Loading...
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/pfn_t.h>
54#include <linux/writeback.h>
55#include <linux/memcontrol.h>
56#include <linux/mmu_notifier.h>
57#include <linux/kallsyms.h>
58#include <linux/swapops.h>
59#include <linux/elf.h>
60#include <linux/gfp.h>
61#include <linux/migrate.h>
62#include <linux/string.h>
63#include <linux/dma-debug.h>
64#include <linux/debugfs.h>
65#include <linux/userfaultfd_k.h>
66
67#include <asm/io.h>
68#include <asm/mmu_context.h>
69#include <asm/pgalloc.h>
70#include <asm/uaccess.h>
71#include <asm/tlb.h>
72#include <asm/tlbflush.h>
73#include <asm/pgtable.h>
74
75#include "internal.h"
76
77#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
78#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
79#endif
80
81#ifndef CONFIG_NEED_MULTIPLE_NODES
82/* use the per-pgdat data instead for discontigmem - mbligh */
83unsigned long max_mapnr;
84struct page *mem_map;
85
86EXPORT_SYMBOL(max_mapnr);
87EXPORT_SYMBOL(mem_map);
88#endif
89
90/*
91 * A number of key systems in x86 including ioremap() rely on the assumption
92 * that high_memory defines the upper bound on direct map memory, then end
93 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
94 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
95 * and ZONE_HIGHMEM.
96 */
97void * high_memory;
98
99EXPORT_SYMBOL(high_memory);
100
101/*
102 * Randomize the address space (stacks, mmaps, brk, etc.).
103 *
104 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
105 * as ancient (libc5 based) binaries can segfault. )
106 */
107int randomize_va_space __read_mostly =
108#ifdef CONFIG_COMPAT_BRK
109 1;
110#else
111 2;
112#endif
113
114static int __init disable_randmaps(char *s)
115{
116 randomize_va_space = 0;
117 return 1;
118}
119__setup("norandmaps", disable_randmaps);
120
121unsigned long zero_pfn __read_mostly;
122unsigned long highest_memmap_pfn __read_mostly;
123
124EXPORT_SYMBOL(zero_pfn);
125
126/*
127 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
128 */
129static int __init init_zero_pfn(void)
130{
131 zero_pfn = page_to_pfn(ZERO_PAGE(0));
132 return 0;
133}
134core_initcall(init_zero_pfn);
135
136
137#if defined(SPLIT_RSS_COUNTING)
138
139void sync_mm_rss(struct mm_struct *mm)
140{
141 int i;
142
143 for (i = 0; i < NR_MM_COUNTERS; i++) {
144 if (current->rss_stat.count[i]) {
145 add_mm_counter(mm, i, current->rss_stat.count[i]);
146 current->rss_stat.count[i] = 0;
147 }
148 }
149 current->rss_stat.events = 0;
150}
151
152static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
153{
154 struct task_struct *task = current;
155
156 if (likely(task->mm == mm))
157 task->rss_stat.count[member] += val;
158 else
159 add_mm_counter(mm, member, val);
160}
161#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
162#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
163
164/* sync counter once per 64 page faults */
165#define TASK_RSS_EVENTS_THRESH (64)
166static void check_sync_rss_stat(struct task_struct *task)
167{
168 if (unlikely(task != current))
169 return;
170 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
171 sync_mm_rss(task->mm);
172}
173#else /* SPLIT_RSS_COUNTING */
174
175#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
176#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
177
178static void check_sync_rss_stat(struct task_struct *task)
179{
180}
181
182#endif /* SPLIT_RSS_COUNTING */
183
184#ifdef HAVE_GENERIC_MMU_GATHER
185
186static bool tlb_next_batch(struct mmu_gather *tlb)
187{
188 struct mmu_gather_batch *batch;
189
190 batch = tlb->active;
191 if (batch->next) {
192 tlb->active = batch->next;
193 return true;
194 }
195
196 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
197 return false;
198
199 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
200 if (!batch)
201 return false;
202
203 tlb->batch_count++;
204 batch->next = NULL;
205 batch->nr = 0;
206 batch->max = MAX_GATHER_BATCH;
207
208 tlb->active->next = batch;
209 tlb->active = batch;
210
211 return true;
212}
213
214/* tlb_gather_mmu
215 * Called to initialize an (on-stack) mmu_gather structure for page-table
216 * tear-down from @mm. The @fullmm argument is used when @mm is without
217 * users and we're going to destroy the full address space (exit/execve).
218 */
219void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
220{
221 tlb->mm = mm;
222
223 /* Is it from 0 to ~0? */
224 tlb->fullmm = !(start | (end+1));
225 tlb->need_flush_all = 0;
226 tlb->local.next = NULL;
227 tlb->local.nr = 0;
228 tlb->local.max = ARRAY_SIZE(tlb->__pages);
229 tlb->active = &tlb->local;
230 tlb->batch_count = 0;
231
232#ifdef CONFIG_HAVE_RCU_TABLE_FREE
233 tlb->batch = NULL;
234#endif
235
236 __tlb_reset_range(tlb);
237}
238
239static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
240{
241 if (!tlb->end)
242 return;
243
244 tlb_flush(tlb);
245 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
246#ifdef CONFIG_HAVE_RCU_TABLE_FREE
247 tlb_table_flush(tlb);
248#endif
249 __tlb_reset_range(tlb);
250}
251
252static void tlb_flush_mmu_free(struct mmu_gather *tlb)
253{
254 struct mmu_gather_batch *batch;
255
256 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
257 free_pages_and_swap_cache(batch->pages, batch->nr);
258 batch->nr = 0;
259 }
260 tlb->active = &tlb->local;
261}
262
263void tlb_flush_mmu(struct mmu_gather *tlb)
264{
265 tlb_flush_mmu_tlbonly(tlb);
266 tlb_flush_mmu_free(tlb);
267}
268
269/* tlb_finish_mmu
270 * Called at the end of the shootdown operation to free up any resources
271 * that were required.
272 */
273void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
274{
275 struct mmu_gather_batch *batch, *next;
276
277 tlb_flush_mmu(tlb);
278
279 /* keep the page table cache within bounds */
280 check_pgt_cache();
281
282 for (batch = tlb->local.next; batch; batch = next) {
283 next = batch->next;
284 free_pages((unsigned long)batch, 0);
285 }
286 tlb->local.next = NULL;
287}
288
289/* __tlb_remove_page
290 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
291 * handling the additional races in SMP caused by other CPUs caching valid
292 * mappings in their TLBs. Returns the number of free page slots left.
293 * When out of page slots we must call tlb_flush_mmu().
294 */
295int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
296{
297 struct mmu_gather_batch *batch;
298
299 VM_BUG_ON(!tlb->end);
300
301 batch = tlb->active;
302 batch->pages[batch->nr++] = page;
303 if (batch->nr == batch->max) {
304 if (!tlb_next_batch(tlb))
305 return 0;
306 batch = tlb->active;
307 }
308 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
309
310 return batch->max - batch->nr;
311}
312
313#endif /* HAVE_GENERIC_MMU_GATHER */
314
315#ifdef CONFIG_HAVE_RCU_TABLE_FREE
316
317/*
318 * See the comment near struct mmu_table_batch.
319 */
320
321static void tlb_remove_table_smp_sync(void *arg)
322{
323 /* Simply deliver the interrupt */
324}
325
326static void tlb_remove_table_one(void *table)
327{
328 /*
329 * This isn't an RCU grace period and hence the page-tables cannot be
330 * assumed to be actually RCU-freed.
331 *
332 * It is however sufficient for software page-table walkers that rely on
333 * IRQ disabling. See the comment near struct mmu_table_batch.
334 */
335 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
336 __tlb_remove_table(table);
337}
338
339static void tlb_remove_table_rcu(struct rcu_head *head)
340{
341 struct mmu_table_batch *batch;
342 int i;
343
344 batch = container_of(head, struct mmu_table_batch, rcu);
345
346 for (i = 0; i < batch->nr; i++)
347 __tlb_remove_table(batch->tables[i]);
348
349 free_page((unsigned long)batch);
350}
351
352void tlb_table_flush(struct mmu_gather *tlb)
353{
354 struct mmu_table_batch **batch = &tlb->batch;
355
356 if (*batch) {
357 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
358 *batch = NULL;
359 }
360}
361
362void tlb_remove_table(struct mmu_gather *tlb, void *table)
363{
364 struct mmu_table_batch **batch = &tlb->batch;
365
366 /*
367 * When there's less then two users of this mm there cannot be a
368 * concurrent page-table walk.
369 */
370 if (atomic_read(&tlb->mm->mm_users) < 2) {
371 __tlb_remove_table(table);
372 return;
373 }
374
375 if (*batch == NULL) {
376 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
377 if (*batch == NULL) {
378 tlb_remove_table_one(table);
379 return;
380 }
381 (*batch)->nr = 0;
382 }
383 (*batch)->tables[(*batch)->nr++] = table;
384 if ((*batch)->nr == MAX_TABLE_BATCH)
385 tlb_table_flush(tlb);
386}
387
388#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
389
390/*
391 * Note: this doesn't free the actual pages themselves. That
392 * has been handled earlier when unmapping all the memory regions.
393 */
394static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
395 unsigned long addr)
396{
397 pgtable_t token = pmd_pgtable(*pmd);
398 pmd_clear(pmd);
399 pte_free_tlb(tlb, token, addr);
400 atomic_long_dec(&tlb->mm->nr_ptes);
401}
402
403static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
404 unsigned long addr, unsigned long end,
405 unsigned long floor, unsigned long ceiling)
406{
407 pmd_t *pmd;
408 unsigned long next;
409 unsigned long start;
410
411 start = addr;
412 pmd = pmd_offset(pud, addr);
413 do {
414 next = pmd_addr_end(addr, end);
415 if (pmd_none_or_clear_bad(pmd))
416 continue;
417 free_pte_range(tlb, pmd, addr);
418 } while (pmd++, addr = next, addr != end);
419
420 start &= PUD_MASK;
421 if (start < floor)
422 return;
423 if (ceiling) {
424 ceiling &= PUD_MASK;
425 if (!ceiling)
426 return;
427 }
428 if (end - 1 > ceiling - 1)
429 return;
430
431 pmd = pmd_offset(pud, start);
432 pud_clear(pud);
433 pmd_free_tlb(tlb, pmd, start);
434 mm_dec_nr_pmds(tlb->mm);
435}
436
437static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
438 unsigned long addr, unsigned long end,
439 unsigned long floor, unsigned long ceiling)
440{
441 pud_t *pud;
442 unsigned long next;
443 unsigned long start;
444
445 start = addr;
446 pud = pud_offset(pgd, addr);
447 do {
448 next = pud_addr_end(addr, end);
449 if (pud_none_or_clear_bad(pud))
450 continue;
451 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
452 } while (pud++, addr = next, addr != end);
453
454 start &= PGDIR_MASK;
455 if (start < floor)
456 return;
457 if (ceiling) {
458 ceiling &= PGDIR_MASK;
459 if (!ceiling)
460 return;
461 }
462 if (end - 1 > ceiling - 1)
463 return;
464
465 pud = pud_offset(pgd, start);
466 pgd_clear(pgd);
467 pud_free_tlb(tlb, pud, start);
468}
469
470/*
471 * This function frees user-level page tables of a process.
472 */
473void free_pgd_range(struct mmu_gather *tlb,
474 unsigned long addr, unsigned long end,
475 unsigned long floor, unsigned long ceiling)
476{
477 pgd_t *pgd;
478 unsigned long next;
479
480 /*
481 * The next few lines have given us lots of grief...
482 *
483 * Why are we testing PMD* at this top level? Because often
484 * there will be no work to do at all, and we'd prefer not to
485 * go all the way down to the bottom just to discover that.
486 *
487 * Why all these "- 1"s? Because 0 represents both the bottom
488 * of the address space and the top of it (using -1 for the
489 * top wouldn't help much: the masks would do the wrong thing).
490 * The rule is that addr 0 and floor 0 refer to the bottom of
491 * the address space, but end 0 and ceiling 0 refer to the top
492 * Comparisons need to use "end - 1" and "ceiling - 1" (though
493 * that end 0 case should be mythical).
494 *
495 * Wherever addr is brought up or ceiling brought down, we must
496 * be careful to reject "the opposite 0" before it confuses the
497 * subsequent tests. But what about where end is brought down
498 * by PMD_SIZE below? no, end can't go down to 0 there.
499 *
500 * Whereas we round start (addr) and ceiling down, by different
501 * masks at different levels, in order to test whether a table
502 * now has no other vmas using it, so can be freed, we don't
503 * bother to round floor or end up - the tests don't need that.
504 */
505
506 addr &= PMD_MASK;
507 if (addr < floor) {
508 addr += PMD_SIZE;
509 if (!addr)
510 return;
511 }
512 if (ceiling) {
513 ceiling &= PMD_MASK;
514 if (!ceiling)
515 return;
516 }
517 if (end - 1 > ceiling - 1)
518 end -= PMD_SIZE;
519 if (addr > end - 1)
520 return;
521
522 pgd = pgd_offset(tlb->mm, addr);
523 do {
524 next = pgd_addr_end(addr, end);
525 if (pgd_none_or_clear_bad(pgd))
526 continue;
527 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
528 } while (pgd++, addr = next, addr != end);
529}
530
531void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
532 unsigned long floor, unsigned long ceiling)
533{
534 while (vma) {
535 struct vm_area_struct *next = vma->vm_next;
536 unsigned long addr = vma->vm_start;
537
538 /*
539 * Hide vma from rmap and truncate_pagecache before freeing
540 * pgtables
541 */
542 unlink_anon_vmas(vma);
543 unlink_file_vma(vma);
544
545 if (is_vm_hugetlb_page(vma)) {
546 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
547 floor, next? next->vm_start: ceiling);
548 } else {
549 /*
550 * Optimization: gather nearby vmas into one call down
551 */
552 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
553 && !is_vm_hugetlb_page(next)) {
554 vma = next;
555 next = vma->vm_next;
556 unlink_anon_vmas(vma);
557 unlink_file_vma(vma);
558 }
559 free_pgd_range(tlb, addr, vma->vm_end,
560 floor, next? next->vm_start: ceiling);
561 }
562 vma = next;
563 }
564}
565
566int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
567{
568 spinlock_t *ptl;
569 pgtable_t new = pte_alloc_one(mm, address);
570 if (!new)
571 return -ENOMEM;
572
573 /*
574 * Ensure all pte setup (eg. pte page lock and page clearing) are
575 * visible before the pte is made visible to other CPUs by being
576 * put into page tables.
577 *
578 * The other side of the story is the pointer chasing in the page
579 * table walking code (when walking the page table without locking;
580 * ie. most of the time). Fortunately, these data accesses consist
581 * of a chain of data-dependent loads, meaning most CPUs (alpha
582 * being the notable exception) will already guarantee loads are
583 * seen in-order. See the alpha page table accessors for the
584 * smp_read_barrier_depends() barriers in page table walking code.
585 */
586 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
587
588 ptl = pmd_lock(mm, pmd);
589 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
590 atomic_long_inc(&mm->nr_ptes);
591 pmd_populate(mm, pmd, new);
592 new = NULL;
593 }
594 spin_unlock(ptl);
595 if (new)
596 pte_free(mm, new);
597 return 0;
598}
599
600int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
601{
602 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
603 if (!new)
604 return -ENOMEM;
605
606 smp_wmb(); /* See comment in __pte_alloc */
607
608 spin_lock(&init_mm.page_table_lock);
609 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
610 pmd_populate_kernel(&init_mm, pmd, new);
611 new = NULL;
612 }
613 spin_unlock(&init_mm.page_table_lock);
614 if (new)
615 pte_free_kernel(&init_mm, new);
616 return 0;
617}
618
619static inline void init_rss_vec(int *rss)
620{
621 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
622}
623
624static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
625{
626 int i;
627
628 if (current->mm == mm)
629 sync_mm_rss(mm);
630 for (i = 0; i < NR_MM_COUNTERS; i++)
631 if (rss[i])
632 add_mm_counter(mm, i, rss[i]);
633}
634
635/*
636 * This function is called to print an error when a bad pte
637 * is found. For example, we might have a PFN-mapped pte in
638 * a region that doesn't allow it.
639 *
640 * The calling function must still handle the error.
641 */
642static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
643 pte_t pte, struct page *page)
644{
645 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
646 pud_t *pud = pud_offset(pgd, addr);
647 pmd_t *pmd = pmd_offset(pud, addr);
648 struct address_space *mapping;
649 pgoff_t index;
650 static unsigned long resume;
651 static unsigned long nr_shown;
652 static unsigned long nr_unshown;
653
654 /*
655 * Allow a burst of 60 reports, then keep quiet for that minute;
656 * or allow a steady drip of one report per second.
657 */
658 if (nr_shown == 60) {
659 if (time_before(jiffies, resume)) {
660 nr_unshown++;
661 return;
662 }
663 if (nr_unshown) {
664 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
665 nr_unshown);
666 nr_unshown = 0;
667 }
668 nr_shown = 0;
669 }
670 if (nr_shown++ == 0)
671 resume = jiffies + 60 * HZ;
672
673 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
674 index = linear_page_index(vma, addr);
675
676 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
677 current->comm,
678 (long long)pte_val(pte), (long long)pmd_val(*pmd));
679 if (page)
680 dump_page(page, "bad pte");
681 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
682 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
683 /*
684 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
685 */
686 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
687 vma->vm_file,
688 vma->vm_ops ? vma->vm_ops->fault : NULL,
689 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
690 mapping ? mapping->a_ops->readpage : NULL);
691 dump_stack();
692 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
693}
694
695/*
696 * vm_normal_page -- This function gets the "struct page" associated with a pte.
697 *
698 * "Special" mappings do not wish to be associated with a "struct page" (either
699 * it doesn't exist, or it exists but they don't want to touch it). In this
700 * case, NULL is returned here. "Normal" mappings do have a struct page.
701 *
702 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
703 * pte bit, in which case this function is trivial. Secondly, an architecture
704 * may not have a spare pte bit, which requires a more complicated scheme,
705 * described below.
706 *
707 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
708 * special mapping (even if there are underlying and valid "struct pages").
709 * COWed pages of a VM_PFNMAP are always normal.
710 *
711 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
712 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
713 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
714 * mapping will always honor the rule
715 *
716 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
717 *
718 * And for normal mappings this is false.
719 *
720 * This restricts such mappings to be a linear translation from virtual address
721 * to pfn. To get around this restriction, we allow arbitrary mappings so long
722 * as the vma is not a COW mapping; in that case, we know that all ptes are
723 * special (because none can have been COWed).
724 *
725 *
726 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
727 *
728 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
729 * page" backing, however the difference is that _all_ pages with a struct
730 * page (that is, those where pfn_valid is true) are refcounted and considered
731 * normal pages by the VM. The disadvantage is that pages are refcounted
732 * (which can be slower and simply not an option for some PFNMAP users). The
733 * advantage is that we don't have to follow the strict linearity rule of
734 * PFNMAP mappings in order to support COWable mappings.
735 *
736 */
737#ifdef __HAVE_ARCH_PTE_SPECIAL
738# define HAVE_PTE_SPECIAL 1
739#else
740# define HAVE_PTE_SPECIAL 0
741#endif
742struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
743 pte_t pte)
744{
745 unsigned long pfn = pte_pfn(pte);
746
747 if (HAVE_PTE_SPECIAL) {
748 if (likely(!pte_special(pte)))
749 goto check_pfn;
750 if (vma->vm_ops && vma->vm_ops->find_special_page)
751 return vma->vm_ops->find_special_page(vma, addr);
752 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
753 return NULL;
754 if (!is_zero_pfn(pfn))
755 print_bad_pte(vma, addr, pte, NULL);
756 return NULL;
757 }
758
759 /* !HAVE_PTE_SPECIAL case follows: */
760
761 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
762 if (vma->vm_flags & VM_MIXEDMAP) {
763 if (!pfn_valid(pfn))
764 return NULL;
765 goto out;
766 } else {
767 unsigned long off;
768 off = (addr - vma->vm_start) >> PAGE_SHIFT;
769 if (pfn == vma->vm_pgoff + off)
770 return NULL;
771 if (!is_cow_mapping(vma->vm_flags))
772 return NULL;
773 }
774 }
775
776 if (is_zero_pfn(pfn))
777 return NULL;
778check_pfn:
779 if (unlikely(pfn > highest_memmap_pfn)) {
780 print_bad_pte(vma, addr, pte, NULL);
781 return NULL;
782 }
783
784 /*
785 * NOTE! We still have PageReserved() pages in the page tables.
786 * eg. VDSO mappings can cause them to exist.
787 */
788out:
789 return pfn_to_page(pfn);
790}
791
792#ifdef CONFIG_TRANSPARENT_HUGEPAGE
793struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
794 pmd_t pmd)
795{
796 unsigned long pfn = pmd_pfn(pmd);
797
798 /*
799 * There is no pmd_special() but there may be special pmds, e.g.
800 * in a direct-access (dax) mapping, so let's just replicate the
801 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
802 */
803 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
804 if (vma->vm_flags & VM_MIXEDMAP) {
805 if (!pfn_valid(pfn))
806 return NULL;
807 goto out;
808 } else {
809 unsigned long off;
810 off = (addr - vma->vm_start) >> PAGE_SHIFT;
811 if (pfn == vma->vm_pgoff + off)
812 return NULL;
813 if (!is_cow_mapping(vma->vm_flags))
814 return NULL;
815 }
816 }
817
818 if (is_zero_pfn(pfn))
819 return NULL;
820 if (unlikely(pfn > highest_memmap_pfn))
821 return NULL;
822
823 /*
824 * NOTE! We still have PageReserved() pages in the page tables.
825 * eg. VDSO mappings can cause them to exist.
826 */
827out:
828 return pfn_to_page(pfn);
829}
830#endif
831
832/*
833 * copy one vm_area from one task to the other. Assumes the page tables
834 * already present in the new task to be cleared in the whole range
835 * covered by this vma.
836 */
837
838static inline unsigned long
839copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
840 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
841 unsigned long addr, int *rss)
842{
843 unsigned long vm_flags = vma->vm_flags;
844 pte_t pte = *src_pte;
845 struct page *page;
846
847 /* pte contains position in swap or file, so copy. */
848 if (unlikely(!pte_present(pte))) {
849 swp_entry_t entry = pte_to_swp_entry(pte);
850
851 if (likely(!non_swap_entry(entry))) {
852 if (swap_duplicate(entry) < 0)
853 return entry.val;
854
855 /* make sure dst_mm is on swapoff's mmlist. */
856 if (unlikely(list_empty(&dst_mm->mmlist))) {
857 spin_lock(&mmlist_lock);
858 if (list_empty(&dst_mm->mmlist))
859 list_add(&dst_mm->mmlist,
860 &src_mm->mmlist);
861 spin_unlock(&mmlist_lock);
862 }
863 rss[MM_SWAPENTS]++;
864 } else if (is_migration_entry(entry)) {
865 page = migration_entry_to_page(entry);
866
867 rss[mm_counter(page)]++;
868
869 if (is_write_migration_entry(entry) &&
870 is_cow_mapping(vm_flags)) {
871 /*
872 * COW mappings require pages in both
873 * parent and child to be set to read.
874 */
875 make_migration_entry_read(&entry);
876 pte = swp_entry_to_pte(entry);
877 if (pte_swp_soft_dirty(*src_pte))
878 pte = pte_swp_mksoft_dirty(pte);
879 set_pte_at(src_mm, addr, src_pte, pte);
880 }
881 }
882 goto out_set_pte;
883 }
884
885 /*
886 * If it's a COW mapping, write protect it both
887 * in the parent and the child
888 */
889 if (is_cow_mapping(vm_flags)) {
890 ptep_set_wrprotect(src_mm, addr, src_pte);
891 pte = pte_wrprotect(pte);
892 }
893
894 /*
895 * If it's a shared mapping, mark it clean in
896 * the child
897 */
898 if (vm_flags & VM_SHARED)
899 pte = pte_mkclean(pte);
900 pte = pte_mkold(pte);
901
902 page = vm_normal_page(vma, addr, pte);
903 if (page) {
904 get_page(page);
905 page_dup_rmap(page, false);
906 rss[mm_counter(page)]++;
907 }
908
909out_set_pte:
910 set_pte_at(dst_mm, addr, dst_pte, pte);
911 return 0;
912}
913
914static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
915 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
916 unsigned long addr, unsigned long end)
917{
918 pte_t *orig_src_pte, *orig_dst_pte;
919 pte_t *src_pte, *dst_pte;
920 spinlock_t *src_ptl, *dst_ptl;
921 int progress = 0;
922 int rss[NR_MM_COUNTERS];
923 swp_entry_t entry = (swp_entry_t){0};
924
925again:
926 init_rss_vec(rss);
927
928 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
929 if (!dst_pte)
930 return -ENOMEM;
931 src_pte = pte_offset_map(src_pmd, addr);
932 src_ptl = pte_lockptr(src_mm, src_pmd);
933 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
934 orig_src_pte = src_pte;
935 orig_dst_pte = dst_pte;
936 arch_enter_lazy_mmu_mode();
937
938 do {
939 /*
940 * We are holding two locks at this point - either of them
941 * could generate latencies in another task on another CPU.
942 */
943 if (progress >= 32) {
944 progress = 0;
945 if (need_resched() ||
946 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
947 break;
948 }
949 if (pte_none(*src_pte)) {
950 progress++;
951 continue;
952 }
953 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
954 vma, addr, rss);
955 if (entry.val)
956 break;
957 progress += 8;
958 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
959
960 arch_leave_lazy_mmu_mode();
961 spin_unlock(src_ptl);
962 pte_unmap(orig_src_pte);
963 add_mm_rss_vec(dst_mm, rss);
964 pte_unmap_unlock(orig_dst_pte, dst_ptl);
965 cond_resched();
966
967 if (entry.val) {
968 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
969 return -ENOMEM;
970 progress = 0;
971 }
972 if (addr != end)
973 goto again;
974 return 0;
975}
976
977static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
978 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
979 unsigned long addr, unsigned long end)
980{
981 pmd_t *src_pmd, *dst_pmd;
982 unsigned long next;
983
984 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
985 if (!dst_pmd)
986 return -ENOMEM;
987 src_pmd = pmd_offset(src_pud, addr);
988 do {
989 next = pmd_addr_end(addr, end);
990 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
991 int err;
992 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
993 err = copy_huge_pmd(dst_mm, src_mm,
994 dst_pmd, src_pmd, addr, vma);
995 if (err == -ENOMEM)
996 return -ENOMEM;
997 if (!err)
998 continue;
999 /* fall through */
1000 }
1001 if (pmd_none_or_clear_bad(src_pmd))
1002 continue;
1003 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1004 vma, addr, next))
1005 return -ENOMEM;
1006 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1007 return 0;
1008}
1009
1010static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1011 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1012 unsigned long addr, unsigned long end)
1013{
1014 pud_t *src_pud, *dst_pud;
1015 unsigned long next;
1016
1017 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1018 if (!dst_pud)
1019 return -ENOMEM;
1020 src_pud = pud_offset(src_pgd, addr);
1021 do {
1022 next = pud_addr_end(addr, end);
1023 if (pud_none_or_clear_bad(src_pud))
1024 continue;
1025 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1026 vma, addr, next))
1027 return -ENOMEM;
1028 } while (dst_pud++, src_pud++, addr = next, addr != end);
1029 return 0;
1030}
1031
1032int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1033 struct vm_area_struct *vma)
1034{
1035 pgd_t *src_pgd, *dst_pgd;
1036 unsigned long next;
1037 unsigned long addr = vma->vm_start;
1038 unsigned long end = vma->vm_end;
1039 unsigned long mmun_start; /* For mmu_notifiers */
1040 unsigned long mmun_end; /* For mmu_notifiers */
1041 bool is_cow;
1042 int ret;
1043
1044 /*
1045 * Don't copy ptes where a page fault will fill them correctly.
1046 * Fork becomes much lighter when there are big shared or private
1047 * readonly mappings. The tradeoff is that copy_page_range is more
1048 * efficient than faulting.
1049 */
1050 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1051 !vma->anon_vma)
1052 return 0;
1053
1054 if (is_vm_hugetlb_page(vma))
1055 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1056
1057 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1058 /*
1059 * We do not free on error cases below as remove_vma
1060 * gets called on error from higher level routine
1061 */
1062 ret = track_pfn_copy(vma);
1063 if (ret)
1064 return ret;
1065 }
1066
1067 /*
1068 * We need to invalidate the secondary MMU mappings only when
1069 * there could be a permission downgrade on the ptes of the
1070 * parent mm. And a permission downgrade will only happen if
1071 * is_cow_mapping() returns true.
1072 */
1073 is_cow = is_cow_mapping(vma->vm_flags);
1074 mmun_start = addr;
1075 mmun_end = end;
1076 if (is_cow)
1077 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1078 mmun_end);
1079
1080 ret = 0;
1081 dst_pgd = pgd_offset(dst_mm, addr);
1082 src_pgd = pgd_offset(src_mm, addr);
1083 do {
1084 next = pgd_addr_end(addr, end);
1085 if (pgd_none_or_clear_bad(src_pgd))
1086 continue;
1087 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1088 vma, addr, next))) {
1089 ret = -ENOMEM;
1090 break;
1091 }
1092 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1093
1094 if (is_cow)
1095 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1096 return ret;
1097}
1098
1099static unsigned long zap_pte_range(struct mmu_gather *tlb,
1100 struct vm_area_struct *vma, pmd_t *pmd,
1101 unsigned long addr, unsigned long end,
1102 struct zap_details *details)
1103{
1104 struct mm_struct *mm = tlb->mm;
1105 int force_flush = 0;
1106 int rss[NR_MM_COUNTERS];
1107 spinlock_t *ptl;
1108 pte_t *start_pte;
1109 pte_t *pte;
1110 swp_entry_t entry;
1111
1112again:
1113 init_rss_vec(rss);
1114 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1115 pte = start_pte;
1116 arch_enter_lazy_mmu_mode();
1117 do {
1118 pte_t ptent = *pte;
1119 if (pte_none(ptent)) {
1120 continue;
1121 }
1122
1123 if (pte_present(ptent)) {
1124 struct page *page;
1125
1126 page = vm_normal_page(vma, addr, ptent);
1127 if (unlikely(details) && page) {
1128 /*
1129 * unmap_shared_mapping_pages() wants to
1130 * invalidate cache without truncating:
1131 * unmap shared but keep private pages.
1132 */
1133 if (details->check_mapping &&
1134 details->check_mapping != page->mapping)
1135 continue;
1136 }
1137 ptent = ptep_get_and_clear_full(mm, addr, pte,
1138 tlb->fullmm);
1139 tlb_remove_tlb_entry(tlb, pte, addr);
1140 if (unlikely(!page))
1141 continue;
1142
1143 if (!PageAnon(page)) {
1144 if (pte_dirty(ptent)) {
1145 /*
1146 * oom_reaper cannot tear down dirty
1147 * pages
1148 */
1149 if (unlikely(details && details->ignore_dirty))
1150 continue;
1151 force_flush = 1;
1152 set_page_dirty(page);
1153 }
1154 if (pte_young(ptent) &&
1155 likely(!(vma->vm_flags & VM_SEQ_READ)))
1156 mark_page_accessed(page);
1157 }
1158 rss[mm_counter(page)]--;
1159 page_remove_rmap(page, false);
1160 if (unlikely(page_mapcount(page) < 0))
1161 print_bad_pte(vma, addr, ptent, page);
1162 if (unlikely(!__tlb_remove_page(tlb, page))) {
1163 force_flush = 1;
1164 addr += PAGE_SIZE;
1165 break;
1166 }
1167 continue;
1168 }
1169 /* only check swap_entries if explicitly asked for in details */
1170 if (unlikely(details && !details->check_swap_entries))
1171 continue;
1172
1173 entry = pte_to_swp_entry(ptent);
1174 if (!non_swap_entry(entry))
1175 rss[MM_SWAPENTS]--;
1176 else if (is_migration_entry(entry)) {
1177 struct page *page;
1178
1179 page = migration_entry_to_page(entry);
1180 rss[mm_counter(page)]--;
1181 }
1182 if (unlikely(!free_swap_and_cache(entry)))
1183 print_bad_pte(vma, addr, ptent, NULL);
1184 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1185 } while (pte++, addr += PAGE_SIZE, addr != end);
1186
1187 add_mm_rss_vec(mm, rss);
1188 arch_leave_lazy_mmu_mode();
1189
1190 /* Do the actual TLB flush before dropping ptl */
1191 if (force_flush)
1192 tlb_flush_mmu_tlbonly(tlb);
1193 pte_unmap_unlock(start_pte, ptl);
1194
1195 /*
1196 * If we forced a TLB flush (either due to running out of
1197 * batch buffers or because we needed to flush dirty TLB
1198 * entries before releasing the ptl), free the batched
1199 * memory too. Restart if we didn't do everything.
1200 */
1201 if (force_flush) {
1202 force_flush = 0;
1203 tlb_flush_mmu_free(tlb);
1204
1205 if (addr != end)
1206 goto again;
1207 }
1208
1209 return addr;
1210}
1211
1212static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1213 struct vm_area_struct *vma, pud_t *pud,
1214 unsigned long addr, unsigned long end,
1215 struct zap_details *details)
1216{
1217 pmd_t *pmd;
1218 unsigned long next;
1219
1220 pmd = pmd_offset(pud, addr);
1221 do {
1222 next = pmd_addr_end(addr, end);
1223 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1224 if (next - addr != HPAGE_PMD_SIZE) {
1225 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1226 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1227 split_huge_pmd(vma, pmd, addr);
1228 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1229 goto next;
1230 /* fall through */
1231 }
1232 /*
1233 * Here there can be other concurrent MADV_DONTNEED or
1234 * trans huge page faults running, and if the pmd is
1235 * none or trans huge it can change under us. This is
1236 * because MADV_DONTNEED holds the mmap_sem in read
1237 * mode.
1238 */
1239 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1240 goto next;
1241 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1242next:
1243 cond_resched();
1244 } while (pmd++, addr = next, addr != end);
1245
1246 return addr;
1247}
1248
1249static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1250 struct vm_area_struct *vma, pgd_t *pgd,
1251 unsigned long addr, unsigned long end,
1252 struct zap_details *details)
1253{
1254 pud_t *pud;
1255 unsigned long next;
1256
1257 pud = pud_offset(pgd, addr);
1258 do {
1259 next = pud_addr_end(addr, end);
1260 if (pud_none_or_clear_bad(pud))
1261 continue;
1262 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1263 } while (pud++, addr = next, addr != end);
1264
1265 return addr;
1266}
1267
1268void unmap_page_range(struct mmu_gather *tlb,
1269 struct vm_area_struct *vma,
1270 unsigned long addr, unsigned long end,
1271 struct zap_details *details)
1272{
1273 pgd_t *pgd;
1274 unsigned long next;
1275
1276 BUG_ON(addr >= end);
1277 tlb_start_vma(tlb, vma);
1278 pgd = pgd_offset(vma->vm_mm, addr);
1279 do {
1280 next = pgd_addr_end(addr, end);
1281 if (pgd_none_or_clear_bad(pgd))
1282 continue;
1283 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1284 } while (pgd++, addr = next, addr != end);
1285 tlb_end_vma(tlb, vma);
1286}
1287
1288
1289static void unmap_single_vma(struct mmu_gather *tlb,
1290 struct vm_area_struct *vma, unsigned long start_addr,
1291 unsigned long end_addr,
1292 struct zap_details *details)
1293{
1294 unsigned long start = max(vma->vm_start, start_addr);
1295 unsigned long end;
1296
1297 if (start >= vma->vm_end)
1298 return;
1299 end = min(vma->vm_end, end_addr);
1300 if (end <= vma->vm_start)
1301 return;
1302
1303 if (vma->vm_file)
1304 uprobe_munmap(vma, start, end);
1305
1306 if (unlikely(vma->vm_flags & VM_PFNMAP))
1307 untrack_pfn(vma, 0, 0);
1308
1309 if (start != end) {
1310 if (unlikely(is_vm_hugetlb_page(vma))) {
1311 /*
1312 * It is undesirable to test vma->vm_file as it
1313 * should be non-null for valid hugetlb area.
1314 * However, vm_file will be NULL in the error
1315 * cleanup path of mmap_region. When
1316 * hugetlbfs ->mmap method fails,
1317 * mmap_region() nullifies vma->vm_file
1318 * before calling this function to clean up.
1319 * Since no pte has actually been setup, it is
1320 * safe to do nothing in this case.
1321 */
1322 if (vma->vm_file) {
1323 i_mmap_lock_write(vma->vm_file->f_mapping);
1324 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1325 i_mmap_unlock_write(vma->vm_file->f_mapping);
1326 }
1327 } else
1328 unmap_page_range(tlb, vma, start, end, details);
1329 }
1330}
1331
1332/**
1333 * unmap_vmas - unmap a range of memory covered by a list of vma's
1334 * @tlb: address of the caller's struct mmu_gather
1335 * @vma: the starting vma
1336 * @start_addr: virtual address at which to start unmapping
1337 * @end_addr: virtual address at which to end unmapping
1338 *
1339 * Unmap all pages in the vma list.
1340 *
1341 * Only addresses between `start' and `end' will be unmapped.
1342 *
1343 * The VMA list must be sorted in ascending virtual address order.
1344 *
1345 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1346 * range after unmap_vmas() returns. So the only responsibility here is to
1347 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1348 * drops the lock and schedules.
1349 */
1350void unmap_vmas(struct mmu_gather *tlb,
1351 struct vm_area_struct *vma, unsigned long start_addr,
1352 unsigned long end_addr)
1353{
1354 struct mm_struct *mm = vma->vm_mm;
1355
1356 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1357 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1358 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1359 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1360}
1361
1362/**
1363 * zap_page_range - remove user pages in a given range
1364 * @vma: vm_area_struct holding the applicable pages
1365 * @start: starting address of pages to zap
1366 * @size: number of bytes to zap
1367 * @details: details of shared cache invalidation
1368 *
1369 * Caller must protect the VMA list
1370 */
1371void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1372 unsigned long size, struct zap_details *details)
1373{
1374 struct mm_struct *mm = vma->vm_mm;
1375 struct mmu_gather tlb;
1376 unsigned long end = start + size;
1377
1378 lru_add_drain();
1379 tlb_gather_mmu(&tlb, mm, start, end);
1380 update_hiwater_rss(mm);
1381 mmu_notifier_invalidate_range_start(mm, start, end);
1382 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1383 unmap_single_vma(&tlb, vma, start, end, details);
1384 mmu_notifier_invalidate_range_end(mm, start, end);
1385 tlb_finish_mmu(&tlb, start, end);
1386}
1387
1388/**
1389 * zap_page_range_single - remove user pages in a given range
1390 * @vma: vm_area_struct holding the applicable pages
1391 * @address: starting address of pages to zap
1392 * @size: number of bytes to zap
1393 * @details: details of shared cache invalidation
1394 *
1395 * The range must fit into one VMA.
1396 */
1397static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1398 unsigned long size, struct zap_details *details)
1399{
1400 struct mm_struct *mm = vma->vm_mm;
1401 struct mmu_gather tlb;
1402 unsigned long end = address + size;
1403
1404 lru_add_drain();
1405 tlb_gather_mmu(&tlb, mm, address, end);
1406 update_hiwater_rss(mm);
1407 mmu_notifier_invalidate_range_start(mm, address, end);
1408 unmap_single_vma(&tlb, vma, address, end, details);
1409 mmu_notifier_invalidate_range_end(mm, address, end);
1410 tlb_finish_mmu(&tlb, address, end);
1411}
1412
1413/**
1414 * zap_vma_ptes - remove ptes mapping the vma
1415 * @vma: vm_area_struct holding ptes to be zapped
1416 * @address: starting address of pages to zap
1417 * @size: number of bytes to zap
1418 *
1419 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1420 *
1421 * The entire address range must be fully contained within the vma.
1422 *
1423 * Returns 0 if successful.
1424 */
1425int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1426 unsigned long size)
1427{
1428 if (address < vma->vm_start || address + size > vma->vm_end ||
1429 !(vma->vm_flags & VM_PFNMAP))
1430 return -1;
1431 zap_page_range_single(vma, address, size, NULL);
1432 return 0;
1433}
1434EXPORT_SYMBOL_GPL(zap_vma_ptes);
1435
1436pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1437 spinlock_t **ptl)
1438{
1439 pgd_t * pgd = pgd_offset(mm, addr);
1440 pud_t * pud = pud_alloc(mm, pgd, addr);
1441 if (pud) {
1442 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1443 if (pmd) {
1444 VM_BUG_ON(pmd_trans_huge(*pmd));
1445 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1446 }
1447 }
1448 return NULL;
1449}
1450
1451/*
1452 * This is the old fallback for page remapping.
1453 *
1454 * For historical reasons, it only allows reserved pages. Only
1455 * old drivers should use this, and they needed to mark their
1456 * pages reserved for the old functions anyway.
1457 */
1458static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1459 struct page *page, pgprot_t prot)
1460{
1461 struct mm_struct *mm = vma->vm_mm;
1462 int retval;
1463 pte_t *pte;
1464 spinlock_t *ptl;
1465
1466 retval = -EINVAL;
1467 if (PageAnon(page))
1468 goto out;
1469 retval = -ENOMEM;
1470 flush_dcache_page(page);
1471 pte = get_locked_pte(mm, addr, &ptl);
1472 if (!pte)
1473 goto out;
1474 retval = -EBUSY;
1475 if (!pte_none(*pte))
1476 goto out_unlock;
1477
1478 /* Ok, finally just insert the thing.. */
1479 get_page(page);
1480 inc_mm_counter_fast(mm, mm_counter_file(page));
1481 page_add_file_rmap(page);
1482 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1483
1484 retval = 0;
1485 pte_unmap_unlock(pte, ptl);
1486 return retval;
1487out_unlock:
1488 pte_unmap_unlock(pte, ptl);
1489out:
1490 return retval;
1491}
1492
1493/**
1494 * vm_insert_page - insert single page into user vma
1495 * @vma: user vma to map to
1496 * @addr: target user address of this page
1497 * @page: source kernel page
1498 *
1499 * This allows drivers to insert individual pages they've allocated
1500 * into a user vma.
1501 *
1502 * The page has to be a nice clean _individual_ kernel allocation.
1503 * If you allocate a compound page, you need to have marked it as
1504 * such (__GFP_COMP), or manually just split the page up yourself
1505 * (see split_page()).
1506 *
1507 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1508 * took an arbitrary page protection parameter. This doesn't allow
1509 * that. Your vma protection will have to be set up correctly, which
1510 * means that if you want a shared writable mapping, you'd better
1511 * ask for a shared writable mapping!
1512 *
1513 * The page does not need to be reserved.
1514 *
1515 * Usually this function is called from f_op->mmap() handler
1516 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1517 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1518 * function from other places, for example from page-fault handler.
1519 */
1520int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1521 struct page *page)
1522{
1523 if (addr < vma->vm_start || addr >= vma->vm_end)
1524 return -EFAULT;
1525 if (!page_count(page))
1526 return -EINVAL;
1527 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1528 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1529 BUG_ON(vma->vm_flags & VM_PFNMAP);
1530 vma->vm_flags |= VM_MIXEDMAP;
1531 }
1532 return insert_page(vma, addr, page, vma->vm_page_prot);
1533}
1534EXPORT_SYMBOL(vm_insert_page);
1535
1536static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1537 pfn_t pfn, pgprot_t prot)
1538{
1539 struct mm_struct *mm = vma->vm_mm;
1540 int retval;
1541 pte_t *pte, entry;
1542 spinlock_t *ptl;
1543
1544 retval = -ENOMEM;
1545 pte = get_locked_pte(mm, addr, &ptl);
1546 if (!pte)
1547 goto out;
1548 retval = -EBUSY;
1549 if (!pte_none(*pte))
1550 goto out_unlock;
1551
1552 /* Ok, finally just insert the thing.. */
1553 if (pfn_t_devmap(pfn))
1554 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1555 else
1556 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1557 set_pte_at(mm, addr, pte, entry);
1558 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1559
1560 retval = 0;
1561out_unlock:
1562 pte_unmap_unlock(pte, ptl);
1563out:
1564 return retval;
1565}
1566
1567/**
1568 * vm_insert_pfn - insert single pfn into user vma
1569 * @vma: user vma to map to
1570 * @addr: target user address of this page
1571 * @pfn: source kernel pfn
1572 *
1573 * Similar to vm_insert_page, this allows drivers to insert individual pages
1574 * they've allocated into a user vma. Same comments apply.
1575 *
1576 * This function should only be called from a vm_ops->fault handler, and
1577 * in that case the handler should return NULL.
1578 *
1579 * vma cannot be a COW mapping.
1580 *
1581 * As this is called only for pages that do not currently exist, we
1582 * do not need to flush old virtual caches or the TLB.
1583 */
1584int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1585 unsigned long pfn)
1586{
1587 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1588}
1589EXPORT_SYMBOL(vm_insert_pfn);
1590
1591/**
1592 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1593 * @vma: user vma to map to
1594 * @addr: target user address of this page
1595 * @pfn: source kernel pfn
1596 * @pgprot: pgprot flags for the inserted page
1597 *
1598 * This is exactly like vm_insert_pfn, except that it allows drivers to
1599 * to override pgprot on a per-page basis.
1600 *
1601 * This only makes sense for IO mappings, and it makes no sense for
1602 * cow mappings. In general, using multiple vmas is preferable;
1603 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1604 * impractical.
1605 */
1606int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1607 unsigned long pfn, pgprot_t pgprot)
1608{
1609 int ret;
1610 /*
1611 * Technically, architectures with pte_special can avoid all these
1612 * restrictions (same for remap_pfn_range). However we would like
1613 * consistency in testing and feature parity among all, so we should
1614 * try to keep these invariants in place for everybody.
1615 */
1616 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1617 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1618 (VM_PFNMAP|VM_MIXEDMAP));
1619 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1620 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1621
1622 if (addr < vma->vm_start || addr >= vma->vm_end)
1623 return -EFAULT;
1624 if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
1625 return -EINVAL;
1626
1627 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1628
1629 return ret;
1630}
1631EXPORT_SYMBOL(vm_insert_pfn_prot);
1632
1633int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1634 pfn_t pfn)
1635{
1636 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1637
1638 if (addr < vma->vm_start || addr >= vma->vm_end)
1639 return -EFAULT;
1640
1641 /*
1642 * If we don't have pte special, then we have to use the pfn_valid()
1643 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1644 * refcount the page if pfn_valid is true (hence insert_page rather
1645 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1646 * without pte special, it would there be refcounted as a normal page.
1647 */
1648 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1649 struct page *page;
1650
1651 /*
1652 * At this point we are committed to insert_page()
1653 * regardless of whether the caller specified flags that
1654 * result in pfn_t_has_page() == false.
1655 */
1656 page = pfn_to_page(pfn_t_to_pfn(pfn));
1657 return insert_page(vma, addr, page, vma->vm_page_prot);
1658 }
1659 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1660}
1661EXPORT_SYMBOL(vm_insert_mixed);
1662
1663/*
1664 * maps a range of physical memory into the requested pages. the old
1665 * mappings are removed. any references to nonexistent pages results
1666 * in null mappings (currently treated as "copy-on-access")
1667 */
1668static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1669 unsigned long addr, unsigned long end,
1670 unsigned long pfn, pgprot_t prot)
1671{
1672 pte_t *pte;
1673 spinlock_t *ptl;
1674
1675 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1676 if (!pte)
1677 return -ENOMEM;
1678 arch_enter_lazy_mmu_mode();
1679 do {
1680 BUG_ON(!pte_none(*pte));
1681 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1682 pfn++;
1683 } while (pte++, addr += PAGE_SIZE, addr != end);
1684 arch_leave_lazy_mmu_mode();
1685 pte_unmap_unlock(pte - 1, ptl);
1686 return 0;
1687}
1688
1689static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1690 unsigned long addr, unsigned long end,
1691 unsigned long pfn, pgprot_t prot)
1692{
1693 pmd_t *pmd;
1694 unsigned long next;
1695
1696 pfn -= addr >> PAGE_SHIFT;
1697 pmd = pmd_alloc(mm, pud, addr);
1698 if (!pmd)
1699 return -ENOMEM;
1700 VM_BUG_ON(pmd_trans_huge(*pmd));
1701 do {
1702 next = pmd_addr_end(addr, end);
1703 if (remap_pte_range(mm, pmd, addr, next,
1704 pfn + (addr >> PAGE_SHIFT), prot))
1705 return -ENOMEM;
1706 } while (pmd++, addr = next, addr != end);
1707 return 0;
1708}
1709
1710static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1711 unsigned long addr, unsigned long end,
1712 unsigned long pfn, pgprot_t prot)
1713{
1714 pud_t *pud;
1715 unsigned long next;
1716
1717 pfn -= addr >> PAGE_SHIFT;
1718 pud = pud_alloc(mm, pgd, addr);
1719 if (!pud)
1720 return -ENOMEM;
1721 do {
1722 next = pud_addr_end(addr, end);
1723 if (remap_pmd_range(mm, pud, addr, next,
1724 pfn + (addr >> PAGE_SHIFT), prot))
1725 return -ENOMEM;
1726 } while (pud++, addr = next, addr != end);
1727 return 0;
1728}
1729
1730/**
1731 * remap_pfn_range - remap kernel memory to userspace
1732 * @vma: user vma to map to
1733 * @addr: target user address to start at
1734 * @pfn: physical address of kernel memory
1735 * @size: size of map area
1736 * @prot: page protection flags for this mapping
1737 *
1738 * Note: this is only safe if the mm semaphore is held when called.
1739 */
1740int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1741 unsigned long pfn, unsigned long size, pgprot_t prot)
1742{
1743 pgd_t *pgd;
1744 unsigned long next;
1745 unsigned long end = addr + PAGE_ALIGN(size);
1746 struct mm_struct *mm = vma->vm_mm;
1747 int err;
1748
1749 /*
1750 * Physically remapped pages are special. Tell the
1751 * rest of the world about it:
1752 * VM_IO tells people not to look at these pages
1753 * (accesses can have side effects).
1754 * VM_PFNMAP tells the core MM that the base pages are just
1755 * raw PFN mappings, and do not have a "struct page" associated
1756 * with them.
1757 * VM_DONTEXPAND
1758 * Disable vma merging and expanding with mremap().
1759 * VM_DONTDUMP
1760 * Omit vma from core dump, even when VM_IO turned off.
1761 *
1762 * There's a horrible special case to handle copy-on-write
1763 * behaviour that some programs depend on. We mark the "original"
1764 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1765 * See vm_normal_page() for details.
1766 */
1767 if (is_cow_mapping(vma->vm_flags)) {
1768 if (addr != vma->vm_start || end != vma->vm_end)
1769 return -EINVAL;
1770 vma->vm_pgoff = pfn;
1771 }
1772
1773 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1774 if (err)
1775 return -EINVAL;
1776
1777 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1778
1779 BUG_ON(addr >= end);
1780 pfn -= addr >> PAGE_SHIFT;
1781 pgd = pgd_offset(mm, addr);
1782 flush_cache_range(vma, addr, end);
1783 do {
1784 next = pgd_addr_end(addr, end);
1785 err = remap_pud_range(mm, pgd, addr, next,
1786 pfn + (addr >> PAGE_SHIFT), prot);
1787 if (err)
1788 break;
1789 } while (pgd++, addr = next, addr != end);
1790
1791 if (err)
1792 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1793
1794 return err;
1795}
1796EXPORT_SYMBOL(remap_pfn_range);
1797
1798/**
1799 * vm_iomap_memory - remap memory to userspace
1800 * @vma: user vma to map to
1801 * @start: start of area
1802 * @len: size of area
1803 *
1804 * This is a simplified io_remap_pfn_range() for common driver use. The
1805 * driver just needs to give us the physical memory range to be mapped,
1806 * we'll figure out the rest from the vma information.
1807 *
1808 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1809 * whatever write-combining details or similar.
1810 */
1811int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1812{
1813 unsigned long vm_len, pfn, pages;
1814
1815 /* Check that the physical memory area passed in looks valid */
1816 if (start + len < start)
1817 return -EINVAL;
1818 /*
1819 * You *really* shouldn't map things that aren't page-aligned,
1820 * but we've historically allowed it because IO memory might
1821 * just have smaller alignment.
1822 */
1823 len += start & ~PAGE_MASK;
1824 pfn = start >> PAGE_SHIFT;
1825 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1826 if (pfn + pages < pfn)
1827 return -EINVAL;
1828
1829 /* We start the mapping 'vm_pgoff' pages into the area */
1830 if (vma->vm_pgoff > pages)
1831 return -EINVAL;
1832 pfn += vma->vm_pgoff;
1833 pages -= vma->vm_pgoff;
1834
1835 /* Can we fit all of the mapping? */
1836 vm_len = vma->vm_end - vma->vm_start;
1837 if (vm_len >> PAGE_SHIFT > pages)
1838 return -EINVAL;
1839
1840 /* Ok, let it rip */
1841 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1842}
1843EXPORT_SYMBOL(vm_iomap_memory);
1844
1845static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1846 unsigned long addr, unsigned long end,
1847 pte_fn_t fn, void *data)
1848{
1849 pte_t *pte;
1850 int err;
1851 pgtable_t token;
1852 spinlock_t *uninitialized_var(ptl);
1853
1854 pte = (mm == &init_mm) ?
1855 pte_alloc_kernel(pmd, addr) :
1856 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1857 if (!pte)
1858 return -ENOMEM;
1859
1860 BUG_ON(pmd_huge(*pmd));
1861
1862 arch_enter_lazy_mmu_mode();
1863
1864 token = pmd_pgtable(*pmd);
1865
1866 do {
1867 err = fn(pte++, token, addr, data);
1868 if (err)
1869 break;
1870 } while (addr += PAGE_SIZE, addr != end);
1871
1872 arch_leave_lazy_mmu_mode();
1873
1874 if (mm != &init_mm)
1875 pte_unmap_unlock(pte-1, ptl);
1876 return err;
1877}
1878
1879static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1880 unsigned long addr, unsigned long end,
1881 pte_fn_t fn, void *data)
1882{
1883 pmd_t *pmd;
1884 unsigned long next;
1885 int err;
1886
1887 BUG_ON(pud_huge(*pud));
1888
1889 pmd = pmd_alloc(mm, pud, addr);
1890 if (!pmd)
1891 return -ENOMEM;
1892 do {
1893 next = pmd_addr_end(addr, end);
1894 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1895 if (err)
1896 break;
1897 } while (pmd++, addr = next, addr != end);
1898 return err;
1899}
1900
1901static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1902 unsigned long addr, unsigned long end,
1903 pte_fn_t fn, void *data)
1904{
1905 pud_t *pud;
1906 unsigned long next;
1907 int err;
1908
1909 pud = pud_alloc(mm, pgd, addr);
1910 if (!pud)
1911 return -ENOMEM;
1912 do {
1913 next = pud_addr_end(addr, end);
1914 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1915 if (err)
1916 break;
1917 } while (pud++, addr = next, addr != end);
1918 return err;
1919}
1920
1921/*
1922 * Scan a region of virtual memory, filling in page tables as necessary
1923 * and calling a provided function on each leaf page table.
1924 */
1925int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1926 unsigned long size, pte_fn_t fn, void *data)
1927{
1928 pgd_t *pgd;
1929 unsigned long next;
1930 unsigned long end = addr + size;
1931 int err;
1932
1933 if (WARN_ON(addr >= end))
1934 return -EINVAL;
1935
1936 pgd = pgd_offset(mm, addr);
1937 do {
1938 next = pgd_addr_end(addr, end);
1939 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1940 if (err)
1941 break;
1942 } while (pgd++, addr = next, addr != end);
1943
1944 return err;
1945}
1946EXPORT_SYMBOL_GPL(apply_to_page_range);
1947
1948/*
1949 * handle_pte_fault chooses page fault handler according to an entry which was
1950 * read non-atomically. Before making any commitment, on those architectures
1951 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1952 * parts, do_swap_page must check under lock before unmapping the pte and
1953 * proceeding (but do_wp_page is only called after already making such a check;
1954 * and do_anonymous_page can safely check later on).
1955 */
1956static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1957 pte_t *page_table, pte_t orig_pte)
1958{
1959 int same = 1;
1960#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1961 if (sizeof(pte_t) > sizeof(unsigned long)) {
1962 spinlock_t *ptl = pte_lockptr(mm, pmd);
1963 spin_lock(ptl);
1964 same = pte_same(*page_table, orig_pte);
1965 spin_unlock(ptl);
1966 }
1967#endif
1968 pte_unmap(page_table);
1969 return same;
1970}
1971
1972static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1973{
1974 debug_dma_assert_idle(src);
1975
1976 /*
1977 * If the source page was a PFN mapping, we don't have
1978 * a "struct page" for it. We do a best-effort copy by
1979 * just copying from the original user address. If that
1980 * fails, we just zero-fill it. Live with it.
1981 */
1982 if (unlikely(!src)) {
1983 void *kaddr = kmap_atomic(dst);
1984 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1985
1986 /*
1987 * This really shouldn't fail, because the page is there
1988 * in the page tables. But it might just be unreadable,
1989 * in which case we just give up and fill the result with
1990 * zeroes.
1991 */
1992 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1993 clear_page(kaddr);
1994 kunmap_atomic(kaddr);
1995 flush_dcache_page(dst);
1996 } else
1997 copy_user_highpage(dst, src, va, vma);
1998}
1999
2000static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2001{
2002 struct file *vm_file = vma->vm_file;
2003
2004 if (vm_file)
2005 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2006
2007 /*
2008 * Special mappings (e.g. VDSO) do not have any file so fake
2009 * a default GFP_KERNEL for them.
2010 */
2011 return GFP_KERNEL;
2012}
2013
2014/*
2015 * Notify the address space that the page is about to become writable so that
2016 * it can prohibit this or wait for the page to get into an appropriate state.
2017 *
2018 * We do this without the lock held, so that it can sleep if it needs to.
2019 */
2020static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2021 unsigned long address)
2022{
2023 struct vm_fault vmf;
2024 int ret;
2025
2026 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2027 vmf.pgoff = page->index;
2028 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2029 vmf.gfp_mask = __get_fault_gfp_mask(vma);
2030 vmf.page = page;
2031 vmf.cow_page = NULL;
2032
2033 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2034 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2035 return ret;
2036 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2037 lock_page(page);
2038 if (!page->mapping) {
2039 unlock_page(page);
2040 return 0; /* retry */
2041 }
2042 ret |= VM_FAULT_LOCKED;
2043 } else
2044 VM_BUG_ON_PAGE(!PageLocked(page), page);
2045 return ret;
2046}
2047
2048/*
2049 * Handle write page faults for pages that can be reused in the current vma
2050 *
2051 * This can happen either due to the mapping being with the VM_SHARED flag,
2052 * or due to us being the last reference standing to the page. In either
2053 * case, all we need to do here is to mark the page as writable and update
2054 * any related book-keeping.
2055 */
2056static inline int wp_page_reuse(struct mm_struct *mm,
2057 struct vm_area_struct *vma, unsigned long address,
2058 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2059 struct page *page, int page_mkwrite,
2060 int dirty_shared)
2061 __releases(ptl)
2062{
2063 pte_t entry;
2064 /*
2065 * Clear the pages cpupid information as the existing
2066 * information potentially belongs to a now completely
2067 * unrelated process.
2068 */
2069 if (page)
2070 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2071
2072 flush_cache_page(vma, address, pte_pfn(orig_pte));
2073 entry = pte_mkyoung(orig_pte);
2074 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2075 if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2076 update_mmu_cache(vma, address, page_table);
2077 pte_unmap_unlock(page_table, ptl);
2078
2079 if (dirty_shared) {
2080 struct address_space *mapping;
2081 int dirtied;
2082
2083 if (!page_mkwrite)
2084 lock_page(page);
2085
2086 dirtied = set_page_dirty(page);
2087 VM_BUG_ON_PAGE(PageAnon(page), page);
2088 mapping = page->mapping;
2089 unlock_page(page);
2090 put_page(page);
2091
2092 if ((dirtied || page_mkwrite) && mapping) {
2093 /*
2094 * Some device drivers do not set page.mapping
2095 * but still dirty their pages
2096 */
2097 balance_dirty_pages_ratelimited(mapping);
2098 }
2099
2100 if (!page_mkwrite)
2101 file_update_time(vma->vm_file);
2102 }
2103
2104 return VM_FAULT_WRITE;
2105}
2106
2107/*
2108 * Handle the case of a page which we actually need to copy to a new page.
2109 *
2110 * Called with mmap_sem locked and the old page referenced, but
2111 * without the ptl held.
2112 *
2113 * High level logic flow:
2114 *
2115 * - Allocate a page, copy the content of the old page to the new one.
2116 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2117 * - Take the PTL. If the pte changed, bail out and release the allocated page
2118 * - If the pte is still the way we remember it, update the page table and all
2119 * relevant references. This includes dropping the reference the page-table
2120 * held to the old page, as well as updating the rmap.
2121 * - In any case, unlock the PTL and drop the reference we took to the old page.
2122 */
2123static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2124 unsigned long address, pte_t *page_table, pmd_t *pmd,
2125 pte_t orig_pte, struct page *old_page)
2126{
2127 struct page *new_page = NULL;
2128 spinlock_t *ptl = NULL;
2129 pte_t entry;
2130 int page_copied = 0;
2131 const unsigned long mmun_start = address & PAGE_MASK; /* For mmu_notifiers */
2132 const unsigned long mmun_end = mmun_start + PAGE_SIZE; /* For mmu_notifiers */
2133 struct mem_cgroup *memcg;
2134
2135 if (unlikely(anon_vma_prepare(vma)))
2136 goto oom;
2137
2138 if (is_zero_pfn(pte_pfn(orig_pte))) {
2139 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2140 if (!new_page)
2141 goto oom;
2142 } else {
2143 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2144 if (!new_page)
2145 goto oom;
2146 cow_user_page(new_page, old_page, address, vma);
2147 }
2148
2149 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2150 goto oom_free_new;
2151
2152 __SetPageUptodate(new_page);
2153
2154 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2155
2156 /*
2157 * Re-check the pte - we dropped the lock
2158 */
2159 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2160 if (likely(pte_same(*page_table, orig_pte))) {
2161 if (old_page) {
2162 if (!PageAnon(old_page)) {
2163 dec_mm_counter_fast(mm,
2164 mm_counter_file(old_page));
2165 inc_mm_counter_fast(mm, MM_ANONPAGES);
2166 }
2167 } else {
2168 inc_mm_counter_fast(mm, MM_ANONPAGES);
2169 }
2170 flush_cache_page(vma, address, pte_pfn(orig_pte));
2171 entry = mk_pte(new_page, vma->vm_page_prot);
2172 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2173 /*
2174 * Clear the pte entry and flush it first, before updating the
2175 * pte with the new entry. This will avoid a race condition
2176 * seen in the presence of one thread doing SMC and another
2177 * thread doing COW.
2178 */
2179 ptep_clear_flush_notify(vma, address, page_table);
2180 page_add_new_anon_rmap(new_page, vma, address, false);
2181 mem_cgroup_commit_charge(new_page, memcg, false, false);
2182 lru_cache_add_active_or_unevictable(new_page, vma);
2183 /*
2184 * We call the notify macro here because, when using secondary
2185 * mmu page tables (such as kvm shadow page tables), we want the
2186 * new page to be mapped directly into the secondary page table.
2187 */
2188 set_pte_at_notify(mm, address, page_table, entry);
2189 update_mmu_cache(vma, address, page_table);
2190 if (old_page) {
2191 /*
2192 * Only after switching the pte to the new page may
2193 * we remove the mapcount here. Otherwise another
2194 * process may come and find the rmap count decremented
2195 * before the pte is switched to the new page, and
2196 * "reuse" the old page writing into it while our pte
2197 * here still points into it and can be read by other
2198 * threads.
2199 *
2200 * The critical issue is to order this
2201 * page_remove_rmap with the ptp_clear_flush above.
2202 * Those stores are ordered by (if nothing else,)
2203 * the barrier present in the atomic_add_negative
2204 * in page_remove_rmap.
2205 *
2206 * Then the TLB flush in ptep_clear_flush ensures that
2207 * no process can access the old page before the
2208 * decremented mapcount is visible. And the old page
2209 * cannot be reused until after the decremented
2210 * mapcount is visible. So transitively, TLBs to
2211 * old page will be flushed before it can be reused.
2212 */
2213 page_remove_rmap(old_page, false);
2214 }
2215
2216 /* Free the old page.. */
2217 new_page = old_page;
2218 page_copied = 1;
2219 } else {
2220 mem_cgroup_cancel_charge(new_page, memcg, false);
2221 }
2222
2223 if (new_page)
2224 put_page(new_page);
2225
2226 pte_unmap_unlock(page_table, ptl);
2227 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2228 if (old_page) {
2229 /*
2230 * Don't let another task, with possibly unlocked vma,
2231 * keep the mlocked page.
2232 */
2233 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2234 lock_page(old_page); /* LRU manipulation */
2235 if (PageMlocked(old_page))
2236 munlock_vma_page(old_page);
2237 unlock_page(old_page);
2238 }
2239 put_page(old_page);
2240 }
2241 return page_copied ? VM_FAULT_WRITE : 0;
2242oom_free_new:
2243 put_page(new_page);
2244oom:
2245 if (old_page)
2246 put_page(old_page);
2247 return VM_FAULT_OOM;
2248}
2249
2250/*
2251 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2252 * mapping
2253 */
2254static int wp_pfn_shared(struct mm_struct *mm,
2255 struct vm_area_struct *vma, unsigned long address,
2256 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2257 pmd_t *pmd)
2258{
2259 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2260 struct vm_fault vmf = {
2261 .page = NULL,
2262 .pgoff = linear_page_index(vma, address),
2263 .virtual_address = (void __user *)(address & PAGE_MASK),
2264 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2265 };
2266 int ret;
2267
2268 pte_unmap_unlock(page_table, ptl);
2269 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2270 if (ret & VM_FAULT_ERROR)
2271 return ret;
2272 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2273 /*
2274 * We might have raced with another page fault while we
2275 * released the pte_offset_map_lock.
2276 */
2277 if (!pte_same(*page_table, orig_pte)) {
2278 pte_unmap_unlock(page_table, ptl);
2279 return 0;
2280 }
2281 }
2282 return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2283 NULL, 0, 0);
2284}
2285
2286static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2287 unsigned long address, pte_t *page_table,
2288 pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2289 struct page *old_page)
2290 __releases(ptl)
2291{
2292 int page_mkwrite = 0;
2293
2294 get_page(old_page);
2295
2296 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2297 int tmp;
2298
2299 pte_unmap_unlock(page_table, ptl);
2300 tmp = do_page_mkwrite(vma, old_page, address);
2301 if (unlikely(!tmp || (tmp &
2302 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2303 put_page(old_page);
2304 return tmp;
2305 }
2306 /*
2307 * Since we dropped the lock we need to revalidate
2308 * the PTE as someone else may have changed it. If
2309 * they did, we just return, as we can count on the
2310 * MMU to tell us if they didn't also make it writable.
2311 */
2312 page_table = pte_offset_map_lock(mm, pmd, address,
2313 &ptl);
2314 if (!pte_same(*page_table, orig_pte)) {
2315 unlock_page(old_page);
2316 pte_unmap_unlock(page_table, ptl);
2317 put_page(old_page);
2318 return 0;
2319 }
2320 page_mkwrite = 1;
2321 }
2322
2323 return wp_page_reuse(mm, vma, address, page_table, ptl,
2324 orig_pte, old_page, page_mkwrite, 1);
2325}
2326
2327/*
2328 * This routine handles present pages, when users try to write
2329 * to a shared page. It is done by copying the page to a new address
2330 * and decrementing the shared-page counter for the old page.
2331 *
2332 * Note that this routine assumes that the protection checks have been
2333 * done by the caller (the low-level page fault routine in most cases).
2334 * Thus we can safely just mark it writable once we've done any necessary
2335 * COW.
2336 *
2337 * We also mark the page dirty at this point even though the page will
2338 * change only once the write actually happens. This avoids a few races,
2339 * and potentially makes it more efficient.
2340 *
2341 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2342 * but allow concurrent faults), with pte both mapped and locked.
2343 * We return with mmap_sem still held, but pte unmapped and unlocked.
2344 */
2345static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2346 unsigned long address, pte_t *page_table, pmd_t *pmd,
2347 spinlock_t *ptl, pte_t orig_pte)
2348 __releases(ptl)
2349{
2350 struct page *old_page;
2351
2352 old_page = vm_normal_page(vma, address, orig_pte);
2353 if (!old_page) {
2354 /*
2355 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2356 * VM_PFNMAP VMA.
2357 *
2358 * We should not cow pages in a shared writeable mapping.
2359 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2360 */
2361 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2362 (VM_WRITE|VM_SHARED))
2363 return wp_pfn_shared(mm, vma, address, page_table, ptl,
2364 orig_pte, pmd);
2365
2366 pte_unmap_unlock(page_table, ptl);
2367 return wp_page_copy(mm, vma, address, page_table, pmd,
2368 orig_pte, old_page);
2369 }
2370
2371 /*
2372 * Take out anonymous pages first, anonymous shared vmas are
2373 * not dirty accountable.
2374 */
2375 if (PageAnon(old_page) && !PageKsm(old_page)) {
2376 int total_mapcount;
2377 if (!trylock_page(old_page)) {
2378 get_page(old_page);
2379 pte_unmap_unlock(page_table, ptl);
2380 lock_page(old_page);
2381 page_table = pte_offset_map_lock(mm, pmd, address,
2382 &ptl);
2383 if (!pte_same(*page_table, orig_pte)) {
2384 unlock_page(old_page);
2385 pte_unmap_unlock(page_table, ptl);
2386 put_page(old_page);
2387 return 0;
2388 }
2389 put_page(old_page);
2390 }
2391 if (reuse_swap_page(old_page, &total_mapcount)) {
2392 if (total_mapcount == 1) {
2393 /*
2394 * The page is all ours. Move it to
2395 * our anon_vma so the rmap code will
2396 * not search our parent or siblings.
2397 * Protected against the rmap code by
2398 * the page lock.
2399 */
2400 page_move_anon_rmap(compound_head(old_page),
2401 vma, address);
2402 }
2403 unlock_page(old_page);
2404 return wp_page_reuse(mm, vma, address, page_table, ptl,
2405 orig_pte, old_page, 0, 0);
2406 }
2407 unlock_page(old_page);
2408 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2409 (VM_WRITE|VM_SHARED))) {
2410 return wp_page_shared(mm, vma, address, page_table, pmd,
2411 ptl, orig_pte, old_page);
2412 }
2413
2414 /*
2415 * Ok, we need to copy. Oh, well..
2416 */
2417 get_page(old_page);
2418
2419 pte_unmap_unlock(page_table, ptl);
2420 return wp_page_copy(mm, vma, address, page_table, pmd,
2421 orig_pte, old_page);
2422}
2423
2424static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2425 unsigned long start_addr, unsigned long end_addr,
2426 struct zap_details *details)
2427{
2428 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2429}
2430
2431static inline void unmap_mapping_range_tree(struct rb_root *root,
2432 struct zap_details *details)
2433{
2434 struct vm_area_struct *vma;
2435 pgoff_t vba, vea, zba, zea;
2436
2437 vma_interval_tree_foreach(vma, root,
2438 details->first_index, details->last_index) {
2439
2440 vba = vma->vm_pgoff;
2441 vea = vba + vma_pages(vma) - 1;
2442 zba = details->first_index;
2443 if (zba < vba)
2444 zba = vba;
2445 zea = details->last_index;
2446 if (zea > vea)
2447 zea = vea;
2448
2449 unmap_mapping_range_vma(vma,
2450 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2451 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2452 details);
2453 }
2454}
2455
2456/**
2457 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2458 * address_space corresponding to the specified page range in the underlying
2459 * file.
2460 *
2461 * @mapping: the address space containing mmaps to be unmapped.
2462 * @holebegin: byte in first page to unmap, relative to the start of
2463 * the underlying file. This will be rounded down to a PAGE_SIZE
2464 * boundary. Note that this is different from truncate_pagecache(), which
2465 * must keep the partial page. In contrast, we must get rid of
2466 * partial pages.
2467 * @holelen: size of prospective hole in bytes. This will be rounded
2468 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2469 * end of the file.
2470 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2471 * but 0 when invalidating pagecache, don't throw away private data.
2472 */
2473void unmap_mapping_range(struct address_space *mapping,
2474 loff_t const holebegin, loff_t const holelen, int even_cows)
2475{
2476 struct zap_details details = { };
2477 pgoff_t hba = holebegin >> PAGE_SHIFT;
2478 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2479
2480 /* Check for overflow. */
2481 if (sizeof(holelen) > sizeof(hlen)) {
2482 long long holeend =
2483 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2484 if (holeend & ~(long long)ULONG_MAX)
2485 hlen = ULONG_MAX - hba + 1;
2486 }
2487
2488 details.check_mapping = even_cows? NULL: mapping;
2489 details.first_index = hba;
2490 details.last_index = hba + hlen - 1;
2491 if (details.last_index < details.first_index)
2492 details.last_index = ULONG_MAX;
2493
2494
2495 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2496 i_mmap_lock_write(mapping);
2497 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2498 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2499 i_mmap_unlock_write(mapping);
2500}
2501EXPORT_SYMBOL(unmap_mapping_range);
2502
2503/*
2504 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2505 * but allow concurrent faults), and pte mapped but not yet locked.
2506 * We return with pte unmapped and unlocked.
2507 *
2508 * We return with the mmap_sem locked or unlocked in the same cases
2509 * as does filemap_fault().
2510 */
2511static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2512 unsigned long address, pte_t *page_table, pmd_t *pmd,
2513 unsigned int flags, pte_t orig_pte)
2514{
2515 spinlock_t *ptl;
2516 struct page *page, *swapcache;
2517 struct mem_cgroup *memcg;
2518 swp_entry_t entry;
2519 pte_t pte;
2520 int locked;
2521 int exclusive = 0;
2522 int ret = 0;
2523
2524 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2525 goto out;
2526
2527 entry = pte_to_swp_entry(orig_pte);
2528 if (unlikely(non_swap_entry(entry))) {
2529 if (is_migration_entry(entry)) {
2530 migration_entry_wait(mm, pmd, address);
2531 } else if (is_hwpoison_entry(entry)) {
2532 ret = VM_FAULT_HWPOISON;
2533 } else {
2534 print_bad_pte(vma, address, orig_pte, NULL);
2535 ret = VM_FAULT_SIGBUS;
2536 }
2537 goto out;
2538 }
2539 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2540 page = lookup_swap_cache(entry);
2541 if (!page) {
2542 page = swapin_readahead(entry,
2543 GFP_HIGHUSER_MOVABLE, vma, address);
2544 if (!page) {
2545 /*
2546 * Back out if somebody else faulted in this pte
2547 * while we released the pte lock.
2548 */
2549 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2550 if (likely(pte_same(*page_table, orig_pte)))
2551 ret = VM_FAULT_OOM;
2552 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2553 goto unlock;
2554 }
2555
2556 /* Had to read the page from swap area: Major fault */
2557 ret = VM_FAULT_MAJOR;
2558 count_vm_event(PGMAJFAULT);
2559 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2560 } else if (PageHWPoison(page)) {
2561 /*
2562 * hwpoisoned dirty swapcache pages are kept for killing
2563 * owner processes (which may be unknown at hwpoison time)
2564 */
2565 ret = VM_FAULT_HWPOISON;
2566 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2567 swapcache = page;
2568 goto out_release;
2569 }
2570
2571 swapcache = page;
2572 locked = lock_page_or_retry(page, mm, flags);
2573
2574 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2575 if (!locked) {
2576 ret |= VM_FAULT_RETRY;
2577 goto out_release;
2578 }
2579
2580 /*
2581 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2582 * release the swapcache from under us. The page pin, and pte_same
2583 * test below, are not enough to exclude that. Even if it is still
2584 * swapcache, we need to check that the page's swap has not changed.
2585 */
2586 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2587 goto out_page;
2588
2589 page = ksm_might_need_to_copy(page, vma, address);
2590 if (unlikely(!page)) {
2591 ret = VM_FAULT_OOM;
2592 page = swapcache;
2593 goto out_page;
2594 }
2595
2596 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) {
2597 ret = VM_FAULT_OOM;
2598 goto out_page;
2599 }
2600
2601 /*
2602 * Back out if somebody else already faulted in this pte.
2603 */
2604 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2605 if (unlikely(!pte_same(*page_table, orig_pte)))
2606 goto out_nomap;
2607
2608 if (unlikely(!PageUptodate(page))) {
2609 ret = VM_FAULT_SIGBUS;
2610 goto out_nomap;
2611 }
2612
2613 /*
2614 * The page isn't present yet, go ahead with the fault.
2615 *
2616 * Be careful about the sequence of operations here.
2617 * To get its accounting right, reuse_swap_page() must be called
2618 * while the page is counted on swap but not yet in mapcount i.e.
2619 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2620 * must be called after the swap_free(), or it will never succeed.
2621 */
2622
2623 inc_mm_counter_fast(mm, MM_ANONPAGES);
2624 dec_mm_counter_fast(mm, MM_SWAPENTS);
2625 pte = mk_pte(page, vma->vm_page_prot);
2626 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2627 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2628 flags &= ~FAULT_FLAG_WRITE;
2629 ret |= VM_FAULT_WRITE;
2630 exclusive = RMAP_EXCLUSIVE;
2631 }
2632 flush_icache_page(vma, page);
2633 if (pte_swp_soft_dirty(orig_pte))
2634 pte = pte_mksoft_dirty(pte);
2635 set_pte_at(mm, address, page_table, pte);
2636 if (page == swapcache) {
2637 do_page_add_anon_rmap(page, vma, address, exclusive);
2638 mem_cgroup_commit_charge(page, memcg, true, false);
2639 } else { /* ksm created a completely new copy */
2640 page_add_new_anon_rmap(page, vma, address, false);
2641 mem_cgroup_commit_charge(page, memcg, false, false);
2642 lru_cache_add_active_or_unevictable(page, vma);
2643 }
2644
2645 swap_free(entry);
2646 if (mem_cgroup_swap_full(page) ||
2647 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2648 try_to_free_swap(page);
2649 unlock_page(page);
2650 if (page != swapcache) {
2651 /*
2652 * Hold the lock to avoid the swap entry to be reused
2653 * until we take the PT lock for the pte_same() check
2654 * (to avoid false positives from pte_same). For
2655 * further safety release the lock after the swap_free
2656 * so that the swap count won't change under a
2657 * parallel locked swapcache.
2658 */
2659 unlock_page(swapcache);
2660 put_page(swapcache);
2661 }
2662
2663 if (flags & FAULT_FLAG_WRITE) {
2664 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2665 if (ret & VM_FAULT_ERROR)
2666 ret &= VM_FAULT_ERROR;
2667 goto out;
2668 }
2669
2670 /* No need to invalidate - it was non-present before */
2671 update_mmu_cache(vma, address, page_table);
2672unlock:
2673 pte_unmap_unlock(page_table, ptl);
2674out:
2675 return ret;
2676out_nomap:
2677 mem_cgroup_cancel_charge(page, memcg, false);
2678 pte_unmap_unlock(page_table, ptl);
2679out_page:
2680 unlock_page(page);
2681out_release:
2682 put_page(page);
2683 if (page != swapcache) {
2684 unlock_page(swapcache);
2685 put_page(swapcache);
2686 }
2687 return ret;
2688}
2689
2690/*
2691 * This is like a special single-page "expand_{down|up}wards()",
2692 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2693 * doesn't hit another vma.
2694 */
2695static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2696{
2697 address &= PAGE_MASK;
2698 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2699 struct vm_area_struct *prev = vma->vm_prev;
2700
2701 /*
2702 * Is there a mapping abutting this one below?
2703 *
2704 * That's only ok if it's the same stack mapping
2705 * that has gotten split..
2706 */
2707 if (prev && prev->vm_end == address)
2708 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2709
2710 return expand_downwards(vma, address - PAGE_SIZE);
2711 }
2712 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2713 struct vm_area_struct *next = vma->vm_next;
2714
2715 /* As VM_GROWSDOWN but s/below/above/ */
2716 if (next && next->vm_start == address + PAGE_SIZE)
2717 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2718
2719 return expand_upwards(vma, address + PAGE_SIZE);
2720 }
2721 return 0;
2722}
2723
2724/*
2725 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2726 * but allow concurrent faults), and pte mapped but not yet locked.
2727 * We return with mmap_sem still held, but pte unmapped and unlocked.
2728 */
2729static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2730 unsigned long address, pte_t *page_table, pmd_t *pmd,
2731 unsigned int flags)
2732{
2733 struct mem_cgroup *memcg;
2734 struct page *page;
2735 spinlock_t *ptl;
2736 pte_t entry;
2737
2738 pte_unmap(page_table);
2739
2740 /* File mapping without ->vm_ops ? */
2741 if (vma->vm_flags & VM_SHARED)
2742 return VM_FAULT_SIGBUS;
2743
2744 /* Check if we need to add a guard page to the stack */
2745 if (check_stack_guard_page(vma, address) < 0)
2746 return VM_FAULT_SIGSEGV;
2747
2748 /* Use the zero-page for reads */
2749 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
2750 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2751 vma->vm_page_prot));
2752 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2753 if (!pte_none(*page_table))
2754 goto unlock;
2755 /* Deliver the page fault to userland, check inside PT lock */
2756 if (userfaultfd_missing(vma)) {
2757 pte_unmap_unlock(page_table, ptl);
2758 return handle_userfault(vma, address, flags,
2759 VM_UFFD_MISSING);
2760 }
2761 goto setpte;
2762 }
2763
2764 /* Allocate our own private page. */
2765 if (unlikely(anon_vma_prepare(vma)))
2766 goto oom;
2767 page = alloc_zeroed_user_highpage_movable(vma, address);
2768 if (!page)
2769 goto oom;
2770
2771 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false))
2772 goto oom_free_page;
2773
2774 /*
2775 * The memory barrier inside __SetPageUptodate makes sure that
2776 * preceeding stores to the page contents become visible before
2777 * the set_pte_at() write.
2778 */
2779 __SetPageUptodate(page);
2780
2781 entry = mk_pte(page, vma->vm_page_prot);
2782 if (vma->vm_flags & VM_WRITE)
2783 entry = pte_mkwrite(pte_mkdirty(entry));
2784
2785 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2786 if (!pte_none(*page_table))
2787 goto release;
2788
2789 /* Deliver the page fault to userland, check inside PT lock */
2790 if (userfaultfd_missing(vma)) {
2791 pte_unmap_unlock(page_table, ptl);
2792 mem_cgroup_cancel_charge(page, memcg, false);
2793 put_page(page);
2794 return handle_userfault(vma, address, flags,
2795 VM_UFFD_MISSING);
2796 }
2797
2798 inc_mm_counter_fast(mm, MM_ANONPAGES);
2799 page_add_new_anon_rmap(page, vma, address, false);
2800 mem_cgroup_commit_charge(page, memcg, false, false);
2801 lru_cache_add_active_or_unevictable(page, vma);
2802setpte:
2803 set_pte_at(mm, address, page_table, entry);
2804
2805 /* No need to invalidate - it was non-present before */
2806 update_mmu_cache(vma, address, page_table);
2807unlock:
2808 pte_unmap_unlock(page_table, ptl);
2809 return 0;
2810release:
2811 mem_cgroup_cancel_charge(page, memcg, false);
2812 put_page(page);
2813 goto unlock;
2814oom_free_page:
2815 put_page(page);
2816oom:
2817 return VM_FAULT_OOM;
2818}
2819
2820/*
2821 * The mmap_sem must have been held on entry, and may have been
2822 * released depending on flags and vma->vm_ops->fault() return value.
2823 * See filemap_fault() and __lock_page_retry().
2824 */
2825static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2826 pgoff_t pgoff, unsigned int flags,
2827 struct page *cow_page, struct page **page)
2828{
2829 struct vm_fault vmf;
2830 int ret;
2831
2832 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2833 vmf.pgoff = pgoff;
2834 vmf.flags = flags;
2835 vmf.page = NULL;
2836 vmf.gfp_mask = __get_fault_gfp_mask(vma);
2837 vmf.cow_page = cow_page;
2838
2839 ret = vma->vm_ops->fault(vma, &vmf);
2840 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2841 return ret;
2842 if (!vmf.page)
2843 goto out;
2844
2845 if (unlikely(PageHWPoison(vmf.page))) {
2846 if (ret & VM_FAULT_LOCKED)
2847 unlock_page(vmf.page);
2848 put_page(vmf.page);
2849 return VM_FAULT_HWPOISON;
2850 }
2851
2852 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2853 lock_page(vmf.page);
2854 else
2855 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2856
2857 out:
2858 *page = vmf.page;
2859 return ret;
2860}
2861
2862/**
2863 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2864 *
2865 * @vma: virtual memory area
2866 * @address: user virtual address
2867 * @page: page to map
2868 * @pte: pointer to target page table entry
2869 * @write: true, if new entry is writable
2870 * @anon: true, if it's anonymous page
2871 *
2872 * Caller must hold page table lock relevant for @pte.
2873 *
2874 * Target users are page handler itself and implementations of
2875 * vm_ops->map_pages.
2876 */
2877void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2878 struct page *page, pte_t *pte, bool write, bool anon)
2879{
2880 pte_t entry;
2881
2882 flush_icache_page(vma, page);
2883 entry = mk_pte(page, vma->vm_page_prot);
2884 if (write)
2885 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2886 if (anon) {
2887 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2888 page_add_new_anon_rmap(page, vma, address, false);
2889 } else {
2890 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
2891 page_add_file_rmap(page);
2892 }
2893 set_pte_at(vma->vm_mm, address, pte, entry);
2894
2895 /* no need to invalidate: a not-present page won't be cached */
2896 update_mmu_cache(vma, address, pte);
2897}
2898
2899static unsigned long fault_around_bytes __read_mostly =
2900 rounddown_pow_of_two(65536);
2901
2902#ifdef CONFIG_DEBUG_FS
2903static int fault_around_bytes_get(void *data, u64 *val)
2904{
2905 *val = fault_around_bytes;
2906 return 0;
2907}
2908
2909/*
2910 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2911 * rounded down to nearest page order. It's what do_fault_around() expects to
2912 * see.
2913 */
2914static int fault_around_bytes_set(void *data, u64 val)
2915{
2916 if (val / PAGE_SIZE > PTRS_PER_PTE)
2917 return -EINVAL;
2918 if (val > PAGE_SIZE)
2919 fault_around_bytes = rounddown_pow_of_two(val);
2920 else
2921 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2922 return 0;
2923}
2924DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2925 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2926
2927static int __init fault_around_debugfs(void)
2928{
2929 void *ret;
2930
2931 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2932 &fault_around_bytes_fops);
2933 if (!ret)
2934 pr_warn("Failed to create fault_around_bytes in debugfs");
2935 return 0;
2936}
2937late_initcall(fault_around_debugfs);
2938#endif
2939
2940/*
2941 * do_fault_around() tries to map few pages around the fault address. The hope
2942 * is that the pages will be needed soon and this will lower the number of
2943 * faults to handle.
2944 *
2945 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2946 * not ready to be mapped: not up-to-date, locked, etc.
2947 *
2948 * This function is called with the page table lock taken. In the split ptlock
2949 * case the page table lock only protects only those entries which belong to
2950 * the page table corresponding to the fault address.
2951 *
2952 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2953 * only once.
2954 *
2955 * fault_around_pages() defines how many pages we'll try to map.
2956 * do_fault_around() expects it to return a power of two less than or equal to
2957 * PTRS_PER_PTE.
2958 *
2959 * The virtual address of the area that we map is naturally aligned to the
2960 * fault_around_pages() value (and therefore to page order). This way it's
2961 * easier to guarantee that we don't cross page table boundaries.
2962 */
2963static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2964 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2965{
2966 unsigned long start_addr, nr_pages, mask;
2967 pgoff_t max_pgoff;
2968 struct vm_fault vmf;
2969 int off;
2970
2971 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2972 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2973
2974 start_addr = max(address & mask, vma->vm_start);
2975 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2976 pte -= off;
2977 pgoff -= off;
2978
2979 /*
2980 * max_pgoff is either end of page table or end of vma
2981 * or fault_around_pages() from pgoff, depending what is nearest.
2982 */
2983 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2984 PTRS_PER_PTE - 1;
2985 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2986 pgoff + nr_pages - 1);
2987
2988 /* Check if it makes any sense to call ->map_pages */
2989 while (!pte_none(*pte)) {
2990 if (++pgoff > max_pgoff)
2991 return;
2992 start_addr += PAGE_SIZE;
2993 if (start_addr >= vma->vm_end)
2994 return;
2995 pte++;
2996 }
2997
2998 vmf.virtual_address = (void __user *) start_addr;
2999 vmf.pte = pte;
3000 vmf.pgoff = pgoff;
3001 vmf.max_pgoff = max_pgoff;
3002 vmf.flags = flags;
3003 vmf.gfp_mask = __get_fault_gfp_mask(vma);
3004 vma->vm_ops->map_pages(vma, &vmf);
3005}
3006
3007static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3008 unsigned long address, pmd_t *pmd,
3009 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3010{
3011 struct page *fault_page;
3012 spinlock_t *ptl;
3013 pte_t *pte;
3014 int ret = 0;
3015
3016 /*
3017 * Let's call ->map_pages() first and use ->fault() as fallback
3018 * if page by the offset is not ready to be mapped (cold cache or
3019 * something).
3020 */
3021 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3022 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3023 do_fault_around(vma, address, pte, pgoff, flags);
3024 if (!pte_same(*pte, orig_pte))
3025 goto unlock_out;
3026 pte_unmap_unlock(pte, ptl);
3027 }
3028
3029 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3030 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3031 return ret;
3032
3033 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3034 if (unlikely(!pte_same(*pte, orig_pte))) {
3035 pte_unmap_unlock(pte, ptl);
3036 unlock_page(fault_page);
3037 put_page(fault_page);
3038 return ret;
3039 }
3040 do_set_pte(vma, address, fault_page, pte, false, false);
3041 unlock_page(fault_page);
3042unlock_out:
3043 pte_unmap_unlock(pte, ptl);
3044 return ret;
3045}
3046
3047static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3048 unsigned long address, pmd_t *pmd,
3049 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3050{
3051 struct page *fault_page, *new_page;
3052 struct mem_cgroup *memcg;
3053 spinlock_t *ptl;
3054 pte_t *pte;
3055 int ret;
3056
3057 if (unlikely(anon_vma_prepare(vma)))
3058 return VM_FAULT_OOM;
3059
3060 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3061 if (!new_page)
3062 return VM_FAULT_OOM;
3063
3064 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) {
3065 put_page(new_page);
3066 return VM_FAULT_OOM;
3067 }
3068
3069 ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
3070 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3071 goto uncharge_out;
3072
3073 if (fault_page)
3074 copy_user_highpage(new_page, fault_page, address, vma);
3075 __SetPageUptodate(new_page);
3076
3077 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3078 if (unlikely(!pte_same(*pte, orig_pte))) {
3079 pte_unmap_unlock(pte, ptl);
3080 if (fault_page) {
3081 unlock_page(fault_page);
3082 put_page(fault_page);
3083 } else {
3084 /*
3085 * The fault handler has no page to lock, so it holds
3086 * i_mmap_lock for read to protect against truncate.
3087 */
3088 i_mmap_unlock_read(vma->vm_file->f_mapping);
3089 }
3090 goto uncharge_out;
3091 }
3092 do_set_pte(vma, address, new_page, pte, true, true);
3093 mem_cgroup_commit_charge(new_page, memcg, false, false);
3094 lru_cache_add_active_or_unevictable(new_page, vma);
3095 pte_unmap_unlock(pte, ptl);
3096 if (fault_page) {
3097 unlock_page(fault_page);
3098 put_page(fault_page);
3099 } else {
3100 /*
3101 * The fault handler has no page to lock, so it holds
3102 * i_mmap_lock for read to protect against truncate.
3103 */
3104 i_mmap_unlock_read(vma->vm_file->f_mapping);
3105 }
3106 return ret;
3107uncharge_out:
3108 mem_cgroup_cancel_charge(new_page, memcg, false);
3109 put_page(new_page);
3110 return ret;
3111}
3112
3113static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3114 unsigned long address, pmd_t *pmd,
3115 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3116{
3117 struct page *fault_page;
3118 struct address_space *mapping;
3119 spinlock_t *ptl;
3120 pte_t *pte;
3121 int dirtied = 0;
3122 int ret, tmp;
3123
3124 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3125 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3126 return ret;
3127
3128 /*
3129 * Check if the backing address space wants to know that the page is
3130 * about to become writable
3131 */
3132 if (vma->vm_ops->page_mkwrite) {
3133 unlock_page(fault_page);
3134 tmp = do_page_mkwrite(vma, fault_page, address);
3135 if (unlikely(!tmp ||
3136 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3137 put_page(fault_page);
3138 return tmp;
3139 }
3140 }
3141
3142 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3143 if (unlikely(!pte_same(*pte, orig_pte))) {
3144 pte_unmap_unlock(pte, ptl);
3145 unlock_page(fault_page);
3146 put_page(fault_page);
3147 return ret;
3148 }
3149 do_set_pte(vma, address, fault_page, pte, true, false);
3150 pte_unmap_unlock(pte, ptl);
3151
3152 if (set_page_dirty(fault_page))
3153 dirtied = 1;
3154 /*
3155 * Take a local copy of the address_space - page.mapping may be zeroed
3156 * by truncate after unlock_page(). The address_space itself remains
3157 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3158 * release semantics to prevent the compiler from undoing this copying.
3159 */
3160 mapping = page_rmapping(fault_page);
3161 unlock_page(fault_page);
3162 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3163 /*
3164 * Some device drivers do not set page.mapping but still
3165 * dirty their pages
3166 */
3167 balance_dirty_pages_ratelimited(mapping);
3168 }
3169
3170 if (!vma->vm_ops->page_mkwrite)
3171 file_update_time(vma->vm_file);
3172
3173 return ret;
3174}
3175
3176/*
3177 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3178 * but allow concurrent faults).
3179 * The mmap_sem may have been released depending on flags and our
3180 * return value. See filemap_fault() and __lock_page_or_retry().
3181 */
3182static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3183 unsigned long address, pte_t *page_table, pmd_t *pmd,
3184 unsigned int flags, pte_t orig_pte)
3185{
3186 pgoff_t pgoff = linear_page_index(vma, address);
3187
3188 pte_unmap(page_table);
3189 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3190 if (!vma->vm_ops->fault)
3191 return VM_FAULT_SIGBUS;
3192 if (!(flags & FAULT_FLAG_WRITE))
3193 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3194 orig_pte);
3195 if (!(vma->vm_flags & VM_SHARED))
3196 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3197 orig_pte);
3198 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3199}
3200
3201static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3202 unsigned long addr, int page_nid,
3203 int *flags)
3204{
3205 get_page(page);
3206
3207 count_vm_numa_event(NUMA_HINT_FAULTS);
3208 if (page_nid == numa_node_id()) {
3209 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3210 *flags |= TNF_FAULT_LOCAL;
3211 }
3212
3213 return mpol_misplaced(page, vma, addr);
3214}
3215
3216static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3217 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3218{
3219 struct page *page = NULL;
3220 spinlock_t *ptl;
3221 int page_nid = -1;
3222 int last_cpupid;
3223 int target_nid;
3224 bool migrated = false;
3225 bool was_writable = pte_write(pte);
3226 int flags = 0;
3227
3228 /* A PROT_NONE fault should not end up here */
3229 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3230
3231 /*
3232 * The "pte" at this point cannot be used safely without
3233 * validation through pte_unmap_same(). It's of NUMA type but
3234 * the pfn may be screwed if the read is non atomic.
3235 *
3236 * We can safely just do a "set_pte_at()", because the old
3237 * page table entry is not accessible, so there would be no
3238 * concurrent hardware modifications to the PTE.
3239 */
3240 ptl = pte_lockptr(mm, pmd);
3241 spin_lock(ptl);
3242 if (unlikely(!pte_same(*ptep, pte))) {
3243 pte_unmap_unlock(ptep, ptl);
3244 goto out;
3245 }
3246
3247 /* Make it present again */
3248 pte = pte_modify(pte, vma->vm_page_prot);
3249 pte = pte_mkyoung(pte);
3250 if (was_writable)
3251 pte = pte_mkwrite(pte);
3252 set_pte_at(mm, addr, ptep, pte);
3253 update_mmu_cache(vma, addr, ptep);
3254
3255 page = vm_normal_page(vma, addr, pte);
3256 if (!page) {
3257 pte_unmap_unlock(ptep, ptl);
3258 return 0;
3259 }
3260
3261 /* TODO: handle PTE-mapped THP */
3262 if (PageCompound(page)) {
3263 pte_unmap_unlock(ptep, ptl);
3264 return 0;
3265 }
3266
3267 /*
3268 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3269 * much anyway since they can be in shared cache state. This misses
3270 * the case where a mapping is writable but the process never writes
3271 * to it but pte_write gets cleared during protection updates and
3272 * pte_dirty has unpredictable behaviour between PTE scan updates,
3273 * background writeback, dirty balancing and application behaviour.
3274 */
3275 if (!(vma->vm_flags & VM_WRITE))
3276 flags |= TNF_NO_GROUP;
3277
3278 /*
3279 * Flag if the page is shared between multiple address spaces. This
3280 * is later used when determining whether to group tasks together
3281 */
3282 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3283 flags |= TNF_SHARED;
3284
3285 last_cpupid = page_cpupid_last(page);
3286 page_nid = page_to_nid(page);
3287 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3288 pte_unmap_unlock(ptep, ptl);
3289 if (target_nid == -1) {
3290 put_page(page);
3291 goto out;
3292 }
3293
3294 /* Migrate to the requested node */
3295 migrated = migrate_misplaced_page(page, vma, target_nid);
3296 if (migrated) {
3297 page_nid = target_nid;
3298 flags |= TNF_MIGRATED;
3299 } else
3300 flags |= TNF_MIGRATE_FAIL;
3301
3302out:
3303 if (page_nid != -1)
3304 task_numa_fault(last_cpupid, page_nid, 1, flags);
3305 return 0;
3306}
3307
3308static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3309 unsigned long address, pmd_t *pmd, unsigned int flags)
3310{
3311 if (vma_is_anonymous(vma))
3312 return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags);
3313 if (vma->vm_ops->pmd_fault)
3314 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3315 return VM_FAULT_FALLBACK;
3316}
3317
3318static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3319 unsigned long address, pmd_t *pmd, pmd_t orig_pmd,
3320 unsigned int flags)
3321{
3322 if (vma_is_anonymous(vma))
3323 return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd);
3324 if (vma->vm_ops->pmd_fault)
3325 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3326 return VM_FAULT_FALLBACK;
3327}
3328
3329/*
3330 * These routines also need to handle stuff like marking pages dirty
3331 * and/or accessed for architectures that don't do it in hardware (most
3332 * RISC architectures). The early dirtying is also good on the i386.
3333 *
3334 * There is also a hook called "update_mmu_cache()" that architectures
3335 * with external mmu caches can use to update those (ie the Sparc or
3336 * PowerPC hashed page tables that act as extended TLBs).
3337 *
3338 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3339 * but allow concurrent faults), and pte mapped but not yet locked.
3340 * We return with pte unmapped and unlocked.
3341 *
3342 * The mmap_sem may have been released depending on flags and our
3343 * return value. See filemap_fault() and __lock_page_or_retry().
3344 */
3345static int handle_pte_fault(struct mm_struct *mm,
3346 struct vm_area_struct *vma, unsigned long address,
3347 pte_t *pte, pmd_t *pmd, unsigned int flags)
3348{
3349 pte_t entry;
3350 spinlock_t *ptl;
3351
3352 /*
3353 * some architectures can have larger ptes than wordsize,
3354 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3355 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3356 * The code below just needs a consistent view for the ifs and
3357 * we later double check anyway with the ptl lock held. So here
3358 * a barrier will do.
3359 */
3360 entry = *pte;
3361 barrier();
3362 if (!pte_present(entry)) {
3363 if (pte_none(entry)) {
3364 if (vma_is_anonymous(vma))
3365 return do_anonymous_page(mm, vma, address,
3366 pte, pmd, flags);
3367 else
3368 return do_fault(mm, vma, address, pte, pmd,
3369 flags, entry);
3370 }
3371 return do_swap_page(mm, vma, address,
3372 pte, pmd, flags, entry);
3373 }
3374
3375 if (pte_protnone(entry))
3376 return do_numa_page(mm, vma, address, entry, pte, pmd);
3377
3378 ptl = pte_lockptr(mm, pmd);
3379 spin_lock(ptl);
3380 if (unlikely(!pte_same(*pte, entry)))
3381 goto unlock;
3382 if (flags & FAULT_FLAG_WRITE) {
3383 if (!pte_write(entry))
3384 return do_wp_page(mm, vma, address,
3385 pte, pmd, ptl, entry);
3386 entry = pte_mkdirty(entry);
3387 }
3388 entry = pte_mkyoung(entry);
3389 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3390 update_mmu_cache(vma, address, pte);
3391 } else {
3392 /*
3393 * This is needed only for protection faults but the arch code
3394 * is not yet telling us if this is a protection fault or not.
3395 * This still avoids useless tlb flushes for .text page faults
3396 * with threads.
3397 */
3398 if (flags & FAULT_FLAG_WRITE)
3399 flush_tlb_fix_spurious_fault(vma, address);
3400 }
3401unlock:
3402 pte_unmap_unlock(pte, ptl);
3403 return 0;
3404}
3405
3406/*
3407 * By the time we get here, we already hold the mm semaphore
3408 *
3409 * The mmap_sem may have been released depending on flags and our
3410 * return value. See filemap_fault() and __lock_page_or_retry().
3411 */
3412static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3413 unsigned long address, unsigned int flags)
3414{
3415 pgd_t *pgd;
3416 pud_t *pud;
3417 pmd_t *pmd;
3418 pte_t *pte;
3419
3420 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3421 flags & FAULT_FLAG_INSTRUCTION,
3422 flags & FAULT_FLAG_REMOTE))
3423 return VM_FAULT_SIGSEGV;
3424
3425 if (unlikely(is_vm_hugetlb_page(vma)))
3426 return hugetlb_fault(mm, vma, address, flags);
3427
3428 pgd = pgd_offset(mm, address);
3429 pud = pud_alloc(mm, pgd, address);
3430 if (!pud)
3431 return VM_FAULT_OOM;
3432 pmd = pmd_alloc(mm, pud, address);
3433 if (!pmd)
3434 return VM_FAULT_OOM;
3435 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3436 int ret = create_huge_pmd(mm, vma, address, pmd, flags);
3437 if (!(ret & VM_FAULT_FALLBACK))
3438 return ret;
3439 } else {
3440 pmd_t orig_pmd = *pmd;
3441 int ret;
3442
3443 barrier();
3444 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3445 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3446
3447 if (pmd_protnone(orig_pmd))
3448 return do_huge_pmd_numa_page(mm, vma, address,
3449 orig_pmd, pmd);
3450
3451 if (dirty && !pmd_write(orig_pmd)) {
3452 ret = wp_huge_pmd(mm, vma, address, pmd,
3453 orig_pmd, flags);
3454 if (!(ret & VM_FAULT_FALLBACK))
3455 return ret;
3456 } else {
3457 huge_pmd_set_accessed(mm, vma, address, pmd,
3458 orig_pmd, dirty);
3459 return 0;
3460 }
3461 }
3462 }
3463
3464 /*
3465 * Use pte_alloc() instead of pte_alloc_map, because we can't
3466 * run pte_offset_map on the pmd, if an huge pmd could
3467 * materialize from under us from a different thread.
3468 */
3469 if (unlikely(pte_alloc(mm, pmd, address)))
3470 return VM_FAULT_OOM;
3471 /*
3472 * If a huge pmd materialized under us just retry later. Use
3473 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3474 * didn't become pmd_trans_huge under us and then back to pmd_none, as
3475 * a result of MADV_DONTNEED running immediately after a huge pmd fault
3476 * in a different thread of this mm, in turn leading to a misleading
3477 * pmd_trans_huge() retval. All we have to ensure is that it is a
3478 * regular pmd that we can walk with pte_offset_map() and we can do that
3479 * through an atomic read in C, which is what pmd_trans_unstable()
3480 * provides.
3481 */
3482 if (unlikely(pmd_trans_unstable(pmd) || pmd_devmap(*pmd)))
3483 return 0;
3484 /*
3485 * A regular pmd is established and it can't morph into a huge pmd
3486 * from under us anymore at this point because we hold the mmap_sem
3487 * read mode and khugepaged takes it in write mode. So now it's
3488 * safe to run pte_offset_map().
3489 */
3490 pte = pte_offset_map(pmd, address);
3491
3492 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3493}
3494
3495/*
3496 * By the time we get here, we already hold the mm semaphore
3497 *
3498 * The mmap_sem may have been released depending on flags and our
3499 * return value. See filemap_fault() and __lock_page_or_retry().
3500 */
3501int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3502 unsigned long address, unsigned int flags)
3503{
3504 int ret;
3505
3506 __set_current_state(TASK_RUNNING);
3507
3508 count_vm_event(PGFAULT);
3509 mem_cgroup_count_vm_event(mm, PGFAULT);
3510
3511 /* do counter updates before entering really critical section. */
3512 check_sync_rss_stat(current);
3513
3514 /*
3515 * Enable the memcg OOM handling for faults triggered in user
3516 * space. Kernel faults are handled more gracefully.
3517 */
3518 if (flags & FAULT_FLAG_USER)
3519 mem_cgroup_oom_enable();
3520
3521 ret = __handle_mm_fault(mm, vma, address, flags);
3522
3523 if (flags & FAULT_FLAG_USER) {
3524 mem_cgroup_oom_disable();
3525 /*
3526 * The task may have entered a memcg OOM situation but
3527 * if the allocation error was handled gracefully (no
3528 * VM_FAULT_OOM), there is no need to kill anything.
3529 * Just clean up the OOM state peacefully.
3530 */
3531 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3532 mem_cgroup_oom_synchronize(false);
3533 }
3534
3535 return ret;
3536}
3537EXPORT_SYMBOL_GPL(handle_mm_fault);
3538
3539#ifndef __PAGETABLE_PUD_FOLDED
3540/*
3541 * Allocate page upper directory.
3542 * We've already handled the fast-path in-line.
3543 */
3544int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3545{
3546 pud_t *new = pud_alloc_one(mm, address);
3547 if (!new)
3548 return -ENOMEM;
3549
3550 smp_wmb(); /* See comment in __pte_alloc */
3551
3552 spin_lock(&mm->page_table_lock);
3553 if (pgd_present(*pgd)) /* Another has populated it */
3554 pud_free(mm, new);
3555 else
3556 pgd_populate(mm, pgd, new);
3557 spin_unlock(&mm->page_table_lock);
3558 return 0;
3559}
3560#endif /* __PAGETABLE_PUD_FOLDED */
3561
3562#ifndef __PAGETABLE_PMD_FOLDED
3563/*
3564 * Allocate page middle directory.
3565 * We've already handled the fast-path in-line.
3566 */
3567int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3568{
3569 pmd_t *new = pmd_alloc_one(mm, address);
3570 if (!new)
3571 return -ENOMEM;
3572
3573 smp_wmb(); /* See comment in __pte_alloc */
3574
3575 spin_lock(&mm->page_table_lock);
3576#ifndef __ARCH_HAS_4LEVEL_HACK
3577 if (!pud_present(*pud)) {
3578 mm_inc_nr_pmds(mm);
3579 pud_populate(mm, pud, new);
3580 } else /* Another has populated it */
3581 pmd_free(mm, new);
3582#else
3583 if (!pgd_present(*pud)) {
3584 mm_inc_nr_pmds(mm);
3585 pgd_populate(mm, pud, new);
3586 } else /* Another has populated it */
3587 pmd_free(mm, new);
3588#endif /* __ARCH_HAS_4LEVEL_HACK */
3589 spin_unlock(&mm->page_table_lock);
3590 return 0;
3591}
3592#endif /* __PAGETABLE_PMD_FOLDED */
3593
3594static int __follow_pte(struct mm_struct *mm, unsigned long address,
3595 pte_t **ptepp, spinlock_t **ptlp)
3596{
3597 pgd_t *pgd;
3598 pud_t *pud;
3599 pmd_t *pmd;
3600 pte_t *ptep;
3601
3602 pgd = pgd_offset(mm, address);
3603 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3604 goto out;
3605
3606 pud = pud_offset(pgd, address);
3607 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3608 goto out;
3609
3610 pmd = pmd_offset(pud, address);
3611 VM_BUG_ON(pmd_trans_huge(*pmd));
3612 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3613 goto out;
3614
3615 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3616 if (pmd_huge(*pmd))
3617 goto out;
3618
3619 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3620 if (!ptep)
3621 goto out;
3622 if (!pte_present(*ptep))
3623 goto unlock;
3624 *ptepp = ptep;
3625 return 0;
3626unlock:
3627 pte_unmap_unlock(ptep, *ptlp);
3628out:
3629 return -EINVAL;
3630}
3631
3632static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3633 pte_t **ptepp, spinlock_t **ptlp)
3634{
3635 int res;
3636
3637 /* (void) is needed to make gcc happy */
3638 (void) __cond_lock(*ptlp,
3639 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3640 return res;
3641}
3642
3643/**
3644 * follow_pfn - look up PFN at a user virtual address
3645 * @vma: memory mapping
3646 * @address: user virtual address
3647 * @pfn: location to store found PFN
3648 *
3649 * Only IO mappings and raw PFN mappings are allowed.
3650 *
3651 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3652 */
3653int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3654 unsigned long *pfn)
3655{
3656 int ret = -EINVAL;
3657 spinlock_t *ptl;
3658 pte_t *ptep;
3659
3660 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3661 return ret;
3662
3663 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3664 if (ret)
3665 return ret;
3666 *pfn = pte_pfn(*ptep);
3667 pte_unmap_unlock(ptep, ptl);
3668 return 0;
3669}
3670EXPORT_SYMBOL(follow_pfn);
3671
3672#ifdef CONFIG_HAVE_IOREMAP_PROT
3673int follow_phys(struct vm_area_struct *vma,
3674 unsigned long address, unsigned int flags,
3675 unsigned long *prot, resource_size_t *phys)
3676{
3677 int ret = -EINVAL;
3678 pte_t *ptep, pte;
3679 spinlock_t *ptl;
3680
3681 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3682 goto out;
3683
3684 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3685 goto out;
3686 pte = *ptep;
3687
3688 if ((flags & FOLL_WRITE) && !pte_write(pte))
3689 goto unlock;
3690
3691 *prot = pgprot_val(pte_pgprot(pte));
3692 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3693
3694 ret = 0;
3695unlock:
3696 pte_unmap_unlock(ptep, ptl);
3697out:
3698 return ret;
3699}
3700
3701int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3702 void *buf, int len, int write)
3703{
3704 resource_size_t phys_addr;
3705 unsigned long prot = 0;
3706 void __iomem *maddr;
3707 int offset = addr & (PAGE_SIZE-1);
3708
3709 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3710 return -EINVAL;
3711
3712 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3713 if (write)
3714 memcpy_toio(maddr + offset, buf, len);
3715 else
3716 memcpy_fromio(buf, maddr + offset, len);
3717 iounmap(maddr);
3718
3719 return len;
3720}
3721EXPORT_SYMBOL_GPL(generic_access_phys);
3722#endif
3723
3724/*
3725 * Access another process' address space as given in mm. If non-NULL, use the
3726 * given task for page fault accounting.
3727 */
3728static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3729 unsigned long addr, void *buf, int len, int write)
3730{
3731 struct vm_area_struct *vma;
3732 void *old_buf = buf;
3733
3734 down_read(&mm->mmap_sem);
3735 /* ignore errors, just check how much was successfully transferred */
3736 while (len) {
3737 int bytes, ret, offset;
3738 void *maddr;
3739 struct page *page = NULL;
3740
3741 ret = get_user_pages_remote(tsk, mm, addr, 1,
3742 write, 1, &page, &vma);
3743 if (ret <= 0) {
3744#ifndef CONFIG_HAVE_IOREMAP_PROT
3745 break;
3746#else
3747 /*
3748 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3749 * we can access using slightly different code.
3750 */
3751 vma = find_vma(mm, addr);
3752 if (!vma || vma->vm_start > addr)
3753 break;
3754 if (vma->vm_ops && vma->vm_ops->access)
3755 ret = vma->vm_ops->access(vma, addr, buf,
3756 len, write);
3757 if (ret <= 0)
3758 break;
3759 bytes = ret;
3760#endif
3761 } else {
3762 bytes = len;
3763 offset = addr & (PAGE_SIZE-1);
3764 if (bytes > PAGE_SIZE-offset)
3765 bytes = PAGE_SIZE-offset;
3766
3767 maddr = kmap(page);
3768 if (write) {
3769 copy_to_user_page(vma, page, addr,
3770 maddr + offset, buf, bytes);
3771 set_page_dirty_lock(page);
3772 } else {
3773 copy_from_user_page(vma, page, addr,
3774 buf, maddr + offset, bytes);
3775 }
3776 kunmap(page);
3777 put_page(page);
3778 }
3779 len -= bytes;
3780 buf += bytes;
3781 addr += bytes;
3782 }
3783 up_read(&mm->mmap_sem);
3784
3785 return buf - old_buf;
3786}
3787
3788/**
3789 * access_remote_vm - access another process' address space
3790 * @mm: the mm_struct of the target address space
3791 * @addr: start address to access
3792 * @buf: source or destination buffer
3793 * @len: number of bytes to transfer
3794 * @write: whether the access is a write
3795 *
3796 * The caller must hold a reference on @mm.
3797 */
3798int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3799 void *buf, int len, int write)
3800{
3801 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3802}
3803
3804/*
3805 * Access another process' address space.
3806 * Source/target buffer must be kernel space,
3807 * Do not walk the page table directly, use get_user_pages
3808 */
3809int access_process_vm(struct task_struct *tsk, unsigned long addr,
3810 void *buf, int len, int write)
3811{
3812 struct mm_struct *mm;
3813 int ret;
3814
3815 mm = get_task_mm(tsk);
3816 if (!mm)
3817 return 0;
3818
3819 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3820 mmput(mm);
3821
3822 return ret;
3823}
3824
3825/*
3826 * Print the name of a VMA.
3827 */
3828void print_vma_addr(char *prefix, unsigned long ip)
3829{
3830 struct mm_struct *mm = current->mm;
3831 struct vm_area_struct *vma;
3832
3833 /*
3834 * Do not print if we are in atomic
3835 * contexts (in exception stacks, etc.):
3836 */
3837 if (preempt_count())
3838 return;
3839
3840 down_read(&mm->mmap_sem);
3841 vma = find_vma(mm, ip);
3842 if (vma && vma->vm_file) {
3843 struct file *f = vma->vm_file;
3844 char *buf = (char *)__get_free_page(GFP_KERNEL);
3845 if (buf) {
3846 char *p;
3847
3848 p = file_path(f, buf, PAGE_SIZE);
3849 if (IS_ERR(p))
3850 p = "?";
3851 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3852 vma->vm_start,
3853 vma->vm_end - vma->vm_start);
3854 free_page((unsigned long)buf);
3855 }
3856 }
3857 up_read(&mm->mmap_sem);
3858}
3859
3860#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3861void __might_fault(const char *file, int line)
3862{
3863 /*
3864 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3865 * holding the mmap_sem, this is safe because kernel memory doesn't
3866 * get paged out, therefore we'll never actually fault, and the
3867 * below annotations will generate false positives.
3868 */
3869 if (segment_eq(get_fs(), KERNEL_DS))
3870 return;
3871 if (pagefault_disabled())
3872 return;
3873 __might_sleep(file, line, 0);
3874#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3875 if (current->mm)
3876 might_lock_read(¤t->mm->mmap_sem);
3877#endif
3878}
3879EXPORT_SYMBOL(__might_fault);
3880#endif
3881
3882#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3883static void clear_gigantic_page(struct page *page,
3884 unsigned long addr,
3885 unsigned int pages_per_huge_page)
3886{
3887 int i;
3888 struct page *p = page;
3889
3890 might_sleep();
3891 for (i = 0; i < pages_per_huge_page;
3892 i++, p = mem_map_next(p, page, i)) {
3893 cond_resched();
3894 clear_user_highpage(p, addr + i * PAGE_SIZE);
3895 }
3896}
3897void clear_huge_page(struct page *page,
3898 unsigned long addr, unsigned int pages_per_huge_page)
3899{
3900 int i;
3901
3902 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3903 clear_gigantic_page(page, addr, pages_per_huge_page);
3904 return;
3905 }
3906
3907 might_sleep();
3908 for (i = 0; i < pages_per_huge_page; i++) {
3909 cond_resched();
3910 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3911 }
3912}
3913
3914static void copy_user_gigantic_page(struct page *dst, struct page *src,
3915 unsigned long addr,
3916 struct vm_area_struct *vma,
3917 unsigned int pages_per_huge_page)
3918{
3919 int i;
3920 struct page *dst_base = dst;
3921 struct page *src_base = src;
3922
3923 for (i = 0; i < pages_per_huge_page; ) {
3924 cond_resched();
3925 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3926
3927 i++;
3928 dst = mem_map_next(dst, dst_base, i);
3929 src = mem_map_next(src, src_base, i);
3930 }
3931}
3932
3933void copy_user_huge_page(struct page *dst, struct page *src,
3934 unsigned long addr, struct vm_area_struct *vma,
3935 unsigned int pages_per_huge_page)
3936{
3937 int i;
3938
3939 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3940 copy_user_gigantic_page(dst, src, addr, vma,
3941 pages_per_huge_page);
3942 return;
3943 }
3944
3945 might_sleep();
3946 for (i = 0; i < pages_per_huge_page; i++) {
3947 cond_resched();
3948 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3949 }
3950}
3951#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3952
3953#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3954
3955static struct kmem_cache *page_ptl_cachep;
3956
3957void __init ptlock_cache_init(void)
3958{
3959 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3960 SLAB_PANIC, NULL);
3961}
3962
3963bool ptlock_alloc(struct page *page)
3964{
3965 spinlock_t *ptl;
3966
3967 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3968 if (!ptl)
3969 return false;
3970 page->ptl = ptl;
3971 return true;
3972}
3973
3974void ptlock_free(struct page *page)
3975{
3976 kmem_cache_free(page_ptl_cachep, page->ptl);
3977}
3978#endif
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
61#include <asm/io.h>
62#include <asm/pgalloc.h>
63#include <asm/uaccess.h>
64#include <asm/tlb.h>
65#include <asm/tlbflush.h>
66#include <asm/pgtable.h>
67
68#include "internal.h"
69
70#ifndef CONFIG_NEED_MULTIPLE_NODES
71/* use the per-pgdat data instead for discontigmem - mbligh */
72unsigned long max_mapnr;
73struct page *mem_map;
74
75EXPORT_SYMBOL(max_mapnr);
76EXPORT_SYMBOL(mem_map);
77#endif
78
79unsigned long num_physpages;
80/*
81 * A number of key systems in x86 including ioremap() rely on the assumption
82 * that high_memory defines the upper bound on direct map memory, then end
83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85 * and ZONE_HIGHMEM.
86 */
87void * high_memory;
88
89EXPORT_SYMBOL(num_physpages);
90EXPORT_SYMBOL(high_memory);
91
92/*
93 * Randomize the address space (stacks, mmaps, brk, etc.).
94 *
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
97 */
98int randomize_va_space __read_mostly =
99#ifdef CONFIG_COMPAT_BRK
100 1;
101#else
102 2;
103#endif
104
105static int __init disable_randmaps(char *s)
106{
107 randomize_va_space = 0;
108 return 1;
109}
110__setup("norandmaps", disable_randmaps);
111
112unsigned long zero_pfn __read_mostly;
113unsigned long highest_memmap_pfn __read_mostly;
114
115/*
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117 */
118static int __init init_zero_pfn(void)
119{
120 zero_pfn = page_to_pfn(ZERO_PAGE(0));
121 return 0;
122}
123core_initcall(init_zero_pfn);
124
125
126#if defined(SPLIT_RSS_COUNTING)
127
128void sync_mm_rss(struct mm_struct *mm)
129{
130 int i;
131
132 for (i = 0; i < NR_MM_COUNTERS; i++) {
133 if (current->rss_stat.count[i]) {
134 add_mm_counter(mm, i, current->rss_stat.count[i]);
135 current->rss_stat.count[i] = 0;
136 }
137 }
138 current->rss_stat.events = 0;
139}
140
141static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142{
143 struct task_struct *task = current;
144
145 if (likely(task->mm == mm))
146 task->rss_stat.count[member] += val;
147 else
148 add_mm_counter(mm, member, val);
149}
150#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
152
153/* sync counter once per 64 page faults */
154#define TASK_RSS_EVENTS_THRESH (64)
155static void check_sync_rss_stat(struct task_struct *task)
156{
157 if (unlikely(task != current))
158 return;
159 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160 sync_mm_rss(task->mm);
161}
162#else /* SPLIT_RSS_COUNTING */
163
164#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
165#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
166
167static void check_sync_rss_stat(struct task_struct *task)
168{
169}
170
171#endif /* SPLIT_RSS_COUNTING */
172
173#ifdef HAVE_GENERIC_MMU_GATHER
174
175static int tlb_next_batch(struct mmu_gather *tlb)
176{
177 struct mmu_gather_batch *batch;
178
179 batch = tlb->active;
180 if (batch->next) {
181 tlb->active = batch->next;
182 return 1;
183 }
184
185 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
186 if (!batch)
187 return 0;
188
189 batch->next = NULL;
190 batch->nr = 0;
191 batch->max = MAX_GATHER_BATCH;
192
193 tlb->active->next = batch;
194 tlb->active = batch;
195
196 return 1;
197}
198
199/* tlb_gather_mmu
200 * Called to initialize an (on-stack) mmu_gather structure for page-table
201 * tear-down from @mm. The @fullmm argument is used when @mm is without
202 * users and we're going to destroy the full address space (exit/execve).
203 */
204void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
205{
206 tlb->mm = mm;
207
208 tlb->fullmm = fullmm;
209 tlb->need_flush = 0;
210 tlb->fast_mode = (num_possible_cpus() == 1);
211 tlb->local.next = NULL;
212 tlb->local.nr = 0;
213 tlb->local.max = ARRAY_SIZE(tlb->__pages);
214 tlb->active = &tlb->local;
215
216#ifdef CONFIG_HAVE_RCU_TABLE_FREE
217 tlb->batch = NULL;
218#endif
219}
220
221void tlb_flush_mmu(struct mmu_gather *tlb)
222{
223 struct mmu_gather_batch *batch;
224
225 if (!tlb->need_flush)
226 return;
227 tlb->need_flush = 0;
228 tlb_flush(tlb);
229#ifdef CONFIG_HAVE_RCU_TABLE_FREE
230 tlb_table_flush(tlb);
231#endif
232
233 if (tlb_fast_mode(tlb))
234 return;
235
236 for (batch = &tlb->local; batch; batch = batch->next) {
237 free_pages_and_swap_cache(batch->pages, batch->nr);
238 batch->nr = 0;
239 }
240 tlb->active = &tlb->local;
241}
242
243/* tlb_finish_mmu
244 * Called at the end of the shootdown operation to free up any resources
245 * that were required.
246 */
247void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
248{
249 struct mmu_gather_batch *batch, *next;
250
251 tlb_flush_mmu(tlb);
252
253 /* keep the page table cache within bounds */
254 check_pgt_cache();
255
256 for (batch = tlb->local.next; batch; batch = next) {
257 next = batch->next;
258 free_pages((unsigned long)batch, 0);
259 }
260 tlb->local.next = NULL;
261}
262
263/* __tlb_remove_page
264 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
265 * handling the additional races in SMP caused by other CPUs caching valid
266 * mappings in their TLBs. Returns the number of free page slots left.
267 * When out of page slots we must call tlb_flush_mmu().
268 */
269int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
270{
271 struct mmu_gather_batch *batch;
272
273 VM_BUG_ON(!tlb->need_flush);
274
275 if (tlb_fast_mode(tlb)) {
276 free_page_and_swap_cache(page);
277 return 1; /* avoid calling tlb_flush_mmu() */
278 }
279
280 batch = tlb->active;
281 batch->pages[batch->nr++] = page;
282 if (batch->nr == batch->max) {
283 if (!tlb_next_batch(tlb))
284 return 0;
285 batch = tlb->active;
286 }
287 VM_BUG_ON(batch->nr > batch->max);
288
289 return batch->max - batch->nr;
290}
291
292#endif /* HAVE_GENERIC_MMU_GATHER */
293
294#ifdef CONFIG_HAVE_RCU_TABLE_FREE
295
296/*
297 * See the comment near struct mmu_table_batch.
298 */
299
300static void tlb_remove_table_smp_sync(void *arg)
301{
302 /* Simply deliver the interrupt */
303}
304
305static void tlb_remove_table_one(void *table)
306{
307 /*
308 * This isn't an RCU grace period and hence the page-tables cannot be
309 * assumed to be actually RCU-freed.
310 *
311 * It is however sufficient for software page-table walkers that rely on
312 * IRQ disabling. See the comment near struct mmu_table_batch.
313 */
314 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
315 __tlb_remove_table(table);
316}
317
318static void tlb_remove_table_rcu(struct rcu_head *head)
319{
320 struct mmu_table_batch *batch;
321 int i;
322
323 batch = container_of(head, struct mmu_table_batch, rcu);
324
325 for (i = 0; i < batch->nr; i++)
326 __tlb_remove_table(batch->tables[i]);
327
328 free_page((unsigned long)batch);
329}
330
331void tlb_table_flush(struct mmu_gather *tlb)
332{
333 struct mmu_table_batch **batch = &tlb->batch;
334
335 if (*batch) {
336 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
337 *batch = NULL;
338 }
339}
340
341void tlb_remove_table(struct mmu_gather *tlb, void *table)
342{
343 struct mmu_table_batch **batch = &tlb->batch;
344
345 tlb->need_flush = 1;
346
347 /*
348 * When there's less then two users of this mm there cannot be a
349 * concurrent page-table walk.
350 */
351 if (atomic_read(&tlb->mm->mm_users) < 2) {
352 __tlb_remove_table(table);
353 return;
354 }
355
356 if (*batch == NULL) {
357 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
358 if (*batch == NULL) {
359 tlb_remove_table_one(table);
360 return;
361 }
362 (*batch)->nr = 0;
363 }
364 (*batch)->tables[(*batch)->nr++] = table;
365 if ((*batch)->nr == MAX_TABLE_BATCH)
366 tlb_table_flush(tlb);
367}
368
369#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
370
371/*
372 * If a p?d_bad entry is found while walking page tables, report
373 * the error, before resetting entry to p?d_none. Usually (but
374 * very seldom) called out from the p?d_none_or_clear_bad macros.
375 */
376
377void pgd_clear_bad(pgd_t *pgd)
378{
379 pgd_ERROR(*pgd);
380 pgd_clear(pgd);
381}
382
383void pud_clear_bad(pud_t *pud)
384{
385 pud_ERROR(*pud);
386 pud_clear(pud);
387}
388
389void pmd_clear_bad(pmd_t *pmd)
390{
391 pmd_ERROR(*pmd);
392 pmd_clear(pmd);
393}
394
395/*
396 * Note: this doesn't free the actual pages themselves. That
397 * has been handled earlier when unmapping all the memory regions.
398 */
399static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
400 unsigned long addr)
401{
402 pgtable_t token = pmd_pgtable(*pmd);
403 pmd_clear(pmd);
404 pte_free_tlb(tlb, token, addr);
405 tlb->mm->nr_ptes--;
406}
407
408static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
409 unsigned long addr, unsigned long end,
410 unsigned long floor, unsigned long ceiling)
411{
412 pmd_t *pmd;
413 unsigned long next;
414 unsigned long start;
415
416 start = addr;
417 pmd = pmd_offset(pud, addr);
418 do {
419 next = pmd_addr_end(addr, end);
420 if (pmd_none_or_clear_bad(pmd))
421 continue;
422 free_pte_range(tlb, pmd, addr);
423 } while (pmd++, addr = next, addr != end);
424
425 start &= PUD_MASK;
426 if (start < floor)
427 return;
428 if (ceiling) {
429 ceiling &= PUD_MASK;
430 if (!ceiling)
431 return;
432 }
433 if (end - 1 > ceiling - 1)
434 return;
435
436 pmd = pmd_offset(pud, start);
437 pud_clear(pud);
438 pmd_free_tlb(tlb, pmd, start);
439}
440
441static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
442 unsigned long addr, unsigned long end,
443 unsigned long floor, unsigned long ceiling)
444{
445 pud_t *pud;
446 unsigned long next;
447 unsigned long start;
448
449 start = addr;
450 pud = pud_offset(pgd, addr);
451 do {
452 next = pud_addr_end(addr, end);
453 if (pud_none_or_clear_bad(pud))
454 continue;
455 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
456 } while (pud++, addr = next, addr != end);
457
458 start &= PGDIR_MASK;
459 if (start < floor)
460 return;
461 if (ceiling) {
462 ceiling &= PGDIR_MASK;
463 if (!ceiling)
464 return;
465 }
466 if (end - 1 > ceiling - 1)
467 return;
468
469 pud = pud_offset(pgd, start);
470 pgd_clear(pgd);
471 pud_free_tlb(tlb, pud, start);
472}
473
474/*
475 * This function frees user-level page tables of a process.
476 *
477 * Must be called with pagetable lock held.
478 */
479void free_pgd_range(struct mmu_gather *tlb,
480 unsigned long addr, unsigned long end,
481 unsigned long floor, unsigned long ceiling)
482{
483 pgd_t *pgd;
484 unsigned long next;
485
486 /*
487 * The next few lines have given us lots of grief...
488 *
489 * Why are we testing PMD* at this top level? Because often
490 * there will be no work to do at all, and we'd prefer not to
491 * go all the way down to the bottom just to discover that.
492 *
493 * Why all these "- 1"s? Because 0 represents both the bottom
494 * of the address space and the top of it (using -1 for the
495 * top wouldn't help much: the masks would do the wrong thing).
496 * The rule is that addr 0 and floor 0 refer to the bottom of
497 * the address space, but end 0 and ceiling 0 refer to the top
498 * Comparisons need to use "end - 1" and "ceiling - 1" (though
499 * that end 0 case should be mythical).
500 *
501 * Wherever addr is brought up or ceiling brought down, we must
502 * be careful to reject "the opposite 0" before it confuses the
503 * subsequent tests. But what about where end is brought down
504 * by PMD_SIZE below? no, end can't go down to 0 there.
505 *
506 * Whereas we round start (addr) and ceiling down, by different
507 * masks at different levels, in order to test whether a table
508 * now has no other vmas using it, so can be freed, we don't
509 * bother to round floor or end up - the tests don't need that.
510 */
511
512 addr &= PMD_MASK;
513 if (addr < floor) {
514 addr += PMD_SIZE;
515 if (!addr)
516 return;
517 }
518 if (ceiling) {
519 ceiling &= PMD_MASK;
520 if (!ceiling)
521 return;
522 }
523 if (end - 1 > ceiling - 1)
524 end -= PMD_SIZE;
525 if (addr > end - 1)
526 return;
527
528 pgd = pgd_offset(tlb->mm, addr);
529 do {
530 next = pgd_addr_end(addr, end);
531 if (pgd_none_or_clear_bad(pgd))
532 continue;
533 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
534 } while (pgd++, addr = next, addr != end);
535}
536
537void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
538 unsigned long floor, unsigned long ceiling)
539{
540 while (vma) {
541 struct vm_area_struct *next = vma->vm_next;
542 unsigned long addr = vma->vm_start;
543
544 /*
545 * Hide vma from rmap and truncate_pagecache before freeing
546 * pgtables
547 */
548 unlink_anon_vmas(vma);
549 unlink_file_vma(vma);
550
551 if (is_vm_hugetlb_page(vma)) {
552 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
553 floor, next? next->vm_start: ceiling);
554 } else {
555 /*
556 * Optimization: gather nearby vmas into one call down
557 */
558 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
559 && !is_vm_hugetlb_page(next)) {
560 vma = next;
561 next = vma->vm_next;
562 unlink_anon_vmas(vma);
563 unlink_file_vma(vma);
564 }
565 free_pgd_range(tlb, addr, vma->vm_end,
566 floor, next? next->vm_start: ceiling);
567 }
568 vma = next;
569 }
570}
571
572int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
573 pmd_t *pmd, unsigned long address)
574{
575 pgtable_t new = pte_alloc_one(mm, address);
576 int wait_split_huge_page;
577 if (!new)
578 return -ENOMEM;
579
580 /*
581 * Ensure all pte setup (eg. pte page lock and page clearing) are
582 * visible before the pte is made visible to other CPUs by being
583 * put into page tables.
584 *
585 * The other side of the story is the pointer chasing in the page
586 * table walking code (when walking the page table without locking;
587 * ie. most of the time). Fortunately, these data accesses consist
588 * of a chain of data-dependent loads, meaning most CPUs (alpha
589 * being the notable exception) will already guarantee loads are
590 * seen in-order. See the alpha page table accessors for the
591 * smp_read_barrier_depends() barriers in page table walking code.
592 */
593 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
594
595 spin_lock(&mm->page_table_lock);
596 wait_split_huge_page = 0;
597 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
598 mm->nr_ptes++;
599 pmd_populate(mm, pmd, new);
600 new = NULL;
601 } else if (unlikely(pmd_trans_splitting(*pmd)))
602 wait_split_huge_page = 1;
603 spin_unlock(&mm->page_table_lock);
604 if (new)
605 pte_free(mm, new);
606 if (wait_split_huge_page)
607 wait_split_huge_page(vma->anon_vma, pmd);
608 return 0;
609}
610
611int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
612{
613 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
614 if (!new)
615 return -ENOMEM;
616
617 smp_wmb(); /* See comment in __pte_alloc */
618
619 spin_lock(&init_mm.page_table_lock);
620 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
621 pmd_populate_kernel(&init_mm, pmd, new);
622 new = NULL;
623 } else
624 VM_BUG_ON(pmd_trans_splitting(*pmd));
625 spin_unlock(&init_mm.page_table_lock);
626 if (new)
627 pte_free_kernel(&init_mm, new);
628 return 0;
629}
630
631static inline void init_rss_vec(int *rss)
632{
633 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
634}
635
636static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
637{
638 int i;
639
640 if (current->mm == mm)
641 sync_mm_rss(mm);
642 for (i = 0; i < NR_MM_COUNTERS; i++)
643 if (rss[i])
644 add_mm_counter(mm, i, rss[i]);
645}
646
647/*
648 * This function is called to print an error when a bad pte
649 * is found. For example, we might have a PFN-mapped pte in
650 * a region that doesn't allow it.
651 *
652 * The calling function must still handle the error.
653 */
654static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
655 pte_t pte, struct page *page)
656{
657 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
658 pud_t *pud = pud_offset(pgd, addr);
659 pmd_t *pmd = pmd_offset(pud, addr);
660 struct address_space *mapping;
661 pgoff_t index;
662 static unsigned long resume;
663 static unsigned long nr_shown;
664 static unsigned long nr_unshown;
665
666 /*
667 * Allow a burst of 60 reports, then keep quiet for that minute;
668 * or allow a steady drip of one report per second.
669 */
670 if (nr_shown == 60) {
671 if (time_before(jiffies, resume)) {
672 nr_unshown++;
673 return;
674 }
675 if (nr_unshown) {
676 printk(KERN_ALERT
677 "BUG: Bad page map: %lu messages suppressed\n",
678 nr_unshown);
679 nr_unshown = 0;
680 }
681 nr_shown = 0;
682 }
683 if (nr_shown++ == 0)
684 resume = jiffies + 60 * HZ;
685
686 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
687 index = linear_page_index(vma, addr);
688
689 printk(KERN_ALERT
690 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
691 current->comm,
692 (long long)pte_val(pte), (long long)pmd_val(*pmd));
693 if (page)
694 dump_page(page);
695 printk(KERN_ALERT
696 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
697 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
698 /*
699 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
700 */
701 if (vma->vm_ops)
702 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
703 (unsigned long)vma->vm_ops->fault);
704 if (vma->vm_file && vma->vm_file->f_op)
705 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
706 (unsigned long)vma->vm_file->f_op->mmap);
707 dump_stack();
708 add_taint(TAINT_BAD_PAGE);
709}
710
711static inline int is_cow_mapping(vm_flags_t flags)
712{
713 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
714}
715
716#ifndef is_zero_pfn
717static inline int is_zero_pfn(unsigned long pfn)
718{
719 return pfn == zero_pfn;
720}
721#endif
722
723#ifndef my_zero_pfn
724static inline unsigned long my_zero_pfn(unsigned long addr)
725{
726 return zero_pfn;
727}
728#endif
729
730/*
731 * vm_normal_page -- This function gets the "struct page" associated with a pte.
732 *
733 * "Special" mappings do not wish to be associated with a "struct page" (either
734 * it doesn't exist, or it exists but they don't want to touch it). In this
735 * case, NULL is returned here. "Normal" mappings do have a struct page.
736 *
737 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
738 * pte bit, in which case this function is trivial. Secondly, an architecture
739 * may not have a spare pte bit, which requires a more complicated scheme,
740 * described below.
741 *
742 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
743 * special mapping (even if there are underlying and valid "struct pages").
744 * COWed pages of a VM_PFNMAP are always normal.
745 *
746 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
747 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
748 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
749 * mapping will always honor the rule
750 *
751 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
752 *
753 * And for normal mappings this is false.
754 *
755 * This restricts such mappings to be a linear translation from virtual address
756 * to pfn. To get around this restriction, we allow arbitrary mappings so long
757 * as the vma is not a COW mapping; in that case, we know that all ptes are
758 * special (because none can have been COWed).
759 *
760 *
761 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
762 *
763 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
764 * page" backing, however the difference is that _all_ pages with a struct
765 * page (that is, those where pfn_valid is true) are refcounted and considered
766 * normal pages by the VM. The disadvantage is that pages are refcounted
767 * (which can be slower and simply not an option for some PFNMAP users). The
768 * advantage is that we don't have to follow the strict linearity rule of
769 * PFNMAP mappings in order to support COWable mappings.
770 *
771 */
772#ifdef __HAVE_ARCH_PTE_SPECIAL
773# define HAVE_PTE_SPECIAL 1
774#else
775# define HAVE_PTE_SPECIAL 0
776#endif
777struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
778 pte_t pte)
779{
780 unsigned long pfn = pte_pfn(pte);
781
782 if (HAVE_PTE_SPECIAL) {
783 if (likely(!pte_special(pte)))
784 goto check_pfn;
785 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
786 return NULL;
787 if (!is_zero_pfn(pfn))
788 print_bad_pte(vma, addr, pte, NULL);
789 return NULL;
790 }
791
792 /* !HAVE_PTE_SPECIAL case follows: */
793
794 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
795 if (vma->vm_flags & VM_MIXEDMAP) {
796 if (!pfn_valid(pfn))
797 return NULL;
798 goto out;
799 } else {
800 unsigned long off;
801 off = (addr - vma->vm_start) >> PAGE_SHIFT;
802 if (pfn == vma->vm_pgoff + off)
803 return NULL;
804 if (!is_cow_mapping(vma->vm_flags))
805 return NULL;
806 }
807 }
808
809 if (is_zero_pfn(pfn))
810 return NULL;
811check_pfn:
812 if (unlikely(pfn > highest_memmap_pfn)) {
813 print_bad_pte(vma, addr, pte, NULL);
814 return NULL;
815 }
816
817 /*
818 * NOTE! We still have PageReserved() pages in the page tables.
819 * eg. VDSO mappings can cause them to exist.
820 */
821out:
822 return pfn_to_page(pfn);
823}
824
825/*
826 * copy one vm_area from one task to the other. Assumes the page tables
827 * already present in the new task to be cleared in the whole range
828 * covered by this vma.
829 */
830
831static inline unsigned long
832copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
833 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
834 unsigned long addr, int *rss)
835{
836 unsigned long vm_flags = vma->vm_flags;
837 pte_t pte = *src_pte;
838 struct page *page;
839
840 /* pte contains position in swap or file, so copy. */
841 if (unlikely(!pte_present(pte))) {
842 if (!pte_file(pte)) {
843 swp_entry_t entry = pte_to_swp_entry(pte);
844
845 if (swap_duplicate(entry) < 0)
846 return entry.val;
847
848 /* make sure dst_mm is on swapoff's mmlist. */
849 if (unlikely(list_empty(&dst_mm->mmlist))) {
850 spin_lock(&mmlist_lock);
851 if (list_empty(&dst_mm->mmlist))
852 list_add(&dst_mm->mmlist,
853 &src_mm->mmlist);
854 spin_unlock(&mmlist_lock);
855 }
856 if (likely(!non_swap_entry(entry)))
857 rss[MM_SWAPENTS]++;
858 else if (is_migration_entry(entry)) {
859 page = migration_entry_to_page(entry);
860
861 if (PageAnon(page))
862 rss[MM_ANONPAGES]++;
863 else
864 rss[MM_FILEPAGES]++;
865
866 if (is_write_migration_entry(entry) &&
867 is_cow_mapping(vm_flags)) {
868 /*
869 * COW mappings require pages in both
870 * parent and child to be set to read.
871 */
872 make_migration_entry_read(&entry);
873 pte = swp_entry_to_pte(entry);
874 set_pte_at(src_mm, addr, src_pte, pte);
875 }
876 }
877 }
878 goto out_set_pte;
879 }
880
881 /*
882 * If it's a COW mapping, write protect it both
883 * in the parent and the child
884 */
885 if (is_cow_mapping(vm_flags)) {
886 ptep_set_wrprotect(src_mm, addr, src_pte);
887 pte = pte_wrprotect(pte);
888 }
889
890 /*
891 * If it's a shared mapping, mark it clean in
892 * the child
893 */
894 if (vm_flags & VM_SHARED)
895 pte = pte_mkclean(pte);
896 pte = pte_mkold(pte);
897
898 page = vm_normal_page(vma, addr, pte);
899 if (page) {
900 get_page(page);
901 page_dup_rmap(page);
902 if (PageAnon(page))
903 rss[MM_ANONPAGES]++;
904 else
905 rss[MM_FILEPAGES]++;
906 }
907
908out_set_pte:
909 set_pte_at(dst_mm, addr, dst_pte, pte);
910 return 0;
911}
912
913int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
914 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
915 unsigned long addr, unsigned long end)
916{
917 pte_t *orig_src_pte, *orig_dst_pte;
918 pte_t *src_pte, *dst_pte;
919 spinlock_t *src_ptl, *dst_ptl;
920 int progress = 0;
921 int rss[NR_MM_COUNTERS];
922 swp_entry_t entry = (swp_entry_t){0};
923
924again:
925 init_rss_vec(rss);
926
927 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
928 if (!dst_pte)
929 return -ENOMEM;
930 src_pte = pte_offset_map(src_pmd, addr);
931 src_ptl = pte_lockptr(src_mm, src_pmd);
932 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
933 orig_src_pte = src_pte;
934 orig_dst_pte = dst_pte;
935 arch_enter_lazy_mmu_mode();
936
937 do {
938 /*
939 * We are holding two locks at this point - either of them
940 * could generate latencies in another task on another CPU.
941 */
942 if (progress >= 32) {
943 progress = 0;
944 if (need_resched() ||
945 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
946 break;
947 }
948 if (pte_none(*src_pte)) {
949 progress++;
950 continue;
951 }
952 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
953 vma, addr, rss);
954 if (entry.val)
955 break;
956 progress += 8;
957 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
958
959 arch_leave_lazy_mmu_mode();
960 spin_unlock(src_ptl);
961 pte_unmap(orig_src_pte);
962 add_mm_rss_vec(dst_mm, rss);
963 pte_unmap_unlock(orig_dst_pte, dst_ptl);
964 cond_resched();
965
966 if (entry.val) {
967 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
968 return -ENOMEM;
969 progress = 0;
970 }
971 if (addr != end)
972 goto again;
973 return 0;
974}
975
976static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
977 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
978 unsigned long addr, unsigned long end)
979{
980 pmd_t *src_pmd, *dst_pmd;
981 unsigned long next;
982
983 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
984 if (!dst_pmd)
985 return -ENOMEM;
986 src_pmd = pmd_offset(src_pud, addr);
987 do {
988 next = pmd_addr_end(addr, end);
989 if (pmd_trans_huge(*src_pmd)) {
990 int err;
991 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
992 err = copy_huge_pmd(dst_mm, src_mm,
993 dst_pmd, src_pmd, addr, vma);
994 if (err == -ENOMEM)
995 return -ENOMEM;
996 if (!err)
997 continue;
998 /* fall through */
999 }
1000 if (pmd_none_or_clear_bad(src_pmd))
1001 continue;
1002 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1003 vma, addr, next))
1004 return -ENOMEM;
1005 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1006 return 0;
1007}
1008
1009static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1010 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1011 unsigned long addr, unsigned long end)
1012{
1013 pud_t *src_pud, *dst_pud;
1014 unsigned long next;
1015
1016 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1017 if (!dst_pud)
1018 return -ENOMEM;
1019 src_pud = pud_offset(src_pgd, addr);
1020 do {
1021 next = pud_addr_end(addr, end);
1022 if (pud_none_or_clear_bad(src_pud))
1023 continue;
1024 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1025 vma, addr, next))
1026 return -ENOMEM;
1027 } while (dst_pud++, src_pud++, addr = next, addr != end);
1028 return 0;
1029}
1030
1031int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1032 struct vm_area_struct *vma)
1033{
1034 pgd_t *src_pgd, *dst_pgd;
1035 unsigned long next;
1036 unsigned long addr = vma->vm_start;
1037 unsigned long end = vma->vm_end;
1038 int ret;
1039
1040 /*
1041 * Don't copy ptes where a page fault will fill them correctly.
1042 * Fork becomes much lighter when there are big shared or private
1043 * readonly mappings. The tradeoff is that copy_page_range is more
1044 * efficient than faulting.
1045 */
1046 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1047 if (!vma->anon_vma)
1048 return 0;
1049 }
1050
1051 if (is_vm_hugetlb_page(vma))
1052 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1053
1054 if (unlikely(is_pfn_mapping(vma))) {
1055 /*
1056 * We do not free on error cases below as remove_vma
1057 * gets called on error from higher level routine
1058 */
1059 ret = track_pfn_vma_copy(vma);
1060 if (ret)
1061 return ret;
1062 }
1063
1064 /*
1065 * We need to invalidate the secondary MMU mappings only when
1066 * there could be a permission downgrade on the ptes of the
1067 * parent mm. And a permission downgrade will only happen if
1068 * is_cow_mapping() returns true.
1069 */
1070 if (is_cow_mapping(vma->vm_flags))
1071 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1072
1073 ret = 0;
1074 dst_pgd = pgd_offset(dst_mm, addr);
1075 src_pgd = pgd_offset(src_mm, addr);
1076 do {
1077 next = pgd_addr_end(addr, end);
1078 if (pgd_none_or_clear_bad(src_pgd))
1079 continue;
1080 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1081 vma, addr, next))) {
1082 ret = -ENOMEM;
1083 break;
1084 }
1085 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1086
1087 if (is_cow_mapping(vma->vm_flags))
1088 mmu_notifier_invalidate_range_end(src_mm,
1089 vma->vm_start, end);
1090 return ret;
1091}
1092
1093static unsigned long zap_pte_range(struct mmu_gather *tlb,
1094 struct vm_area_struct *vma, pmd_t *pmd,
1095 unsigned long addr, unsigned long end,
1096 struct zap_details *details)
1097{
1098 struct mm_struct *mm = tlb->mm;
1099 int force_flush = 0;
1100 int rss[NR_MM_COUNTERS];
1101 spinlock_t *ptl;
1102 pte_t *start_pte;
1103 pte_t *pte;
1104
1105again:
1106 init_rss_vec(rss);
1107 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1108 pte = start_pte;
1109 arch_enter_lazy_mmu_mode();
1110 do {
1111 pte_t ptent = *pte;
1112 if (pte_none(ptent)) {
1113 continue;
1114 }
1115
1116 if (pte_present(ptent)) {
1117 struct page *page;
1118
1119 page = vm_normal_page(vma, addr, ptent);
1120 if (unlikely(details) && page) {
1121 /*
1122 * unmap_shared_mapping_pages() wants to
1123 * invalidate cache without truncating:
1124 * unmap shared but keep private pages.
1125 */
1126 if (details->check_mapping &&
1127 details->check_mapping != page->mapping)
1128 continue;
1129 /*
1130 * Each page->index must be checked when
1131 * invalidating or truncating nonlinear.
1132 */
1133 if (details->nonlinear_vma &&
1134 (page->index < details->first_index ||
1135 page->index > details->last_index))
1136 continue;
1137 }
1138 ptent = ptep_get_and_clear_full(mm, addr, pte,
1139 tlb->fullmm);
1140 tlb_remove_tlb_entry(tlb, pte, addr);
1141 if (unlikely(!page))
1142 continue;
1143 if (unlikely(details) && details->nonlinear_vma
1144 && linear_page_index(details->nonlinear_vma,
1145 addr) != page->index)
1146 set_pte_at(mm, addr, pte,
1147 pgoff_to_pte(page->index));
1148 if (PageAnon(page))
1149 rss[MM_ANONPAGES]--;
1150 else {
1151 if (pte_dirty(ptent))
1152 set_page_dirty(page);
1153 if (pte_young(ptent) &&
1154 likely(!VM_SequentialReadHint(vma)))
1155 mark_page_accessed(page);
1156 rss[MM_FILEPAGES]--;
1157 }
1158 page_remove_rmap(page);
1159 if (unlikely(page_mapcount(page) < 0))
1160 print_bad_pte(vma, addr, ptent, page);
1161 force_flush = !__tlb_remove_page(tlb, page);
1162 if (force_flush)
1163 break;
1164 continue;
1165 }
1166 /*
1167 * If details->check_mapping, we leave swap entries;
1168 * if details->nonlinear_vma, we leave file entries.
1169 */
1170 if (unlikely(details))
1171 continue;
1172 if (pte_file(ptent)) {
1173 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1174 print_bad_pte(vma, addr, ptent, NULL);
1175 } else {
1176 swp_entry_t entry = pte_to_swp_entry(ptent);
1177
1178 if (!non_swap_entry(entry))
1179 rss[MM_SWAPENTS]--;
1180 else if (is_migration_entry(entry)) {
1181 struct page *page;
1182
1183 page = migration_entry_to_page(entry);
1184
1185 if (PageAnon(page))
1186 rss[MM_ANONPAGES]--;
1187 else
1188 rss[MM_FILEPAGES]--;
1189 }
1190 if (unlikely(!free_swap_and_cache(entry)))
1191 print_bad_pte(vma, addr, ptent, NULL);
1192 }
1193 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1194 } while (pte++, addr += PAGE_SIZE, addr != end);
1195
1196 add_mm_rss_vec(mm, rss);
1197 arch_leave_lazy_mmu_mode();
1198 pte_unmap_unlock(start_pte, ptl);
1199
1200 /*
1201 * mmu_gather ran out of room to batch pages, we break out of
1202 * the PTE lock to avoid doing the potential expensive TLB invalidate
1203 * and page-free while holding it.
1204 */
1205 if (force_flush) {
1206 force_flush = 0;
1207 tlb_flush_mmu(tlb);
1208 if (addr != end)
1209 goto again;
1210 }
1211
1212 return addr;
1213}
1214
1215static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1216 struct vm_area_struct *vma, pud_t *pud,
1217 unsigned long addr, unsigned long end,
1218 struct zap_details *details)
1219{
1220 pmd_t *pmd;
1221 unsigned long next;
1222
1223 pmd = pmd_offset(pud, addr);
1224 do {
1225 next = pmd_addr_end(addr, end);
1226 if (pmd_trans_huge(*pmd)) {
1227 if (next - addr != HPAGE_PMD_SIZE) {
1228#ifdef CONFIG_DEBUG_VM
1229 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1230 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1231 __func__, addr, end,
1232 vma->vm_start,
1233 vma->vm_end);
1234 BUG();
1235 }
1236#endif
1237 split_huge_page_pmd(vma->vm_mm, pmd);
1238 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1239 goto next;
1240 /* fall through */
1241 }
1242 /*
1243 * Here there can be other concurrent MADV_DONTNEED or
1244 * trans huge page faults running, and if the pmd is
1245 * none or trans huge it can change under us. This is
1246 * because MADV_DONTNEED holds the mmap_sem in read
1247 * mode.
1248 */
1249 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1250 goto next;
1251 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1252next:
1253 cond_resched();
1254 } while (pmd++, addr = next, addr != end);
1255
1256 return addr;
1257}
1258
1259static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1260 struct vm_area_struct *vma, pgd_t *pgd,
1261 unsigned long addr, unsigned long end,
1262 struct zap_details *details)
1263{
1264 pud_t *pud;
1265 unsigned long next;
1266
1267 pud = pud_offset(pgd, addr);
1268 do {
1269 next = pud_addr_end(addr, end);
1270 if (pud_none_or_clear_bad(pud))
1271 continue;
1272 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1273 } while (pud++, addr = next, addr != end);
1274
1275 return addr;
1276}
1277
1278static void unmap_page_range(struct mmu_gather *tlb,
1279 struct vm_area_struct *vma,
1280 unsigned long addr, unsigned long end,
1281 struct zap_details *details)
1282{
1283 pgd_t *pgd;
1284 unsigned long next;
1285
1286 if (details && !details->check_mapping && !details->nonlinear_vma)
1287 details = NULL;
1288
1289 BUG_ON(addr >= end);
1290 mem_cgroup_uncharge_start();
1291 tlb_start_vma(tlb, vma);
1292 pgd = pgd_offset(vma->vm_mm, addr);
1293 do {
1294 next = pgd_addr_end(addr, end);
1295 if (pgd_none_or_clear_bad(pgd))
1296 continue;
1297 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1298 } while (pgd++, addr = next, addr != end);
1299 tlb_end_vma(tlb, vma);
1300 mem_cgroup_uncharge_end();
1301}
1302
1303
1304static void unmap_single_vma(struct mmu_gather *tlb,
1305 struct vm_area_struct *vma, unsigned long start_addr,
1306 unsigned long end_addr,
1307 struct zap_details *details)
1308{
1309 unsigned long start = max(vma->vm_start, start_addr);
1310 unsigned long end;
1311
1312 if (start >= vma->vm_end)
1313 return;
1314 end = min(vma->vm_end, end_addr);
1315 if (end <= vma->vm_start)
1316 return;
1317
1318 if (vma->vm_file)
1319 uprobe_munmap(vma, start, end);
1320
1321 if (unlikely(is_pfn_mapping(vma)))
1322 untrack_pfn_vma(vma, 0, 0);
1323
1324 if (start != end) {
1325 if (unlikely(is_vm_hugetlb_page(vma))) {
1326 /*
1327 * It is undesirable to test vma->vm_file as it
1328 * should be non-null for valid hugetlb area.
1329 * However, vm_file will be NULL in the error
1330 * cleanup path of do_mmap_pgoff. When
1331 * hugetlbfs ->mmap method fails,
1332 * do_mmap_pgoff() nullifies vma->vm_file
1333 * before calling this function to clean up.
1334 * Since no pte has actually been setup, it is
1335 * safe to do nothing in this case.
1336 */
1337 if (vma->vm_file)
1338 unmap_hugepage_range(vma, start, end, NULL);
1339 } else
1340 unmap_page_range(tlb, vma, start, end, details);
1341 }
1342}
1343
1344/**
1345 * unmap_vmas - unmap a range of memory covered by a list of vma's
1346 * @tlb: address of the caller's struct mmu_gather
1347 * @vma: the starting vma
1348 * @start_addr: virtual address at which to start unmapping
1349 * @end_addr: virtual address at which to end unmapping
1350 *
1351 * Unmap all pages in the vma list.
1352 *
1353 * Only addresses between `start' and `end' will be unmapped.
1354 *
1355 * The VMA list must be sorted in ascending virtual address order.
1356 *
1357 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1358 * range after unmap_vmas() returns. So the only responsibility here is to
1359 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1360 * drops the lock and schedules.
1361 */
1362void unmap_vmas(struct mmu_gather *tlb,
1363 struct vm_area_struct *vma, unsigned long start_addr,
1364 unsigned long end_addr)
1365{
1366 struct mm_struct *mm = vma->vm_mm;
1367
1368 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1369 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1370 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1371 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1372}
1373
1374/**
1375 * zap_page_range - remove user pages in a given range
1376 * @vma: vm_area_struct holding the applicable pages
1377 * @start: starting address of pages to zap
1378 * @size: number of bytes to zap
1379 * @details: details of nonlinear truncation or shared cache invalidation
1380 *
1381 * Caller must protect the VMA list
1382 */
1383void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1384 unsigned long size, struct zap_details *details)
1385{
1386 struct mm_struct *mm = vma->vm_mm;
1387 struct mmu_gather tlb;
1388 unsigned long end = start + size;
1389
1390 lru_add_drain();
1391 tlb_gather_mmu(&tlb, mm, 0);
1392 update_hiwater_rss(mm);
1393 mmu_notifier_invalidate_range_start(mm, start, end);
1394 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1395 unmap_single_vma(&tlb, vma, start, end, details);
1396 mmu_notifier_invalidate_range_end(mm, start, end);
1397 tlb_finish_mmu(&tlb, start, end);
1398}
1399
1400/**
1401 * zap_page_range_single - remove user pages in a given range
1402 * @vma: vm_area_struct holding the applicable pages
1403 * @address: starting address of pages to zap
1404 * @size: number of bytes to zap
1405 * @details: details of nonlinear truncation or shared cache invalidation
1406 *
1407 * The range must fit into one VMA.
1408 */
1409static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1410 unsigned long size, struct zap_details *details)
1411{
1412 struct mm_struct *mm = vma->vm_mm;
1413 struct mmu_gather tlb;
1414 unsigned long end = address + size;
1415
1416 lru_add_drain();
1417 tlb_gather_mmu(&tlb, mm, 0);
1418 update_hiwater_rss(mm);
1419 mmu_notifier_invalidate_range_start(mm, address, end);
1420 unmap_single_vma(&tlb, vma, address, end, details);
1421 mmu_notifier_invalidate_range_end(mm, address, end);
1422 tlb_finish_mmu(&tlb, address, end);
1423}
1424
1425/**
1426 * zap_vma_ptes - remove ptes mapping the vma
1427 * @vma: vm_area_struct holding ptes to be zapped
1428 * @address: starting address of pages to zap
1429 * @size: number of bytes to zap
1430 *
1431 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1432 *
1433 * The entire address range must be fully contained within the vma.
1434 *
1435 * Returns 0 if successful.
1436 */
1437int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1438 unsigned long size)
1439{
1440 if (address < vma->vm_start || address + size > vma->vm_end ||
1441 !(vma->vm_flags & VM_PFNMAP))
1442 return -1;
1443 zap_page_range_single(vma, address, size, NULL);
1444 return 0;
1445}
1446EXPORT_SYMBOL_GPL(zap_vma_ptes);
1447
1448/**
1449 * follow_page - look up a page descriptor from a user-virtual address
1450 * @vma: vm_area_struct mapping @address
1451 * @address: virtual address to look up
1452 * @flags: flags modifying lookup behaviour
1453 *
1454 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1455 *
1456 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1457 * an error pointer if there is a mapping to something not represented
1458 * by a page descriptor (see also vm_normal_page()).
1459 */
1460struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1461 unsigned int flags)
1462{
1463 pgd_t *pgd;
1464 pud_t *pud;
1465 pmd_t *pmd;
1466 pte_t *ptep, pte;
1467 spinlock_t *ptl;
1468 struct page *page;
1469 struct mm_struct *mm = vma->vm_mm;
1470
1471 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1472 if (!IS_ERR(page)) {
1473 BUG_ON(flags & FOLL_GET);
1474 goto out;
1475 }
1476
1477 page = NULL;
1478 pgd = pgd_offset(mm, address);
1479 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1480 goto no_page_table;
1481
1482 pud = pud_offset(pgd, address);
1483 if (pud_none(*pud))
1484 goto no_page_table;
1485 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1486 BUG_ON(flags & FOLL_GET);
1487 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1488 goto out;
1489 }
1490 if (unlikely(pud_bad(*pud)))
1491 goto no_page_table;
1492
1493 pmd = pmd_offset(pud, address);
1494 if (pmd_none(*pmd))
1495 goto no_page_table;
1496 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1497 BUG_ON(flags & FOLL_GET);
1498 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1499 goto out;
1500 }
1501 if (pmd_trans_huge(*pmd)) {
1502 if (flags & FOLL_SPLIT) {
1503 split_huge_page_pmd(mm, pmd);
1504 goto split_fallthrough;
1505 }
1506 spin_lock(&mm->page_table_lock);
1507 if (likely(pmd_trans_huge(*pmd))) {
1508 if (unlikely(pmd_trans_splitting(*pmd))) {
1509 spin_unlock(&mm->page_table_lock);
1510 wait_split_huge_page(vma->anon_vma, pmd);
1511 } else {
1512 page = follow_trans_huge_pmd(mm, address,
1513 pmd, flags);
1514 spin_unlock(&mm->page_table_lock);
1515 goto out;
1516 }
1517 } else
1518 spin_unlock(&mm->page_table_lock);
1519 /* fall through */
1520 }
1521split_fallthrough:
1522 if (unlikely(pmd_bad(*pmd)))
1523 goto no_page_table;
1524
1525 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1526
1527 pte = *ptep;
1528 if (!pte_present(pte))
1529 goto no_page;
1530 if ((flags & FOLL_WRITE) && !pte_write(pte))
1531 goto unlock;
1532
1533 page = vm_normal_page(vma, address, pte);
1534 if (unlikely(!page)) {
1535 if ((flags & FOLL_DUMP) ||
1536 !is_zero_pfn(pte_pfn(pte)))
1537 goto bad_page;
1538 page = pte_page(pte);
1539 }
1540
1541 if (flags & FOLL_GET)
1542 get_page_foll(page);
1543 if (flags & FOLL_TOUCH) {
1544 if ((flags & FOLL_WRITE) &&
1545 !pte_dirty(pte) && !PageDirty(page))
1546 set_page_dirty(page);
1547 /*
1548 * pte_mkyoung() would be more correct here, but atomic care
1549 * is needed to avoid losing the dirty bit: it is easier to use
1550 * mark_page_accessed().
1551 */
1552 mark_page_accessed(page);
1553 }
1554 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1555 /*
1556 * The preliminary mapping check is mainly to avoid the
1557 * pointless overhead of lock_page on the ZERO_PAGE
1558 * which might bounce very badly if there is contention.
1559 *
1560 * If the page is already locked, we don't need to
1561 * handle it now - vmscan will handle it later if and
1562 * when it attempts to reclaim the page.
1563 */
1564 if (page->mapping && trylock_page(page)) {
1565 lru_add_drain(); /* push cached pages to LRU */
1566 /*
1567 * Because we lock page here and migration is
1568 * blocked by the pte's page reference, we need
1569 * only check for file-cache page truncation.
1570 */
1571 if (page->mapping)
1572 mlock_vma_page(page);
1573 unlock_page(page);
1574 }
1575 }
1576unlock:
1577 pte_unmap_unlock(ptep, ptl);
1578out:
1579 return page;
1580
1581bad_page:
1582 pte_unmap_unlock(ptep, ptl);
1583 return ERR_PTR(-EFAULT);
1584
1585no_page:
1586 pte_unmap_unlock(ptep, ptl);
1587 if (!pte_none(pte))
1588 return page;
1589
1590no_page_table:
1591 /*
1592 * When core dumping an enormous anonymous area that nobody
1593 * has touched so far, we don't want to allocate unnecessary pages or
1594 * page tables. Return error instead of NULL to skip handle_mm_fault,
1595 * then get_dump_page() will return NULL to leave a hole in the dump.
1596 * But we can only make this optimization where a hole would surely
1597 * be zero-filled if handle_mm_fault() actually did handle it.
1598 */
1599 if ((flags & FOLL_DUMP) &&
1600 (!vma->vm_ops || !vma->vm_ops->fault))
1601 return ERR_PTR(-EFAULT);
1602 return page;
1603}
1604
1605static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1606{
1607 return stack_guard_page_start(vma, addr) ||
1608 stack_guard_page_end(vma, addr+PAGE_SIZE);
1609}
1610
1611/**
1612 * __get_user_pages() - pin user pages in memory
1613 * @tsk: task_struct of target task
1614 * @mm: mm_struct of target mm
1615 * @start: starting user address
1616 * @nr_pages: number of pages from start to pin
1617 * @gup_flags: flags modifying pin behaviour
1618 * @pages: array that receives pointers to the pages pinned.
1619 * Should be at least nr_pages long. Or NULL, if caller
1620 * only intends to ensure the pages are faulted in.
1621 * @vmas: array of pointers to vmas corresponding to each page.
1622 * Or NULL if the caller does not require them.
1623 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1624 *
1625 * Returns number of pages pinned. This may be fewer than the number
1626 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1627 * were pinned, returns -errno. Each page returned must be released
1628 * with a put_page() call when it is finished with. vmas will only
1629 * remain valid while mmap_sem is held.
1630 *
1631 * Must be called with mmap_sem held for read or write.
1632 *
1633 * __get_user_pages walks a process's page tables and takes a reference to
1634 * each struct page that each user address corresponds to at a given
1635 * instant. That is, it takes the page that would be accessed if a user
1636 * thread accesses the given user virtual address at that instant.
1637 *
1638 * This does not guarantee that the page exists in the user mappings when
1639 * __get_user_pages returns, and there may even be a completely different
1640 * page there in some cases (eg. if mmapped pagecache has been invalidated
1641 * and subsequently re faulted). However it does guarantee that the page
1642 * won't be freed completely. And mostly callers simply care that the page
1643 * contains data that was valid *at some point in time*. Typically, an IO
1644 * or similar operation cannot guarantee anything stronger anyway because
1645 * locks can't be held over the syscall boundary.
1646 *
1647 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1648 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1649 * appropriate) must be called after the page is finished with, and
1650 * before put_page is called.
1651 *
1652 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1653 * or mmap_sem contention, and if waiting is needed to pin all pages,
1654 * *@nonblocking will be set to 0.
1655 *
1656 * In most cases, get_user_pages or get_user_pages_fast should be used
1657 * instead of __get_user_pages. __get_user_pages should be used only if
1658 * you need some special @gup_flags.
1659 */
1660int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1661 unsigned long start, int nr_pages, unsigned int gup_flags,
1662 struct page **pages, struct vm_area_struct **vmas,
1663 int *nonblocking)
1664{
1665 int i;
1666 unsigned long vm_flags;
1667
1668 if (nr_pages <= 0)
1669 return 0;
1670
1671 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1672
1673 /*
1674 * Require read or write permissions.
1675 * If FOLL_FORCE is set, we only require the "MAY" flags.
1676 */
1677 vm_flags = (gup_flags & FOLL_WRITE) ?
1678 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1679 vm_flags &= (gup_flags & FOLL_FORCE) ?
1680 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1681 i = 0;
1682
1683 do {
1684 struct vm_area_struct *vma;
1685
1686 vma = find_extend_vma(mm, start);
1687 if (!vma && in_gate_area(mm, start)) {
1688 unsigned long pg = start & PAGE_MASK;
1689 pgd_t *pgd;
1690 pud_t *pud;
1691 pmd_t *pmd;
1692 pte_t *pte;
1693
1694 /* user gate pages are read-only */
1695 if (gup_flags & FOLL_WRITE)
1696 return i ? : -EFAULT;
1697 if (pg > TASK_SIZE)
1698 pgd = pgd_offset_k(pg);
1699 else
1700 pgd = pgd_offset_gate(mm, pg);
1701 BUG_ON(pgd_none(*pgd));
1702 pud = pud_offset(pgd, pg);
1703 BUG_ON(pud_none(*pud));
1704 pmd = pmd_offset(pud, pg);
1705 if (pmd_none(*pmd))
1706 return i ? : -EFAULT;
1707 VM_BUG_ON(pmd_trans_huge(*pmd));
1708 pte = pte_offset_map(pmd, pg);
1709 if (pte_none(*pte)) {
1710 pte_unmap(pte);
1711 return i ? : -EFAULT;
1712 }
1713 vma = get_gate_vma(mm);
1714 if (pages) {
1715 struct page *page;
1716
1717 page = vm_normal_page(vma, start, *pte);
1718 if (!page) {
1719 if (!(gup_flags & FOLL_DUMP) &&
1720 is_zero_pfn(pte_pfn(*pte)))
1721 page = pte_page(*pte);
1722 else {
1723 pte_unmap(pte);
1724 return i ? : -EFAULT;
1725 }
1726 }
1727 pages[i] = page;
1728 get_page(page);
1729 }
1730 pte_unmap(pte);
1731 goto next_page;
1732 }
1733
1734 if (!vma ||
1735 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1736 !(vm_flags & vma->vm_flags))
1737 return i ? : -EFAULT;
1738
1739 if (is_vm_hugetlb_page(vma)) {
1740 i = follow_hugetlb_page(mm, vma, pages, vmas,
1741 &start, &nr_pages, i, gup_flags);
1742 continue;
1743 }
1744
1745 do {
1746 struct page *page;
1747 unsigned int foll_flags = gup_flags;
1748
1749 /*
1750 * If we have a pending SIGKILL, don't keep faulting
1751 * pages and potentially allocating memory.
1752 */
1753 if (unlikely(fatal_signal_pending(current)))
1754 return i ? i : -ERESTARTSYS;
1755
1756 cond_resched();
1757 while (!(page = follow_page(vma, start, foll_flags))) {
1758 int ret;
1759 unsigned int fault_flags = 0;
1760
1761 /* For mlock, just skip the stack guard page. */
1762 if (foll_flags & FOLL_MLOCK) {
1763 if (stack_guard_page(vma, start))
1764 goto next_page;
1765 }
1766 if (foll_flags & FOLL_WRITE)
1767 fault_flags |= FAULT_FLAG_WRITE;
1768 if (nonblocking)
1769 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1770 if (foll_flags & FOLL_NOWAIT)
1771 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1772
1773 ret = handle_mm_fault(mm, vma, start,
1774 fault_flags);
1775
1776 if (ret & VM_FAULT_ERROR) {
1777 if (ret & VM_FAULT_OOM)
1778 return i ? i : -ENOMEM;
1779 if (ret & (VM_FAULT_HWPOISON |
1780 VM_FAULT_HWPOISON_LARGE)) {
1781 if (i)
1782 return i;
1783 else if (gup_flags & FOLL_HWPOISON)
1784 return -EHWPOISON;
1785 else
1786 return -EFAULT;
1787 }
1788 if (ret & VM_FAULT_SIGBUS)
1789 return i ? i : -EFAULT;
1790 BUG();
1791 }
1792
1793 if (tsk) {
1794 if (ret & VM_FAULT_MAJOR)
1795 tsk->maj_flt++;
1796 else
1797 tsk->min_flt++;
1798 }
1799
1800 if (ret & VM_FAULT_RETRY) {
1801 if (nonblocking)
1802 *nonblocking = 0;
1803 return i;
1804 }
1805
1806 /*
1807 * The VM_FAULT_WRITE bit tells us that
1808 * do_wp_page has broken COW when necessary,
1809 * even if maybe_mkwrite decided not to set
1810 * pte_write. We can thus safely do subsequent
1811 * page lookups as if they were reads. But only
1812 * do so when looping for pte_write is futile:
1813 * in some cases userspace may also be wanting
1814 * to write to the gotten user page, which a
1815 * read fault here might prevent (a readonly
1816 * page might get reCOWed by userspace write).
1817 */
1818 if ((ret & VM_FAULT_WRITE) &&
1819 !(vma->vm_flags & VM_WRITE))
1820 foll_flags &= ~FOLL_WRITE;
1821
1822 cond_resched();
1823 }
1824 if (IS_ERR(page))
1825 return i ? i : PTR_ERR(page);
1826 if (pages) {
1827 pages[i] = page;
1828
1829 flush_anon_page(vma, page, start);
1830 flush_dcache_page(page);
1831 }
1832next_page:
1833 if (vmas)
1834 vmas[i] = vma;
1835 i++;
1836 start += PAGE_SIZE;
1837 nr_pages--;
1838 } while (nr_pages && start < vma->vm_end);
1839 } while (nr_pages);
1840 return i;
1841}
1842EXPORT_SYMBOL(__get_user_pages);
1843
1844/*
1845 * fixup_user_fault() - manually resolve a user page fault
1846 * @tsk: the task_struct to use for page fault accounting, or
1847 * NULL if faults are not to be recorded.
1848 * @mm: mm_struct of target mm
1849 * @address: user address
1850 * @fault_flags:flags to pass down to handle_mm_fault()
1851 *
1852 * This is meant to be called in the specific scenario where for locking reasons
1853 * we try to access user memory in atomic context (within a pagefault_disable()
1854 * section), this returns -EFAULT, and we want to resolve the user fault before
1855 * trying again.
1856 *
1857 * Typically this is meant to be used by the futex code.
1858 *
1859 * The main difference with get_user_pages() is that this function will
1860 * unconditionally call handle_mm_fault() which will in turn perform all the
1861 * necessary SW fixup of the dirty and young bits in the PTE, while
1862 * handle_mm_fault() only guarantees to update these in the struct page.
1863 *
1864 * This is important for some architectures where those bits also gate the
1865 * access permission to the page because they are maintained in software. On
1866 * such architectures, gup() will not be enough to make a subsequent access
1867 * succeed.
1868 *
1869 * This should be called with the mm_sem held for read.
1870 */
1871int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1872 unsigned long address, unsigned int fault_flags)
1873{
1874 struct vm_area_struct *vma;
1875 int ret;
1876
1877 vma = find_extend_vma(mm, address);
1878 if (!vma || address < vma->vm_start)
1879 return -EFAULT;
1880
1881 ret = handle_mm_fault(mm, vma, address, fault_flags);
1882 if (ret & VM_FAULT_ERROR) {
1883 if (ret & VM_FAULT_OOM)
1884 return -ENOMEM;
1885 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1886 return -EHWPOISON;
1887 if (ret & VM_FAULT_SIGBUS)
1888 return -EFAULT;
1889 BUG();
1890 }
1891 if (tsk) {
1892 if (ret & VM_FAULT_MAJOR)
1893 tsk->maj_flt++;
1894 else
1895 tsk->min_flt++;
1896 }
1897 return 0;
1898}
1899
1900/*
1901 * get_user_pages() - pin user pages in memory
1902 * @tsk: the task_struct to use for page fault accounting, or
1903 * NULL if faults are not to be recorded.
1904 * @mm: mm_struct of target mm
1905 * @start: starting user address
1906 * @nr_pages: number of pages from start to pin
1907 * @write: whether pages will be written to by the caller
1908 * @force: whether to force write access even if user mapping is
1909 * readonly. This will result in the page being COWed even
1910 * in MAP_SHARED mappings. You do not want this.
1911 * @pages: array that receives pointers to the pages pinned.
1912 * Should be at least nr_pages long. Or NULL, if caller
1913 * only intends to ensure the pages are faulted in.
1914 * @vmas: array of pointers to vmas corresponding to each page.
1915 * Or NULL if the caller does not require them.
1916 *
1917 * Returns number of pages pinned. This may be fewer than the number
1918 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1919 * were pinned, returns -errno. Each page returned must be released
1920 * with a put_page() call when it is finished with. vmas will only
1921 * remain valid while mmap_sem is held.
1922 *
1923 * Must be called with mmap_sem held for read or write.
1924 *
1925 * get_user_pages walks a process's page tables and takes a reference to
1926 * each struct page that each user address corresponds to at a given
1927 * instant. That is, it takes the page that would be accessed if a user
1928 * thread accesses the given user virtual address at that instant.
1929 *
1930 * This does not guarantee that the page exists in the user mappings when
1931 * get_user_pages returns, and there may even be a completely different
1932 * page there in some cases (eg. if mmapped pagecache has been invalidated
1933 * and subsequently re faulted). However it does guarantee that the page
1934 * won't be freed completely. And mostly callers simply care that the page
1935 * contains data that was valid *at some point in time*. Typically, an IO
1936 * or similar operation cannot guarantee anything stronger anyway because
1937 * locks can't be held over the syscall boundary.
1938 *
1939 * If write=0, the page must not be written to. If the page is written to,
1940 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1941 * after the page is finished with, and before put_page is called.
1942 *
1943 * get_user_pages is typically used for fewer-copy IO operations, to get a
1944 * handle on the memory by some means other than accesses via the user virtual
1945 * addresses. The pages may be submitted for DMA to devices or accessed via
1946 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1947 * use the correct cache flushing APIs.
1948 *
1949 * See also get_user_pages_fast, for performance critical applications.
1950 */
1951int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1952 unsigned long start, int nr_pages, int write, int force,
1953 struct page **pages, struct vm_area_struct **vmas)
1954{
1955 int flags = FOLL_TOUCH;
1956
1957 if (pages)
1958 flags |= FOLL_GET;
1959 if (write)
1960 flags |= FOLL_WRITE;
1961 if (force)
1962 flags |= FOLL_FORCE;
1963
1964 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1965 NULL);
1966}
1967EXPORT_SYMBOL(get_user_pages);
1968
1969/**
1970 * get_dump_page() - pin user page in memory while writing it to core dump
1971 * @addr: user address
1972 *
1973 * Returns struct page pointer of user page pinned for dump,
1974 * to be freed afterwards by page_cache_release() or put_page().
1975 *
1976 * Returns NULL on any kind of failure - a hole must then be inserted into
1977 * the corefile, to preserve alignment with its headers; and also returns
1978 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1979 * allowing a hole to be left in the corefile to save diskspace.
1980 *
1981 * Called without mmap_sem, but after all other threads have been killed.
1982 */
1983#ifdef CONFIG_ELF_CORE
1984struct page *get_dump_page(unsigned long addr)
1985{
1986 struct vm_area_struct *vma;
1987 struct page *page;
1988
1989 if (__get_user_pages(current, current->mm, addr, 1,
1990 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1991 NULL) < 1)
1992 return NULL;
1993 flush_cache_page(vma, addr, page_to_pfn(page));
1994 return page;
1995}
1996#endif /* CONFIG_ELF_CORE */
1997
1998pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1999 spinlock_t **ptl)
2000{
2001 pgd_t * pgd = pgd_offset(mm, addr);
2002 pud_t * pud = pud_alloc(mm, pgd, addr);
2003 if (pud) {
2004 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2005 if (pmd) {
2006 VM_BUG_ON(pmd_trans_huge(*pmd));
2007 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2008 }
2009 }
2010 return NULL;
2011}
2012
2013/*
2014 * This is the old fallback for page remapping.
2015 *
2016 * For historical reasons, it only allows reserved pages. Only
2017 * old drivers should use this, and they needed to mark their
2018 * pages reserved for the old functions anyway.
2019 */
2020static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2021 struct page *page, pgprot_t prot)
2022{
2023 struct mm_struct *mm = vma->vm_mm;
2024 int retval;
2025 pte_t *pte;
2026 spinlock_t *ptl;
2027
2028 retval = -EINVAL;
2029 if (PageAnon(page))
2030 goto out;
2031 retval = -ENOMEM;
2032 flush_dcache_page(page);
2033 pte = get_locked_pte(mm, addr, &ptl);
2034 if (!pte)
2035 goto out;
2036 retval = -EBUSY;
2037 if (!pte_none(*pte))
2038 goto out_unlock;
2039
2040 /* Ok, finally just insert the thing.. */
2041 get_page(page);
2042 inc_mm_counter_fast(mm, MM_FILEPAGES);
2043 page_add_file_rmap(page);
2044 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2045
2046 retval = 0;
2047 pte_unmap_unlock(pte, ptl);
2048 return retval;
2049out_unlock:
2050 pte_unmap_unlock(pte, ptl);
2051out:
2052 return retval;
2053}
2054
2055/**
2056 * vm_insert_page - insert single page into user vma
2057 * @vma: user vma to map to
2058 * @addr: target user address of this page
2059 * @page: source kernel page
2060 *
2061 * This allows drivers to insert individual pages they've allocated
2062 * into a user vma.
2063 *
2064 * The page has to be a nice clean _individual_ kernel allocation.
2065 * If you allocate a compound page, you need to have marked it as
2066 * such (__GFP_COMP), or manually just split the page up yourself
2067 * (see split_page()).
2068 *
2069 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2070 * took an arbitrary page protection parameter. This doesn't allow
2071 * that. Your vma protection will have to be set up correctly, which
2072 * means that if you want a shared writable mapping, you'd better
2073 * ask for a shared writable mapping!
2074 *
2075 * The page does not need to be reserved.
2076 */
2077int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2078 struct page *page)
2079{
2080 if (addr < vma->vm_start || addr >= vma->vm_end)
2081 return -EFAULT;
2082 if (!page_count(page))
2083 return -EINVAL;
2084 vma->vm_flags |= VM_INSERTPAGE;
2085 return insert_page(vma, addr, page, vma->vm_page_prot);
2086}
2087EXPORT_SYMBOL(vm_insert_page);
2088
2089static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2090 unsigned long pfn, pgprot_t prot)
2091{
2092 struct mm_struct *mm = vma->vm_mm;
2093 int retval;
2094 pte_t *pte, entry;
2095 spinlock_t *ptl;
2096
2097 retval = -ENOMEM;
2098 pte = get_locked_pte(mm, addr, &ptl);
2099 if (!pte)
2100 goto out;
2101 retval = -EBUSY;
2102 if (!pte_none(*pte))
2103 goto out_unlock;
2104
2105 /* Ok, finally just insert the thing.. */
2106 entry = pte_mkspecial(pfn_pte(pfn, prot));
2107 set_pte_at(mm, addr, pte, entry);
2108 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2109
2110 retval = 0;
2111out_unlock:
2112 pte_unmap_unlock(pte, ptl);
2113out:
2114 return retval;
2115}
2116
2117/**
2118 * vm_insert_pfn - insert single pfn into user vma
2119 * @vma: user vma to map to
2120 * @addr: target user address of this page
2121 * @pfn: source kernel pfn
2122 *
2123 * Similar to vm_inert_page, this allows drivers to insert individual pages
2124 * they've allocated into a user vma. Same comments apply.
2125 *
2126 * This function should only be called from a vm_ops->fault handler, and
2127 * in that case the handler should return NULL.
2128 *
2129 * vma cannot be a COW mapping.
2130 *
2131 * As this is called only for pages that do not currently exist, we
2132 * do not need to flush old virtual caches or the TLB.
2133 */
2134int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2135 unsigned long pfn)
2136{
2137 int ret;
2138 pgprot_t pgprot = vma->vm_page_prot;
2139 /*
2140 * Technically, architectures with pte_special can avoid all these
2141 * restrictions (same for remap_pfn_range). However we would like
2142 * consistency in testing and feature parity among all, so we should
2143 * try to keep these invariants in place for everybody.
2144 */
2145 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2146 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2147 (VM_PFNMAP|VM_MIXEDMAP));
2148 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2149 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2150
2151 if (addr < vma->vm_start || addr >= vma->vm_end)
2152 return -EFAULT;
2153 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2154 return -EINVAL;
2155
2156 ret = insert_pfn(vma, addr, pfn, pgprot);
2157
2158 if (ret)
2159 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2160
2161 return ret;
2162}
2163EXPORT_SYMBOL(vm_insert_pfn);
2164
2165int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2166 unsigned long pfn)
2167{
2168 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2169
2170 if (addr < vma->vm_start || addr >= vma->vm_end)
2171 return -EFAULT;
2172
2173 /*
2174 * If we don't have pte special, then we have to use the pfn_valid()
2175 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2176 * refcount the page if pfn_valid is true (hence insert_page rather
2177 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2178 * without pte special, it would there be refcounted as a normal page.
2179 */
2180 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2181 struct page *page;
2182
2183 page = pfn_to_page(pfn);
2184 return insert_page(vma, addr, page, vma->vm_page_prot);
2185 }
2186 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2187}
2188EXPORT_SYMBOL(vm_insert_mixed);
2189
2190/*
2191 * maps a range of physical memory into the requested pages. the old
2192 * mappings are removed. any references to nonexistent pages results
2193 * in null mappings (currently treated as "copy-on-access")
2194 */
2195static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2196 unsigned long addr, unsigned long end,
2197 unsigned long pfn, pgprot_t prot)
2198{
2199 pte_t *pte;
2200 spinlock_t *ptl;
2201
2202 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2203 if (!pte)
2204 return -ENOMEM;
2205 arch_enter_lazy_mmu_mode();
2206 do {
2207 BUG_ON(!pte_none(*pte));
2208 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2209 pfn++;
2210 } while (pte++, addr += PAGE_SIZE, addr != end);
2211 arch_leave_lazy_mmu_mode();
2212 pte_unmap_unlock(pte - 1, ptl);
2213 return 0;
2214}
2215
2216static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2217 unsigned long addr, unsigned long end,
2218 unsigned long pfn, pgprot_t prot)
2219{
2220 pmd_t *pmd;
2221 unsigned long next;
2222
2223 pfn -= addr >> PAGE_SHIFT;
2224 pmd = pmd_alloc(mm, pud, addr);
2225 if (!pmd)
2226 return -ENOMEM;
2227 VM_BUG_ON(pmd_trans_huge(*pmd));
2228 do {
2229 next = pmd_addr_end(addr, end);
2230 if (remap_pte_range(mm, pmd, addr, next,
2231 pfn + (addr >> PAGE_SHIFT), prot))
2232 return -ENOMEM;
2233 } while (pmd++, addr = next, addr != end);
2234 return 0;
2235}
2236
2237static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2238 unsigned long addr, unsigned long end,
2239 unsigned long pfn, pgprot_t prot)
2240{
2241 pud_t *pud;
2242 unsigned long next;
2243
2244 pfn -= addr >> PAGE_SHIFT;
2245 pud = pud_alloc(mm, pgd, addr);
2246 if (!pud)
2247 return -ENOMEM;
2248 do {
2249 next = pud_addr_end(addr, end);
2250 if (remap_pmd_range(mm, pud, addr, next,
2251 pfn + (addr >> PAGE_SHIFT), prot))
2252 return -ENOMEM;
2253 } while (pud++, addr = next, addr != end);
2254 return 0;
2255}
2256
2257/**
2258 * remap_pfn_range - remap kernel memory to userspace
2259 * @vma: user vma to map to
2260 * @addr: target user address to start at
2261 * @pfn: physical address of kernel memory
2262 * @size: size of map area
2263 * @prot: page protection flags for this mapping
2264 *
2265 * Note: this is only safe if the mm semaphore is held when called.
2266 */
2267int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2268 unsigned long pfn, unsigned long size, pgprot_t prot)
2269{
2270 pgd_t *pgd;
2271 unsigned long next;
2272 unsigned long end = addr + PAGE_ALIGN(size);
2273 struct mm_struct *mm = vma->vm_mm;
2274 int err;
2275
2276 /*
2277 * Physically remapped pages are special. Tell the
2278 * rest of the world about it:
2279 * VM_IO tells people not to look at these pages
2280 * (accesses can have side effects).
2281 * VM_RESERVED is specified all over the place, because
2282 * in 2.4 it kept swapout's vma scan off this vma; but
2283 * in 2.6 the LRU scan won't even find its pages, so this
2284 * flag means no more than count its pages in reserved_vm,
2285 * and omit it from core dump, even when VM_IO turned off.
2286 * VM_PFNMAP tells the core MM that the base pages are just
2287 * raw PFN mappings, and do not have a "struct page" associated
2288 * with them.
2289 *
2290 * There's a horrible special case to handle copy-on-write
2291 * behaviour that some programs depend on. We mark the "original"
2292 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2293 */
2294 if (addr == vma->vm_start && end == vma->vm_end) {
2295 vma->vm_pgoff = pfn;
2296 vma->vm_flags |= VM_PFN_AT_MMAP;
2297 } else if (is_cow_mapping(vma->vm_flags))
2298 return -EINVAL;
2299
2300 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2301
2302 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2303 if (err) {
2304 /*
2305 * To indicate that track_pfn related cleanup is not
2306 * needed from higher level routine calling unmap_vmas
2307 */
2308 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2309 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2310 return -EINVAL;
2311 }
2312
2313 BUG_ON(addr >= end);
2314 pfn -= addr >> PAGE_SHIFT;
2315 pgd = pgd_offset(mm, addr);
2316 flush_cache_range(vma, addr, end);
2317 do {
2318 next = pgd_addr_end(addr, end);
2319 err = remap_pud_range(mm, pgd, addr, next,
2320 pfn + (addr >> PAGE_SHIFT), prot);
2321 if (err)
2322 break;
2323 } while (pgd++, addr = next, addr != end);
2324
2325 if (err)
2326 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2327
2328 return err;
2329}
2330EXPORT_SYMBOL(remap_pfn_range);
2331
2332static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2333 unsigned long addr, unsigned long end,
2334 pte_fn_t fn, void *data)
2335{
2336 pte_t *pte;
2337 int err;
2338 pgtable_t token;
2339 spinlock_t *uninitialized_var(ptl);
2340
2341 pte = (mm == &init_mm) ?
2342 pte_alloc_kernel(pmd, addr) :
2343 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2344 if (!pte)
2345 return -ENOMEM;
2346
2347 BUG_ON(pmd_huge(*pmd));
2348
2349 arch_enter_lazy_mmu_mode();
2350
2351 token = pmd_pgtable(*pmd);
2352
2353 do {
2354 err = fn(pte++, token, addr, data);
2355 if (err)
2356 break;
2357 } while (addr += PAGE_SIZE, addr != end);
2358
2359 arch_leave_lazy_mmu_mode();
2360
2361 if (mm != &init_mm)
2362 pte_unmap_unlock(pte-1, ptl);
2363 return err;
2364}
2365
2366static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2367 unsigned long addr, unsigned long end,
2368 pte_fn_t fn, void *data)
2369{
2370 pmd_t *pmd;
2371 unsigned long next;
2372 int err;
2373
2374 BUG_ON(pud_huge(*pud));
2375
2376 pmd = pmd_alloc(mm, pud, addr);
2377 if (!pmd)
2378 return -ENOMEM;
2379 do {
2380 next = pmd_addr_end(addr, end);
2381 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2382 if (err)
2383 break;
2384 } while (pmd++, addr = next, addr != end);
2385 return err;
2386}
2387
2388static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2389 unsigned long addr, unsigned long end,
2390 pte_fn_t fn, void *data)
2391{
2392 pud_t *pud;
2393 unsigned long next;
2394 int err;
2395
2396 pud = pud_alloc(mm, pgd, addr);
2397 if (!pud)
2398 return -ENOMEM;
2399 do {
2400 next = pud_addr_end(addr, end);
2401 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2402 if (err)
2403 break;
2404 } while (pud++, addr = next, addr != end);
2405 return err;
2406}
2407
2408/*
2409 * Scan a region of virtual memory, filling in page tables as necessary
2410 * and calling a provided function on each leaf page table.
2411 */
2412int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2413 unsigned long size, pte_fn_t fn, void *data)
2414{
2415 pgd_t *pgd;
2416 unsigned long next;
2417 unsigned long end = addr + size;
2418 int err;
2419
2420 BUG_ON(addr >= end);
2421 pgd = pgd_offset(mm, addr);
2422 do {
2423 next = pgd_addr_end(addr, end);
2424 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2425 if (err)
2426 break;
2427 } while (pgd++, addr = next, addr != end);
2428
2429 return err;
2430}
2431EXPORT_SYMBOL_GPL(apply_to_page_range);
2432
2433/*
2434 * handle_pte_fault chooses page fault handler according to an entry
2435 * which was read non-atomically. Before making any commitment, on
2436 * those architectures or configurations (e.g. i386 with PAE) which
2437 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2438 * must check under lock before unmapping the pte and proceeding
2439 * (but do_wp_page is only called after already making such a check;
2440 * and do_anonymous_page can safely check later on).
2441 */
2442static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2443 pte_t *page_table, pte_t orig_pte)
2444{
2445 int same = 1;
2446#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2447 if (sizeof(pte_t) > sizeof(unsigned long)) {
2448 spinlock_t *ptl = pte_lockptr(mm, pmd);
2449 spin_lock(ptl);
2450 same = pte_same(*page_table, orig_pte);
2451 spin_unlock(ptl);
2452 }
2453#endif
2454 pte_unmap(page_table);
2455 return same;
2456}
2457
2458static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2459{
2460 /*
2461 * If the source page was a PFN mapping, we don't have
2462 * a "struct page" for it. We do a best-effort copy by
2463 * just copying from the original user address. If that
2464 * fails, we just zero-fill it. Live with it.
2465 */
2466 if (unlikely(!src)) {
2467 void *kaddr = kmap_atomic(dst);
2468 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2469
2470 /*
2471 * This really shouldn't fail, because the page is there
2472 * in the page tables. But it might just be unreadable,
2473 * in which case we just give up and fill the result with
2474 * zeroes.
2475 */
2476 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2477 clear_page(kaddr);
2478 kunmap_atomic(kaddr);
2479 flush_dcache_page(dst);
2480 } else
2481 copy_user_highpage(dst, src, va, vma);
2482}
2483
2484/*
2485 * This routine handles present pages, when users try to write
2486 * to a shared page. It is done by copying the page to a new address
2487 * and decrementing the shared-page counter for the old page.
2488 *
2489 * Note that this routine assumes that the protection checks have been
2490 * done by the caller (the low-level page fault routine in most cases).
2491 * Thus we can safely just mark it writable once we've done any necessary
2492 * COW.
2493 *
2494 * We also mark the page dirty at this point even though the page will
2495 * change only once the write actually happens. This avoids a few races,
2496 * and potentially makes it more efficient.
2497 *
2498 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2499 * but allow concurrent faults), with pte both mapped and locked.
2500 * We return with mmap_sem still held, but pte unmapped and unlocked.
2501 */
2502static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2503 unsigned long address, pte_t *page_table, pmd_t *pmd,
2504 spinlock_t *ptl, pte_t orig_pte)
2505 __releases(ptl)
2506{
2507 struct page *old_page, *new_page;
2508 pte_t entry;
2509 int ret = 0;
2510 int page_mkwrite = 0;
2511 struct page *dirty_page = NULL;
2512
2513 old_page = vm_normal_page(vma, address, orig_pte);
2514 if (!old_page) {
2515 /*
2516 * VM_MIXEDMAP !pfn_valid() case
2517 *
2518 * We should not cow pages in a shared writeable mapping.
2519 * Just mark the pages writable as we can't do any dirty
2520 * accounting on raw pfn maps.
2521 */
2522 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2523 (VM_WRITE|VM_SHARED))
2524 goto reuse;
2525 goto gotten;
2526 }
2527
2528 /*
2529 * Take out anonymous pages first, anonymous shared vmas are
2530 * not dirty accountable.
2531 */
2532 if (PageAnon(old_page) && !PageKsm(old_page)) {
2533 if (!trylock_page(old_page)) {
2534 page_cache_get(old_page);
2535 pte_unmap_unlock(page_table, ptl);
2536 lock_page(old_page);
2537 page_table = pte_offset_map_lock(mm, pmd, address,
2538 &ptl);
2539 if (!pte_same(*page_table, orig_pte)) {
2540 unlock_page(old_page);
2541 goto unlock;
2542 }
2543 page_cache_release(old_page);
2544 }
2545 if (reuse_swap_page(old_page)) {
2546 /*
2547 * The page is all ours. Move it to our anon_vma so
2548 * the rmap code will not search our parent or siblings.
2549 * Protected against the rmap code by the page lock.
2550 */
2551 page_move_anon_rmap(old_page, vma, address);
2552 unlock_page(old_page);
2553 goto reuse;
2554 }
2555 unlock_page(old_page);
2556 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2557 (VM_WRITE|VM_SHARED))) {
2558 /*
2559 * Only catch write-faults on shared writable pages,
2560 * read-only shared pages can get COWed by
2561 * get_user_pages(.write=1, .force=1).
2562 */
2563 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2564 struct vm_fault vmf;
2565 int tmp;
2566
2567 vmf.virtual_address = (void __user *)(address &
2568 PAGE_MASK);
2569 vmf.pgoff = old_page->index;
2570 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2571 vmf.page = old_page;
2572
2573 /*
2574 * Notify the address space that the page is about to
2575 * become writable so that it can prohibit this or wait
2576 * for the page to get into an appropriate state.
2577 *
2578 * We do this without the lock held, so that it can
2579 * sleep if it needs to.
2580 */
2581 page_cache_get(old_page);
2582 pte_unmap_unlock(page_table, ptl);
2583
2584 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2585 if (unlikely(tmp &
2586 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2587 ret = tmp;
2588 goto unwritable_page;
2589 }
2590 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2591 lock_page(old_page);
2592 if (!old_page->mapping) {
2593 ret = 0; /* retry the fault */
2594 unlock_page(old_page);
2595 goto unwritable_page;
2596 }
2597 } else
2598 VM_BUG_ON(!PageLocked(old_page));
2599
2600 /*
2601 * Since we dropped the lock we need to revalidate
2602 * the PTE as someone else may have changed it. If
2603 * they did, we just return, as we can count on the
2604 * MMU to tell us if they didn't also make it writable.
2605 */
2606 page_table = pte_offset_map_lock(mm, pmd, address,
2607 &ptl);
2608 if (!pte_same(*page_table, orig_pte)) {
2609 unlock_page(old_page);
2610 goto unlock;
2611 }
2612
2613 page_mkwrite = 1;
2614 }
2615 dirty_page = old_page;
2616 get_page(dirty_page);
2617
2618reuse:
2619 flush_cache_page(vma, address, pte_pfn(orig_pte));
2620 entry = pte_mkyoung(orig_pte);
2621 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2622 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2623 update_mmu_cache(vma, address, page_table);
2624 pte_unmap_unlock(page_table, ptl);
2625 ret |= VM_FAULT_WRITE;
2626
2627 if (!dirty_page)
2628 return ret;
2629
2630 /*
2631 * Yes, Virginia, this is actually required to prevent a race
2632 * with clear_page_dirty_for_io() from clearing the page dirty
2633 * bit after it clear all dirty ptes, but before a racing
2634 * do_wp_page installs a dirty pte.
2635 *
2636 * __do_fault is protected similarly.
2637 */
2638 if (!page_mkwrite) {
2639 wait_on_page_locked(dirty_page);
2640 set_page_dirty_balance(dirty_page, page_mkwrite);
2641 }
2642 put_page(dirty_page);
2643 if (page_mkwrite) {
2644 struct address_space *mapping = dirty_page->mapping;
2645
2646 set_page_dirty(dirty_page);
2647 unlock_page(dirty_page);
2648 page_cache_release(dirty_page);
2649 if (mapping) {
2650 /*
2651 * Some device drivers do not set page.mapping
2652 * but still dirty their pages
2653 */
2654 balance_dirty_pages_ratelimited(mapping);
2655 }
2656 }
2657
2658 /* file_update_time outside page_lock */
2659 if (vma->vm_file)
2660 file_update_time(vma->vm_file);
2661
2662 return ret;
2663 }
2664
2665 /*
2666 * Ok, we need to copy. Oh, well..
2667 */
2668 page_cache_get(old_page);
2669gotten:
2670 pte_unmap_unlock(page_table, ptl);
2671
2672 if (unlikely(anon_vma_prepare(vma)))
2673 goto oom;
2674
2675 if (is_zero_pfn(pte_pfn(orig_pte))) {
2676 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2677 if (!new_page)
2678 goto oom;
2679 } else {
2680 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2681 if (!new_page)
2682 goto oom;
2683 cow_user_page(new_page, old_page, address, vma);
2684 }
2685 __SetPageUptodate(new_page);
2686
2687 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2688 goto oom_free_new;
2689
2690 /*
2691 * Re-check the pte - we dropped the lock
2692 */
2693 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2694 if (likely(pte_same(*page_table, orig_pte))) {
2695 if (old_page) {
2696 if (!PageAnon(old_page)) {
2697 dec_mm_counter_fast(mm, MM_FILEPAGES);
2698 inc_mm_counter_fast(mm, MM_ANONPAGES);
2699 }
2700 } else
2701 inc_mm_counter_fast(mm, MM_ANONPAGES);
2702 flush_cache_page(vma, address, pte_pfn(orig_pte));
2703 entry = mk_pte(new_page, vma->vm_page_prot);
2704 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2705 /*
2706 * Clear the pte entry and flush it first, before updating the
2707 * pte with the new entry. This will avoid a race condition
2708 * seen in the presence of one thread doing SMC and another
2709 * thread doing COW.
2710 */
2711 ptep_clear_flush(vma, address, page_table);
2712 page_add_new_anon_rmap(new_page, vma, address);
2713 /*
2714 * We call the notify macro here because, when using secondary
2715 * mmu page tables (such as kvm shadow page tables), we want the
2716 * new page to be mapped directly into the secondary page table.
2717 */
2718 set_pte_at_notify(mm, address, page_table, entry);
2719 update_mmu_cache(vma, address, page_table);
2720 if (old_page) {
2721 /*
2722 * Only after switching the pte to the new page may
2723 * we remove the mapcount here. Otherwise another
2724 * process may come and find the rmap count decremented
2725 * before the pte is switched to the new page, and
2726 * "reuse" the old page writing into it while our pte
2727 * here still points into it and can be read by other
2728 * threads.
2729 *
2730 * The critical issue is to order this
2731 * page_remove_rmap with the ptp_clear_flush above.
2732 * Those stores are ordered by (if nothing else,)
2733 * the barrier present in the atomic_add_negative
2734 * in page_remove_rmap.
2735 *
2736 * Then the TLB flush in ptep_clear_flush ensures that
2737 * no process can access the old page before the
2738 * decremented mapcount is visible. And the old page
2739 * cannot be reused until after the decremented
2740 * mapcount is visible. So transitively, TLBs to
2741 * old page will be flushed before it can be reused.
2742 */
2743 page_remove_rmap(old_page);
2744 }
2745
2746 /* Free the old page.. */
2747 new_page = old_page;
2748 ret |= VM_FAULT_WRITE;
2749 } else
2750 mem_cgroup_uncharge_page(new_page);
2751
2752 if (new_page)
2753 page_cache_release(new_page);
2754unlock:
2755 pte_unmap_unlock(page_table, ptl);
2756 if (old_page) {
2757 /*
2758 * Don't let another task, with possibly unlocked vma,
2759 * keep the mlocked page.
2760 */
2761 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2762 lock_page(old_page); /* LRU manipulation */
2763 munlock_vma_page(old_page);
2764 unlock_page(old_page);
2765 }
2766 page_cache_release(old_page);
2767 }
2768 return ret;
2769oom_free_new:
2770 page_cache_release(new_page);
2771oom:
2772 if (old_page) {
2773 if (page_mkwrite) {
2774 unlock_page(old_page);
2775 page_cache_release(old_page);
2776 }
2777 page_cache_release(old_page);
2778 }
2779 return VM_FAULT_OOM;
2780
2781unwritable_page:
2782 page_cache_release(old_page);
2783 return ret;
2784}
2785
2786static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2787 unsigned long start_addr, unsigned long end_addr,
2788 struct zap_details *details)
2789{
2790 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2791}
2792
2793static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2794 struct zap_details *details)
2795{
2796 struct vm_area_struct *vma;
2797 struct prio_tree_iter iter;
2798 pgoff_t vba, vea, zba, zea;
2799
2800 vma_prio_tree_foreach(vma, &iter, root,
2801 details->first_index, details->last_index) {
2802
2803 vba = vma->vm_pgoff;
2804 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2805 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2806 zba = details->first_index;
2807 if (zba < vba)
2808 zba = vba;
2809 zea = details->last_index;
2810 if (zea > vea)
2811 zea = vea;
2812
2813 unmap_mapping_range_vma(vma,
2814 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2815 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2816 details);
2817 }
2818}
2819
2820static inline void unmap_mapping_range_list(struct list_head *head,
2821 struct zap_details *details)
2822{
2823 struct vm_area_struct *vma;
2824
2825 /*
2826 * In nonlinear VMAs there is no correspondence between virtual address
2827 * offset and file offset. So we must perform an exhaustive search
2828 * across *all* the pages in each nonlinear VMA, not just the pages
2829 * whose virtual address lies outside the file truncation point.
2830 */
2831 list_for_each_entry(vma, head, shared.vm_set.list) {
2832 details->nonlinear_vma = vma;
2833 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2834 }
2835}
2836
2837/**
2838 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2839 * @mapping: the address space containing mmaps to be unmapped.
2840 * @holebegin: byte in first page to unmap, relative to the start of
2841 * the underlying file. This will be rounded down to a PAGE_SIZE
2842 * boundary. Note that this is different from truncate_pagecache(), which
2843 * must keep the partial page. In contrast, we must get rid of
2844 * partial pages.
2845 * @holelen: size of prospective hole in bytes. This will be rounded
2846 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2847 * end of the file.
2848 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2849 * but 0 when invalidating pagecache, don't throw away private data.
2850 */
2851void unmap_mapping_range(struct address_space *mapping,
2852 loff_t const holebegin, loff_t const holelen, int even_cows)
2853{
2854 struct zap_details details;
2855 pgoff_t hba = holebegin >> PAGE_SHIFT;
2856 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2857
2858 /* Check for overflow. */
2859 if (sizeof(holelen) > sizeof(hlen)) {
2860 long long holeend =
2861 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2862 if (holeend & ~(long long)ULONG_MAX)
2863 hlen = ULONG_MAX - hba + 1;
2864 }
2865
2866 details.check_mapping = even_cows? NULL: mapping;
2867 details.nonlinear_vma = NULL;
2868 details.first_index = hba;
2869 details.last_index = hba + hlen - 1;
2870 if (details.last_index < details.first_index)
2871 details.last_index = ULONG_MAX;
2872
2873
2874 mutex_lock(&mapping->i_mmap_mutex);
2875 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2876 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2877 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2878 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2879 mutex_unlock(&mapping->i_mmap_mutex);
2880}
2881EXPORT_SYMBOL(unmap_mapping_range);
2882
2883/*
2884 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2885 * but allow concurrent faults), and pte mapped but not yet locked.
2886 * We return with mmap_sem still held, but pte unmapped and unlocked.
2887 */
2888static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2889 unsigned long address, pte_t *page_table, pmd_t *pmd,
2890 unsigned int flags, pte_t orig_pte)
2891{
2892 spinlock_t *ptl;
2893 struct page *page, *swapcache = NULL;
2894 swp_entry_t entry;
2895 pte_t pte;
2896 int locked;
2897 struct mem_cgroup *ptr;
2898 int exclusive = 0;
2899 int ret = 0;
2900
2901 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2902 goto out;
2903
2904 entry = pte_to_swp_entry(orig_pte);
2905 if (unlikely(non_swap_entry(entry))) {
2906 if (is_migration_entry(entry)) {
2907 migration_entry_wait(mm, pmd, address);
2908 } else if (is_hwpoison_entry(entry)) {
2909 ret = VM_FAULT_HWPOISON;
2910 } else {
2911 print_bad_pte(vma, address, orig_pte, NULL);
2912 ret = VM_FAULT_SIGBUS;
2913 }
2914 goto out;
2915 }
2916 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2917 page = lookup_swap_cache(entry);
2918 if (!page) {
2919 page = swapin_readahead(entry,
2920 GFP_HIGHUSER_MOVABLE, vma, address);
2921 if (!page) {
2922 /*
2923 * Back out if somebody else faulted in this pte
2924 * while we released the pte lock.
2925 */
2926 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2927 if (likely(pte_same(*page_table, orig_pte)))
2928 ret = VM_FAULT_OOM;
2929 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2930 goto unlock;
2931 }
2932
2933 /* Had to read the page from swap area: Major fault */
2934 ret = VM_FAULT_MAJOR;
2935 count_vm_event(PGMAJFAULT);
2936 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2937 } else if (PageHWPoison(page)) {
2938 /*
2939 * hwpoisoned dirty swapcache pages are kept for killing
2940 * owner processes (which may be unknown at hwpoison time)
2941 */
2942 ret = VM_FAULT_HWPOISON;
2943 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2944 goto out_release;
2945 }
2946
2947 locked = lock_page_or_retry(page, mm, flags);
2948
2949 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2950 if (!locked) {
2951 ret |= VM_FAULT_RETRY;
2952 goto out_release;
2953 }
2954
2955 /*
2956 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2957 * release the swapcache from under us. The page pin, and pte_same
2958 * test below, are not enough to exclude that. Even if it is still
2959 * swapcache, we need to check that the page's swap has not changed.
2960 */
2961 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2962 goto out_page;
2963
2964 if (ksm_might_need_to_copy(page, vma, address)) {
2965 swapcache = page;
2966 page = ksm_does_need_to_copy(page, vma, address);
2967
2968 if (unlikely(!page)) {
2969 ret = VM_FAULT_OOM;
2970 page = swapcache;
2971 swapcache = NULL;
2972 goto out_page;
2973 }
2974 }
2975
2976 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2977 ret = VM_FAULT_OOM;
2978 goto out_page;
2979 }
2980
2981 /*
2982 * Back out if somebody else already faulted in this pte.
2983 */
2984 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2985 if (unlikely(!pte_same(*page_table, orig_pte)))
2986 goto out_nomap;
2987
2988 if (unlikely(!PageUptodate(page))) {
2989 ret = VM_FAULT_SIGBUS;
2990 goto out_nomap;
2991 }
2992
2993 /*
2994 * The page isn't present yet, go ahead with the fault.
2995 *
2996 * Be careful about the sequence of operations here.
2997 * To get its accounting right, reuse_swap_page() must be called
2998 * while the page is counted on swap but not yet in mapcount i.e.
2999 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3000 * must be called after the swap_free(), or it will never succeed.
3001 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3002 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3003 * in page->private. In this case, a record in swap_cgroup is silently
3004 * discarded at swap_free().
3005 */
3006
3007 inc_mm_counter_fast(mm, MM_ANONPAGES);
3008 dec_mm_counter_fast(mm, MM_SWAPENTS);
3009 pte = mk_pte(page, vma->vm_page_prot);
3010 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3011 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3012 flags &= ~FAULT_FLAG_WRITE;
3013 ret |= VM_FAULT_WRITE;
3014 exclusive = 1;
3015 }
3016 flush_icache_page(vma, page);
3017 set_pte_at(mm, address, page_table, pte);
3018 do_page_add_anon_rmap(page, vma, address, exclusive);
3019 /* It's better to call commit-charge after rmap is established */
3020 mem_cgroup_commit_charge_swapin(page, ptr);
3021
3022 swap_free(entry);
3023 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3024 try_to_free_swap(page);
3025 unlock_page(page);
3026 if (swapcache) {
3027 /*
3028 * Hold the lock to avoid the swap entry to be reused
3029 * until we take the PT lock for the pte_same() check
3030 * (to avoid false positives from pte_same). For
3031 * further safety release the lock after the swap_free
3032 * so that the swap count won't change under a
3033 * parallel locked swapcache.
3034 */
3035 unlock_page(swapcache);
3036 page_cache_release(swapcache);
3037 }
3038
3039 if (flags & FAULT_FLAG_WRITE) {
3040 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3041 if (ret & VM_FAULT_ERROR)
3042 ret &= VM_FAULT_ERROR;
3043 goto out;
3044 }
3045
3046 /* No need to invalidate - it was non-present before */
3047 update_mmu_cache(vma, address, page_table);
3048unlock:
3049 pte_unmap_unlock(page_table, ptl);
3050out:
3051 return ret;
3052out_nomap:
3053 mem_cgroup_cancel_charge_swapin(ptr);
3054 pte_unmap_unlock(page_table, ptl);
3055out_page:
3056 unlock_page(page);
3057out_release:
3058 page_cache_release(page);
3059 if (swapcache) {
3060 unlock_page(swapcache);
3061 page_cache_release(swapcache);
3062 }
3063 return ret;
3064}
3065
3066/*
3067 * This is like a special single-page "expand_{down|up}wards()",
3068 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3069 * doesn't hit another vma.
3070 */
3071static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3072{
3073 address &= PAGE_MASK;
3074 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3075 struct vm_area_struct *prev = vma->vm_prev;
3076
3077 /*
3078 * Is there a mapping abutting this one below?
3079 *
3080 * That's only ok if it's the same stack mapping
3081 * that has gotten split..
3082 */
3083 if (prev && prev->vm_end == address)
3084 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3085
3086 expand_downwards(vma, address - PAGE_SIZE);
3087 }
3088 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3089 struct vm_area_struct *next = vma->vm_next;
3090
3091 /* As VM_GROWSDOWN but s/below/above/ */
3092 if (next && next->vm_start == address + PAGE_SIZE)
3093 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3094
3095 expand_upwards(vma, address + PAGE_SIZE);
3096 }
3097 return 0;
3098}
3099
3100/*
3101 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3102 * but allow concurrent faults), and pte mapped but not yet locked.
3103 * We return with mmap_sem still held, but pte unmapped and unlocked.
3104 */
3105static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3106 unsigned long address, pte_t *page_table, pmd_t *pmd,
3107 unsigned int flags)
3108{
3109 struct page *page;
3110 spinlock_t *ptl;
3111 pte_t entry;
3112
3113 pte_unmap(page_table);
3114
3115 /* Check if we need to add a guard page to the stack */
3116 if (check_stack_guard_page(vma, address) < 0)
3117 return VM_FAULT_SIGBUS;
3118
3119 /* Use the zero-page for reads */
3120 if (!(flags & FAULT_FLAG_WRITE)) {
3121 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3122 vma->vm_page_prot));
3123 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3124 if (!pte_none(*page_table))
3125 goto unlock;
3126 goto setpte;
3127 }
3128
3129 /* Allocate our own private page. */
3130 if (unlikely(anon_vma_prepare(vma)))
3131 goto oom;
3132 page = alloc_zeroed_user_highpage_movable(vma, address);
3133 if (!page)
3134 goto oom;
3135 __SetPageUptodate(page);
3136
3137 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3138 goto oom_free_page;
3139
3140 entry = mk_pte(page, vma->vm_page_prot);
3141 if (vma->vm_flags & VM_WRITE)
3142 entry = pte_mkwrite(pte_mkdirty(entry));
3143
3144 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3145 if (!pte_none(*page_table))
3146 goto release;
3147
3148 inc_mm_counter_fast(mm, MM_ANONPAGES);
3149 page_add_new_anon_rmap(page, vma, address);
3150setpte:
3151 set_pte_at(mm, address, page_table, entry);
3152
3153 /* No need to invalidate - it was non-present before */
3154 update_mmu_cache(vma, address, page_table);
3155unlock:
3156 pte_unmap_unlock(page_table, ptl);
3157 return 0;
3158release:
3159 mem_cgroup_uncharge_page(page);
3160 page_cache_release(page);
3161 goto unlock;
3162oom_free_page:
3163 page_cache_release(page);
3164oom:
3165 return VM_FAULT_OOM;
3166}
3167
3168/*
3169 * __do_fault() tries to create a new page mapping. It aggressively
3170 * tries to share with existing pages, but makes a separate copy if
3171 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3172 * the next page fault.
3173 *
3174 * As this is called only for pages that do not currently exist, we
3175 * do not need to flush old virtual caches or the TLB.
3176 *
3177 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3178 * but allow concurrent faults), and pte neither mapped nor locked.
3179 * We return with mmap_sem still held, but pte unmapped and unlocked.
3180 */
3181static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3182 unsigned long address, pmd_t *pmd,
3183 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3184{
3185 pte_t *page_table;
3186 spinlock_t *ptl;
3187 struct page *page;
3188 struct page *cow_page;
3189 pte_t entry;
3190 int anon = 0;
3191 struct page *dirty_page = NULL;
3192 struct vm_fault vmf;
3193 int ret;
3194 int page_mkwrite = 0;
3195
3196 /*
3197 * If we do COW later, allocate page befor taking lock_page()
3198 * on the file cache page. This will reduce lock holding time.
3199 */
3200 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3201
3202 if (unlikely(anon_vma_prepare(vma)))
3203 return VM_FAULT_OOM;
3204
3205 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3206 if (!cow_page)
3207 return VM_FAULT_OOM;
3208
3209 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3210 page_cache_release(cow_page);
3211 return VM_FAULT_OOM;
3212 }
3213 } else
3214 cow_page = NULL;
3215
3216 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3217 vmf.pgoff = pgoff;
3218 vmf.flags = flags;
3219 vmf.page = NULL;
3220
3221 ret = vma->vm_ops->fault(vma, &vmf);
3222 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3223 VM_FAULT_RETRY)))
3224 goto uncharge_out;
3225
3226 if (unlikely(PageHWPoison(vmf.page))) {
3227 if (ret & VM_FAULT_LOCKED)
3228 unlock_page(vmf.page);
3229 ret = VM_FAULT_HWPOISON;
3230 goto uncharge_out;
3231 }
3232
3233 /*
3234 * For consistency in subsequent calls, make the faulted page always
3235 * locked.
3236 */
3237 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3238 lock_page(vmf.page);
3239 else
3240 VM_BUG_ON(!PageLocked(vmf.page));
3241
3242 /*
3243 * Should we do an early C-O-W break?
3244 */
3245 page = vmf.page;
3246 if (flags & FAULT_FLAG_WRITE) {
3247 if (!(vma->vm_flags & VM_SHARED)) {
3248 page = cow_page;
3249 anon = 1;
3250 copy_user_highpage(page, vmf.page, address, vma);
3251 __SetPageUptodate(page);
3252 } else {
3253 /*
3254 * If the page will be shareable, see if the backing
3255 * address space wants to know that the page is about
3256 * to become writable
3257 */
3258 if (vma->vm_ops->page_mkwrite) {
3259 int tmp;
3260
3261 unlock_page(page);
3262 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3263 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3264 if (unlikely(tmp &
3265 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3266 ret = tmp;
3267 goto unwritable_page;
3268 }
3269 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3270 lock_page(page);
3271 if (!page->mapping) {
3272 ret = 0; /* retry the fault */
3273 unlock_page(page);
3274 goto unwritable_page;
3275 }
3276 } else
3277 VM_BUG_ON(!PageLocked(page));
3278 page_mkwrite = 1;
3279 }
3280 }
3281
3282 }
3283
3284 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3285
3286 /*
3287 * This silly early PAGE_DIRTY setting removes a race
3288 * due to the bad i386 page protection. But it's valid
3289 * for other architectures too.
3290 *
3291 * Note that if FAULT_FLAG_WRITE is set, we either now have
3292 * an exclusive copy of the page, or this is a shared mapping,
3293 * so we can make it writable and dirty to avoid having to
3294 * handle that later.
3295 */
3296 /* Only go through if we didn't race with anybody else... */
3297 if (likely(pte_same(*page_table, orig_pte))) {
3298 flush_icache_page(vma, page);
3299 entry = mk_pte(page, vma->vm_page_prot);
3300 if (flags & FAULT_FLAG_WRITE)
3301 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3302 if (anon) {
3303 inc_mm_counter_fast(mm, MM_ANONPAGES);
3304 page_add_new_anon_rmap(page, vma, address);
3305 } else {
3306 inc_mm_counter_fast(mm, MM_FILEPAGES);
3307 page_add_file_rmap(page);
3308 if (flags & FAULT_FLAG_WRITE) {
3309 dirty_page = page;
3310 get_page(dirty_page);
3311 }
3312 }
3313 set_pte_at(mm, address, page_table, entry);
3314
3315 /* no need to invalidate: a not-present page won't be cached */
3316 update_mmu_cache(vma, address, page_table);
3317 } else {
3318 if (cow_page)
3319 mem_cgroup_uncharge_page(cow_page);
3320 if (anon)
3321 page_cache_release(page);
3322 else
3323 anon = 1; /* no anon but release faulted_page */
3324 }
3325
3326 pte_unmap_unlock(page_table, ptl);
3327
3328 if (dirty_page) {
3329 struct address_space *mapping = page->mapping;
3330
3331 if (set_page_dirty(dirty_page))
3332 page_mkwrite = 1;
3333 unlock_page(dirty_page);
3334 put_page(dirty_page);
3335 if (page_mkwrite && mapping) {
3336 /*
3337 * Some device drivers do not set page.mapping but still
3338 * dirty their pages
3339 */
3340 balance_dirty_pages_ratelimited(mapping);
3341 }
3342
3343 /* file_update_time outside page_lock */
3344 if (vma->vm_file)
3345 file_update_time(vma->vm_file);
3346 } else {
3347 unlock_page(vmf.page);
3348 if (anon)
3349 page_cache_release(vmf.page);
3350 }
3351
3352 return ret;
3353
3354unwritable_page:
3355 page_cache_release(page);
3356 return ret;
3357uncharge_out:
3358 /* fs's fault handler get error */
3359 if (cow_page) {
3360 mem_cgroup_uncharge_page(cow_page);
3361 page_cache_release(cow_page);
3362 }
3363 return ret;
3364}
3365
3366static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3367 unsigned long address, pte_t *page_table, pmd_t *pmd,
3368 unsigned int flags, pte_t orig_pte)
3369{
3370 pgoff_t pgoff = (((address & PAGE_MASK)
3371 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3372
3373 pte_unmap(page_table);
3374 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3375}
3376
3377/*
3378 * Fault of a previously existing named mapping. Repopulate the pte
3379 * from the encoded file_pte if possible. This enables swappable
3380 * nonlinear vmas.
3381 *
3382 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3383 * but allow concurrent faults), and pte mapped but not yet locked.
3384 * We return with mmap_sem still held, but pte unmapped and unlocked.
3385 */
3386static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3387 unsigned long address, pte_t *page_table, pmd_t *pmd,
3388 unsigned int flags, pte_t orig_pte)
3389{
3390 pgoff_t pgoff;
3391
3392 flags |= FAULT_FLAG_NONLINEAR;
3393
3394 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3395 return 0;
3396
3397 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3398 /*
3399 * Page table corrupted: show pte and kill process.
3400 */
3401 print_bad_pte(vma, address, orig_pte, NULL);
3402 return VM_FAULT_SIGBUS;
3403 }
3404
3405 pgoff = pte_to_pgoff(orig_pte);
3406 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3407}
3408
3409/*
3410 * These routines also need to handle stuff like marking pages dirty
3411 * and/or accessed for architectures that don't do it in hardware (most
3412 * RISC architectures). The early dirtying is also good on the i386.
3413 *
3414 * There is also a hook called "update_mmu_cache()" that architectures
3415 * with external mmu caches can use to update those (ie the Sparc or
3416 * PowerPC hashed page tables that act as extended TLBs).
3417 *
3418 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3419 * but allow concurrent faults), and pte mapped but not yet locked.
3420 * We return with mmap_sem still held, but pte unmapped and unlocked.
3421 */
3422int handle_pte_fault(struct mm_struct *mm,
3423 struct vm_area_struct *vma, unsigned long address,
3424 pte_t *pte, pmd_t *pmd, unsigned int flags)
3425{
3426 pte_t entry;
3427 spinlock_t *ptl;
3428
3429 entry = *pte;
3430 if (!pte_present(entry)) {
3431 if (pte_none(entry)) {
3432 if (vma->vm_ops) {
3433 if (likely(vma->vm_ops->fault))
3434 return do_linear_fault(mm, vma, address,
3435 pte, pmd, flags, entry);
3436 }
3437 return do_anonymous_page(mm, vma, address,
3438 pte, pmd, flags);
3439 }
3440 if (pte_file(entry))
3441 return do_nonlinear_fault(mm, vma, address,
3442 pte, pmd, flags, entry);
3443 return do_swap_page(mm, vma, address,
3444 pte, pmd, flags, entry);
3445 }
3446
3447 ptl = pte_lockptr(mm, pmd);
3448 spin_lock(ptl);
3449 if (unlikely(!pte_same(*pte, entry)))
3450 goto unlock;
3451 if (flags & FAULT_FLAG_WRITE) {
3452 if (!pte_write(entry))
3453 return do_wp_page(mm, vma, address,
3454 pte, pmd, ptl, entry);
3455 entry = pte_mkdirty(entry);
3456 }
3457 entry = pte_mkyoung(entry);
3458 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3459 update_mmu_cache(vma, address, pte);
3460 } else {
3461 /*
3462 * This is needed only for protection faults but the arch code
3463 * is not yet telling us if this is a protection fault or not.
3464 * This still avoids useless tlb flushes for .text page faults
3465 * with threads.
3466 */
3467 if (flags & FAULT_FLAG_WRITE)
3468 flush_tlb_fix_spurious_fault(vma, address);
3469 }
3470unlock:
3471 pte_unmap_unlock(pte, ptl);
3472 return 0;
3473}
3474
3475/*
3476 * By the time we get here, we already hold the mm semaphore
3477 */
3478int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3479 unsigned long address, unsigned int flags)
3480{
3481 pgd_t *pgd;
3482 pud_t *pud;
3483 pmd_t *pmd;
3484 pte_t *pte;
3485
3486 __set_current_state(TASK_RUNNING);
3487
3488 count_vm_event(PGFAULT);
3489 mem_cgroup_count_vm_event(mm, PGFAULT);
3490
3491 /* do counter updates before entering really critical section. */
3492 check_sync_rss_stat(current);
3493
3494 if (unlikely(is_vm_hugetlb_page(vma)))
3495 return hugetlb_fault(mm, vma, address, flags);
3496
3497retry:
3498 pgd = pgd_offset(mm, address);
3499 pud = pud_alloc(mm, pgd, address);
3500 if (!pud)
3501 return VM_FAULT_OOM;
3502 pmd = pmd_alloc(mm, pud, address);
3503 if (!pmd)
3504 return VM_FAULT_OOM;
3505 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3506 if (!vma->vm_ops)
3507 return do_huge_pmd_anonymous_page(mm, vma, address,
3508 pmd, flags);
3509 } else {
3510 pmd_t orig_pmd = *pmd;
3511 int ret;
3512
3513 barrier();
3514 if (pmd_trans_huge(orig_pmd)) {
3515 if (flags & FAULT_FLAG_WRITE &&
3516 !pmd_write(orig_pmd) &&
3517 !pmd_trans_splitting(orig_pmd)) {
3518 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3519 orig_pmd);
3520 /*
3521 * If COW results in an oom, the huge pmd will
3522 * have been split, so retry the fault on the
3523 * pte for a smaller charge.
3524 */
3525 if (unlikely(ret & VM_FAULT_OOM))
3526 goto retry;
3527 return ret;
3528 }
3529 return 0;
3530 }
3531 }
3532
3533 /*
3534 * Use __pte_alloc instead of pte_alloc_map, because we can't
3535 * run pte_offset_map on the pmd, if an huge pmd could
3536 * materialize from under us from a different thread.
3537 */
3538 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3539 return VM_FAULT_OOM;
3540 /* if an huge pmd materialized from under us just retry later */
3541 if (unlikely(pmd_trans_huge(*pmd)))
3542 return 0;
3543 /*
3544 * A regular pmd is established and it can't morph into a huge pmd
3545 * from under us anymore at this point because we hold the mmap_sem
3546 * read mode and khugepaged takes it in write mode. So now it's
3547 * safe to run pte_offset_map().
3548 */
3549 pte = pte_offset_map(pmd, address);
3550
3551 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3552}
3553
3554#ifndef __PAGETABLE_PUD_FOLDED
3555/*
3556 * Allocate page upper directory.
3557 * We've already handled the fast-path in-line.
3558 */
3559int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3560{
3561 pud_t *new = pud_alloc_one(mm, address);
3562 if (!new)
3563 return -ENOMEM;
3564
3565 smp_wmb(); /* See comment in __pte_alloc */
3566
3567 spin_lock(&mm->page_table_lock);
3568 if (pgd_present(*pgd)) /* Another has populated it */
3569 pud_free(mm, new);
3570 else
3571 pgd_populate(mm, pgd, new);
3572 spin_unlock(&mm->page_table_lock);
3573 return 0;
3574}
3575#endif /* __PAGETABLE_PUD_FOLDED */
3576
3577#ifndef __PAGETABLE_PMD_FOLDED
3578/*
3579 * Allocate page middle directory.
3580 * We've already handled the fast-path in-line.
3581 */
3582int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3583{
3584 pmd_t *new = pmd_alloc_one(mm, address);
3585 if (!new)
3586 return -ENOMEM;
3587
3588 smp_wmb(); /* See comment in __pte_alloc */
3589
3590 spin_lock(&mm->page_table_lock);
3591#ifndef __ARCH_HAS_4LEVEL_HACK
3592 if (pud_present(*pud)) /* Another has populated it */
3593 pmd_free(mm, new);
3594 else
3595 pud_populate(mm, pud, new);
3596#else
3597 if (pgd_present(*pud)) /* Another has populated it */
3598 pmd_free(mm, new);
3599 else
3600 pgd_populate(mm, pud, new);
3601#endif /* __ARCH_HAS_4LEVEL_HACK */
3602 spin_unlock(&mm->page_table_lock);
3603 return 0;
3604}
3605#endif /* __PAGETABLE_PMD_FOLDED */
3606
3607int make_pages_present(unsigned long addr, unsigned long end)
3608{
3609 int ret, len, write;
3610 struct vm_area_struct * vma;
3611
3612 vma = find_vma(current->mm, addr);
3613 if (!vma)
3614 return -ENOMEM;
3615 /*
3616 * We want to touch writable mappings with a write fault in order
3617 * to break COW, except for shared mappings because these don't COW
3618 * and we would not want to dirty them for nothing.
3619 */
3620 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3621 BUG_ON(addr >= end);
3622 BUG_ON(end > vma->vm_end);
3623 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3624 ret = get_user_pages(current, current->mm, addr,
3625 len, write, 0, NULL, NULL);
3626 if (ret < 0)
3627 return ret;
3628 return ret == len ? 0 : -EFAULT;
3629}
3630
3631#if !defined(__HAVE_ARCH_GATE_AREA)
3632
3633#if defined(AT_SYSINFO_EHDR)
3634static struct vm_area_struct gate_vma;
3635
3636static int __init gate_vma_init(void)
3637{
3638 gate_vma.vm_mm = NULL;
3639 gate_vma.vm_start = FIXADDR_USER_START;
3640 gate_vma.vm_end = FIXADDR_USER_END;
3641 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3642 gate_vma.vm_page_prot = __P101;
3643
3644 return 0;
3645}
3646__initcall(gate_vma_init);
3647#endif
3648
3649struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3650{
3651#ifdef AT_SYSINFO_EHDR
3652 return &gate_vma;
3653#else
3654 return NULL;
3655#endif
3656}
3657
3658int in_gate_area_no_mm(unsigned long addr)
3659{
3660#ifdef AT_SYSINFO_EHDR
3661 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3662 return 1;
3663#endif
3664 return 0;
3665}
3666
3667#endif /* __HAVE_ARCH_GATE_AREA */
3668
3669static int __follow_pte(struct mm_struct *mm, unsigned long address,
3670 pte_t **ptepp, spinlock_t **ptlp)
3671{
3672 pgd_t *pgd;
3673 pud_t *pud;
3674 pmd_t *pmd;
3675 pte_t *ptep;
3676
3677 pgd = pgd_offset(mm, address);
3678 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3679 goto out;
3680
3681 pud = pud_offset(pgd, address);
3682 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3683 goto out;
3684
3685 pmd = pmd_offset(pud, address);
3686 VM_BUG_ON(pmd_trans_huge(*pmd));
3687 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3688 goto out;
3689
3690 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3691 if (pmd_huge(*pmd))
3692 goto out;
3693
3694 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3695 if (!ptep)
3696 goto out;
3697 if (!pte_present(*ptep))
3698 goto unlock;
3699 *ptepp = ptep;
3700 return 0;
3701unlock:
3702 pte_unmap_unlock(ptep, *ptlp);
3703out:
3704 return -EINVAL;
3705}
3706
3707static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3708 pte_t **ptepp, spinlock_t **ptlp)
3709{
3710 int res;
3711
3712 /* (void) is needed to make gcc happy */
3713 (void) __cond_lock(*ptlp,
3714 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3715 return res;
3716}
3717
3718/**
3719 * follow_pfn - look up PFN at a user virtual address
3720 * @vma: memory mapping
3721 * @address: user virtual address
3722 * @pfn: location to store found PFN
3723 *
3724 * Only IO mappings and raw PFN mappings are allowed.
3725 *
3726 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3727 */
3728int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3729 unsigned long *pfn)
3730{
3731 int ret = -EINVAL;
3732 spinlock_t *ptl;
3733 pte_t *ptep;
3734
3735 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3736 return ret;
3737
3738 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3739 if (ret)
3740 return ret;
3741 *pfn = pte_pfn(*ptep);
3742 pte_unmap_unlock(ptep, ptl);
3743 return 0;
3744}
3745EXPORT_SYMBOL(follow_pfn);
3746
3747#ifdef CONFIG_HAVE_IOREMAP_PROT
3748int follow_phys(struct vm_area_struct *vma,
3749 unsigned long address, unsigned int flags,
3750 unsigned long *prot, resource_size_t *phys)
3751{
3752 int ret = -EINVAL;
3753 pte_t *ptep, pte;
3754 spinlock_t *ptl;
3755
3756 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3757 goto out;
3758
3759 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3760 goto out;
3761 pte = *ptep;
3762
3763 if ((flags & FOLL_WRITE) && !pte_write(pte))
3764 goto unlock;
3765
3766 *prot = pgprot_val(pte_pgprot(pte));
3767 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3768
3769 ret = 0;
3770unlock:
3771 pte_unmap_unlock(ptep, ptl);
3772out:
3773 return ret;
3774}
3775
3776int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3777 void *buf, int len, int write)
3778{
3779 resource_size_t phys_addr;
3780 unsigned long prot = 0;
3781 void __iomem *maddr;
3782 int offset = addr & (PAGE_SIZE-1);
3783
3784 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3785 return -EINVAL;
3786
3787 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3788 if (write)
3789 memcpy_toio(maddr + offset, buf, len);
3790 else
3791 memcpy_fromio(buf, maddr + offset, len);
3792 iounmap(maddr);
3793
3794 return len;
3795}
3796#endif
3797
3798/*
3799 * Access another process' address space as given in mm. If non-NULL, use the
3800 * given task for page fault accounting.
3801 */
3802static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3803 unsigned long addr, void *buf, int len, int write)
3804{
3805 struct vm_area_struct *vma;
3806 void *old_buf = buf;
3807
3808 down_read(&mm->mmap_sem);
3809 /* ignore errors, just check how much was successfully transferred */
3810 while (len) {
3811 int bytes, ret, offset;
3812 void *maddr;
3813 struct page *page = NULL;
3814
3815 ret = get_user_pages(tsk, mm, addr, 1,
3816 write, 1, &page, &vma);
3817 if (ret <= 0) {
3818 /*
3819 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3820 * we can access using slightly different code.
3821 */
3822#ifdef CONFIG_HAVE_IOREMAP_PROT
3823 vma = find_vma(mm, addr);
3824 if (!vma || vma->vm_start > addr)
3825 break;
3826 if (vma->vm_ops && vma->vm_ops->access)
3827 ret = vma->vm_ops->access(vma, addr, buf,
3828 len, write);
3829 if (ret <= 0)
3830#endif
3831 break;
3832 bytes = ret;
3833 } else {
3834 bytes = len;
3835 offset = addr & (PAGE_SIZE-1);
3836 if (bytes > PAGE_SIZE-offset)
3837 bytes = PAGE_SIZE-offset;
3838
3839 maddr = kmap(page);
3840 if (write) {
3841 copy_to_user_page(vma, page, addr,
3842 maddr + offset, buf, bytes);
3843 set_page_dirty_lock(page);
3844 } else {
3845 copy_from_user_page(vma, page, addr,
3846 buf, maddr + offset, bytes);
3847 }
3848 kunmap(page);
3849 page_cache_release(page);
3850 }
3851 len -= bytes;
3852 buf += bytes;
3853 addr += bytes;
3854 }
3855 up_read(&mm->mmap_sem);
3856
3857 return buf - old_buf;
3858}
3859
3860/**
3861 * access_remote_vm - access another process' address space
3862 * @mm: the mm_struct of the target address space
3863 * @addr: start address to access
3864 * @buf: source or destination buffer
3865 * @len: number of bytes to transfer
3866 * @write: whether the access is a write
3867 *
3868 * The caller must hold a reference on @mm.
3869 */
3870int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3871 void *buf, int len, int write)
3872{
3873 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3874}
3875
3876/*
3877 * Access another process' address space.
3878 * Source/target buffer must be kernel space,
3879 * Do not walk the page table directly, use get_user_pages
3880 */
3881int access_process_vm(struct task_struct *tsk, unsigned long addr,
3882 void *buf, int len, int write)
3883{
3884 struct mm_struct *mm;
3885 int ret;
3886
3887 mm = get_task_mm(tsk);
3888 if (!mm)
3889 return 0;
3890
3891 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3892 mmput(mm);
3893
3894 return ret;
3895}
3896
3897/*
3898 * Print the name of a VMA.
3899 */
3900void print_vma_addr(char *prefix, unsigned long ip)
3901{
3902 struct mm_struct *mm = current->mm;
3903 struct vm_area_struct *vma;
3904
3905 /*
3906 * Do not print if we are in atomic
3907 * contexts (in exception stacks, etc.):
3908 */
3909 if (preempt_count())
3910 return;
3911
3912 down_read(&mm->mmap_sem);
3913 vma = find_vma(mm, ip);
3914 if (vma && vma->vm_file) {
3915 struct file *f = vma->vm_file;
3916 char *buf = (char *)__get_free_page(GFP_KERNEL);
3917 if (buf) {
3918 char *p, *s;
3919
3920 p = d_path(&f->f_path, buf, PAGE_SIZE);
3921 if (IS_ERR(p))
3922 p = "?";
3923 s = strrchr(p, '/');
3924 if (s)
3925 p = s+1;
3926 printk("%s%s[%lx+%lx]", prefix, p,
3927 vma->vm_start,
3928 vma->vm_end - vma->vm_start);
3929 free_page((unsigned long)buf);
3930 }
3931 }
3932 up_read(¤t->mm->mmap_sem);
3933}
3934
3935#ifdef CONFIG_PROVE_LOCKING
3936void might_fault(void)
3937{
3938 /*
3939 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3940 * holding the mmap_sem, this is safe because kernel memory doesn't
3941 * get paged out, therefore we'll never actually fault, and the
3942 * below annotations will generate false positives.
3943 */
3944 if (segment_eq(get_fs(), KERNEL_DS))
3945 return;
3946
3947 might_sleep();
3948 /*
3949 * it would be nicer only to annotate paths which are not under
3950 * pagefault_disable, however that requires a larger audit and
3951 * providing helpers like get_user_atomic.
3952 */
3953 if (!in_atomic() && current->mm)
3954 might_lock_read(¤t->mm->mmap_sem);
3955}
3956EXPORT_SYMBOL(might_fault);
3957#endif
3958
3959#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3960static void clear_gigantic_page(struct page *page,
3961 unsigned long addr,
3962 unsigned int pages_per_huge_page)
3963{
3964 int i;
3965 struct page *p = page;
3966
3967 might_sleep();
3968 for (i = 0; i < pages_per_huge_page;
3969 i++, p = mem_map_next(p, page, i)) {
3970 cond_resched();
3971 clear_user_highpage(p, addr + i * PAGE_SIZE);
3972 }
3973}
3974void clear_huge_page(struct page *page,
3975 unsigned long addr, unsigned int pages_per_huge_page)
3976{
3977 int i;
3978
3979 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3980 clear_gigantic_page(page, addr, pages_per_huge_page);
3981 return;
3982 }
3983
3984 might_sleep();
3985 for (i = 0; i < pages_per_huge_page; i++) {
3986 cond_resched();
3987 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3988 }
3989}
3990
3991static void copy_user_gigantic_page(struct page *dst, struct page *src,
3992 unsigned long addr,
3993 struct vm_area_struct *vma,
3994 unsigned int pages_per_huge_page)
3995{
3996 int i;
3997 struct page *dst_base = dst;
3998 struct page *src_base = src;
3999
4000 for (i = 0; i < pages_per_huge_page; ) {
4001 cond_resched();
4002 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4003
4004 i++;
4005 dst = mem_map_next(dst, dst_base, i);
4006 src = mem_map_next(src, src_base, i);
4007 }
4008}
4009
4010void copy_user_huge_page(struct page *dst, struct page *src,
4011 unsigned long addr, struct vm_area_struct *vma,
4012 unsigned int pages_per_huge_page)
4013{
4014 int i;
4015
4016 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4017 copy_user_gigantic_page(dst, src, addr, vma,
4018 pages_per_huge_page);
4019 return;
4020 }
4021
4022 might_sleep();
4023 for (i = 0; i < pages_per_huge_page; i++) {
4024 cond_resched();
4025 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4026 }
4027}
4028#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */