blob: 47248dc0b9e1a92e272f517ef76ec2c8b4ac0cb9 [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 if (*batch == NULL) {
377 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
378 if (*batch == NULL) {
379 tlb_remove_table_one(table);
380 return;
381 }
382 (*batch)->nr = 0;
383 }
384 (*batch)->tables[(*batch)->nr++] = table;
385 if ((*batch)->nr == MAX_TABLE_BATCH)
386 tlb_table_flush(tlb);
387}
388
Peter Zijlstra9547d012011-05-24 17:12:14 -0700389#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700390
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 * Note: this doesn't free the actual pages themselves. That
393 * has been handled earlier when unmapping all the memory regions.
394 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000395static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
396 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800398 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700399 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000400 pte_free_tlb(tlb, token, addr);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800401 atomic_long_dec(&tlb->mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402}
403
Hugh Dickinse0da3822005-04-19 13:29:15 -0700404static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
405 unsigned long addr, unsigned long end,
406 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407{
408 pmd_t *pmd;
409 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700410 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Hugh Dickinse0da3822005-04-19 13:29:15 -0700412 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 do {
415 next = pmd_addr_end(addr, end);
416 if (pmd_none_or_clear_bad(pmd))
417 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000418 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 } while (pmd++, addr = next, addr != end);
420
Hugh Dickinse0da3822005-04-19 13:29:15 -0700421 start &= PUD_MASK;
422 if (start < floor)
423 return;
424 if (ceiling) {
425 ceiling &= PUD_MASK;
426 if (!ceiling)
427 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700429 if (end - 1 > ceiling - 1)
430 return;
431
432 pmd = pmd_offset(pud, start);
433 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000434 pmd_free_tlb(tlb, pmd, start);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -0800435 mm_dec_nr_pmds(tlb->mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436}
437
Hugh Dickinse0da3822005-04-19 13:29:15 -0700438static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
439 unsigned long addr, unsigned long end,
440 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441{
442 pud_t *pud;
443 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700444 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445
Hugh Dickinse0da3822005-04-19 13:29:15 -0700446 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 do {
449 next = pud_addr_end(addr, end);
450 if (pud_none_or_clear_bad(pud))
451 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700452 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 } while (pud++, addr = next, addr != end);
454
Hugh Dickinse0da3822005-04-19 13:29:15 -0700455 start &= PGDIR_MASK;
456 if (start < floor)
457 return;
458 if (ceiling) {
459 ceiling &= PGDIR_MASK;
460 if (!ceiling)
461 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700463 if (end - 1 > ceiling - 1)
464 return;
465
466 pud = pud_offset(pgd, start);
467 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000468 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469}
470
471/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700472 * This function frees user-level page tables of a process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 */
Jan Beulich42b77722008-07-23 21:27:10 -0700474void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700475 unsigned long addr, unsigned long end,
476 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477{
478 pgd_t *pgd;
479 unsigned long next;
480
Hugh Dickinse0da3822005-04-19 13:29:15 -0700481 /*
482 * The next few lines have given us lots of grief...
483 *
484 * Why are we testing PMD* at this top level? Because often
485 * there will be no work to do at all, and we'd prefer not to
486 * go all the way down to the bottom just to discover that.
487 *
488 * Why all these "- 1"s? Because 0 represents both the bottom
489 * of the address space and the top of it (using -1 for the
490 * top wouldn't help much: the masks would do the wrong thing).
491 * The rule is that addr 0 and floor 0 refer to the bottom of
492 * the address space, but end 0 and ceiling 0 refer to the top
493 * Comparisons need to use "end - 1" and "ceiling - 1" (though
494 * that end 0 case should be mythical).
495 *
496 * Wherever addr is brought up or ceiling brought down, we must
497 * be careful to reject "the opposite 0" before it confuses the
498 * subsequent tests. But what about where end is brought down
499 * by PMD_SIZE below? no, end can't go down to 0 there.
500 *
501 * Whereas we round start (addr) and ceiling down, by different
502 * masks at different levels, in order to test whether a table
503 * now has no other vmas using it, so can be freed, we don't
504 * bother to round floor or end up - the tests don't need that.
505 */
506
507 addr &= PMD_MASK;
508 if (addr < floor) {
509 addr += PMD_SIZE;
510 if (!addr)
511 return;
512 }
513 if (ceiling) {
514 ceiling &= PMD_MASK;
515 if (!ceiling)
516 return;
517 }
518 if (end - 1 > ceiling - 1)
519 end -= PMD_SIZE;
520 if (addr > end - 1)
521 return;
522
Jan Beulich42b77722008-07-23 21:27:10 -0700523 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524 do {
525 next = pgd_addr_end(addr, end);
526 if (pgd_none_or_clear_bad(pgd))
527 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700528 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700530}
531
Jan Beulich42b77722008-07-23 21:27:10 -0700532void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700533 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700534{
535 while (vma) {
536 struct vm_area_struct *next = vma->vm_next;
537 unsigned long addr = vma->vm_start;
538
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700539 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000540 * Hide vma from rmap and truncate_pagecache before freeing
541 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700542 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800543 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700544 unlink_file_vma(vma);
545
David Gibson9da61ae2006-03-22 00:08:57 -0800546 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700547 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700548 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700549 } else {
550 /*
551 * Optimization: gather nearby vmas into one call down
552 */
553 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800554 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700555 vma = next;
556 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800557 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700558 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700559 }
560 free_pgd_range(tlb, addr, vma->vm_end,
561 floor, next? next->vm_start: ceiling);
562 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700563 vma = next;
564 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565}
566
Kirill A. Shutemov3ed3a4f2016-03-17 14:19:11 -0700567int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800569 spinlock_t *ptl;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800570 pgtable_t new = pte_alloc_one(mm, address);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700571 if (!new)
572 return -ENOMEM;
573
Nick Piggin362a61a2008-05-14 06:37:36 +0200574 /*
575 * Ensure all pte setup (eg. pte page lock and page clearing) are
576 * visible before the pte is made visible to other CPUs by being
577 * put into page tables.
578 *
579 * The other side of the story is the pointer chasing in the page
580 * table walking code (when walking the page table without locking;
581 * ie. most of the time). Fortunately, these data accesses consist
582 * of a chain of data-dependent loads, meaning most CPUs (alpha
583 * being the notable exception) will already guarantee loads are
584 * seen in-order. See the alpha page table accessors for the
585 * smp_read_barrier_depends() barriers in page table walking code.
586 */
587 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
588
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800589 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800590 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800591 atomic_long_inc(&mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800593 new = NULL;
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -0800594 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800595 spin_unlock(ptl);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800596 if (new)
597 pte_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700598 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599}
600
Hugh Dickins1bb36302005-10-29 18:16:22 -0700601int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700603 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
604 if (!new)
605 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606
Nick Piggin362a61a2008-05-14 06:37:36 +0200607 smp_wmb(); /* See comment in __pte_alloc */
608
Hugh Dickins1bb36302005-10-29 18:16:22 -0700609 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800610 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700611 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800612 new = NULL;
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -0800613 }
Hugh Dickins1bb36302005-10-29 18:16:22 -0700614 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800615 if (new)
616 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700617 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618}
619
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800620static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700621{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800622 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
623}
624
625static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
626{
627 int i;
628
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800629 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700630 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800631 for (i = 0; i < NR_MM_COUNTERS; i++)
632 if (rss[i])
633 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700634}
635
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800637 * This function is called to print an error when a bad pte
638 * is found. For example, we might have a PFN-mapped pte in
639 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700640 *
641 * The calling function must still handle the error.
642 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800643static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
644 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700645{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800646 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
647 pud_t *pud = pud_offset(pgd, addr);
648 pmd_t *pmd = pmd_offset(pud, addr);
649 struct address_space *mapping;
650 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800651 static unsigned long resume;
652 static unsigned long nr_shown;
653 static unsigned long nr_unshown;
654
655 /*
656 * Allow a burst of 60 reports, then keep quiet for that minute;
657 * or allow a steady drip of one report per second.
658 */
659 if (nr_shown == 60) {
660 if (time_before(jiffies, resume)) {
661 nr_unshown++;
662 return;
663 }
664 if (nr_unshown) {
Joe Perches11705322016-03-17 14:19:50 -0700665 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
666 nr_unshown);
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800667 nr_unshown = 0;
668 }
669 nr_shown = 0;
670 }
671 if (nr_shown++ == 0)
672 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800673
674 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
675 index = linear_page_index(vma, addr);
676
Joe Perches11705322016-03-17 14:19:50 -0700677 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
678 current->comm,
679 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800680 if (page)
Dave Hansenf0b791a2014-01-23 15:52:49 -0800681 dump_page(page, "bad pte");
Joe Perches11705322016-03-17 14:19:50 -0700682 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
683 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
Hugh Dickins3dc14742009-01-06 14:40:08 -0800684 /*
685 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
686 */
Konstantin Khlebnikov26825822015-04-15 16:15:08 -0700687 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
688 vma->vm_file,
689 vma->vm_ops ? vma->vm_ops->fault : NULL,
690 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
691 mapping ? mapping->a_ops->readpage : NULL);
Nick Pigginb5810032005-10-29 18:16:12 -0700692 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030693 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700694}
695
696/*
Nick Piggin7e675132008-04-28 02:13:00 -0700697 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800698 *
Nick Piggin7e675132008-04-28 02:13:00 -0700699 * "Special" mappings do not wish to be associated with a "struct page" (either
700 * it doesn't exist, or it exists but they don't want to touch it). In this
701 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700702 *
Nick Piggin7e675132008-04-28 02:13:00 -0700703 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
704 * pte bit, in which case this function is trivial. Secondly, an architecture
705 * may not have a spare pte bit, which requires a more complicated scheme,
706 * described below.
707 *
708 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
709 * special mapping (even if there are underlying and valid "struct pages").
710 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800711 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700712 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
713 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700714 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
715 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800716 *
717 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
718 *
Nick Piggin7e675132008-04-28 02:13:00 -0700719 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700720 *
Nick Piggin7e675132008-04-28 02:13:00 -0700721 * This restricts such mappings to be a linear translation from virtual address
722 * to pfn. To get around this restriction, we allow arbitrary mappings so long
723 * as the vma is not a COW mapping; in that case, we know that all ptes are
724 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700725 *
726 *
Nick Piggin7e675132008-04-28 02:13:00 -0700727 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
728 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700729 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
730 * page" backing, however the difference is that _all_ pages with a struct
731 * page (that is, those where pfn_valid is true) are refcounted and considered
732 * normal pages by the VM. The disadvantage is that pages are refcounted
733 * (which can be slower and simply not an option for some PFNMAP users). The
734 * advantage is that we don't have to follow the strict linearity rule of
735 * PFNMAP mappings in order to support COWable mappings.
736 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800737 */
Nick Piggin7e675132008-04-28 02:13:00 -0700738#ifdef __HAVE_ARCH_PTE_SPECIAL
739# define HAVE_PTE_SPECIAL 1
740#else
741# define HAVE_PTE_SPECIAL 0
742#endif
743struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
744 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800745{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800746 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700747
748 if (HAVE_PTE_SPECIAL) {
Hugh Dickinsb38af472014-08-29 15:18:44 -0700749 if (likely(!pte_special(pte)))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800750 goto check_pfn;
David Vrabel667a0a02014-12-18 14:48:15 +0000751 if (vma->vm_ops && vma->vm_ops->find_special_page)
752 return vma->vm_ops->find_special_page(vma, addr);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700753 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
754 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700755 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800756 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700757 return NULL;
758 }
759
760 /* !HAVE_PTE_SPECIAL case follows: */
761
Jared Hulbertb379d792008-04-28 02:12:58 -0700762 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
763 if (vma->vm_flags & VM_MIXEDMAP) {
764 if (!pfn_valid(pfn))
765 return NULL;
766 goto out;
767 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700768 unsigned long off;
769 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700770 if (pfn == vma->vm_pgoff + off)
771 return NULL;
772 if (!is_cow_mapping(vma->vm_flags))
773 return NULL;
774 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800775 }
776
Hugh Dickinsb38af472014-08-29 15:18:44 -0700777 if (is_zero_pfn(pfn))
778 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800779check_pfn:
780 if (unlikely(pfn > highest_memmap_pfn)) {
781 print_bad_pte(vma, addr, pte, NULL);
782 return NULL;
783 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800784
785 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700786 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700787 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800788 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700789out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800790 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800791}
792
Gerald Schaefer28093f92016-04-28 16:18:35 -0700793#ifdef CONFIG_TRANSPARENT_HUGEPAGE
794struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
795 pmd_t pmd)
796{
797 unsigned long pfn = pmd_pfn(pmd);
798
799 /*
800 * There is no pmd_special() but there may be special pmds, e.g.
801 * in a direct-access (dax) mapping, so let's just replicate the
802 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
803 */
804 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
805 if (vma->vm_flags & VM_MIXEDMAP) {
806 if (!pfn_valid(pfn))
807 return NULL;
808 goto out;
809 } else {
810 unsigned long off;
811 off = (addr - vma->vm_start) >> PAGE_SHIFT;
812 if (pfn == vma->vm_pgoff + off)
813 return NULL;
814 if (!is_cow_mapping(vma->vm_flags))
815 return NULL;
816 }
817 }
818
819 if (is_zero_pfn(pfn))
820 return NULL;
821 if (unlikely(pfn > highest_memmap_pfn))
822 return NULL;
823
824 /*
825 * NOTE! We still have PageReserved() pages in the page tables.
826 * eg. VDSO mappings can cause them to exist.
827 */
828out:
829 return pfn_to_page(pfn);
830}
831#endif
832
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800833/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 * copy one vm_area from one task to the other. Assumes the page tables
835 * already present in the new task to be cleared in the whole range
836 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 */
838
Hugh Dickins570a335b2009-12-14 17:58:46 -0800839static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700841 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700842 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843{
Nick Pigginb5810032005-10-29 18:16:12 -0700844 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 pte_t pte = *src_pte;
846 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847
848 /* pte contains position in swap or file, so copy. */
849 if (unlikely(!pte_present(pte))) {
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800850 swp_entry_t entry = pte_to_swp_entry(pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700851
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800852 if (likely(!non_swap_entry(entry))) {
853 if (swap_duplicate(entry) < 0)
854 return entry.val;
Hugh Dickins570a335b2009-12-14 17:58:46 -0800855
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800856 /* make sure dst_mm is on swapoff's mmlist. */
857 if (unlikely(list_empty(&dst_mm->mmlist))) {
858 spin_lock(&mmlist_lock);
859 if (list_empty(&dst_mm->mmlist))
860 list_add(&dst_mm->mmlist,
861 &src_mm->mmlist);
862 spin_unlock(&mmlist_lock);
863 }
864 rss[MM_SWAPENTS]++;
865 } else if (is_migration_entry(entry)) {
866 page = migration_entry_to_page(entry);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800867
Jerome Marchandeca56ff2016-01-14 15:19:26 -0800868 rss[mm_counter(page)]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800869
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800870 if (is_write_migration_entry(entry) &&
871 is_cow_mapping(vm_flags)) {
872 /*
873 * COW mappings require pages in both
874 * parent and child to be set to read.
875 */
876 make_migration_entry_read(&entry);
877 pte = swp_entry_to_pte(entry);
878 if (pte_swp_soft_dirty(*src_pte))
879 pte = pte_swp_mksoft_dirty(pte);
880 set_pte_at(src_mm, addr, src_pte, pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700881 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700883 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 }
885
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 /*
887 * If it's a COW mapping, write protect it both
888 * in the parent and the child
889 */
Linus Torvalds67121172005-12-11 20:38:17 -0800890 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700892 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 }
894
895 /*
896 * If it's a shared mapping, mark it clean in
897 * the child
898 */
899 if (vm_flags & VM_SHARED)
900 pte = pte_mkclean(pte);
901 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800902
903 page = vm_normal_page(vma, addr, pte);
904 if (page) {
905 get_page(page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -0800906 page_dup_rmap(page, false);
Jerome Marchandeca56ff2016-01-14 15:19:26 -0800907 rss[mm_counter(page)]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800908 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700909
910out_set_pte:
911 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800912 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913}
914
Jerome Marchand21bda262014-08-06 16:06:56 -0700915static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800916 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
917 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700919 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700921 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700922 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800923 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800924 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925
926again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800927 init_rss_vec(rss);
928
Hugh Dickinsc74df322005-10-29 18:16:23 -0700929 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 if (!dst_pte)
931 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700932 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700933 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700934 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700935 orig_src_pte = src_pte;
936 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700937 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 do {
940 /*
941 * We are holding two locks at this point - either of them
942 * could generate latencies in another task on another CPU.
943 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700944 if (progress >= 32) {
945 progress = 0;
946 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100947 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700948 break;
949 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 if (pte_none(*src_pte)) {
951 progress++;
952 continue;
953 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800954 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
955 vma, addr, rss);
956 if (entry.val)
957 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 progress += 8;
959 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700961 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700962 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700963 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800964 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700965 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700966 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800967
968 if (entry.val) {
969 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
970 return -ENOMEM;
971 progress = 0;
972 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 if (addr != end)
974 goto again;
975 return 0;
976}
977
978static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
979 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
980 unsigned long addr, unsigned long end)
981{
982 pmd_t *src_pmd, *dst_pmd;
983 unsigned long next;
984
985 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
986 if (!dst_pmd)
987 return -ENOMEM;
988 src_pmd = pmd_offset(src_pud, addr);
989 do {
990 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -0800991 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800992 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800993 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800994 err = copy_huge_pmd(dst_mm, src_mm,
995 dst_pmd, src_pmd, addr, vma);
996 if (err == -ENOMEM)
997 return -ENOMEM;
998 if (!err)
999 continue;
1000 /* fall through */
1001 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 if (pmd_none_or_clear_bad(src_pmd))
1003 continue;
1004 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1005 vma, addr, next))
1006 return -ENOMEM;
1007 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1008 return 0;
1009}
1010
1011static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1012 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1013 unsigned long addr, unsigned long end)
1014{
1015 pud_t *src_pud, *dst_pud;
1016 unsigned long next;
1017
1018 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1019 if (!dst_pud)
1020 return -ENOMEM;
1021 src_pud = pud_offset(src_pgd, addr);
1022 do {
1023 next = pud_addr_end(addr, end);
1024 if (pud_none_or_clear_bad(src_pud))
1025 continue;
1026 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1027 vma, addr, next))
1028 return -ENOMEM;
1029 } while (dst_pud++, src_pud++, addr = next, addr != end);
1030 return 0;
1031}
1032
1033int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1034 struct vm_area_struct *vma)
1035{
1036 pgd_t *src_pgd, *dst_pgd;
1037 unsigned long next;
1038 unsigned long addr = vma->vm_start;
1039 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001040 unsigned long mmun_start; /* For mmu_notifiers */
1041 unsigned long mmun_end; /* For mmu_notifiers */
1042 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001043 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044
Nick Piggind9928952005-08-28 16:49:11 +10001045 /*
1046 * Don't copy ptes where a page fault will fill them correctly.
1047 * Fork becomes much lighter when there are big shared or private
1048 * readonly mappings. The tradeoff is that copy_page_range is more
1049 * efficient than faulting.
1050 */
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08001051 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1052 !vma->anon_vma)
1053 return 0;
Nick Piggind9928952005-08-28 16:49:11 +10001054
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 if (is_vm_hugetlb_page(vma))
1056 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1057
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001058 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001059 /*
1060 * We do not free on error cases below as remove_vma
1061 * gets called on error from higher level routine
1062 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001063 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001064 if (ret)
1065 return ret;
1066 }
1067
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001068 /*
1069 * We need to invalidate the secondary MMU mappings only when
1070 * there could be a permission downgrade on the ptes of the
1071 * parent mm. And a permission downgrade will only happen if
1072 * is_cow_mapping() returns true.
1073 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001074 is_cow = is_cow_mapping(vma->vm_flags);
1075 mmun_start = addr;
1076 mmun_end = end;
1077 if (is_cow)
1078 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1079 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001080
1081 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082 dst_pgd = pgd_offset(dst_mm, addr);
1083 src_pgd = pgd_offset(src_mm, addr);
1084 do {
1085 next = pgd_addr_end(addr, end);
1086 if (pgd_none_or_clear_bad(src_pgd))
1087 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001088 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1089 vma, addr, next))) {
1090 ret = -ENOMEM;
1091 break;
1092 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001094
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001095 if (is_cow)
1096 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001097 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098}
1099
Robin Holt51c6f662005-11-13 16:06:42 -08001100static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001101 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001103 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104{
Nick Pigginb5810032005-10-29 18:16:12 -07001105 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001106 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001107 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001108 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001109 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001110 pte_t *pte;
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001111 swp_entry_t entry;
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001112 struct page *pending_page = NULL;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001113
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001114again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001115 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001116 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1117 pte = start_pte;
Mel Gorman5a1eef72017-08-02 13:31:52 -07001118 flush_tlb_batched_pending(mm);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001119 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 do {
1121 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001122 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001124 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001125
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001127 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001128
Linus Torvalds6aab3412005-11-28 14:34:23 -08001129 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 if (unlikely(details) && page) {
1131 /*
1132 * unmap_shared_mapping_pages() wants to
1133 * invalidate cache without truncating:
1134 * unmap shared but keep private pages.
1135 */
1136 if (details->check_mapping &&
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07001137 details->check_mapping != page_rmapping(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 }
Nick Pigginb5810032005-10-29 18:16:12 -07001140 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001141 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 tlb_remove_tlb_entry(tlb, pte, addr);
1143 if (unlikely(!page))
1144 continue;
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001145
1146 if (!PageAnon(page)) {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001147 if (pte_dirty(ptent)) {
Michal Hockoaac45362016-03-25 14:20:24 -07001148 /*
1149 * oom_reaper cannot tear down dirty
1150 * pages
1151 */
1152 if (unlikely(details && details->ignore_dirty))
1153 continue;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001154 force_flush = 1;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001155 set_page_dirty(page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001156 }
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001157 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001158 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001159 mark_page_accessed(page);
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001160 }
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001161 rss[mm_counter(page)]--;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001162 page_remove_rmap(page, false);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001163 if (unlikely(page_mapcount(page) < 0))
1164 print_bad_pte(vma, addr, ptent, page);
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001165 if (unlikely(__tlb_remove_page(tlb, page))) {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001166 force_flush = 1;
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001167 pending_page = page;
Will Deaconce9ec372014-10-28 13:16:28 -07001168 addr += PAGE_SIZE;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001169 break;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001170 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 continue;
1172 }
Michal Hockoaac45362016-03-25 14:20:24 -07001173 /* only check swap_entries if explicitly asked for in details */
1174 if (unlikely(details && !details->check_swap_entries))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 continue;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001176
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001177 entry = pte_to_swp_entry(ptent);
1178 if (!non_swap_entry(entry))
1179 rss[MM_SWAPENTS]--;
1180 else if (is_migration_entry(entry)) {
1181 struct page *page;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001182
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001183 page = migration_entry_to_page(entry);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001184 rss[mm_counter(page)]--;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001185 }
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001186 if (unlikely(!free_swap_and_cache(entry)))
1187 print_bad_pte(vma, addr, ptent, NULL);
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001188 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001189 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001190
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001191 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001192 arch_leave_lazy_mmu_mode();
Robin Holt51c6f662005-11-13 16:06:42 -08001193
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001194 /* Do the actual TLB flush before dropping ptl */
Will Deaconfb7332a2014-10-29 10:03:09 +00001195 if (force_flush)
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001196 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001197 pte_unmap_unlock(start_pte, ptl);
1198
1199 /*
1200 * If we forced a TLB flush (either due to running out of
1201 * batch buffers or because we needed to flush dirty TLB
1202 * entries before releasing the ptl), free the batched
1203 * memory too. Restart if we didn't do everything.
1204 */
1205 if (force_flush) {
1206 force_flush = 0;
1207 tlb_flush_mmu_free(tlb);
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001208 if (pending_page) {
1209 /* remove the page with new size */
1210 __tlb_remove_pte_page(tlb, pending_page);
1211 pending_page = NULL;
1212 }
Linus Torvalds2b047252013-08-15 11:42:25 -07001213 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001214 goto again;
1215 }
1216
Robin Holt51c6f662005-11-13 16:06:42 -08001217 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218}
1219
Robin Holt51c6f662005-11-13 16:06:42 -08001220static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001221 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001223 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224{
1225 pmd_t *pmd;
1226 unsigned long next;
1227
1228 pmd = pmd_offset(pud, addr);
1229 do {
1230 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -08001231 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001232 if (next - addr != HPAGE_PMD_SIZE) {
Hugh Dickins68428392016-05-05 16:22:09 -07001233 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1234 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001235 split_huge_pmd(vma, pmd, addr);
Shaohua Lif21760b2012-01-12 17:19:16 -08001236 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001237 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001238 /* fall through */
1239 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001240 /*
1241 * Here there can be other concurrent MADV_DONTNEED or
1242 * trans huge page faults running, and if the pmd is
1243 * none or trans huge it can change under us. This is
1244 * because MADV_DONTNEED holds the mmap_sem in read
1245 * mode.
1246 */
1247 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1248 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001249 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001250next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001251 cond_resched();
1252 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001253
1254 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255}
1256
Robin Holt51c6f662005-11-13 16:06:42 -08001257static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001258 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001260 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261{
1262 pud_t *pud;
1263 unsigned long next;
1264
1265 pud = pud_offset(pgd, addr);
1266 do {
1267 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001268 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001270 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1271 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001272
1273 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274}
1275
Michal Hockoaac45362016-03-25 14:20:24 -07001276void unmap_page_range(struct mmu_gather *tlb,
Al Viro038c7aa2012-03-05 13:25:09 -05001277 struct vm_area_struct *vma,
1278 unsigned long addr, unsigned long end,
1279 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280{
1281 pgd_t *pgd;
1282 unsigned long next;
1283
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 BUG_ON(addr >= end);
1285 tlb_start_vma(tlb, vma);
1286 pgd = pgd_offset(vma->vm_mm, addr);
1287 do {
1288 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001289 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001291 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1292 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 tlb_end_vma(tlb, vma);
1294}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001295
Al Virof5cc4ee2012-03-05 14:14:20 -05001296
1297static void unmap_single_vma(struct mmu_gather *tlb,
1298 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001299 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001300 struct zap_details *details)
1301{
1302 unsigned long start = max(vma->vm_start, start_addr);
1303 unsigned long end;
1304
1305 if (start >= vma->vm_end)
1306 return;
1307 end = min(vma->vm_end, end_addr);
1308 if (end <= vma->vm_start)
1309 return;
1310
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301311 if (vma->vm_file)
1312 uprobe_munmap(vma, start, end);
1313
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001314 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001315 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001316
1317 if (start != end) {
1318 if (unlikely(is_vm_hugetlb_page(vma))) {
1319 /*
1320 * It is undesirable to test vma->vm_file as it
1321 * should be non-null for valid hugetlb area.
1322 * However, vm_file will be NULL in the error
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001323 * cleanup path of mmap_region. When
Al Virof5cc4ee2012-03-05 14:14:20 -05001324 * hugetlbfs ->mmap method fails,
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001325 * mmap_region() nullifies vma->vm_file
Al Virof5cc4ee2012-03-05 14:14:20 -05001326 * before calling this function to clean up.
1327 * Since no pte has actually been setup, it is
1328 * safe to do nothing in this case.
1329 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001330 if (vma->vm_file) {
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001331 i_mmap_lock_write(vma->vm_file->f_mapping);
Mel Gormand8333522012-07-31 16:46:20 -07001332 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001333 i_mmap_unlock_write(vma->vm_file->f_mapping);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001334 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001335 } else
1336 unmap_page_range(tlb, vma, start, end, details);
1337 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338}
1339
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340/**
1341 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001342 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 * @vma: the starting vma
1344 * @start_addr: virtual address at which to start unmapping
1345 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001347 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 * Only addresses between `start' and `end' will be unmapped.
1350 *
1351 * The VMA list must be sorted in ascending virtual address order.
1352 *
1353 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1354 * range after unmap_vmas() returns. So the only responsibility here is to
1355 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1356 * drops the lock and schedules.
1357 */
Al Viro6e8bb012012-03-05 13:41:15 -05001358void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001360 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001362 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001364 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001365 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001366 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001367 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368}
1369
1370/**
1371 * zap_page_range - remove user pages in a given range
1372 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001373 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001375 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001376 *
1377 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001379void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 unsigned long size, struct zap_details *details)
1381{
1382 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001383 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001384 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001387 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001388 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001389 mmu_notifier_invalidate_range_start(mm, start, end);
1390 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001391 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001392 mmu_notifier_invalidate_range_end(mm, start, end);
1393 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394}
1395
Jack Steinerc627f9c2008-07-29 22:33:53 -07001396/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001397 * zap_page_range_single - remove user pages in a given range
1398 * @vma: vm_area_struct holding the applicable pages
1399 * @address: starting address of pages to zap
1400 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001401 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001402 *
1403 * The range must fit into one VMA.
1404 */
1405static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1406 unsigned long size, struct zap_details *details)
1407{
1408 struct mm_struct *mm = vma->vm_mm;
1409 struct mmu_gather tlb;
1410 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001411
1412 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001413 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001414 update_hiwater_rss(mm);
1415 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001416 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001417 mmu_notifier_invalidate_range_end(mm, address, end);
1418 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419}
1420
Jack Steinerc627f9c2008-07-29 22:33:53 -07001421/**
1422 * zap_vma_ptes - remove ptes mapping the vma
1423 * @vma: vm_area_struct holding ptes to be zapped
1424 * @address: starting address of pages to zap
1425 * @size: number of bytes to zap
1426 *
1427 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1428 *
1429 * The entire address range must be fully contained within the vma.
1430 *
1431 * Returns 0 if successful.
1432 */
1433int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1434 unsigned long size)
1435{
1436 if (address < vma->vm_start || address + size > vma->vm_end ||
1437 !(vma->vm_flags & VM_PFNMAP))
1438 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001439 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001440 return 0;
1441}
1442EXPORT_SYMBOL_GPL(zap_vma_ptes);
1443
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001444pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001445 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001446{
1447 pgd_t * pgd = pgd_offset(mm, addr);
1448 pud_t * pud = pud_alloc(mm, pgd, addr);
1449 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001450 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001451 if (pmd) {
1452 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001453 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001454 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001455 }
1456 return NULL;
1457}
1458
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001460 * This is the old fallback for page remapping.
1461 *
1462 * For historical reasons, it only allows reserved pages. Only
1463 * old drivers should use this, and they needed to mark their
1464 * pages reserved for the old functions anyway.
1465 */
Nick Piggin423bad62008-04-28 02:13:01 -07001466static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1467 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001468{
Nick Piggin423bad62008-04-28 02:13:01 -07001469 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001470 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001471 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001472 spinlock_t *ptl;
1473
Linus Torvalds238f58d2005-11-29 13:01:56 -08001474 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001475 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001476 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001477 retval = -ENOMEM;
1478 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001479 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001480 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001481 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001482 retval = -EBUSY;
1483 if (!pte_none(*pte))
1484 goto out_unlock;
1485
1486 /* Ok, finally just insert the thing.. */
1487 get_page(page);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001488 inc_mm_counter_fast(mm, mm_counter_file(page));
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001489 page_add_file_rmap(page, false);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001490 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1491
1492 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001493 pte_unmap_unlock(pte, ptl);
1494 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001495out_unlock:
1496 pte_unmap_unlock(pte, ptl);
1497out:
1498 return retval;
1499}
1500
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001501/**
1502 * vm_insert_page - insert single page into user vma
1503 * @vma: user vma to map to
1504 * @addr: target user address of this page
1505 * @page: source kernel page
1506 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001507 * This allows drivers to insert individual pages they've allocated
1508 * into a user vma.
1509 *
1510 * The page has to be a nice clean _individual_ kernel allocation.
1511 * If you allocate a compound page, you need to have marked it as
1512 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001513 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001514 *
1515 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1516 * took an arbitrary page protection parameter. This doesn't allow
1517 * that. Your vma protection will have to be set up correctly, which
1518 * means that if you want a shared writable mapping, you'd better
1519 * ask for a shared writable mapping!
1520 *
1521 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001522 *
1523 * Usually this function is called from f_op->mmap() handler
1524 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1525 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1526 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001527 */
Nick Piggin423bad62008-04-28 02:13:01 -07001528int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1529 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001530{
1531 if (addr < vma->vm_start || addr >= vma->vm_end)
1532 return -EFAULT;
1533 if (!page_count(page))
1534 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001535 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1536 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1537 BUG_ON(vma->vm_flags & VM_PFNMAP);
1538 vma->vm_flags |= VM_MIXEDMAP;
1539 }
Nick Piggin423bad62008-04-28 02:13:01 -07001540 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001541}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001542EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001543
Nick Piggin423bad62008-04-28 02:13:01 -07001544static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001545 pfn_t pfn, pgprot_t prot)
Nick Piggin423bad62008-04-28 02:13:01 -07001546{
1547 struct mm_struct *mm = vma->vm_mm;
1548 int retval;
1549 pte_t *pte, entry;
1550 spinlock_t *ptl;
1551
1552 retval = -ENOMEM;
1553 pte = get_locked_pte(mm, addr, &ptl);
1554 if (!pte)
1555 goto out;
1556 retval = -EBUSY;
1557 if (!pte_none(*pte))
1558 goto out_unlock;
1559
1560 /* Ok, finally just insert the thing.. */
Dan Williams01c8f1c2016-01-15 16:56:40 -08001561 if (pfn_t_devmap(pfn))
1562 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1563 else
1564 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
Nick Piggin423bad62008-04-28 02:13:01 -07001565 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001566 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07001567
1568 retval = 0;
1569out_unlock:
1570 pte_unmap_unlock(pte, ptl);
1571out:
1572 return retval;
1573}
1574
Nick Piggine0dc0d82007-02-12 00:51:36 -08001575/**
1576 * vm_insert_pfn - insert single pfn into user vma
1577 * @vma: user vma to map to
1578 * @addr: target user address of this page
1579 * @pfn: source kernel pfn
1580 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07001581 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08001582 * they've allocated into a user vma. Same comments apply.
1583 *
1584 * This function should only be called from a vm_ops->fault handler, and
1585 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001586 *
1587 * vma cannot be a COW mapping.
1588 *
1589 * As this is called only for pages that do not currently exist, we
1590 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001591 */
1592int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07001593 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001594{
Andy Lutomirski1745cbc2015-12-29 20:12:20 -08001595 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1596}
1597EXPORT_SYMBOL(vm_insert_pfn);
1598
1599/**
1600 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1601 * @vma: user vma to map to
1602 * @addr: target user address of this page
1603 * @pfn: source kernel pfn
1604 * @pgprot: pgprot flags for the inserted page
1605 *
1606 * This is exactly like vm_insert_pfn, except that it allows drivers to
1607 * to override pgprot on a per-page basis.
1608 *
1609 * This only makes sense for IO mappings, and it makes no sense for
1610 * cow mappings. In general, using multiple vmas is preferable;
1611 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1612 * impractical.
1613 */
1614int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1615 unsigned long pfn, pgprot_t pgprot)
1616{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001617 int ret;
Nick Piggin7e675132008-04-28 02:13:00 -07001618 /*
1619 * Technically, architectures with pte_special can avoid all these
1620 * restrictions (same for remap_pfn_range). However we would like
1621 * consistency in testing and feature parity among all, so we should
1622 * try to keep these invariants in place for everybody.
1623 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001624 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1625 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1626 (VM_PFNMAP|VM_MIXEDMAP));
1627 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1628 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001629
Nick Piggin423bad62008-04-28 02:13:01 -07001630 if (addr < vma->vm_start || addr >= vma->vm_end)
1631 return -EFAULT;
Dan Williamsf25748e32016-01-15 16:56:43 -08001632 if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001633 return -EINVAL;
1634
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001635 if (!pfn_modify_allowed(pfn, pgprot))
1636 return -EACCES;
1637
Dan Williams01c8f1c2016-01-15 16:56:40 -08001638 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001639
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001640 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001641}
Andy Lutomirski1745cbc2015-12-29 20:12:20 -08001642EXPORT_SYMBOL(vm_insert_pfn_prot);
Nick Piggine0dc0d82007-02-12 00:51:36 -08001643
Nick Piggin423bad62008-04-28 02:13:01 -07001644int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001645 pfn_t pfn)
Nick Piggin423bad62008-04-28 02:13:01 -07001646{
Dan Williams87744ab2016-10-07 17:00:18 -07001647 pgprot_t pgprot = vma->vm_page_prot;
1648
Nick Piggin423bad62008-04-28 02:13:01 -07001649 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1650
1651 if (addr < vma->vm_start || addr >= vma->vm_end)
1652 return -EFAULT;
Dan Williams87744ab2016-10-07 17:00:18 -07001653 if (track_pfn_insert(vma, &pgprot, pfn))
1654 return -EINVAL;
Nick Piggin423bad62008-04-28 02:13:01 -07001655
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001656 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1657 return -EACCES;
1658
Nick Piggin423bad62008-04-28 02:13:01 -07001659 /*
1660 * If we don't have pte special, then we have to use the pfn_valid()
1661 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1662 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001663 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1664 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07001665 */
Dan Williams03fc2da2016-01-26 09:48:05 -08001666 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
Nick Piggin423bad62008-04-28 02:13:01 -07001667 struct page *page;
1668
Dan Williams03fc2da2016-01-26 09:48:05 -08001669 /*
1670 * At this point we are committed to insert_page()
1671 * regardless of whether the caller specified flags that
1672 * result in pfn_t_has_page() == false.
1673 */
1674 page = pfn_to_page(pfn_t_to_pfn(pfn));
Dan Williams87744ab2016-10-07 17:00:18 -07001675 return insert_page(vma, addr, page, pgprot);
Nick Piggin423bad62008-04-28 02:13:01 -07001676 }
Dan Williams87744ab2016-10-07 17:00:18 -07001677 return insert_pfn(vma, addr, pfn, pgprot);
Nick Piggin423bad62008-04-28 02:13:01 -07001678}
1679EXPORT_SYMBOL(vm_insert_mixed);
1680
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001681/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 * maps a range of physical memory into the requested pages. the old
1683 * mappings are removed. any references to nonexistent pages results
1684 * in null mappings (currently treated as "copy-on-access")
1685 */
1686static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1687 unsigned long addr, unsigned long end,
1688 unsigned long pfn, pgprot_t prot)
1689{
1690 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001691 spinlock_t *ptl;
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001692 int err = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693
Hugh Dickinsc74df322005-10-29 18:16:23 -07001694 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 if (!pte)
1696 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001697 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698 do {
1699 BUG_ON(!pte_none(*pte));
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001700 if (!pfn_modify_allowed(pfn, prot)) {
1701 err = -EACCES;
1702 break;
1703 }
Nick Piggin7e675132008-04-28 02:13:00 -07001704 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 pfn++;
1706 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001707 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001708 pte_unmap_unlock(pte - 1, ptl);
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001709 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710}
1711
1712static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1713 unsigned long addr, unsigned long end,
1714 unsigned long pfn, pgprot_t prot)
1715{
1716 pmd_t *pmd;
1717 unsigned long next;
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001718 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719
1720 pfn -= addr >> PAGE_SHIFT;
1721 pmd = pmd_alloc(mm, pud, addr);
1722 if (!pmd)
1723 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001724 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 do {
1726 next = pmd_addr_end(addr, end);
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001727 err = remap_pte_range(mm, pmd, addr, next,
1728 pfn + (addr >> PAGE_SHIFT), prot);
1729 if (err)
1730 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 } while (pmd++, addr = next, addr != end);
1732 return 0;
1733}
1734
1735static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1736 unsigned long addr, unsigned long end,
1737 unsigned long pfn, pgprot_t prot)
1738{
1739 pud_t *pud;
1740 unsigned long next;
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001741 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742
1743 pfn -= addr >> PAGE_SHIFT;
1744 pud = pud_alloc(mm, pgd, addr);
1745 if (!pud)
1746 return -ENOMEM;
1747 do {
1748 next = pud_addr_end(addr, end);
Andi Kleen7c5b42f2018-06-13 15:48:27 -07001749 err = remap_pmd_range(mm, pud, addr, next,
1750 pfn + (addr >> PAGE_SHIFT), prot);
1751 if (err)
1752 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753 } while (pud++, addr = next, addr != end);
1754 return 0;
1755}
1756
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001757/**
1758 * remap_pfn_range - remap kernel memory to userspace
1759 * @vma: user vma to map to
1760 * @addr: target user address to start at
1761 * @pfn: physical address of kernel memory
1762 * @size: size of map area
1763 * @prot: page protection flags for this mapping
1764 *
1765 * Note: this is only safe if the mm semaphore is held when called.
1766 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1768 unsigned long pfn, unsigned long size, pgprot_t prot)
1769{
1770 pgd_t *pgd;
1771 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001772 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773 struct mm_struct *mm = vma->vm_mm;
Yongji Xied5957d22016-05-20 16:57:41 -07001774 unsigned long remap_pfn = pfn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775 int err;
1776
1777 /*
1778 * Physically remapped pages are special. Tell the
1779 * rest of the world about it:
1780 * VM_IO tells people not to look at these pages
1781 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08001782 * VM_PFNMAP tells the core MM that the base pages are just
1783 * raw PFN mappings, and do not have a "struct page" associated
1784 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001785 * VM_DONTEXPAND
1786 * Disable vma merging and expanding with mremap().
1787 * VM_DONTDUMP
1788 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001789 *
1790 * There's a horrible special case to handle copy-on-write
1791 * behaviour that some programs depend on. We mark the "original"
1792 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001793 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001795 if (is_cow_mapping(vma->vm_flags)) {
1796 if (addr != vma->vm_start || end != vma->vm_end)
1797 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001798 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001799 }
1800
Yongji Xied5957d22016-05-20 16:57:41 -07001801 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001802 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001803 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001804
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001805 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806
1807 BUG_ON(addr >= end);
1808 pfn -= addr >> PAGE_SHIFT;
1809 pgd = pgd_offset(mm, addr);
1810 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811 do {
1812 next = pgd_addr_end(addr, end);
1813 err = remap_pud_range(mm, pgd, addr, next,
1814 pfn + (addr >> PAGE_SHIFT), prot);
1815 if (err)
1816 break;
1817 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001818
1819 if (err)
Yongji Xied5957d22016-05-20 16:57:41 -07001820 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001821
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 return err;
1823}
1824EXPORT_SYMBOL(remap_pfn_range);
1825
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07001826/**
1827 * vm_iomap_memory - remap memory to userspace
1828 * @vma: user vma to map to
1829 * @start: start of area
1830 * @len: size of area
1831 *
1832 * This is a simplified io_remap_pfn_range() for common driver use. The
1833 * driver just needs to give us the physical memory range to be mapped,
1834 * we'll figure out the rest from the vma information.
1835 *
1836 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1837 * whatever write-combining details or similar.
1838 */
1839int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1840{
1841 unsigned long vm_len, pfn, pages;
1842
1843 /* Check that the physical memory area passed in looks valid */
1844 if (start + len < start)
1845 return -EINVAL;
1846 /*
1847 * You *really* shouldn't map things that aren't page-aligned,
1848 * but we've historically allowed it because IO memory might
1849 * just have smaller alignment.
1850 */
1851 len += start & ~PAGE_MASK;
1852 pfn = start >> PAGE_SHIFT;
1853 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1854 if (pfn + pages < pfn)
1855 return -EINVAL;
1856
1857 /* We start the mapping 'vm_pgoff' pages into the area */
1858 if (vma->vm_pgoff > pages)
1859 return -EINVAL;
1860 pfn += vma->vm_pgoff;
1861 pages -= vma->vm_pgoff;
1862
1863 /* Can we fit all of the mapping? */
1864 vm_len = vma->vm_end - vma->vm_start;
1865 if (vm_len >> PAGE_SHIFT > pages)
1866 return -EINVAL;
1867
1868 /* Ok, let it rip */
1869 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1870}
1871EXPORT_SYMBOL(vm_iomap_memory);
1872
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001873static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1874 unsigned long addr, unsigned long end,
1875 pte_fn_t fn, void *data)
1876{
1877 pte_t *pte;
1878 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001879 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001880 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001881
1882 pte = (mm == &init_mm) ?
1883 pte_alloc_kernel(pmd, addr) :
1884 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1885 if (!pte)
1886 return -ENOMEM;
1887
1888 BUG_ON(pmd_huge(*pmd));
1889
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001890 arch_enter_lazy_mmu_mode();
1891
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001892 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001893
1894 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001895 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001896 if (err)
1897 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001898 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001899
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001900 arch_leave_lazy_mmu_mode();
1901
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001902 if (mm != &init_mm)
1903 pte_unmap_unlock(pte-1, ptl);
1904 return err;
1905}
1906
1907static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1908 unsigned long addr, unsigned long end,
1909 pte_fn_t fn, void *data)
1910{
1911 pmd_t *pmd;
1912 unsigned long next;
1913 int err;
1914
Andi Kleenceb86872008-07-23 21:27:50 -07001915 BUG_ON(pud_huge(*pud));
1916
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001917 pmd = pmd_alloc(mm, pud, addr);
1918 if (!pmd)
1919 return -ENOMEM;
1920 do {
1921 next = pmd_addr_end(addr, end);
1922 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1923 if (err)
1924 break;
1925 } while (pmd++, addr = next, addr != end);
1926 return err;
1927}
1928
1929static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1930 unsigned long addr, unsigned long end,
1931 pte_fn_t fn, void *data)
1932{
1933 pud_t *pud;
1934 unsigned long next;
1935 int err;
1936
1937 pud = pud_alloc(mm, pgd, addr);
1938 if (!pud)
1939 return -ENOMEM;
1940 do {
1941 next = pud_addr_end(addr, end);
1942 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1943 if (err)
1944 break;
1945 } while (pud++, addr = next, addr != end);
1946 return err;
1947}
1948
1949/*
1950 * Scan a region of virtual memory, filling in page tables as necessary
1951 * and calling a provided function on each leaf page table.
1952 */
1953int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1954 unsigned long size, pte_fn_t fn, void *data)
1955{
1956 pgd_t *pgd;
1957 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001958 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001959 int err;
1960
Mika Penttilä9cb65bc2016-03-15 14:56:45 -07001961 if (WARN_ON(addr >= end))
1962 return -EINVAL;
1963
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001964 pgd = pgd_offset(mm, addr);
1965 do {
1966 next = pgd_addr_end(addr, end);
1967 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1968 if (err)
1969 break;
1970 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001971
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001972 return err;
1973}
1974EXPORT_SYMBOL_GPL(apply_to_page_range);
1975
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976/*
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08001977 * handle_pte_fault chooses page fault handler according to an entry which was
1978 * read non-atomically. Before making any commitment, on those architectures
1979 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1980 * parts, do_swap_page must check under lock before unmapping the pte and
1981 * proceeding (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001982 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001983 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001984static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001985 pte_t *page_table, pte_t orig_pte)
1986{
1987 int same = 1;
1988#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1989 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001990 spinlock_t *ptl = pte_lockptr(mm, pmd);
1991 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001992 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001993 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001994 }
1995#endif
1996 pte_unmap(page_table);
1997 return same;
1998}
1999
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002000static 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 -08002001{
Dan Williams0abdd7a2014-01-21 15:48:12 -08002002 debug_dma_assert_idle(src);
2003
Linus Torvalds6aab3412005-11-28 14:34:23 -08002004 /*
2005 * If the source page was a PFN mapping, we don't have
2006 * a "struct page" for it. We do a best-effort copy by
2007 * just copying from the original user address. If that
2008 * fails, we just zero-fill it. Live with it.
2009 */
2010 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002011 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002012 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2013
2014 /*
2015 * This really shouldn't fail, because the page is there
2016 * in the page tables. But it might just be unreadable,
2017 * in which case we just give up and fill the result with
2018 * zeroes.
2019 */
2020 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002021 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002022 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002023 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002024 } else
2025 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002026}
2027
Michal Hockoc20cd452016-01-14 15:20:12 -08002028static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2029{
2030 struct file *vm_file = vma->vm_file;
2031
2032 if (vm_file)
2033 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2034
2035 /*
2036 * Special mappings (e.g. VDSO) do not have any file so fake
2037 * a default GFP_KERNEL for them.
2038 */
2039 return GFP_KERNEL;
2040}
2041
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042/*
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002043 * Notify the address space that the page is about to become writable so that
2044 * it can prohibit this or wait for the page to get into an appropriate state.
2045 *
2046 * We do this without the lock held, so that it can sleep if it needs to.
2047 */
2048static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2049 unsigned long address)
2050{
2051 struct vm_fault vmf;
2052 int ret;
2053
2054 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2055 vmf.pgoff = page->index;
2056 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Michal Hockoc20cd452016-01-14 15:20:12 -08002057 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002058 vmf.page = page;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002059 vmf.cow_page = NULL;
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002060
2061 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2062 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2063 return ret;
2064 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2065 lock_page(page);
2066 if (!page->mapping) {
2067 unlock_page(page);
2068 return 0; /* retry */
2069 }
2070 ret |= VM_FAULT_LOCKED;
2071 } else
2072 VM_BUG_ON_PAGE(!PageLocked(page), page);
2073 return ret;
2074}
2075
2076/*
Shachar Raindel4e047f82015-04-14 15:46:25 -07002077 * Handle write page faults for pages that can be reused in the current vma
2078 *
2079 * This can happen either due to the mapping being with the VM_SHARED flag,
2080 * or due to us being the last reference standing to the page. In either
2081 * case, all we need to do here is to mark the page as writable and update
2082 * any related book-keeping.
2083 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002084static inline int wp_page_reuse(struct fault_env *fe, pte_t orig_pte,
2085 struct page *page, int page_mkwrite, int dirty_shared)
2086 __releases(fe->ptl)
Shachar Raindel4e047f82015-04-14 15:46:25 -07002087{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002088 struct vm_area_struct *vma = fe->vma;
Shachar Raindel4e047f82015-04-14 15:46:25 -07002089 pte_t entry;
2090 /*
2091 * Clear the pages cpupid information as the existing
2092 * information potentially belongs to a now completely
2093 * unrelated process.
2094 */
2095 if (page)
2096 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2097
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002098 flush_cache_page(vma, fe->address, pte_pfn(orig_pte));
Shachar Raindel4e047f82015-04-14 15:46:25 -07002099 entry = pte_mkyoung(orig_pte);
2100 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002101 if (ptep_set_access_flags(vma, fe->address, fe->pte, entry, 1))
2102 update_mmu_cache(vma, fe->address, fe->pte);
2103 pte_unmap_unlock(fe->pte, fe->ptl);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002104
2105 if (dirty_shared) {
2106 struct address_space *mapping;
2107 int dirtied;
2108
2109 if (!page_mkwrite)
2110 lock_page(page);
2111
2112 dirtied = set_page_dirty(page);
2113 VM_BUG_ON_PAGE(PageAnon(page), page);
2114 mapping = page->mapping;
2115 unlock_page(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002116 put_page(page);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002117
2118 if ((dirtied || page_mkwrite) && mapping) {
2119 /*
2120 * Some device drivers do not set page.mapping
2121 * but still dirty their pages
2122 */
2123 balance_dirty_pages_ratelimited(mapping);
2124 }
2125
2126 if (!page_mkwrite)
2127 file_update_time(vma->vm_file);
2128 }
2129
2130 return VM_FAULT_WRITE;
2131}
2132
2133/*
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002134 * Handle the case of a page which we actually need to copy to a new page.
2135 *
2136 * Called with mmap_sem locked and the old page referenced, but
2137 * without the ptl held.
2138 *
2139 * High level logic flow:
2140 *
2141 * - Allocate a page, copy the content of the old page to the new one.
2142 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2143 * - Take the PTL. If the pte changed, bail out and release the allocated page
2144 * - If the pte is still the way we remember it, update the page table and all
2145 * relevant references. This includes dropping the reference the page-table
2146 * held to the old page, as well as updating the rmap.
2147 * - In any case, unlock the PTL and drop the reference we took to the old page.
2148 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002149static int wp_page_copy(struct fault_env *fe, pte_t orig_pte,
2150 struct page *old_page)
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002151{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002152 struct vm_area_struct *vma = fe->vma;
2153 struct mm_struct *mm = vma->vm_mm;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002154 struct page *new_page = NULL;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002155 pte_t entry;
2156 int page_copied = 0;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002157 const unsigned long mmun_start = fe->address & PAGE_MASK;
2158 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002159 struct mem_cgroup *memcg;
2160
2161 if (unlikely(anon_vma_prepare(vma)))
2162 goto oom;
2163
2164 if (is_zero_pfn(pte_pfn(orig_pte))) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002165 new_page = alloc_zeroed_user_highpage_movable(vma, fe->address);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002166 if (!new_page)
2167 goto oom;
2168 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002169 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2170 fe->address);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002171 if (!new_page)
2172 goto oom;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002173 cow_user_page(new_page, old_page, fe->address, vma);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002174 }
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002175
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002176 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002177 goto oom_free_new;
2178
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002179 __SetPageUptodate(new_page);
2180
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002181 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2182
2183 /*
2184 * Re-check the pte - we dropped the lock
2185 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002186 fe->pte = pte_offset_map_lock(mm, fe->pmd, fe->address, &fe->ptl);
2187 if (likely(pte_same(*fe->pte, orig_pte))) {
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002188 if (old_page) {
2189 if (!PageAnon(old_page)) {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08002190 dec_mm_counter_fast(mm,
2191 mm_counter_file(old_page));
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002192 inc_mm_counter_fast(mm, MM_ANONPAGES);
2193 }
2194 } else {
2195 inc_mm_counter_fast(mm, MM_ANONPAGES);
2196 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002197 flush_cache_page(vma, fe->address, pte_pfn(orig_pte));
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002198 entry = mk_pte(new_page, vma->vm_page_prot);
2199 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2200 /*
2201 * Clear the pte entry and flush it first, before updating the
2202 * pte with the new entry. This will avoid a race condition
2203 * seen in the presence of one thread doing SMC and another
2204 * thread doing COW.
2205 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002206 ptep_clear_flush_notify(vma, fe->address, fe->pte);
2207 page_add_new_anon_rmap(new_page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002208 mem_cgroup_commit_charge(new_page, memcg, false, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002209 lru_cache_add_active_or_unevictable(new_page, vma);
2210 /*
2211 * We call the notify macro here because, when using secondary
2212 * mmu page tables (such as kvm shadow page tables), we want the
2213 * new page to be mapped directly into the secondary page table.
2214 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002215 set_pte_at_notify(mm, fe->address, fe->pte, entry);
2216 update_mmu_cache(vma, fe->address, fe->pte);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002217 if (old_page) {
2218 /*
2219 * Only after switching the pte to the new page may
2220 * we remove the mapcount here. Otherwise another
2221 * process may come and find the rmap count decremented
2222 * before the pte is switched to the new page, and
2223 * "reuse" the old page writing into it while our pte
2224 * here still points into it and can be read by other
2225 * threads.
2226 *
2227 * The critical issue is to order this
2228 * page_remove_rmap with the ptp_clear_flush above.
2229 * Those stores are ordered by (if nothing else,)
2230 * the barrier present in the atomic_add_negative
2231 * in page_remove_rmap.
2232 *
2233 * Then the TLB flush in ptep_clear_flush ensures that
2234 * no process can access the old page before the
2235 * decremented mapcount is visible. And the old page
2236 * cannot be reused until after the decremented
2237 * mapcount is visible. So transitively, TLBs to
2238 * old page will be flushed before it can be reused.
2239 */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002240 page_remove_rmap(old_page, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002241 }
2242
2243 /* Free the old page.. */
2244 new_page = old_page;
2245 page_copied = 1;
2246 } else {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002247 mem_cgroup_cancel_charge(new_page, memcg, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002248 }
2249
2250 if (new_page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002251 put_page(new_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002252
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002253 pte_unmap_unlock(fe->pte, fe->ptl);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002254 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2255 if (old_page) {
2256 /*
2257 * Don't let another task, with possibly unlocked vma,
2258 * keep the mlocked page.
2259 */
2260 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2261 lock_page(old_page); /* LRU manipulation */
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002262 if (PageMlocked(old_page))
2263 munlock_vma_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002264 unlock_page(old_page);
2265 }
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002266 put_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002267 }
2268 return page_copied ? VM_FAULT_WRITE : 0;
2269oom_free_new:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002270 put_page(new_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002271oom:
2272 if (old_page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002273 put_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002274 return VM_FAULT_OOM;
2275}
2276
Boaz Harroshdd906182015-04-15 16:15:11 -07002277/*
2278 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2279 * mapping
2280 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002281static int wp_pfn_shared(struct fault_env *fe, pte_t orig_pte)
Boaz Harroshdd906182015-04-15 16:15:11 -07002282{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002283 struct vm_area_struct *vma = fe->vma;
2284
Boaz Harroshdd906182015-04-15 16:15:11 -07002285 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2286 struct vm_fault vmf = {
2287 .page = NULL,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002288 .pgoff = linear_page_index(vma, fe->address),
2289 .virtual_address =
2290 (void __user *)(fe->address & PAGE_MASK),
Boaz Harroshdd906182015-04-15 16:15:11 -07002291 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2292 };
2293 int ret;
2294
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002295 pte_unmap_unlock(fe->pte, fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002296 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2297 if (ret & VM_FAULT_ERROR)
2298 return ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002299 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2300 &fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002301 /*
2302 * We might have raced with another page fault while we
2303 * released the pte_offset_map_lock.
2304 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002305 if (!pte_same(*fe->pte, orig_pte)) {
2306 pte_unmap_unlock(fe->pte, fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002307 return 0;
2308 }
2309 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002310 return wp_page_reuse(fe, orig_pte, NULL, 0, 0);
Boaz Harroshdd906182015-04-15 16:15:11 -07002311}
2312
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002313static int wp_page_shared(struct fault_env *fe, pte_t orig_pte,
2314 struct page *old_page)
2315 __releases(fe->ptl)
Shachar Raindel93e478d2015-04-14 15:46:35 -07002316{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002317 struct vm_area_struct *vma = fe->vma;
Shachar Raindel93e478d2015-04-14 15:46:35 -07002318 int page_mkwrite = 0;
2319
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002320 get_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002321
Shachar Raindel93e478d2015-04-14 15:46:35 -07002322 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2323 int tmp;
2324
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002325 pte_unmap_unlock(fe->pte, fe->ptl);
2326 tmp = do_page_mkwrite(vma, old_page, fe->address);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002327 if (unlikely(!tmp || (tmp &
2328 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002329 put_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002330 return tmp;
2331 }
2332 /*
2333 * Since we dropped the lock we need to revalidate
2334 * the PTE as someone else may have changed it. If
2335 * they did, we just return, as we can count on the
2336 * MMU to tell us if they didn't also make it writable.
2337 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002338 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2339 &fe->ptl);
2340 if (!pte_same(*fe->pte, orig_pte)) {
Shachar Raindel93e478d2015-04-14 15:46:35 -07002341 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002342 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002343 put_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002344 return 0;
2345 }
2346 page_mkwrite = 1;
2347 }
2348
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002349 return wp_page_reuse(fe, orig_pte, old_page, page_mkwrite, 1);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002350}
2351
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002352/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353 * This routine handles present pages, when users try to write
2354 * to a shared page. It is done by copying the page to a new address
2355 * and decrementing the shared-page counter for the old page.
2356 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 * Note that this routine assumes that the protection checks have been
2358 * done by the caller (the low-level page fault routine in most cases).
2359 * Thus we can safely just mark it writable once we've done any necessary
2360 * COW.
2361 *
2362 * We also mark the page dirty at this point even though the page will
2363 * change only once the write actually happens. This avoids a few races,
2364 * and potentially makes it more efficient.
2365 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002366 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2367 * but allow concurrent faults), with pte both mapped and locked.
2368 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002370static int do_wp_page(struct fault_env *fe, pte_t orig_pte)
2371 __releases(fe->ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002373 struct vm_area_struct *vma = fe->vma;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002374 struct page *old_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002376 old_page = vm_normal_page(vma, fe->address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002377 if (!old_page) {
2378 /*
Peter Feiner64e45502014-10-13 15:55:46 -07002379 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2380 * VM_PFNMAP VMA.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002381 *
2382 * We should not cow pages in a shared writeable mapping.
Boaz Harroshdd906182015-04-15 16:15:11 -07002383 * Just mark the pages writable and/or call ops->pfn_mkwrite.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002384 */
2385 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2386 (VM_WRITE|VM_SHARED))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002387 return wp_pfn_shared(fe, orig_pte);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002388
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002389 pte_unmap_unlock(fe->pte, fe->ptl);
2390 return wp_page_copy(fe, orig_pte, old_page);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002391 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002393 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002394 * Take out anonymous pages first, anonymous shared vmas are
2395 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002396 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002397 if (PageAnon(old_page) && !PageKsm(old_page)) {
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002398 int total_mapcount;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002399 if (!trylock_page(old_page)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002400 get_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002401 pte_unmap_unlock(fe->pte, fe->ptl);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002402 lock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002403 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd,
2404 fe->address, &fe->ptl);
2405 if (!pte_same(*fe->pte, orig_pte)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002406 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002407 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002408 put_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002409 return 0;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002410 }
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002411 put_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002412 }
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002413 if (reuse_swap_page(old_page, &total_mapcount)) {
2414 if (total_mapcount == 1) {
2415 /*
2416 * The page is all ours. Move it to
2417 * our anon_vma so the rmap code will
2418 * not search our parent or siblings.
2419 * Protected against the rmap code by
2420 * the page lock.
2421 */
Hugh Dickins5a499732016-07-14 12:07:38 -07002422 page_move_anon_rmap(old_page, vma);
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002423 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002424 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002425 return wp_page_reuse(fe, orig_pte, old_page, 0, 0);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002426 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002427 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002428 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002429 (VM_WRITE|VM_SHARED))) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002430 return wp_page_shared(fe, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432
2433 /*
2434 * Ok, we need to copy. Oh, well..
2435 */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002436 get_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002437
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002438 pte_unmap_unlock(fe->pte, fe->ptl);
2439 return wp_page_copy(fe, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440}
2441
Peter Zijlstra97a89412011-05-24 17:12:04 -07002442static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443 unsigned long start_addr, unsigned long end_addr,
2444 struct zap_details *details)
2445{
Al Virof5cc4ee2012-03-05 14:14:20 -05002446 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447}
2448
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002449static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 struct zap_details *details)
2451{
2452 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453 pgoff_t vba, vea, zba, zea;
2454
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002455 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457
2458 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002459 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 zba = details->first_index;
2461 if (zba < vba)
2462 zba = vba;
2463 zea = details->last_index;
2464 if (zea > vea)
2465 zea = vea;
2466
Peter Zijlstra97a89412011-05-24 17:12:04 -07002467 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2469 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002470 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471 }
2472}
2473
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474/**
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08002475 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2476 * address_space corresponding to the specified page range in the underlying
2477 * file.
2478 *
Martin Waitz3d410882005-06-23 22:05:21 -07002479 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 * @holebegin: byte in first page to unmap, relative to the start of
2481 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002482 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 * must keep the partial page. In contrast, we must get rid of
2484 * partial pages.
2485 * @holelen: size of prospective hole in bytes. This will be rounded
2486 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2487 * end of the file.
2488 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2489 * but 0 when invalidating pagecache, don't throw away private data.
2490 */
2491void unmap_mapping_range(struct address_space *mapping,
2492 loff_t const holebegin, loff_t const holelen, int even_cows)
2493{
Michal Hockoaac45362016-03-25 14:20:24 -07002494 struct zap_details details = { };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495 pgoff_t hba = holebegin >> PAGE_SHIFT;
2496 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2497
2498 /* Check for overflow. */
2499 if (sizeof(holelen) > sizeof(hlen)) {
2500 long long holeend =
2501 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2502 if (holeend & ~(long long)ULONG_MAX)
2503 hlen = ULONG_MAX - hba + 1;
2504 }
2505
2506 details.check_mapping = even_cows? NULL: mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 details.first_index = hba;
2508 details.last_index = hba + hlen - 1;
2509 if (details.last_index < details.first_index)
2510 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002512 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002513 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514 unmap_mapping_range_tree(&mapping->i_mmap, &details);
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002515 i_mmap_unlock_write(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516}
2517EXPORT_SYMBOL(unmap_mapping_range);
2518
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002520 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2521 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002522 * We return with pte unmapped and unlocked.
2523 *
2524 * We return with the mmap_sem locked or unlocked in the same cases
2525 * as does filemap_fault().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002527int do_swap_page(struct fault_env *fe, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002529 struct vm_area_struct *vma = fe->vma;
Hugh Dickins56f31802013-02-22 16:36:10 -08002530 struct page *page, *swapcache;
Johannes Weiner00501b52014-08-08 14:19:20 -07002531 struct mem_cgroup *memcg;
Hugh Dickins65500d22005-10-29 18:15:59 -07002532 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002534 int locked;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002535 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002536 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002538 if (!pte_unmap_same(vma->vm_mm, fe->pmd, fe->pte, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002539 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002540
2541 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002542 if (unlikely(non_swap_entry(entry))) {
2543 if (is_migration_entry(entry)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002544 migration_entry_wait(vma->vm_mm, fe->pmd, fe->address);
Andi Kleend1737fd2009-09-16 11:50:06 +02002545 } else if (is_hwpoison_entry(entry)) {
2546 ret = VM_FAULT_HWPOISON;
2547 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002548 print_bad_pte(vma, fe->address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002549 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002550 }
Christoph Lameter06972122006-06-23 02:03:35 -07002551 goto out;
2552 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002553 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 page = lookup_swap_cache(entry);
2555 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002556 page = swapin_readahead(entry,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002557 GFP_HIGHUSER_MOVABLE, vma, fe->address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558 if (!page) {
2559 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002560 * Back out if somebody else faulted in this pte
2561 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002563 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd,
2564 fe->address, &fe->ptl);
2565 if (likely(pte_same(*fe->pte, orig_pte)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002567 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002568 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 }
2570
2571 /* Had to read the page from swap area: Major fault */
2572 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002573 count_vm_event(PGMAJFAULT);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002574 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002575 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002576 /*
2577 * hwpoisoned dirty swapcache pages are kept for killing
2578 * owner processes (which may be unknown at hwpoison time)
2579 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002580 ret = VM_FAULT_HWPOISON;
2581 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08002582 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02002583 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584 }
2585
Hugh Dickins56f31802013-02-22 16:36:10 -08002586 swapcache = page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002587 locked = lock_page_or_retry(page, vma->vm_mm, fe->flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002588
Balbir Singh20a10222007-11-14 17:00:33 -08002589 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002590 if (!locked) {
2591 ret |= VM_FAULT_RETRY;
2592 goto out_release;
2593 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002595 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002596 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2597 * release the swapcache from under us. The page pin, and pte_same
2598 * test below, are not enough to exclude that. Even if it is still
2599 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002600 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002601 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002602 goto out_page;
2603
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002604 page = ksm_might_need_to_copy(page, vma, fe->address);
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002605 if (unlikely(!page)) {
2606 ret = VM_FAULT_OOM;
2607 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002608 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08002609 }
2610
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002611 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2612 &memcg, false)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002613 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002614 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002615 }
2616
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002618 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002620 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2621 &fe->ptl);
2622 if (unlikely(!pte_same(*fe->pte, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002623 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002624
2625 if (unlikely(!PageUptodate(page))) {
2626 ret = VM_FAULT_SIGBUS;
2627 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 }
2629
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002630 /*
2631 * The page isn't present yet, go ahead with the fault.
2632 *
2633 * Be careful about the sequence of operations here.
2634 * To get its accounting right, reuse_swap_page() must be called
2635 * while the page is counted on swap but not yet in mapcount i.e.
2636 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2637 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002638 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002640 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2641 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642 pte = mk_pte(page, vma->vm_page_prot);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002643 if ((fe->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002645 fe->flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002646 ret |= VM_FAULT_WRITE;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002647 exclusive = RMAP_EXCLUSIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07002650 if (pte_swp_soft_dirty(orig_pte))
2651 pte = pte_mksoft_dirty(pte);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002652 set_pte_at(vma->vm_mm, fe->address, fe->pte, pte);
Johannes Weiner00501b52014-08-08 14:19:20 -07002653 if (page == swapcache) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002654 do_page_add_anon_rmap(page, vma, fe->address, exclusive);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002655 mem_cgroup_commit_charge(page, memcg, true, false);
Minchan Kim1a8018f2016-08-02 14:02:25 -07002656 activate_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -07002657 } else { /* ksm created a completely new copy */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002658 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002659 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002660 lru_cache_add_active_or_unevictable(page, vma);
2661 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002663 swap_free(entry);
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08002664 if (mem_cgroup_swap_full(page) ||
2665 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002666 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002667 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002668 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002669 /*
2670 * Hold the lock to avoid the swap entry to be reused
2671 * until we take the PT lock for the pte_same() check
2672 * (to avoid false positives from pte_same). For
2673 * further safety release the lock after the swap_free
2674 * so that the swap count won't change under a
2675 * parallel locked swapcache.
2676 */
2677 unlock_page(swapcache);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002678 put_page(swapcache);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002679 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002680
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002681 if (fe->flags & FAULT_FLAG_WRITE) {
2682 ret |= do_wp_page(fe, pte);
Hugh Dickins61469f12008-03-04 14:29:04 -08002683 if (ret & VM_FAULT_ERROR)
2684 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685 goto out;
2686 }
2687
2688 /* No need to invalidate - it was non-present before */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002689 update_mmu_cache(vma, fe->address, fe->pte);
Hugh Dickins65500d22005-10-29 18:15:59 -07002690unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002691 pte_unmap_unlock(fe->pte, fe->ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692out:
2693 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002694out_nomap:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002695 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002696 pte_unmap_unlock(fe->pte, fe->ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002697out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002698 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002699out_release:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002700 put_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002701 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002702 unlock_page(swapcache);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002703 put_page(swapcache);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002704 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002705 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706}
2707
2708/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002709 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2710 * but allow concurrent faults), and pte mapped but not yet locked.
2711 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002713static int do_anonymous_page(struct fault_env *fe)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002715 struct vm_area_struct *vma = fe->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -07002716 struct mem_cgroup *memcg;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002717 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03002720 /* File mapping without ->vm_ops ? */
2721 if (vma->vm_flags & VM_SHARED)
2722 return VM_FAULT_SIGBUS;
2723
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002724 /*
2725 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2726 * pte_offset_map() on pmds where a huge pmd might be created
2727 * from a different thread.
2728 *
2729 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2730 * parallel threads are excluded by other means.
2731 *
2732 * Here we only have down_read(mmap_sem).
2733 */
2734 if (pte_alloc(vma->vm_mm, fe->pmd, fe->address))
2735 return VM_FAULT_OOM;
2736
2737 /* See the comment in pte_alloc_one_map() */
2738 if (unlikely(pmd_trans_unstable(fe->pmd)))
2739 return 0;
2740
Linus Torvalds11ac5522010-08-14 11:44:56 -07002741 /* Use the zero-page for reads */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002742 if (!(fe->flags & FAULT_FLAG_WRITE) &&
2743 !mm_forbids_zeropage(vma->vm_mm)) {
2744 entry = pte_mkspecial(pfn_pte(my_zero_pfn(fe->address),
Hugh Dickins62eede62009-09-21 17:03:34 -07002745 vma->vm_page_prot));
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002746 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2747 &fe->ptl);
2748 if (!pte_none(*fe->pte))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002749 goto unlock;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002750 /* Deliver the page fault to userland, check inside PT lock */
2751 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002752 pte_unmap_unlock(fe->pte, fe->ptl);
2753 return handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002754 }
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002755 goto setpte;
2756 }
2757
Nick Piggin557ed1f2007-10-16 01:24:40 -07002758 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002759 if (unlikely(anon_vma_prepare(vma)))
2760 goto oom;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002761 page = alloc_zeroed_user_highpage_movable(vma, fe->address);
Nick Piggin557ed1f2007-10-16 01:24:40 -07002762 if (!page)
2763 goto oom;
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002764
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002765 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002766 goto oom_free_page;
2767
Minchan Kim52f37622013-04-29 15:08:15 -07002768 /*
2769 * The memory barrier inside __SetPageUptodate makes sure that
2770 * preceeding stores to the page contents become visible before
2771 * the set_pte_at() write.
2772 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08002773 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774
Nick Piggin557ed1f2007-10-16 01:24:40 -07002775 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002776 if (vma->vm_flags & VM_WRITE)
2777 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002778
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002779 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2780 &fe->ptl);
2781 if (!pte_none(*fe->pte))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002782 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002783
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002784 /* Deliver the page fault to userland, check inside PT lock */
2785 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002786 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002787 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002788 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002789 return handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002790 }
2791
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002792 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2793 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002794 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002795 lru_cache_add_active_or_unevictable(page, vma);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002796setpte:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002797 set_pte_at(vma->vm_mm, fe->address, fe->pte, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798
2799 /* No need to invalidate - it was non-present before */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002800 update_mmu_cache(vma, fe->address, fe->pte);
Hugh Dickins65500d22005-10-29 18:15:59 -07002801unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002802 pte_unmap_unlock(fe->pte, fe->ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002803 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002804release:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002805 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002806 put_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002807 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002808oom_free_page:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002809 put_page(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002810oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 return VM_FAULT_OOM;
2812}
2813
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002814/*
2815 * The mmap_sem must have been held on entry, and may have been
2816 * released depending on flags and vma->vm_ops->fault() return value.
2817 * See filemap_fault() and __lock_page_retry().
2818 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002819static int __do_fault(struct fault_env *fe, pgoff_t pgoff,
2820 struct page *cow_page, struct page **page, void **entry)
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002821{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002822 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002823 struct vm_fault vmf;
2824 int ret;
2825
Michal Hocko5cf3e5f2019-01-08 15:23:07 -08002826 /*
2827 * Preallocate pte before we take page_lock because this might lead to
2828 * deadlocks for memcg reclaim which waits for pages under writeback:
2829 * lock_page(A)
2830 * SetPageWriteback(A)
2831 * unlock_page(A)
2832 * lock_page(B)
2833 * lock_page(B)
2834 * pte_alloc_pne
2835 * shrink_page_list
2836 * wait_on_page_writeback(A)
2837 * SetPageWriteback(B)
2838 * unlock_page(B)
2839 * # flush A, B to clear the writeback
2840 */
2841 if (pmd_none(*fe->pmd) && !fe->prealloc_pte) {
2842 fe->prealloc_pte = pte_alloc_one(vma->vm_mm, fe->address);
2843 if (!fe->prealloc_pte)
2844 return VM_FAULT_OOM;
2845 smp_wmb(); /* See comment in __pte_alloc() */
2846 }
2847
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002848 vmf.virtual_address = (void __user *)(fe->address & PAGE_MASK);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002849 vmf.pgoff = pgoff;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002850 vmf.flags = fe->flags;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002851 vmf.page = NULL;
Michal Hockoc20cd452016-01-14 15:20:12 -08002852 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002853 vmf.cow_page = cow_page;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002854
2855 ret = vma->vm_ops->fault(vma, &vmf);
2856 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2857 return ret;
Jan Karabc2466e2016-05-12 18:29:19 +02002858 if (ret & VM_FAULT_DAX_LOCKED) {
2859 *entry = vmf.entry;
2860 return ret;
2861 }
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002862
2863 if (unlikely(PageHWPoison(vmf.page))) {
2864 if (ret & VM_FAULT_LOCKED)
2865 unlock_page(vmf.page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002866 put_page(vmf.page);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002867 return VM_FAULT_HWPOISON;
2868 }
2869
2870 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2871 lock_page(vmf.page);
2872 else
2873 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2874
2875 *page = vmf.page;
2876 return ret;
2877}
2878
Ross Zwislerf2562ed2018-02-23 14:05:27 -08002879/*
2880 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
2881 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
2882 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
2883 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
2884 */
2885static int pmd_devmap_trans_unstable(pmd_t *pmd)
2886{
2887 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
2888}
2889
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002890static int pte_alloc_one_map(struct fault_env *fe)
2891{
2892 struct vm_area_struct *vma = fe->vma;
2893
2894 if (!pmd_none(*fe->pmd))
2895 goto map_pte;
2896 if (fe->prealloc_pte) {
2897 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
2898 if (unlikely(!pmd_none(*fe->pmd))) {
2899 spin_unlock(fe->ptl);
2900 goto map_pte;
2901 }
2902
2903 atomic_long_inc(&vma->vm_mm->nr_ptes);
2904 pmd_populate(vma->vm_mm, fe->pmd, fe->prealloc_pte);
2905 spin_unlock(fe->ptl);
2906 fe->prealloc_pte = 0;
2907 } else if (unlikely(pte_alloc(vma->vm_mm, fe->pmd, fe->address))) {
2908 return VM_FAULT_OOM;
2909 }
2910map_pte:
2911 /*
2912 * If a huge pmd materialized under us just retry later. Use
Ross Zwislerf2562ed2018-02-23 14:05:27 -08002913 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
2914 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
2915 * under us and then back to pmd_none, as a result of MADV_DONTNEED
2916 * running immediately after a huge pmd fault in a different thread of
2917 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
2918 * All we have to ensure is that it is a regular pmd that we can walk
2919 * with pte_offset_map() and we can do that through an atomic read in
2920 * C, which is what pmd_trans_unstable() provides.
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002921 */
Ross Zwislerf2562ed2018-02-23 14:05:27 -08002922 if (pmd_devmap_trans_unstable(fe->pmd))
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002923 return VM_FAULT_NOPAGE;
2924
Ross Zwislerf2562ed2018-02-23 14:05:27 -08002925 /*
2926 * At this point we know that our vmf->pmd points to a page of ptes
2927 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
2928 * for the duration of the fault. If a racing MADV_DONTNEED runs and
2929 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
2930 * be valid and we will re-check to make sure the vmf->pte isn't
2931 * pte_none() under vmf->ptl protection when we return to
2932 * alloc_set_pte().
2933 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002934 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2935 &fe->ptl);
2936 return 0;
2937}
2938
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07002939#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002940
2941#define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
2942static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
2943 unsigned long haddr)
2944{
2945 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
2946 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
2947 return false;
2948 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
2949 return false;
2950 return true;
2951}
2952
2953static int do_set_pmd(struct fault_env *fe, struct page *page)
2954{
2955 struct vm_area_struct *vma = fe->vma;
2956 bool write = fe->flags & FAULT_FLAG_WRITE;
2957 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
2958 pmd_t entry;
2959 int i, ret;
2960
2961 if (!transhuge_vma_suitable(vma, haddr))
2962 return VM_FAULT_FALLBACK;
2963
2964 ret = VM_FAULT_FALLBACK;
2965 page = compound_head(page);
2966
2967 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
2968 if (unlikely(!pmd_none(*fe->pmd)))
2969 goto out;
2970
2971 for (i = 0; i < HPAGE_PMD_NR; i++)
2972 flush_icache_page(vma, page + i);
2973
2974 entry = mk_huge_pmd(page, vma->vm_page_prot);
2975 if (write)
2976 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2977
2978 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
2979 page_add_file_rmap(page, true);
2980
2981 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
2982
2983 update_mmu_cache_pmd(vma, haddr, fe->pmd);
2984
2985 /* fault is handled */
2986 ret = 0;
Kirill A. Shutemov95ecedc2016-07-26 15:25:31 -07002987 count_vm_event(THP_FILE_MAPPED);
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002988out:
2989 spin_unlock(fe->ptl);
2990 return ret;
2991}
2992#else
2993static int do_set_pmd(struct fault_env *fe, struct page *page)
2994{
2995 BUILD_BUG();
2996 return 0;
2997}
2998#endif
2999
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003000/**
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003001 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3002 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003003 *
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003004 * @fe: fault environment
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003005 * @memcg: memcg to charge page (only for private mappings)
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003006 * @page: page to map
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003007 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003008 * Caller must take care of unlocking fe->ptl, if fe->pte is non-NULL on return.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003009 *
3010 * Target users are page handler itself and implementations of
3011 * vm_ops->map_pages.
3012 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003013int alloc_set_pte(struct fault_env *fe, struct mem_cgroup *memcg,
3014 struct page *page)
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003015{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003016 struct vm_area_struct *vma = fe->vma;
3017 bool write = fe->flags & FAULT_FLAG_WRITE;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003018 pte_t entry;
Kirill A. Shutemov10102452016-07-26 15:25:29 -07003019 int ret;
3020
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07003021 if (pmd_none(*fe->pmd) && PageTransCompound(page) &&
3022 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
Kirill A. Shutemov10102452016-07-26 15:25:29 -07003023 /* THP on COW? */
3024 VM_BUG_ON_PAGE(memcg, page);
3025
3026 ret = do_set_pmd(fe, page);
3027 if (ret != VM_FAULT_FALLBACK)
3028 return ret;
3029 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003030
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003031 if (!fe->pte) {
Kirill A. Shutemov10102452016-07-26 15:25:29 -07003032 ret = pte_alloc_one_map(fe);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003033 if (ret)
3034 return ret;
3035 }
3036
3037 /* Re-check under ptl */
3038 if (unlikely(!pte_none(*fe->pte)))
3039 return VM_FAULT_NOPAGE;
3040
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003041 flush_icache_page(vma, page);
3042 entry = mk_pte(page, vma->vm_page_prot);
3043 if (write)
3044 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003045 /* copy-on-write page */
3046 if (write && !(vma->vm_flags & VM_SHARED)) {
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003047 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003048 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003049 mem_cgroup_commit_charge(page, memcg, false, false);
3050 lru_cache_add_active_or_unevictable(page, vma);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003051 } else {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08003052 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07003053 page_add_file_rmap(page, false);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003054 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003055 set_pte_at(vma->vm_mm, fe->address, fe->pte, entry);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003056
3057 /* no need to invalidate: a not-present page won't be cached */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003058 update_mmu_cache(vma, fe->address, fe->pte);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003059
3060 return 0;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003061}
3062
Kirill A. Shutemov3a910532014-08-06 16:08:07 -07003063static unsigned long fault_around_bytes __read_mostly =
3064 rounddown_pow_of_two(65536);
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003065
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003066#ifdef CONFIG_DEBUG_FS
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003067static int fault_around_bytes_get(void *data, u64 *val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003068{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003069 *val = fault_around_bytes;
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003070 return 0;
3071}
3072
Andrey Ryabininb4903d62014-07-30 16:08:35 -07003073/*
3074 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3075 * rounded down to nearest page order. It's what do_fault_around() expects to
3076 * see.
3077 */
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003078static int fault_around_bytes_set(void *data, u64 val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003079{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003080 if (val / PAGE_SIZE > PTRS_PER_PTE)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003081 return -EINVAL;
Andrey Ryabininb4903d62014-07-30 16:08:35 -07003082 if (val > PAGE_SIZE)
3083 fault_around_bytes = rounddown_pow_of_two(val);
3084 else
3085 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003086 return 0;
3087}
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003088DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3089 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003090
3091static int __init fault_around_debugfs(void)
3092{
3093 void *ret;
3094
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003095 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3096 &fault_around_bytes_fops);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003097 if (!ret)
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003098 pr_warn("Failed to create fault_around_bytes in debugfs");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003099 return 0;
3100}
3101late_initcall(fault_around_debugfs);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003102#endif
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003103
Kirill A. Shutemov1fdb4122014-06-04 16:10:55 -07003104/*
3105 * do_fault_around() tries to map few pages around the fault address. The hope
3106 * is that the pages will be needed soon and this will lower the number of
3107 * faults to handle.
3108 *
3109 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3110 * not ready to be mapped: not up-to-date, locked, etc.
3111 *
3112 * This function is called with the page table lock taken. In the split ptlock
3113 * case the page table lock only protects only those entries which belong to
3114 * the page table corresponding to the fault address.
3115 *
3116 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3117 * only once.
3118 *
3119 * fault_around_pages() defines how many pages we'll try to map.
3120 * do_fault_around() expects it to return a power of two less than or equal to
3121 * PTRS_PER_PTE.
3122 *
3123 * The virtual address of the area that we map is naturally aligned to the
3124 * fault_around_pages() value (and therefore to page order). This way it's
3125 * easier to guarantee that we don't cross page table boundaries.
3126 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003127static int do_fault_around(struct fault_env *fe, pgoff_t start_pgoff)
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003128{
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003129 unsigned long address = fe->address, nr_pages, mask;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003130 pgoff_t end_pgoff;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003131 int off, ret = 0;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003132
Jason Low4db0c3c2015-04-15 16:14:08 -07003133 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07003134 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3135
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003136 fe->address = max(address & mask, fe->vma->vm_start);
3137 off = ((address - fe->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003138 start_pgoff -= off;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003139
3140 /*
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003141 * end_pgoff is either end of page table or end of vma
3142 * or fault_around_pages() from start_pgoff, depending what is nearest.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003143 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003144 end_pgoff = start_pgoff -
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003145 ((fe->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003146 PTRS_PER_PTE - 1;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003147 end_pgoff = min3(end_pgoff, vma_pages(fe->vma) + fe->vma->vm_pgoff - 1,
3148 start_pgoff + nr_pages - 1);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003149
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003150 if (pmd_none(*fe->pmd)) {
3151 fe->prealloc_pte = pte_alloc_one(fe->vma->vm_mm, fe->address);
Vegard Nossumc5f88bd2016-08-02 14:02:22 -07003152 if (!fe->prealloc_pte)
3153 goto out;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003154 smp_wmb(); /* See comment in __pte_alloc() */
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003155 }
3156
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003157 fe->vma->vm_ops->map_pages(fe, start_pgoff, end_pgoff);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003158
3159 /* preallocated pagetable is unused: free it */
3160 if (fe->prealloc_pte) {
3161 pte_free(fe->vma->vm_mm, fe->prealloc_pte);
3162 fe->prealloc_pte = 0;
3163 }
3164 /* Huge page is mapped? Page fault is solved */
3165 if (pmd_trans_huge(*fe->pmd)) {
3166 ret = VM_FAULT_NOPAGE;
3167 goto out;
3168 }
3169
3170 /* ->map_pages() haven't done anything useful. Cold page cache? */
3171 if (!fe->pte)
3172 goto out;
3173
3174 /* check if the page fault is solved */
3175 fe->pte -= (fe->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3176 if (!pte_none(*fe->pte))
3177 ret = VM_FAULT_NOPAGE;
3178 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003179out:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003180 fe->address = address;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003181 fe->pte = NULL;
3182 return ret;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003183}
3184
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003185static int do_read_fault(struct fault_env *fe, pgoff_t pgoff)
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003186{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003187 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003188 struct page *fault_page;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003189 int ret = 0;
3190
3191 /*
3192 * Let's call ->map_pages() first and use ->fault() as fallback
3193 * if page by the offset is not ready to be mapped (cold cache or
3194 * something).
3195 */
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08003196 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003197 ret = do_fault_around(fe, pgoff);
3198 if (ret)
3199 return ret;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003200 }
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003201
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003202 ret = __do_fault(fe, pgoff, NULL, &fault_page, NULL);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003203 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3204 return ret;
3205
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003206 ret |= alloc_set_pte(fe, NULL, fault_page);
3207 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003208 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003209 unlock_page(fault_page);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003210 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3211 put_page(fault_page);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003212 return ret;
3213}
3214
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003215static int do_cow_fault(struct fault_env *fe, pgoff_t pgoff)
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003216{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003217 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003218 struct page *fault_page, *new_page;
Jan Karabc2466e2016-05-12 18:29:19 +02003219 void *fault_entry;
Johannes Weiner00501b52014-08-08 14:19:20 -07003220 struct mem_cgroup *memcg;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003221 int ret;
3222
3223 if (unlikely(anon_vma_prepare(vma)))
3224 return VM_FAULT_OOM;
3225
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003226 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, fe->address);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003227 if (!new_page)
3228 return VM_FAULT_OOM;
3229
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003230 if (mem_cgroup_try_charge(new_page, vma->vm_mm, GFP_KERNEL,
3231 &memcg, false)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003232 put_page(new_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003233 return VM_FAULT_OOM;
3234 }
3235
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003236 ret = __do_fault(fe, pgoff, new_page, &fault_page, &fault_entry);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003237 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3238 goto uncharge_out;
3239
Jan Karabc2466e2016-05-12 18:29:19 +02003240 if (!(ret & VM_FAULT_DAX_LOCKED))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003241 copy_user_highpage(new_page, fault_page, fe->address, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003242 __SetPageUptodate(new_page);
3243
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003244 ret |= alloc_set_pte(fe, memcg, new_page);
3245 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003246 pte_unmap_unlock(fe->pte, fe->ptl);
Jan Karabc2466e2016-05-12 18:29:19 +02003247 if (!(ret & VM_FAULT_DAX_LOCKED)) {
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003248 unlock_page(fault_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003249 put_page(fault_page);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003250 } else {
Jan Karabc2466e2016-05-12 18:29:19 +02003251 dax_unlock_mapping_entry(vma->vm_file->f_mapping, pgoff);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003252 }
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003253 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3254 goto uncharge_out;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003255 return ret;
3256uncharge_out:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08003257 mem_cgroup_cancel_charge(new_page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003258 put_page(new_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003259 return ret;
3260}
3261
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003262static int do_shared_fault(struct fault_env *fe, pgoff_t pgoff)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003264 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003265 struct page *fault_page;
3266 struct address_space *mapping;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003267 int dirtied = 0;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003268 int ret, tmp;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003269
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003270 ret = __do_fault(fe, pgoff, NULL, &fault_page, NULL);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07003271 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003272 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273
3274 /*
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003275 * Check if the backing address space wants to know that the page is
3276 * about to become writable
Linus Torvalds1da177e2005-04-16 15:20:36 -07003277 */
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003278 if (vma->vm_ops->page_mkwrite) {
3279 unlock_page(fault_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003280 tmp = do_page_mkwrite(vma, fault_page, fe->address);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003281 if (unlikely(!tmp ||
3282 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003283 put_page(fault_page);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003284 return tmp;
3285 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286 }
3287
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003288 ret |= alloc_set_pte(fe, NULL, fault_page);
3289 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003290 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003291 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3292 VM_FAULT_RETRY))) {
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003293 unlock_page(fault_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003294 put_page(fault_page);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003295 return ret;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003296 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003297
3298 if (set_page_dirty(fault_page))
3299 dirtied = 1;
Andrew Mortond82fa872014-12-18 16:17:29 -08003300 /*
3301 * Take a local copy of the address_space - page.mapping may be zeroed
3302 * by truncate after unlock_page(). The address_space itself remains
3303 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3304 * release semantics to prevent the compiler from undoing this copying.
3305 */
Kirill A. Shutemov1c290f62016-01-15 16:52:07 -08003306 mapping = page_rmapping(fault_page);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003307 unlock_page(fault_page);
3308 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3309 /*
3310 * Some device drivers do not set page.mapping but still
3311 * dirty their pages
3312 */
3313 balance_dirty_pages_ratelimited(mapping);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003314 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003315
Johannes Weiner74ec6752015-02-10 14:11:28 -08003316 if (!vma->vm_ops->page_mkwrite)
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003317 file_update_time(vma->vm_file);
3318
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003319 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003320}
Nick Piggind00806b2007-07-19 01:46:57 -07003321
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003322/*
3323 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3324 * but allow concurrent faults).
3325 * The mmap_sem may have been released depending on flags and our
3326 * return value. See filemap_fault() and __lock_page_or_retry().
3327 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003328static int do_fault(struct fault_env *fe)
Nick Piggin54cb8822007-07-19 01:46:59 -07003329{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003330 struct vm_area_struct *vma = fe->vma;
3331 pgoff_t pgoff = linear_page_index(vma, fe->address);
Hugh Dickinseede34e2017-01-07 15:37:31 -08003332 int ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003333
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03003334 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3335 if (!vma->vm_ops->fault)
Hugh Dickinseede34e2017-01-07 15:37:31 -08003336 ret = VM_FAULT_SIGBUS;
3337 else if (!(fe->flags & FAULT_FLAG_WRITE))
3338 ret = do_read_fault(fe, pgoff);
3339 else if (!(vma->vm_flags & VM_SHARED))
3340 ret = do_cow_fault(fe, pgoff);
3341 else
3342 ret = do_shared_fault(fe, pgoff);
3343
3344 /* preallocated pagetable is unused: free it */
3345 if (fe->prealloc_pte) {
3346 pte_free(vma->vm_mm, fe->prealloc_pte);
3347 fe->prealloc_pte = 0;
3348 }
3349 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003350}
3351
Rashika Kheriab19a9932014-04-03 14:48:02 -07003352static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
Rik van Riel04bb2f92013-10-07 11:29:36 +01003353 unsigned long addr, int page_nid,
3354 int *flags)
Mel Gorman9532fec2012-11-15 01:24:32 +00003355{
3356 get_page(page);
3357
3358 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003359 if (page_nid == numa_node_id()) {
Mel Gorman9532fec2012-11-15 01:24:32 +00003360 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003361 *flags |= TNF_FAULT_LOCAL;
3362 }
Mel Gorman9532fec2012-11-15 01:24:32 +00003363
3364 return mpol_misplaced(page, vma, addr);
3365}
3366
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003367static int do_numa_page(struct fault_env *fe, pte_t pte)
Mel Gormand10e63f2012-10-25 14:16:31 +02003368{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003369 struct vm_area_struct *vma = fe->vma;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003370 struct page *page = NULL;
Mel Gorman8191acb2013-10-07 11:28:45 +01003371 int page_nid = -1;
Peter Zijlstra90572892013-10-07 11:29:20 +01003372 int last_cpupid;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003373 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003374 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07003375 bool was_writable = pte_write(pte);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003376 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003377
3378 /*
3379 * The "pte" at this point cannot be used safely without
3380 * validation through pte_unmap_same(). It's of NUMA type but
3381 * the pfn may be screwed if the read is non atomic.
3382 *
Mel Gorman4d942462015-02-12 14:58:28 -08003383 * We can safely just do a "set_pte_at()", because the old
3384 * page table entry is not accessible, so there would be no
3385 * concurrent hardware modifications to the PTE.
Mel Gormand10e63f2012-10-25 14:16:31 +02003386 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003387 fe->ptl = pte_lockptr(vma->vm_mm, fe->pmd);
3388 spin_lock(fe->ptl);
3389 if (unlikely(!pte_same(*fe->pte, pte))) {
3390 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003391 goto out;
3392 }
3393
Mel Gorman4d942462015-02-12 14:58:28 -08003394 /* Make it present again */
3395 pte = pte_modify(pte, vma->vm_page_prot);
3396 pte = pte_mkyoung(pte);
Mel Gormanb191f9b2015-03-25 15:55:40 -07003397 if (was_writable)
3398 pte = pte_mkwrite(pte);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003399 set_pte_at(vma->vm_mm, fe->address, fe->pte, pte);
3400 update_mmu_cache(vma, fe->address, fe->pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003401
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003402 page = vm_normal_page(vma, fe->address, pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003403 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003404 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gormand10e63f2012-10-25 14:16:31 +02003405 return 0;
3406 }
3407
Kirill A. Shutemove81c4802016-01-15 16:53:49 -08003408 /* TODO: handle PTE-mapped THP */
3409 if (PageCompound(page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003410 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemove81c4802016-01-15 16:53:49 -08003411 return 0;
3412 }
3413
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003414 /*
Mel Gormanbea66fb2015-03-25 15:55:37 -07003415 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3416 * much anyway since they can be in shared cache state. This misses
3417 * the case where a mapping is writable but the process never writes
3418 * to it but pte_write gets cleared during protection updates and
3419 * pte_dirty has unpredictable behaviour between PTE scan updates,
3420 * background writeback, dirty balancing and application behaviour.
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003421 */
Rik van Rield59dc7b2016-09-08 21:30:53 -04003422 if (!pte_write(pte))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003423 flags |= TNF_NO_GROUP;
3424
Rik van Rieldabe1d92013-10-07 11:29:34 +01003425 /*
3426 * Flag if the page is shared between multiple address spaces. This
3427 * is later used when determining whether to group tasks together
3428 */
3429 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3430 flags |= TNF_SHARED;
3431
Peter Zijlstra90572892013-10-07 11:29:20 +01003432 last_cpupid = page_cpupid_last(page);
Mel Gorman8191acb2013-10-07 11:28:45 +01003433 page_nid = page_to_nid(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003434 target_nid = numa_migrate_prep(page, vma, fe->address, page_nid,
3435 &flags);
3436 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003437 if (target_nid == -1) {
Mel Gorman4daae3b2012-11-02 11:33:45 +00003438 put_page(page);
3439 goto out;
3440 }
3441
3442 /* Migrate to the requested node */
Mel Gorman1bc115d2013-10-07 11:29:05 +01003443 migrated = migrate_misplaced_page(page, vma, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003444 if (migrated) {
Mel Gorman8191acb2013-10-07 11:28:45 +01003445 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003446 flags |= TNF_MIGRATED;
Mel Gorman074c2382015-03-25 15:55:42 -07003447 } else
3448 flags |= TNF_MIGRATE_FAIL;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003449
3450out:
Mel Gorman8191acb2013-10-07 11:28:45 +01003451 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003452 task_numa_fault(last_cpupid, page_nid, 1, flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003453 return 0;
3454}
3455
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003456static int create_huge_pmd(struct fault_env *fe)
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003457{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003458 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovfb6dd5f2015-09-09 15:39:35 -07003459 if (vma_is_anonymous(vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003460 return do_huge_pmd_anonymous_page(fe);
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003461 if (vma->vm_ops->pmd_fault)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003462 return vma->vm_ops->pmd_fault(vma, fe->address, fe->pmd,
3463 fe->flags);
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003464 return VM_FAULT_FALLBACK;
3465}
3466
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003467static int wp_huge_pmd(struct fault_env *fe, pmd_t orig_pmd)
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003468{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003469 if (vma_is_anonymous(fe->vma))
3470 return do_huge_pmd_wp_page(fe, orig_pmd);
3471 if (fe->vma->vm_ops->pmd_fault)
3472 return fe->vma->vm_ops->pmd_fault(fe->vma, fe->address, fe->pmd,
3473 fe->flags);
Kirill A. Shutemovaf9e4d52016-07-26 15:25:40 -07003474
3475 /* COW handled on pte level: split pmd */
3476 VM_BUG_ON_VMA(fe->vma->vm_flags & VM_SHARED, fe->vma);
3477 split_huge_pmd(fe->vma, fe->pmd, fe->address);
3478
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003479 return VM_FAULT_FALLBACK;
3480}
3481
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003482static inline bool vma_is_accessible(struct vm_area_struct *vma)
3483{
3484 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3485}
3486
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487/*
3488 * These routines also need to handle stuff like marking pages dirty
3489 * and/or accessed for architectures that don't do it in hardware (most
3490 * RISC architectures). The early dirtying is also good on the i386.
3491 *
3492 * There is also a hook called "update_mmu_cache()" that architectures
3493 * with external mmu caches can use to update those (ie the Sparc or
3494 * PowerPC hashed page tables that act as extended TLBs).
3495 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003496 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3497 * concurrent faults).
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003498 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003499 * The mmap_sem may have been released depending on flags and our return value.
3500 * See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003502static int handle_pte_fault(struct fault_env *fe)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503{
3504 pte_t entry;
3505
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003506 if (unlikely(pmd_none(*fe->pmd))) {
3507 /*
3508 * Leave __pte_alloc() until later: because vm_ops->fault may
3509 * want to allocate huge page, and if we expose page table
3510 * for an instant, it will be difficult to retract from
3511 * concurrent faults and from rmap lookups.
3512 */
3513 fe->pte = NULL;
3514 } else {
3515 /* See comment in pte_alloc_one_map() */
Ross Zwislerf2562ed2018-02-23 14:05:27 -08003516 if (pmd_devmap_trans_unstable(fe->pmd))
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003517 return 0;
3518 /*
3519 * A regular pmd is established and it can't morph into a huge
3520 * pmd from under us anymore at this point because we hold the
3521 * mmap_sem read mode and khugepaged takes it in write mode.
3522 * So now it's safe to run pte_offset_map().
3523 */
3524 fe->pte = pte_offset_map(fe->pmd, fe->address);
3525
3526 entry = *fe->pte;
3527
3528 /*
3529 * some architectures can have larger ptes than wordsize,
3530 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3531 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3532 * atomic accesses. The code below just needs a consistent
3533 * view for the ifs and we later double check anyway with the
3534 * ptl lock held. So here a barrier will do.
3535 */
3536 barrier();
Hugh Dickins65500d22005-10-29 18:15:59 -07003537 if (pte_none(entry)) {
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003538 pte_unmap(fe->pte);
3539 fe->pte = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07003540 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541 }
3542
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003543 if (!fe->pte) {
3544 if (vma_is_anonymous(fe->vma))
3545 return do_anonymous_page(fe);
3546 else
3547 return do_fault(fe);
3548 }
3549
3550 if (!pte_present(entry))
3551 return do_swap_page(fe, entry);
3552
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003553 if (pte_protnone(entry) && vma_is_accessible(fe->vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003554 return do_numa_page(fe, entry);
Mel Gormand10e63f2012-10-25 14:16:31 +02003555
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003556 fe->ptl = pte_lockptr(fe->vma->vm_mm, fe->pmd);
3557 spin_lock(fe->ptl);
3558 if (unlikely(!pte_same(*fe->pte, entry)))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003559 goto unlock;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003560 if (fe->flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003561 if (!pte_write(entry))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003562 return do_wp_page(fe, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563 entry = pte_mkdirty(entry);
3564 }
3565 entry = pte_mkyoung(entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003566 if (ptep_set_access_flags(fe->vma, fe->address, fe->pte, entry,
3567 fe->flags & FAULT_FLAG_WRITE)) {
3568 update_mmu_cache(fe->vma, fe->address, fe->pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003569 } else {
3570 /*
3571 * This is needed only for protection faults but the arch code
3572 * is not yet telling us if this is a protection fault or not.
3573 * This still avoids useless tlb flushes for .text page faults
3574 * with threads.
3575 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003576 if (fe->flags & FAULT_FLAG_WRITE)
3577 flush_tlb_fix_spurious_fault(fe->vma, fe->address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003578 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003579unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003580 pte_unmap_unlock(fe->pte, fe->ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003581 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582}
3583
3584/*
3585 * By the time we get here, we already hold the mm semaphore
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003586 *
3587 * The mmap_sem may have been released depending on flags and our
3588 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 */
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003590static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3591 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003593 struct fault_env fe = {
3594 .vma = vma,
3595 .address = address,
3596 .flags = flags,
3597 };
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003598 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599 pgd_t *pgd;
3600 pud_t *pud;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 pud = pud_alloc(mm, pgd, address);
3604 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003605 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003606 fe.pmd = pmd_alloc(mm, pud, address);
3607 if (!fe.pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003608 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003609 if (pmd_none(*fe.pmd) && transparent_hugepage_enabled(vma)) {
3610 int ret = create_huge_pmd(&fe);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003611 if (!(ret & VM_FAULT_FALLBACK))
3612 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003613 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003614 pmd_t orig_pmd = *fe.pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003615 int ret;
3616
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003617 barrier();
Dan Williams5c7fb562016-01-15 16:56:52 -08003618 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003619 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003620 return do_huge_pmd_numa_page(&fe, orig_pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02003621
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003622 if ((fe.flags & FAULT_FLAG_WRITE) &&
3623 !pmd_write(orig_pmd)) {
3624 ret = wp_huge_pmd(&fe, orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003625 if (!(ret & VM_FAULT_FALLBACK))
3626 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003627 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003628 huge_pmd_set_accessed(&fe, orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003629 return 0;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003630 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003631 }
3632 }
3633
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003634 return handle_pte_fault(&fe);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635}
3636
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003637/*
3638 * By the time we get here, we already hold the mm semaphore
3639 *
3640 * The mmap_sem may have been released depending on flags and our
3641 * return value. See filemap_fault() and __lock_page_or_retry().
3642 */
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003643int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3644 unsigned int flags)
Johannes Weiner519e5242013-09-12 15:13:42 -07003645{
3646 int ret;
3647
3648 __set_current_state(TASK_RUNNING);
3649
3650 count_vm_event(PGFAULT);
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003651 mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
Johannes Weiner519e5242013-09-12 15:13:42 -07003652
3653 /* do counter updates before entering really critical section. */
3654 check_sync_rss_stat(current);
3655
Laurent Dufour3c8381d2017-09-08 16:13:12 -07003656 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3657 flags & FAULT_FLAG_INSTRUCTION,
3658 flags & FAULT_FLAG_REMOTE))
3659 return VM_FAULT_SIGSEGV;
3660
Johannes Weiner519e5242013-09-12 15:13:42 -07003661 /*
3662 * Enable the memcg OOM handling for faults triggered in user
3663 * space. Kernel faults are handled more gracefully.
3664 */
3665 if (flags & FAULT_FLAG_USER)
Johannes Weiner49426422013-10-16 13:46:59 -07003666 mem_cgroup_oom_enable();
Johannes Weiner519e5242013-09-12 15:13:42 -07003667
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003668 if (unlikely(is_vm_hugetlb_page(vma)))
3669 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3670 else
3671 ret = __handle_mm_fault(vma, address, flags);
Johannes Weiner519e5242013-09-12 15:13:42 -07003672
Johannes Weiner49426422013-10-16 13:46:59 -07003673 if (flags & FAULT_FLAG_USER) {
3674 mem_cgroup_oom_disable();
3675 /*
3676 * The task may have entered a memcg OOM situation but
3677 * if the allocation error was handled gracefully (no
3678 * VM_FAULT_OOM), there is no need to kill anything.
3679 * Just clean up the OOM state peacefully.
3680 */
3681 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3682 mem_cgroup_oom_synchronize(false);
3683 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003684
Michal Hocko3f70dc32016-10-07 16:59:06 -07003685 /*
3686 * This mm has been already reaped by the oom reaper and so the
3687 * refault cannot be trusted in general. Anonymous refaults would
3688 * lose data and give a zero page instead e.g. This is especially
3689 * problem for use_mm() because regular tasks will just die and
3690 * the corrupted data will not be visible anywhere while kthread
3691 * will outlive the oom victim and potentially propagate the date
3692 * further.
3693 */
3694 if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
Michal Hocko59ee25d2017-08-18 15:16:12 -07003695 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags))) {
3696
3697 /*
3698 * We are going to enforce SIGBUS but the PF path might have
3699 * dropped the mmap_sem already so take it again so that
3700 * we do not break expectations of all arch specific PF paths
3701 * and g-u-p
3702 */
3703 if (ret & VM_FAULT_RETRY)
3704 down_read(&vma->vm_mm->mmap_sem);
Michal Hocko3f70dc32016-10-07 16:59:06 -07003705 ret = VM_FAULT_SIGBUS;
Michal Hocko59ee25d2017-08-18 15:16:12 -07003706 }
Michal Hocko3f70dc32016-10-07 16:59:06 -07003707
Johannes Weiner519e5242013-09-12 15:13:42 -07003708 return ret;
3709}
Jesse Barnese1d6d012014-12-12 16:55:27 -08003710EXPORT_SYMBOL_GPL(handle_mm_fault);
Johannes Weiner519e5242013-09-12 15:13:42 -07003711
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712#ifndef __PAGETABLE_PUD_FOLDED
3713/*
3714 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003715 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003717int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003719 pud_t *new = pud_alloc_one(mm, address);
3720 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003721 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003722
Nick Piggin362a61a2008-05-14 06:37:36 +02003723 smp_wmb(); /* See comment in __pte_alloc */
3724
Hugh Dickins872fec12005-10-29 18:16:21 -07003725 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003726 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003727 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003728 else
3729 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003730 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003731 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732}
3733#endif /* __PAGETABLE_PUD_FOLDED */
3734
3735#ifndef __PAGETABLE_PMD_FOLDED
3736/*
3737 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003738 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003740int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003741{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003742 pmd_t *new = pmd_alloc_one(mm, address);
3743 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003744 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745
Nick Piggin362a61a2008-05-14 06:37:36 +02003746 smp_wmb(); /* See comment in __pte_alloc */
3747
Hugh Dickins872fec12005-10-29 18:16:21 -07003748 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003749#ifndef __ARCH_HAS_4LEVEL_HACK
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003750 if (!pud_present(*pud)) {
3751 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003752 pud_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003753 } else /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003754 pmd_free(mm, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003755#else
3756 if (!pgd_present(*pud)) {
3757 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003758 pgd_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003759 } else /* Another has populated it */
3760 pmd_free(mm, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003762 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003763 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764}
3765#endif /* __PAGETABLE_PMD_FOLDED */
3766
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003767static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003768 pte_t **ptepp, spinlock_t **ptlp)
3769{
3770 pgd_t *pgd;
3771 pud_t *pud;
3772 pmd_t *pmd;
3773 pte_t *ptep;
3774
3775 pgd = pgd_offset(mm, address);
3776 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3777 goto out;
3778
3779 pud = pud_offset(pgd, address);
3780 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3781 goto out;
3782
3783 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003784 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003785 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3786 goto out;
3787
3788 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3789 if (pmd_huge(*pmd))
3790 goto out;
3791
3792 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3793 if (!ptep)
3794 goto out;
3795 if (!pte_present(*ptep))
3796 goto unlock;
3797 *ptepp = ptep;
3798 return 0;
3799unlock:
3800 pte_unmap_unlock(ptep, *ptlp);
3801out:
3802 return -EINVAL;
3803}
3804
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003805static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3806 pte_t **ptepp, spinlock_t **ptlp)
3807{
3808 int res;
3809
3810 /* (void) is needed to make gcc happy */
3811 (void) __cond_lock(*ptlp,
3812 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3813 return res;
3814}
3815
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003816/**
3817 * follow_pfn - look up PFN at a user virtual address
3818 * @vma: memory mapping
3819 * @address: user virtual address
3820 * @pfn: location to store found PFN
3821 *
3822 * Only IO mappings and raw PFN mappings are allowed.
3823 *
3824 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3825 */
3826int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3827 unsigned long *pfn)
3828{
3829 int ret = -EINVAL;
3830 spinlock_t *ptl;
3831 pte_t *ptep;
3832
3833 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3834 return ret;
3835
3836 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3837 if (ret)
3838 return ret;
3839 *pfn = pte_pfn(*ptep);
3840 pte_unmap_unlock(ptep, ptl);
3841 return 0;
3842}
3843EXPORT_SYMBOL(follow_pfn);
3844
Rik van Riel28b2ee22008-07-23 21:27:05 -07003845#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003846int follow_phys(struct vm_area_struct *vma,
3847 unsigned long address, unsigned int flags,
3848 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003849{
Johannes Weiner03668a42009-06-16 15:32:34 -07003850 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003851 pte_t *ptep, pte;
3852 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003853
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003854 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3855 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003856
Johannes Weiner03668a42009-06-16 15:32:34 -07003857 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003858 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003859 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003860
Rik van Riel28b2ee22008-07-23 21:27:05 -07003861 if ((flags & FOLL_WRITE) && !pte_write(pte))
3862 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003863
3864 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003865 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003866
Johannes Weiner03668a42009-06-16 15:32:34 -07003867 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003868unlock:
3869 pte_unmap_unlock(ptep, ptl);
3870out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003871 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003872}
3873
3874int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3875 void *buf, int len, int write)
3876{
3877 resource_size_t phys_addr;
3878 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003879 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003880 int offset = addr & (PAGE_SIZE-1);
3881
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003882 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003883 return -EINVAL;
3884
Grazvydas Ignotas9cb12d72015-02-12 15:00:19 -08003885 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
jie@chenjie6@huwei.comaf669a02018-08-10 17:23:06 -07003886 if (!maddr)
3887 return -ENOMEM;
3888
Rik van Riel28b2ee22008-07-23 21:27:05 -07003889 if (write)
3890 memcpy_toio(maddr + offset, buf, len);
3891 else
3892 memcpy_fromio(buf, maddr + offset, len);
3893 iounmap(maddr);
3894
3895 return len;
3896}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02003897EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003898#endif
3899
David Howells0ec76a12006-09-27 01:50:15 -07003900/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003901 * Access another process' address space as given in mm. If non-NULL, use the
3902 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003903 */
Eric W. Biedermane71b4e02016-11-22 12:06:50 -06003904int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003905 unsigned long addr, void *buf, int len, unsigned int gup_flags)
David Howells0ec76a12006-09-27 01:50:15 -07003906{
David Howells0ec76a12006-09-27 01:50:15 -07003907 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003908 void *old_buf = buf;
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003909 int write = gup_flags & FOLL_WRITE;
David Howells0ec76a12006-09-27 01:50:15 -07003910
David Howells0ec76a12006-09-27 01:50:15 -07003911 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003912 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003913 while (len) {
3914 int bytes, ret, offset;
3915 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003916 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003917
Dave Hansen1e987792016-02-12 13:01:54 -08003918 ret = get_user_pages_remote(tsk, mm, addr, 1,
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003919 gup_flags, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003920 if (ret <= 0) {
Rik van Rieldbffcd02014-08-06 16:08:12 -07003921#ifndef CONFIG_HAVE_IOREMAP_PROT
3922 break;
3923#else
Rik van Riel28b2ee22008-07-23 21:27:05 -07003924 /*
3925 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3926 * we can access using slightly different code.
3927 */
Rik van Riel28b2ee22008-07-23 21:27:05 -07003928 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003929 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003930 break;
3931 if (vma->vm_ops && vma->vm_ops->access)
3932 ret = vma->vm_ops->access(vma, addr, buf,
3933 len, write);
3934 if (ret <= 0)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003935 break;
3936 bytes = ret;
Rik van Rieldbffcd02014-08-06 16:08:12 -07003937#endif
David Howells0ec76a12006-09-27 01:50:15 -07003938 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003939 bytes = len;
3940 offset = addr & (PAGE_SIZE-1);
3941 if (bytes > PAGE_SIZE-offset)
3942 bytes = PAGE_SIZE-offset;
3943
3944 maddr = kmap(page);
3945 if (write) {
3946 copy_to_user_page(vma, page, addr,
3947 maddr + offset, buf, bytes);
3948 set_page_dirty_lock(page);
3949 } else {
3950 copy_from_user_page(vma, page, addr,
3951 buf, maddr + offset, bytes);
3952 }
3953 kunmap(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003954 put_page(page);
David Howells0ec76a12006-09-27 01:50:15 -07003955 }
David Howells0ec76a12006-09-27 01:50:15 -07003956 len -= bytes;
3957 buf += bytes;
3958 addr += bytes;
3959 }
3960 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003961
3962 return buf - old_buf;
3963}
Andi Kleen03252912008-01-30 13:33:18 +01003964
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003965/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003966 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003967 * @mm: the mm_struct of the target address space
3968 * @addr: start address to access
3969 * @buf: source or destination buffer
3970 * @len: number of bytes to transfer
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003971 * @gup_flags: flags modifying lookup behaviour
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003972 *
3973 * The caller must hold a reference on @mm.
3974 */
3975int access_remote_vm(struct mm_struct *mm, unsigned long addr,
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003976 void *buf, int len, unsigned int gup_flags)
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003977{
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003978 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003979}
3980
Andi Kleen03252912008-01-30 13:33:18 +01003981/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003982 * Access another process' address space.
3983 * Source/target buffer must be kernel space,
3984 * Do not walk the page table directly, use get_user_pages
3985 */
3986int access_process_vm(struct task_struct *tsk, unsigned long addr,
Lorenzo Stoakesf307ab62016-10-13 01:20:20 +01003987 void *buf, int len, unsigned int gup_flags)
Stephen Wilson206cb632011-03-13 15:49:19 -04003988{
3989 struct mm_struct *mm;
3990 int ret;
3991
3992 mm = get_task_mm(tsk);
3993 if (!mm)
3994 return 0;
3995
Lorenzo Stoakesf307ab62016-10-13 01:20:20 +01003996 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003997
Stephen Wilson206cb632011-03-13 15:49:19 -04003998 mmput(mm);
3999
4000 return ret;
4001}
4002
Andi Kleen03252912008-01-30 13:33:18 +01004003/*
4004 * Print the name of a VMA.
4005 */
4006void print_vma_addr(char *prefix, unsigned long ip)
4007{
4008 struct mm_struct *mm = current->mm;
4009 struct vm_area_struct *vma;
4010
Ingo Molnare8bff742008-02-13 20:21:06 +01004011 /*
4012 * Do not print if we are in atomic
4013 * contexts (in exception stacks, etc.):
4014 */
4015 if (preempt_count())
4016 return;
4017
Andi Kleen03252912008-01-30 13:33:18 +01004018 down_read(&mm->mmap_sem);
4019 vma = find_vma(mm, ip);
4020 if (vma && vma->vm_file) {
4021 struct file *f = vma->vm_file;
4022 char *buf = (char *)__get_free_page(GFP_KERNEL);
4023 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004024 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01004025
Miklos Szeredi9bf39ab2015-06-19 10:29:13 +02004026 p = file_path(f, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01004027 if (IS_ERR(p))
4028 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08004029 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01004030 vma->vm_start,
4031 vma->vm_end - vma->vm_start);
4032 free_page((unsigned long)buf);
4033 }
4034 }
Jeff Liu51a07e52012-07-31 16:43:18 -07004035 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01004036}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004037
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004038#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
David Hildenbrand9ec23532015-05-11 17:52:07 +02004039void __might_fault(const char *file, int line)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004040{
Peter Zijlstra95156f02009-01-12 13:02:11 +01004041 /*
4042 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4043 * holding the mmap_sem, this is safe because kernel memory doesn't
4044 * get paged out, therefore we'll never actually fault, and the
4045 * below annotations will generate false positives.
4046 */
4047 if (segment_eq(get_fs(), KERNEL_DS))
4048 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02004049 if (pagefault_disabled())
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004050 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02004051 __might_sleep(file, line, 0);
4052#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03004053 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004054 might_lock_read(&current->mm->mmap_sem);
David Hildenbrand9ec23532015-05-11 17:52:07 +02004055#endif
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004056}
David Hildenbrand9ec23532015-05-11 17:52:07 +02004057EXPORT_SYMBOL(__might_fault);
Nick Piggin3ee1afa2008-09-10 13:37:17 +02004058#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004059
4060#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4061static void clear_gigantic_page(struct page *page,
4062 unsigned long addr,
4063 unsigned int pages_per_huge_page)
4064{
4065 int i;
4066 struct page *p = page;
4067
4068 might_sleep();
4069 for (i = 0; i < pages_per_huge_page;
4070 i++, p = mem_map_next(p, page, i)) {
4071 cond_resched();
4072 clear_user_highpage(p, addr + i * PAGE_SIZE);
4073 }
4074}
4075void clear_huge_page(struct page *page,
4076 unsigned long addr, unsigned int pages_per_huge_page)
4077{
4078 int i;
4079
4080 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4081 clear_gigantic_page(page, addr, pages_per_huge_page);
4082 return;
4083 }
4084
4085 might_sleep();
4086 for (i = 0; i < pages_per_huge_page; i++) {
4087 cond_resched();
4088 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4089 }
4090}
4091
4092static void copy_user_gigantic_page(struct page *dst, struct page *src,
4093 unsigned long addr,
4094 struct vm_area_struct *vma,
4095 unsigned int pages_per_huge_page)
4096{
4097 int i;
4098 struct page *dst_base = dst;
4099 struct page *src_base = src;
4100
4101 for (i = 0; i < pages_per_huge_page; ) {
4102 cond_resched();
4103 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4104
4105 i++;
4106 dst = mem_map_next(dst, dst_base, i);
4107 src = mem_map_next(src, src_base, i);
4108 }
4109}
4110
4111void copy_user_huge_page(struct page *dst, struct page *src,
4112 unsigned long addr, struct vm_area_struct *vma,
4113 unsigned int pages_per_huge_page)
4114{
4115 int i;
4116
4117 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4118 copy_user_gigantic_page(dst, src, addr, vma,
4119 pages_per_huge_page);
4120 return;
4121 }
4122
4123 might_sleep();
4124 for (i = 0; i < pages_per_huge_page; i++) {
4125 cond_resched();
4126 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4127 }
4128}
4129#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004130
Olof Johansson40b64ac2013-12-20 14:28:05 -08004131#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004132
4133static struct kmem_cache *page_ptl_cachep;
4134
4135void __init ptlock_cache_init(void)
4136{
4137 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4138 SLAB_PANIC, NULL);
4139}
4140
Peter Zijlstra539edb52013-11-14 14:31:52 -08004141bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004142{
4143 spinlock_t *ptl;
4144
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004145 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004146 if (!ptl)
4147 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08004148 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004149 return true;
4150}
4151
Peter Zijlstra539edb52013-11-14 14:31:52 -08004152void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004153{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004154 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004155}
4156#endif