blob: 54f3a9b0095600749fda793d7e6067bd2c7f997d [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;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700223 tlb->local.next = NULL;
224 tlb->local.nr = 0;
225 tlb->local.max = ARRAY_SIZE(tlb->__pages);
226 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800227 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700228
229#ifdef CONFIG_HAVE_RCU_TABLE_FREE
230 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800231#endif
Will Deaconfb7332a2014-10-29 10:03:09 +0000232
233 __tlb_reset_range(tlb);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700234}
235
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700236static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700237{
Will Deacon721c21c2015-01-12 19:10:55 +0000238 if (!tlb->end)
239 return;
240
Peter Zijlstra9547d012011-05-24 17:12:14 -0700241 tlb_flush(tlb);
Joerg Roedel34ee6452014-11-13 13:46:09 +1100242 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700243#ifdef CONFIG_HAVE_RCU_TABLE_FREE
244 tlb_table_flush(tlb);
245#endif
Will Deaconfb7332a2014-10-29 10:03:09 +0000246 __tlb_reset_range(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700247}
248
249static void tlb_flush_mmu_free(struct mmu_gather *tlb)
250{
251 struct mmu_gather_batch *batch;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700252
Will Deacon721c21c2015-01-12 19:10:55 +0000253 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
Peter Zijlstra9547d012011-05-24 17:12:14 -0700254 free_pages_and_swap_cache(batch->pages, batch->nr);
255 batch->nr = 0;
256 }
257 tlb->active = &tlb->local;
258}
259
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700260void tlb_flush_mmu(struct mmu_gather *tlb)
261{
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700262 tlb_flush_mmu_tlbonly(tlb);
263 tlb_flush_mmu_free(tlb);
264}
265
Peter Zijlstra9547d012011-05-24 17:12:14 -0700266/* tlb_finish_mmu
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
269 */
270void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
271{
272 struct mmu_gather_batch *batch, *next;
273
274 tlb_flush_mmu(tlb);
275
276 /* keep the page table cache within bounds */
277 check_pgt_cache();
278
279 for (batch = tlb->local.next; batch; batch = next) {
280 next = batch->next;
281 free_pages((unsigned long)batch, 0);
282 }
283 tlb->local.next = NULL;
284}
285
286/* __tlb_remove_page
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
291 */
292int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
293{
294 struct mmu_gather_batch *batch;
295
Will Deaconfb7332a2014-10-29 10:03:09 +0000296 VM_BUG_ON(!tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700297
Peter Zijlstra9547d012011-05-24 17:12:14 -0700298 batch = tlb->active;
299 batch->pages[batch->nr++] = page;
300 if (batch->nr == batch->max) {
301 if (!tlb_next_batch(tlb))
302 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700303 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700304 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800305 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700306
307 return batch->max - batch->nr;
308}
309
310#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800311
Peter Zijlstra26723912011-05-24 17:12:00 -0700312#ifdef CONFIG_HAVE_RCU_TABLE_FREE
313
314/*
315 * See the comment near struct mmu_table_batch.
316 */
317
318static void tlb_remove_table_smp_sync(void *arg)
319{
320 /* Simply deliver the interrupt */
321}
322
323static void tlb_remove_table_one(void *table)
324{
325 /*
326 * This isn't an RCU grace period and hence the page-tables cannot be
327 * assumed to be actually RCU-freed.
328 *
329 * It is however sufficient for software page-table walkers that rely on
330 * IRQ disabling. See the comment near struct mmu_table_batch.
331 */
332 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
333 __tlb_remove_table(table);
334}
335
336static void tlb_remove_table_rcu(struct rcu_head *head)
337{
338 struct mmu_table_batch *batch;
339 int i;
340
341 batch = container_of(head, struct mmu_table_batch, rcu);
342
343 for (i = 0; i < batch->nr; i++)
344 __tlb_remove_table(batch->tables[i]);
345
346 free_page((unsigned long)batch);
347}
348
349void tlb_table_flush(struct mmu_gather *tlb)
350{
351 struct mmu_table_batch **batch = &tlb->batch;
352
353 if (*batch) {
354 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
355 *batch = NULL;
356 }
357}
358
359void tlb_remove_table(struct mmu_gather *tlb, void *table)
360{
361 struct mmu_table_batch **batch = &tlb->batch;
362
Peter Zijlstra26723912011-05-24 17:12:00 -0700363 /*
364 * When there's less then two users of this mm there cannot be a
365 * concurrent page-table walk.
366 */
367 if (atomic_read(&tlb->mm->mm_users) < 2) {
368 __tlb_remove_table(table);
369 return;
370 }
371
372 if (*batch == NULL) {
373 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
374 if (*batch == NULL) {
375 tlb_remove_table_one(table);
376 return;
377 }
378 (*batch)->nr = 0;
379 }
380 (*batch)->tables[(*batch)->nr++] = table;
381 if ((*batch)->nr == MAX_TABLE_BATCH)
382 tlb_table_flush(tlb);
383}
384
Peter Zijlstra9547d012011-05-24 17:12:14 -0700385#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700386
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 * Note: this doesn't free the actual pages themselves. That
389 * has been handled earlier when unmapping all the memory regions.
390 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000391static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
392 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800394 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700395 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000396 pte_free_tlb(tlb, token, addr);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800397 atomic_long_dec(&tlb->mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398}
399
Hugh Dickinse0da3822005-04-19 13:29:15 -0700400static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
401 unsigned long addr, unsigned long end,
402 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403{
404 pmd_t *pmd;
405 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700406 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
Hugh Dickinse0da3822005-04-19 13:29:15 -0700408 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 do {
411 next = pmd_addr_end(addr, end);
412 if (pmd_none_or_clear_bad(pmd))
413 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000414 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 } while (pmd++, addr = next, addr != end);
416
Hugh Dickinse0da3822005-04-19 13:29:15 -0700417 start &= PUD_MASK;
418 if (start < floor)
419 return;
420 if (ceiling) {
421 ceiling &= PUD_MASK;
422 if (!ceiling)
423 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700425 if (end - 1 > ceiling - 1)
426 return;
427
428 pmd = pmd_offset(pud, start);
429 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000430 pmd_free_tlb(tlb, pmd, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431}
432
Hugh Dickinse0da3822005-04-19 13:29:15 -0700433static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
434 unsigned long addr, unsigned long end,
435 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436{
437 pud_t *pud;
438 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700439 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440
Hugh Dickinse0da3822005-04-19 13:29:15 -0700441 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 do {
444 next = pud_addr_end(addr, end);
445 if (pud_none_or_clear_bad(pud))
446 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700447 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 } while (pud++, addr = next, addr != end);
449
Hugh Dickinse0da3822005-04-19 13:29:15 -0700450 start &= PGDIR_MASK;
451 if (start < floor)
452 return;
453 if (ceiling) {
454 ceiling &= PGDIR_MASK;
455 if (!ceiling)
456 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700458 if (end - 1 > ceiling - 1)
459 return;
460
461 pud = pud_offset(pgd, start);
462 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000463 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464}
465
466/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700467 * This function frees user-level page tables of a process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 */
Jan Beulich42b77722008-07-23 21:27:10 -0700469void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700470 unsigned long addr, unsigned long end,
471 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472{
473 pgd_t *pgd;
474 unsigned long next;
475
Hugh Dickinse0da3822005-04-19 13:29:15 -0700476 /*
477 * The next few lines have given us lots of grief...
478 *
479 * Why are we testing PMD* at this top level? Because often
480 * there will be no work to do at all, and we'd prefer not to
481 * go all the way down to the bottom just to discover that.
482 *
483 * Why all these "- 1"s? Because 0 represents both the bottom
484 * of the address space and the top of it (using -1 for the
485 * top wouldn't help much: the masks would do the wrong thing).
486 * The rule is that addr 0 and floor 0 refer to the bottom of
487 * the address space, but end 0 and ceiling 0 refer to the top
488 * Comparisons need to use "end - 1" and "ceiling - 1" (though
489 * that end 0 case should be mythical).
490 *
491 * Wherever addr is brought up or ceiling brought down, we must
492 * be careful to reject "the opposite 0" before it confuses the
493 * subsequent tests. But what about where end is brought down
494 * by PMD_SIZE below? no, end can't go down to 0 there.
495 *
496 * Whereas we round start (addr) and ceiling down, by different
497 * masks at different levels, in order to test whether a table
498 * now has no other vmas using it, so can be freed, we don't
499 * bother to round floor or end up - the tests don't need that.
500 */
501
502 addr &= PMD_MASK;
503 if (addr < floor) {
504 addr += PMD_SIZE;
505 if (!addr)
506 return;
507 }
508 if (ceiling) {
509 ceiling &= PMD_MASK;
510 if (!ceiling)
511 return;
512 }
513 if (end - 1 > ceiling - 1)
514 end -= PMD_SIZE;
515 if (addr > end - 1)
516 return;
517
Jan Beulich42b77722008-07-23 21:27:10 -0700518 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 do {
520 next = pgd_addr_end(addr, end);
521 if (pgd_none_or_clear_bad(pgd))
522 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700523 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700525}
526
Jan Beulich42b77722008-07-23 21:27:10 -0700527void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700528 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700529{
530 while (vma) {
531 struct vm_area_struct *next = vma->vm_next;
532 unsigned long addr = vma->vm_start;
533
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700534 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000535 * Hide vma from rmap and truncate_pagecache before freeing
536 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700537 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800538 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700539 unlink_file_vma(vma);
540
David Gibson9da61ae2006-03-22 00:08:57 -0800541 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700542 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700543 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700544 } else {
545 /*
546 * Optimization: gather nearby vmas into one call down
547 */
548 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800549 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700550 vma = next;
551 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800552 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700553 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700554 }
555 free_pgd_range(tlb, addr, vma->vm_end,
556 floor, next? next->vm_start: ceiling);
557 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700558 vma = next;
559 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560}
561
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800562int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
563 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800565 spinlock_t *ptl;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800566 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800567 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700568 if (!new)
569 return -ENOMEM;
570
Nick Piggin362a61a2008-05-14 06:37:36 +0200571 /*
572 * Ensure all pte setup (eg. pte page lock and page clearing) are
573 * visible before the pte is made visible to other CPUs by being
574 * put into page tables.
575 *
576 * The other side of the story is the pointer chasing in the page
577 * table walking code (when walking the page table without locking;
578 * ie. most of the time). Fortunately, these data accesses consist
579 * of a chain of data-dependent loads, meaning most CPUs (alpha
580 * being the notable exception) will already guarantee loads are
581 * seen in-order. See the alpha page table accessors for the
582 * smp_read_barrier_depends() barriers in page table walking code.
583 */
584 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
585
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800586 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800587 wait_split_huge_page = 0;
588 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800589 atomic_long_inc(&mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800591 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800592 } else if (unlikely(pmd_trans_splitting(*pmd)))
593 wait_split_huge_page = 1;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800594 spin_unlock(ptl);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800595 if (new)
596 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800597 if (wait_split_huge_page)
598 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700599 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600}
601
Hugh Dickins1bb36302005-10-29 18:16:22 -0700602int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700604 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
605 if (!new)
606 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607
Nick Piggin362a61a2008-05-14 06:37:36 +0200608 smp_wmb(); /* See comment in __pte_alloc */
609
Hugh Dickins1bb36302005-10-29 18:16:22 -0700610 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800611 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700612 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800613 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800614 } else
615 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700616 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800617 if (new)
618 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700619 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620}
621
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800622static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700623{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800624 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
625}
626
627static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
628{
629 int i;
630
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800631 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700632 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800633 for (i = 0; i < NR_MM_COUNTERS; i++)
634 if (rss[i])
635 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700636}
637
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800639 * This function is called to print an error when a bad pte
640 * is found. For example, we might have a PFN-mapped pte in
641 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700642 *
643 * The calling function must still handle the error.
644 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800645static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
646 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700647{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800648 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
649 pud_t *pud = pud_offset(pgd, addr);
650 pmd_t *pmd = pmd_offset(pud, addr);
651 struct address_space *mapping;
652 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800653 static unsigned long resume;
654 static unsigned long nr_shown;
655 static unsigned long nr_unshown;
656
657 /*
658 * Allow a burst of 60 reports, then keep quiet for that minute;
659 * or allow a steady drip of one report per second.
660 */
661 if (nr_shown == 60) {
662 if (time_before(jiffies, resume)) {
663 nr_unshown++;
664 return;
665 }
666 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800667 printk(KERN_ALERT
668 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800669 nr_unshown);
670 nr_unshown = 0;
671 }
672 nr_shown = 0;
673 }
674 if (nr_shown++ == 0)
675 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800676
677 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
678 index = linear_page_index(vma, addr);
679
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800680 printk(KERN_ALERT
681 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800682 current->comm,
683 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800684 if (page)
Dave Hansenf0b791a2014-01-23 15:52:49 -0800685 dump_page(page, "bad pte");
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800686 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800687 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
688 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
689 /*
690 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
691 */
692 if (vma->vm_ops)
Joe Perches071361d2012-12-12 10:19:12 -0800693 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
694 vma->vm_ops->fault);
Al Viro72c2d532013-09-22 16:27:52 -0400695 if (vma->vm_file)
Joe Perches071361d2012-12-12 10:19:12 -0800696 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
697 vma->vm_file->f_op->mmap);
Nick Pigginb5810032005-10-29 18:16:12 -0700698 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030699 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700700}
701
702/*
Nick Piggin7e675132008-04-28 02:13:00 -0700703 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800704 *
Nick Piggin7e675132008-04-28 02:13:00 -0700705 * "Special" mappings do not wish to be associated with a "struct page" (either
706 * it doesn't exist, or it exists but they don't want to touch it). In this
707 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700708 *
Nick Piggin7e675132008-04-28 02:13:00 -0700709 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
710 * pte bit, in which case this function is trivial. Secondly, an architecture
711 * may not have a spare pte bit, which requires a more complicated scheme,
712 * described below.
713 *
714 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
715 * special mapping (even if there are underlying and valid "struct pages").
716 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800717 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700718 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
719 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700720 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
721 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800722 *
723 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
724 *
Nick Piggin7e675132008-04-28 02:13:00 -0700725 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700726 *
Nick Piggin7e675132008-04-28 02:13:00 -0700727 * This restricts such mappings to be a linear translation from virtual address
728 * to pfn. To get around this restriction, we allow arbitrary mappings so long
729 * as the vma is not a COW mapping; in that case, we know that all ptes are
730 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700731 *
732 *
Nick Piggin7e675132008-04-28 02:13:00 -0700733 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
734 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700735 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
736 * page" backing, however the difference is that _all_ pages with a struct
737 * page (that is, those where pfn_valid is true) are refcounted and considered
738 * normal pages by the VM. The disadvantage is that pages are refcounted
739 * (which can be slower and simply not an option for some PFNMAP users). The
740 * advantage is that we don't have to follow the strict linearity rule of
741 * PFNMAP mappings in order to support COWable mappings.
742 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800743 */
Nick Piggin7e675132008-04-28 02:13:00 -0700744#ifdef __HAVE_ARCH_PTE_SPECIAL
745# define HAVE_PTE_SPECIAL 1
746#else
747# define HAVE_PTE_SPECIAL 0
748#endif
749struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
750 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800751{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800752 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700753
754 if (HAVE_PTE_SPECIAL) {
Hugh Dickinsb38af472014-08-29 15:18:44 -0700755 if (likely(!pte_special(pte)))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800756 goto check_pfn;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700757 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
758 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700759 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800760 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700761 return NULL;
762 }
763
764 /* !HAVE_PTE_SPECIAL case follows: */
765
Jared Hulbertb379d792008-04-28 02:12:58 -0700766 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
767 if (vma->vm_flags & VM_MIXEDMAP) {
768 if (!pfn_valid(pfn))
769 return NULL;
770 goto out;
771 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700772 unsigned long off;
773 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700774 if (pfn == vma->vm_pgoff + off)
775 return NULL;
776 if (!is_cow_mapping(vma->vm_flags))
777 return NULL;
778 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800779 }
780
Hugh Dickinsb38af472014-08-29 15:18:44 -0700781 if (is_zero_pfn(pfn))
782 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800783check_pfn:
784 if (unlikely(pfn > highest_memmap_pfn)) {
785 print_bad_pte(vma, addr, pte, NULL);
786 return NULL;
787 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800788
789 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700790 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700791 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800792 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700793out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800794 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800795}
796
797/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 * copy one vm_area from one task to the other. Assumes the page tables
799 * already present in the new task to be cleared in the whole range
800 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 */
802
Hugh Dickins570a335b2009-12-14 17:58:46 -0800803static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700805 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700806 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807{
Nick Pigginb5810032005-10-29 18:16:12 -0700808 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 pte_t pte = *src_pte;
810 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811
812 /* pte contains position in swap or file, so copy. */
813 if (unlikely(!pte_present(pte))) {
814 if (!pte_file(pte)) {
Christoph Lameter06972122006-06-23 02:03:35 -0700815 swp_entry_t entry = pte_to_swp_entry(pte);
816
Hugh Dickins2022b4d12014-12-02 15:59:39 -0800817 if (likely(!non_swap_entry(entry))) {
818 if (swap_duplicate(entry) < 0)
819 return entry.val;
Hugh Dickins570a335b2009-12-14 17:58:46 -0800820
Hugh Dickins2022b4d12014-12-02 15:59:39 -0800821 /* make sure dst_mm is on swapoff's mmlist. */
822 if (unlikely(list_empty(&dst_mm->mmlist))) {
823 spin_lock(&mmlist_lock);
824 if (list_empty(&dst_mm->mmlist))
825 list_add(&dst_mm->mmlist,
826 &src_mm->mmlist);
827 spin_unlock(&mmlist_lock);
828 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800829 rss[MM_SWAPENTS]++;
Hugh Dickins2022b4d12014-12-02 15:59:39 -0800830 } else if (is_migration_entry(entry)) {
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800831 page = migration_entry_to_page(entry);
832
833 if (PageAnon(page))
834 rss[MM_ANONPAGES]++;
835 else
836 rss[MM_FILEPAGES]++;
837
838 if (is_write_migration_entry(entry) &&
839 is_cow_mapping(vm_flags)) {
840 /*
841 * COW mappings require pages in both
842 * parent and child to be set to read.
843 */
844 make_migration_entry_read(&entry);
845 pte = swp_entry_to_pte(entry);
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700846 if (pte_swp_soft_dirty(*src_pte))
847 pte = pte_swp_mksoft_dirty(pte);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800848 set_pte_at(src_mm, addr, src_pte, pte);
849 }
Christoph Lameter06972122006-06-23 02:03:35 -0700850 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700852 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 }
854
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 /*
856 * If it's a COW mapping, write protect it both
857 * in the parent and the child
858 */
Linus Torvalds67121172005-12-11 20:38:17 -0800859 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700861 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 }
863
864 /*
865 * If it's a shared mapping, mark it clean in
866 * the child
867 */
868 if (vm_flags & VM_SHARED)
869 pte = pte_mkclean(pte);
870 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800871
872 page = vm_normal_page(vma, addr, pte);
873 if (page) {
874 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700875 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800876 if (PageAnon(page))
877 rss[MM_ANONPAGES]++;
878 else
879 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800880 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700881
882out_set_pte:
883 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800884 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885}
886
Jerome Marchand21bda262014-08-06 16:06:56 -0700887static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800888 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
889 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700891 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700893 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700894 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800895 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800896 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897
898again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800899 init_rss_vec(rss);
900
Hugh Dickinsc74df322005-10-29 18:16:23 -0700901 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 if (!dst_pte)
903 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700904 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700905 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700906 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700907 orig_src_pte = src_pte;
908 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700909 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 do {
912 /*
913 * We are holding two locks at this point - either of them
914 * could generate latencies in another task on another CPU.
915 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700916 if (progress >= 32) {
917 progress = 0;
918 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100919 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700920 break;
921 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 if (pte_none(*src_pte)) {
923 progress++;
924 continue;
925 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800926 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
927 vma, addr, rss);
928 if (entry.val)
929 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 progress += 8;
931 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700933 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700934 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700935 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800936 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700937 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700938 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800939
940 if (entry.val) {
941 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
942 return -ENOMEM;
943 progress = 0;
944 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 if (addr != end)
946 goto again;
947 return 0;
948}
949
950static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
951 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
952 unsigned long addr, unsigned long end)
953{
954 pmd_t *src_pmd, *dst_pmd;
955 unsigned long next;
956
957 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
958 if (!dst_pmd)
959 return -ENOMEM;
960 src_pmd = pmd_offset(src_pud, addr);
961 do {
962 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800963 if (pmd_trans_huge(*src_pmd)) {
964 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800965 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800966 err = copy_huge_pmd(dst_mm, src_mm,
967 dst_pmd, src_pmd, addr, vma);
968 if (err == -ENOMEM)
969 return -ENOMEM;
970 if (!err)
971 continue;
972 /* fall through */
973 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 if (pmd_none_or_clear_bad(src_pmd))
975 continue;
976 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
977 vma, addr, next))
978 return -ENOMEM;
979 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
980 return 0;
981}
982
983static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
984 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
985 unsigned long addr, unsigned long end)
986{
987 pud_t *src_pud, *dst_pud;
988 unsigned long next;
989
990 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
991 if (!dst_pud)
992 return -ENOMEM;
993 src_pud = pud_offset(src_pgd, addr);
994 do {
995 next = pud_addr_end(addr, end);
996 if (pud_none_or_clear_bad(src_pud))
997 continue;
998 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
999 vma, addr, next))
1000 return -ENOMEM;
1001 } while (dst_pud++, src_pud++, addr = next, addr != end);
1002 return 0;
1003}
1004
1005int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1006 struct vm_area_struct *vma)
1007{
1008 pgd_t *src_pgd, *dst_pgd;
1009 unsigned long next;
1010 unsigned long addr = vma->vm_start;
1011 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001012 unsigned long mmun_start; /* For mmu_notifiers */
1013 unsigned long mmun_end; /* For mmu_notifiers */
1014 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001015 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016
Nick Piggind9928952005-08-28 16:49:11 +10001017 /*
1018 * Don't copy ptes where a page fault will fill them correctly.
1019 * Fork becomes much lighter when there are big shared or private
1020 * readonly mappings. The tradeoff is that copy_page_range is more
1021 * efficient than faulting.
1022 */
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001023 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1024 VM_PFNMAP | VM_MIXEDMAP))) {
Nick Piggind9928952005-08-28 16:49:11 +10001025 if (!vma->anon_vma)
1026 return 0;
1027 }
1028
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 if (is_vm_hugetlb_page(vma))
1030 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1031
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001032 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001033 /*
1034 * We do not free on error cases below as remove_vma
1035 * gets called on error from higher level routine
1036 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001037 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001038 if (ret)
1039 return ret;
1040 }
1041
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001042 /*
1043 * We need to invalidate the secondary MMU mappings only when
1044 * there could be a permission downgrade on the ptes of the
1045 * parent mm. And a permission downgrade will only happen if
1046 * is_cow_mapping() returns true.
1047 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001048 is_cow = is_cow_mapping(vma->vm_flags);
1049 mmun_start = addr;
1050 mmun_end = end;
1051 if (is_cow)
1052 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1053 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001054
1055 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056 dst_pgd = pgd_offset(dst_mm, addr);
1057 src_pgd = pgd_offset(src_mm, addr);
1058 do {
1059 next = pgd_addr_end(addr, end);
1060 if (pgd_none_or_clear_bad(src_pgd))
1061 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001062 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1063 vma, addr, next))) {
1064 ret = -ENOMEM;
1065 break;
1066 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001068
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001069 if (is_cow)
1070 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001071 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072}
1073
Robin Holt51c6f662005-11-13 16:06:42 -08001074static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001075 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001077 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078{
Nick Pigginb5810032005-10-29 18:16:12 -07001079 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001080 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001081 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001082 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001083 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001084 pte_t *pte;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001085
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001086again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001087 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001088 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1089 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001090 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 do {
1092 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001093 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001095 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001096
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001098 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001099
Linus Torvalds6aab3412005-11-28 14:34:23 -08001100 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 if (unlikely(details) && page) {
1102 /*
1103 * unmap_shared_mapping_pages() wants to
1104 * invalidate cache without truncating:
1105 * unmap shared but keep private pages.
1106 */
1107 if (details->check_mapping &&
1108 details->check_mapping != page->mapping)
1109 continue;
1110 /*
1111 * Each page->index must be checked when
1112 * invalidating or truncating nonlinear.
1113 */
1114 if (details->nonlinear_vma &&
1115 (page->index < details->first_index ||
1116 page->index > details->last_index))
1117 continue;
1118 }
Nick Pigginb5810032005-10-29 18:16:12 -07001119 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001120 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 tlb_remove_tlb_entry(tlb, pte, addr);
1122 if (unlikely(!page))
1123 continue;
1124 if (unlikely(details) && details->nonlinear_vma
1125 && linear_page_index(details->nonlinear_vma,
Cyrill Gorcunov41bb3472013-08-13 16:00:51 -07001126 addr) != page->index) {
1127 pte_t ptfile = pgoff_to_pte(page->index);
1128 if (pte_soft_dirty(ptent))
Peter Feinerdbab31a2014-09-25 16:05:29 -07001129 ptfile = pte_file_mksoft_dirty(ptfile);
Cyrill Gorcunov41bb3472013-08-13 16:00:51 -07001130 set_pte_at(mm, addr, pte, ptfile);
1131 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001133 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001134 else {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001135 if (pte_dirty(ptent)) {
1136 force_flush = 1;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001137 set_page_dirty(page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001138 }
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001139 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001140 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001141 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001142 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001143 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001144 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001145 if (unlikely(page_mapcount(page) < 0))
1146 print_bad_pte(vma, addr, ptent, page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001147 if (unlikely(!__tlb_remove_page(tlb, page))) {
1148 force_flush = 1;
Will Deaconce9ec372014-10-28 13:16:28 -07001149 addr += PAGE_SIZE;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001150 break;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001151 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 continue;
1153 }
1154 /*
1155 * If details->check_mapping, we leave swap entries;
1156 * if details->nonlinear_vma, we leave file entries.
1157 */
1158 if (unlikely(details))
1159 continue;
Hugh Dickins2509ef22009-01-06 14:40:10 -08001160 if (pte_file(ptent)) {
1161 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1162 print_bad_pte(vma, addr, ptent, NULL);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001163 } else {
1164 swp_entry_t entry = pte_to_swp_entry(ptent);
1165
1166 if (!non_swap_entry(entry))
1167 rss[MM_SWAPENTS]--;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001168 else if (is_migration_entry(entry)) {
1169 struct page *page;
1170
1171 page = migration_entry_to_page(entry);
1172
1173 if (PageAnon(page))
1174 rss[MM_ANONPAGES]--;
1175 else
1176 rss[MM_FILEPAGES]--;
1177 }
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001178 if (unlikely(!free_swap_and_cache(entry)))
1179 print_bad_pte(vma, addr, ptent, NULL);
1180 }
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001181 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001182 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001183
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001184 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001185 arch_leave_lazy_mmu_mode();
Robin Holt51c6f662005-11-13 16:06:42 -08001186
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001187 /* Do the actual TLB flush before dropping ptl */
Will Deaconfb7332a2014-10-29 10:03:09 +00001188 if (force_flush)
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001189 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001190 pte_unmap_unlock(start_pte, ptl);
1191
1192 /*
1193 * If we forced a TLB flush (either due to running out of
1194 * batch buffers or because we needed to flush dirty TLB
1195 * entries before releasing the ptl), free the batched
1196 * memory too. Restart if we didn't do everything.
1197 */
1198 if (force_flush) {
1199 force_flush = 0;
1200 tlb_flush_mmu_free(tlb);
Linus Torvalds2b047252013-08-15 11:42:25 -07001201
1202 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001203 goto again;
1204 }
1205
Robin Holt51c6f662005-11-13 16:06:42 -08001206 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207}
1208
Robin Holt51c6f662005-11-13 16:06:42 -08001209static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001210 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001212 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213{
1214 pmd_t *pmd;
1215 unsigned long next;
1216
1217 pmd = pmd_offset(pud, addr);
1218 do {
1219 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001220 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001221 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001222#ifdef CONFIG_DEBUG_VM
1223 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1224 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1225 __func__, addr, end,
1226 vma->vm_start,
1227 vma->vm_end);
1228 BUG();
1229 }
1230#endif
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001231 split_huge_page_pmd(vma, addr, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001232 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001233 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001234 /* fall through */
1235 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001236 /*
1237 * Here there can be other concurrent MADV_DONTNEED or
1238 * trans huge page faults running, and if the pmd is
1239 * none or trans huge it can change under us. This is
1240 * because MADV_DONTNEED holds the mmap_sem in read
1241 * mode.
1242 */
1243 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1244 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001245 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001246next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001247 cond_resched();
1248 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001249
1250 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251}
1252
Robin Holt51c6f662005-11-13 16:06:42 -08001253static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001254 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001256 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257{
1258 pud_t *pud;
1259 unsigned long next;
1260
1261 pud = pud_offset(pgd, addr);
1262 do {
1263 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001264 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001266 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1267 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001268
1269 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270}
1271
Al Viro038c7aa2012-03-05 13:25:09 -05001272static void unmap_page_range(struct mmu_gather *tlb,
1273 struct vm_area_struct *vma,
1274 unsigned long addr, unsigned long end,
1275 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276{
1277 pgd_t *pgd;
1278 unsigned long next;
1279
1280 if (details && !details->check_mapping && !details->nonlinear_vma)
1281 details = NULL;
1282
1283 BUG_ON(addr >= end);
1284 tlb_start_vma(tlb, vma);
1285 pgd = pgd_offset(vma->vm_mm, addr);
1286 do {
1287 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001288 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001290 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1291 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 tlb_end_vma(tlb, vma);
1293}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294
Al Virof5cc4ee2012-03-05 14:14:20 -05001295
1296static void unmap_single_vma(struct mmu_gather *tlb,
1297 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001298 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001299 struct zap_details *details)
1300{
1301 unsigned long start = max(vma->vm_start, start_addr);
1302 unsigned long end;
1303
1304 if (start >= vma->vm_end)
1305 return;
1306 end = min(vma->vm_end, end_addr);
1307 if (end <= vma->vm_start)
1308 return;
1309
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301310 if (vma->vm_file)
1311 uprobe_munmap(vma, start, end);
1312
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001313 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001314 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001315
1316 if (start != end) {
1317 if (unlikely(is_vm_hugetlb_page(vma))) {
1318 /*
1319 * It is undesirable to test vma->vm_file as it
1320 * should be non-null for valid hugetlb area.
1321 * However, vm_file will be NULL in the error
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001322 * cleanup path of mmap_region. When
Al Virof5cc4ee2012-03-05 14:14:20 -05001323 * hugetlbfs ->mmap method fails,
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001324 * mmap_region() nullifies vma->vm_file
Al Virof5cc4ee2012-03-05 14:14:20 -05001325 * before calling this function to clean up.
1326 * Since no pte has actually been setup, it is
1327 * safe to do nothing in this case.
1328 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001329 if (vma->vm_file) {
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001330 i_mmap_lock_write(vma->vm_file->f_mapping);
Mel Gormand8333522012-07-31 16:46:20 -07001331 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001332 i_mmap_unlock_write(vma->vm_file->f_mapping);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001333 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001334 } else
1335 unmap_page_range(tlb, vma, start, end, details);
1336 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337}
1338
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339/**
1340 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001341 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 * @vma: the starting vma
1343 * @start_addr: virtual address at which to start unmapping
1344 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001346 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 * Only addresses between `start' and `end' will be unmapped.
1349 *
1350 * The VMA list must be sorted in ascending virtual address order.
1351 *
1352 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1353 * range after unmap_vmas() returns. So the only responsibility here is to
1354 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1355 * drops the lock and schedules.
1356 */
Al Viro6e8bb012012-03-05 13:41:15 -05001357void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001359 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001361 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001363 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001364 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001365 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001366 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367}
1368
1369/**
1370 * zap_page_range - remove user pages in a given range
1371 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001372 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 * @size: number of bytes to zap
1374 * @details: details of nonlinear truncation or shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001375 *
1376 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001378void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 unsigned long size, struct zap_details *details)
1380{
1381 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001382 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001383 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001386 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001387 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001388 mmu_notifier_invalidate_range_start(mm, start, end);
1389 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001390 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001391 mmu_notifier_invalidate_range_end(mm, start, end);
1392 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393}
1394
Jack Steinerc627f9c2008-07-29 22:33:53 -07001395/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001396 * zap_page_range_single - remove user pages in a given range
1397 * @vma: vm_area_struct holding the applicable pages
1398 * @address: starting address of pages to zap
1399 * @size: number of bytes to zap
1400 * @details: details of nonlinear truncation or shared cache invalidation
1401 *
1402 * The range must fit into one VMA.
1403 */
1404static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1405 unsigned long size, struct zap_details *details)
1406{
1407 struct mm_struct *mm = vma->vm_mm;
1408 struct mmu_gather tlb;
1409 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001410
1411 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001412 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001413 update_hiwater_rss(mm);
1414 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001415 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001416 mmu_notifier_invalidate_range_end(mm, address, end);
1417 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418}
1419
Jack Steinerc627f9c2008-07-29 22:33:53 -07001420/**
1421 * zap_vma_ptes - remove ptes mapping the vma
1422 * @vma: vm_area_struct holding ptes to be zapped
1423 * @address: starting address of pages to zap
1424 * @size: number of bytes to zap
1425 *
1426 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1427 *
1428 * The entire address range must be fully contained within the vma.
1429 *
1430 * Returns 0 if successful.
1431 */
1432int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1433 unsigned long size)
1434{
1435 if (address < vma->vm_start || address + size > vma->vm_end ||
1436 !(vma->vm_flags & VM_PFNMAP))
1437 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001438 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001439 return 0;
1440}
1441EXPORT_SYMBOL_GPL(zap_vma_ptes);
1442
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001443pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001444 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001445{
1446 pgd_t * pgd = pgd_offset(mm, addr);
1447 pud_t * pud = pud_alloc(mm, pgd, addr);
1448 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001449 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001450 if (pmd) {
1451 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001452 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001453 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001454 }
1455 return NULL;
1456}
1457
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001459 * This is the old fallback for page remapping.
1460 *
1461 * For historical reasons, it only allows reserved pages. Only
1462 * old drivers should use this, and they needed to mark their
1463 * pages reserved for the old functions anyway.
1464 */
Nick Piggin423bad62008-04-28 02:13:01 -07001465static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1466 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001467{
Nick Piggin423bad62008-04-28 02:13:01 -07001468 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001469 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001470 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001471 spinlock_t *ptl;
1472
Linus Torvalds238f58d2005-11-29 13:01:56 -08001473 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001474 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001475 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001476 retval = -ENOMEM;
1477 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001478 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001479 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001480 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001481 retval = -EBUSY;
1482 if (!pte_none(*pte))
1483 goto out_unlock;
1484
1485 /* Ok, finally just insert the thing.. */
1486 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08001487 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001488 page_add_file_rmap(page);
1489 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1490
1491 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001492 pte_unmap_unlock(pte, ptl);
1493 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001494out_unlock:
1495 pte_unmap_unlock(pte, ptl);
1496out:
1497 return retval;
1498}
1499
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001500/**
1501 * vm_insert_page - insert single page into user vma
1502 * @vma: user vma to map to
1503 * @addr: target user address of this page
1504 * @page: source kernel page
1505 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001506 * This allows drivers to insert individual pages they've allocated
1507 * into a user vma.
1508 *
1509 * The page has to be a nice clean _individual_ kernel allocation.
1510 * If you allocate a compound page, you need to have marked it as
1511 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001512 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001513 *
1514 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1515 * took an arbitrary page protection parameter. This doesn't allow
1516 * that. Your vma protection will have to be set up correctly, which
1517 * means that if you want a shared writable mapping, you'd better
1518 * ask for a shared writable mapping!
1519 *
1520 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001521 *
1522 * Usually this function is called from f_op->mmap() handler
1523 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1524 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1525 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001526 */
Nick Piggin423bad62008-04-28 02:13:01 -07001527int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1528 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001529{
1530 if (addr < vma->vm_start || addr >= vma->vm_end)
1531 return -EFAULT;
1532 if (!page_count(page))
1533 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001534 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1535 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1536 BUG_ON(vma->vm_flags & VM_PFNMAP);
1537 vma->vm_flags |= VM_MIXEDMAP;
1538 }
Nick Piggin423bad62008-04-28 02:13:01 -07001539 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001540}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001541EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001542
Nick Piggin423bad62008-04-28 02:13:01 -07001543static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1544 unsigned long pfn, pgprot_t prot)
1545{
1546 struct mm_struct *mm = vma->vm_mm;
1547 int retval;
1548 pte_t *pte, entry;
1549 spinlock_t *ptl;
1550
1551 retval = -ENOMEM;
1552 pte = get_locked_pte(mm, addr, &ptl);
1553 if (!pte)
1554 goto out;
1555 retval = -EBUSY;
1556 if (!pte_none(*pte))
1557 goto out_unlock;
1558
1559 /* Ok, finally just insert the thing.. */
1560 entry = pte_mkspecial(pfn_pte(pfn, prot));
1561 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001562 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07001563
1564 retval = 0;
1565out_unlock:
1566 pte_unmap_unlock(pte, ptl);
1567out:
1568 return retval;
1569}
1570
Nick Piggine0dc0d82007-02-12 00:51:36 -08001571/**
1572 * vm_insert_pfn - insert single pfn into user vma
1573 * @vma: user vma to map to
1574 * @addr: target user address of this page
1575 * @pfn: source kernel pfn
1576 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07001577 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08001578 * they've allocated into a user vma. Same comments apply.
1579 *
1580 * This function should only be called from a vm_ops->fault handler, and
1581 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001582 *
1583 * vma cannot be a COW mapping.
1584 *
1585 * As this is called only for pages that do not currently exist, we
1586 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001587 */
1588int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07001589 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001590{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001591 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001592 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07001593 /*
1594 * Technically, architectures with pte_special can avoid all these
1595 * restrictions (same for remap_pfn_range). However we would like
1596 * consistency in testing and feature parity among all, so we should
1597 * try to keep these invariants in place for everybody.
1598 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001599 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1600 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1601 (VM_PFNMAP|VM_MIXEDMAP));
1602 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1603 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001604
Nick Piggin423bad62008-04-28 02:13:01 -07001605 if (addr < vma->vm_start || addr >= vma->vm_end)
1606 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07001607 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001608 return -EINVAL;
1609
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001610 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001611
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001612 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001613}
1614EXPORT_SYMBOL(vm_insert_pfn);
1615
Nick Piggin423bad62008-04-28 02:13:01 -07001616int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1617 unsigned long pfn)
1618{
1619 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1620
1621 if (addr < vma->vm_start || addr >= vma->vm_end)
1622 return -EFAULT;
1623
1624 /*
1625 * If we don't have pte special, then we have to use the pfn_valid()
1626 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1627 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001628 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1629 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07001630 */
1631 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1632 struct page *page;
1633
1634 page = pfn_to_page(pfn);
1635 return insert_page(vma, addr, page, vma->vm_page_prot);
1636 }
1637 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1638}
1639EXPORT_SYMBOL(vm_insert_mixed);
1640
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001641/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 * maps a range of physical memory into the requested pages. the old
1643 * mappings are removed. any references to nonexistent pages results
1644 * in null mappings (currently treated as "copy-on-access")
1645 */
1646static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1647 unsigned long addr, unsigned long end,
1648 unsigned long pfn, pgprot_t prot)
1649{
1650 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001651 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652
Hugh Dickinsc74df322005-10-29 18:16:23 -07001653 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 if (!pte)
1655 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001656 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 do {
1658 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07001659 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660 pfn++;
1661 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001662 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001663 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664 return 0;
1665}
1666
1667static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1668 unsigned long addr, unsigned long end,
1669 unsigned long pfn, pgprot_t prot)
1670{
1671 pmd_t *pmd;
1672 unsigned long next;
1673
1674 pfn -= addr >> PAGE_SHIFT;
1675 pmd = pmd_alloc(mm, pud, addr);
1676 if (!pmd)
1677 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001678 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 do {
1680 next = pmd_addr_end(addr, end);
1681 if (remap_pte_range(mm, pmd, addr, next,
1682 pfn + (addr >> PAGE_SHIFT), prot))
1683 return -ENOMEM;
1684 } while (pmd++, addr = next, addr != end);
1685 return 0;
1686}
1687
1688static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1689 unsigned long addr, unsigned long end,
1690 unsigned long pfn, pgprot_t prot)
1691{
1692 pud_t *pud;
1693 unsigned long next;
1694
1695 pfn -= addr >> PAGE_SHIFT;
1696 pud = pud_alloc(mm, pgd, addr);
1697 if (!pud)
1698 return -ENOMEM;
1699 do {
1700 next = pud_addr_end(addr, end);
1701 if (remap_pmd_range(mm, pud, addr, next,
1702 pfn + (addr >> PAGE_SHIFT), prot))
1703 return -ENOMEM;
1704 } while (pud++, addr = next, addr != end);
1705 return 0;
1706}
1707
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001708/**
1709 * remap_pfn_range - remap kernel memory to userspace
1710 * @vma: user vma to map to
1711 * @addr: target user address to start at
1712 * @pfn: physical address of kernel memory
1713 * @size: size of map area
1714 * @prot: page protection flags for this mapping
1715 *
1716 * Note: this is only safe if the mm semaphore is held when called.
1717 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1719 unsigned long pfn, unsigned long size, pgprot_t prot)
1720{
1721 pgd_t *pgd;
1722 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001723 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 struct mm_struct *mm = vma->vm_mm;
1725 int err;
1726
1727 /*
1728 * Physically remapped pages are special. Tell the
1729 * rest of the world about it:
1730 * VM_IO tells people not to look at these pages
1731 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08001732 * VM_PFNMAP tells the core MM that the base pages are just
1733 * raw PFN mappings, and do not have a "struct page" associated
1734 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001735 * VM_DONTEXPAND
1736 * Disable vma merging and expanding with mremap().
1737 * VM_DONTDUMP
1738 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001739 *
1740 * There's a horrible special case to handle copy-on-write
1741 * behaviour that some programs depend on. We mark the "original"
1742 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001743 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001745 if (is_cow_mapping(vma->vm_flags)) {
1746 if (addr != vma->vm_start || end != vma->vm_end)
1747 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001748 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001749 }
1750
1751 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1752 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001753 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001754
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001755 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756
1757 BUG_ON(addr >= end);
1758 pfn -= addr >> PAGE_SHIFT;
1759 pgd = pgd_offset(mm, addr);
1760 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 do {
1762 next = pgd_addr_end(addr, end);
1763 err = remap_pud_range(mm, pgd, addr, next,
1764 pfn + (addr >> PAGE_SHIFT), prot);
1765 if (err)
1766 break;
1767 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001768
1769 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07001770 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001771
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 return err;
1773}
1774EXPORT_SYMBOL(remap_pfn_range);
1775
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07001776/**
1777 * vm_iomap_memory - remap memory to userspace
1778 * @vma: user vma to map to
1779 * @start: start of area
1780 * @len: size of area
1781 *
1782 * This is a simplified io_remap_pfn_range() for common driver use. The
1783 * driver just needs to give us the physical memory range to be mapped,
1784 * we'll figure out the rest from the vma information.
1785 *
1786 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1787 * whatever write-combining details or similar.
1788 */
1789int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1790{
1791 unsigned long vm_len, pfn, pages;
1792
1793 /* Check that the physical memory area passed in looks valid */
1794 if (start + len < start)
1795 return -EINVAL;
1796 /*
1797 * You *really* shouldn't map things that aren't page-aligned,
1798 * but we've historically allowed it because IO memory might
1799 * just have smaller alignment.
1800 */
1801 len += start & ~PAGE_MASK;
1802 pfn = start >> PAGE_SHIFT;
1803 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1804 if (pfn + pages < pfn)
1805 return -EINVAL;
1806
1807 /* We start the mapping 'vm_pgoff' pages into the area */
1808 if (vma->vm_pgoff > pages)
1809 return -EINVAL;
1810 pfn += vma->vm_pgoff;
1811 pages -= vma->vm_pgoff;
1812
1813 /* Can we fit all of the mapping? */
1814 vm_len = vma->vm_end - vma->vm_start;
1815 if (vm_len >> PAGE_SHIFT > pages)
1816 return -EINVAL;
1817
1818 /* Ok, let it rip */
1819 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1820}
1821EXPORT_SYMBOL(vm_iomap_memory);
1822
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001823static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1824 unsigned long addr, unsigned long end,
1825 pte_fn_t fn, void *data)
1826{
1827 pte_t *pte;
1828 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001829 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001830 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001831
1832 pte = (mm == &init_mm) ?
1833 pte_alloc_kernel(pmd, addr) :
1834 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1835 if (!pte)
1836 return -ENOMEM;
1837
1838 BUG_ON(pmd_huge(*pmd));
1839
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001840 arch_enter_lazy_mmu_mode();
1841
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001842 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001843
1844 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001845 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001846 if (err)
1847 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001848 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001849
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001850 arch_leave_lazy_mmu_mode();
1851
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001852 if (mm != &init_mm)
1853 pte_unmap_unlock(pte-1, ptl);
1854 return err;
1855}
1856
1857static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1858 unsigned long addr, unsigned long end,
1859 pte_fn_t fn, void *data)
1860{
1861 pmd_t *pmd;
1862 unsigned long next;
1863 int err;
1864
Andi Kleenceb86872008-07-23 21:27:50 -07001865 BUG_ON(pud_huge(*pud));
1866
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001867 pmd = pmd_alloc(mm, pud, addr);
1868 if (!pmd)
1869 return -ENOMEM;
1870 do {
1871 next = pmd_addr_end(addr, end);
1872 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1873 if (err)
1874 break;
1875 } while (pmd++, addr = next, addr != end);
1876 return err;
1877}
1878
1879static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1880 unsigned long addr, unsigned long end,
1881 pte_fn_t fn, void *data)
1882{
1883 pud_t *pud;
1884 unsigned long next;
1885 int err;
1886
1887 pud = pud_alloc(mm, pgd, addr);
1888 if (!pud)
1889 return -ENOMEM;
1890 do {
1891 next = pud_addr_end(addr, end);
1892 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1893 if (err)
1894 break;
1895 } while (pud++, addr = next, addr != end);
1896 return err;
1897}
1898
1899/*
1900 * Scan a region of virtual memory, filling in page tables as necessary
1901 * and calling a provided function on each leaf page table.
1902 */
1903int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1904 unsigned long size, pte_fn_t fn, void *data)
1905{
1906 pgd_t *pgd;
1907 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001908 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001909 int err;
1910
1911 BUG_ON(addr >= end);
1912 pgd = pgd_offset(mm, addr);
1913 do {
1914 next = pgd_addr_end(addr, end);
1915 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1916 if (err)
1917 break;
1918 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001919
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001920 return err;
1921}
1922EXPORT_SYMBOL_GPL(apply_to_page_range);
1923
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001925 * handle_pte_fault chooses page fault handler according to an entry
1926 * which was read non-atomically. Before making any commitment, on
1927 * those architectures or configurations (e.g. i386 with PAE) which
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001928 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001929 * must check under lock before unmapping the pte and proceeding
1930 * (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001931 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001932 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001933static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001934 pte_t *page_table, pte_t orig_pte)
1935{
1936 int same = 1;
1937#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1938 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001939 spinlock_t *ptl = pte_lockptr(mm, pmd);
1940 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001941 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001942 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001943 }
1944#endif
1945 pte_unmap(page_table);
1946 return same;
1947}
1948
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001949static 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 -08001950{
Dan Williams0abdd7a2014-01-21 15:48:12 -08001951 debug_dma_assert_idle(src);
1952
Linus Torvalds6aab3412005-11-28 14:34:23 -08001953 /*
1954 * If the source page was a PFN mapping, we don't have
1955 * a "struct page" for it. We do a best-effort copy by
1956 * just copying from the original user address. If that
1957 * fails, we just zero-fill it. Live with it.
1958 */
1959 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08001960 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08001961 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1962
1963 /*
1964 * This really shouldn't fail, because the page is there
1965 * in the page tables. But it might just be unreadable,
1966 * in which case we just give up and fill the result with
1967 * zeroes.
1968 */
1969 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07001970 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08001971 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07001972 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08001973 } else
1974 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08001975}
1976
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977/*
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001978 * Notify the address space that the page is about to become writable so that
1979 * it can prohibit this or wait for the page to get into an appropriate state.
1980 *
1981 * We do this without the lock held, so that it can sleep if it needs to.
1982 */
1983static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1984 unsigned long address)
1985{
1986 struct vm_fault vmf;
1987 int ret;
1988
1989 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
1990 vmf.pgoff = page->index;
1991 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
1992 vmf.page = page;
1993
1994 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
1995 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
1996 return ret;
1997 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
1998 lock_page(page);
1999 if (!page->mapping) {
2000 unlock_page(page);
2001 return 0; /* retry */
2002 }
2003 ret |= VM_FAULT_LOCKED;
2004 } else
2005 VM_BUG_ON_PAGE(!PageLocked(page), page);
2006 return ret;
2007}
2008
2009/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 * This routine handles present pages, when users try to write
2011 * to a shared page. It is done by copying the page to a new address
2012 * and decrementing the shared-page counter for the old page.
2013 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014 * Note that this routine assumes that the protection checks have been
2015 * done by the caller (the low-level page fault routine in most cases).
2016 * Thus we can safely just mark it writable once we've done any necessary
2017 * COW.
2018 *
2019 * We also mark the page dirty at this point even though the page will
2020 * change only once the write actually happens. This avoids a few races,
2021 * and potentially makes it more efficient.
2022 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002023 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2024 * but allow concurrent faults), with pte both mapped and locked.
2025 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002027static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2028 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002029 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002030 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031{
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002032 struct page *old_page, *new_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 pte_t entry;
Michel Lespinasseb009c022011-01-13 15:46:07 -08002034 int ret = 0;
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002035 int page_mkwrite = 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002036 struct page *dirty_page = NULL;
David Rientjes17569542012-11-16 14:14:48 -08002037 unsigned long mmun_start = 0; /* For mmu_notifiers */
2038 unsigned long mmun_end = 0; /* For mmu_notifiers */
Johannes Weiner00501b52014-08-08 14:19:20 -07002039 struct mem_cgroup *memcg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040
Linus Torvalds6aab3412005-11-28 14:34:23 -08002041 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002042 if (!old_page) {
2043 /*
Peter Feiner64e45502014-10-13 15:55:46 -07002044 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2045 * VM_PFNMAP VMA.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002046 *
2047 * We should not cow pages in a shared writeable mapping.
2048 * Just mark the pages writable as we can't do any dirty
2049 * accounting on raw pfn maps.
2050 */
2051 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2052 (VM_WRITE|VM_SHARED))
2053 goto reuse;
Linus Torvalds6aab3412005-11-28 14:34:23 -08002054 goto gotten;
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002055 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002057 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002058 * Take out anonymous pages first, anonymous shared vmas are
2059 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002060 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002061 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002062 if (!trylock_page(old_page)) {
2063 page_cache_get(old_page);
2064 pte_unmap_unlock(page_table, ptl);
2065 lock_page(old_page);
2066 page_table = pte_offset_map_lock(mm, pmd, address,
2067 &ptl);
2068 if (!pte_same(*page_table, orig_pte)) {
2069 unlock_page(old_page);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002070 goto unlock;
2071 }
2072 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002073 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002074 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002075 /*
2076 * The page is all ours. Move it to our anon_vma so
2077 * the rmap code will not search our parent or siblings.
2078 * Protected against the rmap code by the page lock.
2079 */
2080 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002081 unlock_page(old_page);
2082 goto reuse;
2083 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002084 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002085 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002086 (VM_WRITE|VM_SHARED))) {
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002087 /*
2088 * Only catch write-faults on shared writable pages,
2089 * read-only shared pages can get COWed by
2090 * get_user_pages(.write=1, .force=1).
2091 */
David Howells9637a5e2006-06-23 02:03:43 -07002092 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
Nick Pigginc2ec1752009-03-31 15:23:21 -07002093 int tmp;
David Howells9637a5e2006-06-23 02:03:43 -07002094 page_cache_get(old_page);
2095 pte_unmap_unlock(page_table, ptl);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002096 tmp = do_page_mkwrite(vma, old_page, address);
2097 if (unlikely(!tmp || (tmp &
2098 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2099 page_cache_release(old_page);
2100 return tmp;
Nick Pigginc2ec1752009-03-31 15:23:21 -07002101 }
David Howells9637a5e2006-06-23 02:03:43 -07002102 /*
2103 * Since we dropped the lock we need to revalidate
2104 * the PTE as someone else may have changed it. If
2105 * they did, we just return, as we can count on the
2106 * MMU to tell us if they didn't also make it writable.
2107 */
2108 page_table = pte_offset_map_lock(mm, pmd, address,
2109 &ptl);
Nick Pigginb827e492009-04-30 15:08:16 -07002110 if (!pte_same(*page_table, orig_pte)) {
2111 unlock_page(old_page);
David Howells9637a5e2006-06-23 02:03:43 -07002112 goto unlock;
Nick Pigginb827e492009-04-30 15:08:16 -07002113 }
Peter Zijlstraa200ee12007-10-08 18:54:37 +02002114
2115 page_mkwrite = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002117 dirty_page = old_page;
2118 get_page(dirty_page);
David Howells9637a5e2006-06-23 02:03:43 -07002119
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002120reuse:
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002121 /*
2122 * Clear the pages cpupid information as the existing
2123 * information potentially belongs to a now completely
2124 * unrelated process.
2125 */
2126 if (old_page)
2127 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2128
David Howells9637a5e2006-06-23 02:03:43 -07002129 flush_cache_page(vma, address, pte_pfn(orig_pte));
2130 entry = pte_mkyoung(orig_pte);
2131 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
KAMEZAWA Hiroyuki954ffcb2007-10-16 01:25:44 -07002132 if (ptep_set_access_flags(vma, address, page_table, entry,1))
Russell King4b3073e2009-12-18 16:40:18 +00002133 update_mmu_cache(vma, address, page_table);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002134 pte_unmap_unlock(page_table, ptl);
David Howells9637a5e2006-06-23 02:03:43 -07002135 ret |= VM_FAULT_WRITE;
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002136
2137 if (!dirty_page)
2138 return ret;
2139
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002140 if (!page_mkwrite) {
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002141 struct address_space *mapping;
2142 int dirtied;
2143
2144 lock_page(dirty_page);
2145 dirtied = set_page_dirty(dirty_page);
2146 VM_BUG_ON_PAGE(PageAnon(dirty_page), dirty_page);
2147 mapping = dirty_page->mapping;
2148 unlock_page(dirty_page);
2149
2150 if (dirtied && mapping) {
2151 /*
2152 * Some device drivers do not set page.mapping
2153 * but still dirty their pages
2154 */
2155 balance_dirty_pages_ratelimited(mapping);
2156 }
2157
Jan Kara41c4d252012-06-12 16:20:28 +02002158 /* file_update_time outside page_lock */
2159 if (vma->vm_file)
2160 file_update_time(vma->vm_file);
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002161 }
2162 put_page(dirty_page);
2163 if (page_mkwrite) {
2164 struct address_space *mapping = dirty_page->mapping;
2165
2166 set_page_dirty(dirty_page);
2167 unlock_page(dirty_page);
2168 page_cache_release(dirty_page);
2169 if (mapping) {
2170 /*
2171 * Some device drivers do not set page.mapping
2172 * but still dirty their pages
2173 */
2174 balance_dirty_pages_ratelimited(mapping);
2175 }
2176 }
2177
Michel Lespinasse72ddc8f2011-01-13 15:46:08 -08002178 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180
2181 /*
2182 * Ok, we need to copy. Oh, well..
2183 */
Nick Pigginb5810032005-10-29 18:16:12 -07002184 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08002185gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002186 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187
2188 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07002189 goto oom;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002190
Hugh Dickins62eede62009-09-21 17:03:34 -07002191 if (is_zero_pfn(pte_pfn(orig_pte))) {
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002192 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2193 if (!new_page)
2194 goto oom;
2195 } else {
2196 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2197 if (!new_page)
2198 goto oom;
2199 cow_user_page(new_page, old_page, address, vma);
2200 }
2201 __SetPageUptodate(new_page);
2202
Johannes Weiner00501b52014-08-08 14:19:20 -07002203 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002204 goto oom_free_new;
2205
Haggai Eran6bdb9132012-10-08 16:33:35 -07002206 mmun_start = address & PAGE_MASK;
David Rientjes17569542012-11-16 14:14:48 -08002207 mmun_end = mmun_start + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -07002208 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2209
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 /*
2211 * Re-check the pte - we dropped the lock
2212 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002213 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07002214 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002215 if (old_page) {
Hugh Dickins920fc352005-11-21 21:32:17 -08002216 if (!PageAnon(old_page)) {
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002217 dec_mm_counter_fast(mm, MM_FILEPAGES);
2218 inc_mm_counter_fast(mm, MM_ANONPAGES);
Hugh Dickins920fc352005-11-21 21:32:17 -08002219 }
2220 } else
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002221 inc_mm_counter_fast(mm, MM_ANONPAGES);
Ben Collinseca35132005-11-29 11:45:26 -08002222 flush_cache_page(vma, address, pte_pfn(orig_pte));
Hugh Dickins65500d22005-10-29 18:15:59 -07002223 entry = mk_pte(new_page, vma->vm_page_prot);
2224 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Siddha, Suresh B4ce072f2006-09-29 01:58:42 -07002225 /*
2226 * Clear the pte entry and flush it first, before updating the
2227 * pte with the new entry. This will avoid a race condition
2228 * seen in the presence of one thread doing SMC and another
2229 * thread doing COW.
2230 */
Joerg Roedel34ee6452014-11-13 13:46:09 +11002231 ptep_clear_flush_notify(vma, address, page_table);
Nick Piggin9617d952006-01-06 00:11:12 -08002232 page_add_new_anon_rmap(new_page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002233 mem_cgroup_commit_charge(new_page, memcg, false);
2234 lru_cache_add_active_or_unevictable(new_page, vma);
Izik Eidus828502d2009-09-21 17:01:51 -07002235 /*
2236 * We call the notify macro here because, when using secondary
2237 * mmu page tables (such as kvm shadow page tables), we want the
2238 * new page to be mapped directly into the secondary page table.
2239 */
2240 set_pte_at_notify(mm, address, page_table, entry);
Russell King4b3073e2009-12-18 16:40:18 +00002241 update_mmu_cache(vma, address, page_table);
Nick Piggin945754a2008-06-23 14:30:30 +02002242 if (old_page) {
2243 /*
2244 * Only after switching the pte to the new page may
2245 * we remove the mapcount here. Otherwise another
2246 * process may come and find the rmap count decremented
2247 * before the pte is switched to the new page, and
2248 * "reuse" the old page writing into it while our pte
2249 * here still points into it and can be read by other
2250 * threads.
2251 *
2252 * The critical issue is to order this
2253 * page_remove_rmap with the ptp_clear_flush above.
2254 * Those stores are ordered by (if nothing else,)
2255 * the barrier present in the atomic_add_negative
2256 * in page_remove_rmap.
2257 *
2258 * Then the TLB flush in ptep_clear_flush ensures that
2259 * no process can access the old page before the
2260 * decremented mapcount is visible. And the old page
2261 * cannot be reused until after the decremented
2262 * mapcount is visible. So transitively, TLBs to
2263 * old page will be flushed before it can be reused.
2264 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08002265 page_remove_rmap(old_page);
Nick Piggin945754a2008-06-23 14:30:30 +02002266 }
2267
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 /* Free the old page.. */
2269 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002270 ret |= VM_FAULT_WRITE;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002271 } else
Johannes Weiner00501b52014-08-08 14:19:20 -07002272 mem_cgroup_cancel_charge(new_page, memcg);
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002273
Haggai Eran6bdb9132012-10-08 16:33:35 -07002274 if (new_page)
2275 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002276unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002277 pte_unmap_unlock(page_table, ptl);
David Rientjes17569542012-11-16 14:14:48 -08002278 if (mmun_end > mmun_start)
Haggai Eran6bdb9132012-10-08 16:33:35 -07002279 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Michel Lespinassee15f8c02011-02-10 15:01:32 -08002280 if (old_page) {
2281 /*
2282 * Don't let another task, with possibly unlocked vma,
2283 * keep the mlocked page.
2284 */
2285 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2286 lock_page(old_page); /* LRU manipulation */
2287 munlock_vma_page(old_page);
2288 unlock_page(old_page);
2289 }
2290 page_cache_release(old_page);
2291 }
Nick Pigginf33ea7f2005-08-03 20:24:01 +10002292 return ret;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002293oom_free_new:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002294 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002295oom:
Dominik Dingel66521d52012-12-12 13:52:37 -08002296 if (old_page)
Hugh Dickins920fc352005-11-21 21:32:17 -08002297 page_cache_release(old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 return VM_FAULT_OOM;
2299}
2300
Peter Zijlstra97a89412011-05-24 17:12:04 -07002301static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302 unsigned long start_addr, unsigned long end_addr,
2303 struct zap_details *details)
2304{
Al Virof5cc4ee2012-03-05 14:14:20 -05002305 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306}
2307
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002308static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309 struct zap_details *details)
2310{
2311 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 pgoff_t vba, vea, zba, zea;
2313
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002314 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316
2317 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002318 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2320 zba = details->first_index;
2321 if (zba < vba)
2322 zba = vba;
2323 zea = details->last_index;
2324 if (zea > vea)
2325 zea = vea;
2326
Peter Zijlstra97a89412011-05-24 17:12:04 -07002327 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2329 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002330 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 }
2332}
2333
2334static inline void unmap_mapping_range_list(struct list_head *head,
2335 struct zap_details *details)
2336{
2337 struct vm_area_struct *vma;
2338
2339 /*
2340 * In nonlinear VMAs there is no correspondence between virtual address
2341 * offset and file offset. So we must perform an exhaustive search
2342 * across *all* the pages in each nonlinear VMA, not just the pages
2343 * whose virtual address lies outside the file truncation point.
2344 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002345 list_for_each_entry(vma, head, shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 details->nonlinear_vma = vma;
Peter Zijlstra97a89412011-05-24 17:12:04 -07002347 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 }
2349}
2350
2351/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002352 * 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 -07002353 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 * @holebegin: byte in first page to unmap, relative to the start of
2355 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002356 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 * must keep the partial page. In contrast, we must get rid of
2358 * partial pages.
2359 * @holelen: size of prospective hole in bytes. This will be rounded
2360 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2361 * end of the file.
2362 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2363 * but 0 when invalidating pagecache, don't throw away private data.
2364 */
2365void unmap_mapping_range(struct address_space *mapping,
2366 loff_t const holebegin, loff_t const holelen, int even_cows)
2367{
2368 struct zap_details details;
2369 pgoff_t hba = holebegin >> PAGE_SHIFT;
2370 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2371
2372 /* Check for overflow. */
2373 if (sizeof(holelen) > sizeof(hlen)) {
2374 long long holeend =
2375 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2376 if (holeend & ~(long long)ULONG_MAX)
2377 hlen = ULONG_MAX - hba + 1;
2378 }
2379
2380 details.check_mapping = even_cows? NULL: mapping;
2381 details.nonlinear_vma = NULL;
2382 details.first_index = hba;
2383 details.last_index = hba + hlen - 1;
2384 if (details.last_index < details.first_index)
2385 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386
Peter Zijlstra97a89412011-05-24 17:12:04 -07002387
Kirill A. Shutemov48ec8332014-12-22 21:01:54 +02002388 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002389 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2391 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2392 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
Kirill A. Shutemov48ec8332014-12-22 21:01:54 +02002393 i_mmap_unlock_write(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394}
2395EXPORT_SYMBOL(unmap_mapping_range);
2396
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002398 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2399 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002400 * We return with pte unmapped and unlocked.
2401 *
2402 * We return with the mmap_sem locked or unlocked in the same cases
2403 * as does filemap_fault().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002405static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2406 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002407 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002409 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08002410 struct page *page, *swapcache;
Johannes Weiner00501b52014-08-08 14:19:20 -07002411 struct mem_cgroup *memcg;
Hugh Dickins65500d22005-10-29 18:15:59 -07002412 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002414 int locked;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002415 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002416 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002418 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002419 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002420
2421 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002422 if (unlikely(non_swap_entry(entry))) {
2423 if (is_migration_entry(entry)) {
2424 migration_entry_wait(mm, pmd, address);
2425 } else if (is_hwpoison_entry(entry)) {
2426 ret = VM_FAULT_HWPOISON;
2427 } else {
2428 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002429 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002430 }
Christoph Lameter06972122006-06-23 02:03:35 -07002431 goto out;
2432 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002433 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 page = lookup_swap_cache(entry);
2435 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002436 page = swapin_readahead(entry,
2437 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438 if (!page) {
2439 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002440 * Back out if somebody else faulted in this pte
2441 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002443 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 if (likely(pte_same(*page_table, orig_pte)))
2445 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002446 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002447 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 }
2449
2450 /* Had to read the page from swap area: Major fault */
2451 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002452 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002453 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002454 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002455 /*
2456 * hwpoisoned dirty swapcache pages are kept for killing
2457 * owner processes (which may be unknown at hwpoison time)
2458 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002459 ret = VM_FAULT_HWPOISON;
2460 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08002461 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02002462 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463 }
2464
Hugh Dickins56f31802013-02-22 16:36:10 -08002465 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002466 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002467
Balbir Singh20a10222007-11-14 17:00:33 -08002468 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002469 if (!locked) {
2470 ret |= VM_FAULT_RETRY;
2471 goto out_release;
2472 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002474 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002475 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2476 * release the swapcache from under us. The page pin, and pte_same
2477 * test below, are not enough to exclude that. Even if it is still
2478 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002479 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002480 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002481 goto out_page;
2482
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002483 page = ksm_might_need_to_copy(page, vma, address);
2484 if (unlikely(!page)) {
2485 ret = VM_FAULT_OOM;
2486 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002487 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08002488 }
2489
Johannes Weiner00501b52014-08-08 14:19:20 -07002490 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002491 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002492 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002493 }
2494
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002496 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002498 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002499 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002500 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002501
2502 if (unlikely(!PageUptodate(page))) {
2503 ret = VM_FAULT_SIGBUS;
2504 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505 }
2506
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002507 /*
2508 * The page isn't present yet, go ahead with the fault.
2509 *
2510 * Be careful about the sequence of operations here.
2511 * To get its accounting right, reuse_swap_page() must be called
2512 * while the page is counted on swap but not yet in mapcount i.e.
2513 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2514 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002515 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002517 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08002518 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002520 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002522 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002523 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002524 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07002527 if (pte_swp_soft_dirty(orig_pte))
2528 pte = pte_mksoft_dirty(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 set_pte_at(mm, address, page_table, pte);
Johannes Weiner00501b52014-08-08 14:19:20 -07002530 if (page == swapcache) {
Johannes Weineraf347702013-02-22 16:32:20 -08002531 do_page_add_anon_rmap(page, vma, address, exclusive);
Johannes Weiner00501b52014-08-08 14:19:20 -07002532 mem_cgroup_commit_charge(page, memcg, true);
2533 } else { /* ksm created a completely new copy */
Hugh Dickins56f31802013-02-22 16:36:10 -08002534 page_add_new_anon_rmap(page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002535 mem_cgroup_commit_charge(page, memcg, false);
2536 lru_cache_add_active_or_unevictable(page, vma);
2537 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002539 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07002540 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002541 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002542 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002543 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002544 /*
2545 * Hold the lock to avoid the swap entry to be reused
2546 * until we take the PT lock for the pte_same() check
2547 * (to avoid false positives from pte_same). For
2548 * further safety release the lock after the swap_free
2549 * so that the swap count won't change under a
2550 * parallel locked swapcache.
2551 */
2552 unlock_page(swapcache);
2553 page_cache_release(swapcache);
2554 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002555
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002556 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08002557 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2558 if (ret & VM_FAULT_ERROR)
2559 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560 goto out;
2561 }
2562
2563 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002564 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002565unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002566 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567out:
2568 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002569out_nomap:
Johannes Weiner00501b52014-08-08 14:19:20 -07002570 mem_cgroup_cancel_charge(page, memcg);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002571 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002572out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002573 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002574out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002575 page_cache_release(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002576 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002577 unlock_page(swapcache);
2578 page_cache_release(swapcache);
2579 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002580 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581}
2582
2583/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002584 * This is like a special single-page "expand_{down|up}wards()",
2585 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07002586 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07002587 */
2588static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2589{
2590 address &= PAGE_MASK;
2591 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002592 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07002593
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002594 /*
2595 * Is there a mapping abutting this one below?
2596 *
2597 * That's only ok if it's the same stack mapping
2598 * that has gotten split..
2599 */
2600 if (prev && prev->vm_end == address)
2601 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2602
Linus Torvaldsfee7e492015-01-06 13:00:05 -08002603 return expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002604 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002605 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2606 struct vm_area_struct *next = vma->vm_next;
2607
2608 /* As VM_GROWSDOWN but s/below/above/ */
2609 if (next && next->vm_start == address + PAGE_SIZE)
2610 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2611
Linus Torvaldsfee7e492015-01-06 13:00:05 -08002612 return expand_upwards(vma, address + PAGE_SIZE);
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002613 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07002614 return 0;
2615}
2616
2617/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002618 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2619 * but allow concurrent faults), and pte mapped but not yet locked.
2620 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002622static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2623 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002624 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625{
Johannes Weiner00501b52014-08-08 14:19:20 -07002626 struct mem_cgroup *memcg;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002627 struct page *page;
2628 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630
Linus Torvalds11ac5522010-08-14 11:44:56 -07002631 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002632
Linus Torvalds11ac5522010-08-14 11:44:56 -07002633 /* Check if we need to add a guard page to the stack */
2634 if (check_stack_guard_page(vma, address) < 0)
2635 return VM_FAULT_SIGBUS;
2636
2637 /* Use the zero-page for reads */
Dominik Dingel593befa2014-10-23 12:07:44 +02002638 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
Hugh Dickins62eede62009-09-21 17:03:34 -07002639 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2640 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07002641 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002642 if (!pte_none(*page_table))
2643 goto unlock;
2644 goto setpte;
2645 }
2646
Nick Piggin557ed1f2007-10-16 01:24:40 -07002647 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002648 if (unlikely(anon_vma_prepare(vma)))
2649 goto oom;
2650 page = alloc_zeroed_user_highpage_movable(vma, address);
2651 if (!page)
2652 goto oom;
Minchan Kim52f37622013-04-29 15:08:15 -07002653 /*
2654 * The memory barrier inside __SetPageUptodate makes sure that
2655 * preceeding stores to the page contents become visible before
2656 * the set_pte_at() write.
2657 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08002658 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659
Johannes Weiner00501b52014-08-08 14:19:20 -07002660 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg))
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002661 goto oom_free_page;
2662
Nick Piggin557ed1f2007-10-16 01:24:40 -07002663 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002664 if (vma->vm_flags & VM_WRITE)
2665 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002666
Nick Piggin557ed1f2007-10-16 01:24:40 -07002667 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07002668 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002669 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002670
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002671 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07002672 page_add_new_anon_rmap(page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002673 mem_cgroup_commit_charge(page, memcg, false);
2674 lru_cache_add_active_or_unevictable(page, vma);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002675setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07002676 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677
2678 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002679 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002680unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002681 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002682 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002683release:
Johannes Weiner00501b52014-08-08 14:19:20 -07002684 mem_cgroup_cancel_charge(page, memcg);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002685 page_cache_release(page);
2686 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002687oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002688 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002689oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 return VM_FAULT_OOM;
2691}
2692
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002693/*
2694 * The mmap_sem must have been held on entry, and may have been
2695 * released depending on flags and vma->vm_ops->fault() return value.
2696 * See filemap_fault() and __lock_page_retry().
2697 */
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002698static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2699 pgoff_t pgoff, unsigned int flags, struct page **page)
2700{
2701 struct vm_fault vmf;
2702 int ret;
2703
2704 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2705 vmf.pgoff = pgoff;
2706 vmf.flags = flags;
2707 vmf.page = NULL;
2708
2709 ret = vma->vm_ops->fault(vma, &vmf);
2710 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2711 return ret;
2712
2713 if (unlikely(PageHWPoison(vmf.page))) {
2714 if (ret & VM_FAULT_LOCKED)
2715 unlock_page(vmf.page);
2716 page_cache_release(vmf.page);
2717 return VM_FAULT_HWPOISON;
2718 }
2719
2720 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2721 lock_page(vmf.page);
2722 else
2723 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2724
2725 *page = vmf.page;
2726 return ret;
2727}
2728
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002729/**
2730 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2731 *
2732 * @vma: virtual memory area
2733 * @address: user virtual address
2734 * @page: page to map
2735 * @pte: pointer to target page table entry
2736 * @write: true, if new entry is writable
2737 * @anon: true, if it's anonymous page
2738 *
2739 * Caller must hold page table lock relevant for @pte.
2740 *
2741 * Target users are page handler itself and implementations of
2742 * vm_ops->map_pages.
2743 */
2744void do_set_pte(struct vm_area_struct *vma, unsigned long address,
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002745 struct page *page, pte_t *pte, bool write, bool anon)
2746{
2747 pte_t entry;
2748
2749 flush_icache_page(vma, page);
2750 entry = mk_pte(page, vma->vm_page_prot);
2751 if (write)
2752 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2753 else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
Cyrill Gorcunov9aed8612014-08-06 16:07:05 -07002754 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002755 if (anon) {
2756 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2757 page_add_new_anon_rmap(page, vma, address);
2758 } else {
2759 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2760 page_add_file_rmap(page);
2761 }
2762 set_pte_at(vma->vm_mm, address, pte, entry);
2763
2764 /* no need to invalidate: a not-present page won't be cached */
2765 update_mmu_cache(vma, address, pte);
2766}
2767
Kirill A. Shutemov3a910532014-08-06 16:08:07 -07002768static unsigned long fault_around_bytes __read_mostly =
2769 rounddown_pow_of_two(65536);
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002770
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002771#ifdef CONFIG_DEBUG_FS
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002772static int fault_around_bytes_get(void *data, u64 *val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002773{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002774 *val = fault_around_bytes;
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002775 return 0;
2776}
2777
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002778/*
2779 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2780 * rounded down to nearest page order. It's what do_fault_around() expects to
2781 * see.
2782 */
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002783static int fault_around_bytes_set(void *data, u64 val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002784{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002785 if (val / PAGE_SIZE > PTRS_PER_PTE)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002786 return -EINVAL;
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002787 if (val > PAGE_SIZE)
2788 fault_around_bytes = rounddown_pow_of_two(val);
2789 else
2790 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002791 return 0;
2792}
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002793DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2794 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002795
2796static int __init fault_around_debugfs(void)
2797{
2798 void *ret;
2799
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002800 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2801 &fault_around_bytes_fops);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002802 if (!ret)
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002803 pr_warn("Failed to create fault_around_bytes in debugfs");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002804 return 0;
2805}
2806late_initcall(fault_around_debugfs);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002807#endif
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002808
Kirill A. Shutemov1fdb4122014-06-04 16:10:55 -07002809/*
2810 * do_fault_around() tries to map few pages around the fault address. The hope
2811 * is that the pages will be needed soon and this will lower the number of
2812 * faults to handle.
2813 *
2814 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2815 * not ready to be mapped: not up-to-date, locked, etc.
2816 *
2817 * This function is called with the page table lock taken. In the split ptlock
2818 * case the page table lock only protects only those entries which belong to
2819 * the page table corresponding to the fault address.
2820 *
2821 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2822 * only once.
2823 *
2824 * fault_around_pages() defines how many pages we'll try to map.
2825 * do_fault_around() expects it to return a power of two less than or equal to
2826 * PTRS_PER_PTE.
2827 *
2828 * The virtual address of the area that we map is naturally aligned to the
2829 * fault_around_pages() value (and therefore to page order). This way it's
2830 * easier to guarantee that we don't cross page table boundaries.
2831 */
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002832static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2833 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2834{
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002835 unsigned long start_addr, nr_pages, mask;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002836 pgoff_t max_pgoff;
2837 struct vm_fault vmf;
2838 int off;
2839
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002840 nr_pages = ACCESS_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2841 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2842
2843 start_addr = max(address & mask, vma->vm_start);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002844 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2845 pte -= off;
2846 pgoff -= off;
2847
2848 /*
2849 * max_pgoff is either end of page table or end of vma
Kirill A. Shutemov850e9c62014-06-04 16:10:45 -07002850 * or fault_around_pages() from pgoff, depending what is nearest.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002851 */
2852 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2853 PTRS_PER_PTE - 1;
2854 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002855 pgoff + nr_pages - 1);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002856
2857 /* Check if it makes any sense to call ->map_pages */
2858 while (!pte_none(*pte)) {
2859 if (++pgoff > max_pgoff)
2860 return;
2861 start_addr += PAGE_SIZE;
2862 if (start_addr >= vma->vm_end)
2863 return;
2864 pte++;
2865 }
2866
2867 vmf.virtual_address = (void __user *) start_addr;
2868 vmf.pte = pte;
2869 vmf.pgoff = pgoff;
2870 vmf.max_pgoff = max_pgoff;
2871 vmf.flags = flags;
2872 vma->vm_ops->map_pages(vma, &vmf);
2873}
2874
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002875static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2876 unsigned long address, pmd_t *pmd,
2877 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2878{
2879 struct page *fault_page;
2880 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002881 pte_t *pte;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002882 int ret = 0;
2883
2884 /*
2885 * Let's call ->map_pages() first and use ->fault() as fallback
2886 * if page by the offset is not ready to be mapped (cold cache or
2887 * something).
2888 */
Konstantin Khlebnikovc1186782014-07-23 14:00:08 -07002889 if (vma->vm_ops->map_pages && !(flags & FAULT_FLAG_NONLINEAR) &&
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002890 fault_around_bytes >> PAGE_SHIFT > 1) {
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002891 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2892 do_fault_around(vma, address, pte, pgoff, flags);
2893 if (!pte_same(*pte, orig_pte))
2894 goto unlock_out;
2895 pte_unmap_unlock(pte, ptl);
2896 }
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002897
2898 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2899 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2900 return ret;
2901
2902 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2903 if (unlikely(!pte_same(*pte, orig_pte))) {
2904 pte_unmap_unlock(pte, ptl);
2905 unlock_page(fault_page);
2906 page_cache_release(fault_page);
2907 return ret;
2908 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002909 do_set_pte(vma, address, fault_page, pte, false, false);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002910 unlock_page(fault_page);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002911unlock_out:
2912 pte_unmap_unlock(pte, ptl);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002913 return ret;
2914}
2915
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002916static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2917 unsigned long address, pmd_t *pmd,
2918 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2919{
2920 struct page *fault_page, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07002921 struct mem_cgroup *memcg;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002922 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002923 pte_t *pte;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002924 int ret;
2925
2926 if (unlikely(anon_vma_prepare(vma)))
2927 return VM_FAULT_OOM;
2928
2929 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2930 if (!new_page)
2931 return VM_FAULT_OOM;
2932
Johannes Weiner00501b52014-08-08 14:19:20 -07002933 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) {
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002934 page_cache_release(new_page);
2935 return VM_FAULT_OOM;
2936 }
2937
2938 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2939 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2940 goto uncharge_out;
2941
2942 copy_user_highpage(new_page, fault_page, address, vma);
2943 __SetPageUptodate(new_page);
2944
2945 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2946 if (unlikely(!pte_same(*pte, orig_pte))) {
2947 pte_unmap_unlock(pte, ptl);
2948 unlock_page(fault_page);
2949 page_cache_release(fault_page);
2950 goto uncharge_out;
2951 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002952 do_set_pte(vma, address, new_page, pte, true, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07002953 mem_cgroup_commit_charge(new_page, memcg, false);
2954 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002955 pte_unmap_unlock(pte, ptl);
2956 unlock_page(fault_page);
2957 page_cache_release(fault_page);
2958 return ret;
2959uncharge_out:
Johannes Weiner00501b52014-08-08 14:19:20 -07002960 mem_cgroup_cancel_charge(new_page, memcg);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002961 page_cache_release(new_page);
2962 return ret;
2963}
2964
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002965static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01002966 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07002967 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002968{
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002969 struct page *fault_page;
2970 struct address_space *mapping;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002971 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002972 pte_t *pte;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002973 int dirtied = 0;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002974 int ret, tmp;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07002975
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002976 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2977 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002978 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979
2980 /*
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002981 * Check if the backing address space wants to know that the page is
2982 * about to become writable
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983 */
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002984 if (vma->vm_ops->page_mkwrite) {
2985 unlock_page(fault_page);
2986 tmp = do_page_mkwrite(vma, fault_page, address);
2987 if (unlikely(!tmp ||
2988 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2989 page_cache_release(fault_page);
2990 return tmp;
2991 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992 }
2993
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07002994 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2995 if (unlikely(!pte_same(*pte, orig_pte))) {
2996 pte_unmap_unlock(pte, ptl);
2997 unlock_page(fault_page);
2998 page_cache_release(fault_page);
2999 return ret;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003000 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003001 do_set_pte(vma, address, fault_page, pte, true, false);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003002 pte_unmap_unlock(pte, ptl);
3003
3004 if (set_page_dirty(fault_page))
3005 dirtied = 1;
Andrew Mortond82fa872014-12-18 16:17:29 -08003006 /*
3007 * Take a local copy of the address_space - page.mapping may be zeroed
3008 * by truncate after unlock_page(). The address_space itself remains
3009 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3010 * release semantics to prevent the compiler from undoing this copying.
3011 */
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003012 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
Christian Borntraegere37c6982014-12-07 21:41:33 +01003205 /*
3206 * some architectures can have larger ptes than wordsize,
3207 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3208 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3209 * The code below just needs a consistent view for the ifs and
3210 * we later double check anyway with the ptl lock held. So here
3211 * a barrier will do.
3212 */
3213 entry = *pte;
3214 barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003216 if (pte_none(entry)) {
Jes Sorensenf4b81802006-09-27 01:50:10 -07003217 if (vma->vm_ops) {
Nick Piggin3c18ddd2008-04-28 02:12:10 -07003218 if (likely(vma->vm_ops->fault))
Nick Piggin54cb8822007-07-19 01:46:59 -07003219 return do_linear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003220 pte, pmd, flags, entry);
Jes Sorensenf4b81802006-09-27 01:50:10 -07003221 }
3222 return do_anonymous_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003223 pte, pmd, flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003224 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225 if (pte_file(entry))
Nick Piggind0217ac2007-07-19 01:47:03 -07003226 return do_nonlinear_fault(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003227 pte, pmd, flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003228 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003229 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 }
3231
Mel Gormand10e63f2012-10-25 14:16:31 +02003232 if (pte_numa(entry))
3233 return do_numa_page(mm, vma, address, entry, pte, pmd);
3234
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003235 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003236 spin_lock(ptl);
3237 if (unlikely(!pte_same(*pte, entry)))
3238 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003239 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003240 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003241 return do_wp_page(mm, vma, address,
3242 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003243 entry = pte_mkdirty(entry);
3244 }
3245 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003246 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003247 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003248 } else {
3249 /*
3250 * This is needed only for protection faults but the arch code
3251 * is not yet telling us if this is a protection fault or not.
3252 * This still avoids useless tlb flushes for .text page faults
3253 * with threads.
3254 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003255 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003256 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003257 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003258unlock:
3259 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003260 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261}
3262
3263/*
3264 * By the time we get here, we already hold the mm semaphore
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003265 *
3266 * The mmap_sem may have been released depending on flags and our
3267 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003269static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3270 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271{
3272 pgd_t *pgd;
3273 pud_t *pud;
3274 pmd_t *pmd;
3275 pte_t *pte;
3276
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003277 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003278 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003279
Linus Torvalds1da177e2005-04-16 15:20:36 -07003280 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281 pud = pud_alloc(mm, pgd, address);
3282 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003283 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284 pmd = pmd_alloc(mm, pud, address);
3285 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003286 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003287 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003288 int ret = VM_FAULT_FALLBACK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003289 if (!vma->vm_ops)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003290 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3291 pmd, flags);
3292 if (!(ret & VM_FAULT_FALLBACK))
3293 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003294 } else {
3295 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003296 int ret;
3297
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003298 barrier();
3299 if (pmd_trans_huge(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003300 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3301
Linus Torvaldse53289c2013-01-09 08:36:54 -08003302 /*
3303 * If the pmd is splitting, return and retry the
3304 * the fault. Alternative: wait until the split
3305 * is done, and goto retry.
3306 */
3307 if (pmd_trans_splitting(orig_pmd))
3308 return 0;
3309
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003310 if (pmd_numa(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003311 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003312 orig_pmd, pmd);
3313
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003314 if (dirty && !pmd_write(orig_pmd)) {
David Rientjes1f1d06c2012-05-29 15:06:23 -07003315 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3316 orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003317 if (!(ret & VM_FAULT_FALLBACK))
3318 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003319 } else {
3320 huge_pmd_set_accessed(mm, vma, address, pmd,
3321 orig_pmd, dirty);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003322 return 0;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003323 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003324 }
3325 }
3326
3327 /*
3328 * Use __pte_alloc instead of pte_alloc_map, because we can't
3329 * run pte_offset_map on the pmd, if an huge pmd could
3330 * materialize from under us from a different thread.
3331 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003332 if (unlikely(pmd_none(*pmd)) &&
3333 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003334 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003335 /* if an huge pmd materialized from under us just retry later */
3336 if (unlikely(pmd_trans_huge(*pmd)))
3337 return 0;
3338 /*
3339 * A regular pmd is established and it can't morph into a huge pmd
3340 * from under us anymore at this point because we hold the mmap_sem
3341 * read mode and khugepaged takes it in write mode. So now it's
3342 * safe to run pte_offset_map().
3343 */
3344 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003346 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347}
3348
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003349/*
3350 * By the time we get here, we already hold the mm semaphore
3351 *
3352 * The mmap_sem may have been released depending on flags and our
3353 * return value. See filemap_fault() and __lock_page_or_retry().
3354 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003355int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3356 unsigned long address, unsigned int flags)
3357{
3358 int ret;
3359
3360 __set_current_state(TASK_RUNNING);
3361
3362 count_vm_event(PGFAULT);
3363 mem_cgroup_count_vm_event(mm, PGFAULT);
3364
3365 /* do counter updates before entering really critical section. */
3366 check_sync_rss_stat(current);
3367
3368 /*
3369 * Enable the memcg OOM handling for faults triggered in user
3370 * space. Kernel faults are handled more gracefully.
3371 */
3372 if (flags & FAULT_FLAG_USER)
Johannes Weiner49426422013-10-16 13:46:59 -07003373 mem_cgroup_oom_enable();
Johannes Weiner519e5242013-09-12 15:13:42 -07003374
3375 ret = __handle_mm_fault(mm, vma, address, flags);
3376
Johannes Weiner49426422013-10-16 13:46:59 -07003377 if (flags & FAULT_FLAG_USER) {
3378 mem_cgroup_oom_disable();
3379 /*
3380 * The task may have entered a memcg OOM situation but
3381 * if the allocation error was handled gracefully (no
3382 * VM_FAULT_OOM), there is no need to kill anything.
3383 * Just clean up the OOM state peacefully.
3384 */
3385 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3386 mem_cgroup_oom_synchronize(false);
3387 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003388
Johannes Weiner519e5242013-09-12 15:13:42 -07003389 return ret;
3390}
Jesse Barnese1d6d012014-12-12 16:55:27 -08003391EXPORT_SYMBOL_GPL(handle_mm_fault);
Johannes Weiner519e5242013-09-12 15:13:42 -07003392
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393#ifndef __PAGETABLE_PUD_FOLDED
3394/*
3395 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003396 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003398int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003400 pud_t *new = pud_alloc_one(mm, address);
3401 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003402 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403
Nick Piggin362a61a2008-05-14 06:37:36 +02003404 smp_wmb(); /* See comment in __pte_alloc */
3405
Hugh Dickins872fec12005-10-29 18:16:21 -07003406 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003407 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003408 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003409 else
3410 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003411 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003412 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413}
3414#endif /* __PAGETABLE_PUD_FOLDED */
3415
3416#ifndef __PAGETABLE_PMD_FOLDED
3417/*
3418 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003419 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003421int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003423 pmd_t *new = pmd_alloc_one(mm, address);
3424 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003425 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426
Nick Piggin362a61a2008-05-14 06:37:36 +02003427 smp_wmb(); /* See comment in __pte_alloc */
3428
Hugh Dickins872fec12005-10-29 18:16:21 -07003429 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07003431 if (pud_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003432 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003433 else
3434 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07003436 if (pgd_present(*pud)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003437 pmd_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003438 else
3439 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003441 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003442 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003443}
3444#endif /* __PAGETABLE_PMD_FOLDED */
3445
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003446static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003447 pte_t **ptepp, spinlock_t **ptlp)
3448{
3449 pgd_t *pgd;
3450 pud_t *pud;
3451 pmd_t *pmd;
3452 pte_t *ptep;
3453
3454 pgd = pgd_offset(mm, address);
3455 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3456 goto out;
3457
3458 pud = pud_offset(pgd, address);
3459 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3460 goto out;
3461
3462 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003463 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003464 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3465 goto out;
3466
3467 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3468 if (pmd_huge(*pmd))
3469 goto out;
3470
3471 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3472 if (!ptep)
3473 goto out;
3474 if (!pte_present(*ptep))
3475 goto unlock;
3476 *ptepp = ptep;
3477 return 0;
3478unlock:
3479 pte_unmap_unlock(ptep, *ptlp);
3480out:
3481 return -EINVAL;
3482}
3483
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003484static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3485 pte_t **ptepp, spinlock_t **ptlp)
3486{
3487 int res;
3488
3489 /* (void) is needed to make gcc happy */
3490 (void) __cond_lock(*ptlp,
3491 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3492 return res;
3493}
3494
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003495/**
3496 * follow_pfn - look up PFN at a user virtual address
3497 * @vma: memory mapping
3498 * @address: user virtual address
3499 * @pfn: location to store found PFN
3500 *
3501 * Only IO mappings and raw PFN mappings are allowed.
3502 *
3503 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3504 */
3505int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3506 unsigned long *pfn)
3507{
3508 int ret = -EINVAL;
3509 spinlock_t *ptl;
3510 pte_t *ptep;
3511
3512 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3513 return ret;
3514
3515 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3516 if (ret)
3517 return ret;
3518 *pfn = pte_pfn(*ptep);
3519 pte_unmap_unlock(ptep, ptl);
3520 return 0;
3521}
3522EXPORT_SYMBOL(follow_pfn);
3523
Rik van Riel28b2ee22008-07-23 21:27:05 -07003524#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003525int follow_phys(struct vm_area_struct *vma,
3526 unsigned long address, unsigned int flags,
3527 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003528{
Johannes Weiner03668a42009-06-16 15:32:34 -07003529 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003530 pte_t *ptep, pte;
3531 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003532
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003533 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3534 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003535
Johannes Weiner03668a42009-06-16 15:32:34 -07003536 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003537 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003538 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003539
Rik van Riel28b2ee22008-07-23 21:27:05 -07003540 if ((flags & FOLL_WRITE) && !pte_write(pte))
3541 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003542
3543 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003544 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003545
Johannes Weiner03668a42009-06-16 15:32:34 -07003546 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003547unlock:
3548 pte_unmap_unlock(ptep, ptl);
3549out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003550 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003551}
3552
3553int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3554 void *buf, int len, int write)
3555{
3556 resource_size_t phys_addr;
3557 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003558 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003559 int offset = addr & (PAGE_SIZE-1);
3560
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003561 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003562 return -EINVAL;
3563
3564 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3565 if (write)
3566 memcpy_toio(maddr + offset, buf, len);
3567 else
3568 memcpy_fromio(buf, maddr + offset, len);
3569 iounmap(maddr);
3570
3571 return len;
3572}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02003573EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003574#endif
3575
David Howells0ec76a12006-09-27 01:50:15 -07003576/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003577 * Access another process' address space as given in mm. If non-NULL, use the
3578 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003579 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003580static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3581 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003582{
David Howells0ec76a12006-09-27 01:50:15 -07003583 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003584 void *old_buf = buf;
3585
David Howells0ec76a12006-09-27 01:50:15 -07003586 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003587 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003588 while (len) {
3589 int bytes, ret, offset;
3590 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003591 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003592
3593 ret = get_user_pages(tsk, mm, addr, 1,
3594 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003595 if (ret <= 0) {
Rik van Rieldbffcd02014-08-06 16:08:12 -07003596#ifndef CONFIG_HAVE_IOREMAP_PROT
3597 break;
3598#else
Rik van Riel28b2ee22008-07-23 21:27:05 -07003599 /*
3600 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3601 * we can access using slightly different code.
3602 */
Rik van Riel28b2ee22008-07-23 21:27:05 -07003603 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003604 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003605 break;
3606 if (vma->vm_ops && vma->vm_ops->access)
3607 ret = vma->vm_ops->access(vma, addr, buf,
3608 len, write);
3609 if (ret <= 0)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003610 break;
3611 bytes = ret;
Rik van Rieldbffcd02014-08-06 16:08:12 -07003612#endif
David Howells0ec76a12006-09-27 01:50:15 -07003613 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003614 bytes = len;
3615 offset = addr & (PAGE_SIZE-1);
3616 if (bytes > PAGE_SIZE-offset)
3617 bytes = PAGE_SIZE-offset;
3618
3619 maddr = kmap(page);
3620 if (write) {
3621 copy_to_user_page(vma, page, addr,
3622 maddr + offset, buf, bytes);
3623 set_page_dirty_lock(page);
3624 } else {
3625 copy_from_user_page(vma, page, addr,
3626 buf, maddr + offset, bytes);
3627 }
3628 kunmap(page);
3629 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003630 }
David Howells0ec76a12006-09-27 01:50:15 -07003631 len -= bytes;
3632 buf += bytes;
3633 addr += bytes;
3634 }
3635 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003636
3637 return buf - old_buf;
3638}
Andi Kleen03252912008-01-30 13:33:18 +01003639
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003640/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003641 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003642 * @mm: the mm_struct of the target address space
3643 * @addr: start address to access
3644 * @buf: source or destination buffer
3645 * @len: number of bytes to transfer
3646 * @write: whether the access is a write
3647 *
3648 * The caller must hold a reference on @mm.
3649 */
3650int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3651 void *buf, int len, int write)
3652{
3653 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3654}
3655
Andi Kleen03252912008-01-30 13:33:18 +01003656/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003657 * Access another process' address space.
3658 * Source/target buffer must be kernel space,
3659 * Do not walk the page table directly, use get_user_pages
3660 */
3661int access_process_vm(struct task_struct *tsk, unsigned long addr,
3662 void *buf, int len, int write)
3663{
3664 struct mm_struct *mm;
3665 int ret;
3666
3667 mm = get_task_mm(tsk);
3668 if (!mm)
3669 return 0;
3670
3671 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3672 mmput(mm);
3673
3674 return ret;
3675}
3676
Andi Kleen03252912008-01-30 13:33:18 +01003677/*
3678 * Print the name of a VMA.
3679 */
3680void print_vma_addr(char *prefix, unsigned long ip)
3681{
3682 struct mm_struct *mm = current->mm;
3683 struct vm_area_struct *vma;
3684
Ingo Molnare8bff742008-02-13 20:21:06 +01003685 /*
3686 * Do not print if we are in atomic
3687 * contexts (in exception stacks, etc.):
3688 */
3689 if (preempt_count())
3690 return;
3691
Andi Kleen03252912008-01-30 13:33:18 +01003692 down_read(&mm->mmap_sem);
3693 vma = find_vma(mm, ip);
3694 if (vma && vma->vm_file) {
3695 struct file *f = vma->vm_file;
3696 char *buf = (char *)__get_free_page(GFP_KERNEL);
3697 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003698 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01003699
Jan Blunckcf28b482008-02-14 19:38:44 -08003700 p = d_path(&f->f_path, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003701 if (IS_ERR(p))
3702 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003703 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01003704 vma->vm_start,
3705 vma->vm_end - vma->vm_start);
3706 free_page((unsigned long)buf);
3707 }
3708 }
Jeff Liu51a07e52012-07-31 16:43:18 -07003709 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01003710}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003711
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003712#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003713void might_fault(void)
3714{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003715 /*
3716 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3717 * holding the mmap_sem, this is safe because kernel memory doesn't
3718 * get paged out, therefore we'll never actually fault, and the
3719 * below annotations will generate false positives.
3720 */
3721 if (segment_eq(get_fs(), KERNEL_DS))
3722 return;
3723
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003724 /*
3725 * it would be nicer only to annotate paths which are not under
3726 * pagefault_disable, however that requires a larger audit and
3727 * providing helpers like get_user_atomic.
3728 */
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003729 if (in_atomic())
3730 return;
3731
3732 __might_sleep(__FILE__, __LINE__, 0);
3733
3734 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003735 might_lock_read(&current->mm->mmap_sem);
3736}
3737EXPORT_SYMBOL(might_fault);
3738#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003739
3740#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3741static void clear_gigantic_page(struct page *page,
3742 unsigned long addr,
3743 unsigned int pages_per_huge_page)
3744{
3745 int i;
3746 struct page *p = page;
3747
3748 might_sleep();
3749 for (i = 0; i < pages_per_huge_page;
3750 i++, p = mem_map_next(p, page, i)) {
3751 cond_resched();
3752 clear_user_highpage(p, addr + i * PAGE_SIZE);
3753 }
3754}
3755void clear_huge_page(struct page *page,
3756 unsigned long addr, unsigned int pages_per_huge_page)
3757{
3758 int i;
3759
3760 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3761 clear_gigantic_page(page, addr, pages_per_huge_page);
3762 return;
3763 }
3764
3765 might_sleep();
3766 for (i = 0; i < pages_per_huge_page; i++) {
3767 cond_resched();
3768 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3769 }
3770}
3771
3772static void copy_user_gigantic_page(struct page *dst, struct page *src,
3773 unsigned long addr,
3774 struct vm_area_struct *vma,
3775 unsigned int pages_per_huge_page)
3776{
3777 int i;
3778 struct page *dst_base = dst;
3779 struct page *src_base = src;
3780
3781 for (i = 0; i < pages_per_huge_page; ) {
3782 cond_resched();
3783 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3784
3785 i++;
3786 dst = mem_map_next(dst, dst_base, i);
3787 src = mem_map_next(src, src_base, i);
3788 }
3789}
3790
3791void copy_user_huge_page(struct page *dst, struct page *src,
3792 unsigned long addr, struct vm_area_struct *vma,
3793 unsigned int pages_per_huge_page)
3794{
3795 int i;
3796
3797 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3798 copy_user_gigantic_page(dst, src, addr, vma,
3799 pages_per_huge_page);
3800 return;
3801 }
3802
3803 might_sleep();
3804 for (i = 0; i < pages_per_huge_page; i++) {
3805 cond_resched();
3806 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3807 }
3808}
3809#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003810
Olof Johansson40b64ac2013-12-20 14:28:05 -08003811#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003812
3813static struct kmem_cache *page_ptl_cachep;
3814
3815void __init ptlock_cache_init(void)
3816{
3817 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3818 SLAB_PANIC, NULL);
3819}
3820
Peter Zijlstra539edb52013-11-14 14:31:52 -08003821bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003822{
3823 spinlock_t *ptl;
3824
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003825 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003826 if (!ptl)
3827 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08003828 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003829 return true;
3830}
3831
Peter Zijlstra539edb52013-11-14 14:31:52 -08003832void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003833{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003834 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003835}
3836#endif