blob: 829d437354022959eeca37082f695e2ee73025ed [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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>
Hugh Dickins9a840892009-09-21 17:02:01 -070048#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049#include <linux/rmap.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040050#include <linux/export.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070051#include <linux/delayacct.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052#include <linux/init.h>
Peter Zijlstraedc79b22006-09-25 23:30:58 -070053#include <linux/writeback.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080054#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070055#include <linux/mmu_notifier.h>
Hugh Dickins3dc14742009-01-06 14:40:08 -080056#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090059#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Alexey Dobriyan6952b612009-09-18 23:55:55 +040061#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070062#include <asm/pgalloc.h>
63#include <asm/uaccess.h>
64#include <asm/tlb.h>
65#include <asm/tlbflush.h>
66#include <asm/pgtable.h>
67
Jan Beulich42b77722008-07-23 21:27:10 -070068#include "internal.h"
69
Andy Whitcroftd41dee32005-06-23 00:07:54 -070070#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070071/* use the per-pgdat data instead for discontigmem - mbligh */
72unsigned long max_mapnr;
73struct page *mem_map;
74
75EXPORT_SYMBOL(max_mapnr);
76EXPORT_SYMBOL(mem_map);
77#endif
78
79unsigned long num_physpages;
80/*
81 * A number of key systems in x86 including ioremap() rely on the assumption
82 * that high_memory defines the upper bound on direct map memory, then end
83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85 * and ZONE_HIGHMEM.
86 */
87void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
89EXPORT_SYMBOL(num_physpages);
90EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
Ingo Molnar32a93232008-02-06 22:39:44 +010092/*
93 * Randomize the address space (stacks, mmaps, brk, etc.).
94 *
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
97 */
98int randomize_va_space __read_mostly =
99#ifdef CONFIG_COMPAT_BRK
100 1;
101#else
102 2;
103#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100104
105static int __init disable_randmaps(char *s)
106{
107 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800108 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100109}
110__setup("norandmaps", disable_randmaps);
111
Hugh Dickins62eede62009-09-21 17:03:34 -0700112unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700113unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700114
115/*
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117 */
118static int __init init_zero_pfn(void)
119{
120 zero_pfn = page_to_pfn(ZERO_PAGE(0));
121 return 0;
122}
123core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100124
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800125
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800126#if defined(SPLIT_RSS_COUNTING)
127
KAMEZAWA Hiroyukia3a2e762010-04-06 14:34:42 -0700128static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800129{
130 int i;
131
132 for (i = 0; i < NR_MM_COUNTERS; i++) {
133 if (task->rss_stat.count[i]) {
134 add_mm_counter(mm, i, task->rss_stat.count[i]);
135 task->rss_stat.count[i] = 0;
136 }
137 }
138 task->rss_stat.events = 0;
139}
140
141static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142{
143 struct task_struct *task = current;
144
145 if (likely(task->mm == mm))
146 task->rss_stat.count[member] += val;
147 else
148 add_mm_counter(mm, member, val);
149}
150#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
152
153/* sync counter once per 64 page faults */
154#define TASK_RSS_EVENTS_THRESH (64)
155static void check_sync_rss_stat(struct task_struct *task)
156{
157 if (unlikely(task != current))
158 return;
159 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160 __sync_task_rss_stat(task, task->mm);
161}
162
163unsigned long get_mm_counter(struct mm_struct *mm, int member)
164{
165 long val = 0;
166
167 /*
168 * Don't use task->mm here...for avoiding to use task_get_mm()..
169 * The caller must guarantee task->mm is not invalid.
170 */
171 val = atomic_long_read(&mm->rss_stat.count[member]);
172 /*
173 * counter is updated in asynchronous manner and may go to minus.
174 * But it's never be expected number for users.
175 */
176 if (val < 0)
177 return 0;
178 return (unsigned long)val;
179}
180
181void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182{
183 __sync_task_rss_stat(task, mm);
184}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700185#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800186
187#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
189
190static void check_sync_rss_stat(struct task_struct *task)
191{
192}
193
Peter Zijlstra9547d012011-05-24 17:12:14 -0700194#endif /* SPLIT_RSS_COUNTING */
195
196#ifdef HAVE_GENERIC_MMU_GATHER
197
198static int tlb_next_batch(struct mmu_gather *tlb)
199{
200 struct mmu_gather_batch *batch;
201
202 batch = tlb->active;
203 if (batch->next) {
204 tlb->active = batch->next;
205 return 1;
206 }
207
208 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
209 if (!batch)
210 return 0;
211
212 batch->next = NULL;
213 batch->nr = 0;
214 batch->max = MAX_GATHER_BATCH;
215
216 tlb->active->next = batch;
217 tlb->active = batch;
218
219 return 1;
220}
221
222/* tlb_gather_mmu
223 * Called to initialize an (on-stack) mmu_gather structure for page-table
224 * tear-down from @mm. The @fullmm argument is used when @mm is without
225 * users and we're going to destroy the full address space (exit/execve).
226 */
227void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
228{
229 tlb->mm = mm;
230
231 tlb->fullmm = fullmm;
232 tlb->need_flush = 0;
233 tlb->fast_mode = (num_possible_cpus() == 1);
234 tlb->local.next = NULL;
235 tlb->local.nr = 0;
236 tlb->local.max = ARRAY_SIZE(tlb->__pages);
237 tlb->active = &tlb->local;
238
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800241#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700242}
243
244void tlb_flush_mmu(struct mmu_gather *tlb)
245{
246 struct mmu_gather_batch *batch;
247
248 if (!tlb->need_flush)
249 return;
250 tlb->need_flush = 0;
251 tlb_flush(tlb);
252#ifdef CONFIG_HAVE_RCU_TABLE_FREE
253 tlb_table_flush(tlb);
254#endif
255
256 if (tlb_fast_mode(tlb))
257 return;
258
259 for (batch = &tlb->local; batch; batch = batch->next) {
260 free_pages_and_swap_cache(batch->pages, batch->nr);
261 batch->nr = 0;
262 }
263 tlb->active = &tlb->local;
264}
265
266/* tlb_finish_mmu
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
269 */
270void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
271{
272 struct mmu_gather_batch *batch, *next;
273
274 tlb_flush_mmu(tlb);
275
276 /* keep the page table cache within bounds */
277 check_pgt_cache();
278
279 for (batch = tlb->local.next; batch; batch = next) {
280 next = batch->next;
281 free_pages((unsigned long)batch, 0);
282 }
283 tlb->local.next = NULL;
284}
285
286/* __tlb_remove_page
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
291 */
292int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
293{
294 struct mmu_gather_batch *batch;
295
296 tlb->need_flush = 1;
297
298 if (tlb_fast_mode(tlb)) {
299 free_page_and_swap_cache(page);
300 return 1; /* avoid calling tlb_flush_mmu() */
301 }
302
303 batch = tlb->active;
304 batch->pages[batch->nr++] = page;
305 if (batch->nr == batch->max) {
306 if (!tlb_next_batch(tlb))
307 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700308 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700309 }
310 VM_BUG_ON(batch->nr > batch->max);
311
312 return batch->max - batch->nr;
313}
314
315#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800316
Peter Zijlstra26723912011-05-24 17:12:00 -0700317#ifdef CONFIG_HAVE_RCU_TABLE_FREE
318
319/*
320 * See the comment near struct mmu_table_batch.
321 */
322
323static void tlb_remove_table_smp_sync(void *arg)
324{
325 /* Simply deliver the interrupt */
326}
327
328static void tlb_remove_table_one(void *table)
329{
330 /*
331 * This isn't an RCU grace period and hence the page-tables cannot be
332 * assumed to be actually RCU-freed.
333 *
334 * It is however sufficient for software page-table walkers that rely on
335 * IRQ disabling. See the comment near struct mmu_table_batch.
336 */
337 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
338 __tlb_remove_table(table);
339}
340
341static void tlb_remove_table_rcu(struct rcu_head *head)
342{
343 struct mmu_table_batch *batch;
344 int i;
345
346 batch = container_of(head, struct mmu_table_batch, rcu);
347
348 for (i = 0; i < batch->nr; i++)
349 __tlb_remove_table(batch->tables[i]);
350
351 free_page((unsigned long)batch);
352}
353
354void tlb_table_flush(struct mmu_gather *tlb)
355{
356 struct mmu_table_batch **batch = &tlb->batch;
357
358 if (*batch) {
359 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
360 *batch = NULL;
361 }
362}
363
364void tlb_remove_table(struct mmu_gather *tlb, void *table)
365{
366 struct mmu_table_batch **batch = &tlb->batch;
367
368 tlb->need_flush = 1;
369
370 /*
371 * When there's less then two users of this mm there cannot be a
372 * concurrent page-table walk.
373 */
374 if (atomic_read(&tlb->mm->mm_users) < 2) {
375 __tlb_remove_table(table);
376 return;
377 }
378
379 if (*batch == NULL) {
380 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
381 if (*batch == NULL) {
382 tlb_remove_table_one(table);
383 return;
384 }
385 (*batch)->nr = 0;
386 }
387 (*batch)->tables[(*batch)->nr++] = table;
388 if ((*batch)->nr == MAX_TABLE_BATCH)
389 tlb_table_flush(tlb);
390}
391
Peter Zijlstra9547d012011-05-24 17:12:14 -0700392#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700393
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394/*
395 * If a p?d_bad entry is found while walking page tables, report
396 * the error, before resetting entry to p?d_none. Usually (but
397 * very seldom) called out from the p?d_none_or_clear_bad macros.
398 */
399
400void pgd_clear_bad(pgd_t *pgd)
401{
402 pgd_ERROR(*pgd);
403 pgd_clear(pgd);
404}
405
406void pud_clear_bad(pud_t *pud)
407{
408 pud_ERROR(*pud);
409 pud_clear(pud);
410}
411
412void pmd_clear_bad(pmd_t *pmd)
413{
414 pmd_ERROR(*pmd);
415 pmd_clear(pmd);
416}
417
418/*
419 * Note: this doesn't free the actual pages themselves. That
420 * has been handled earlier when unmapping all the memory regions.
421 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000422static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
423 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800425 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700426 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000427 pte_free_tlb(tlb, token, addr);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700428 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429}
430
Hugh Dickinse0da3822005-04-19 13:29:15 -0700431static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
432 unsigned long addr, unsigned long end,
433 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434{
435 pmd_t *pmd;
436 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700437 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438
Hugh Dickinse0da3822005-04-19 13:29:15 -0700439 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441 do {
442 next = pmd_addr_end(addr, end);
443 if (pmd_none_or_clear_bad(pmd))
444 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000445 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 } while (pmd++, addr = next, addr != end);
447
Hugh Dickinse0da3822005-04-19 13:29:15 -0700448 start &= PUD_MASK;
449 if (start < floor)
450 return;
451 if (ceiling) {
452 ceiling &= PUD_MASK;
453 if (!ceiling)
454 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700456 if (end - 1 > ceiling - 1)
457 return;
458
459 pmd = pmd_offset(pud, start);
460 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000461 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462}
463
Hugh Dickinse0da3822005-04-19 13:29:15 -0700464static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
465 unsigned long addr, unsigned long end,
466 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467{
468 pud_t *pud;
469 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700470 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471
Hugh Dickinse0da3822005-04-19 13:29:15 -0700472 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 do {
475 next = pud_addr_end(addr, end);
476 if (pud_none_or_clear_bad(pud))
477 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700478 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479 } while (pud++, addr = next, addr != end);
480
Hugh Dickinse0da3822005-04-19 13:29:15 -0700481 start &= PGDIR_MASK;
482 if (start < floor)
483 return;
484 if (ceiling) {
485 ceiling &= PGDIR_MASK;
486 if (!ceiling)
487 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700489 if (end - 1 > ceiling - 1)
490 return;
491
492 pud = pud_offset(pgd, start);
493 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000494 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495}
496
497/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700498 * This function frees user-level page tables of a process.
499 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 * Must be called with pagetable lock held.
501 */
Jan Beulich42b77722008-07-23 21:27:10 -0700502void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700503 unsigned long addr, unsigned long end,
504 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505{
506 pgd_t *pgd;
507 unsigned long next;
508
Hugh Dickinse0da3822005-04-19 13:29:15 -0700509 /*
510 * The next few lines have given us lots of grief...
511 *
512 * Why are we testing PMD* at this top level? Because often
513 * there will be no work to do at all, and we'd prefer not to
514 * go all the way down to the bottom just to discover that.
515 *
516 * Why all these "- 1"s? Because 0 represents both the bottom
517 * of the address space and the top of it (using -1 for the
518 * top wouldn't help much: the masks would do the wrong thing).
519 * The rule is that addr 0 and floor 0 refer to the bottom of
520 * the address space, but end 0 and ceiling 0 refer to the top
521 * Comparisons need to use "end - 1" and "ceiling - 1" (though
522 * that end 0 case should be mythical).
523 *
524 * Wherever addr is brought up or ceiling brought down, we must
525 * be careful to reject "the opposite 0" before it confuses the
526 * subsequent tests. But what about where end is brought down
527 * by PMD_SIZE below? no, end can't go down to 0 there.
528 *
529 * Whereas we round start (addr) and ceiling down, by different
530 * masks at different levels, in order to test whether a table
531 * now has no other vmas using it, so can be freed, we don't
532 * bother to round floor or end up - the tests don't need that.
533 */
534
535 addr &= PMD_MASK;
536 if (addr < floor) {
537 addr += PMD_SIZE;
538 if (!addr)
539 return;
540 }
541 if (ceiling) {
542 ceiling &= PMD_MASK;
543 if (!ceiling)
544 return;
545 }
546 if (end - 1 > ceiling - 1)
547 end -= PMD_SIZE;
548 if (addr > end - 1)
549 return;
550
Jan Beulich42b77722008-07-23 21:27:10 -0700551 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 do {
553 next = pgd_addr_end(addr, end);
554 if (pgd_none_or_clear_bad(pgd))
555 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700556 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700558}
559
Jan Beulich42b77722008-07-23 21:27:10 -0700560void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700561 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700562{
563 while (vma) {
564 struct vm_area_struct *next = vma->vm_next;
565 unsigned long addr = vma->vm_start;
566
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700567 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000568 * Hide vma from rmap and truncate_pagecache before freeing
569 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700570 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800571 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700572 unlink_file_vma(vma);
573
David Gibson9da61ae2006-03-22 00:08:57 -0800574 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700575 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700576 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700577 } else {
578 /*
579 * Optimization: gather nearby vmas into one call down
580 */
581 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800582 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700583 vma = next;
584 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800585 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700586 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700587 }
588 free_pgd_range(tlb, addr, vma->vm_end,
589 floor, next? next->vm_start: ceiling);
590 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700591 vma = next;
592 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593}
594
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800595int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
596 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800598 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800599 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700600 if (!new)
601 return -ENOMEM;
602
Nick Piggin362a61a2008-05-14 06:37:36 +0200603 /*
604 * Ensure all pte setup (eg. pte page lock and page clearing) are
605 * visible before the pte is made visible to other CPUs by being
606 * put into page tables.
607 *
608 * The other side of the story is the pointer chasing in the page
609 * table walking code (when walking the page table without locking;
610 * ie. most of the time). Fortunately, these data accesses consist
611 * of a chain of data-dependent loads, meaning most CPUs (alpha
612 * being the notable exception) will already guarantee loads are
613 * seen in-order. See the alpha page table accessors for the
614 * smp_read_barrier_depends() barriers in page table walking code.
615 */
616 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
617
Hugh Dickinsc74df322005-10-29 18:16:23 -0700618 spin_lock(&mm->page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800619 wait_split_huge_page = 0;
620 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 mm->nr_ptes++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800623 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800624 } else if (unlikely(pmd_trans_splitting(*pmd)))
625 wait_split_huge_page = 1;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700626 spin_unlock(&mm->page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800627 if (new)
628 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800629 if (wait_split_huge_page)
630 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700631 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632}
633
Hugh Dickins1bb36302005-10-29 18:16:22 -0700634int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700636 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
637 if (!new)
638 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
Nick Piggin362a61a2008-05-14 06:37:36 +0200640 smp_wmb(); /* See comment in __pte_alloc */
641
Hugh Dickins1bb36302005-10-29 18:16:22 -0700642 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800643 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700644 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800645 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800646 } else
647 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700648 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800649 if (new)
650 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700651 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652}
653
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800654static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700655{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800656 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
657}
658
659static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
660{
661 int i;
662
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800663 if (current->mm == mm)
664 sync_mm_rss(current, mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800665 for (i = 0; i < NR_MM_COUNTERS; i++)
666 if (rss[i])
667 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700668}
669
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800671 * This function is called to print an error when a bad pte
672 * is found. For example, we might have a PFN-mapped pte in
673 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700674 *
675 * The calling function must still handle the error.
676 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800677static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
678 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700679{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800680 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
681 pud_t *pud = pud_offset(pgd, addr);
682 pmd_t *pmd = pmd_offset(pud, addr);
683 struct address_space *mapping;
684 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800685 static unsigned long resume;
686 static unsigned long nr_shown;
687 static unsigned long nr_unshown;
688
689 /*
690 * Allow a burst of 60 reports, then keep quiet for that minute;
691 * or allow a steady drip of one report per second.
692 */
693 if (nr_shown == 60) {
694 if (time_before(jiffies, resume)) {
695 nr_unshown++;
696 return;
697 }
698 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800699 printk(KERN_ALERT
700 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800701 nr_unshown);
702 nr_unshown = 0;
703 }
704 nr_shown = 0;
705 }
706 if (nr_shown++ == 0)
707 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800708
709 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
710 index = linear_page_index(vma, addr);
711
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800712 printk(KERN_ALERT
713 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800714 current->comm,
715 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800716 if (page)
717 dump_page(page);
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800718 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800719 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
720 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
721 /*
722 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
723 */
724 if (vma->vm_ops)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800725 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800726 (unsigned long)vma->vm_ops->fault);
727 if (vma->vm_file && vma->vm_file->f_op)
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800728 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800729 (unsigned long)vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700730 dump_stack();
Hugh Dickins3dc14742009-01-06 14:40:08 -0800731 add_taint(TAINT_BAD_PAGE);
Nick Pigginb5810032005-10-29 18:16:12 -0700732}
733
KOSAKI Motohiroca16d142011-05-26 19:16:19 +0900734static inline int is_cow_mapping(vm_flags_t flags)
Linus Torvalds67121172005-12-11 20:38:17 -0800735{
736 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
737}
738
Hugh Dickins62eede62009-09-21 17:03:34 -0700739#ifndef is_zero_pfn
740static inline int is_zero_pfn(unsigned long pfn)
741{
742 return pfn == zero_pfn;
743}
744#endif
745
746#ifndef my_zero_pfn
747static inline unsigned long my_zero_pfn(unsigned long addr)
748{
749 return zero_pfn;
750}
751#endif
752
Nick Pigginb5810032005-10-29 18:16:12 -0700753/*
Nick Piggin7e675132008-04-28 02:13:00 -0700754 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800755 *
Nick Piggin7e675132008-04-28 02:13:00 -0700756 * "Special" mappings do not wish to be associated with a "struct page" (either
757 * it doesn't exist, or it exists but they don't want to touch it). In this
758 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700759 *
Nick Piggin7e675132008-04-28 02:13:00 -0700760 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
761 * pte bit, in which case this function is trivial. Secondly, an architecture
762 * may not have a spare pte bit, which requires a more complicated scheme,
763 * described below.
764 *
765 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
766 * special mapping (even if there are underlying and valid "struct pages").
767 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800768 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700769 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
770 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700771 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
772 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800773 *
774 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
775 *
Nick Piggin7e675132008-04-28 02:13:00 -0700776 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700777 *
Nick Piggin7e675132008-04-28 02:13:00 -0700778 * This restricts such mappings to be a linear translation from virtual address
779 * to pfn. To get around this restriction, we allow arbitrary mappings so long
780 * as the vma is not a COW mapping; in that case, we know that all ptes are
781 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700782 *
783 *
Nick Piggin7e675132008-04-28 02:13:00 -0700784 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
785 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700786 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
787 * page" backing, however the difference is that _all_ pages with a struct
788 * page (that is, those where pfn_valid is true) are refcounted and considered
789 * normal pages by the VM. The disadvantage is that pages are refcounted
790 * (which can be slower and simply not an option for some PFNMAP users). The
791 * advantage is that we don't have to follow the strict linearity rule of
792 * PFNMAP mappings in order to support COWable mappings.
793 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800794 */
Nick Piggin7e675132008-04-28 02:13:00 -0700795#ifdef __HAVE_ARCH_PTE_SPECIAL
796# define HAVE_PTE_SPECIAL 1
797#else
798# define HAVE_PTE_SPECIAL 0
799#endif
800struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
801 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800802{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800803 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700804
805 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800806 if (likely(!pte_special(pte)))
807 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700808 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
809 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700810 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800811 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700812 return NULL;
813 }
814
815 /* !HAVE_PTE_SPECIAL case follows: */
816
Jared Hulbertb379d792008-04-28 02:12:58 -0700817 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
818 if (vma->vm_flags & VM_MIXEDMAP) {
819 if (!pfn_valid(pfn))
820 return NULL;
821 goto out;
822 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700823 unsigned long off;
824 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700825 if (pfn == vma->vm_pgoff + off)
826 return NULL;
827 if (!is_cow_mapping(vma->vm_flags))
828 return NULL;
829 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800830 }
831
Hugh Dickins62eede62009-09-21 17:03:34 -0700832 if (is_zero_pfn(pfn))
833 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800834check_pfn:
835 if (unlikely(pfn > highest_memmap_pfn)) {
836 print_bad_pte(vma, addr, pte, NULL);
837 return NULL;
838 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800839
840 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700841 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700842 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800843 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700844out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800845 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800846}
847
848/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849 * copy one vm_area from one task to the other. Assumes the page tables
850 * already present in the new task to be cleared in the whole range
851 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 */
853
Hugh Dickins570a3352009-12-14 17:58:46 -0800854static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700856 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700857 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858{
Nick Pigginb5810032005-10-29 18:16:12 -0700859 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 pte_t pte = *src_pte;
861 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862
863 /* pte contains position in swap or file, so copy. */
864 if (unlikely(!pte_present(pte))) {
865 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700866 swp_entry_t entry = pte_to_swp_entry(pte);
867
Hugh Dickins570a3352009-12-14 17:58:46 -0800868 if (swap_duplicate(entry) < 0)
869 return entry.val;
870
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 /* make sure dst_mm is on swapoff's mmlist. */
872 if (unlikely(list_empty(&dst_mm->mmlist))) {
873 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700874 if (list_empty(&dst_mm->mmlist))
875 list_add(&dst_mm->mmlist,
876 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 spin_unlock(&mmlist_lock);
878 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800879 if (likely(!non_swap_entry(entry)))
880 rss[MM_SWAPENTS]++;
881 else if (is_write_migration_entry(entry) &&
Christoph Lameter06972122006-06-23 02:03:35 -0700882 is_cow_mapping(vm_flags)) {
883 /*
884 * COW mappings require pages in both parent
885 * and child to be set to read.
886 */
887 make_migration_entry_read(&entry);
888 pte = swp_entry_to_pte(entry);
889 set_pte_at(src_mm, addr, src_pte, pte);
890 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700892 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 }
894
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 /*
896 * If it's a COW mapping, write protect it both
897 * in the parent and the child
898 */
Linus Torvalds67121172005-12-11 20:38:17 -0800899 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700901 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 }
903
904 /*
905 * If it's a shared mapping, mark it clean in
906 * the child
907 */
908 if (vm_flags & VM_SHARED)
909 pte = pte_mkclean(pte);
910 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800911
912 page = vm_normal_page(vma, addr, pte);
913 if (page) {
914 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700915 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800916 if (PageAnon(page))
917 rss[MM_ANONPAGES]++;
918 else
919 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800920 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700921
922out_set_pte:
923 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a3352009-12-14 17:58:46 -0800924 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925}
926
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800927int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
928 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
929 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700931 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700933 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700934 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800935 int rss[NR_MM_COUNTERS];
Hugh Dickins570a3352009-12-14 17:58:46 -0800936 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937
938again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800939 init_rss_vec(rss);
940
Hugh Dickinsc74df322005-10-29 18:16:23 -0700941 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942 if (!dst_pte)
943 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700944 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700945 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700946 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700947 orig_src_pte = src_pte;
948 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700949 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 do {
952 /*
953 * We are holding two locks at this point - either of them
954 * could generate latencies in another task on another CPU.
955 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700956 if (progress >= 32) {
957 progress = 0;
958 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100959 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700960 break;
961 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 if (pte_none(*src_pte)) {
963 progress++;
964 continue;
965 }
Hugh Dickins570a3352009-12-14 17:58:46 -0800966 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
967 vma, addr, rss);
968 if (entry.val)
969 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 progress += 8;
971 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700973 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700974 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700975 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800976 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700977 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700978 cond_resched();
Hugh Dickins570a3352009-12-14 17:58:46 -0800979
980 if (entry.val) {
981 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
982 return -ENOMEM;
983 progress = 0;
984 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 if (addr != end)
986 goto again;
987 return 0;
988}
989
990static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
991 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
992 unsigned long addr, unsigned long end)
993{
994 pmd_t *src_pmd, *dst_pmd;
995 unsigned long next;
996
997 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
998 if (!dst_pmd)
999 return -ENOMEM;
1000 src_pmd = pmd_offset(src_pud, addr);
1001 do {
1002 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001003 if (pmd_trans_huge(*src_pmd)) {
1004 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -08001005 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001006 err = copy_huge_pmd(dst_mm, src_mm,
1007 dst_pmd, src_pmd, addr, vma);
1008 if (err == -ENOMEM)
1009 return -ENOMEM;
1010 if (!err)
1011 continue;
1012 /* fall through */
1013 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 if (pmd_none_or_clear_bad(src_pmd))
1015 continue;
1016 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1017 vma, addr, next))
1018 return -ENOMEM;
1019 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1020 return 0;
1021}
1022
1023static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1024 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1025 unsigned long addr, unsigned long end)
1026{
1027 pud_t *src_pud, *dst_pud;
1028 unsigned long next;
1029
1030 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1031 if (!dst_pud)
1032 return -ENOMEM;
1033 src_pud = pud_offset(src_pgd, addr);
1034 do {
1035 next = pud_addr_end(addr, end);
1036 if (pud_none_or_clear_bad(src_pud))
1037 continue;
1038 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1039 vma, addr, next))
1040 return -ENOMEM;
1041 } while (dst_pud++, src_pud++, addr = next, addr != end);
1042 return 0;
1043}
1044
1045int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1046 struct vm_area_struct *vma)
1047{
1048 pgd_t *src_pgd, *dst_pgd;
1049 unsigned long next;
1050 unsigned long addr = vma->vm_start;
1051 unsigned long end = vma->vm_end;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001052 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053
Nick Piggind9928952005-08-28 16:49:11 +10001054 /*
1055 * Don't copy ptes where a page fault will fill them correctly.
1056 * Fork becomes much lighter when there are big shared or private
1057 * readonly mappings. The tradeoff is that copy_page_range is more
1058 * efficient than faulting.
1059 */
Linus Torvalds4d7672b2005-12-16 10:21:23 -08001060 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
Nick Piggind9928952005-08-28 16:49:11 +10001061 if (!vma->anon_vma)
1062 return 0;
1063 }
1064
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 if (is_vm_hugetlb_page(vma))
1066 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1067
venkatesh.pallipadi@intel.com34801ba2008-12-19 13:47:29 -08001068 if (unlikely(is_pfn_mapping(vma))) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001069 /*
1070 * We do not free on error cases below as remove_vma
1071 * gets called on error from higher level routine
1072 */
1073 ret = track_pfn_vma_copy(vma);
1074 if (ret)
1075 return ret;
1076 }
1077
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001078 /*
1079 * We need to invalidate the secondary MMU mappings only when
1080 * there could be a permission downgrade on the ptes of the
1081 * parent mm. And a permission downgrade will only happen if
1082 * is_cow_mapping() returns true.
1083 */
1084 if (is_cow_mapping(vma->vm_flags))
1085 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1086
1087 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 dst_pgd = pgd_offset(dst_mm, addr);
1089 src_pgd = pgd_offset(src_mm, addr);
1090 do {
1091 next = pgd_addr_end(addr, end);
1092 if (pgd_none_or_clear_bad(src_pgd))
1093 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001094 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1095 vma, addr, next))) {
1096 ret = -ENOMEM;
1097 break;
1098 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001100
1101 if (is_cow_mapping(vma->vm_flags))
1102 mmu_notifier_invalidate_range_end(src_mm,
1103 vma->vm_start, end);
1104 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105}
1106
Robin Holt51c6f662005-11-13 16:06:42 -08001107static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001108 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001110 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111{
Nick Pigginb5810032005-10-29 18:16:12 -07001112 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001113 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001114 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001115 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001116 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001117 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001118
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001119again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001120 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001121 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1122 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001123 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 do {
1125 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001126 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001128 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001129
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001131 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001132
Linus Torvalds6aab3412005-11-28 14:34:23 -08001133 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 if (unlikely(details) && page) {
1135 /*
1136 * unmap_shared_mapping_pages() wants to
1137 * invalidate cache without truncating:
1138 * unmap shared but keep private pages.
1139 */
1140 if (details->check_mapping &&
1141 details->check_mapping != page->mapping)
1142 continue;
1143 /*
1144 * Each page->index must be checked when
1145 * invalidating or truncating nonlinear.
1146 */
1147 if (details->nonlinear_vma &&
1148 (page->index < details->first_index ||
1149 page->index > details->last_index))
1150 continue;
1151 }
Nick Pigginb5810032005-10-29 18:16:12 -07001152 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001153 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 tlb_remove_tlb_entry(tlb, pte, addr);
1155 if (unlikely(!page))
1156 continue;
1157 if (unlikely(details) && details->nonlinear_vma
1158 && linear_page_index(details->nonlinear_vma,
1159 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -07001160 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001163 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001164 else {
1165 if (pte_dirty(ptent))
1166 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001167 if (pte_young(ptent) &&
1168 likely(!VM_SequentialReadHint(vma)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001169 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001170 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001171 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001172 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001173 if (unlikely(page_mapcount(page) < 0))
1174 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001175 force_flush = !__tlb_remove_page(tlb, page);
1176 if (force_flush)
1177 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 continue;
1179 }
1180 /*
1181 * If details->check_mapping, we leave swap entries;
1182 * if details->nonlinear_vma, we leave file entries.
1183 */
1184 if (unlikely(details))
1185 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001186 if (pte_file(ptent)) {
1187 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1188 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001189 } else {
1190 swp_entry_t entry = pte_to_swp_entry(ptent);
1191
1192 if (!non_swap_entry(entry))
1193 rss[MM_SWAPENTS]--;
1194 if (unlikely(!free_swap_and_cache(entry)))
1195 print_bad_pte(vma, addr, ptent, NULL);
1196 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001197 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001198 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001199
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001200 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001201 arch_leave_lazy_mmu_mode();
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001202 pte_unmap_unlock(start_pte, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001203
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001204 /*
1205 * mmu_gather ran out of room to batch pages, we break out of
1206 * the PTE lock to avoid doing the potential expensive TLB invalidate
1207 * and page-free while holding it.
1208 */
1209 if (force_flush) {
1210 force_flush = 0;
1211 tlb_flush_mmu(tlb);
1212 if (addr != end)
1213 goto again;
1214 }
1215
Robin Holt51c6f662005-11-13 16:06:42 -08001216 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217}
1218
Robin Holt51c6f662005-11-13 16:06:42 -08001219static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001220 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001222 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223{
1224 pmd_t *pmd;
1225 unsigned long next;
1226
1227 pmd = pmd_offset(pud, addr);
1228 do {
1229 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001230 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli14d1a552011-01-13 15:47:15 -08001231 if (next-addr != HPAGE_PMD_SIZE) {
1232 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001233 split_huge_page_pmd(vma->vm_mm, pmd);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001234 } else if (zap_huge_pmd(tlb, vma, pmd))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001235 continue;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001236 /* fall through */
1237 }
Peter Zijlstra97a89412011-05-24 17:12:04 -07001238 if (pmd_none_or_clear_bad(pmd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001240 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1241 cond_resched();
1242 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001243
1244 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245}
1246
Robin Holt51c6f662005-11-13 16:06:42 -08001247static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001248 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001250 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251{
1252 pud_t *pud;
1253 unsigned long next;
1254
1255 pud = pud_offset(pgd, addr);
1256 do {
1257 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001258 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001260 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1261 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001262
1263 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264}
1265
Robin Holt51c6f662005-11-13 16:06:42 -08001266static unsigned long unmap_page_range(struct mmu_gather *tlb,
1267 struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001269 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270{
1271 pgd_t *pgd;
1272 unsigned long next;
1273
1274 if (details && !details->check_mapping && !details->nonlinear_vma)
1275 details = NULL;
1276
1277 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001278 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279 tlb_start_vma(tlb, vma);
1280 pgd = pgd_offset(vma->vm_mm, addr);
1281 do {
1282 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001283 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001285 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1286 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001288 mem_cgroup_uncharge_end();
Robin Holt51c6f662005-11-13 16:06:42 -08001289
1290 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291}
1292
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293/**
1294 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001295 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 * @vma: the starting vma
1297 * @start_addr: virtual address at which to start unmapping
1298 * @end_addr: virtual address at which to end unmapping
1299 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1300 * @details: details of nonlinear truncation or shared cache invalidation
1301 *
Hugh Dickinsee39b372005-04-19 13:29:15 -07001302 * Returns the end address of the unmapping (restart addr if interrupted).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001304 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306 * Only addresses between `start' and `end' will be unmapped.
1307 *
1308 * The VMA list must be sorted in ascending virtual address order.
1309 *
1310 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1311 * range after unmap_vmas() returns. So the only responsibility here is to
1312 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1313 * drops the lock and schedules.
1314 */
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001315unsigned long unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 struct vm_area_struct *vma, unsigned long start_addr,
1317 unsigned long end_addr, unsigned long *nr_accounted,
1318 struct zap_details *details)
1319{
Hugh Dickinsee39b372005-04-19 13:29:15 -07001320 unsigned long start = start_addr;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001321 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001323 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 unsigned long end;
1326
1327 start = max(vma->vm_start, start_addr);
1328 if (start >= vma->vm_end)
1329 continue;
1330 end = min(vma->vm_end, end_addr);
1331 if (end <= vma->vm_start)
1332 continue;
1333
1334 if (vma->vm_flags & VM_ACCOUNT)
1335 *nr_accounted += (end - start) >> PAGE_SHIFT;
1336
venkatesh.pallipadi@intel.com34801ba2008-12-19 13:47:29 -08001337 if (unlikely(is_pfn_mapping(vma)))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001338 untrack_pfn_vma(vma, 0, 0);
1339
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 while (start != end) {
Robin Holt51c6f662005-11-13 16:06:42 -08001341 if (unlikely(is_vm_hugetlb_page(vma))) {
Andi Kleena137e1c2008-07-23 21:27:43 -07001342 /*
1343 * It is undesirable to test vma->vm_file as it
1344 * should be non-null for valid hugetlb area.
1345 * However, vm_file will be NULL in the error
1346 * cleanup path of do_mmap_pgoff. When
1347 * hugetlbfs ->mmap method fails,
1348 * do_mmap_pgoff() nullifies vma->vm_file
1349 * before calling this function to clean up.
1350 * Since no pte has actually been setup, it is
1351 * safe to do nothing in this case.
1352 */
Peter Zijlstra97a89412011-05-24 17:12:04 -07001353 if (vma->vm_file)
Andi Kleena137e1c2008-07-23 21:27:43 -07001354 unmap_hugepage_range(vma, start, end, NULL);
Andi Kleena137e1c2008-07-23 21:27:43 -07001355
Robin Holt51c6f662005-11-13 16:06:42 -08001356 start = end;
1357 } else
Peter Zijlstra97a89412011-05-24 17:12:04 -07001358 start = unmap_page_range(tlb, vma, start, end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 }
1360 }
Peter Zijlstra97a89412011-05-24 17:12:04 -07001361
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001362 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Hugh Dickinsee39b372005-04-19 13:29:15 -07001363 return start; /* which is now the end (or restart) address */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364}
1365
1366/**
1367 * zap_page_range - remove user pages in a given range
1368 * @vma: vm_area_struct holding the applicable pages
1369 * @address: starting address of pages to zap
1370 * @size: number of bytes to zap
1371 * @details: details of nonlinear truncation or shared cache invalidation
1372 */
Hugh Dickinsee39b372005-04-19 13:29:15 -07001373unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 unsigned long size, struct zap_details *details)
1375{
1376 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001377 struct mmu_gather tlb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 unsigned long end = address + size;
1379 unsigned long nr_accounted = 0;
1380
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 lru_add_drain();
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001382 tlb_gather_mmu(&tlb, mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001383 update_hiwater_rss(mm);
Hugh Dickins508034a2005-10-29 18:16:30 -07001384 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001385 tlb_finish_mmu(&tlb, address, end);
Hugh Dickinsee39b372005-04-19 13:29:15 -07001386 return end;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387}
1388
Jack Steinerc627f9c2008-07-29 22:33:53 -07001389/**
1390 * zap_vma_ptes - remove ptes mapping the vma
1391 * @vma: vm_area_struct holding ptes to be zapped
1392 * @address: starting address of pages to zap
1393 * @size: number of bytes to zap
1394 *
1395 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1396 *
1397 * The entire address range must be fully contained within the vma.
1398 *
1399 * Returns 0 if successful.
1400 */
1401int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1402 unsigned long size)
1403{
1404 if (address < vma->vm_start || address + size > vma->vm_end ||
1405 !(vma->vm_flags & VM_PFNMAP))
1406 return -1;
1407 zap_page_range(vma, address, size, NULL);
1408 return 0;
1409}
1410EXPORT_SYMBOL_GPL(zap_vma_ptes);
1411
Johannes Weiner142762b2010-05-24 14:32:39 -07001412/**
1413 * follow_page - look up a page descriptor from a user-virtual address
1414 * @vma: vm_area_struct mapping @address
1415 * @address: virtual address to look up
1416 * @flags: flags modifying lookup behaviour
1417 *
1418 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1419 *
1420 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1421 * an error pointer if there is a mapping to something not represented
1422 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423 */
Linus Torvalds6aab3412005-11-28 14:34:23 -08001424struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001425 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426{
1427 pgd_t *pgd;
1428 pud_t *pud;
1429 pmd_t *pmd;
1430 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001431 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001433 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001435 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1436 if (!IS_ERR(page)) {
1437 BUG_ON(flags & FOLL_GET);
1438 goto out;
1439 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001441 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442 pgd = pgd_offset(mm, address);
1443 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001444 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445
1446 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001447 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001448 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001449 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Andi Kleenceb86872008-07-23 21:27:50 -07001450 BUG_ON(flags & FOLL_GET);
1451 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1452 goto out;
1453 }
1454 if (unlikely(pud_bad(*pud)))
1455 goto no_page_table;
1456
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001458 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001459 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001460 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001461 BUG_ON(flags & FOLL_GET);
1462 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1463 goto out;
1464 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001465 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001466 if (flags & FOLL_SPLIT) {
1467 split_huge_page_pmd(mm, pmd);
1468 goto split_fallthrough;
1469 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001470 spin_lock(&mm->page_table_lock);
1471 if (likely(pmd_trans_huge(*pmd))) {
1472 if (unlikely(pmd_trans_splitting(*pmd))) {
1473 spin_unlock(&mm->page_table_lock);
1474 wait_split_huge_page(vma->anon_vma, pmd);
1475 } else {
1476 page = follow_trans_huge_pmd(mm, address,
1477 pmd, flags);
1478 spin_unlock(&mm->page_table_lock);
1479 goto out;
1480 }
1481 } else
1482 spin_unlock(&mm->page_table_lock);
1483 /* fall through */
1484 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001485split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001486 if (unlikely(pmd_bad(*pmd)))
1487 goto no_page_table;
1488
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001489 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490
1491 pte = *ptep;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001492 if (!pte_present(pte))
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001493 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001494 if ((flags & FOLL_WRITE) && !pte_write(pte))
1495 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001496
Linus Torvalds6aab3412005-11-28 14:34:23 -08001497 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001498 if (unlikely(!page)) {
1499 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001500 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001501 goto bad_page;
1502 page = pte_page(pte);
1503 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001504
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001505 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001506 get_page_foll(page);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001507 if (flags & FOLL_TOUCH) {
1508 if ((flags & FOLL_WRITE) &&
1509 !pte_dirty(pte) && !PageDirty(page))
1510 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001511 /*
1512 * pte_mkyoung() would be more correct here, but atomic care
1513 * is needed to avoid losing the dirty bit: it is easier to use
1514 * mark_page_accessed().
1515 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001516 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001518 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001519 /*
1520 * The preliminary mapping check is mainly to avoid the
1521 * pointless overhead of lock_page on the ZERO_PAGE
1522 * which might bounce very badly if there is contention.
1523 *
1524 * If the page is already locked, we don't need to
1525 * handle it now - vmscan will handle it later if and
1526 * when it attempts to reclaim the page.
1527 */
1528 if (page->mapping && trylock_page(page)) {
1529 lru_add_drain(); /* push cached pages to LRU */
1530 /*
1531 * Because we lock page here and migration is
1532 * blocked by the pte's page reference, we need
1533 * only check for file-cache page truncation.
1534 */
1535 if (page->mapping)
1536 mlock_vma_page(page);
1537 unlock_page(page);
1538 }
1539 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001540unlock:
1541 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001543 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001545bad_page:
1546 pte_unmap_unlock(ptep, ptl);
1547 return ERR_PTR(-EFAULT);
1548
1549no_page:
1550 pte_unmap_unlock(ptep, ptl);
1551 if (!pte_none(pte))
1552 return page;
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001553
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001554no_page_table:
1555 /*
1556 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001557 * has touched so far, we don't want to allocate unnecessary pages or
1558 * page tables. Return error instead of NULL to skip handle_mm_fault,
1559 * then get_dump_page() will return NULL to leave a hole in the dump.
1560 * But we can only make this optimization where a hole would surely
1561 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001562 */
Hugh Dickins8e4b9a62009-09-21 17:03:26 -07001563 if ((flags & FOLL_DUMP) &&
1564 (!vma->vm_ops || !vma->vm_ops->fault))
1565 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001566 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567}
1568
Linus Torvalds95042f92011-04-12 14:15:51 -07001569static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1570{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001571 return stack_guard_page_start(vma, addr) ||
1572 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001573}
1574
Huang Ying0014bd92011-01-30 11:15:47 +08001575/**
1576 * __get_user_pages() - pin user pages in memory
1577 * @tsk: task_struct of target task
1578 * @mm: mm_struct of target mm
1579 * @start: starting user address
1580 * @nr_pages: number of pages from start to pin
1581 * @gup_flags: flags modifying pin behaviour
1582 * @pages: array that receives pointers to the pages pinned.
1583 * Should be at least nr_pages long. Or NULL, if caller
1584 * only intends to ensure the pages are faulted in.
1585 * @vmas: array of pointers to vmas corresponding to each page.
1586 * Or NULL if the caller does not require them.
1587 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1588 *
1589 * Returns number of pages pinned. This may be fewer than the number
1590 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1591 * were pinned, returns -errno. Each page returned must be released
1592 * with a put_page() call when it is finished with. vmas will only
1593 * remain valid while mmap_sem is held.
1594 *
1595 * Must be called with mmap_sem held for read or write.
1596 *
1597 * __get_user_pages walks a process's page tables and takes a reference to
1598 * each struct page that each user address corresponds to at a given
1599 * instant. That is, it takes the page that would be accessed if a user
1600 * thread accesses the given user virtual address at that instant.
1601 *
1602 * This does not guarantee that the page exists in the user mappings when
1603 * __get_user_pages returns, and there may even be a completely different
1604 * page there in some cases (eg. if mmapped pagecache has been invalidated
1605 * and subsequently re faulted). However it does guarantee that the page
1606 * won't be freed completely. And mostly callers simply care that the page
1607 * contains data that was valid *at some point in time*. Typically, an IO
1608 * or similar operation cannot guarantee anything stronger anyway because
1609 * locks can't be held over the syscall boundary.
1610 *
1611 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1612 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1613 * appropriate) must be called after the page is finished with, and
1614 * before put_page is called.
1615 *
1616 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1617 * or mmap_sem contention, and if waiting is needed to pin all pages,
1618 * *@nonblocking will be set to 0.
1619 *
1620 * In most cases, get_user_pages or get_user_pages_fast should be used
1621 * instead of __get_user_pages. __get_user_pages should be used only if
1622 * you need some special @gup_flags.
1623 */
Nick Pigginb291f002008-10-18 20:26:44 -07001624int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001625 unsigned long start, int nr_pages, unsigned int gup_flags,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001626 struct page **pages, struct vm_area_struct **vmas,
1627 int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628{
1629 int i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001630 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631
Peter Zijlstra9d737772009-06-25 11:58:55 +02001632 if (nr_pages <= 0)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001633 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001634
1635 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1636
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 /*
1638 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001639 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001641 vm_flags = (gup_flags & FOLL_WRITE) ?
1642 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1643 vm_flags &= (gup_flags & FOLL_FORCE) ?
1644 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 i = 0;
1646
1647 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001648 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649
1650 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001651 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 pgd_t *pgd;
1654 pud_t *pud;
1655 pmd_t *pmd;
1656 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001657
1658 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001659 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660 return i ? : -EFAULT;
1661 if (pg > TASK_SIZE)
1662 pgd = pgd_offset_k(pg);
1663 else
1664 pgd = pgd_offset_gate(mm, pg);
1665 BUG_ON(pgd_none(*pgd));
1666 pud = pud_offset(pgd, pg);
1667 BUG_ON(pud_none(*pud));
1668 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001669 if (pmd_none(*pmd))
1670 return i ? : -EFAULT;
Andrea Arcangelif66055a2011-01-13 15:46:54 -08001671 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001673 if (pte_none(*pte)) {
1674 pte_unmap(pte);
1675 return i ? : -EFAULT;
1676 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001677 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 if (pages) {
Hugh Dickinsde512572010-07-30 10:58:26 -07001679 struct page *page;
1680
Linus Torvalds95042f92011-04-12 14:15:51 -07001681 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde512572010-07-30 10:58:26 -07001682 if (!page) {
1683 if (!(gup_flags & FOLL_DUMP) &&
1684 is_zero_pfn(pte_pfn(*pte)))
1685 page = pte_page(*pte);
1686 else {
1687 pte_unmap(pte);
1688 return i ? : -EFAULT;
1689 }
1690 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001691 pages[i] = page;
Hugh Dickinsde512572010-07-30 10:58:26 -07001692 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 }
1694 pte_unmap(pte);
Linus Torvalds95042f92011-04-12 14:15:51 -07001695 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 }
1697
Nick Pigginb291f002008-10-18 20:26:44 -07001698 if (!vma ||
1699 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001700 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 return i ? : -EFAULT;
1702
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001703 if (is_vm_hugetlb_page(vma)) {
1704 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001705 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001706 continue;
1707 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001708
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001710 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001711 unsigned int foll_flags = gup_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712
Ethan Solomita462e00c2007-07-15 23:38:16 -07001713 /*
Ying Han47792802009-01-06 14:40:18 -08001714 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001715 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001716 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001717 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001718 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001719
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001720 cond_resched();
Linus Torvalds6aab3412005-11-28 14:34:23 -08001721 while (!(page = follow_page(vma, start, foll_flags))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001722 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001723 unsigned int fault_flags = 0;
1724
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001725 /* For mlock, just skip the stack guard page. */
1726 if (foll_flags & FOLL_MLOCK) {
1727 if (stack_guard_page(vma, start))
1728 goto next_page;
1729 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001730 if (foll_flags & FOLL_WRITE)
1731 fault_flags |= FAULT_FLAG_WRITE;
1732 if (nonblocking)
1733 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001734 if (foll_flags & FOLL_NOWAIT)
1735 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001736
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001737 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001738 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001739
Nick Piggin83c54072007-07-19 01:47:05 -07001740 if (ret & VM_FAULT_ERROR) {
1741 if (ret & VM_FAULT_OOM)
1742 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001743 if (ret & (VM_FAULT_HWPOISON |
1744 VM_FAULT_HWPOISON_LARGE)) {
1745 if (i)
1746 return i;
1747 else if (gup_flags & FOLL_HWPOISON)
1748 return -EHWPOISON;
1749 else
1750 return -EFAULT;
1751 }
1752 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001753 return i ? i : -EFAULT;
1754 BUG();
1755 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001756
1757 if (tsk) {
1758 if (ret & VM_FAULT_MAJOR)
1759 tsk->maj_flt++;
1760 else
1761 tsk->min_flt++;
1762 }
Nick Piggin83c54072007-07-19 01:47:05 -07001763
Michel Lespinasse53a77062011-01-13 15:46:14 -08001764 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001765 if (nonblocking)
1766 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001767 return i;
1768 }
1769
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001770 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001771 * The VM_FAULT_WRITE bit tells us that
1772 * do_wp_page has broken COW when necessary,
1773 * even if maybe_mkwrite decided not to set
1774 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001775 * page lookups as if they were reads. But only
1776 * do so when looping for pte_write is futile:
1777 * in some cases userspace may also be wanting
1778 * to write to the gotten user page, which a
1779 * read fault here might prevent (a readonly
1780 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001781 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001782 if ((ret & VM_FAULT_WRITE) &&
1783 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001784 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001785
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001786 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001788 if (IS_ERR(page))
1789 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001791 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001792
Russell Kinga6f36be2006-12-30 22:24:19 +00001793 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001794 flush_dcache_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001796next_page:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 if (vmas)
1798 vmas[i] = vma;
1799 i++;
1800 start += PAGE_SIZE;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001801 nr_pages--;
1802 } while (nr_pages && start < vma->vm_end);
1803 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804 return i;
1805}
Huang Ying0014bd92011-01-30 11:15:47 +08001806EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001807
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -07001808/*
1809 * fixup_user_fault() - manually resolve a user page fault
1810 * @tsk: the task_struct to use for page fault accounting, or
1811 * NULL if faults are not to be recorded.
1812 * @mm: mm_struct of target mm
1813 * @address: user address
1814 * @fault_flags:flags to pass down to handle_mm_fault()
1815 *
1816 * This is meant to be called in the specific scenario where for locking reasons
1817 * we try to access user memory in atomic context (within a pagefault_disable()
1818 * section), this returns -EFAULT, and we want to resolve the user fault before
1819 * trying again.
1820 *
1821 * Typically this is meant to be used by the futex code.
1822 *
1823 * The main difference with get_user_pages() is that this function will
1824 * unconditionally call handle_mm_fault() which will in turn perform all the
1825 * necessary SW fixup of the dirty and young bits in the PTE, while
1826 * handle_mm_fault() only guarantees to update these in the struct page.
1827 *
1828 * This is important for some architectures where those bits also gate the
1829 * access permission to the page because they are maintained in software. On
1830 * such architectures, gup() will not be enough to make a subsequent access
1831 * succeed.
1832 *
1833 * This should be called with the mm_sem held for read.
1834 */
1835int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1836 unsigned long address, unsigned int fault_flags)
1837{
1838 struct vm_area_struct *vma;
1839 int ret;
1840
1841 vma = find_extend_vma(mm, address);
1842 if (!vma || address < vma->vm_start)
1843 return -EFAULT;
1844
1845 ret = handle_mm_fault(mm, vma, address, fault_flags);
1846 if (ret & VM_FAULT_ERROR) {
1847 if (ret & VM_FAULT_OOM)
1848 return -ENOMEM;
1849 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1850 return -EHWPOISON;
1851 if (ret & VM_FAULT_SIGBUS)
1852 return -EFAULT;
1853 BUG();
1854 }
1855 if (tsk) {
1856 if (ret & VM_FAULT_MAJOR)
1857 tsk->maj_flt++;
1858 else
1859 tsk->min_flt++;
1860 }
1861 return 0;
1862}
1863
1864/*
Nick Piggind2bf6be2009-06-16 15:31:39 -07001865 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001866 * @tsk: the task_struct to use for page fault accounting, or
1867 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001868 * @mm: mm_struct of target mm
1869 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001870 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001871 * @write: whether pages will be written to by the caller
1872 * @force: whether to force write access even if user mapping is
1873 * readonly. This will result in the page being COWed even
1874 * in MAP_SHARED mappings. You do not want this.
1875 * @pages: array that receives pointers to the pages pinned.
1876 * Should be at least nr_pages long. Or NULL, if caller
1877 * only intends to ensure the pages are faulted in.
1878 * @vmas: array of pointers to vmas corresponding to each page.
1879 * Or NULL if the caller does not require them.
1880 *
1881 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001882 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001883 * were pinned, returns -errno. Each page returned must be released
1884 * with a put_page() call when it is finished with. vmas will only
1885 * remain valid while mmap_sem is held.
1886 *
1887 * Must be called with mmap_sem held for read or write.
1888 *
1889 * get_user_pages walks a process's page tables and takes a reference to
1890 * each struct page that each user address corresponds to at a given
1891 * instant. That is, it takes the page that would be accessed if a user
1892 * thread accesses the given user virtual address at that instant.
1893 *
1894 * This does not guarantee that the page exists in the user mappings when
1895 * get_user_pages returns, and there may even be a completely different
1896 * page there in some cases (eg. if mmapped pagecache has been invalidated
1897 * and subsequently re faulted). However it does guarantee that the page
1898 * won't be freed completely. And mostly callers simply care that the page
1899 * contains data that was valid *at some point in time*. Typically, an IO
1900 * or similar operation cannot guarantee anything stronger anyway because
1901 * locks can't be held over the syscall boundary.
1902 *
1903 * If write=0, the page must not be written to. If the page is written to,
1904 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1905 * after the page is finished with, and before put_page is called.
1906 *
1907 * get_user_pages is typically used for fewer-copy IO operations, to get a
1908 * handle on the memory by some means other than accesses via the user virtual
1909 * addresses. The pages may be submitted for DMA to devices or accessed via
1910 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1911 * use the correct cache flushing APIs.
1912 *
1913 * See also get_user_pages_fast, for performance critical applications.
1914 */
Nick Pigginb291f002008-10-18 20:26:44 -07001915int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Peter Zijlstra9d737772009-06-25 11:58:55 +02001916 unsigned long start, int nr_pages, int write, int force,
Nick Pigginb291f002008-10-18 20:26:44 -07001917 struct page **pages, struct vm_area_struct **vmas)
1918{
Hugh Dickins58fa8792009-09-21 17:03:31 -07001919 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07001920
Hugh Dickins58fa8792009-09-21 17:03:31 -07001921 if (pages)
1922 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07001923 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001924 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07001925 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07001926 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07001927
Michel Lespinasse53a77062011-01-13 15:46:14 -08001928 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1929 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07001930}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931EXPORT_SYMBOL(get_user_pages);
1932
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001933/**
1934 * get_dump_page() - pin user page in memory while writing it to core dump
1935 * @addr: user address
1936 *
1937 * Returns struct page pointer of user page pinned for dump,
1938 * to be freed afterwards by page_cache_release() or put_page().
1939 *
1940 * Returns NULL on any kind of failure - a hole must then be inserted into
1941 * the corefile, to preserve alignment with its headers; and also returns
1942 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1943 * allowing a hole to be left in the corefile to save diskspace.
1944 *
1945 * Called without mmap_sem, but after all other threads have been killed.
1946 */
1947#ifdef CONFIG_ELF_CORE
1948struct page *get_dump_page(unsigned long addr)
1949{
1950 struct vm_area_struct *vma;
1951 struct page *page;
1952
1953 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001954 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1955 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001956 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07001957 flush_cache_page(vma, addr, page_to_pfn(page));
1958 return page;
1959}
1960#endif /* CONFIG_ELF_CORE */
1961
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001962pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001963 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001964{
1965 pgd_t * pgd = pgd_offset(mm, addr);
1966 pud_t * pud = pud_alloc(mm, pgd, addr);
1967 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001968 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055a2011-01-13 15:46:54 -08001969 if (pmd) {
1970 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001971 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055a2011-01-13 15:46:54 -08001972 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001973 }
1974 return NULL;
1975}
1976
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001978 * This is the old fallback for page remapping.
1979 *
1980 * For historical reasons, it only allows reserved pages. Only
1981 * old drivers should use this, and they needed to mark their
1982 * pages reserved for the old functions anyway.
1983 */
Nick Piggin423bad62008-04-28 02:13:01 -07001984static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1985 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001986{
Nick Piggin423bad62008-04-28 02:13:01 -07001987 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001988 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001989 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001990 spinlock_t *ptl;
1991
Linus Torvalds238f58d2005-11-29 13:01:56 -08001992 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001993 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001994 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001995 retval = -ENOMEM;
1996 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001997 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001998 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001999 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002000 retval = -EBUSY;
2001 if (!pte_none(*pte))
2002 goto out_unlock;
2003
2004 /* Ok, finally just insert the thing.. */
2005 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002006 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002007 page_add_file_rmap(page);
2008 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2009
2010 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002011 pte_unmap_unlock(pte, ptl);
2012 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002013out_unlock:
2014 pte_unmap_unlock(pte, ptl);
2015out:
2016 return retval;
2017}
2018
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002019/**
2020 * vm_insert_page - insert single page into user vma
2021 * @vma: user vma to map to
2022 * @addr: target user address of this page
2023 * @page: source kernel page
2024 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002025 * This allows drivers to insert individual pages they've allocated
2026 * into a user vma.
2027 *
2028 * The page has to be a nice clean _individual_ kernel allocation.
2029 * If you allocate a compound page, you need to have marked it as
2030 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08002031 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002032 *
2033 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2034 * took an arbitrary page protection parameter. This doesn't allow
2035 * that. Your vma protection will have to be set up correctly, which
2036 * means that if you want a shared writable mapping, you'd better
2037 * ask for a shared writable mapping!
2038 *
2039 * The page does not need to be reserved.
2040 */
Nick Piggin423bad62008-04-28 02:13:01 -07002041int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2042 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002043{
2044 if (addr < vma->vm_start || addr >= vma->vm_end)
2045 return -EFAULT;
2046 if (!page_count(page))
2047 return -EINVAL;
Linus Torvalds4d7672b2005-12-16 10:21:23 -08002048 vma->vm_flags |= VM_INSERTPAGE;
Nick Piggin423bad62008-04-28 02:13:01 -07002049 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002050}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002051EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002052
Nick Piggin423bad62008-04-28 02:13:01 -07002053static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2054 unsigned long pfn, pgprot_t prot)
2055{
2056 struct mm_struct *mm = vma->vm_mm;
2057 int retval;
2058 pte_t *pte, entry;
2059 spinlock_t *ptl;
2060
2061 retval = -ENOMEM;
2062 pte = get_locked_pte(mm, addr, &ptl);
2063 if (!pte)
2064 goto out;
2065 retval = -EBUSY;
2066 if (!pte_none(*pte))
2067 goto out_unlock;
2068
2069 /* Ok, finally just insert the thing.. */
2070 entry = pte_mkspecial(pfn_pte(pfn, prot));
2071 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002072 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07002073
2074 retval = 0;
2075out_unlock:
2076 pte_unmap_unlock(pte, ptl);
2077out:
2078 return retval;
2079}
2080
Nick Piggine0dc0d82007-02-12 00:51:36 -08002081/**
2082 * vm_insert_pfn - insert single pfn into user vma
2083 * @vma: user vma to map to
2084 * @addr: target user address of this page
2085 * @pfn: source kernel pfn
2086 *
2087 * Similar to vm_inert_page, this allows drivers to insert individual pages
2088 * they've allocated into a user vma. Same comments apply.
2089 *
2090 * This function should only be called from a vm_ops->fault handler, and
2091 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002092 *
2093 * vma cannot be a COW mapping.
2094 *
2095 * As this is called only for pages that do not currently exist, we
2096 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002097 */
2098int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07002099 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002100{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002101 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002102 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002103 /*
2104 * Technically, architectures with pte_special can avoid all these
2105 * restrictions (same for remap_pfn_range). However we would like
2106 * consistency in testing and feature parity among all, so we should
2107 * try to keep these invariants in place for everybody.
2108 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002109 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2110 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2111 (VM_PFNMAP|VM_MIXEDMAP));
2112 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2113 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002114
Nick Piggin423bad62008-04-28 02:13:01 -07002115 if (addr < vma->vm_start || addr >= vma->vm_end)
2116 return -EFAULT;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002117 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002118 return -EINVAL;
2119
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002120 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002121
2122 if (ret)
2123 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2124
2125 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002126}
2127EXPORT_SYMBOL(vm_insert_pfn);
2128
Nick Piggin423bad62008-04-28 02:13:01 -07002129int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2130 unsigned long pfn)
2131{
2132 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2133
2134 if (addr < vma->vm_start || addr >= vma->vm_end)
2135 return -EFAULT;
2136
2137 /*
2138 * If we don't have pte special, then we have to use the pfn_valid()
2139 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2140 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002141 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2142 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07002143 */
2144 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2145 struct page *page;
2146
2147 page = pfn_to_page(pfn);
2148 return insert_page(vma, addr, page, vma->vm_page_prot);
2149 }
2150 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2151}
2152EXPORT_SYMBOL(vm_insert_mixed);
2153
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002154/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155 * maps a range of physical memory into the requested pages. the old
2156 * mappings are removed. any references to nonexistent pages results
2157 * in null mappings (currently treated as "copy-on-access")
2158 */
2159static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2160 unsigned long addr, unsigned long end,
2161 unsigned long pfn, pgprot_t prot)
2162{
2163 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002164 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165
Hugh Dickinsc74df322005-10-29 18:16:23 -07002166 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 if (!pte)
2168 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002169 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170 do {
2171 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002172 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173 pfn++;
2174 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002175 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002176 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 return 0;
2178}
2179
2180static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2181 unsigned long addr, unsigned long end,
2182 unsigned long pfn, pgprot_t prot)
2183{
2184 pmd_t *pmd;
2185 unsigned long next;
2186
2187 pfn -= addr >> PAGE_SHIFT;
2188 pmd = pmd_alloc(mm, pud, addr);
2189 if (!pmd)
2190 return -ENOMEM;
Andrea Arcangelif66055a2011-01-13 15:46:54 -08002191 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192 do {
2193 next = pmd_addr_end(addr, end);
2194 if (remap_pte_range(mm, pmd, addr, next,
2195 pfn + (addr >> PAGE_SHIFT), prot))
2196 return -ENOMEM;
2197 } while (pmd++, addr = next, addr != end);
2198 return 0;
2199}
2200
2201static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2202 unsigned long addr, unsigned long end,
2203 unsigned long pfn, pgprot_t prot)
2204{
2205 pud_t *pud;
2206 unsigned long next;
2207
2208 pfn -= addr >> PAGE_SHIFT;
2209 pud = pud_alloc(mm, pgd, addr);
2210 if (!pud)
2211 return -ENOMEM;
2212 do {
2213 next = pud_addr_end(addr, end);
2214 if (remap_pmd_range(mm, pud, addr, next,
2215 pfn + (addr >> PAGE_SHIFT), prot))
2216 return -ENOMEM;
2217 } while (pud++, addr = next, addr != end);
2218 return 0;
2219}
2220
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002221/**
2222 * remap_pfn_range - remap kernel memory to userspace
2223 * @vma: user vma to map to
2224 * @addr: target user address to start at
2225 * @pfn: physical address of kernel memory
2226 * @size: size of map area
2227 * @prot: page protection flags for this mapping
2228 *
2229 * Note: this is only safe if the mm semaphore is held when called.
2230 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2232 unsigned long pfn, unsigned long size, pgprot_t prot)
2233{
2234 pgd_t *pgd;
2235 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002236 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 struct mm_struct *mm = vma->vm_mm;
2238 int err;
2239
2240 /*
2241 * Physically remapped pages are special. Tell the
2242 * rest of the world about it:
2243 * VM_IO tells people not to look at these pages
2244 * (accesses can have side effects).
Hugh Dickins0b14c172005-11-21 21:32:15 -08002245 * VM_RESERVED is specified all over the place, because
2246 * in 2.4 it kept swapout's vma scan off this vma; but
2247 * in 2.6 the LRU scan won't even find its pages, so this
2248 * flag means no more than count its pages in reserved_vm,
2249 * and omit it from core dump, even when VM_IO turned off.
Linus Torvalds6aab3412005-11-28 14:34:23 -08002250 * VM_PFNMAP tells the core MM that the base pages are just
2251 * raw PFN mappings, and do not have a "struct page" associated
2252 * with them.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002253 *
2254 * There's a horrible special case to handle copy-on-write
2255 * behaviour that some programs depend on. We mark the "original"
2256 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 */
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07002258 if (addr == vma->vm_start && end == vma->vm_end) {
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002259 vma->vm_pgoff = pfn;
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07002260 vma->vm_flags |= VM_PFN_AT_MMAP;
Pallipadi, Venkatesh4bb9c5c2009-03-12 17:45:27 -07002261 } else if (is_cow_mapping(vma->vm_flags))
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002262 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002263
Linus Torvalds6aab3412005-11-28 14:34:23 -08002264 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002266 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08002267 if (err) {
2268 /*
2269 * To indicate that track_pfn related cleanup is not
2270 * needed from higher level routine calling unmap_vmas
2271 */
2272 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
Pallipadi, Venkatesh895791d2009-03-13 16:35:44 -07002273 vma->vm_flags &= ~VM_PFN_AT_MMAP;
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002274 return -EINVAL;
venkatesh.pallipadi@intel.coma3670612009-01-09 16:13:09 -08002275 }
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002276
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 BUG_ON(addr >= end);
2278 pfn -= addr >> PAGE_SHIFT;
2279 pgd = pgd_offset(mm, addr);
2280 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 do {
2282 next = pgd_addr_end(addr, end);
2283 err = remap_pud_range(mm, pgd, addr, next,
2284 pfn + (addr >> PAGE_SHIFT), prot);
2285 if (err)
2286 break;
2287 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002288
2289 if (err)
2290 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2291
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292 return err;
2293}
2294EXPORT_SYMBOL(remap_pfn_range);
2295
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002296static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2297 unsigned long addr, unsigned long end,
2298 pte_fn_t fn, void *data)
2299{
2300 pte_t *pte;
2301 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002302 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002303 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002304
2305 pte = (mm == &init_mm) ?
2306 pte_alloc_kernel(pmd, addr) :
2307 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2308 if (!pte)
2309 return -ENOMEM;
2310
2311 BUG_ON(pmd_huge(*pmd));
2312
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002313 arch_enter_lazy_mmu_mode();
2314
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002315 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002316
2317 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002318 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002319 if (err)
2320 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002321 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002322
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002323 arch_leave_lazy_mmu_mode();
2324
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002325 if (mm != &init_mm)
2326 pte_unmap_unlock(pte-1, ptl);
2327 return err;
2328}
2329
2330static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2331 unsigned long addr, unsigned long end,
2332 pte_fn_t fn, void *data)
2333{
2334 pmd_t *pmd;
2335 unsigned long next;
2336 int err;
2337
Andi Kleenceb86872008-07-23 21:27:50 -07002338 BUG_ON(pud_huge(*pud));
2339
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002340 pmd = pmd_alloc(mm, pud, addr);
2341 if (!pmd)
2342 return -ENOMEM;
2343 do {
2344 next = pmd_addr_end(addr, end);
2345 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2346 if (err)
2347 break;
2348 } while (pmd++, addr = next, addr != end);
2349 return err;
2350}
2351
2352static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2353 unsigned long addr, unsigned long end,
2354 pte_fn_t fn, void *data)
2355{
2356 pud_t *pud;
2357 unsigned long next;
2358 int err;
2359
2360 pud = pud_alloc(mm, pgd, addr);
2361 if (!pud)
2362 return -ENOMEM;
2363 do {
2364 next = pud_addr_end(addr, end);
2365 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2366 if (err)
2367 break;
2368 } while (pud++, addr = next, addr != end);
2369 return err;
2370}
2371
2372/*
2373 * Scan a region of virtual memory, filling in page tables as necessary
2374 * and calling a provided function on each leaf page table.
2375 */
2376int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2377 unsigned long size, pte_fn_t fn, void *data)
2378{
2379 pgd_t *pgd;
2380 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002381 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002382 int err;
2383
2384 BUG_ON(addr >= end);
2385 pgd = pgd_offset(mm, addr);
2386 do {
2387 next = pgd_addr_end(addr, end);
2388 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2389 if (err)
2390 break;
2391 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002392
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002393 return err;
2394}
2395EXPORT_SYMBOL_GPL(apply_to_page_range);
2396
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002398 * handle_pte_fault chooses page fault handler according to an entry
2399 * which was read non-atomically. Before making any commitment, on
2400 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002401 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002402 * must check under lock before unmapping the pte and proceeding
2403 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002404 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002405 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002406static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002407 pte_t *page_table, pte_t orig_pte)
2408{
2409 int same = 1;
2410#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2411 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002412 spinlock_t *ptl = pte_lockptr(mm, pmd);
2413 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002414 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002415 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002416 }
2417#endif
2418 pte_unmap(page_table);
2419 return same;
2420}
2421
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002422static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
Linus Torvalds6aab3412005-11-28 14:34:23 -08002423{
2424 /*
2425 * If the source page was a PFN mapping, we don't have
2426 * a "struct page" for it. We do a best-effort copy by
2427 * just copying from the original user address. If that
2428 * fails, we just zero-fill it. Live with it.
2429 */
2430 if (unlikely(!src)) {
2431 void *kaddr = kmap_atomic(dst, KM_USER0);
Linus Torvalds5d2a2db2005-11-29 14:07:55 -08002432 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2433
2434 /*
2435 * This really shouldn't fail, because the page is there
2436 * in the page tables. But it might just be unreadable,
2437 * in which case we just give up and fill the result with
2438 * zeroes.
2439 */
2440 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002441 clear_page(kaddr);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002442 kunmap_atomic(kaddr, KM_USER0);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002443 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002444 } else
2445 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002446}
2447
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 * This routine handles present pages, when users try to write
2450 * to a shared page. It is done by copying the page to a new address
2451 * and decrementing the shared-page counter for the old page.
2452 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453 * Note that this routine assumes that the protection checks have been
2454 * done by the caller (the low-level page fault routine in most cases).
2455 * Thus we can safely just mark it writable once we've done any necessary
2456 * COW.
2457 *
2458 * We also mark the page dirty at this point even though the page will
2459 * change only once the write actually happens. This avoids a few races,
2460 * and potentially makes it more efficient.
2461 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002462 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2463 * but allow concurrent faults), with pte both mapped and locked.
2464 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002466static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2467 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002468 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002469 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470{
Hugh Dickinse5bbe4d2005-11-29 16:54:51 +00002471 struct page *old_page, *new_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002473 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002474 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002475 struct page *dirty_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476
Linus Torvalds6aab3412005-11-28 14:34:23 -08002477 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002478 if (!old_page) {
2479 /*
2480 * VM_MIXEDMAP !pfn_valid() case
2481 *
2482 * We should not cow pages in a shared writeable mapping.
2483 * Just mark the pages writable as we can't do any dirty
2484 * accounting on raw pfn maps.
2485 */
2486 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2487 (VM_WRITE|VM_SHARED))
2488 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002489 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002490 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002492 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002493 * Take out anonymous pages first, anonymous shared vmas are
2494 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002495 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002496 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002497 if (!trylock_page(old_page)) {
2498 page_cache_get(old_page);
2499 pte_unmap_unlock(page_table, ptl);
2500 lock_page(old_page);
2501 page_table = pte_offset_map_lock(mm, pmd, address,
2502 &ptl);
2503 if (!pte_same(*page_table, orig_pte)) {
2504 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002505 goto unlock;
2506 }
2507 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002508 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002509 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002510 /*
2511 * The page is all ours. Move it to our anon_vma so
2512 * the rmap code will not search our parent or siblings.
2513 * Protected against the rmap code by the page lock.
2514 */
2515 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002516 unlock_page(old_page);
2517 goto reuse;
2518 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002519 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002520 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002521 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002522 /*
2523 * Only catch write-faults on shared writable pages,
2524 * read-only shared pages can get COWed by
2525 * get_user_pages(.write=1, .force=1).
2526 */
David Howells9637a5e2006-06-23 02:03:43 -07002527 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002528 struct vm_fault vmf;
2529 int tmp;
2530
2531 vmf.virtual_address = (void __user *)(address &
2532 PAGE_MASK);
2533 vmf.pgoff = old_page->index;
2534 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2535 vmf.page = old_page;
2536
David Howells9637a5e2006-06-23 02:03:43 -07002537 /*
2538 * Notify the address space that the page is about to
2539 * become writable so that it can prohibit this or wait
2540 * for the page to get into an appropriate state.
2541 *
2542 * We do this without the lock held, so that it can
2543 * sleep if it needs to.
2544 */
2545 page_cache_get(old_page);
2546 pte_unmap_unlock(page_table, ptl);
2547
Nick Pigginc2ec1752009-03-31 15:23:21 -07002548 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2549 if (unlikely(tmp &
2550 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2551 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002552 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002553 }
Nick Pigginb827e492009-04-30 15:08:16 -07002554 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2555 lock_page(old_page);
2556 if (!old_page->mapping) {
2557 ret = 0; /* retry the fault */
2558 unlock_page(old_page);
2559 goto unwritable_page;
2560 }
2561 } else
2562 VM_BUG_ON(!PageLocked(old_page));
David Howells9637a5e2006-06-23 02:03:43 -07002563
David Howells9637a5e2006-06-23 02:03:43 -07002564 /*
2565 * Since we dropped the lock we need to revalidate
2566 * the PTE as someone else may have changed it. If
2567 * they did, we just return, as we can count on the
2568 * MMU to tell us if they didn't also make it writable.
2569 */
2570 page_table = pte_offset_map_lock(mm, pmd, address,
2571 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002572 if (!pte_same(*page_table, orig_pte)) {
2573 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002574 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002575 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002576
2577 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002579 dirty_page = old_page;
2580 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002581
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002582reuse:
David Howells9637a5e2006-06-23 02:03:43 -07002583 flush_cache_page(vma, address, pte_pfn(orig_pte));
2584 entry = pte_mkyoung(orig_pte);
2585 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002586 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002587 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002588 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002589 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002590
2591 if (!dirty_page)
2592 return ret;
2593
2594 /*
2595 * Yes, Virginia, this is actually required to prevent a race
2596 * with clear_page_dirty_for_io() from clearing the page dirty
2597 * bit after it clear all dirty ptes, but before a racing
2598 * do_wp_page installs a dirty pte.
2599 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002600 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002601 */
2602 if (!page_mkwrite) {
2603 wait_on_page_locked(dirty_page);
2604 set_page_dirty_balance(dirty_page, page_mkwrite);
2605 }
2606 put_page(dirty_page);
2607 if (page_mkwrite) {
2608 struct address_space *mapping = dirty_page->mapping;
2609
2610 set_page_dirty(dirty_page);
2611 unlock_page(dirty_page);
2612 page_cache_release(dirty_page);
2613 if (mapping) {
2614 /*
2615 * Some device drivers do not set page.mapping
2616 * but still dirty their pages
2617 */
2618 balance_dirty_pages_ratelimited(mapping);
2619 }
2620 }
2621
2622 /* file_update_time outside page_lock */
2623 if (vma->vm_file)
2624 file_update_time(vma->vm_file);
2625
2626 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628
2629 /*
2630 * Ok, we need to copy. Oh, well..
2631 */
Nick Pigginb5810032005-10-29 18:16:12 -07002632 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002633gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002634 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635
2636 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002637 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002638
Hugh Dickins62eede62009-09-21 17:03:34 -07002639 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002640 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2641 if (!new_page)
2642 goto oom;
2643 } else {
2644 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2645 if (!new_page)
2646 goto oom;
2647 cow_user_page(new_page, old_page, address, vma);
2648 }
2649 __SetPageUptodate(new_page);
2650
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002651 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002652 goto oom_free_new;
2653
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 /*
2655 * Re-check the pte - we dropped the lock
2656 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002657 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002658 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002659 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002660 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002661 dec_mm_counter_fast(mm, MM_FILEPAGES);
2662 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002663 }
2664 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002665 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002666 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002667 entry = mk_pte(new_page, vma->vm_page_prot);
2668 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002669 /*
2670 * Clear the pte entry and flush it first, before updating the
2671 * pte with the new entry. This will avoid a race condition
2672 * seen in the presence of one thread doing SMC and another
2673 * thread doing COW.
2674 */
Izik Eidus828502d2009-09-21 17:01:51 -07002675 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002676 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002677 /*
2678 * We call the notify macro here because, when using secondary
2679 * mmu page tables (such as kvm shadow page tables), we want the
2680 * new page to be mapped directly into the secondary page table.
2681 */
2682 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002683 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002684 if (old_page) {
2685 /*
2686 * Only after switching the pte to the new page may
2687 * we remove the mapcount here. Otherwise another
2688 * process may come and find the rmap count decremented
2689 * before the pte is switched to the new page, and
2690 * "reuse" the old page writing into it while our pte
2691 * here still points into it and can be read by other
2692 * threads.
2693 *
2694 * The critical issue is to order this
2695 * page_remove_rmap with the ptp_clear_flush above.
2696 * Those stores are ordered by (if nothing else,)
2697 * the barrier present in the atomic_add_negative
2698 * in page_remove_rmap.
2699 *
2700 * Then the TLB flush in ptep_clear_flush ensures that
2701 * no process can access the old page before the
2702 * decremented mapcount is visible. And the old page
2703 * cannot be reused until after the decremented
2704 * mapcount is visible. So transitively, TLBs to
2705 * old page will be flushed before it can be reused.
2706 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002707 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002708 }
2709
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 /* Free the old page.. */
2711 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002712 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002713 } else
2714 mem_cgroup_uncharge_page(new_page);
2715
Hugh Dickins920fc352005-11-21 21:32:17 -08002716 if (new_page)
2717 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002718unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002719 pte_unmap_unlock(page_table, ptl);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002720 if (old_page) {
2721 /*
2722 * Don't let another task, with possibly unlocked vma,
2723 * keep the mlocked page.
2724 */
2725 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2726 lock_page(old_page); /* LRU manipulation */
2727 munlock_vma_page(old_page);
2728 unlock_page(old_page);
2729 }
2730 page_cache_release(old_page);
2731 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002732 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002733oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002734 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002735oom:
Nick Pigginb827e492009-04-30 15:08:16 -07002736 if (old_page) {
2737 if (page_mkwrite) {
2738 unlock_page(old_page);
2739 page_cache_release(old_page);
2740 }
Hugh Dickins920fc352005-11-21 21:32:17 -08002741 page_cache_release(old_page);
Nick Pigginb827e492009-04-30 15:08:16 -07002742 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002744
2745unwritable_page:
2746 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002747 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748}
2749
Peter Zijlstra97a89412011-05-24 17:12:04 -07002750static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751 unsigned long start_addr, unsigned long end_addr,
2752 struct zap_details *details)
2753{
Peter Zijlstra97a89412011-05-24 17:12:04 -07002754 zap_page_range(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755}
2756
2757static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2758 struct zap_details *details)
2759{
2760 struct vm_area_struct *vma;
2761 struct prio_tree_iter iter;
2762 pgoff_t vba, vea, zba, zea;
2763
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 vma_prio_tree_foreach(vma, &iter, root,
2765 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766
2767 vba = vma->vm_pgoff;
2768 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2769 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2770 zba = details->first_index;
2771 if (zba < vba)
2772 zba = vba;
2773 zea = details->last_index;
2774 if (zea > vea)
2775 zea = vea;
2776
Peter Zijlstra97a89412011-05-24 17:12:04 -07002777 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2779 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002780 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 }
2782}
2783
2784static inline void unmap_mapping_range_list(struct list_head *head,
2785 struct zap_details *details)
2786{
2787 struct vm_area_struct *vma;
2788
2789 /*
2790 * In nonlinear VMAs there is no correspondence between virtual address
2791 * offset and file offset. So we must perform an exhaustive search
2792 * across *all* the pages in each nonlinear VMA, not just the pages
2793 * whose virtual address lies outside the file truncation point.
2794 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795 list_for_each_entry(vma, head, shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002797 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 }
2799}
2800
2801/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002802 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
Martin Waitz3d410882005-06-23 22:05:21 -07002803 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 * @holebegin: byte in first page to unmap, relative to the start of
2805 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002806 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807 * must keep the partial page. In contrast, we must get rid of
2808 * partial pages.
2809 * @holelen: size of prospective hole in bytes. This will be rounded
2810 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2811 * end of the file.
2812 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2813 * but 0 when invalidating pagecache, don't throw away private data.
2814 */
2815void unmap_mapping_range(struct address_space *mapping,
2816 loff_t const holebegin, loff_t const holelen, int even_cows)
2817{
2818 struct zap_details details;
2819 pgoff_t hba = holebegin >> PAGE_SHIFT;
2820 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2821
2822 /* Check for overflow. */
2823 if (sizeof(holelen) > sizeof(hlen)) {
2824 long long holeend =
2825 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2826 if (holeend & ~(long long)ULONG_MAX)
2827 hlen = ULONG_MAX - hba + 1;
2828 }
2829
2830 details.check_mapping = even_cows? NULL: mapping;
2831 details.nonlinear_vma = NULL;
2832 details.first_index = hba;
2833 details.last_index = hba + hlen - 1;
2834 if (details.last_index < details.first_index)
2835 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836
Peter Zijlstra97a89412011-05-24 17:12:04 -07002837
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002838 mutex_lock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2840 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2841 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2842 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002843 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844}
2845EXPORT_SYMBOL(unmap_mapping_range);
2846
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002848 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2849 * but allow concurrent faults), and pte mapped but not yet locked.
2850 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002852static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2853 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002854 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002856 spinlock_t *ptl;
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002857 struct page *page, *swapcache = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07002858 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002860 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07002861 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002862 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002863 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002865 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002866 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002867
2868 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002869 if (unlikely(non_swap_entry(entry))) {
2870 if (is_migration_entry(entry)) {
2871 migration_entry_wait(mm, pmd, address);
2872 } else if (is_hwpoison_entry(entry)) {
2873 ret = VM_FAULT_HWPOISON;
2874 } else {
2875 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002876 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002877 }
Christoph Lameter06972122006-06-23 02:03:35 -07002878 goto out;
2879 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002880 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 page = lookup_swap_cache(entry);
2882 if (!page) {
Hugh Dickinsa5c9b692009-06-23 12:36:58 -07002883 grab_swap_token(mm); /* Contend for token _before_ read-in */
Hugh Dickins02098fe2008-02-04 22:28:42 -08002884 page = swapin_readahead(entry,
2885 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886 if (!page) {
2887 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002888 * Back out if somebody else faulted in this pte
2889 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002891 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 if (likely(pte_same(*page_table, orig_pte)))
2893 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002894 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002895 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896 }
2897
2898 /* Had to read the page from swap area: Major fault */
2899 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002900 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002901 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002902 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002903 /*
2904 * hwpoisoned dirty swapcache pages are kept for killing
2905 * owner processes (which may be unknown at hwpoison time)
2906 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002907 ret = VM_FAULT_HWPOISON;
2908 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Andi Kleen4779cb32009-10-14 01:51:41 +02002909 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 }
2911
Michel Lespinassed065bd82010-10-26 14:21:57 -07002912 locked = lock_page_or_retry(page, mm, flags);
Balbir Singh20a10222007-11-14 17:00:33 -08002913 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002914 if (!locked) {
2915 ret |= VM_FAULT_RETRY;
2916 goto out_release;
2917 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002919 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002920 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2921 * release the swapcache from under us. The page pin, and pte_same
2922 * test below, are not enough to exclude that. Even if it is still
2923 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002924 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002925 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002926 goto out_page;
2927
2928 if (ksm_might_need_to_copy(page, vma, address)) {
2929 swapcache = page;
2930 page = ksm_does_need_to_copy(page, vma, address);
2931
2932 if (unlikely(!page)) {
2933 ret = VM_FAULT_OOM;
2934 page = swapcache;
2935 swapcache = NULL;
2936 goto out_page;
2937 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08002938 }
2939
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002940 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002941 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002942 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002943 }
2944
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002946 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002948 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002949 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002950 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002951
2952 if (unlikely(!PageUptodate(page))) {
2953 ret = VM_FAULT_SIGBUS;
2954 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955 }
2956
KAMEZAWA Hiroyuki8c7c6e32009-01-07 18:08:00 -08002957 /*
2958 * The page isn't present yet, go ahead with the fault.
2959 *
2960 * Be careful about the sequence of operations here.
2961 * To get its accounting right, reuse_swap_page() must be called
2962 * while the page is counted on swap but not yet in mapcount i.e.
2963 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2964 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08002965 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2966 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2967 * in page->private. In this case, a record in swap_cgroup is silently
2968 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e32009-01-07 18:08:00 -08002969 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002971 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08002972 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002974 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002976 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002977 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002978 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980 flush_icache_page(vma, page);
2981 set_pte_at(mm, address, page_table, pte);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002982 do_page_add_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08002983 /* It's better to call commit-charge after rmap is established */
2984 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002986 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07002987 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002988 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002989 unlock_page(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002990 if (swapcache) {
2991 /*
2992 * Hold the lock to avoid the swap entry to be reused
2993 * until we take the PT lock for the pte_same() check
2994 * (to avoid false positives from pte_same). For
2995 * further safety release the lock after the swap_free
2996 * so that the swap count won't change under a
2997 * parallel locked swapcache.
2998 */
2999 unlock_page(swapcache);
3000 page_cache_release(swapcache);
3001 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003002
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003003 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08003004 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3005 if (ret & VM_FAULT_ERROR)
3006 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 goto out;
3008 }
3009
3010 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003011 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003012unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003013 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014out:
3015 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003016out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08003017 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003018 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07003019out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003020 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02003021out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003022 page_cache_release(page);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003023 if (swapcache) {
3024 unlock_page(swapcache);
3025 page_cache_release(swapcache);
3026 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003027 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028}
3029
3030/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003031 * This is like a special single-page "expand_{down|up}wards()",
3032 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07003033 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07003034 */
3035static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3036{
3037 address &= PAGE_MASK;
3038 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003039 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07003040
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003041 /*
3042 * Is there a mapping abutting this one below?
3043 *
3044 * That's only ok if it's the same stack mapping
3045 * that has gotten split..
3046 */
3047 if (prev && prev->vm_end == address)
3048 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3049
Michal Hockod05f3162011-05-24 17:11:44 -07003050 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003051 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003052 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3053 struct vm_area_struct *next = vma->vm_next;
3054
3055 /* As VM_GROWSDOWN but s/below/above/ */
3056 if (next && next->vm_start == address + PAGE_SIZE)
3057 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3058
3059 expand_upwards(vma, address + PAGE_SIZE);
3060 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07003061 return 0;
3062}
3063
3064/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003065 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3066 * but allow concurrent faults), and pte mapped but not yet locked.
3067 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003069static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3070 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003071 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003073 struct page *page;
3074 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076
Linus Torvalds11ac5522010-08-14 11:44:56 -07003077 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003078
Linus Torvalds11ac5522010-08-14 11:44:56 -07003079 /* Check if we need to add a guard page to the stack */
3080 if (check_stack_guard_page(vma, address) < 0)
3081 return VM_FAULT_SIGBUS;
3082
3083 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003084 if (!(flags & FAULT_FLAG_WRITE)) {
3085 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3086 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003087 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003088 if (!pte_none(*page_table))
3089 goto unlock;
3090 goto setpte;
3091 }
3092
Nick Piggin557ed1f2007-10-16 01:24:40 -07003093 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003094 if (unlikely(anon_vma_prepare(vma)))
3095 goto oom;
3096 page = alloc_zeroed_user_highpage_movable(vma, address);
3097 if (!page)
3098 goto oom;
Nick Piggin0ed361d2008-02-04 22:29:34 -08003099 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003101 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003102 goto oom_free_page;
3103
Nick Piggin557ed1f2007-10-16 01:24:40 -07003104 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003105 if (vma->vm_flags & VM_WRITE)
3106 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003107
Nick Piggin557ed1f2007-10-16 01:24:40 -07003108 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003109 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003110 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003111
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003112 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003113 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003114setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003115 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116
3117 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003118 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003119unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003120 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003121 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003122release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003123 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003124 page_cache_release(page);
3125 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003126oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003127 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003128oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129 return VM_FAULT_OOM;
3130}
3131
3132/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003133 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003135 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3136 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137 *
3138 * As this is called only for pages that do not currently exist, we
3139 * do not need to flush old virtual caches or the TLB.
3140 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003141 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003142 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003143 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003144 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003145static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003146 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003147 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003149 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003150 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003151 struct page *page;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003152 struct page *cow_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003153 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154 int anon = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003155 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003156 struct vm_fault vmf;
3157 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003158 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003159
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003160 /*
3161 * If we do COW later, allocate page befor taking lock_page()
3162 * on the file cache page. This will reduce lock holding time.
3163 */
3164 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3165
3166 if (unlikely(anon_vma_prepare(vma)))
3167 return VM_FAULT_OOM;
3168
3169 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3170 if (!cow_page)
3171 return VM_FAULT_OOM;
3172
3173 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3174 page_cache_release(cow_page);
3175 return VM_FAULT_OOM;
3176 }
3177 } else
3178 cow_page = NULL;
3179
Nick Piggind0217ac2007-07-19 01:47:03 -07003180 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3181 vmf.pgoff = pgoff;
3182 vmf.flags = flags;
3183 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003185 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003186 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3187 VM_FAULT_RETRY)))
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003188 goto uncharge_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189
Andi Kleena3b947e2009-09-16 11:50:08 +02003190 if (unlikely(PageHWPoison(vmf.page))) {
3191 if (ret & VM_FAULT_LOCKED)
3192 unlock_page(vmf.page);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003193 ret = VM_FAULT_HWPOISON;
3194 goto uncharge_out;
Andi Kleena3b947e2009-09-16 11:50:08 +02003195 }
3196
Nick Piggind00806b2007-07-19 01:46:57 -07003197 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003198 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003199 * locked.
3200 */
Nick Piggin83c54072007-07-19 01:47:05 -07003201 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003202 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003203 else
Nick Piggind0217ac2007-07-19 01:47:03 -07003204 VM_BUG_ON(!PageLocked(vmf.page));
Nick Piggind00806b2007-07-19 01:46:57 -07003205
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206 /*
3207 * Should we do an early C-O-W break?
3208 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003209 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003210 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003211 if (!(vma->vm_flags & VM_SHARED)) {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003212 page = cow_page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003213 anon = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003214 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003215 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003216 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003217 /*
3218 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003219 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003220 * to become writable
3221 */
Mark Fasheh69676142007-07-19 01:47:00 -07003222 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003223 int tmp;
3224
Mark Fasheh69676142007-07-19 01:47:00 -07003225 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003226 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003227 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3228 if (unlikely(tmp &
3229 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3230 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003231 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003232 }
Nick Pigginb827e492009-04-30 15:08:16 -07003233 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3234 lock_page(page);
3235 if (!page->mapping) {
3236 ret = 0; /* retry the fault */
3237 unlock_page(page);
3238 goto unwritable_page;
3239 }
3240 } else
3241 VM_BUG_ON(!PageLocked(page));
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003242 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003243 }
3244 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003245
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 }
3247
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003248 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003249
3250 /*
3251 * This silly early PAGE_DIRTY setting removes a race
3252 * due to the bad i386 page protection. But it's valid
3253 * for other architectures too.
3254 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003255 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 * an exclusive copy of the page, or this is a shared mapping,
3257 * so we can make it writable and dirty to avoid having to
3258 * handle that later.
3259 */
3260 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003261 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003262 flush_icache_page(vma, page);
3263 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003264 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003265 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003266 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003267 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003268 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003269 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003270 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003271 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003272 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003273 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003274 get_page(dirty_page);
3275 }
Hugh Dickins42946212005-10-29 18:16:05 -07003276 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003277 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003278
3279 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003280 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281 } else {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003282 if (cow_page)
3283 mem_cgroup_uncharge_page(cow_page);
Nick Piggind00806b2007-07-19 01:46:57 -07003284 if (anon)
3285 page_cache_release(page);
3286 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003287 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003288 }
3289
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003290 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003291
Nick Pigginb827e492009-04-30 15:08:16 -07003292 if (dirty_page) {
3293 struct address_space *mapping = page->mapping;
3294
3295 if (set_page_dirty(dirty_page))
3296 page_mkwrite = 1;
3297 unlock_page(dirty_page);
3298 put_page(dirty_page);
3299 if (page_mkwrite && mapping) {
3300 /*
3301 * Some device drivers do not set page.mapping but still
3302 * dirty their pages
3303 */
3304 balance_dirty_pages_ratelimited(mapping);
3305 }
3306
3307 /* file_update_time outside page_lock */
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003308 if (vma->vm_file)
3309 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003310 } else {
3311 unlock_page(vmf.page);
3312 if (anon)
3313 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003314 }
Nick Piggind00806b2007-07-19 01:46:57 -07003315
Nick Piggin83c54072007-07-19 01:47:05 -07003316 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003317
3318unwritable_page:
3319 page_cache_release(page);
3320 return ret;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003321uncharge_out:
3322 /* fs's fault handler get error */
3323 if (cow_page) {
3324 mem_cgroup_uncharge_page(cow_page);
3325 page_cache_release(cow_page);
3326 }
3327 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003328}
Nick Piggind00806b2007-07-19 01:46:57 -07003329
Nick Piggin54cb8822007-07-19 01:46:59 -07003330static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3331 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003332 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003333{
3334 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003335 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003336
Hugh Dickins16abfa02007-10-04 16:56:06 +01003337 pte_unmap(page_table);
3338 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003339}
3340
Jes Sorensenf4b81802006-09-27 01:50:10 -07003341/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003342 * Fault of a previously existing named mapping. Repopulate the pte
3343 * from the encoded file_pte if possible. This enables swappable
3344 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003345 *
3346 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3347 * but allow concurrent faults), and pte mapped but not yet locked.
3348 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003350static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003351 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003352 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353{
Hugh Dickins65500d22005-10-29 18:15:59 -07003354 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003355
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003356 flags |= FAULT_FLAG_NONLINEAR;
3357
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003358 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003359 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360
Hugh Dickins2509ef22009-01-06 14:40:10 -08003361 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003362 /*
3363 * Page table corrupted: show pte and kill process.
3364 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003365 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003366 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003367 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003368
3369 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003370 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371}
3372
3373/*
3374 * These routines also need to handle stuff like marking pages dirty
3375 * and/or accessed for architectures that don't do it in hardware (most
3376 * RISC architectures). The early dirtying is also good on the i386.
3377 *
3378 * There is also a hook called "update_mmu_cache()" that architectures
3379 * with external mmu caches can use to update those (ie the Sparc or
3380 * PowerPC hashed page tables that act as extended TLBs).
3381 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003382 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3383 * but allow concurrent faults), and pte mapped but not yet locked.
3384 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003386int handle_pte_fault(struct mm_struct *mm,
3387 struct vm_area_struct *vma, unsigned long address,
3388 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389{
3390 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003391 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003393 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003395 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003396 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003397 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003398 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003399 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003400 }
3401 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003402 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003403 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003405 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003406 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003407 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003408 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409 }
3410
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003411 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003412 spin_lock(ptl);
3413 if (unlikely(!pte_same(*pte, entry)))
3414 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003415 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003417 return do_wp_page(mm, vma, address,
3418 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419 entry = pte_mkdirty(entry);
3420 }
3421 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003422 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003423 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003424 } else {
3425 /*
3426 * This is needed only for protection faults but the arch code
3427 * is not yet telling us if this is a protection fault or not.
3428 * This still avoids useless tlb flushes for .text page faults
3429 * with threads.
3430 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003431 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003432 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003433 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003434unlock:
3435 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003436 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437}
3438
3439/*
3440 * By the time we get here, we already hold the mm semaphore
3441 */
Nick Piggin83c54072007-07-19 01:47:05 -07003442int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvaldsd06063c2009-04-10 09:01:23 -07003443 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444{
3445 pgd_t *pgd;
3446 pud_t *pud;
3447 pmd_t *pmd;
3448 pte_t *pte;
3449
3450 __set_current_state(TASK_RUNNING);
3451
Christoph Lameterf8891e52006-06-30 01:55:45 -07003452 count_vm_event(PGFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003453 mem_cgroup_count_vm_event(mm, PGFAULT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003455 /* do counter updates before entering really critical section. */
3456 check_sync_rss_stat(current);
3457
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003458 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003459 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003460
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462 pud = pud_alloc(mm, pgd, address);
3463 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003464 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465 pmd = pmd_alloc(mm, pud, address);
3466 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003467 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003468 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3469 if (!vma->vm_ops)
3470 return do_huge_pmd_anonymous_page(mm, vma, address,
3471 pmd, flags);
3472 } else {
3473 pmd_t orig_pmd = *pmd;
3474 barrier();
3475 if (pmd_trans_huge(orig_pmd)) {
3476 if (flags & FAULT_FLAG_WRITE &&
3477 !pmd_write(orig_pmd) &&
3478 !pmd_trans_splitting(orig_pmd))
3479 return do_huge_pmd_wp_page(mm, vma, address,
3480 pmd, orig_pmd);
3481 return 0;
3482 }
3483 }
3484
3485 /*
3486 * Use __pte_alloc instead of pte_alloc_map, because we can't
3487 * run pte_offset_map on the pmd, if an huge pmd could
3488 * materialize from under us from a different thread.
3489 */
Mel Gormancc036382011-04-27 15:26:56 -07003490 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003491 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003492 /* if an huge pmd materialized from under us just retry later */
3493 if (unlikely(pmd_trans_huge(*pmd)))
3494 return 0;
3495 /*
3496 * A regular pmd is established and it can't morph into a huge pmd
3497 * from under us anymore at this point because we hold the mmap_sem
3498 * read mode and khugepaged takes it in write mode. So now it's
3499 * safe to run pte_offset_map().
3500 */
3501 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003503 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504}
3505
3506#ifndef __PAGETABLE_PUD_FOLDED
3507/*
3508 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003509 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003511int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003513 pud_t *new = pud_alloc_one(mm, address);
3514 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003515 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516
Nick Piggin362a61a2008-05-14 06:37:36 +02003517 smp_wmb(); /* See comment in __pte_alloc */
3518
Hugh Dickins872fec12005-10-29 18:16:21 -07003519 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003520 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003521 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003522 else
3523 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003524 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003525 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526}
3527#endif /* __PAGETABLE_PUD_FOLDED */
3528
3529#ifndef __PAGETABLE_PMD_FOLDED
3530/*
3531 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003532 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003534int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003536 pmd_t *new = pmd_alloc_one(mm, address);
3537 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003538 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539
Nick Piggin362a61a2008-05-14 06:37:36 +02003540 smp_wmb(); /* See comment in __pte_alloc */
3541
Hugh Dickins872fec12005-10-29 18:16:21 -07003542 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003543#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003544 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003545 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003546 else
3547 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003549 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003550 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003551 else
3552 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003554 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003555 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556}
3557#endif /* __PAGETABLE_PMD_FOLDED */
3558
3559int make_pages_present(unsigned long addr, unsigned long end)
3560{
3561 int ret, len, write;
3562 struct vm_area_struct * vma;
3563
3564 vma = find_vma(current->mm, addr);
3565 if (!vma)
KOSAKI Motohiroa4770972008-08-04 13:41:14 -07003566 return -ENOMEM;
Michel Lespinasse5ecfda02011-01-13 15:46:09 -08003567 /*
3568 * We want to touch writable mappings with a write fault in order
3569 * to break COW, except for shared mappings because these don't COW
3570 * and we would not want to dirty them for nothing.
3571 */
3572 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
Eric Sesterhenn5bcb28b2006-03-26 18:30:52 +02003573 BUG_ON(addr >= end);
3574 BUG_ON(end > vma->vm_end);
Rolf Eike Beer68e116a2007-07-15 23:38:03 -07003575 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576 ret = get_user_pages(current, current->mm, addr,
3577 len, write, 0, NULL, NULL);
Lee Schermerhornc11d69d2008-10-18 20:26:56 -07003578 if (ret < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579 return ret;
Lee Schermerhorn9978ad52008-10-18 20:26:56 -07003580 return ret == len ? 0 : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581}
3582
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583#if !defined(__HAVE_ARCH_GATE_AREA)
3584
3585#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003586static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587
3588static int __init gate_vma_init(void)
3589{
3590 gate_vma.vm_mm = NULL;
3591 gate_vma.vm_start = FIXADDR_USER_START;
3592 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003593 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3594 gate_vma.vm_page_prot = __P101;
Roland McGrathf47aef52007-01-26 00:56:49 -08003595 /*
3596 * Make sure the vDSO gets into every core dump.
3597 * Dumping its contents makes post-mortem fully interpretable later
3598 * without matching up the same kernel and hardware config to see
3599 * what PC values meant.
3600 */
3601 gate_vma.vm_flags |= VM_ALWAYSDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602 return 0;
3603}
3604__initcall(gate_vma_init);
3605#endif
3606
Stephen Wilson31db58b2011-03-13 15:49:15 -04003607struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608{
3609#ifdef AT_SYSINFO_EHDR
3610 return &gate_vma;
3611#else
3612 return NULL;
3613#endif
3614}
3615
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003616int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617{
3618#ifdef AT_SYSINFO_EHDR
3619 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3620 return 1;
3621#endif
3622 return 0;
3623}
3624
3625#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003626
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003627static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003628 pte_t **ptepp, spinlock_t **ptlp)
3629{
3630 pgd_t *pgd;
3631 pud_t *pud;
3632 pmd_t *pmd;
3633 pte_t *ptep;
3634
3635 pgd = pgd_offset(mm, address);
3636 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3637 goto out;
3638
3639 pud = pud_offset(pgd, address);
3640 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3641 goto out;
3642
3643 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055a2011-01-13 15:46:54 -08003644 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003645 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3646 goto out;
3647
3648 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3649 if (pmd_huge(*pmd))
3650 goto out;
3651
3652 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3653 if (!ptep)
3654 goto out;
3655 if (!pte_present(*ptep))
3656 goto unlock;
3657 *ptepp = ptep;
3658 return 0;
3659unlock:
3660 pte_unmap_unlock(ptep, *ptlp);
3661out:
3662 return -EINVAL;
3663}
3664
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003665static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3666 pte_t **ptepp, spinlock_t **ptlp)
3667{
3668 int res;
3669
3670 /* (void) is needed to make gcc happy */
3671 (void) __cond_lock(*ptlp,
3672 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3673 return res;
3674}
3675
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003676/**
3677 * follow_pfn - look up PFN at a user virtual address
3678 * @vma: memory mapping
3679 * @address: user virtual address
3680 * @pfn: location to store found PFN
3681 *
3682 * Only IO mappings and raw PFN mappings are allowed.
3683 *
3684 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3685 */
3686int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3687 unsigned long *pfn)
3688{
3689 int ret = -EINVAL;
3690 spinlock_t *ptl;
3691 pte_t *ptep;
3692
3693 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3694 return ret;
3695
3696 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3697 if (ret)
3698 return ret;
3699 *pfn = pte_pfn(*ptep);
3700 pte_unmap_unlock(ptep, ptl);
3701 return 0;
3702}
3703EXPORT_SYMBOL(follow_pfn);
3704
Rik van Riel28b2ee22008-07-23 21:27:05 -07003705#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003706int follow_phys(struct vm_area_struct *vma,
3707 unsigned long address, unsigned int flags,
3708 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003709{
Johannes Weiner03668a42009-06-16 15:32:34 -07003710 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003711 pte_t *ptep, pte;
3712 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003713
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003714 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3715 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003716
Johannes Weiner03668a42009-06-16 15:32:34 -07003717 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003718 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003719 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003720
Rik van Riel28b2ee22008-07-23 21:27:05 -07003721 if ((flags & FOLL_WRITE) && !pte_write(pte))
3722 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003723
3724 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003725 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003726
Johannes Weiner03668a42009-06-16 15:32:34 -07003727 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003728unlock:
3729 pte_unmap_unlock(ptep, ptl);
3730out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003731 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003732}
3733
3734int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3735 void *buf, int len, int write)
3736{
3737 resource_size_t phys_addr;
3738 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003739 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003740 int offset = addr & (PAGE_SIZE-1);
3741
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003742 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003743 return -EINVAL;
3744
3745 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3746 if (write)
3747 memcpy_toio(maddr + offset, buf, len);
3748 else
3749 memcpy_fromio(buf, maddr + offset, len);
3750 iounmap(maddr);
3751
3752 return len;
3753}
3754#endif
3755
David Howells0ec76a12006-09-27 01:50:15 -07003756/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003757 * Access another process' address space as given in mm. If non-NULL, use the
3758 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003759 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003760static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3761 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003762{
David Howells0ec76a12006-09-27 01:50:15 -07003763 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003764 void *old_buf = buf;
3765
David Howells0ec76a12006-09-27 01:50:15 -07003766 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003767 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003768 while (len) {
3769 int bytes, ret, offset;
3770 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003771 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003772
3773 ret = get_user_pages(tsk, mm, addr, 1,
3774 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003775 if (ret <= 0) {
3776 /*
3777 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3778 * we can access using slightly different code.
3779 */
3780#ifdef CONFIG_HAVE_IOREMAP_PROT
3781 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003782 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003783 break;
3784 if (vma->vm_ops && vma->vm_ops->access)
3785 ret = vma->vm_ops->access(vma, addr, buf,
3786 len, write);
3787 if (ret <= 0)
3788#endif
3789 break;
3790 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07003791 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003792 bytes = len;
3793 offset = addr & (PAGE_SIZE-1);
3794 if (bytes > PAGE_SIZE-offset)
3795 bytes = PAGE_SIZE-offset;
3796
3797 maddr = kmap(page);
3798 if (write) {
3799 copy_to_user_page(vma, page, addr,
3800 maddr + offset, buf, bytes);
3801 set_page_dirty_lock(page);
3802 } else {
3803 copy_from_user_page(vma, page, addr,
3804 buf, maddr + offset, bytes);
3805 }
3806 kunmap(page);
3807 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003808 }
David Howells0ec76a12006-09-27 01:50:15 -07003809 len -= bytes;
3810 buf += bytes;
3811 addr += bytes;
3812 }
3813 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003814
3815 return buf - old_buf;
3816}
Andi Kleen03252912008-01-30 13:33:18 +01003817
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003818/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003819 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003820 * @mm: the mm_struct of the target address space
3821 * @addr: start address to access
3822 * @buf: source or destination buffer
3823 * @len: number of bytes to transfer
3824 * @write: whether the access is a write
3825 *
3826 * The caller must hold a reference on @mm.
3827 */
3828int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3829 void *buf, int len, int write)
3830{
3831 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3832}
3833
Andi Kleen03252912008-01-30 13:33:18 +01003834/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003835 * Access another process' address space.
3836 * Source/target buffer must be kernel space,
3837 * Do not walk the page table directly, use get_user_pages
3838 */
3839int access_process_vm(struct task_struct *tsk, unsigned long addr,
3840 void *buf, int len, int write)
3841{
3842 struct mm_struct *mm;
3843 int ret;
3844
3845 mm = get_task_mm(tsk);
3846 if (!mm)
3847 return 0;
3848
3849 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3850 mmput(mm);
3851
3852 return ret;
3853}
3854
Andi Kleen03252912008-01-30 13:33:18 +01003855/*
3856 * Print the name of a VMA.
3857 */
3858void print_vma_addr(char *prefix, unsigned long ip)
3859{
3860 struct mm_struct *mm = current->mm;
3861 struct vm_area_struct *vma;
3862
Ingo Molnare8bff742008-02-13 20:21:06 +01003863 /*
3864 * Do not print if we are in atomic
3865 * contexts (in exception stacks, etc.):
3866 */
3867 if (preempt_count())
3868 return;
3869
Andi Kleen03252912008-01-30 13:33:18 +01003870 down_read(&mm->mmap_sem);
3871 vma = find_vma(mm, ip);
3872 if (vma && vma->vm_file) {
3873 struct file *f = vma->vm_file;
3874 char *buf = (char *)__get_free_page(GFP_KERNEL);
3875 if (buf) {
3876 char *p, *s;
3877
Jan Blunckcf28b482008-02-14 19:38:44 -08003878 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003879 if (IS_ERR(p))
3880 p = "?";
3881 s = strrchr(p, '/');
3882 if (s)
3883 p = s+1;
3884 printk("%s%s[%lx+%lx]", prefix, p,
3885 vma->vm_start,
3886 vma->vm_end - vma->vm_start);
3887 free_page((unsigned long)buf);
3888 }
3889 }
3890 up_read(&current->mm->mmap_sem);
3891}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003892
3893#ifdef CONFIG_PROVE_LOCKING
3894void might_fault(void)
3895{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003896 /*
3897 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3898 * holding the mmap_sem, this is safe because kernel memory doesn't
3899 * get paged out, therefore we'll never actually fault, and the
3900 * below annotations will generate false positives.
3901 */
3902 if (segment_eq(get_fs(), KERNEL_DS))
3903 return;
3904
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003905 might_sleep();
3906 /*
3907 * it would be nicer only to annotate paths which are not under
3908 * pagefault_disable, however that requires a larger audit and
3909 * providing helpers like get_user_atomic.
3910 */
3911 if (!in_atomic() && current->mm)
3912 might_lock_read(&current->mm->mmap_sem);
3913}
3914EXPORT_SYMBOL(might_fault);
3915#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003916
3917#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3918static void clear_gigantic_page(struct page *page,
3919 unsigned long addr,
3920 unsigned int pages_per_huge_page)
3921{
3922 int i;
3923 struct page *p = page;
3924
3925 might_sleep();
3926 for (i = 0; i < pages_per_huge_page;
3927 i++, p = mem_map_next(p, page, i)) {
3928 cond_resched();
3929 clear_user_highpage(p, addr + i * PAGE_SIZE);
3930 }
3931}
3932void clear_huge_page(struct page *page,
3933 unsigned long addr, unsigned int pages_per_huge_page)
3934{
3935 int i;
3936
3937 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3938 clear_gigantic_page(page, addr, pages_per_huge_page);
3939 return;
3940 }
3941
3942 might_sleep();
3943 for (i = 0; i < pages_per_huge_page; i++) {
3944 cond_resched();
3945 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3946 }
3947}
3948
3949static void copy_user_gigantic_page(struct page *dst, struct page *src,
3950 unsigned long addr,
3951 struct vm_area_struct *vma,
3952 unsigned int pages_per_huge_page)
3953{
3954 int i;
3955 struct page *dst_base = dst;
3956 struct page *src_base = src;
3957
3958 for (i = 0; i < pages_per_huge_page; ) {
3959 cond_resched();
3960 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3961
3962 i++;
3963 dst = mem_map_next(dst, dst_base, i);
3964 src = mem_map_next(src, src_base, i);
3965 }
3966}
3967
3968void copy_user_huge_page(struct page *dst, struct page *src,
3969 unsigned long addr, struct vm_area_struct *vma,
3970 unsigned int pages_per_huge_page)
3971{
3972 int i;
3973
3974 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3975 copy_user_gigantic_page(dst, src, addr, vma,
3976 pages_per_huge_page);
3977 return;
3978 }
3979
3980 might_sleep();
3981 for (i = 0; i < pages_per_huge_page; i++) {
3982 cond_resched();
3983 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3984 }
3985}
3986#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */