]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blame - mm/memory.c
mei: me: fix hardware reset flow
[mirror_ubuntu-zesty-kernel.git] / mm / memory.c
CommitLineData
1da177e4
LT
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 (alex@cconcepts.co.uk)
34 *
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
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>
9a840895 48#include <linux/ksm.h>
1da177e4 49#include <linux/rmap.h>
b95f1b31 50#include <linux/export.h>
0ff92245 51#include <linux/delayacct.h>
1da177e4 52#include <linux/init.h>
edc79b2a 53#include <linux/writeback.h>
8a9f3ccd 54#include <linux/memcontrol.h>
cddb8a5c 55#include <linux/mmu_notifier.h>
3dc14741
HD
56#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
5a0e3ad6 59#include <linux/gfp.h>
4daae3b4 60#include <linux/migrate.h>
2fbc57c5 61#include <linux/string.h>
1da177e4 62
6952b61d 63#include <asm/io.h>
1da177e4
LT
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
42b77728
JB
70#include "internal.h"
71
75980e97
PZ
72#ifdef LAST_NID_NOT_IN_PAGE_FLAGS
73#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
74#endif
75
d41dee36 76#ifndef CONFIG_NEED_MULTIPLE_NODES
1da177e4
LT
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
1da177e4
LT
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;
1da177e4 93
1da177e4 94EXPORT_SYMBOL(high_memory);
1da177e4 95
32a93233
IM
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
a62eaf15
AK
108
109static int __init disable_randmaps(char *s)
110{
111 randomize_va_space = 0;
9b41046c 112 return 1;
a62eaf15
AK
113}
114__setup("norandmaps", disable_randmaps);
115
62eede62 116unsigned long zero_pfn __read_mostly;
03f6462a 117unsigned long highest_memmap_pfn __read_mostly;
a13ea5b7
HD
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);
a62eaf15 128
d559db08 129
34e55232
KH
130#if defined(SPLIT_RSS_COUNTING)
131
ea48cf78 132void sync_mm_rss(struct mm_struct *mm)
34e55232
KH
133{
134 int i;
135
136 for (i = 0; i < NR_MM_COUNTERS; i++) {
05af2e10
DR
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;
34e55232
KH
140 }
141 }
05af2e10 142 current->rss_stat.events = 0;
34e55232
KH
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))
ea48cf78 164 sync_mm_rss(task->mm);
34e55232 165}
9547d01b 166#else /* SPLIT_RSS_COUNTING */
34e55232
KH
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
9547d01b
PZ
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
53a59fc6
MH
189 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
190 return 0;
191
9547d01b
PZ
192 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
193 if (!batch)
194 return 0;
195
53a59fc6 196 tlb->batch_count++;
9547d01b
PZ
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, bool fullmm)
213{
214 tlb->mm = mm;
215
216 tlb->fullmm = fullmm;
1de14c3c 217 tlb->need_flush_all = 0;
597e1c35
AS
218 tlb->start = -1UL;
219 tlb->end = 0;
9547d01b 220 tlb->need_flush = 0;
9547d01b
PZ
221 tlb->local.next = NULL;
222 tlb->local.nr = 0;
223 tlb->local.max = ARRAY_SIZE(tlb->__pages);
224 tlb->active = &tlb->local;
53a59fc6 225 tlb->batch_count = 0;
9547d01b
PZ
226
227#ifdef CONFIG_HAVE_RCU_TABLE_FREE
228 tlb->batch = NULL;
229#endif
230}
231
232void tlb_flush_mmu(struct mmu_gather *tlb)
233{
234 struct mmu_gather_batch *batch;
235
236 if (!tlb->need_flush)
237 return;
238 tlb->need_flush = 0;
239 tlb_flush(tlb);
240#ifdef CONFIG_HAVE_RCU_TABLE_FREE
241 tlb_table_flush(tlb);
34e55232
KH
242#endif
243
9547d01b
PZ
244 for (batch = &tlb->local; batch; batch = batch->next) {
245 free_pages_and_swap_cache(batch->pages, batch->nr);
246 batch->nr = 0;
247 }
248 tlb->active = &tlb->local;
249}
250
251/* tlb_finish_mmu
252 * Called at the end of the shootdown operation to free up any resources
253 * that were required.
254 */
255void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
256{
257 struct mmu_gather_batch *batch, *next;
258
597e1c35
AS
259 tlb->start = start;
260 tlb->end = end;
9547d01b
PZ
261 tlb_flush_mmu(tlb);
262
263 /* keep the page table cache within bounds */
264 check_pgt_cache();
265
266 for (batch = tlb->local.next; batch; batch = next) {
267 next = batch->next;
268 free_pages((unsigned long)batch, 0);
269 }
270 tlb->local.next = NULL;
271}
272
273/* __tlb_remove_page
274 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
275 * handling the additional races in SMP caused by other CPUs caching valid
276 * mappings in their TLBs. Returns the number of free page slots left.
277 * When out of page slots we must call tlb_flush_mmu().
278 */
279int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
280{
281 struct mmu_gather_batch *batch;
282
f21760b1 283 VM_BUG_ON(!tlb->need_flush);
9547d01b 284
9547d01b
PZ
285 batch = tlb->active;
286 batch->pages[batch->nr++] = page;
287 if (batch->nr == batch->max) {
288 if (!tlb_next_batch(tlb))
289 return 0;
0b43c3aa 290 batch = tlb->active;
9547d01b
PZ
291 }
292 VM_BUG_ON(batch->nr > batch->max);
293
294 return batch->max - batch->nr;
295}
296
297#endif /* HAVE_GENERIC_MMU_GATHER */
298
26723911
PZ
299#ifdef CONFIG_HAVE_RCU_TABLE_FREE
300
301/*
302 * See the comment near struct mmu_table_batch.
303 */
304
305static void tlb_remove_table_smp_sync(void *arg)
306{
307 /* Simply deliver the interrupt */
308}
309
310static void tlb_remove_table_one(void *table)
311{
312 /*
313 * This isn't an RCU grace period and hence the page-tables cannot be
314 * assumed to be actually RCU-freed.
315 *
316 * It is however sufficient for software page-table walkers that rely on
317 * IRQ disabling. See the comment near struct mmu_table_batch.
318 */
319 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
320 __tlb_remove_table(table);
321}
322
323static void tlb_remove_table_rcu(struct rcu_head *head)
324{
325 struct mmu_table_batch *batch;
326 int i;
327
328 batch = container_of(head, struct mmu_table_batch, rcu);
329
330 for (i = 0; i < batch->nr; i++)
331 __tlb_remove_table(batch->tables[i]);
332
333 free_page((unsigned long)batch);
334}
335
336void tlb_table_flush(struct mmu_gather *tlb)
337{
338 struct mmu_table_batch **batch = &tlb->batch;
339
340 if (*batch) {
341 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
342 *batch = NULL;
343 }
344}
345
346void tlb_remove_table(struct mmu_gather *tlb, void *table)
347{
348 struct mmu_table_batch **batch = &tlb->batch;
349
350 tlb->need_flush = 1;
351
352 /*
353 * When there's less then two users of this mm there cannot be a
354 * concurrent page-table walk.
355 */
356 if (atomic_read(&tlb->mm->mm_users) < 2) {
357 __tlb_remove_table(table);
358 return;
359 }
360
361 if (*batch == NULL) {
362 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
363 if (*batch == NULL) {
364 tlb_remove_table_one(table);
365 return;
366 }
367 (*batch)->nr = 0;
368 }
369 (*batch)->tables[(*batch)->nr++] = table;
370 if ((*batch)->nr == MAX_TABLE_BATCH)
371 tlb_table_flush(tlb);
372}
373
9547d01b 374#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
26723911 375
1da177e4
LT
376/*
377 * If a p?d_bad entry is found while walking page tables, report
378 * the error, before resetting entry to p?d_none. Usually (but
379 * very seldom) called out from the p?d_none_or_clear_bad macros.
380 */
381
382void pgd_clear_bad(pgd_t *pgd)
383{
384 pgd_ERROR(*pgd);
385 pgd_clear(pgd);
386}
387
388void pud_clear_bad(pud_t *pud)
389{
390 pud_ERROR(*pud);
391 pud_clear(pud);
392}
393
394void pmd_clear_bad(pmd_t *pmd)
395{
396 pmd_ERROR(*pmd);
397 pmd_clear(pmd);
398}
399
400/*
401 * Note: this doesn't free the actual pages themselves. That
402 * has been handled earlier when unmapping all the memory regions.
403 */
9e1b32ca
BH
404static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
405 unsigned long addr)
1da177e4 406{
2f569afd 407 pgtable_t token = pmd_pgtable(*pmd);
e0da382c 408 pmd_clear(pmd);
9e1b32ca 409 pte_free_tlb(tlb, token, addr);
e0da382c 410 tlb->mm->nr_ptes--;
1da177e4
LT
411}
412
e0da382c
HD
413static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
414 unsigned long addr, unsigned long end,
415 unsigned long floor, unsigned long ceiling)
1da177e4
LT
416{
417 pmd_t *pmd;
418 unsigned long next;
e0da382c 419 unsigned long start;
1da177e4 420
e0da382c 421 start = addr;
1da177e4 422 pmd = pmd_offset(pud, addr);
1da177e4
LT
423 do {
424 next = pmd_addr_end(addr, end);
425 if (pmd_none_or_clear_bad(pmd))
426 continue;
9e1b32ca 427 free_pte_range(tlb, pmd, addr);
1da177e4
LT
428 } while (pmd++, addr = next, addr != end);
429
e0da382c
HD
430 start &= PUD_MASK;
431 if (start < floor)
432 return;
433 if (ceiling) {
434 ceiling &= PUD_MASK;
435 if (!ceiling)
436 return;
1da177e4 437 }
e0da382c
HD
438 if (end - 1 > ceiling - 1)
439 return;
440
441 pmd = pmd_offset(pud, start);
442 pud_clear(pud);
9e1b32ca 443 pmd_free_tlb(tlb, pmd, start);
1da177e4
LT
444}
445
e0da382c
HD
446static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
447 unsigned long addr, unsigned long end,
448 unsigned long floor, unsigned long ceiling)
1da177e4
LT
449{
450 pud_t *pud;
451 unsigned long next;
e0da382c 452 unsigned long start;
1da177e4 453
e0da382c 454 start = addr;
1da177e4 455 pud = pud_offset(pgd, addr);
1da177e4
LT
456 do {
457 next = pud_addr_end(addr, end);
458 if (pud_none_or_clear_bad(pud))
459 continue;
e0da382c 460 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
1da177e4
LT
461 } while (pud++, addr = next, addr != end);
462
e0da382c
HD
463 start &= PGDIR_MASK;
464 if (start < floor)
465 return;
466 if (ceiling) {
467 ceiling &= PGDIR_MASK;
468 if (!ceiling)
469 return;
1da177e4 470 }
e0da382c
HD
471 if (end - 1 > ceiling - 1)
472 return;
473
474 pud = pud_offset(pgd, start);
475 pgd_clear(pgd);
9e1b32ca 476 pud_free_tlb(tlb, pud, start);
1da177e4
LT
477}
478
479/*
e0da382c
HD
480 * This function frees user-level page tables of a process.
481 *
1da177e4
LT
482 * Must be called with pagetable lock held.
483 */
42b77728 484void free_pgd_range(struct mmu_gather *tlb,
e0da382c
HD
485 unsigned long addr, unsigned long end,
486 unsigned long floor, unsigned long ceiling)
1da177e4
LT
487{
488 pgd_t *pgd;
489 unsigned long next;
e0da382c
HD
490
491 /*
492 * The next few lines have given us lots of grief...
493 *
494 * Why are we testing PMD* at this top level? Because often
495 * there will be no work to do at all, and we'd prefer not to
496 * go all the way down to the bottom just to discover that.
497 *
498 * Why all these "- 1"s? Because 0 represents both the bottom
499 * of the address space and the top of it (using -1 for the
500 * top wouldn't help much: the masks would do the wrong thing).
501 * The rule is that addr 0 and floor 0 refer to the bottom of
502 * the address space, but end 0 and ceiling 0 refer to the top
503 * Comparisons need to use "end - 1" and "ceiling - 1" (though
504 * that end 0 case should be mythical).
505 *
506 * Wherever addr is brought up or ceiling brought down, we must
507 * be careful to reject "the opposite 0" before it confuses the
508 * subsequent tests. But what about where end is brought down
509 * by PMD_SIZE below? no, end can't go down to 0 there.
510 *
511 * Whereas we round start (addr) and ceiling down, by different
512 * masks at different levels, in order to test whether a table
513 * now has no other vmas using it, so can be freed, we don't
514 * bother to round floor or end up - the tests don't need that.
515 */
1da177e4 516
e0da382c
HD
517 addr &= PMD_MASK;
518 if (addr < floor) {
519 addr += PMD_SIZE;
520 if (!addr)
521 return;
522 }
523 if (ceiling) {
524 ceiling &= PMD_MASK;
525 if (!ceiling)
526 return;
527 }
528 if (end - 1 > ceiling - 1)
529 end -= PMD_SIZE;
530 if (addr > end - 1)
531 return;
532
42b77728 533 pgd = pgd_offset(tlb->mm, addr);
1da177e4
LT
534 do {
535 next = pgd_addr_end(addr, end);
536 if (pgd_none_or_clear_bad(pgd))
537 continue;
42b77728 538 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
1da177e4 539 } while (pgd++, addr = next, addr != end);
e0da382c
HD
540}
541
42b77728 542void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
3bf5ee95 543 unsigned long floor, unsigned long ceiling)
e0da382c
HD
544{
545 while (vma) {
546 struct vm_area_struct *next = vma->vm_next;
547 unsigned long addr = vma->vm_start;
548
8f4f8c16 549 /*
25d9e2d1 550 * Hide vma from rmap and truncate_pagecache before freeing
551 * pgtables
8f4f8c16 552 */
5beb4930 553 unlink_anon_vmas(vma);
8f4f8c16
HD
554 unlink_file_vma(vma);
555
9da61aef 556 if (is_vm_hugetlb_page(vma)) {
3bf5ee95 557 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
e0da382c 558 floor, next? next->vm_start: ceiling);
3bf5ee95
HD
559 } else {
560 /*
561 * Optimization: gather nearby vmas into one call down
562 */
563 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
4866920b 564 && !is_vm_hugetlb_page(next)) {
3bf5ee95
HD
565 vma = next;
566 next = vma->vm_next;
5beb4930 567 unlink_anon_vmas(vma);
8f4f8c16 568 unlink_file_vma(vma);
3bf5ee95
HD
569 }
570 free_pgd_range(tlb, addr, vma->vm_end,
571 floor, next? next->vm_start: ceiling);
572 }
e0da382c
HD
573 vma = next;
574 }
1da177e4
LT
575}
576
8ac1f832
AA
577int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
578 pmd_t *pmd, unsigned long address)
1da177e4 579{
2f569afd 580 pgtable_t new = pte_alloc_one(mm, address);
8ac1f832 581 int wait_split_huge_page;
1bb3630e
HD
582 if (!new)
583 return -ENOMEM;
584
362a61ad
NP
585 /*
586 * Ensure all pte setup (eg. pte page lock and page clearing) are
587 * visible before the pte is made visible to other CPUs by being
588 * put into page tables.
589 *
590 * The other side of the story is the pointer chasing in the page
591 * table walking code (when walking the page table without locking;
592 * ie. most of the time). Fortunately, these data accesses consist
593 * of a chain of data-dependent loads, meaning most CPUs (alpha
594 * being the notable exception) will already guarantee loads are
595 * seen in-order. See the alpha page table accessors for the
596 * smp_read_barrier_depends() barriers in page table walking code.
597 */
598 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
599
c74df32c 600 spin_lock(&mm->page_table_lock);
8ac1f832
AA
601 wait_split_huge_page = 0;
602 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
1da177e4 603 mm->nr_ptes++;
1da177e4 604 pmd_populate(mm, pmd, new);
2f569afd 605 new = NULL;
8ac1f832
AA
606 } else if (unlikely(pmd_trans_splitting(*pmd)))
607 wait_split_huge_page = 1;
c74df32c 608 spin_unlock(&mm->page_table_lock);
2f569afd
MS
609 if (new)
610 pte_free(mm, new);
8ac1f832
AA
611 if (wait_split_huge_page)
612 wait_split_huge_page(vma->anon_vma, pmd);
1bb3630e 613 return 0;
1da177e4
LT
614}
615
1bb3630e 616int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
1da177e4 617{
1bb3630e
HD
618 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
619 if (!new)
620 return -ENOMEM;
621
362a61ad
NP
622 smp_wmb(); /* See comment in __pte_alloc */
623
1bb3630e 624 spin_lock(&init_mm.page_table_lock);
8ac1f832 625 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
1bb3630e 626 pmd_populate_kernel(&init_mm, pmd, new);
2f569afd 627 new = NULL;
8ac1f832
AA
628 } else
629 VM_BUG_ON(pmd_trans_splitting(*pmd));
1bb3630e 630 spin_unlock(&init_mm.page_table_lock);
2f569afd
MS
631 if (new)
632 pte_free_kernel(&init_mm, new);
1bb3630e 633 return 0;
1da177e4
LT
634}
635
d559db08
KH
636static inline void init_rss_vec(int *rss)
637{
638 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
639}
640
641static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
ae859762 642{
d559db08
KH
643 int i;
644
34e55232 645 if (current->mm == mm)
05af2e10 646 sync_mm_rss(mm);
d559db08
KH
647 for (i = 0; i < NR_MM_COUNTERS; i++)
648 if (rss[i])
649 add_mm_counter(mm, i, rss[i]);
ae859762
HD
650}
651
b5810039 652/*
6aab341e
LT
653 * This function is called to print an error when a bad pte
654 * is found. For example, we might have a PFN-mapped pte in
655 * a region that doesn't allow it.
b5810039
NP
656 *
657 * The calling function must still handle the error.
658 */
3dc14741
HD
659static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
660 pte_t pte, struct page *page)
b5810039 661{
3dc14741
HD
662 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
663 pud_t *pud = pud_offset(pgd, addr);
664 pmd_t *pmd = pmd_offset(pud, addr);
665 struct address_space *mapping;
666 pgoff_t index;
d936cf9b
HD
667 static unsigned long resume;
668 static unsigned long nr_shown;
669 static unsigned long nr_unshown;
670
671 /*
672 * Allow a burst of 60 reports, then keep quiet for that minute;
673 * or allow a steady drip of one report per second.
674 */
675 if (nr_shown == 60) {
676 if (time_before(jiffies, resume)) {
677 nr_unshown++;
678 return;
679 }
680 if (nr_unshown) {
1e9e6365
HD
681 printk(KERN_ALERT
682 "BUG: Bad page map: %lu messages suppressed\n",
d936cf9b
HD
683 nr_unshown);
684 nr_unshown = 0;
685 }
686 nr_shown = 0;
687 }
688 if (nr_shown++ == 0)
689 resume = jiffies + 60 * HZ;
3dc14741
HD
690
691 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
692 index = linear_page_index(vma, addr);
693
1e9e6365
HD
694 printk(KERN_ALERT
695 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
3dc14741
HD
696 current->comm,
697 (long long)pte_val(pte), (long long)pmd_val(*pmd));
718a3821
WF
698 if (page)
699 dump_page(page);
1e9e6365 700 printk(KERN_ALERT
3dc14741
HD
701 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
702 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
703 /*
704 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
705 */
706 if (vma->vm_ops)
071361d3
JP
707 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
708 vma->vm_ops->fault);
3dc14741 709 if (vma->vm_file && vma->vm_file->f_op)
071361d3
JP
710 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
711 vma->vm_file->f_op->mmap);
b5810039 712 dump_stack();
373d4d09 713 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
b5810039
NP
714}
715
2ec74c3e 716static inline bool is_cow_mapping(vm_flags_t flags)
67121172
LT
717{
718 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
719}
720
ee498ed7 721/*
7e675137 722 * vm_normal_page -- This function gets the "struct page" associated with a pte.
6aab341e 723 *
7e675137
NP
724 * "Special" mappings do not wish to be associated with a "struct page" (either
725 * it doesn't exist, or it exists but they don't want to touch it). In this
726 * case, NULL is returned here. "Normal" mappings do have a struct page.
b379d790 727 *
7e675137
NP
728 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
729 * pte bit, in which case this function is trivial. Secondly, an architecture
730 * may not have a spare pte bit, which requires a more complicated scheme,
731 * described below.
732 *
733 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
734 * special mapping (even if there are underlying and valid "struct pages").
735 * COWed pages of a VM_PFNMAP are always normal.
6aab341e 736 *
b379d790
JH
737 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
738 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
7e675137
NP
739 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
740 * mapping will always honor the rule
6aab341e
LT
741 *
742 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
743 *
7e675137
NP
744 * And for normal mappings this is false.
745 *
746 * This restricts such mappings to be a linear translation from virtual address
747 * to pfn. To get around this restriction, we allow arbitrary mappings so long
748 * as the vma is not a COW mapping; in that case, we know that all ptes are
749 * special (because none can have been COWed).
b379d790 750 *
b379d790 751 *
7e675137 752 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
b379d790
JH
753 *
754 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
755 * page" backing, however the difference is that _all_ pages with a struct
756 * page (that is, those where pfn_valid is true) are refcounted and considered
757 * normal pages by the VM. The disadvantage is that pages are refcounted
758 * (which can be slower and simply not an option for some PFNMAP users). The
759 * advantage is that we don't have to follow the strict linearity rule of
760 * PFNMAP mappings in order to support COWable mappings.
761 *
ee498ed7 762 */
7e675137
NP
763#ifdef __HAVE_ARCH_PTE_SPECIAL
764# define HAVE_PTE_SPECIAL 1
765#else
766# define HAVE_PTE_SPECIAL 0
767#endif
768struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
769 pte_t pte)
ee498ed7 770{
22b31eec 771 unsigned long pfn = pte_pfn(pte);
7e675137
NP
772
773 if (HAVE_PTE_SPECIAL) {
22b31eec
HD
774 if (likely(!pte_special(pte)))
775 goto check_pfn;
a13ea5b7
HD
776 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
777 return NULL;
62eede62 778 if (!is_zero_pfn(pfn))
22b31eec 779 print_bad_pte(vma, addr, pte, NULL);
7e675137
NP
780 return NULL;
781 }
782
783 /* !HAVE_PTE_SPECIAL case follows: */
784
b379d790
JH
785 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
786 if (vma->vm_flags & VM_MIXEDMAP) {
787 if (!pfn_valid(pfn))
788 return NULL;
789 goto out;
790 } else {
7e675137
NP
791 unsigned long off;
792 off = (addr - vma->vm_start) >> PAGE_SHIFT;
b379d790
JH
793 if (pfn == vma->vm_pgoff + off)
794 return NULL;
795 if (!is_cow_mapping(vma->vm_flags))
796 return NULL;
797 }
6aab341e
LT
798 }
799
62eede62
HD
800 if (is_zero_pfn(pfn))
801 return NULL;
22b31eec
HD
802check_pfn:
803 if (unlikely(pfn > highest_memmap_pfn)) {
804 print_bad_pte(vma, addr, pte, NULL);
805 return NULL;
806 }
6aab341e
LT
807
808 /*
7e675137 809 * NOTE! We still have PageReserved() pages in the page tables.
7e675137 810 * eg. VDSO mappings can cause them to exist.
6aab341e 811 */
b379d790 812out:
6aab341e 813 return pfn_to_page(pfn);
ee498ed7
HD
814}
815
1da177e4
LT
816/*
817 * copy one vm_area from one task to the other. Assumes the page tables
818 * already present in the new task to be cleared in the whole range
819 * covered by this vma.
1da177e4
LT
820 */
821
570a335b 822static inline unsigned long
1da177e4 823copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
b5810039 824 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
8c103762 825 unsigned long addr, int *rss)
1da177e4 826{
b5810039 827 unsigned long vm_flags = vma->vm_flags;
1da177e4
LT
828 pte_t pte = *src_pte;
829 struct page *page;
1da177e4
LT
830
831 /* pte contains position in swap or file, so copy. */
832 if (unlikely(!pte_present(pte))) {
833 if (!pte_file(pte)) {
0697212a
CL
834 swp_entry_t entry = pte_to_swp_entry(pte);
835
570a335b
HD
836 if (swap_duplicate(entry) < 0)
837 return entry.val;
838
1da177e4
LT
839 /* make sure dst_mm is on swapoff's mmlist. */
840 if (unlikely(list_empty(&dst_mm->mmlist))) {
841 spin_lock(&mmlist_lock);
f412ac08
HD
842 if (list_empty(&dst_mm->mmlist))
843 list_add(&dst_mm->mmlist,
844 &src_mm->mmlist);
1da177e4
LT
845 spin_unlock(&mmlist_lock);
846 }
b084d435
KH
847 if (likely(!non_swap_entry(entry)))
848 rss[MM_SWAPENTS]++;
9f9f1acd
KK
849 else if (is_migration_entry(entry)) {
850 page = migration_entry_to_page(entry);
851
852 if (PageAnon(page))
853 rss[MM_ANONPAGES]++;
854 else
855 rss[MM_FILEPAGES]++;
856
857 if (is_write_migration_entry(entry) &&
858 is_cow_mapping(vm_flags)) {
859 /*
860 * COW mappings require pages in both
861 * parent and child to be set to read.
862 */
863 make_migration_entry_read(&entry);
864 pte = swp_entry_to_pte(entry);
865 set_pte_at(src_mm, addr, src_pte, pte);
866 }
0697212a 867 }
1da177e4 868 }
ae859762 869 goto out_set_pte;
1da177e4
LT
870 }
871
1da177e4
LT
872 /*
873 * If it's a COW mapping, write protect it both
874 * in the parent and the child
875 */
67121172 876 if (is_cow_mapping(vm_flags)) {
1da177e4 877 ptep_set_wrprotect(src_mm, addr, src_pte);
3dc90795 878 pte = pte_wrprotect(pte);
1da177e4
LT
879 }
880
881 /*
882 * If it's a shared mapping, mark it clean in
883 * the child
884 */
885 if (vm_flags & VM_SHARED)
886 pte = pte_mkclean(pte);
887 pte = pte_mkold(pte);
6aab341e
LT
888
889 page = vm_normal_page(vma, addr, pte);
890 if (page) {
891 get_page(page);
21333b2b 892 page_dup_rmap(page);
d559db08
KH
893 if (PageAnon(page))
894 rss[MM_ANONPAGES]++;
895 else
896 rss[MM_FILEPAGES]++;
6aab341e 897 }
ae859762
HD
898
899out_set_pte:
900 set_pte_at(dst_mm, addr, dst_pte, pte);
570a335b 901 return 0;
1da177e4
LT
902}
903
71e3aac0
AA
904int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
905 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
906 unsigned long addr, unsigned long end)
1da177e4 907{
c36987e2 908 pte_t *orig_src_pte, *orig_dst_pte;
1da177e4 909 pte_t *src_pte, *dst_pte;
c74df32c 910 spinlock_t *src_ptl, *dst_ptl;
e040f218 911 int progress = 0;
d559db08 912 int rss[NR_MM_COUNTERS];
570a335b 913 swp_entry_t entry = (swp_entry_t){0};
1da177e4
LT
914
915again:
d559db08
KH
916 init_rss_vec(rss);
917
c74df32c 918 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1da177e4
LT
919 if (!dst_pte)
920 return -ENOMEM;
ece0e2b6 921 src_pte = pte_offset_map(src_pmd, addr);
4c21e2f2 922 src_ptl = pte_lockptr(src_mm, src_pmd);
f20dc5f7 923 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
c36987e2
DN
924 orig_src_pte = src_pte;
925 orig_dst_pte = dst_pte;
6606c3e0 926 arch_enter_lazy_mmu_mode();
1da177e4 927
1da177e4
LT
928 do {
929 /*
930 * We are holding two locks at this point - either of them
931 * could generate latencies in another task on another CPU.
932 */
e040f218
HD
933 if (progress >= 32) {
934 progress = 0;
935 if (need_resched() ||
95c354fe 936 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
e040f218
HD
937 break;
938 }
1da177e4
LT
939 if (pte_none(*src_pte)) {
940 progress++;
941 continue;
942 }
570a335b
HD
943 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
944 vma, addr, rss);
945 if (entry.val)
946 break;
1da177e4
LT
947 progress += 8;
948 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1da177e4 949
6606c3e0 950 arch_leave_lazy_mmu_mode();
c74df32c 951 spin_unlock(src_ptl);
ece0e2b6 952 pte_unmap(orig_src_pte);
d559db08 953 add_mm_rss_vec(dst_mm, rss);
c36987e2 954 pte_unmap_unlock(orig_dst_pte, dst_ptl);
c74df32c 955 cond_resched();
570a335b
HD
956
957 if (entry.val) {
958 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
959 return -ENOMEM;
960 progress = 0;
961 }
1da177e4
LT
962 if (addr != end)
963 goto again;
964 return 0;
965}
966
967static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
968 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
969 unsigned long addr, unsigned long end)
970{
971 pmd_t *src_pmd, *dst_pmd;
972 unsigned long next;
973
974 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
975 if (!dst_pmd)
976 return -ENOMEM;
977 src_pmd = pmd_offset(src_pud, addr);
978 do {
979 next = pmd_addr_end(addr, end);
71e3aac0
AA
980 if (pmd_trans_huge(*src_pmd)) {
981 int err;
14d1a55c 982 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
71e3aac0
AA
983 err = copy_huge_pmd(dst_mm, src_mm,
984 dst_pmd, src_pmd, addr, vma);
985 if (err == -ENOMEM)
986 return -ENOMEM;
987 if (!err)
988 continue;
989 /* fall through */
990 }
1da177e4
LT
991 if (pmd_none_or_clear_bad(src_pmd))
992 continue;
993 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
994 vma, addr, next))
995 return -ENOMEM;
996 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
997 return 0;
998}
999
1000static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1001 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1002 unsigned long addr, unsigned long end)
1003{
1004 pud_t *src_pud, *dst_pud;
1005 unsigned long next;
1006
1007 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1008 if (!dst_pud)
1009 return -ENOMEM;
1010 src_pud = pud_offset(src_pgd, addr);
1011 do {
1012 next = pud_addr_end(addr, end);
1013 if (pud_none_or_clear_bad(src_pud))
1014 continue;
1015 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1016 vma, addr, next))
1017 return -ENOMEM;
1018 } while (dst_pud++, src_pud++, addr = next, addr != end);
1019 return 0;
1020}
1021
1022int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023 struct vm_area_struct *vma)
1024{
1025 pgd_t *src_pgd, *dst_pgd;
1026 unsigned long next;
1027 unsigned long addr = vma->vm_start;
1028 unsigned long end = vma->vm_end;
2ec74c3e
SG
1029 unsigned long mmun_start; /* For mmu_notifiers */
1030 unsigned long mmun_end; /* For mmu_notifiers */
1031 bool is_cow;
cddb8a5c 1032 int ret;
1da177e4 1033
d992895b
NP
1034 /*
1035 * Don't copy ptes where a page fault will fill them correctly.
1036 * Fork becomes much lighter when there are big shared or private
1037 * readonly mappings. The tradeoff is that copy_page_range is more
1038 * efficient than faulting.
1039 */
4b6e1e37
KK
1040 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1041 VM_PFNMAP | VM_MIXEDMAP))) {
d992895b
NP
1042 if (!vma->anon_vma)
1043 return 0;
1044 }
1045
1da177e4
LT
1046 if (is_vm_hugetlb_page(vma))
1047 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1048
b3b9c293 1049 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
2ab64037 1050 /*
1051 * We do not free on error cases below as remove_vma
1052 * gets called on error from higher level routine
1053 */
5180da41 1054 ret = track_pfn_copy(vma);
2ab64037 1055 if (ret)
1056 return ret;
1057 }
1058
cddb8a5c
AA
1059 /*
1060 * We need to invalidate the secondary MMU mappings only when
1061 * there could be a permission downgrade on the ptes of the
1062 * parent mm. And a permission downgrade will only happen if
1063 * is_cow_mapping() returns true.
1064 */
2ec74c3e
SG
1065 is_cow = is_cow_mapping(vma->vm_flags);
1066 mmun_start = addr;
1067 mmun_end = end;
1068 if (is_cow)
1069 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1070 mmun_end);
cddb8a5c
AA
1071
1072 ret = 0;
1da177e4
LT
1073 dst_pgd = pgd_offset(dst_mm, addr);
1074 src_pgd = pgd_offset(src_mm, addr);
1075 do {
1076 next = pgd_addr_end(addr, end);
1077 if (pgd_none_or_clear_bad(src_pgd))
1078 continue;
cddb8a5c
AA
1079 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1080 vma, addr, next))) {
1081 ret = -ENOMEM;
1082 break;
1083 }
1da177e4 1084 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
cddb8a5c 1085
2ec74c3e
SG
1086 if (is_cow)
1087 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
cddb8a5c 1088 return ret;
1da177e4
LT
1089}
1090
51c6f666 1091static unsigned long zap_pte_range(struct mmu_gather *tlb,
b5810039 1092 struct vm_area_struct *vma, pmd_t *pmd,
1da177e4 1093 unsigned long addr, unsigned long end,
97a89413 1094 struct zap_details *details)
1da177e4 1095{
b5810039 1096 struct mm_struct *mm = tlb->mm;
d16dfc55 1097 int force_flush = 0;
d559db08 1098 int rss[NR_MM_COUNTERS];
97a89413 1099 spinlock_t *ptl;
5f1a1907 1100 pte_t *start_pte;
97a89413 1101 pte_t *pte;
e6c495a9 1102 unsigned long range_start = addr;
d559db08 1103
d16dfc55 1104again:
e303297e 1105 init_rss_vec(rss);
5f1a1907
SR
1106 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1107 pte = start_pte;
6606c3e0 1108 arch_enter_lazy_mmu_mode();
1da177e4
LT
1109 do {
1110 pte_t ptent = *pte;
51c6f666 1111 if (pte_none(ptent)) {
1da177e4 1112 continue;
51c6f666 1113 }
6f5e6b9e 1114
1da177e4 1115 if (pte_present(ptent)) {
ee498ed7 1116 struct page *page;
51c6f666 1117
6aab341e 1118 page = vm_normal_page(vma, addr, ptent);
1da177e4
LT
1119 if (unlikely(details) && page) {
1120 /*
1121 * unmap_shared_mapping_pages() wants to
1122 * invalidate cache without truncating:
1123 * unmap shared but keep private pages.
1124 */
1125 if (details->check_mapping &&
1126 details->check_mapping != page->mapping)
1127 continue;
1128 /*
1129 * Each page->index must be checked when
1130 * invalidating or truncating nonlinear.
1131 */
1132 if (details->nonlinear_vma &&
1133 (page->index < details->first_index ||
1134 page->index > details->last_index))
1135 continue;
1136 }
b5810039 1137 ptent = ptep_get_and_clear_full(mm, addr, pte,
a600388d 1138 tlb->fullmm);
1da177e4
LT
1139 tlb_remove_tlb_entry(tlb, pte, addr);
1140 if (unlikely(!page))
1141 continue;
1142 if (unlikely(details) && details->nonlinear_vma
1143 && linear_page_index(details->nonlinear_vma,
1144 addr) != page->index)
b5810039 1145 set_pte_at(mm, addr, pte,
1da177e4 1146 pgoff_to_pte(page->index));
1da177e4 1147 if (PageAnon(page))
d559db08 1148 rss[MM_ANONPAGES]--;
6237bcd9
HD
1149 else {
1150 if (pte_dirty(ptent))
1151 set_page_dirty(page);
4917e5d0 1152 if (pte_young(ptent) &&
64363aad 1153 likely(!(vma->vm_flags & VM_SEQ_READ)))
bf3f3bc5 1154 mark_page_accessed(page);
d559db08 1155 rss[MM_FILEPAGES]--;
6237bcd9 1156 }
edc315fd 1157 page_remove_rmap(page);
3dc14741
HD
1158 if (unlikely(page_mapcount(page) < 0))
1159 print_bad_pte(vma, addr, ptent, page);
d16dfc55
PZ
1160 force_flush = !__tlb_remove_page(tlb, page);
1161 if (force_flush)
1162 break;
1da177e4
LT
1163 continue;
1164 }
1165 /*
1166 * If details->check_mapping, we leave swap entries;
1167 * if details->nonlinear_vma, we leave file entries.
1168 */
1169 if (unlikely(details))
1170 continue;
2509ef26
HD
1171 if (pte_file(ptent)) {
1172 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1173 print_bad_pte(vma, addr, ptent, NULL);
b084d435
KH
1174 } else {
1175 swp_entry_t entry = pte_to_swp_entry(ptent);
1176
1177 if (!non_swap_entry(entry))
1178 rss[MM_SWAPENTS]--;
9f9f1acd
KK
1179 else if (is_migration_entry(entry)) {
1180 struct page *page;
1181
1182 page = migration_entry_to_page(entry);
1183
1184 if (PageAnon(page))
1185 rss[MM_ANONPAGES]--;
1186 else
1187 rss[MM_FILEPAGES]--;
1188 }
b084d435
KH
1189 if (unlikely(!free_swap_and_cache(entry)))
1190 print_bad_pte(vma, addr, ptent, NULL);
1191 }
9888a1ca 1192 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
97a89413 1193 } while (pte++, addr += PAGE_SIZE, addr != end);
ae859762 1194
d559db08 1195 add_mm_rss_vec(mm, rss);
6606c3e0 1196 arch_leave_lazy_mmu_mode();
5f1a1907 1197 pte_unmap_unlock(start_pte, ptl);
51c6f666 1198
d16dfc55
PZ
1199 /*
1200 * mmu_gather ran out of room to batch pages, we break out of
1201 * the PTE lock to avoid doing the potential expensive TLB invalidate
1202 * and page-free while holding it.
1203 */
1204 if (force_flush) {
1205 force_flush = 0;
597e1c35
AS
1206
1207#ifdef HAVE_GENERIC_MMU_GATHER
e6c495a9
VG
1208 tlb->start = range_start;
1209 tlb->end = addr;
597e1c35 1210#endif
d16dfc55 1211 tlb_flush_mmu(tlb);
e6c495a9
VG
1212 if (addr != end) {
1213 range_start = addr;
d16dfc55 1214 goto again;
e6c495a9 1215 }
d16dfc55
PZ
1216 }
1217
51c6f666 1218 return addr;
1da177e4
LT
1219}
1220
51c6f666 1221static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
b5810039 1222 struct vm_area_struct *vma, pud_t *pud,
1da177e4 1223 unsigned long addr, unsigned long end,
97a89413 1224 struct zap_details *details)
1da177e4
LT
1225{
1226 pmd_t *pmd;
1227 unsigned long next;
1228
1229 pmd = pmd_offset(pud, addr);
1230 do {
1231 next = pmd_addr_end(addr, end);
71e3aac0 1232 if (pmd_trans_huge(*pmd)) {
1a5a9906 1233 if (next - addr != HPAGE_PMD_SIZE) {
e0897d75
DR
1234#ifdef CONFIG_DEBUG_VM
1235 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1236 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1237 __func__, addr, end,
1238 vma->vm_start,
1239 vma->vm_end);
1240 BUG();
1241 }
1242#endif
e180377f 1243 split_huge_page_pmd(vma, addr, pmd);
f21760b1 1244 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1a5a9906 1245 goto next;
71e3aac0
AA
1246 /* fall through */
1247 }
1a5a9906
AA
1248 /*
1249 * Here there can be other concurrent MADV_DONTNEED or
1250 * trans huge page faults running, and if the pmd is
1251 * none or trans huge it can change under us. This is
1252 * because MADV_DONTNEED holds the mmap_sem in read
1253 * mode.
1254 */
1255 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1256 goto next;
97a89413 1257 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1a5a9906 1258next:
97a89413
PZ
1259 cond_resched();
1260 } while (pmd++, addr = next, addr != end);
51c6f666
RH
1261
1262 return addr;
1da177e4
LT
1263}
1264
51c6f666 1265static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
b5810039 1266 struct vm_area_struct *vma, pgd_t *pgd,
1da177e4 1267 unsigned long addr, unsigned long end,
97a89413 1268 struct zap_details *details)
1da177e4
LT
1269{
1270 pud_t *pud;
1271 unsigned long next;
1272
1273 pud = pud_offset(pgd, addr);
1274 do {
1275 next = pud_addr_end(addr, end);
97a89413 1276 if (pud_none_or_clear_bad(pud))
1da177e4 1277 continue;
97a89413
PZ
1278 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1279 } while (pud++, addr = next, addr != end);
51c6f666
RH
1280
1281 return addr;
1da177e4
LT
1282}
1283
038c7aa1
AV
1284static void unmap_page_range(struct mmu_gather *tlb,
1285 struct vm_area_struct *vma,
1286 unsigned long addr, unsigned long end,
1287 struct zap_details *details)
1da177e4
LT
1288{
1289 pgd_t *pgd;
1290 unsigned long next;
1291
1292 if (details && !details->check_mapping && !details->nonlinear_vma)
1293 details = NULL;
1294
1295 BUG_ON(addr >= end);
569b846d 1296 mem_cgroup_uncharge_start();
1da177e4
LT
1297 tlb_start_vma(tlb, vma);
1298 pgd = pgd_offset(vma->vm_mm, addr);
1299 do {
1300 next = pgd_addr_end(addr, end);
97a89413 1301 if (pgd_none_or_clear_bad(pgd))
1da177e4 1302 continue;
97a89413
PZ
1303 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1304 } while (pgd++, addr = next, addr != end);
1da177e4 1305 tlb_end_vma(tlb, vma);
569b846d 1306 mem_cgroup_uncharge_end();
1da177e4 1307}
51c6f666 1308
f5cc4eef
AV
1309
1310static void unmap_single_vma(struct mmu_gather *tlb,
1311 struct vm_area_struct *vma, unsigned long start_addr,
4f74d2c8 1312 unsigned long end_addr,
f5cc4eef
AV
1313 struct zap_details *details)
1314{
1315 unsigned long start = max(vma->vm_start, start_addr);
1316 unsigned long end;
1317
1318 if (start >= vma->vm_end)
1319 return;
1320 end = min(vma->vm_end, end_addr);
1321 if (end <= vma->vm_start)
1322 return;
1323
cbc91f71
SD
1324 if (vma->vm_file)
1325 uprobe_munmap(vma, start, end);
1326
b3b9c293 1327 if (unlikely(vma->vm_flags & VM_PFNMAP))
5180da41 1328 untrack_pfn(vma, 0, 0);
f5cc4eef
AV
1329
1330 if (start != end) {
1331 if (unlikely(is_vm_hugetlb_page(vma))) {
1332 /*
1333 * It is undesirable to test vma->vm_file as it
1334 * should be non-null for valid hugetlb area.
1335 * However, vm_file will be NULL in the error
1336 * cleanup path of do_mmap_pgoff. When
1337 * hugetlbfs ->mmap method fails,
1338 * do_mmap_pgoff() nullifies vma->vm_file
1339 * before calling this function to clean up.
1340 * Since no pte has actually been setup, it is
1341 * safe to do nothing in this case.
1342 */
24669e58
AK
1343 if (vma->vm_file) {
1344 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
d833352a 1345 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
24669e58
AK
1346 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1347 }
f5cc4eef
AV
1348 } else
1349 unmap_page_range(tlb, vma, start, end, details);
1350 }
1da177e4
LT
1351}
1352
1da177e4
LT
1353/**
1354 * unmap_vmas - unmap a range of memory covered by a list of vma's
0164f69d 1355 * @tlb: address of the caller's struct mmu_gather
1da177e4
LT
1356 * @vma: the starting vma
1357 * @start_addr: virtual address at which to start unmapping
1358 * @end_addr: virtual address at which to end unmapping
1da177e4 1359 *
508034a3 1360 * Unmap all pages in the vma list.
1da177e4 1361 *
1da177e4
LT
1362 * Only addresses between `start' and `end' will be unmapped.
1363 *
1364 * The VMA list must be sorted in ascending virtual address order.
1365 *
1366 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1367 * range after unmap_vmas() returns. So the only responsibility here is to
1368 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1369 * drops the lock and schedules.
1370 */
6e8bb019 1371void unmap_vmas(struct mmu_gather *tlb,
1da177e4 1372 struct vm_area_struct *vma, unsigned long start_addr,
4f74d2c8 1373 unsigned long end_addr)
1da177e4 1374{
cddb8a5c 1375 struct mm_struct *mm = vma->vm_mm;
1da177e4 1376
cddb8a5c 1377 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
f5cc4eef 1378 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
4f74d2c8 1379 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
cddb8a5c 1380 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1da177e4
LT
1381}
1382
1383/**
1384 * zap_page_range - remove user pages in a given range
1385 * @vma: vm_area_struct holding the applicable pages
eb4546bb 1386 * @start: starting address of pages to zap
1da177e4
LT
1387 * @size: number of bytes to zap
1388 * @details: details of nonlinear truncation or shared cache invalidation
f5cc4eef
AV
1389 *
1390 * Caller must protect the VMA list
1da177e4 1391 */
7e027b14 1392void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1da177e4
LT
1393 unsigned long size, struct zap_details *details)
1394{
1395 struct mm_struct *mm = vma->vm_mm;
d16dfc55 1396 struct mmu_gather tlb;
7e027b14 1397 unsigned long end = start + size;
1da177e4 1398
1da177e4 1399 lru_add_drain();
d16dfc55 1400 tlb_gather_mmu(&tlb, mm, 0);
365e9c87 1401 update_hiwater_rss(mm);
7e027b14
LT
1402 mmu_notifier_invalidate_range_start(mm, start, end);
1403 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
4f74d2c8 1404 unmap_single_vma(&tlb, vma, start, end, details);
7e027b14
LT
1405 mmu_notifier_invalidate_range_end(mm, start, end);
1406 tlb_finish_mmu(&tlb, start, end);
1da177e4
LT
1407}
1408
f5cc4eef
AV
1409/**
1410 * zap_page_range_single - remove user pages in a given range
1411 * @vma: vm_area_struct holding the applicable pages
1412 * @address: starting address of pages to zap
1413 * @size: number of bytes to zap
1414 * @details: details of nonlinear truncation or shared cache invalidation
1415 *
1416 * The range must fit into one VMA.
1da177e4 1417 */
f5cc4eef 1418static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1da177e4
LT
1419 unsigned long size, struct zap_details *details)
1420{
1421 struct mm_struct *mm = vma->vm_mm;
d16dfc55 1422 struct mmu_gather tlb;
1da177e4 1423 unsigned long end = address + size;
1da177e4 1424
1da177e4 1425 lru_add_drain();
d16dfc55 1426 tlb_gather_mmu(&tlb, mm, 0);
365e9c87 1427 update_hiwater_rss(mm);
f5cc4eef 1428 mmu_notifier_invalidate_range_start(mm, address, end);
4f74d2c8 1429 unmap_single_vma(&tlb, vma, address, end, details);
f5cc4eef 1430 mmu_notifier_invalidate_range_end(mm, address, end);
d16dfc55 1431 tlb_finish_mmu(&tlb, address, end);
1da177e4
LT
1432}
1433
c627f9cc
JS
1434/**
1435 * zap_vma_ptes - remove ptes mapping the vma
1436 * @vma: vm_area_struct holding ptes to be zapped
1437 * @address: starting address of pages to zap
1438 * @size: number of bytes to zap
1439 *
1440 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1441 *
1442 * The entire address range must be fully contained within the vma.
1443 *
1444 * Returns 0 if successful.
1445 */
1446int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1447 unsigned long size)
1448{
1449 if (address < vma->vm_start || address + size > vma->vm_end ||
1450 !(vma->vm_flags & VM_PFNMAP))
1451 return -1;
f5cc4eef 1452 zap_page_range_single(vma, address, size, NULL);
c627f9cc
JS
1453 return 0;
1454}
1455EXPORT_SYMBOL_GPL(zap_vma_ptes);
1456
142762bd 1457/**
240aadee 1458 * follow_page_mask - look up a page descriptor from a user-virtual address
142762bd
JW
1459 * @vma: vm_area_struct mapping @address
1460 * @address: virtual address to look up
1461 * @flags: flags modifying lookup behaviour
240aadee 1462 * @page_mask: on output, *page_mask is set according to the size of the page
142762bd
JW
1463 *
1464 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1465 *
1466 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1467 * an error pointer if there is a mapping to something not represented
1468 * by a page descriptor (see also vm_normal_page()).
1da177e4 1469 */
240aadee
ML
1470struct page *follow_page_mask(struct vm_area_struct *vma,
1471 unsigned long address, unsigned int flags,
1472 unsigned int *page_mask)
1da177e4
LT
1473{
1474 pgd_t *pgd;
1475 pud_t *pud;
1476 pmd_t *pmd;
1477 pte_t *ptep, pte;
deceb6cd 1478 spinlock_t *ptl;
1da177e4 1479 struct page *page;
6aab341e 1480 struct mm_struct *mm = vma->vm_mm;
1da177e4 1481
240aadee
ML
1482 *page_mask = 0;
1483
deceb6cd
HD
1484 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1485 if (!IS_ERR(page)) {
1486 BUG_ON(flags & FOLL_GET);
1487 goto out;
1488 }
1da177e4 1489
deceb6cd 1490 page = NULL;
1da177e4
LT
1491 pgd = pgd_offset(mm, address);
1492 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
deceb6cd 1493 goto no_page_table;
1da177e4
LT
1494
1495 pud = pud_offset(pgd, address);
ceb86879 1496 if (pud_none(*pud))
deceb6cd 1497 goto no_page_table;
8a07651e 1498 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
ceb86879
AK
1499 BUG_ON(flags & FOLL_GET);
1500 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1501 goto out;
1502 }
1503 if (unlikely(pud_bad(*pud)))
1504 goto no_page_table;
1505
1da177e4 1506 pmd = pmd_offset(pud, address);
aeed5fce 1507 if (pmd_none(*pmd))
deceb6cd 1508 goto no_page_table;
71e3aac0 1509 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
deceb6cd
HD
1510 BUG_ON(flags & FOLL_GET);
1511 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1da177e4 1512 goto out;
deceb6cd 1513 }
0b9d7052
AA
1514 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1515 goto no_page_table;
71e3aac0 1516 if (pmd_trans_huge(*pmd)) {
500d65d4 1517 if (flags & FOLL_SPLIT) {
e180377f 1518 split_huge_page_pmd(vma, address, pmd);
500d65d4
AA
1519 goto split_fallthrough;
1520 }
71e3aac0
AA
1521 spin_lock(&mm->page_table_lock);
1522 if (likely(pmd_trans_huge(*pmd))) {
1523 if (unlikely(pmd_trans_splitting(*pmd))) {
1524 spin_unlock(&mm->page_table_lock);
1525 wait_split_huge_page(vma->anon_vma, pmd);
1526 } else {
b676b293 1527 page = follow_trans_huge_pmd(vma, address,
71e3aac0
AA
1528 pmd, flags);
1529 spin_unlock(&mm->page_table_lock);
240aadee 1530 *page_mask = HPAGE_PMD_NR - 1;
71e3aac0
AA
1531 goto out;
1532 }
1533 } else
1534 spin_unlock(&mm->page_table_lock);
1535 /* fall through */
1536 }
500d65d4 1537split_fallthrough:
aeed5fce
HD
1538 if (unlikely(pmd_bad(*pmd)))
1539 goto no_page_table;
1540
deceb6cd 1541 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1da177e4
LT
1542
1543 pte = *ptep;
5117b3b8
HD
1544 if (!pte_present(pte)) {
1545 swp_entry_t entry;
1546 /*
1547 * KSM's break_ksm() relies upon recognizing a ksm page
1548 * even while it is being migrated, so for that case we
1549 * need migration_entry_wait().
1550 */
1551 if (likely(!(flags & FOLL_MIGRATION)))
1552 goto no_page;
1553 if (pte_none(pte) || pte_file(pte))
1554 goto no_page;
1555 entry = pte_to_swp_entry(pte);
1556 if (!is_migration_entry(entry))
1557 goto no_page;
1558 pte_unmap_unlock(ptep, ptl);
1559 migration_entry_wait(mm, pmd, address);
1560 goto split_fallthrough;
1561 }
0b9d7052
AA
1562 if ((flags & FOLL_NUMA) && pte_numa(pte))
1563 goto no_page;
deceb6cd
HD
1564 if ((flags & FOLL_WRITE) && !pte_write(pte))
1565 goto unlock;
a13ea5b7 1566
6aab341e 1567 page = vm_normal_page(vma, address, pte);
a13ea5b7
HD
1568 if (unlikely(!page)) {
1569 if ((flags & FOLL_DUMP) ||
62eede62 1570 !is_zero_pfn(pte_pfn(pte)))
a13ea5b7
HD
1571 goto bad_page;
1572 page = pte_page(pte);
1573 }
1da177e4 1574
deceb6cd 1575 if (flags & FOLL_GET)
70b50f94 1576 get_page_foll(page);
deceb6cd
HD
1577 if (flags & FOLL_TOUCH) {
1578 if ((flags & FOLL_WRITE) &&
1579 !pte_dirty(pte) && !PageDirty(page))
1580 set_page_dirty(page);
bd775c42
KM
1581 /*
1582 * pte_mkyoung() would be more correct here, but atomic care
1583 * is needed to avoid losing the dirty bit: it is easier to use
1584 * mark_page_accessed().
1585 */
deceb6cd
HD
1586 mark_page_accessed(page);
1587 }
a1fde08c 1588 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
110d74a9
ML
1589 /*
1590 * The preliminary mapping check is mainly to avoid the
1591 * pointless overhead of lock_page on the ZERO_PAGE
1592 * which might bounce very badly if there is contention.
1593 *
1594 * If the page is already locked, we don't need to
1595 * handle it now - vmscan will handle it later if and
1596 * when it attempts to reclaim the page.
1597 */
1598 if (page->mapping && trylock_page(page)) {
1599 lru_add_drain(); /* push cached pages to LRU */
1600 /*
e6c509f8
HD
1601 * Because we lock page here, and migration is
1602 * blocked by the pte's page reference, and we
1603 * know the page is still mapped, we don't even
1604 * need to check for file-cache page truncation.
110d74a9 1605 */
e6c509f8 1606 mlock_vma_page(page);
110d74a9
ML
1607 unlock_page(page);
1608 }
1609 }
deceb6cd
HD
1610unlock:
1611 pte_unmap_unlock(ptep, ptl);
1da177e4 1612out:
deceb6cd 1613 return page;
1da177e4 1614
89f5b7da
LT
1615bad_page:
1616 pte_unmap_unlock(ptep, ptl);
1617 return ERR_PTR(-EFAULT);
1618
1619no_page:
1620 pte_unmap_unlock(ptep, ptl);
1621 if (!pte_none(pte))
1622 return page;
8e4b9a60 1623
deceb6cd
HD
1624no_page_table:
1625 /*
1626 * When core dumping an enormous anonymous area that nobody
8e4b9a60
HD
1627 * has touched so far, we don't want to allocate unnecessary pages or
1628 * page tables. Return error instead of NULL to skip handle_mm_fault,
1629 * then get_dump_page() will return NULL to leave a hole in the dump.
1630 * But we can only make this optimization where a hole would surely
1631 * be zero-filled if handle_mm_fault() actually did handle it.
deceb6cd 1632 */
8e4b9a60
HD
1633 if ((flags & FOLL_DUMP) &&
1634 (!vma->vm_ops || !vma->vm_ops->fault))
1635 return ERR_PTR(-EFAULT);
deceb6cd 1636 return page;
1da177e4
LT
1637}
1638
95042f9e
LT
1639static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1640{
a09a79f6
MP
1641 return stack_guard_page_start(vma, addr) ||
1642 stack_guard_page_end(vma, addr+PAGE_SIZE);
95042f9e
LT
1643}
1644
0014bd99
HY
1645/**
1646 * __get_user_pages() - pin user pages in memory
1647 * @tsk: task_struct of target task
1648 * @mm: mm_struct of target mm
1649 * @start: starting user address
1650 * @nr_pages: number of pages from start to pin
1651 * @gup_flags: flags modifying pin behaviour
1652 * @pages: array that receives pointers to the pages pinned.
1653 * Should be at least nr_pages long. Or NULL, if caller
1654 * only intends to ensure the pages are faulted in.
1655 * @vmas: array of pointers to vmas corresponding to each page.
1656 * Or NULL if the caller does not require them.
1657 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1658 *
1659 * Returns number of pages pinned. This may be fewer than the number
1660 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1661 * were pinned, returns -errno. Each page returned must be released
1662 * with a put_page() call when it is finished with. vmas will only
1663 * remain valid while mmap_sem is held.
1664 *
1665 * Must be called with mmap_sem held for read or write.
1666 *
1667 * __get_user_pages walks a process's page tables and takes a reference to
1668 * each struct page that each user address corresponds to at a given
1669 * instant. That is, it takes the page that would be accessed if a user
1670 * thread accesses the given user virtual address at that instant.
1671 *
1672 * This does not guarantee that the page exists in the user mappings when
1673 * __get_user_pages returns, and there may even be a completely different
1674 * page there in some cases (eg. if mmapped pagecache has been invalidated
1675 * and subsequently re faulted). However it does guarantee that the page
1676 * won't be freed completely. And mostly callers simply care that the page
1677 * contains data that was valid *at some point in time*. Typically, an IO
1678 * or similar operation cannot guarantee anything stronger anyway because
1679 * locks can't be held over the syscall boundary.
1680 *
1681 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1682 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1683 * appropriate) must be called after the page is finished with, and
1684 * before put_page is called.
1685 *
1686 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1687 * or mmap_sem contention, and if waiting is needed to pin all pages,
1688 * *@nonblocking will be set to 0.
1689 *
1690 * In most cases, get_user_pages or get_user_pages_fast should be used
1691 * instead of __get_user_pages. __get_user_pages should be used only if
1692 * you need some special @gup_flags.
1693 */
28a35716
ML
1694long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1695 unsigned long start, unsigned long nr_pages,
1696 unsigned int gup_flags, struct page **pages,
1697 struct vm_area_struct **vmas, int *nonblocking)
1da177e4 1698{
28a35716 1699 long i;
58fa879e 1700 unsigned long vm_flags;
240aadee 1701 unsigned int page_mask;
1da177e4 1702
28a35716 1703 if (!nr_pages)
900cf086 1704 return 0;
58fa879e
HD
1705
1706 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1707
1da177e4
LT
1708 /*
1709 * Require read or write permissions.
58fa879e 1710 * If FOLL_FORCE is set, we only require the "MAY" flags.
1da177e4 1711 */
58fa879e
HD
1712 vm_flags = (gup_flags & FOLL_WRITE) ?
1713 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1714 vm_flags &= (gup_flags & FOLL_FORCE) ?
1715 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
0b9d7052
AA
1716
1717 /*
1718 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1719 * would be called on PROT_NONE ranges. We must never invoke
1720 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1721 * page faults would unprotect the PROT_NONE ranges if
1722 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1723 * bitflag. So to avoid that, don't set FOLL_NUMA if
1724 * FOLL_FORCE is set.
1725 */
1726 if (!(gup_flags & FOLL_FORCE))
1727 gup_flags |= FOLL_NUMA;
1728
1da177e4
LT
1729 i = 0;
1730
1731 do {
deceb6cd 1732 struct vm_area_struct *vma;
1da177e4
LT
1733
1734 vma = find_extend_vma(mm, start);
e7f22e20 1735 if (!vma && in_gate_area(mm, start)) {
1da177e4 1736 unsigned long pg = start & PAGE_MASK;
1da177e4
LT
1737 pgd_t *pgd;
1738 pud_t *pud;
1739 pmd_t *pmd;
1740 pte_t *pte;
b291f000
NP
1741
1742 /* user gate pages are read-only */
58fa879e 1743 if (gup_flags & FOLL_WRITE)
1da177e4
LT
1744 return i ? : -EFAULT;
1745 if (pg > TASK_SIZE)
1746 pgd = pgd_offset_k(pg);
1747 else
1748 pgd = pgd_offset_gate(mm, pg);
1749 BUG_ON(pgd_none(*pgd));
1750 pud = pud_offset(pgd, pg);
1751 BUG_ON(pud_none(*pud));
1752 pmd = pmd_offset(pud, pg);
690dbe1c
HD
1753 if (pmd_none(*pmd))
1754 return i ? : -EFAULT;
f66055ab 1755 VM_BUG_ON(pmd_trans_huge(*pmd));
1da177e4 1756 pte = pte_offset_map(pmd, pg);
690dbe1c
HD
1757 if (pte_none(*pte)) {
1758 pte_unmap(pte);
1759 return i ? : -EFAULT;
1760 }
95042f9e 1761 vma = get_gate_vma(mm);
1da177e4 1762 if (pages) {
de51257a
HD
1763 struct page *page;
1764
95042f9e 1765 page = vm_normal_page(vma, start, *pte);
de51257a
HD
1766 if (!page) {
1767 if (!(gup_flags & FOLL_DUMP) &&
1768 is_zero_pfn(pte_pfn(*pte)))
1769 page = pte_page(*pte);
1770 else {
1771 pte_unmap(pte);
1772 return i ? : -EFAULT;
1773 }
1774 }
6aab341e 1775 pages[i] = page;
de51257a 1776 get_page(page);
1da177e4
LT
1777 }
1778 pte_unmap(pte);
240aadee 1779 page_mask = 0;
95042f9e 1780 goto next_page;
1da177e4
LT
1781 }
1782
b291f000
NP
1783 if (!vma ||
1784 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1c3aff1c 1785 !(vm_flags & vma->vm_flags))
1da177e4
LT
1786 return i ? : -EFAULT;
1787
2a15efc9
HD
1788 if (is_vm_hugetlb_page(vma)) {
1789 i = follow_hugetlb_page(mm, vma, pages, vmas,
58fa879e 1790 &start, &nr_pages, i, gup_flags);
2a15efc9
HD
1791 continue;
1792 }
deceb6cd 1793
1da177e4 1794 do {
08ef4729 1795 struct page *page;
58fa879e 1796 unsigned int foll_flags = gup_flags;
240aadee 1797 unsigned int page_increm;
1da177e4 1798
462e00cc 1799 /*
4779280d 1800 * If we have a pending SIGKILL, don't keep faulting
1c3aff1c 1801 * pages and potentially allocating memory.
462e00cc 1802 */
1c3aff1c 1803 if (unlikely(fatal_signal_pending(current)))
4779280d 1804 return i ? i : -ERESTARTSYS;
462e00cc 1805
deceb6cd 1806 cond_resched();
240aadee
ML
1807 while (!(page = follow_page_mask(vma, start,
1808 foll_flags, &page_mask))) {
deceb6cd 1809 int ret;
53a7706d
ML
1810 unsigned int fault_flags = 0;
1811
a09a79f6
MP
1812 /* For mlock, just skip the stack guard page. */
1813 if (foll_flags & FOLL_MLOCK) {
1814 if (stack_guard_page(vma, start))
1815 goto next_page;
1816 }
53a7706d
ML
1817 if (foll_flags & FOLL_WRITE)
1818 fault_flags |= FAULT_FLAG_WRITE;
1819 if (nonblocking)
1820 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
318b275f
GN
1821 if (foll_flags & FOLL_NOWAIT)
1822 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
d06063cc 1823
d26ed650 1824 ret = handle_mm_fault(mm, vma, start,
53a7706d 1825 fault_flags);
d26ed650 1826
83c54070
NP
1827 if (ret & VM_FAULT_ERROR) {
1828 if (ret & VM_FAULT_OOM)
1829 return i ? i : -ENOMEM;
69ebb83e
HY
1830 if (ret & (VM_FAULT_HWPOISON |
1831 VM_FAULT_HWPOISON_LARGE)) {
1832 if (i)
1833 return i;
1834 else if (gup_flags & FOLL_HWPOISON)
1835 return -EHWPOISON;
1836 else
1837 return -EFAULT;
1838 }
1839 if (ret & VM_FAULT_SIGBUS)
83c54070
NP
1840 return i ? i : -EFAULT;
1841 BUG();
1842 }
e7f22e20
SW
1843
1844 if (tsk) {
1845 if (ret & VM_FAULT_MAJOR)
1846 tsk->maj_flt++;
1847 else
1848 tsk->min_flt++;
1849 }
83c54070 1850
53a7706d 1851 if (ret & VM_FAULT_RETRY) {
318b275f
GN
1852 if (nonblocking)
1853 *nonblocking = 0;
53a7706d
ML
1854 return i;
1855 }
1856
a68d2ebc 1857 /*
83c54070
NP
1858 * The VM_FAULT_WRITE bit tells us that
1859 * do_wp_page has broken COW when necessary,
1860 * even if maybe_mkwrite decided not to set
1861 * pte_write. We can thus safely do subsequent
878b63ac
HD
1862 * page lookups as if they were reads. But only
1863 * do so when looping for pte_write is futile:
1864 * in some cases userspace may also be wanting
1865 * to write to the gotten user page, which a
1866 * read fault here might prevent (a readonly
1867 * page might get reCOWed by userspace write).
a68d2ebc 1868 */
878b63ac
HD
1869 if ((ret & VM_FAULT_WRITE) &&
1870 !(vma->vm_flags & VM_WRITE))
deceb6cd 1871 foll_flags &= ~FOLL_WRITE;
83c54070 1872
7f7bbbe5 1873 cond_resched();
1da177e4 1874 }
89f5b7da
LT
1875 if (IS_ERR(page))
1876 return i ? i : PTR_ERR(page);
1da177e4 1877 if (pages) {
08ef4729 1878 pages[i] = page;
03beb076 1879
a6f36be3 1880 flush_anon_page(vma, page, start);
08ef4729 1881 flush_dcache_page(page);
240aadee 1882 page_mask = 0;
1da177e4 1883 }
95042f9e 1884next_page:
240aadee 1885 if (vmas) {
1da177e4 1886 vmas[i] = vma;
240aadee
ML
1887 page_mask = 0;
1888 }
1889 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1890 if (page_increm > nr_pages)
1891 page_increm = nr_pages;
1892 i += page_increm;
1893 start += page_increm * PAGE_SIZE;
1894 nr_pages -= page_increm;
9d73777e
PZ
1895 } while (nr_pages && start < vma->vm_end);
1896 } while (nr_pages);
1da177e4
LT
1897 return i;
1898}
0014bd99 1899EXPORT_SYMBOL(__get_user_pages);
b291f000 1900
2efaca92
BH
1901/*
1902 * fixup_user_fault() - manually resolve a user page fault
1903 * @tsk: the task_struct to use for page fault accounting, or
1904 * NULL if faults are not to be recorded.
1905 * @mm: mm_struct of target mm
1906 * @address: user address
1907 * @fault_flags:flags to pass down to handle_mm_fault()
1908 *
1909 * This is meant to be called in the specific scenario where for locking reasons
1910 * we try to access user memory in atomic context (within a pagefault_disable()
1911 * section), this returns -EFAULT, and we want to resolve the user fault before
1912 * trying again.
1913 *
1914 * Typically this is meant to be used by the futex code.
1915 *
1916 * The main difference with get_user_pages() is that this function will
1917 * unconditionally call handle_mm_fault() which will in turn perform all the
1918 * necessary SW fixup of the dirty and young bits in the PTE, while
1919 * handle_mm_fault() only guarantees to update these in the struct page.
1920 *
1921 * This is important for some architectures where those bits also gate the
1922 * access permission to the page because they are maintained in software. On
1923 * such architectures, gup() will not be enough to make a subsequent access
1924 * succeed.
1925 *
1926 * This should be called with the mm_sem held for read.
1927 */
1928int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1929 unsigned long address, unsigned int fault_flags)
1930{
1931 struct vm_area_struct *vma;
1932 int ret;
1933
1934 vma = find_extend_vma(mm, address);
1935 if (!vma || address < vma->vm_start)
1936 return -EFAULT;
1937
1938 ret = handle_mm_fault(mm, vma, address, fault_flags);
1939 if (ret & VM_FAULT_ERROR) {
1940 if (ret & VM_FAULT_OOM)
1941 return -ENOMEM;
1942 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1943 return -EHWPOISON;
1944 if (ret & VM_FAULT_SIGBUS)
1945 return -EFAULT;
1946 BUG();
1947 }
1948 if (tsk) {
1949 if (ret & VM_FAULT_MAJOR)
1950 tsk->maj_flt++;
1951 else
1952 tsk->min_flt++;
1953 }
1954 return 0;
1955}
1956
1957/*
d2bf6be8 1958 * get_user_pages() - pin user pages in memory
e7f22e20
SW
1959 * @tsk: the task_struct to use for page fault accounting, or
1960 * NULL if faults are not to be recorded.
d2bf6be8
NP
1961 * @mm: mm_struct of target mm
1962 * @start: starting user address
9d73777e 1963 * @nr_pages: number of pages from start to pin
d2bf6be8
NP
1964 * @write: whether pages will be written to by the caller
1965 * @force: whether to force write access even if user mapping is
1966 * readonly. This will result in the page being COWed even
1967 * in MAP_SHARED mappings. You do not want this.
1968 * @pages: array that receives pointers to the pages pinned.
1969 * Should be at least nr_pages long. Or NULL, if caller
1970 * only intends to ensure the pages are faulted in.
1971 * @vmas: array of pointers to vmas corresponding to each page.
1972 * Or NULL if the caller does not require them.
1973 *
1974 * Returns number of pages pinned. This may be fewer than the number
9d73777e 1975 * requested. If nr_pages is 0 or negative, returns 0. If no pages
d2bf6be8
NP
1976 * were pinned, returns -errno. Each page returned must be released
1977 * with a put_page() call when it is finished with. vmas will only
1978 * remain valid while mmap_sem is held.
1979 *
1980 * Must be called with mmap_sem held for read or write.
1981 *
1982 * get_user_pages walks a process's page tables and takes a reference to
1983 * each struct page that each user address corresponds to at a given
1984 * instant. That is, it takes the page that would be accessed if a user
1985 * thread accesses the given user virtual address at that instant.
1986 *
1987 * This does not guarantee that the page exists in the user mappings when
1988 * get_user_pages returns, and there may even be a completely different
1989 * page there in some cases (eg. if mmapped pagecache has been invalidated
1990 * and subsequently re faulted). However it does guarantee that the page
1991 * won't be freed completely. And mostly callers simply care that the page
1992 * contains data that was valid *at some point in time*. Typically, an IO
1993 * or similar operation cannot guarantee anything stronger anyway because
1994 * locks can't be held over the syscall boundary.
1995 *
1996 * If write=0, the page must not be written to. If the page is written to,
1997 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1998 * after the page is finished with, and before put_page is called.
1999 *
2000 * get_user_pages is typically used for fewer-copy IO operations, to get a
2001 * handle on the memory by some means other than accesses via the user virtual
2002 * addresses. The pages may be submitted for DMA to devices or accessed via
2003 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2004 * use the correct cache flushing APIs.
2005 *
2006 * See also get_user_pages_fast, for performance critical applications.
2007 */
28a35716
ML
2008long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2009 unsigned long start, unsigned long nr_pages, int write,
2010 int force, struct page **pages, struct vm_area_struct **vmas)
b291f000 2011{
58fa879e 2012 int flags = FOLL_TOUCH;
b291f000 2013
58fa879e
HD
2014 if (pages)
2015 flags |= FOLL_GET;
b291f000 2016 if (write)
58fa879e 2017 flags |= FOLL_WRITE;
b291f000 2018 if (force)
58fa879e 2019 flags |= FOLL_FORCE;
b291f000 2020
53a7706d
ML
2021 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2022 NULL);
b291f000 2023}
1da177e4
LT
2024EXPORT_SYMBOL(get_user_pages);
2025
f3e8fccd
HD
2026/**
2027 * get_dump_page() - pin user page in memory while writing it to core dump
2028 * @addr: user address
2029 *
2030 * Returns struct page pointer of user page pinned for dump,
2031 * to be freed afterwards by page_cache_release() or put_page().
2032 *
2033 * Returns NULL on any kind of failure - a hole must then be inserted into
2034 * the corefile, to preserve alignment with its headers; and also returns
2035 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2036 * allowing a hole to be left in the corefile to save diskspace.
2037 *
2038 * Called without mmap_sem, but after all other threads have been killed.
2039 */
2040#ifdef CONFIG_ELF_CORE
2041struct page *get_dump_page(unsigned long addr)
2042{
2043 struct vm_area_struct *vma;
2044 struct page *page;
2045
2046 if (__get_user_pages(current, current->mm, addr, 1,
53a7706d
ML
2047 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2048 NULL) < 1)
f3e8fccd 2049 return NULL;
f3e8fccd
HD
2050 flush_cache_page(vma, addr, page_to_pfn(page));
2051 return page;
2052}
2053#endif /* CONFIG_ELF_CORE */
2054
25ca1d6c 2055pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
920c7a5d 2056 spinlock_t **ptl)
c9cfcddf
LT
2057{
2058 pgd_t * pgd = pgd_offset(mm, addr);
2059 pud_t * pud = pud_alloc(mm, pgd, addr);
2060 if (pud) {
49c91fb0 2061 pmd_t * pmd = pmd_alloc(mm, pud, addr);
f66055ab
AA
2062 if (pmd) {
2063 VM_BUG_ON(pmd_trans_huge(*pmd));
c9cfcddf 2064 return pte_alloc_map_lock(mm, pmd, addr, ptl);
f66055ab 2065 }
c9cfcddf
LT
2066 }
2067 return NULL;
2068}
2069
238f58d8
LT
2070/*
2071 * This is the old fallback for page remapping.
2072 *
2073 * For historical reasons, it only allows reserved pages. Only
2074 * old drivers should use this, and they needed to mark their
2075 * pages reserved for the old functions anyway.
2076 */
423bad60
NP
2077static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2078 struct page *page, pgprot_t prot)
238f58d8 2079{
423bad60 2080 struct mm_struct *mm = vma->vm_mm;
238f58d8 2081 int retval;
c9cfcddf 2082 pte_t *pte;
8a9f3ccd
BS
2083 spinlock_t *ptl;
2084
238f58d8 2085 retval = -EINVAL;
a145dd41 2086 if (PageAnon(page))
5b4e655e 2087 goto out;
238f58d8
LT
2088 retval = -ENOMEM;
2089 flush_dcache_page(page);
c9cfcddf 2090 pte = get_locked_pte(mm, addr, &ptl);
238f58d8 2091 if (!pte)
5b4e655e 2092 goto out;
238f58d8
LT
2093 retval = -EBUSY;
2094 if (!pte_none(*pte))
2095 goto out_unlock;
2096
2097 /* Ok, finally just insert the thing.. */
2098 get_page(page);
34e55232 2099 inc_mm_counter_fast(mm, MM_FILEPAGES);
238f58d8
LT
2100 page_add_file_rmap(page);
2101 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2102
2103 retval = 0;
8a9f3ccd
BS
2104 pte_unmap_unlock(pte, ptl);
2105 return retval;
238f58d8
LT
2106out_unlock:
2107 pte_unmap_unlock(pte, ptl);
2108out:
2109 return retval;
2110}
2111
bfa5bf6d
REB
2112/**
2113 * vm_insert_page - insert single page into user vma
2114 * @vma: user vma to map to
2115 * @addr: target user address of this page
2116 * @page: source kernel page
2117 *
a145dd41
LT
2118 * This allows drivers to insert individual pages they've allocated
2119 * into a user vma.
2120 *
2121 * The page has to be a nice clean _individual_ kernel allocation.
2122 * If you allocate a compound page, you need to have marked it as
2123 * such (__GFP_COMP), or manually just split the page up yourself
8dfcc9ba 2124 * (see split_page()).
a145dd41
LT
2125 *
2126 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2127 * took an arbitrary page protection parameter. This doesn't allow
2128 * that. Your vma protection will have to be set up correctly, which
2129 * means that if you want a shared writable mapping, you'd better
2130 * ask for a shared writable mapping!
2131 *
2132 * The page does not need to be reserved.
4b6e1e37
KK
2133 *
2134 * Usually this function is called from f_op->mmap() handler
2135 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2136 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2137 * function from other places, for example from page-fault handler.
a145dd41 2138 */
423bad60
NP
2139int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2140 struct page *page)
a145dd41
LT
2141{
2142 if (addr < vma->vm_start || addr >= vma->vm_end)
2143 return -EFAULT;
2144 if (!page_count(page))
2145 return -EINVAL;
4b6e1e37
KK
2146 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2147 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2148 BUG_ON(vma->vm_flags & VM_PFNMAP);
2149 vma->vm_flags |= VM_MIXEDMAP;
2150 }
423bad60 2151 return insert_page(vma, addr, page, vma->vm_page_prot);
a145dd41 2152}
e3c3374f 2153EXPORT_SYMBOL(vm_insert_page);
a145dd41 2154
423bad60
NP
2155static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2156 unsigned long pfn, pgprot_t prot)
2157{
2158 struct mm_struct *mm = vma->vm_mm;
2159 int retval;
2160 pte_t *pte, entry;
2161 spinlock_t *ptl;
2162
2163 retval = -ENOMEM;
2164 pte = get_locked_pte(mm, addr, &ptl);
2165 if (!pte)
2166 goto out;
2167 retval = -EBUSY;
2168 if (!pte_none(*pte))
2169 goto out_unlock;
2170
2171 /* Ok, finally just insert the thing.. */
2172 entry = pte_mkspecial(pfn_pte(pfn, prot));
2173 set_pte_at(mm, addr, pte, entry);
4b3073e1 2174 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
423bad60
NP
2175
2176 retval = 0;
2177out_unlock:
2178 pte_unmap_unlock(pte, ptl);
2179out:
2180 return retval;
2181}
2182
e0dc0d8f
NP
2183/**
2184 * vm_insert_pfn - insert single pfn into user vma
2185 * @vma: user vma to map to
2186 * @addr: target user address of this page
2187 * @pfn: source kernel pfn
2188 *
c462f179 2189 * Similar to vm_insert_page, this allows drivers to insert individual pages
e0dc0d8f
NP
2190 * they've allocated into a user vma. Same comments apply.
2191 *
2192 * This function should only be called from a vm_ops->fault handler, and
2193 * in that case the handler should return NULL.
0d71d10a
NP
2194 *
2195 * vma cannot be a COW mapping.
2196 *
2197 * As this is called only for pages that do not currently exist, we
2198 * do not need to flush old virtual caches or the TLB.
e0dc0d8f
NP
2199 */
2200int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
423bad60 2201 unsigned long pfn)
e0dc0d8f 2202{
2ab64037 2203 int ret;
e4b866ed 2204 pgprot_t pgprot = vma->vm_page_prot;
7e675137
NP
2205 /*
2206 * Technically, architectures with pte_special can avoid all these
2207 * restrictions (same for remap_pfn_range). However we would like
2208 * consistency in testing and feature parity among all, so we should
2209 * try to keep these invariants in place for everybody.
2210 */
b379d790
JH
2211 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2212 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2213 (VM_PFNMAP|VM_MIXEDMAP));
2214 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2215 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
e0dc0d8f 2216
423bad60
NP
2217 if (addr < vma->vm_start || addr >= vma->vm_end)
2218 return -EFAULT;
5180da41 2219 if (track_pfn_insert(vma, &pgprot, pfn))
2ab64037 2220 return -EINVAL;
2221
e4b866ed 2222 ret = insert_pfn(vma, addr, pfn, pgprot);
2ab64037 2223
2ab64037 2224 return ret;
423bad60
NP
2225}
2226EXPORT_SYMBOL(vm_insert_pfn);
e0dc0d8f 2227
423bad60
NP
2228int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2229 unsigned long pfn)
2230{
2231 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
e0dc0d8f 2232
423bad60
NP
2233 if (addr < vma->vm_start || addr >= vma->vm_end)
2234 return -EFAULT;
e0dc0d8f 2235
423bad60
NP
2236 /*
2237 * If we don't have pte special, then we have to use the pfn_valid()
2238 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2239 * refcount the page if pfn_valid is true (hence insert_page rather
62eede62
HD
2240 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2241 * without pte special, it would there be refcounted as a normal page.
423bad60
NP
2242 */
2243 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2244 struct page *page;
2245
2246 page = pfn_to_page(pfn);
2247 return insert_page(vma, addr, page, vma->vm_page_prot);
2248 }
2249 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
e0dc0d8f 2250}
423bad60 2251EXPORT_SYMBOL(vm_insert_mixed);
e0dc0d8f 2252
1da177e4
LT
2253/*
2254 * maps a range of physical memory into the requested pages. the old
2255 * mappings are removed. any references to nonexistent pages results
2256 * in null mappings (currently treated as "copy-on-access")
2257 */
2258static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2259 unsigned long addr, unsigned long end,
2260 unsigned long pfn, pgprot_t prot)
2261{
2262 pte_t *pte;
c74df32c 2263 spinlock_t *ptl;
1da177e4 2264
c74df32c 2265 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1da177e4
LT
2266 if (!pte)
2267 return -ENOMEM;
6606c3e0 2268 arch_enter_lazy_mmu_mode();
1da177e4
LT
2269 do {
2270 BUG_ON(!pte_none(*pte));
7e675137 2271 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1da177e4
LT
2272 pfn++;
2273 } while (pte++, addr += PAGE_SIZE, addr != end);
6606c3e0 2274 arch_leave_lazy_mmu_mode();
c74df32c 2275 pte_unmap_unlock(pte - 1, ptl);
1da177e4
LT
2276 return 0;
2277}
2278
2279static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2280 unsigned long addr, unsigned long end,
2281 unsigned long pfn, pgprot_t prot)
2282{
2283 pmd_t *pmd;
2284 unsigned long next;
2285
2286 pfn -= addr >> PAGE_SHIFT;
2287 pmd = pmd_alloc(mm, pud, addr);
2288 if (!pmd)
2289 return -ENOMEM;
f66055ab 2290 VM_BUG_ON(pmd_trans_huge(*pmd));
1da177e4
LT
2291 do {
2292 next = pmd_addr_end(addr, end);
2293 if (remap_pte_range(mm, pmd, addr, next,
2294 pfn + (addr >> PAGE_SHIFT), prot))
2295 return -ENOMEM;
2296 } while (pmd++, addr = next, addr != end);
2297 return 0;
2298}
2299
2300static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2301 unsigned long addr, unsigned long end,
2302 unsigned long pfn, pgprot_t prot)
2303{
2304 pud_t *pud;
2305 unsigned long next;
2306
2307 pfn -= addr >> PAGE_SHIFT;
2308 pud = pud_alloc(mm, pgd, addr);
2309 if (!pud)
2310 return -ENOMEM;
2311 do {
2312 next = pud_addr_end(addr, end);
2313 if (remap_pmd_range(mm, pud, addr, next,
2314 pfn + (addr >> PAGE_SHIFT), prot))
2315 return -ENOMEM;
2316 } while (pud++, addr = next, addr != end);
2317 return 0;
2318}
2319
bfa5bf6d
REB
2320/**
2321 * remap_pfn_range - remap kernel memory to userspace
2322 * @vma: user vma to map to
2323 * @addr: target user address to start at
2324 * @pfn: physical address of kernel memory
2325 * @size: size of map area
2326 * @prot: page protection flags for this mapping
2327 *
2328 * Note: this is only safe if the mm semaphore is held when called.
2329 */
1da177e4
LT
2330int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2331 unsigned long pfn, unsigned long size, pgprot_t prot)
2332{
2333 pgd_t *pgd;
2334 unsigned long next;
2d15cab8 2335 unsigned long end = addr + PAGE_ALIGN(size);
1da177e4
LT
2336 struct mm_struct *mm = vma->vm_mm;
2337 int err;
2338
2339 /*
2340 * Physically remapped pages are special. Tell the
2341 * rest of the world about it:
2342 * VM_IO tells people not to look at these pages
2343 * (accesses can have side effects).
6aab341e
LT
2344 * VM_PFNMAP tells the core MM that the base pages are just
2345 * raw PFN mappings, and do not have a "struct page" associated
2346 * with them.
314e51b9
KK
2347 * VM_DONTEXPAND
2348 * Disable vma merging and expanding with mremap().
2349 * VM_DONTDUMP
2350 * Omit vma from core dump, even when VM_IO turned off.
fb155c16
LT
2351 *
2352 * There's a horrible special case to handle copy-on-write
2353 * behaviour that some programs depend on. We mark the "original"
2354 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
b3b9c293 2355 * See vm_normal_page() for details.
1da177e4 2356 */
b3b9c293
KK
2357 if (is_cow_mapping(vma->vm_flags)) {
2358 if (addr != vma->vm_start || end != vma->vm_end)
2359 return -EINVAL;
fb155c16 2360 vma->vm_pgoff = pfn;
b3b9c293
KK
2361 }
2362
2363 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2364 if (err)
3c8bb73a 2365 return -EINVAL;
fb155c16 2366
314e51b9 2367 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1da177e4
LT
2368
2369 BUG_ON(addr >= end);
2370 pfn -= addr >> PAGE_SHIFT;
2371 pgd = pgd_offset(mm, addr);
2372 flush_cache_range(vma, addr, end);
1da177e4
LT
2373 do {
2374 next = pgd_addr_end(addr, end);
2375 err = remap_pud_range(mm, pgd, addr, next,
2376 pfn + (addr >> PAGE_SHIFT), prot);
2377 if (err)
2378 break;
2379 } while (pgd++, addr = next, addr != end);
2ab64037 2380
2381 if (err)
5180da41 2382 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2ab64037 2383
1da177e4
LT
2384 return err;
2385}
2386EXPORT_SYMBOL(remap_pfn_range);
2387
b4cbb197
LT
2388/**
2389 * vm_iomap_memory - remap memory to userspace
2390 * @vma: user vma to map to
2391 * @start: start of area
2392 * @len: size of area
2393 *
2394 * This is a simplified io_remap_pfn_range() for common driver use. The
2395 * driver just needs to give us the physical memory range to be mapped,
2396 * we'll figure out the rest from the vma information.
2397 *
2398 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2399 * whatever write-combining details or similar.
2400 */
2401int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2402{
2403 unsigned long vm_len, pfn, pages;
2404
2405 /* Check that the physical memory area passed in looks valid */
2406 if (start + len < start)
2407 return -EINVAL;
2408 /*
2409 * You *really* shouldn't map things that aren't page-aligned,
2410 * but we've historically allowed it because IO memory might
2411 * just have smaller alignment.
2412 */
2413 len += start & ~PAGE_MASK;
2414 pfn = start >> PAGE_SHIFT;
2415 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2416 if (pfn + pages < pfn)
2417 return -EINVAL;
2418
2419 /* We start the mapping 'vm_pgoff' pages into the area */
2420 if (vma->vm_pgoff > pages)
2421 return -EINVAL;
2422 pfn += vma->vm_pgoff;
2423 pages -= vma->vm_pgoff;
2424
2425 /* Can we fit all of the mapping? */
2426 vm_len = vma->vm_end - vma->vm_start;
2427 if (vm_len >> PAGE_SHIFT > pages)
2428 return -EINVAL;
2429
2430 /* Ok, let it rip */
2431 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2432}
2433EXPORT_SYMBOL(vm_iomap_memory);
2434
aee16b3c
JF
2435static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2436 unsigned long addr, unsigned long end,
2437 pte_fn_t fn, void *data)
2438{
2439 pte_t *pte;
2440 int err;
2f569afd 2441 pgtable_t token;
94909914 2442 spinlock_t *uninitialized_var(ptl);
aee16b3c
JF
2443
2444 pte = (mm == &init_mm) ?
2445 pte_alloc_kernel(pmd, addr) :
2446 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2447 if (!pte)
2448 return -ENOMEM;
2449
2450 BUG_ON(pmd_huge(*pmd));
2451
38e0edb1
JF
2452 arch_enter_lazy_mmu_mode();
2453
2f569afd 2454 token = pmd_pgtable(*pmd);
aee16b3c
JF
2455
2456 do {
c36987e2 2457 err = fn(pte++, token, addr, data);
aee16b3c
JF
2458 if (err)
2459 break;
c36987e2 2460 } while (addr += PAGE_SIZE, addr != end);
aee16b3c 2461
38e0edb1
JF
2462 arch_leave_lazy_mmu_mode();
2463
aee16b3c
JF
2464 if (mm != &init_mm)
2465 pte_unmap_unlock(pte-1, ptl);
2466 return err;
2467}
2468
2469static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2470 unsigned long addr, unsigned long end,
2471 pte_fn_t fn, void *data)
2472{
2473 pmd_t *pmd;
2474 unsigned long next;
2475 int err;
2476
ceb86879
AK
2477 BUG_ON(pud_huge(*pud));
2478
aee16b3c
JF
2479 pmd = pmd_alloc(mm, pud, addr);
2480 if (!pmd)
2481 return -ENOMEM;
2482 do {
2483 next = pmd_addr_end(addr, end);
2484 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2485 if (err)
2486 break;
2487 } while (pmd++, addr = next, addr != end);
2488 return err;
2489}
2490
2491static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2492 unsigned long addr, unsigned long end,
2493 pte_fn_t fn, void *data)
2494{
2495 pud_t *pud;
2496 unsigned long next;
2497 int err;
2498
2499 pud = pud_alloc(mm, pgd, addr);
2500 if (!pud)
2501 return -ENOMEM;
2502 do {
2503 next = pud_addr_end(addr, end);
2504 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2505 if (err)
2506 break;
2507 } while (pud++, addr = next, addr != end);
2508 return err;
2509}
2510
2511/*
2512 * Scan a region of virtual memory, filling in page tables as necessary
2513 * and calling a provided function on each leaf page table.
2514 */
2515int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2516 unsigned long size, pte_fn_t fn, void *data)
2517{
2518 pgd_t *pgd;
2519 unsigned long next;
57250a5b 2520 unsigned long end = addr + size;
aee16b3c
JF
2521 int err;
2522
2523 BUG_ON(addr >= end);
2524 pgd = pgd_offset(mm, addr);
2525 do {
2526 next = pgd_addr_end(addr, end);
2527 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2528 if (err)
2529 break;
2530 } while (pgd++, addr = next, addr != end);
57250a5b 2531
aee16b3c
JF
2532 return err;
2533}
2534EXPORT_SYMBOL_GPL(apply_to_page_range);
2535
8f4e2101
HD
2536/*
2537 * handle_pte_fault chooses page fault handler according to an entry
2538 * which was read non-atomically. Before making any commitment, on
2539 * those architectures or configurations (e.g. i386 with PAE) which
a335b2e1 2540 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
8f4e2101
HD
2541 * must check under lock before unmapping the pte and proceeding
2542 * (but do_wp_page is only called after already making such a check;
a335b2e1 2543 * and do_anonymous_page can safely check later on).
8f4e2101 2544 */
4c21e2f2 2545static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
8f4e2101
HD
2546 pte_t *page_table, pte_t orig_pte)
2547{
2548 int same = 1;
2549#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2550 if (sizeof(pte_t) > sizeof(unsigned long)) {
4c21e2f2
HD
2551 spinlock_t *ptl = pte_lockptr(mm, pmd);
2552 spin_lock(ptl);
8f4e2101 2553 same = pte_same(*page_table, orig_pte);
4c21e2f2 2554 spin_unlock(ptl);
8f4e2101
HD
2555 }
2556#endif
2557 pte_unmap(page_table);
2558 return same;
2559}
2560
9de455b2 2561static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
6aab341e
LT
2562{
2563 /*
2564 * If the source page was a PFN mapping, we don't have
2565 * a "struct page" for it. We do a best-effort copy by
2566 * just copying from the original user address. If that
2567 * fails, we just zero-fill it. Live with it.
2568 */
2569 if (unlikely(!src)) {
9b04c5fe 2570 void *kaddr = kmap_atomic(dst);
5d2a2dbb
LT
2571 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2572
2573 /*
2574 * This really shouldn't fail, because the page is there
2575 * in the page tables. But it might just be unreadable,
2576 * in which case we just give up and fill the result with
2577 * zeroes.
2578 */
2579 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
3ecb01df 2580 clear_page(kaddr);
9b04c5fe 2581 kunmap_atomic(kaddr);
c4ec7b0d 2582 flush_dcache_page(dst);
0ed361de
NP
2583 } else
2584 copy_user_highpage(dst, src, va, vma);
6aab341e
LT
2585}
2586
1da177e4
LT
2587/*
2588 * This routine handles present pages, when users try to write
2589 * to a shared page. It is done by copying the page to a new address
2590 * and decrementing the shared-page counter for the old page.
2591 *
1da177e4
LT
2592 * Note that this routine assumes that the protection checks have been
2593 * done by the caller (the low-level page fault routine in most cases).
2594 * Thus we can safely just mark it writable once we've done any necessary
2595 * COW.
2596 *
2597 * We also mark the page dirty at this point even though the page will
2598 * change only once the write actually happens. This avoids a few races,
2599 * and potentially makes it more efficient.
2600 *
8f4e2101
HD
2601 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2602 * but allow concurrent faults), with pte both mapped and locked.
2603 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 2604 */
65500d23
HD
2605static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2606 unsigned long address, pte_t *page_table, pmd_t *pmd,
8f4e2101 2607 spinlock_t *ptl, pte_t orig_pte)
e6219ec8 2608 __releases(ptl)
1da177e4 2609{
2ec74c3e 2610 struct page *old_page, *new_page = NULL;
1da177e4 2611 pte_t entry;
b009c024 2612 int ret = 0;
a200ee18 2613 int page_mkwrite = 0;
d08b3851 2614 struct page *dirty_page = NULL;
1756954c
DR
2615 unsigned long mmun_start = 0; /* For mmu_notifiers */
2616 unsigned long mmun_end = 0; /* For mmu_notifiers */
1da177e4 2617
6aab341e 2618 old_page = vm_normal_page(vma, address, orig_pte);
251b97f5
PZ
2619 if (!old_page) {
2620 /*
2621 * VM_MIXEDMAP !pfn_valid() case
2622 *
2623 * We should not cow pages in a shared writeable mapping.
2624 * Just mark the pages writable as we can't do any dirty
2625 * accounting on raw pfn maps.
2626 */
2627 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2628 (VM_WRITE|VM_SHARED))
2629 goto reuse;
6aab341e 2630 goto gotten;
251b97f5 2631 }
1da177e4 2632
d08b3851 2633 /*
ee6a6457
PZ
2634 * Take out anonymous pages first, anonymous shared vmas are
2635 * not dirty accountable.
d08b3851 2636 */
9a840895 2637 if (PageAnon(old_page) && !PageKsm(old_page)) {
ab967d86
HD
2638 if (!trylock_page(old_page)) {
2639 page_cache_get(old_page);
2640 pte_unmap_unlock(page_table, ptl);
2641 lock_page(old_page);
2642 page_table = pte_offset_map_lock(mm, pmd, address,
2643 &ptl);
2644 if (!pte_same(*page_table, orig_pte)) {
2645 unlock_page(old_page);
ab967d86
HD
2646 goto unlock;
2647 }
2648 page_cache_release(old_page);
ee6a6457 2649 }
b009c024 2650 if (reuse_swap_page(old_page)) {
c44b6743
RR
2651 /*
2652 * The page is all ours. Move it to our anon_vma so
2653 * the rmap code will not search our parent or siblings.
2654 * Protected against the rmap code by the page lock.
2655 */
2656 page_move_anon_rmap(old_page, vma, address);
b009c024
ML
2657 unlock_page(old_page);
2658 goto reuse;
2659 }
ab967d86 2660 unlock_page(old_page);
ee6a6457 2661 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
d08b3851 2662 (VM_WRITE|VM_SHARED))) {
ee6a6457
PZ
2663 /*
2664 * Only catch write-faults on shared writable pages,
2665 * read-only shared pages can get COWed by
2666 * get_user_pages(.write=1, .force=1).
2667 */
9637a5ef 2668 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
c2ec175c
NP
2669 struct vm_fault vmf;
2670 int tmp;
2671
2672 vmf.virtual_address = (void __user *)(address &
2673 PAGE_MASK);
2674 vmf.pgoff = old_page->index;
2675 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2676 vmf.page = old_page;
2677
9637a5ef
DH
2678 /*
2679 * Notify the address space that the page is about to
2680 * become writable so that it can prohibit this or wait
2681 * for the page to get into an appropriate state.
2682 *
2683 * We do this without the lock held, so that it can
2684 * sleep if it needs to.
2685 */
2686 page_cache_get(old_page);
2687 pte_unmap_unlock(page_table, ptl);
2688
c2ec175c
NP
2689 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2690 if (unlikely(tmp &
2691 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2692 ret = tmp;
9637a5ef 2693 goto unwritable_page;
c2ec175c 2694 }
b827e496
NP
2695 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2696 lock_page(old_page);
2697 if (!old_page->mapping) {
2698 ret = 0; /* retry the fault */
2699 unlock_page(old_page);
2700 goto unwritable_page;
2701 }
2702 } else
2703 VM_BUG_ON(!PageLocked(old_page));
9637a5ef 2704
9637a5ef
DH
2705 /*
2706 * Since we dropped the lock we need to revalidate
2707 * the PTE as someone else may have changed it. If
2708 * they did, we just return, as we can count on the
2709 * MMU to tell us if they didn't also make it writable.
2710 */
2711 page_table = pte_offset_map_lock(mm, pmd, address,
2712 &ptl);
b827e496
NP
2713 if (!pte_same(*page_table, orig_pte)) {
2714 unlock_page(old_page);
9637a5ef 2715 goto unlock;
b827e496 2716 }
a200ee18
PZ
2717
2718 page_mkwrite = 1;
1da177e4 2719 }
d08b3851
PZ
2720 dirty_page = old_page;
2721 get_page(dirty_page);
9637a5ef 2722
251b97f5 2723reuse:
9637a5ef
DH
2724 flush_cache_page(vma, address, pte_pfn(orig_pte));
2725 entry = pte_mkyoung(orig_pte);
2726 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
954ffcb3 2727 if (ptep_set_access_flags(vma, address, page_table, entry,1))
4b3073e1 2728 update_mmu_cache(vma, address, page_table);
72ddc8f7 2729 pte_unmap_unlock(page_table, ptl);
9637a5ef 2730 ret |= VM_FAULT_WRITE;
72ddc8f7
ML
2731
2732 if (!dirty_page)
2733 return ret;
2734
2735 /*
2736 * Yes, Virginia, this is actually required to prevent a race
2737 * with clear_page_dirty_for_io() from clearing the page dirty
2738 * bit after it clear all dirty ptes, but before a racing
2739 * do_wp_page installs a dirty pte.
2740 *
a335b2e1 2741 * __do_fault is protected similarly.
72ddc8f7
ML
2742 */
2743 if (!page_mkwrite) {
2744 wait_on_page_locked(dirty_page);
2745 set_page_dirty_balance(dirty_page, page_mkwrite);
41c4d25f
JK
2746 /* file_update_time outside page_lock */
2747 if (vma->vm_file)
2748 file_update_time(vma->vm_file);
72ddc8f7
ML
2749 }
2750 put_page(dirty_page);
2751 if (page_mkwrite) {
2752 struct address_space *mapping = dirty_page->mapping;
2753
2754 set_page_dirty(dirty_page);
2755 unlock_page(dirty_page);
2756 page_cache_release(dirty_page);
2757 if (mapping) {
2758 /*
2759 * Some device drivers do not set page.mapping
2760 * but still dirty their pages
2761 */
2762 balance_dirty_pages_ratelimited(mapping);
2763 }
2764 }
2765
72ddc8f7 2766 return ret;
1da177e4 2767 }
1da177e4
LT
2768
2769 /*
2770 * Ok, we need to copy. Oh, well..
2771 */
b5810039 2772 page_cache_get(old_page);
920fc356 2773gotten:
8f4e2101 2774 pte_unmap_unlock(page_table, ptl);
1da177e4
LT
2775
2776 if (unlikely(anon_vma_prepare(vma)))
65500d23 2777 goto oom;
a13ea5b7 2778
62eede62 2779 if (is_zero_pfn(pte_pfn(orig_pte))) {
a13ea5b7
HD
2780 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2781 if (!new_page)
2782 goto oom;
2783 } else {
2784 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2785 if (!new_page)
2786 goto oom;
2787 cow_user_page(new_page, old_page, address, vma);
2788 }
2789 __SetPageUptodate(new_page);
2790
2c26fdd7 2791 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
8a9f3ccd
BS
2792 goto oom_free_new;
2793
6bdb913f 2794 mmun_start = address & PAGE_MASK;
1756954c 2795 mmun_end = mmun_start + PAGE_SIZE;
6bdb913f
HE
2796 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2797
1da177e4
LT
2798 /*
2799 * Re-check the pte - we dropped the lock
2800 */
8f4e2101 2801 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
65500d23 2802 if (likely(pte_same(*page_table, orig_pte))) {
920fc356 2803 if (old_page) {
920fc356 2804 if (!PageAnon(old_page)) {
34e55232
KH
2805 dec_mm_counter_fast(mm, MM_FILEPAGES);
2806 inc_mm_counter_fast(mm, MM_ANONPAGES);
920fc356
HD
2807 }
2808 } else
34e55232 2809 inc_mm_counter_fast(mm, MM_ANONPAGES);
eca35133 2810 flush_cache_page(vma, address, pte_pfn(orig_pte));
65500d23
HD
2811 entry = mk_pte(new_page, vma->vm_page_prot);
2812 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4ce072f1
SS
2813 /*
2814 * Clear the pte entry and flush it first, before updating the
2815 * pte with the new entry. This will avoid a race condition
2816 * seen in the presence of one thread doing SMC and another
2817 * thread doing COW.
2818 */
828502d3 2819 ptep_clear_flush(vma, address, page_table);
9617d95e 2820 page_add_new_anon_rmap(new_page, vma, address);
828502d3
IE
2821 /*
2822 * We call the notify macro here because, when using secondary
2823 * mmu page tables (such as kvm shadow page tables), we want the
2824 * new page to be mapped directly into the secondary page table.
2825 */
2826 set_pte_at_notify(mm, address, page_table, entry);
4b3073e1 2827 update_mmu_cache(vma, address, page_table);
945754a1
NP
2828 if (old_page) {
2829 /*
2830 * Only after switching the pte to the new page may
2831 * we remove the mapcount here. Otherwise another
2832 * process may come and find the rmap count decremented
2833 * before the pte is switched to the new page, and
2834 * "reuse" the old page writing into it while our pte
2835 * here still points into it and can be read by other
2836 * threads.
2837 *
2838 * The critical issue is to order this
2839 * page_remove_rmap with the ptp_clear_flush above.
2840 * Those stores are ordered by (if nothing else,)
2841 * the barrier present in the atomic_add_negative
2842 * in page_remove_rmap.
2843 *
2844 * Then the TLB flush in ptep_clear_flush ensures that
2845 * no process can access the old page before the
2846 * decremented mapcount is visible. And the old page
2847 * cannot be reused until after the decremented
2848 * mapcount is visible. So transitively, TLBs to
2849 * old page will be flushed before it can be reused.
2850 */
edc315fd 2851 page_remove_rmap(old_page);
945754a1
NP
2852 }
2853
1da177e4
LT
2854 /* Free the old page.. */
2855 new_page = old_page;
f33ea7f4 2856 ret |= VM_FAULT_WRITE;
8a9f3ccd
BS
2857 } else
2858 mem_cgroup_uncharge_page(new_page);
2859
6bdb913f
HE
2860 if (new_page)
2861 page_cache_release(new_page);
65500d23 2862unlock:
8f4e2101 2863 pte_unmap_unlock(page_table, ptl);
1756954c 2864 if (mmun_end > mmun_start)
6bdb913f 2865 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
e15f8c01
ML
2866 if (old_page) {
2867 /*
2868 * Don't let another task, with possibly unlocked vma,
2869 * keep the mlocked page.
2870 */
2871 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2872 lock_page(old_page); /* LRU manipulation */
2873 munlock_vma_page(old_page);
2874 unlock_page(old_page);
2875 }
2876 page_cache_release(old_page);
2877 }
f33ea7f4 2878 return ret;
8a9f3ccd 2879oom_free_new:
6dbf6d3b 2880 page_cache_release(new_page);
65500d23 2881oom:
66521d5a 2882 if (old_page)
920fc356 2883 page_cache_release(old_page);
1da177e4 2884 return VM_FAULT_OOM;
9637a5ef
DH
2885
2886unwritable_page:
2887 page_cache_release(old_page);
c2ec175c 2888 return ret;
1da177e4
LT
2889}
2890
97a89413 2891static void unmap_mapping_range_vma(struct vm_area_struct *vma,
1da177e4
LT
2892 unsigned long start_addr, unsigned long end_addr,
2893 struct zap_details *details)
2894{
f5cc4eef 2895 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
1da177e4
LT
2896}
2897
6b2dbba8 2898static inline void unmap_mapping_range_tree(struct rb_root *root,
1da177e4
LT
2899 struct zap_details *details)
2900{
2901 struct vm_area_struct *vma;
1da177e4
LT
2902 pgoff_t vba, vea, zba, zea;
2903
6b2dbba8 2904 vma_interval_tree_foreach(vma, root,
1da177e4 2905 details->first_index, details->last_index) {
1da177e4
LT
2906
2907 vba = vma->vm_pgoff;
d6e93217 2908 vea = vba + vma_pages(vma) - 1;
1da177e4
LT
2909 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2910 zba = details->first_index;
2911 if (zba < vba)
2912 zba = vba;
2913 zea = details->last_index;
2914 if (zea > vea)
2915 zea = vea;
2916
97a89413 2917 unmap_mapping_range_vma(vma,
1da177e4
LT
2918 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2919 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
97a89413 2920 details);
1da177e4
LT
2921 }
2922}
2923
2924static inline void unmap_mapping_range_list(struct list_head *head,
2925 struct zap_details *details)
2926{
2927 struct vm_area_struct *vma;
2928
2929 /*
2930 * In nonlinear VMAs there is no correspondence between virtual address
2931 * offset and file offset. So we must perform an exhaustive search
2932 * across *all* the pages in each nonlinear VMA, not just the pages
2933 * whose virtual address lies outside the file truncation point.
2934 */
6b2dbba8 2935 list_for_each_entry(vma, head, shared.nonlinear) {
1da177e4 2936 details->nonlinear_vma = vma;
97a89413 2937 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
1da177e4
LT
2938 }
2939}
2940
2941/**
72fd4a35 2942 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
3d41088f 2943 * @mapping: the address space containing mmaps to be unmapped.
1da177e4
LT
2944 * @holebegin: byte in first page to unmap, relative to the start of
2945 * the underlying file. This will be rounded down to a PAGE_SIZE
25d9e2d1 2946 * boundary. Note that this is different from truncate_pagecache(), which
1da177e4
LT
2947 * must keep the partial page. In contrast, we must get rid of
2948 * partial pages.
2949 * @holelen: size of prospective hole in bytes. This will be rounded
2950 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2951 * end of the file.
2952 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2953 * but 0 when invalidating pagecache, don't throw away private data.
2954 */
2955void unmap_mapping_range(struct address_space *mapping,
2956 loff_t const holebegin, loff_t const holelen, int even_cows)
2957{
2958 struct zap_details details;
2959 pgoff_t hba = holebegin >> PAGE_SHIFT;
2960 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2961
2962 /* Check for overflow. */
2963 if (sizeof(holelen) > sizeof(hlen)) {
2964 long long holeend =
2965 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2966 if (holeend & ~(long long)ULONG_MAX)
2967 hlen = ULONG_MAX - hba + 1;
2968 }
2969
2970 details.check_mapping = even_cows? NULL: mapping;
2971 details.nonlinear_vma = NULL;
2972 details.first_index = hba;
2973 details.last_index = hba + hlen - 1;
2974 if (details.last_index < details.first_index)
2975 details.last_index = ULONG_MAX;
1da177e4 2976
1da177e4 2977
3d48ae45 2978 mutex_lock(&mapping->i_mmap_mutex);
6b2dbba8 2979 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
1da177e4
LT
2980 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2981 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2982 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
3d48ae45 2983 mutex_unlock(&mapping->i_mmap_mutex);
1da177e4
LT
2984}
2985EXPORT_SYMBOL(unmap_mapping_range);
2986
1da177e4 2987/*
8f4e2101
HD
2988 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2989 * but allow concurrent faults), and pte mapped but not yet locked.
2990 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 2991 */
65500d23
HD
2992static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2993 unsigned long address, pte_t *page_table, pmd_t *pmd,
30c9f3a9 2994 unsigned int flags, pte_t orig_pte)
1da177e4 2995{
8f4e2101 2996 spinlock_t *ptl;
56f31801 2997 struct page *page, *swapcache;
65500d23 2998 swp_entry_t entry;
1da177e4 2999 pte_t pte;
d065bd81 3000 int locked;
56039efa 3001 struct mem_cgroup *ptr;
ad8c2ee8 3002 int exclusive = 0;
83c54070 3003 int ret = 0;
1da177e4 3004
4c21e2f2 3005 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
8f4e2101 3006 goto out;
65500d23
HD
3007
3008 entry = pte_to_swp_entry(orig_pte);
d1737fdb
AK
3009 if (unlikely(non_swap_entry(entry))) {
3010 if (is_migration_entry(entry)) {
3011 migration_entry_wait(mm, pmd, address);
3012 } else if (is_hwpoison_entry(entry)) {
3013 ret = VM_FAULT_HWPOISON;
3014 } else {
3015 print_bad_pte(vma, address, orig_pte, NULL);
d99be1a8 3016 ret = VM_FAULT_SIGBUS;
d1737fdb 3017 }
0697212a
CL
3018 goto out;
3019 }
0ff92245 3020 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
1da177e4
LT
3021 page = lookup_swap_cache(entry);
3022 if (!page) {
02098fea
HD
3023 page = swapin_readahead(entry,
3024 GFP_HIGHUSER_MOVABLE, vma, address);
1da177e4
LT
3025 if (!page) {
3026 /*
8f4e2101
HD
3027 * Back out if somebody else faulted in this pte
3028 * while we released the pte lock.
1da177e4 3029 */
8f4e2101 3030 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1da177e4
LT
3031 if (likely(pte_same(*page_table, orig_pte)))
3032 ret = VM_FAULT_OOM;
0ff92245 3033 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
65500d23 3034 goto unlock;
1da177e4
LT
3035 }
3036
3037 /* Had to read the page from swap area: Major fault */
3038 ret = VM_FAULT_MAJOR;
f8891e5e 3039 count_vm_event(PGMAJFAULT);
456f998e 3040 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
d1737fdb 3041 } else if (PageHWPoison(page)) {
71f72525
WF
3042 /*
3043 * hwpoisoned dirty swapcache pages are kept for killing
3044 * owner processes (which may be unknown at hwpoison time)
3045 */
d1737fdb
AK
3046 ret = VM_FAULT_HWPOISON;
3047 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
56f31801 3048 swapcache = page;
4779cb31 3049 goto out_release;
1da177e4
LT
3050 }
3051
56f31801 3052 swapcache = page;
d065bd81 3053 locked = lock_page_or_retry(page, mm, flags);
e709ffd6 3054
073e587e 3055 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
d065bd81
ML
3056 if (!locked) {
3057 ret |= VM_FAULT_RETRY;
3058 goto out_release;
3059 }
073e587e 3060
4969c119 3061 /*
31c4a3d3
HD
3062 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3063 * release the swapcache from under us. The page pin, and pte_same
3064 * test below, are not enough to exclude that. Even if it is still
3065 * swapcache, we need to check that the page's swap has not changed.
4969c119 3066 */
31c4a3d3 3067 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
4969c119
AA
3068 goto out_page;
3069
cbf86cfe
HD
3070 page = ksm_might_need_to_copy(page, vma, address);
3071 if (unlikely(!page)) {
3072 ret = VM_FAULT_OOM;
3073 page = swapcache;
cbf86cfe 3074 goto out_page;
5ad64688
HD
3075 }
3076
2c26fdd7 3077 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
8a9f3ccd 3078 ret = VM_FAULT_OOM;
bc43f75c 3079 goto out_page;
8a9f3ccd
BS
3080 }
3081
1da177e4 3082 /*
8f4e2101 3083 * Back out if somebody else already faulted in this pte.
1da177e4 3084 */
8f4e2101 3085 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
9e9bef07 3086 if (unlikely(!pte_same(*page_table, orig_pte)))
b8107480 3087 goto out_nomap;
b8107480
KK
3088
3089 if (unlikely(!PageUptodate(page))) {
3090 ret = VM_FAULT_SIGBUS;
3091 goto out_nomap;
1da177e4
LT
3092 }
3093
8c7c6e34
KH
3094 /*
3095 * The page isn't present yet, go ahead with the fault.
3096 *
3097 * Be careful about the sequence of operations here.
3098 * To get its accounting right, reuse_swap_page() must be called
3099 * while the page is counted on swap but not yet in mapcount i.e.
3100 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3101 * must be called after the swap_free(), or it will never succeed.
03f3c433
KH
3102 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3103 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3104 * in page->private. In this case, a record in swap_cgroup is silently
3105 * discarded at swap_free().
8c7c6e34 3106 */
1da177e4 3107
34e55232 3108 inc_mm_counter_fast(mm, MM_ANONPAGES);
b084d435 3109 dec_mm_counter_fast(mm, MM_SWAPENTS);
1da177e4 3110 pte = mk_pte(page, vma->vm_page_prot);
30c9f3a9 3111 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
1da177e4 3112 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
30c9f3a9 3113 flags &= ~FAULT_FLAG_WRITE;
9a5b489b 3114 ret |= VM_FAULT_WRITE;
ad8c2ee8 3115 exclusive = 1;
1da177e4 3116 }
1da177e4
LT
3117 flush_icache_page(vma, page);
3118 set_pte_at(mm, address, page_table, pte);
56f31801 3119 if (page == swapcache)
af34770e 3120 do_page_add_anon_rmap(page, vma, address, exclusive);
56f31801
HD
3121 else /* ksm created a completely new copy */
3122 page_add_new_anon_rmap(page, vma, address);
03f3c433
KH
3123 /* It's better to call commit-charge after rmap is established */
3124 mem_cgroup_commit_charge_swapin(page, ptr);
1da177e4 3125
c475a8ab 3126 swap_free(entry);
b291f000 3127 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
a2c43eed 3128 try_to_free_swap(page);
c475a8ab 3129 unlock_page(page);
56f31801 3130 if (page != swapcache) {
4969c119
AA
3131 /*
3132 * Hold the lock to avoid the swap entry to be reused
3133 * until we take the PT lock for the pte_same() check
3134 * (to avoid false positives from pte_same). For
3135 * further safety release the lock after the swap_free
3136 * so that the swap count won't change under a
3137 * parallel locked swapcache.
3138 */
3139 unlock_page(swapcache);
3140 page_cache_release(swapcache);
3141 }
c475a8ab 3142
30c9f3a9 3143 if (flags & FAULT_FLAG_WRITE) {
61469f1d
HD
3144 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3145 if (ret & VM_FAULT_ERROR)
3146 ret &= VM_FAULT_ERROR;
1da177e4
LT
3147 goto out;
3148 }
3149
3150 /* No need to invalidate - it was non-present before */
4b3073e1 3151 update_mmu_cache(vma, address, page_table);
65500d23 3152unlock:
8f4e2101 3153 pte_unmap_unlock(page_table, ptl);
1da177e4
LT
3154out:
3155 return ret;
b8107480 3156out_nomap:
7a81b88c 3157 mem_cgroup_cancel_charge_swapin(ptr);
8f4e2101 3158 pte_unmap_unlock(page_table, ptl);
bc43f75c 3159out_page:
b8107480 3160 unlock_page(page);
4779cb31 3161out_release:
b8107480 3162 page_cache_release(page);
56f31801 3163 if (page != swapcache) {
4969c119
AA
3164 unlock_page(swapcache);
3165 page_cache_release(swapcache);
3166 }
65500d23 3167 return ret;
1da177e4
LT
3168}
3169
320b2b8d 3170/*
8ca3eb08
TL
3171 * This is like a special single-page "expand_{down|up}wards()",
3172 * except we must first make sure that 'address{-|+}PAGE_SIZE'
320b2b8d 3173 * doesn't hit another vma.
320b2b8d
LT
3174 */
3175static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3176{
3177 address &= PAGE_MASK;
3178 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
0e8e50e2
LT
3179 struct vm_area_struct *prev = vma->vm_prev;
3180
3181 /*
3182 * Is there a mapping abutting this one below?
3183 *
3184 * That's only ok if it's the same stack mapping
3185 * that has gotten split..
3186 */
3187 if (prev && prev->vm_end == address)
3188 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
320b2b8d 3189
d05f3169 3190 expand_downwards(vma, address - PAGE_SIZE);
320b2b8d 3191 }
8ca3eb08
TL
3192 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3193 struct vm_area_struct *next = vma->vm_next;
3194
3195 /* As VM_GROWSDOWN but s/below/above/ */
3196 if (next && next->vm_start == address + PAGE_SIZE)
3197 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3198
3199 expand_upwards(vma, address + PAGE_SIZE);
3200 }
320b2b8d
LT
3201 return 0;
3202}
3203
1da177e4 3204/*
8f4e2101
HD
3205 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3206 * but allow concurrent faults), and pte mapped but not yet locked.
3207 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 3208 */
65500d23
HD
3209static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3210 unsigned long address, pte_t *page_table, pmd_t *pmd,
30c9f3a9 3211 unsigned int flags)
1da177e4 3212{
8f4e2101
HD
3213 struct page *page;
3214 spinlock_t *ptl;
1da177e4 3215 pte_t entry;
1da177e4 3216
11ac5524
LT
3217 pte_unmap(page_table);
3218
3219 /* Check if we need to add a guard page to the stack */
3220 if (check_stack_guard_page(vma, address) < 0)
320b2b8d
LT
3221 return VM_FAULT_SIGBUS;
3222
11ac5524 3223 /* Use the zero-page for reads */
62eede62
HD
3224 if (!(flags & FAULT_FLAG_WRITE)) {
3225 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3226 vma->vm_page_prot));
11ac5524 3227 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
a13ea5b7
HD
3228 if (!pte_none(*page_table))
3229 goto unlock;
3230 goto setpte;
3231 }
3232
557ed1fa 3233 /* Allocate our own private page. */
557ed1fa
NP
3234 if (unlikely(anon_vma_prepare(vma)))
3235 goto oom;
3236 page = alloc_zeroed_user_highpage_movable(vma, address);
3237 if (!page)
3238 goto oom;
52f37629
MK
3239 /*
3240 * The memory barrier inside __SetPageUptodate makes sure that
3241 * preceeding stores to the page contents become visible before
3242 * the set_pte_at() write.
3243 */
0ed361de 3244 __SetPageUptodate(page);
8f4e2101 3245
2c26fdd7 3246 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
8a9f3ccd
BS
3247 goto oom_free_page;
3248
557ed1fa 3249 entry = mk_pte(page, vma->vm_page_prot);
1ac0cb5d
HD
3250 if (vma->vm_flags & VM_WRITE)
3251 entry = pte_mkwrite(pte_mkdirty(entry));
1da177e4 3252
557ed1fa 3253 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1c2fb7a4 3254 if (!pte_none(*page_table))
557ed1fa 3255 goto release;
9ba69294 3256
34e55232 3257 inc_mm_counter_fast(mm, MM_ANONPAGES);
557ed1fa 3258 page_add_new_anon_rmap(page, vma, address);
a13ea5b7 3259setpte:
65500d23 3260 set_pte_at(mm, address, page_table, entry);
1da177e4
LT
3261
3262 /* No need to invalidate - it was non-present before */
4b3073e1 3263 update_mmu_cache(vma, address, page_table);
65500d23 3264unlock:
8f4e2101 3265 pte_unmap_unlock(page_table, ptl);
83c54070 3266 return 0;
8f4e2101 3267release:
8a9f3ccd 3268 mem_cgroup_uncharge_page(page);
8f4e2101
HD
3269 page_cache_release(page);
3270 goto unlock;
8a9f3ccd 3271oom_free_page:
6dbf6d3b 3272 page_cache_release(page);
65500d23 3273oom:
1da177e4
LT
3274 return VM_FAULT_OOM;
3275}
3276
3277/*
54cb8821 3278 * __do_fault() tries to create a new page mapping. It aggressively
1da177e4 3279 * tries to share with existing pages, but makes a separate copy if
54cb8821
NP
3280 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3281 * the next page fault.
1da177e4
LT
3282 *
3283 * As this is called only for pages that do not currently exist, we
3284 * do not need to flush old virtual caches or the TLB.
3285 *
8f4e2101 3286 * We enter with non-exclusive mmap_sem (to exclude vma changes,
16abfa08 3287 * but allow concurrent faults), and pte neither mapped nor locked.
8f4e2101 3288 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 3289 */
54cb8821 3290static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
16abfa08 3291 unsigned long address, pmd_t *pmd,
54cb8821 3292 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
1da177e4 3293{
16abfa08 3294 pte_t *page_table;
8f4e2101 3295 spinlock_t *ptl;
d0217ac0 3296 struct page *page;
1d65f86d 3297 struct page *cow_page;
1da177e4 3298 pte_t entry;
1da177e4 3299 int anon = 0;
d08b3851 3300 struct page *dirty_page = NULL;
d0217ac0
NP
3301 struct vm_fault vmf;
3302 int ret;
a200ee18 3303 int page_mkwrite = 0;
54cb8821 3304
1d65f86d
KH
3305 /*
3306 * If we do COW later, allocate page befor taking lock_page()
3307 * on the file cache page. This will reduce lock holding time.
3308 */
3309 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3310
3311 if (unlikely(anon_vma_prepare(vma)))
3312 return VM_FAULT_OOM;
3313
3314 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3315 if (!cow_page)
3316 return VM_FAULT_OOM;
3317
3318 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3319 page_cache_release(cow_page);
3320 return VM_FAULT_OOM;
3321 }
3322 } else
3323 cow_page = NULL;
3324
d0217ac0
NP
3325 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3326 vmf.pgoff = pgoff;
3327 vmf.flags = flags;
3328 vmf.page = NULL;
1da177e4 3329
3c18ddd1 3330 ret = vma->vm_ops->fault(vma, &vmf);
d065bd81
ML
3331 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3332 VM_FAULT_RETRY)))
1d65f86d 3333 goto uncharge_out;
1da177e4 3334
a3b947ea
AK
3335 if (unlikely(PageHWPoison(vmf.page))) {
3336 if (ret & VM_FAULT_LOCKED)
3337 unlock_page(vmf.page);
1d65f86d
KH
3338 ret = VM_FAULT_HWPOISON;
3339 goto uncharge_out;
a3b947ea
AK
3340 }
3341
d00806b1 3342 /*
d0217ac0 3343 * For consistency in subsequent calls, make the faulted page always
d00806b1
NP
3344 * locked.
3345 */
83c54070 3346 if (unlikely(!(ret & VM_FAULT_LOCKED)))
d0217ac0 3347 lock_page(vmf.page);
54cb8821 3348 else
d0217ac0 3349 VM_BUG_ON(!PageLocked(vmf.page));
d00806b1 3350
1da177e4
LT
3351 /*
3352 * Should we do an early C-O-W break?
3353 */
d0217ac0 3354 page = vmf.page;
54cb8821 3355 if (flags & FAULT_FLAG_WRITE) {
9637a5ef 3356 if (!(vma->vm_flags & VM_SHARED)) {
1d65f86d 3357 page = cow_page;
54cb8821 3358 anon = 1;
d0217ac0 3359 copy_user_highpage(page, vmf.page, address, vma);
0ed361de 3360 __SetPageUptodate(page);
9637a5ef 3361 } else {
54cb8821
NP
3362 /*
3363 * If the page will be shareable, see if the backing
9637a5ef 3364 * address space wants to know that the page is about
54cb8821
NP
3365 * to become writable
3366 */
69676147 3367 if (vma->vm_ops->page_mkwrite) {
c2ec175c
NP
3368 int tmp;
3369
69676147 3370 unlock_page(page);
b827e496 3371 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
c2ec175c
NP
3372 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3373 if (unlikely(tmp &
3374 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3375 ret = tmp;
b827e496 3376 goto unwritable_page;
d0217ac0 3377 }
b827e496
NP
3378 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3379 lock_page(page);
3380 if (!page->mapping) {
3381 ret = 0; /* retry the fault */
3382 unlock_page(page);
3383 goto unwritable_page;
3384 }
3385 } else
3386 VM_BUG_ON(!PageLocked(page));
a200ee18 3387 page_mkwrite = 1;
9637a5ef
DH
3388 }
3389 }
54cb8821 3390
1da177e4
LT
3391 }
3392
8f4e2101 3393 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1da177e4
LT
3394
3395 /*
3396 * This silly early PAGE_DIRTY setting removes a race
3397 * due to the bad i386 page protection. But it's valid
3398 * for other architectures too.
3399 *
30c9f3a9 3400 * Note that if FAULT_FLAG_WRITE is set, we either now have
1da177e4
LT
3401 * an exclusive copy of the page, or this is a shared mapping,
3402 * so we can make it writable and dirty to avoid having to
3403 * handle that later.
3404 */
3405 /* Only go through if we didn't race with anybody else... */
1c2fb7a4 3406 if (likely(pte_same(*page_table, orig_pte))) {
d00806b1
NP
3407 flush_icache_page(vma, page);
3408 entry = mk_pte(page, vma->vm_page_prot);
54cb8821 3409 if (flags & FAULT_FLAG_WRITE)
1da177e4 3410 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1da177e4 3411 if (anon) {
34e55232 3412 inc_mm_counter_fast(mm, MM_ANONPAGES);
64d6519d 3413 page_add_new_anon_rmap(page, vma, address);
f57e88a8 3414 } else {
34e55232 3415 inc_mm_counter_fast(mm, MM_FILEPAGES);
d00806b1 3416 page_add_file_rmap(page);
54cb8821 3417 if (flags & FAULT_FLAG_WRITE) {
d00806b1 3418 dirty_page = page;
d08b3851
PZ
3419 get_page(dirty_page);
3420 }
4294621f 3421 }
64d6519d 3422 set_pte_at(mm, address, page_table, entry);
d00806b1
NP
3423
3424 /* no need to invalidate: a not-present page won't be cached */
4b3073e1 3425 update_mmu_cache(vma, address, page_table);
1da177e4 3426 } else {
1d65f86d
KH
3427 if (cow_page)
3428 mem_cgroup_uncharge_page(cow_page);
d00806b1
NP
3429 if (anon)
3430 page_cache_release(page);
3431 else
54cb8821 3432 anon = 1; /* no anon but release faulted_page */
1da177e4
LT
3433 }
3434
8f4e2101 3435 pte_unmap_unlock(page_table, ptl);
d00806b1 3436
b827e496
NP
3437 if (dirty_page) {
3438 struct address_space *mapping = page->mapping;
41c4d25f 3439 int dirtied = 0;
8f7b3d15 3440
b827e496 3441 if (set_page_dirty(dirty_page))
41c4d25f 3442 dirtied = 1;
b827e496 3443 unlock_page(dirty_page);
d08b3851 3444 put_page(dirty_page);
41c4d25f 3445 if ((dirtied || page_mkwrite) && mapping) {
b827e496
NP
3446 /*
3447 * Some device drivers do not set page.mapping but still
3448 * dirty their pages
3449 */
3450 balance_dirty_pages_ratelimited(mapping);
3451 }
3452
3453 /* file_update_time outside page_lock */
41c4d25f 3454 if (vma->vm_file && !page_mkwrite)
b827e496
NP
3455 file_update_time(vma->vm_file);
3456 } else {
3457 unlock_page(vmf.page);
3458 if (anon)
3459 page_cache_release(vmf.page);
d08b3851 3460 }
d00806b1 3461
83c54070 3462 return ret;
b827e496
NP
3463
3464unwritable_page:
3465 page_cache_release(page);
3466 return ret;
1d65f86d
KH
3467uncharge_out:
3468 /* fs's fault handler get error */
3469 if (cow_page) {
3470 mem_cgroup_uncharge_page(cow_page);
3471 page_cache_release(cow_page);
3472 }
3473 return ret;
54cb8821 3474}
d00806b1 3475
54cb8821
NP
3476static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3477 unsigned long address, pte_t *page_table, pmd_t *pmd,
30c9f3a9 3478 unsigned int flags, pte_t orig_pte)
54cb8821
NP
3479{
3480 pgoff_t pgoff = (((address & PAGE_MASK)
0da7e01f 3481 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
54cb8821 3482
16abfa08
HD
3483 pte_unmap(page_table);
3484 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
54cb8821
NP
3485}
3486
1da177e4
LT
3487/*
3488 * Fault of a previously existing named mapping. Repopulate the pte
3489 * from the encoded file_pte if possible. This enables swappable
3490 * nonlinear vmas.
8f4e2101
HD
3491 *
3492 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3493 * but allow concurrent faults), and pte mapped but not yet locked.
3494 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 3495 */
d0217ac0 3496static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
65500d23 3497 unsigned long address, pte_t *page_table, pmd_t *pmd,
30c9f3a9 3498 unsigned int flags, pte_t orig_pte)
1da177e4 3499{
65500d23 3500 pgoff_t pgoff;
1da177e4 3501
30c9f3a9
LT
3502 flags |= FAULT_FLAG_NONLINEAR;
3503
4c21e2f2 3504 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
83c54070 3505 return 0;
1da177e4 3506
2509ef26 3507 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
65500d23
HD
3508 /*
3509 * Page table corrupted: show pte and kill process.
3510 */
3dc14741 3511 print_bad_pte(vma, address, orig_pte, NULL);
d99be1a8 3512 return VM_FAULT_SIGBUS;
65500d23 3513 }
65500d23
HD
3514
3515 pgoff = pte_to_pgoff(orig_pte);
16abfa08 3516 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
1da177e4
LT
3517}
3518
9532fec1
MG
3519int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3520 unsigned long addr, int current_nid)
3521{
3522 get_page(page);
3523
3524 count_vm_numa_event(NUMA_HINT_FAULTS);
3525 if (current_nid == numa_node_id())
3526 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3527
3528 return mpol_misplaced(page, vma, addr);
3529}
3530
d10e63f2
MG
3531int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3532 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3533{
4daae3b4 3534 struct page *page = NULL;
d10e63f2 3535 spinlock_t *ptl;
cbee9f88
PZ
3536 int current_nid = -1;
3537 int target_nid;
b8593bfd 3538 bool migrated = false;
d10e63f2
MG
3539
3540 /*
3541 * The "pte" at this point cannot be used safely without
3542 * validation through pte_unmap_same(). It's of NUMA type but
3543 * the pfn may be screwed if the read is non atomic.
3544 *
3545 * ptep_modify_prot_start is not called as this is clearing
3546 * the _PAGE_NUMA bit and it is not really expected that there
3547 * would be concurrent hardware modifications to the PTE.
3548 */
3549 ptl = pte_lockptr(mm, pmd);
3550 spin_lock(ptl);
4daae3b4
MG
3551 if (unlikely(!pte_same(*ptep, pte))) {
3552 pte_unmap_unlock(ptep, ptl);
3553 goto out;
3554 }
3555
d10e63f2
MG
3556 pte = pte_mknonnuma(pte);
3557 set_pte_at(mm, addr, ptep, pte);
3558 update_mmu_cache(vma, addr, ptep);
3559
3560 page = vm_normal_page(vma, addr, pte);
3561 if (!page) {
3562 pte_unmap_unlock(ptep, ptl);
3563 return 0;
3564 }
3565
4daae3b4 3566 current_nid = page_to_nid(page);
9532fec1 3567 target_nid = numa_migrate_prep(page, vma, addr, current_nid);
d10e63f2 3568 pte_unmap_unlock(ptep, ptl);
4daae3b4
MG
3569 if (target_nid == -1) {
3570 /*
3571 * Account for the fault against the current node if it not
3572 * being replaced regardless of where the page is located.
3573 */
3574 current_nid = numa_node_id();
3575 put_page(page);
3576 goto out;
3577 }
3578
3579 /* Migrate to the requested node */
b8593bfd
MG
3580 migrated = migrate_misplaced_page(page, target_nid);
3581 if (migrated)
4daae3b4
MG
3582 current_nid = target_nid;
3583
3584out:
9532fec1 3585 if (current_nid != -1)
b8593bfd 3586 task_numa_fault(current_nid, 1, migrated);
d10e63f2
MG
3587 return 0;
3588}
3589
3590/* NUMA hinting page fault entry point for regular pmds */
3591#ifdef CONFIG_NUMA_BALANCING
3592static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3593 unsigned long addr, pmd_t *pmdp)
3594{
3595 pmd_t pmd;
3596 pte_t *pte, *orig_pte;
3597 unsigned long _addr = addr & PMD_MASK;
3598 unsigned long offset;
3599 spinlock_t *ptl;
3600 bool numa = false;
03c5a6e1 3601 int local_nid = numa_node_id();
d10e63f2
MG
3602
3603 spin_lock(&mm->page_table_lock);
3604 pmd = *pmdp;
3605 if (pmd_numa(pmd)) {
3606 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3607 numa = true;
3608 }
3609 spin_unlock(&mm->page_table_lock);
3610
3611 if (!numa)
3612 return 0;
3613
3614 /* we're in a page fault so some vma must be in the range */
3615 BUG_ON(!vma);
3616 BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3617 offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3618 VM_BUG_ON(offset >= PMD_SIZE);
3619 orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3620 pte += offset >> PAGE_SHIFT;
3621 for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3622 pte_t pteval = *pte;
3623 struct page *page;
9532fec1
MG
3624 int curr_nid = local_nid;
3625 int target_nid;
b8593bfd 3626 bool migrated;
d10e63f2
MG
3627 if (!pte_present(pteval))
3628 continue;
3629 if (!pte_numa(pteval))
3630 continue;
3631 if (addr >= vma->vm_end) {
3632 vma = find_vma(mm, addr);
3633 /* there's a pte present so there must be a vma */
3634 BUG_ON(!vma);
3635 BUG_ON(addr < vma->vm_start);
3636 }
3637 if (pte_numa(pteval)) {
3638 pteval = pte_mknonnuma(pteval);
3639 set_pte_at(mm, addr, pte, pteval);
3640 }
3641 page = vm_normal_page(vma, addr, pteval);
3642 if (unlikely(!page))
3643 continue;
cbee9f88
PZ
3644 /* only check non-shared pages */
3645 if (unlikely(page_mapcount(page) != 1))
3646 continue;
cbee9f88 3647
9532fec1
MG
3648 /*
3649 * Note that the NUMA fault is later accounted to either
3650 * the node that is currently running or where the page is
3651 * migrated to.
3652 */
3653 curr_nid = local_nid;
3654 target_nid = numa_migrate_prep(page, vma, addr,
3655 page_to_nid(page));
3656 if (target_nid == -1) {
3657 put_page(page);
3658 continue;
3659 }
cbee9f88 3660
9532fec1
MG
3661 /* Migrate to the requested node */
3662 pte_unmap_unlock(pte, ptl);
b8593bfd
MG
3663 migrated = migrate_misplaced_page(page, target_nid);
3664 if (migrated)
9532fec1 3665 curr_nid = target_nid;
b8593bfd 3666 task_numa_fault(curr_nid, 1, migrated);
03c5a6e1 3667
cbee9f88 3668 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
d10e63f2
MG
3669 }
3670 pte_unmap_unlock(orig_pte, ptl);
3671
3672 return 0;
3673}
3674#else
3675static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3676 unsigned long addr, pmd_t *pmdp)
3677{
3678 BUG();
b3dd2070 3679 return 0;
d10e63f2
MG
3680}
3681#endif /* CONFIG_NUMA_BALANCING */
3682
1da177e4
LT
3683/*
3684 * These routines also need to handle stuff like marking pages dirty
3685 * and/or accessed for architectures that don't do it in hardware (most
3686 * RISC architectures). The early dirtying is also good on the i386.
3687 *
3688 * There is also a hook called "update_mmu_cache()" that architectures
3689 * with external mmu caches can use to update those (ie the Sparc or
3690 * PowerPC hashed page tables that act as extended TLBs).
3691 *
c74df32c
HD
3692 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3693 * but allow concurrent faults), and pte mapped but not yet locked.
3694 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 3695 */
71e3aac0
AA
3696int handle_pte_fault(struct mm_struct *mm,
3697 struct vm_area_struct *vma, unsigned long address,
3698 pte_t *pte, pmd_t *pmd, unsigned int flags)
1da177e4
LT
3699{
3700 pte_t entry;
8f4e2101 3701 spinlock_t *ptl;
1da177e4 3702
8dab5241 3703 entry = *pte;
1da177e4 3704 if (!pte_present(entry)) {
65500d23 3705 if (pte_none(entry)) {
f4b81804 3706 if (vma->vm_ops) {
3c18ddd1 3707 if (likely(vma->vm_ops->fault))
54cb8821 3708 return do_linear_fault(mm, vma, address,
30c9f3a9 3709 pte, pmd, flags, entry);
f4b81804
JS
3710 }
3711 return do_anonymous_page(mm, vma, address,
30c9f3a9 3712 pte, pmd, flags);
65500d23 3713 }
1da177e4 3714 if (pte_file(entry))
d0217ac0 3715 return do_nonlinear_fault(mm, vma, address,
30c9f3a9 3716 pte, pmd, flags, entry);
65500d23 3717 return do_swap_page(mm, vma, address,
30c9f3a9 3718 pte, pmd, flags, entry);
1da177e4
LT
3719 }
3720
d10e63f2
MG
3721 if (pte_numa(entry))
3722 return do_numa_page(mm, vma, address, entry, pte, pmd);
3723
4c21e2f2 3724 ptl = pte_lockptr(mm, pmd);
8f4e2101
HD
3725 spin_lock(ptl);
3726 if (unlikely(!pte_same(*pte, entry)))
3727 goto unlock;
30c9f3a9 3728 if (flags & FAULT_FLAG_WRITE) {
1da177e4 3729 if (!pte_write(entry))
8f4e2101
HD
3730 return do_wp_page(mm, vma, address,
3731 pte, pmd, ptl, entry);
1da177e4
LT
3732 entry = pte_mkdirty(entry);
3733 }
3734 entry = pte_mkyoung(entry);
30c9f3a9 3735 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
4b3073e1 3736 update_mmu_cache(vma, address, pte);
1a44e149
AA
3737 } else {
3738 /*
3739 * This is needed only for protection faults but the arch code
3740 * is not yet telling us if this is a protection fault or not.
3741 * This still avoids useless tlb flushes for .text page faults
3742 * with threads.
3743 */
30c9f3a9 3744 if (flags & FAULT_FLAG_WRITE)
61c77326 3745 flush_tlb_fix_spurious_fault(vma, address);
1a44e149 3746 }
8f4e2101
HD
3747unlock:
3748 pte_unmap_unlock(pte, ptl);
83c54070 3749 return 0;
1da177e4
LT
3750}
3751
3752/*
3753 * By the time we get here, we already hold the mm semaphore
3754 */
83c54070 3755int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
d06063cc 3756 unsigned long address, unsigned int flags)
1da177e4
LT
3757{
3758 pgd_t *pgd;
3759 pud_t *pud;
3760 pmd_t *pmd;
3761 pte_t *pte;
3762
3763 __set_current_state(TASK_RUNNING);
3764
f8891e5e 3765 count_vm_event(PGFAULT);
456f998e 3766 mem_cgroup_count_vm_event(mm, PGFAULT);
1da177e4 3767
34e55232
KH
3768 /* do counter updates before entering really critical section. */
3769 check_sync_rss_stat(current);
3770
ac9b9c66 3771 if (unlikely(is_vm_hugetlb_page(vma)))
30c9f3a9 3772 return hugetlb_fault(mm, vma, address, flags);
1da177e4 3773
1f1d06c3 3774retry:
1da177e4 3775 pgd = pgd_offset(mm, address);
1da177e4
LT
3776 pud = pud_alloc(mm, pgd, address);
3777 if (!pud)
c74df32c 3778 return VM_FAULT_OOM;
1da177e4
LT
3779 pmd = pmd_alloc(mm, pud, address);
3780 if (!pmd)
c74df32c 3781 return VM_FAULT_OOM;
71e3aac0
AA
3782 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3783 if (!vma->vm_ops)
3784 return do_huge_pmd_anonymous_page(mm, vma, address,
3785 pmd, flags);
3786 } else {
3787 pmd_t orig_pmd = *pmd;
1f1d06c3
DR
3788 int ret;
3789
71e3aac0
AA
3790 barrier();
3791 if (pmd_trans_huge(orig_pmd)) {
a1dd450b
WD
3792 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3793
e53289c0
LT
3794 /*
3795 * If the pmd is splitting, return and retry the
3796 * the fault. Alternative: wait until the split
3797 * is done, and goto retry.
3798 */
3799 if (pmd_trans_splitting(orig_pmd))
3800 return 0;
3801
3d59eebc 3802 if (pmd_numa(orig_pmd))
4daae3b4 3803 return do_huge_pmd_numa_page(mm, vma, address,
d10e63f2
MG
3804 orig_pmd, pmd);
3805
3d59eebc 3806 if (dirty && !pmd_write(orig_pmd)) {
1f1d06c3
DR
3807 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3808 orig_pmd);
3809 /*
3810 * If COW results in an oom, the huge pmd will
3811 * have been split, so retry the fault on the
3812 * pte for a smaller charge.
3813 */
3814 if (unlikely(ret & VM_FAULT_OOM))
3815 goto retry;
3816 return ret;
a1dd450b
WD
3817 } else {
3818 huge_pmd_set_accessed(mm, vma, address, pmd,
3819 orig_pmd, dirty);
1f1d06c3 3820 }
d10e63f2 3821
71e3aac0
AA
3822 return 0;
3823 }
3824 }
3825
d10e63f2
MG
3826 if (pmd_numa(*pmd))
3827 return do_pmd_numa_page(mm, vma, address, pmd);
3828
71e3aac0
AA
3829 /*
3830 * Use __pte_alloc instead of pte_alloc_map, because we can't
3831 * run pte_offset_map on the pmd, if an huge pmd could
3832 * materialize from under us from a different thread.
3833 */
4fd01770
MG
3834 if (unlikely(pmd_none(*pmd)) &&
3835 unlikely(__pte_alloc(mm, vma, pmd, address)))
c74df32c 3836 return VM_FAULT_OOM;
71e3aac0
AA
3837 /* if an huge pmd materialized from under us just retry later */
3838 if (unlikely(pmd_trans_huge(*pmd)))
3839 return 0;
3840 /*
3841 * A regular pmd is established and it can't morph into a huge pmd
3842 * from under us anymore at this point because we hold the mmap_sem
3843 * read mode and khugepaged takes it in write mode. So now it's
3844 * safe to run pte_offset_map().
3845 */
3846 pte = pte_offset_map(pmd, address);
1da177e4 3847
30c9f3a9 3848 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
1da177e4
LT
3849}
3850
3851#ifndef __PAGETABLE_PUD_FOLDED
3852/*
3853 * Allocate page upper directory.
872fec16 3854 * We've already handled the fast-path in-line.
1da177e4 3855 */
1bb3630e 3856int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1da177e4 3857{
c74df32c
HD
3858 pud_t *new = pud_alloc_one(mm, address);
3859 if (!new)
1bb3630e 3860 return -ENOMEM;
1da177e4 3861
362a61ad
NP
3862 smp_wmb(); /* See comment in __pte_alloc */
3863
872fec16 3864 spin_lock(&mm->page_table_lock);
1bb3630e 3865 if (pgd_present(*pgd)) /* Another has populated it */
5e541973 3866 pud_free(mm, new);
1bb3630e
HD
3867 else
3868 pgd_populate(mm, pgd, new);
c74df32c 3869 spin_unlock(&mm->page_table_lock);
1bb3630e 3870 return 0;
1da177e4
LT
3871}
3872#endif /* __PAGETABLE_PUD_FOLDED */
3873
3874#ifndef __PAGETABLE_PMD_FOLDED
3875/*
3876 * Allocate page middle directory.
872fec16 3877 * We've already handled the fast-path in-line.
1da177e4 3878 */
1bb3630e 3879int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
1da177e4 3880{
c74df32c
HD
3881 pmd_t *new = pmd_alloc_one(mm, address);
3882 if (!new)
1bb3630e 3883 return -ENOMEM;
1da177e4 3884
362a61ad
NP
3885 smp_wmb(); /* See comment in __pte_alloc */
3886
872fec16 3887 spin_lock(&mm->page_table_lock);
1da177e4 3888#ifndef __ARCH_HAS_4LEVEL_HACK
1bb3630e 3889 if (pud_present(*pud)) /* Another has populated it */
5e541973 3890 pmd_free(mm, new);
1bb3630e
HD
3891 else
3892 pud_populate(mm, pud, new);
1da177e4 3893#else
1bb3630e 3894 if (pgd_present(*pud)) /* Another has populated it */
5e541973 3895 pmd_free(mm, new);
1bb3630e
HD
3896 else
3897 pgd_populate(mm, pud, new);
1da177e4 3898#endif /* __ARCH_HAS_4LEVEL_HACK */
c74df32c 3899 spin_unlock(&mm->page_table_lock);
1bb3630e 3900 return 0;
e0f39591 3901}
1da177e4
LT
3902#endif /* __PAGETABLE_PMD_FOLDED */
3903
1da177e4
LT
3904#if !defined(__HAVE_ARCH_GATE_AREA)
3905
3906#if defined(AT_SYSINFO_EHDR)
5ce7852c 3907static struct vm_area_struct gate_vma;
1da177e4
LT
3908
3909static int __init gate_vma_init(void)
3910{
3911 gate_vma.vm_mm = NULL;
3912 gate_vma.vm_start = FIXADDR_USER_START;
3913 gate_vma.vm_end = FIXADDR_USER_END;
b6558c4a
RM
3914 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3915 gate_vma.vm_page_prot = __P101;
909af768 3916
1da177e4
LT
3917 return 0;
3918}
3919__initcall(gate_vma_init);
3920#endif
3921
31db58b3 3922struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
1da177e4
LT
3923{
3924#ifdef AT_SYSINFO_EHDR
3925 return &gate_vma;
3926#else
3927 return NULL;
3928#endif
3929}
3930
cae5d390 3931int in_gate_area_no_mm(unsigned long addr)
1da177e4
LT
3932{
3933#ifdef AT_SYSINFO_EHDR
3934 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3935 return 1;
3936#endif
3937 return 0;
3938}
3939
3940#endif /* __HAVE_ARCH_GATE_AREA */
0ec76a11 3941
1b36ba81 3942static int __follow_pte(struct mm_struct *mm, unsigned long address,
f8ad0f49
JW
3943 pte_t **ptepp, spinlock_t **ptlp)
3944{
3945 pgd_t *pgd;
3946 pud_t *pud;
3947 pmd_t *pmd;
3948 pte_t *ptep;
3949
3950 pgd = pgd_offset(mm, address);
3951 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3952 goto out;
3953
3954 pud = pud_offset(pgd, address);
3955 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3956 goto out;
3957
3958 pmd = pmd_offset(pud, address);
f66055ab 3959 VM_BUG_ON(pmd_trans_huge(*pmd));
f8ad0f49
JW
3960 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3961 goto out;
3962
3963 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3964 if (pmd_huge(*pmd))
3965 goto out;
3966
3967 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3968 if (!ptep)
3969 goto out;
3970 if (!pte_present(*ptep))
3971 goto unlock;
3972 *ptepp = ptep;
3973 return 0;
3974unlock:
3975 pte_unmap_unlock(ptep, *ptlp);
3976out:
3977 return -EINVAL;
3978}
3979
1b36ba81
NK
3980static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3981 pte_t **ptepp, spinlock_t **ptlp)
3982{
3983 int res;
3984
3985 /* (void) is needed to make gcc happy */
3986 (void) __cond_lock(*ptlp,
3987 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3988 return res;
3989}
3990
3b6748e2
JW
3991/**
3992 * follow_pfn - look up PFN at a user virtual address
3993 * @vma: memory mapping
3994 * @address: user virtual address
3995 * @pfn: location to store found PFN
3996 *
3997 * Only IO mappings and raw PFN mappings are allowed.
3998 *
3999 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4000 */
4001int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4002 unsigned long *pfn)
4003{
4004 int ret = -EINVAL;
4005 spinlock_t *ptl;
4006 pte_t *ptep;
4007
4008 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4009 return ret;
4010
4011 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4012 if (ret)
4013 return ret;
4014 *pfn = pte_pfn(*ptep);
4015 pte_unmap_unlock(ptep, ptl);
4016 return 0;
4017}
4018EXPORT_SYMBOL(follow_pfn);
4019
28b2ee20 4020#ifdef CONFIG_HAVE_IOREMAP_PROT
d87fe660 4021int follow_phys(struct vm_area_struct *vma,
4022 unsigned long address, unsigned int flags,
4023 unsigned long *prot, resource_size_t *phys)
28b2ee20 4024{
03668a4d 4025 int ret = -EINVAL;
28b2ee20
RR
4026 pte_t *ptep, pte;
4027 spinlock_t *ptl;
28b2ee20 4028
d87fe660 4029 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4030 goto out;
28b2ee20 4031
03668a4d 4032 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
d87fe660 4033 goto out;
28b2ee20 4034 pte = *ptep;
03668a4d 4035
28b2ee20
RR
4036 if ((flags & FOLL_WRITE) && !pte_write(pte))
4037 goto unlock;
28b2ee20
RR
4038
4039 *prot = pgprot_val(pte_pgprot(pte));
03668a4d 4040 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
28b2ee20 4041
03668a4d 4042 ret = 0;
28b2ee20
RR
4043unlock:
4044 pte_unmap_unlock(ptep, ptl);
4045out:
d87fe660 4046 return ret;
28b2ee20
RR
4047}
4048
4049int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4050 void *buf, int len, int write)
4051{
4052 resource_size_t phys_addr;
4053 unsigned long prot = 0;
2bc7273b 4054 void __iomem *maddr;
28b2ee20
RR
4055 int offset = addr & (PAGE_SIZE-1);
4056
d87fe660 4057 if (follow_phys(vma, addr, write, &prot, &phys_addr))
28b2ee20
RR
4058 return -EINVAL;
4059
4060 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4061 if (write)
4062 memcpy_toio(maddr + offset, buf, len);
4063 else
4064 memcpy_fromio(buf, maddr + offset, len);
4065 iounmap(maddr);
4066
4067 return len;
4068}
4069#endif
4070
0ec76a11 4071/*
206cb636
SW
4072 * Access another process' address space as given in mm. If non-NULL, use the
4073 * given task for page fault accounting.
0ec76a11 4074 */
206cb636
SW
4075static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4076 unsigned long addr, void *buf, int len, int write)
0ec76a11 4077{
0ec76a11 4078 struct vm_area_struct *vma;
0ec76a11
DH
4079 void *old_buf = buf;
4080
0ec76a11 4081 down_read(&mm->mmap_sem);
183ff22b 4082 /* ignore errors, just check how much was successfully transferred */
0ec76a11
DH
4083 while (len) {
4084 int bytes, ret, offset;
4085 void *maddr;
28b2ee20 4086 struct page *page = NULL;
0ec76a11
DH
4087
4088 ret = get_user_pages(tsk, mm, addr, 1,
4089 write, 1, &page, &vma);
28b2ee20
RR
4090 if (ret <= 0) {
4091 /*
4092 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4093 * we can access using slightly different code.
4094 */
4095#ifdef CONFIG_HAVE_IOREMAP_PROT
4096 vma = find_vma(mm, addr);
fe936dfc 4097 if (!vma || vma->vm_start > addr)
28b2ee20
RR
4098 break;
4099 if (vma->vm_ops && vma->vm_ops->access)
4100 ret = vma->vm_ops->access(vma, addr, buf,
4101 len, write);
4102 if (ret <= 0)
4103#endif
4104 break;
4105 bytes = ret;
0ec76a11 4106 } else {
28b2ee20
RR
4107 bytes = len;
4108 offset = addr & (PAGE_SIZE-1);
4109 if (bytes > PAGE_SIZE-offset)
4110 bytes = PAGE_SIZE-offset;
4111
4112 maddr = kmap(page);
4113 if (write) {
4114 copy_to_user_page(vma, page, addr,
4115 maddr + offset, buf, bytes);
4116 set_page_dirty_lock(page);
4117 } else {
4118 copy_from_user_page(vma, page, addr,
4119 buf, maddr + offset, bytes);
4120 }
4121 kunmap(page);
4122 page_cache_release(page);
0ec76a11 4123 }
0ec76a11
DH
4124 len -= bytes;
4125 buf += bytes;
4126 addr += bytes;
4127 }
4128 up_read(&mm->mmap_sem);
0ec76a11
DH
4129
4130 return buf - old_buf;
4131}
03252919 4132
5ddd36b9 4133/**
ae91dbfc 4134 * access_remote_vm - access another process' address space
5ddd36b9
SW
4135 * @mm: the mm_struct of the target address space
4136 * @addr: start address to access
4137 * @buf: source or destination buffer
4138 * @len: number of bytes to transfer
4139 * @write: whether the access is a write
4140 *
4141 * The caller must hold a reference on @mm.
4142 */
4143int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4144 void *buf, int len, int write)
4145{
4146 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4147}
4148
206cb636
SW
4149/*
4150 * Access another process' address space.
4151 * Source/target buffer must be kernel space,
4152 * Do not walk the page table directly, use get_user_pages
4153 */
4154int access_process_vm(struct task_struct *tsk, unsigned long addr,
4155 void *buf, int len, int write)
4156{
4157 struct mm_struct *mm;
4158 int ret;
4159
4160 mm = get_task_mm(tsk);
4161 if (!mm)
4162 return 0;
4163
4164 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4165 mmput(mm);
4166
4167 return ret;
4168}
4169
03252919
AK
4170/*
4171 * Print the name of a VMA.
4172 */
4173void print_vma_addr(char *prefix, unsigned long ip)
4174{
4175 struct mm_struct *mm = current->mm;
4176 struct vm_area_struct *vma;
4177
e8bff74a
IM
4178 /*
4179 * Do not print if we are in atomic
4180 * contexts (in exception stacks, etc.):
4181 */
4182 if (preempt_count())
4183 return;
4184
03252919
AK
4185 down_read(&mm->mmap_sem);
4186 vma = find_vma(mm, ip);
4187 if (vma && vma->vm_file) {
4188 struct file *f = vma->vm_file;
4189 char *buf = (char *)__get_free_page(GFP_KERNEL);
4190 if (buf) {
2fbc57c5 4191 char *p;
03252919 4192
cf28b486 4193 p = d_path(&f->f_path, buf, PAGE_SIZE);
03252919
AK
4194 if (IS_ERR(p))
4195 p = "?";
2fbc57c5 4196 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
03252919
AK
4197 vma->vm_start,
4198 vma->vm_end - vma->vm_start);
4199 free_page((unsigned long)buf);
4200 }
4201 }
51a07e50 4202 up_read(&mm->mmap_sem);
03252919 4203}
3ee1afa3 4204
662bbcb2 4205#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3ee1afa3
NP
4206void might_fault(void)
4207{
95156f00
PZ
4208 /*
4209 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4210 * holding the mmap_sem, this is safe because kernel memory doesn't
4211 * get paged out, therefore we'll never actually fault, and the
4212 * below annotations will generate false positives.
4213 */
4214 if (segment_eq(get_fs(), KERNEL_DS))
4215 return;
4216
3ee1afa3
NP
4217 /*
4218 * it would be nicer only to annotate paths which are not under
4219 * pagefault_disable, however that requires a larger audit and
4220 * providing helpers like get_user_atomic.
4221 */
662bbcb2
MT
4222 if (in_atomic())
4223 return;
4224
4225 __might_sleep(__FILE__, __LINE__, 0);
4226
4227 if (current->mm)
3ee1afa3
NP
4228 might_lock_read(&current->mm->mmap_sem);
4229}
4230EXPORT_SYMBOL(might_fault);
4231#endif
47ad8475
AA
4232
4233#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4234static void clear_gigantic_page(struct page *page,
4235 unsigned long addr,
4236 unsigned int pages_per_huge_page)
4237{
4238 int i;
4239 struct page *p = page;
4240
4241 might_sleep();
4242 for (i = 0; i < pages_per_huge_page;
4243 i++, p = mem_map_next(p, page, i)) {
4244 cond_resched();
4245 clear_user_highpage(p, addr + i * PAGE_SIZE);
4246 }
4247}
4248void clear_huge_page(struct page *page,
4249 unsigned long addr, unsigned int pages_per_huge_page)
4250{
4251 int i;
4252
4253 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4254 clear_gigantic_page(page, addr, pages_per_huge_page);
4255 return;
4256 }
4257
4258 might_sleep();
4259 for (i = 0; i < pages_per_huge_page; i++) {
4260 cond_resched();
4261 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4262 }
4263}
4264
4265static void copy_user_gigantic_page(struct page *dst, struct page *src,
4266 unsigned long addr,
4267 struct vm_area_struct *vma,
4268 unsigned int pages_per_huge_page)
4269{
4270 int i;
4271 struct page *dst_base = dst;
4272 struct page *src_base = src;
4273
4274 for (i = 0; i < pages_per_huge_page; ) {
4275 cond_resched();
4276 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4277
4278 i++;
4279 dst = mem_map_next(dst, dst_base, i);
4280 src = mem_map_next(src, src_base, i);
4281 }
4282}
4283
4284void copy_user_huge_page(struct page *dst, struct page *src,
4285 unsigned long addr, struct vm_area_struct *vma,
4286 unsigned int pages_per_huge_page)
4287{
4288 int i;
4289
4290 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4291 copy_user_gigantic_page(dst, src, addr, vma,
4292 pages_per_huge_page);
4293 return;
4294 }
4295
4296 might_sleep();
4297 for (i = 0; i < pages_per_huge_page; i++) {
4298 cond_resched();
4299 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4300 }
4301}
4302#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */