]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - mm/memory.c
mm: memory: merge shared-writable dirtying branches in do_wp_page()
[mirror_ubuntu-bionic-kernel.git] / mm / memory.c
1 /*
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>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60 #include <linux/migrate.h>
61 #include <linux/string.h>
62 #include <linux/dma-debug.h>
63 #include <linux/debugfs.h>
64
65 #include <asm/io.h>
66 #include <asm/pgalloc.h>
67 #include <asm/uaccess.h>
68 #include <asm/tlb.h>
69 #include <asm/tlbflush.h>
70 #include <asm/pgtable.h>
71
72 #include "internal.h"
73
74 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
76 #endif
77
78 #ifndef CONFIG_NEED_MULTIPLE_NODES
79 /* use the per-pgdat data instead for discontigmem - mbligh */
80 unsigned long max_mapnr;
81 struct page *mem_map;
82
83 EXPORT_SYMBOL(max_mapnr);
84 EXPORT_SYMBOL(mem_map);
85 #endif
86
87 /*
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92 * and ZONE_HIGHMEM.
93 */
94 void * high_memory;
95
96 EXPORT_SYMBOL(high_memory);
97
98 /*
99 * Randomize the address space (stacks, mmaps, brk, etc.).
100 *
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
103 */
104 int randomize_va_space __read_mostly =
105 #ifdef CONFIG_COMPAT_BRK
106 1;
107 #else
108 2;
109 #endif
110
111 static int __init disable_randmaps(char *s)
112 {
113 randomize_va_space = 0;
114 return 1;
115 }
116 __setup("norandmaps", disable_randmaps);
117
118 unsigned long zero_pfn __read_mostly;
119 unsigned long highest_memmap_pfn __read_mostly;
120
121 EXPORT_SYMBOL(zero_pfn);
122
123 /*
124 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
125 */
126 static int __init init_zero_pfn(void)
127 {
128 zero_pfn = page_to_pfn(ZERO_PAGE(0));
129 return 0;
130 }
131 core_initcall(init_zero_pfn);
132
133
134 #if defined(SPLIT_RSS_COUNTING)
135
136 void sync_mm_rss(struct mm_struct *mm)
137 {
138 int i;
139
140 for (i = 0; i < NR_MM_COUNTERS; i++) {
141 if (current->rss_stat.count[i]) {
142 add_mm_counter(mm, i, current->rss_stat.count[i]);
143 current->rss_stat.count[i] = 0;
144 }
145 }
146 current->rss_stat.events = 0;
147 }
148
149 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
150 {
151 struct task_struct *task = current;
152
153 if (likely(task->mm == mm))
154 task->rss_stat.count[member] += val;
155 else
156 add_mm_counter(mm, member, val);
157 }
158 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
159 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
160
161 /* sync counter once per 64 page faults */
162 #define TASK_RSS_EVENTS_THRESH (64)
163 static void check_sync_rss_stat(struct task_struct *task)
164 {
165 if (unlikely(task != current))
166 return;
167 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
168 sync_mm_rss(task->mm);
169 }
170 #else /* SPLIT_RSS_COUNTING */
171
172 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
173 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
174
175 static void check_sync_rss_stat(struct task_struct *task)
176 {
177 }
178
179 #endif /* SPLIT_RSS_COUNTING */
180
181 #ifdef HAVE_GENERIC_MMU_GATHER
182
183 static int tlb_next_batch(struct mmu_gather *tlb)
184 {
185 struct mmu_gather_batch *batch;
186
187 batch = tlb->active;
188 if (batch->next) {
189 tlb->active = batch->next;
190 return 1;
191 }
192
193 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
194 return 0;
195
196 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
197 if (!batch)
198 return 0;
199
200 tlb->batch_count++;
201 batch->next = NULL;
202 batch->nr = 0;
203 batch->max = MAX_GATHER_BATCH;
204
205 tlb->active->next = batch;
206 tlb->active = batch;
207
208 return 1;
209 }
210
211 /* tlb_gather_mmu
212 * Called to initialize an (on-stack) mmu_gather structure for page-table
213 * tear-down from @mm. The @fullmm argument is used when @mm is without
214 * users and we're going to destroy the full address space (exit/execve).
215 */
216 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
217 {
218 tlb->mm = mm;
219
220 /* Is it from 0 to ~0? */
221 tlb->fullmm = !(start | (end+1));
222 tlb->need_flush_all = 0;
223 tlb->local.next = NULL;
224 tlb->local.nr = 0;
225 tlb->local.max = ARRAY_SIZE(tlb->__pages);
226 tlb->active = &tlb->local;
227 tlb->batch_count = 0;
228
229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
230 tlb->batch = NULL;
231 #endif
232
233 __tlb_reset_range(tlb);
234 }
235
236 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
237 {
238 if (!tlb->end)
239 return;
240
241 tlb_flush(tlb);
242 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
243 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
244 tlb_table_flush(tlb);
245 #endif
246 __tlb_reset_range(tlb);
247 }
248
249 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
250 {
251 struct mmu_gather_batch *batch;
252
253 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
254 free_pages_and_swap_cache(batch->pages, batch->nr);
255 batch->nr = 0;
256 }
257 tlb->active = &tlb->local;
258 }
259
260 void tlb_flush_mmu(struct mmu_gather *tlb)
261 {
262 tlb_flush_mmu_tlbonly(tlb);
263 tlb_flush_mmu_free(tlb);
264 }
265
266 /* tlb_finish_mmu
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
269 */
270 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
271 {
272 struct mmu_gather_batch *batch, *next;
273
274 tlb_flush_mmu(tlb);
275
276 /* keep the page table cache within bounds */
277 check_pgt_cache();
278
279 for (batch = tlb->local.next; batch; batch = next) {
280 next = batch->next;
281 free_pages((unsigned long)batch, 0);
282 }
283 tlb->local.next = NULL;
284 }
285
286 /* __tlb_remove_page
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
291 */
292 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
293 {
294 struct mmu_gather_batch *batch;
295
296 VM_BUG_ON(!tlb->end);
297
298 batch = tlb->active;
299 batch->pages[batch->nr++] = page;
300 if (batch->nr == batch->max) {
301 if (!tlb_next_batch(tlb))
302 return 0;
303 batch = tlb->active;
304 }
305 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
306
307 return batch->max - batch->nr;
308 }
309
310 #endif /* HAVE_GENERIC_MMU_GATHER */
311
312 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
313
314 /*
315 * See the comment near struct mmu_table_batch.
316 */
317
318 static void tlb_remove_table_smp_sync(void *arg)
319 {
320 /* Simply deliver the interrupt */
321 }
322
323 static void tlb_remove_table_one(void *table)
324 {
325 /*
326 * This isn't an RCU grace period and hence the page-tables cannot be
327 * assumed to be actually RCU-freed.
328 *
329 * It is however sufficient for software page-table walkers that rely on
330 * IRQ disabling. See the comment near struct mmu_table_batch.
331 */
332 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
333 __tlb_remove_table(table);
334 }
335
336 static void tlb_remove_table_rcu(struct rcu_head *head)
337 {
338 struct mmu_table_batch *batch;
339 int i;
340
341 batch = container_of(head, struct mmu_table_batch, rcu);
342
343 for (i = 0; i < batch->nr; i++)
344 __tlb_remove_table(batch->tables[i]);
345
346 free_page((unsigned long)batch);
347 }
348
349 void tlb_table_flush(struct mmu_gather *tlb)
350 {
351 struct mmu_table_batch **batch = &tlb->batch;
352
353 if (*batch) {
354 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
355 *batch = NULL;
356 }
357 }
358
359 void tlb_remove_table(struct mmu_gather *tlb, void *table)
360 {
361 struct mmu_table_batch **batch = &tlb->batch;
362
363 /*
364 * When there's less then two users of this mm there cannot be a
365 * concurrent page-table walk.
366 */
367 if (atomic_read(&tlb->mm->mm_users) < 2) {
368 __tlb_remove_table(table);
369 return;
370 }
371
372 if (*batch == NULL) {
373 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
374 if (*batch == NULL) {
375 tlb_remove_table_one(table);
376 return;
377 }
378 (*batch)->nr = 0;
379 }
380 (*batch)->tables[(*batch)->nr++] = table;
381 if ((*batch)->nr == MAX_TABLE_BATCH)
382 tlb_table_flush(tlb);
383 }
384
385 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
386
387 /*
388 * Note: this doesn't free the actual pages themselves. That
389 * has been handled earlier when unmapping all the memory regions.
390 */
391 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
392 unsigned long addr)
393 {
394 pgtable_t token = pmd_pgtable(*pmd);
395 pmd_clear(pmd);
396 pte_free_tlb(tlb, token, addr);
397 atomic_long_dec(&tlb->mm->nr_ptes);
398 }
399
400 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
401 unsigned long addr, unsigned long end,
402 unsigned long floor, unsigned long ceiling)
403 {
404 pmd_t *pmd;
405 unsigned long next;
406 unsigned long start;
407
408 start = addr;
409 pmd = pmd_offset(pud, addr);
410 do {
411 next = pmd_addr_end(addr, end);
412 if (pmd_none_or_clear_bad(pmd))
413 continue;
414 free_pte_range(tlb, pmd, addr);
415 } while (pmd++, addr = next, addr != end);
416
417 start &= PUD_MASK;
418 if (start < floor)
419 return;
420 if (ceiling) {
421 ceiling &= PUD_MASK;
422 if (!ceiling)
423 return;
424 }
425 if (end - 1 > ceiling - 1)
426 return;
427
428 pmd = pmd_offset(pud, start);
429 pud_clear(pud);
430 pmd_free_tlb(tlb, pmd, start);
431 }
432
433 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
434 unsigned long addr, unsigned long end,
435 unsigned long floor, unsigned long ceiling)
436 {
437 pud_t *pud;
438 unsigned long next;
439 unsigned long start;
440
441 start = addr;
442 pud = pud_offset(pgd, addr);
443 do {
444 next = pud_addr_end(addr, end);
445 if (pud_none_or_clear_bad(pud))
446 continue;
447 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
448 } while (pud++, addr = next, addr != end);
449
450 start &= PGDIR_MASK;
451 if (start < floor)
452 return;
453 if (ceiling) {
454 ceiling &= PGDIR_MASK;
455 if (!ceiling)
456 return;
457 }
458 if (end - 1 > ceiling - 1)
459 return;
460
461 pud = pud_offset(pgd, start);
462 pgd_clear(pgd);
463 pud_free_tlb(tlb, pud, start);
464 }
465
466 /*
467 * This function frees user-level page tables of a process.
468 */
469 void free_pgd_range(struct mmu_gather *tlb,
470 unsigned long addr, unsigned long end,
471 unsigned long floor, unsigned long ceiling)
472 {
473 pgd_t *pgd;
474 unsigned long next;
475
476 /*
477 * The next few lines have given us lots of grief...
478 *
479 * Why are we testing PMD* at this top level? Because often
480 * there will be no work to do at all, and we'd prefer not to
481 * go all the way down to the bottom just to discover that.
482 *
483 * Why all these "- 1"s? Because 0 represents both the bottom
484 * of the address space and the top of it (using -1 for the
485 * top wouldn't help much: the masks would do the wrong thing).
486 * The rule is that addr 0 and floor 0 refer to the bottom of
487 * the address space, but end 0 and ceiling 0 refer to the top
488 * Comparisons need to use "end - 1" and "ceiling - 1" (though
489 * that end 0 case should be mythical).
490 *
491 * Wherever addr is brought up or ceiling brought down, we must
492 * be careful to reject "the opposite 0" before it confuses the
493 * subsequent tests. But what about where end is brought down
494 * by PMD_SIZE below? no, end can't go down to 0 there.
495 *
496 * Whereas we round start (addr) and ceiling down, by different
497 * masks at different levels, in order to test whether a table
498 * now has no other vmas using it, so can be freed, we don't
499 * bother to round floor or end up - the tests don't need that.
500 */
501
502 addr &= PMD_MASK;
503 if (addr < floor) {
504 addr += PMD_SIZE;
505 if (!addr)
506 return;
507 }
508 if (ceiling) {
509 ceiling &= PMD_MASK;
510 if (!ceiling)
511 return;
512 }
513 if (end - 1 > ceiling - 1)
514 end -= PMD_SIZE;
515 if (addr > end - 1)
516 return;
517
518 pgd = pgd_offset(tlb->mm, addr);
519 do {
520 next = pgd_addr_end(addr, end);
521 if (pgd_none_or_clear_bad(pgd))
522 continue;
523 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
524 } while (pgd++, addr = next, addr != end);
525 }
526
527 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
528 unsigned long floor, unsigned long ceiling)
529 {
530 while (vma) {
531 struct vm_area_struct *next = vma->vm_next;
532 unsigned long addr = vma->vm_start;
533
534 /*
535 * Hide vma from rmap and truncate_pagecache before freeing
536 * pgtables
537 */
538 unlink_anon_vmas(vma);
539 unlink_file_vma(vma);
540
541 if (is_vm_hugetlb_page(vma)) {
542 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
543 floor, next? next->vm_start: ceiling);
544 } else {
545 /*
546 * Optimization: gather nearby vmas into one call down
547 */
548 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
549 && !is_vm_hugetlb_page(next)) {
550 vma = next;
551 next = vma->vm_next;
552 unlink_anon_vmas(vma);
553 unlink_file_vma(vma);
554 }
555 free_pgd_range(tlb, addr, vma->vm_end,
556 floor, next? next->vm_start: ceiling);
557 }
558 vma = next;
559 }
560 }
561
562 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
563 pmd_t *pmd, unsigned long address)
564 {
565 spinlock_t *ptl;
566 pgtable_t new = pte_alloc_one(mm, address);
567 int wait_split_huge_page;
568 if (!new)
569 return -ENOMEM;
570
571 /*
572 * Ensure all pte setup (eg. pte page lock and page clearing) are
573 * visible before the pte is made visible to other CPUs by being
574 * put into page tables.
575 *
576 * The other side of the story is the pointer chasing in the page
577 * table walking code (when walking the page table without locking;
578 * ie. most of the time). Fortunately, these data accesses consist
579 * of a chain of data-dependent loads, meaning most CPUs (alpha
580 * being the notable exception) will already guarantee loads are
581 * seen in-order. See the alpha page table accessors for the
582 * smp_read_barrier_depends() barriers in page table walking code.
583 */
584 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
585
586 ptl = pmd_lock(mm, pmd);
587 wait_split_huge_page = 0;
588 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
589 atomic_long_inc(&mm->nr_ptes);
590 pmd_populate(mm, pmd, new);
591 new = NULL;
592 } else if (unlikely(pmd_trans_splitting(*pmd)))
593 wait_split_huge_page = 1;
594 spin_unlock(ptl);
595 if (new)
596 pte_free(mm, new);
597 if (wait_split_huge_page)
598 wait_split_huge_page(vma->anon_vma, pmd);
599 return 0;
600 }
601
602 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
603 {
604 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
605 if (!new)
606 return -ENOMEM;
607
608 smp_wmb(); /* See comment in __pte_alloc */
609
610 spin_lock(&init_mm.page_table_lock);
611 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
612 pmd_populate_kernel(&init_mm, pmd, new);
613 new = NULL;
614 } else
615 VM_BUG_ON(pmd_trans_splitting(*pmd));
616 spin_unlock(&init_mm.page_table_lock);
617 if (new)
618 pte_free_kernel(&init_mm, new);
619 return 0;
620 }
621
622 static inline void init_rss_vec(int *rss)
623 {
624 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
625 }
626
627 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
628 {
629 int i;
630
631 if (current->mm == mm)
632 sync_mm_rss(mm);
633 for (i = 0; i < NR_MM_COUNTERS; i++)
634 if (rss[i])
635 add_mm_counter(mm, i, rss[i]);
636 }
637
638 /*
639 * This function is called to print an error when a bad pte
640 * is found. For example, we might have a PFN-mapped pte in
641 * a region that doesn't allow it.
642 *
643 * The calling function must still handle the error.
644 */
645 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
646 pte_t pte, struct page *page)
647 {
648 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
649 pud_t *pud = pud_offset(pgd, addr);
650 pmd_t *pmd = pmd_offset(pud, addr);
651 struct address_space *mapping;
652 pgoff_t index;
653 static unsigned long resume;
654 static unsigned long nr_shown;
655 static unsigned long nr_unshown;
656
657 /*
658 * Allow a burst of 60 reports, then keep quiet for that minute;
659 * or allow a steady drip of one report per second.
660 */
661 if (nr_shown == 60) {
662 if (time_before(jiffies, resume)) {
663 nr_unshown++;
664 return;
665 }
666 if (nr_unshown) {
667 printk(KERN_ALERT
668 "BUG: Bad page map: %lu messages suppressed\n",
669 nr_unshown);
670 nr_unshown = 0;
671 }
672 nr_shown = 0;
673 }
674 if (nr_shown++ == 0)
675 resume = jiffies + 60 * HZ;
676
677 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
678 index = linear_page_index(vma, addr);
679
680 printk(KERN_ALERT
681 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
682 current->comm,
683 (long long)pte_val(pte), (long long)pmd_val(*pmd));
684 if (page)
685 dump_page(page, "bad pte");
686 printk(KERN_ALERT
687 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
688 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
689 /*
690 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
691 */
692 if (vma->vm_ops)
693 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
694 vma->vm_ops->fault);
695 if (vma->vm_file)
696 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
697 vma->vm_file->f_op->mmap);
698 dump_stack();
699 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
700 }
701
702 /*
703 * vm_normal_page -- This function gets the "struct page" associated with a pte.
704 *
705 * "Special" mappings do not wish to be associated with a "struct page" (either
706 * it doesn't exist, or it exists but they don't want to touch it). In this
707 * case, NULL is returned here. "Normal" mappings do have a struct page.
708 *
709 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
710 * pte bit, in which case this function is trivial. Secondly, an architecture
711 * may not have a spare pte bit, which requires a more complicated scheme,
712 * described below.
713 *
714 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
715 * special mapping (even if there are underlying and valid "struct pages").
716 * COWed pages of a VM_PFNMAP are always normal.
717 *
718 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
719 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
720 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
721 * mapping will always honor the rule
722 *
723 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
724 *
725 * And for normal mappings this is false.
726 *
727 * This restricts such mappings to be a linear translation from virtual address
728 * to pfn. To get around this restriction, we allow arbitrary mappings so long
729 * as the vma is not a COW mapping; in that case, we know that all ptes are
730 * special (because none can have been COWed).
731 *
732 *
733 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
734 *
735 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
736 * page" backing, however the difference is that _all_ pages with a struct
737 * page (that is, those where pfn_valid is true) are refcounted and considered
738 * normal pages by the VM. The disadvantage is that pages are refcounted
739 * (which can be slower and simply not an option for some PFNMAP users). The
740 * advantage is that we don't have to follow the strict linearity rule of
741 * PFNMAP mappings in order to support COWable mappings.
742 *
743 */
744 #ifdef __HAVE_ARCH_PTE_SPECIAL
745 # define HAVE_PTE_SPECIAL 1
746 #else
747 # define HAVE_PTE_SPECIAL 0
748 #endif
749 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
750 pte_t pte)
751 {
752 unsigned long pfn = pte_pfn(pte);
753
754 if (HAVE_PTE_SPECIAL) {
755 if (likely(!pte_special(pte)))
756 goto check_pfn;
757 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
758 return NULL;
759 if (!is_zero_pfn(pfn))
760 print_bad_pte(vma, addr, pte, NULL);
761 return NULL;
762 }
763
764 /* !HAVE_PTE_SPECIAL case follows: */
765
766 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
767 if (vma->vm_flags & VM_MIXEDMAP) {
768 if (!pfn_valid(pfn))
769 return NULL;
770 goto out;
771 } else {
772 unsigned long off;
773 off = (addr - vma->vm_start) >> PAGE_SHIFT;
774 if (pfn == vma->vm_pgoff + off)
775 return NULL;
776 if (!is_cow_mapping(vma->vm_flags))
777 return NULL;
778 }
779 }
780
781 if (is_zero_pfn(pfn))
782 return NULL;
783 check_pfn:
784 if (unlikely(pfn > highest_memmap_pfn)) {
785 print_bad_pte(vma, addr, pte, NULL);
786 return NULL;
787 }
788
789 /*
790 * NOTE! We still have PageReserved() pages in the page tables.
791 * eg. VDSO mappings can cause them to exist.
792 */
793 out:
794 return pfn_to_page(pfn);
795 }
796
797 /*
798 * copy one vm_area from one task to the other. Assumes the page tables
799 * already present in the new task to be cleared in the whole range
800 * covered by this vma.
801 */
802
803 static inline unsigned long
804 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
805 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
806 unsigned long addr, int *rss)
807 {
808 unsigned long vm_flags = vma->vm_flags;
809 pte_t pte = *src_pte;
810 struct page *page;
811
812 /* pte contains position in swap or file, so copy. */
813 if (unlikely(!pte_present(pte))) {
814 swp_entry_t entry = pte_to_swp_entry(pte);
815
816 if (likely(!non_swap_entry(entry))) {
817 if (swap_duplicate(entry) < 0)
818 return entry.val;
819
820 /* make sure dst_mm is on swapoff's mmlist. */
821 if (unlikely(list_empty(&dst_mm->mmlist))) {
822 spin_lock(&mmlist_lock);
823 if (list_empty(&dst_mm->mmlist))
824 list_add(&dst_mm->mmlist,
825 &src_mm->mmlist);
826 spin_unlock(&mmlist_lock);
827 }
828 rss[MM_SWAPENTS]++;
829 } else if (is_migration_entry(entry)) {
830 page = migration_entry_to_page(entry);
831
832 if (PageAnon(page))
833 rss[MM_ANONPAGES]++;
834 else
835 rss[MM_FILEPAGES]++;
836
837 if (is_write_migration_entry(entry) &&
838 is_cow_mapping(vm_flags)) {
839 /*
840 * COW mappings require pages in both
841 * parent and child to be set to read.
842 */
843 make_migration_entry_read(&entry);
844 pte = swp_entry_to_pte(entry);
845 if (pte_swp_soft_dirty(*src_pte))
846 pte = pte_swp_mksoft_dirty(pte);
847 set_pte_at(src_mm, addr, src_pte, pte);
848 }
849 }
850 goto out_set_pte;
851 }
852
853 /*
854 * If it's a COW mapping, write protect it both
855 * in the parent and the child
856 */
857 if (is_cow_mapping(vm_flags)) {
858 ptep_set_wrprotect(src_mm, addr, src_pte);
859 pte = pte_wrprotect(pte);
860 }
861
862 /*
863 * If it's a shared mapping, mark it clean in
864 * the child
865 */
866 if (vm_flags & VM_SHARED)
867 pte = pte_mkclean(pte);
868 pte = pte_mkold(pte);
869
870 page = vm_normal_page(vma, addr, pte);
871 if (page) {
872 get_page(page);
873 page_dup_rmap(page);
874 if (PageAnon(page))
875 rss[MM_ANONPAGES]++;
876 else
877 rss[MM_FILEPAGES]++;
878 }
879
880 out_set_pte:
881 set_pte_at(dst_mm, addr, dst_pte, pte);
882 return 0;
883 }
884
885 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
886 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
887 unsigned long addr, unsigned long end)
888 {
889 pte_t *orig_src_pte, *orig_dst_pte;
890 pte_t *src_pte, *dst_pte;
891 spinlock_t *src_ptl, *dst_ptl;
892 int progress = 0;
893 int rss[NR_MM_COUNTERS];
894 swp_entry_t entry = (swp_entry_t){0};
895
896 again:
897 init_rss_vec(rss);
898
899 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
900 if (!dst_pte)
901 return -ENOMEM;
902 src_pte = pte_offset_map(src_pmd, addr);
903 src_ptl = pte_lockptr(src_mm, src_pmd);
904 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
905 orig_src_pte = src_pte;
906 orig_dst_pte = dst_pte;
907 arch_enter_lazy_mmu_mode();
908
909 do {
910 /*
911 * We are holding two locks at this point - either of them
912 * could generate latencies in another task on another CPU.
913 */
914 if (progress >= 32) {
915 progress = 0;
916 if (need_resched() ||
917 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
918 break;
919 }
920 if (pte_none(*src_pte)) {
921 progress++;
922 continue;
923 }
924 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
925 vma, addr, rss);
926 if (entry.val)
927 break;
928 progress += 8;
929 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
930
931 arch_leave_lazy_mmu_mode();
932 spin_unlock(src_ptl);
933 pte_unmap(orig_src_pte);
934 add_mm_rss_vec(dst_mm, rss);
935 pte_unmap_unlock(orig_dst_pte, dst_ptl);
936 cond_resched();
937
938 if (entry.val) {
939 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
940 return -ENOMEM;
941 progress = 0;
942 }
943 if (addr != end)
944 goto again;
945 return 0;
946 }
947
948 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
949 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
950 unsigned long addr, unsigned long end)
951 {
952 pmd_t *src_pmd, *dst_pmd;
953 unsigned long next;
954
955 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
956 if (!dst_pmd)
957 return -ENOMEM;
958 src_pmd = pmd_offset(src_pud, addr);
959 do {
960 next = pmd_addr_end(addr, end);
961 if (pmd_trans_huge(*src_pmd)) {
962 int err;
963 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
964 err = copy_huge_pmd(dst_mm, src_mm,
965 dst_pmd, src_pmd, addr, vma);
966 if (err == -ENOMEM)
967 return -ENOMEM;
968 if (!err)
969 continue;
970 /* fall through */
971 }
972 if (pmd_none_or_clear_bad(src_pmd))
973 continue;
974 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
975 vma, addr, next))
976 return -ENOMEM;
977 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
978 return 0;
979 }
980
981 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
982 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
983 unsigned long addr, unsigned long end)
984 {
985 pud_t *src_pud, *dst_pud;
986 unsigned long next;
987
988 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
989 if (!dst_pud)
990 return -ENOMEM;
991 src_pud = pud_offset(src_pgd, addr);
992 do {
993 next = pud_addr_end(addr, end);
994 if (pud_none_or_clear_bad(src_pud))
995 continue;
996 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
997 vma, addr, next))
998 return -ENOMEM;
999 } while (dst_pud++, src_pud++, addr = next, addr != end);
1000 return 0;
1001 }
1002
1003 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1004 struct vm_area_struct *vma)
1005 {
1006 pgd_t *src_pgd, *dst_pgd;
1007 unsigned long next;
1008 unsigned long addr = vma->vm_start;
1009 unsigned long end = vma->vm_end;
1010 unsigned long mmun_start; /* For mmu_notifiers */
1011 unsigned long mmun_end; /* For mmu_notifiers */
1012 bool is_cow;
1013 int ret;
1014
1015 /*
1016 * Don't copy ptes where a page fault will fill them correctly.
1017 * Fork becomes much lighter when there are big shared or private
1018 * readonly mappings. The tradeoff is that copy_page_range is more
1019 * efficient than faulting.
1020 */
1021 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1022 !vma->anon_vma)
1023 return 0;
1024
1025 if (is_vm_hugetlb_page(vma))
1026 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1027
1028 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1029 /*
1030 * We do not free on error cases below as remove_vma
1031 * gets called on error from higher level routine
1032 */
1033 ret = track_pfn_copy(vma);
1034 if (ret)
1035 return ret;
1036 }
1037
1038 /*
1039 * We need to invalidate the secondary MMU mappings only when
1040 * there could be a permission downgrade on the ptes of the
1041 * parent mm. And a permission downgrade will only happen if
1042 * is_cow_mapping() returns true.
1043 */
1044 is_cow = is_cow_mapping(vma->vm_flags);
1045 mmun_start = addr;
1046 mmun_end = end;
1047 if (is_cow)
1048 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1049 mmun_end);
1050
1051 ret = 0;
1052 dst_pgd = pgd_offset(dst_mm, addr);
1053 src_pgd = pgd_offset(src_mm, addr);
1054 do {
1055 next = pgd_addr_end(addr, end);
1056 if (pgd_none_or_clear_bad(src_pgd))
1057 continue;
1058 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1059 vma, addr, next))) {
1060 ret = -ENOMEM;
1061 break;
1062 }
1063 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1064
1065 if (is_cow)
1066 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1067 return ret;
1068 }
1069
1070 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1071 struct vm_area_struct *vma, pmd_t *pmd,
1072 unsigned long addr, unsigned long end,
1073 struct zap_details *details)
1074 {
1075 struct mm_struct *mm = tlb->mm;
1076 int force_flush = 0;
1077 int rss[NR_MM_COUNTERS];
1078 spinlock_t *ptl;
1079 pte_t *start_pte;
1080 pte_t *pte;
1081 swp_entry_t entry;
1082
1083 again:
1084 init_rss_vec(rss);
1085 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1086 pte = start_pte;
1087 arch_enter_lazy_mmu_mode();
1088 do {
1089 pte_t ptent = *pte;
1090 if (pte_none(ptent)) {
1091 continue;
1092 }
1093
1094 if (pte_present(ptent)) {
1095 struct page *page;
1096
1097 page = vm_normal_page(vma, addr, ptent);
1098 if (unlikely(details) && page) {
1099 /*
1100 * unmap_shared_mapping_pages() wants to
1101 * invalidate cache without truncating:
1102 * unmap shared but keep private pages.
1103 */
1104 if (details->check_mapping &&
1105 details->check_mapping != page->mapping)
1106 continue;
1107 }
1108 ptent = ptep_get_and_clear_full(mm, addr, pte,
1109 tlb->fullmm);
1110 tlb_remove_tlb_entry(tlb, pte, addr);
1111 if (unlikely(!page))
1112 continue;
1113 if (PageAnon(page))
1114 rss[MM_ANONPAGES]--;
1115 else {
1116 if (pte_dirty(ptent)) {
1117 force_flush = 1;
1118 set_page_dirty(page);
1119 }
1120 if (pte_young(ptent) &&
1121 likely(!(vma->vm_flags & VM_SEQ_READ)))
1122 mark_page_accessed(page);
1123 rss[MM_FILEPAGES]--;
1124 }
1125 page_remove_rmap(page);
1126 if (unlikely(page_mapcount(page) < 0))
1127 print_bad_pte(vma, addr, ptent, page);
1128 if (unlikely(!__tlb_remove_page(tlb, page))) {
1129 force_flush = 1;
1130 addr += PAGE_SIZE;
1131 break;
1132 }
1133 continue;
1134 }
1135 /* If details->check_mapping, we leave swap entries. */
1136 if (unlikely(details))
1137 continue;
1138
1139 entry = pte_to_swp_entry(ptent);
1140 if (!non_swap_entry(entry))
1141 rss[MM_SWAPENTS]--;
1142 else if (is_migration_entry(entry)) {
1143 struct page *page;
1144
1145 page = migration_entry_to_page(entry);
1146
1147 if (PageAnon(page))
1148 rss[MM_ANONPAGES]--;
1149 else
1150 rss[MM_FILEPAGES]--;
1151 }
1152 if (unlikely(!free_swap_and_cache(entry)))
1153 print_bad_pte(vma, addr, ptent, NULL);
1154 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1155 } while (pte++, addr += PAGE_SIZE, addr != end);
1156
1157 add_mm_rss_vec(mm, rss);
1158 arch_leave_lazy_mmu_mode();
1159
1160 /* Do the actual TLB flush before dropping ptl */
1161 if (force_flush)
1162 tlb_flush_mmu_tlbonly(tlb);
1163 pte_unmap_unlock(start_pte, ptl);
1164
1165 /*
1166 * If we forced a TLB flush (either due to running out of
1167 * batch buffers or because we needed to flush dirty TLB
1168 * entries before releasing the ptl), free the batched
1169 * memory too. Restart if we didn't do everything.
1170 */
1171 if (force_flush) {
1172 force_flush = 0;
1173 tlb_flush_mmu_free(tlb);
1174
1175 if (addr != end)
1176 goto again;
1177 }
1178
1179 return addr;
1180 }
1181
1182 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1183 struct vm_area_struct *vma, pud_t *pud,
1184 unsigned long addr, unsigned long end,
1185 struct zap_details *details)
1186 {
1187 pmd_t *pmd;
1188 unsigned long next;
1189
1190 pmd = pmd_offset(pud, addr);
1191 do {
1192 next = pmd_addr_end(addr, end);
1193 if (pmd_trans_huge(*pmd)) {
1194 if (next - addr != HPAGE_PMD_SIZE) {
1195 #ifdef CONFIG_DEBUG_VM
1196 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1197 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1198 __func__, addr, end,
1199 vma->vm_start,
1200 vma->vm_end);
1201 BUG();
1202 }
1203 #endif
1204 split_huge_page_pmd(vma, addr, pmd);
1205 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1206 goto next;
1207 /* fall through */
1208 }
1209 /*
1210 * Here there can be other concurrent MADV_DONTNEED or
1211 * trans huge page faults running, and if the pmd is
1212 * none or trans huge it can change under us. This is
1213 * because MADV_DONTNEED holds the mmap_sem in read
1214 * mode.
1215 */
1216 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1217 goto next;
1218 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1219 next:
1220 cond_resched();
1221 } while (pmd++, addr = next, addr != end);
1222
1223 return addr;
1224 }
1225
1226 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1227 struct vm_area_struct *vma, pgd_t *pgd,
1228 unsigned long addr, unsigned long end,
1229 struct zap_details *details)
1230 {
1231 pud_t *pud;
1232 unsigned long next;
1233
1234 pud = pud_offset(pgd, addr);
1235 do {
1236 next = pud_addr_end(addr, end);
1237 if (pud_none_or_clear_bad(pud))
1238 continue;
1239 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1240 } while (pud++, addr = next, addr != end);
1241
1242 return addr;
1243 }
1244
1245 static void unmap_page_range(struct mmu_gather *tlb,
1246 struct vm_area_struct *vma,
1247 unsigned long addr, unsigned long end,
1248 struct zap_details *details)
1249 {
1250 pgd_t *pgd;
1251 unsigned long next;
1252
1253 if (details && !details->check_mapping)
1254 details = NULL;
1255
1256 BUG_ON(addr >= end);
1257 tlb_start_vma(tlb, vma);
1258 pgd = pgd_offset(vma->vm_mm, addr);
1259 do {
1260 next = pgd_addr_end(addr, end);
1261 if (pgd_none_or_clear_bad(pgd))
1262 continue;
1263 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1264 } while (pgd++, addr = next, addr != end);
1265 tlb_end_vma(tlb, vma);
1266 }
1267
1268
1269 static void unmap_single_vma(struct mmu_gather *tlb,
1270 struct vm_area_struct *vma, unsigned long start_addr,
1271 unsigned long end_addr,
1272 struct zap_details *details)
1273 {
1274 unsigned long start = max(vma->vm_start, start_addr);
1275 unsigned long end;
1276
1277 if (start >= vma->vm_end)
1278 return;
1279 end = min(vma->vm_end, end_addr);
1280 if (end <= vma->vm_start)
1281 return;
1282
1283 if (vma->vm_file)
1284 uprobe_munmap(vma, start, end);
1285
1286 if (unlikely(vma->vm_flags & VM_PFNMAP))
1287 untrack_pfn(vma, 0, 0);
1288
1289 if (start != end) {
1290 if (unlikely(is_vm_hugetlb_page(vma))) {
1291 /*
1292 * It is undesirable to test vma->vm_file as it
1293 * should be non-null for valid hugetlb area.
1294 * However, vm_file will be NULL in the error
1295 * cleanup path of mmap_region. When
1296 * hugetlbfs ->mmap method fails,
1297 * mmap_region() nullifies vma->vm_file
1298 * before calling this function to clean up.
1299 * Since no pte has actually been setup, it is
1300 * safe to do nothing in this case.
1301 */
1302 if (vma->vm_file) {
1303 i_mmap_lock_write(vma->vm_file->f_mapping);
1304 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1305 i_mmap_unlock_write(vma->vm_file->f_mapping);
1306 }
1307 } else
1308 unmap_page_range(tlb, vma, start, end, details);
1309 }
1310 }
1311
1312 /**
1313 * unmap_vmas - unmap a range of memory covered by a list of vma's
1314 * @tlb: address of the caller's struct mmu_gather
1315 * @vma: the starting vma
1316 * @start_addr: virtual address at which to start unmapping
1317 * @end_addr: virtual address at which to end unmapping
1318 *
1319 * Unmap all pages in the vma list.
1320 *
1321 * Only addresses between `start' and `end' will be unmapped.
1322 *
1323 * The VMA list must be sorted in ascending virtual address order.
1324 *
1325 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1326 * range after unmap_vmas() returns. So the only responsibility here is to
1327 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1328 * drops the lock and schedules.
1329 */
1330 void unmap_vmas(struct mmu_gather *tlb,
1331 struct vm_area_struct *vma, unsigned long start_addr,
1332 unsigned long end_addr)
1333 {
1334 struct mm_struct *mm = vma->vm_mm;
1335
1336 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1337 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1338 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1339 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1340 }
1341
1342 /**
1343 * zap_page_range - remove user pages in a given range
1344 * @vma: vm_area_struct holding the applicable pages
1345 * @start: starting address of pages to zap
1346 * @size: number of bytes to zap
1347 * @details: details of shared cache invalidation
1348 *
1349 * Caller must protect the VMA list
1350 */
1351 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1352 unsigned long size, struct zap_details *details)
1353 {
1354 struct mm_struct *mm = vma->vm_mm;
1355 struct mmu_gather tlb;
1356 unsigned long end = start + size;
1357
1358 lru_add_drain();
1359 tlb_gather_mmu(&tlb, mm, start, end);
1360 update_hiwater_rss(mm);
1361 mmu_notifier_invalidate_range_start(mm, start, end);
1362 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1363 unmap_single_vma(&tlb, vma, start, end, details);
1364 mmu_notifier_invalidate_range_end(mm, start, end);
1365 tlb_finish_mmu(&tlb, start, end);
1366 }
1367
1368 /**
1369 * zap_page_range_single - remove user pages in a given range
1370 * @vma: vm_area_struct holding the applicable pages
1371 * @address: starting address of pages to zap
1372 * @size: number of bytes to zap
1373 * @details: details of shared cache invalidation
1374 *
1375 * The range must fit into one VMA.
1376 */
1377 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1378 unsigned long size, struct zap_details *details)
1379 {
1380 struct mm_struct *mm = vma->vm_mm;
1381 struct mmu_gather tlb;
1382 unsigned long end = address + size;
1383
1384 lru_add_drain();
1385 tlb_gather_mmu(&tlb, mm, address, end);
1386 update_hiwater_rss(mm);
1387 mmu_notifier_invalidate_range_start(mm, address, end);
1388 unmap_single_vma(&tlb, vma, address, end, details);
1389 mmu_notifier_invalidate_range_end(mm, address, end);
1390 tlb_finish_mmu(&tlb, address, end);
1391 }
1392
1393 /**
1394 * zap_vma_ptes - remove ptes mapping the vma
1395 * @vma: vm_area_struct holding ptes to be zapped
1396 * @address: starting address of pages to zap
1397 * @size: number of bytes to zap
1398 *
1399 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1400 *
1401 * The entire address range must be fully contained within the vma.
1402 *
1403 * Returns 0 if successful.
1404 */
1405 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1406 unsigned long size)
1407 {
1408 if (address < vma->vm_start || address + size > vma->vm_end ||
1409 !(vma->vm_flags & VM_PFNMAP))
1410 return -1;
1411 zap_page_range_single(vma, address, size, NULL);
1412 return 0;
1413 }
1414 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1415
1416 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1417 spinlock_t **ptl)
1418 {
1419 pgd_t * pgd = pgd_offset(mm, addr);
1420 pud_t * pud = pud_alloc(mm, pgd, addr);
1421 if (pud) {
1422 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1423 if (pmd) {
1424 VM_BUG_ON(pmd_trans_huge(*pmd));
1425 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1426 }
1427 }
1428 return NULL;
1429 }
1430
1431 /*
1432 * This is the old fallback for page remapping.
1433 *
1434 * For historical reasons, it only allows reserved pages. Only
1435 * old drivers should use this, and they needed to mark their
1436 * pages reserved for the old functions anyway.
1437 */
1438 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1439 struct page *page, pgprot_t prot)
1440 {
1441 struct mm_struct *mm = vma->vm_mm;
1442 int retval;
1443 pte_t *pte;
1444 spinlock_t *ptl;
1445
1446 retval = -EINVAL;
1447 if (PageAnon(page))
1448 goto out;
1449 retval = -ENOMEM;
1450 flush_dcache_page(page);
1451 pte = get_locked_pte(mm, addr, &ptl);
1452 if (!pte)
1453 goto out;
1454 retval = -EBUSY;
1455 if (!pte_none(*pte))
1456 goto out_unlock;
1457
1458 /* Ok, finally just insert the thing.. */
1459 get_page(page);
1460 inc_mm_counter_fast(mm, MM_FILEPAGES);
1461 page_add_file_rmap(page);
1462 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1463
1464 retval = 0;
1465 pte_unmap_unlock(pte, ptl);
1466 return retval;
1467 out_unlock:
1468 pte_unmap_unlock(pte, ptl);
1469 out:
1470 return retval;
1471 }
1472
1473 /**
1474 * vm_insert_page - insert single page into user vma
1475 * @vma: user vma to map to
1476 * @addr: target user address of this page
1477 * @page: source kernel page
1478 *
1479 * This allows drivers to insert individual pages they've allocated
1480 * into a user vma.
1481 *
1482 * The page has to be a nice clean _individual_ kernel allocation.
1483 * If you allocate a compound page, you need to have marked it as
1484 * such (__GFP_COMP), or manually just split the page up yourself
1485 * (see split_page()).
1486 *
1487 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1488 * took an arbitrary page protection parameter. This doesn't allow
1489 * that. Your vma protection will have to be set up correctly, which
1490 * means that if you want a shared writable mapping, you'd better
1491 * ask for a shared writable mapping!
1492 *
1493 * The page does not need to be reserved.
1494 *
1495 * Usually this function is called from f_op->mmap() handler
1496 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1497 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1498 * function from other places, for example from page-fault handler.
1499 */
1500 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1501 struct page *page)
1502 {
1503 if (addr < vma->vm_start || addr >= vma->vm_end)
1504 return -EFAULT;
1505 if (!page_count(page))
1506 return -EINVAL;
1507 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1508 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1509 BUG_ON(vma->vm_flags & VM_PFNMAP);
1510 vma->vm_flags |= VM_MIXEDMAP;
1511 }
1512 return insert_page(vma, addr, page, vma->vm_page_prot);
1513 }
1514 EXPORT_SYMBOL(vm_insert_page);
1515
1516 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1517 unsigned long pfn, pgprot_t prot)
1518 {
1519 struct mm_struct *mm = vma->vm_mm;
1520 int retval;
1521 pte_t *pte, entry;
1522 spinlock_t *ptl;
1523
1524 retval = -ENOMEM;
1525 pte = get_locked_pte(mm, addr, &ptl);
1526 if (!pte)
1527 goto out;
1528 retval = -EBUSY;
1529 if (!pte_none(*pte))
1530 goto out_unlock;
1531
1532 /* Ok, finally just insert the thing.. */
1533 entry = pte_mkspecial(pfn_pte(pfn, prot));
1534 set_pte_at(mm, addr, pte, entry);
1535 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1536
1537 retval = 0;
1538 out_unlock:
1539 pte_unmap_unlock(pte, ptl);
1540 out:
1541 return retval;
1542 }
1543
1544 /**
1545 * vm_insert_pfn - insert single pfn into user vma
1546 * @vma: user vma to map to
1547 * @addr: target user address of this page
1548 * @pfn: source kernel pfn
1549 *
1550 * Similar to vm_insert_page, this allows drivers to insert individual pages
1551 * they've allocated into a user vma. Same comments apply.
1552 *
1553 * This function should only be called from a vm_ops->fault handler, and
1554 * in that case the handler should return NULL.
1555 *
1556 * vma cannot be a COW mapping.
1557 *
1558 * As this is called only for pages that do not currently exist, we
1559 * do not need to flush old virtual caches or the TLB.
1560 */
1561 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1562 unsigned long pfn)
1563 {
1564 int ret;
1565 pgprot_t pgprot = vma->vm_page_prot;
1566 /*
1567 * Technically, architectures with pte_special can avoid all these
1568 * restrictions (same for remap_pfn_range). However we would like
1569 * consistency in testing and feature parity among all, so we should
1570 * try to keep these invariants in place for everybody.
1571 */
1572 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1573 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1574 (VM_PFNMAP|VM_MIXEDMAP));
1575 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1576 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1577
1578 if (addr < vma->vm_start || addr >= vma->vm_end)
1579 return -EFAULT;
1580 if (track_pfn_insert(vma, &pgprot, pfn))
1581 return -EINVAL;
1582
1583 ret = insert_pfn(vma, addr, pfn, pgprot);
1584
1585 return ret;
1586 }
1587 EXPORT_SYMBOL(vm_insert_pfn);
1588
1589 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1590 unsigned long pfn)
1591 {
1592 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1593
1594 if (addr < vma->vm_start || addr >= vma->vm_end)
1595 return -EFAULT;
1596
1597 /*
1598 * If we don't have pte special, then we have to use the pfn_valid()
1599 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1600 * refcount the page if pfn_valid is true (hence insert_page rather
1601 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1602 * without pte special, it would there be refcounted as a normal page.
1603 */
1604 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1605 struct page *page;
1606
1607 page = pfn_to_page(pfn);
1608 return insert_page(vma, addr, page, vma->vm_page_prot);
1609 }
1610 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1611 }
1612 EXPORT_SYMBOL(vm_insert_mixed);
1613
1614 /*
1615 * maps a range of physical memory into the requested pages. the old
1616 * mappings are removed. any references to nonexistent pages results
1617 * in null mappings (currently treated as "copy-on-access")
1618 */
1619 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1620 unsigned long addr, unsigned long end,
1621 unsigned long pfn, pgprot_t prot)
1622 {
1623 pte_t *pte;
1624 spinlock_t *ptl;
1625
1626 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1627 if (!pte)
1628 return -ENOMEM;
1629 arch_enter_lazy_mmu_mode();
1630 do {
1631 BUG_ON(!pte_none(*pte));
1632 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1633 pfn++;
1634 } while (pte++, addr += PAGE_SIZE, addr != end);
1635 arch_leave_lazy_mmu_mode();
1636 pte_unmap_unlock(pte - 1, ptl);
1637 return 0;
1638 }
1639
1640 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1641 unsigned long addr, unsigned long end,
1642 unsigned long pfn, pgprot_t prot)
1643 {
1644 pmd_t *pmd;
1645 unsigned long next;
1646
1647 pfn -= addr >> PAGE_SHIFT;
1648 pmd = pmd_alloc(mm, pud, addr);
1649 if (!pmd)
1650 return -ENOMEM;
1651 VM_BUG_ON(pmd_trans_huge(*pmd));
1652 do {
1653 next = pmd_addr_end(addr, end);
1654 if (remap_pte_range(mm, pmd, addr, next,
1655 pfn + (addr >> PAGE_SHIFT), prot))
1656 return -ENOMEM;
1657 } while (pmd++, addr = next, addr != end);
1658 return 0;
1659 }
1660
1661 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1662 unsigned long addr, unsigned long end,
1663 unsigned long pfn, pgprot_t prot)
1664 {
1665 pud_t *pud;
1666 unsigned long next;
1667
1668 pfn -= addr >> PAGE_SHIFT;
1669 pud = pud_alloc(mm, pgd, addr);
1670 if (!pud)
1671 return -ENOMEM;
1672 do {
1673 next = pud_addr_end(addr, end);
1674 if (remap_pmd_range(mm, pud, addr, next,
1675 pfn + (addr >> PAGE_SHIFT), prot))
1676 return -ENOMEM;
1677 } while (pud++, addr = next, addr != end);
1678 return 0;
1679 }
1680
1681 /**
1682 * remap_pfn_range - remap kernel memory to userspace
1683 * @vma: user vma to map to
1684 * @addr: target user address to start at
1685 * @pfn: physical address of kernel memory
1686 * @size: size of map area
1687 * @prot: page protection flags for this mapping
1688 *
1689 * Note: this is only safe if the mm semaphore is held when called.
1690 */
1691 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1692 unsigned long pfn, unsigned long size, pgprot_t prot)
1693 {
1694 pgd_t *pgd;
1695 unsigned long next;
1696 unsigned long end = addr + PAGE_ALIGN(size);
1697 struct mm_struct *mm = vma->vm_mm;
1698 int err;
1699
1700 /*
1701 * Physically remapped pages are special. Tell the
1702 * rest of the world about it:
1703 * VM_IO tells people not to look at these pages
1704 * (accesses can have side effects).
1705 * VM_PFNMAP tells the core MM that the base pages are just
1706 * raw PFN mappings, and do not have a "struct page" associated
1707 * with them.
1708 * VM_DONTEXPAND
1709 * Disable vma merging and expanding with mremap().
1710 * VM_DONTDUMP
1711 * Omit vma from core dump, even when VM_IO turned off.
1712 *
1713 * There's a horrible special case to handle copy-on-write
1714 * behaviour that some programs depend on. We mark the "original"
1715 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1716 * See vm_normal_page() for details.
1717 */
1718 if (is_cow_mapping(vma->vm_flags)) {
1719 if (addr != vma->vm_start || end != vma->vm_end)
1720 return -EINVAL;
1721 vma->vm_pgoff = pfn;
1722 }
1723
1724 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1725 if (err)
1726 return -EINVAL;
1727
1728 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1729
1730 BUG_ON(addr >= end);
1731 pfn -= addr >> PAGE_SHIFT;
1732 pgd = pgd_offset(mm, addr);
1733 flush_cache_range(vma, addr, end);
1734 do {
1735 next = pgd_addr_end(addr, end);
1736 err = remap_pud_range(mm, pgd, addr, next,
1737 pfn + (addr >> PAGE_SHIFT), prot);
1738 if (err)
1739 break;
1740 } while (pgd++, addr = next, addr != end);
1741
1742 if (err)
1743 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1744
1745 return err;
1746 }
1747 EXPORT_SYMBOL(remap_pfn_range);
1748
1749 /**
1750 * vm_iomap_memory - remap memory to userspace
1751 * @vma: user vma to map to
1752 * @start: start of area
1753 * @len: size of area
1754 *
1755 * This is a simplified io_remap_pfn_range() for common driver use. The
1756 * driver just needs to give us the physical memory range to be mapped,
1757 * we'll figure out the rest from the vma information.
1758 *
1759 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1760 * whatever write-combining details or similar.
1761 */
1762 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1763 {
1764 unsigned long vm_len, pfn, pages;
1765
1766 /* Check that the physical memory area passed in looks valid */
1767 if (start + len < start)
1768 return -EINVAL;
1769 /*
1770 * You *really* shouldn't map things that aren't page-aligned,
1771 * but we've historically allowed it because IO memory might
1772 * just have smaller alignment.
1773 */
1774 len += start & ~PAGE_MASK;
1775 pfn = start >> PAGE_SHIFT;
1776 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1777 if (pfn + pages < pfn)
1778 return -EINVAL;
1779
1780 /* We start the mapping 'vm_pgoff' pages into the area */
1781 if (vma->vm_pgoff > pages)
1782 return -EINVAL;
1783 pfn += vma->vm_pgoff;
1784 pages -= vma->vm_pgoff;
1785
1786 /* Can we fit all of the mapping? */
1787 vm_len = vma->vm_end - vma->vm_start;
1788 if (vm_len >> PAGE_SHIFT > pages)
1789 return -EINVAL;
1790
1791 /* Ok, let it rip */
1792 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1793 }
1794 EXPORT_SYMBOL(vm_iomap_memory);
1795
1796 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1797 unsigned long addr, unsigned long end,
1798 pte_fn_t fn, void *data)
1799 {
1800 pte_t *pte;
1801 int err;
1802 pgtable_t token;
1803 spinlock_t *uninitialized_var(ptl);
1804
1805 pte = (mm == &init_mm) ?
1806 pte_alloc_kernel(pmd, addr) :
1807 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1808 if (!pte)
1809 return -ENOMEM;
1810
1811 BUG_ON(pmd_huge(*pmd));
1812
1813 arch_enter_lazy_mmu_mode();
1814
1815 token = pmd_pgtable(*pmd);
1816
1817 do {
1818 err = fn(pte++, token, addr, data);
1819 if (err)
1820 break;
1821 } while (addr += PAGE_SIZE, addr != end);
1822
1823 arch_leave_lazy_mmu_mode();
1824
1825 if (mm != &init_mm)
1826 pte_unmap_unlock(pte-1, ptl);
1827 return err;
1828 }
1829
1830 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1831 unsigned long addr, unsigned long end,
1832 pte_fn_t fn, void *data)
1833 {
1834 pmd_t *pmd;
1835 unsigned long next;
1836 int err;
1837
1838 BUG_ON(pud_huge(*pud));
1839
1840 pmd = pmd_alloc(mm, pud, addr);
1841 if (!pmd)
1842 return -ENOMEM;
1843 do {
1844 next = pmd_addr_end(addr, end);
1845 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1846 if (err)
1847 break;
1848 } while (pmd++, addr = next, addr != end);
1849 return err;
1850 }
1851
1852 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1853 unsigned long addr, unsigned long end,
1854 pte_fn_t fn, void *data)
1855 {
1856 pud_t *pud;
1857 unsigned long next;
1858 int err;
1859
1860 pud = pud_alloc(mm, pgd, addr);
1861 if (!pud)
1862 return -ENOMEM;
1863 do {
1864 next = pud_addr_end(addr, end);
1865 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1866 if (err)
1867 break;
1868 } while (pud++, addr = next, addr != end);
1869 return err;
1870 }
1871
1872 /*
1873 * Scan a region of virtual memory, filling in page tables as necessary
1874 * and calling a provided function on each leaf page table.
1875 */
1876 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1877 unsigned long size, pte_fn_t fn, void *data)
1878 {
1879 pgd_t *pgd;
1880 unsigned long next;
1881 unsigned long end = addr + size;
1882 int err;
1883
1884 BUG_ON(addr >= end);
1885 pgd = pgd_offset(mm, addr);
1886 do {
1887 next = pgd_addr_end(addr, end);
1888 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1889 if (err)
1890 break;
1891 } while (pgd++, addr = next, addr != end);
1892
1893 return err;
1894 }
1895 EXPORT_SYMBOL_GPL(apply_to_page_range);
1896
1897 /*
1898 * handle_pte_fault chooses page fault handler according to an entry which was
1899 * read non-atomically. Before making any commitment, on those architectures
1900 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1901 * parts, do_swap_page must check under lock before unmapping the pte and
1902 * proceeding (but do_wp_page is only called after already making such a check;
1903 * and do_anonymous_page can safely check later on).
1904 */
1905 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1906 pte_t *page_table, pte_t orig_pte)
1907 {
1908 int same = 1;
1909 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1910 if (sizeof(pte_t) > sizeof(unsigned long)) {
1911 spinlock_t *ptl = pte_lockptr(mm, pmd);
1912 spin_lock(ptl);
1913 same = pte_same(*page_table, orig_pte);
1914 spin_unlock(ptl);
1915 }
1916 #endif
1917 pte_unmap(page_table);
1918 return same;
1919 }
1920
1921 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1922 {
1923 debug_dma_assert_idle(src);
1924
1925 /*
1926 * If the source page was a PFN mapping, we don't have
1927 * a "struct page" for it. We do a best-effort copy by
1928 * just copying from the original user address. If that
1929 * fails, we just zero-fill it. Live with it.
1930 */
1931 if (unlikely(!src)) {
1932 void *kaddr = kmap_atomic(dst);
1933 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1934
1935 /*
1936 * This really shouldn't fail, because the page is there
1937 * in the page tables. But it might just be unreadable,
1938 * in which case we just give up and fill the result with
1939 * zeroes.
1940 */
1941 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1942 clear_page(kaddr);
1943 kunmap_atomic(kaddr);
1944 flush_dcache_page(dst);
1945 } else
1946 copy_user_highpage(dst, src, va, vma);
1947 }
1948
1949 /*
1950 * Notify the address space that the page is about to become writable so that
1951 * it can prohibit this or wait for the page to get into an appropriate state.
1952 *
1953 * We do this without the lock held, so that it can sleep if it needs to.
1954 */
1955 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1956 unsigned long address)
1957 {
1958 struct vm_fault vmf;
1959 int ret;
1960
1961 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
1962 vmf.pgoff = page->index;
1963 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
1964 vmf.page = page;
1965
1966 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
1967 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
1968 return ret;
1969 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
1970 lock_page(page);
1971 if (!page->mapping) {
1972 unlock_page(page);
1973 return 0; /* retry */
1974 }
1975 ret |= VM_FAULT_LOCKED;
1976 } else
1977 VM_BUG_ON_PAGE(!PageLocked(page), page);
1978 return ret;
1979 }
1980
1981 /*
1982 * This routine handles present pages, when users try to write
1983 * to a shared page. It is done by copying the page to a new address
1984 * and decrementing the shared-page counter for the old page.
1985 *
1986 * Note that this routine assumes that the protection checks have been
1987 * done by the caller (the low-level page fault routine in most cases).
1988 * Thus we can safely just mark it writable once we've done any necessary
1989 * COW.
1990 *
1991 * We also mark the page dirty at this point even though the page will
1992 * change only once the write actually happens. This avoids a few races,
1993 * and potentially makes it more efficient.
1994 *
1995 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1996 * but allow concurrent faults), with pte both mapped and locked.
1997 * We return with mmap_sem still held, but pte unmapped and unlocked.
1998 */
1999 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2000 unsigned long address, pte_t *page_table, pmd_t *pmd,
2001 spinlock_t *ptl, pte_t orig_pte)
2002 __releases(ptl)
2003 {
2004 struct page *old_page, *new_page = NULL;
2005 pte_t entry;
2006 int ret = 0;
2007 int page_mkwrite = 0;
2008 bool dirty_shared = false;
2009 unsigned long mmun_start = 0; /* For mmu_notifiers */
2010 unsigned long mmun_end = 0; /* For mmu_notifiers */
2011 struct mem_cgroup *memcg;
2012
2013 old_page = vm_normal_page(vma, address, orig_pte);
2014 if (!old_page) {
2015 /*
2016 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2017 * VM_PFNMAP VMA.
2018 *
2019 * We should not cow pages in a shared writeable mapping.
2020 * Just mark the pages writable as we can't do any dirty
2021 * accounting on raw pfn maps.
2022 */
2023 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2024 (VM_WRITE|VM_SHARED))
2025 goto reuse;
2026 goto gotten;
2027 }
2028
2029 /*
2030 * Take out anonymous pages first, anonymous shared vmas are
2031 * not dirty accountable.
2032 */
2033 if (PageAnon(old_page) && !PageKsm(old_page)) {
2034 if (!trylock_page(old_page)) {
2035 page_cache_get(old_page);
2036 pte_unmap_unlock(page_table, ptl);
2037 lock_page(old_page);
2038 page_table = pte_offset_map_lock(mm, pmd, address,
2039 &ptl);
2040 if (!pte_same(*page_table, orig_pte)) {
2041 unlock_page(old_page);
2042 goto unlock;
2043 }
2044 page_cache_release(old_page);
2045 }
2046 if (reuse_swap_page(old_page)) {
2047 /*
2048 * The page is all ours. Move it to our anon_vma so
2049 * the rmap code will not search our parent or siblings.
2050 * Protected against the rmap code by the page lock.
2051 */
2052 page_move_anon_rmap(old_page, vma, address);
2053 unlock_page(old_page);
2054 goto reuse;
2055 }
2056 unlock_page(old_page);
2057 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2058 (VM_WRITE|VM_SHARED))) {
2059 page_cache_get(old_page);
2060 /*
2061 * Only catch write-faults on shared writable pages,
2062 * read-only shared pages can get COWed by
2063 * get_user_pages(.write=1, .force=1).
2064 */
2065 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2066 int tmp;
2067
2068 pte_unmap_unlock(page_table, ptl);
2069 tmp = do_page_mkwrite(vma, old_page, address);
2070 if (unlikely(!tmp || (tmp &
2071 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2072 page_cache_release(old_page);
2073 return tmp;
2074 }
2075 /*
2076 * Since we dropped the lock we need to revalidate
2077 * the PTE as someone else may have changed it. If
2078 * they did, we just return, as we can count on the
2079 * MMU to tell us if they didn't also make it writable.
2080 */
2081 page_table = pte_offset_map_lock(mm, pmd, address,
2082 &ptl);
2083 if (!pte_same(*page_table, orig_pte)) {
2084 unlock_page(old_page);
2085 goto unlock;
2086 }
2087 page_mkwrite = 1;
2088 }
2089
2090 dirty_shared = true;
2091
2092 reuse:
2093 /*
2094 * Clear the pages cpupid information as the existing
2095 * information potentially belongs to a now completely
2096 * unrelated process.
2097 */
2098 if (old_page)
2099 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2100
2101 flush_cache_page(vma, address, pte_pfn(orig_pte));
2102 entry = pte_mkyoung(orig_pte);
2103 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2104 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2105 update_mmu_cache(vma, address, page_table);
2106 pte_unmap_unlock(page_table, ptl);
2107 ret |= VM_FAULT_WRITE;
2108
2109 if (dirty_shared) {
2110 struct address_space *mapping;
2111 int dirtied;
2112
2113 if (!page_mkwrite)
2114 lock_page(old_page);
2115
2116 dirtied = set_page_dirty(old_page);
2117 VM_BUG_ON_PAGE(PageAnon(old_page), old_page);
2118 mapping = old_page->mapping;
2119 unlock_page(old_page);
2120 page_cache_release(old_page);
2121
2122 if ((dirtied || page_mkwrite) && mapping) {
2123 /*
2124 * Some device drivers do not set page.mapping
2125 * but still dirty their pages
2126 */
2127 balance_dirty_pages_ratelimited(mapping);
2128 }
2129
2130 if (!page_mkwrite)
2131 file_update_time(vma->vm_file);
2132 }
2133
2134 return ret;
2135 }
2136
2137 /*
2138 * Ok, we need to copy. Oh, well..
2139 */
2140 page_cache_get(old_page);
2141 gotten:
2142 pte_unmap_unlock(page_table, ptl);
2143
2144 if (unlikely(anon_vma_prepare(vma)))
2145 goto oom;
2146
2147 if (is_zero_pfn(pte_pfn(orig_pte))) {
2148 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2149 if (!new_page)
2150 goto oom;
2151 } else {
2152 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2153 if (!new_page)
2154 goto oom;
2155 cow_user_page(new_page, old_page, address, vma);
2156 }
2157 __SetPageUptodate(new_page);
2158
2159 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg))
2160 goto oom_free_new;
2161
2162 mmun_start = address & PAGE_MASK;
2163 mmun_end = mmun_start + PAGE_SIZE;
2164 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2165
2166 /*
2167 * Re-check the pte - we dropped the lock
2168 */
2169 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2170 if (likely(pte_same(*page_table, orig_pte))) {
2171 if (old_page) {
2172 if (!PageAnon(old_page)) {
2173 dec_mm_counter_fast(mm, MM_FILEPAGES);
2174 inc_mm_counter_fast(mm, MM_ANONPAGES);
2175 }
2176 } else
2177 inc_mm_counter_fast(mm, MM_ANONPAGES);
2178 flush_cache_page(vma, address, pte_pfn(orig_pte));
2179 entry = mk_pte(new_page, vma->vm_page_prot);
2180 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2181 /*
2182 * Clear the pte entry and flush it first, before updating the
2183 * pte with the new entry. This will avoid a race condition
2184 * seen in the presence of one thread doing SMC and another
2185 * thread doing COW.
2186 */
2187 ptep_clear_flush_notify(vma, address, page_table);
2188 page_add_new_anon_rmap(new_page, vma, address);
2189 mem_cgroup_commit_charge(new_page, memcg, false);
2190 lru_cache_add_active_or_unevictable(new_page, vma);
2191 /*
2192 * We call the notify macro here because, when using secondary
2193 * mmu page tables (such as kvm shadow page tables), we want the
2194 * new page to be mapped directly into the secondary page table.
2195 */
2196 set_pte_at_notify(mm, address, page_table, entry);
2197 update_mmu_cache(vma, address, page_table);
2198 if (old_page) {
2199 /*
2200 * Only after switching the pte to the new page may
2201 * we remove the mapcount here. Otherwise another
2202 * process may come and find the rmap count decremented
2203 * before the pte is switched to the new page, and
2204 * "reuse" the old page writing into it while our pte
2205 * here still points into it and can be read by other
2206 * threads.
2207 *
2208 * The critical issue is to order this
2209 * page_remove_rmap with the ptp_clear_flush above.
2210 * Those stores are ordered by (if nothing else,)
2211 * the barrier present in the atomic_add_negative
2212 * in page_remove_rmap.
2213 *
2214 * Then the TLB flush in ptep_clear_flush ensures that
2215 * no process can access the old page before the
2216 * decremented mapcount is visible. And the old page
2217 * cannot be reused until after the decremented
2218 * mapcount is visible. So transitively, TLBs to
2219 * old page will be flushed before it can be reused.
2220 */
2221 page_remove_rmap(old_page);
2222 }
2223
2224 /* Free the old page.. */
2225 new_page = old_page;
2226 ret |= VM_FAULT_WRITE;
2227 } else
2228 mem_cgroup_cancel_charge(new_page, memcg);
2229
2230 if (new_page)
2231 page_cache_release(new_page);
2232 unlock:
2233 pte_unmap_unlock(page_table, ptl);
2234 if (mmun_end > mmun_start)
2235 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2236 if (old_page) {
2237 /*
2238 * Don't let another task, with possibly unlocked vma,
2239 * keep the mlocked page.
2240 */
2241 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2242 lock_page(old_page); /* LRU manipulation */
2243 munlock_vma_page(old_page);
2244 unlock_page(old_page);
2245 }
2246 page_cache_release(old_page);
2247 }
2248 return ret;
2249 oom_free_new:
2250 page_cache_release(new_page);
2251 oom:
2252 if (old_page)
2253 page_cache_release(old_page);
2254 return VM_FAULT_OOM;
2255 }
2256
2257 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2258 unsigned long start_addr, unsigned long end_addr,
2259 struct zap_details *details)
2260 {
2261 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2262 }
2263
2264 static inline void unmap_mapping_range_tree(struct rb_root *root,
2265 struct zap_details *details)
2266 {
2267 struct vm_area_struct *vma;
2268 pgoff_t vba, vea, zba, zea;
2269
2270 vma_interval_tree_foreach(vma, root,
2271 details->first_index, details->last_index) {
2272
2273 vba = vma->vm_pgoff;
2274 vea = vba + vma_pages(vma) - 1;
2275 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2276 zba = details->first_index;
2277 if (zba < vba)
2278 zba = vba;
2279 zea = details->last_index;
2280 if (zea > vea)
2281 zea = vea;
2282
2283 unmap_mapping_range_vma(vma,
2284 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2285 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2286 details);
2287 }
2288 }
2289
2290 /**
2291 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2292 * address_space corresponding to the specified page range in the underlying
2293 * file.
2294 *
2295 * @mapping: the address space containing mmaps to be unmapped.
2296 * @holebegin: byte in first page to unmap, relative to the start of
2297 * the underlying file. This will be rounded down to a PAGE_SIZE
2298 * boundary. Note that this is different from truncate_pagecache(), which
2299 * must keep the partial page. In contrast, we must get rid of
2300 * partial pages.
2301 * @holelen: size of prospective hole in bytes. This will be rounded
2302 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2303 * end of the file.
2304 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2305 * but 0 when invalidating pagecache, don't throw away private data.
2306 */
2307 void unmap_mapping_range(struct address_space *mapping,
2308 loff_t const holebegin, loff_t const holelen, int even_cows)
2309 {
2310 struct zap_details details;
2311 pgoff_t hba = holebegin >> PAGE_SHIFT;
2312 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2313
2314 /* Check for overflow. */
2315 if (sizeof(holelen) > sizeof(hlen)) {
2316 long long holeend =
2317 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2318 if (holeend & ~(long long)ULONG_MAX)
2319 hlen = ULONG_MAX - hba + 1;
2320 }
2321
2322 details.check_mapping = even_cows? NULL: mapping;
2323 details.first_index = hba;
2324 details.last_index = hba + hlen - 1;
2325 if (details.last_index < details.first_index)
2326 details.last_index = ULONG_MAX;
2327
2328
2329 i_mmap_lock_write(mapping);
2330 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2331 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2332 i_mmap_unlock_write(mapping);
2333 }
2334 EXPORT_SYMBOL(unmap_mapping_range);
2335
2336 /*
2337 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2338 * but allow concurrent faults), and pte mapped but not yet locked.
2339 * We return with pte unmapped and unlocked.
2340 *
2341 * We return with the mmap_sem locked or unlocked in the same cases
2342 * as does filemap_fault().
2343 */
2344 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2345 unsigned long address, pte_t *page_table, pmd_t *pmd,
2346 unsigned int flags, pte_t orig_pte)
2347 {
2348 spinlock_t *ptl;
2349 struct page *page, *swapcache;
2350 struct mem_cgroup *memcg;
2351 swp_entry_t entry;
2352 pte_t pte;
2353 int locked;
2354 int exclusive = 0;
2355 int ret = 0;
2356
2357 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2358 goto out;
2359
2360 entry = pte_to_swp_entry(orig_pte);
2361 if (unlikely(non_swap_entry(entry))) {
2362 if (is_migration_entry(entry)) {
2363 migration_entry_wait(mm, pmd, address);
2364 } else if (is_hwpoison_entry(entry)) {
2365 ret = VM_FAULT_HWPOISON;
2366 } else {
2367 print_bad_pte(vma, address, orig_pte, NULL);
2368 ret = VM_FAULT_SIGBUS;
2369 }
2370 goto out;
2371 }
2372 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2373 page = lookup_swap_cache(entry);
2374 if (!page) {
2375 page = swapin_readahead(entry,
2376 GFP_HIGHUSER_MOVABLE, vma, address);
2377 if (!page) {
2378 /*
2379 * Back out if somebody else faulted in this pte
2380 * while we released the pte lock.
2381 */
2382 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2383 if (likely(pte_same(*page_table, orig_pte)))
2384 ret = VM_FAULT_OOM;
2385 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2386 goto unlock;
2387 }
2388
2389 /* Had to read the page from swap area: Major fault */
2390 ret = VM_FAULT_MAJOR;
2391 count_vm_event(PGMAJFAULT);
2392 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2393 } else if (PageHWPoison(page)) {
2394 /*
2395 * hwpoisoned dirty swapcache pages are kept for killing
2396 * owner processes (which may be unknown at hwpoison time)
2397 */
2398 ret = VM_FAULT_HWPOISON;
2399 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2400 swapcache = page;
2401 goto out_release;
2402 }
2403
2404 swapcache = page;
2405 locked = lock_page_or_retry(page, mm, flags);
2406
2407 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2408 if (!locked) {
2409 ret |= VM_FAULT_RETRY;
2410 goto out_release;
2411 }
2412
2413 /*
2414 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2415 * release the swapcache from under us. The page pin, and pte_same
2416 * test below, are not enough to exclude that. Even if it is still
2417 * swapcache, we need to check that the page's swap has not changed.
2418 */
2419 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2420 goto out_page;
2421
2422 page = ksm_might_need_to_copy(page, vma, address);
2423 if (unlikely(!page)) {
2424 ret = VM_FAULT_OOM;
2425 page = swapcache;
2426 goto out_page;
2427 }
2428
2429 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) {
2430 ret = VM_FAULT_OOM;
2431 goto out_page;
2432 }
2433
2434 /*
2435 * Back out if somebody else already faulted in this pte.
2436 */
2437 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2438 if (unlikely(!pte_same(*page_table, orig_pte)))
2439 goto out_nomap;
2440
2441 if (unlikely(!PageUptodate(page))) {
2442 ret = VM_FAULT_SIGBUS;
2443 goto out_nomap;
2444 }
2445
2446 /*
2447 * The page isn't present yet, go ahead with the fault.
2448 *
2449 * Be careful about the sequence of operations here.
2450 * To get its accounting right, reuse_swap_page() must be called
2451 * while the page is counted on swap but not yet in mapcount i.e.
2452 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2453 * must be called after the swap_free(), or it will never succeed.
2454 */
2455
2456 inc_mm_counter_fast(mm, MM_ANONPAGES);
2457 dec_mm_counter_fast(mm, MM_SWAPENTS);
2458 pte = mk_pte(page, vma->vm_page_prot);
2459 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2460 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2461 flags &= ~FAULT_FLAG_WRITE;
2462 ret |= VM_FAULT_WRITE;
2463 exclusive = 1;
2464 }
2465 flush_icache_page(vma, page);
2466 if (pte_swp_soft_dirty(orig_pte))
2467 pte = pte_mksoft_dirty(pte);
2468 set_pte_at(mm, address, page_table, pte);
2469 if (page == swapcache) {
2470 do_page_add_anon_rmap(page, vma, address, exclusive);
2471 mem_cgroup_commit_charge(page, memcg, true);
2472 } else { /* ksm created a completely new copy */
2473 page_add_new_anon_rmap(page, vma, address);
2474 mem_cgroup_commit_charge(page, memcg, false);
2475 lru_cache_add_active_or_unevictable(page, vma);
2476 }
2477
2478 swap_free(entry);
2479 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2480 try_to_free_swap(page);
2481 unlock_page(page);
2482 if (page != swapcache) {
2483 /*
2484 * Hold the lock to avoid the swap entry to be reused
2485 * until we take the PT lock for the pte_same() check
2486 * (to avoid false positives from pte_same). For
2487 * further safety release the lock after the swap_free
2488 * so that the swap count won't change under a
2489 * parallel locked swapcache.
2490 */
2491 unlock_page(swapcache);
2492 page_cache_release(swapcache);
2493 }
2494
2495 if (flags & FAULT_FLAG_WRITE) {
2496 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2497 if (ret & VM_FAULT_ERROR)
2498 ret &= VM_FAULT_ERROR;
2499 goto out;
2500 }
2501
2502 /* No need to invalidate - it was non-present before */
2503 update_mmu_cache(vma, address, page_table);
2504 unlock:
2505 pte_unmap_unlock(page_table, ptl);
2506 out:
2507 return ret;
2508 out_nomap:
2509 mem_cgroup_cancel_charge(page, memcg);
2510 pte_unmap_unlock(page_table, ptl);
2511 out_page:
2512 unlock_page(page);
2513 out_release:
2514 page_cache_release(page);
2515 if (page != swapcache) {
2516 unlock_page(swapcache);
2517 page_cache_release(swapcache);
2518 }
2519 return ret;
2520 }
2521
2522 /*
2523 * This is like a special single-page "expand_{down|up}wards()",
2524 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2525 * doesn't hit another vma.
2526 */
2527 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2528 {
2529 address &= PAGE_MASK;
2530 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2531 struct vm_area_struct *prev = vma->vm_prev;
2532
2533 /*
2534 * Is there a mapping abutting this one below?
2535 *
2536 * That's only ok if it's the same stack mapping
2537 * that has gotten split..
2538 */
2539 if (prev && prev->vm_end == address)
2540 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2541
2542 return expand_downwards(vma, address - PAGE_SIZE);
2543 }
2544 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2545 struct vm_area_struct *next = vma->vm_next;
2546
2547 /* As VM_GROWSDOWN but s/below/above/ */
2548 if (next && next->vm_start == address + PAGE_SIZE)
2549 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2550
2551 return expand_upwards(vma, address + PAGE_SIZE);
2552 }
2553 return 0;
2554 }
2555
2556 /*
2557 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2558 * but allow concurrent faults), and pte mapped but not yet locked.
2559 * We return with mmap_sem still held, but pte unmapped and unlocked.
2560 */
2561 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2562 unsigned long address, pte_t *page_table, pmd_t *pmd,
2563 unsigned int flags)
2564 {
2565 struct mem_cgroup *memcg;
2566 struct page *page;
2567 spinlock_t *ptl;
2568 pte_t entry;
2569
2570 pte_unmap(page_table);
2571
2572 /* Check if we need to add a guard page to the stack */
2573 if (check_stack_guard_page(vma, address) < 0)
2574 return VM_FAULT_SIGSEGV;
2575
2576 /* Use the zero-page for reads */
2577 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
2578 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2579 vma->vm_page_prot));
2580 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2581 if (!pte_none(*page_table))
2582 goto unlock;
2583 goto setpte;
2584 }
2585
2586 /* Allocate our own private page. */
2587 if (unlikely(anon_vma_prepare(vma)))
2588 goto oom;
2589 page = alloc_zeroed_user_highpage_movable(vma, address);
2590 if (!page)
2591 goto oom;
2592 /*
2593 * The memory barrier inside __SetPageUptodate makes sure that
2594 * preceeding stores to the page contents become visible before
2595 * the set_pte_at() write.
2596 */
2597 __SetPageUptodate(page);
2598
2599 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg))
2600 goto oom_free_page;
2601
2602 entry = mk_pte(page, vma->vm_page_prot);
2603 if (vma->vm_flags & VM_WRITE)
2604 entry = pte_mkwrite(pte_mkdirty(entry));
2605
2606 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2607 if (!pte_none(*page_table))
2608 goto release;
2609
2610 inc_mm_counter_fast(mm, MM_ANONPAGES);
2611 page_add_new_anon_rmap(page, vma, address);
2612 mem_cgroup_commit_charge(page, memcg, false);
2613 lru_cache_add_active_or_unevictable(page, vma);
2614 setpte:
2615 set_pte_at(mm, address, page_table, entry);
2616
2617 /* No need to invalidate - it was non-present before */
2618 update_mmu_cache(vma, address, page_table);
2619 unlock:
2620 pte_unmap_unlock(page_table, ptl);
2621 return 0;
2622 release:
2623 mem_cgroup_cancel_charge(page, memcg);
2624 page_cache_release(page);
2625 goto unlock;
2626 oom_free_page:
2627 page_cache_release(page);
2628 oom:
2629 return VM_FAULT_OOM;
2630 }
2631
2632 /*
2633 * The mmap_sem must have been held on entry, and may have been
2634 * released depending on flags and vma->vm_ops->fault() return value.
2635 * See filemap_fault() and __lock_page_retry().
2636 */
2637 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2638 pgoff_t pgoff, unsigned int flags, struct page **page)
2639 {
2640 struct vm_fault vmf;
2641 int ret;
2642
2643 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2644 vmf.pgoff = pgoff;
2645 vmf.flags = flags;
2646 vmf.page = NULL;
2647
2648 ret = vma->vm_ops->fault(vma, &vmf);
2649 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2650 return ret;
2651
2652 if (unlikely(PageHWPoison(vmf.page))) {
2653 if (ret & VM_FAULT_LOCKED)
2654 unlock_page(vmf.page);
2655 page_cache_release(vmf.page);
2656 return VM_FAULT_HWPOISON;
2657 }
2658
2659 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2660 lock_page(vmf.page);
2661 else
2662 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2663
2664 *page = vmf.page;
2665 return ret;
2666 }
2667
2668 /**
2669 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2670 *
2671 * @vma: virtual memory area
2672 * @address: user virtual address
2673 * @page: page to map
2674 * @pte: pointer to target page table entry
2675 * @write: true, if new entry is writable
2676 * @anon: true, if it's anonymous page
2677 *
2678 * Caller must hold page table lock relevant for @pte.
2679 *
2680 * Target users are page handler itself and implementations of
2681 * vm_ops->map_pages.
2682 */
2683 void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2684 struct page *page, pte_t *pte, bool write, bool anon)
2685 {
2686 pte_t entry;
2687
2688 flush_icache_page(vma, page);
2689 entry = mk_pte(page, vma->vm_page_prot);
2690 if (write)
2691 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2692 if (anon) {
2693 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2694 page_add_new_anon_rmap(page, vma, address);
2695 } else {
2696 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2697 page_add_file_rmap(page);
2698 }
2699 set_pte_at(vma->vm_mm, address, pte, entry);
2700
2701 /* no need to invalidate: a not-present page won't be cached */
2702 update_mmu_cache(vma, address, pte);
2703 }
2704
2705 static unsigned long fault_around_bytes __read_mostly =
2706 rounddown_pow_of_two(65536);
2707
2708 #ifdef CONFIG_DEBUG_FS
2709 static int fault_around_bytes_get(void *data, u64 *val)
2710 {
2711 *val = fault_around_bytes;
2712 return 0;
2713 }
2714
2715 /*
2716 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2717 * rounded down to nearest page order. It's what do_fault_around() expects to
2718 * see.
2719 */
2720 static int fault_around_bytes_set(void *data, u64 val)
2721 {
2722 if (val / PAGE_SIZE > PTRS_PER_PTE)
2723 return -EINVAL;
2724 if (val > PAGE_SIZE)
2725 fault_around_bytes = rounddown_pow_of_two(val);
2726 else
2727 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2728 return 0;
2729 }
2730 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2731 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2732
2733 static int __init fault_around_debugfs(void)
2734 {
2735 void *ret;
2736
2737 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2738 &fault_around_bytes_fops);
2739 if (!ret)
2740 pr_warn("Failed to create fault_around_bytes in debugfs");
2741 return 0;
2742 }
2743 late_initcall(fault_around_debugfs);
2744 #endif
2745
2746 /*
2747 * do_fault_around() tries to map few pages around the fault address. The hope
2748 * is that the pages will be needed soon and this will lower the number of
2749 * faults to handle.
2750 *
2751 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2752 * not ready to be mapped: not up-to-date, locked, etc.
2753 *
2754 * This function is called with the page table lock taken. In the split ptlock
2755 * case the page table lock only protects only those entries which belong to
2756 * the page table corresponding to the fault address.
2757 *
2758 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2759 * only once.
2760 *
2761 * fault_around_pages() defines how many pages we'll try to map.
2762 * do_fault_around() expects it to return a power of two less than or equal to
2763 * PTRS_PER_PTE.
2764 *
2765 * The virtual address of the area that we map is naturally aligned to the
2766 * fault_around_pages() value (and therefore to page order). This way it's
2767 * easier to guarantee that we don't cross page table boundaries.
2768 */
2769 static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2770 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2771 {
2772 unsigned long start_addr, nr_pages, mask;
2773 pgoff_t max_pgoff;
2774 struct vm_fault vmf;
2775 int off;
2776
2777 nr_pages = ACCESS_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2778 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2779
2780 start_addr = max(address & mask, vma->vm_start);
2781 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2782 pte -= off;
2783 pgoff -= off;
2784
2785 /*
2786 * max_pgoff is either end of page table or end of vma
2787 * or fault_around_pages() from pgoff, depending what is nearest.
2788 */
2789 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2790 PTRS_PER_PTE - 1;
2791 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2792 pgoff + nr_pages - 1);
2793
2794 /* Check if it makes any sense to call ->map_pages */
2795 while (!pte_none(*pte)) {
2796 if (++pgoff > max_pgoff)
2797 return;
2798 start_addr += PAGE_SIZE;
2799 if (start_addr >= vma->vm_end)
2800 return;
2801 pte++;
2802 }
2803
2804 vmf.virtual_address = (void __user *) start_addr;
2805 vmf.pte = pte;
2806 vmf.pgoff = pgoff;
2807 vmf.max_pgoff = max_pgoff;
2808 vmf.flags = flags;
2809 vma->vm_ops->map_pages(vma, &vmf);
2810 }
2811
2812 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2813 unsigned long address, pmd_t *pmd,
2814 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2815 {
2816 struct page *fault_page;
2817 spinlock_t *ptl;
2818 pte_t *pte;
2819 int ret = 0;
2820
2821 /*
2822 * Let's call ->map_pages() first and use ->fault() as fallback
2823 * if page by the offset is not ready to be mapped (cold cache or
2824 * something).
2825 */
2826 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
2827 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2828 do_fault_around(vma, address, pte, pgoff, flags);
2829 if (!pte_same(*pte, orig_pte))
2830 goto unlock_out;
2831 pte_unmap_unlock(pte, ptl);
2832 }
2833
2834 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2835 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2836 return ret;
2837
2838 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2839 if (unlikely(!pte_same(*pte, orig_pte))) {
2840 pte_unmap_unlock(pte, ptl);
2841 unlock_page(fault_page);
2842 page_cache_release(fault_page);
2843 return ret;
2844 }
2845 do_set_pte(vma, address, fault_page, pte, false, false);
2846 unlock_page(fault_page);
2847 unlock_out:
2848 pte_unmap_unlock(pte, ptl);
2849 return ret;
2850 }
2851
2852 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2853 unsigned long address, pmd_t *pmd,
2854 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2855 {
2856 struct page *fault_page, *new_page;
2857 struct mem_cgroup *memcg;
2858 spinlock_t *ptl;
2859 pte_t *pte;
2860 int ret;
2861
2862 if (unlikely(anon_vma_prepare(vma)))
2863 return VM_FAULT_OOM;
2864
2865 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2866 if (!new_page)
2867 return VM_FAULT_OOM;
2868
2869 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) {
2870 page_cache_release(new_page);
2871 return VM_FAULT_OOM;
2872 }
2873
2874 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2875 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2876 goto uncharge_out;
2877
2878 copy_user_highpage(new_page, fault_page, address, vma);
2879 __SetPageUptodate(new_page);
2880
2881 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2882 if (unlikely(!pte_same(*pte, orig_pte))) {
2883 pte_unmap_unlock(pte, ptl);
2884 unlock_page(fault_page);
2885 page_cache_release(fault_page);
2886 goto uncharge_out;
2887 }
2888 do_set_pte(vma, address, new_page, pte, true, true);
2889 mem_cgroup_commit_charge(new_page, memcg, false);
2890 lru_cache_add_active_or_unevictable(new_page, vma);
2891 pte_unmap_unlock(pte, ptl);
2892 unlock_page(fault_page);
2893 page_cache_release(fault_page);
2894 return ret;
2895 uncharge_out:
2896 mem_cgroup_cancel_charge(new_page, memcg);
2897 page_cache_release(new_page);
2898 return ret;
2899 }
2900
2901 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2902 unsigned long address, pmd_t *pmd,
2903 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2904 {
2905 struct page *fault_page;
2906 struct address_space *mapping;
2907 spinlock_t *ptl;
2908 pte_t *pte;
2909 int dirtied = 0;
2910 int ret, tmp;
2911
2912 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2913 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2914 return ret;
2915
2916 /*
2917 * Check if the backing address space wants to know that the page is
2918 * about to become writable
2919 */
2920 if (vma->vm_ops->page_mkwrite) {
2921 unlock_page(fault_page);
2922 tmp = do_page_mkwrite(vma, fault_page, address);
2923 if (unlikely(!tmp ||
2924 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2925 page_cache_release(fault_page);
2926 return tmp;
2927 }
2928 }
2929
2930 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2931 if (unlikely(!pte_same(*pte, orig_pte))) {
2932 pte_unmap_unlock(pte, ptl);
2933 unlock_page(fault_page);
2934 page_cache_release(fault_page);
2935 return ret;
2936 }
2937 do_set_pte(vma, address, fault_page, pte, true, false);
2938 pte_unmap_unlock(pte, ptl);
2939
2940 if (set_page_dirty(fault_page))
2941 dirtied = 1;
2942 /*
2943 * Take a local copy of the address_space - page.mapping may be zeroed
2944 * by truncate after unlock_page(). The address_space itself remains
2945 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2946 * release semantics to prevent the compiler from undoing this copying.
2947 */
2948 mapping = fault_page->mapping;
2949 unlock_page(fault_page);
2950 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
2951 /*
2952 * Some device drivers do not set page.mapping but still
2953 * dirty their pages
2954 */
2955 balance_dirty_pages_ratelimited(mapping);
2956 }
2957
2958 if (!vma->vm_ops->page_mkwrite)
2959 file_update_time(vma->vm_file);
2960
2961 return ret;
2962 }
2963
2964 /*
2965 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2966 * but allow concurrent faults).
2967 * The mmap_sem may have been released depending on flags and our
2968 * return value. See filemap_fault() and __lock_page_or_retry().
2969 */
2970 static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2971 unsigned long address, pte_t *page_table, pmd_t *pmd,
2972 unsigned int flags, pte_t orig_pte)
2973 {
2974 pgoff_t pgoff = (((address & PAGE_MASK)
2975 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2976
2977 pte_unmap(page_table);
2978 if (!(flags & FAULT_FLAG_WRITE))
2979 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
2980 orig_pte);
2981 if (!(vma->vm_flags & VM_SHARED))
2982 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
2983 orig_pte);
2984 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2985 }
2986
2987 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
2988 unsigned long addr, int page_nid,
2989 int *flags)
2990 {
2991 get_page(page);
2992
2993 count_vm_numa_event(NUMA_HINT_FAULTS);
2994 if (page_nid == numa_node_id()) {
2995 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
2996 *flags |= TNF_FAULT_LOCAL;
2997 }
2998
2999 return mpol_misplaced(page, vma, addr);
3000 }
3001
3002 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3003 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3004 {
3005 struct page *page = NULL;
3006 spinlock_t *ptl;
3007 int page_nid = -1;
3008 int last_cpupid;
3009 int target_nid;
3010 bool migrated = false;
3011 int flags = 0;
3012
3013 /*
3014 * The "pte" at this point cannot be used safely without
3015 * validation through pte_unmap_same(). It's of NUMA type but
3016 * the pfn may be screwed if the read is non atomic.
3017 *
3018 * ptep_modify_prot_start is not called as this is clearing
3019 * the _PAGE_NUMA bit and it is not really expected that there
3020 * would be concurrent hardware modifications to the PTE.
3021 */
3022 ptl = pte_lockptr(mm, pmd);
3023 spin_lock(ptl);
3024 if (unlikely(!pte_same(*ptep, pte))) {
3025 pte_unmap_unlock(ptep, ptl);
3026 goto out;
3027 }
3028
3029 pte = pte_mknonnuma(pte);
3030 set_pte_at(mm, addr, ptep, pte);
3031 update_mmu_cache(vma, addr, ptep);
3032
3033 page = vm_normal_page(vma, addr, pte);
3034 if (!page) {
3035 pte_unmap_unlock(ptep, ptl);
3036 return 0;
3037 }
3038 BUG_ON(is_zero_pfn(page_to_pfn(page)));
3039
3040 /*
3041 * Avoid grouping on DSO/COW pages in specific and RO pages
3042 * in general, RO pages shouldn't hurt as much anyway since
3043 * they can be in shared cache state.
3044 */
3045 if (!pte_write(pte))
3046 flags |= TNF_NO_GROUP;
3047
3048 /*
3049 * Flag if the page is shared between multiple address spaces. This
3050 * is later used when determining whether to group tasks together
3051 */
3052 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3053 flags |= TNF_SHARED;
3054
3055 last_cpupid = page_cpupid_last(page);
3056 page_nid = page_to_nid(page);
3057 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3058 pte_unmap_unlock(ptep, ptl);
3059 if (target_nid == -1) {
3060 put_page(page);
3061 goto out;
3062 }
3063
3064 /* Migrate to the requested node */
3065 migrated = migrate_misplaced_page(page, vma, target_nid);
3066 if (migrated) {
3067 page_nid = target_nid;
3068 flags |= TNF_MIGRATED;
3069 }
3070
3071 out:
3072 if (page_nid != -1)
3073 task_numa_fault(last_cpupid, page_nid, 1, flags);
3074 return 0;
3075 }
3076
3077 /*
3078 * These routines also need to handle stuff like marking pages dirty
3079 * and/or accessed for architectures that don't do it in hardware (most
3080 * RISC architectures). The early dirtying is also good on the i386.
3081 *
3082 * There is also a hook called "update_mmu_cache()" that architectures
3083 * with external mmu caches can use to update those (ie the Sparc or
3084 * PowerPC hashed page tables that act as extended TLBs).
3085 *
3086 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3087 * but allow concurrent faults), and pte mapped but not yet locked.
3088 * We return with pte unmapped and unlocked.
3089 *
3090 * The mmap_sem may have been released depending on flags and our
3091 * return value. See filemap_fault() and __lock_page_or_retry().
3092 */
3093 static int handle_pte_fault(struct mm_struct *mm,
3094 struct vm_area_struct *vma, unsigned long address,
3095 pte_t *pte, pmd_t *pmd, unsigned int flags)
3096 {
3097 pte_t entry;
3098 spinlock_t *ptl;
3099
3100 /*
3101 * some architectures can have larger ptes than wordsize,
3102 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3103 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3104 * The code below just needs a consistent view for the ifs and
3105 * we later double check anyway with the ptl lock held. So here
3106 * a barrier will do.
3107 */
3108 entry = *pte;
3109 barrier();
3110 if (!pte_present(entry)) {
3111 if (pte_none(entry)) {
3112 if (vma->vm_ops) {
3113 if (likely(vma->vm_ops->fault))
3114 return do_fault(mm, vma, address, pte,
3115 pmd, flags, entry);
3116 }
3117 return do_anonymous_page(mm, vma, address,
3118 pte, pmd, flags);
3119 }
3120 return do_swap_page(mm, vma, address,
3121 pte, pmd, flags, entry);
3122 }
3123
3124 if (pte_numa(entry))
3125 return do_numa_page(mm, vma, address, entry, pte, pmd);
3126
3127 ptl = pte_lockptr(mm, pmd);
3128 spin_lock(ptl);
3129 if (unlikely(!pte_same(*pte, entry)))
3130 goto unlock;
3131 if (flags & FAULT_FLAG_WRITE) {
3132 if (!pte_write(entry))
3133 return do_wp_page(mm, vma, address,
3134 pte, pmd, ptl, entry);
3135 entry = pte_mkdirty(entry);
3136 }
3137 entry = pte_mkyoung(entry);
3138 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3139 update_mmu_cache(vma, address, pte);
3140 } else {
3141 /*
3142 * This is needed only for protection faults but the arch code
3143 * is not yet telling us if this is a protection fault or not.
3144 * This still avoids useless tlb flushes for .text page faults
3145 * with threads.
3146 */
3147 if (flags & FAULT_FLAG_WRITE)
3148 flush_tlb_fix_spurious_fault(vma, address);
3149 }
3150 unlock:
3151 pte_unmap_unlock(pte, ptl);
3152 return 0;
3153 }
3154
3155 /*
3156 * By the time we get here, we already hold the mm semaphore
3157 *
3158 * The mmap_sem may have been released depending on flags and our
3159 * return value. See filemap_fault() and __lock_page_or_retry().
3160 */
3161 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3162 unsigned long address, unsigned int flags)
3163 {
3164 pgd_t *pgd;
3165 pud_t *pud;
3166 pmd_t *pmd;
3167 pte_t *pte;
3168
3169 if (unlikely(is_vm_hugetlb_page(vma)))
3170 return hugetlb_fault(mm, vma, address, flags);
3171
3172 pgd = pgd_offset(mm, address);
3173 pud = pud_alloc(mm, pgd, address);
3174 if (!pud)
3175 return VM_FAULT_OOM;
3176 pmd = pmd_alloc(mm, pud, address);
3177 if (!pmd)
3178 return VM_FAULT_OOM;
3179 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3180 int ret = VM_FAULT_FALLBACK;
3181 if (!vma->vm_ops)
3182 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3183 pmd, flags);
3184 if (!(ret & VM_FAULT_FALLBACK))
3185 return ret;
3186 } else {
3187 pmd_t orig_pmd = *pmd;
3188 int ret;
3189
3190 barrier();
3191 if (pmd_trans_huge(orig_pmd)) {
3192 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3193
3194 /*
3195 * If the pmd is splitting, return and retry the
3196 * the fault. Alternative: wait until the split
3197 * is done, and goto retry.
3198 */
3199 if (pmd_trans_splitting(orig_pmd))
3200 return 0;
3201
3202 if (pmd_numa(orig_pmd))
3203 return do_huge_pmd_numa_page(mm, vma, address,
3204 orig_pmd, pmd);
3205
3206 if (dirty && !pmd_write(orig_pmd)) {
3207 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3208 orig_pmd);
3209 if (!(ret & VM_FAULT_FALLBACK))
3210 return ret;
3211 } else {
3212 huge_pmd_set_accessed(mm, vma, address, pmd,
3213 orig_pmd, dirty);
3214 return 0;
3215 }
3216 }
3217 }
3218
3219 /*
3220 * Use __pte_alloc instead of pte_alloc_map, because we can't
3221 * run pte_offset_map on the pmd, if an huge pmd could
3222 * materialize from under us from a different thread.
3223 */
3224 if (unlikely(pmd_none(*pmd)) &&
3225 unlikely(__pte_alloc(mm, vma, pmd, address)))
3226 return VM_FAULT_OOM;
3227 /* if an huge pmd materialized from under us just retry later */
3228 if (unlikely(pmd_trans_huge(*pmd)))
3229 return 0;
3230 /*
3231 * A regular pmd is established and it can't morph into a huge pmd
3232 * from under us anymore at this point because we hold the mmap_sem
3233 * read mode and khugepaged takes it in write mode. So now it's
3234 * safe to run pte_offset_map().
3235 */
3236 pte = pte_offset_map(pmd, address);
3237
3238 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3239 }
3240
3241 /*
3242 * By the time we get here, we already hold the mm semaphore
3243 *
3244 * The mmap_sem may have been released depending on flags and our
3245 * return value. See filemap_fault() and __lock_page_or_retry().
3246 */
3247 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3248 unsigned long address, unsigned int flags)
3249 {
3250 int ret;
3251
3252 __set_current_state(TASK_RUNNING);
3253
3254 count_vm_event(PGFAULT);
3255 mem_cgroup_count_vm_event(mm, PGFAULT);
3256
3257 /* do counter updates before entering really critical section. */
3258 check_sync_rss_stat(current);
3259
3260 /*
3261 * Enable the memcg OOM handling for faults triggered in user
3262 * space. Kernel faults are handled more gracefully.
3263 */
3264 if (flags & FAULT_FLAG_USER)
3265 mem_cgroup_oom_enable();
3266
3267 ret = __handle_mm_fault(mm, vma, address, flags);
3268
3269 if (flags & FAULT_FLAG_USER) {
3270 mem_cgroup_oom_disable();
3271 /*
3272 * The task may have entered a memcg OOM situation but
3273 * if the allocation error was handled gracefully (no
3274 * VM_FAULT_OOM), there is no need to kill anything.
3275 * Just clean up the OOM state peacefully.
3276 */
3277 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3278 mem_cgroup_oom_synchronize(false);
3279 }
3280
3281 return ret;
3282 }
3283 EXPORT_SYMBOL_GPL(handle_mm_fault);
3284
3285 #ifndef __PAGETABLE_PUD_FOLDED
3286 /*
3287 * Allocate page upper directory.
3288 * We've already handled the fast-path in-line.
3289 */
3290 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3291 {
3292 pud_t *new = pud_alloc_one(mm, address);
3293 if (!new)
3294 return -ENOMEM;
3295
3296 smp_wmb(); /* See comment in __pte_alloc */
3297
3298 spin_lock(&mm->page_table_lock);
3299 if (pgd_present(*pgd)) /* Another has populated it */
3300 pud_free(mm, new);
3301 else
3302 pgd_populate(mm, pgd, new);
3303 spin_unlock(&mm->page_table_lock);
3304 return 0;
3305 }
3306 #endif /* __PAGETABLE_PUD_FOLDED */
3307
3308 #ifndef __PAGETABLE_PMD_FOLDED
3309 /*
3310 * Allocate page middle directory.
3311 * We've already handled the fast-path in-line.
3312 */
3313 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3314 {
3315 pmd_t *new = pmd_alloc_one(mm, address);
3316 if (!new)
3317 return -ENOMEM;
3318
3319 smp_wmb(); /* See comment in __pte_alloc */
3320
3321 spin_lock(&mm->page_table_lock);
3322 #ifndef __ARCH_HAS_4LEVEL_HACK
3323 if (pud_present(*pud)) /* Another has populated it */
3324 pmd_free(mm, new);
3325 else
3326 pud_populate(mm, pud, new);
3327 #else
3328 if (pgd_present(*pud)) /* Another has populated it */
3329 pmd_free(mm, new);
3330 else
3331 pgd_populate(mm, pud, new);
3332 #endif /* __ARCH_HAS_4LEVEL_HACK */
3333 spin_unlock(&mm->page_table_lock);
3334 return 0;
3335 }
3336 #endif /* __PAGETABLE_PMD_FOLDED */
3337
3338 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3339 pte_t **ptepp, spinlock_t **ptlp)
3340 {
3341 pgd_t *pgd;
3342 pud_t *pud;
3343 pmd_t *pmd;
3344 pte_t *ptep;
3345
3346 pgd = pgd_offset(mm, address);
3347 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3348 goto out;
3349
3350 pud = pud_offset(pgd, address);
3351 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3352 goto out;
3353
3354 pmd = pmd_offset(pud, address);
3355 VM_BUG_ON(pmd_trans_huge(*pmd));
3356 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3357 goto out;
3358
3359 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3360 if (pmd_huge(*pmd))
3361 goto out;
3362
3363 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3364 if (!ptep)
3365 goto out;
3366 if (!pte_present(*ptep))
3367 goto unlock;
3368 *ptepp = ptep;
3369 return 0;
3370 unlock:
3371 pte_unmap_unlock(ptep, *ptlp);
3372 out:
3373 return -EINVAL;
3374 }
3375
3376 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3377 pte_t **ptepp, spinlock_t **ptlp)
3378 {
3379 int res;
3380
3381 /* (void) is needed to make gcc happy */
3382 (void) __cond_lock(*ptlp,
3383 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3384 return res;
3385 }
3386
3387 /**
3388 * follow_pfn - look up PFN at a user virtual address
3389 * @vma: memory mapping
3390 * @address: user virtual address
3391 * @pfn: location to store found PFN
3392 *
3393 * Only IO mappings and raw PFN mappings are allowed.
3394 *
3395 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3396 */
3397 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3398 unsigned long *pfn)
3399 {
3400 int ret = -EINVAL;
3401 spinlock_t *ptl;
3402 pte_t *ptep;
3403
3404 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3405 return ret;
3406
3407 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3408 if (ret)
3409 return ret;
3410 *pfn = pte_pfn(*ptep);
3411 pte_unmap_unlock(ptep, ptl);
3412 return 0;
3413 }
3414 EXPORT_SYMBOL(follow_pfn);
3415
3416 #ifdef CONFIG_HAVE_IOREMAP_PROT
3417 int follow_phys(struct vm_area_struct *vma,
3418 unsigned long address, unsigned int flags,
3419 unsigned long *prot, resource_size_t *phys)
3420 {
3421 int ret = -EINVAL;
3422 pte_t *ptep, pte;
3423 spinlock_t *ptl;
3424
3425 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3426 goto out;
3427
3428 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3429 goto out;
3430 pte = *ptep;
3431
3432 if ((flags & FOLL_WRITE) && !pte_write(pte))
3433 goto unlock;
3434
3435 *prot = pgprot_val(pte_pgprot(pte));
3436 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3437
3438 ret = 0;
3439 unlock:
3440 pte_unmap_unlock(ptep, ptl);
3441 out:
3442 return ret;
3443 }
3444
3445 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3446 void *buf, int len, int write)
3447 {
3448 resource_size_t phys_addr;
3449 unsigned long prot = 0;
3450 void __iomem *maddr;
3451 int offset = addr & (PAGE_SIZE-1);
3452
3453 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3454 return -EINVAL;
3455
3456 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3457 if (write)
3458 memcpy_toio(maddr + offset, buf, len);
3459 else
3460 memcpy_fromio(buf, maddr + offset, len);
3461 iounmap(maddr);
3462
3463 return len;
3464 }
3465 EXPORT_SYMBOL_GPL(generic_access_phys);
3466 #endif
3467
3468 /*
3469 * Access another process' address space as given in mm. If non-NULL, use the
3470 * given task for page fault accounting.
3471 */
3472 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3473 unsigned long addr, void *buf, int len, int write)
3474 {
3475 struct vm_area_struct *vma;
3476 void *old_buf = buf;
3477
3478 down_read(&mm->mmap_sem);
3479 /* ignore errors, just check how much was successfully transferred */
3480 while (len) {
3481 int bytes, ret, offset;
3482 void *maddr;
3483 struct page *page = NULL;
3484
3485 ret = get_user_pages(tsk, mm, addr, 1,
3486 write, 1, &page, &vma);
3487 if (ret <= 0) {
3488 #ifndef CONFIG_HAVE_IOREMAP_PROT
3489 break;
3490 #else
3491 /*
3492 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3493 * we can access using slightly different code.
3494 */
3495 vma = find_vma(mm, addr);
3496 if (!vma || vma->vm_start > addr)
3497 break;
3498 if (vma->vm_ops && vma->vm_ops->access)
3499 ret = vma->vm_ops->access(vma, addr, buf,
3500 len, write);
3501 if (ret <= 0)
3502 break;
3503 bytes = ret;
3504 #endif
3505 } else {
3506 bytes = len;
3507 offset = addr & (PAGE_SIZE-1);
3508 if (bytes > PAGE_SIZE-offset)
3509 bytes = PAGE_SIZE-offset;
3510
3511 maddr = kmap(page);
3512 if (write) {
3513 copy_to_user_page(vma, page, addr,
3514 maddr + offset, buf, bytes);
3515 set_page_dirty_lock(page);
3516 } else {
3517 copy_from_user_page(vma, page, addr,
3518 buf, maddr + offset, bytes);
3519 }
3520 kunmap(page);
3521 page_cache_release(page);
3522 }
3523 len -= bytes;
3524 buf += bytes;
3525 addr += bytes;
3526 }
3527 up_read(&mm->mmap_sem);
3528
3529 return buf - old_buf;
3530 }
3531
3532 /**
3533 * access_remote_vm - access another process' address space
3534 * @mm: the mm_struct of the target address space
3535 * @addr: start address to access
3536 * @buf: source or destination buffer
3537 * @len: number of bytes to transfer
3538 * @write: whether the access is a write
3539 *
3540 * The caller must hold a reference on @mm.
3541 */
3542 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3543 void *buf, int len, int write)
3544 {
3545 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3546 }
3547
3548 /*
3549 * Access another process' address space.
3550 * Source/target buffer must be kernel space,
3551 * Do not walk the page table directly, use get_user_pages
3552 */
3553 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3554 void *buf, int len, int write)
3555 {
3556 struct mm_struct *mm;
3557 int ret;
3558
3559 mm = get_task_mm(tsk);
3560 if (!mm)
3561 return 0;
3562
3563 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3564 mmput(mm);
3565
3566 return ret;
3567 }
3568
3569 /*
3570 * Print the name of a VMA.
3571 */
3572 void print_vma_addr(char *prefix, unsigned long ip)
3573 {
3574 struct mm_struct *mm = current->mm;
3575 struct vm_area_struct *vma;
3576
3577 /*
3578 * Do not print if we are in atomic
3579 * contexts (in exception stacks, etc.):
3580 */
3581 if (preempt_count())
3582 return;
3583
3584 down_read(&mm->mmap_sem);
3585 vma = find_vma(mm, ip);
3586 if (vma && vma->vm_file) {
3587 struct file *f = vma->vm_file;
3588 char *buf = (char *)__get_free_page(GFP_KERNEL);
3589 if (buf) {
3590 char *p;
3591
3592 p = d_path(&f->f_path, buf, PAGE_SIZE);
3593 if (IS_ERR(p))
3594 p = "?";
3595 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3596 vma->vm_start,
3597 vma->vm_end - vma->vm_start);
3598 free_page((unsigned long)buf);
3599 }
3600 }
3601 up_read(&mm->mmap_sem);
3602 }
3603
3604 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3605 void might_fault(void)
3606 {
3607 /*
3608 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3609 * holding the mmap_sem, this is safe because kernel memory doesn't
3610 * get paged out, therefore we'll never actually fault, and the
3611 * below annotations will generate false positives.
3612 */
3613 if (segment_eq(get_fs(), KERNEL_DS))
3614 return;
3615
3616 /*
3617 * it would be nicer only to annotate paths which are not under
3618 * pagefault_disable, however that requires a larger audit and
3619 * providing helpers like get_user_atomic.
3620 */
3621 if (in_atomic())
3622 return;
3623
3624 __might_sleep(__FILE__, __LINE__, 0);
3625
3626 if (current->mm)
3627 might_lock_read(&current->mm->mmap_sem);
3628 }
3629 EXPORT_SYMBOL(might_fault);
3630 #endif
3631
3632 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3633 static void clear_gigantic_page(struct page *page,
3634 unsigned long addr,
3635 unsigned int pages_per_huge_page)
3636 {
3637 int i;
3638 struct page *p = page;
3639
3640 might_sleep();
3641 for (i = 0; i < pages_per_huge_page;
3642 i++, p = mem_map_next(p, page, i)) {
3643 cond_resched();
3644 clear_user_highpage(p, addr + i * PAGE_SIZE);
3645 }
3646 }
3647 void clear_huge_page(struct page *page,
3648 unsigned long addr, unsigned int pages_per_huge_page)
3649 {
3650 int i;
3651
3652 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3653 clear_gigantic_page(page, addr, pages_per_huge_page);
3654 return;
3655 }
3656
3657 might_sleep();
3658 for (i = 0; i < pages_per_huge_page; i++) {
3659 cond_resched();
3660 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3661 }
3662 }
3663
3664 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3665 unsigned long addr,
3666 struct vm_area_struct *vma,
3667 unsigned int pages_per_huge_page)
3668 {
3669 int i;
3670 struct page *dst_base = dst;
3671 struct page *src_base = src;
3672
3673 for (i = 0; i < pages_per_huge_page; ) {
3674 cond_resched();
3675 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3676
3677 i++;
3678 dst = mem_map_next(dst, dst_base, i);
3679 src = mem_map_next(src, src_base, i);
3680 }
3681 }
3682
3683 void copy_user_huge_page(struct page *dst, struct page *src,
3684 unsigned long addr, struct vm_area_struct *vma,
3685 unsigned int pages_per_huge_page)
3686 {
3687 int i;
3688
3689 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3690 copy_user_gigantic_page(dst, src, addr, vma,
3691 pages_per_huge_page);
3692 return;
3693 }
3694
3695 might_sleep();
3696 for (i = 0; i < pages_per_huge_page; i++) {
3697 cond_resched();
3698 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3699 }
3700 }
3701 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3702
3703 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3704
3705 static struct kmem_cache *page_ptl_cachep;
3706
3707 void __init ptlock_cache_init(void)
3708 {
3709 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3710 SLAB_PANIC, NULL);
3711 }
3712
3713 bool ptlock_alloc(struct page *page)
3714 {
3715 spinlock_t *ptl;
3716
3717 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3718 if (!ptl)
3719 return false;
3720 page->ptl = ptl;
3721 return true;
3722 }
3723
3724 void ptlock_free(struct page *page)
3725 {
3726 kmem_cache_free(page_ptl_cachep, page->ptl);
3727 }
3728 #endif