blob: be6a0c0d4ae081d48edd26f09d37f67cc70b1c52 [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>
Mel Gorman4daae3b2012-11-02 11:33:45 +000060#include <linux/migrate.h>
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -080061#include <linux/string.h>
Dan Williams0abdd7a2014-01-21 15:48:12 -080062#include <linux/dma-debug.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
Alexey Dobriyan6952b612009-09-18 23:55:55 +040064#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070065#include <asm/pgalloc.h>
66#include <asm/uaccess.h>
67#include <asm/tlb.h>
68#include <asm/tlbflush.h>
69#include <asm/pgtable.h>
70
Jan Beulich42b77722008-07-23 21:27:10 -070071#include "internal.h"
72
Peter Zijlstra90572892013-10-07 11:29:20 +010073#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
74#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
Peter Zijlstra75980e92013-02-22 16:34:32 -080075#endif
76
Andy Whitcroftd41dee32005-06-23 00:07:54 -070077#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070078/* use the per-pgdat data instead for discontigmem - mbligh */
79unsigned long max_mapnr;
80struct page *mem_map;
81
82EXPORT_SYMBOL(max_mapnr);
83EXPORT_SYMBOL(mem_map);
84#endif
85
Linus Torvalds1da177e2005-04-16 15:20:36 -070086/*
87 * A number of key systems in x86 including ioremap() rely on the assumption
88 * that high_memory defines the upper bound on direct map memory, then end
89 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
90 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
91 * and ZONE_HIGHMEM.
92 */
93void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070094
Linus Torvalds1da177e2005-04-16 15:20:36 -070095EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
Ingo Molnar32a93232008-02-06 22:39:44 +010097/*
98 * Randomize the address space (stacks, mmaps, brk, etc.).
99 *
100 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
101 * as ancient (libc5 based) binaries can segfault. )
102 */
103int randomize_va_space __read_mostly =
104#ifdef CONFIG_COMPAT_BRK
105 1;
106#else
107 2;
108#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100109
110static int __init disable_randmaps(char *s)
111{
112 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800113 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100114}
115__setup("norandmaps", disable_randmaps);
116
Hugh Dickins62eede62009-09-21 17:03:34 -0700117unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700118unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700119
120/*
121 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
122 */
123static int __init init_zero_pfn(void)
124{
125 zero_pfn = page_to_pfn(ZERO_PAGE(0));
126 return 0;
127}
128core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100129
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800130
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800131#if defined(SPLIT_RSS_COUNTING)
132
David Rientjesea48cf72012-03-21 16:34:13 -0700133void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800134{
135 int i;
136
137 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700138 if (current->rss_stat.count[i]) {
139 add_mm_counter(mm, i, current->rss_stat.count[i]);
140 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800141 }
142 }
David Rientjes05af2e12012-03-21 16:34:13 -0700143 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800144}
145
146static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
147{
148 struct task_struct *task = current;
149
150 if (likely(task->mm == mm))
151 task->rss_stat.count[member] += val;
152 else
153 add_mm_counter(mm, member, val);
154}
155#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
156#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
157
158/* sync counter once per 64 page faults */
159#define TASK_RSS_EVENTS_THRESH (64)
160static void check_sync_rss_stat(struct task_struct *task)
161{
162 if (unlikely(task != current))
163 return;
164 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700165 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800166}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700167#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800168
169#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
170#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
171
172static void check_sync_rss_stat(struct task_struct *task)
173{
174}
175
Peter Zijlstra9547d012011-05-24 17:12:14 -0700176#endif /* SPLIT_RSS_COUNTING */
177
178#ifdef HAVE_GENERIC_MMU_GATHER
179
180static int tlb_next_batch(struct mmu_gather *tlb)
181{
182 struct mmu_gather_batch *batch;
183
184 batch = tlb->active;
185 if (batch->next) {
186 tlb->active = batch->next;
187 return 1;
188 }
189
Michal Hocko53a59fc2013-01-04 15:35:12 -0800190 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
191 return 0;
192
Peter Zijlstra9547d012011-05-24 17:12:14 -0700193 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
194 if (!batch)
195 return 0;
196
Michal Hocko53a59fc2013-01-04 15:35:12 -0800197 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700198 batch->next = NULL;
199 batch->nr = 0;
200 batch->max = MAX_GATHER_BATCH;
201
202 tlb->active->next = batch;
203 tlb->active = batch;
204
205 return 1;
206}
207
208/* tlb_gather_mmu
209 * Called to initialize an (on-stack) mmu_gather structure for page-table
210 * tear-down from @mm. The @fullmm argument is used when @mm is without
211 * users and we're going to destroy the full address space (exit/execve).
212 */
Linus Torvalds2b047252013-08-15 11:42:25 -0700213void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700214{
215 tlb->mm = mm;
216
Linus Torvalds2b047252013-08-15 11:42:25 -0700217 /* Is it from 0 to ~0? */
218 tlb->fullmm = !(start | (end+1));
Dave Hansen1de14c32013-04-12 16:23:54 -0700219 tlb->need_flush_all = 0;
Linus Torvalds2b047252013-08-15 11:42:25 -0700220 tlb->start = start;
221 tlb->end = end;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700222 tlb->need_flush = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700223 tlb->local.next = NULL;
224 tlb->local.nr = 0;
225 tlb->local.max = ARRAY_SIZE(tlb->__pages);
226 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800227 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700228
229#ifdef CONFIG_HAVE_RCU_TABLE_FREE
230 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800231#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700232}
233
234void tlb_flush_mmu(struct mmu_gather *tlb)
235{
236 struct mmu_gather_batch *batch;
237
238 if (!tlb->need_flush)
239 return;
240 tlb->need_flush = 0;
241 tlb_flush(tlb);
242#ifdef CONFIG_HAVE_RCU_TABLE_FREE
243 tlb_table_flush(tlb);
244#endif
245
Peter Zijlstra9547d012011-05-24 17:12:14 -0700246 for (batch = &tlb->local; batch; batch = batch->next) {
247 free_pages_and_swap_cache(batch->pages, batch->nr);
248 batch->nr = 0;
249 }
250 tlb->active = &tlb->local;
251}
252
253/* tlb_finish_mmu
254 * Called at the end of the shootdown operation to free up any resources
255 * that were required.
256 */
257void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
258{
259 struct mmu_gather_batch *batch, *next;
260
261 tlb_flush_mmu(tlb);
262
263 /* keep the page table cache within bounds */
264 check_pgt_cache();
265
266 for (batch = tlb->local.next; batch; batch = next) {
267 next = batch->next;
268 free_pages((unsigned long)batch, 0);
269 }
270 tlb->local.next = NULL;
271}
272
273/* __tlb_remove_page
274 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
275 * handling the additional races in SMP caused by other CPUs caching valid
276 * mappings in their TLBs. Returns the number of free page slots left.
277 * When out of page slots we must call tlb_flush_mmu().
278 */
279int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
280{
281 struct mmu_gather_batch *batch;
282
Shaohua Lif21760b2012-01-12 17:19:16 -0800283 VM_BUG_ON(!tlb->need_flush);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700284
Peter Zijlstra9547d012011-05-24 17:12:14 -0700285 batch = tlb->active;
286 batch->pages[batch->nr++] = page;
287 if (batch->nr == batch->max) {
288 if (!tlb_next_batch(tlb))
289 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700290 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700291 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800292 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700293
294 return batch->max - batch->nr;
295}
296
297#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800298
Peter Zijlstra26723912011-05-24 17:12:00 -0700299#ifdef CONFIG_HAVE_RCU_TABLE_FREE
300
301/*
302 * See the comment near struct mmu_table_batch.
303 */
304
305static void tlb_remove_table_smp_sync(void *arg)
306{
307 /* Simply deliver the interrupt */
308}
309
310static void tlb_remove_table_one(void *table)
311{
312 /*
313 * This isn't an RCU grace period and hence the page-tables cannot be
314 * assumed to be actually RCU-freed.
315 *
316 * It is however sufficient for software page-table walkers that rely on
317 * IRQ disabling. See the comment near struct mmu_table_batch.
318 */
319 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
320 __tlb_remove_table(table);
321}
322
323static void tlb_remove_table_rcu(struct rcu_head *head)
324{
325 struct mmu_table_batch *batch;
326 int i;
327
328 batch = container_of(head, struct mmu_table_batch, rcu);
329
330 for (i = 0; i < batch->nr; i++)
331 __tlb_remove_table(batch->tables[i]);
332
333 free_page((unsigned long)batch);
334}
335
336void tlb_table_flush(struct mmu_gather *tlb)
337{
338 struct mmu_table_batch **batch = &tlb->batch;
339
340 if (*batch) {
341 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
342 *batch = NULL;
343 }
344}
345
346void tlb_remove_table(struct mmu_gather *tlb, void *table)
347{
348 struct mmu_table_batch **batch = &tlb->batch;
349
350 tlb->need_flush = 1;
351
352 /*
353 * When there's less then two users of this mm there cannot be a
354 * concurrent page-table walk.
355 */
356 if (atomic_read(&tlb->mm->mm_users) < 2) {
357 __tlb_remove_table(table);
358 return;
359 }
360
361 if (*batch == NULL) {
362 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
363 if (*batch == NULL) {
364 tlb_remove_table_one(table);
365 return;
366 }
367 (*batch)->nr = 0;
368 }
369 (*batch)->tables[(*batch)->nr++] = table;
370 if ((*batch)->nr == MAX_TABLE_BATCH)
371 tlb_table_flush(tlb);
372}
373
Peter Zijlstra9547d012011-05-24 17:12:14 -0700374#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700375
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 * Note: this doesn't free the actual pages themselves. That
378 * has been handled earlier when unmapping all the memory regions.
379 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000380static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
381 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800383 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700384 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000385 pte_free_tlb(tlb, token, addr);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800386 atomic_long_dec(&tlb->mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387}
388
Hugh Dickinse0da3822005-04-19 13:29:15 -0700389static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
390 unsigned long addr, unsigned long end,
391 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392{
393 pmd_t *pmd;
394 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700395 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396
Hugh Dickinse0da3822005-04-19 13:29:15 -0700397 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 do {
400 next = pmd_addr_end(addr, end);
401 if (pmd_none_or_clear_bad(pmd))
402 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000403 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404 } while (pmd++, addr = next, addr != end);
405
Hugh Dickinse0da3822005-04-19 13:29:15 -0700406 start &= PUD_MASK;
407 if (start < floor)
408 return;
409 if (ceiling) {
410 ceiling &= PUD_MASK;
411 if (!ceiling)
412 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700414 if (end - 1 > ceiling - 1)
415 return;
416
417 pmd = pmd_offset(pud, start);
418 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000419 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420}
421
Hugh Dickinse0da3822005-04-19 13:29:15 -0700422static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
423 unsigned long addr, unsigned long end,
424 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425{
426 pud_t *pud;
427 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700428 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429
Hugh Dickinse0da3822005-04-19 13:29:15 -0700430 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 do {
433 next = pud_addr_end(addr, end);
434 if (pud_none_or_clear_bad(pud))
435 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700436 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 } while (pud++, addr = next, addr != end);
438
Hugh Dickinse0da3822005-04-19 13:29:15 -0700439 start &= PGDIR_MASK;
440 if (start < floor)
441 return;
442 if (ceiling) {
443 ceiling &= PGDIR_MASK;
444 if (!ceiling)
445 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700447 if (end - 1 > ceiling - 1)
448 return;
449
450 pud = pud_offset(pgd, start);
451 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000452 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453}
454
455/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700456 * This function frees user-level page tables of a process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 */
Jan Beulich42b77722008-07-23 21:27:10 -0700458void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700459 unsigned long addr, unsigned long end,
460 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461{
462 pgd_t *pgd;
463 unsigned long next;
464
Hugh Dickinse0da3822005-04-19 13:29:15 -0700465 /*
466 * The next few lines have given us lots of grief...
467 *
468 * Why are we testing PMD* at this top level? Because often
469 * there will be no work to do at all, and we'd prefer not to
470 * go all the way down to the bottom just to discover that.
471 *
472 * Why all these "- 1"s? Because 0 represents both the bottom
473 * of the address space and the top of it (using -1 for the
474 * top wouldn't help much: the masks would do the wrong thing).
475 * The rule is that addr 0 and floor 0 refer to the bottom of
476 * the address space, but end 0 and ceiling 0 refer to the top
477 * Comparisons need to use "end - 1" and "ceiling - 1" (though
478 * that end 0 case should be mythical).
479 *
480 * Wherever addr is brought up or ceiling brought down, we must
481 * be careful to reject "the opposite 0" before it confuses the
482 * subsequent tests. But what about where end is brought down
483 * by PMD_SIZE below? no, end can't go down to 0 there.
484 *
485 * Whereas we round start (addr) and ceiling down, by different
486 * masks at different levels, in order to test whether a table
487 * now has no other vmas using it, so can be freed, we don't
488 * bother to round floor or end up - the tests don't need that.
489 */
490
491 addr &= PMD_MASK;
492 if (addr < floor) {
493 addr += PMD_SIZE;
494 if (!addr)
495 return;
496 }
497 if (ceiling) {
498 ceiling &= PMD_MASK;
499 if (!ceiling)
500 return;
501 }
502 if (end - 1 > ceiling - 1)
503 end -= PMD_SIZE;
504 if (addr > end - 1)
505 return;
506
Jan Beulich42b77722008-07-23 21:27:10 -0700507 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 do {
509 next = pgd_addr_end(addr, end);
510 if (pgd_none_or_clear_bad(pgd))
511 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700512 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700514}
515
Jan Beulich42b77722008-07-23 21:27:10 -0700516void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700517 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700518{
519 while (vma) {
520 struct vm_area_struct *next = vma->vm_next;
521 unsigned long addr = vma->vm_start;
522
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700523 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000524 * Hide vma from rmap and truncate_pagecache before freeing
525 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700526 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800527 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700528 unlink_file_vma(vma);
529
David Gibson9da61ae2006-03-22 00:08:57 -0800530 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700531 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700532 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700533 } else {
534 /*
535 * Optimization: gather nearby vmas into one call down
536 */
537 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800538 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700539 vma = next;
540 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800541 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700542 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700543 }
544 free_pgd_range(tlb, addr, vma->vm_end,
545 floor, next? next->vm_start: ceiling);
546 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700547 vma = next;
548 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549}
550
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800551int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
552 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800554 spinlock_t *ptl;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800555 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800556 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700557 if (!new)
558 return -ENOMEM;
559
Nick Piggin362a61a2008-05-14 06:37:36 +0200560 /*
561 * Ensure all pte setup (eg. pte page lock and page clearing) are
562 * visible before the pte is made visible to other CPUs by being
563 * put into page tables.
564 *
565 * The other side of the story is the pointer chasing in the page
566 * table walking code (when walking the page table without locking;
567 * ie. most of the time). Fortunately, these data accesses consist
568 * of a chain of data-dependent loads, meaning most CPUs (alpha
569 * being the notable exception) will already guarantee loads are
570 * seen in-order. See the alpha page table accessors for the
571 * smp_read_barrier_depends() barriers in page table walking code.
572 */
573 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
574
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800575 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800576 wait_split_huge_page = 0;
577 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800578 atomic_long_inc(&mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800580 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800581 } else if (unlikely(pmd_trans_splitting(*pmd)))
582 wait_split_huge_page = 1;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800583 spin_unlock(ptl);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800584 if (new)
585 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800586 if (wait_split_huge_page)
587 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700588 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589}
590
Hugh Dickins1bb36302005-10-29 18:16:22 -0700591int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700593 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
594 if (!new)
595 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596
Nick Piggin362a61a2008-05-14 06:37:36 +0200597 smp_wmb(); /* See comment in __pte_alloc */
598
Hugh Dickins1bb36302005-10-29 18:16:22 -0700599 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800600 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700601 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800602 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800603 } else
604 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700605 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800606 if (new)
607 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700608 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609}
610
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800611static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700612{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800613 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
614}
615
616static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
617{
618 int i;
619
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800620 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700621 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800622 for (i = 0; i < NR_MM_COUNTERS; i++)
623 if (rss[i])
624 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700625}
626
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800628 * This function is called to print an error when a bad pte
629 * is found. For example, we might have a PFN-mapped pte in
630 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700631 *
632 * The calling function must still handle the error.
633 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800634static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
635 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700636{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800637 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
638 pud_t *pud = pud_offset(pgd, addr);
639 pmd_t *pmd = pmd_offset(pud, addr);
640 struct address_space *mapping;
641 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800642 static unsigned long resume;
643 static unsigned long nr_shown;
644 static unsigned long nr_unshown;
645
646 /*
647 * Allow a burst of 60 reports, then keep quiet for that minute;
648 * or allow a steady drip of one report per second.
649 */
650 if (nr_shown == 60) {
651 if (time_before(jiffies, resume)) {
652 nr_unshown++;
653 return;
654 }
655 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800656 printk(KERN_ALERT
657 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800658 nr_unshown);
659 nr_unshown = 0;
660 }
661 nr_shown = 0;
662 }
663 if (nr_shown++ == 0)
664 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800665
666 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
667 index = linear_page_index(vma, addr);
668
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800669 printk(KERN_ALERT
670 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800671 current->comm,
672 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800673 if (page)
Dave Hansenf0b791a2014-01-23 15:52:49 -0800674 dump_page(page, "bad pte");
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800675 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800676 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
677 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
678 /*
679 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
680 */
681 if (vma->vm_ops)
Joe Perches071361d2012-12-12 10:19:12 -0800682 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
683 vma->vm_ops->fault);
Al Viro72c2d532013-09-22 16:27:52 -0400684 if (vma->vm_file)
Joe Perches071361d2012-12-12 10:19:12 -0800685 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
686 vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700687 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030688 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700689}
690
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700691static inline bool is_cow_mapping(vm_flags_t flags)
Linus Torvalds67121172005-12-11 20:38:17 -0800692{
693 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
694}
695
Nick Pigginb5810032005-10-29 18:16:12 -0700696/*
Nick Piggin7e675132008-04-28 02:13:00 -0700697 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800698 *
Nick Piggin7e675132008-04-28 02:13:00 -0700699 * "Special" mappings do not wish to be associated with a "struct page" (either
700 * it doesn't exist, or it exists but they don't want to touch it). In this
701 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700702 *
Nick Piggin7e675132008-04-28 02:13:00 -0700703 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
704 * pte bit, in which case this function is trivial. Secondly, an architecture
705 * may not have a spare pte bit, which requires a more complicated scheme,
706 * described below.
707 *
708 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
709 * special mapping (even if there are underlying and valid "struct pages").
710 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800711 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700712 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
713 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700714 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
715 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800716 *
717 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
718 *
Nick Piggin7e675132008-04-28 02:13:00 -0700719 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700720 *
Nick Piggin7e675132008-04-28 02:13:00 -0700721 * This restricts such mappings to be a linear translation from virtual address
722 * to pfn. To get around this restriction, we allow arbitrary mappings so long
723 * as the vma is not a COW mapping; in that case, we know that all ptes are
724 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700725 *
726 *
Nick Piggin7e675132008-04-28 02:13:00 -0700727 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
728 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700729 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
730 * page" backing, however the difference is that _all_ pages with a struct
731 * page (that is, those where pfn_valid is true) are refcounted and considered
732 * normal pages by the VM. The disadvantage is that pages are refcounted
733 * (which can be slower and simply not an option for some PFNMAP users). The
734 * advantage is that we don't have to follow the strict linearity rule of
735 * PFNMAP mappings in order to support COWable mappings.
736 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800737 */
Nick Piggin7e675132008-04-28 02:13:00 -0700738#ifdef __HAVE_ARCH_PTE_SPECIAL
739# define HAVE_PTE_SPECIAL 1
740#else
741# define HAVE_PTE_SPECIAL 0
742#endif
743struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
744 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800745{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800746 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700747
748 if (HAVE_PTE_SPECIAL) {
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800749 if (likely(!pte_special(pte)))
750 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700751 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
752 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700753 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800754 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700755 return NULL;
756 }
757
758 /* !HAVE_PTE_SPECIAL case follows: */
759
Jared Hulbertb379d792008-04-28 02:12:58 -0700760 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
761 if (vma->vm_flags & VM_MIXEDMAP) {
762 if (!pfn_valid(pfn))
763 return NULL;
764 goto out;
765 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700766 unsigned long off;
767 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700768 if (pfn == vma->vm_pgoff + off)
769 return NULL;
770 if (!is_cow_mapping(vma->vm_flags))
771 return NULL;
772 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800773 }
774
Hugh Dickins62eede62009-09-21 17:03:34 -0700775 if (is_zero_pfn(pfn))
776 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800777check_pfn:
778 if (unlikely(pfn > highest_memmap_pfn)) {
779 print_bad_pte(vma, addr, pte, NULL);
780 return NULL;
781 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800782
783 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700784 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700785 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800786 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700787out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800788 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800789}
790
791/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 * copy one vm_area from one task to the other. Assumes the page tables
793 * already present in the new task to be cleared in the whole range
794 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 */
796
Hugh Dickins570a335b2009-12-14 17:58:46 -0800797static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700799 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700800 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801{
Nick Pigginb5810032005-10-29 18:16:12 -0700802 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 pte_t pte = *src_pte;
804 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805
806 /* pte contains position in swap or file, so copy. */
807 if (unlikely(!pte_present(pte))) {
808 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700809 swp_entry_t entry = pte_to_swp_entry(pte);
810
Hugh Dickins570a335b2009-12-14 17:58:46 -0800811 if (swap_duplicate(entry) < 0)
812 return entry.val;
813
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 /* make sure dst_mm is on swapoff's mmlist. */
815 if (unlikely(list_empty(&dst_mm->mmlist))) {
816 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700817 if (list_empty(&dst_mm->mmlist))
818 list_add(&dst_mm->mmlist,
819 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 spin_unlock(&mmlist_lock);
821 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800822 if (likely(!non_swap_entry(entry)))
823 rss[MM_SWAPENTS]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800824 else if (is_migration_entry(entry)) {
825 page = migration_entry_to_page(entry);
826
827 if (PageAnon(page))
828 rss[MM_ANONPAGES]++;
829 else
830 rss[MM_FILEPAGES]++;
831
832 if (is_write_migration_entry(entry) &&
833 is_cow_mapping(vm_flags)) {
834 /*
835 * COW mappings require pages in both
836 * parent and child to be set to read.
837 */
838 make_migration_entry_read(&entry);
839 pte = swp_entry_to_pte(entry);
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700840 if (pte_swp_soft_dirty(*src_pte))
841 pte = pte_swp_mksoft_dirty(pte);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800842 set_pte_at(src_mm, addr, src_pte, pte);
843 }
Christoph Lameter06972122006-06-23 02:03:35 -0700844 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700846 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 }
848
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849 /*
850 * If it's a COW mapping, write protect it both
851 * in the parent and the child
852 */
Linus Torvalds67121172005-12-11 20:38:17 -0800853 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700855 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 }
857
858 /*
859 * If it's a shared mapping, mark it clean in
860 * the child
861 */
862 if (vm_flags & VM_SHARED)
863 pte = pte_mkclean(pte);
864 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800865
866 page = vm_normal_page(vma, addr, pte);
867 if (page) {
868 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700869 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800870 if (PageAnon(page))
871 rss[MM_ANONPAGES]++;
872 else
873 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800874 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700875
876out_set_pte:
877 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800878 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879}
880
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800881int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
882 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
883 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700885 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700887 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700888 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800889 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800890 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891
892again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800893 init_rss_vec(rss);
894
Hugh Dickinsc74df322005-10-29 18:16:23 -0700895 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 if (!dst_pte)
897 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700898 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700899 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700900 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700901 orig_src_pte = src_pte;
902 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700903 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 do {
906 /*
907 * We are holding two locks at this point - either of them
908 * could generate latencies in another task on another CPU.
909 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700910 if (progress >= 32) {
911 progress = 0;
912 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100913 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700914 break;
915 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 if (pte_none(*src_pte)) {
917 progress++;
918 continue;
919 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800920 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
921 vma, addr, rss);
922 if (entry.val)
923 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 progress += 8;
925 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700927 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700928 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700929 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800930 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700931 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700932 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800933
934 if (entry.val) {
935 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
936 return -ENOMEM;
937 progress = 0;
938 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 if (addr != end)
940 goto again;
941 return 0;
942}
943
944static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
945 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
946 unsigned long addr, unsigned long end)
947{
948 pmd_t *src_pmd, *dst_pmd;
949 unsigned long next;
950
951 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
952 if (!dst_pmd)
953 return -ENOMEM;
954 src_pmd = pmd_offset(src_pud, addr);
955 do {
956 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800957 if (pmd_trans_huge(*src_pmd)) {
958 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800959 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800960 err = copy_huge_pmd(dst_mm, src_mm,
961 dst_pmd, src_pmd, addr, vma);
962 if (err == -ENOMEM)
963 return -ENOMEM;
964 if (!err)
965 continue;
966 /* fall through */
967 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 if (pmd_none_or_clear_bad(src_pmd))
969 continue;
970 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
971 vma, addr, next))
972 return -ENOMEM;
973 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
974 return 0;
975}
976
977static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
978 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
979 unsigned long addr, unsigned long end)
980{
981 pud_t *src_pud, *dst_pud;
982 unsigned long next;
983
984 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
985 if (!dst_pud)
986 return -ENOMEM;
987 src_pud = pud_offset(src_pgd, addr);
988 do {
989 next = pud_addr_end(addr, end);
990 if (pud_none_or_clear_bad(src_pud))
991 continue;
992 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
993 vma, addr, next))
994 return -ENOMEM;
995 } while (dst_pud++, src_pud++, addr = next, addr != end);
996 return 0;
997}
998
999int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1000 struct vm_area_struct *vma)
1001{
1002 pgd_t *src_pgd, *dst_pgd;
1003 unsigned long next;
1004 unsigned long addr = vma->vm_start;
1005 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001006 unsigned long mmun_start; /* For mmu_notifiers */
1007 unsigned long mmun_end; /* For mmu_notifiers */
1008 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001009 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010
Nick Piggind9928952005-08-28 16:49:11 +10001011 /*
1012 * Don't copy ptes where a page fault will fill them correctly.
1013 * Fork becomes much lighter when there are big shared or private
1014 * readonly mappings. The tradeoff is that copy_page_range is more
1015 * efficient than faulting.
1016 */
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001017 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1018 VM_PFNMAP | VM_MIXEDMAP))) {
Nick Piggind9928952005-08-28 16:49:11 +10001019 if (!vma->anon_vma)
1020 return 0;
1021 }
1022
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 if (is_vm_hugetlb_page(vma))
1024 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1025
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001026 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001027 /*
1028 * We do not free on error cases below as remove_vma
1029 * gets called on error from higher level routine
1030 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001031 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001032 if (ret)
1033 return ret;
1034 }
1035
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001036 /*
1037 * We need to invalidate the secondary MMU mappings only when
1038 * there could be a permission downgrade on the ptes of the
1039 * parent mm. And a permission downgrade will only happen if
1040 * is_cow_mapping() returns true.
1041 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001042 is_cow = is_cow_mapping(vma->vm_flags);
1043 mmun_start = addr;
1044 mmun_end = end;
1045 if (is_cow)
1046 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1047 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001048
1049 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050 dst_pgd = pgd_offset(dst_mm, addr);
1051 src_pgd = pgd_offset(src_mm, addr);
1052 do {
1053 next = pgd_addr_end(addr, end);
1054 if (pgd_none_or_clear_bad(src_pgd))
1055 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001056 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1057 vma, addr, next))) {
1058 ret = -ENOMEM;
1059 break;
1060 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001062
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001063 if (is_cow)
1064 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001065 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066}
1067
Robin Holt51c6f662005-11-13 16:06:42 -08001068static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001069 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001071 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072{
Nick Pigginb5810032005-10-29 18:16:12 -07001073 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001074 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001075 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001076 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001077 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001078 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001079
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001080again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001081 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001082 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1083 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001084 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 do {
1086 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001087 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001089 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001090
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001092 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001093
Linus Torvalds6aab3412005-11-28 14:34:23 -08001094 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 if (unlikely(details) && page) {
1096 /*
1097 * unmap_shared_mapping_pages() wants to
1098 * invalidate cache without truncating:
1099 * unmap shared but keep private pages.
1100 */
1101 if (details->check_mapping &&
1102 details->check_mapping != page->mapping)
1103 continue;
1104 /*
1105 * Each page->index must be checked when
1106 * invalidating or truncating nonlinear.
1107 */
1108 if (details->nonlinear_vma &&
1109 (page->index < details->first_index ||
1110 page->index > details->last_index))
1111 continue;
1112 }
Nick Pigginb5810032005-10-29 18:16:12 -07001113 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001114 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 tlb_remove_tlb_entry(tlb, pte, addr);
1116 if (unlikely(!page))
1117 continue;
1118 if (unlikely(details) && details->nonlinear_vma
1119 && linear_page_index(details->nonlinear_vma,
Cyrill Gorcunov41bb3472013-08-13 16:00:51 -07001120 addr) != page->index) {
1121 pte_t ptfile = pgoff_to_pte(page->index);
1122 if (pte_soft_dirty(ptent))
1123 pte_file_mksoft_dirty(ptfile);
1124 set_pte_at(mm, addr, pte, ptfile);
1125 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001127 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001128 else {
1129 if (pte_dirty(ptent))
1130 set_page_dirty(page);
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001131 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001132 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001133 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001134 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001135 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001136 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001137 if (unlikely(page_mapcount(page) < 0))
1138 print_bad_pte(vma, addr, ptent, page);
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001139 force_flush = !__tlb_remove_page(tlb, page);
1140 if (force_flush)
1141 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 continue;
1143 }
1144 /*
1145 * If details->check_mapping, we leave swap entries;
1146 * if details->nonlinear_vma, we leave file entries.
1147 */
1148 if (unlikely(details))
1149 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001150 if (pte_file(ptent)) {
1151 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1152 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001153 } else {
1154 swp_entry_t entry = pte_to_swp_entry(ptent);
1155
1156 if (!non_swap_entry(entry))
1157 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001158 else if (is_migration_entry(entry)) {
1159 struct page *page;
1160
1161 page = migration_entry_to_page(entry);
1162
1163 if (PageAnon(page))
1164 rss[MM_ANONPAGES]--;
1165 else
1166 rss[MM_FILEPAGES]--;
1167 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001168 if (unlikely(!free_swap_and_cache(entry)))
1169 print_bad_pte(vma, addr, ptent, NULL);
1170 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001171 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001172 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001173
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001174 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001175 arch_leave_lazy_mmu_mode();
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001176 pte_unmap_unlock(start_pte, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -08001177
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001178 /*
1179 * mmu_gather ran out of room to batch pages, we break out of
1180 * the PTE lock to avoid doing the potential expensive TLB invalidate
1181 * and page-free while holding it.
1182 */
1183 if (force_flush) {
Linus Torvalds2b047252013-08-15 11:42:25 -07001184 unsigned long old_end;
1185
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001186 force_flush = 0;
Alex Shi597e1c32012-06-28 09:02:21 +08001187
Linus Torvalds2b047252013-08-15 11:42:25 -07001188 /*
1189 * Flush the TLB just for the previous segment,
1190 * then update the range to be the remaining
1191 * TLB range.
1192 */
1193 old_end = tlb->end;
Vineet Guptae6c495a2013-07-03 15:03:31 -07001194 tlb->end = addr;
Linus Torvalds2b047252013-08-15 11:42:25 -07001195
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001196 tlb_flush_mmu(tlb);
Linus Torvalds2b047252013-08-15 11:42:25 -07001197
1198 tlb->start = addr;
1199 tlb->end = old_end;
1200
1201 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001202 goto again;
1203 }
1204
Robin Holt51c6f662005-11-13 16:06:42 -08001205 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206}
1207
Robin Holt51c6f662005-11-13 16:06:42 -08001208static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001209 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001211 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212{
1213 pmd_t *pmd;
1214 unsigned long next;
1215
1216 pmd = pmd_offset(pud, addr);
1217 do {
1218 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001219 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001220 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001221#ifdef CONFIG_DEBUG_VM
1222 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1223 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1224 __func__, addr, end,
1225 vma->vm_start,
1226 vma->vm_end);
1227 BUG();
1228 }
1229#endif
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001230 split_huge_page_pmd(vma, addr, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001231 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001232 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001233 /* fall through */
1234 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001235 /*
1236 * Here there can be other concurrent MADV_DONTNEED or
1237 * trans huge page faults running, and if the pmd is
1238 * none or trans huge it can change under us. This is
1239 * because MADV_DONTNEED holds the mmap_sem in read
1240 * mode.
1241 */
1242 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1243 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001244 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001245next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001246 cond_resched();
1247 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001248
1249 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250}
1251
Robin Holt51c6f662005-11-13 16:06:42 -08001252static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001253 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001255 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256{
1257 pud_t *pud;
1258 unsigned long next;
1259
1260 pud = pud_offset(pgd, addr);
1261 do {
1262 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001263 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001265 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1266 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001267
1268 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269}
1270
Al Viro038c7aa2012-03-05 13:25:09 -05001271static void unmap_page_range(struct mmu_gather *tlb,
1272 struct vm_area_struct *vma,
1273 unsigned long addr, unsigned long end,
1274 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275{
1276 pgd_t *pgd;
1277 unsigned long next;
1278
1279 if (details && !details->check_mapping && !details->nonlinear_vma)
1280 details = NULL;
1281
1282 BUG_ON(addr >= end);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001283 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 tlb_start_vma(tlb, vma);
1285 pgd = pgd_offset(vma->vm_mm, addr);
1286 do {
1287 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001288 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001290 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1291 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 tlb_end_vma(tlb, vma);
KAMEZAWA Hiroyuki569b8462009-12-15 16:47:03 -08001293 mem_cgroup_uncharge_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001295
Al Virof5cc4ee2012-03-05 14:14:20 -05001296
1297static void unmap_single_vma(struct mmu_gather *tlb,
1298 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001299 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001300 struct zap_details *details)
1301{
1302 unsigned long start = max(vma->vm_start, start_addr);
1303 unsigned long end;
1304
1305 if (start >= vma->vm_end)
1306 return;
1307 end = min(vma->vm_end, end_addr);
1308 if (end <= vma->vm_start)
1309 return;
1310
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301311 if (vma->vm_file)
1312 uprobe_munmap(vma, start, end);
1313
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001314 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001315 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001316
1317 if (start != end) {
1318 if (unlikely(is_vm_hugetlb_page(vma))) {
1319 /*
1320 * It is undesirable to test vma->vm_file as it
1321 * should be non-null for valid hugetlb area.
1322 * However, vm_file will be NULL in the error
1323 * cleanup path of do_mmap_pgoff. When
1324 * hugetlbfs ->mmap method fails,
1325 * do_mmap_pgoff() nullifies vma->vm_file
1326 * before calling this function to clean up.
1327 * Since no pte has actually been setup, it is
1328 * safe to do nothing in this case.
1329 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001330 if (vma->vm_file) {
1331 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Mel Gormand8333522012-07-31 16:46:20 -07001332 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001333 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1334 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001335 } else
1336 unmap_page_range(tlb, vma, start, end, details);
1337 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338}
1339
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340/**
1341 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001342 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 * @vma: the starting vma
1344 * @start_addr: virtual address at which to start unmapping
1345 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001347 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 * Only addresses between `start' and `end' will be unmapped.
1350 *
1351 * The VMA list must be sorted in ascending virtual address order.
1352 *
1353 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1354 * range after unmap_vmas() returns. So the only responsibility here is to
1355 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1356 * drops the lock and schedules.
1357 */
Al Viro6e8bb012012-03-05 13:41:15 -05001358void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001360 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001362 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001364 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001365 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001366 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001367 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368}
1369
1370/**
1371 * zap_page_range - remove user pages in a given range
1372 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001373 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 * @size: number of bytes to zap
1375 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001376 *
1377 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001379void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 unsigned long size, struct zap_details *details)
1381{
1382 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001383 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001384 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001387 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001388 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001389 mmu_notifier_invalidate_range_start(mm, start, end);
1390 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001391 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001392 mmu_notifier_invalidate_range_end(mm, start, end);
1393 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394}
1395
Jack Steinerc627f9c2008-07-29 22:33:53 -07001396/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001397 * zap_page_range_single - remove user pages in a given range
1398 * @vma: vm_area_struct holding the applicable pages
1399 * @address: starting address of pages to zap
1400 * @size: number of bytes to zap
1401 * @details: details of nonlinear truncation or shared cache invalidation
1402 *
1403 * The range must fit into one VMA.
1404 */
1405static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1406 unsigned long size, struct zap_details *details)
1407{
1408 struct mm_struct *mm = vma->vm_mm;
1409 struct mmu_gather tlb;
1410 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001411
1412 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001413 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001414 update_hiwater_rss(mm);
1415 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001416 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001417 mmu_notifier_invalidate_range_end(mm, address, end);
1418 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419}
1420
Jack Steinerc627f9c2008-07-29 22:33:53 -07001421/**
1422 * zap_vma_ptes - remove ptes mapping the vma
1423 * @vma: vm_area_struct holding ptes to be zapped
1424 * @address: starting address of pages to zap
1425 * @size: number of bytes to zap
1426 *
1427 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1428 *
1429 * The entire address range must be fully contained within the vma.
1430 *
1431 * Returns 0 if successful.
1432 */
1433int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1434 unsigned long size)
1435{
1436 if (address < vma->vm_start || address + size > vma->vm_end ||
1437 !(vma->vm_flags & VM_PFNMAP))
1438 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001439 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001440 return 0;
1441}
1442EXPORT_SYMBOL_GPL(zap_vma_ptes);
1443
Johannes Weiner142762b2010-05-24 14:32:39 -07001444/**
Michel Lespinasse240aade2013-02-22 16:35:56 -08001445 * follow_page_mask - look up a page descriptor from a user-virtual address
Johannes Weiner142762b2010-05-24 14:32:39 -07001446 * @vma: vm_area_struct mapping @address
1447 * @address: virtual address to look up
1448 * @flags: flags modifying lookup behaviour
Michel Lespinasse240aade2013-02-22 16:35:56 -08001449 * @page_mask: on output, *page_mask is set according to the size of the page
Johannes Weiner142762b2010-05-24 14:32:39 -07001450 *
1451 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1452 *
1453 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1454 * an error pointer if there is a mapping to something not represented
1455 * by a page descriptor (see also vm_normal_page()).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 */
Michel Lespinasse240aade2013-02-22 16:35:56 -08001457struct page *follow_page_mask(struct vm_area_struct *vma,
1458 unsigned long address, unsigned int flags,
1459 unsigned int *page_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460{
1461 pgd_t *pgd;
1462 pud_t *pud;
1463 pmd_t *pmd;
1464 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001465 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466 struct page *page;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001467 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468
Michel Lespinasse240aade2013-02-22 16:35:56 -08001469 *page_mask = 0;
1470
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001471 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1472 if (!IS_ERR(page)) {
1473 BUG_ON(flags & FOLL_GET);
1474 goto out;
1475 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001477 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 pgd = pgd_offset(mm, address);
1479 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001480 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481
1482 pud = pud_offset(pgd, address);
Andi Kleenceb86872008-07-23 21:27:50 -07001483 if (pud_none(*pud))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001484 goto no_page_table;
Hugh Dickins8a076512011-01-13 15:46:52 -08001485 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie632a932013-09-11 14:22:04 -07001486 if (flags & FOLL_GET)
1487 goto out;
Andi Kleenceb86872008-07-23 21:27:50 -07001488 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1489 goto out;
1490 }
1491 if (unlikely(pud_bad(*pud)))
1492 goto no_page_table;
1493
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 pmd = pmd_offset(pud, address);
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001495 if (pmd_none(*pmd))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001496 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001497 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001498 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
Naoya Horiguchie632a932013-09-11 14:22:04 -07001499 if (flags & FOLL_GET) {
1500 /*
1501 * Refcount on tail pages are not well-defined and
1502 * shouldn't be taken. The caller should handle a NULL
1503 * return when trying to follow tail pages.
1504 */
1505 if (PageHead(page))
1506 get_page(page);
1507 else {
1508 page = NULL;
1509 goto out;
1510 }
1511 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001512 goto out;
1513 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001514 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1515 goto no_page_table;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001516 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001517 if (flags & FOLL_SPLIT) {
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001518 split_huge_page_pmd(vma, address, pmd);
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001519 goto split_fallthrough;
1520 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001521 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001522 if (likely(pmd_trans_huge(*pmd))) {
1523 if (unlikely(pmd_trans_splitting(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001524 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001525 wait_split_huge_page(vma->anon_vma, pmd);
1526 } else {
David Rientjesb676b292012-10-08 16:34:03 -07001527 page = follow_trans_huge_pmd(vma, address,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001528 pmd, flags);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001529 spin_unlock(ptl);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001530 *page_mask = HPAGE_PMD_NR - 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001531 goto out;
1532 }
1533 } else
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001534 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001535 /* fall through */
1536 }
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001537split_fallthrough:
Hugh Dickinsaeed5fc2008-05-06 20:49:23 +01001538 if (unlikely(pmd_bad(*pmd)))
1539 goto no_page_table;
1540
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001541 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542
1543 pte = *ptep;
Hugh Dickins5117b3b2013-02-22 16:36:07 -08001544 if (!pte_present(pte)) {
1545 swp_entry_t entry;
1546 /*
1547 * KSM's break_ksm() relies upon recognizing a ksm page
1548 * even while it is being migrated, so for that case we
1549 * need migration_entry_wait().
1550 */
1551 if (likely(!(flags & FOLL_MIGRATION)))
1552 goto no_page;
1553 if (pte_none(pte) || pte_file(pte))
1554 goto no_page;
1555 entry = pte_to_swp_entry(pte);
1556 if (!is_migration_entry(entry))
1557 goto no_page;
1558 pte_unmap_unlock(ptep, ptl);
1559 migration_entry_wait(mm, pmd, address);
1560 goto split_fallthrough;
1561 }
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001562 if ((flags & FOLL_NUMA) && pte_numa(pte))
1563 goto no_page;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001564 if ((flags & FOLL_WRITE) && !pte_write(pte))
1565 goto unlock;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001566
Linus Torvalds6aab3412005-11-28 14:34:23 -08001567 page = vm_normal_page(vma, address, pte);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001568 if (unlikely(!page)) {
1569 if ((flags & FOLL_DUMP) ||
Hugh Dickins62eede62009-09-21 17:03:34 -07001570 !is_zero_pfn(pte_pfn(pte)))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07001571 goto bad_page;
1572 page = pte_page(pte);
1573 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001574
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001575 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001576 get_page_foll(page);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001577 if (flags & FOLL_TOUCH) {
1578 if ((flags & FOLL_WRITE) &&
1579 !pte_dirty(pte) && !PageDirty(page))
1580 set_page_dirty(page);
KOSAKI Motohirobd775c42009-03-31 15:19:37 -07001581 /*
1582 * pte_mkyoung() would be more correct here, but atomic care
1583 * is needed to avoid losing the dirty bit: it is easier to use
1584 * mark_page_accessed().
1585 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001586 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 }
Linus Torvaldsa1fde082011-05-04 21:30:28 -07001588 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001589 /*
1590 * The preliminary mapping check is mainly to avoid the
1591 * pointless overhead of lock_page on the ZERO_PAGE
1592 * which might bounce very badly if there is contention.
1593 *
1594 * If the page is already locked, we don't need to
1595 * handle it now - vmscan will handle it later if and
1596 * when it attempts to reclaim the page.
1597 */
1598 if (page->mapping && trylock_page(page)) {
1599 lru_add_drain(); /* push cached pages to LRU */
1600 /*
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001601 * Because we lock page here, and migration is
1602 * blocked by the pte's page reference, and we
1603 * know the page is still mapped, we don't even
1604 * need to check for file-cache page truncation.
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001605 */
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001606 mlock_vma_page(page);
Michel Lespinasse110d74a2011-01-13 15:46:11 -08001607 unlock_page(page);
1608 }
1609 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001610unlock:
1611 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001613 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001615bad_page:
1616 pte_unmap_unlock(ptep, ptl);
1617 return ERR_PTR(-EFAULT);
1618
1619no_page:
1620 pte_unmap_unlock(ptep, ptl);
1621 if (!pte_none(pte))
1622 return page;
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001623
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001624no_page_table:
1625 /*
1626 * When core dumping an enormous anonymous area that nobody
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001627 * has touched so far, we don't want to allocate unnecessary pages or
1628 * page tables. Return error instead of NULL to skip handle_mm_fault,
1629 * then get_dump_page() will return NULL to leave a hole in the dump.
1630 * But we can only make this optimization where a hole would surely
1631 * be zero-filled if handle_mm_fault() actually did handle it.
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001632 */
Hugh Dickins8e4b9a602009-09-21 17:03:26 -07001633 if ((flags & FOLL_DUMP) &&
1634 (!vma->vm_ops || !vma->vm_ops->fault))
1635 return ERR_PTR(-EFAULT);
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001636 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637}
1638
Linus Torvalds95042f92011-04-12 14:15:51 -07001639static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1640{
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001641 return stack_guard_page_start(vma, addr) ||
1642 stack_guard_page_end(vma, addr+PAGE_SIZE);
Linus Torvalds95042f92011-04-12 14:15:51 -07001643}
1644
Huang Ying0014bd92011-01-30 11:15:47 +08001645/**
1646 * __get_user_pages() - pin user pages in memory
1647 * @tsk: task_struct of target task
1648 * @mm: mm_struct of target mm
1649 * @start: starting user address
1650 * @nr_pages: number of pages from start to pin
1651 * @gup_flags: flags modifying pin behaviour
1652 * @pages: array that receives pointers to the pages pinned.
1653 * Should be at least nr_pages long. Or NULL, if caller
1654 * only intends to ensure the pages are faulted in.
1655 * @vmas: array of pointers to vmas corresponding to each page.
1656 * Or NULL if the caller does not require them.
1657 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1658 *
1659 * Returns number of pages pinned. This may be fewer than the number
1660 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1661 * were pinned, returns -errno. Each page returned must be released
1662 * with a put_page() call when it is finished with. vmas will only
1663 * remain valid while mmap_sem is held.
1664 *
1665 * Must be called with mmap_sem held for read or write.
1666 *
1667 * __get_user_pages walks a process's page tables and takes a reference to
1668 * each struct page that each user address corresponds to at a given
1669 * instant. That is, it takes the page that would be accessed if a user
1670 * thread accesses the given user virtual address at that instant.
1671 *
1672 * This does not guarantee that the page exists in the user mappings when
1673 * __get_user_pages returns, and there may even be a completely different
1674 * page there in some cases (eg. if mmapped pagecache has been invalidated
1675 * and subsequently re faulted). However it does guarantee that the page
1676 * won't be freed completely. And mostly callers simply care that the page
1677 * contains data that was valid *at some point in time*. Typically, an IO
1678 * or similar operation cannot guarantee anything stronger anyway because
1679 * locks can't be held over the syscall boundary.
1680 *
1681 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1682 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1683 * appropriate) must be called after the page is finished with, and
1684 * before put_page is called.
1685 *
1686 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1687 * or mmap_sem contention, and if waiting is needed to pin all pages,
1688 * *@nonblocking will be set to 0.
1689 *
1690 * In most cases, get_user_pages or get_user_pages_fast should be used
1691 * instead of __get_user_pages. __get_user_pages should be used only if
1692 * you need some special @gup_flags.
1693 */
Michel Lespinasse28a35712013-02-22 16:35:55 -08001694long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1695 unsigned long start, unsigned long nr_pages,
1696 unsigned int gup_flags, struct page **pages,
1697 struct vm_area_struct **vmas, int *nonblocking)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698{
Michel Lespinasse28a35712013-02-22 16:35:55 -08001699 long i;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001700 unsigned long vm_flags;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001701 unsigned int page_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702
Michel Lespinasse28a35712013-02-22 16:35:55 -08001703 if (!nr_pages)
Jonathan Corbet900cf082008-02-11 16:17:33 -07001704 return 0;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001705
1706 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1707
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 /*
1709 * Require read or write permissions.
Hugh Dickins58fa8792009-09-21 17:03:31 -07001710 * If FOLL_FORCE is set, we only require the "MAY" flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001712 vm_flags = (gup_flags & FOLL_WRITE) ?
1713 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1714 vm_flags &= (gup_flags & FOLL_FORCE) ?
1715 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Andrea Arcangeli0b9d7052012-10-05 21:36:27 +02001716
1717 /*
1718 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1719 * would be called on PROT_NONE ranges. We must never invoke
1720 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1721 * page faults would unprotect the PROT_NONE ranges if
1722 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1723 * bitflag. So to avoid that, don't set FOLL_NUMA if
1724 * FOLL_FORCE is set.
1725 */
1726 if (!(gup_flags & FOLL_FORCE))
1727 gup_flags |= FOLL_NUMA;
1728
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 i = 0;
1730
1731 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001732 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733
1734 vma = find_extend_vma(mm, start);
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001735 if (!vma && in_gate_area(mm, start)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736 unsigned long pg = start & PAGE_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 pgd_t *pgd;
1738 pud_t *pud;
1739 pmd_t *pmd;
1740 pte_t *pte;
Nick Pigginb291f002008-10-18 20:26:44 -07001741
1742 /* user gate pages are read-only */
Hugh Dickins58fa8792009-09-21 17:03:31 -07001743 if (gup_flags & FOLL_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 return i ? : -EFAULT;
1745 if (pg > TASK_SIZE)
1746 pgd = pgd_offset_k(pg);
1747 else
1748 pgd = pgd_offset_gate(mm, pg);
1749 BUG_ON(pgd_none(*pgd));
1750 pud = pud_offset(pgd, pg);
1751 BUG_ON(pud_none(*pud));
1752 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001753 if (pmd_none(*pmd))
1754 return i ? : -EFAULT;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001755 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -07001757 if (pte_none(*pte)) {
1758 pte_unmap(pte);
1759 return i ? : -EFAULT;
1760 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001761 vma = get_gate_vma(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762 if (pages) {
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001763 struct page *page;
1764
Linus Torvalds95042f92011-04-12 14:15:51 -07001765 page = vm_normal_page(vma, start, *pte);
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001766 if (!page) {
1767 if (!(gup_flags & FOLL_DUMP) &&
1768 is_zero_pfn(pte_pfn(*pte)))
1769 page = pte_page(*pte);
1770 else {
1771 pte_unmap(pte);
1772 return i ? : -EFAULT;
1773 }
1774 }
Linus Torvalds6aab3412005-11-28 14:34:23 -08001775 pages[i] = page;
Hugh Dickinsde51257a2010-07-30 10:58:26 -07001776 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 }
1778 pte_unmap(pte);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001779 page_mask = 0;
Linus Torvalds95042f92011-04-12 14:15:51 -07001780 goto next_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 }
1782
Nick Pigginb291f002008-10-18 20:26:44 -07001783 if (!vma ||
1784 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001785 !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 return i ? : -EFAULT;
1787
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001788 if (is_vm_hugetlb_page(vma)) {
1789 i = follow_hugetlb_page(mm, vma, pages, vmas,
Hugh Dickins58fa8792009-09-21 17:03:31 -07001790 &start, &nr_pages, i, gup_flags);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07001791 continue;
1792 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001793
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001795 struct page *page;
Hugh Dickins58fa8792009-09-21 17:03:31 -07001796 unsigned int foll_flags = gup_flags;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001797 unsigned int page_increm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798
Ethan Solomita462e00c2007-07-15 23:38:16 -07001799 /*
Ying Han47792802009-01-06 14:40:18 -08001800 * If we have a pending SIGKILL, don't keep faulting
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001801 * pages and potentially allocating memory.
Ethan Solomita462e00c2007-07-15 23:38:16 -07001802 */
Hugh Dickins1c3aff12009-09-21 17:03:24 -07001803 if (unlikely(fatal_signal_pending(current)))
Ying Han47792802009-01-06 14:40:18 -08001804 return i ? i : -ERESTARTSYS;
Ethan Solomita462e00c2007-07-15 23:38:16 -07001805
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001806 cond_resched();
Michel Lespinasse240aade2013-02-22 16:35:56 -08001807 while (!(page = follow_page_mask(vma, start,
1808 foll_flags, &page_mask))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001809 int ret;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001810 unsigned int fault_flags = 0;
1811
Mikulas Patockaa09a79f2011-05-09 13:01:09 +02001812 /* For mlock, just skip the stack guard page. */
1813 if (foll_flags & FOLL_MLOCK) {
1814 if (stack_guard_page(vma, start))
1815 goto next_page;
1816 }
Michel Lespinasse53a77062011-01-13 15:46:14 -08001817 if (foll_flags & FOLL_WRITE)
1818 fault_flags |= FAULT_FLAG_WRITE;
1819 if (nonblocking)
1820 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
Gleb Natapov318b2752011-03-22 16:30:51 -07001821 if (foll_flags & FOLL_NOWAIT)
1822 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
Linus Torvaldsd06063c2009-04-10 09:01:23 -07001823
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001824 ret = handle_mm_fault(mm, vma, start,
Michel Lespinasse53a77062011-01-13 15:46:14 -08001825 fault_flags);
Hugh Dickinsd26ed652009-06-23 13:52:49 +01001826
Nick Piggin83c54072007-07-19 01:47:05 -07001827 if (ret & VM_FAULT_ERROR) {
1828 if (ret & VM_FAULT_OOM)
1829 return i ? i : -ENOMEM;
Huang Ying69ebb832011-01-30 11:15:48 +08001830 if (ret & (VM_FAULT_HWPOISON |
1831 VM_FAULT_HWPOISON_LARGE)) {
1832 if (i)
1833 return i;
1834 else if (gup_flags & FOLL_HWPOISON)
1835 return -EHWPOISON;
1836 else
1837 return -EFAULT;
1838 }
1839 if (ret & VM_FAULT_SIGBUS)
Nick Piggin83c54072007-07-19 01:47:05 -07001840 return i ? i : -EFAULT;
1841 BUG();
1842 }
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001843
1844 if (tsk) {
1845 if (ret & VM_FAULT_MAJOR)
1846 tsk->maj_flt++;
1847 else
1848 tsk->min_flt++;
1849 }
Nick Piggin83c54072007-07-19 01:47:05 -07001850
Michel Lespinasse53a77062011-01-13 15:46:14 -08001851 if (ret & VM_FAULT_RETRY) {
Gleb Natapov318b2752011-03-22 16:30:51 -07001852 if (nonblocking)
1853 *nonblocking = 0;
Michel Lespinasse53a77062011-01-13 15:46:14 -08001854 return i;
1855 }
1856
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001857 /*
Nick Piggin83c54072007-07-19 01:47:05 -07001858 * The VM_FAULT_WRITE bit tells us that
1859 * do_wp_page has broken COW when necessary,
1860 * even if maybe_mkwrite decided not to set
1861 * pte_write. We can thus safely do subsequent
Hugh Dickins878b63a2009-01-06 14:39:32 -08001862 * page lookups as if they were reads. But only
1863 * do so when looping for pte_write is futile:
1864 * in some cases userspace may also be wanting
1865 * to write to the gotten user page, which a
1866 * read fault here might prevent (a readonly
1867 * page might get reCOWed by userspace write).
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001868 */
Hugh Dickins878b63a2009-01-06 14:39:32 -08001869 if ((ret & VM_FAULT_WRITE) &&
1870 !(vma->vm_flags & VM_WRITE))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001871 foll_flags &= ~FOLL_WRITE;
Nick Piggin83c54072007-07-19 01:47:05 -07001872
Benjamin Herrenschmidt7f7bbbe2006-10-06 00:43:53 -07001873 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874 }
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001875 if (IS_ERR(page))
1876 return i ? i : PTR_ERR(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001878 pages[i] = page;
James Bottomley03beb072006-03-26 01:36:57 -08001879
Russell Kinga6f36be2006-12-30 22:24:19 +00001880 flush_anon_page(vma, page, start);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001881 flush_dcache_page(page);
Michel Lespinasse240aade2013-02-22 16:35:56 -08001882 page_mask = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 }
Linus Torvalds95042f92011-04-12 14:15:51 -07001884next_page:
Michel Lespinasse240aade2013-02-22 16:35:56 -08001885 if (vmas) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 vmas[i] = vma;
Michel Lespinasse240aade2013-02-22 16:35:56 -08001887 page_mask = 0;
1888 }
1889 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1890 if (page_increm > nr_pages)
1891 page_increm = nr_pages;
1892 i += page_increm;
1893 start += page_increm * PAGE_SIZE;
1894 nr_pages -= page_increm;
Peter Zijlstra9d737772009-06-25 11:58:55 +02001895 } while (nr_pages && start < vma->vm_end);
1896 } while (nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 return i;
1898}
Huang Ying0014bd92011-01-30 11:15:47 +08001899EXPORT_SYMBOL(__get_user_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001900
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -07001901/*
1902 * fixup_user_fault() - manually resolve a user page fault
1903 * @tsk: the task_struct to use for page fault accounting, or
1904 * NULL if faults are not to be recorded.
1905 * @mm: mm_struct of target mm
1906 * @address: user address
1907 * @fault_flags:flags to pass down to handle_mm_fault()
1908 *
1909 * This is meant to be called in the specific scenario where for locking reasons
1910 * we try to access user memory in atomic context (within a pagefault_disable()
1911 * section), this returns -EFAULT, and we want to resolve the user fault before
1912 * trying again.
1913 *
1914 * Typically this is meant to be used by the futex code.
1915 *
1916 * The main difference with get_user_pages() is that this function will
1917 * unconditionally call handle_mm_fault() which will in turn perform all the
1918 * necessary SW fixup of the dirty and young bits in the PTE, while
1919 * handle_mm_fault() only guarantees to update these in the struct page.
1920 *
1921 * This is important for some architectures where those bits also gate the
1922 * access permission to the page because they are maintained in software. On
1923 * such architectures, gup() will not be enough to make a subsequent access
1924 * succeed.
1925 *
1926 * This should be called with the mm_sem held for read.
1927 */
1928int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1929 unsigned long address, unsigned int fault_flags)
1930{
1931 struct vm_area_struct *vma;
1932 int ret;
1933
1934 vma = find_extend_vma(mm, address);
1935 if (!vma || address < vma->vm_start)
1936 return -EFAULT;
1937
1938 ret = handle_mm_fault(mm, vma, address, fault_flags);
1939 if (ret & VM_FAULT_ERROR) {
1940 if (ret & VM_FAULT_OOM)
1941 return -ENOMEM;
1942 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1943 return -EHWPOISON;
1944 if (ret & VM_FAULT_SIGBUS)
1945 return -EFAULT;
1946 BUG();
1947 }
1948 if (tsk) {
1949 if (ret & VM_FAULT_MAJOR)
1950 tsk->maj_flt++;
1951 else
1952 tsk->min_flt++;
1953 }
1954 return 0;
1955}
1956
1957/*
Nick Piggind2bf6be2009-06-16 15:31:39 -07001958 * get_user_pages() - pin user pages in memory
Stephen Wilsone7f22e22011-03-13 15:49:18 -04001959 * @tsk: the task_struct to use for page fault accounting, or
1960 * NULL if faults are not to be recorded.
Nick Piggind2bf6be2009-06-16 15:31:39 -07001961 * @mm: mm_struct of target mm
1962 * @start: starting user address
Peter Zijlstra9d737772009-06-25 11:58:55 +02001963 * @nr_pages: number of pages from start to pin
Nick Piggind2bf6be2009-06-16 15:31:39 -07001964 * @write: whether pages will be written to by the caller
1965 * @force: whether to force write access even if user mapping is
1966 * readonly. This will result in the page being COWed even
1967 * in MAP_SHARED mappings. You do not want this.
1968 * @pages: array that receives pointers to the pages pinned.
1969 * Should be at least nr_pages long. Or NULL, if caller
1970 * only intends to ensure the pages are faulted in.
1971 * @vmas: array of pointers to vmas corresponding to each page.
1972 * Or NULL if the caller does not require them.
1973 *
1974 * Returns number of pages pinned. This may be fewer than the number
Peter Zijlstra9d737772009-06-25 11:58:55 +02001975 * requested. If nr_pages is 0 or negative, returns 0. If no pages
Nick Piggind2bf6be2009-06-16 15:31:39 -07001976 * were pinned, returns -errno. Each page returned must be released
1977 * with a put_page() call when it is finished with. vmas will only
1978 * remain valid while mmap_sem is held.
1979 *
1980 * Must be called with mmap_sem held for read or write.
1981 *
1982 * get_user_pages walks a process's page tables and takes a reference to
1983 * each struct page that each user address corresponds to at a given
1984 * instant. That is, it takes the page that would be accessed if a user
1985 * thread accesses the given user virtual address at that instant.
1986 *
1987 * This does not guarantee that the page exists in the user mappings when
1988 * get_user_pages returns, and there may even be a completely different
1989 * page there in some cases (eg. if mmapped pagecache has been invalidated
1990 * and subsequently re faulted). However it does guarantee that the page
1991 * won't be freed completely. And mostly callers simply care that the page
1992 * contains data that was valid *at some point in time*. Typically, an IO
1993 * or similar operation cannot guarantee anything stronger anyway because
1994 * locks can't be held over the syscall boundary.
1995 *
1996 * If write=0, the page must not be written to. If the page is written to,
1997 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1998 * after the page is finished with, and before put_page is called.
1999 *
2000 * get_user_pages is typically used for fewer-copy IO operations, to get a
2001 * handle on the memory by some means other than accesses via the user virtual
2002 * addresses. The pages may be submitted for DMA to devices or accessed via
2003 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2004 * use the correct cache flushing APIs.
2005 *
2006 * See also get_user_pages_fast, for performance critical applications.
2007 */
Michel Lespinasse28a35712013-02-22 16:35:55 -08002008long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2009 unsigned long start, unsigned long nr_pages, int write,
2010 int force, struct page **pages, struct vm_area_struct **vmas)
Nick Pigginb291f002008-10-18 20:26:44 -07002011{
Hugh Dickins58fa8792009-09-21 17:03:31 -07002012 int flags = FOLL_TOUCH;
Nick Pigginb291f002008-10-18 20:26:44 -07002013
Hugh Dickins58fa8792009-09-21 17:03:31 -07002014 if (pages)
2015 flags |= FOLL_GET;
Nick Pigginb291f002008-10-18 20:26:44 -07002016 if (write)
Hugh Dickins58fa8792009-09-21 17:03:31 -07002017 flags |= FOLL_WRITE;
Nick Pigginb291f002008-10-18 20:26:44 -07002018 if (force)
Hugh Dickins58fa8792009-09-21 17:03:31 -07002019 flags |= FOLL_FORCE;
Nick Pigginb291f002008-10-18 20:26:44 -07002020
Michel Lespinasse53a77062011-01-13 15:46:14 -08002021 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2022 NULL);
Nick Pigginb291f002008-10-18 20:26:44 -07002023}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024EXPORT_SYMBOL(get_user_pages);
2025
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002026/**
2027 * get_dump_page() - pin user page in memory while writing it to core dump
2028 * @addr: user address
2029 *
2030 * Returns struct page pointer of user page pinned for dump,
2031 * to be freed afterwards by page_cache_release() or put_page().
2032 *
2033 * Returns NULL on any kind of failure - a hole must then be inserted into
2034 * the corefile, to preserve alignment with its headers; and also returns
2035 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2036 * allowing a hole to be left in the corefile to save diskspace.
2037 *
2038 * Called without mmap_sem, but after all other threads have been killed.
2039 */
2040#ifdef CONFIG_ELF_CORE
2041struct page *get_dump_page(unsigned long addr)
2042{
2043 struct vm_area_struct *vma;
2044 struct page *page;
2045
2046 if (__get_user_pages(current, current->mm, addr, 1,
Michel Lespinasse53a77062011-01-13 15:46:14 -08002047 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2048 NULL) < 1)
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002049 return NULL;
Hugh Dickinsf3e8fcc2009-09-21 17:03:25 -07002050 flush_cache_page(vma, addr, page_to_pfn(page));
2051 return page;
2052}
2053#endif /* CONFIG_ELF_CORE */
2054
Namhyung Kim25ca1d62010-10-26 14:21:59 -07002055pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08002056 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002057{
2058 pgd_t * pgd = pgd_offset(mm, addr);
2059 pud_t * pud = pud_alloc(mm, pgd, addr);
2060 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05002061 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002062 if (pmd) {
2063 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002064 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002065 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002066 }
2067 return NULL;
2068}
2069
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08002071 * This is the old fallback for page remapping.
2072 *
2073 * For historical reasons, it only allows reserved pages. Only
2074 * old drivers should use this, and they needed to mark their
2075 * pages reserved for the old functions anyway.
2076 */
Nick Piggin423bad62008-04-28 02:13:01 -07002077static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2078 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08002079{
Nick Piggin423bad62008-04-28 02:13:01 -07002080 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002081 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002082 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002083 spinlock_t *ptl;
2084
Linus Torvalds238f58d2005-11-29 13:01:56 -08002085 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002086 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002087 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002088 retval = -ENOMEM;
2089 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08002090 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002091 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07002092 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002093 retval = -EBUSY;
2094 if (!pte_none(*pte))
2095 goto out_unlock;
2096
2097 /* Ok, finally just insert the thing.. */
2098 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002099 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08002100 page_add_file_rmap(page);
2101 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2102
2103 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002104 pte_unmap_unlock(pte, ptl);
2105 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08002106out_unlock:
2107 pte_unmap_unlock(pte, ptl);
2108out:
2109 return retval;
2110}
2111
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002112/**
2113 * vm_insert_page - insert single page into user vma
2114 * @vma: user vma to map to
2115 * @addr: target user address of this page
2116 * @page: source kernel page
2117 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002118 * This allows drivers to insert individual pages they've allocated
2119 * into a user vma.
2120 *
2121 * The page has to be a nice clean _individual_ kernel allocation.
2122 * If you allocate a compound page, you need to have marked it as
2123 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08002124 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002125 *
2126 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2127 * took an arbitrary page protection parameter. This doesn't allow
2128 * that. Your vma protection will have to be set up correctly, which
2129 * means that if you want a shared writable mapping, you'd better
2130 * ask for a shared writable mapping!
2131 *
2132 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002133 *
2134 * Usually this function is called from f_op->mmap() handler
2135 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2136 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2137 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002138 */
Nick Piggin423bad62008-04-28 02:13:01 -07002139int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2140 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002141{
2142 if (addr < vma->vm_start || addr >= vma->vm_end)
2143 return -EFAULT;
2144 if (!page_count(page))
2145 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07002146 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2147 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2148 BUG_ON(vma->vm_flags & VM_PFNMAP);
2149 vma->vm_flags |= VM_MIXEDMAP;
2150 }
Nick Piggin423bad62008-04-28 02:13:01 -07002151 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002152}
Linus Torvaldse3c33742005-12-03 20:48:11 -08002153EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002154
Nick Piggin423bad62008-04-28 02:13:01 -07002155static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2156 unsigned long pfn, pgprot_t prot)
2157{
2158 struct mm_struct *mm = vma->vm_mm;
2159 int retval;
2160 pte_t *pte, entry;
2161 spinlock_t *ptl;
2162
2163 retval = -ENOMEM;
2164 pte = get_locked_pte(mm, addr, &ptl);
2165 if (!pte)
2166 goto out;
2167 retval = -EBUSY;
2168 if (!pte_none(*pte))
2169 goto out_unlock;
2170
2171 /* Ok, finally just insert the thing.. */
2172 entry = pte_mkspecial(pfn_pte(pfn, prot));
2173 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002174 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07002175
2176 retval = 0;
2177out_unlock:
2178 pte_unmap_unlock(pte, ptl);
2179out:
2180 return retval;
2181}
2182
Nick Piggine0dc0d82007-02-12 00:51:36 -08002183/**
2184 * vm_insert_pfn - insert single pfn into user vma
2185 * @vma: user vma to map to
2186 * @addr: target user address of this page
2187 * @pfn: source kernel pfn
2188 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07002189 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08002190 * they've allocated into a user vma. Same comments apply.
2191 *
2192 * This function should only be called from a vm_ops->fault handler, and
2193 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07002194 *
2195 * vma cannot be a COW mapping.
2196 *
2197 * As this is called only for pages that do not currently exist, we
2198 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08002199 */
2200int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07002201 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08002202{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002203 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002204 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07002205 /*
2206 * Technically, architectures with pte_special can avoid all these
2207 * restrictions (same for remap_pfn_range). However we would like
2208 * consistency in testing and feature parity among all, so we should
2209 * try to keep these invariants in place for everybody.
2210 */
Jared Hulbertb379d792008-04-28 02:12:58 -07002211 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2212 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2213 (VM_PFNMAP|VM_MIXEDMAP));
2214 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2215 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08002216
Nick Piggin423bad62008-04-28 02:13:01 -07002217 if (addr < vma->vm_start || addr >= vma->vm_end)
2218 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07002219 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002220 return -EINVAL;
2221
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08002222 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002223
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002224 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08002225}
2226EXPORT_SYMBOL(vm_insert_pfn);
2227
Nick Piggin423bad62008-04-28 02:13:01 -07002228int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2229 unsigned long pfn)
2230{
2231 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2232
2233 if (addr < vma->vm_start || addr >= vma->vm_end)
2234 return -EFAULT;
2235
2236 /*
2237 * If we don't have pte special, then we have to use the pfn_valid()
2238 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2239 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07002240 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2241 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07002242 */
2243 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2244 struct page *page;
2245
2246 page = pfn_to_page(pfn);
2247 return insert_page(vma, addr, page, vma->vm_page_prot);
2248 }
2249 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2250}
2251EXPORT_SYMBOL(vm_insert_mixed);
2252
Linus Torvaldsa145dd42005-11-30 09:35:19 -08002253/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254 * maps a range of physical memory into the requested pages. the old
2255 * mappings are removed. any references to nonexistent pages results
2256 * in null mappings (currently treated as "copy-on-access")
2257 */
2258static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2259 unsigned long addr, unsigned long end,
2260 unsigned long pfn, pgprot_t prot)
2261{
2262 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002263 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264
Hugh Dickinsc74df322005-10-29 18:16:23 -07002265 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 if (!pte)
2267 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002268 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 do {
2270 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07002271 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 pfn++;
2273 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07002274 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07002275 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 return 0;
2277}
2278
2279static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2280 unsigned long addr, unsigned long end,
2281 unsigned long pfn, pgprot_t prot)
2282{
2283 pmd_t *pmd;
2284 unsigned long next;
2285
2286 pfn -= addr >> PAGE_SHIFT;
2287 pmd = pmd_alloc(mm, pud, addr);
2288 if (!pmd)
2289 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08002290 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 do {
2292 next = pmd_addr_end(addr, end);
2293 if (remap_pte_range(mm, pmd, addr, next,
2294 pfn + (addr >> PAGE_SHIFT), prot))
2295 return -ENOMEM;
2296 } while (pmd++, addr = next, addr != end);
2297 return 0;
2298}
2299
2300static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2301 unsigned long addr, unsigned long end,
2302 unsigned long pfn, pgprot_t prot)
2303{
2304 pud_t *pud;
2305 unsigned long next;
2306
2307 pfn -= addr >> PAGE_SHIFT;
2308 pud = pud_alloc(mm, pgd, addr);
2309 if (!pud)
2310 return -ENOMEM;
2311 do {
2312 next = pud_addr_end(addr, end);
2313 if (remap_pmd_range(mm, pud, addr, next,
2314 pfn + (addr >> PAGE_SHIFT), prot))
2315 return -ENOMEM;
2316 } while (pud++, addr = next, addr != end);
2317 return 0;
2318}
2319
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07002320/**
2321 * remap_pfn_range - remap kernel memory to userspace
2322 * @vma: user vma to map to
2323 * @addr: target user address to start at
2324 * @pfn: physical address of kernel memory
2325 * @size: size of map area
2326 * @prot: page protection flags for this mapping
2327 *
2328 * Note: this is only safe if the mm semaphore is held when called.
2329 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2331 unsigned long pfn, unsigned long size, pgprot_t prot)
2332{
2333 pgd_t *pgd;
2334 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07002335 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 struct mm_struct *mm = vma->vm_mm;
2337 int err;
2338
2339 /*
2340 * Physically remapped pages are special. Tell the
2341 * rest of the world about it:
2342 * VM_IO tells people not to look at these pages
2343 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08002344 * VM_PFNMAP tells the core MM that the base pages are just
2345 * raw PFN mappings, and do not have a "struct page" associated
2346 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002347 * VM_DONTEXPAND
2348 * Disable vma merging and expanding with mremap().
2349 * VM_DONTDUMP
2350 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002351 *
2352 * There's a horrible special case to handle copy-on-write
2353 * behaviour that some programs depend on. We mark the "original"
2354 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002355 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002357 if (is_cow_mapping(vma->vm_flags)) {
2358 if (addr != vma->vm_start || end != vma->vm_end)
2359 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002360 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002361 }
2362
2363 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2364 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08002365 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08002366
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07002367 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368
2369 BUG_ON(addr >= end);
2370 pfn -= addr >> PAGE_SHIFT;
2371 pgd = pgd_offset(mm, addr);
2372 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 do {
2374 next = pgd_addr_end(addr, end);
2375 err = remap_pud_range(mm, pgd, addr, next,
2376 pfn + (addr >> PAGE_SHIFT), prot);
2377 if (err)
2378 break;
2379 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002380
2381 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07002382 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08002383
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 return err;
2385}
2386EXPORT_SYMBOL(remap_pfn_range);
2387
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07002388/**
2389 * vm_iomap_memory - remap memory to userspace
2390 * @vma: user vma to map to
2391 * @start: start of area
2392 * @len: size of area
2393 *
2394 * This is a simplified io_remap_pfn_range() for common driver use. The
2395 * driver just needs to give us the physical memory range to be mapped,
2396 * we'll figure out the rest from the vma information.
2397 *
2398 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2399 * whatever write-combining details or similar.
2400 */
2401int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2402{
2403 unsigned long vm_len, pfn, pages;
2404
2405 /* Check that the physical memory area passed in looks valid */
2406 if (start + len < start)
2407 return -EINVAL;
2408 /*
2409 * You *really* shouldn't map things that aren't page-aligned,
2410 * but we've historically allowed it because IO memory might
2411 * just have smaller alignment.
2412 */
2413 len += start & ~PAGE_MASK;
2414 pfn = start >> PAGE_SHIFT;
2415 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2416 if (pfn + pages < pfn)
2417 return -EINVAL;
2418
2419 /* We start the mapping 'vm_pgoff' pages into the area */
2420 if (vma->vm_pgoff > pages)
2421 return -EINVAL;
2422 pfn += vma->vm_pgoff;
2423 pages -= vma->vm_pgoff;
2424
2425 /* Can we fit all of the mapping? */
2426 vm_len = vma->vm_end - vma->vm_start;
2427 if (vm_len >> PAGE_SHIFT > pages)
2428 return -EINVAL;
2429
2430 /* Ok, let it rip */
2431 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2432}
2433EXPORT_SYMBOL(vm_iomap_memory);
2434
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002435static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2436 unsigned long addr, unsigned long end,
2437 pte_fn_t fn, void *data)
2438{
2439 pte_t *pte;
2440 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002441 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07002442 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002443
2444 pte = (mm == &init_mm) ?
2445 pte_alloc_kernel(pmd, addr) :
2446 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2447 if (!pte)
2448 return -ENOMEM;
2449
2450 BUG_ON(pmd_huge(*pmd));
2451
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002452 arch_enter_lazy_mmu_mode();
2453
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08002454 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002455
2456 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002457 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002458 if (err)
2459 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07002460 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002461
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08002462 arch_leave_lazy_mmu_mode();
2463
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002464 if (mm != &init_mm)
2465 pte_unmap_unlock(pte-1, ptl);
2466 return err;
2467}
2468
2469static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2470 unsigned long addr, unsigned long end,
2471 pte_fn_t fn, void *data)
2472{
2473 pmd_t *pmd;
2474 unsigned long next;
2475 int err;
2476
Andi Kleenceb86872008-07-23 21:27:50 -07002477 BUG_ON(pud_huge(*pud));
2478
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002479 pmd = pmd_alloc(mm, pud, addr);
2480 if (!pmd)
2481 return -ENOMEM;
2482 do {
2483 next = pmd_addr_end(addr, end);
2484 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2485 if (err)
2486 break;
2487 } while (pmd++, addr = next, addr != end);
2488 return err;
2489}
2490
2491static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2492 unsigned long addr, unsigned long end,
2493 pte_fn_t fn, void *data)
2494{
2495 pud_t *pud;
2496 unsigned long next;
2497 int err;
2498
2499 pud = pud_alloc(mm, pgd, addr);
2500 if (!pud)
2501 return -ENOMEM;
2502 do {
2503 next = pud_addr_end(addr, end);
2504 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2505 if (err)
2506 break;
2507 } while (pud++, addr = next, addr != end);
2508 return err;
2509}
2510
2511/*
2512 * Scan a region of virtual memory, filling in page tables as necessary
2513 * and calling a provided function on each leaf page table.
2514 */
2515int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2516 unsigned long size, pte_fn_t fn, void *data)
2517{
2518 pgd_t *pgd;
2519 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002520 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002521 int err;
2522
2523 BUG_ON(addr >= end);
2524 pgd = pgd_offset(mm, addr);
2525 do {
2526 next = pgd_addr_end(addr, end);
2527 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2528 if (err)
2529 break;
2530 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07002531
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07002532 return err;
2533}
2534EXPORT_SYMBOL_GPL(apply_to_page_range);
2535
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002537 * handle_pte_fault chooses page fault handler according to an entry
2538 * which was read non-atomically. Before making any commitment, on
2539 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002540 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002541 * must check under lock before unmapping the pte and proceeding
2542 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002543 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002544 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002545static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002546 pte_t *page_table, pte_t orig_pte)
2547{
2548 int same = 1;
2549#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2550 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002551 spinlock_t *ptl = pte_lockptr(mm, pmd);
2552 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002553 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002554 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002555 }
2556#endif
2557 pte_unmap(page_table);
2558 return same;
2559}
2560
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002561static 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 -08002562{
Dan Williams0abdd7a2014-01-21 15:48:12 -08002563 debug_dma_assert_idle(src);
2564
Linus Torvalds6aab3412005-11-28 14:34:23 -08002565 /*
2566 * If the source page was a PFN mapping, we don't have
2567 * a "struct page" for it. We do a best-effort copy by
2568 * just copying from the original user address. If that
2569 * fails, we just zero-fill it. Live with it.
2570 */
2571 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002572 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002573 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2574
2575 /*
2576 * This really shouldn't fail, because the page is there
2577 * in the page tables. But it might just be unreadable,
2578 * in which case we just give up and fill the result with
2579 * zeroes.
2580 */
2581 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002582 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002583 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002584 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002585 } else
2586 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002587}
2588
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 * This routine handles present pages, when users try to write
2591 * to a shared page. It is done by copying the page to a new address
2592 * and decrementing the shared-page counter for the old page.
2593 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594 * Note that this routine assumes that the protection checks have been
2595 * done by the caller (the low-level page fault routine in most cases).
2596 * Thus we can safely just mark it writable once we've done any necessary
2597 * COW.
2598 *
2599 * We also mark the page dirty at this point even though the page will
2600 * change only once the write actually happens. This avoids a few races,
2601 * and potentially makes it more efficient.
2602 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002603 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2604 * but allow concurrent faults), with pte both mapped and locked.
2605 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002607static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2608 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002609 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002610 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611{
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002612 struct page *old_page, *new_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002614 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002615 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002616 struct page *dirty_page = NULL;
David Rientjes17569542012-11-16 14:14:48 -08002617 unsigned long mmun_start = 0; /* For mmu_notifiers */
2618 unsigned long mmun_end = 0; /* For mmu_notifiers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619
Linus Torvalds6aab3412005-11-28 14:34:23 -08002620 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002621 if (!old_page) {
2622 /*
2623 * VM_MIXEDMAP !pfn_valid() case
2624 *
2625 * We should not cow pages in a shared writeable mapping.
2626 * Just mark the pages writable as we can't do any dirty
2627 * accounting on raw pfn maps.
2628 */
2629 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2630 (VM_WRITE|VM_SHARED))
2631 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002632 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002633 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002635 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002636 * Take out anonymous pages first, anonymous shared vmas are
2637 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002638 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002639 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002640 if (!trylock_page(old_page)) {
2641 page_cache_get(old_page);
2642 pte_unmap_unlock(page_table, ptl);
2643 lock_page(old_page);
2644 page_table = pte_offset_map_lock(mm, pmd, address,
2645 &ptl);
2646 if (!pte_same(*page_table, orig_pte)) {
2647 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002648 goto unlock;
2649 }
2650 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002651 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002652 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002653 /*
2654 * The page is all ours. Move it to our anon_vma so
2655 * the rmap code will not search our parent or siblings.
2656 * Protected against the rmap code by the page lock.
2657 */
2658 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002659 unlock_page(old_page);
2660 goto reuse;
2661 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002662 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002663 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002664 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002665 /*
2666 * Only catch write-faults on shared writable pages,
2667 * read-only shared pages can get COWed by
2668 * get_user_pages(.write=1, .force=1).
2669 */
David Howells9637a5e2006-06-23 02:03:43 -07002670 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002671 struct vm_fault vmf;
2672 int tmp;
2673
2674 vmf.virtual_address = (void __user *)(address &
2675 PAGE_MASK);
2676 vmf.pgoff = old_page->index;
2677 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2678 vmf.page = old_page;
2679
David Howells9637a5e2006-06-23 02:03:43 -07002680 /*
2681 * Notify the address space that the page is about to
2682 * become writable so that it can prohibit this or wait
2683 * for the page to get into an appropriate state.
2684 *
2685 * We do this without the lock held, so that it can
2686 * sleep if it needs to.
2687 */
2688 page_cache_get(old_page);
2689 pte_unmap_unlock(page_table, ptl);
2690
Nick Pigginc2ec1752009-03-31 15:23:21 -07002691 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2692 if (unlikely(tmp &
2693 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2694 ret = tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002695 goto unwritable_page;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002696 }
Nick Pigginb827e492009-04-30 15:08:16 -07002697 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2698 lock_page(old_page);
2699 if (!old_page->mapping) {
2700 ret = 0; /* retry the fault */
2701 unlock_page(old_page);
2702 goto unwritable_page;
2703 }
2704 } else
Sasha Levin309381fea2014-01-23 15:52:54 -08002705 VM_BUG_ON_PAGE(!PageLocked(old_page), old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002706
David Howells9637a5e2006-06-23 02:03:43 -07002707 /*
2708 * Since we dropped the lock we need to revalidate
2709 * the PTE as someone else may have changed it. If
2710 * they did, we just return, as we can count on the
2711 * MMU to tell us if they didn't also make it writable.
2712 */
2713 page_table = pte_offset_map_lock(mm, pmd, address,
2714 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002715 if (!pte_same(*page_table, orig_pte)) {
2716 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002717 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002718 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002719
2720 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002722 dirty_page = old_page;
2723 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002724
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002725reuse:
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002726 /*
2727 * Clear the pages cpupid information as the existing
2728 * information potentially belongs to a now completely
2729 * unrelated process.
2730 */
2731 if (old_page)
2732 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2733
David Howells9637a5e2006-06-23 02:03:43 -07002734 flush_cache_page(vma, address, pte_pfn(orig_pte));
2735 entry = pte_mkyoung(orig_pte);
2736 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002737 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002738 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002739 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002740 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002741
2742 if (!dirty_page)
2743 return ret;
2744
2745 /*
2746 * Yes, Virginia, this is actually required to prevent a race
2747 * with clear_page_dirty_for_io() from clearing the page dirty
2748 * bit after it clear all dirty ptes, but before a racing
2749 * do_wp_page installs a dirty pte.
2750 *
Ryota Ozakia335b2e2011-02-10 13:56:28 +09002751 * __do_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002752 */
2753 if (!page_mkwrite) {
2754 wait_on_page_locked(dirty_page);
2755 set_page_dirty_balance(dirty_page, page_mkwrite);
Jan Kara41c4d252012-06-12 16:20:28 +02002756 /* file_update_time outside page_lock */
2757 if (vma->vm_file)
2758 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002759 }
2760 put_page(dirty_page);
2761 if (page_mkwrite) {
2762 struct address_space *mapping = dirty_page->mapping;
2763
2764 set_page_dirty(dirty_page);
2765 unlock_page(dirty_page);
2766 page_cache_release(dirty_page);
2767 if (mapping) {
2768 /*
2769 * Some device drivers do not set page.mapping
2770 * but still dirty their pages
2771 */
2772 balance_dirty_pages_ratelimited(mapping);
2773 }
2774 }
2775
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002776 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778
2779 /*
2780 * Ok, we need to copy. Oh, well..
2781 */
Nick Pigginb5810032005-10-29 18:16:12 -07002782 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002783gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002784 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785
2786 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002787 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002788
Hugh Dickins62eede62009-09-21 17:03:34 -07002789 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002790 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2791 if (!new_page)
2792 goto oom;
2793 } else {
2794 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2795 if (!new_page)
2796 goto oom;
2797 cow_user_page(new_page, old_page, address, vma);
2798 }
2799 __SetPageUptodate(new_page);
2800
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08002801 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002802 goto oom_free_new;
2803
Haggai Eran6bdb9132012-10-08 16:33:35 -07002804 mmun_start = address & PAGE_MASK;
David Rientjes17569542012-11-16 14:14:48 -08002805 mmun_end = mmun_start + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -07002806 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2807
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 /*
2809 * Re-check the pte - we dropped the lock
2810 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002811 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002812 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002813 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002814 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002815 dec_mm_counter_fast(mm, MM_FILEPAGES);
2816 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002817 }
2818 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002819 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002820 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002821 entry = mk_pte(new_page, vma->vm_page_prot);
2822 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002823 /*
2824 * Clear the pte entry and flush it first, before updating the
2825 * pte with the new entry. This will avoid a race condition
2826 * seen in the presence of one thread doing SMC and another
2827 * thread doing COW.
2828 */
Izik Eidus828502d2009-09-21 17:01:51 -07002829 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002830 page_add_new_anon_rmap(new_page, vma, address);
Izik Eidus828502d2009-09-21 17:01:51 -07002831 /*
2832 * We call the notify macro here because, when using secondary
2833 * mmu page tables (such as kvm shadow page tables), we want the
2834 * new page to be mapped directly into the secondary page table.
2835 */
2836 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002837 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002838 if (old_page) {
2839 /*
2840 * Only after switching the pte to the new page may
2841 * we remove the mapcount here. Otherwise another
2842 * process may come and find the rmap count decremented
2843 * before the pte is switched to the new page, and
2844 * "reuse" the old page writing into it while our pte
2845 * here still points into it and can be read by other
2846 * threads.
2847 *
2848 * The critical issue is to order this
2849 * page_remove_rmap with the ptp_clear_flush above.
2850 * Those stores are ordered by (if nothing else,)
2851 * the barrier present in the atomic_add_negative
2852 * in page_remove_rmap.
2853 *
2854 * Then the TLB flush in ptep_clear_flush ensures that
2855 * no process can access the old page before the
2856 * decremented mapcount is visible. And the old page
2857 * cannot be reused until after the decremented
2858 * mapcount is visible. So transitively, TLBs to
2859 * old page will be flushed before it can be reused.
2860 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002861 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002862 }
2863
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 /* Free the old page.. */
2865 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002866 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002867 } else
2868 mem_cgroup_uncharge_page(new_page);
2869
Haggai Eran6bdb9132012-10-08 16:33:35 -07002870 if (new_page)
2871 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002872unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002873 pte_unmap_unlock(page_table, ptl);
David Rientjes17569542012-11-16 14:14:48 -08002874 if (mmun_end > mmun_start)
Haggai Eran6bdb9132012-10-08 16:33:35 -07002875 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002876 if (old_page) {
2877 /*
2878 * Don't let another task, with possibly unlocked vma,
2879 * keep the mlocked page.
2880 */
2881 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2882 lock_page(old_page); /* LRU manipulation */
2883 munlock_vma_page(old_page);
2884 unlock_page(old_page);
2885 }
2886 page_cache_release(old_page);
2887 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002888 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002889oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002890 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002891oom:
Dominik Dingel66521d52012-12-12 13:52:37 -08002892 if (old_page)
Hugh Dickins920fc352005-11-21 21:32:17 -08002893 page_cache_release(old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894 return VM_FAULT_OOM;
David Howells9637a5e2006-06-23 02:03:43 -07002895
2896unwritable_page:
2897 page_cache_release(old_page);
Nick Pigginc2ec1752009-03-31 15:23:21 -07002898 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899}
2900
Peter Zijlstra97a89412011-05-24 17:12:04 -07002901static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902 unsigned long start_addr, unsigned long end_addr,
2903 struct zap_details *details)
2904{
Al Virof5cc4ee2012-03-05 14:14:20 -05002905 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906}
2907
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002908static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 struct zap_details *details)
2910{
2911 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912 pgoff_t vba, vea, zba, zea;
2913
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002914 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916
2917 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002918 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2920 zba = details->first_index;
2921 if (zba < vba)
2922 zba = vba;
2923 zea = details->last_index;
2924 if (zea > vea)
2925 zea = vea;
2926
Peter Zijlstra97a89412011-05-24 17:12:04 -07002927 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2929 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002930 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 }
2932}
2933
2934static inline void unmap_mapping_range_list(struct list_head *head,
2935 struct zap_details *details)
2936{
2937 struct vm_area_struct *vma;
2938
2939 /*
2940 * In nonlinear VMAs there is no correspondence between virtual address
2941 * offset and file offset. So we must perform an exhaustive search
2942 * across *all* the pages in each nonlinear VMA, not just the pages
2943 * whose virtual address lies outside the file truncation point.
2944 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002945 list_for_each_entry(vma, head, shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002947 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 }
2949}
2950
2951/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002952 * 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 -07002953 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954 * @holebegin: byte in first page to unmap, relative to the start of
2955 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002956 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 * must keep the partial page. In contrast, we must get rid of
2958 * partial pages.
2959 * @holelen: size of prospective hole in bytes. This will be rounded
2960 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2961 * end of the file.
2962 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2963 * but 0 when invalidating pagecache, don't throw away private data.
2964 */
2965void unmap_mapping_range(struct address_space *mapping,
2966 loff_t const holebegin, loff_t const holelen, int even_cows)
2967{
2968 struct zap_details details;
2969 pgoff_t hba = holebegin >> PAGE_SHIFT;
2970 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2971
2972 /* Check for overflow. */
2973 if (sizeof(holelen) > sizeof(hlen)) {
2974 long long holeend =
2975 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2976 if (holeend & ~(long long)ULONG_MAX)
2977 hlen = ULONG_MAX - hba + 1;
2978 }
2979
2980 details.check_mapping = even_cows? NULL: mapping;
2981 details.nonlinear_vma = NULL;
2982 details.first_index = hba;
2983 details.last_index = hba + hlen - 1;
2984 if (details.last_index < details.first_index)
2985 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986
Peter Zijlstra97a89412011-05-24 17:12:04 -07002987
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002988 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002989 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2991 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2992 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002993 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994}
2995EXPORT_SYMBOL(unmap_mapping_range);
2996
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002998 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2999 * but allow concurrent faults), and pte mapped but not yet locked.
3000 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003002static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3003 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003004 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003005{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003006 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08003007 struct page *page, *swapcache;
Hugh Dickins65500d22005-10-29 18:15:59 -07003008 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07003010 int locked;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -07003011 struct mem_cgroup *ptr;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003012 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07003013 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003015 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003016 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07003017
3018 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02003019 if (unlikely(non_swap_entry(entry))) {
3020 if (is_migration_entry(entry)) {
3021 migration_entry_wait(mm, pmd, address);
3022 } else if (is_hwpoison_entry(entry)) {
3023 ret = VM_FAULT_HWPOISON;
3024 } else {
3025 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003026 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02003027 }
Christoph Lameter06972122006-06-23 02:03:35 -07003028 goto out;
3029 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07003030 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 page = lookup_swap_cache(entry);
3032 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08003033 page = swapin_readahead(entry,
3034 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 if (!page) {
3036 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003037 * Back out if somebody else faulted in this pte
3038 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003040 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 if (likely(pte_same(*page_table, orig_pte)))
3042 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07003043 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07003044 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 }
3046
3047 /* Had to read the page from swap area: Major fault */
3048 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07003049 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07003050 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02003051 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01003052 /*
3053 * hwpoisoned dirty swapcache pages are kept for killing
3054 * owner processes (which may be unknown at hwpoison time)
3055 */
Andi Kleend1737fd2009-09-16 11:50:06 +02003056 ret = VM_FAULT_HWPOISON;
3057 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08003058 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02003059 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 }
3061
Hugh Dickins56f31802013-02-22 16:36:10 -08003062 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07003063 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07003064
Balbir Singh20a10222007-11-14 17:00:33 -08003065 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003066 if (!locked) {
3067 ret |= VM_FAULT_RETRY;
3068 goto out_release;
3069 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003071 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07003072 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3073 * release the swapcache from under us. The page pin, and pte_same
3074 * test below, are not enough to exclude that. Even if it is still
3075 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003076 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07003077 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003078 goto out_page;
3079
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08003080 page = ksm_might_need_to_copy(page, vma, address);
3081 if (unlikely(!page)) {
3082 ret = VM_FAULT_OOM;
3083 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08003084 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08003085 }
3086
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003087 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003088 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07003089 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07003090 }
3091
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003093 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003094 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003095 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07003096 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07003097 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003098
3099 if (unlikely(!PageUptodate(page))) {
3100 ret = VM_FAULT_SIGBUS;
3101 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003102 }
3103
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003104 /*
3105 * The page isn't present yet, go ahead with the fault.
3106 *
3107 * Be careful about the sequence of operations here.
3108 * To get its accounting right, reuse_swap_page() must be called
3109 * while the page is counted on swap but not yet in mapcount i.e.
3110 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3111 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003112 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3113 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3114 * in page->private. In this case, a record in swap_cgroup is silently
3115 * discarded at swap_free().
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08003116 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003118 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08003119 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003121 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003123 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07003124 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07003125 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07003128 if (pte_swp_soft_dirty(orig_pte))
3129 pte = pte_mksoft_dirty(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003130 set_pte_at(mm, address, page_table, pte);
Hugh Dickins56f31802013-02-22 16:36:10 -08003131 if (page == swapcache)
Johannes Weineraf347702013-02-22 16:32:20 -08003132 do_page_add_anon_rmap(page, vma, address, exclusive);
Hugh Dickins56f31802013-02-22 16:36:10 -08003133 else /* ksm created a completely new copy */
3134 page_add_new_anon_rmap(page, vma, address);
KAMEZAWA Hiroyuki03f3c432009-01-07 18:08:31 -08003135 /* It's better to call commit-charge after rmap is established */
3136 mem_cgroup_commit_charge_swapin(page, ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003138 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07003139 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08003140 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003141 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08003142 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003143 /*
3144 * Hold the lock to avoid the swap entry to be reused
3145 * until we take the PT lock for the pte_same() check
3146 * (to avoid false positives from pte_same). For
3147 * further safety release the lock after the swap_free
3148 * so that the swap count won't change under a
3149 * parallel locked swapcache.
3150 */
3151 unlock_page(swapcache);
3152 page_cache_release(swapcache);
3153 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07003154
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003155 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08003156 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3157 if (ret & VM_FAULT_ERROR)
3158 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003159 goto out;
3160 }
3161
3162 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003163 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003164unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003165 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166out:
3167 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07003168out_nomap:
KAMEZAWA Hiroyuki7a81b882009-01-07 18:07:48 -08003169 mem_cgroup_cancel_charge_swapin(ptr);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003170 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07003171out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003172 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02003173out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07003174 page_cache_release(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08003175 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07003176 unlock_page(swapcache);
3177 page_cache_release(swapcache);
3178 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003179 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180}
3181
3182/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003183 * This is like a special single-page "expand_{down|up}wards()",
3184 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07003185 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07003186 */
3187static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3188{
3189 address &= PAGE_MASK;
3190 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003191 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07003192
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07003193 /*
3194 * Is there a mapping abutting this one below?
3195 *
3196 * That's only ok if it's the same stack mapping
3197 * that has gotten split..
3198 */
3199 if (prev && prev->vm_end == address)
3200 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3201
Michal Hockod05f3162011-05-24 17:11:44 -07003202 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003203 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07003204 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3205 struct vm_area_struct *next = vma->vm_next;
3206
3207 /* As VM_GROWSDOWN but s/below/above/ */
3208 if (next && next->vm_start == address + PAGE_SIZE)
3209 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3210
3211 expand_upwards(vma, address + PAGE_SIZE);
3212 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07003213 return 0;
3214}
3215
3216/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003217 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3218 * but allow concurrent faults), and pte mapped but not yet locked.
3219 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220 */
Hugh Dickins65500d22005-10-29 18:15:59 -07003221static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3222 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003223 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003224{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003225 struct page *page;
3226 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003228
Linus Torvalds11ac5522010-08-14 11:44:56 -07003229 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07003230
Linus Torvalds11ac5522010-08-14 11:44:56 -07003231 /* Check if we need to add a guard page to the stack */
3232 if (check_stack_guard_page(vma, address) < 0)
3233 return VM_FAULT_SIGBUS;
3234
3235 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07003236 if (!(flags & FAULT_FLAG_WRITE)) {
3237 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3238 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07003239 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003240 if (!pte_none(*page_table))
3241 goto unlock;
3242 goto setpte;
3243 }
3244
Nick Piggin557ed1f2007-10-16 01:24:40 -07003245 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07003246 if (unlikely(anon_vma_prepare(vma)))
3247 goto oom;
3248 page = alloc_zeroed_user_highpage_movable(vma, address);
3249 if (!page)
3250 goto oom;
Minchan Kim52f37622013-04-29 15:08:15 -07003251 /*
3252 * The memory barrier inside __SetPageUptodate makes sure that
3253 * preceeding stores to the page contents become visible before
3254 * the set_pte_at() write.
3255 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08003256 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257
KAMEZAWA Hiroyuki2c26fdd2009-01-07 18:08:10 -08003258 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003259 goto oom_free_page;
3260
Nick Piggin557ed1f2007-10-16 01:24:40 -07003261 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07003262 if (vma->vm_flags & VM_WRITE)
3263 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003264
Nick Piggin557ed1f2007-10-16 01:24:40 -07003265 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003266 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07003267 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07003268
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003269 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07003270 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07003271setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07003272 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273
3274 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00003275 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07003276unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003277 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003278 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003279release:
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003280 mem_cgroup_uncharge_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003281 page_cache_release(page);
3282 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08003283oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08003284 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07003285oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286 return VM_FAULT_OOM;
3287}
3288
3289/*
Nick Piggin54cb8822007-07-19 01:46:59 -07003290 * __do_fault() tries to create a new page mapping. It aggressively
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291 * tries to share with existing pages, but makes a separate copy if
Nick Piggin54cb8822007-07-19 01:46:59 -07003292 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3293 * the next page fault.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 *
3295 * As this is called only for pages that do not currently exist, we
3296 * do not need to flush old virtual caches or the TLB.
3297 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003298 * We enter with non-exclusive mmap_sem (to exclude vma changes,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003299 * but allow concurrent faults), and pte neither mapped nor locked.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003300 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003302static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003303 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003304 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305{
Hugh Dickins16abfa02007-10-04 16:56:06 +01003306 pte_t *page_table;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003307 spinlock_t *ptl;
Nick Piggind0217ac2007-07-19 01:47:03 -07003308 struct page *page;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003309 struct page *cow_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311 int anon = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003312 struct page *dirty_page = NULL;
Nick Piggind0217ac2007-07-19 01:47:03 -07003313 struct vm_fault vmf;
3314 int ret;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003315 int page_mkwrite = 0;
Nick Piggin54cb8822007-07-19 01:46:59 -07003316
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003317 /*
3318 * If we do COW later, allocate page befor taking lock_page()
3319 * on the file cache page. This will reduce lock holding time.
3320 */
3321 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3322
3323 if (unlikely(anon_vma_prepare(vma)))
3324 return VM_FAULT_OOM;
3325
3326 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3327 if (!cow_page)
3328 return VM_FAULT_OOM;
3329
3330 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3331 page_cache_release(cow_page);
3332 return VM_FAULT_OOM;
3333 }
3334 } else
3335 cow_page = NULL;
3336
Nick Piggind0217ac2007-07-19 01:47:03 -07003337 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3338 vmf.pgoff = pgoff;
3339 vmf.flags = flags;
3340 vmf.page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003341
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003342 ret = vma->vm_ops->fault(vma, &vmf);
Michel Lespinassed065bd82010-10-26 14:21:57 -07003343 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3344 VM_FAULT_RETRY)))
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003345 goto uncharge_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346
Andi Kleena3b947e2009-09-16 11:50:08 +02003347 if (unlikely(PageHWPoison(vmf.page))) {
3348 if (ret & VM_FAULT_LOCKED)
3349 unlock_page(vmf.page);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003350 ret = VM_FAULT_HWPOISON;
3351 goto uncharge_out;
Andi Kleena3b947e2009-09-16 11:50:08 +02003352 }
3353
Nick Piggind00806b2007-07-19 01:46:57 -07003354 /*
Nick Piggind0217ac2007-07-19 01:47:03 -07003355 * For consistency in subsequent calls, make the faulted page always
Nick Piggind00806b2007-07-19 01:46:57 -07003356 * locked.
3357 */
Nick Piggin83c54072007-07-19 01:47:05 -07003358 if (unlikely(!(ret & VM_FAULT_LOCKED)))
Nick Piggind0217ac2007-07-19 01:47:03 -07003359 lock_page(vmf.page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003360 else
Sasha Levin309381fea2014-01-23 15:52:54 -08003361 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
Nick Piggind00806b2007-07-19 01:46:57 -07003362
Linus Torvalds1da177e2005-04-16 15:20:36 -07003363 /*
3364 * Should we do an early C-O-W break?
3365 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003366 page = vmf.page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003367 if (flags & FAULT_FLAG_WRITE) {
David Howells9637a5e2006-06-23 02:03:43 -07003368 if (!(vma->vm_flags & VM_SHARED)) {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003369 page = cow_page;
Nick Piggin54cb8822007-07-19 01:46:59 -07003370 anon = 1;
Nick Piggind0217ac2007-07-19 01:47:03 -07003371 copy_user_highpage(page, vmf.page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08003372 __SetPageUptodate(page);
David Howells9637a5e2006-06-23 02:03:43 -07003373 } else {
Nick Piggin54cb8822007-07-19 01:46:59 -07003374 /*
3375 * If the page will be shareable, see if the backing
David Howells9637a5e2006-06-23 02:03:43 -07003376 * address space wants to know that the page is about
Nick Piggin54cb8822007-07-19 01:46:59 -07003377 * to become writable
3378 */
Mark Fasheh69676142007-07-19 01:47:00 -07003379 if (vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07003380 int tmp;
3381
Mark Fasheh69676142007-07-19 01:47:00 -07003382 unlock_page(page);
Nick Pigginb827e492009-04-30 15:08:16 -07003383 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Nick Pigginc2ec1752009-03-31 15:23:21 -07003384 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3385 if (unlikely(tmp &
3386 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3387 ret = tmp;
Nick Pigginb827e492009-04-30 15:08:16 -07003388 goto unwritable_page;
Mark Fasheh69676142007-07-19 01:47:00 -07003389 }
Nick Pigginb827e492009-04-30 15:08:16 -07003390 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3391 lock_page(page);
3392 if (!page->mapping) {
3393 ret = 0; /* retry the fault */
3394 unlock_page(page);
3395 goto unwritable_page;
3396 }
3397 } else
Sasha Levin309381fea2014-01-23 15:52:54 -08003398 VM_BUG_ON_PAGE(!PageLocked(page), page);
Peter Zijlstraa200ee12007-10-08 18:54:37 +02003399 page_mkwrite = 1;
David Howells9637a5e2006-06-23 02:03:43 -07003400 }
3401 }
Nick Piggin54cb8822007-07-19 01:46:59 -07003402
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403 }
3404
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003405 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406
3407 /*
3408 * This silly early PAGE_DIRTY setting removes a race
3409 * due to the bad i386 page protection. But it's valid
3410 * for other architectures too.
3411 *
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003412 * Note that if FAULT_FLAG_WRITE is set, we either now have
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413 * an exclusive copy of the page, or this is a shared mapping,
3414 * so we can make it writable and dirty to avoid having to
3415 * handle that later.
3416 */
3417 /* Only go through if we didn't race with anybody else... */
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07003418 if (likely(pte_same(*page_table, orig_pte))) {
Nick Piggind00806b2007-07-19 01:46:57 -07003419 flush_icache_page(vma, page);
3420 entry = mk_pte(page, vma->vm_page_prot);
Nick Piggin54cb8822007-07-19 01:46:59 -07003421 if (flags & FAULT_FLAG_WRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Cyrill Gorcunov41bb3472013-08-13 16:00:51 -07003423 else if (pte_file(orig_pte) && pte_file_soft_dirty(orig_pte))
3424 pte_mksoft_dirty(entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425 if (anon) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003426 inc_mm_counter_fast(mm, MM_ANONPAGES);
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003427 page_add_new_anon_rmap(page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08003428 } else {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08003429 inc_mm_counter_fast(mm, MM_FILEPAGES);
Nick Piggind00806b2007-07-19 01:46:57 -07003430 page_add_file_rmap(page);
Nick Piggin54cb8822007-07-19 01:46:59 -07003431 if (flags & FAULT_FLAG_WRITE) {
Nick Piggind00806b2007-07-19 01:46:57 -07003432 dirty_page = page;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003433 get_page(dirty_page);
3434 }
Hugh Dickins42946212005-10-29 18:16:05 -07003435 }
Lee Schermerhorn64d65192008-10-18 20:26:52 -07003436 set_pte_at(mm, address, page_table, entry);
Nick Piggind00806b2007-07-19 01:46:57 -07003437
3438 /* no need to invalidate: a not-present page won't be cached */
Russell King4b3073e2009-12-18 16:40:18 +00003439 update_mmu_cache(vma, address, page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440 } else {
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003441 if (cow_page)
3442 mem_cgroup_uncharge_page(cow_page);
Nick Piggind00806b2007-07-19 01:46:57 -07003443 if (anon)
3444 page_cache_release(page);
3445 else
Nick Piggin54cb8822007-07-19 01:46:59 -07003446 anon = 1; /* no anon but release faulted_page */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447 }
3448
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003449 pte_unmap_unlock(page_table, ptl);
Nick Piggind00806b2007-07-19 01:46:57 -07003450
Nick Pigginb827e492009-04-30 15:08:16 -07003451 if (dirty_page) {
3452 struct address_space *mapping = page->mapping;
Jan Kara41c4d252012-06-12 16:20:28 +02003453 int dirtied = 0;
Nick Pigginb827e492009-04-30 15:08:16 -07003454
3455 if (set_page_dirty(dirty_page))
Jan Kara41c4d252012-06-12 16:20:28 +02003456 dirtied = 1;
Nick Pigginb827e492009-04-30 15:08:16 -07003457 unlock_page(dirty_page);
3458 put_page(dirty_page);
Jan Kara41c4d252012-06-12 16:20:28 +02003459 if ((dirtied || page_mkwrite) && mapping) {
Nick Pigginb827e492009-04-30 15:08:16 -07003460 /*
3461 * Some device drivers do not set page.mapping but still
3462 * dirty their pages
3463 */
3464 balance_dirty_pages_ratelimited(mapping);
3465 }
3466
3467 /* file_update_time outside page_lock */
Jan Kara41c4d252012-06-12 16:20:28 +02003468 if (vma->vm_file && !page_mkwrite)
Anton Salikhmetov8f7b3d12008-01-23 02:21:18 +03003469 file_update_time(vma->vm_file);
Nick Pigginb827e492009-04-30 15:08:16 -07003470 } else {
3471 unlock_page(vmf.page);
3472 if (anon)
3473 page_cache_release(vmf.page);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003474 }
Nick Piggind00806b2007-07-19 01:46:57 -07003475
Nick Piggin83c54072007-07-19 01:47:05 -07003476 return ret;
Nick Pigginb827e492009-04-30 15:08:16 -07003477
3478unwritable_page:
3479 page_cache_release(page);
3480 return ret;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003481uncharge_out:
3482 /* fs's fault handler get error */
3483 if (cow_page) {
3484 mem_cgroup_uncharge_page(cow_page);
3485 page_cache_release(cow_page);
3486 }
3487 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003488}
Nick Piggind00806b2007-07-19 01:46:57 -07003489
Nick Piggin54cb8822007-07-19 01:46:59 -07003490static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3491 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003492 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003493{
3494 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003495 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003496
Hugh Dickins16abfa02007-10-04 16:56:06 +01003497 pte_unmap(page_table);
3498 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003499}
3500
Jes Sorensenf4b81802006-09-27 01:50:10 -07003501/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502 * Fault of a previously existing named mapping. Repopulate the pte
3503 * from the encoded file_pte if possible. This enables swappable
3504 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003505 *
3506 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3507 * but allow concurrent faults), and pte mapped but not yet locked.
3508 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003510static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003511 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003512 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513{
Hugh Dickins65500d22005-10-29 18:15:59 -07003514 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003516 flags |= FAULT_FLAG_NONLINEAR;
3517
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003518 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003519 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520
Hugh Dickins2509ef22009-01-06 14:40:10 -08003521 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003522 /*
3523 * Page table corrupted: show pte and kill process.
3524 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003525 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003526 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003527 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003528
3529 pgoff = pte_to_pgoff(orig_pte);
Hugh Dickins16abfa02007-10-04 16:56:06 +01003530 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531}
3532
Mel Gorman9532fec2012-11-15 01:24:32 +00003533int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
Rik van Riel04bb2f92013-10-07 11:29:36 +01003534 unsigned long addr, int page_nid,
3535 int *flags)
Mel Gorman9532fec2012-11-15 01:24:32 +00003536{
3537 get_page(page);
3538
3539 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003540 if (page_nid == numa_node_id()) {
Mel Gorman9532fec2012-11-15 01:24:32 +00003541 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003542 *flags |= TNF_FAULT_LOCAL;
3543 }
Mel Gorman9532fec2012-11-15 01:24:32 +00003544
3545 return mpol_misplaced(page, vma, addr);
3546}
3547
Mel Gormand10e63f2012-10-25 14:16:31 +02003548int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3549 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3550{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003551 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003552 spinlock_t *ptl;
Mel Gorman8191acb2013-10-07 11:28:45 +01003553 int page_nid = -1;
Peter Zijlstra90572892013-10-07 11:29:20 +01003554 int last_cpupid;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003555 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003556 bool migrated = false;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003557 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003558
3559 /*
3560 * The "pte" at this point cannot be used safely without
3561 * validation through pte_unmap_same(). It's of NUMA type but
3562 * the pfn may be screwed if the read is non atomic.
3563 *
3564 * ptep_modify_prot_start is not called as this is clearing
3565 * the _PAGE_NUMA bit and it is not really expected that there
3566 * would be concurrent hardware modifications to the PTE.
3567 */
3568 ptl = pte_lockptr(mm, pmd);
3569 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003570 if (unlikely(!pte_same(*ptep, pte))) {
3571 pte_unmap_unlock(ptep, ptl);
3572 goto out;
3573 }
3574
Mel Gormand10e63f2012-10-25 14:16:31 +02003575 pte = pte_mknonnuma(pte);
3576 set_pte_at(mm, addr, ptep, pte);
3577 update_mmu_cache(vma, addr, ptep);
3578
3579 page = vm_normal_page(vma, addr, pte);
3580 if (!page) {
3581 pte_unmap_unlock(ptep, ptl);
3582 return 0;
3583 }
Mel Gormana1a46182013-10-07 11:28:50 +01003584 BUG_ON(is_zero_pfn(page_to_pfn(page)));
Mel Gormand10e63f2012-10-25 14:16:31 +02003585
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003586 /*
3587 * Avoid grouping on DSO/COW pages in specific and RO pages
3588 * in general, RO pages shouldn't hurt as much anyway since
3589 * they can be in shared cache state.
3590 */
3591 if (!pte_write(pte))
3592 flags |= TNF_NO_GROUP;
3593
Rik van Rieldabe1d92013-10-07 11:29:34 +01003594 /*
3595 * Flag if the page is shared between multiple address spaces. This
3596 * is later used when determining whether to group tasks together
3597 */
3598 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3599 flags |= TNF_SHARED;
3600
Peter Zijlstra90572892013-10-07 11:29:20 +01003601 last_cpupid = page_cpupid_last(page);
Mel Gorman8191acb2013-10-07 11:28:45 +01003602 page_nid = page_to_nid(page);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003603 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003604 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003605 if (target_nid == -1) {
Mel Gorman4daae3b2012-11-02 11:33:45 +00003606 put_page(page);
3607 goto out;
3608 }
3609
3610 /* Migrate to the requested node */
Mel Gorman1bc115d2013-10-07 11:29:05 +01003611 migrated = migrate_misplaced_page(page, vma, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003612 if (migrated) {
Mel Gorman8191acb2013-10-07 11:28:45 +01003613 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003614 flags |= TNF_MIGRATED;
3615 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00003616
3617out:
Mel Gorman8191acb2013-10-07 11:28:45 +01003618 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003619 task_numa_fault(last_cpupid, page_nid, 1, flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003620 return 0;
3621}
3622
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623/*
3624 * These routines also need to handle stuff like marking pages dirty
3625 * and/or accessed for architectures that don't do it in hardware (most
3626 * RISC architectures). The early dirtying is also good on the i386.
3627 *
3628 * There is also a hook called "update_mmu_cache()" that architectures
3629 * with external mmu caches can use to update those (ie the Sparc or
3630 * PowerPC hashed page tables that act as extended TLBs).
3631 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003632 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3633 * but allow concurrent faults), and pte mapped but not yet locked.
3634 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635 */
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003636static int handle_pte_fault(struct mm_struct *mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003637 struct vm_area_struct *vma, unsigned long address,
3638 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639{
3640 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003641 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07003643 entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003645 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003646 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003647 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003648 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003649 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003650 }
3651 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003652 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003653 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003655 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003656 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003657 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003658 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659 }
3660
Mel Gormand10e63f2012-10-25 14:16:31 +02003661 if (pte_numa(entry))
3662 return do_numa_page(mm, vma, address, entry, pte, pmd);
3663
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003664 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003665 spin_lock(ptl);
3666 if (unlikely(!pte_same(*pte, entry)))
3667 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003668 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003670 return do_wp_page(mm, vma, address,
3671 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003672 entry = pte_mkdirty(entry);
3673 }
3674 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003675 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003676 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003677 } else {
3678 /*
3679 * This is needed only for protection faults but the arch code
3680 * is not yet telling us if this is a protection fault or not.
3681 * This still avoids useless tlb flushes for .text page faults
3682 * with threads.
3683 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003684 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003685 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003686 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003687unlock:
3688 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003689 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003690}
3691
3692/*
3693 * By the time we get here, we already hold the mm semaphore
3694 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003695static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3696 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697{
3698 pgd_t *pgd;
3699 pud_t *pud;
3700 pmd_t *pmd;
3701 pte_t *pte;
3702
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003703 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003704 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705
David Rientjes1f1d06c2012-05-29 15:06:23 -07003706retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708 pud = pud_alloc(mm, pgd, address);
3709 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003710 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711 pmd = pmd_alloc(mm, pud, address);
3712 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003713 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003714 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003715 int ret = VM_FAULT_FALLBACK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003716 if (!vma->vm_ops)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003717 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3718 pmd, flags);
3719 if (!(ret & VM_FAULT_FALLBACK))
3720 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003721 } else {
3722 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003723 int ret;
3724
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003725 barrier();
3726 if (pmd_trans_huge(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003727 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3728
Linus Torvaldse53289c2013-01-09 08:36:54 -08003729 /*
3730 * If the pmd is splitting, return and retry the
3731 * the fault. Alternative: wait until the split
3732 * is done, and goto retry.
3733 */
3734 if (pmd_trans_splitting(orig_pmd))
3735 return 0;
3736
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003737 if (pmd_numa(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003738 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003739 orig_pmd, pmd);
3740
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003741 if (dirty && !pmd_write(orig_pmd)) {
David Rientjes1f1d06c2012-05-29 15:06:23 -07003742 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3743 orig_pmd);
3744 /*
3745 * If COW results in an oom, the huge pmd will
3746 * have been split, so retry the fault on the
3747 * pte for a smaller charge.
3748 */
3749 if (unlikely(ret & VM_FAULT_OOM))
3750 goto retry;
3751 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003752 } else {
3753 huge_pmd_set_accessed(mm, vma, address, pmd,
3754 orig_pmd, dirty);
David Rientjes1f1d06c2012-05-29 15:06:23 -07003755 }
Mel Gormand10e63f2012-10-25 14:16:31 +02003756
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003757 return 0;
3758 }
3759 }
3760
Mel Gorman0f19c172013-10-07 11:29:25 +01003761 /* THP should already have been handled */
3762 BUG_ON(pmd_numa(*pmd));
Mel Gormand10e63f2012-10-25 14:16:31 +02003763
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003764 /*
3765 * Use __pte_alloc instead of pte_alloc_map, because we can't
3766 * run pte_offset_map on the pmd, if an huge pmd could
3767 * materialize from under us from a different thread.
3768 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003769 if (unlikely(pmd_none(*pmd)) &&
3770 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003771 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003772 /* if an huge pmd materialized from under us just retry later */
3773 if (unlikely(pmd_trans_huge(*pmd)))
3774 return 0;
3775 /*
3776 * A regular pmd is established and it can't morph into a huge pmd
3777 * from under us anymore at this point because we hold the mmap_sem
3778 * read mode and khugepaged takes it in write mode. So now it's
3779 * safe to run pte_offset_map().
3780 */
3781 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003782
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003783 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003784}
3785
Johannes Weiner519e5242013-09-12 15:13:42 -07003786int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3787 unsigned long address, unsigned int flags)
3788{
3789 int ret;
3790
3791 __set_current_state(TASK_RUNNING);
3792
3793 count_vm_event(PGFAULT);
3794 mem_cgroup_count_vm_event(mm, PGFAULT);
3795
3796 /* do counter updates before entering really critical section. */
3797 check_sync_rss_stat(current);
3798
3799 /*
3800 * Enable the memcg OOM handling for faults triggered in user
3801 * space. Kernel faults are handled more gracefully.
3802 */
3803 if (flags & FAULT_FLAG_USER)
Johannes Weiner49426422013-10-16 13:46:59 -07003804 mem_cgroup_oom_enable();
Johannes Weiner519e5242013-09-12 15:13:42 -07003805
3806 ret = __handle_mm_fault(mm, vma, address, flags);
3807
Johannes Weiner49426422013-10-16 13:46:59 -07003808 if (flags & FAULT_FLAG_USER) {
3809 mem_cgroup_oom_disable();
3810 /*
3811 * The task may have entered a memcg OOM situation but
3812 * if the allocation error was handled gracefully (no
3813 * VM_FAULT_OOM), there is no need to kill anything.
3814 * Just clean up the OOM state peacefully.
3815 */
3816 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3817 mem_cgroup_oom_synchronize(false);
3818 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003819
Johannes Weiner519e5242013-09-12 15:13:42 -07003820 return ret;
3821}
3822
Linus Torvalds1da177e2005-04-16 15:20:36 -07003823#ifndef __PAGETABLE_PUD_FOLDED
3824/*
3825 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003826 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003827 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003828int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003830 pud_t *new = pud_alloc_one(mm, address);
3831 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003832 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833
Nick Piggin362a61a2008-05-14 06:37:36 +02003834 smp_wmb(); /* See comment in __pte_alloc */
3835
Hugh Dickins872fec12005-10-29 18:16:21 -07003836 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003837 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003838 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003839 else
3840 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003841 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003842 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003843}
3844#endif /* __PAGETABLE_PUD_FOLDED */
3845
3846#ifndef __PAGETABLE_PMD_FOLDED
3847/*
3848 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003849 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003851int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003853 pmd_t *new = pmd_alloc_one(mm, address);
3854 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003855 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856
Nick Piggin362a61a2008-05-14 06:37:36 +02003857 smp_wmb(); /* See comment in __pte_alloc */
3858
Hugh Dickins872fec12005-10-29 18:16:21 -07003859 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003861 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003862 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003863 else
3864 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003865#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003866 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003867 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003868 else
3869 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003871 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003872 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873}
3874#endif /* __PAGETABLE_PMD_FOLDED */
3875
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876#if !defined(__HAVE_ARCH_GATE_AREA)
3877
3878#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07003879static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880
3881static int __init gate_vma_init(void)
3882{
3883 gate_vma.vm_mm = NULL;
3884 gate_vma.vm_start = FIXADDR_USER_START;
3885 gate_vma.vm_end = FIXADDR_USER_END;
Roland McGrathb6558c42007-01-26 00:56:47 -08003886 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3887 gate_vma.vm_page_prot = __P101;
Jason Baron909af762012-03-23 15:02:51 -07003888
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889 return 0;
3890}
3891__initcall(gate_vma_init);
3892#endif
3893
Stephen Wilson31db58b2011-03-13 15:49:15 -04003894struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003895{
3896#ifdef AT_SYSINFO_EHDR
3897 return &gate_vma;
3898#else
3899 return NULL;
3900#endif
3901}
3902
Stephen Wilsoncae5d392011-03-13 15:49:17 -04003903int in_gate_area_no_mm(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904{
3905#ifdef AT_SYSINFO_EHDR
3906 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3907 return 1;
3908#endif
3909 return 0;
3910}
3911
3912#endif /* __HAVE_ARCH_GATE_AREA */
David Howells0ec76a12006-09-27 01:50:15 -07003913
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003914static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003915 pte_t **ptepp, spinlock_t **ptlp)
3916{
3917 pgd_t *pgd;
3918 pud_t *pud;
3919 pmd_t *pmd;
3920 pte_t *ptep;
3921
3922 pgd = pgd_offset(mm, address);
3923 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3924 goto out;
3925
3926 pud = pud_offset(pgd, address);
3927 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3928 goto out;
3929
3930 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003931 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003932 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3933 goto out;
3934
3935 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3936 if (pmd_huge(*pmd))
3937 goto out;
3938
3939 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3940 if (!ptep)
3941 goto out;
3942 if (!pte_present(*ptep))
3943 goto unlock;
3944 *ptepp = ptep;
3945 return 0;
3946unlock:
3947 pte_unmap_unlock(ptep, *ptlp);
3948out:
3949 return -EINVAL;
3950}
3951
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003952static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3953 pte_t **ptepp, spinlock_t **ptlp)
3954{
3955 int res;
3956
3957 /* (void) is needed to make gcc happy */
3958 (void) __cond_lock(*ptlp,
3959 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3960 return res;
3961}
3962
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003963/**
3964 * follow_pfn - look up PFN at a user virtual address
3965 * @vma: memory mapping
3966 * @address: user virtual address
3967 * @pfn: location to store found PFN
3968 *
3969 * Only IO mappings and raw PFN mappings are allowed.
3970 *
3971 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3972 */
3973int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3974 unsigned long *pfn)
3975{
3976 int ret = -EINVAL;
3977 spinlock_t *ptl;
3978 pte_t *ptep;
3979
3980 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3981 return ret;
3982
3983 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3984 if (ret)
3985 return ret;
3986 *pfn = pte_pfn(*ptep);
3987 pte_unmap_unlock(ptep, ptl);
3988 return 0;
3989}
3990EXPORT_SYMBOL(follow_pfn);
3991
Rik van Riel28b2ee22008-07-23 21:27:05 -07003992#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003993int follow_phys(struct vm_area_struct *vma,
3994 unsigned long address, unsigned int flags,
3995 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003996{
Johannes Weiner03668a42009-06-16 15:32:34 -07003997 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003998 pte_t *ptep, pte;
3999 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004000
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004001 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4002 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004003
Johannes Weiner03668a42009-06-16 15:32:34 -07004004 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004005 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004006 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07004007
Rik van Riel28b2ee22008-07-23 21:27:05 -07004008 if ((flags & FOLL_WRITE) && !pte_write(pte))
4009 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004010
4011 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07004012 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004013
Johannes Weiner03668a42009-06-16 15:32:34 -07004014 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004015unlock:
4016 pte_unmap_unlock(ptep, ptl);
4017out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004018 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004019}
4020
4021int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4022 void *buf, int len, int write)
4023{
4024 resource_size_t phys_addr;
4025 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08004026 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004027 int offset = addr & (PAGE_SIZE-1);
4028
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08004029 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07004030 return -EINVAL;
4031
4032 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4033 if (write)
4034 memcpy_toio(maddr + offset, buf, len);
4035 else
4036 memcpy_fromio(buf, maddr + offset, len);
4037 iounmap(maddr);
4038
4039 return len;
4040}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02004041EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07004042#endif
4043
David Howells0ec76a12006-09-27 01:50:15 -07004044/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004045 * Access another process' address space as given in mm. If non-NULL, use the
4046 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07004047 */
Stephen Wilson206cb632011-03-13 15:49:19 -04004048static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4049 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07004050{
David Howells0ec76a12006-09-27 01:50:15 -07004051 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07004052 void *old_buf = buf;
4053
David Howells0ec76a12006-09-27 01:50:15 -07004054 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02004055 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07004056 while (len) {
4057 int bytes, ret, offset;
4058 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07004059 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07004060
4061 ret = get_user_pages(tsk, mm, addr, 1,
4062 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07004063 if (ret <= 0) {
4064 /*
4065 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4066 * we can access using slightly different code.
4067 */
4068#ifdef CONFIG_HAVE_IOREMAP_PROT
4069 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07004070 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07004071 break;
4072 if (vma->vm_ops && vma->vm_ops->access)
4073 ret = vma->vm_ops->access(vma, addr, buf,
4074 len, write);
4075 if (ret <= 0)
4076#endif
4077 break;
4078 bytes = ret;
David Howells0ec76a12006-09-27 01:50:15 -07004079 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07004080 bytes = len;
4081 offset = addr & (PAGE_SIZE-1);
4082 if (bytes > PAGE_SIZE-offset)
4083 bytes = PAGE_SIZE-offset;
4084
4085 maddr = kmap(page);
4086 if (write) {
4087 copy_to_user_page(vma, page, addr,
4088 maddr + offset, buf, bytes);
4089 set_page_dirty_lock(page);
4090 } else {
4091 copy_from_user_page(vma, page, addr,
4092 buf, maddr + offset, bytes);
4093 }
4094 kunmap(page);
4095 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07004096 }
David Howells0ec76a12006-09-27 01:50:15 -07004097 len -= bytes;
4098 buf += bytes;
4099 addr += bytes;
4100 }
4101 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07004102
4103 return buf - old_buf;
4104}
Andi Kleen03252912008-01-30 13:33:18 +01004105
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004106/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07004107 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04004108 * @mm: the mm_struct of the target address space
4109 * @addr: start address to access
4110 * @buf: source or destination buffer
4111 * @len: number of bytes to transfer
4112 * @write: whether the access is a write
4113 *
4114 * The caller must hold a reference on @mm.
4115 */
4116int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4117 void *buf, int len, int write)
4118{
4119 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4120}
4121
Andi Kleen03252912008-01-30 13:33:18 +01004122/*
Stephen Wilson206cb632011-03-13 15:49:19 -04004123 * Access another process' address space.
4124 * Source/target buffer must be kernel space,
4125 * Do not walk the page table directly, use get_user_pages
4126 */
4127int access_process_vm(struct task_struct *tsk, unsigned long addr,
4128 void *buf, int len, int write)
4129{
4130 struct mm_struct *mm;
4131 int ret;
4132
4133 mm = get_task_mm(tsk);
4134 if (!mm)
4135 return 0;
4136
4137 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4138 mmput(mm);
4139
4140 return ret;
4141}
4142
Andi Kleen03252912008-01-30 13:33:18 +01004143/*
4144 * Print the name of a VMA.
4145 */
4146void print_vma_addr(char *prefix, unsigned long ip)
4147{
4148 struct mm_struct *mm = current->mm;
4149 struct vm_area_struct *vma;
4150
Ingo Molnare8bff742008-02-13 20:21:06 +01004151 /*
4152 * Do not print if we are in atomic
4153 * contexts (in exception stacks, etc.):
4154 */
4155 if (preempt_count())
4156 return;
4157
Andi Kleen03252912008-01-30 13:33:18 +01004158 down_read(&mm->mmap_sem);
4159 vma = find_vma(mm, ip);
4160 if (vma && vma->vm_file) {
4161 struct file *f = vma->vm_file;
4162 char *buf = (char *)__get_free_page(GFP_KERNEL);
4163 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004164 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01004165
Jan Blunckcf28b482008-02-14 19:38:44 -08004166 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01004167 if (IS_ERR(p))
4168 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004169 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01004170 vma->vm_start,
4171 vma->vm_end - vma->vm_start);
4172 free_page((unsigned long)buf);
4173 }
4174 }
Jeff Liu51a07e52012-07-31 16:43:18 -07004175 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01004176}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004177
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004178#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004179void might_fault(void)
4180{
Peter Zijlstra95156f02009-01-12 13:02:11 +01004181 /*
4182 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4183 * holding the mmap_sem, this is safe because kernel memory doesn't
4184 * get paged out, therefore we'll never actually fault, and the
4185 * below annotations will generate false positives.
4186 */
4187 if (segment_eq(get_fs(), KERNEL_DS))
4188 return;
4189
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004190 /*
4191 * it would be nicer only to annotate paths which are not under
4192 * pagefault_disable, however that requires a larger audit and
4193 * providing helpers like get_user_atomic.
4194 */
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004195 if (in_atomic())
4196 return;
4197
4198 __might_sleep(__FILE__, __LINE__, 0);
4199
4200 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004201 might_lock_read(&current->mm->mmap_sem);
4202}
4203EXPORT_SYMBOL(might_fault);
4204#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004205
4206#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4207static void clear_gigantic_page(struct page *page,
4208 unsigned long addr,
4209 unsigned int pages_per_huge_page)
4210{
4211 int i;
4212 struct page *p = page;
4213
4214 might_sleep();
4215 for (i = 0; i < pages_per_huge_page;
4216 i++, p = mem_map_next(p, page, i)) {
4217 cond_resched();
4218 clear_user_highpage(p, addr + i * PAGE_SIZE);
4219 }
4220}
4221void clear_huge_page(struct page *page,
4222 unsigned long addr, unsigned int pages_per_huge_page)
4223{
4224 int i;
4225
4226 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4227 clear_gigantic_page(page, addr, pages_per_huge_page);
4228 return;
4229 }
4230
4231 might_sleep();
4232 for (i = 0; i < pages_per_huge_page; i++) {
4233 cond_resched();
4234 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4235 }
4236}
4237
4238static void copy_user_gigantic_page(struct page *dst, struct page *src,
4239 unsigned long addr,
4240 struct vm_area_struct *vma,
4241 unsigned int pages_per_huge_page)
4242{
4243 int i;
4244 struct page *dst_base = dst;
4245 struct page *src_base = src;
4246
4247 for (i = 0; i < pages_per_huge_page; ) {
4248 cond_resched();
4249 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4250
4251 i++;
4252 dst = mem_map_next(dst, dst_base, i);
4253 src = mem_map_next(src, src_base, i);
4254 }
4255}
4256
4257void copy_user_huge_page(struct page *dst, struct page *src,
4258 unsigned long addr, struct vm_area_struct *vma,
4259 unsigned int pages_per_huge_page)
4260{
4261 int i;
4262
4263 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4264 copy_user_gigantic_page(dst, src, addr, vma,
4265 pages_per_huge_page);
4266 return;
4267 }
4268
4269 might_sleep();
4270 for (i = 0; i < pages_per_huge_page; i++) {
4271 cond_resched();
4272 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4273 }
4274}
4275#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004276
Olof Johansson40b64ac2013-12-20 14:28:05 -08004277#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004278
4279static struct kmem_cache *page_ptl_cachep;
4280
4281void __init ptlock_cache_init(void)
4282{
4283 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4284 SLAB_PANIC, NULL);
4285}
4286
Peter Zijlstra539edb52013-11-14 14:31:52 -08004287bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004288{
4289 spinlock_t *ptl;
4290
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004291 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004292 if (!ptl)
4293 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08004294 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004295 return true;
4296}
4297
Peter Zijlstra539edb52013-11-14 14:31:52 -08004298void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004299{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004300 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004301}
4302#endif