @page_mask: on output, *page_mask is set according to the size of the page
1449 *
1450 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1451 *
1452 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1453 * an error pointer if there is a mapping to something not represented
1454 * by a page descriptor (see also vm_normal_page()).
1455 */
1456struct page *follow_page_mask(struct vm_area_struct *vma,
1457 unsigned long address, unsigned int flags,
1458 unsigned int *page_mask)
1459{
1460 pgd_t *pgd;
1461 pud_t *pud;
1462 pmd_t *pmd;
1463 pte_t *ptep, pte;
1464 spinlock_t *ptl;
1465 struct page *page;
1466 struct mm_struct *mm = vma->vm_mm;
1467
1468 *page_mask = 0;
1469
1470 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1471 if (!IS_ERR(page)) {
1472 BUG_ON(flags & FOLL_GET);
1473 goto out;
1474 }
1475
1476 page = NULL;
1477 pgd = pgd_offset(mm, address);
1478 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1479 goto no_page_table;
1480
1481 pud = pud_offset(pgd, address);
1482 if (pud_none(*pud))
1483 goto no_page_table;
1484 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1485 if (flags & FOLL_GET)
1486 goto out;
1487 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1488 goto out;
1489 }
1490 if (unlikely(pud_bad(*pud)))
1491 goto no_page_table;
1492
1493 pmd = pmd_offset(pud, address);
1494 if (pmd_none(*pmd))
1495 goto no_page_table;
1496 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1497 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1498 if (flags & FOLL_GET) {
1499 /*
1500 * Refcount on tail pages are not well-defined and
1501 * shouldn't be taken. The caller should handle a NULL
1502 * return when trying to follow tail pages.
1503 */
1504 if (PageHead(page))
1505 get_page(page);
1506 else {
1507 page = NULL;
1508 goto out;
1509 }
1510 }
1511 goto out;
1512 }
1513 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1514 goto no_page_table;
1515 if (pmd_trans_huge(*pmd)) {
1516 if (flags & FOLL_SPLIT) {
1517 split_huge_page_pmd(vma, address, pmd);
1518 goto split_fallthrough;
1519 }
1520 ptl = pmd_lock(mm, pmd);
1521 if (likely(pmd_trans_huge(*pmd))) {
1522 if (unlikely(pmd_trans_splitting(*pmd))) {
1523 spin_unlock(ptl);
1524 wait_split_huge_page(vma->anon_vma, pmd);
1525 } else {
1526 page = follow_trans_huge_pmd(vma, address,
1527 pmd, flags);
1528 spin_unlock(ptl);
1529 *page_mask = HPAGE_PMD_NR - 1;
1530 goto out;
1531 }
1532 } else
1533 spin_unlock(ptl);
1534 /* fall through */
1535 }
1536split_fallthrough:
1537 if (unlikely(pmd_bad(*pmd)))
1538 goto no_page_table;
1539
1540 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1541
1542 pte = *ptep;
1543 if (!pte_present(pte)) {
1544 swp_entry_t entry;
1545 /*
1546 * KSM's break_ksm() relies upon recognizing a ksm page
1547 * even while it is being migrated, so for that case we
1548 * need migration_entry_wait().
1549 */
1550 if (likely(!(flags & FOLL_MIGRATION)))
1551 goto no_page;
1552 if (pte_none(pte) || pte_file(pte))
1553 goto no_page;
1554 entry = pte_to_swp_entry(pte);
1555 if (!is_migration_entry(entry))
1556 goto no_page;
1557 pte_unmap_unlock(ptep, ptl);
1558 migration_entry_wait(mm, pmd, address);
1559 goto split_fallthrough;
1560 }
1561 if ((flags & FOLL_NUMA) && pte_numa(pte))
1562 goto no_page;
1563 if ((flags & FOLL_WRITE) && !pte_write(pte))
1564 goto unlock;
1565
1566 page = vm_normal_page(vma, address, pte);
1567 if (unlikely(!page)) {
1568 if ((flags & FOLL_DUMP) ||
1569 !is_zero_pfn(pte_pfn(pte)))
1570 goto bad_page;
1571 page = pte_page(pte);
1572 }
1573
1574 if (flags & FOLL_GET)
1575 get_page_foll(page);
1576 if (flags & FOLL_TOUCH) {
1577 if ((flags & FOLL_WRITE) &&
1578 !pte_dirty(pte) && !PageDirty(page))
1579 set_page_dirty(page);
1580 /*
1581 * pte_mkyoung() would be more correct here, but atomic care
1582 * is needed to avoid losing the dirty bit: it is easier to use
1583 * mark_page_accessed().
1584 */
1585 mark_page_accessed(page);
1586 }
1587 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1588 /*
1589 * The preliminary mapping check is mainly to avoid the
1590 * pointless overhead of lock_page on the ZERO_PAGE
1591 * which might bounce very badly if there is contention.
1592 *
1593 * If the page is already locked, we don't need to
1594 * handle it now - vmscan will handle it later if and
1595 * when it attempts to reclaim the page.
1596 */
1597 if (page->mapping && trylock_page(page)) {
1598 lru_add_drain(); /* push cached pages to LRU */
1599 /*
1600 * Because we lock page here, and migration is
1601 * blocked by the pte's page reference, and we
1602 * know the page is still mapped, we don't even
1603 * need to check for file-cache page truncation.
1604 */
1605 mlock_vma_page(page);
1606 unlock_page(page);
1607 }
1608 }
1609unlock:
1610 pte_unmap_unlock(ptep, ptl);
1611out:
1612 return page;
1613
1614bad_page:
1615 pte_unmap_unlock(ptep, ptl);
1616 return ERR_PTR(-EFAULT);
1617
1618no_page:
1619 pte_unmap_unlock(ptep, ptl);
1620 if (!pte_none(pte))
1621 return page;
1622
1623no_page_table:
1624 /*
1625 * When core dumping an enormous anonymous area that nobody
1626 * has touched so far, we don't want to allocate unnecessary pages or
1627 * page tables. Return error instead of NULL to skip handle_mm_fault,
1628 * then get_dump_page() will return NULL to leave a hole in the dump.
1629 * But we can only make this optimization where a hole would surely
1630 * be zero-filled if handle_mm_fault() actually did handle it.
1631 */
1632 if ((flags & FOLL_DUMP) &&
1633 (!vma->vm_ops || !vma->vm_ops->fault))
1634 return ERR_PTR(-EFAULT);
1635 return page;
1636}
1637
1638static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1639{
1640 return stack_guard_page_start(vma, addr) ||
1641 stack_guard_page_end(vma, addr+PAGE_SIZE);
1642}
1643
1644/**
1645 * __get_user_pages() - pin user pages in memory
1646 * @tsk: task_struct of target task
1647 * @mm: mm_struct of target mm
1648 * @start: starting user address
1649 * @nr_pages: number of pages from start to pin
1650 * @gup_flags: flags modifying pin behaviour
1651 * @pages: array that receives pointers to the pages pinned.
1652 * Should be at least nr_pages long. Or NULL, if caller
1653 * only intends to ensure the pages are faulted in.
1654 * @vmas: array of pointers to vmas corresponding to each page.
1655 * Or NULL if the caller does not require them.
1656 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1657 *
1658 * Returns number of pages pinned. This may be fewer than the number
1659 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1660 * were pinned, returns -errno. Each page returned must be released
1661 * with a put_page() call when it is finished with. vmas will only
1662 * remain valid while mmap_sem is held.
1663 *
1664 * Must be called with mmap_sem held for read or write.
1665 *
1666 * __get_user_pages walks a process's page tables and takes a reference to
1667 * each struct page that each user address corresponds to at a given
1668 * instant. That is, it takes the page that would be accessed if a user
1669 * thread accesses the given user virtual address at that instant.
1670 *
1671 * This does not guarantee that the page exists in the user mappings when
1672 * __get_user_pages returns, and there may even be a completely different
1673 * page there in some cases (eg. if mmapped pagecache has been invalidated
1674 * and subsequently re faulted). However it does guarantee that the page
1675 * won't be freed completely. And mostly callers simply care that the page
1676 * contains data that was valid *at some point in time*. Typically, an IO
1677 * or similar operation cannot guarantee anything stronger anyway because
1678 * locks can't be held over the syscall boundary.
1679 *
1680 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1681 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1682 * appropriate) must be called after the page is finished with, and
1683 * before put_page is called.
1684 *
1685 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1686 * or mmap_sem contention, and if waiting is needed to pin all pages,
1687 * *@nonblocking will be set to 0.
1688 *
1689 * In most cases, get_user_pages or get_user_pages_fast should be used
1690 * instead of __get_user_pages. __get_user_pages should be used only if
1691 * you need some special @gup_flags.
1692 */
1693long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1694 unsigned long start, unsigned long nr_pages,
1695 unsigned int gup_flags, struct page **pages,
1696 struct vm_area_struct **vmas, int *nonblocking)
1697{
1698 long i;
1699 unsigned long vm_flags;
1700 unsigned int page_mask;
1701
1702 if (!nr_pages)
1703 return 0;
1704
1705 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1706
1707 /*
1708 * Require read or write permissions.
1709 * If FOLL_FORCE is set, we only require the "MAY" flags.
1710 */
1711 vm_flags = (gup_flags & FOLL_WRITE) ?
1712 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1713 vm_flags &= (gup_flags & FOLL_FORCE) ?
1714 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1715
1716 /*
1717 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1718 * would be called on PROT_NONE ranges. We must never invoke
1719 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1720 * page faults would unprotect the PROT_NONE ranges if
1721 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1722 * bitflag. So to avoid that, don't set FOLL_NUMA if
1723 * FOLL_FORCE is set.
1724 */
1725 if (!(gup_flags & FOLL_FORCE))
1726 gup_flags |= FOLL_NUMA;
1727
1728 i = 0;
1729
1730 do {
1731 struct vm_area_struct *vma;
1732
1733 vma = find_extend_vma(mm, start);
1734 if (!vma && in_gate_area(mm, start)) {
1735 unsigned long pg = start & PAGE_MASK;
1736 pgd_t *pgd;
1737 pud_t *pud;
1738 pmd_t *pmd;
1739 pte_t *pte;
1740
1741 /* user gate pages are read-only */
1742 if (gup_flags & FOLL_WRITE)
1743 return i ? : -EFAULT;
1744 if (pg > TASK_SIZE)
1745 pgd = pgd_offset_k(pg);
1746 else
1747 pgd = pgd_offset_gate(mm, pg);
1748 BUG_ON(pgd_none(*pgd));
1749 pud = pud_offset(pgd, pg);
1750 BUG_ON(pud_none(*pud));
1751 pmd = pmd_offset(pud, pg);
1752 if (pmd_none(*pmd))
1753 return i ? : -EFAULT;
1754 VM_BUG_ON(pmd_trans_huge(*pmd));
1755 pte = pte_offset_map(pmd, pg);
1756 if (pte_none(*pte)) {
1757 pte_unmap(pte);
1758 return i ? : -EFAULT;
1759 }
1760 vma = get_gate_vma(mm);
1761 if (pages) {
1762 struct page *page;
1763
1764 page = vm_normal_page(vma, start, *pte);
1765 if (!page) {
1766 if (!(gup_flags & FOLL_DUMP) &&
1767 is_zero_pfn(pte_pfn(*pte)))
1768 page = pte_page(*pte);
1769 else {
1770 pte_unmap(pte);
1771 return i ? : -EFAULT;
1772 }
1773 }
1774 pages[i] = page;
1775 get_page(page);
1776 }
1777 pte_unmap(pte);
1778 page_mask = 0;
1779 goto next_page;
1780 }
1781
1782 if (!vma ||
1783 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1784 !(vm_flags & vma->vm_flags))
1785 return i ? : -EFAULT;
1786
1787 if (is_vm_hugetlb_page(vma)) {
1788 i = follow_hugetlb_page(mm, vma, pages, vmas,
1789 &start, &nr_pages, i, gup_flags);
1790 continue;
1791 }
1792
1793 do {
1794 struct page *page;
1795 unsigned int foll_flags = gup_flags;
1796 unsigned int page_increm;
1797
1798 /*
1799 * If we have a pending SIGKILL, don't keep faulting
1800 * pages and potentially allocating memory.
1801 */
1802 if (unlikely(fatal_signal_pending(current)))
1803 return i ? i : -ERESTARTSYS;
1804
1805 cond_resched();
1806 while (!(page = follow_page_mask(vma, start,
1807 foll_flags, &page_mask))) {
1808 int ret;
1809 unsigned int fault_flags = 0;
1810
1811 /* For mlock, just skip the stack guard page. */
1812 if (foll_flags & FOLL_MLOCK) {
1813 if (stack_guard_page(vma, start))
1814 goto next_page;
1815 }
1816 if (foll_flags & FOLL_WRITE)
1817 fault_flags |= FAULT_FLAG_WRITE;
1818 if (nonblocking)
1819 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1820 if (foll_flags & FOLL_NOWAIT)
1821 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1822
1823 ret = handle_mm_fault(mm, vma, start,
1824 fault_flags);
1825
1826 if (ret & VM_FAULT_ERROR) {
1827 if (ret & VM_FAULT_OOM)
1828 return i ? i : -ENOMEM;
1829 if (ret & (VM_FAULT_HWPOISON |
1830 VM_FAULT_HWPOISON_LARGE)) {
1831 if (i)
1832 return i;
1833 else if (gup_flags & FOLL_HWPOISON)
1834 return -EHWPOISON;
1835 else
1836 return -EFAULT;
1837 }
1838 if (ret & VM_FAULT_SIGBUS)
1839 return i ? i : -EFAULT;
1840 BUG();
1841 }
1842
1843 if (tsk) {
1844 if (ret & VM_FAULT_MAJOR)
1845 tsk->maj_flt++;
1846 else
1847 tsk->min_flt++;
1848 }
1849
1850 if (ret & VM_FAULT_RETRY) {
1851 if (nonblocking)
1852 *nonblocking = 0;
1853 return i;
1854 }
1855
1856 /*
1857 * The VM_FAULT_WRITE bit tells us that
1858 * do_wp_page has broken COW when necessary,
1859 * even if maybe_mkwrite decided not to set
1860 * pte_write. We can thus safely do subsequent
1861 * page lookups as if they were reads. But only
1862 * do so when looping for pte_write is futile:
1863 * in some cases userspace may also be wanting
1864 * to write to the gotten user page, which a
1865 * read fault here might prevent (a readonly
1866 * page might get reCOWed by userspace write).
1867 */
1868 if ((ret & VM_FAULT_WRITE) &&
1869 !(vma->vm_flags & VM_WRITE))
1870 foll_flags &= ~FOLL_WRITE;
1871
1872 cond_resched();
1873 }
1874 if (IS_ERR(page))
1875 return i ? i : PTR_ERR(page);
1876 if (pages) {
1877 pages[i] = page;
1878
1879 flush_anon_page(vma, page, start);
1880 flush_dcache_page(page);
1881 page_mask = 0;
1882 }
1883next_page:
1884 if (vmas) {
1885 vmas[i] = vma;
1886 page_mask = 0;
1887 }
1888 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1889 if (page_increm > nr_pages)
1890 page_increm = nr_pages;
1891 i += page_increm;
1892 start += page_increm * PAGE_SIZE;
1893 nr_pages -= page_increm;
1894 } while (nr_pages && start < vma->vm_end);
1895 } while (nr_pages);
1896 return i;
1897}
1898EXPORT_SYMBOL(__get_user_pages);
1899
1900/*
1901 * fixup_user_fault() - manually resolve a user page fault
1902 * @tsk: the task_struct to use for page fault accounting, or
1903 * NULL if faults are not to be recorded.
1904 * @mm: mm_struct of target mm
1905 * @address: user address
1906 * @fault_flags:flags to pass down to handle_mm_fault()
1907 *
1908 * This is meant to be called in the specific scenario where for locking reasons
1909 * we try to access user memory in atomic context (within a pagefault_disable()
1910 * section), this returns -EFAULT, and we want to resolve the user fault before
1911 * trying again.
1912 *
1913 * Typically this is meant to be used by the futex code.
1914 *
1915 * The main difference with get_user_pages() is that this function will
1916 * unconditionally call handle_mm_fault() which will in turn perform all the
1917 * necessary SW fixup of the dirty and young bits in the PTE, while
1918 * handle_mm_fault() only guarantees to update these in the struct page.
1919 *
1920 * This is important for some architectures where those bits also gate the
1921 * access permission to the page because they are maintained in software. On
1922 * such architectures, gup() will not be enough to make a subsequent access
1923 * succeed.
1924 *
1925 * This should be called with the mm_sem held for read.
1926 */
1927int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1928 unsigned long address, unsigned int fault_flags)
1929{
1930 struct vm_area_struct *vma;
1931 int ret;
1932
1933 vma = find_extend_vma(mm, address);
1934 if (!vma || address < vma->vm_start)
1935 return -EFAULT;
1936
1937 ret = handle_mm_fault(mm, vma, address, fault_flags);
1938 if (ret & VM_FAULT_ERROR) {
1939 if (ret & VM_FAULT_OOM)
1940 return -ENOMEM;
1941 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1942 return -EHWPOISON;
1943 if (ret & VM_FAULT_SIGBUS)
1944 return -EFAULT;
1945 BUG();
1946 }
1947 if (tsk) {
1948 if (ret & VM_FAULT_MAJOR)
1949 tsk->maj_flt++;
1950 else
1951 tsk->min_flt++;
1952 }
1953 return 0;
1954}
1955
1956/*
1957 * get_user_pages() - pin user pages in memory
1958 * @tsk: the task_struct to use for page fault accounting, or
1959 * NULL if faults are not to be recorded.
1960 * @mm: mm_struct of target mm
1961 * @start: starting user address
1962 * @nr_pages: number of pages from start to pin
1963 * @write: whether pages will be written to by the caller
1964 * @force: whether to force write access even if user mapping is
1965 * readonly. This will result in the page being COWed even
1966 * in MAP_SHARED mappings. You do not want this.
1967 * @pages: array that receives pointers to the pages pinned.
1968 * Should be at least nr_pages long. Or NULL, if caller
1969 * only intends to ensure the pages are faulted in.
1970 * @vmas: array of pointers to vmas corresponding to each page.
1971 * Or NULL if the caller does not require them.
1972 *
1973 * Returns number of pages pinned. This may be fewer than the number
1974 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1975 * were pinned, returns -errno. Each page returned must be released
1976 * with a put_page() call when it is finished with. vmas will only
1977 * remain valid while mmap_sem is held.
1978 *
1979 * Must be called with mmap_sem held for read or write.
1980 *
1981 * get_user_pages walks a process's page tables and takes a reference to
1982 * each struct page that each user address corresponds to at a given
1983 * instant. That is, it takes the page that would be accessed if a user
1984 * thread accesses the given user virtual address at that instant.
1985 *
1986 * This does not guarantee that the page exists in the user mappings when
1987 * get_user_pages returns, and there may even be a completely different
1988 * page there in some cases (eg. if mmapped pagecache has been invalidated
1989 * and subsequently re faulted). However it does guarantee that the page
1990 * won't be freed completely. And mostly callers simply care that the page
1991 * contains data that was valid *at some point in time*. Typically, an IO
1992 * or similar operation cannot guarantee anything stronger anyway because
1993 * locks can't be held over the syscall boundary.
1994 *
1995 * If write=0, the page must not be written to. If the page is written to,
1996 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1997 * after the page is finished with, and before put_page is called.
1998 *
1999 * get_user_pages is typically used for fewer-copy IO operations, to get a
2000 * handle on the memory by some means other than accesses via the user virtual
2001 * addresses. The pages may be submitted for DMA to devices or accessed via
2002 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2003 * use the correct cache flushing APIs.
2004 *
2005 * See also get_user_pages_fast, for performance critical applications.
2006 */
2007long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2008 unsigned long start, unsigned long nr_pages, int write,
2009 int force, struct page **pages, struct vm_area_struct **vmas)
2010{
2011 int flags = FOLL_TOUCH;
2012
2013 if (pages)
2014 flags |= FOLL_GET;
2015 if (write)
2016 flags |= FOLL_WRITE;
2017 if (force)
2018 flags |= FOLL_FORCE;
2019
2020 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2021 NULL);
2022}
2023EXPORT_SYMBOL(get_user_pages);
2024
2025/**
2026 * get_dump_page() - pin user page in memory while writing it to core dump
2027 * @addr: user address
2028 *
2029 * Returns struct page pointer of user page pinned for dump,
2030 * to be freed afterwards by page_cache_release() or put_page().
2031 *
2032 * Returns NULL on any kind of failure - a hole must then be inserted into
2033 * the corefile, to preserve alignment with its headers; and also returns
2034 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2035 * allowing a hole to be left in the corefile to save diskspace.
2036 *
2037 * Called without mmap_sem, but after all other threads have been killed.
2038 */
2039#ifdef CONFIG_ELF_CORE
2040struct page *get_dump_page(unsigned long addr)
2041{
2042 struct vm_area_struct *vma;
2043 struct page *page;
2044
2045 if (__get_user_pages(current, current->mm, addr, 1,
2046 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2047 NULL) < 1)
2048 return NULL;
2049 flush_cache_page(vma, addr, page_to_pfn(page));
2050 return page;
2051}
2052#endif /* CONFIG_ELF_CORE */
2053
2054pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2055 spinlock_t **ptl)
2056{
2057 pgd_t * pgd = pgd_offset(mm, addr);
2058 pud_t * pud = pud_alloc(mm, pgd, addr);
2059 if (pud) {
2060 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2061 if (pmd) {
2062 VM_BUG_ON(pmd_trans_huge(*pmd));
2063 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2064 }
2065 }
2066 return NULL;
2067}
2068
2069/*
2070 * This is the old fallback for page remapping.
2071 *
2072 * For historical reasons, it only allows reserved pages. Only
2073 * old drivers should use this, and they needed to mark their
2074 * pages reserved for the old functions anyway.
2075 */
2076static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2077 struct page *page, pgprot_t prot)
2078{
2079 struct mm_struct *mm = vma->vm_mm;
2080 int retval;
2081 pte_t *pte;
2082 spinlock_t *ptl;
2083
2084 retval = -EINVAL;
2085 if (PageAnon(page))
2086 goto out;
2087 retval = -ENOMEM;
2088 flush_dcache_page(page);
2089 pte = get_locked_pte(mm, addr, &ptl);
2090 if (!pte)
2091 goto out;
2092 retval = -EBUSY;
2093 if (!pte_none(*pte))
2094 goto out_unlock;
2095
2096 /* Ok, finally just insert the thing.. */
2097 get_page(page);
2098 inc_mm_counter_fast(mm, MM_FILEPAGES);
2099 page_add_file_rmap(page);
2100 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2101
2102 retval = 0;
2103 pte_unmap_unlock(pte, ptl);
2104 return retval;
2105out_unlock:
2106 pte_unmap_unlock(pte, ptl);
2107out:
2108 return retval;
2109}
2110
2111/**
2112 * vm_insert_page - insert single page into user vma
2113 * @vma: user vma to map to
2114 * @addr: target user address of this page
2115 * @page: source kernel page
2116 *
2117 * This allows drivers to insert individual pages they've allocated
2118 * into a user vma.
2119 *
2120 * The page has to be a nice clean _individual_ kernel allocation.
2121 * If you allocate a compound page, you need to have marked it as
2122 * such (__GFP_COMP), or manually just split the page up yourself
2123 * (see split_page()).
2124 *
2125 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2126 * took an arbitrary page protection parameter. This doesn't allow
2127 * that. Your vma protection will have to be set up correctly, which
2128 * means that if you want a shared writable mapping, you'd better
2129 * ask for a shared writable mapping!
2130 *
2131 * The page does not need to be reserved.
2132 *
2133 * Usually this function is called from f_op->mmap() handler
2134 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2135 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2136 * function from other places, for example from page-fault handler.
2137 */
2138int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2139 struct page *page)
2140{
2141 if (addr < vma->vm_start || addr >= vma->vm_end)
2142 return -EFAULT;
2143 if (!page_count(page))
2144 return -EINVAL;
2145 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2146 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2147 BUG_ON(vma->vm_flags & VM_PFNMAP);
2148 vma->vm_flags |= VM_MIXEDMAP;
2149 }
2150 return insert_page(vma, addr, page, vma->vm_page_prot);
2151}
2152EXPORT_SYMBOL(vm_insert_page);
2153
2154static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2155 unsigned long pfn, pgprot_t prot)
2156{
2157 struct mm_struct *mm = vma->vm_mm;
2158 int retval;
2159 pte_t *pte, entry;
2160 spinlock_t *ptl;
2161
2162 retval = -ENOMEM;
2163 pte = get_locked_pte(mm, addr, &ptl);
2164 if (!pte)
2165 goto out;
2166 retval = -EBUSY;
2167 if (!pte_none(*pte))
2168 goto out_unlock;
2169
2170 /* Ok, finally just insert the thing.. */
2171 entry = pte_mkspecial(pfn_pte(pfn, prot));
2172 set_pte_at(mm, addr, pte, entry);
2173 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2174
2175 retval = 0;
2176out_unlock:
2177 pte_unmap_unlock(pte, ptl);
2178out:
2179 return retval;
2180}
2181
2182/**
2183 * vm_insert_pfn - insert single pfn into user vma
2184 * @vma: user vma to map to
2185 * @addr: target user address of this page
2186 * @pfn: source kernel pfn
2187 *
2188 * Similar to vm_insert_page, this allows drivers to insert individual pages
2189 * they've allocated into a user vma. Same comments apply.
2190 *
2191 * This function should only be called from a vm_ops->fault handler, and
2192 * in that case the handler should return NULL.
2193 *
2194 * vma cannot be a COW mapping.
2195 *
2196 * As this is called only for pages that do not currently exist, we
2197 * do not need to flush old virtual caches or the TLB.
2198 */
2199int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2200 unsigned long pfn)
2201{
2202 int ret;
2203 pgprot_t pgprot = vma->vm_page_prot;
2204 /*
2205 * Technically, architectures with pte_special can avoid all these
2206 * restrictions (same for remap_pfn_range). However we would like
2207 * consistency in testing and feature parity among all, so we should
2208 * try to keep these invariants in place for everybody.
2209 */
2210 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2211 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2212 (VM_PFNMAP|VM_MIXEDMAP));
2213 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2214 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2215
2216 if (addr < vma->vm_start || addr >= vma->vm_end)
2217 return -EFAULT;
2218 if (track_pfn_insert(vma, &pgprot, pfn))
2219 return -EINVAL;
2220
2221 ret = insert_pfn(vma, addr, pfn, pgprot);
2222
2223 return ret;
2224}
2225EXPORT_SYMBOL(vm_insert_pfn);
2226
2227int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2228 unsigned long pfn)
2229{
2230 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2231
2232 if (addr < vma->vm_start || addr >= vma->vm_end)
2233 return -EFAULT;
2234
2235 /*
2236 * If we don't have pte special, then we have to use the pfn_valid()
2237 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2238 * refcount the page if pfn_valid is true (hence insert_page rather
2239 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2240 * without pte special, it would there be refcounted as a normal page.
2241 */
2242 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2243 struct page *page;
2244
2245 page = pfn_to_page(pfn);
2246 return insert_page(vma, addr, page, vma->vm_page_prot);
2247 }
2248 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2249}
2250EXPORT_SYMBOL(vm_insert_mixed);
2251
2252/*
2253 * maps a range of physical memory into the requested pages. the old
2254 * mappings are removed. any references to nonexistent pages results
2255 * in null mappings (currently treated as "copy-on-access")
2256 */
2257static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2258 unsigned long addr, unsigned long end,
2259 unsigned long pfn, pgprot_t prot)
2260{
2261 pte_t *pte;
2262 spinlock_t *ptl;
2263
2264 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2265 if (!pte)
2266 return -ENOMEM;
2267 arch_enter_lazy_mmu_mode();
2268 do {
2269 BUG_ON(!pte_none(*pte));
2270 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2271 pfn++;
2272 } while (pte++, addr += PAGE_SIZE, addr != end);
2273 arch_leave_lazy_mmu_mode();
2274 pte_unmap_unlock(pte - 1, ptl);
2275 return 0;
2276}
2277
2278static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2279 unsigned long addr, unsigned long end,
2280 unsigned long pfn, pgprot_t prot)
2281{
2282 pmd_t *pmd;
2283 unsigned long next;
2284
2285 pfn -= addr >> PAGE_SHIFT;
2286 pmd = pmd_alloc(mm, pud, addr);
2287 if (!pmd)
2288 return -ENOMEM;
2289 VM_BUG_ON(pmd_trans_huge(*pmd));
2290 do {
2291 next = pmd_addr_end(addr, end);
2292 if (remap_pte_range(mm, pmd, addr, next,
2293 pfn + (addr >> PAGE_SHIFT), prot))
2294 return -ENOMEM;
2295 } while (pmd++, addr = next, addr != end);
2296 return 0;
2297}
2298
2299static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2300 unsigned long addr, unsigned long end,
2301 unsigned long pfn, pgprot_t prot)
2302{
2303 pud_t *pud;
2304 unsigned long next;
2305
2306 pfn -= addr >> PAGE_SHIFT;
2307 pud = pud_alloc(mm, pgd, addr);
2308 if (!pud)
2309 return -ENOMEM;
2310 do {
2311 next = pud_addr_end(addr, end);
2312 if (remap_pmd_range(mm, pud, addr, next,
2313 pfn + (addr >> PAGE_SHIFT), prot))
2314 return -ENOMEM;
2315 } while (pud++, addr = next, addr != end);
2316 return 0;
2317}
2318
2319/**
2320 * remap_pfn_range - remap kernel memory to userspace
2321 * @vma: user vma to map to
2322 * @addr: target user address to start at
2323 * @pfn: physical address of kernel memory
2324 * @size: size of map area
2325 * @prot: page protection flags for this mapping
2326 *
2327 * Note: this is only safe if the mm semaphore is held when called.
2328 */
2329int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2330 unsigned long pfn, unsigned long size, pgprot_t prot)
2331{
2332 pgd_t *pgd;
2333 unsigned long next;
2334 unsigned long end = addr + PAGE_ALIGN(size);
2335 struct mm_struct *mm = vma->vm_mm;
2336 int err;
2337
2338 /*
2339 * Physically remapped pages are special. Tell the
2340 * rest of the world about it:
2341 * VM_IO tells people not to look at these pages
2342 * (accesses can have side effects).
2343 * VM_PFNMAP tells the core MM that the base pages are just
2344 * raw PFN mappings, and do not have a "struct page" associated
2345 * with them.
2346 * VM_DONTEXPAND
2347 * Disable vma merging and expanding with mremap().
2348 * VM_DONTDUMP
2349 * Omit vma from core dump, even when VM_IO turned off.
2350 *
2351 * There's a horrible special case to handle copy-on-write
2352 * behaviour that some programs depend on. We mark the "original"
2353 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2354 * See vm_normal_page() for details.
2355 */
2356 if (is_cow_mapping(vma->vm_flags)) {
2357 if (addr != vma->vm_start || end != vma->vm_end)
2358 return -EINVAL;
2359 vma->vm_pgoff = pfn;
2360 }
2361
2362 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2363 if (err)
2364 return -EINVAL;
2365
2366 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2367
2368 BUG_ON(addr >= end);
2369 pfn -= addr >> PAGE_SHIFT;
2370 pgd = pgd_offset(mm, addr);
2371 flush_cache_range(vma, addr, end);
2372 do {
2373 next = pgd_addr_end(addr, end);
2374 err = remap_pud_range(mm, pgd, addr, next,
2375 pfn + (addr >> PAGE_SHIFT), prot);
2376 if (err)
2377 break;
2378 } while (pgd++, addr = next, addr != end);
2379
2380 if (err)
2381 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2382
2383 return err;
2384}
2385EXPORT_SYMBOL(remap_pfn_range);
2386
2387/**
2388 * vm_iomap_memory - remap memory to userspace
2389 * @vma: user vma to map to
2390 * @start: start of area
2391 * @len: size of area
2392 *
2393 * This is a simplified io_remap_pfn_range() for common driver use. The
2394 * driver just needs to give us the physical memory range to be mapped,
2395 * we'll figure out the rest from the vma information.
2396 *
2397 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2398 * whatever write-combining details or similar.
2399 */
2400int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2401{
2402 unsigned long vm_len, pfn, pages;
2403
2404 /* Check that the physical memory area passed in looks valid */
2405 if (start + len < start)
2406 return -EINVAL;
2407 /*
2408 * You *really* shouldn't map things that aren't page-aligned,
2409 * but we've historically allowed it because IO memory might
2410 * just have smaller alignment.
2411 */
2412 len += start & ~PAGE_MASK;
2413 pfn = start >> PAGE_SHIFT;
2414 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2415 if (pfn + pages < pfn)
2416 return -EINVAL;
2417
2418 /* We start the mapping 'vm_pgoff' pages into the area */
2419 if (vma->vm_pgoff > pages)
2420 return -EINVAL;
2421 pfn += vma->vm_pgoff;
2422 pages -= vma->vm_pgoff;
2423
2424 /* Can we fit all of the mapping? */
2425 vm_len = vma->vm_end - vma->vm_start;
2426 if (vm_len >> PAGE_SHIFT > pages)
2427 return -EINVAL;
2428
2429 /* Ok, let it rip */
2430 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2431}
2432EXPORT_SYMBOL(vm_iomap_memory);
2433
2434static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2435 unsigned long addr, unsigned long end,
2436 pte_fn_t fn, void *data)
2437{
2438 pte_t *pte;
2439 int err;
2440 pgtable_t token;
2441 spinlock_t *uninitialized_var(ptl);
2442
2443 pte = (mm == &init_mm) ?
2444 pte_alloc_kernel(pmd, addr) :
2445 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2446 if (!pte)
2447 return -ENOMEM;
2448
2449 BUG_ON(pmd_huge(*pmd));
2450
2451 arch_enter_lazy_mmu_mode();
2452
2453 token = pmd_pgtable(*pmd);
2454
2455 do {
2456 err = fn(pte++, token, addr, data);
2457 if (err)
2458 break;
2459 } while (addr += PAGE_SIZE, addr != end);
2460
2461 arch_leave_lazy_mmu_mode();
2462
2463 if (mm != &init_mm)
2464 pte_unmap_unlock(pte-1, ptl);
2465 return err;
2466}
2467
2468static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2469 unsigned long addr, unsigned long end,
2470 pte_fn_t fn, void *data)
2471{
2472 pmd_t *pmd;
2473 unsigned long next;
2474 int err;
2475
2476 BUG_ON(pud_huge(*pud));
2477
2478 pmd = pmd_alloc(mm, pud, addr);
2479 if (!pmd)
2480 return -ENOMEM;
2481 do {
2482 next = pmd_addr_end(addr, end);
2483 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2484 if (err)
2485 break;
2486 } while (pmd++, addr = next, addr != end);
2487 return err;
2488}
2489
2490static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2491 unsigned long addr, unsigned long end,
2492 pte_fn_t fn, void *data)
2493{
2494 pud_t *pud;
2495 unsigned long next;
2496 int err;
2497
2498 pud = pud_alloc(mm, pgd, addr);
2499 if (!pud)
2500 return -ENOMEM;
2501 do {
2502 next = pud_addr_end(addr, end);
2503 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2504 if (err)
2505 break;
2506 } while (pud++, addr = next, addr != end);
2507 return err;
2508}
2509
2510/*
2511 * Scan a region of virtual memory, filling in page tables as necessary
2512 * and calling a provided function on each leaf page table.
2513 */
2514int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2515 unsigned long size, pte_fn_t fn, void *data)
2516{
2517 pgd_t *pgd;
2518 unsigned long next;
2519 unsigned long end = addr + size;
2520 int err;
2521
2522 BUG_ON(addr >= end);
2523 pgd = pgd_offset(mm, addr);
2524 do {
2525 next = pgd_addr_end(addr, end);
2526 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2527 if (err)
2528 break;
2529 } while (pgd++, addr = next, addr != end);
2530
2531 return err;
2532}
2533EXPORT_SYMBOL_GPL(apply_to_page_range);
2534
2535/*
2536 * handle_pte_fault chooses page fault handler according to an entry
2537 * which was read non-atomically. Before making any commitment, on
2538 * those architectures or configurations (e.g. i386 with PAE) which
2539 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2540 * must check under lock before unmapping the pte and proceeding
2541 * (but do_wp_page is only called after already making such a check;
2542 * and do_anonymous_page can safely check later on).
2543 */
2544static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2545 pte_t *page_table, pte_t orig_pte)
2546{
2547 int same = 1;
2548#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2549 if (sizeof(pte_t) > sizeof(unsigned long)) {
2550 spinlock_t *ptl = pte_lockptr(mm, pmd);
2551 spin_lock(ptl);
2552 same = pte_same(*page_table, orig_pte);
2553 spin_unlock(ptl);
2554 }
2555#endif
2556 pte_unmap(page_table);
2557 return same;
2558}
2559
2560static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2561{
2562 /*
2563 * If the source page was a PFN mapping, we don't have
2564 * a "struct page" for it. We do a best-effort copy by
2565 * just copying from the original user address. If that
2566 * fails, we just zero-fill it. Live with it.
2567 */
2568 if (unlikely(!src)) {
2569 void *kaddr = kmap_atomic(dst);
2570 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2571
2572 /*
2573 * This really shouldn't fail, because the page is there
2574 * in the page tables. But it might just be unreadable,
2575 * in which case we just give up and fill the result with
2576 * zeroes.
2577 */
2578 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2579 clear_page(kaddr);
2580 kunmap_atomic(kaddr);
2581 flush_dcache_page(dst);
2582 } else
2583 copy_user_highpage(dst, src, va, vma);
2584}
2585
2586/*
2587 * This routine handles present pages, when users try to write
2588 * to a shared page. It is done by copying the page to a new address
2589 * and decrementing the shared-page counter for the old page.
2590 *
2591 * Note that this routine assumes that the protection checks have been
2592 * done by the caller (the low-level page fault routine in most cases).
2593 * Thus we can safely just mark it writable once we've done any necessary
2594 * COW.
2595 *
2596 * We also mark the page dirty at this point even though the page will
2597 * change only once the write actually happens. This avoids a few races,
2598 * and potentially makes it more efficient.
2599 *
2600 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2601 * but allow concurrent faults), with pte both mapped and locked.
2602 * We return with mmap_sem still held, but pte unmapped and unlocked.
2603 */
2604static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2605 unsigned long address, pte_t *page_table, pmd_t *pmd,
2606 spinlock_t *ptl, pte_t orig_pte)
2607 __releases(ptl)
2608{
2609 struct page *old_page, *new_page = NULL;
2610 pte_t entry;
2611 int ret = 0;
2612 int page_mkwrite = 0;
2613 struct page *dirty_page = NULL;
2614 unsigned long mmun_start = 0; /* For mmu_notifiers */
2615 unsigned long mmun_end = 0; /* For mmu_notifiers */
2616
2617 old_page = vm_normal_page(vma, address, orig_pte);
2618 if (!old_page) {
2619 /*
2620 * VM_MIXEDMAP !pfn_valid() case
2621 *
2622 * We should not cow pages in a shared writeable mapping.
2623 * Just mark the pages writable as we can't do any dirty
2624 * accounting on raw pfn maps.
2625 */
2626 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2627 (VM_WRITE|VM_SHARED))
2628 goto reuse;
2629 goto gotten;
2630 }
2631
2632 /*
2633 * Take out anonymous pages first, anonymous shared vmas are
2634 * not dirty accountable.
2635 */
2636 if (PageAnon(old_page) && !PageKsm(old_page)) {
2637 if (!trylock_page(old_page)) {
2638 page_cache_get(old_page);
2639 pte_unmap_unlock(page_table, ptl);
2640 lock_page(old_page);
2641 page_table = pte_offset_map_lock(mm, pmd, address,
2642 &ptl);
2643 if (!pte_same(*page_table, orig_pte)) {
2644 unlock_page(old_page);
2645 goto unlock;
2646 }
2647 page_cache_release(old_page);
2648 }
2649 if (reuse_swap_page(old_page)) {
2650 /*
2651 * The page is all ours. Move it to our anon_vma so
2652 * the rmap code will not search our parent or siblings.
2653 * Protected against the rmap code by the page lock.
2654 */
2655 page_move_anon_rmap(old_page, vma, address);
2656 unlock_page(old_page);
2657 goto reuse;
2658 }
2659 unlock_page(old_page);
2660 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2661 (VM_WRITE|VM_SHARED))) {
2662 /*
2663 * Only catch write-faults on shared writable pages,
2664 * read-only shared pages can get COWed by
2665 * get_user_pages(.write=1, .force=1).
2666 */
2667 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2668 struct vm_fault vmf;
2669 int tmp;
2670
2671 vmf.virtual_address = (void __user *)(address &
2672 PAGE_MASK);
2673 vmf.pgoff = old_page->index;
2674 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2675 vmf.page = old_page;
2676
2677 /*
2678 * Notify the address space that the page is about to
2679 * become writable so that it can prohibit this or wait
2680 * for the page to get into an appropriate state.
2681 *
2682 * We do this without the lock held, so that it can
2683 * sleep if it needs to.
2684 */
2685 page_cache_get(old_page);
2686 pte_unmap_unlock(page_table, ptl);
2687
2688 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2689 if (unlikely(tmp &
2690 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2691 ret = tmp;
2692 goto unwritable_page;
2693 }
2694 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2695 lock_page(old_page);
2696 if (!old_page->mapping) {
2697 ret = 0; /* retry the fault */
2698 unlock_page(old_page);
2699 goto unwritable_page;
2700 }
2701 } else
2702 VM_BUG_ON(!PageLocked(old_page));
2703
2704 /*
2705 * Since we dropped the lock we need to revalidate
2706 * the PTE as someone else may have changed it. If
2707 * they did, we just return, as we can count on the
2708 * MMU to tell us if they didn't also make it writable.
2709 */
2710 page_table = pte_offset_map_lock(mm, pmd, address,
2711 &ptl);
2712 if (!pte_same(*page_table, orig_pte)) {
2713 unlock_page(old_page);
2714 goto unlock;
2715 }
2716
2717 page_mkwrite = 1;
2718 }
2719 dirty_page = old_page;
2720 get_page(dirty_page);
2721
2722reuse:
2723 /*
2724 * Clear the pages cpupid information as the existing
2725 * information potentially belongs to a now completely
2726 * unrelated process.
2727 */
2728 if (old_page)
2729 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2730
2731 flush_cache_page(vma, address, pte_pfn(orig_pte));
2732 entry = pte_mkyoung(orig_pte);
2733 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2734 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2735 update_mmu_cache(vma, address, page_table);
2736 pte_unmap_unlock(page_table, ptl);
2737 ret |= VM_FAULT_WRITE;
2738
2739 if (!dirty_page)
2740 return ret;
2741
2742 /*
2743 * Yes, Virginia, this is actually required to prevent a race
2744 * with clear_page_dirty_for_io() from clearing the page dirty
2745 * bit after it clear all dirty ptes, but before a racing
2746 * do_wp_page installs a dirty pte.
2747 *
2748 * __do_fault is protected similarly.
2749 */
2750 if (!page_mkwrite) {
2751 wait_on_page_locked(dirty_page);
2752 set_page_dirty_balance(dirty_page, page_mkwrite);
2753 /* file_update_time outside page_lock */
2754 if (vma->vm_file)
2755 file_update_time(vma->vm_file);
2756 }
2757 put_page(dirty_page);
2758 if (page_mkwrite) {
2759 struct address_space *mapping = dirty_page->mapping;
2760
2761 set_page_dirty(dirty_page);
2762 unlock_page(dirty_page);
2763 page_cache_release(dirty_page);
2764 if (mapping) {
2765 /*
2766 * Some device drivers do not set page.mapping
2767 * but still dirty their pages
2768 */
2769 balance_dirty_pages_ratelimited(mapping);
2770 }
2771 }
2772
2773 return ret;
2774 }
2775
2776 /*
2777 * Ok, we need to copy. Oh, well..
2778 */
2779 page_cache_get(old_page);
2780gotten:
2781 pte_unmap_unlock(page_table, ptl);
2782
2783 if (unlikely(anon_vma_prepare(vma)))
2784 goto oom;
2785
2786 if (is_zero_pfn(pte_pfn(orig_pte))) {
2787 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2788 if (!new_page)
2789 goto oom;
2790 } else {
2791 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2792 if (!new_page)
2793 goto oom;
2794 cow_user_page(new_page, old_page, address, vma);
2795 }
2796 __SetPageUptodate(new_page);
2797
2798 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2799 goto oom_free_new;
2800
2801 mmun_start = address & PAGE_MASK;
2802 mmun_end = mmun_start + PAGE_SIZE;
2803 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2804
2805 /*
2806 * Re-check the pte - we dropped the lock
2807 */
2808 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2809 if (likely(pte_same(*page_table, orig_pte))) {
2810 if (old_page) {
2811 if (!PageAnon(old_page)) {
2812 dec_mm_counter_fast(mm, MM_FILEPAGES);
2813 inc_mm_counter_fast(mm, MM_ANONPAGES);
2814 }
2815 } else
2816 inc_mm_counter_fast(mm, MM_ANONPAGES);
2817 flush_cache_page(vma, address, pte_pfn(orig_pte));
2818 entry = mk_pte(new_page, vma->vm_page_prot);
2819 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2820 /*
2821 * Clear the pte entry and flush it first, before updating the
2822 * pte with the new entry. This will avoid a race condition
2823 * seen in the presence of one thread doing SMC and another
2824 * thread doing COW.
2825 */
2826 ptep_clear_flush(vma, address, page_table);
2827 page_add_new_anon_rmap(new_page, vma, address);
2828 /*
2829 * We call the notify macro here because, when using secondary
2830 * mmu page tables (such as kvm shadow page tables), we want the
2831 * new page to be mapped directly into the secondary page table.
2832 */
2833 set_pte_at_notify(mm, address, page_table, entry);
2834 update_mmu_cache(vma, address, page_table);
2835 if (old_page) {
2836 /*
2837 * Only after switching the pte to the new page may
2838 * we remove the mapcount here. Otherwise another
2839 * process may come and find the rmap count decremented
2840 * before the pte is switched to the new page, and
2841 * "reuse" the old page writing into it while our pte
2842 * here still points into it and can be read by other
2843 * threads.
2844 *
2845 * The critical issue is to order this
2846 * page_remove_rmap with the ptp_clear_flush above.
2847 * Those stores are ordered by (if nothing else,)
2848 * the barrier present in the atomic_add_negative
2849 * in page_remove_rmap.
2850 *
2851 * Then the TLB flush in ptep_clear_flush ensures that
2852 * no process can access the old page before the
2853 * decremented mapcount is visible. And the old page
2854 * cannot be reused until after the decremented
2855 * mapcount is visible. So transitively, TLBs to
2856 * old page will be flushed before it can be reused.
2857 */
2858 page_remove_rmap(old_page);
2859 }
2860
2861 /* Free the old page.. */
2862 new_page = old_page;
2863 ret |= VM_FAULT_WRITE;
2864 } else
2865 mem_cgroup_uncharge_page(new_page);
2866
2867 if (new_page)
2868 page_cache_release(new_page);
2869unlock:
2870 pte_unmap_unlock(page_table, ptl);
2871 if (mmun_end > mmun_start)
2872 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2873 if (old_page) {
2874 /*
2875 * Don't let another task, with possibly unlocked vma,
2876 * keep the mlocked page.
2877 */
2878 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2879 lock_page(old_page); /* LRU manipulation */
2880 munlock_vma_page(old_page);
2881 unlock_page(old_page);
2882 }
2883 page_cache_release(old_page);
2884 }
2885 return ret;
2886oom_free_new:
2887 page_cache_release(new_page);
2888oom:
2889 if (old_page)
2890 page_cache_release(old_page);
2891 return VM_FAULT_OOM;
2892
2893unwritable_page:
2894 page_cache_release(old_page);
2895 return ret;
2896}
2897
2898static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2899 unsigned long start_addr, unsigned long end_addr,
2900 struct zap_details *details)
2901{
2902 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2903}
2904
2905static inline void unmap_mapping_range_tree(struct rb_root *root,
2906 struct zap_details *details)
2907{
2908 struct vm_area_struct *vma;
2909 pgoff_t vba, vea, zba, zea;
2910
2911 vma_interval_tree_foreach(vma, root,
2912 details->first_index, details->last_index) {
2913
2914 vba = vma->vm_pgoff;
2915 vea = vba + vma_pages(vma) - 1;
2916 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2917 zba = details->first_index;
2918 if (zba < vba)
2919 zba = vba;
2920 zea = details->last_index;
2921 if (zea > vea)
2922 zea = vea;
2923
2924 unmap_mapping_range_vma(vma,
2925 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2926 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2927 details);
2928 }
2929}
2930
2931static inline void unmap_mapping_range_list(struct list_head *head,
2932 struct zap_details *details)
2933{
2934 struct vm_area_struct *vma;
2935
2936 /*
2937 * In nonlinear VMAs there is no correspondence between virtual address
2938 * offset and file offset. So we must perform an exhaustive search
2939 * across *all* the pages in each nonlinear VMA, not just the pages
2940 * whose virtual address lies outside the file truncation point.
2941 */
2942 list_for_each_entry(vma, head, shared.nonlinear) {
2943 details->nonlinear_vma = vma;
2944 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2945 }
2946}
2947
2948/**
2949 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2950 * @mapping: the address space containing mmaps to be unmapped.
2951 * @holebegin: byte in first page to unmap, relative to the start of
2952 * the underlying file. This will be rounded down to a PAGE_SIZE
2953 * boundary. Note that this is different from truncate_pagecache(), which
2954 * must keep the partial page. In contrast, we must get rid of
2955 * partial pages.
2956 * @holelen: size of prospective hole in bytes. This will be rounded
2957 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2958 * end of the file.
2959 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2960 * but 0 when invalidating pagecache, don't throw away private data.
2961 */
2962void unmap_mapping_range(struct address_space *mapping,
2963 loff_t const holebegin, loff_t const holelen, int even_cows)
2964{
2965 struct zap_details details;
2966 pgoff_t hba = holebegin >> PAGE_SHIFT;
2967 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2968
2969 /* Check for overflow. */
2970 if (sizeof(holelen) > sizeof(hlen)) {
2971 long long holeend =
2972 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2973 if (holeend & ~(long long)ULONG_MAX)
2974 hlen = ULONG_MAX - hba + 1;
2975 }
2976
2977 details.check_mapping = even_cows? NULL: mapping;
2978 details.nonlinear_vma = NULL;
2979 details.first_index = hba;
2980 details.last_index = hba + hlen - 1;
2981 if (details.last_index < details.first_index)
2982 details.last_index = ULONG_MAX;
2983
2984
2985 mutex_lock(&mapping->i_mmap_mutex);
2986 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2987 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2988 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2989 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2990 mutex_unlock(&mapping->i_mmap_mutex);
2991}
2992EXPORT_SYMBOL(unmap_mapping_range);
2993
2994/*
2995 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2996 * but allow concurrent faults), and pte mapped but not yet locked.
2997 * We return with mmap_sem still held, but pte unmapped and unlocked.
2998 */
2999static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3000 unsigned long address, pte_t *page_table, pmd_t *pmd,
3001 unsigned int flags, pte_t orig_pte)
3002{
3003 spinlock_t *ptl;
3004 struct page *page, *swapcache;
3005 swp_entry_t entry;
3006 pte_t pte;
3007 int locked;
3008 struct mem_cgroup *ptr;
3009 int exclusive = 0;
3010 int ret = 0;
3011
3012 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3013 goto out;
3014
3015 entry = pte_to_swp_entry(orig_pte);
3016 if (unlikely(non_swap_entry(entry))) {
3017 if (is_migration_entry(entry)) {
3018 migration_entry_wait(mm, pmd, address);
3019 } else if (is_hwpoison_entry(entry)) {
3020 ret = VM_FAULT_HWPOISON;
3021 } else {
3022 print_bad_pte(vma, address, orig_pte, NULL);
3023 ret = VM_FAULT_SIGBUS;
3024 }
3025 goto out;
3026 }
3027 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3028 page = lookup_swap_cache(entry);
3029 if (!page) {
3030 page = swapin_readahead(entry,
3031 GFP_HIGHUSER_MOVABLE, vma, address);
3032 if (!page) {
3033 /*
3034 * Back out if somebody else faulted in this pte
3035 * while we released the pte lock.
3036 */
3037 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3038 if (likely(pte_same(*page_table, orig_pte)))
3039 ret = VM_FAULT_OOM;
3040 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3041 goto unlock;
3042 }
3043
3044 /* Had to read the page from swap area: Major fault */
3045 ret = VM_FAULT_MAJOR;
3046 count_vm_event(PGMAJFAULT);
3047 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3048 } else if (PageHWPoison(page)) {
3049 /*
3050 * hwpoisoned dirty swapcache pages are kept for killing
3051 * owner processes (which may be unknown at hwpoison time)
3052 */
3053 ret = VM_FAULT_HWPOISON;
3054 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3055 swapcache = page;
3056 goto out_release;
3057 }
3058
3059 swapcache = page;
3060 locked = lock_page_or_retry(page, mm, flags);
3061
3062 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3063 if (!locked) {
3064 ret |= VM_FAULT_RETRY;
3065 goto out_release;
3066 }
3067
3068 /*
3069 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3070 * release the swapcache from under us. The page pin, and pte_same
3071 * test below, are not enough to exclude that. Even if it is still
3072 * swapcache, we need to check that the page's swap has not changed.
3073 */
3074 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3075 goto out_page;
3076
3077 page = ksm_might_need_to_copy(page, vma, address);
3078 if (unlikely(!page)) {
3079 ret = VM_FAULT_OOM;
3080 page = swapcache;
3081 goto out_page;
3082 }
3083
3084 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3085 ret = VM_FAULT_OOM;
3086 goto out_page;
3087 }
3088
3089 /*
3090 * Back out if somebody else already faulted in this pte.
3091 */
3092 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3093 if (unlikely(!pte_same(*page_table, orig_pte)))
3094 goto out_nomap;
3095
3096 if (unlikely(!PageUptodate(page))) {
3097 ret = VM_FAULT_SIGBUS;
3098 goto out_nomap;
3099 }
3100
3101 /*
3102 * The page isn't present yet, go ahead with the fault.
3103 *
3104 * Be careful about the sequence of operations here.
3105 * To get its accounting right, reuse_swap_page() must be called
3106 * while the page is counted on swap but not yet in mapcount i.e.
3107 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3108 * must be called after the swap_free(), or it will never succeed.
3109 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3110 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3111 * in page->private. In this case, a record in swap_cgroup is silently
3112 * discarded at swap_free().
3113 */
3114
3115 inc_mm_counter_fast(mm, MM_ANONPAGES);
3116 dec_mm_counter_fast(mm, MM_SWAPENTS);
3117 pte = mk_pte(page, vma->vm_page_prot);
3118 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3119 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3120 flags &= ~FAULT_FLAG_WRITE;
3121 ret |= VM_FAULT_WRITE;
3122 exclusive = 1;
3123 }
3124 flush_icache_page(vma, page);
3125 if (pte_swp_soft_dirty(orig_pte))
3126 pte = pte_mksoft_dirty(pte);
3127 set_pte_at(mm, address, page_table, pte);
3128 if (page == swapcache)
3129 do_page_add_anon_rmap(page, vma, address, exclusive);
3130 else /* ksm created a completely new copy */
3131 page_add_new_anon_rmap(page, vma, address);
3132 /* It's better to call commit-charge after rmap is established */
3133 mem_cgroup_commit_charge_swapin(page, ptr);
3134
3135 swap_free(entry);
3136 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3137 try_to_free_swap(page);
3138 unlock_page(page);
3139 if (page != swapcache) {
3140 /*
3141 * Hold the lock to avoid the swap entry to be reused
3142 * until we take the PT lock for the pte_same() check
3143 * (to avoid false positives from pte_same). For
3144 * further safety release the lock after the swap_free
3145 * so that the swap count won't change under a
3146 * parallel locked swapcache.
3147 */
3148 unlock_page(swapcache);
3149 page_cache_release(swapcache);
3150 }
3151
3152 if (flags & FAULT_FLAG_WRITE) {
3153 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3154 if (ret & VM_FAULT_ERROR)
3155 ret &= VM_FAULT_ERROR;
3156 goto out;
3157 }
3158
3159 /* No need to invalidate - it was non-present before */
3160 update_mmu_cache(vma, address, page_table);
3161unlock:
3162 pte_unmap_unlock(page_table, ptl);
3163out:
3164 return ret;
3165out_nomap:
3166 mem_cgroup_cancel_charge_swapin(ptr);
3167 pte_unmap_unlock(page_table, ptl);
3168out_page:
3169 unlock_page(page);
3170out_release:
3171 page_cache_release(page);
3172 if (page != swapcache) {
3173 unlock_page(swapcache);
3174 page_cache_release(swapcache);
3175 }
3176 return ret;
3177}
3178
3179/*
3180 * This is like a special single-page "expand_{down|up}wards()",
3181 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3182 * doesn't hit another vma.
3183 */
3184static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3185{
3186 address &= PAGE_MASK;
3187 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3188 struct vm_area_struct *prev = vma->vm_prev;
3189
3190 /*
3191 * Is there a mapping abutting this one below?
3192 *
3193 * That's only ok if it's the same stack mapping
3194 * that has gotten split..
3195 */
3196 if (prev && prev->vm_end == address)
3197 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3198
3199 expand_downwards(vma, address - PAGE_SIZE);
3200 }
3201 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3202 struct vm_area_struct *next = vma->vm_next;
3203
3204 /* As VM_GROWSDOWN but s/below/above/ */
3205 if (next && next->vm_start == address + PAGE_SIZE)
3206 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3207
3208 expand_upwards(vma, address + PAGE_SIZE);
3209 }
3210 return 0;
3211}
3212
3213/*
3214 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3215 * but allow concurrent faults), and pte mapped but not yet locked.
3216 * We return with mmap_sem still held, but pte unmapped and unlocked.
3217 */
3218static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3219 unsigned long address, pte_t *page_table, pmd_t *pmd,
3220 unsigned int flags)
3221{
3222 struct page *page;
3223 spinlock_t *ptl;
3224 pte_t entry;
3225
3226 pte_unmap(page_table);
3227
3228 /* Check if we need to add a guard page to the stack */
3229 if (check_stack_guard_page(vma, address) < 0)
3230 return VM_FAULT_SIGBUS;
3231
3232 /* Use the zero-page for reads */
3233 if (!(flags & FAULT_FLAG_WRITE)) {
3234 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3235 vma->vm_page_prot));
3236 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3237 if (!pte_none(*page_table))
3238 goto unlock;
3239 goto setpte;
3240 }
3241
3242 /* Allocate our own private page. */
3243 if (unlikely(anon_vma_prepare(vma)))
3244 goto oom;
3245 page = alloc_zeroed_user_highpage_movable(vma, address);
3246 if (!page)
3247 goto oom;
3248 /*
3249 * The memory barrier inside __SetPageUptodate makes sure that
3250 * preceeding stores to the page contents become visible before
3251 * the set_pte_at() write.
3252 */
3253 __SetPageUptodate(page);
3254
3255 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3256 goto oom_free_page;
3257
3258 entry = mk_pte(page, vma->vm_page_prot);
3259 if (vma->vm_flags & VM_WRITE)
3260 entry = pte_mkwrite(pte_mkdirty(entry));
3261
3262 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3263 if (!pte_none(*page_table))
3264 goto release;
3265
3266 inc_mm_counter_fast(mm, MM_ANONPAGES);
3267 page_add_new_anon_rmap(page, vma, address);
3268setpte:
3269 set_pte_at(mm, address, page_table, entry);
3270
3271 /* No need to invalidate - it was non-present before */
3272 update_mmu_cache(vma, address, page_table);
3273unlock:
3274 pte_unmap_unlock(page_table, ptl);
3275 return 0;
3276release:
3277 mem_cgroup_uncharge_page(page);
3278 page_cache_release(page);
3279 goto unlock;
3280oom_free_page:
3281 page_cache_release(page);
3282oom:
3283 return VM_FAULT_OOM;
3284}
3285
3286/*
3287 * __do_fault() tries to create a new page mapping. It aggressively
3288 * tries to share with existing pages, but makes a separate copy if
3289 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3290 * the next page fault.
3291 *
3292 * As this is called only for pages that do not currently exist, we
3293 * do not need to flush old virtual caches or the TLB.
3294 *
3295 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3296 * but allow concurrent faults), and pte neither mapped nor locked.
3297 * We return with mmap_sem still held, but pte unmapped and unlocked.
3298 */
3299static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3300 unsigned long address, pmd_t *pmd,
3301 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3302{
3303 pte_t *page_table;
3304 spinlock_t *ptl;
3305 struct page *page;
3306 struct page *cow_page;
3307 pte_t entry;
3308 int anon = 0;
3309 struct page *dirty_page = NULL;
3310 struct vm_fault vmf;
3311 int ret;
3312 int page_mkwrite = 0;
3313
3314 /*
3315 * If we do COW later, allocate page befor taking lock_page()
3316 * on the file cache page. This will reduce lock holding time.
3317 */
3318 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3319
3320 if (unlikely(anon_vma_prepare(vma)))
3321 return VM_FAULT_OOM;
3322
3323 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3324 if (!cow_page)
3325 return VM_FAULT_OOM;
3326
3327 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3328 page_cache_release(cow_page);
3329 return VM_FAULT_OOM;
3330 }
3331 } else
3332 cow_page = NULL;
3333
3334 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3335 vmf.pgoff = pgoff;
3336 vmf.flags = flags;
3337 vmf.page = NULL;
3338
3339 ret = vma->vm_ops->fault(vma, &vmf);
3340 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3341 VM_FAULT_RETRY)))
3342 goto uncharge_out;
3343
3344 if (unlikely(PageHWPoison(vmf.page))) {
3345 if (ret & VM_FAULT_LOCKED)
3346 unlock_page(vmf.page);
3347 ret = VM_FAULT_HWPOISON;
3348 goto uncharge_out;
3349 }
3350
3351 /*
3352 * For consistency in subsequent calls, make the faulted page always
3353 * locked.
3354 */
3355 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3356 lock_page(vmf.page);
3357 else
3358 VM_BUG_ON(!PageLocked(vmf.page));
3359
3360 /*
3361 * Should we do an early C-O-W break?
3362 */
3363 page = vmf.page;
3364 if (flags & FAULT_FLAG_WRITE) {
3365 if (!(vma->vm_flags & VM_SHARED)) {
3366 page = cow_page;
3367 anon = 1;
3368 copy_user_highpage(page, vmf.page, address, vma);
3369 __SetPageUptodate(page);
3370 } else {
3371 /*
3372 * If the page will be shareable, see if the backing
3373 * address space wants to know that the page is about
3374 * to become writable
3375 */
3376 if (vma->vm_ops->page_mkwrite) {
3377 int tmp;
3378
3379 unlock_page(page);
3380 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3381 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3382 if (unlikely(tmp &
3383 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3384 ret = tmp;
3385 goto unwritable_page;
3386 }
3387 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3388 lock_page(page);
3389 if (!page->mapping) {
3390 ret = 0; /* retry the fault */
3391 unlock_page(page);
3392 goto unwritable_page;
3393 }
3394 } else
3395 VM_BUG_ON(!PageLocked(page));
3396 page_mkwrite = 1;
3397 }
3398 }
3399
3400 }
3401
3402 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3403
3404 /*
3405 * This silly early PAGE_DIRTY setting removes a race
3406 * due to the bad i386 page protection. But it's valid
3407 * for other architectures too.
3408 *
3409 * Note that if FAULT_FLAG_WRITE is set, we either now have
3410 * an exclusive copy of the page, or this is a shared mapping,
3411 * so we can make it writable and dirty to avoid having to
3412 * handle that later.
3413 */
3414 /* Only go through if we didn't race with anybody else... */
3415 if (likely(pte_same(*page_table, orig_pte))) {
3416 flush_icache_page(vma, page);
3417 entry = mk_pte(page, vma->vm_page_prot);
3418 if (flags & FAULT_FLAG_WRITE)
3419 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3420 else if (pte_file(orig_pte) && pte_file_soft_dirty(orig_pte))
3421 pte_mksoft_dirty(entry);
3422 if (anon) {
3423 inc_mm_counter_fast(mm, MM_ANONPAGES);
3424 page_add_new_anon_rmap(page, vma, address);
3425 } else {
3426 inc_mm_counter_fast(mm, MM_FILEPAGES);
3427 page_add_file_rmap(page);
3428 if (flags & FAULT_FLAG_WRITE) {
3429 dirty_page = page;
3430 get_page(dirty_page);
3431 }
3432 }
3433 set_pte_at(mm, address, page_table, entry);
3434
3435 /* no need to invalidate: a not-present page won't be cached */
3436 update_mmu_cache(vma, address, page_table);
3437 } else {
3438 if (cow_page)
3439 mem_cgroup_uncharge_page(cow_page);
3440 if (anon)
3441 page_cache_release(page);
3442 else
3443 anon = 1; /* no anon but release faulted_page */
3444 }
3445
3446 pte_unmap_unlock(page_table, ptl);
3447
3448 if (dirty_page) {
3449 struct address_space *mapping = page->mapping;
3450 int dirtied = 0;
3451
3452 if (set_page_dirty(dirty_page))
3453 dirtied = 1;
3454 unlock_page(dirty_page);
3455 put_page(dirty_page);
3456 if ((dirtied || page_mkwrite) && mapping) {
3457 /*
3458 * Some device drivers do not set page.mapping but still
3459 * dirty their pages
3460 */
3461 balance_dirty_pages_ratelimited(mapping);
3462 }
3463
3464 /* file_update_time outside page_lock */
3465 if (vma->vm_file && !page_mkwrite)
3466 file_update_time(vma->vm_file);
3467 } else {
3468 unlock_page(vmf.page);
3469 if (anon)
3470 page_cache_release(vmf.page);
3471 }
3472
3473 return ret;
3474
3475unwritable_page:
3476 page_cache_release(page);
3477 return ret;
3478uncharge_out:
3479 /* fs's fault handler get error */
3480 if (cow_page) {
3481 mem_cgroup_uncharge_page(cow_page);
3482 page_cache_release(cow_page);
3483 }
3484 return ret;
3485}
3486
3487static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3488 unsigned long address, pte_t *page_table, pmd_t *pmd,
3489 unsigned int flags, pte_t orig_pte)
3490{
3491 pgoff_t pgoff = (((address & PAGE_MASK)
3492 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3493
3494 pte_unmap(page_table);
3495 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3496}
3497
3498/*
3499 * Fault of a previously existing named mapping. Repopulate the pte
3500 * from the encoded file_pte if possible. This enables swappable
3501 * nonlinear vmas.
3502 *
3503 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3504 * but allow concurrent faults), and pte mapped but not yet locked.
3505 * We return with mmap_sem still held, but pte unmapped and unlocked.
3506 */
3507static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3508 unsigned long address, pte_t *page_table, pmd_t *pmd,
3509 unsigned int flags, pte_t orig_pte)
3510{
3511 pgoff_t pgoff;
3512
3513 flags |= FAULT_FLAG_NONLINEAR;
3514
3515 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3516 return 0;
3517
3518 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3519 /*
3520 * Page table corrupted: show pte and kill process.
3521 */
3522 print_bad_pte(vma, address, orig_pte, NULL);
3523 return VM_FAULT_SIGBUS;
3524 }
3525
3526 pgoff = pte_to_pgoff(orig_pte);
3527 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3528}
3529
3530int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3531 unsigned long addr, int page_nid,
3532 int *flags)
3533{
3534 get_page(page);
3535
3536 count_vm_numa_event(NUMA_HINT_FAULTS);
3537 if (page_nid == numa_node_id()) {
3538 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3539 *flags |= TNF_FAULT_LOCAL;
3540 }
3541
3542 return mpol_misplaced(page, vma, addr);
3543}
3544
3545int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3546 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3547{
3548 struct page *page = NULL;
3549 spinlock_t *ptl;
3550 int page_nid = -1;
3551 int last_cpupid;
3552 int target_nid;
3553 bool migrated = false;
3554 int flags = 0;
3555
3556 /*
3557 * The "pte" at this point cannot be used safely without
3558 * validation through pte_unmap_same(). It's of NUMA type but
3559 * the pfn may be screwed if the read is non atomic.
3560 *
3561 * ptep_modify_prot_start is not called as this is clearing
3562 * the _PAGE_NUMA bit and it is not really expected that there
3563 * would be concurrent hardware modifications to the PTE.
3564 */
3565 ptl = pte_lockptr(mm, pmd);
3566 spin_lock(ptl);
3567 if (unlikely(!pte_same(*ptep, pte))) {
3568 pte_unmap_unlock(ptep, ptl);
3569 goto out;
3570 }
3571
3572 pte = pte_mknonnuma(pte);
3573 set_pte_at(mm, addr, ptep, pte);
3574 update_mmu_cache(vma, addr, ptep);
3575
3576 page = vm_normal_page(vma, addr, pte);
3577 if (!page) {
3578 pte_unmap_unlock(ptep, ptl);
3579 return 0;
3580 }
3581 BUG_ON(is_zero_pfn(page_to_pfn(page)));
3582
3583 /*
3584 * Avoid grouping on DSO/COW pages in specific and RO pages
3585 * in general, RO pages shouldn't hurt as much anyway since
3586 * they can be in shared cache state.
3587 */
3588 if (!pte_write(pte))
3589 flags |= TNF_NO_GROUP;
3590
3591 /*
3592 * Flag if the page is shared between multiple address spaces. This
3593 * is later used when determining whether to group tasks together
3594 */
3595 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3596 flags |= TNF_SHARED;
3597
3598 last_cpupid = page_cpupid_last(page);
3599 page_nid = page_to_nid(page);
3600 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3601 pte_unmap_unlock(ptep, ptl);
3602 if (target_nid == -1) {
3603 put_page(page);
3604 goto out;
3605 }
3606
3607 /* Migrate to the requested node */
3608 migrated = migrate_misplaced_page(page, vma, target_nid);
3609 if (migrated) {
3610 page_nid = target_nid;
3611 flags |= TNF_MIGRATED;
3612 }
3613
3614out:
3615 if (page_nid != -1)
3616 task_numa_fault(last_cpupid, page_nid, 1, flags);
3617 return 0;
3618}
3619
3620/*
3621 * These routines also need to handle stuff like marking pages dirty
3622 * and/or accessed for architectures that don't do it in hardware (most
3623 * RISC architectures). The early dirtying is also good on the i386.
3624 *
3625 * There is also a hook called "update_mmu_cache()" that architectures
3626 * with external mmu caches can use to update those (ie the Sparc or
3627 * PowerPC hashed page tables that act as extended TLBs).
3628 *
3629 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3630 * but allow concurrent faults), and pte mapped but not yet locked.
3631 * We return with mmap_sem still held, but pte unmapped and unlocked.
3632 */
3633static int handle_pte_fault(struct mm_struct *mm,
3634 struct vm_area_struct *vma, unsigned long address,
3635 pte_t *pte, pmd_t *pmd, unsigned int flags)
3636{
3637 pte_t entry;
3638 spinlock_t *ptl;
3639
3640 entry = *pte;
3641 if (!pte_present(entry)) {
3642 if (pte_none(entry)) {
3643 if (vma->vm_ops) {
3644 if (likely(vma->vm_ops->fault))
3645 return do_linear_fault(mm, vma, address,
3646 pte, pmd, flags, entry);
3647 }
3648 return do_anonymous_page(mm, vma, address,
3649 pte, pmd, flags);
3650 }
3651 if (pte_file(entry))
3652 return do_nonlinear_fault(mm, vma, address,
3653 pte, pmd, flags, entry);
3654 return do_swap_page(mm, vma, address,
3655 pte, pmd, flags, entry);
3656 }
3657
3658 if (pte_numa(entry))
3659 return do_numa_page(mm, vma, address, entry, pte, pmd);
3660
3661 ptl = pte_lockptr(mm, pmd);
3662 spin_lock(ptl);
3663 if (unlikely(!pte_same(*pte, entry)))
3664 goto unlock;
3665 if (flags & FAULT_FLAG_WRITE) {
3666 if (!pte_write(entry))
3667 return do_wp_page(mm, vma, address,
3668 pte, pmd, ptl, entry);
3669 entry = pte_mkdirty(entry);
3670 }
3671 entry = pte_mkyoung(entry);
3672 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3673 update_mmu_cache(vma, address, pte);
3674 } else {
3675 /*
3676 * This is needed only for protection faults but the arch code
3677 * is not yet telling us if this is a protection fault or not.
3678 * This still avoids useless tlb flushes for .text page faults
3679 * with threads.
3680 */
3681 if (flags & FAULT_FLAG_WRITE)
3682 flush_tlb_fix_spurious_fault(vma, address);
3683 }
3684unlock:
3685 pte_unmap_unlock(pte, ptl);
3686 return 0;
3687}
3688
3689/*
3690 * By the time we get here, we already hold the mm semaphore
3691 */
3692static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3693 unsigned long address, unsigned int flags)
3694{
3695 pgd_t *pgd;
3696 pud_t *pud;
3697 pmd_t *pmd;
3698 pte_t *pte;
3699
3700 if (unlikely(is_vm_hugetlb_page(vma)))
3701 return hugetlb_fault(mm, vma, address, flags);
3702
3703retry:
3704 pgd = pgd_offset(mm, address);
3705 pud = pud_alloc(mm, pgd, address);
3706 if (!pud)
3707 return VM_FAULT_OOM;
3708 pmd = pmd_alloc(mm, pud, address);
3709 if (!pmd)
3710 return VM_FAULT_OOM;
3711 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3712 int ret = VM_FAULT_FALLBACK;
3713 if (!vma->vm_ops)
3714 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3715 pmd, flags);
3716 if (!(ret & VM_FAULT_FALLBACK))
3717 return ret;
3718 } else {
3719 pmd_t orig_pmd = *pmd;
3720 int ret;
3721
3722 barrier();
3723 if (pmd_trans_huge(orig_pmd)) {
3724 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3725
3726 /*
3727 * If the pmd is splitting, return and retry the
3728 * the fault. Alternative: wait until the split
3729 * is done, and goto retry.
3730 */
3731 if (pmd_trans_splitting(orig_pmd))
3732 return 0;
3733
3734 if (pmd_numa(orig_pmd))
3735 return do_huge_pmd_numa_page(mm, vma, address,
3736 orig_pmd, pmd);
3737
3738 if (dirty && !pmd_write(orig_pmd)) {
3739 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3740 orig_pmd);
3741 /*
3742 * If COW results in an oom, the huge pmd will
3743 * have been split, so retry the fault on the
3744 * pte for a smaller charge.
3745 */
3746 if (unlikely(ret & VM_FAULT_OOM))
3747 goto retry;
3748 return ret;
3749 } else {
3750 huge_pmd_set_accessed(mm, vma, address, pmd,
3751 orig_pmd, dirty);
3752 }
3753
3754 return 0;
3755 }
3756 }
3757
3758 /* THP should already have been handled */
3759 BUG_ON(pmd_numa(*pmd));
3760
3761 /*
3762 * Use __pte_alloc instead of pte_alloc_map, because we can't
3763 * run pte_offset_map on the pmd, if an huge pmd could
3764 * materialize from under us from a different thread.
3765 */
3766 if (unlikely(pmd_none(*pmd)) &&
3767 unlikely(__pte_alloc(mm, vma, pmd, address)))
3768 return VM_FAULT_OOM;
3769 /* if an huge pmd materialized from under us just retry later */
3770 if (unlikely(pmd_trans_huge(*pmd)))
3771 return 0;
3772 /*
3773 * A regular pmd is established and it can't morph into a huge pmd
3774 * from under us anymore at this point because we hold the mmap_sem
3775 * read mode and khugepaged takes it in write mode. So now it's
3776 * safe to run pte_offset_map().
3777 */
3778 pte = pte_offset_map(pmd, address);
3779
3780 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3781}
3782
3783int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3784 unsigned long address, unsigned int flags)
3785{
3786 int ret;
3787
3788 __set_current_state(TASK_RUNNING);
3789
3790 count_vm_event(PGFAULT);
3791 mem_cgroup_count_vm_event(mm, PGFAULT);
3792
3793 /* do counter updates before entering really critical section. */
3794 check_sync_rss_stat(current);
3795
3796 /*
3797 * Enable the memcg OOM handling for faults triggered in user
3798 * space. Kernel faults are handled more gracefully.
3799 */
3800 if (flags & FAULT_FLAG_USER)
3801 mem_cgroup_oom_enable();
3802
3803 ret = __handle_mm_fault(mm, vma, address, flags);
3804
3805 if (flags & FAULT_FLAG_USER) {
3806 mem_cgroup_oom_disable();
3807 /*
3808 * The task may have entered a memcg OOM situation but
3809 * if the allocation error was handled gracefully (no
3810 * VM_FAULT_OOM), there is no need to kill anything.
3811 * Just clean up the OOM state peacefully.
3812 */
3813 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3814 mem_cgroup_oom_synchronize(false);
3815 }
3816
3817 return ret;
3818}
3819
3820#ifndef __PAGETABLE_PUD_FOLDED
3821/*
3822 * Allocate page upper directory.
3823 * We've already handled the fast-path in-line.
3824 */
3825int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3826{
3827 pud_t *new = pud_alloc_one(mm, address);
3828 if (!new)
3829 return -ENOMEM;
3830
3831 smp_wmb(); /* See comment in __pte_alloc */
3832
3833 spin_lock(&mm->page_table_lock);
3834 if (pgd_present(*pgd)) /* Another has populated it */
3835 pud_free(mm, new);
3836 else
3837 pgd_populate(mm, pgd, new);
3838 spin_unlock(&mm->page_table_lock);
3839 return 0;
3840}
3841#endif /* __PAGETABLE_PUD_FOLDED */
3842
3843#ifndef __PAGETABLE_PMD_FOLDED
3844/*
3845 * Allocate page middle directory.
3846 * We've already handled the fast-path in-line.
3847 */
3848int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3849{
3850 pmd_t *new = pmd_alloc_one(mm, address);
3851 if (!new)
3852 return -ENOMEM;
3853
3854 smp_wmb(); /* See comment in __pte_alloc */
3855
3856 spin_lock(&mm->page_table_lock);
3857#ifndef __ARCH_HAS_4LEVEL_HACK
3858 if (pud_present(*pud)) /* Another has populated it */
3859 pmd_free(mm, new);
3860 else
3861 pud_populate(mm, pud, new);
3862#else
3863 if (pgd_present(*pud)) /* Another has populated it */
3864 pmd_free(mm, new);
3865 else
3866 pgd_populate(mm, pud, new);
3867#endif /* __ARCH_HAS_4LEVEL_HACK */
3868 spin_unlock(&mm->page_table_lock);
3869 return 0;
3870}
3871#endif /* __PAGETABLE_PMD_FOLDED */
3872
3873#if !defined(__HAVE_ARCH_GATE_AREA)
3874
3875#if defined(AT_SYSINFO_EHDR)
3876static struct vm_area_struct gate_vma;
3877
3878static int __init gate_vma_init(void)
3879{
3880 gate_vma.vm_mm = NULL;
3881 gate_vma.vm_start = FIXADDR_USER_START;
3882 gate_vma.vm_end = FIXADDR_USER_END;
3883 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3884 gate_vma.vm_page_prot = __P101;
3885
3886 return 0;
3887}
3888__initcall(gate_vma_init);
3889#endif
3890
3891struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3892{
3893#ifdef AT_SYSINFO_EHDR
3894 return &gate_vma;
3895#else
3896 return NULL;
3897#endif
3898}
3899
3900int in_gate_area_no_mm(unsigned long addr)
3901{
3902#ifdef AT_SYSINFO_EHDR
3903 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3904 return 1;
3905#endif
3906 return 0;
3907}
3908
3909#endif /* __HAVE_ARCH_GATE_AREA */
3910
3911static int __follow_pte(struct mm_struct *mm, unsigned long address,
3912 pte_t **ptepp, spinlock_t **ptlp)
3913{
3914 pgd_t *pgd;
3915 pud_t *pud;
3916 pmd_t *pmd;
3917 pte_t *ptep;
3918
3919 pgd = pgd_offset(mm, address);
3920 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3921 goto out;
3922
3923 pud = pud_offset(pgd, address);
3924 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3925 goto out;
3926
3927 pmd = pmd_offset(pud, address);
3928 VM_BUG_ON(pmd_trans_huge(*pmd));
3929 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3930 goto out;
3931
3932 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3933 if (pmd_huge(*pmd))
3934 goto out;
3935
3936 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3937 if (!ptep)
3938 goto out;
3939 if (!pte_present(*ptep))
3940 goto unlock;
3941 *ptepp = ptep;
3942 return 0;
3943unlock:
3944 pte_unmap_unlock(ptep, *ptlp);
3945out:
3946 return -EINVAL;
3947}
3948
3949static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3950 pte_t **ptepp, spinlock_t **ptlp)
3951{
3952 int res;
3953
3954 /* (void) is needed to make gcc happy */
3955 (void) __cond_lock(*ptlp,
3956 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3957 return res;
3958}
3959
3960/**
3961 * follow_pfn - look up PFN at a user virtual address
3962 * @vma: memory mapping
3963 * @address: user virtual address
3964 * @pfn: location to store found PFN
3965 *
3966 * Only IO mappings and raw PFN mappings are allowed.
3967 *
3968 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3969 */
3970int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3971 unsigned long *pfn)
3972{
3973 int ret = -EINVAL;
3974 spinlock_t *ptl;
3975 pte_t *ptep;
3976
3977 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3978 return ret;
3979
3980 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3981 if (ret)
3982 return ret;
3983 *pfn = pte_pfn(*ptep);
3984 pte_unmap_unlock(ptep, ptl);
3985 return 0;
3986}
3987EXPORT_SYMBOL(follow_pfn);
3988
3989#ifdef CONFIG_HAVE_IOREMAP_PROT
3990int follow_phys(struct vm_area_struct *vma,
3991 unsigned long address, unsigned int flags,
3992 unsigned long *prot, resource_size_t *phys)
3993{
3994 int ret = -EINVAL;
3995 pte_t *ptep, pte;
3996 spinlock_t *ptl;
3997
3998 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3999 goto out;
4000
4001 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4002 goto out;
4003 pte = *ptep;
4004
4005 if ((flags & FOLL_WRITE) && !pte_write(pte))
4006 goto unlock;
4007
4008 *prot = pgprot_val(pte_pgprot(pte));
4009 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4010
4011 ret = 0;
4012unlock:
4013 pte_unmap_unlock(ptep, ptl);
4014out:
4015 return ret;
4016}
4017
4018int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4019 void *buf, int len, int write)
4020{
4021 resource_size_t phys_addr;
4022 unsigned long prot = 0;
4023 void __iomem *maddr;
4024 int offset = addr & (PAGE_SIZE-1);
4025
4026 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4027 return -EINVAL;
4028
4029 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4030 if (write)
4031 memcpy_toio(maddr + offset, buf, len);
4032 else
4033 memcpy_fromio(buf, maddr + offset, len);
4034 iounmap(maddr);
4035
4036 return len;
4037}
4038EXPORT_SYMBOL_GPL(generic_access_phys);
4039#endif
4040
4041/*
4042 * Access another process' address space as given in mm. If non-NULL, use the
4043 * given task for page fault accounting.
4044 */
4045static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4046 unsigned long addr, void *buf, int len, int write)
4047{
4048 struct vm_area_struct *vma;
4049 void *old_buf = buf;
4050
4051 down_read(&mm->mmap_sem);
4052 /* ignore errors, just check how much was successfully transferred */
4053 while (len) {
4054 int bytes, ret, offset;
4055 void *maddr;
4056 struct page *page = NULL;
4057
4058 ret = get_user_pages(tsk, mm, addr, 1,
4059 write, 1, &page, &vma);
4060 if (ret <= 0) {
4061 /*
4062 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4063 * we can access using slightly different code.
4064 */
4065#ifdef CONFIG_HAVE_IOREMAP_PROT
4066 vma = find_vma(mm, addr);
4067 if (!vma || vma->vm_start > addr)
4068 break;
4069 if (vma->vm_ops && vma->vm_ops->access)
4070 ret = vma->vm_ops->access(vma, addr, buf,
4071 len, write);
4072 if (ret <= 0)
4073#endif
4074 break;
4075 bytes = ret;
4076 } else {
4077 bytes = len;
4078 offset = addr & (PAGE_SIZE-1);
4079 if (bytes > PAGE_SIZE-offset)
4080 bytes = PAGE_SIZE-offset;
4081
4082 maddr = kmap(page);
4083 if (write) {
4084 copy_to_user_page(vma, page, addr,
4085 maddr + offset, buf, bytes);
4086 set_page_dirty_lock(page);
4087 } else {
4088 copy_from_user_page(vma, page, addr,
4089 buf, maddr + offset, bytes);
4090 }
4091 kunmap(page);
4092 page_cache_release(page);
4093 }
4094 len -= bytes;
4095 buf += bytes;
4096 addr += bytes;
4097 }
4098 up_read(&mm->mmap_sem);
4099
4100 return buf - old_buf;
4101}
4102
4103/**
4104 * access_remote_vm - access another process' address space
4105 * @mm: the mm_struct of the target address space
4106 * @addr: start address to access
4107 * @buf: source or destination buffer
4108 * @len: number of bytes to transfer
4109 * @write: whether the access is a write
4110 *
4111 * The caller must hold a reference on @mm.
4112 */
4113int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4114 void *buf, int len, int write)
4115{
4116 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4117}
4118
4119/*
4120 * Access another process' address space.
4121 * Source/target buffer must be kernel space,
4122 * Do not walk the page table directly, use get_user_pages
4123 */
4124int access_process_vm(struct task_struct *tsk, unsigned long addr,
4125 void *buf, int len, int write)
4126{
4127 struct mm_struct *mm;
4128 int ret;
4129
4130 mm = get_task_mm(tsk);
4131 if (!mm)
4132 return 0;
4133
4134 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4135 mmput(mm);
4136
4137 return ret;
4138}
4139
4140/*
4141 * Print the name of a VMA.
4142 */
4143void print_vma_addr(char *prefix, unsigned long ip)
4144{
4145 struct mm_struct *mm = current->mm;
4146 struct vm_area_struct *vma;
4147
4148 /*
4149 * Do not print if we are in atomic
4150 * contexts (in exception stacks, etc.):
4151 */
4152 if (preempt_count())
4153 return;
4154
4155 down_read(&mm->mmap_sem);
4156 vma = find_vma(mm, ip);
4157 if (vma && vma->vm_file) {
4158 struct file *f = vma->vm_file;
4159 char *buf = (char *)__get_free_page(GFP_KERNEL);
4160 if (buf) {
4161 char *p;
4162
4163 p = d_path(&f->f_path, buf, PAGE_SIZE);
4164 if (IS_ERR(p))
4165 p = "?";
4166 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4167 vma->vm_start,
4168 vma->vm_end - vma->vm_start);
4169 free_page((unsigned long)buf);
4170 }
4171 }
4172 up_read(&mm->mmap_sem);
4173}
4174
4175#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4176void might_fault(void)
4177{
4178 /*
4179 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4180 * holding the mmap_sem, this is safe because kernel memory doesn't
4181 * get paged out, therefore we'll never actually fault, and the
4182 * below annotations will generate false positives.
4183 */
4184 if (segment_eq(get_fs(), KERNEL_DS))
4185 return;
4186
4187 /*
4188 * it would be nicer only to annotate paths which are not under
4189 * pagefault_disable, however that requires a larger audit and
4190 * providing helpers like get_user_atomic.
4191 */
4192 if (in_atomic())
4193 return;
4194
4195 __might_sleep(__FILE__, __LINE__, 0);
4196
4197 if (current->mm)
4198 might_lock_read(&current->mm->mmap_sem);
4199}
4200EXPORT_SYMBOL(might_fault);
4201#endif
4202
4203#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4204static void clear_gigantic_page(struct page *page,
4205 unsigned long addr,
4206 unsigned int pages_per_huge_page)
4207{
4208 int i;
4209 struct page *p = page;
4210
4211 might_sleep();
4212 for (i = 0; i < pages_per_huge_page;
4213 i++, p = mem_map_next(p, page, i)) {
4214 cond_resched();
4215 clear_user_highpage(p, addr + i * PAGE_SIZE);
4216 }
4217}
4218void clear_huge_page(struct page *page,
4219 unsigned long addr, unsigned int pages_per_huge_page)
4220{
4221 int i;
4222
4223 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4224 clear_gigantic_page(page, addr, pages_per_huge_page);
4225 return;
4226 }
4227
4228 might_sleep();
4229 for (i = 0; i < pages_per_huge_page; i++) {
4230 cond_resched();
4231 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4232 }
4233}
4234
4235static void copy_user_gigantic_page(struct page *dst, struct page *src,
4236 unsigned long addr,
4237 struct vm_area_struct *vma,
4238 unsigned int pages_per_huge_page)
4239{
4240 int i;
4241 struct page *dst_base = dst;
4242 struct page *src_base = src;
4243
4244 for (i = 0; i < pages_per_huge_page; ) {
4245 cond_resched();
4246 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4247
4248 i++;
4249 dst = mem_map_next(dst, dst_base, i);
4250 src = mem_map_next(src, src_base, i);
4251 }
4252}
4253
4254void copy_user_huge_page(struct page *dst, struct page *src,
4255 unsigned long addr, struct vm_area_struct *vma,
4256 unsigned int pages_per_huge_page)
4257{
4258 int i;
4259
4260 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4261 copy_user_gigantic_page(dst, src, addr, vma,
4262 pages_per_huge_page);
4263 return;
4264 }
4265
4266 might_sleep();
4267 for (i = 0; i < pages_per_huge_page; i++) {
4268 cond_resched();
4269 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4270 }
4271}
4272#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4273
4274#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4275bool ptlock_alloc(struct page *page)
4276{
4277 spinlock_t *ptl;
4278
4279 ptl = kmalloc(sizeof(spinlock_t), GFP_KERNEL);
4280 if (!ptl)
4281 return false;
4282 page->ptl = ptl;
4283 return true;
4284}
4285
4286void ptlock_free(struct page *page)
4287{
4288 kfree(page->ptl);
4289}
4290#endif