blob: adeac306610f7d7895e9d361d11649fb31f0a914 [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>
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -070063#include <linux/debugfs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
Alexey Dobriyan6952b612009-09-18 23:55:55 +040065#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070066#include <asm/pgalloc.h>
67#include <asm/uaccess.h>
68#include <asm/tlb.h>
69#include <asm/tlbflush.h>
70#include <asm/pgtable.h>
71
Jan Beulich42b77722008-07-23 21:27:10 -070072#include "internal.h"
73
Peter Zijlstra90572892013-10-07 11:29:20 +010074#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
Peter Zijlstra75980e92013-02-22 16:34:32 -080076#endif
77
Andy Whitcroftd41dee32005-06-23 00:07:54 -070078#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070079/* use the per-pgdat data instead for discontigmem - mbligh */
80unsigned long max_mapnr;
81struct page *mem_map;
82
83EXPORT_SYMBOL(max_mapnr);
84EXPORT_SYMBOL(mem_map);
85#endif
86
Linus Torvalds1da177e2005-04-16 15:20:36 -070087/*
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92 * and ZONE_HIGHMEM.
93 */
94void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070095
Linus Torvalds1da177e2005-04-16 15:20:36 -070096EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
Ingo Molnar32a93232008-02-06 22:39:44 +010098/*
99 * Randomize the address space (stacks, mmaps, brk, etc.).
100 *
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
103 */
104int randomize_va_space __read_mostly =
105#ifdef CONFIG_COMPAT_BRK
106 1;
107#else
108 2;
109#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100110
111static int __init disable_randmaps(char *s)
112{
113 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800114 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100115}
116__setup("norandmaps", disable_randmaps);
117
Hugh Dickins62eede62009-09-21 17:03:34 -0700118unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700119unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700120
121/*
122 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123 */
124static int __init init_zero_pfn(void)
125{
126 zero_pfn = page_to_pfn(ZERO_PAGE(0));
127 return 0;
128}
129core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100130
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800131
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800132#if defined(SPLIT_RSS_COUNTING)
133
David Rientjesea48cf72012-03-21 16:34:13 -0700134void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800135{
136 int i;
137
138 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700139 if (current->rss_stat.count[i]) {
140 add_mm_counter(mm, i, current->rss_stat.count[i]);
141 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800142 }
143 }
David Rientjes05af2e12012-03-21 16:34:13 -0700144 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800145}
146
147static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148{
149 struct task_struct *task = current;
150
151 if (likely(task->mm == mm))
152 task->rss_stat.count[member] += val;
153 else
154 add_mm_counter(mm, member, val);
155}
156#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159/* sync counter once per 64 page faults */
160#define TASK_RSS_EVENTS_THRESH (64)
161static void check_sync_rss_stat(struct task_struct *task)
162{
163 if (unlikely(task != current))
164 return;
165 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700166 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800167}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700168#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800169
170#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173static void check_sync_rss_stat(struct task_struct *task)
174{
175}
176
Peter Zijlstra9547d012011-05-24 17:12:14 -0700177#endif /* SPLIT_RSS_COUNTING */
178
179#ifdef HAVE_GENERIC_MMU_GATHER
180
181static int tlb_next_batch(struct mmu_gather *tlb)
182{
183 struct mmu_gather_batch *batch;
184
185 batch = tlb->active;
186 if (batch->next) {
187 tlb->active = batch->next;
188 return 1;
189 }
190
Michal Hocko53a59fc2013-01-04 15:35:12 -0800191 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192 return 0;
193
Peter Zijlstra9547d012011-05-24 17:12:14 -0700194 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195 if (!batch)
196 return 0;
197
Michal Hocko53a59fc2013-01-04 15:35:12 -0800198 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700199 batch->next = NULL;
200 batch->nr = 0;
201 batch->max = MAX_GATHER_BATCH;
202
203 tlb->active->next = batch;
204 tlb->active = batch;
205
206 return 1;
207}
208
209/* tlb_gather_mmu
210 * Called to initialize an (on-stack) mmu_gather structure for page-table
211 * tear-down from @mm. The @fullmm argument is used when @mm is without
212 * users and we're going to destroy the full address space (exit/execve).
213 */
Linus Torvalds2b047252013-08-15 11:42:25 -0700214void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700215{
216 tlb->mm = mm;
217
Linus Torvalds2b047252013-08-15 11:42:25 -0700218 /* Is it from 0 to ~0? */
219 tlb->fullmm = !(start | (end+1));
Dave Hansen1de14c32013-04-12 16:23:54 -0700220 tlb->need_flush_all = 0;
Linus Torvalds2b047252013-08-15 11:42:25 -0700221 tlb->start = start;
222 tlb->end = end;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700223 tlb->need_flush = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700224 tlb->local.next = NULL;
225 tlb->local.nr = 0;
226 tlb->local.max = ARRAY_SIZE(tlb->__pages);
227 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800228 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700229
230#ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800232#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700233}
234
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700235static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700236{
Peter Zijlstra9547d012011-05-24 17:12:14 -0700237 tlb->need_flush = 0;
238 tlb_flush(tlb);
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb_table_flush(tlb);
241#endif
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700242}
243
244static void tlb_flush_mmu_free(struct mmu_gather *tlb)
245{
246 struct mmu_gather_batch *batch;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700247
Peter Zijlstra9547d012011-05-24 17:12:14 -0700248 for (batch = &tlb->local; batch; batch = batch->next) {
249 free_pages_and_swap_cache(batch->pages, batch->nr);
250 batch->nr = 0;
251 }
252 tlb->active = &tlb->local;
253}
254
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700255void tlb_flush_mmu(struct mmu_gather *tlb)
256{
257 if (!tlb->need_flush)
258 return;
259 tlb_flush_mmu_tlbonly(tlb);
260 tlb_flush_mmu_free(tlb);
261}
262
Peter Zijlstra9547d012011-05-24 17:12:14 -0700263/* tlb_finish_mmu
264 * Called at the end of the shootdown operation to free up any resources
265 * that were required.
266 */
267void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
268{
269 struct mmu_gather_batch *batch, *next;
270
271 tlb_flush_mmu(tlb);
272
273 /* keep the page table cache within bounds */
274 check_pgt_cache();
275
276 for (batch = tlb->local.next; batch; batch = next) {
277 next = batch->next;
278 free_pages((unsigned long)batch, 0);
279 }
280 tlb->local.next = NULL;
281}
282
283/* __tlb_remove_page
284 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
285 * handling the additional races in SMP caused by other CPUs caching valid
286 * mappings in their TLBs. Returns the number of free page slots left.
287 * When out of page slots we must call tlb_flush_mmu().
288 */
289int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
290{
291 struct mmu_gather_batch *batch;
292
Shaohua Lif21760b2012-01-12 17:19:16 -0800293 VM_BUG_ON(!tlb->need_flush);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700294
Peter Zijlstra9547d012011-05-24 17:12:14 -0700295 batch = tlb->active;
296 batch->pages[batch->nr++] = page;
297 if (batch->nr == batch->max) {
298 if (!tlb_next_batch(tlb))
299 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700300 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700301 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800302 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700303
304 return batch->max - batch->nr;
305}
306
307#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800308
Peter Zijlstra26723912011-05-24 17:12:00 -0700309#ifdef CONFIG_HAVE_RCU_TABLE_FREE
310
311/*
312 * See the comment near struct mmu_table_batch.
313 */
314
315static void tlb_remove_table_smp_sync(void *arg)
316{
317 /* Simply deliver the interrupt */
318}
319
320static void tlb_remove_table_one(void *table)
321{
322 /*
323 * This isn't an RCU grace period and hence the page-tables cannot be
324 * assumed to be actually RCU-freed.
325 *
326 * It is however sufficient for software page-table walkers that rely on
327 * IRQ disabling. See the comment near struct mmu_table_batch.
328 */
329 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
330 __tlb_remove_table(table);
331}
332
333static void tlb_remove_table_rcu(struct rcu_head *head)
334{
335 struct mmu_table_batch *batch;
336 int i;
337
338 batch = container_of(head, struct mmu_table_batch, rcu);
339
340 for (i = 0; i < batch->nr; i++)
341 __tlb_remove_table(batch->tables[i]);
342
343 free_page((unsigned long)batch);
344}
345
346void tlb_table_flush(struct mmu_gather *tlb)
347{
348 struct mmu_table_batch **batch = &tlb->batch;
349
350 if (*batch) {
351 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
352 *batch = NULL;
353 }
354}
355
356void tlb_remove_table(struct mmu_gather *tlb, void *table)
357{
358 struct mmu_table_batch **batch = &tlb->batch;
359
360 tlb->need_flush = 1;
361
362 /*
363 * When there's less then two users of this mm there cannot be a
364 * concurrent page-table walk.
365 */
366 if (atomic_read(&tlb->mm->mm_users) < 2) {
367 __tlb_remove_table(table);
368 return;
369 }
370
371 if (*batch == NULL) {
372 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
373 if (*batch == NULL) {
374 tlb_remove_table_one(table);
375 return;
376 }
377 (*batch)->nr = 0;
378 }
379 (*batch)->tables[(*batch)->nr++] = table;
380 if ((*batch)->nr == MAX_TABLE_BATCH)
381 tlb_table_flush(tlb);
382}
383
Peter Zijlstra9547d012011-05-24 17:12:14 -0700384#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700385
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 * Note: this doesn't free the actual pages themselves. That
388 * has been handled earlier when unmapping all the memory regions.
389 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000390static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
391 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800393 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700394 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000395 pte_free_tlb(tlb, token, addr);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800396 atomic_long_dec(&tlb->mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397}
398
Hugh Dickinse0da3822005-04-19 13:29:15 -0700399static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
400 unsigned long addr, unsigned long end,
401 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402{
403 pmd_t *pmd;
404 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700405 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406
Hugh Dickinse0da3822005-04-19 13:29:15 -0700407 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409 do {
410 next = pmd_addr_end(addr, end);
411 if (pmd_none_or_clear_bad(pmd))
412 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000413 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 } while (pmd++, addr = next, addr != end);
415
Hugh Dickinse0da3822005-04-19 13:29:15 -0700416 start &= PUD_MASK;
417 if (start < floor)
418 return;
419 if (ceiling) {
420 ceiling &= PUD_MASK;
421 if (!ceiling)
422 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700424 if (end - 1 > ceiling - 1)
425 return;
426
427 pmd = pmd_offset(pud, start);
428 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000429 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430}
431
Hugh Dickinse0da3822005-04-19 13:29:15 -0700432static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
433 unsigned long addr, unsigned long end,
434 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435{
436 pud_t *pud;
437 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700438 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439
Hugh Dickinse0da3822005-04-19 13:29:15 -0700440 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 do {
443 next = pud_addr_end(addr, end);
444 if (pud_none_or_clear_bad(pud))
445 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700446 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 } while (pud++, addr = next, addr != end);
448
Hugh Dickinse0da3822005-04-19 13:29:15 -0700449 start &= PGDIR_MASK;
450 if (start < floor)
451 return;
452 if (ceiling) {
453 ceiling &= PGDIR_MASK;
454 if (!ceiling)
455 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700457 if (end - 1 > ceiling - 1)
458 return;
459
460 pud = pud_offset(pgd, start);
461 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000462 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463}
464
465/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700466 * This function frees user-level page tables of a process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 */
Jan Beulich42b77722008-07-23 21:27:10 -0700468void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700469 unsigned long addr, unsigned long end,
470 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471{
472 pgd_t *pgd;
473 unsigned long next;
474
Hugh Dickinse0da3822005-04-19 13:29:15 -0700475 /*
476 * The next few lines have given us lots of grief...
477 *
478 * Why are we testing PMD* at this top level? Because often
479 * there will be no work to do at all, and we'd prefer not to
480 * go all the way down to the bottom just to discover that.
481 *
482 * Why all these "- 1"s? Because 0 represents both the bottom
483 * of the address space and the top of it (using -1 for the
484 * top wouldn't help much: the masks would do the wrong thing).
485 * The rule is that addr 0 and floor 0 refer to the bottom of
486 * the address space, but end 0 and ceiling 0 refer to the top
487 * Comparisons need to use "end - 1" and "ceiling - 1" (though
488 * that end 0 case should be mythical).
489 *
490 * Wherever addr is brought up or ceiling brought down, we must
491 * be careful to reject "the opposite 0" before it confuses the
492 * subsequent tests. But what about where end is brought down
493 * by PMD_SIZE below? no, end can't go down to 0 there.
494 *
495 * Whereas we round start (addr) and ceiling down, by different
496 * masks at different levels, in order to test whether a table
497 * now has no other vmas using it, so can be freed, we don't
498 * bother to round floor or end up - the tests don't need that.
499 */
500
501 addr &= PMD_MASK;
502 if (addr < floor) {
503 addr += PMD_SIZE;
504 if (!addr)
505 return;
506 }
507 if (ceiling) {
508 ceiling &= PMD_MASK;
509 if (!ceiling)
510 return;
511 }
512 if (end - 1 > ceiling - 1)
513 end -= PMD_SIZE;
514 if (addr > end - 1)
515 return;
516
Jan Beulich42b77722008-07-23 21:27:10 -0700517 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 do {
519 next = pgd_addr_end(addr, end);
520 if (pgd_none_or_clear_bad(pgd))
521 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700522 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700524}
525
Jan Beulich42b77722008-07-23 21:27:10 -0700526void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700527 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700528{
529 while (vma) {
530 struct vm_area_struct *next = vma->vm_next;
531 unsigned long addr = vma->vm_start;
532
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700533 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000534 * Hide vma from rmap and truncate_pagecache before freeing
535 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700536 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800537 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700538 unlink_file_vma(vma);
539
David Gibson9da61ae2006-03-22 00:08:57 -0800540 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700541 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700542 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700543 } else {
544 /*
545 * Optimization: gather nearby vmas into one call down
546 */
547 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800548 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700549 vma = next;
550 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800551 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700552 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700553 }
554 free_pgd_range(tlb, addr, vma->vm_end,
555 floor, next? next->vm_start: ceiling);
556 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700557 vma = next;
558 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559}
560
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800561int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
562 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800564 spinlock_t *ptl;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800565 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800566 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700567 if (!new)
568 return -ENOMEM;
569
Nick Piggin362a61a2008-05-14 06:37:36 +0200570 /*
571 * Ensure all pte setup (eg. pte page lock and page clearing) are
572 * visible before the pte is made visible to other CPUs by being
573 * put into page tables.
574 *
575 * The other side of the story is the pointer chasing in the page
576 * table walking code (when walking the page table without locking;
577 * ie. most of the time). Fortunately, these data accesses consist
578 * of a chain of data-dependent loads, meaning most CPUs (alpha
579 * being the notable exception) will already guarantee loads are
580 * seen in-order. See the alpha page table accessors for the
581 * smp_read_barrier_depends() barriers in page table walking code.
582 */
583 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
584
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800585 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800586 wait_split_huge_page = 0;
587 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800588 atomic_long_inc(&mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800590 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800591 } else if (unlikely(pmd_trans_splitting(*pmd)))
592 wait_split_huge_page = 1;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800593 spin_unlock(ptl);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800594 if (new)
595 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800596 if (wait_split_huge_page)
597 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700598 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599}
600
Hugh Dickins1bb36302005-10-29 18:16:22 -0700601int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700603 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
604 if (!new)
605 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606
Nick Piggin362a61a2008-05-14 06:37:36 +0200607 smp_wmb(); /* See comment in __pte_alloc */
608
Hugh Dickins1bb36302005-10-29 18:16:22 -0700609 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800610 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700611 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800612 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800613 } else
614 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700615 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800616 if (new)
617 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700618 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619}
620
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800621static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700622{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800623 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
624}
625
626static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
627{
628 int i;
629
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800630 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700631 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800632 for (i = 0; i < NR_MM_COUNTERS; i++)
633 if (rss[i])
634 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700635}
636
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800638 * This function is called to print an error when a bad pte
639 * is found. For example, we might have a PFN-mapped pte in
640 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700641 *
642 * The calling function must still handle the error.
643 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800644static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
645 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700646{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800647 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
648 pud_t *pud = pud_offset(pgd, addr);
649 pmd_t *pmd = pmd_offset(pud, addr);
650 struct address_space *mapping;
651 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800652 static unsigned long resume;
653 static unsigned long nr_shown;
654 static unsigned long nr_unshown;
655
656 /*
657 * Allow a burst of 60 reports, then keep quiet for that minute;
658 * or allow a steady drip of one report per second.
659 */
660 if (nr_shown == 60) {
661 if (time_before(jiffies, resume)) {
662 nr_unshown++;
663 return;
664 }
665 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800666 printk(KERN_ALERT
667 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800668 nr_unshown);
669 nr_unshown = 0;
670 }
671 nr_shown = 0;
672 }
673 if (nr_shown++ == 0)
674 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800675
676 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
677 index = linear_page_index(vma, addr);
678
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800679 printk(KERN_ALERT
680 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800681 current->comm,
682 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800683 if (page)
Dave Hansenf0b791a2014-01-23 15:52:49 -0800684 dump_page(page, "bad pte");
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800685 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800686 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
687 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
688 /*
689 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
690 */
691 if (vma->vm_ops)
Joe Perches071361d2012-12-12 10:19:12 -0800692 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
693 vma->vm_ops->fault);
Al Viro72c2d532013-09-22 16:27:52 -0400694 if (vma->vm_file)
Joe Perches071361d2012-12-12 10:19:12 -0800695 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
696 vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700697 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030698 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700699}
700
701/*
Nick Piggin7e675132008-04-28 02:13:00 -0700702 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800703 *
Nick Piggin7e675132008-04-28 02:13:00 -0700704 * "Special" mappings do not wish to be associated with a "struct page" (either
705 * it doesn't exist, or it exists but they don't want to touch it). In this
706 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700707 *
Nick Piggin7e675132008-04-28 02:13:00 -0700708 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
709 * pte bit, in which case this function is trivial. Secondly, an architecture
710 * may not have a spare pte bit, which requires a more complicated scheme,
711 * described below.
712 *
713 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
714 * special mapping (even if there are underlying and valid "struct pages").
715 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800716 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700717 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
718 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700719 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
720 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800721 *
722 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
723 *
Nick Piggin7e675132008-04-28 02:13:00 -0700724 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700725 *
Nick Piggin7e675132008-04-28 02:13:00 -0700726 * This restricts such mappings to be a linear translation from virtual address
727 * to pfn. To get around this restriction, we allow arbitrary mappings so long
728 * as the vma is not a COW mapping; in that case, we know that all ptes are
729 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700730 *
731 *
Nick Piggin7e675132008-04-28 02:13:00 -0700732 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
733 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700734 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
735 * page" backing, however the difference is that _all_ pages with a struct
736 * page (that is, those where pfn_valid is true) are refcounted and considered
737 * normal pages by the VM. The disadvantage is that pages are refcounted
738 * (which can be slower and simply not an option for some PFNMAP users). The
739 * advantage is that we don't have to follow the strict linearity rule of
740 * PFNMAP mappings in order to support COWable mappings.
741 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800742 */
Nick Piggin7e675132008-04-28 02:13:00 -0700743#ifdef __HAVE_ARCH_PTE_SPECIAL
744# define HAVE_PTE_SPECIAL 1
745#else
746# define HAVE_PTE_SPECIAL 0
747#endif
748struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
749 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800750{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800751 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700752
753 if (HAVE_PTE_SPECIAL) {
Hugh Dickinsb38af472014-08-29 15:18:44 -0700754 if (likely(!pte_special(pte)))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800755 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700756 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
757 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700758 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800759 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700760 return NULL;
761 }
762
763 /* !HAVE_PTE_SPECIAL case follows: */
764
Jared Hulbertb379d792008-04-28 02:12:58 -0700765 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
766 if (vma->vm_flags & VM_MIXEDMAP) {
767 if (!pfn_valid(pfn))
768 return NULL;
769 goto out;
770 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700771 unsigned long off;
772 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700773 if (pfn == vma->vm_pgoff + off)
774 return NULL;
775 if (!is_cow_mapping(vma->vm_flags))
776 return NULL;
777 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800778 }
779
Hugh Dickinsb38af472014-08-29 15:18:44 -0700780 if (is_zero_pfn(pfn))
781 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800782check_pfn:
783 if (unlikely(pfn > highest_memmap_pfn)) {
784 print_bad_pte(vma, addr, pte, NULL);
785 return NULL;
786 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800787
788 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700789 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700790 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800791 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700792out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800793 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800794}
795
796/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 * copy one vm_area from one task to the other. Assumes the page tables
798 * already present in the new task to be cleared in the whole range
799 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 */
801
Hugh Dickins570a335b2009-12-14 17:58:46 -0800802static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700804 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700805 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806{
Nick Pigginb5810032005-10-29 18:16:12 -0700807 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 pte_t pte = *src_pte;
809 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810
811 /* pte contains position in swap or file, so copy. */
812 if (unlikely(!pte_present(pte))) {
813 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700814 swp_entry_t entry = pte_to_swp_entry(pte);
815
Hugh Dickins570a335b2009-12-14 17:58:46 -0800816 if (swap_duplicate(entry) < 0)
817 return entry.val;
818
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 /* make sure dst_mm is on swapoff's mmlist. */
820 if (unlikely(list_empty(&dst_mm->mmlist))) {
821 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700822 if (list_empty(&dst_mm->mmlist))
823 list_add(&dst_mm->mmlist,
824 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 spin_unlock(&mmlist_lock);
826 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800827 if (likely(!non_swap_entry(entry)))
828 rss[MM_SWAPENTS]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800829 else if (is_migration_entry(entry)) {
830 page = migration_entry_to_page(entry);
831
832 if (PageAnon(page))
833 rss[MM_ANONPAGES]++;
834 else
835 rss[MM_FILEPAGES]++;
836
837 if (is_write_migration_entry(entry) &&
838 is_cow_mapping(vm_flags)) {
839 /*
840 * COW mappings require pages in both
841 * parent and child to be set to read.
842 */
843 make_migration_entry_read(&entry);
844 pte = swp_entry_to_pte(entry);
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700845 if (pte_swp_soft_dirty(*src_pte))
846 pte = pte_swp_mksoft_dirty(pte);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800847 set_pte_at(src_mm, addr, src_pte, pte);
848 }
Christoph Lameter06972122006-06-23 02:03:35 -0700849 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700851 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 }
853
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 /*
855 * If it's a COW mapping, write protect it both
856 * in the parent and the child
857 */
Linus Torvalds67121172005-12-11 20:38:17 -0800858 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700860 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 }
862
863 /*
864 * If it's a shared mapping, mark it clean in
865 * the child
866 */
867 if (vm_flags & VM_SHARED)
868 pte = pte_mkclean(pte);
869 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800870
871 page = vm_normal_page(vma, addr, pte);
872 if (page) {
873 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700874 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800875 if (PageAnon(page))
876 rss[MM_ANONPAGES]++;
877 else
878 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800879 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700880
881out_set_pte:
882 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800883 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884}
885
Jerome Marchand21bda262014-08-06 16:06:56 -0700886static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800887 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
888 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700890 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700892 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700893 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800894 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800895 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896
897again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800898 init_rss_vec(rss);
899
Hugh Dickinsc74df322005-10-29 18:16:23 -0700900 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 if (!dst_pte)
902 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700903 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700904 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700905 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700906 orig_src_pte = src_pte;
907 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700908 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 do {
911 /*
912 * We are holding two locks at this point - either of them
913 * could generate latencies in another task on another CPU.
914 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700915 if (progress >= 32) {
916 progress = 0;
917 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100918 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700919 break;
920 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 if (pte_none(*src_pte)) {
922 progress++;
923 continue;
924 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800925 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
926 vma, addr, rss);
927 if (entry.val)
928 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 progress += 8;
930 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700932 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700933 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700934 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800935 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700936 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700937 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800938
939 if (entry.val) {
940 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
941 return -ENOMEM;
942 progress = 0;
943 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 if (addr != end)
945 goto again;
946 return 0;
947}
948
949static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
950 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
951 unsigned long addr, unsigned long end)
952{
953 pmd_t *src_pmd, *dst_pmd;
954 unsigned long next;
955
956 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
957 if (!dst_pmd)
958 return -ENOMEM;
959 src_pmd = pmd_offset(src_pud, addr);
960 do {
961 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800962 if (pmd_trans_huge(*src_pmd)) {
963 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800964 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800965 err = copy_huge_pmd(dst_mm, src_mm,
966 dst_pmd, src_pmd, addr, vma);
967 if (err == -ENOMEM)
968 return -ENOMEM;
969 if (!err)
970 continue;
971 /* fall through */
972 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 if (pmd_none_or_clear_bad(src_pmd))
974 continue;
975 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
976 vma, addr, next))
977 return -ENOMEM;
978 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
979 return 0;
980}
981
982static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
983 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
984 unsigned long addr, unsigned long end)
985{
986 pud_t *src_pud, *dst_pud;
987 unsigned long next;
988
989 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
990 if (!dst_pud)
991 return -ENOMEM;
992 src_pud = pud_offset(src_pgd, addr);
993 do {
994 next = pud_addr_end(addr, end);
995 if (pud_none_or_clear_bad(src_pud))
996 continue;
997 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
998 vma, addr, next))
999 return -ENOMEM;
1000 } while (dst_pud++, src_pud++, addr = next, addr != end);
1001 return 0;
1002}
1003
1004int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1005 struct vm_area_struct *vma)
1006{
1007 pgd_t *src_pgd, *dst_pgd;
1008 unsigned long next;
1009 unsigned long addr = vma->vm_start;
1010 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001011 unsigned long mmun_start; /* For mmu_notifiers */
1012 unsigned long mmun_end; /* For mmu_notifiers */
1013 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001014 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015
Nick Piggind9928952005-08-28 16:49:11 +10001016 /*
1017 * Don't copy ptes where a page fault will fill them correctly.
1018 * Fork becomes much lighter when there are big shared or private
1019 * readonly mappings. The tradeoff is that copy_page_range is more
1020 * efficient than faulting.
1021 */
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001022 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1023 VM_PFNMAP | VM_MIXEDMAP))) {
Nick Piggind9928952005-08-28 16:49:11 +10001024 if (!vma->anon_vma)
1025 return 0;
1026 }
1027
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 if (is_vm_hugetlb_page(vma))
1029 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1030
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001031 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001032 /*
1033 * We do not free on error cases below as remove_vma
1034 * gets called on error from higher level routine
1035 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001036 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001037 if (ret)
1038 return ret;
1039 }
1040
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001041 /*
1042 * We need to invalidate the secondary MMU mappings only when
1043 * there could be a permission downgrade on the ptes of the
1044 * parent mm. And a permission downgrade will only happen if
1045 * is_cow_mapping() returns true.
1046 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001047 is_cow = is_cow_mapping(vma->vm_flags);
1048 mmun_start = addr;
1049 mmun_end = end;
1050 if (is_cow)
1051 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1052 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001053
1054 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 dst_pgd = pgd_offset(dst_mm, addr);
1056 src_pgd = pgd_offset(src_mm, addr);
1057 do {
1058 next = pgd_addr_end(addr, end);
1059 if (pgd_none_or_clear_bad(src_pgd))
1060 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001061 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1062 vma, addr, next))) {
1063 ret = -ENOMEM;
1064 break;
1065 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001067
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001068 if (is_cow)
1069 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001070 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071}
1072
Robin Holt51c6f662005-11-13 16:06:42 -08001073static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001074 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001076 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077{
Nick Pigginb5810032005-10-29 18:16:12 -07001078 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001079 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001080 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001081 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001082 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001083 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001084
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001085again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001086 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001087 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1088 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001089 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 do {
1091 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001092 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001094 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001095
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001097 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001098
Linus Torvalds6aab3412005-11-28 14:34:23 -08001099 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 if (unlikely(details) && page) {
1101 /*
1102 * unmap_shared_mapping_pages() wants to
1103 * invalidate cache without truncating:
1104 * unmap shared but keep private pages.
1105 */
1106 if (details->check_mapping &&
1107 details->check_mapping != page->mapping)
1108 continue;
1109 /*
1110 * Each page->index must be checked when
1111 * invalidating or truncating nonlinear.
1112 */
1113 if (details->nonlinear_vma &&
1114 (page->index < details->first_index ||
1115 page->index > details->last_index))
1116 continue;
1117 }
Nick Pigginb5810032005-10-29 18:16:12 -07001118 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001119 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 tlb_remove_tlb_entry(tlb, pte, addr);
1121 if (unlikely(!page))
1122 continue;
1123 if (unlikely(details) && details->nonlinear_vma
1124 && linear_page_index(details->nonlinear_vma,
Cyrill Gorcunov41bb3472013-08-13 16:00:51 -07001125 addr) != page->index) {
1126 pte_t ptfile = pgoff_to_pte(page->index);
1127 if (pte_soft_dirty(ptent))
1128 pte_file_mksoft_dirty(ptfile);
1129 set_pte_at(mm, addr, pte, ptfile);
1130 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001132 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001133 else {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001134 if (pte_dirty(ptent)) {
1135 force_flush = 1;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001136 set_page_dirty(page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001137 }
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001138 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001139 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001140 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001141 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001142 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001143 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001144 if (unlikely(page_mapcount(page) < 0))
1145 print_bad_pte(vma, addr, ptent, page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001146 if (unlikely(!__tlb_remove_page(tlb, page))) {
1147 force_flush = 1;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001148 break;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001149 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 continue;
1151 }
1152 /*
1153 * If details->check_mapping, we leave swap entries;
1154 * if details->nonlinear_vma, we leave file entries.
1155 */
1156 if (unlikely(details))
1157 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001158 if (pte_file(ptent)) {
1159 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1160 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001161 } else {
1162 swp_entry_t entry = pte_to_swp_entry(ptent);
1163
1164 if (!non_swap_entry(entry))
1165 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001166 else if (is_migration_entry(entry)) {
1167 struct page *page;
1168
1169 page = migration_entry_to_page(entry);
1170
1171 if (PageAnon(page))
1172 rss[MM_ANONPAGES]--;
1173 else
1174 rss[MM_FILEPAGES]--;
1175 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001176 if (unlikely(!free_swap_and_cache(entry)))
1177 print_bad_pte(vma, addr, ptent, NULL);
1178 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001179 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001180 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001181
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001182 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001183 arch_leave_lazy_mmu_mode();
Robin Holt51c6f662005-11-13 16:06:42 -08001184
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001185 /* Do the actual TLB flush before dropping ptl */
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001186 if (force_flush) {
Linus Torvalds2b047252013-08-15 11:42:25 -07001187 unsigned long old_end;
1188
Linus Torvalds2b047252013-08-15 11:42:25 -07001189 /*
1190 * Flush the TLB just for the previous segment,
1191 * then update the range to be the remaining
1192 * TLB range.
1193 */
1194 old_end = tlb->end;
Vineet Guptae6c495a2013-07-03 15:03:31 -07001195 tlb->end = addr;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001196 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds2b047252013-08-15 11:42:25 -07001197 tlb->start = addr;
1198 tlb->end = old_end;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001199 }
1200 pte_unmap_unlock(start_pte, ptl);
1201
1202 /*
1203 * If we forced a TLB flush (either due to running out of
1204 * batch buffers or because we needed to flush dirty TLB
1205 * entries before releasing the ptl), free the batched
1206 * memory too. Restart if we didn't do everything.
1207 */
1208 if (force_flush) {
1209 force_flush = 0;
1210 tlb_flush_mmu_free(tlb);
Linus Torvalds2b047252013-08-15 11:42:25 -07001211
1212 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001213 goto again;
1214 }
1215
Robin Holt51c6f662005-11-13 16:06:42 -08001216 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217}
1218
Robin Holt51c6f662005-11-13 16:06:42 -08001219static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001220 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001222 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223{
1224 pmd_t *pmd;
1225 unsigned long next;
1226
1227 pmd = pmd_offset(pud, addr);
1228 do {
1229 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001230 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001231 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001232#ifdef CONFIG_DEBUG_VM
1233 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1234 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1235 __func__, addr, end,
1236 vma->vm_start,
1237 vma->vm_end);
1238 BUG();
1239 }
1240#endif
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001241 split_huge_page_pmd(vma, addr, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001242 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001243 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001244 /* fall through */
1245 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001246 /*
1247 * Here there can be other concurrent MADV_DONTNEED or
1248 * trans huge page faults running, and if the pmd is
1249 * none or trans huge it can change under us. This is
1250 * because MADV_DONTNEED holds the mmap_sem in read
1251 * mode.
1252 */
1253 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1254 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001255 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001256next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001257 cond_resched();
1258 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001259
1260 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261}
1262
Robin Holt51c6f662005-11-13 16:06:42 -08001263static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001264 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001266 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267{
1268 pud_t *pud;
1269 unsigned long next;
1270
1271 pud = pud_offset(pgd, addr);
1272 do {
1273 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001274 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001276 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1277 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001278
1279 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280}
1281
Al Viro038c7aa2012-03-05 13:25:09 -05001282static void unmap_page_range(struct mmu_gather *tlb,
1283 struct vm_area_struct *vma,
1284 unsigned long addr, unsigned long end,
1285 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286{
1287 pgd_t *pgd;
1288 unsigned long next;
1289
1290 if (details && !details->check_mapping && !details->nonlinear_vma)
1291 details = NULL;
1292
1293 BUG_ON(addr >= end);
1294 tlb_start_vma(tlb, vma);
1295 pgd = pgd_offset(vma->vm_mm, addr);
1296 do {
1297 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001298 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001300 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1301 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 tlb_end_vma(tlb, vma);
1303}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304
Al Virof5cc4ee2012-03-05 14:14:20 -05001305
1306static void unmap_single_vma(struct mmu_gather *tlb,
1307 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001308 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001309 struct zap_details *details)
1310{
1311 unsigned long start = max(vma->vm_start, start_addr);
1312 unsigned long end;
1313
1314 if (start >= vma->vm_end)
1315 return;
1316 end = min(vma->vm_end, end_addr);
1317 if (end <= vma->vm_start)
1318 return;
1319
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301320 if (vma->vm_file)
1321 uprobe_munmap(vma, start, end);
1322
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001323 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001324 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001325
1326 if (start != end) {
1327 if (unlikely(is_vm_hugetlb_page(vma))) {
1328 /*
1329 * It is undesirable to test vma->vm_file as it
1330 * should be non-null for valid hugetlb area.
1331 * However, vm_file will be NULL in the error
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001332 * cleanup path of mmap_region. When
Al Virof5cc4ee2012-03-05 14:14:20 -05001333 * hugetlbfs ->mmap method fails,
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001334 * mmap_region() nullifies vma->vm_file
Al Virof5cc4ee2012-03-05 14:14:20 -05001335 * before calling this function to clean up.
1336 * Since no pte has actually been setup, it is
1337 * safe to do nothing in this case.
1338 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001339 if (vma->vm_file) {
1340 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Mel Gormand8333522012-07-31 16:46:20 -07001341 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001342 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1343 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001344 } else
1345 unmap_page_range(tlb, vma, start, end, details);
1346 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347}
1348
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349/**
1350 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001351 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 * @vma: the starting vma
1353 * @start_addr: virtual address at which to start unmapping
1354 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001356 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 * Only addresses between `start' and `end' will be unmapped.
1359 *
1360 * The VMA list must be sorted in ascending virtual address order.
1361 *
1362 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1363 * range after unmap_vmas() returns. So the only responsibility here is to
1364 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1365 * drops the lock and schedules.
1366 */
Al Viro6e8bb012012-03-05 13:41:15 -05001367void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001369 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001371 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001373 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001374 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001375 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001376 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377}
1378
1379/**
1380 * zap_page_range - remove user pages in a given range
1381 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001382 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 * @size: number of bytes to zap
1384 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001385 *
1386 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001388void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 unsigned long size, struct zap_details *details)
1390{
1391 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001392 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001393 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001396 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001397 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001398 mmu_notifier_invalidate_range_start(mm, start, end);
1399 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001400 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001401 mmu_notifier_invalidate_range_end(mm, start, end);
1402 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403}
1404
Jack Steinerc627f9c2008-07-29 22:33:53 -07001405/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001406 * zap_page_range_single - remove user pages in a given range
1407 * @vma: vm_area_struct holding the applicable pages
1408 * @address: starting address of pages to zap
1409 * @size: number of bytes to zap
1410 * @details: details of nonlinear truncation or shared cache invalidation
1411 *
1412 * The range must fit into one VMA.
1413 */
1414static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1415 unsigned long size, struct zap_details *details)
1416{
1417 struct mm_struct *mm = vma->vm_mm;
1418 struct mmu_gather tlb;
1419 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001420
1421 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001422 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001423 update_hiwater_rss(mm);
1424 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001425 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001426 mmu_notifier_invalidate_range_end(mm, address, end);
1427 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428}
1429
Jack Steinerc627f9c2008-07-29 22:33:53 -07001430/**
1431 * zap_vma_ptes - remove ptes mapping the vma
1432 * @vma: vm_area_struct holding ptes to be zapped
1433 * @address: starting address of pages to zap
1434 * @size: number of bytes to zap
1435 *
1436 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1437 *
1438 * The entire address range must be fully contained within the vma.
1439 *
1440 * Returns 0 if successful.
1441 */
1442int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1443 unsigned long size)
1444{
1445 if (address < vma->vm_start || address + size > vma->vm_end ||
1446 !(vma->vm_flags & VM_PFNMAP))
1447 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001448 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001449 return 0;
1450}
1451EXPORT_SYMBOL_GPL(zap_vma_ptes);
1452
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001453pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001454 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001455{
1456 pgd_t * pgd = pgd_offset(mm, addr);
1457 pud_t * pud = pud_alloc(mm, pgd, addr);
1458 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001459 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001460 if (pmd) {
1461 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001462 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001463 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001464 }
1465 return NULL;
1466}
1467
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001469 * This is the old fallback for page remapping.
1470 *
1471 * For historical reasons, it only allows reserved pages. Only
1472 * old drivers should use this, and they needed to mark their
1473 * pages reserved for the old functions anyway.
1474 */
Nick Piggin423bad602008-04-28 02:13:01 -07001475static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1476 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001477{
Nick Piggin423bad602008-04-28 02:13:01 -07001478 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001479 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001480 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001481 spinlock_t *ptl;
1482
Linus Torvalds238f58d2005-11-29 13:01:56 -08001483 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001484 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001485 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001486 retval = -ENOMEM;
1487 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001488 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001489 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001490 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001491 retval = -EBUSY;
1492 if (!pte_none(*pte))
1493 goto out_unlock;
1494
1495 /* Ok, finally just insert the thing.. */
1496 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08001497 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001498 page_add_file_rmap(page);
1499 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1500
1501 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001502 pte_unmap_unlock(pte, ptl);
1503 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001504out_unlock:
1505 pte_unmap_unlock(pte, ptl);
1506out:
1507 return retval;
1508}
1509
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001510/**
1511 * vm_insert_page - insert single page into user vma
1512 * @vma: user vma to map to
1513 * @addr: target user address of this page
1514 * @page: source kernel page
1515 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001516 * This allows drivers to insert individual pages they've allocated
1517 * into a user vma.
1518 *
1519 * The page has to be a nice clean _individual_ kernel allocation.
1520 * If you allocate a compound page, you need to have marked it as
1521 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001522 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001523 *
1524 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1525 * took an arbitrary page protection parameter. This doesn't allow
1526 * that. Your vma protection will have to be set up correctly, which
1527 * means that if you want a shared writable mapping, you'd better
1528 * ask for a shared writable mapping!
1529 *
1530 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001531 *
1532 * Usually this function is called from f_op->mmap() handler
1533 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1534 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1535 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001536 */
Nick Piggin423bad602008-04-28 02:13:01 -07001537int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1538 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001539{
1540 if (addr < vma->vm_start || addr >= vma->vm_end)
1541 return -EFAULT;
1542 if (!page_count(page))
1543 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001544 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1545 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1546 BUG_ON(vma->vm_flags & VM_PFNMAP);
1547 vma->vm_flags |= VM_MIXEDMAP;
1548 }
Nick Piggin423bad602008-04-28 02:13:01 -07001549 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001550}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001551EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001552
Nick Piggin423bad602008-04-28 02:13:01 -07001553static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1554 unsigned long pfn, pgprot_t prot)
1555{
1556 struct mm_struct *mm = vma->vm_mm;
1557 int retval;
1558 pte_t *pte, entry;
1559 spinlock_t *ptl;
1560
1561 retval = -ENOMEM;
1562 pte = get_locked_pte(mm, addr, &ptl);
1563 if (!pte)
1564 goto out;
1565 retval = -EBUSY;
1566 if (!pte_none(*pte))
1567 goto out_unlock;
1568
1569 /* Ok, finally just insert the thing.. */
1570 entry = pte_mkspecial(pfn_pte(pfn, prot));
1571 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001572 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad602008-04-28 02:13:01 -07001573
1574 retval = 0;
1575out_unlock:
1576 pte_unmap_unlock(pte, ptl);
1577out:
1578 return retval;
1579}
1580
Nick Piggine0dc0d82007-02-12 00:51:36 -08001581/**
1582 * vm_insert_pfn - insert single pfn into user vma
1583 * @vma: user vma to map to
1584 * @addr: target user address of this page
1585 * @pfn: source kernel pfn
1586 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07001587 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08001588 * they've allocated into a user vma. Same comments apply.
1589 *
1590 * This function should only be called from a vm_ops->fault handler, and
1591 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001592 *
1593 * vma cannot be a COW mapping.
1594 *
1595 * As this is called only for pages that do not currently exist, we
1596 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001597 */
1598int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad602008-04-28 02:13:01 -07001599 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001600{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001601 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001602 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07001603 /*
1604 * Technically, architectures with pte_special can avoid all these
1605 * restrictions (same for remap_pfn_range). However we would like
1606 * consistency in testing and feature parity among all, so we should
1607 * try to keep these invariants in place for everybody.
1608 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001609 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1610 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1611 (VM_PFNMAP|VM_MIXEDMAP));
1612 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1613 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001614
Nick Piggin423bad602008-04-28 02:13:01 -07001615 if (addr < vma->vm_start || addr >= vma->vm_end)
1616 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07001617 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001618 return -EINVAL;
1619
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001620 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001621
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001622 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001623}
1624EXPORT_SYMBOL(vm_insert_pfn);
1625
Nick Piggin423bad602008-04-28 02:13:01 -07001626int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1627 unsigned long pfn)
1628{
1629 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1630
1631 if (addr < vma->vm_start || addr >= vma->vm_end)
1632 return -EFAULT;
1633
1634 /*
1635 * If we don't have pte special, then we have to use the pfn_valid()
1636 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1637 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001638 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1639 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad602008-04-28 02:13:01 -07001640 */
1641 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1642 struct page *page;
1643
1644 page = pfn_to_page(pfn);
1645 return insert_page(vma, addr, page, vma->vm_page_prot);
1646 }
1647 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1648}
1649EXPORT_SYMBOL(vm_insert_mixed);
1650
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001651/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652 * maps a range of physical memory into the requested pages. the old
1653 * mappings are removed. any references to nonexistent pages results
1654 * in null mappings (currently treated as "copy-on-access")
1655 */
1656static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1657 unsigned long addr, unsigned long end,
1658 unsigned long pfn, pgprot_t prot)
1659{
1660 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001661 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662
Hugh Dickinsc74df322005-10-29 18:16:23 -07001663 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664 if (!pte)
1665 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001666 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 do {
1668 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07001669 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 pfn++;
1671 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001672 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001673 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674 return 0;
1675}
1676
1677static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1678 unsigned long addr, unsigned long end,
1679 unsigned long pfn, pgprot_t prot)
1680{
1681 pmd_t *pmd;
1682 unsigned long next;
1683
1684 pfn -= addr >> PAGE_SHIFT;
1685 pmd = pmd_alloc(mm, pud, addr);
1686 if (!pmd)
1687 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001688 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 do {
1690 next = pmd_addr_end(addr, end);
1691 if (remap_pte_range(mm, pmd, addr, next,
1692 pfn + (addr >> PAGE_SHIFT), prot))
1693 return -ENOMEM;
1694 } while (pmd++, addr = next, addr != end);
1695 return 0;
1696}
1697
1698static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1699 unsigned long addr, unsigned long end,
1700 unsigned long pfn, pgprot_t prot)
1701{
1702 pud_t *pud;
1703 unsigned long next;
1704
1705 pfn -= addr >> PAGE_SHIFT;
1706 pud = pud_alloc(mm, pgd, addr);
1707 if (!pud)
1708 return -ENOMEM;
1709 do {
1710 next = pud_addr_end(addr, end);
1711 if (remap_pmd_range(mm, pud, addr, next,
1712 pfn + (addr >> PAGE_SHIFT), prot))
1713 return -ENOMEM;
1714 } while (pud++, addr = next, addr != end);
1715 return 0;
1716}
1717
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001718/**
1719 * remap_pfn_range - remap kernel memory to userspace
1720 * @vma: user vma to map to
1721 * @addr: target user address to start at
1722 * @pfn: physical address of kernel memory
1723 * @size: size of map area
1724 * @prot: page protection flags for this mapping
1725 *
1726 * Note: this is only safe if the mm semaphore is held when called.
1727 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1729 unsigned long pfn, unsigned long size, pgprot_t prot)
1730{
1731 pgd_t *pgd;
1732 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001733 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 struct mm_struct *mm = vma->vm_mm;
1735 int err;
1736
1737 /*
1738 * Physically remapped pages are special. Tell the
1739 * rest of the world about it:
1740 * VM_IO tells people not to look at these pages
1741 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08001742 * VM_PFNMAP tells the core MM that the base pages are just
1743 * raw PFN mappings, and do not have a "struct page" associated
1744 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001745 * VM_DONTEXPAND
1746 * Disable vma merging and expanding with mremap().
1747 * VM_DONTDUMP
1748 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001749 *
1750 * There's a horrible special case to handle copy-on-write
1751 * behaviour that some programs depend on. We mark the "original"
1752 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001753 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001755 if (is_cow_mapping(vma->vm_flags)) {
1756 if (addr != vma->vm_start || end != vma->vm_end)
1757 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001758 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001759 }
1760
1761 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1762 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001763 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001764
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001765 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766
1767 BUG_ON(addr >= end);
1768 pfn -= addr >> PAGE_SHIFT;
1769 pgd = pgd_offset(mm, addr);
1770 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771 do {
1772 next = pgd_addr_end(addr, end);
1773 err = remap_pud_range(mm, pgd, addr, next,
1774 pfn + (addr >> PAGE_SHIFT), prot);
1775 if (err)
1776 break;
1777 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001778
1779 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07001780 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001781
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782 return err;
1783}
1784EXPORT_SYMBOL(remap_pfn_range);
1785
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07001786/**
1787 * vm_iomap_memory - remap memory to userspace
1788 * @vma: user vma to map to
1789 * @start: start of area
1790 * @len: size of area
1791 *
1792 * This is a simplified io_remap_pfn_range() for common driver use. The
1793 * driver just needs to give us the physical memory range to be mapped,
1794 * we'll figure out the rest from the vma information.
1795 *
1796 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1797 * whatever write-combining details or similar.
1798 */
1799int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1800{
1801 unsigned long vm_len, pfn, pages;
1802
1803 /* Check that the physical memory area passed in looks valid */
1804 if (start + len < start)
1805 return -EINVAL;
1806 /*
1807 * You *really* shouldn't map things that aren't page-aligned,
1808 * but we've historically allowed it because IO memory might
1809 * just have smaller alignment.
1810 */
1811 len += start & ~PAGE_MASK;
1812 pfn = start >> PAGE_SHIFT;
1813 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1814 if (pfn + pages < pfn)
1815 return -EINVAL;
1816
1817 /* We start the mapping 'vm_pgoff' pages into the area */
1818 if (vma->vm_pgoff > pages)
1819 return -EINVAL;
1820 pfn += vma->vm_pgoff;
1821 pages -= vma->vm_pgoff;
1822
1823 /* Can we fit all of the mapping? */
1824 vm_len = vma->vm_end - vma->vm_start;
1825 if (vm_len >> PAGE_SHIFT > pages)
1826 return -EINVAL;
1827
1828 /* Ok, let it rip */
1829 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1830}
1831EXPORT_SYMBOL(vm_iomap_memory);
1832
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001833static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1834 unsigned long addr, unsigned long end,
1835 pte_fn_t fn, void *data)
1836{
1837 pte_t *pte;
1838 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001839 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001840 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001841
1842 pte = (mm == &init_mm) ?
1843 pte_alloc_kernel(pmd, addr) :
1844 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1845 if (!pte)
1846 return -ENOMEM;
1847
1848 BUG_ON(pmd_huge(*pmd));
1849
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001850 arch_enter_lazy_mmu_mode();
1851
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001852 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001853
1854 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001855 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001856 if (err)
1857 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001858 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001859
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001860 arch_leave_lazy_mmu_mode();
1861
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001862 if (mm != &init_mm)
1863 pte_unmap_unlock(pte-1, ptl);
1864 return err;
1865}
1866
1867static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1868 unsigned long addr, unsigned long end,
1869 pte_fn_t fn, void *data)
1870{
1871 pmd_t *pmd;
1872 unsigned long next;
1873 int err;
1874
Andi Kleenceb86872008-07-23 21:27:50 -07001875 BUG_ON(pud_huge(*pud));
1876
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001877 pmd = pmd_alloc(mm, pud, addr);
1878 if (!pmd)
1879 return -ENOMEM;
1880 do {
1881 next = pmd_addr_end(addr, end);
1882 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1883 if (err)
1884 break;
1885 } while (pmd++, addr = next, addr != end);
1886 return err;
1887}
1888
1889static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1890 unsigned long addr, unsigned long end,
1891 pte_fn_t fn, void *data)
1892{
1893 pud_t *pud;
1894 unsigned long next;
1895 int err;
1896
1897 pud = pud_alloc(mm, pgd, addr);
1898 if (!pud)
1899 return -ENOMEM;
1900 do {
1901 next = pud_addr_end(addr, end);
1902 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1903 if (err)
1904 break;
1905 } while (pud++, addr = next, addr != end);
1906 return err;
1907}
1908
1909/*
1910 * Scan a region of virtual memory, filling in page tables as necessary
1911 * and calling a provided function on each leaf page table.
1912 */
1913int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1914 unsigned long size, pte_fn_t fn, void *data)
1915{
1916 pgd_t *pgd;
1917 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001918 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001919 int err;
1920
1921 BUG_ON(addr >= end);
1922 pgd = pgd_offset(mm, addr);
1923 do {
1924 next = pgd_addr_end(addr, end);
1925 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1926 if (err)
1927 break;
1928 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001929
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001930 return err;
1931}
1932EXPORT_SYMBOL_GPL(apply_to_page_range);
1933
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001935 * handle_pte_fault chooses page fault handler according to an entry
1936 * which was read non-atomically. Before making any commitment, on
1937 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001938 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001939 * must check under lock before unmapping the pte and proceeding
1940 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001941 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001942 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001943static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001944 pte_t *page_table, pte_t orig_pte)
1945{
1946 int same = 1;
1947#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1948 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001949 spinlock_t *ptl = pte_lockptr(mm, pmd);
1950 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001951 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001952 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001953 }
1954#endif
1955 pte_unmap(page_table);
1956 return same;
1957}
1958
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001959static 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 -08001960{
Dan Williams0abdd7a2014-01-21 15:48:12 -08001961 debug_dma_assert_idle(src);
1962
Linus Torvalds6aab3412005-11-28 14:34:23 -08001963 /*
1964 * If the source page was a PFN mapping, we don't have
1965 * a "struct page" for it. We do a best-effort copy by
1966 * just copying from the original user address. If that
1967 * fails, we just zero-fill it. Live with it.
1968 */
1969 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08001970 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08001971 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1972
1973 /*
1974 * This really shouldn't fail, because the page is there
1975 * in the page tables. But it might just be unreadable,
1976 * in which case we just give up and fill the result with
1977 * zeroes.
1978 */
1979 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07001980 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08001981 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07001982 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08001983 } else
1984 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08001985}
1986
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987/*
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001988 * Notify the address space that the page is about to become writable so that
1989 * it can prohibit this or wait for the page to get into an appropriate state.
1990 *
1991 * We do this without the lock held, so that it can sleep if it needs to.
1992 */
1993static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1994 unsigned long address)
1995{
1996 struct vm_fault vmf;
1997 int ret;
1998
1999 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2000 vmf.pgoff = page->index;
2001 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2002 vmf.page = page;
2003
2004 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2005 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2006 return ret;
2007 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2008 lock_page(page);
2009 if (!page->mapping) {
2010 unlock_page(page);
2011 return 0; /* retry */
2012 }
2013 ret |= VM_FAULT_LOCKED;
2014 } else
2015 VM_BUG_ON_PAGE(!PageLocked(page), page);
2016 return ret;
2017}
2018
2019/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020 * This routine handles present pages, when users try to write
2021 * to a shared page. It is done by copying the page to a new address
2022 * and decrementing the shared-page counter for the old page.
2023 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024 * Note that this routine assumes that the protection checks have been
2025 * done by the caller (the low-level page fault routine in most cases).
2026 * Thus we can safely just mark it writable once we've done any necessary
2027 * COW.
2028 *
2029 * We also mark the page dirty at this point even though the page will
2030 * change only once the write actually happens. This avoids a few races,
2031 * and potentially makes it more efficient.
2032 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002033 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2034 * but allow concurrent faults), with pte both mapped and locked.
2035 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002037static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2038 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002039 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002040 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041{
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002042 struct page *old_page, *new_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002044 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002045 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002046 struct page *dirty_page = NULL;
David Rientjes17569542012-11-16 14:14:48 -08002047 unsigned long mmun_start = 0; /* For mmu_notifiers */
2048 unsigned long mmun_end = 0; /* For mmu_notifiers */
Johannes Weiner00501b52014-08-08 14:19:20 -07002049 struct mem_cgroup *memcg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050
Linus Torvalds6aab3412005-11-28 14:34:23 -08002051 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002052 if (!old_page) {
2053 /*
2054 * VM_MIXEDMAP !pfn_valid() case
2055 *
2056 * We should not cow pages in a shared writeable mapping.
2057 * Just mark the pages writable as we can't do any dirty
2058 * accounting on raw pfn maps.
2059 */
2060 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2061 (VM_WRITE|VM_SHARED))
2062 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002063 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002064 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002066 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002067 * Take out anonymous pages first, anonymous shared vmas are
2068 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002069 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002070 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002071 if (!trylock_page(old_page)) {
2072 page_cache_get(old_page);
2073 pte_unmap_unlock(page_table, ptl);
2074 lock_page(old_page);
2075 page_table = pte_offset_map_lock(mm, pmd, address,
2076 &ptl);
2077 if (!pte_same(*page_table, orig_pte)) {
2078 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002079 goto unlock;
2080 }
2081 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002082 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002083 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002084 /*
2085 * The page is all ours. Move it to our anon_vma so
2086 * the rmap code will not search our parent or siblings.
2087 * Protected against the rmap code by the page lock.
2088 */
2089 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002090 unlock_page(old_page);
2091 goto reuse;
2092 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002093 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002094 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002095 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002096 /*
2097 * Only catch write-faults on shared writable pages,
2098 * read-only shared pages can get COWed by
2099 * get_user_pages(.write=1, .force=1).
2100 */
David Howells9637a5e2006-06-23 02:03:43 -07002101 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002102 int tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002103 page_cache_get(old_page);
2104 pte_unmap_unlock(page_table, ptl);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002105 tmp = do_page_mkwrite(vma, old_page, address);
2106 if (unlikely(!tmp || (tmp &
2107 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2108 page_cache_release(old_page);
2109 return tmp;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002110 }
David Howells9637a5e2006-06-23 02:03:43 -07002111 /*
2112 * Since we dropped the lock we need to revalidate
2113 * the PTE as someone else may have changed it. If
2114 * they did, we just return, as we can count on the
2115 * MMU to tell us if they didn't also make it writable.
2116 */
2117 page_table = pte_offset_map_lock(mm, pmd, address,
2118 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002119 if (!pte_same(*page_table, orig_pte)) {
2120 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002121 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002122 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002123
2124 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002126 dirty_page = old_page;
2127 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002128
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002129reuse:
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002130 /*
2131 * Clear the pages cpupid information as the existing
2132 * information potentially belongs to a now completely
2133 * unrelated process.
2134 */
2135 if (old_page)
2136 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2137
David Howells9637a5e2006-06-23 02:03:43 -07002138 flush_cache_page(vma, address, pte_pfn(orig_pte));
2139 entry = pte_mkyoung(orig_pte);
2140 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002141 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002142 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002143 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002144 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002145
2146 if (!dirty_page)
2147 return ret;
2148
2149 /*
2150 * Yes, Virginia, this is actually required to prevent a race
2151 * with clear_page_dirty_for_io() from clearing the page dirty
2152 * bit after it clear all dirty ptes, but before a racing
2153 * do_wp_page installs a dirty pte.
2154 *
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002155 * do_shared_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002156 */
2157 if (!page_mkwrite) {
2158 wait_on_page_locked(dirty_page);
Miklos Szeredied6d7c82014-04-07 15:37:51 -07002159 set_page_dirty_balance(dirty_page);
Jan Kara41c4d252012-06-12 16:20:28 +02002160 /* file_update_time outside page_lock */
2161 if (vma->vm_file)
2162 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002163 }
2164 put_page(dirty_page);
2165 if (page_mkwrite) {
2166 struct address_space *mapping = dirty_page->mapping;
2167
2168 set_page_dirty(dirty_page);
2169 unlock_page(dirty_page);
2170 page_cache_release(dirty_page);
2171 if (mapping) {
2172 /*
2173 * Some device drivers do not set page.mapping
2174 * but still dirty their pages
2175 */
2176 balance_dirty_pages_ratelimited(mapping);
2177 }
2178 }
2179
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002180 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182
2183 /*
2184 * Ok, we need to copy. Oh, well..
2185 */
Nick Pigginb5810032005-10-29 18:16:12 -07002186 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002187gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002188 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189
2190 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002191 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002192
Hugh Dickins62eede62009-09-21 17:03:34 -07002193 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002194 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2195 if (!new_page)
2196 goto oom;
2197 } else {
2198 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2199 if (!new_page)
2200 goto oom;
2201 cow_user_page(new_page, old_page, address, vma);
2202 }
2203 __SetPageUptodate(new_page);
2204
Johannes Weiner00501b52014-08-08 14:19:20 -07002205 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002206 goto oom_free_new;
2207
Haggai Eran6bdb9132012-10-08 16:33:35 -07002208 mmun_start = address & PAGE_MASK;
David Rientjes17569542012-11-16 14:14:48 -08002209 mmun_end = mmun_start + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -07002210 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2211
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 /*
2213 * Re-check the pte - we dropped the lock
2214 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002215 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002216 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002217 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002218 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002219 dec_mm_counter_fast(mm, MM_FILEPAGES);
2220 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002221 }
2222 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002223 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002224 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002225 entry = mk_pte(new_page, vma->vm_page_prot);
2226 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002227 /*
2228 * Clear the pte entry and flush it first, before updating the
2229 * pte with the new entry. This will avoid a race condition
2230 * seen in the presence of one thread doing SMC and another
2231 * thread doing COW.
2232 */
Izik Eidus828502d2009-09-21 17:01:51 -07002233 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002234 page_add_new_anon_rmap(new_page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002235 mem_cgroup_commit_charge(new_page, memcg, false);
2236 lru_cache_add_active_or_unevictable(new_page, vma);
Izik Eidus828502d2009-09-21 17:01:51 -07002237 /*
2238 * We call the notify macro here because, when using secondary
2239 * mmu page tables (such as kvm shadow page tables), we want the
2240 * new page to be mapped directly into the secondary page table.
2241 */
2242 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002243 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002244 if (old_page) {
2245 /*
2246 * Only after switching the pte to the new page may
2247 * we remove the mapcount here. Otherwise another
2248 * process may come and find the rmap count decremented
2249 * before the pte is switched to the new page, and
2250 * "reuse" the old page writing into it while our pte
2251 * here still points into it and can be read by other
2252 * threads.
2253 *
2254 * The critical issue is to order this
2255 * page_remove_rmap with the ptp_clear_flush above.
2256 * Those stores are ordered by (if nothing else,)
2257 * the barrier present in the atomic_add_negative
2258 * in page_remove_rmap.
2259 *
2260 * Then the TLB flush in ptep_clear_flush ensures that
2261 * no process can access the old page before the
2262 * decremented mapcount is visible. And the old page
2263 * cannot be reused until after the decremented
2264 * mapcount is visible. So transitively, TLBs to
2265 * old page will be flushed before it can be reused.
2266 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002267 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002268 }
2269
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 /* Free the old page.. */
2271 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002272 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002273 } else
Johannes Weiner00501b52014-08-08 14:19:20 -07002274 mem_cgroup_cancel_charge(new_page, memcg);
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002275
Haggai Eran6bdb9132012-10-08 16:33:35 -07002276 if (new_page)
2277 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002278unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002279 pte_unmap_unlock(page_table, ptl);
David Rientjes17569542012-11-16 14:14:48 -08002280 if (mmun_end > mmun_start)
Haggai Eran6bdb9132012-10-08 16:33:35 -07002281 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002282 if (old_page) {
2283 /*
2284 * Don't let another task, with possibly unlocked vma,
2285 * keep the mlocked page.
2286 */
2287 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2288 lock_page(old_page); /* LRU manipulation */
2289 munlock_vma_page(old_page);
2290 unlock_page(old_page);
2291 }
2292 page_cache_release(old_page);
2293 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002294 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002295oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002296 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002297oom:
Dominik Dingel66521d52012-12-12 13:52:37 -08002298 if (old_page)
Hugh Dickins920fc352005-11-21 21:32:17 -08002299 page_cache_release(old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 return VM_FAULT_OOM;
2301}
2302
Peter Zijlstra97a89412011-05-24 17:12:04 -07002303static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 unsigned long start_addr, unsigned long end_addr,
2305 struct zap_details *details)
2306{
Al Virof5cc4ee2012-03-05 14:14:20 -05002307 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308}
2309
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002310static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311 struct zap_details *details)
2312{
2313 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314 pgoff_t vba, vea, zba, zea;
2315
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002316 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318
2319 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002320 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2322 zba = details->first_index;
2323 if (zba < vba)
2324 zba = vba;
2325 zea = details->last_index;
2326 if (zea > vea)
2327 zea = vea;
2328
Peter Zijlstra97a89412011-05-24 17:12:04 -07002329 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2331 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002332 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333 }
2334}
2335
2336static inline void unmap_mapping_range_list(struct list_head *head,
2337 struct zap_details *details)
2338{
2339 struct vm_area_struct *vma;
2340
2341 /*
2342 * In nonlinear VMAs there is no correspondence between virtual address
2343 * offset and file offset. So we must perform an exhaustive search
2344 * across *all* the pages in each nonlinear VMA, not just the pages
2345 * whose virtual address lies outside the file truncation point.
2346 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002347 list_for_each_entry(vma, head, shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002349 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 }
2351}
2352
2353/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002354 * 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 -07002355 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 * @holebegin: byte in first page to unmap, relative to the start of
2357 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002358 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 * must keep the partial page. In contrast, we must get rid of
2360 * partial pages.
2361 * @holelen: size of prospective hole in bytes. This will be rounded
2362 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2363 * end of the file.
2364 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2365 * but 0 when invalidating pagecache, don't throw away private data.
2366 */
2367void unmap_mapping_range(struct address_space *mapping,
2368 loff_t const holebegin, loff_t const holelen, int even_cows)
2369{
2370 struct zap_details details;
2371 pgoff_t hba = holebegin >> PAGE_SHIFT;
2372 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2373
2374 /* Check for overflow. */
2375 if (sizeof(holelen) > sizeof(hlen)) {
2376 long long holeend =
2377 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2378 if (holeend & ~(long long)ULONG_MAX)
2379 hlen = ULONG_MAX - hba + 1;
2380 }
2381
2382 details.check_mapping = even_cows? NULL: mapping;
2383 details.nonlinear_vma = NULL;
2384 details.first_index = hba;
2385 details.last_index = hba + hlen - 1;
2386 if (details.last_index < details.first_index)
2387 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388
Peter Zijlstra97a89412011-05-24 17:12:04 -07002389
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002390 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002391 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2393 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2394 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002395 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396}
2397EXPORT_SYMBOL(unmap_mapping_range);
2398
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002400 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2401 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002402 * We return with pte unmapped and unlocked.
2403 *
2404 * We return with the mmap_sem locked or unlocked in the same cases
2405 * as does filemap_fault().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002407static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2408 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002409 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002411 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08002412 struct page *page, *swapcache;
Johannes Weiner00501b52014-08-08 14:19:20 -07002413 struct mem_cgroup *memcg;
Hugh Dickins65500d22005-10-29 18:15:59 -07002414 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002416 int locked;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002417 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002418 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002420 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002421 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002422
2423 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002424 if (unlikely(non_swap_entry(entry))) {
2425 if (is_migration_entry(entry)) {
2426 migration_entry_wait(mm, pmd, address);
2427 } else if (is_hwpoison_entry(entry)) {
2428 ret = VM_FAULT_HWPOISON;
2429 } else {
2430 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002431 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002432 }
Christoph Lameter06972122006-06-23 02:03:35 -07002433 goto out;
2434 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002435 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436 page = lookup_swap_cache(entry);
2437 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002438 page = swapin_readahead(entry,
2439 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440 if (!page) {
2441 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002442 * Back out if somebody else faulted in this pte
2443 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002445 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 if (likely(pte_same(*page_table, orig_pte)))
2447 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002448 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002449 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 }
2451
2452 /* Had to read the page from swap area: Major fault */
2453 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002454 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002455 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002456 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002457 /*
2458 * hwpoisoned dirty swapcache pages are kept for killing
2459 * owner processes (which may be unknown at hwpoison time)
2460 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002461 ret = VM_FAULT_HWPOISON;
2462 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08002463 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02002464 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 }
2466
Hugh Dickins56f31802013-02-22 16:36:10 -08002467 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002468 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002469
Balbir Singh20a10222007-11-14 17:00:33 -08002470 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002471 if (!locked) {
2472 ret |= VM_FAULT_RETRY;
2473 goto out_release;
2474 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002476 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002477 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2478 * release the swapcache from under us. The page pin, and pte_same
2479 * test below, are not enough to exclude that. Even if it is still
2480 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002481 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002482 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002483 goto out_page;
2484
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002485 page = ksm_might_need_to_copy(page, vma, address);
2486 if (unlikely(!page)) {
2487 ret = VM_FAULT_OOM;
2488 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002489 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08002490 }
2491
Johannes Weiner00501b52014-08-08 14:19:20 -07002492 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002493 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002494 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002495 }
2496
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002498 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002500 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002501 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002502 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002503
2504 if (unlikely(!PageUptodate(page))) {
2505 ret = VM_FAULT_SIGBUS;
2506 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 }
2508
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002509 /*
2510 * The page isn't present yet, go ahead with the fault.
2511 *
2512 * Be careful about the sequence of operations here.
2513 * To get its accounting right, reuse_swap_page() must be called
2514 * while the page is counted on swap but not yet in mapcount i.e.
2515 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2516 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002517 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002519 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08002520 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002522 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002524 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002525 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002526 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07002529 if (pte_swp_soft_dirty(orig_pte))
2530 pte = pte_mksoft_dirty(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 set_pte_at(mm, address, page_table, pte);
Johannes Weiner00501b52014-08-08 14:19:20 -07002532 if (page == swapcache) {
Johannes Weineraf347702013-02-22 16:32:20 -08002533 do_page_add_anon_rmap(page, vma, address, exclusive);
Johannes Weiner00501b52014-08-08 14:19:20 -07002534 mem_cgroup_commit_charge(page, memcg, true);
2535 } else { /* ksm created a completely new copy */
Hugh Dickins56f31802013-02-22 16:36:10 -08002536 page_add_new_anon_rmap(page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002537 mem_cgroup_commit_charge(page, memcg, false);
2538 lru_cache_add_active_or_unevictable(page, vma);
2539 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002541 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07002542 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002543 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002544 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002545 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002546 /*
2547 * Hold the lock to avoid the swap entry to be reused
2548 * until we take the PT lock for the pte_same() check
2549 * (to avoid false positives from pte_same). For
2550 * further safety release the lock after the swap_free
2551 * so that the swap count won't change under a
2552 * parallel locked swapcache.
2553 */
2554 unlock_page(swapcache);
2555 page_cache_release(swapcache);
2556 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002557
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002558 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08002559 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2560 if (ret & VM_FAULT_ERROR)
2561 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 goto out;
2563 }
2564
2565 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002566 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002567unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002568 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569out:
2570 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002571out_nomap:
Johannes Weiner00501b52014-08-08 14:19:20 -07002572 mem_cgroup_cancel_charge(page, memcg);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002573 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002574out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002575 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002576out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002577 page_cache_release(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002578 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002579 unlock_page(swapcache);
2580 page_cache_release(swapcache);
2581 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002582 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583}
2584
2585/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002586 * This is like a special single-page "expand_{down|up}wards()",
2587 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07002588 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07002589 */
2590static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2591{
2592 address &= PAGE_MASK;
2593 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002594 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07002595
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002596 /*
2597 * Is there a mapping abutting this one below?
2598 *
2599 * That's only ok if it's the same stack mapping
2600 * that has gotten split..
2601 */
2602 if (prev && prev->vm_end == address)
2603 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2604
Michal Hockod05f3162011-05-24 17:11:44 -07002605 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002606 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002607 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2608 struct vm_area_struct *next = vma->vm_next;
2609
2610 /* As VM_GROWSDOWN but s/below/above/ */
2611 if (next && next->vm_start == address + PAGE_SIZE)
2612 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2613
2614 expand_upwards(vma, address + PAGE_SIZE);
2615 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07002616 return 0;
2617}
2618
2619/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002620 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2621 * but allow concurrent faults), and pte mapped but not yet locked.
2622 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002624static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2625 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002626 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627{
Johannes Weiner00501b52014-08-08 14:19:20 -07002628 struct mem_cgroup *memcg;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002629 struct page *page;
2630 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632
Linus Torvalds11ac5522010-08-14 11:44:56 -07002633 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002634
Linus Torvalds11ac5522010-08-14 11:44:56 -07002635 /* Check if we need to add a guard page to the stack */
2636 if (check_stack_guard_page(vma, address) < 0)
2637 return VM_FAULT_SIGBUS;
2638
2639 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07002640 if (!(flags & FAULT_FLAG_WRITE)) {
2641 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2642 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07002643 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002644 if (!pte_none(*page_table))
2645 goto unlock;
2646 goto setpte;
2647 }
2648
Nick Piggin557ed1f2007-10-16 01:24:40 -07002649 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002650 if (unlikely(anon_vma_prepare(vma)))
2651 goto oom;
2652 page = alloc_zeroed_user_highpage_movable(vma, address);
2653 if (!page)
2654 goto oom;
Minchan Kim52f37622013-04-29 15:08:15 -07002655 /*
2656 * The memory barrier inside __SetPageUptodate makes sure that
2657 * preceeding stores to the page contents become visible before
2658 * the set_pte_at() write.
2659 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08002660 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661
Johannes Weiner00501b52014-08-08 14:19:20 -07002662 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002663 goto oom_free_page;
2664
Nick Piggin557ed1f2007-10-16 01:24:40 -07002665 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002666 if (vma->vm_flags & VM_WRITE)
2667 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002668
Nick Piggin557ed1f2007-10-16 01:24:40 -07002669 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07002670 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002671 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002672
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002673 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07002674 page_add_new_anon_rmap(page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002675 mem_cgroup_commit_charge(page, memcg, false);
2676 lru_cache_add_active_or_unevictable(page, vma);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002677setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07002678 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679
2680 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002681 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002682unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002683 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002684 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002685release:
Johannes Weiner00501b52014-08-08 14:19:20 -07002686 mem_cgroup_cancel_charge(page, memcg);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002687 page_cache_release(page);
2688 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002689oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002690 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002691oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692 return VM_FAULT_OOM;
2693}
2694
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002695/*
2696 * The mmap_sem must have been held on entry, and may have been
2697 * released depending on flags and vma->vm_ops->fault() return value.
2698 * See filemap_fault() and __lock_page_retry().
2699 */
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002700static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2701 pgoff_t pgoff, unsigned int flags, struct page **page)
2702{
2703 struct vm_fault vmf;
2704 int ret;
2705
2706 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2707 vmf.pgoff = pgoff;
2708 vmf.flags = flags;
2709 vmf.page = NULL;
2710
2711 ret = vma->vm_ops->fault(vma, &vmf);
2712 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2713 return ret;
2714
2715 if (unlikely(PageHWPoison(vmf.page))) {
2716 if (ret & VM_FAULT_LOCKED)
2717 unlock_page(vmf.page);
2718 page_cache_release(vmf.page);
2719 return VM_FAULT_HWPOISON;
2720 }
2721
2722 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2723 lock_page(vmf.page);
2724 else
2725 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2726
2727 *page = vmf.page;
2728 return ret;
2729}
2730
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002731/**
2732 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2733 *
2734 * @vma: virtual memory area
2735 * @address: user virtual address
2736 * @page: page to map
2737 * @pte: pointer to target page table entry
2738 * @write: true, if new entry is writable
2739 * @anon: true, if it's anonymous page
2740 *
2741 * Caller must hold page table lock relevant for @pte.
2742 *
2743 * Target users are page handler itself and implementations of
2744 * vm_ops->map_pages.
2745 */
2746void do_set_pte(struct vm_area_struct *vma, unsigned long address,
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002747 struct page *page, pte_t *pte, bool write, bool anon)
2748{
2749 pte_t entry;
2750
2751 flush_icache_page(vma, page);
2752 entry = mk_pte(page, vma->vm_page_prot);
2753 if (write)
2754 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2755 else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
Cyrill Gorcunov9aed8612014-08-06 16:07:05 -07002756 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002757 if (anon) {
2758 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2759 page_add_new_anon_rmap(page, vma, address);
2760 } else {
2761 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2762 page_add_file_rmap(page);
2763 }
2764 set_pte_at(vma->vm_mm, address, pte, entry);
2765
2766 /* no need to invalidate: a not-present page won't be cached */
2767 update_mmu_cache(vma, address, pte);
2768}
2769
Kirill A. Shutemov3a910532014-08-06 16:08:07 -07002770static unsigned long fault_around_bytes __read_mostly =
2771 rounddown_pow_of_two(65536);
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002772
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002773#ifdef CONFIG_DEBUG_FS
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002774static int fault_around_bytes_get(void *data, u64 *val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002775{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002776 *val = fault_around_bytes;
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002777 return 0;
2778}
2779
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002780/*
2781 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2782 * rounded down to nearest page order. It's what do_fault_around() expects to
2783 * see.
2784 */
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002785static int fault_around_bytes_set(void *data, u64 val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002786{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002787 if (val / PAGE_SIZE > PTRS_PER_PTE)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002788 return -EINVAL;
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002789 if (val > PAGE_SIZE)
2790 fault_around_bytes = rounddown_pow_of_two(val);
2791 else
2792 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002793 return 0;
2794}
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002795DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2796 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002797
2798static int __init fault_around_debugfs(void)
2799{
2800 void *ret;
2801
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002802 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2803 &fault_around_bytes_fops);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002804 if (!ret)
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002805 pr_warn("Failed to create fault_around_bytes in debugfs");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002806 return 0;
2807}
2808late_initcall(fault_around_debugfs);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002809#endif
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002810
Kirill A. Shutemov1fdb4122014-06-04 16:10:55 -07002811/*
2812 * do_fault_around() tries to map few pages around the fault address. The hope
2813 * is that the pages will be needed soon and this will lower the number of
2814 * faults to handle.
2815 *
2816 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2817 * not ready to be mapped: not up-to-date, locked, etc.
2818 *
2819 * This function is called with the page table lock taken. In the split ptlock
2820 * case the page table lock only protects only those entries which belong to
2821 * the page table corresponding to the fault address.
2822 *
2823 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2824 * only once.
2825 *
2826 * fault_around_pages() defines how many pages we'll try to map.
2827 * do_fault_around() expects it to return a power of two less than or equal to
2828 * PTRS_PER_PTE.
2829 *
2830 * The virtual address of the area that we map is naturally aligned to the
2831 * fault_around_pages() value (and therefore to page order). This way it's
2832 * easier to guarantee that we don't cross page table boundaries.
2833 */
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002834static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2835 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2836{
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002837 unsigned long start_addr, nr_pages, mask;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002838 pgoff_t max_pgoff;
2839 struct vm_fault vmf;
2840 int off;
2841
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002842 nr_pages = ACCESS_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2843 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2844
2845 start_addr = max(address & mask, vma->vm_start);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002846 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2847 pte -= off;
2848 pgoff -= off;
2849
2850 /*
2851 * max_pgoff is either end of page table or end of vma
Kirill A. Shutemov850e9c62014-06-04 16:10:45 -07002852 * or fault_around_pages() from pgoff, depending what is nearest.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002853 */
2854 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2855 PTRS_PER_PTE - 1;
2856 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002857 pgoff + nr_pages - 1);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002858
2859 /* Check if it makes any sense to call ->map_pages */
2860 while (!pte_none(*pte)) {
2861 if (++pgoff > max_pgoff)
2862 return;
2863 start_addr += PAGE_SIZE;
2864 if (start_addr >= vma->vm_end)
2865 return;
2866 pte++;
2867 }
2868
2869 vmf.virtual_address = (void __user *) start_addr;
2870 vmf.pte = pte;
2871 vmf.pgoff = pgoff;
2872 vmf.max_pgoff = max_pgoff;
2873 vmf.flags = flags;
2874 vma->vm_ops->map_pages(vma, &vmf);
2875}
2876
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002877static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2878 unsigned long address, pmd_t *pmd,
2879 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2880{
2881 struct page *fault_page;
2882 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002883 pte_t *pte;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002884 int ret = 0;
2885
2886 /*
2887 * Let's call ->map_pages() first and use ->fault() as fallback
2888 * if page by the offset is not ready to be mapped (cold cache or
2889 * something).
2890 */
Konstantin Khlebnikovc1186782014-07-23 14:00:08 -07002891 if (vma->vm_ops->map_pages && !(flags & FAULT_FLAG_NONLINEAR) &&
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002892 fault_around_bytes >> PAGE_SHIFT > 1) {
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002893 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2894 do_fault_around(vma, address, pte, pgoff, flags);
2895 if (!pte_same(*pte, orig_pte))
2896 goto unlock_out;
2897 pte_unmap_unlock(pte, ptl);
2898 }
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002899
2900 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2901 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2902 return ret;
2903
2904 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2905 if (unlikely(!pte_same(*pte, orig_pte))) {
2906 pte_unmap_unlock(pte, ptl);
2907 unlock_page(fault_page);
2908 page_cache_release(fault_page);
2909 return ret;
2910 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002911 do_set_pte(vma, address, fault_page, pte, false, false);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002912 unlock_page(fault_page);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002913unlock_out:
2914 pte_unmap_unlock(pte, ptl);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002915 return ret;
2916}
2917
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002918static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2919 unsigned long address, pmd_t *pmd,
2920 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2921{
2922 struct page *fault_page, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07002923 struct mem_cgroup *memcg;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002924 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002925 pte_t *pte;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002926 int ret;
2927
2928 if (unlikely(anon_vma_prepare(vma)))
2929 return VM_FAULT_OOM;
2930
2931 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2932 if (!new_page)
2933 return VM_FAULT_OOM;
2934
Johannes Weiner00501b52014-08-08 14:19:20 -07002935 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) {
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002936 page_cache_release(new_page);
2937 return VM_FAULT_OOM;
2938 }
2939
2940 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2941 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2942 goto uncharge_out;
2943
2944 copy_user_highpage(new_page, fault_page, address, vma);
2945 __SetPageUptodate(new_page);
2946
2947 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2948 if (unlikely(!pte_same(*pte, orig_pte))) {
2949 pte_unmap_unlock(pte, ptl);
2950 unlock_page(fault_page);
2951 page_cache_release(fault_page);
2952 goto uncharge_out;
2953 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002954 do_set_pte(vma, address, new_page, pte, true, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07002955 mem_cgroup_commit_charge(new_page, memcg, false);
2956 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002957 pte_unmap_unlock(pte, ptl);
2958 unlock_page(fault_page);
2959 page_cache_release(fault_page);
2960 return ret;
2961uncharge_out:
Johannes Weiner00501b52014-08-08 14:19:20 -07002962 mem_cgroup_cancel_charge(new_page, memcg);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002963 page_cache_release(new_page);
2964 return ret;
2965}
2966
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002967static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01002968 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07002969 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970{
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002971 struct page *fault_page;
2972 struct address_space *mapping;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002973 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002974 pte_t *pte;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002975 int dirtied = 0;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002976 int ret, tmp;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07002977
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002978 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2979 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002980 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981
2982 /*
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002983 * Check if the backing address space wants to know that the page is
2984 * about to become writable
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985 */
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002986 if (vma->vm_ops->page_mkwrite) {
2987 unlock_page(fault_page);
2988 tmp = do_page_mkwrite(vma, fault_page, address);
2989 if (unlikely(!tmp ||
2990 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2991 page_cache_release(fault_page);
2992 return tmp;
2993 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 }
2995
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002996 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2997 if (unlikely(!pte_same(*pte, orig_pte))) {
2998 pte_unmap_unlock(pte, ptl);
2999 unlock_page(fault_page);
3000 page_cache_release(fault_page);
3001 return ret;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003002 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003003 do_set_pte(vma, address, fault_page, pte, true, false);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003004 pte_unmap_unlock(pte, ptl);
3005
3006 if (set_page_dirty(fault_page))
3007 dirtied = 1;
3008 mapping = fault_page->mapping;
3009 unlock_page(fault_page);
3010 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3011 /*
3012 * Some device drivers do not set page.mapping but still
3013 * dirty their pages
3014 */
3015 balance_dirty_pages_ratelimited(mapping);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003016 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003017
3018 /* file_update_time outside page_lock */
3019 if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3020 file_update_time(vma->vm_file);
3021
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003022 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003023}
Nick Piggind00806b2007-07-19 01:46:57 -07003024
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003025/*
3026 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3027 * but allow concurrent faults).
3028 * The mmap_sem may have been released depending on flags and our
3029 * return value. See filemap_fault() and __lock_page_or_retry().
3030 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003031static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3032 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003033 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003034{
3035 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003036 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003037
Hugh Dickins16abfa02007-10-04 16:56:06 +01003038 pte_unmap(page_table);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003039 if (!(flags & FAULT_FLAG_WRITE))
3040 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3041 orig_pte);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003042 if (!(vma->vm_flags & VM_SHARED))
3043 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3044 orig_pte);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003045 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003046}
3047
Jes Sorensenf4b81802006-09-27 01:50:10 -07003048/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049 * Fault of a previously existing named mapping. Repopulate the pte
3050 * from the encoded file_pte if possible. This enables swappable
3051 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003052 *
3053 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3054 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003055 * We return with pte unmapped and unlocked.
3056 * The mmap_sem may have been released depending on flags and our
3057 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003058 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003059static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003060 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003061 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062{
Hugh Dickins65500d22005-10-29 18:15:59 -07003063 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003065 flags |= FAULT_FLAG_NONLINEAR;
3066
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003067 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003068 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003069
Hugh Dickins2509ef22009-01-06 14:40:10 -08003070 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003071 /*
3072 * Page table corrupted: show pte and kill process.
3073 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003074 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003075 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003076 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003077
3078 pgoff = pte_to_pgoff(orig_pte);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003079 if (!(flags & FAULT_FLAG_WRITE))
3080 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3081 orig_pte);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003082 if (!(vma->vm_flags & VM_SHARED))
3083 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3084 orig_pte);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003085 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086}
3087
Rashika Kheriab19a9932014-04-03 14:48:02 -07003088static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
Rik van Riel04bb2f92013-10-07 11:29:36 +01003089 unsigned long addr, int page_nid,
3090 int *flags)
Mel Gorman9532fec2012-11-15 01:24:32 +00003091{
3092 get_page(page);
3093
3094 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003095 if (page_nid == numa_node_id()) {
Mel Gorman9532fec2012-11-15 01:24:32 +00003096 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003097 *flags |= TNF_FAULT_LOCAL;
3098 }
Mel Gorman9532fec2012-11-15 01:24:32 +00003099
3100 return mpol_misplaced(page, vma, addr);
3101}
3102
Rashika Kheriab19a9932014-04-03 14:48:02 -07003103static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gormand10e63f2012-10-25 14:16:31 +02003104 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3105{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003106 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003107 spinlock_t *ptl;
Mel Gorman8191acb2013-10-07 11:28:45 +01003108 int page_nid = -1;
Peter Zijlstra90572892013-10-07 11:29:20 +01003109 int last_cpupid;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003110 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003111 bool migrated = false;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003112 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003113
3114 /*
3115 * The "pte" at this point cannot be used safely without
3116 * validation through pte_unmap_same(). It's of NUMA type but
3117 * the pfn may be screwed if the read is non atomic.
3118 *
3119 * ptep_modify_prot_start is not called as this is clearing
3120 * the _PAGE_NUMA bit and it is not really expected that there
3121 * would be concurrent hardware modifications to the PTE.
3122 */
3123 ptl = pte_lockptr(mm, pmd);
3124 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003125 if (unlikely(!pte_same(*ptep, pte))) {
3126 pte_unmap_unlock(ptep, ptl);
3127 goto out;
3128 }
3129
Mel Gormand10e63f2012-10-25 14:16:31 +02003130 pte = pte_mknonnuma(pte);
3131 set_pte_at(mm, addr, ptep, pte);
3132 update_mmu_cache(vma, addr, ptep);
3133
3134 page = vm_normal_page(vma, addr, pte);
3135 if (!page) {
3136 pte_unmap_unlock(ptep, ptl);
3137 return 0;
3138 }
Mel Gormana1a46182013-10-07 11:28:50 +01003139 BUG_ON(is_zero_pfn(page_to_pfn(page)));
Mel Gormand10e63f2012-10-25 14:16:31 +02003140
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003141 /*
3142 * Avoid grouping on DSO/COW pages in specific and RO pages
3143 * in general, RO pages shouldn't hurt as much anyway since
3144 * they can be in shared cache state.
3145 */
3146 if (!pte_write(pte))
3147 flags |= TNF_NO_GROUP;
3148
Rik van Rieldabe1d92013-10-07 11:29:34 +01003149 /*
3150 * Flag if the page is shared between multiple address spaces. This
3151 * is later used when determining whether to group tasks together
3152 */
3153 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3154 flags |= TNF_SHARED;
3155
Peter Zijlstra90572892013-10-07 11:29:20 +01003156 last_cpupid = page_cpupid_last(page);
Mel Gorman8191acb2013-10-07 11:28:45 +01003157 page_nid = page_to_nid(page);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003158 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003159 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003160 if (target_nid == -1) {
Mel Gorman4daae3b2012-11-02 11:33:45 +00003161 put_page(page);
3162 goto out;
3163 }
3164
3165 /* Migrate to the requested node */
Mel Gorman1bc115d2013-10-07 11:29:05 +01003166 migrated = migrate_misplaced_page(page, vma, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003167 if (migrated) {
Mel Gorman8191acb2013-10-07 11:28:45 +01003168 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003169 flags |= TNF_MIGRATED;
3170 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00003171
3172out:
Mel Gorman8191acb2013-10-07 11:28:45 +01003173 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003174 task_numa_fault(last_cpupid, page_nid, 1, flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003175 return 0;
3176}
3177
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178/*
3179 * These routines also need to handle stuff like marking pages dirty
3180 * and/or accessed for architectures that don't do it in hardware (most
3181 * RISC architectures). The early dirtying is also good on the i386.
3182 *
3183 * There is also a hook called "update_mmu_cache()" that architectures
3184 * with external mmu caches can use to update those (ie the Sparc or
3185 * PowerPC hashed page tables that act as extended TLBs).
3186 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003187 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3188 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003189 * We return with pte unmapped and unlocked.
3190 *
3191 * The mmap_sem may have been released depending on flags and our
3192 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003193 */
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003194static int handle_pte_fault(struct mm_struct *mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003195 struct vm_area_struct *vma, unsigned long address,
3196 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197{
3198 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003199 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200
Hugh Dickinsc0d73262014-08-06 16:05:08 -07003201 entry = ACCESS_ONCE(*pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003202 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003203 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003204 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003205 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003206 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003207 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003208 }
3209 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003210 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003211 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003212 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003213 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003214 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003215 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003216 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003217 }
3218
Mel Gormand10e63f2012-10-25 14:16:31 +02003219 if (pte_numa(entry))
3220 return do_numa_page(mm, vma, address, entry, pte, pmd);
3221
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003222 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003223 spin_lock(ptl);
3224 if (unlikely(!pte_same(*pte, entry)))
3225 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003226 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003228 return do_wp_page(mm, vma, address,
3229 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 entry = pte_mkdirty(entry);
3231 }
3232 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003233 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003234 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003235 } else {
3236 /*
3237 * This is needed only for protection faults but the arch code
3238 * is not yet telling us if this is a protection fault or not.
3239 * This still avoids useless tlb flushes for .text page faults
3240 * with threads.
3241 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003242 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003243 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003244 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003245unlock:
3246 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003247 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248}
3249
3250/*
3251 * By the time we get here, we already hold the mm semaphore
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003252 *
3253 * The mmap_sem may have been released depending on flags and our
3254 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003256static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3257 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003258{
3259 pgd_t *pgd;
3260 pud_t *pud;
3261 pmd_t *pmd;
3262 pte_t *pte;
3263
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003264 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003265 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003266
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 pud = pud_alloc(mm, pgd, address);
3269 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003270 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271 pmd = pmd_alloc(mm, pud, address);
3272 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003273 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003274 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003275 int ret = VM_FAULT_FALLBACK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003276 if (!vma->vm_ops)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003277 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3278 pmd, flags);
3279 if (!(ret & VM_FAULT_FALLBACK))
3280 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003281 } else {
3282 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003283 int ret;
3284
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003285 barrier();
3286 if (pmd_trans_huge(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003287 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3288
Linus Torvaldse53289c2013-01-09 08:36:54 -08003289 /*
3290 * If the pmd is splitting, return and retry the
3291 * the fault. Alternative: wait until the split
3292 * is done, and goto retry.
3293 */
3294 if (pmd_trans_splitting(orig_pmd))
3295 return 0;
3296
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003297 if (pmd_numa(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003298 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003299 orig_pmd, pmd);
3300
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003301 if (dirty && !pmd_write(orig_pmd)) {
David Rientjes1f1d06c2012-05-29 15:06:23 -07003302 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3303 orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003304 if (!(ret & VM_FAULT_FALLBACK))
3305 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003306 } else {
3307 huge_pmd_set_accessed(mm, vma, address, pmd,
3308 orig_pmd, dirty);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003309 return 0;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003310 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003311 }
3312 }
3313
3314 /*
3315 * Use __pte_alloc instead of pte_alloc_map, because we can't
3316 * run pte_offset_map on the pmd, if an huge pmd could
3317 * materialize from under us from a different thread.
3318 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003319 if (unlikely(pmd_none(*pmd)) &&
3320 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003321 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003322 /* if an huge pmd materialized from under us just retry later */
3323 if (unlikely(pmd_trans_huge(*pmd)))
3324 return 0;
3325 /*
3326 * A regular pmd is established and it can't morph into a huge pmd
3327 * from under us anymore at this point because we hold the mmap_sem
3328 * read mode and khugepaged takes it in write mode. So now it's
3329 * safe to run pte_offset_map().
3330 */
3331 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003332
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003333 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334}
3335
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003336/*
3337 * By the time we get here, we already hold the mm semaphore
3338 *
3339 * The mmap_sem may have been released depending on flags and our
3340 * return value. See filemap_fault() and __lock_page_or_retry().
3341 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003342int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3343 unsigned long address, unsigned int flags)
3344{
3345 int ret;
3346
3347 __set_current_state(TASK_RUNNING);
3348
3349 count_vm_event(PGFAULT);
3350 mem_cgroup_count_vm_event(mm, PGFAULT);
3351
3352 /* do counter updates before entering really critical section. */
3353 check_sync_rss_stat(current);
3354
3355 /*
3356 * Enable the memcg OOM handling for faults triggered in user
3357 * space. Kernel faults are handled more gracefully.
3358 */
3359 if (flags & FAULT_FLAG_USER)
Johannes Weiner49426422013-10-16 13:46:59 -07003360 mem_cgroup_oom_enable();
Johannes Weiner519e5242013-09-12 15:13:42 -07003361
3362 ret = __handle_mm_fault(mm, vma, address, flags);
3363
Johannes Weiner49426422013-10-16 13:46:59 -07003364 if (flags & FAULT_FLAG_USER) {
3365 mem_cgroup_oom_disable();
3366 /*
3367 * The task may have entered a memcg OOM situation but
3368 * if the allocation error was handled gracefully (no
3369 * VM_FAULT_OOM), there is no need to kill anything.
3370 * Just clean up the OOM state peacefully.
3371 */
3372 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3373 mem_cgroup_oom_synchronize(false);
3374 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003375
Johannes Weiner519e5242013-09-12 15:13:42 -07003376 return ret;
3377}
3378
Linus Torvalds1da177e2005-04-16 15:20:36 -07003379#ifndef __PAGETABLE_PUD_FOLDED
3380/*
3381 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003382 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003384int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003386 pud_t *new = pud_alloc_one(mm, address);
3387 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003388 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389
Nick Piggin362a61a2008-05-14 06:37:36 +02003390 smp_wmb(); /* See comment in __pte_alloc */
3391
Hugh Dickins872fec12005-10-29 18:16:21 -07003392 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003393 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003394 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003395 else
3396 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003397 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003398 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399}
3400#endif /* __PAGETABLE_PUD_FOLDED */
3401
3402#ifndef __PAGETABLE_PMD_FOLDED
3403/*
3404 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003405 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003407int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003409 pmd_t *new = pmd_alloc_one(mm, address);
3410 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003411 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412
Nick Piggin362a61a2008-05-14 06:37:36 +02003413 smp_wmb(); /* See comment in __pte_alloc */
3414
Hugh Dickins872fec12005-10-29 18:16:21 -07003415 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003417 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003418 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003419 else
3420 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003422 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003423 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003424 else
3425 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003427 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003428 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429}
3430#endif /* __PAGETABLE_PMD_FOLDED */
3431
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003432static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003433 pte_t **ptepp, spinlock_t **ptlp)
3434{
3435 pgd_t *pgd;
3436 pud_t *pud;
3437 pmd_t *pmd;
3438 pte_t *ptep;
3439
3440 pgd = pgd_offset(mm, address);
3441 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3442 goto out;
3443
3444 pud = pud_offset(pgd, address);
3445 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3446 goto out;
3447
3448 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003449 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003450 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3451 goto out;
3452
3453 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3454 if (pmd_huge(*pmd))
3455 goto out;
3456
3457 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3458 if (!ptep)
3459 goto out;
3460 if (!pte_present(*ptep))
3461 goto unlock;
3462 *ptepp = ptep;
3463 return 0;
3464unlock:
3465 pte_unmap_unlock(ptep, *ptlp);
3466out:
3467 return -EINVAL;
3468}
3469
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003470static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3471 pte_t **ptepp, spinlock_t **ptlp)
3472{
3473 int res;
3474
3475 /* (void) is needed to make gcc happy */
3476 (void) __cond_lock(*ptlp,
3477 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3478 return res;
3479}
3480
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003481/**
3482 * follow_pfn - look up PFN at a user virtual address
3483 * @vma: memory mapping
3484 * @address: user virtual address
3485 * @pfn: location to store found PFN
3486 *
3487 * Only IO mappings and raw PFN mappings are allowed.
3488 *
3489 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3490 */
3491int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3492 unsigned long *pfn)
3493{
3494 int ret = -EINVAL;
3495 spinlock_t *ptl;
3496 pte_t *ptep;
3497
3498 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3499 return ret;
3500
3501 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3502 if (ret)
3503 return ret;
3504 *pfn = pte_pfn(*ptep);
3505 pte_unmap_unlock(ptep, ptl);
3506 return 0;
3507}
3508EXPORT_SYMBOL(follow_pfn);
3509
Rik van Riel28b2ee22008-07-23 21:27:05 -07003510#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003511int follow_phys(struct vm_area_struct *vma,
3512 unsigned long address, unsigned int flags,
3513 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003514{
Johannes Weiner03668a42009-06-16 15:32:34 -07003515 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003516 pte_t *ptep, pte;
3517 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003518
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003519 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3520 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003521
Johannes Weiner03668a42009-06-16 15:32:34 -07003522 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003523 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003524 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003525
Rik van Riel28b2ee22008-07-23 21:27:05 -07003526 if ((flags & FOLL_WRITE) && !pte_write(pte))
3527 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003528
3529 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003530 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003531
Johannes Weiner03668a42009-06-16 15:32:34 -07003532 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003533unlock:
3534 pte_unmap_unlock(ptep, ptl);
3535out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003536 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003537}
3538
3539int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3540 void *buf, int len, int write)
3541{
3542 resource_size_t phys_addr;
3543 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003544 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003545 int offset = addr & (PAGE_SIZE-1);
3546
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003547 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003548 return -EINVAL;
3549
3550 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3551 if (write)
3552 memcpy_toio(maddr + offset, buf, len);
3553 else
3554 memcpy_fromio(buf, maddr + offset, len);
3555 iounmap(maddr);
3556
3557 return len;
3558}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02003559EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003560#endif
3561
David Howells0ec76a12006-09-27 01:50:15 -07003562/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003563 * Access another process' address space as given in mm. If non-NULL, use the
3564 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003565 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003566static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3567 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003568{
David Howells0ec76a12006-09-27 01:50:15 -07003569 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003570 void *old_buf = buf;
3571
David Howells0ec76a12006-09-27 01:50:15 -07003572 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003573 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003574 while (len) {
3575 int bytes, ret, offset;
3576 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003577 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003578
3579 ret = get_user_pages(tsk, mm, addr, 1,
3580 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003581 if (ret <= 0) {
Rik van Rieldbffcd02014-08-06 16:08:12 -07003582#ifndef CONFIG_HAVE_IOREMAP_PROT
3583 break;
3584#else
Rik van Riel28b2ee22008-07-23 21:27:05 -07003585 /*
3586 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3587 * we can access using slightly different code.
3588 */
Rik van Riel28b2ee22008-07-23 21:27:05 -07003589 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003590 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003591 break;
3592 if (vma->vm_ops && vma->vm_ops->access)
3593 ret = vma->vm_ops->access(vma, addr, buf,
3594 len, write);
3595 if (ret <= 0)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003596 break;
3597 bytes = ret;
Rik van Rieldbffcd02014-08-06 16:08:12 -07003598#endif
David Howells0ec76a12006-09-27 01:50:15 -07003599 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003600 bytes = len;
3601 offset = addr & (PAGE_SIZE-1);
3602 if (bytes > PAGE_SIZE-offset)
3603 bytes = PAGE_SIZE-offset;
3604
3605 maddr = kmap(page);
3606 if (write) {
3607 copy_to_user_page(vma, page, addr,
3608 maddr + offset, buf, bytes);
3609 set_page_dirty_lock(page);
3610 } else {
3611 copy_from_user_page(vma, page, addr,
3612 buf, maddr + offset, bytes);
3613 }
3614 kunmap(page);
3615 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003616 }
David Howells0ec76a12006-09-27 01:50:15 -07003617 len -= bytes;
3618 buf += bytes;
3619 addr += bytes;
3620 }
3621 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003622
3623 return buf - old_buf;
3624}
Andi Kleen03252912008-01-30 13:33:18 +01003625
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003626/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003627 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003628 * @mm: the mm_struct of the target address space
3629 * @addr: start address to access
3630 * @buf: source or destination buffer
3631 * @len: number of bytes to transfer
3632 * @write: whether the access is a write
3633 *
3634 * The caller must hold a reference on @mm.
3635 */
3636int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3637 void *buf, int len, int write)
3638{
3639 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3640}
3641
Andi Kleen03252912008-01-30 13:33:18 +01003642/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003643 * Access another process' address space.
3644 * Source/target buffer must be kernel space,
3645 * Do not walk the page table directly, use get_user_pages
3646 */
3647int access_process_vm(struct task_struct *tsk, unsigned long addr,
3648 void *buf, int len, int write)
3649{
3650 struct mm_struct *mm;
3651 int ret;
3652
3653 mm = get_task_mm(tsk);
3654 if (!mm)
3655 return 0;
3656
3657 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3658 mmput(mm);
3659
3660 return ret;
3661}
3662
Andi Kleen03252912008-01-30 13:33:18 +01003663/*
3664 * Print the name of a VMA.
3665 */
3666void print_vma_addr(char *prefix, unsigned long ip)
3667{
3668 struct mm_struct *mm = current->mm;
3669 struct vm_area_struct *vma;
3670
Ingo Molnare8bff742008-02-13 20:21:06 +01003671 /*
3672 * Do not print if we are in atomic
3673 * contexts (in exception stacks, etc.):
3674 */
3675 if (preempt_count())
3676 return;
3677
Andi Kleen03252912008-01-30 13:33:18 +01003678 down_read(&mm->mmap_sem);
3679 vma = find_vma(mm, ip);
3680 if (vma && vma->vm_file) {
3681 struct file *f = vma->vm_file;
3682 char *buf = (char *)__get_free_page(GFP_KERNEL);
3683 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003684 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01003685
Jan Blunckcf28b482008-02-14 19:38:44 -08003686 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003687 if (IS_ERR(p))
3688 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003689 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01003690 vma->vm_start,
3691 vma->vm_end - vma->vm_start);
3692 free_page((unsigned long)buf);
3693 }
3694 }
Jeff Liu51a07e52012-07-31 16:43:18 -07003695 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01003696}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003697
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003698#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003699void might_fault(void)
3700{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003701 /*
3702 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3703 * holding the mmap_sem, this is safe because kernel memory doesn't
3704 * get paged out, therefore we'll never actually fault, and the
3705 * below annotations will generate false positives.
3706 */
3707 if (segment_eq(get_fs(), KERNEL_DS))
3708 return;
3709
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003710 /*
3711 * it would be nicer only to annotate paths which are not under
3712 * pagefault_disable, however that requires a larger audit and
3713 * providing helpers like get_user_atomic.
3714 */
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003715 if (in_atomic())
3716 return;
3717
3718 __might_sleep(__FILE__, __LINE__, 0);
3719
3720 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003721 might_lock_read(&current->mm->mmap_sem);
3722}
3723EXPORT_SYMBOL(might_fault);
3724#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003725
3726#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3727static void clear_gigantic_page(struct page *page,
3728 unsigned long addr,
3729 unsigned int pages_per_huge_page)
3730{
3731 int i;
3732 struct page *p = page;
3733
3734 might_sleep();
3735 for (i = 0; i < pages_per_huge_page;
3736 i++, p = mem_map_next(p, page, i)) {
3737 cond_resched();
3738 clear_user_highpage(p, addr + i * PAGE_SIZE);
3739 }
3740}
3741void clear_huge_page(struct page *page,
3742 unsigned long addr, unsigned int pages_per_huge_page)
3743{
3744 int i;
3745
3746 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3747 clear_gigantic_page(page, addr, pages_per_huge_page);
3748 return;
3749 }
3750
3751 might_sleep();
3752 for (i = 0; i < pages_per_huge_page; i++) {
3753 cond_resched();
3754 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3755 }
3756}
3757
3758static void copy_user_gigantic_page(struct page *dst, struct page *src,
3759 unsigned long addr,
3760 struct vm_area_struct *vma,
3761 unsigned int pages_per_huge_page)
3762{
3763 int i;
3764 struct page *dst_base = dst;
3765 struct page *src_base = src;
3766
3767 for (i = 0; i < pages_per_huge_page; ) {
3768 cond_resched();
3769 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3770
3771 i++;
3772 dst = mem_map_next(dst, dst_base, i);
3773 src = mem_map_next(src, src_base, i);
3774 }
3775}
3776
3777void copy_user_huge_page(struct page *dst, struct page *src,
3778 unsigned long addr, struct vm_area_struct *vma,
3779 unsigned int pages_per_huge_page)
3780{
3781 int i;
3782
3783 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3784 copy_user_gigantic_page(dst, src, addr, vma,
3785 pages_per_huge_page);
3786 return;
3787 }
3788
3789 might_sleep();
3790 for (i = 0; i < pages_per_huge_page; i++) {
3791 cond_resched();
3792 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3793 }
3794}
3795#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003796
Olof Johansson40b64ac2013-12-20 14:28:05 -08003797#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003798
3799static struct kmem_cache *page_ptl_cachep;
3800
3801void __init ptlock_cache_init(void)
3802{
3803 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3804 SLAB_PANIC, NULL);
3805}
3806
Peter Zijlstra539edb52013-11-14 14:31:52 -08003807bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003808{
3809 spinlock_t *ptl;
3810
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003811 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003812 if (!ptl)
3813 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08003814 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003815 return true;
3816}
3817
Peter Zijlstra539edb52013-11-14 14:31:52 -08003818void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003819{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003820 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003821}
3822#endif