blob: 88f8d6a2af05ce2dd0e98e7bd605bfaa15deef28 [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>
Dan Williams01c8f1c2016-01-15 16:56:40 -080053#include <linux/pfn_t.h>
Peter Zijlstraedc79b22006-09-25 23:30:58 -070054#include <linux/writeback.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080055#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070056#include <linux/mmu_notifier.h>
Hugh Dickins3dc14742009-01-06 14:40:08 -080057#include <linux/kallsyms.h>
58#include <linux/swapops.h>
59#include <linux/elf.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090060#include <linux/gfp.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000061#include <linux/migrate.h>
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -080062#include <linux/string.h>
Dan Williams0abdd7a2014-01-21 15:48:12 -080063#include <linux/dma-debug.h>
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -070064#include <linux/debugfs.h>
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -070065#include <linux/userfaultfd_k.h>
Jan Karabc2466e2016-05-12 18:29:19 +020066#include <linux/dax.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070067
Alexey Dobriyan6952b612009-09-18 23:55:55 +040068#include <asm/io.h>
Dave Hansen33a709b2016-02-12 13:02:19 -080069#include <asm/mmu_context.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070070#include <asm/pgalloc.h>
71#include <asm/uaccess.h>
72#include <asm/tlb.h>
73#include <asm/tlbflush.h>
74#include <asm/pgtable.h>
75
Jan Beulich42b77722008-07-23 21:27:10 -070076#include "internal.h"
77
Arnd Bergmann73184542018-02-16 16:25:53 +010078#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
Peter Zijlstra90572892013-10-07 11:29:20 +010079#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
Peter Zijlstra75980e92013-02-22 16:34:32 -080080#endif
81
Andy Whitcroftd41dee32005-06-23 00:07:54 -070082#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070083/* use the per-pgdat data instead for discontigmem - mbligh */
84unsigned long max_mapnr;
85struct page *mem_map;
86
87EXPORT_SYMBOL(max_mapnr);
88EXPORT_SYMBOL(mem_map);
89#endif
90
Linus Torvalds1da177e2005-04-16 15:20:36 -070091/*
92 * A number of key systems in x86 including ioremap() rely on the assumption
93 * that high_memory defines the upper bound on direct map memory, then end
94 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
95 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
96 * and ZONE_HIGHMEM.
97 */
98void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101
Ingo Molnar32a93232008-02-06 22:39:44 +0100102/*
103 * Randomize the address space (stacks, mmaps, brk, etc.).
104 *
105 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
106 * as ancient (libc5 based) binaries can segfault. )
107 */
108int randomize_va_space __read_mostly =
109#ifdef CONFIG_COMPAT_BRK
110 1;
111#else
112 2;
113#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100114
115static int __init disable_randmaps(char *s)
116{
117 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800118 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100119}
120__setup("norandmaps", disable_randmaps);
121
Hugh Dickins62eede62009-09-21 17:03:34 -0700122unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700123unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700124
Ard Biesheuvel0b700682014-09-12 22:17:23 +0200125EXPORT_SYMBOL(zero_pfn);
126
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700127/*
128 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
129 */
130static int __init init_zero_pfn(void)
131{
132 zero_pfn = page_to_pfn(ZERO_PAGE(0));
133 return 0;
134}
135core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100136
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800137
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800138#if defined(SPLIT_RSS_COUNTING)
139
David Rientjesea48cf72012-03-21 16:34:13 -0700140void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800141{
142 int i;
143
144 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700145 if (current->rss_stat.count[i]) {
146 add_mm_counter(mm, i, current->rss_stat.count[i]);
147 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800148 }
149 }
David Rientjes05af2e12012-03-21 16:34:13 -0700150 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800151}
152
153static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
154{
155 struct task_struct *task = current;
156
157 if (likely(task->mm == mm))
158 task->rss_stat.count[member] += val;
159 else
160 add_mm_counter(mm, member, val);
161}
162#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
163#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
164
165/* sync counter once per 64 page faults */
166#define TASK_RSS_EVENTS_THRESH (64)
167static void check_sync_rss_stat(struct task_struct *task)
168{
169 if (unlikely(task != current))
170 return;
171 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700172 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800173}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700174#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800175
176#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
177#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
178
179static void check_sync_rss_stat(struct task_struct *task)
180{
181}
182
Peter Zijlstra9547d012011-05-24 17:12:14 -0700183#endif /* SPLIT_RSS_COUNTING */
184
185#ifdef HAVE_GENERIC_MMU_GATHER
186
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700187static bool tlb_next_batch(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700188{
189 struct mmu_gather_batch *batch;
190
191 batch = tlb->active;
192 if (batch->next) {
193 tlb->active = batch->next;
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700194 return true;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700195 }
196
Michal Hocko53a59fc2013-01-04 15:35:12 -0800197 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700198 return false;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800199
Peter Zijlstra9547d012011-05-24 17:12:14 -0700200 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
201 if (!batch)
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700202 return false;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700203
Michal Hocko53a59fc2013-01-04 15:35:12 -0800204 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700205 batch->next = NULL;
206 batch->nr = 0;
207 batch->max = MAX_GATHER_BATCH;
208
209 tlb->active->next = batch;
210 tlb->active = batch;
211
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700212 return true;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700213}
214
215/* tlb_gather_mmu
216 * Called to initialize an (on-stack) mmu_gather structure for page-table
217 * tear-down from @mm. The @fullmm argument is used when @mm is without
218 * users and we're going to destroy the full address space (exit/execve).
219 */
Linus Torvalds2b047252013-08-15 11:42:25 -0700220void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700221{
222 tlb->mm = mm;
223
Linus Torvalds2b047252013-08-15 11:42:25 -0700224 /* Is it from 0 to ~0? */
225 tlb->fullmm = !(start | (end+1));
Dave Hansen1de14c32013-04-12 16:23:54 -0700226 tlb->need_flush_all = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700227 tlb->local.next = NULL;
228 tlb->local.nr = 0;
229 tlb->local.max = ARRAY_SIZE(tlb->__pages);
230 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800231 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700232
233#ifdef CONFIG_HAVE_RCU_TABLE_FREE
234 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800235#endif
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -0700236 tlb->page_size = 0;
Will Deaconfb7332a2014-10-29 10:03:09 +0000237
238 __tlb_reset_range(tlb);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700239}
240
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700241static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700242{
Will Deacon721c21c2015-01-12 19:10:55 +0000243 if (!tlb->end)
244 return;
245
Peter Zijlstra9547d012011-05-24 17:12:14 -0700246 tlb_flush(tlb);
Joerg Roedel34ee6452014-11-13 13:46:09 +1100247 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700248#ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 tlb_table_flush(tlb);
250#endif
Will Deaconfb7332a2014-10-29 10:03:09 +0000251 __tlb_reset_range(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700252}
253
254static void tlb_flush_mmu_free(struct mmu_gather *tlb)
255{
256 struct mmu_gather_batch *batch;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700257
Will Deacon721c21c2015-01-12 19:10:55 +0000258 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
Peter Zijlstra9547d012011-05-24 17:12:14 -0700259 free_pages_and_swap_cache(batch->pages, batch->nr);
260 batch->nr = 0;
261 }
262 tlb->active = &tlb->local;
263}
264
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700265void tlb_flush_mmu(struct mmu_gather *tlb)
266{
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700267 tlb_flush_mmu_tlbonly(tlb);
268 tlb_flush_mmu_free(tlb);
269}
270
Peter Zijlstra9547d012011-05-24 17:12:14 -0700271/* tlb_finish_mmu
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
274 */
275void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
276{
277 struct mmu_gather_batch *batch, *next;
278
279 tlb_flush_mmu(tlb);
280
281 /* keep the page table cache within bounds */
282 check_pgt_cache();
283
284 for (batch = tlb->local.next; batch; batch = next) {
285 next = batch->next;
286 free_pages((unsigned long)batch, 0);
287 }
288 tlb->local.next = NULL;
289}
290
291/* __tlb_remove_page
292 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
293 * handling the additional races in SMP caused by other CPUs caching valid
294 * mappings in their TLBs. Returns the number of free page slots left.
295 * When out of page slots we must call tlb_flush_mmu().
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -0700296 *returns true if the caller should flush.
Peter Zijlstra9547d012011-05-24 17:12:14 -0700297 */
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -0700298bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700299{
300 struct mmu_gather_batch *batch;
301
Will Deaconfb7332a2014-10-29 10:03:09 +0000302 VM_BUG_ON(!tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700303
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -0700304 if (!tlb->page_size)
305 tlb->page_size = page_size;
306 else {
307 if (page_size != tlb->page_size)
308 return true;
309 }
310
Peter Zijlstra9547d012011-05-24 17:12:14 -0700311 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700312 if (batch->nr == batch->max) {
313 if (!tlb_next_batch(tlb))
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -0700314 return true;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700315 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700316 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800317 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700318
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -0700319 batch->pages[batch->nr++] = page;
320 return false;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700321}
322
323#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800324
Peter Zijlstra26723912011-05-24 17:12:00 -0700325#ifdef CONFIG_HAVE_RCU_TABLE_FREE
326
327/*
328 * See the comment near struct mmu_table_batch.
329 */
330
331static void tlb_remove_table_smp_sync(void *arg)
332{
333 /* Simply deliver the interrupt */
334}
335
336static void tlb_remove_table_one(void *table)
337{
338 /*
339 * This isn't an RCU grace period and hence the page-tables cannot be
340 * assumed to be actually RCU-freed.
341 *
342 * It is however sufficient for software page-table walkers that rely on
343 * IRQ disabling. See the comment near struct mmu_table_batch.
344 */
345 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
346 __tlb_remove_table(table);
347}
348
349static void tlb_remove_table_rcu(struct rcu_head *head)
350{
351 struct mmu_table_batch *batch;
352 int i;
353
354 batch = container_of(head, struct mmu_table_batch, rcu);
355
356 for (i = 0; i < batch->nr; i++)
357 __tlb_remove_table(batch->tables[i]);
358
359 free_page((unsigned long)batch);
360}
361
362void tlb_table_flush(struct mmu_gather *tlb)
363{
364 struct mmu_table_batch **batch = &tlb->batch;
365
366 if (*batch) {
367 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
368 *batch = NULL;
369 }
370}
371
372void tlb_remove_table(struct mmu_gather *tlb, void *table)
373{
374 struct mmu_table_batch **batch = &tlb->batch;
375
Peter Zijlstra26723912011-05-24 17:12:00 -0700376 /*
377 * When there's less then two users of this mm there cannot be a
378 * concurrent page-table walk.
379 */
380 if (atomic_read(&tlb->mm->mm_users) < 2) {
381 __tlb_remove_table(table);
382 return;
383 }
384
385 if (*batch == NULL) {
386 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
387 if (*batch == NULL) {
388 tlb_remove_table_one(table);
389 return;
390 }
391 (*batch)->nr = 0;
392 }
393 (*batch)->tables[(*batch)->nr++] = table;
394 if ((*batch)->nr == MAX_TABLE_BATCH)
395 tlb_table_flush(tlb);
396}
397
Peter Zijlstra9547d012011-05-24 17:12:14 -0700398#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700399
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 * Note: this doesn't free the actual pages themselves. That
402 * has been handled earlier when unmapping all the memory regions.
403 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000404static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
405 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800407 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700408 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000409 pte_free_tlb(tlb, token, addr);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800410 atomic_long_dec(&tlb->mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411}
412
Hugh Dickinse0da3822005-04-19 13:29:15 -0700413static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
414 unsigned long addr, unsigned long end,
415 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416{
417 pmd_t *pmd;
418 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700419 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420
Hugh Dickinse0da3822005-04-19 13:29:15 -0700421 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 do {
424 next = pmd_addr_end(addr, end);
425 if (pmd_none_or_clear_bad(pmd))
426 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000427 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 } while (pmd++, addr = next, addr != end);
429
Hugh Dickinse0da3822005-04-19 13:29:15 -0700430 start &= PUD_MASK;
431 if (start < floor)
432 return;
433 if (ceiling) {
434 ceiling &= PUD_MASK;
435 if (!ceiling)
436 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700438 if (end - 1 > ceiling - 1)
439 return;
440
441 pmd = pmd_offset(pud, start);
442 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000443 pmd_free_tlb(tlb, pmd, start);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -0800444 mm_dec_nr_pmds(tlb->mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445}
446
Hugh Dickinse0da3822005-04-19 13:29:15 -0700447static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
448 unsigned long addr, unsigned long end,
449 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450{
451 pud_t *pud;
452 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700453 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454
Hugh Dickinse0da3822005-04-19 13:29:15 -0700455 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 do {
458 next = pud_addr_end(addr, end);
459 if (pud_none_or_clear_bad(pud))
460 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700461 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 } while (pud++, addr = next, addr != end);
463
Hugh Dickinse0da3822005-04-19 13:29:15 -0700464 start &= PGDIR_MASK;
465 if (start < floor)
466 return;
467 if (ceiling) {
468 ceiling &= PGDIR_MASK;
469 if (!ceiling)
470 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700472 if (end - 1 > ceiling - 1)
473 return;
474
475 pud = pud_offset(pgd, start);
476 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000477 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478}
479
480/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700481 * This function frees user-level page tables of a process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 */
Jan Beulich42b77722008-07-23 21:27:10 -0700483void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700484 unsigned long addr, unsigned long end,
485 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486{
487 pgd_t *pgd;
488 unsigned long next;
489
Hugh Dickinse0da3822005-04-19 13:29:15 -0700490 /*
491 * The next few lines have given us lots of grief...
492 *
493 * Why are we testing PMD* at this top level? Because often
494 * there will be no work to do at all, and we'd prefer not to
495 * go all the way down to the bottom just to discover that.
496 *
497 * Why all these "- 1"s? Because 0 represents both the bottom
498 * of the address space and the top of it (using -1 for the
499 * top wouldn't help much: the masks would do the wrong thing).
500 * The rule is that addr 0 and floor 0 refer to the bottom of
501 * the address space, but end 0 and ceiling 0 refer to the top
502 * Comparisons need to use "end - 1" and "ceiling - 1" (though
503 * that end 0 case should be mythical).
504 *
505 * Wherever addr is brought up or ceiling brought down, we must
506 * be careful to reject "the opposite 0" before it confuses the
507 * subsequent tests. But what about where end is brought down
508 * by PMD_SIZE below? no, end can't go down to 0 there.
509 *
510 * Whereas we round start (addr) and ceiling down, by different
511 * masks at different levels, in order to test whether a table
512 * now has no other vmas using it, so can be freed, we don't
513 * bother to round floor or end up - the tests don't need that.
514 */
515
516 addr &= PMD_MASK;
517 if (addr < floor) {
518 addr += PMD_SIZE;
519 if (!addr)
520 return;
521 }
522 if (ceiling) {
523 ceiling &= PMD_MASK;
524 if (!ceiling)
525 return;
526 }
527 if (end - 1 > ceiling - 1)
528 end -= PMD_SIZE;
529 if (addr > end - 1)
530 return;
531
Jan Beulich42b77722008-07-23 21:27:10 -0700532 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 do {
534 next = pgd_addr_end(addr, end);
535 if (pgd_none_or_clear_bad(pgd))
536 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700537 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700539}
540
Jan Beulich42b77722008-07-23 21:27:10 -0700541void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700542 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700543{
544 while (vma) {
545 struct vm_area_struct *next = vma->vm_next;
546 unsigned long addr = vma->vm_start;
547
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700548 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000549 * Hide vma from rmap and truncate_pagecache before freeing
550 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700551 */
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);
554
David Gibson9da61ae2006-03-22 00:08:57 -0800555 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700556 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700557 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700558 } else {
559 /*
560 * Optimization: gather nearby vmas into one call down
561 */
562 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800563 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700564 vma = next;
565 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800566 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700567 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700568 }
569 free_pgd_range(tlb, addr, vma->vm_end,
570 floor, next? next->vm_start: ceiling);
571 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700572 vma = next;
573 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574}
575
Kirill A. Shutemov3ed3a4f2016-03-17 14:19:11 -0700576int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800578 spinlock_t *ptl;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800579 pgtable_t new = pte_alloc_one(mm, address);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700580 if (!new)
581 return -ENOMEM;
582
Nick Piggin362a61a2008-05-14 06:37:36 +0200583 /*
584 * Ensure all pte setup (eg. pte page lock and page clearing) are
585 * visible before the pte is made visible to other CPUs by being
586 * put into page tables.
587 *
588 * The other side of the story is the pointer chasing in the page
589 * table walking code (when walking the page table without locking;
590 * ie. most of the time). Fortunately, these data accesses consist
591 * of a chain of data-dependent loads, meaning most CPUs (alpha
592 * being the notable exception) will already guarantee loads are
593 * seen in-order. See the alpha page table accessors for the
594 * smp_read_barrier_depends() barriers in page table walking code.
595 */
596 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
597
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800598 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800599 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800600 atomic_long_inc(&mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800602 new = NULL;
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -0800603 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800604 spin_unlock(ptl);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800605 if (new)
606 pte_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700607 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608}
609
Hugh Dickins1bb36302005-10-29 18:16:22 -0700610int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700612 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
613 if (!new)
614 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615
Nick Piggin362a61a2008-05-14 06:37:36 +0200616 smp_wmb(); /* See comment in __pte_alloc */
617
Hugh Dickins1bb36302005-10-29 18:16:22 -0700618 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800619 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700620 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800621 new = NULL;
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -0800622 }
Hugh Dickins1bb36302005-10-29 18:16:22 -0700623 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800624 if (new)
625 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700626 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627}
628
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800629static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700630{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800631 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
632}
633
634static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
635{
636 int i;
637
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800638 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700639 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800640 for (i = 0; i < NR_MM_COUNTERS; i++)
641 if (rss[i])
642 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700643}
644
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800646 * This function is called to print an error when a bad pte
647 * is found. For example, we might have a PFN-mapped pte in
648 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700649 *
650 * The calling function must still handle the error.
651 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800652static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
653 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700654{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800655 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
656 pud_t *pud = pud_offset(pgd, addr);
657 pmd_t *pmd = pmd_offset(pud, addr);
658 struct address_space *mapping;
659 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800660 static unsigned long resume;
661 static unsigned long nr_shown;
662 static unsigned long nr_unshown;
663
664 /*
665 * Allow a burst of 60 reports, then keep quiet for that minute;
666 * or allow a steady drip of one report per second.
667 */
668 if (nr_shown == 60) {
669 if (time_before(jiffies, resume)) {
670 nr_unshown++;
671 return;
672 }
673 if (nr_unshown) {
Joe Perches11705322016-03-17 14:19:50 -0700674 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
675 nr_unshown);
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800676 nr_unshown = 0;
677 }
678 nr_shown = 0;
679 }
680 if (nr_shown++ == 0)
681 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800682
683 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
684 index = linear_page_index(vma, addr);
685
Joe Perches11705322016-03-17 14:19:50 -0700686 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
687 current->comm,
688 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800689 if (page)
Dave Hansenf0b791a2014-01-23 15:52:49 -0800690 dump_page(page, "bad pte");
Joe Perches11705322016-03-17 14:19:50 -0700691 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
692 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
Hugh Dickins3dc14742009-01-06 14:40:08 -0800693 /*
694 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
695 */
Konstantin Khlebnikov26825822015-04-15 16:15:08 -0700696 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
697 vma->vm_file,
698 vma->vm_ops ? vma->vm_ops->fault : NULL,
699 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
700 mapping ? mapping->a_ops->readpage : NULL);
Nick Pigginb5810032005-10-29 18:16:12 -0700701 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030702 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700703}
704
705/*
Nick Piggin7e675132008-04-28 02:13:00 -0700706 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800707 *
Nick Piggin7e675132008-04-28 02:13:00 -0700708 * "Special" mappings do not wish to be associated with a "struct page" (either
709 * it doesn't exist, or it exists but they don't want to touch it). In this
710 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700711 *
Nick Piggin7e675132008-04-28 02:13:00 -0700712 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
713 * pte bit, in which case this function is trivial. Secondly, an architecture
714 * may not have a spare pte bit, which requires a more complicated scheme,
715 * described below.
716 *
717 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
718 * special mapping (even if there are underlying and valid "struct pages").
719 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800720 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700721 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
722 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700723 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
724 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800725 *
726 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
727 *
Nick Piggin7e675132008-04-28 02:13:00 -0700728 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700729 *
Nick Piggin7e675132008-04-28 02:13:00 -0700730 * This restricts such mappings to be a linear translation from virtual address
731 * to pfn. To get around this restriction, we allow arbitrary mappings so long
732 * as the vma is not a COW mapping; in that case, we know that all ptes are
733 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700734 *
735 *
Nick Piggin7e675132008-04-28 02:13:00 -0700736 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
737 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700738 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
739 * page" backing, however the difference is that _all_ pages with a struct
740 * page (that is, those where pfn_valid is true) are refcounted and considered
741 * normal pages by the VM. The disadvantage is that pages are refcounted
742 * (which can be slower and simply not an option for some PFNMAP users). The
743 * advantage is that we don't have to follow the strict linearity rule of
744 * PFNMAP mappings in order to support COWable mappings.
745 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800746 */
Nick Piggin7e675132008-04-28 02:13:00 -0700747#ifdef __HAVE_ARCH_PTE_SPECIAL
748# define HAVE_PTE_SPECIAL 1
749#else
750# define HAVE_PTE_SPECIAL 0
751#endif
752struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
753 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800754{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800755 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700756
757 if (HAVE_PTE_SPECIAL) {
Hugh Dickinsb38af472014-08-29 15:18:44 -0700758 if (likely(!pte_special(pte)))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800759 goto check_pfn;
David Vrabel667a0a02014-12-18 14:48:15 +0000760 if (vma->vm_ops && vma->vm_ops->find_special_page)
761 return vma->vm_ops->find_special_page(vma, addr);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700762 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
763 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700764 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800765 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700766 return NULL;
767 }
768
769 /* !HAVE_PTE_SPECIAL case follows: */
770
Jared Hulbertb379d792008-04-28 02:12:58 -0700771 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
772 if (vma->vm_flags & VM_MIXEDMAP) {
773 if (!pfn_valid(pfn))
774 return NULL;
775 goto out;
776 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700777 unsigned long off;
778 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700779 if (pfn == vma->vm_pgoff + off)
780 return NULL;
781 if (!is_cow_mapping(vma->vm_flags))
782 return NULL;
783 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800784 }
785
Hugh Dickinsb38af472014-08-29 15:18:44 -0700786 if (is_zero_pfn(pfn))
787 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800788check_pfn:
789 if (unlikely(pfn > highest_memmap_pfn)) {
790 print_bad_pte(vma, addr, pte, NULL);
791 return NULL;
792 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800793
794 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700795 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700796 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800797 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700798out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800799 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800800}
801
Gerald Schaefer28093f92016-04-28 16:18:35 -0700802#ifdef CONFIG_TRANSPARENT_HUGEPAGE
803struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
804 pmd_t pmd)
805{
806 unsigned long pfn = pmd_pfn(pmd);
807
808 /*
809 * There is no pmd_special() but there may be special pmds, e.g.
810 * in a direct-access (dax) mapping, so let's just replicate the
811 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
812 */
813 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
814 if (vma->vm_flags & VM_MIXEDMAP) {
815 if (!pfn_valid(pfn))
816 return NULL;
817 goto out;
818 } else {
819 unsigned long off;
820 off = (addr - vma->vm_start) >> PAGE_SHIFT;
821 if (pfn == vma->vm_pgoff + off)
822 return NULL;
823 if (!is_cow_mapping(vma->vm_flags))
824 return NULL;
825 }
826 }
827
828 if (is_zero_pfn(pfn))
829 return NULL;
830 if (unlikely(pfn > highest_memmap_pfn))
831 return NULL;
832
833 /*
834 * NOTE! We still have PageReserved() pages in the page tables.
835 * eg. VDSO mappings can cause them to exist.
836 */
837out:
838 return pfn_to_page(pfn);
839}
840#endif
841
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800842/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 * copy one vm_area from one task to the other. Assumes the page tables
844 * already present in the new task to be cleared in the whole range
845 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 */
847
Hugh Dickins570a335b2009-12-14 17:58:46 -0800848static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700850 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700851 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852{
Nick Pigginb5810032005-10-29 18:16:12 -0700853 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 pte_t pte = *src_pte;
855 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856
857 /* pte contains position in swap or file, so copy. */
858 if (unlikely(!pte_present(pte))) {
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800859 swp_entry_t entry = pte_to_swp_entry(pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700860
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800861 if (likely(!non_swap_entry(entry))) {
862 if (swap_duplicate(entry) < 0)
863 return entry.val;
Hugh Dickins570a335b2009-12-14 17:58:46 -0800864
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800865 /* make sure dst_mm is on swapoff's mmlist. */
866 if (unlikely(list_empty(&dst_mm->mmlist))) {
867 spin_lock(&mmlist_lock);
868 if (list_empty(&dst_mm->mmlist))
869 list_add(&dst_mm->mmlist,
870 &src_mm->mmlist);
871 spin_unlock(&mmlist_lock);
872 }
873 rss[MM_SWAPENTS]++;
874 } else if (is_migration_entry(entry)) {
875 page = migration_entry_to_page(entry);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800876
Jerome Marchandeca56ff2016-01-14 15:19:26 -0800877 rss[mm_counter(page)]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800878
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800879 if (is_write_migration_entry(entry) &&
880 is_cow_mapping(vm_flags)) {
881 /*
882 * COW mappings require pages in both
883 * parent and child to be set to read.
884 */
885 make_migration_entry_read(&entry);
886 pte = swp_entry_to_pte(entry);
887 if (pte_swp_soft_dirty(*src_pte))
888 pte = pte_swp_mksoft_dirty(pte);
889 set_pte_at(src_mm, addr, src_pte, pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700890 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700892 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 }
894
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 /*
896 * If it's a COW mapping, write protect it both
897 * in the parent and the child
898 */
Linus Torvalds67121172005-12-11 20:38:17 -0800899 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700901 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 }
903
904 /*
905 * If it's a shared mapping, mark it clean in
906 * the child
907 */
908 if (vm_flags & VM_SHARED)
909 pte = pte_mkclean(pte);
910 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800911
912 page = vm_normal_page(vma, addr, pte);
913 if (page) {
914 get_page(page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -0800915 page_dup_rmap(page, false);
Jerome Marchandeca56ff2016-01-14 15:19:26 -0800916 rss[mm_counter(page)]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800917 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700918
919out_set_pte:
920 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800921 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922}
923
Jerome Marchand21bda262014-08-06 16:06:56 -0700924static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800925 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
926 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700928 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700930 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700931 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800932 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800933 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934
935again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800936 init_rss_vec(rss);
937
Hugh Dickinsc74df322005-10-29 18:16:23 -0700938 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 if (!dst_pte)
940 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700941 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700942 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700943 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700944 orig_src_pte = src_pte;
945 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700946 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 do {
949 /*
950 * We are holding two locks at this point - either of them
951 * could generate latencies in another task on another CPU.
952 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700953 if (progress >= 32) {
954 progress = 0;
955 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100956 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700957 break;
958 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 if (pte_none(*src_pte)) {
960 progress++;
961 continue;
962 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800963 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
964 vma, addr, rss);
965 if (entry.val)
966 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 progress += 8;
968 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700970 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700971 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700972 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800973 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700974 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700975 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800976
977 if (entry.val) {
978 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
979 return -ENOMEM;
980 progress = 0;
981 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 if (addr != end)
983 goto again;
984 return 0;
985}
986
987static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
988 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
989 unsigned long addr, unsigned long end)
990{
991 pmd_t *src_pmd, *dst_pmd;
992 unsigned long next;
993
994 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
995 if (!dst_pmd)
996 return -ENOMEM;
997 src_pmd = pmd_offset(src_pud, addr);
998 do {
999 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -08001000 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001001 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -08001002 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001003 err = copy_huge_pmd(dst_mm, src_mm,
1004 dst_pmd, src_pmd, addr, vma);
1005 if (err == -ENOMEM)
1006 return -ENOMEM;
1007 if (!err)
1008 continue;
1009 /* fall through */
1010 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011 if (pmd_none_or_clear_bad(src_pmd))
1012 continue;
1013 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1014 vma, addr, next))
1015 return -ENOMEM;
1016 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1017 return 0;
1018}
1019
1020static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1021 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1022 unsigned long addr, unsigned long end)
1023{
1024 pud_t *src_pud, *dst_pud;
1025 unsigned long next;
1026
1027 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1028 if (!dst_pud)
1029 return -ENOMEM;
1030 src_pud = pud_offset(src_pgd, addr);
1031 do {
1032 next = pud_addr_end(addr, end);
1033 if (pud_none_or_clear_bad(src_pud))
1034 continue;
1035 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1036 vma, addr, next))
1037 return -ENOMEM;
1038 } while (dst_pud++, src_pud++, addr = next, addr != end);
1039 return 0;
1040}
1041
1042int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1043 struct vm_area_struct *vma)
1044{
1045 pgd_t *src_pgd, *dst_pgd;
1046 unsigned long next;
1047 unsigned long addr = vma->vm_start;
1048 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001049 unsigned long mmun_start; /* For mmu_notifiers */
1050 unsigned long mmun_end; /* For mmu_notifiers */
1051 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001052 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053
Nick Piggind9928952005-08-28 16:49:11 +10001054 /*
1055 * Don't copy ptes where a page fault will fill them correctly.
1056 * Fork becomes much lighter when there are big shared or private
1057 * readonly mappings. The tradeoff is that copy_page_range is more
1058 * efficient than faulting.
1059 */
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08001060 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1061 !vma->anon_vma)
1062 return 0;
Nick Piggind9928952005-08-28 16:49:11 +10001063
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 if (is_vm_hugetlb_page(vma))
1065 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1066
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001067 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001068 /*
1069 * We do not free on error cases below as remove_vma
1070 * gets called on error from higher level routine
1071 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001072 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001073 if (ret)
1074 return ret;
1075 }
1076
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001077 /*
1078 * We need to invalidate the secondary MMU mappings only when
1079 * there could be a permission downgrade on the ptes of the
1080 * parent mm. And a permission downgrade will only happen if
1081 * is_cow_mapping() returns true.
1082 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001083 is_cow = is_cow_mapping(vma->vm_flags);
1084 mmun_start = addr;
1085 mmun_end = end;
1086 if (is_cow)
1087 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1088 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001089
1090 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 dst_pgd = pgd_offset(dst_mm, addr);
1092 src_pgd = pgd_offset(src_mm, addr);
1093 do {
1094 next = pgd_addr_end(addr, end);
1095 if (pgd_none_or_clear_bad(src_pgd))
1096 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001097 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1098 vma, addr, next))) {
1099 ret = -ENOMEM;
1100 break;
1101 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001103
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001104 if (is_cow)
1105 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001106 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107}
1108
Robin Holt51c6f662005-11-13 16:06:42 -08001109static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001110 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001112 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113{
Nick Pigginb5810032005-10-29 18:16:12 -07001114 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001115 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001116 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001117 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001118 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001119 pte_t *pte;
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001120 swp_entry_t entry;
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001121 struct page *pending_page = NULL;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001122
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001123again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001124 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001125 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1126 pte = start_pte;
Mel Gorman5a1eef72017-08-02 13:31:52 -07001127 flush_tlb_batched_pending(mm);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001128 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 do {
1130 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001131 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001133 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001134
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001136 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001137
Linus Torvalds6aab3412005-11-28 14:34:23 -08001138 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 if (unlikely(details) && page) {
1140 /*
1141 * unmap_shared_mapping_pages() wants to
1142 * invalidate cache without truncating:
1143 * unmap shared but keep private pages.
1144 */
1145 if (details->check_mapping &&
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07001146 details->check_mapping != page_rmapping(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 }
Nick Pigginb5810032005-10-29 18:16:12 -07001149 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001150 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 tlb_remove_tlb_entry(tlb, pte, addr);
1152 if (unlikely(!page))
1153 continue;
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001154
1155 if (!PageAnon(page)) {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001156 if (pte_dirty(ptent)) {
Michal Hockoaac45362016-03-25 14:20:24 -07001157 /*
1158 * oom_reaper cannot tear down dirty
1159 * pages
1160 */
1161 if (unlikely(details && details->ignore_dirty))
1162 continue;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001163 force_flush = 1;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001164 set_page_dirty(page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001165 }
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001166 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001167 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001168 mark_page_accessed(page);
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001169 }
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001170 rss[mm_counter(page)]--;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001171 page_remove_rmap(page, false);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001172 if (unlikely(page_mapcount(page) < 0))
1173 print_bad_pte(vma, addr, ptent, page);
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001174 if (unlikely(__tlb_remove_page(tlb, page))) {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001175 force_flush = 1;
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001176 pending_page = page;
Will Deaconce9ec372014-10-28 13:16:28 -07001177 addr += PAGE_SIZE;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001178 break;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001179 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 continue;
1181 }
Michal Hockoaac45362016-03-25 14:20:24 -07001182 /* only check swap_entries if explicitly asked for in details */
1183 if (unlikely(details && !details->check_swap_entries))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 continue;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001185
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001186 entry = pte_to_swp_entry(ptent);
1187 if (!non_swap_entry(entry))
1188 rss[MM_SWAPENTS]--;
1189 else if (is_migration_entry(entry)) {
1190 struct page *page;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001191
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001192 page = migration_entry_to_page(entry);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001193 rss[mm_counter(page)]--;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001194 }
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001195 if (unlikely(!free_swap_and_cache(entry)))
1196 print_bad_pte(vma, addr, ptent, NULL);
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001197 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001198 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001199
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001200 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001201 arch_leave_lazy_mmu_mode();
Robin Holt51c6f662005-11-13 16:06:42 -08001202
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001203 /* Do the actual TLB flush before dropping ptl */
Will Deaconfb7332a2014-10-29 10:03:09 +00001204 if (force_flush)
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001205 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001206 pte_unmap_unlock(start_pte, ptl);
1207
1208 /*
1209 * If we forced a TLB flush (either due to running out of
1210 * batch buffers or because we needed to flush dirty TLB
1211 * entries before releasing the ptl), free the batched
1212 * memory too. Restart if we didn't do everything.
1213 */
1214 if (force_flush) {
1215 force_flush = 0;
1216 tlb_flush_mmu_free(tlb);
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001217 if (pending_page) {
1218 /* remove the page with new size */
1219 __tlb_remove_pte_page(tlb, pending_page);
1220 pending_page = NULL;
1221 }
Linus Torvalds2b047252013-08-15 11:42:25 -07001222 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001223 goto again;
1224 }
1225
Robin Holt51c6f662005-11-13 16:06:42 -08001226 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227}
1228
Robin Holt51c6f662005-11-13 16:06:42 -08001229static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001230 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001232 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233{
1234 pmd_t *pmd;
1235 unsigned long next;
1236
1237 pmd = pmd_offset(pud, addr);
1238 do {
1239 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -08001240 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001241 if (next - addr != HPAGE_PMD_SIZE) {
Hugh Dickins68428392016-05-05 16:22:09 -07001242 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1243 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001244 split_huge_pmd(vma, pmd, addr);
Shaohua Lif21760b2012-01-12 17:19:16 -08001245 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001246 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001247 /* fall through */
1248 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001249 /*
1250 * Here there can be other concurrent MADV_DONTNEED or
1251 * trans huge page faults running, and if the pmd is
1252 * none or trans huge it can change under us. This is
1253 * because MADV_DONTNEED holds the mmap_sem in read
1254 * mode.
1255 */
1256 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1257 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001258 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001259next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001260 cond_resched();
1261 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001262
1263 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264}
1265
Robin Holt51c6f662005-11-13 16:06:42 -08001266static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001267 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001269 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270{
1271 pud_t *pud;
1272 unsigned long next;
1273
1274 pud = pud_offset(pgd, addr);
1275 do {
1276 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001277 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001279 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1280 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001281
1282 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283}
1284
Michal Hockoaac45362016-03-25 14:20:24 -07001285void unmap_page_range(struct mmu_gather *tlb,
Al Viro038c7aa2012-03-05 13:25:09 -05001286 struct vm_area_struct *vma,
1287 unsigned long addr, unsigned long end,
1288 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289{
1290 pgd_t *pgd;
1291 unsigned long next;
1292
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 BUG_ON(addr >= end);
1294 tlb_start_vma(tlb, vma);
1295 pgd = pgd_offset(vma->vm_mm, addr);
1296 do {
1297 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001298 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001300 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1301 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 tlb_end_vma(tlb, vma);
1303}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304
Al Virof5cc4ee2012-03-05 14:14:20 -05001305
1306static void unmap_single_vma(struct mmu_gather *tlb,
1307 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001308 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001309 struct zap_details *details)
1310{
1311 unsigned long start = max(vma->vm_start, start_addr);
1312 unsigned long end;
1313
1314 if (start >= vma->vm_end)
1315 return;
1316 end = min(vma->vm_end, end_addr);
1317 if (end <= vma->vm_start)
1318 return;
1319
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301320 if (vma->vm_file)
1321 uprobe_munmap(vma, start, end);
1322
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001323 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001324 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001325
1326 if (start != end) {
1327 if (unlikely(is_vm_hugetlb_page(vma))) {
1328 /*
1329 * It is undesirable to test vma->vm_file as it
1330 * should be non-null for valid hugetlb area.
1331 * However, vm_file will be NULL in the error
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001332 * cleanup path of mmap_region. When
Al Virof5cc4ee2012-03-05 14:14:20 -05001333 * hugetlbfs ->mmap method fails,
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001334 * mmap_region() nullifies vma->vm_file
Al Virof5cc4ee2012-03-05 14:14:20 -05001335 * before calling this function to clean up.
1336 * Since no pte has actually been setup, it is
1337 * safe to do nothing in this case.
1338 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001339 if (vma->vm_file) {
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001340 i_mmap_lock_write(vma->vm_file->f_mapping);
Mel Gormand8333522012-07-31 16:46:20 -07001341 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001342 i_mmap_unlock_write(vma->vm_file->f_mapping);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001343 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001344 } else
1345 unmap_page_range(tlb, vma, start, end, details);
1346 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347}
1348
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349/**
1350 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001351 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 * @vma: the starting vma
1353 * @start_addr: virtual address at which to start unmapping
1354 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001356 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 * Only addresses between `start' and `end' will be unmapped.
1359 *
1360 * The VMA list must be sorted in ascending virtual address order.
1361 *
1362 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1363 * range after unmap_vmas() returns. So the only responsibility here is to
1364 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1365 * drops the lock and schedules.
1366 */
Al Viro6e8bb012012-03-05 13:41:15 -05001367void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001369 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001371 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001373 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001374 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001375 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001376 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377}
1378
1379/**
1380 * zap_page_range - remove user pages in a given range
1381 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001382 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001384 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001385 *
1386 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001388void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 unsigned long size, struct zap_details *details)
1390{
1391 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001392 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001393 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001396 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001397 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001398 mmu_notifier_invalidate_range_start(mm, start, end);
1399 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001400 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001401 mmu_notifier_invalidate_range_end(mm, start, end);
1402 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403}
1404
Jack Steinerc627f9c2008-07-29 22:33:53 -07001405/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001406 * zap_page_range_single - remove user pages in a given range
1407 * @vma: vm_area_struct holding the applicable pages
1408 * @address: starting address of pages to zap
1409 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001410 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001411 *
1412 * The range must fit into one VMA.
1413 */
1414static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1415 unsigned long size, struct zap_details *details)
1416{
1417 struct mm_struct *mm = vma->vm_mm;
1418 struct mmu_gather tlb;
1419 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001420
1421 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001422 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001423 update_hiwater_rss(mm);
1424 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001425 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001426 mmu_notifier_invalidate_range_end(mm, address, end);
1427 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428}
1429
Jack Steinerc627f9c2008-07-29 22:33:53 -07001430/**
1431 * zap_vma_ptes - remove ptes mapping the vma
1432 * @vma: vm_area_struct holding ptes to be zapped
1433 * @address: starting address of pages to zap
1434 * @size: number of bytes to zap
1435 *
1436 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1437 *
1438 * The entire address range must be fully contained within the vma.
1439 *
1440 * Returns 0 if successful.
1441 */
1442int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1443 unsigned long size)
1444{
1445 if (address < vma->vm_start || address + size > vma->vm_end ||
1446 !(vma->vm_flags & VM_PFNMAP))
1447 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001448 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001449 return 0;
1450}
1451EXPORT_SYMBOL_GPL(zap_vma_ptes);
1452
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001453pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001454 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001455{
1456 pgd_t * pgd = pgd_offset(mm, addr);
1457 pud_t * pud = pud_alloc(mm, pgd, addr);
1458 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001459 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001460 if (pmd) {
1461 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001462 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001463 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001464 }
1465 return NULL;
1466}
1467
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001469 * This is the old fallback for page remapping.
1470 *
1471 * For historical reasons, it only allows reserved pages. Only
1472 * old drivers should use this, and they needed to mark their
1473 * pages reserved for the old functions anyway.
1474 */
Nick Piggin423bad62008-04-28 02:13:01 -07001475static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1476 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001477{
Nick Piggin423bad62008-04-28 02:13:01 -07001478 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001479 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001480 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001481 spinlock_t *ptl;
1482
Linus Torvalds238f58d2005-11-29 13:01:56 -08001483 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001484 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001485 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001486 retval = -ENOMEM;
1487 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001488 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001489 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001490 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001491 retval = -EBUSY;
1492 if (!pte_none(*pte))
1493 goto out_unlock;
1494
1495 /* Ok, finally just insert the thing.. */
1496 get_page(page);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001497 inc_mm_counter_fast(mm, mm_counter_file(page));
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001498 page_add_file_rmap(page, false);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001499 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1500
1501 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001502 pte_unmap_unlock(pte, ptl);
1503 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001504out_unlock:
1505 pte_unmap_unlock(pte, ptl);
1506out:
1507 return retval;
1508}
1509
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001510/**
1511 * vm_insert_page - insert single page into user vma
1512 * @vma: user vma to map to
1513 * @addr: target user address of this page
1514 * @page: source kernel page
1515 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001516 * This allows drivers to insert individual pages they've allocated
1517 * into a user vma.
1518 *
1519 * The page has to be a nice clean _individual_ kernel allocation.
1520 * If you allocate a compound page, you need to have marked it as
1521 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001522 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001523 *
1524 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1525 * took an arbitrary page protection parameter. This doesn't allow
1526 * that. Your vma protection will have to be set up correctly, which
1527 * means that if you want a shared writable mapping, you'd better
1528 * ask for a shared writable mapping!
1529 *
1530 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001531 *
1532 * Usually this function is called from f_op->mmap() handler
1533 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1534 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1535 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001536 */
Nick Piggin423bad62008-04-28 02:13:01 -07001537int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1538 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001539{
1540 if (addr < vma->vm_start || addr >= vma->vm_end)
1541 return -EFAULT;
1542 if (!page_count(page))
1543 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001544 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1545 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1546 BUG_ON(vma->vm_flags & VM_PFNMAP);
1547 vma->vm_flags |= VM_MIXEDMAP;
1548 }
Nick Piggin423bad62008-04-28 02:13:01 -07001549 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001550}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001551EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001552
Nick Piggin423bad62008-04-28 02:13:01 -07001553static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001554 pfn_t pfn, pgprot_t prot)
Nick Piggin423bad62008-04-28 02:13:01 -07001555{
1556 struct mm_struct *mm = vma->vm_mm;
1557 int retval;
1558 pte_t *pte, entry;
1559 spinlock_t *ptl;
1560
1561 retval = -ENOMEM;
1562 pte = get_locked_pte(mm, addr, &ptl);
1563 if (!pte)
1564 goto out;
1565 retval = -EBUSY;
1566 if (!pte_none(*pte))
1567 goto out_unlock;
1568
1569 /* Ok, finally just insert the thing.. */
Dan Williams01c8f1c2016-01-15 16:56:40 -08001570 if (pfn_t_devmap(pfn))
1571 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1572 else
1573 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
Nick Piggin423bad62008-04-28 02:13:01 -07001574 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001575 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07001576
1577 retval = 0;
1578out_unlock:
1579 pte_unmap_unlock(pte, ptl);
1580out:
1581 return retval;
1582}
1583
Nick Piggine0dc0d82007-02-12 00:51:36 -08001584/**
1585 * vm_insert_pfn - insert single pfn into user vma
1586 * @vma: user vma to map to
1587 * @addr: target user address of this page
1588 * @pfn: source kernel pfn
1589 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07001590 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08001591 * they've allocated into a user vma. Same comments apply.
1592 *
1593 * This function should only be called from a vm_ops->fault handler, and
1594 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001595 *
1596 * vma cannot be a COW mapping.
1597 *
1598 * As this is called only for pages that do not currently exist, we
1599 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001600 */
1601int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07001602 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001603{
Andy Lutomirski1745cbc2015-12-29 20:12:20 -08001604 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1605}
1606EXPORT_SYMBOL(vm_insert_pfn);
1607
1608/**
1609 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1610 * @vma: user vma to map to
1611 * @addr: target user address of this page
1612 * @pfn: source kernel pfn
1613 * @pgprot: pgprot flags for the inserted page
1614 *
1615 * This is exactly like vm_insert_pfn, except that it allows drivers to
1616 * to override pgprot on a per-page basis.
1617 *
1618 * This only makes sense for IO mappings, and it makes no sense for
1619 * cow mappings. In general, using multiple vmas is preferable;
1620 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1621 * impractical.
1622 */
1623int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1624 unsigned long pfn, pgprot_t pgprot)
1625{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001626 int ret;
Nick Piggin7e675132008-04-28 02:13:00 -07001627 /*
1628 * Technically, architectures with pte_special can avoid all these
1629 * restrictions (same for remap_pfn_range). However we would like
1630 * consistency in testing and feature parity among all, so we should
1631 * try to keep these invariants in place for everybody.
1632 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001633 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1634 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1635 (VM_PFNMAP|VM_MIXEDMAP));
1636 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1637 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001638
Nick Piggin423bad62008-04-28 02:13:01 -07001639 if (addr < vma->vm_start || addr >= vma->vm_end)
1640 return -EFAULT;
Dan Williamsf25748e32016-01-15 16:56:43 -08001641 if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001642 return -EINVAL;
1643
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001644 if (!pfn_modify_allowed(pfn, pgprot))
1645 return -EACCES;
1646
Dan Williams01c8f1c2016-01-15 16:56:40 -08001647 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001648
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001649 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001650}
Andy Lutomirski1745cbc2015-12-29 20:12:20 -08001651EXPORT_SYMBOL(vm_insert_pfn_prot);
Nick Piggine0dc0d82007-02-12 00:51:36 -08001652
Nick Piggin423bad62008-04-28 02:13:01 -07001653int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001654 pfn_t pfn)
Nick Piggin423bad62008-04-28 02:13:01 -07001655{
Dan Williams87744ab2016-10-07 17:00:18 -07001656 pgprot_t pgprot = vma->vm_page_prot;
1657
Nick Piggin423bad62008-04-28 02:13:01 -07001658 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1659
1660 if (addr < vma->vm_start || addr >= vma->vm_end)
1661 return -EFAULT;
Dan Williams87744ab2016-10-07 17:00:18 -07001662 if (track_pfn_insert(vma, &pgprot, pfn))
1663 return -EINVAL;
Nick Piggin423bad62008-04-28 02:13:01 -07001664
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001665 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1666 return -EACCES;
1667
Nick Piggin423bad62008-04-28 02:13:01 -07001668 /*
1669 * If we don't have pte special, then we have to use the pfn_valid()
1670 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1671 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001672 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1673 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07001674 */
Dan Williams03fc2da2016-01-26 09:48:05 -08001675 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
Nick Piggin423bad62008-04-28 02:13:01 -07001676 struct page *page;
1677
Dan Williams03fc2da2016-01-26 09:48:05 -08001678 /*
1679 * At this point we are committed to insert_page()
1680 * regardless of whether the caller specified flags that
1681 * result in pfn_t_has_page() == false.
1682 */
1683 page = pfn_to_page(pfn_t_to_pfn(pfn));
Dan Williams87744ab2016-10-07 17:00:18 -07001684 return insert_page(vma, addr, page, pgprot);
Nick Piggin423bad62008-04-28 02:13:01 -07001685 }
Dan Williams87744ab2016-10-07 17:00:18 -07001686 return insert_pfn(vma, addr, pfn, pgprot);
Nick Piggin423bad62008-04-28 02:13:01 -07001687}
1688EXPORT_SYMBOL(vm_insert_mixed);
1689
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001690/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691 * maps a range of physical memory into the requested pages. the old
1692 * mappings are removed. any references to nonexistent pages results
1693 * in null mappings (currently treated as "copy-on-access")
1694 */
1695static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1696 unsigned long addr, unsigned long end,
1697 unsigned long pfn, pgprot_t prot)
1698{
1699 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001700 spinlock_t *ptl;
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001701 int err = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702
Hugh Dickinsc74df322005-10-29 18:16:23 -07001703 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 if (!pte)
1705 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001706 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707 do {
1708 BUG_ON(!pte_none(*pte));
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001709 if (!pfn_modify_allowed(pfn, prot)) {
1710 err = -EACCES;
1711 break;
1712 }
Nick Piggin7e675132008-04-28 02:13:00 -07001713 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 pfn++;
1715 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001716 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001717 pte_unmap_unlock(pte - 1, ptl);
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001718 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719}
1720
1721static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1722 unsigned long addr, unsigned long end,
1723 unsigned long pfn, pgprot_t prot)
1724{
1725 pmd_t *pmd;
1726 unsigned long next;
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001727 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728
1729 pfn -= addr >> PAGE_SHIFT;
1730 pmd = pmd_alloc(mm, pud, addr);
1731 if (!pmd)
1732 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001733 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 do {
1735 next = pmd_addr_end(addr, end);
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001736 err = remap_pte_range(mm, pmd, addr, next,
1737 pfn + (addr >> PAGE_SHIFT), prot);
1738 if (err)
1739 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740 } while (pmd++, addr = next, addr != end);
1741 return 0;
1742}
1743
1744static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1745 unsigned long addr, unsigned long end,
1746 unsigned long pfn, pgprot_t prot)
1747{
1748 pud_t *pud;
1749 unsigned long next;
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001750 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751
1752 pfn -= addr >> PAGE_SHIFT;
1753 pud = pud_alloc(mm, pgd, addr);
1754 if (!pud)
1755 return -ENOMEM;
1756 do {
1757 next = pud_addr_end(addr, end);
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001758 err = remap_pmd_range(mm, pud, addr, next,
1759 pfn + (addr >> PAGE_SHIFT), prot);
1760 if (err)
1761 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762 } while (pud++, addr = next, addr != end);
1763 return 0;
1764}
1765
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001766/**
1767 * remap_pfn_range - remap kernel memory to userspace
1768 * @vma: user vma to map to
1769 * @addr: target user address to start at
1770 * @pfn: physical address of kernel memory
1771 * @size: size of map area
1772 * @prot: page protection flags for this mapping
1773 *
1774 * Note: this is only safe if the mm semaphore is held when called.
1775 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1777 unsigned long pfn, unsigned long size, pgprot_t prot)
1778{
1779 pgd_t *pgd;
1780 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001781 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782 struct mm_struct *mm = vma->vm_mm;
Yongji Xied5957d22016-05-20 16:57:41 -07001783 unsigned long remap_pfn = pfn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 int err;
1785
1786 /*
1787 * Physically remapped pages are special. Tell the
1788 * rest of the world about it:
1789 * VM_IO tells people not to look at these pages
1790 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08001791 * VM_PFNMAP tells the core MM that the base pages are just
1792 * raw PFN mappings, and do not have a "struct page" associated
1793 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001794 * VM_DONTEXPAND
1795 * Disable vma merging and expanding with mremap().
1796 * VM_DONTDUMP
1797 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001798 *
1799 * There's a horrible special case to handle copy-on-write
1800 * behaviour that some programs depend on. We mark the "original"
1801 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001802 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001804 if (is_cow_mapping(vma->vm_flags)) {
1805 if (addr != vma->vm_start || end != vma->vm_end)
1806 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001807 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001808 }
1809
Yongji Xied5957d22016-05-20 16:57:41 -07001810 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001811 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001812 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001813
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001814 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815
1816 BUG_ON(addr >= end);
1817 pfn -= addr >> PAGE_SHIFT;
1818 pgd = pgd_offset(mm, addr);
1819 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 do {
1821 next = pgd_addr_end(addr, end);
1822 err = remap_pud_range(mm, pgd, addr, next,
1823 pfn + (addr >> PAGE_SHIFT), prot);
1824 if (err)
1825 break;
1826 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001827
1828 if (err)
Yongji Xied5957d22016-05-20 16:57:41 -07001829 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001830
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 return err;
1832}
1833EXPORT_SYMBOL(remap_pfn_range);
1834
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07001835/**
1836 * vm_iomap_memory - remap memory to userspace
1837 * @vma: user vma to map to
1838 * @start: start of area
1839 * @len: size of area
1840 *
1841 * This is a simplified io_remap_pfn_range() for common driver use. The
1842 * driver just needs to give us the physical memory range to be mapped,
1843 * we'll figure out the rest from the vma information.
1844 *
1845 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1846 * whatever write-combining details or similar.
1847 */
1848int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1849{
1850 unsigned long vm_len, pfn, pages;
1851
1852 /* Check that the physical memory area passed in looks valid */
1853 if (start + len < start)
1854 return -EINVAL;
1855 /*
1856 * You *really* shouldn't map things that aren't page-aligned,
1857 * but we've historically allowed it because IO memory might
1858 * just have smaller alignment.
1859 */
1860 len += start & ~PAGE_MASK;
1861 pfn = start >> PAGE_SHIFT;
1862 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1863 if (pfn + pages < pfn)
1864 return -EINVAL;
1865
1866 /* We start the mapping 'vm_pgoff' pages into the area */
1867 if (vma->vm_pgoff > pages)
1868 return -EINVAL;
1869 pfn += vma->vm_pgoff;
1870 pages -= vma->vm_pgoff;
1871
1872 /* Can we fit all of the mapping? */
1873 vm_len = vma->vm_end - vma->vm_start;
1874 if (vm_len >> PAGE_SHIFT > pages)
1875 return -EINVAL;
1876
1877 /* Ok, let it rip */
1878 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1879}
1880EXPORT_SYMBOL(vm_iomap_memory);
1881
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001882static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1883 unsigned long addr, unsigned long end,
1884 pte_fn_t fn, void *data)
1885{
1886 pte_t *pte;
1887 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001888 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001889 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001890
1891 pte = (mm == &init_mm) ?
1892 pte_alloc_kernel(pmd, addr) :
1893 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1894 if (!pte)
1895 return -ENOMEM;
1896
1897 BUG_ON(pmd_huge(*pmd));
1898
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001899 arch_enter_lazy_mmu_mode();
1900
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001901 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001902
1903 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001904 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001905 if (err)
1906 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001907 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001908
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001909 arch_leave_lazy_mmu_mode();
1910
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001911 if (mm != &init_mm)
1912 pte_unmap_unlock(pte-1, ptl);
1913 return err;
1914}
1915
1916static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1917 unsigned long addr, unsigned long end,
1918 pte_fn_t fn, void *data)
1919{
1920 pmd_t *pmd;
1921 unsigned long next;
1922 int err;
1923
Andi Kleenceb86872008-07-23 21:27:50 -07001924 BUG_ON(pud_huge(*pud));
1925
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001926 pmd = pmd_alloc(mm, pud, addr);
1927 if (!pmd)
1928 return -ENOMEM;
1929 do {
1930 next = pmd_addr_end(addr, end);
1931 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1932 if (err)
1933 break;
1934 } while (pmd++, addr = next, addr != end);
1935 return err;
1936}
1937
1938static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1939 unsigned long addr, unsigned long end,
1940 pte_fn_t fn, void *data)
1941{
1942 pud_t *pud;
1943 unsigned long next;
1944 int err;
1945
1946 pud = pud_alloc(mm, pgd, addr);
1947 if (!pud)
1948 return -ENOMEM;
1949 do {
1950 next = pud_addr_end(addr, end);
1951 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1952 if (err)
1953 break;
1954 } while (pud++, addr = next, addr != end);
1955 return err;
1956}
1957
1958/*
1959 * Scan a region of virtual memory, filling in page tables as necessary
1960 * and calling a provided function on each leaf page table.
1961 */
1962int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1963 unsigned long size, pte_fn_t fn, void *data)
1964{
1965 pgd_t *pgd;
1966 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001967 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001968 int err;
1969
Mika Penttilä9cb65bc2016-03-15 14:56:45 -07001970 if (WARN_ON(addr >= end))
1971 return -EINVAL;
1972
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001973 pgd = pgd_offset(mm, addr);
1974 do {
1975 next = pgd_addr_end(addr, end);
1976 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1977 if (err)
1978 break;
1979 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001980
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001981 return err;
1982}
1983EXPORT_SYMBOL_GPL(apply_to_page_range);
1984
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985/*
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08001986 * handle_pte_fault chooses page fault handler according to an entry which was
1987 * read non-atomically. Before making any commitment, on those architectures
1988 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1989 * parts, do_swap_page must check under lock before unmapping the pte and
1990 * proceeding (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001991 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001992 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001993static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001994 pte_t *page_table, pte_t orig_pte)
1995{
1996 int same = 1;
1997#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1998 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001999 spinlock_t *ptl = pte_lockptr(mm, pmd);
2000 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002001 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002002 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002003 }
2004#endif
2005 pte_unmap(page_table);
2006 return same;
2007}
2008
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002009static 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 -08002010{
Dan Williams0abdd7a2014-01-21 15:48:12 -08002011 debug_dma_assert_idle(src);
2012
Linus Torvalds6aab3412005-11-28 14:34:23 -08002013 /*
2014 * If the source page was a PFN mapping, we don't have
2015 * a "struct page" for it. We do a best-effort copy by
2016 * just copying from the original user address. If that
2017 * fails, we just zero-fill it. Live with it.
2018 */
2019 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002020 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002021 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2022
2023 /*
2024 * This really shouldn't fail, because the page is there
2025 * in the page tables. But it might just be unreadable,
2026 * in which case we just give up and fill the result with
2027 * zeroes.
2028 */
2029 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002030 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002031 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002032 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002033 } else
2034 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002035}
2036
Michal Hockoc20cd452016-01-14 15:20:12 -08002037static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2038{
2039 struct file *vm_file = vma->vm_file;
2040
2041 if (vm_file)
2042 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2043
2044 /*
2045 * Special mappings (e.g. VDSO) do not have any file so fake
2046 * a default GFP_KERNEL for them.
2047 */
2048 return GFP_KERNEL;
2049}
2050
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051/*
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002052 * Notify the address space that the page is about to become writable so that
2053 * it can prohibit this or wait for the page to get into an appropriate state.
2054 *
2055 * We do this without the lock held, so that it can sleep if it needs to.
2056 */
2057static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2058 unsigned long address)
2059{
2060 struct vm_fault vmf;
2061 int ret;
2062
2063 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2064 vmf.pgoff = page->index;
2065 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Michal Hockoc20cd452016-01-14 15:20:12 -08002066 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002067 vmf.page = page;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002068 vmf.cow_page = NULL;
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002069
2070 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2071 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2072 return ret;
2073 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2074 lock_page(page);
2075 if (!page->mapping) {
2076 unlock_page(page);
2077 return 0; /* retry */
2078 }
2079 ret |= VM_FAULT_LOCKED;
2080 } else
2081 VM_BUG_ON_PAGE(!PageLocked(page), page);
2082 return ret;
2083}
2084
2085/*
Shachar Raindel4e047f82015-04-14 15:46:25 -07002086 * Handle write page faults for pages that can be reused in the current vma
2087 *
2088 * This can happen either due to the mapping being with the VM_SHARED flag,
2089 * or due to us being the last reference standing to the page. In either
2090 * case, all we need to do here is to mark the page as writable and update
2091 * any related book-keeping.
2092 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002093static inline int wp_page_reuse(struct fault_env *fe, pte_t orig_pte,
2094 struct page *page, int page_mkwrite, int dirty_shared)
2095 __releases(fe->ptl)
Shachar Raindel4e047f82015-04-14 15:46:25 -07002096{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002097 struct vm_area_struct *vma = fe->vma;
Shachar Raindel4e047f82015-04-14 15:46:25 -07002098 pte_t entry;
2099 /*
2100 * Clear the pages cpupid information as the existing
2101 * information potentially belongs to a now completely
2102 * unrelated process.
2103 */
2104 if (page)
2105 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2106
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002107 flush_cache_page(vma, fe->address, pte_pfn(orig_pte));
Shachar Raindel4e047f82015-04-14 15:46:25 -07002108 entry = pte_mkyoung(orig_pte);
2109 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002110 if (ptep_set_access_flags(vma, fe->address, fe->pte, entry, 1))
2111 update_mmu_cache(vma, fe->address, fe->pte);
2112 pte_unmap_unlock(fe->pte, fe->ptl);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002113
2114 if (dirty_shared) {
2115 struct address_space *mapping;
2116 int dirtied;
2117
2118 if (!page_mkwrite)
2119 lock_page(page);
2120
2121 dirtied = set_page_dirty(page);
2122 VM_BUG_ON_PAGE(PageAnon(page), page);
2123 mapping = page->mapping;
2124 unlock_page(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002125 put_page(page);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002126
2127 if ((dirtied || page_mkwrite) && mapping) {
2128 /*
2129 * Some device drivers do not set page.mapping
2130 * but still dirty their pages
2131 */
2132 balance_dirty_pages_ratelimited(mapping);
2133 }
2134
2135 if (!page_mkwrite)
2136 file_update_time(vma->vm_file);
2137 }
2138
2139 return VM_FAULT_WRITE;
2140}
2141
2142/*
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002143 * Handle the case of a page which we actually need to copy to a new page.
2144 *
2145 * Called with mmap_sem locked and the old page referenced, but
2146 * without the ptl held.
2147 *
2148 * High level logic flow:
2149 *
2150 * - Allocate a page, copy the content of the old page to the new one.
2151 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2152 * - Take the PTL. If the pte changed, bail out and release the allocated page
2153 * - If the pte is still the way we remember it, update the page table and all
2154 * relevant references. This includes dropping the reference the page-table
2155 * held to the old page, as well as updating the rmap.
2156 * - In any case, unlock the PTL and drop the reference we took to the old page.
2157 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002158static int wp_page_copy(struct fault_env *fe, pte_t orig_pte,
2159 struct page *old_page)
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002160{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002161 struct vm_area_struct *vma = fe->vma;
2162 struct mm_struct *mm = vma->vm_mm;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002163 struct page *new_page = NULL;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002164 pte_t entry;
2165 int page_copied = 0;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002166 const unsigned long mmun_start = fe->address & PAGE_MASK;
2167 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002168 struct mem_cgroup *memcg;
2169
2170 if (unlikely(anon_vma_prepare(vma)))
2171 goto oom;
2172
2173 if (is_zero_pfn(pte_pfn(orig_pte))) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002174 new_page = alloc_zeroed_user_highpage_movable(vma, fe->address);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002175 if (!new_page)
2176 goto oom;
2177 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002178 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2179 fe->address);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002180 if (!new_page)
2181 goto oom;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002182 cow_user_page(new_page, old_page, fe->address, vma);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002183 }
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002184
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002185 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002186 goto oom_free_new;
2187
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002188 __SetPageUptodate(new_page);
2189
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002190 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2191
2192 /*
2193 * Re-check the pte - we dropped the lock
2194 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002195 fe->pte = pte_offset_map_lock(mm, fe->pmd, fe->address, &fe->ptl);
2196 if (likely(pte_same(*fe->pte, orig_pte))) {
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002197 if (old_page) {
2198 if (!PageAnon(old_page)) {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08002199 dec_mm_counter_fast(mm,
2200 mm_counter_file(old_page));
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002201 inc_mm_counter_fast(mm, MM_ANONPAGES);
2202 }
2203 } else {
2204 inc_mm_counter_fast(mm, MM_ANONPAGES);
2205 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002206 flush_cache_page(vma, fe->address, pte_pfn(orig_pte));
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002207 entry = mk_pte(new_page, vma->vm_page_prot);
2208 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2209 /*
2210 * Clear the pte entry and flush it first, before updating the
2211 * pte with the new entry. This will avoid a race condition
2212 * seen in the presence of one thread doing SMC and another
2213 * thread doing COW.
2214 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002215 ptep_clear_flush_notify(vma, fe->address, fe->pte);
2216 page_add_new_anon_rmap(new_page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002217 mem_cgroup_commit_charge(new_page, memcg, false, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002218 lru_cache_add_active_or_unevictable(new_page, vma);
2219 /*
2220 * We call the notify macro here because, when using secondary
2221 * mmu page tables (such as kvm shadow page tables), we want the
2222 * new page to be mapped directly into the secondary page table.
2223 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002224 set_pte_at_notify(mm, fe->address, fe->pte, entry);
2225 update_mmu_cache(vma, fe->address, fe->pte);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002226 if (old_page) {
2227 /*
2228 * Only after switching the pte to the new page may
2229 * we remove the mapcount here. Otherwise another
2230 * process may come and find the rmap count decremented
2231 * before the pte is switched to the new page, and
2232 * "reuse" the old page writing into it while our pte
2233 * here still points into it and can be read by other
2234 * threads.
2235 *
2236 * The critical issue is to order this
2237 * page_remove_rmap with the ptp_clear_flush above.
2238 * Those stores are ordered by (if nothing else,)
2239 * the barrier present in the atomic_add_negative
2240 * in page_remove_rmap.
2241 *
2242 * Then the TLB flush in ptep_clear_flush ensures that
2243 * no process can access the old page before the
2244 * decremented mapcount is visible. And the old page
2245 * cannot be reused until after the decremented
2246 * mapcount is visible. So transitively, TLBs to
2247 * old page will be flushed before it can be reused.
2248 */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002249 page_remove_rmap(old_page, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002250 }
2251
2252 /* Free the old page.. */
2253 new_page = old_page;
2254 page_copied = 1;
2255 } else {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002256 mem_cgroup_cancel_charge(new_page, memcg, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002257 }
2258
2259 if (new_page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002260 put_page(new_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002261
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002262 pte_unmap_unlock(fe->pte, fe->ptl);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002263 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2264 if (old_page) {
2265 /*
2266 * Don't let another task, with possibly unlocked vma,
2267 * keep the mlocked page.
2268 */
2269 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2270 lock_page(old_page); /* LRU manipulation */
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002271 if (PageMlocked(old_page))
2272 munlock_vma_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002273 unlock_page(old_page);
2274 }
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002275 put_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002276 }
2277 return page_copied ? VM_FAULT_WRITE : 0;
2278oom_free_new:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002279 put_page(new_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002280oom:
2281 if (old_page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002282 put_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002283 return VM_FAULT_OOM;
2284}
2285
Boaz Harroshdd906182015-04-15 16:15:11 -07002286/*
2287 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2288 * mapping
2289 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002290static int wp_pfn_shared(struct fault_env *fe, pte_t orig_pte)
Boaz Harroshdd906182015-04-15 16:15:11 -07002291{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002292 struct vm_area_struct *vma = fe->vma;
2293
Boaz Harroshdd906182015-04-15 16:15:11 -07002294 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2295 struct vm_fault vmf = {
2296 .page = NULL,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002297 .pgoff = linear_page_index(vma, fe->address),
2298 .virtual_address =
2299 (void __user *)(fe->address & PAGE_MASK),
Boaz Harroshdd906182015-04-15 16:15:11 -07002300 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2301 };
2302 int ret;
2303
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002304 pte_unmap_unlock(fe->pte, fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002305 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2306 if (ret & VM_FAULT_ERROR)
2307 return ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002308 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2309 &fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002310 /*
2311 * We might have raced with another page fault while we
2312 * released the pte_offset_map_lock.
2313 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002314 if (!pte_same(*fe->pte, orig_pte)) {
2315 pte_unmap_unlock(fe->pte, fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002316 return 0;
2317 }
2318 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002319 return wp_page_reuse(fe, orig_pte, NULL, 0, 0);
Boaz Harroshdd906182015-04-15 16:15:11 -07002320}
2321
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002322static int wp_page_shared(struct fault_env *fe, pte_t orig_pte,
2323 struct page *old_page)
2324 __releases(fe->ptl)
Shachar Raindel93e478d2015-04-14 15:46:35 -07002325{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002326 struct vm_area_struct *vma = fe->vma;
Shachar Raindel93e478d2015-04-14 15:46:35 -07002327 int page_mkwrite = 0;
2328
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002329 get_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002330
Shachar Raindel93e478d2015-04-14 15:46:35 -07002331 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2332 int tmp;
2333
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002334 pte_unmap_unlock(fe->pte, fe->ptl);
2335 tmp = do_page_mkwrite(vma, old_page, fe->address);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002336 if (unlikely(!tmp || (tmp &
2337 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002338 put_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002339 return tmp;
2340 }
2341 /*
2342 * Since we dropped the lock we need to revalidate
2343 * the PTE as someone else may have changed it. If
2344 * they did, we just return, as we can count on the
2345 * MMU to tell us if they didn't also make it writable.
2346 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002347 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2348 &fe->ptl);
2349 if (!pte_same(*fe->pte, orig_pte)) {
Shachar Raindel93e478d2015-04-14 15:46:35 -07002350 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002351 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002352 put_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002353 return 0;
2354 }
2355 page_mkwrite = 1;
2356 }
2357
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002358 return wp_page_reuse(fe, orig_pte, old_page, page_mkwrite, 1);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002359}
2360
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002361/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 * This routine handles present pages, when users try to write
2363 * to a shared page. It is done by copying the page to a new address
2364 * and decrementing the shared-page counter for the old page.
2365 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 * Note that this routine assumes that the protection checks have been
2367 * done by the caller (the low-level page fault routine in most cases).
2368 * Thus we can safely just mark it writable once we've done any necessary
2369 * COW.
2370 *
2371 * We also mark the page dirty at this point even though the page will
2372 * change only once the write actually happens. This avoids a few races,
2373 * and potentially makes it more efficient.
2374 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002375 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2376 * but allow concurrent faults), with pte both mapped and locked.
2377 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002379static int do_wp_page(struct fault_env *fe, pte_t orig_pte)
2380 __releases(fe->ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002382 struct vm_area_struct *vma = fe->vma;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002383 struct page *old_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002385 old_page = vm_normal_page(vma, fe->address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002386 if (!old_page) {
2387 /*
Peter Feiner64e45502014-10-13 15:55:46 -07002388 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2389 * VM_PFNMAP VMA.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002390 *
2391 * We should not cow pages in a shared writeable mapping.
Boaz Harroshdd906182015-04-15 16:15:11 -07002392 * Just mark the pages writable and/or call ops->pfn_mkwrite.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002393 */
2394 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2395 (VM_WRITE|VM_SHARED))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002396 return wp_pfn_shared(fe, orig_pte);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002397
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002398 pte_unmap_unlock(fe->pte, fe->ptl);
2399 return wp_page_copy(fe, orig_pte, old_page);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002400 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002402 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002403 * Take out anonymous pages first, anonymous shared vmas are
2404 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002405 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002406 if (PageAnon(old_page) && !PageKsm(old_page)) {
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002407 int total_mapcount;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002408 if (!trylock_page(old_page)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002409 get_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002410 pte_unmap_unlock(fe->pte, fe->ptl);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002411 lock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002412 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd,
2413 fe->address, &fe->ptl);
2414 if (!pte_same(*fe->pte, orig_pte)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002415 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002416 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002417 put_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002418 return 0;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002419 }
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002420 put_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002421 }
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002422 if (reuse_swap_page(old_page, &total_mapcount)) {
2423 if (total_mapcount == 1) {
2424 /*
2425 * The page is all ours. Move it to
2426 * our anon_vma so the rmap code will
2427 * not search our parent or siblings.
2428 * Protected against the rmap code by
2429 * the page lock.
2430 */
Hugh Dickins5a499732016-07-14 12:07:38 -07002431 page_move_anon_rmap(old_page, vma);
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002432 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002433 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002434 return wp_page_reuse(fe, orig_pte, old_page, 0, 0);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002435 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002436 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002437 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002438 (VM_WRITE|VM_SHARED))) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002439 return wp_page_shared(fe, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441
2442 /*
2443 * Ok, we need to copy. Oh, well..
2444 */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002445 get_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002446
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002447 pte_unmap_unlock(fe->pte, fe->ptl);
2448 return wp_page_copy(fe, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449}
2450
Peter Zijlstra97a89412011-05-24 17:12:04 -07002451static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 unsigned long start_addr, unsigned long end_addr,
2453 struct zap_details *details)
2454{
Al Virof5cc4ee2012-03-05 14:14:20 -05002455 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456}
2457
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002458static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 struct zap_details *details)
2460{
2461 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 pgoff_t vba, vea, zba, zea;
2463
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002464 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466
2467 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002468 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469 zba = details->first_index;
2470 if (zba < vba)
2471 zba = vba;
2472 zea = details->last_index;
2473 if (zea > vea)
2474 zea = vea;
2475
Peter Zijlstra97a89412011-05-24 17:12:04 -07002476 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2478 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002479 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 }
2481}
2482
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483/**
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08002484 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2485 * address_space corresponding to the specified page range in the underlying
2486 * file.
2487 *
Martin Waitz3d410882005-06-23 22:05:21 -07002488 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 * @holebegin: byte in first page to unmap, relative to the start of
2490 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002491 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 * must keep the partial page. In contrast, we must get rid of
2493 * partial pages.
2494 * @holelen: size of prospective hole in bytes. This will be rounded
2495 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2496 * end of the file.
2497 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2498 * but 0 when invalidating pagecache, don't throw away private data.
2499 */
2500void unmap_mapping_range(struct address_space *mapping,
2501 loff_t const holebegin, loff_t const holelen, int even_cows)
2502{
Michal Hockoaac45362016-03-25 14:20:24 -07002503 struct zap_details details = { };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504 pgoff_t hba = holebegin >> PAGE_SHIFT;
2505 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2506
2507 /* Check for overflow. */
2508 if (sizeof(holelen) > sizeof(hlen)) {
2509 long long holeend =
2510 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2511 if (holeend & ~(long long)ULONG_MAX)
2512 hlen = ULONG_MAX - hba + 1;
2513 }
2514
2515 details.check_mapping = even_cows? NULL: mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 details.first_index = hba;
2517 details.last_index = hba + hlen - 1;
2518 if (details.last_index < details.first_index)
2519 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002521 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002522 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 unmap_mapping_range_tree(&mapping->i_mmap, &details);
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002524 i_mmap_unlock_write(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525}
2526EXPORT_SYMBOL(unmap_mapping_range);
2527
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002529 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2530 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002531 * We return with pte unmapped and unlocked.
2532 *
2533 * We return with the mmap_sem locked or unlocked in the same cases
2534 * as does filemap_fault().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002536int do_swap_page(struct fault_env *fe, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002538 struct vm_area_struct *vma = fe->vma;
Hugh Dickins56f31802013-02-22 16:36:10 -08002539 struct page *page, *swapcache;
Johannes Weiner00501b52014-08-08 14:19:20 -07002540 struct mem_cgroup *memcg;
Hugh Dickins65500d22005-10-29 18:15:59 -07002541 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002543 int locked;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002544 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002545 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002547 if (!pte_unmap_same(vma->vm_mm, fe->pmd, fe->pte, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002548 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002549
2550 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002551 if (unlikely(non_swap_entry(entry))) {
2552 if (is_migration_entry(entry)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002553 migration_entry_wait(vma->vm_mm, fe->pmd, fe->address);
Andi Kleend1737fd2009-09-16 11:50:06 +02002554 } else if (is_hwpoison_entry(entry)) {
2555 ret = VM_FAULT_HWPOISON;
2556 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002557 print_bad_pte(vma, fe->address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002558 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002559 }
Christoph Lameter06972122006-06-23 02:03:35 -07002560 goto out;
2561 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002562 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563 page = lookup_swap_cache(entry);
2564 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002565 page = swapin_readahead(entry,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002566 GFP_HIGHUSER_MOVABLE, vma, fe->address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 if (!page) {
2568 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002569 * Back out if somebody else faulted in this pte
2570 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002572 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd,
2573 fe->address, &fe->ptl);
2574 if (likely(pte_same(*fe->pte, orig_pte)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002576 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002577 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578 }
2579
2580 /* Had to read the page from swap area: Major fault */
2581 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002582 count_vm_event(PGMAJFAULT);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002583 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002584 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002585 /*
2586 * hwpoisoned dirty swapcache pages are kept for killing
2587 * owner processes (which may be unknown at hwpoison time)
2588 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002589 ret = VM_FAULT_HWPOISON;
2590 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08002591 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02002592 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593 }
2594
Hugh Dickins56f31802013-02-22 16:36:10 -08002595 swapcache = page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002596 locked = lock_page_or_retry(page, vma->vm_mm, fe->flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002597
Balbir Singh20a10222007-11-14 17:00:33 -08002598 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002599 if (!locked) {
2600 ret |= VM_FAULT_RETRY;
2601 goto out_release;
2602 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002604 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002605 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2606 * release the swapcache from under us. The page pin, and pte_same
2607 * test below, are not enough to exclude that. Even if it is still
2608 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002609 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002610 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002611 goto out_page;
2612
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002613 page = ksm_might_need_to_copy(page, vma, fe->address);
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002614 if (unlikely(!page)) {
2615 ret = VM_FAULT_OOM;
2616 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002617 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08002618 }
2619
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002620 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2621 &memcg, false)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002622 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002623 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002624 }
2625
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002627 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002629 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2630 &fe->ptl);
2631 if (unlikely(!pte_same(*fe->pte, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002632 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002633
2634 if (unlikely(!PageUptodate(page))) {
2635 ret = VM_FAULT_SIGBUS;
2636 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 }
2638
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002639 /*
2640 * The page isn't present yet, go ahead with the fault.
2641 *
2642 * Be careful about the sequence of operations here.
2643 * To get its accounting right, reuse_swap_page() must be called
2644 * while the page is counted on swap but not yet in mapcount i.e.
2645 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2646 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002647 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002649 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2650 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 pte = mk_pte(page, vma->vm_page_prot);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002652 if ((fe->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002654 fe->flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002655 ret |= VM_FAULT_WRITE;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002656 exclusive = RMAP_EXCLUSIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07002659 if (pte_swp_soft_dirty(orig_pte))
2660 pte = pte_mksoft_dirty(pte);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002661 set_pte_at(vma->vm_mm, fe->address, fe->pte, pte);
Johannes Weiner00501b52014-08-08 14:19:20 -07002662 if (page == swapcache) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002663 do_page_add_anon_rmap(page, vma, fe->address, exclusive);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002664 mem_cgroup_commit_charge(page, memcg, true, false);
Minchan Kim1a8018f2016-08-02 14:02:25 -07002665 activate_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -07002666 } else { /* ksm created a completely new copy */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002667 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002668 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002669 lru_cache_add_active_or_unevictable(page, vma);
2670 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002672 swap_free(entry);
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08002673 if (mem_cgroup_swap_full(page) ||
2674 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002675 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002676 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002677 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002678 /*
2679 * Hold the lock to avoid the swap entry to be reused
2680 * until we take the PT lock for the pte_same() check
2681 * (to avoid false positives from pte_same). For
2682 * further safety release the lock after the swap_free
2683 * so that the swap count won't change under a
2684 * parallel locked swapcache.
2685 */
2686 unlock_page(swapcache);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002687 put_page(swapcache);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002688 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002689
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002690 if (fe->flags & FAULT_FLAG_WRITE) {
2691 ret |= do_wp_page(fe, pte);
Hugh Dickins61469f12008-03-04 14:29:04 -08002692 if (ret & VM_FAULT_ERROR)
2693 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 goto out;
2695 }
2696
2697 /* No need to invalidate - it was non-present before */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002698 update_mmu_cache(vma, fe->address, fe->pte);
Hugh Dickins65500d22005-10-29 18:15:59 -07002699unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002700 pte_unmap_unlock(fe->pte, fe->ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701out:
2702 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002703out_nomap:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002704 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002705 pte_unmap_unlock(fe->pte, fe->ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002706out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002707 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002708out_release:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002709 put_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002710 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002711 unlock_page(swapcache);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002712 put_page(swapcache);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002713 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002714 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715}
2716
2717/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002718 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2719 * but allow concurrent faults), and pte mapped but not yet locked.
2720 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002722static int do_anonymous_page(struct fault_env *fe)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002724 struct vm_area_struct *vma = fe->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -07002725 struct mem_cgroup *memcg;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002726 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03002729 /* File mapping without ->vm_ops ? */
2730 if (vma->vm_flags & VM_SHARED)
2731 return VM_FAULT_SIGBUS;
2732
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002733 /*
2734 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2735 * pte_offset_map() on pmds where a huge pmd might be created
2736 * from a different thread.
2737 *
2738 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2739 * parallel threads are excluded by other means.
2740 *
2741 * Here we only have down_read(mmap_sem).
2742 */
2743 if (pte_alloc(vma->vm_mm, fe->pmd, fe->address))
2744 return VM_FAULT_OOM;
2745
2746 /* See the comment in pte_alloc_one_map() */
2747 if (unlikely(pmd_trans_unstable(fe->pmd)))
2748 return 0;
2749
Linus Torvalds11ac5522010-08-14 11:44:56 -07002750 /* Use the zero-page for reads */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002751 if (!(fe->flags & FAULT_FLAG_WRITE) &&
2752 !mm_forbids_zeropage(vma->vm_mm)) {
2753 entry = pte_mkspecial(pfn_pte(my_zero_pfn(fe->address),
Hugh Dickins62eede62009-09-21 17:03:34 -07002754 vma->vm_page_prot));
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002755 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2756 &fe->ptl);
2757 if (!pte_none(*fe->pte))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002758 goto unlock;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002759 /* Deliver the page fault to userland, check inside PT lock */
2760 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002761 pte_unmap_unlock(fe->pte, fe->ptl);
2762 return handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002763 }
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002764 goto setpte;
2765 }
2766
Nick Piggin557ed1f2007-10-16 01:24:40 -07002767 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002768 if (unlikely(anon_vma_prepare(vma)))
2769 goto oom;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002770 page = alloc_zeroed_user_highpage_movable(vma, fe->address);
Nick Piggin557ed1f2007-10-16 01:24:40 -07002771 if (!page)
2772 goto oom;
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002773
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002774 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002775 goto oom_free_page;
2776
Minchan Kim52f37622013-04-29 15:08:15 -07002777 /*
2778 * The memory barrier inside __SetPageUptodate makes sure that
2779 * preceeding stores to the page contents become visible before
2780 * the set_pte_at() write.
2781 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08002782 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783
Nick Piggin557ed1f2007-10-16 01:24:40 -07002784 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002785 if (vma->vm_flags & VM_WRITE)
2786 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002787
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002788 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2789 &fe->ptl);
2790 if (!pte_none(*fe->pte))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002791 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002792
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002793 /* Deliver the page fault to userland, check inside PT lock */
2794 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002795 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002796 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002797 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002798 return handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002799 }
2800
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002801 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2802 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002803 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002804 lru_cache_add_active_or_unevictable(page, vma);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002805setpte:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002806 set_pte_at(vma->vm_mm, fe->address, fe->pte, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807
2808 /* No need to invalidate - it was non-present before */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002809 update_mmu_cache(vma, fe->address, fe->pte);
Hugh Dickins65500d22005-10-29 18:15:59 -07002810unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002811 pte_unmap_unlock(fe->pte, fe->ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002812 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002813release:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002814 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002815 put_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002816 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002817oom_free_page:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002818 put_page(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002819oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820 return VM_FAULT_OOM;
2821}
2822
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002823/*
2824 * The mmap_sem must have been held on entry, and may have been
2825 * released depending on flags and vma->vm_ops->fault() return value.
2826 * See filemap_fault() and __lock_page_retry().
2827 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002828static int __do_fault(struct fault_env *fe, pgoff_t pgoff,
2829 struct page *cow_page, struct page **page, void **entry)
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002830{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002831 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002832 struct vm_fault vmf;
2833 int ret;
2834
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002835 vmf.virtual_address = (void __user *)(fe->address & PAGE_MASK);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002836 vmf.pgoff = pgoff;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002837 vmf.flags = fe->flags;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002838 vmf.page = NULL;
Michal Hockoc20cd452016-01-14 15:20:12 -08002839 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002840 vmf.cow_page = cow_page;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002841
2842 ret = vma->vm_ops->fault(vma, &vmf);
2843 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2844 return ret;
Jan Karabc2466e2016-05-12 18:29:19 +02002845 if (ret & VM_FAULT_DAX_LOCKED) {
2846 *entry = vmf.entry;
2847 return ret;
2848 }
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002849
2850 if (unlikely(PageHWPoison(vmf.page))) {
2851 if (ret & VM_FAULT_LOCKED)
2852 unlock_page(vmf.page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002853 put_page(vmf.page);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002854 return VM_FAULT_HWPOISON;
2855 }
2856
2857 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2858 lock_page(vmf.page);
2859 else
2860 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2861
2862 *page = vmf.page;
2863 return ret;
2864}
2865
Ross Zwislerf2562ed2018-02-23 14:05:27 -08002866/*
2867 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
2868 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
2869 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
2870 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
2871 */
2872static int pmd_devmap_trans_unstable(pmd_t *pmd)
2873{
2874 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
2875}
2876
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002877static int pte_alloc_one_map(struct fault_env *fe)
2878{
2879 struct vm_area_struct *vma = fe->vma;
2880
2881 if (!pmd_none(*fe->pmd))
2882 goto map_pte;
2883 if (fe->prealloc_pte) {
2884 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
2885 if (unlikely(!pmd_none(*fe->pmd))) {
2886 spin_unlock(fe->ptl);
2887 goto map_pte;
2888 }
2889
2890 atomic_long_inc(&vma->vm_mm->nr_ptes);
2891 pmd_populate(vma->vm_mm, fe->pmd, fe->prealloc_pte);
2892 spin_unlock(fe->ptl);
2893 fe->prealloc_pte = 0;
2894 } else if (unlikely(pte_alloc(vma->vm_mm, fe->pmd, fe->address))) {
2895 return VM_FAULT_OOM;
2896 }
2897map_pte:
2898 /*
2899 * If a huge pmd materialized under us just retry later. Use
Ross Zwislerf2562ed2018-02-23 14:05:27 -08002900 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
2901 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
2902 * under us and then back to pmd_none, as a result of MADV_DONTNEED
2903 * running immediately after a huge pmd fault in a different thread of
2904 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
2905 * All we have to ensure is that it is a regular pmd that we can walk
2906 * with pte_offset_map() and we can do that through an atomic read in
2907 * C, which is what pmd_trans_unstable() provides.
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002908 */
Ross Zwislerf2562ed2018-02-23 14:05:27 -08002909 if (pmd_devmap_trans_unstable(fe->pmd))
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002910 return VM_FAULT_NOPAGE;
2911
Ross Zwislerf2562ed2018-02-23 14:05:27 -08002912 /*
2913 * At this point we know that our vmf->pmd points to a page of ptes
2914 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
2915 * for the duration of the fault. If a racing MADV_DONTNEED runs and
2916 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
2917 * be valid and we will re-check to make sure the vmf->pte isn't
2918 * pte_none() under vmf->ptl protection when we return to
2919 * alloc_set_pte().
2920 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002921 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2922 &fe->ptl);
2923 return 0;
2924}
2925
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07002926#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002927
2928#define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
2929static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
2930 unsigned long haddr)
2931{
2932 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
2933 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
2934 return false;
2935 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
2936 return false;
2937 return true;
2938}
2939
2940static int do_set_pmd(struct fault_env *fe, struct page *page)
2941{
2942 struct vm_area_struct *vma = fe->vma;
2943 bool write = fe->flags & FAULT_FLAG_WRITE;
2944 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
2945 pmd_t entry;
2946 int i, ret;
2947
2948 if (!transhuge_vma_suitable(vma, haddr))
2949 return VM_FAULT_FALLBACK;
2950
2951 ret = VM_FAULT_FALLBACK;
2952 page = compound_head(page);
2953
2954 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
2955 if (unlikely(!pmd_none(*fe->pmd)))
2956 goto out;
2957
2958 for (i = 0; i < HPAGE_PMD_NR; i++)
2959 flush_icache_page(vma, page + i);
2960
2961 entry = mk_huge_pmd(page, vma->vm_page_prot);
2962 if (write)
2963 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2964
2965 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
2966 page_add_file_rmap(page, true);
2967
2968 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
2969
2970 update_mmu_cache_pmd(vma, haddr, fe->pmd);
2971
2972 /* fault is handled */
2973 ret = 0;
Kirill A. Shutemov95ecedc2016-07-26 15:25:31 -07002974 count_vm_event(THP_FILE_MAPPED);
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002975out:
2976 spin_unlock(fe->ptl);
2977 return ret;
2978}
2979#else
2980static int do_set_pmd(struct fault_env *fe, struct page *page)
2981{
2982 BUILD_BUG();
2983 return 0;
2984}
2985#endif
2986
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002987/**
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002988 * alloc_set_pte - setup new PTE entry for given page and add reverse page
2989 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002990 *
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002991 * @fe: fault environment
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002992 * @memcg: memcg to charge page (only for private mappings)
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002993 * @page: page to map
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002994 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002995 * Caller must take care of unlocking fe->ptl, if fe->pte is non-NULL on return.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002996 *
2997 * Target users are page handler itself and implementations of
2998 * vm_ops->map_pages.
2999 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003000int alloc_set_pte(struct fault_env *fe, struct mem_cgroup *memcg,
3001 struct page *page)
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003002{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003003 struct vm_area_struct *vma = fe->vma;
3004 bool write = fe->flags & FAULT_FLAG_WRITE;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003005 pte_t entry;
Kirill A. Shutemov10102452016-07-26 15:25:29 -07003006 int ret;
3007
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07003008 if (pmd_none(*fe->pmd) && PageTransCompound(page) &&
3009 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
Kirill A. Shutemov10102452016-07-26 15:25:29 -07003010 /* THP on COW? */
3011 VM_BUG_ON_PAGE(memcg, page);
3012
3013 ret = do_set_pmd(fe, page);
3014 if (ret != VM_FAULT_FALLBACK)
3015 return ret;
3016 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003017
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003018 if (!fe->pte) {
Kirill A. Shutemov10102452016-07-26 15:25:29 -07003019 ret = pte_alloc_one_map(fe);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003020 if (ret)
3021 return ret;
3022 }
3023
3024 /* Re-check under ptl */
3025 if (unlikely(!pte_none(*fe->pte)))
3026 return VM_FAULT_NOPAGE;
3027
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003028 flush_icache_page(vma, page);
3029 entry = mk_pte(page, vma->vm_page_prot);
3030 if (write)
3031 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003032 /* copy-on-write page */
3033 if (write && !(vma->vm_flags & VM_SHARED)) {
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003034 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003035 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003036 mem_cgroup_commit_charge(page, memcg, false, false);
3037 lru_cache_add_active_or_unevictable(page, vma);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003038 } else {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08003039 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07003040 page_add_file_rmap(page, false);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003041 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003042 set_pte_at(vma->vm_mm, fe->address, fe->pte, entry);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003043
3044 /* no need to invalidate: a not-present page won't be cached */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003045 update_mmu_cache(vma, fe->address, fe->pte);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003046
3047 return 0;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003048}
3049
Kirill A. Shutemov3a910532014-08-06 16:08:07 -07003050static unsigned long fault_around_bytes __read_mostly =
3051 rounddown_pow_of_two(65536);
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003052
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003053#ifdef CONFIG_DEBUG_FS
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003054static int fault_around_bytes_get(void *data, u64 *val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003055{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003056 *val = fault_around_bytes;
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003057 return 0;
3058}
3059
Andrey Ryabininb4903d62014-07-30 16:08:35 -07003060/*
3061 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3062 * rounded down to nearest page order. It's what do_fault_around() expects to
3063 * see.
3064 */
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003065static int fault_around_bytes_set(void *data, u64 val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003066{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003067 if (val / PAGE_SIZE > PTRS_PER_PTE)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003068 return -EINVAL;
Andrey Ryabininb4903d62014-07-30 16:08:35 -07003069 if (val > PAGE_SIZE)
3070 fault_around_bytes = rounddown_pow_of_two(val);
3071 else
3072 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003073 return 0;
3074}
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003075DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3076 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003077
3078static int __init fault_around_debugfs(void)
3079{
3080 void *ret;
3081
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003082 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3083 &fault_around_bytes_fops);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003084 if (!ret)
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003085 pr_warn("Failed to create fault_around_bytes in debugfs");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003086 return 0;
3087}
3088late_initcall(fault_around_debugfs);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003089#endif
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003090
Kirill A. Shutemov1fdb4122014-06-04 16:10:55 -07003091/*
3092 * do_fault_around() tries to map few pages around the fault address. The hope
3093 * is that the pages will be needed soon and this will lower the number of
3094 * faults to handle.
3095 *
3096 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3097 * not ready to be mapped: not up-to-date, locked, etc.
3098 *
3099 * This function is called with the page table lock taken. In the split ptlock
3100 * case the page table lock only protects only those entries which belong to
3101 * the page table corresponding to the fault address.
3102 *
3103 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3104 * only once.
3105 *
3106 * fault_around_pages() defines how many pages we'll try to map.
3107 * do_fault_around() expects it to return a power of two less than or equal to
3108 * PTRS_PER_PTE.
3109 *
3110 * The virtual address of the area that we map is naturally aligned to the
3111 * fault_around_pages() value (and therefore to page order). This way it's
3112 * easier to guarantee that we don't cross page table boundaries.
3113 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003114static int do_fault_around(struct fault_env *fe, pgoff_t start_pgoff)
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003115{
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003116 unsigned long address = fe->address, nr_pages, mask;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003117 pgoff_t end_pgoff;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003118 int off, ret = 0;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003119
Jason Low4db0c3c2015-04-15 16:14:08 -07003120 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07003121 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3122
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003123 fe->address = max(address & mask, fe->vma->vm_start);
3124 off = ((address - fe->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003125 start_pgoff -= off;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003126
3127 /*
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003128 * end_pgoff is either end of page table or end of vma
3129 * or fault_around_pages() from start_pgoff, depending what is nearest.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003130 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003131 end_pgoff = start_pgoff -
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003132 ((fe->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003133 PTRS_PER_PTE - 1;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003134 end_pgoff = min3(end_pgoff, vma_pages(fe->vma) + fe->vma->vm_pgoff - 1,
3135 start_pgoff + nr_pages - 1);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003136
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003137 if (pmd_none(*fe->pmd)) {
3138 fe->prealloc_pte = pte_alloc_one(fe->vma->vm_mm, fe->address);
Vegard Nossumc5f88bd2016-08-02 14:02:22 -07003139 if (!fe->prealloc_pte)
3140 goto out;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003141 smp_wmb(); /* See comment in __pte_alloc() */
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003142 }
3143
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003144 fe->vma->vm_ops->map_pages(fe, start_pgoff, end_pgoff);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003145
3146 /* preallocated pagetable is unused: free it */
3147 if (fe->prealloc_pte) {
3148 pte_free(fe->vma->vm_mm, fe->prealloc_pte);
3149 fe->prealloc_pte = 0;
3150 }
3151 /* Huge page is mapped? Page fault is solved */
3152 if (pmd_trans_huge(*fe->pmd)) {
3153 ret = VM_FAULT_NOPAGE;
3154 goto out;
3155 }
3156
3157 /* ->map_pages() haven't done anything useful. Cold page cache? */
3158 if (!fe->pte)
3159 goto out;
3160
3161 /* check if the page fault is solved */
3162 fe->pte -= (fe->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3163 if (!pte_none(*fe->pte))
3164 ret = VM_FAULT_NOPAGE;
3165 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003166out:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003167 fe->address = address;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003168 fe->pte = NULL;
3169 return ret;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003170}
3171
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003172static int do_read_fault(struct fault_env *fe, pgoff_t pgoff)
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003173{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003174 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003175 struct page *fault_page;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003176 int ret = 0;
3177
3178 /*
3179 * Let's call ->map_pages() first and use ->fault() as fallback
3180 * if page by the offset is not ready to be mapped (cold cache or
3181 * something).
3182 */
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08003183 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003184 ret = do_fault_around(fe, pgoff);
3185 if (ret)
3186 return ret;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003187 }
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003188
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003189 ret = __do_fault(fe, pgoff, NULL, &fault_page, NULL);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003190 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3191 return ret;
3192
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003193 ret |= alloc_set_pte(fe, NULL, fault_page);
3194 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003195 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003196 unlock_page(fault_page);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003197 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3198 put_page(fault_page);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003199 return ret;
3200}
3201
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003202static int do_cow_fault(struct fault_env *fe, pgoff_t pgoff)
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003203{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003204 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003205 struct page *fault_page, *new_page;
Jan Karabc2466e2016-05-12 18:29:19 +02003206 void *fault_entry;
Johannes Weiner00501b52014-08-08 14:19:20 -07003207 struct mem_cgroup *memcg;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003208 int ret;
3209
3210 if (unlikely(anon_vma_prepare(vma)))
3211 return VM_FAULT_OOM;
3212
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003213 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, fe->address);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003214 if (!new_page)
3215 return VM_FAULT_OOM;
3216
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003217 if (mem_cgroup_try_charge(new_page, vma->vm_mm, GFP_KERNEL,
3218 &memcg, false)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003219 put_page(new_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003220 return VM_FAULT_OOM;
3221 }
3222
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003223 ret = __do_fault(fe, pgoff, new_page, &fault_page, &fault_entry);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003224 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3225 goto uncharge_out;
3226
Jan Karabc2466e2016-05-12 18:29:19 +02003227 if (!(ret & VM_FAULT_DAX_LOCKED))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003228 copy_user_highpage(new_page, fault_page, fe->address, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003229 __SetPageUptodate(new_page);
3230
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003231 ret |= alloc_set_pte(fe, memcg, new_page);
3232 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003233 pte_unmap_unlock(fe->pte, fe->ptl);
Jan Karabc2466e2016-05-12 18:29:19 +02003234 if (!(ret & VM_FAULT_DAX_LOCKED)) {
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003235 unlock_page(fault_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003236 put_page(fault_page);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003237 } else {
Jan Karabc2466e2016-05-12 18:29:19 +02003238 dax_unlock_mapping_entry(vma->vm_file->f_mapping, pgoff);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003239 }
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003240 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3241 goto uncharge_out;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003242 return ret;
3243uncharge_out:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08003244 mem_cgroup_cancel_charge(new_page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003245 put_page(new_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003246 return ret;
3247}
3248
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003249static int do_shared_fault(struct fault_env *fe, pgoff_t pgoff)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003250{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003251 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003252 struct page *fault_page;
3253 struct address_space *mapping;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003254 int dirtied = 0;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003255 int ret, tmp;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003256
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003257 ret = __do_fault(fe, pgoff, NULL, &fault_page, NULL);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07003258 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003259 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003260
3261 /*
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003262 * Check if the backing address space wants to know that the page is
3263 * about to become writable
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264 */
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003265 if (vma->vm_ops->page_mkwrite) {
3266 unlock_page(fault_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003267 tmp = do_page_mkwrite(vma, fault_page, fe->address);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003268 if (unlikely(!tmp ||
3269 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003270 put_page(fault_page);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003271 return tmp;
3272 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273 }
3274
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003275 ret |= alloc_set_pte(fe, NULL, fault_page);
3276 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003277 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003278 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3279 VM_FAULT_RETRY))) {
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003280 unlock_page(fault_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003281 put_page(fault_page);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003282 return ret;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003283 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003284
3285 if (set_page_dirty(fault_page))
3286 dirtied = 1;
Andrew Mortond82fa872014-12-18 16:17:29 -08003287 /*
3288 * Take a local copy of the address_space - page.mapping may be zeroed
3289 * by truncate after unlock_page(). The address_space itself remains
3290 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3291 * release semantics to prevent the compiler from undoing this copying.
3292 */
Kirill A. Shutemov1c290f62016-01-15 16:52:07 -08003293 mapping = page_rmapping(fault_page);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003294 unlock_page(fault_page);
3295 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3296 /*
3297 * Some device drivers do not set page.mapping but still
3298 * dirty their pages
3299 */
3300 balance_dirty_pages_ratelimited(mapping);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003301 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003302
Johannes Weiner74ec6752015-02-10 14:11:28 -08003303 if (!vma->vm_ops->page_mkwrite)
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003304 file_update_time(vma->vm_file);
3305
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003306 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003307}
Nick Piggind00806b2007-07-19 01:46:57 -07003308
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003309/*
3310 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3311 * but allow concurrent faults).
3312 * The mmap_sem may have been released depending on flags and our
3313 * return value. See filemap_fault() and __lock_page_or_retry().
3314 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003315static int do_fault(struct fault_env *fe)
Nick Piggin54cb8822007-07-19 01:46:59 -07003316{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003317 struct vm_area_struct *vma = fe->vma;
3318 pgoff_t pgoff = linear_page_index(vma, fe->address);
Nick Piggin54cb8822007-07-19 01:46:59 -07003319
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03003320 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3321 if (!vma->vm_ops->fault)
3322 return VM_FAULT_SIGBUS;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003323 if (!(fe->flags & FAULT_FLAG_WRITE))
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003324 return do_read_fault(fe, pgoff);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003325 if (!(vma->vm_flags & VM_SHARED))
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003326 return do_cow_fault(fe, pgoff);
3327 return do_shared_fault(fe, pgoff);
Nick Piggin54cb8822007-07-19 01:46:59 -07003328}
3329
Rashika Kheriab19a9932014-04-03 14:48:02 -07003330static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
Rik van Riel04bb2f92013-10-07 11:29:36 +01003331 unsigned long addr, int page_nid,
3332 int *flags)
Mel Gorman9532fec2012-11-15 01:24:32 +00003333{
3334 get_page(page);
3335
3336 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003337 if (page_nid == numa_node_id()) {
Mel Gorman9532fec2012-11-15 01:24:32 +00003338 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003339 *flags |= TNF_FAULT_LOCAL;
3340 }
Mel Gorman9532fec2012-11-15 01:24:32 +00003341
3342 return mpol_misplaced(page, vma, addr);
3343}
3344
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003345static int do_numa_page(struct fault_env *fe, pte_t pte)
Mel Gormand10e63f2012-10-25 14:16:31 +02003346{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003347 struct vm_area_struct *vma = fe->vma;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003348 struct page *page = NULL;
Mel Gorman8191acb2013-10-07 11:28:45 +01003349 int page_nid = -1;
Peter Zijlstra90572892013-10-07 11:29:20 +01003350 int last_cpupid;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003351 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003352 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07003353 bool was_writable = pte_write(pte);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003354 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003355
3356 /*
3357 * The "pte" at this point cannot be used safely without
3358 * validation through pte_unmap_same(). It's of NUMA type but
3359 * the pfn may be screwed if the read is non atomic.
3360 *
Mel Gorman4d942462015-02-12 14:58:28 -08003361 * We can safely just do a "set_pte_at()", because the old
3362 * page table entry is not accessible, so there would be no
3363 * concurrent hardware modifications to the PTE.
Mel Gormand10e63f2012-10-25 14:16:31 +02003364 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003365 fe->ptl = pte_lockptr(vma->vm_mm, fe->pmd);
3366 spin_lock(fe->ptl);
3367 if (unlikely(!pte_same(*fe->pte, pte))) {
3368 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003369 goto out;
3370 }
3371
Mel Gorman4d942462015-02-12 14:58:28 -08003372 /* Make it present again */
3373 pte = pte_modify(pte, vma->vm_page_prot);
3374 pte = pte_mkyoung(pte);
Mel Gormanb191f9b2015-03-25 15:55:40 -07003375 if (was_writable)
3376 pte = pte_mkwrite(pte);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003377 set_pte_at(vma->vm_mm, fe->address, fe->pte, pte);
3378 update_mmu_cache(vma, fe->address, fe->pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003379
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003380 page = vm_normal_page(vma, fe->address, pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003381 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003382 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gormand10e63f2012-10-25 14:16:31 +02003383 return 0;
3384 }
3385
Kirill A. Shutemove81c4802016-01-15 16:53:49 -08003386 /* TODO: handle PTE-mapped THP */
3387 if (PageCompound(page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003388 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemove81c4802016-01-15 16:53:49 -08003389 return 0;
3390 }
3391
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003392 /*
Mel Gormanbea66fb2015-03-25 15:55:37 -07003393 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3394 * much anyway since they can be in shared cache state. This misses
3395 * the case where a mapping is writable but the process never writes
3396 * to it but pte_write gets cleared during protection updates and
3397 * pte_dirty has unpredictable behaviour between PTE scan updates,
3398 * background writeback, dirty balancing and application behaviour.
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003399 */
Rik van Rield59dc7b2016-09-08 21:30:53 -04003400 if (!pte_write(pte))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003401 flags |= TNF_NO_GROUP;
3402
Rik van Rieldabe1d92013-10-07 11:29:34 +01003403 /*
3404 * Flag if the page is shared between multiple address spaces. This
3405 * is later used when determining whether to group tasks together
3406 */
3407 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3408 flags |= TNF_SHARED;
3409
Peter Zijlstra90572892013-10-07 11:29:20 +01003410 last_cpupid = page_cpupid_last(page);
Mel Gorman8191acb2013-10-07 11:28:45 +01003411 page_nid = page_to_nid(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003412 target_nid = numa_migrate_prep(page, vma, fe->address, page_nid,
3413 &flags);
3414 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003415 if (target_nid == -1) {
Mel Gorman4daae3b2012-11-02 11:33:45 +00003416 put_page(page);
3417 goto out;
3418 }
3419
3420 /* Migrate to the requested node */
Mel Gorman1bc115d2013-10-07 11:29:05 +01003421 migrated = migrate_misplaced_page(page, vma, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003422 if (migrated) {
Mel Gorman8191acb2013-10-07 11:28:45 +01003423 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003424 flags |= TNF_MIGRATED;
Mel Gorman074c2382015-03-25 15:55:42 -07003425 } else
3426 flags |= TNF_MIGRATE_FAIL;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003427
3428out:
Mel Gorman8191acb2013-10-07 11:28:45 +01003429 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003430 task_numa_fault(last_cpupid, page_nid, 1, flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003431 return 0;
3432}
3433
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003434static int create_huge_pmd(struct fault_env *fe)
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003435{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003436 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovfb6dd5f2015-09-09 15:39:35 -07003437 if (vma_is_anonymous(vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003438 return do_huge_pmd_anonymous_page(fe);
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003439 if (vma->vm_ops->pmd_fault)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003440 return vma->vm_ops->pmd_fault(vma, fe->address, fe->pmd,
3441 fe->flags);
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003442 return VM_FAULT_FALLBACK;
3443}
3444
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003445static int wp_huge_pmd(struct fault_env *fe, pmd_t orig_pmd)
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003446{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003447 if (vma_is_anonymous(fe->vma))
3448 return do_huge_pmd_wp_page(fe, orig_pmd);
3449 if (fe->vma->vm_ops->pmd_fault)
3450 return fe->vma->vm_ops->pmd_fault(fe->vma, fe->address, fe->pmd,
3451 fe->flags);
Kirill A. Shutemovaf9e4d52016-07-26 15:25:40 -07003452
3453 /* COW handled on pte level: split pmd */
3454 VM_BUG_ON_VMA(fe->vma->vm_flags & VM_SHARED, fe->vma);
3455 split_huge_pmd(fe->vma, fe->pmd, fe->address);
3456
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003457 return VM_FAULT_FALLBACK;
3458}
3459
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003460static inline bool vma_is_accessible(struct vm_area_struct *vma)
3461{
3462 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3463}
3464
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465/*
3466 * These routines also need to handle stuff like marking pages dirty
3467 * and/or accessed for architectures that don't do it in hardware (most
3468 * RISC architectures). The early dirtying is also good on the i386.
3469 *
3470 * There is also a hook called "update_mmu_cache()" that architectures
3471 * with external mmu caches can use to update those (ie the Sparc or
3472 * PowerPC hashed page tables that act as extended TLBs).
3473 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003474 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3475 * concurrent faults).
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003476 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003477 * The mmap_sem may have been released depending on flags and our return value.
3478 * See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003480static int handle_pte_fault(struct fault_env *fe)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481{
3482 pte_t entry;
3483
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003484 if (unlikely(pmd_none(*fe->pmd))) {
3485 /*
3486 * Leave __pte_alloc() until later: because vm_ops->fault may
3487 * want to allocate huge page, and if we expose page table
3488 * for an instant, it will be difficult to retract from
3489 * concurrent faults and from rmap lookups.
3490 */
3491 fe->pte = NULL;
3492 } else {
3493 /* See comment in pte_alloc_one_map() */
Ross Zwislerf2562ed2018-02-23 14:05:27 -08003494 if (pmd_devmap_trans_unstable(fe->pmd))
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003495 return 0;
3496 /*
3497 * A regular pmd is established and it can't morph into a huge
3498 * pmd from under us anymore at this point because we hold the
3499 * mmap_sem read mode and khugepaged takes it in write mode.
3500 * So now it's safe to run pte_offset_map().
3501 */
3502 fe->pte = pte_offset_map(fe->pmd, fe->address);
3503
3504 entry = *fe->pte;
3505
3506 /*
3507 * some architectures can have larger ptes than wordsize,
3508 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3509 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3510 * atomic accesses. The code below just needs a consistent
3511 * view for the ifs and we later double check anyway with the
3512 * ptl lock held. So here a barrier will do.
3513 */
3514 barrier();
Hugh Dickins65500d22005-10-29 18:15:59 -07003515 if (pte_none(entry)) {
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003516 pte_unmap(fe->pte);
3517 fe->pte = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07003518 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519 }
3520
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003521 if (!fe->pte) {
3522 if (vma_is_anonymous(fe->vma))
3523 return do_anonymous_page(fe);
3524 else
3525 return do_fault(fe);
3526 }
3527
3528 if (!pte_present(entry))
3529 return do_swap_page(fe, entry);
3530
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003531 if (pte_protnone(entry) && vma_is_accessible(fe->vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003532 return do_numa_page(fe, entry);
Mel Gormand10e63f2012-10-25 14:16:31 +02003533
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003534 fe->ptl = pte_lockptr(fe->vma->vm_mm, fe->pmd);
3535 spin_lock(fe->ptl);
3536 if (unlikely(!pte_same(*fe->pte, entry)))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003537 goto unlock;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003538 if (fe->flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 if (!pte_write(entry))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003540 return do_wp_page(fe, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541 entry = pte_mkdirty(entry);
3542 }
3543 entry = pte_mkyoung(entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003544 if (ptep_set_access_flags(fe->vma, fe->address, fe->pte, entry,
3545 fe->flags & FAULT_FLAG_WRITE)) {
3546 update_mmu_cache(fe->vma, fe->address, fe->pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003547 } else {
3548 /*
3549 * This is needed only for protection faults but the arch code
3550 * is not yet telling us if this is a protection fault or not.
3551 * This still avoids useless tlb flushes for .text page faults
3552 * with threads.
3553 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003554 if (fe->flags & FAULT_FLAG_WRITE)
3555 flush_tlb_fix_spurious_fault(fe->vma, fe->address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003556 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003557unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003558 pte_unmap_unlock(fe->pte, fe->ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003559 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560}
3561
3562/*
3563 * By the time we get here, we already hold the mm semaphore
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003564 *
3565 * The mmap_sem may have been released depending on flags and our
3566 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 */
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003568static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3569 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003571 struct fault_env fe = {
3572 .vma = vma,
3573 .address = address,
3574 .flags = flags,
3575 };
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003576 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577 pgd_t *pgd;
3578 pud_t *pud;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581 pud = pud_alloc(mm, pgd, address);
3582 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003583 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003584 fe.pmd = pmd_alloc(mm, pud, address);
3585 if (!fe.pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003586 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003587 if (pmd_none(*fe.pmd) && transparent_hugepage_enabled(vma)) {
3588 int ret = create_huge_pmd(&fe);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003589 if (!(ret & VM_FAULT_FALLBACK))
3590 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003591 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003592 pmd_t orig_pmd = *fe.pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003593 int ret;
3594
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003595 barrier();
Dan Williams5c7fb562016-01-15 16:56:52 -08003596 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003597 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003598 return do_huge_pmd_numa_page(&fe, orig_pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02003599
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003600 if ((fe.flags & FAULT_FLAG_WRITE) &&
3601 !pmd_write(orig_pmd)) {
3602 ret = wp_huge_pmd(&fe, orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003603 if (!(ret & VM_FAULT_FALLBACK))
3604 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003605 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003606 huge_pmd_set_accessed(&fe, orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003607 return 0;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003608 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003609 }
3610 }
3611
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003612 return handle_pte_fault(&fe);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003613}
3614
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003615/*
3616 * By the time we get here, we already hold the mm semaphore
3617 *
3618 * The mmap_sem may have been released depending on flags and our
3619 * return value. See filemap_fault() and __lock_page_or_retry().
3620 */
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003621int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3622 unsigned int flags)
Johannes Weiner519e5242013-09-12 15:13:42 -07003623{
3624 int ret;
3625
3626 __set_current_state(TASK_RUNNING);
3627
3628 count_vm_event(PGFAULT);
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003629 mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
Johannes Weiner519e5242013-09-12 15:13:42 -07003630
3631 /* do counter updates before entering really critical section. */
3632 check_sync_rss_stat(current);
3633
Laurent Dufour3c8381d2017-09-08 16:13:12 -07003634 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3635 flags & FAULT_FLAG_INSTRUCTION,
3636 flags & FAULT_FLAG_REMOTE))
3637 return VM_FAULT_SIGSEGV;
3638
Johannes Weiner519e5242013-09-12 15:13:42 -07003639 /*
3640 * Enable the memcg OOM handling for faults triggered in user
3641 * space. Kernel faults are handled more gracefully.
3642 */
3643 if (flags & FAULT_FLAG_USER)
Johannes Weiner49426422013-10-16 13:46:59 -07003644 mem_cgroup_oom_enable();
Johannes Weiner519e5242013-09-12 15:13:42 -07003645
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003646 if (unlikely(is_vm_hugetlb_page(vma)))
3647 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3648 else
3649 ret = __handle_mm_fault(vma, address, flags);
Johannes Weiner519e5242013-09-12 15:13:42 -07003650
Johannes Weiner49426422013-10-16 13:46:59 -07003651 if (flags & FAULT_FLAG_USER) {
3652 mem_cgroup_oom_disable();
3653 /*
3654 * The task may have entered a memcg OOM situation but
3655 * if the allocation error was handled gracefully (no
3656 * VM_FAULT_OOM), there is no need to kill anything.
3657 * Just clean up the OOM state peacefully.
3658 */
3659 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3660 mem_cgroup_oom_synchronize(false);
3661 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003662
Michal Hocko3f70dc32016-10-07 16:59:06 -07003663 /*
3664 * This mm has been already reaped by the oom reaper and so the
3665 * refault cannot be trusted in general. Anonymous refaults would
3666 * lose data and give a zero page instead e.g. This is especially
3667 * problem for use_mm() because regular tasks will just die and
3668 * the corrupted data will not be visible anywhere while kthread
3669 * will outlive the oom victim and potentially propagate the date
3670 * further.
3671 */
3672 if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
Michal Hocko59ee25d2017-08-18 15:16:12 -07003673 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags))) {
3674
3675 /*
3676 * We are going to enforce SIGBUS but the PF path might have
3677 * dropped the mmap_sem already so take it again so that
3678 * we do not break expectations of all arch specific PF paths
3679 * and g-u-p
3680 */
3681 if (ret & VM_FAULT_RETRY)
3682 down_read(&vma->vm_mm->mmap_sem);
Michal Hocko3f70dc32016-10-07 16:59:06 -07003683 ret = VM_FAULT_SIGBUS;
Michal Hocko59ee25d2017-08-18 15:16:12 -07003684 }
Michal Hocko3f70dc32016-10-07 16:59:06 -07003685
Johannes Weiner519e5242013-09-12 15:13:42 -07003686 return ret;
3687}
Jesse Barnese1d6d012014-12-12 16:55:27 -08003688EXPORT_SYMBOL_GPL(handle_mm_fault);
Johannes Weiner519e5242013-09-12 15:13:42 -07003689
Linus Torvalds1da177e2005-04-16 15:20:36 -07003690#ifndef __PAGETABLE_PUD_FOLDED
3691/*
3692 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003693 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003695int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003696{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003697 pud_t *new = pud_alloc_one(mm, address);
3698 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003699 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700
Nick Piggin362a61a2008-05-14 06:37:36 +02003701 smp_wmb(); /* See comment in __pte_alloc */
3702
Hugh Dickins872fec12005-10-29 18:16:21 -07003703 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003704 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003705 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003706 else
3707 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003708 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003709 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710}
3711#endif /* __PAGETABLE_PUD_FOLDED */
3712
3713#ifndef __PAGETABLE_PMD_FOLDED
3714/*
3715 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003716 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003718int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003720 pmd_t *new = pmd_alloc_one(mm, address);
3721 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003722 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003723
Nick Piggin362a61a2008-05-14 06:37:36 +02003724 smp_wmb(); /* See comment in __pte_alloc */
3725
Hugh Dickins872fec12005-10-29 18:16:21 -07003726 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727#ifndef __ARCH_HAS_4LEVEL_HACK
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003728 if (!pud_present(*pud)) {
3729 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003730 pud_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003731 } else /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003732 pmd_free(mm, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003733#else
3734 if (!pgd_present(*pud)) {
3735 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003736 pgd_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003737 } else /* Another has populated it */
3738 pmd_free(mm, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003740 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003741 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003742}
3743#endif /* __PAGETABLE_PMD_FOLDED */
3744
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003745static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003746 pte_t **ptepp, spinlock_t **ptlp)
3747{
3748 pgd_t *pgd;
3749 pud_t *pud;
3750 pmd_t *pmd;
3751 pte_t *ptep;
3752
3753 pgd = pgd_offset(mm, address);
3754 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3755 goto out;
3756
3757 pud = pud_offset(pgd, address);
3758 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3759 goto out;
3760
3761 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003762 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003763 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3764 goto out;
3765
3766 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3767 if (pmd_huge(*pmd))
3768 goto out;
3769
3770 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3771 if (!ptep)
3772 goto out;
3773 if (!pte_present(*ptep))
3774 goto unlock;
3775 *ptepp = ptep;
3776 return 0;
3777unlock:
3778 pte_unmap_unlock(ptep, *ptlp);
3779out:
3780 return -EINVAL;
3781}
3782
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003783static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3784 pte_t **ptepp, spinlock_t **ptlp)
3785{
3786 int res;
3787
3788 /* (void) is needed to make gcc happy */
3789 (void) __cond_lock(*ptlp,
3790 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3791 return res;
3792}
3793
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003794/**
3795 * follow_pfn - look up PFN at a user virtual address
3796 * @vma: memory mapping
3797 * @address: user virtual address
3798 * @pfn: location to store found PFN
3799 *
3800 * Only IO mappings and raw PFN mappings are allowed.
3801 *
3802 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3803 */
3804int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3805 unsigned long *pfn)
3806{
3807 int ret = -EINVAL;
3808 spinlock_t *ptl;
3809 pte_t *ptep;
3810
3811 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3812 return ret;
3813
3814 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3815 if (ret)
3816 return ret;
3817 *pfn = pte_pfn(*ptep);
3818 pte_unmap_unlock(ptep, ptl);
3819 return 0;
3820}
3821EXPORT_SYMBOL(follow_pfn);
3822
Rik van Riel28b2ee22008-07-23 21:27:05 -07003823#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003824int follow_phys(struct vm_area_struct *vma,
3825 unsigned long address, unsigned int flags,
3826 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003827{
Johannes Weiner03668a42009-06-16 15:32:34 -07003828 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003829 pte_t *ptep, pte;
3830 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003831
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003832 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3833 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003834
Johannes Weiner03668a42009-06-16 15:32:34 -07003835 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003836 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003837 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003838
Rik van Riel28b2ee22008-07-23 21:27:05 -07003839 if ((flags & FOLL_WRITE) && !pte_write(pte))
3840 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003841
3842 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003843 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003844
Johannes Weiner03668a42009-06-16 15:32:34 -07003845 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003846unlock:
3847 pte_unmap_unlock(ptep, ptl);
3848out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003849 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003850}
3851
3852int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3853 void *buf, int len, int write)
3854{
3855 resource_size_t phys_addr;
3856 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003857 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003858 int offset = addr & (PAGE_SIZE-1);
3859
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003860 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003861 return -EINVAL;
3862
Grazvydas Ignotas9cb12d72015-02-12 15:00:19 -08003863 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003864 if (write)
3865 memcpy_toio(maddr + offset, buf, len);
3866 else
3867 memcpy_fromio(buf, maddr + offset, len);
3868 iounmap(maddr);
3869
3870 return len;
3871}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02003872EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003873#endif
3874
David Howells0ec76a12006-09-27 01:50:15 -07003875/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003876 * Access another process' address space as given in mm. If non-NULL, use the
3877 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003878 */
Eric W. Biedermane71b4e02016-11-22 12:06:50 -06003879int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003880 unsigned long addr, void *buf, int len, unsigned int gup_flags)
David Howells0ec76a12006-09-27 01:50:15 -07003881{
David Howells0ec76a12006-09-27 01:50:15 -07003882 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003883 void *old_buf = buf;
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003884 int write = gup_flags & FOLL_WRITE;
David Howells0ec76a12006-09-27 01:50:15 -07003885
David Howells0ec76a12006-09-27 01:50:15 -07003886 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003887 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003888 while (len) {
3889 int bytes, ret, offset;
3890 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003891 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003892
Dave Hansen1e987792016-02-12 13:01:54 -08003893 ret = get_user_pages_remote(tsk, mm, addr, 1,
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003894 gup_flags, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003895 if (ret <= 0) {
Rik van Rieldbffcd02014-08-06 16:08:12 -07003896#ifndef CONFIG_HAVE_IOREMAP_PROT
3897 break;
3898#else
Rik van Riel28b2ee22008-07-23 21:27:05 -07003899 /*
3900 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3901 * we can access using slightly different code.
3902 */
Rik van Riel28b2ee22008-07-23 21:27:05 -07003903 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003904 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003905 break;
3906 if (vma->vm_ops && vma->vm_ops->access)
3907 ret = vma->vm_ops->access(vma, addr, buf,
3908 len, write);
3909 if (ret <= 0)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003910 break;
3911 bytes = ret;
Rik van Rieldbffcd02014-08-06 16:08:12 -07003912#endif
David Howells0ec76a12006-09-27 01:50:15 -07003913 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003914 bytes = len;
3915 offset = addr & (PAGE_SIZE-1);
3916 if (bytes > PAGE_SIZE-offset)
3917 bytes = PAGE_SIZE-offset;
3918
3919 maddr = kmap(page);
3920 if (write) {
3921 copy_to_user_page(vma, page, addr,
3922 maddr + offset, buf, bytes);
3923 set_page_dirty_lock(page);
3924 } else {
3925 copy_from_user_page(vma, page, addr,
3926 buf, maddr + offset, bytes);
3927 }
3928 kunmap(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003929 put_page(page);
David Howells0ec76a12006-09-27 01:50:15 -07003930 }
David Howells0ec76a12006-09-27 01:50:15 -07003931 len -= bytes;
3932 buf += bytes;
3933 addr += bytes;
3934 }
3935 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003936
3937 return buf - old_buf;
3938}
Andi Kleen03252912008-01-30 13:33:18 +01003939
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003940/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003941 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003942 * @mm: the mm_struct of the target address space
3943 * @addr: start address to access
3944 * @buf: source or destination buffer
3945 * @len: number of bytes to transfer
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003946 * @gup_flags: flags modifying lookup behaviour
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003947 *
3948 * The caller must hold a reference on @mm.
3949 */
3950int access_remote_vm(struct mm_struct *mm, unsigned long addr,
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003951 void *buf, int len, unsigned int gup_flags)
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003952{
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003953 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003954}
3955
Andi Kleen03252912008-01-30 13:33:18 +01003956/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003957 * Access another process' address space.
3958 * Source/target buffer must be kernel space,
3959 * Do not walk the page table directly, use get_user_pages
3960 */
3961int access_process_vm(struct task_struct *tsk, unsigned long addr,
Lorenzo Stoakesf307ab62016-10-13 01:20:20 +01003962 void *buf, int len, unsigned int gup_flags)
Stephen Wilson206cb632011-03-13 15:49:19 -04003963{
3964 struct mm_struct *mm;
3965 int ret;
3966
3967 mm = get_task_mm(tsk);
3968 if (!mm)
3969 return 0;
3970
Lorenzo Stoakesf307ab62016-10-13 01:20:20 +01003971 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003972
Stephen Wilson206cb632011-03-13 15:49:19 -04003973 mmput(mm);
3974
3975 return ret;
3976}
3977
Andi Kleen03252912008-01-30 13:33:18 +01003978/*
3979 * Print the name of a VMA.
3980 */
3981void print_vma_addr(char *prefix, unsigned long ip)
3982{
3983 struct mm_struct *mm = current->mm;
3984 struct vm_area_struct *vma;
3985
Ingo Molnare8bff742008-02-13 20:21:06 +01003986 /*
3987 * Do not print if we are in atomic
3988 * contexts (in exception stacks, etc.):
3989 */
3990 if (preempt_count())
3991 return;
3992
Andi Kleen03252912008-01-30 13:33:18 +01003993 down_read(&mm->mmap_sem);
3994 vma = find_vma(mm, ip);
3995 if (vma && vma->vm_file) {
3996 struct file *f = vma->vm_file;
3997 char *buf = (char *)__get_free_page(GFP_KERNEL);
3998 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003999 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01004000
Miklos Szeredi9bf39ab2015-06-19 10:29:13 +02004001 p = file_path(f, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01004002 if (IS_ERR(p))
4003 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004004 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01004005 vma->vm_start,
4006 vma->vm_end - vma->vm_start);
4007 free_page((unsigned long)buf);
4008 }
4009 }
Jeff Liu51a07e52012-07-31 16:43:18 -07004010 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01004011}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004012
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004013#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
David Hildenbrand9ec23532015-05-11 17:52:07 +02004014void __might_fault(const char *file, int line)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004015{
Peter Zijlstra95156f02009-01-12 13:02:11 +01004016 /*
4017 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4018 * holding the mmap_sem, this is safe because kernel memory doesn't
4019 * get paged out, therefore we'll never actually fault, and the
4020 * below annotations will generate false positives.
4021 */
4022 if (segment_eq(get_fs(), KERNEL_DS))
4023 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02004024 if (pagefault_disabled())
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004025 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02004026 __might_sleep(file, line, 0);
4027#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004028 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004029 might_lock_read(&current->mm->mmap_sem);
David Hildenbrand9ec23532015-05-11 17:52:07 +02004030#endif
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004031}
David Hildenbrand9ec23532015-05-11 17:52:07 +02004032EXPORT_SYMBOL(__might_fault);
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004033#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004034
4035#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4036static void clear_gigantic_page(struct page *page,
4037 unsigned long addr,
4038 unsigned int pages_per_huge_page)
4039{
4040 int i;
4041 struct page *p = page;
4042
4043 might_sleep();
4044 for (i = 0; i < pages_per_huge_page;
4045 i++, p = mem_map_next(p, page, i)) {
4046 cond_resched();
4047 clear_user_highpage(p, addr + i * PAGE_SIZE);
4048 }
4049}
4050void clear_huge_page(struct page *page,
4051 unsigned long addr, unsigned int pages_per_huge_page)
4052{
4053 int i;
4054
4055 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4056 clear_gigantic_page(page, addr, pages_per_huge_page);
4057 return;
4058 }
4059
4060 might_sleep();
4061 for (i = 0; i < pages_per_huge_page; i++) {
4062 cond_resched();
4063 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4064 }
4065}
4066
4067static void copy_user_gigantic_page(struct page *dst, struct page *src,
4068 unsigned long addr,
4069 struct vm_area_struct *vma,
4070 unsigned int pages_per_huge_page)
4071{
4072 int i;
4073 struct page *dst_base = dst;
4074 struct page *src_base = src;
4075
4076 for (i = 0; i < pages_per_huge_page; ) {
4077 cond_resched();
4078 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4079
4080 i++;
4081 dst = mem_map_next(dst, dst_base, i);
4082 src = mem_map_next(src, src_base, i);
4083 }
4084}
4085
4086void copy_user_huge_page(struct page *dst, struct page *src,
4087 unsigned long addr, struct vm_area_struct *vma,
4088 unsigned int pages_per_huge_page)
4089{
4090 int i;
4091
4092 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4093 copy_user_gigantic_page(dst, src, addr, vma,
4094 pages_per_huge_page);
4095 return;
4096 }
4097
4098 might_sleep();
4099 for (i = 0; i < pages_per_huge_page; i++) {
4100 cond_resched();
4101 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4102 }
4103}
4104#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004105
Olof Johansson40b64ac2013-12-20 14:28:05 -08004106#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004107
4108static struct kmem_cache *page_ptl_cachep;
4109
4110void __init ptlock_cache_init(void)
4111{
4112 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4113 SLAB_PANIC, NULL);
4114}
4115
Peter Zijlstra539edb52013-11-14 14:31:52 -08004116bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004117{
4118 spinlock_t *ptl;
4119
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004120 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004121 if (!ptl)
4122 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08004123 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004124 return true;
4125}
4126
Peter Zijlstra539edb52013-11-14 14:31:52 -08004127void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004128{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004129 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004130}
4131#endif