blob: 3e503831e042a6aa7b96d2608ecb570dfffa0aa7 [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
Ard Biesheuvel0b700682014-09-12 22:17:23 +0200121EXPORT_SYMBOL(zero_pfn);
122
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700123/*
124 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
125 */
126static int __init init_zero_pfn(void)
127{
128 zero_pfn = page_to_pfn(ZERO_PAGE(0));
129 return 0;
130}
131core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100132
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800133
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800134#if defined(SPLIT_RSS_COUNTING)
135
David Rientjesea48cf72012-03-21 16:34:13 -0700136void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800137{
138 int i;
139
140 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700141 if (current->rss_stat.count[i]) {
142 add_mm_counter(mm, i, current->rss_stat.count[i]);
143 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800144 }
145 }
David Rientjes05af2e12012-03-21 16:34:13 -0700146 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800147}
148
149static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
150{
151 struct task_struct *task = current;
152
153 if (likely(task->mm == mm))
154 task->rss_stat.count[member] += val;
155 else
156 add_mm_counter(mm, member, val);
157}
158#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
159#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
160
161/* sync counter once per 64 page faults */
162#define TASK_RSS_EVENTS_THRESH (64)
163static void check_sync_rss_stat(struct task_struct *task)
164{
165 if (unlikely(task != current))
166 return;
167 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700168 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800169}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700170#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800171
172#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
173#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
174
175static void check_sync_rss_stat(struct task_struct *task)
176{
177}
178
Peter Zijlstra9547d012011-05-24 17:12:14 -0700179#endif /* SPLIT_RSS_COUNTING */
180
181#ifdef HAVE_GENERIC_MMU_GATHER
182
183static int tlb_next_batch(struct mmu_gather *tlb)
184{
185 struct mmu_gather_batch *batch;
186
187 batch = tlb->active;
188 if (batch->next) {
189 tlb->active = batch->next;
190 return 1;
191 }
192
Michal Hocko53a59fc2013-01-04 15:35:12 -0800193 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
194 return 0;
195
Peter Zijlstra9547d012011-05-24 17:12:14 -0700196 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
197 if (!batch)
198 return 0;
199
Michal Hocko53a59fc2013-01-04 15:35:12 -0800200 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700201 batch->next = NULL;
202 batch->nr = 0;
203 batch->max = MAX_GATHER_BATCH;
204
205 tlb->active->next = batch;
206 tlb->active = batch;
207
208 return 1;
209}
210
211/* tlb_gather_mmu
212 * Called to initialize an (on-stack) mmu_gather structure for page-table
213 * tear-down from @mm. The @fullmm argument is used when @mm is without
214 * users and we're going to destroy the full address space (exit/execve).
215 */
Linus Torvalds2b047252013-08-15 11:42:25 -0700216void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700217{
218 tlb->mm = mm;
219
Linus Torvalds2b047252013-08-15 11:42:25 -0700220 /* Is it from 0 to ~0? */
221 tlb->fullmm = !(start | (end+1));
Dave Hansen1de14c32013-04-12 16:23:54 -0700222 tlb->need_flush_all = 0;
Linus Torvalds2b047252013-08-15 11:42:25 -0700223 tlb->start = start;
224 tlb->end = end;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700225 tlb->need_flush = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700226 tlb->local.next = NULL;
227 tlb->local.nr = 0;
228 tlb->local.max = ARRAY_SIZE(tlb->__pages);
229 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800230 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700231
232#ifdef CONFIG_HAVE_RCU_TABLE_FREE
233 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800234#endif
Peter Zijlstra9547d012011-05-24 17:12:14 -0700235}
236
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700237static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700238{
Peter Zijlstra9547d012011-05-24 17:12:14 -0700239 tlb->need_flush = 0;
240 tlb_flush(tlb);
241#ifdef CONFIG_HAVE_RCU_TABLE_FREE
242 tlb_table_flush(tlb);
243#endif
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700244}
245
246static void tlb_flush_mmu_free(struct mmu_gather *tlb)
247{
248 struct mmu_gather_batch *batch;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700249
Peter Zijlstra9547d012011-05-24 17:12:14 -0700250 for (batch = &tlb->local; batch; batch = batch->next) {
251 free_pages_and_swap_cache(batch->pages, batch->nr);
252 batch->nr = 0;
253 }
254 tlb->active = &tlb->local;
255}
256
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700257void tlb_flush_mmu(struct mmu_gather *tlb)
258{
259 if (!tlb->need_flush)
260 return;
261 tlb_flush_mmu_tlbonly(tlb);
262 tlb_flush_mmu_free(tlb);
263}
264
Peter Zijlstra9547d012011-05-24 17:12:14 -0700265/* tlb_finish_mmu
266 * Called at the end of the shootdown operation to free up any resources
267 * that were required.
268 */
269void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
270{
271 struct mmu_gather_batch *batch, *next;
272
273 tlb_flush_mmu(tlb);
274
275 /* keep the page table cache within bounds */
276 check_pgt_cache();
277
278 for (batch = tlb->local.next; batch; batch = next) {
279 next = batch->next;
280 free_pages((unsigned long)batch, 0);
281 }
282 tlb->local.next = NULL;
283}
284
285/* __tlb_remove_page
286 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
287 * handling the additional races in SMP caused by other CPUs caching valid
288 * mappings in their TLBs. Returns the number of free page slots left.
289 * When out of page slots we must call tlb_flush_mmu().
290 */
291int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
292{
293 struct mmu_gather_batch *batch;
294
Shaohua Lif21760b2012-01-12 17:19:16 -0800295 VM_BUG_ON(!tlb->need_flush);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700296
Peter Zijlstra9547d012011-05-24 17:12:14 -0700297 batch = tlb->active;
298 batch->pages[batch->nr++] = page;
299 if (batch->nr == batch->max) {
300 if (!tlb_next_batch(tlb))
301 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700302 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700303 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800304 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700305
306 return batch->max - batch->nr;
307}
308
309#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800310
Peter Zijlstra26723912011-05-24 17:12:00 -0700311#ifdef CONFIG_HAVE_RCU_TABLE_FREE
312
313/*
314 * See the comment near struct mmu_table_batch.
315 */
316
317static void tlb_remove_table_smp_sync(void *arg)
318{
319 /* Simply deliver the interrupt */
320}
321
322static void tlb_remove_table_one(void *table)
323{
324 /*
325 * This isn't an RCU grace period and hence the page-tables cannot be
326 * assumed to be actually RCU-freed.
327 *
328 * It is however sufficient for software page-table walkers that rely on
329 * IRQ disabling. See the comment near struct mmu_table_batch.
330 */
331 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
332 __tlb_remove_table(table);
333}
334
335static void tlb_remove_table_rcu(struct rcu_head *head)
336{
337 struct mmu_table_batch *batch;
338 int i;
339
340 batch = container_of(head, struct mmu_table_batch, rcu);
341
342 for (i = 0; i < batch->nr; i++)
343 __tlb_remove_table(batch->tables[i]);
344
345 free_page((unsigned long)batch);
346}
347
348void tlb_table_flush(struct mmu_gather *tlb)
349{
350 struct mmu_table_batch **batch = &tlb->batch;
351
352 if (*batch) {
353 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
354 *batch = NULL;
355 }
356}
357
358void tlb_remove_table(struct mmu_gather *tlb, void *table)
359{
360 struct mmu_table_batch **batch = &tlb->batch;
361
362 tlb->need_flush = 1;
363
364 /*
365 * When there's less then two users of this mm there cannot be a
366 * concurrent page-table walk.
367 */
368 if (atomic_read(&tlb->mm->mm_users) < 2) {
369 __tlb_remove_table(table);
370 return;
371 }
372
373 if (*batch == NULL) {
374 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
375 if (*batch == NULL) {
376 tlb_remove_table_one(table);
377 return;
378 }
379 (*batch)->nr = 0;
380 }
381 (*batch)->tables[(*batch)->nr++] = table;
382 if ((*batch)->nr == MAX_TABLE_BATCH)
383 tlb_table_flush(tlb);
384}
385
Peter Zijlstra9547d012011-05-24 17:12:14 -0700386#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700387
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389 * Note: this doesn't free the actual pages themselves. That
390 * has been handled earlier when unmapping all the memory regions.
391 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000392static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
393 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800395 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700396 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000397 pte_free_tlb(tlb, token, addr);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800398 atomic_long_dec(&tlb->mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399}
400
Hugh Dickinse0da3822005-04-19 13:29:15 -0700401static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
402 unsigned long addr, unsigned long end,
403 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404{
405 pmd_t *pmd;
406 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700407 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408
Hugh Dickinse0da3822005-04-19 13:29:15 -0700409 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 do {
412 next = pmd_addr_end(addr, end);
413 if (pmd_none_or_clear_bad(pmd))
414 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000415 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 } while (pmd++, addr = next, addr != end);
417
Hugh Dickinse0da3822005-04-19 13:29:15 -0700418 start &= PUD_MASK;
419 if (start < floor)
420 return;
421 if (ceiling) {
422 ceiling &= PUD_MASK;
423 if (!ceiling)
424 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700426 if (end - 1 > ceiling - 1)
427 return;
428
429 pmd = pmd_offset(pud, start);
430 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000431 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432}
433
Hugh Dickinse0da3822005-04-19 13:29:15 -0700434static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
435 unsigned long addr, unsigned long end,
436 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
438 pud_t *pud;
439 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700440 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441
Hugh Dickinse0da3822005-04-19 13:29:15 -0700442 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 do {
445 next = pud_addr_end(addr, end);
446 if (pud_none_or_clear_bad(pud))
447 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700448 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 } while (pud++, addr = next, addr != end);
450
Hugh Dickinse0da3822005-04-19 13:29:15 -0700451 start &= PGDIR_MASK;
452 if (start < floor)
453 return;
454 if (ceiling) {
455 ceiling &= PGDIR_MASK;
456 if (!ceiling)
457 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700459 if (end - 1 > ceiling - 1)
460 return;
461
462 pud = pud_offset(pgd, start);
463 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000464 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465}
466
467/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700468 * This function frees user-level page tables of a process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 */
Jan Beulich42b77722008-07-23 21:27:10 -0700470void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700471 unsigned long addr, unsigned long end,
472 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473{
474 pgd_t *pgd;
475 unsigned long next;
476
Hugh Dickinse0da3822005-04-19 13:29:15 -0700477 /*
478 * The next few lines have given us lots of grief...
479 *
480 * Why are we testing PMD* at this top level? Because often
481 * there will be no work to do at all, and we'd prefer not to
482 * go all the way down to the bottom just to discover that.
483 *
484 * Why all these "- 1"s? Because 0 represents both the bottom
485 * of the address space and the top of it (using -1 for the
486 * top wouldn't help much: the masks would do the wrong thing).
487 * The rule is that addr 0 and floor 0 refer to the bottom of
488 * the address space, but end 0 and ceiling 0 refer to the top
489 * Comparisons need to use "end - 1" and "ceiling - 1" (though
490 * that end 0 case should be mythical).
491 *
492 * Wherever addr is brought up or ceiling brought down, we must
493 * be careful to reject "the opposite 0" before it confuses the
494 * subsequent tests. But what about where end is brought down
495 * by PMD_SIZE below? no, end can't go down to 0 there.
496 *
497 * Whereas we round start (addr) and ceiling down, by different
498 * masks at different levels, in order to test whether a table
499 * now has no other vmas using it, so can be freed, we don't
500 * bother to round floor or end up - the tests don't need that.
501 */
502
503 addr &= PMD_MASK;
504 if (addr < floor) {
505 addr += PMD_SIZE;
506 if (!addr)
507 return;
508 }
509 if (ceiling) {
510 ceiling &= PMD_MASK;
511 if (!ceiling)
512 return;
513 }
514 if (end - 1 > ceiling - 1)
515 end -= PMD_SIZE;
516 if (addr > end - 1)
517 return;
518
Jan Beulich42b77722008-07-23 21:27:10 -0700519 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 do {
521 next = pgd_addr_end(addr, end);
522 if (pgd_none_or_clear_bad(pgd))
523 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700524 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700526}
527
Jan Beulich42b77722008-07-23 21:27:10 -0700528void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700529 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700530{
531 while (vma) {
532 struct vm_area_struct *next = vma->vm_next;
533 unsigned long addr = vma->vm_start;
534
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700535 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000536 * Hide vma from rmap and truncate_pagecache before freeing
537 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700538 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800539 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700540 unlink_file_vma(vma);
541
David Gibson9da61ae2006-03-22 00:08:57 -0800542 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700543 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700544 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700545 } else {
546 /*
547 * Optimization: gather nearby vmas into one call down
548 */
549 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800550 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700551 vma = next;
552 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800553 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700554 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700555 }
556 free_pgd_range(tlb, addr, vma->vm_end,
557 floor, next? next->vm_start: ceiling);
558 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700559 vma = next;
560 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561}
562
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800563int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
564 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800566 spinlock_t *ptl;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800567 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800568 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700569 if (!new)
570 return -ENOMEM;
571
Nick Piggin362a61a2008-05-14 06:37:36 +0200572 /*
573 * Ensure all pte setup (eg. pte page lock and page clearing) are
574 * visible before the pte is made visible to other CPUs by being
575 * put into page tables.
576 *
577 * The other side of the story is the pointer chasing in the page
578 * table walking code (when walking the page table without locking;
579 * ie. most of the time). Fortunately, these data accesses consist
580 * of a chain of data-dependent loads, meaning most CPUs (alpha
581 * being the notable exception) will already guarantee loads are
582 * seen in-order. See the alpha page table accessors for the
583 * smp_read_barrier_depends() barriers in page table walking code.
584 */
585 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
586
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800587 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800588 wait_split_huge_page = 0;
589 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800590 atomic_long_inc(&mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800592 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800593 } else if (unlikely(pmd_trans_splitting(*pmd)))
594 wait_split_huge_page = 1;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800595 spin_unlock(ptl);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800596 if (new)
597 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800598 if (wait_split_huge_page)
599 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700600 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601}
602
Hugh Dickins1bb36302005-10-29 18:16:22 -0700603int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700605 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
606 if (!new)
607 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608
Nick Piggin362a61a2008-05-14 06:37:36 +0200609 smp_wmb(); /* See comment in __pte_alloc */
610
Hugh Dickins1bb36302005-10-29 18:16:22 -0700611 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800612 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700613 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800614 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800615 } else
616 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700617 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800618 if (new)
619 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700620 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621}
622
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800623static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700624{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800625 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
626}
627
628static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
629{
630 int i;
631
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800632 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700633 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800634 for (i = 0; i < NR_MM_COUNTERS; i++)
635 if (rss[i])
636 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700637}
638
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800640 * This function is called to print an error when a bad pte
641 * is found. For example, we might have a PFN-mapped pte in
642 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700643 *
644 * The calling function must still handle the error.
645 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800646static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
647 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700648{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800649 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
650 pud_t *pud = pud_offset(pgd, addr);
651 pmd_t *pmd = pmd_offset(pud, addr);
652 struct address_space *mapping;
653 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800654 static unsigned long resume;
655 static unsigned long nr_shown;
656 static unsigned long nr_unshown;
657
658 /*
659 * Allow a burst of 60 reports, then keep quiet for that minute;
660 * or allow a steady drip of one report per second.
661 */
662 if (nr_shown == 60) {
663 if (time_before(jiffies, resume)) {
664 nr_unshown++;
665 return;
666 }
667 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800668 printk(KERN_ALERT
669 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800670 nr_unshown);
671 nr_unshown = 0;
672 }
673 nr_shown = 0;
674 }
675 if (nr_shown++ == 0)
676 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800677
678 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
679 index = linear_page_index(vma, addr);
680
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800681 printk(KERN_ALERT
682 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800683 current->comm,
684 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800685 if (page)
Dave Hansenf0b791a2014-01-23 15:52:49 -0800686 dump_page(page, "bad pte");
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800687 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800688 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
689 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
690 /*
691 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
692 */
693 if (vma->vm_ops)
Joe Perches071361d2012-12-12 10:19:12 -0800694 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
695 vma->vm_ops->fault);
Al Viro72c2d532013-09-22 16:27:52 -0400696 if (vma->vm_file)
Joe Perches071361d2012-12-12 10:19:12 -0800697 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
698 vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700699 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030700 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700701}
702
703/*
Nick Piggin7e675132008-04-28 02:13:00 -0700704 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800705 *
Nick Piggin7e675132008-04-28 02:13:00 -0700706 * "Special" mappings do not wish to be associated with a "struct page" (either
707 * it doesn't exist, or it exists but they don't want to touch it). In this
708 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700709 *
Nick Piggin7e675132008-04-28 02:13:00 -0700710 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
711 * pte bit, in which case this function is trivial. Secondly, an architecture
712 * may not have a spare pte bit, which requires a more complicated scheme,
713 * described below.
714 *
715 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
716 * special mapping (even if there are underlying and valid "struct pages").
717 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800718 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700719 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
720 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700721 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
722 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800723 *
724 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
725 *
Nick Piggin7e675132008-04-28 02:13:00 -0700726 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700727 *
Nick Piggin7e675132008-04-28 02:13:00 -0700728 * This restricts such mappings to be a linear translation from virtual address
729 * to pfn. To get around this restriction, we allow arbitrary mappings so long
730 * as the vma is not a COW mapping; in that case, we know that all ptes are
731 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700732 *
733 *
Nick Piggin7e675132008-04-28 02:13:00 -0700734 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
735 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700736 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
737 * page" backing, however the difference is that _all_ pages with a struct
738 * page (that is, those where pfn_valid is true) are refcounted and considered
739 * normal pages by the VM. The disadvantage is that pages are refcounted
740 * (which can be slower and simply not an option for some PFNMAP users). The
741 * advantage is that we don't have to follow the strict linearity rule of
742 * PFNMAP mappings in order to support COWable mappings.
743 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800744 */
Nick Piggin7e675132008-04-28 02:13:00 -0700745#ifdef __HAVE_ARCH_PTE_SPECIAL
746# define HAVE_PTE_SPECIAL 1
747#else
748# define HAVE_PTE_SPECIAL 0
749#endif
750struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
751 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800752{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800753 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700754
755 if (HAVE_PTE_SPECIAL) {
Hugh Dickinsb38af472014-08-29 15:18:44 -0700756 if (likely(!pte_special(pte)))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800757 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700758 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
759 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700760 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800761 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700762 return NULL;
763 }
764
765 /* !HAVE_PTE_SPECIAL case follows: */
766
Jared Hulbertb379d792008-04-28 02:12:58 -0700767 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
768 if (vma->vm_flags & VM_MIXEDMAP) {
769 if (!pfn_valid(pfn))
770 return NULL;
771 goto out;
772 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700773 unsigned long off;
774 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700775 if (pfn == vma->vm_pgoff + off)
776 return NULL;
777 if (!is_cow_mapping(vma->vm_flags))
778 return NULL;
779 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800780 }
781
Hugh Dickinsb38af472014-08-29 15:18:44 -0700782 if (is_zero_pfn(pfn))
783 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800784check_pfn:
785 if (unlikely(pfn > highest_memmap_pfn)) {
786 print_bad_pte(vma, addr, pte, NULL);
787 return NULL;
788 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800789
790 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700791 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700792 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800793 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700794out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800795 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800796}
797
798/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 * copy one vm_area from one task to the other. Assumes the page tables
800 * already present in the new task to be cleared in the whole range
801 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 */
803
Hugh Dickins570a335b2009-12-14 17:58:46 -0800804static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700806 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700807 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808{
Nick Pigginb5810032005-10-29 18:16:12 -0700809 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 pte_t pte = *src_pte;
811 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812
813 /* pte contains position in swap or file, so copy. */
814 if (unlikely(!pte_present(pte))) {
815 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700816 swp_entry_t entry = pte_to_swp_entry(pte);
817
Hugh Dickins570a335b2009-12-14 17:58:46 -0800818 if (swap_duplicate(entry) < 0)
819 return entry.val;
820
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 /* make sure dst_mm is on swapoff's mmlist. */
822 if (unlikely(list_empty(&dst_mm->mmlist))) {
823 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700824 if (list_empty(&dst_mm->mmlist))
825 list_add(&dst_mm->mmlist,
826 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 spin_unlock(&mmlist_lock);
828 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800829 if (likely(!non_swap_entry(entry)))
830 rss[MM_SWAPENTS]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800831 else if (is_migration_entry(entry)) {
832 page = migration_entry_to_page(entry);
833
834 if (PageAnon(page))
835 rss[MM_ANONPAGES]++;
836 else
837 rss[MM_FILEPAGES]++;
838
839 if (is_write_migration_entry(entry) &&
840 is_cow_mapping(vm_flags)) {
841 /*
842 * COW mappings require pages in both
843 * parent and child to be set to read.
844 */
845 make_migration_entry_read(&entry);
846 pte = swp_entry_to_pte(entry);
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700847 if (pte_swp_soft_dirty(*src_pte))
848 pte = pte_swp_mksoft_dirty(pte);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800849 set_pte_at(src_mm, addr, src_pte, pte);
850 }
Christoph Lameter06972122006-06-23 02:03:35 -0700851 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700853 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 }
855
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 /*
857 * If it's a COW mapping, write protect it both
858 * in the parent and the child
859 */
Linus Torvalds67121172005-12-11 20:38:17 -0800860 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700862 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 }
864
865 /*
866 * If it's a shared mapping, mark it clean in
867 * the child
868 */
869 if (vm_flags & VM_SHARED)
870 pte = pte_mkclean(pte);
871 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800872
873 page = vm_normal_page(vma, addr, pte);
874 if (page) {
875 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700876 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800877 if (PageAnon(page))
878 rss[MM_ANONPAGES]++;
879 else
880 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800881 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700882
883out_set_pte:
884 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800885 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886}
887
Jerome Marchand21bda262014-08-06 16:06:56 -0700888static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800889 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
890 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700892 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700894 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700895 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800896 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800897 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898
899again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800900 init_rss_vec(rss);
901
Hugh Dickinsc74df322005-10-29 18:16:23 -0700902 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 if (!dst_pte)
904 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700905 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700906 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700907 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700908 orig_src_pte = src_pte;
909 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700910 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 do {
913 /*
914 * We are holding two locks at this point - either of them
915 * could generate latencies in another task on another CPU.
916 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700917 if (progress >= 32) {
918 progress = 0;
919 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100920 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700921 break;
922 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 if (pte_none(*src_pte)) {
924 progress++;
925 continue;
926 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800927 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
928 vma, addr, rss);
929 if (entry.val)
930 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 progress += 8;
932 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700934 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700935 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700936 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800937 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700938 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700939 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800940
941 if (entry.val) {
942 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
943 return -ENOMEM;
944 progress = 0;
945 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 if (addr != end)
947 goto again;
948 return 0;
949}
950
951static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
952 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
953 unsigned long addr, unsigned long end)
954{
955 pmd_t *src_pmd, *dst_pmd;
956 unsigned long next;
957
958 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
959 if (!dst_pmd)
960 return -ENOMEM;
961 src_pmd = pmd_offset(src_pud, addr);
962 do {
963 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800964 if (pmd_trans_huge(*src_pmd)) {
965 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800966 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800967 err = copy_huge_pmd(dst_mm, src_mm,
968 dst_pmd, src_pmd, addr, vma);
969 if (err == -ENOMEM)
970 return -ENOMEM;
971 if (!err)
972 continue;
973 /* fall through */
974 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 if (pmd_none_or_clear_bad(src_pmd))
976 continue;
977 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
978 vma, addr, next))
979 return -ENOMEM;
980 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
981 return 0;
982}
983
984static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
985 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
986 unsigned long addr, unsigned long end)
987{
988 pud_t *src_pud, *dst_pud;
989 unsigned long next;
990
991 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
992 if (!dst_pud)
993 return -ENOMEM;
994 src_pud = pud_offset(src_pgd, addr);
995 do {
996 next = pud_addr_end(addr, end);
997 if (pud_none_or_clear_bad(src_pud))
998 continue;
999 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1000 vma, addr, next))
1001 return -ENOMEM;
1002 } while (dst_pud++, src_pud++, addr = next, addr != end);
1003 return 0;
1004}
1005
1006int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1007 struct vm_area_struct *vma)
1008{
1009 pgd_t *src_pgd, *dst_pgd;
1010 unsigned long next;
1011 unsigned long addr = vma->vm_start;
1012 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001013 unsigned long mmun_start; /* For mmu_notifiers */
1014 unsigned long mmun_end; /* For mmu_notifiers */
1015 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001016 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017
Nick Piggind9928952005-08-28 16:49:11 +10001018 /*
1019 * Don't copy ptes where a page fault will fill them correctly.
1020 * Fork becomes much lighter when there are big shared or private
1021 * readonly mappings. The tradeoff is that copy_page_range is more
1022 * efficient than faulting.
1023 */
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001024 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1025 VM_PFNMAP | VM_MIXEDMAP))) {
Nick Piggind9928952005-08-28 16:49:11 +10001026 if (!vma->anon_vma)
1027 return 0;
1028 }
1029
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 if (is_vm_hugetlb_page(vma))
1031 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1032
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001033 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001034 /*
1035 * We do not free on error cases below as remove_vma
1036 * gets called on error from higher level routine
1037 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001038 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001039 if (ret)
1040 return ret;
1041 }
1042
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001043 /*
1044 * We need to invalidate the secondary MMU mappings only when
1045 * there could be a permission downgrade on the ptes of the
1046 * parent mm. And a permission downgrade will only happen if
1047 * is_cow_mapping() returns true.
1048 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001049 is_cow = is_cow_mapping(vma->vm_flags);
1050 mmun_start = addr;
1051 mmun_end = end;
1052 if (is_cow)
1053 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1054 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001055
1056 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057 dst_pgd = pgd_offset(dst_mm, addr);
1058 src_pgd = pgd_offset(src_mm, addr);
1059 do {
1060 next = pgd_addr_end(addr, end);
1061 if (pgd_none_or_clear_bad(src_pgd))
1062 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001063 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1064 vma, addr, next))) {
1065 ret = -ENOMEM;
1066 break;
1067 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001069
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001070 if (is_cow)
1071 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001072 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073}
1074
Robin Holt51c6f662005-11-13 16:06:42 -08001075static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001076 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001078 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079{
Nick Pigginb5810032005-10-29 18:16:12 -07001080 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001081 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001082 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001083 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001084 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001085 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001086
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001087again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001088 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001089 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1090 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001091 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 do {
1093 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001094 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001096 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001097
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001099 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001100
Linus Torvalds6aab3412005-11-28 14:34:23 -08001101 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 if (unlikely(details) && page) {
1103 /*
1104 * unmap_shared_mapping_pages() wants to
1105 * invalidate cache without truncating:
1106 * unmap shared but keep private pages.
1107 */
1108 if (details->check_mapping &&
1109 details->check_mapping != page->mapping)
1110 continue;
1111 /*
1112 * Each page->index must be checked when
1113 * invalidating or truncating nonlinear.
1114 */
1115 if (details->nonlinear_vma &&
1116 (page->index < details->first_index ||
1117 page->index > details->last_index))
1118 continue;
1119 }
Nick Pigginb5810032005-10-29 18:16:12 -07001120 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001121 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 tlb_remove_tlb_entry(tlb, pte, addr);
1123 if (unlikely(!page))
1124 continue;
1125 if (unlikely(details) && details->nonlinear_vma
1126 && linear_page_index(details->nonlinear_vma,
Cyrill Gorcunov41bb3472013-08-13 16:00:51 -07001127 addr) != page->index) {
1128 pte_t ptfile = pgoff_to_pte(page->index);
1129 if (pte_soft_dirty(ptent))
Peter Feinerdbab31a2014-09-25 16:05:29 -07001130 ptfile = pte_file_mksoft_dirty(ptfile);
Cyrill Gorcunov41bb3472013-08-13 16:00:51 -07001131 set_pte_at(mm, addr, pte, ptfile);
1132 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001134 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001135 else {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001136 if (pte_dirty(ptent)) {
1137 force_flush = 1;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001138 set_page_dirty(page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001139 }
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001140 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001141 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001142 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001143 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001144 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001145 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001146 if (unlikely(page_mapcount(page) < 0))
1147 print_bad_pte(vma, addr, ptent, page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001148 if (unlikely(!__tlb_remove_page(tlb, page))) {
1149 force_flush = 1;
Will Deaconce9ec372014-10-28 13:16:28 -07001150 addr += PAGE_SIZE;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001151 break;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001152 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153 continue;
1154 }
1155 /*
1156 * If details->check_mapping, we leave swap entries;
1157 * if details->nonlinear_vma, we leave file entries.
1158 */
1159 if (unlikely(details))
1160 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001161 if (pte_file(ptent)) {
1162 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1163 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001164 } else {
1165 swp_entry_t entry = pte_to_swp_entry(ptent);
1166
1167 if (!non_swap_entry(entry))
1168 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001169 else if (is_migration_entry(entry)) {
1170 struct page *page;
1171
1172 page = migration_entry_to_page(entry);
1173
1174 if (PageAnon(page))
1175 rss[MM_ANONPAGES]--;
1176 else
1177 rss[MM_FILEPAGES]--;
1178 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001179 if (unlikely(!free_swap_and_cache(entry)))
1180 print_bad_pte(vma, addr, ptent, NULL);
1181 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001182 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001183 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001184
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001185 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001186 arch_leave_lazy_mmu_mode();
Robin Holt51c6f662005-11-13 16:06:42 -08001187
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001188 /* Do the actual TLB flush before dropping ptl */
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001189 if (force_flush) {
Linus Torvalds2b047252013-08-15 11:42:25 -07001190 unsigned long old_end;
1191
Linus Torvalds2b047252013-08-15 11:42:25 -07001192 /*
1193 * Flush the TLB just for the previous segment,
1194 * then update the range to be the remaining
1195 * TLB range.
1196 */
1197 old_end = tlb->end;
Vineet Guptae6c495a2013-07-03 15:03:31 -07001198 tlb->end = addr;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001199 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds2b047252013-08-15 11:42:25 -07001200 tlb->start = addr;
1201 tlb->end = old_end;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001202 }
1203 pte_unmap_unlock(start_pte, ptl);
1204
1205 /*
1206 * If we forced a TLB flush (either due to running out of
1207 * batch buffers or because we needed to flush dirty TLB
1208 * entries before releasing the ptl), free the batched
1209 * memory too. Restart if we didn't do everything.
1210 */
1211 if (force_flush) {
1212 force_flush = 0;
1213 tlb_flush_mmu_free(tlb);
Linus Torvalds2b047252013-08-15 11:42:25 -07001214
1215 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001216 goto again;
1217 }
1218
Robin Holt51c6f662005-11-13 16:06:42 -08001219 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220}
1221
Robin Holt51c6f662005-11-13 16:06:42 -08001222static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001223 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001225 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226{
1227 pmd_t *pmd;
1228 unsigned long next;
1229
1230 pmd = pmd_offset(pud, addr);
1231 do {
1232 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001233 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001234 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001235#ifdef CONFIG_DEBUG_VM
1236 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1237 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1238 __func__, addr, end,
1239 vma->vm_start,
1240 vma->vm_end);
1241 BUG();
1242 }
1243#endif
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001244 split_huge_page_pmd(vma, addr, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001245 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001246 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001247 /* fall through */
1248 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001249 /*
1250 * Here there can be other concurrent MADV_DONTNEED or
1251 * trans huge page faults running, and if the pmd is
1252 * none or trans huge it can change under us. This is
1253 * because MADV_DONTNEED holds the mmap_sem in read
1254 * mode.
1255 */
1256 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1257 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001258 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001259next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001260 cond_resched();
1261 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001262
1263 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264}
1265
Robin Holt51c6f662005-11-13 16:06:42 -08001266static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001267 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001269 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270{
1271 pud_t *pud;
1272 unsigned long next;
1273
1274 pud = pud_offset(pgd, addr);
1275 do {
1276 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001277 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001279 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1280 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001281
1282 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283}
1284
Al Viro038c7aa2012-03-05 13:25:09 -05001285static void unmap_page_range(struct mmu_gather *tlb,
1286 struct vm_area_struct *vma,
1287 unsigned long addr, unsigned long end,
1288 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289{
1290 pgd_t *pgd;
1291 unsigned long next;
1292
1293 if (details && !details->check_mapping && !details->nonlinear_vma)
1294 details = NULL;
1295
1296 BUG_ON(addr >= end);
1297 tlb_start_vma(tlb, vma);
1298 pgd = pgd_offset(vma->vm_mm, addr);
1299 do {
1300 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001301 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001303 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1304 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 tlb_end_vma(tlb, vma);
1306}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307
Al Virof5cc4ee2012-03-05 14:14:20 -05001308
1309static void unmap_single_vma(struct mmu_gather *tlb,
1310 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001311 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001312 struct zap_details *details)
1313{
1314 unsigned long start = max(vma->vm_start, start_addr);
1315 unsigned long end;
1316
1317 if (start >= vma->vm_end)
1318 return;
1319 end = min(vma->vm_end, end_addr);
1320 if (end <= vma->vm_start)
1321 return;
1322
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301323 if (vma->vm_file)
1324 uprobe_munmap(vma, start, end);
1325
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001326 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001327 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001328
1329 if (start != end) {
1330 if (unlikely(is_vm_hugetlb_page(vma))) {
1331 /*
1332 * It is undesirable to test vma->vm_file as it
1333 * should be non-null for valid hugetlb area.
1334 * However, vm_file will be NULL in the error
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001335 * cleanup path of mmap_region. When
Al Virof5cc4ee2012-03-05 14:14:20 -05001336 * hugetlbfs ->mmap method fails,
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001337 * mmap_region() nullifies vma->vm_file
Al Virof5cc4ee2012-03-05 14:14:20 -05001338 * before calling this function to clean up.
1339 * Since no pte has actually been setup, it is
1340 * safe to do nothing in this case.
1341 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001342 if (vma->vm_file) {
1343 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Mel Gormand8333522012-07-31 16:46:20 -07001344 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001345 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1346 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001347 } else
1348 unmap_page_range(tlb, vma, start, end, details);
1349 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350}
1351
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352/**
1353 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001354 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 * @vma: the starting vma
1356 * @start_addr: virtual address at which to start unmapping
1357 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001359 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 * Only addresses between `start' and `end' will be unmapped.
1362 *
1363 * The VMA list must be sorted in ascending virtual address order.
1364 *
1365 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1366 * range after unmap_vmas() returns. So the only responsibility here is to
1367 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1368 * drops the lock and schedules.
1369 */
Al Viro6e8bb012012-03-05 13:41:15 -05001370void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001372 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001374 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001376 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001377 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001378 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001379 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380}
1381
1382/**
1383 * zap_page_range - remove user pages in a given range
1384 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001385 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 * @size: number of bytes to zap
1387 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001388 *
1389 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001391void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 unsigned long size, struct zap_details *details)
1393{
1394 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001395 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001396 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001399 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001400 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001401 mmu_notifier_invalidate_range_start(mm, start, end);
1402 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001403 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001404 mmu_notifier_invalidate_range_end(mm, start, end);
1405 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406}
1407
Jack Steinerc627f9c2008-07-29 22:33:53 -07001408/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001409 * zap_page_range_single - remove user pages in a given range
1410 * @vma: vm_area_struct holding the applicable pages
1411 * @address: starting address of pages to zap
1412 * @size: number of bytes to zap
1413 * @details: details of nonlinear truncation or shared cache invalidation
1414 *
1415 * The range must fit into one VMA.
1416 */
1417static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1418 unsigned long size, struct zap_details *details)
1419{
1420 struct mm_struct *mm = vma->vm_mm;
1421 struct mmu_gather tlb;
1422 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001423
1424 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001425 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001426 update_hiwater_rss(mm);
1427 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001428 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001429 mmu_notifier_invalidate_range_end(mm, address, end);
1430 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431}
1432
Jack Steinerc627f9c2008-07-29 22:33:53 -07001433/**
1434 * zap_vma_ptes - remove ptes mapping the vma
1435 * @vma: vm_area_struct holding ptes to be zapped
1436 * @address: starting address of pages to zap
1437 * @size: number of bytes to zap
1438 *
1439 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1440 *
1441 * The entire address range must be fully contained within the vma.
1442 *
1443 * Returns 0 if successful.
1444 */
1445int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1446 unsigned long size)
1447{
1448 if (address < vma->vm_start || address + size > vma->vm_end ||
1449 !(vma->vm_flags & VM_PFNMAP))
1450 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001451 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001452 return 0;
1453}
1454EXPORT_SYMBOL_GPL(zap_vma_ptes);
1455
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001456pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001457 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001458{
1459 pgd_t * pgd = pgd_offset(mm, addr);
1460 pud_t * pud = pud_alloc(mm, pgd, addr);
1461 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001462 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001463 if (pmd) {
1464 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001465 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001466 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001467 }
1468 return NULL;
1469}
1470
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001472 * This is the old fallback for page remapping.
1473 *
1474 * For historical reasons, it only allows reserved pages. Only
1475 * old drivers should use this, and they needed to mark their
1476 * pages reserved for the old functions anyway.
1477 */
Nick Piggin423bad62008-04-28 02:13:01 -07001478static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1479 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001480{
Nick Piggin423bad62008-04-28 02:13:01 -07001481 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001482 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001483 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001484 spinlock_t *ptl;
1485
Linus Torvalds238f58d2005-11-29 13:01:56 -08001486 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001487 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001488 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001489 retval = -ENOMEM;
1490 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001491 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001492 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001493 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001494 retval = -EBUSY;
1495 if (!pte_none(*pte))
1496 goto out_unlock;
1497
1498 /* Ok, finally just insert the thing.. */
1499 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08001500 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001501 page_add_file_rmap(page);
1502 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1503
1504 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001505 pte_unmap_unlock(pte, ptl);
1506 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001507out_unlock:
1508 pte_unmap_unlock(pte, ptl);
1509out:
1510 return retval;
1511}
1512
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001513/**
1514 * vm_insert_page - insert single page into user vma
1515 * @vma: user vma to map to
1516 * @addr: target user address of this page
1517 * @page: source kernel page
1518 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001519 * This allows drivers to insert individual pages they've allocated
1520 * into a user vma.
1521 *
1522 * The page has to be a nice clean _individual_ kernel allocation.
1523 * If you allocate a compound page, you need to have marked it as
1524 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001525 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001526 *
1527 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1528 * took an arbitrary page protection parameter. This doesn't allow
1529 * that. Your vma protection will have to be set up correctly, which
1530 * means that if you want a shared writable mapping, you'd better
1531 * ask for a shared writable mapping!
1532 *
1533 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001534 *
1535 * Usually this function is called from f_op->mmap() handler
1536 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1537 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1538 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001539 */
Nick Piggin423bad62008-04-28 02:13:01 -07001540int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1541 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001542{
1543 if (addr < vma->vm_start || addr >= vma->vm_end)
1544 return -EFAULT;
1545 if (!page_count(page))
1546 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001547 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1548 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1549 BUG_ON(vma->vm_flags & VM_PFNMAP);
1550 vma->vm_flags |= VM_MIXEDMAP;
1551 }
Nick Piggin423bad62008-04-28 02:13:01 -07001552 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001553}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001554EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001555
Nick Piggin423bad62008-04-28 02:13:01 -07001556static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1557 unsigned long pfn, pgprot_t prot)
1558{
1559 struct mm_struct *mm = vma->vm_mm;
1560 int retval;
1561 pte_t *pte, entry;
1562 spinlock_t *ptl;
1563
1564 retval = -ENOMEM;
1565 pte = get_locked_pte(mm, addr, &ptl);
1566 if (!pte)
1567 goto out;
1568 retval = -EBUSY;
1569 if (!pte_none(*pte))
1570 goto out_unlock;
1571
1572 /* Ok, finally just insert the thing.. */
1573 entry = pte_mkspecial(pfn_pte(pfn, prot));
1574 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001575 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07001576
1577 retval = 0;
1578out_unlock:
1579 pte_unmap_unlock(pte, ptl);
1580out:
1581 return retval;
1582}
1583
Nick Piggine0dc0d82007-02-12 00:51:36 -08001584/**
1585 * vm_insert_pfn - insert single pfn into user vma
1586 * @vma: user vma to map to
1587 * @addr: target user address of this page
1588 * @pfn: source kernel pfn
1589 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07001590 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08001591 * they've allocated into a user vma. Same comments apply.
1592 *
1593 * This function should only be called from a vm_ops->fault handler, and
1594 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001595 *
1596 * vma cannot be a COW mapping.
1597 *
1598 * As this is called only for pages that do not currently exist, we
1599 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001600 */
1601int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07001602 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001603{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001604 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001605 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07001606 /*
1607 * Technically, architectures with pte_special can avoid all these
1608 * restrictions (same for remap_pfn_range). However we would like
1609 * consistency in testing and feature parity among all, so we should
1610 * try to keep these invariants in place for everybody.
1611 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001612 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1613 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1614 (VM_PFNMAP|VM_MIXEDMAP));
1615 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1616 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001617
Nick Piggin423bad62008-04-28 02:13:01 -07001618 if (addr < vma->vm_start || addr >= vma->vm_end)
1619 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07001620 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001621 return -EINVAL;
1622
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001623 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001624
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001625 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001626}
1627EXPORT_SYMBOL(vm_insert_pfn);
1628
Nick Piggin423bad62008-04-28 02:13:01 -07001629int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1630 unsigned long pfn)
1631{
1632 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1633
1634 if (addr < vma->vm_start || addr >= vma->vm_end)
1635 return -EFAULT;
1636
1637 /*
1638 * If we don't have pte special, then we have to use the pfn_valid()
1639 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1640 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001641 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1642 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07001643 */
1644 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1645 struct page *page;
1646
1647 page = pfn_to_page(pfn);
1648 return insert_page(vma, addr, page, vma->vm_page_prot);
1649 }
1650 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1651}
1652EXPORT_SYMBOL(vm_insert_mixed);
1653
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001654/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 * maps a range of physical memory into the requested pages. the old
1656 * mappings are removed. any references to nonexistent pages results
1657 * in null mappings (currently treated as "copy-on-access")
1658 */
1659static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1660 unsigned long addr, unsigned long end,
1661 unsigned long pfn, pgprot_t prot)
1662{
1663 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001664 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665
Hugh Dickinsc74df322005-10-29 18:16:23 -07001666 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 if (!pte)
1668 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001669 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 do {
1671 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07001672 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 pfn++;
1674 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001675 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001676 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 return 0;
1678}
1679
1680static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1681 unsigned long addr, unsigned long end,
1682 unsigned long pfn, pgprot_t prot)
1683{
1684 pmd_t *pmd;
1685 unsigned long next;
1686
1687 pfn -= addr >> PAGE_SHIFT;
1688 pmd = pmd_alloc(mm, pud, addr);
1689 if (!pmd)
1690 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001691 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 do {
1693 next = pmd_addr_end(addr, end);
1694 if (remap_pte_range(mm, pmd, addr, next,
1695 pfn + (addr >> PAGE_SHIFT), prot))
1696 return -ENOMEM;
1697 } while (pmd++, addr = next, addr != end);
1698 return 0;
1699}
1700
1701static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1702 unsigned long addr, unsigned long end,
1703 unsigned long pfn, pgprot_t prot)
1704{
1705 pud_t *pud;
1706 unsigned long next;
1707
1708 pfn -= addr >> PAGE_SHIFT;
1709 pud = pud_alloc(mm, pgd, addr);
1710 if (!pud)
1711 return -ENOMEM;
1712 do {
1713 next = pud_addr_end(addr, end);
1714 if (remap_pmd_range(mm, pud, addr, next,
1715 pfn + (addr >> PAGE_SHIFT), prot))
1716 return -ENOMEM;
1717 } while (pud++, addr = next, addr != end);
1718 return 0;
1719}
1720
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001721/**
1722 * remap_pfn_range - remap kernel memory to userspace
1723 * @vma: user vma to map to
1724 * @addr: target user address to start at
1725 * @pfn: physical address of kernel memory
1726 * @size: size of map area
1727 * @prot: page protection flags for this mapping
1728 *
1729 * Note: this is only safe if the mm semaphore is held when called.
1730 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1732 unsigned long pfn, unsigned long size, pgprot_t prot)
1733{
1734 pgd_t *pgd;
1735 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001736 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 struct mm_struct *mm = vma->vm_mm;
1738 int err;
1739
1740 /*
1741 * Physically remapped pages are special. Tell the
1742 * rest of the world about it:
1743 * VM_IO tells people not to look at these pages
1744 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08001745 * VM_PFNMAP tells the core MM that the base pages are just
1746 * raw PFN mappings, and do not have a "struct page" associated
1747 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001748 * VM_DONTEXPAND
1749 * Disable vma merging and expanding with mremap().
1750 * VM_DONTDUMP
1751 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001752 *
1753 * There's a horrible special case to handle copy-on-write
1754 * behaviour that some programs depend on. We mark the "original"
1755 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001756 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001758 if (is_cow_mapping(vma->vm_flags)) {
1759 if (addr != vma->vm_start || end != vma->vm_end)
1760 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001761 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001762 }
1763
1764 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1765 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001766 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001767
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001768 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769
1770 BUG_ON(addr >= end);
1771 pfn -= addr >> PAGE_SHIFT;
1772 pgd = pgd_offset(mm, addr);
1773 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 do {
1775 next = pgd_addr_end(addr, end);
1776 err = remap_pud_range(mm, pgd, addr, next,
1777 pfn + (addr >> PAGE_SHIFT), prot);
1778 if (err)
1779 break;
1780 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001781
1782 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07001783 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001784
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 return err;
1786}
1787EXPORT_SYMBOL(remap_pfn_range);
1788
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07001789/**
1790 * vm_iomap_memory - remap memory to userspace
1791 * @vma: user vma to map to
1792 * @start: start of area
1793 * @len: size of area
1794 *
1795 * This is a simplified io_remap_pfn_range() for common driver use. The
1796 * driver just needs to give us the physical memory range to be mapped,
1797 * we'll figure out the rest from the vma information.
1798 *
1799 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1800 * whatever write-combining details or similar.
1801 */
1802int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1803{
1804 unsigned long vm_len, pfn, pages;
1805
1806 /* Check that the physical memory area passed in looks valid */
1807 if (start + len < start)
1808 return -EINVAL;
1809 /*
1810 * You *really* shouldn't map things that aren't page-aligned,
1811 * but we've historically allowed it because IO memory might
1812 * just have smaller alignment.
1813 */
1814 len += start & ~PAGE_MASK;
1815 pfn = start >> PAGE_SHIFT;
1816 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1817 if (pfn + pages < pfn)
1818 return -EINVAL;
1819
1820 /* We start the mapping 'vm_pgoff' pages into the area */
1821 if (vma->vm_pgoff > pages)
1822 return -EINVAL;
1823 pfn += vma->vm_pgoff;
1824 pages -= vma->vm_pgoff;
1825
1826 /* Can we fit all of the mapping? */
1827 vm_len = vma->vm_end - vma->vm_start;
1828 if (vm_len >> PAGE_SHIFT > pages)
1829 return -EINVAL;
1830
1831 /* Ok, let it rip */
1832 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1833}
1834EXPORT_SYMBOL(vm_iomap_memory);
1835
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001836static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1837 unsigned long addr, unsigned long end,
1838 pte_fn_t fn, void *data)
1839{
1840 pte_t *pte;
1841 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001842 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001843 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001844
1845 pte = (mm == &init_mm) ?
1846 pte_alloc_kernel(pmd, addr) :
1847 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1848 if (!pte)
1849 return -ENOMEM;
1850
1851 BUG_ON(pmd_huge(*pmd));
1852
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001853 arch_enter_lazy_mmu_mode();
1854
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001855 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001856
1857 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001858 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001859 if (err)
1860 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001861 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001862
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001863 arch_leave_lazy_mmu_mode();
1864
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001865 if (mm != &init_mm)
1866 pte_unmap_unlock(pte-1, ptl);
1867 return err;
1868}
1869
1870static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1871 unsigned long addr, unsigned long end,
1872 pte_fn_t fn, void *data)
1873{
1874 pmd_t *pmd;
1875 unsigned long next;
1876 int err;
1877
Andi Kleenceb86872008-07-23 21:27:50 -07001878 BUG_ON(pud_huge(*pud));
1879
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001880 pmd = pmd_alloc(mm, pud, addr);
1881 if (!pmd)
1882 return -ENOMEM;
1883 do {
1884 next = pmd_addr_end(addr, end);
1885 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1886 if (err)
1887 break;
1888 } while (pmd++, addr = next, addr != end);
1889 return err;
1890}
1891
1892static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1893 unsigned long addr, unsigned long end,
1894 pte_fn_t fn, void *data)
1895{
1896 pud_t *pud;
1897 unsigned long next;
1898 int err;
1899
1900 pud = pud_alloc(mm, pgd, addr);
1901 if (!pud)
1902 return -ENOMEM;
1903 do {
1904 next = pud_addr_end(addr, end);
1905 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1906 if (err)
1907 break;
1908 } while (pud++, addr = next, addr != end);
1909 return err;
1910}
1911
1912/*
1913 * Scan a region of virtual memory, filling in page tables as necessary
1914 * and calling a provided function on each leaf page table.
1915 */
1916int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1917 unsigned long size, pte_fn_t fn, void *data)
1918{
1919 pgd_t *pgd;
1920 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001921 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001922 int err;
1923
1924 BUG_ON(addr >= end);
1925 pgd = pgd_offset(mm, addr);
1926 do {
1927 next = pgd_addr_end(addr, end);
1928 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1929 if (err)
1930 break;
1931 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001932
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001933 return err;
1934}
1935EXPORT_SYMBOL_GPL(apply_to_page_range);
1936
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001938 * handle_pte_fault chooses page fault handler according to an entry
1939 * which was read non-atomically. Before making any commitment, on
1940 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001941 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001942 * must check under lock before unmapping the pte and proceeding
1943 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001944 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001945 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001946static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001947 pte_t *page_table, pte_t orig_pte)
1948{
1949 int same = 1;
1950#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1951 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001952 spinlock_t *ptl = pte_lockptr(mm, pmd);
1953 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001954 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001955 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001956 }
1957#endif
1958 pte_unmap(page_table);
1959 return same;
1960}
1961
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001962static 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 -08001963{
Dan Williams0abdd7a2014-01-21 15:48:12 -08001964 debug_dma_assert_idle(src);
1965
Linus Torvalds6aab3412005-11-28 14:34:23 -08001966 /*
1967 * If the source page was a PFN mapping, we don't have
1968 * a "struct page" for it. We do a best-effort copy by
1969 * just copying from the original user address. If that
1970 * fails, we just zero-fill it. Live with it.
1971 */
1972 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08001973 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08001974 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1975
1976 /*
1977 * This really shouldn't fail, because the page is there
1978 * in the page tables. But it might just be unreadable,
1979 * in which case we just give up and fill the result with
1980 * zeroes.
1981 */
1982 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07001983 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08001984 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07001985 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08001986 } else
1987 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08001988}
1989
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990/*
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001991 * Notify the address space that the page is about to become writable so that
1992 * it can prohibit this or wait for the page to get into an appropriate state.
1993 *
1994 * We do this without the lock held, so that it can sleep if it needs to.
1995 */
1996static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1997 unsigned long address)
1998{
1999 struct vm_fault vmf;
2000 int ret;
2001
2002 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2003 vmf.pgoff = page->index;
2004 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2005 vmf.page = page;
2006
2007 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2008 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2009 return ret;
2010 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2011 lock_page(page);
2012 if (!page->mapping) {
2013 unlock_page(page);
2014 return 0; /* retry */
2015 }
2016 ret |= VM_FAULT_LOCKED;
2017 } else
2018 VM_BUG_ON_PAGE(!PageLocked(page), page);
2019 return ret;
2020}
2021
2022/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 * This routine handles present pages, when users try to write
2024 * to a shared page. It is done by copying the page to a new address
2025 * and decrementing the shared-page counter for the old page.
2026 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 * Note that this routine assumes that the protection checks have been
2028 * done by the caller (the low-level page fault routine in most cases).
2029 * Thus we can safely just mark it writable once we've done any necessary
2030 * COW.
2031 *
2032 * We also mark the page dirty at this point even though the page will
2033 * change only once the write actually happens. This avoids a few races,
2034 * and potentially makes it more efficient.
2035 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002036 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2037 * but allow concurrent faults), with pte both mapped and locked.
2038 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002040static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2041 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002042 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002043 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044{
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002045 struct page *old_page, *new_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002047 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002048 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002049 struct page *dirty_page = NULL;
David Rientjes17569542012-11-16 14:14:48 -08002050 unsigned long mmun_start = 0; /* For mmu_notifiers */
2051 unsigned long mmun_end = 0; /* For mmu_notifiers */
Johannes Weiner00501b52014-08-08 14:19:20 -07002052 struct mem_cgroup *memcg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053
Linus Torvalds6aab3412005-11-28 14:34:23 -08002054 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002055 if (!old_page) {
2056 /*
Peter Feiner64e45502014-10-13 15:55:46 -07002057 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2058 * VM_PFNMAP VMA.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002059 *
2060 * We should not cow pages in a shared writeable mapping.
2061 * Just mark the pages writable as we can't do any dirty
2062 * accounting on raw pfn maps.
2063 */
2064 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2065 (VM_WRITE|VM_SHARED))
2066 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002067 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002068 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002070 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002071 * Take out anonymous pages first, anonymous shared vmas are
2072 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002073 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002074 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002075 if (!trylock_page(old_page)) {
2076 page_cache_get(old_page);
2077 pte_unmap_unlock(page_table, ptl);
2078 lock_page(old_page);
2079 page_table = pte_offset_map_lock(mm, pmd, address,
2080 &ptl);
2081 if (!pte_same(*page_table, orig_pte)) {
2082 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002083 goto unlock;
2084 }
2085 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002086 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002087 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002088 /*
2089 * The page is all ours. Move it to our anon_vma so
2090 * the rmap code will not search our parent or siblings.
2091 * Protected against the rmap code by the page lock.
2092 */
2093 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002094 unlock_page(old_page);
2095 goto reuse;
2096 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002097 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002098 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002099 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002100 /*
2101 * Only catch write-faults on shared writable pages,
2102 * read-only shared pages can get COWed by
2103 * get_user_pages(.write=1, .force=1).
2104 */
David Howells9637a5e2006-06-23 02:03:43 -07002105 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002106 int tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002107 page_cache_get(old_page);
2108 pte_unmap_unlock(page_table, ptl);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002109 tmp = do_page_mkwrite(vma, old_page, address);
2110 if (unlikely(!tmp || (tmp &
2111 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2112 page_cache_release(old_page);
2113 return tmp;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002114 }
David Howells9637a5e2006-06-23 02:03:43 -07002115 /*
2116 * Since we dropped the lock we need to revalidate
2117 * the PTE as someone else may have changed it. If
2118 * they did, we just return, as we can count on the
2119 * MMU to tell us if they didn't also make it writable.
2120 */
2121 page_table = pte_offset_map_lock(mm, pmd, address,
2122 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002123 if (!pte_same(*page_table, orig_pte)) {
2124 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002125 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002126 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002127
2128 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002130 dirty_page = old_page;
2131 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002132
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002133reuse:
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002134 /*
2135 * Clear the pages cpupid information as the existing
2136 * information potentially belongs to a now completely
2137 * unrelated process.
2138 */
2139 if (old_page)
2140 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2141
David Howells9637a5e2006-06-23 02:03:43 -07002142 flush_cache_page(vma, address, pte_pfn(orig_pte));
2143 entry = pte_mkyoung(orig_pte);
2144 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002145 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002146 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002147 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002148 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002149
2150 if (!dirty_page)
2151 return ret;
2152
2153 /*
2154 * Yes, Virginia, this is actually required to prevent a race
2155 * with clear_page_dirty_for_io() from clearing the page dirty
2156 * bit after it clear all dirty ptes, but before a racing
2157 * do_wp_page installs a dirty pte.
2158 *
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002159 * do_shared_fault is protected similarly.
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002160 */
2161 if (!page_mkwrite) {
2162 wait_on_page_locked(dirty_page);
Miklos Szeredied6d7c82014-04-07 15:37:51 -07002163 set_page_dirty_balance(dirty_page);
Jan Kara41c4d252012-06-12 16:20:28 +02002164 /* file_update_time outside page_lock */
2165 if (vma->vm_file)
2166 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002167 }
2168 put_page(dirty_page);
2169 if (page_mkwrite) {
2170 struct address_space *mapping = dirty_page->mapping;
2171
2172 set_page_dirty(dirty_page);
2173 unlock_page(dirty_page);
2174 page_cache_release(dirty_page);
2175 if (mapping) {
2176 /*
2177 * Some device drivers do not set page.mapping
2178 * but still dirty their pages
2179 */
2180 balance_dirty_pages_ratelimited(mapping);
2181 }
2182 }
2183
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002184 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186
2187 /*
2188 * Ok, we need to copy. Oh, well..
2189 */
Nick Pigginb5810032005-10-29 18:16:12 -07002190 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002191gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002192 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193
2194 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002195 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002196
Hugh Dickins62eede62009-09-21 17:03:34 -07002197 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002198 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2199 if (!new_page)
2200 goto oom;
2201 } else {
2202 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2203 if (!new_page)
2204 goto oom;
2205 cow_user_page(new_page, old_page, address, vma);
2206 }
2207 __SetPageUptodate(new_page);
2208
Johannes Weiner00501b52014-08-08 14:19:20 -07002209 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002210 goto oom_free_new;
2211
Haggai Eran6bdb9132012-10-08 16:33:35 -07002212 mmun_start = address & PAGE_MASK;
David Rientjes17569542012-11-16 14:14:48 -08002213 mmun_end = mmun_start + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -07002214 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2215
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 /*
2217 * Re-check the pte - we dropped the lock
2218 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002219 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002220 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002221 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002222 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002223 dec_mm_counter_fast(mm, MM_FILEPAGES);
2224 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002225 }
2226 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002227 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002228 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002229 entry = mk_pte(new_page, vma->vm_page_prot);
2230 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002231 /*
2232 * Clear the pte entry and flush it first, before updating the
2233 * pte with the new entry. This will avoid a race condition
2234 * seen in the presence of one thread doing SMC and another
2235 * thread doing COW.
2236 */
Izik Eidus828502d2009-09-21 17:01:51 -07002237 ptep_clear_flush(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002238 page_add_new_anon_rmap(new_page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002239 mem_cgroup_commit_charge(new_page, memcg, false);
2240 lru_cache_add_active_or_unevictable(new_page, vma);
Izik Eidus828502d2009-09-21 17:01:51 -07002241 /*
2242 * We call the notify macro here because, when using secondary
2243 * mmu page tables (such as kvm shadow page tables), we want the
2244 * new page to be mapped directly into the secondary page table.
2245 */
2246 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002247 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002248 if (old_page) {
2249 /*
2250 * Only after switching the pte to the new page may
2251 * we remove the mapcount here. Otherwise another
2252 * process may come and find the rmap count decremented
2253 * before the pte is switched to the new page, and
2254 * "reuse" the old page writing into it while our pte
2255 * here still points into it and can be read by other
2256 * threads.
2257 *
2258 * The critical issue is to order this
2259 * page_remove_rmap with the ptp_clear_flush above.
2260 * Those stores are ordered by (if nothing else,)
2261 * the barrier present in the atomic_add_negative
2262 * in page_remove_rmap.
2263 *
2264 * Then the TLB flush in ptep_clear_flush ensures that
2265 * no process can access the old page before the
2266 * decremented mapcount is visible. And the old page
2267 * cannot be reused until after the decremented
2268 * mapcount is visible. So transitively, TLBs to
2269 * old page will be flushed before it can be reused.
2270 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002271 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002272 }
2273
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 /* Free the old page.. */
2275 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002276 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002277 } else
Johannes Weiner00501b52014-08-08 14:19:20 -07002278 mem_cgroup_cancel_charge(new_page, memcg);
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002279
Haggai Eran6bdb9132012-10-08 16:33:35 -07002280 if (new_page)
2281 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002282unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002283 pte_unmap_unlock(page_table, ptl);
David Rientjes17569542012-11-16 14:14:48 -08002284 if (mmun_end > mmun_start)
Haggai Eran6bdb9132012-10-08 16:33:35 -07002285 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002286 if (old_page) {
2287 /*
2288 * Don't let another task, with possibly unlocked vma,
2289 * keep the mlocked page.
2290 */
2291 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2292 lock_page(old_page); /* LRU manipulation */
2293 munlock_vma_page(old_page);
2294 unlock_page(old_page);
2295 }
2296 page_cache_release(old_page);
2297 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002298 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002299oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002300 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002301oom:
Dominik Dingel66521d52012-12-12 13:52:37 -08002302 if (old_page)
Hugh Dickins920fc352005-11-21 21:32:17 -08002303 page_cache_release(old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 return VM_FAULT_OOM;
2305}
2306
Peter Zijlstra97a89412011-05-24 17:12:04 -07002307static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 unsigned long start_addr, unsigned long end_addr,
2309 struct zap_details *details)
2310{
Al Virof5cc4ee2012-03-05 14:14:20 -05002311 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312}
2313
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002314static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315 struct zap_details *details)
2316{
2317 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 pgoff_t vba, vea, zba, zea;
2319
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002320 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322
2323 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002324 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2326 zba = details->first_index;
2327 if (zba < vba)
2328 zba = vba;
2329 zea = details->last_index;
2330 if (zea > vea)
2331 zea = vea;
2332
Peter Zijlstra97a89412011-05-24 17:12:04 -07002333 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2335 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002336 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 }
2338}
2339
2340static inline void unmap_mapping_range_list(struct list_head *head,
2341 struct zap_details *details)
2342{
2343 struct vm_area_struct *vma;
2344
2345 /*
2346 * In nonlinear VMAs there is no correspondence between virtual address
2347 * offset and file offset. So we must perform an exhaustive search
2348 * across *all* the pages in each nonlinear VMA, not just the pages
2349 * whose virtual address lies outside the file truncation point.
2350 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002351 list_for_each_entry(vma, head, shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002353 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 }
2355}
2356
2357/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002358 * 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 -07002359 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360 * @holebegin: byte in first page to unmap, relative to the start of
2361 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002362 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 * must keep the partial page. In contrast, we must get rid of
2364 * partial pages.
2365 * @holelen: size of prospective hole in bytes. This will be rounded
2366 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2367 * end of the file.
2368 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2369 * but 0 when invalidating pagecache, don't throw away private data.
2370 */
2371void unmap_mapping_range(struct address_space *mapping,
2372 loff_t const holebegin, loff_t const holelen, int even_cows)
2373{
2374 struct zap_details details;
2375 pgoff_t hba = holebegin >> PAGE_SHIFT;
2376 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2377
2378 /* Check for overflow. */
2379 if (sizeof(holelen) > sizeof(hlen)) {
2380 long long holeend =
2381 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2382 if (holeend & ~(long long)ULONG_MAX)
2383 hlen = ULONG_MAX - hba + 1;
2384 }
2385
2386 details.check_mapping = even_cows? NULL: mapping;
2387 details.nonlinear_vma = NULL;
2388 details.first_index = hba;
2389 details.last_index = hba + hlen - 1;
2390 if (details.last_index < details.first_index)
2391 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392
Peter Zijlstra97a89412011-05-24 17:12:04 -07002393
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002394 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002395 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2397 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2398 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002399 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400}
2401EXPORT_SYMBOL(unmap_mapping_range);
2402
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002404 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2405 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002406 * We return with pte unmapped and unlocked.
2407 *
2408 * We return with the mmap_sem locked or unlocked in the same cases
2409 * as does filemap_fault().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002411static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2412 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002413 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002415 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08002416 struct page *page, *swapcache;
Johannes Weiner00501b52014-08-08 14:19:20 -07002417 struct mem_cgroup *memcg;
Hugh Dickins65500d22005-10-29 18:15:59 -07002418 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002420 int locked;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002421 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002422 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002424 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002425 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002426
2427 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002428 if (unlikely(non_swap_entry(entry))) {
2429 if (is_migration_entry(entry)) {
2430 migration_entry_wait(mm, pmd, address);
2431 } else if (is_hwpoison_entry(entry)) {
2432 ret = VM_FAULT_HWPOISON;
2433 } else {
2434 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002435 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002436 }
Christoph Lameter06972122006-06-23 02:03:35 -07002437 goto out;
2438 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002439 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440 page = lookup_swap_cache(entry);
2441 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002442 page = swapin_readahead(entry,
2443 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 if (!page) {
2445 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002446 * Back out if somebody else faulted in this pte
2447 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002449 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 if (likely(pte_same(*page_table, orig_pte)))
2451 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002452 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002453 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454 }
2455
2456 /* Had to read the page from swap area: Major fault */
2457 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002458 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002459 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002460 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002461 /*
2462 * hwpoisoned dirty swapcache pages are kept for killing
2463 * owner processes (which may be unknown at hwpoison time)
2464 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002465 ret = VM_FAULT_HWPOISON;
2466 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08002467 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02002468 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469 }
2470
Hugh Dickins56f31802013-02-22 16:36:10 -08002471 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002472 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002473
Balbir Singh20a10222007-11-14 17:00:33 -08002474 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002475 if (!locked) {
2476 ret |= VM_FAULT_RETRY;
2477 goto out_release;
2478 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002480 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002481 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2482 * release the swapcache from under us. The page pin, and pte_same
2483 * test below, are not enough to exclude that. Even if it is still
2484 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002485 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002486 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002487 goto out_page;
2488
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002489 page = ksm_might_need_to_copy(page, vma, address);
2490 if (unlikely(!page)) {
2491 ret = VM_FAULT_OOM;
2492 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002493 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08002494 }
2495
Johannes Weiner00501b52014-08-08 14:19:20 -07002496 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002497 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002498 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002499 }
2500
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002502 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002504 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002505 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002506 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002507
2508 if (unlikely(!PageUptodate(page))) {
2509 ret = VM_FAULT_SIGBUS;
2510 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 }
2512
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002513 /*
2514 * The page isn't present yet, go ahead with the fault.
2515 *
2516 * Be careful about the sequence of operations here.
2517 * To get its accounting right, reuse_swap_page() must be called
2518 * while the page is counted on swap but not yet in mapcount i.e.
2519 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2520 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002521 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002523 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08002524 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002526 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002528 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002529 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002530 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07002533 if (pte_swp_soft_dirty(orig_pte))
2534 pte = pte_mksoft_dirty(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 set_pte_at(mm, address, page_table, pte);
Johannes Weiner00501b52014-08-08 14:19:20 -07002536 if (page == swapcache) {
Johannes Weineraf347702013-02-22 16:32:20 -08002537 do_page_add_anon_rmap(page, vma, address, exclusive);
Johannes Weiner00501b52014-08-08 14:19:20 -07002538 mem_cgroup_commit_charge(page, memcg, true);
2539 } else { /* ksm created a completely new copy */
Hugh Dickins56f31802013-02-22 16:36:10 -08002540 page_add_new_anon_rmap(page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002541 mem_cgroup_commit_charge(page, memcg, false);
2542 lru_cache_add_active_or_unevictable(page, vma);
2543 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002545 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07002546 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002547 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002548 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002549 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002550 /*
2551 * Hold the lock to avoid the swap entry to be reused
2552 * until we take the PT lock for the pte_same() check
2553 * (to avoid false positives from pte_same). For
2554 * further safety release the lock after the swap_free
2555 * so that the swap count won't change under a
2556 * parallel locked swapcache.
2557 */
2558 unlock_page(swapcache);
2559 page_cache_release(swapcache);
2560 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002561
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002562 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08002563 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2564 if (ret & VM_FAULT_ERROR)
2565 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566 goto out;
2567 }
2568
2569 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002570 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002571unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002572 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573out:
2574 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002575out_nomap:
Johannes Weiner00501b52014-08-08 14:19:20 -07002576 mem_cgroup_cancel_charge(page, memcg);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002577 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002578out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002579 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002580out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002581 page_cache_release(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002582 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002583 unlock_page(swapcache);
2584 page_cache_release(swapcache);
2585 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002586 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587}
2588
2589/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002590 * This is like a special single-page "expand_{down|up}wards()",
2591 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07002592 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07002593 */
2594static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2595{
2596 address &= PAGE_MASK;
2597 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002598 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07002599
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002600 /*
2601 * Is there a mapping abutting this one below?
2602 *
2603 * That's only ok if it's the same stack mapping
2604 * that has gotten split..
2605 */
2606 if (prev && prev->vm_end == address)
2607 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2608
Michal Hockod05f3162011-05-24 17:11:44 -07002609 expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002610 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002611 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2612 struct vm_area_struct *next = vma->vm_next;
2613
2614 /* As VM_GROWSDOWN but s/below/above/ */
2615 if (next && next->vm_start == address + PAGE_SIZE)
2616 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2617
2618 expand_upwards(vma, address + PAGE_SIZE);
2619 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07002620 return 0;
2621}
2622
2623/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002624 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2625 * but allow concurrent faults), and pte mapped but not yet locked.
2626 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002628static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2629 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002630 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631{
Johannes Weiner00501b52014-08-08 14:19:20 -07002632 struct mem_cgroup *memcg;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002633 struct page *page;
2634 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636
Linus Torvalds11ac5522010-08-14 11:44:56 -07002637 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002638
Linus Torvalds11ac5522010-08-14 11:44:56 -07002639 /* Check if we need to add a guard page to the stack */
2640 if (check_stack_guard_page(vma, address) < 0)
2641 return VM_FAULT_SIGBUS;
2642
2643 /* Use the zero-page for reads */
Hugh Dickins62eede62009-09-21 17:03:34 -07002644 if (!(flags & FAULT_FLAG_WRITE)) {
2645 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2646 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07002647 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002648 if (!pte_none(*page_table))
2649 goto unlock;
2650 goto setpte;
2651 }
2652
Nick Piggin557ed1f2007-10-16 01:24:40 -07002653 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002654 if (unlikely(anon_vma_prepare(vma)))
2655 goto oom;
2656 page = alloc_zeroed_user_highpage_movable(vma, address);
2657 if (!page)
2658 goto oom;
Minchan Kim52f37622013-04-29 15:08:15 -07002659 /*
2660 * The memory barrier inside __SetPageUptodate makes sure that
2661 * preceeding stores to the page contents become visible before
2662 * the set_pte_at() write.
2663 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08002664 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665
Johannes Weiner00501b52014-08-08 14:19:20 -07002666 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002667 goto oom_free_page;
2668
Nick Piggin557ed1f2007-10-16 01:24:40 -07002669 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002670 if (vma->vm_flags & VM_WRITE)
2671 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002672
Nick Piggin557ed1f2007-10-16 01:24:40 -07002673 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07002674 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002675 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002676
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002677 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07002678 page_add_new_anon_rmap(page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002679 mem_cgroup_commit_charge(page, memcg, false);
2680 lru_cache_add_active_or_unevictable(page, vma);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002681setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07002682 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683
2684 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002685 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002686unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002687 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002688 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002689release:
Johannes Weiner00501b52014-08-08 14:19:20 -07002690 mem_cgroup_cancel_charge(page, memcg);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002691 page_cache_release(page);
2692 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002693oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002694 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002695oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696 return VM_FAULT_OOM;
2697}
2698
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002699/*
2700 * The mmap_sem must have been held on entry, and may have been
2701 * released depending on flags and vma->vm_ops->fault() return value.
2702 * See filemap_fault() and __lock_page_retry().
2703 */
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002704static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2705 pgoff_t pgoff, unsigned int flags, struct page **page)
2706{
2707 struct vm_fault vmf;
2708 int ret;
2709
2710 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2711 vmf.pgoff = pgoff;
2712 vmf.flags = flags;
2713 vmf.page = NULL;
2714
2715 ret = vma->vm_ops->fault(vma, &vmf);
2716 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2717 return ret;
2718
2719 if (unlikely(PageHWPoison(vmf.page))) {
2720 if (ret & VM_FAULT_LOCKED)
2721 unlock_page(vmf.page);
2722 page_cache_release(vmf.page);
2723 return VM_FAULT_HWPOISON;
2724 }
2725
2726 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2727 lock_page(vmf.page);
2728 else
2729 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2730
2731 *page = vmf.page;
2732 return ret;
2733}
2734
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002735/**
2736 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2737 *
2738 * @vma: virtual memory area
2739 * @address: user virtual address
2740 * @page: page to map
2741 * @pte: pointer to target page table entry
2742 * @write: true, if new entry is writable
2743 * @anon: true, if it's anonymous page
2744 *
2745 * Caller must hold page table lock relevant for @pte.
2746 *
2747 * Target users are page handler itself and implementations of
2748 * vm_ops->map_pages.
2749 */
2750void do_set_pte(struct vm_area_struct *vma, unsigned long address,
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002751 struct page *page, pte_t *pte, bool write, bool anon)
2752{
2753 pte_t entry;
2754
2755 flush_icache_page(vma, page);
2756 entry = mk_pte(page, vma->vm_page_prot);
2757 if (write)
2758 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2759 else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
Cyrill Gorcunov9aed8612014-08-06 16:07:05 -07002760 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002761 if (anon) {
2762 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2763 page_add_new_anon_rmap(page, vma, address);
2764 } else {
2765 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2766 page_add_file_rmap(page);
2767 }
2768 set_pte_at(vma->vm_mm, address, pte, entry);
2769
2770 /* no need to invalidate: a not-present page won't be cached */
2771 update_mmu_cache(vma, address, pte);
2772}
2773
Kirill A. Shutemov3a910532014-08-06 16:08:07 -07002774static unsigned long fault_around_bytes __read_mostly =
2775 rounddown_pow_of_two(65536);
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002776
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002777#ifdef CONFIG_DEBUG_FS
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002778static int fault_around_bytes_get(void *data, u64 *val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002779{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002780 *val = fault_around_bytes;
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002781 return 0;
2782}
2783
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002784/*
2785 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2786 * rounded down to nearest page order. It's what do_fault_around() expects to
2787 * see.
2788 */
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002789static int fault_around_bytes_set(void *data, u64 val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002790{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002791 if (val / PAGE_SIZE > PTRS_PER_PTE)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002792 return -EINVAL;
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002793 if (val > PAGE_SIZE)
2794 fault_around_bytes = rounddown_pow_of_two(val);
2795 else
2796 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002797 return 0;
2798}
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002799DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2800 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002801
2802static int __init fault_around_debugfs(void)
2803{
2804 void *ret;
2805
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002806 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2807 &fault_around_bytes_fops);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002808 if (!ret)
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002809 pr_warn("Failed to create fault_around_bytes in debugfs");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002810 return 0;
2811}
2812late_initcall(fault_around_debugfs);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002813#endif
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002814
Kirill A. Shutemov1fdb4122014-06-04 16:10:55 -07002815/*
2816 * do_fault_around() tries to map few pages around the fault address. The hope
2817 * is that the pages will be needed soon and this will lower the number of
2818 * faults to handle.
2819 *
2820 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2821 * not ready to be mapped: not up-to-date, locked, etc.
2822 *
2823 * This function is called with the page table lock taken. In the split ptlock
2824 * case the page table lock only protects only those entries which belong to
2825 * the page table corresponding to the fault address.
2826 *
2827 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2828 * only once.
2829 *
2830 * fault_around_pages() defines how many pages we'll try to map.
2831 * do_fault_around() expects it to return a power of two less than or equal to
2832 * PTRS_PER_PTE.
2833 *
2834 * The virtual address of the area that we map is naturally aligned to the
2835 * fault_around_pages() value (and therefore to page order). This way it's
2836 * easier to guarantee that we don't cross page table boundaries.
2837 */
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002838static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2839 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2840{
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002841 unsigned long start_addr, nr_pages, mask;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002842 pgoff_t max_pgoff;
2843 struct vm_fault vmf;
2844 int off;
2845
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002846 nr_pages = ACCESS_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2847 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2848
2849 start_addr = max(address & mask, vma->vm_start);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002850 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2851 pte -= off;
2852 pgoff -= off;
2853
2854 /*
2855 * max_pgoff is either end of page table or end of vma
Kirill A. Shutemov850e9c62014-06-04 16:10:45 -07002856 * or fault_around_pages() from pgoff, depending what is nearest.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002857 */
2858 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2859 PTRS_PER_PTE - 1;
2860 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002861 pgoff + nr_pages - 1);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002862
2863 /* Check if it makes any sense to call ->map_pages */
2864 while (!pte_none(*pte)) {
2865 if (++pgoff > max_pgoff)
2866 return;
2867 start_addr += PAGE_SIZE;
2868 if (start_addr >= vma->vm_end)
2869 return;
2870 pte++;
2871 }
2872
2873 vmf.virtual_address = (void __user *) start_addr;
2874 vmf.pte = pte;
2875 vmf.pgoff = pgoff;
2876 vmf.max_pgoff = max_pgoff;
2877 vmf.flags = flags;
2878 vma->vm_ops->map_pages(vma, &vmf);
2879}
2880
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002881static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2882 unsigned long address, pmd_t *pmd,
2883 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2884{
2885 struct page *fault_page;
2886 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002887 pte_t *pte;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002888 int ret = 0;
2889
2890 /*
2891 * Let's call ->map_pages() first and use ->fault() as fallback
2892 * if page by the offset is not ready to be mapped (cold cache or
2893 * something).
2894 */
Konstantin Khlebnikovc1186782014-07-23 14:00:08 -07002895 if (vma->vm_ops->map_pages && !(flags & FAULT_FLAG_NONLINEAR) &&
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002896 fault_around_bytes >> PAGE_SHIFT > 1) {
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002897 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2898 do_fault_around(vma, address, pte, pgoff, flags);
2899 if (!pte_same(*pte, orig_pte))
2900 goto unlock_out;
2901 pte_unmap_unlock(pte, ptl);
2902 }
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002903
2904 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2905 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2906 return ret;
2907
2908 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2909 if (unlikely(!pte_same(*pte, orig_pte))) {
2910 pte_unmap_unlock(pte, ptl);
2911 unlock_page(fault_page);
2912 page_cache_release(fault_page);
2913 return ret;
2914 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002915 do_set_pte(vma, address, fault_page, pte, false, false);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002916 unlock_page(fault_page);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002917unlock_out:
2918 pte_unmap_unlock(pte, ptl);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002919 return ret;
2920}
2921
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002922static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2923 unsigned long address, pmd_t *pmd,
2924 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2925{
2926 struct page *fault_page, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07002927 struct mem_cgroup *memcg;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002928 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002929 pte_t *pte;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002930 int ret;
2931
2932 if (unlikely(anon_vma_prepare(vma)))
2933 return VM_FAULT_OOM;
2934
2935 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2936 if (!new_page)
2937 return VM_FAULT_OOM;
2938
Johannes Weiner00501b52014-08-08 14:19:20 -07002939 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) {
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002940 page_cache_release(new_page);
2941 return VM_FAULT_OOM;
2942 }
2943
2944 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2945 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2946 goto uncharge_out;
2947
2948 copy_user_highpage(new_page, fault_page, address, vma);
2949 __SetPageUptodate(new_page);
2950
2951 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2952 if (unlikely(!pte_same(*pte, orig_pte))) {
2953 pte_unmap_unlock(pte, ptl);
2954 unlock_page(fault_page);
2955 page_cache_release(fault_page);
2956 goto uncharge_out;
2957 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002958 do_set_pte(vma, address, new_page, pte, true, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07002959 mem_cgroup_commit_charge(new_page, memcg, false);
2960 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002961 pte_unmap_unlock(pte, ptl);
2962 unlock_page(fault_page);
2963 page_cache_release(fault_page);
2964 return ret;
2965uncharge_out:
Johannes Weiner00501b52014-08-08 14:19:20 -07002966 mem_cgroup_cancel_charge(new_page, memcg);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002967 page_cache_release(new_page);
2968 return ret;
2969}
2970
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002971static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01002972 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07002973 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974{
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002975 struct page *fault_page;
2976 struct address_space *mapping;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002977 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002978 pte_t *pte;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002979 int dirtied = 0;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002980 int ret, tmp;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07002981
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002982 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2983 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002984 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985
2986 /*
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002987 * Check if the backing address space wants to know that the page is
2988 * about to become writable
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989 */
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002990 if (vma->vm_ops->page_mkwrite) {
2991 unlock_page(fault_page);
2992 tmp = do_page_mkwrite(vma, fault_page, address);
2993 if (unlikely(!tmp ||
2994 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2995 page_cache_release(fault_page);
2996 return tmp;
2997 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998 }
2999
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003000 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3001 if (unlikely(!pte_same(*pte, orig_pte))) {
3002 pte_unmap_unlock(pte, ptl);
3003 unlock_page(fault_page);
3004 page_cache_release(fault_page);
3005 return ret;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003006 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003007 do_set_pte(vma, address, fault_page, pte, true, false);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003008 pte_unmap_unlock(pte, ptl);
3009
3010 if (set_page_dirty(fault_page))
3011 dirtied = 1;
3012 mapping = fault_page->mapping;
3013 unlock_page(fault_page);
3014 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3015 /*
3016 * Some device drivers do not set page.mapping but still
3017 * dirty their pages
3018 */
3019 balance_dirty_pages_ratelimited(mapping);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003020 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003021
3022 /* file_update_time outside page_lock */
3023 if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3024 file_update_time(vma->vm_file);
3025
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003026 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003027}
Nick Piggind00806b2007-07-19 01:46:57 -07003028
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003029/*
3030 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3031 * but allow concurrent faults).
3032 * The mmap_sem may have been released depending on flags and our
3033 * return value. See filemap_fault() and __lock_page_or_retry().
3034 */
Nick Piggin54cb8822007-07-19 01:46:59 -07003035static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3036 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003037 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003038{
3039 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003040 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003041
Hugh Dickins16abfa02007-10-04 16:56:06 +01003042 pte_unmap(page_table);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003043 if (!(flags & FAULT_FLAG_WRITE))
3044 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3045 orig_pte);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003046 if (!(vma->vm_flags & VM_SHARED))
3047 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3048 orig_pte);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003049 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003050}
3051
Jes Sorensenf4b81802006-09-27 01:50:10 -07003052/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 * Fault of a previously existing named mapping. Repopulate the pte
3054 * from the encoded file_pte if possible. This enables swappable
3055 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003056 *
3057 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3058 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003059 * We return with pte unmapped and unlocked.
3060 * The mmap_sem may have been released depending on flags and our
3061 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003063static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins65500d22005-10-29 18:15:59 -07003064 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003065 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066{
Hugh Dickins65500d22005-10-29 18:15:59 -07003067 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003069 flags |= FAULT_FLAG_NONLINEAR;
3070
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003071 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Nick Piggin83c54072007-07-19 01:47:05 -07003072 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073
Hugh Dickins2509ef22009-01-06 14:40:10 -08003074 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003075 /*
3076 * Page table corrupted: show pte and kill process.
3077 */
Hugh Dickins3dc14742009-01-06 14:40:08 -08003078 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08003079 return VM_FAULT_SIGBUS;
Hugh Dickins65500d22005-10-29 18:15:59 -07003080 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003081
3082 pgoff = pte_to_pgoff(orig_pte);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003083 if (!(flags & FAULT_FLAG_WRITE))
3084 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3085 orig_pte);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003086 if (!(vma->vm_flags & VM_SHARED))
3087 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3088 orig_pte);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003089 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090}
3091
Rashika Kheriab19a9932014-04-03 14:48:02 -07003092static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
Rik van Riel04bb2f92013-10-07 11:29:36 +01003093 unsigned long addr, int page_nid,
3094 int *flags)
Mel Gorman9532fec2012-11-15 01:24:32 +00003095{
3096 get_page(page);
3097
3098 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003099 if (page_nid == numa_node_id()) {
Mel Gorman9532fec2012-11-15 01:24:32 +00003100 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003101 *flags |= TNF_FAULT_LOCAL;
3102 }
Mel Gorman9532fec2012-11-15 01:24:32 +00003103
3104 return mpol_misplaced(page, vma, addr);
3105}
3106
Rashika Kheriab19a9932014-04-03 14:48:02 -07003107static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gormand10e63f2012-10-25 14:16:31 +02003108 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3109{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003110 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003111 spinlock_t *ptl;
Mel Gorman8191acb2013-10-07 11:28:45 +01003112 int page_nid = -1;
Peter Zijlstra90572892013-10-07 11:29:20 +01003113 int last_cpupid;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003114 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003115 bool migrated = false;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003116 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003117
3118 /*
3119 * The "pte" at this point cannot be used safely without
3120 * validation through pte_unmap_same(). It's of NUMA type but
3121 * the pfn may be screwed if the read is non atomic.
3122 *
3123 * ptep_modify_prot_start is not called as this is clearing
3124 * the _PAGE_NUMA bit and it is not really expected that there
3125 * would be concurrent hardware modifications to the PTE.
3126 */
3127 ptl = pte_lockptr(mm, pmd);
3128 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003129 if (unlikely(!pte_same(*ptep, pte))) {
3130 pte_unmap_unlock(ptep, ptl);
3131 goto out;
3132 }
3133
Mel Gormand10e63f2012-10-25 14:16:31 +02003134 pte = pte_mknonnuma(pte);
3135 set_pte_at(mm, addr, ptep, pte);
3136 update_mmu_cache(vma, addr, ptep);
3137
3138 page = vm_normal_page(vma, addr, pte);
3139 if (!page) {
3140 pte_unmap_unlock(ptep, ptl);
3141 return 0;
3142 }
Mel Gormana1a46182013-10-07 11:28:50 +01003143 BUG_ON(is_zero_pfn(page_to_pfn(page)));
Mel Gormand10e63f2012-10-25 14:16:31 +02003144
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003145 /*
3146 * Avoid grouping on DSO/COW pages in specific and RO pages
3147 * in general, RO pages shouldn't hurt as much anyway since
3148 * they can be in shared cache state.
3149 */
3150 if (!pte_write(pte))
3151 flags |= TNF_NO_GROUP;
3152
Rik van Rieldabe1d92013-10-07 11:29:34 +01003153 /*
3154 * Flag if the page is shared between multiple address spaces. This
3155 * is later used when determining whether to group tasks together
3156 */
3157 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3158 flags |= TNF_SHARED;
3159
Peter Zijlstra90572892013-10-07 11:29:20 +01003160 last_cpupid = page_cpupid_last(page);
Mel Gorman8191acb2013-10-07 11:28:45 +01003161 page_nid = page_to_nid(page);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003162 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003163 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003164 if (target_nid == -1) {
Mel Gorman4daae3b2012-11-02 11:33:45 +00003165 put_page(page);
3166 goto out;
3167 }
3168
3169 /* Migrate to the requested node */
Mel Gorman1bc115d2013-10-07 11:29:05 +01003170 migrated = migrate_misplaced_page(page, vma, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003171 if (migrated) {
Mel Gorman8191acb2013-10-07 11:28:45 +01003172 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003173 flags |= TNF_MIGRATED;
3174 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00003175
3176out:
Mel Gorman8191acb2013-10-07 11:28:45 +01003177 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003178 task_numa_fault(last_cpupid, page_nid, 1, flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003179 return 0;
3180}
3181
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182/*
3183 * These routines also need to handle stuff like marking pages dirty
3184 * and/or accessed for architectures that don't do it in hardware (most
3185 * RISC architectures). The early dirtying is also good on the i386.
3186 *
3187 * There is also a hook called "update_mmu_cache()" that architectures
3188 * with external mmu caches can use to update those (ie the Sparc or
3189 * PowerPC hashed page tables that act as extended TLBs).
3190 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003191 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3192 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003193 * We return with pte unmapped and unlocked.
3194 *
3195 * The mmap_sem may have been released depending on flags and our
3196 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197 */
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003198static int handle_pte_fault(struct mm_struct *mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003199 struct vm_area_struct *vma, unsigned long address,
3200 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201{
3202 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003203 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204
Hugh Dickinsc0d73262014-08-06 16:05:08 -07003205 entry = ACCESS_ONCE(*pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003207 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003208 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003209 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003210 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003211 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003212 }
3213 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003214 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003215 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003217 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003218 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003219 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003220 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003221 }
3222
Mel Gormand10e63f2012-10-25 14:16:31 +02003223 if (pte_numa(entry))
3224 return do_numa_page(mm, vma, address, entry, pte, pmd);
3225
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003226 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003227 spin_lock(ptl);
3228 if (unlikely(!pte_same(*pte, entry)))
3229 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003230 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003232 return do_wp_page(mm, vma, address,
3233 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003234 entry = pte_mkdirty(entry);
3235 }
3236 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003237 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003238 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003239 } else {
3240 /*
3241 * This is needed only for protection faults but the arch code
3242 * is not yet telling us if this is a protection fault or not.
3243 * This still avoids useless tlb flushes for .text page faults
3244 * with threads.
3245 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003246 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003247 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003248 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003249unlock:
3250 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003251 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252}
3253
3254/*
3255 * By the time we get here, we already hold the mm semaphore
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003256 *
3257 * The mmap_sem may have been released depending on flags and our
3258 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003259 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003260static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3261 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003262{
3263 pgd_t *pgd;
3264 pud_t *pud;
3265 pmd_t *pmd;
3266 pte_t *pte;
3267
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003268 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003269 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003272 pud = pud_alloc(mm, pgd, address);
3273 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003274 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003275 pmd = pmd_alloc(mm, pud, address);
3276 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003277 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003278 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003279 int ret = VM_FAULT_FALLBACK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003280 if (!vma->vm_ops)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003281 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3282 pmd, flags);
3283 if (!(ret & VM_FAULT_FALLBACK))
3284 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003285 } else {
3286 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003287 int ret;
3288
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003289 barrier();
3290 if (pmd_trans_huge(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003291 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3292
Linus Torvaldse53289c2013-01-09 08:36:54 -08003293 /*
3294 * If the pmd is splitting, return and retry the
3295 * the fault. Alternative: wait until the split
3296 * is done, and goto retry.
3297 */
3298 if (pmd_trans_splitting(orig_pmd))
3299 return 0;
3300
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003301 if (pmd_numa(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003302 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003303 orig_pmd, pmd);
3304
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003305 if (dirty && !pmd_write(orig_pmd)) {
David Rientjes1f1d06c2012-05-29 15:06:23 -07003306 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3307 orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003308 if (!(ret & VM_FAULT_FALLBACK))
3309 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003310 } else {
3311 huge_pmd_set_accessed(mm, vma, address, pmd,
3312 orig_pmd, dirty);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003313 return 0;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003314 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003315 }
3316 }
3317
3318 /*
3319 * Use __pte_alloc instead of pte_alloc_map, because we can't
3320 * run pte_offset_map on the pmd, if an huge pmd could
3321 * materialize from under us from a different thread.
3322 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003323 if (unlikely(pmd_none(*pmd)) &&
3324 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003325 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003326 /* if an huge pmd materialized from under us just retry later */
3327 if (unlikely(pmd_trans_huge(*pmd)))
3328 return 0;
3329 /*
3330 * A regular pmd is established and it can't morph into a huge pmd
3331 * from under us anymore at this point because we hold the mmap_sem
3332 * read mode and khugepaged takes it in write mode. So now it's
3333 * safe to run pte_offset_map().
3334 */
3335 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003337 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003338}
3339
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003340/*
3341 * By the time we get here, we already hold the mm semaphore
3342 *
3343 * The mmap_sem may have been released depending on flags and our
3344 * return value. See filemap_fault() and __lock_page_or_retry().
3345 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003346int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3347 unsigned long address, unsigned int flags)
3348{
3349 int ret;
3350
3351 __set_current_state(TASK_RUNNING);
3352
3353 count_vm_event(PGFAULT);
3354 mem_cgroup_count_vm_event(mm, PGFAULT);
3355
3356 /* do counter updates before entering really critical section. */
3357 check_sync_rss_stat(current);
3358
3359 /*
3360 * Enable the memcg OOM handling for faults triggered in user
3361 * space. Kernel faults are handled more gracefully.
3362 */
3363 if (flags & FAULT_FLAG_USER)
Johannes Weiner49426422013-10-16 13:46:59 -07003364 mem_cgroup_oom_enable();
Johannes Weiner519e5242013-09-12 15:13:42 -07003365
3366 ret = __handle_mm_fault(mm, vma, address, flags);
3367
Johannes Weiner49426422013-10-16 13:46:59 -07003368 if (flags & FAULT_FLAG_USER) {
3369 mem_cgroup_oom_disable();
3370 /*
3371 * The task may have entered a memcg OOM situation but
3372 * if the allocation error was handled gracefully (no
3373 * VM_FAULT_OOM), there is no need to kill anything.
3374 * Just clean up the OOM state peacefully.
3375 */
3376 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3377 mem_cgroup_oom_synchronize(false);
3378 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003379
Johannes Weiner519e5242013-09-12 15:13:42 -07003380 return ret;
3381}
3382
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383#ifndef __PAGETABLE_PUD_FOLDED
3384/*
3385 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003386 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003388int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003390 pud_t *new = pud_alloc_one(mm, address);
3391 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003392 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393
Nick Piggin362a61a2008-05-14 06:37:36 +02003394 smp_wmb(); /* See comment in __pte_alloc */
3395
Hugh Dickins872fec12005-10-29 18:16:21 -07003396 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003397 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003398 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003399 else
3400 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003401 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003402 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403}
3404#endif /* __PAGETABLE_PUD_FOLDED */
3405
3406#ifndef __PAGETABLE_PMD_FOLDED
3407/*
3408 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003409 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003411int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003413 pmd_t *new = pmd_alloc_one(mm, address);
3414 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003415 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416
Nick Piggin362a61a2008-05-14 06:37:36 +02003417 smp_wmb(); /* See comment in __pte_alloc */
3418
Hugh Dickins872fec12005-10-29 18:16:21 -07003419 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003421 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003422 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003423 else
3424 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003426 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003427 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003428 else
3429 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003431 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003432 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433}
3434#endif /* __PAGETABLE_PMD_FOLDED */
3435
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003436static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003437 pte_t **ptepp, spinlock_t **ptlp)
3438{
3439 pgd_t *pgd;
3440 pud_t *pud;
3441 pmd_t *pmd;
3442 pte_t *ptep;
3443
3444 pgd = pgd_offset(mm, address);
3445 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3446 goto out;
3447
3448 pud = pud_offset(pgd, address);
3449 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3450 goto out;
3451
3452 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003453 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003454 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3455 goto out;
3456
3457 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3458 if (pmd_huge(*pmd))
3459 goto out;
3460
3461 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3462 if (!ptep)
3463 goto out;
3464 if (!pte_present(*ptep))
3465 goto unlock;
3466 *ptepp = ptep;
3467 return 0;
3468unlock:
3469 pte_unmap_unlock(ptep, *ptlp);
3470out:
3471 return -EINVAL;
3472}
3473
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003474static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3475 pte_t **ptepp, spinlock_t **ptlp)
3476{
3477 int res;
3478
3479 /* (void) is needed to make gcc happy */
3480 (void) __cond_lock(*ptlp,
3481 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3482 return res;
3483}
3484
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003485/**
3486 * follow_pfn - look up PFN at a user virtual address
3487 * @vma: memory mapping
3488 * @address: user virtual address
3489 * @pfn: location to store found PFN
3490 *
3491 * Only IO mappings and raw PFN mappings are allowed.
3492 *
3493 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3494 */
3495int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3496 unsigned long *pfn)
3497{
3498 int ret = -EINVAL;
3499 spinlock_t *ptl;
3500 pte_t *ptep;
3501
3502 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3503 return ret;
3504
3505 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3506 if (ret)
3507 return ret;
3508 *pfn = pte_pfn(*ptep);
3509 pte_unmap_unlock(ptep, ptl);
3510 return 0;
3511}
3512EXPORT_SYMBOL(follow_pfn);
3513
Rik van Riel28b2ee22008-07-23 21:27:05 -07003514#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003515int follow_phys(struct vm_area_struct *vma,
3516 unsigned long address, unsigned int flags,
3517 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003518{
Johannes Weiner03668a42009-06-16 15:32:34 -07003519 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003520 pte_t *ptep, pte;
3521 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003522
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003523 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3524 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003525
Johannes Weiner03668a42009-06-16 15:32:34 -07003526 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003527 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003528 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003529
Rik van Riel28b2ee22008-07-23 21:27:05 -07003530 if ((flags & FOLL_WRITE) && !pte_write(pte))
3531 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003532
3533 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003534 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003535
Johannes Weiner03668a42009-06-16 15:32:34 -07003536 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003537unlock:
3538 pte_unmap_unlock(ptep, ptl);
3539out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003540 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003541}
3542
3543int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3544 void *buf, int len, int write)
3545{
3546 resource_size_t phys_addr;
3547 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003548 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003549 int offset = addr & (PAGE_SIZE-1);
3550
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003551 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003552 return -EINVAL;
3553
3554 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3555 if (write)
3556 memcpy_toio(maddr + offset, buf, len);
3557 else
3558 memcpy_fromio(buf, maddr + offset, len);
3559 iounmap(maddr);
3560
3561 return len;
3562}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02003563EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003564#endif
3565
David Howells0ec76a12006-09-27 01:50:15 -07003566/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003567 * Access another process' address space as given in mm. If non-NULL, use the
3568 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003569 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003570static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3571 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003572{
David Howells0ec76a12006-09-27 01:50:15 -07003573 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003574 void *old_buf = buf;
3575
David Howells0ec76a12006-09-27 01:50:15 -07003576 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003577 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003578 while (len) {
3579 int bytes, ret, offset;
3580 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003581 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003582
3583 ret = get_user_pages(tsk, mm, addr, 1,
3584 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003585 if (ret <= 0) {
Rik van Rieldbffcd02014-08-06 16:08:12 -07003586#ifndef CONFIG_HAVE_IOREMAP_PROT
3587 break;
3588#else
Rik van Riel28b2ee22008-07-23 21:27:05 -07003589 /*
3590 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3591 * we can access using slightly different code.
3592 */
Rik van Riel28b2ee22008-07-23 21:27:05 -07003593 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003594 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003595 break;
3596 if (vma->vm_ops && vma->vm_ops->access)
3597 ret = vma->vm_ops->access(vma, addr, buf,
3598 len, write);
3599 if (ret <= 0)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003600 break;
3601 bytes = ret;
Rik van Rieldbffcd02014-08-06 16:08:12 -07003602#endif
David Howells0ec76a12006-09-27 01:50:15 -07003603 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003604 bytes = len;
3605 offset = addr & (PAGE_SIZE-1);
3606 if (bytes > PAGE_SIZE-offset)
3607 bytes = PAGE_SIZE-offset;
3608
3609 maddr = kmap(page);
3610 if (write) {
3611 copy_to_user_page(vma, page, addr,
3612 maddr + offset, buf, bytes);
3613 set_page_dirty_lock(page);
3614 } else {
3615 copy_from_user_page(vma, page, addr,
3616 buf, maddr + offset, bytes);
3617 }
3618 kunmap(page);
3619 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003620 }
David Howells0ec76a12006-09-27 01:50:15 -07003621 len -= bytes;
3622 buf += bytes;
3623 addr += bytes;
3624 }
3625 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003626
3627 return buf - old_buf;
3628}
Andi Kleen03252912008-01-30 13:33:18 +01003629
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003630/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003631 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003632 * @mm: the mm_struct of the target address space
3633 * @addr: start address to access
3634 * @buf: source or destination buffer
3635 * @len: number of bytes to transfer
3636 * @write: whether the access is a write
3637 *
3638 * The caller must hold a reference on @mm.
3639 */
3640int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3641 void *buf, int len, int write)
3642{
3643 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3644}
3645
Andi Kleen03252912008-01-30 13:33:18 +01003646/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003647 * Access another process' address space.
3648 * Source/target buffer must be kernel space,
3649 * Do not walk the page table directly, use get_user_pages
3650 */
3651int access_process_vm(struct task_struct *tsk, unsigned long addr,
3652 void *buf, int len, int write)
3653{
3654 struct mm_struct *mm;
3655 int ret;
3656
3657 mm = get_task_mm(tsk);
3658 if (!mm)
3659 return 0;
3660
3661 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3662 mmput(mm);
3663
3664 return ret;
3665}
3666
Andi Kleen03252912008-01-30 13:33:18 +01003667/*
3668 * Print the name of a VMA.
3669 */
3670void print_vma_addr(char *prefix, unsigned long ip)
3671{
3672 struct mm_struct *mm = current->mm;
3673 struct vm_area_struct *vma;
3674
Ingo Molnare8bff742008-02-13 20:21:06 +01003675 /*
3676 * Do not print if we are in atomic
3677 * contexts (in exception stacks, etc.):
3678 */
3679 if (preempt_count())
3680 return;
3681
Andi Kleen03252912008-01-30 13:33:18 +01003682 down_read(&mm->mmap_sem);
3683 vma = find_vma(mm, ip);
3684 if (vma && vma->vm_file) {
3685 struct file *f = vma->vm_file;
3686 char *buf = (char *)__get_free_page(GFP_KERNEL);
3687 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003688 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01003689
Jan Blunckcf28b482008-02-14 19:38:44 -08003690 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003691 if (IS_ERR(p))
3692 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003693 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01003694 vma->vm_start,
3695 vma->vm_end - vma->vm_start);
3696 free_page((unsigned long)buf);
3697 }
3698 }
Jeff Liu51a07e52012-07-31 16:43:18 -07003699 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01003700}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003701
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003702#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003703void might_fault(void)
3704{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003705 /*
3706 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3707 * holding the mmap_sem, this is safe because kernel memory doesn't
3708 * get paged out, therefore we'll never actually fault, and the
3709 * below annotations will generate false positives.
3710 */
3711 if (segment_eq(get_fs(), KERNEL_DS))
3712 return;
3713
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003714 /*
3715 * it would be nicer only to annotate paths which are not under
3716 * pagefault_disable, however that requires a larger audit and
3717 * providing helpers like get_user_atomic.
3718 */
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003719 if (in_atomic())
3720 return;
3721
3722 __might_sleep(__FILE__, __LINE__, 0);
3723
3724 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003725 might_lock_read(&current->mm->mmap_sem);
3726}
3727EXPORT_SYMBOL(might_fault);
3728#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003729
3730#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3731static void clear_gigantic_page(struct page *page,
3732 unsigned long addr,
3733 unsigned int pages_per_huge_page)
3734{
3735 int i;
3736 struct page *p = page;
3737
3738 might_sleep();
3739 for (i = 0; i < pages_per_huge_page;
3740 i++, p = mem_map_next(p, page, i)) {
3741 cond_resched();
3742 clear_user_highpage(p, addr + i * PAGE_SIZE);
3743 }
3744}
3745void clear_huge_page(struct page *page,
3746 unsigned long addr, unsigned int pages_per_huge_page)
3747{
3748 int i;
3749
3750 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3751 clear_gigantic_page(page, addr, pages_per_huge_page);
3752 return;
3753 }
3754
3755 might_sleep();
3756 for (i = 0; i < pages_per_huge_page; i++) {
3757 cond_resched();
3758 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3759 }
3760}
3761
3762static void copy_user_gigantic_page(struct page *dst, struct page *src,
3763 unsigned long addr,
3764 struct vm_area_struct *vma,
3765 unsigned int pages_per_huge_page)
3766{
3767 int i;
3768 struct page *dst_base = dst;
3769 struct page *src_base = src;
3770
3771 for (i = 0; i < pages_per_huge_page; ) {
3772 cond_resched();
3773 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3774
3775 i++;
3776 dst = mem_map_next(dst, dst_base, i);
3777 src = mem_map_next(src, src_base, i);
3778 }
3779}
3780
3781void copy_user_huge_page(struct page *dst, struct page *src,
3782 unsigned long addr, struct vm_area_struct *vma,
3783 unsigned int pages_per_huge_page)
3784{
3785 int i;
3786
3787 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3788 copy_user_gigantic_page(dst, src, addr, vma,
3789 pages_per_huge_page);
3790 return;
3791 }
3792
3793 might_sleep();
3794 for (i = 0; i < pages_per_huge_page; i++) {
3795 cond_resched();
3796 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3797 }
3798}
3799#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003800
Olof Johansson40b64ac2013-12-20 14:28:05 -08003801#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003802
3803static struct kmem_cache *page_ptl_cachep;
3804
3805void __init ptlock_cache_init(void)
3806{
3807 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3808 SLAB_PANIC, NULL);
3809}
3810
Peter Zijlstra539edb52013-11-14 14:31:52 -08003811bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003812{
3813 spinlock_t *ptl;
3814
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003815 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003816 if (!ptl)
3817 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08003818 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003819 return true;
3820}
3821
Peter Zijlstra539edb52013-11-14 14:31:52 -08003822void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003823{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003824 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003825}
3826#endif