Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/memory.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 */
7
8/*
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
11 */
12
13/*
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
16 *
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
19 * far as I could see.
20 *
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
22 */
23
24/*
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
30 */
31
32/*
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 *
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
38 *
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
40 */
41
42#include <linux/kernel_stat.h>
43#include <linux/mm.h>
44#include <linux/sched/mm.h>
45#include <linux/sched/coredump.h>
46#include <linux/sched/numa_balancing.h>
47#include <linux/sched/task.h>
48#include <linux/hugetlb.h>
49#include <linux/mman.h>
50#include <linux/swap.h>
51#include <linux/highmem.h>
52#include <linux/pagemap.h>
53#include <linux/memremap.h>
54#include <linux/ksm.h>
55#include <linux/rmap.h>
56#include <linux/export.h>
57#include <linux/delayacct.h>
58#include <linux/init.h>
59#include <linux/pfn_t.h>
60#include <linux/writeback.h>
61#include <linux/memcontrol.h>
62#include <linux/mmu_notifier.h>
63#include <linux/swapops.h>
64#include <linux/elf.h>
65#include <linux/gfp.h>
66#include <linux/migrate.h>
67#include <linux/string.h>
68#include <linux/dma-debug.h>
69#include <linux/debugfs.h>
70#include <linux/userfaultfd_k.h>
71#include <linux/dax.h>
72#include <linux/oom.h>
73#include <linux/numa.h>
74
75#include <asm/io.h>
76#include <asm/mmu_context.h>
77#include <asm/pgalloc.h>
78#include <linux/uaccess.h>
79#include <asm/tlb.h>
80#include <asm/tlbflush.h>
81#include <asm/pgtable.h>
82
83#include "internal.h"
84
85#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
86#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
87#endif
88
89#ifndef CONFIG_NEED_MULTIPLE_NODES
90/* use the per-pgdat data instead for discontigmem - mbligh */
91unsigned long max_mapnr;
92EXPORT_SYMBOL(max_mapnr);
93
94struct page *mem_map;
95EXPORT_SYMBOL(mem_map);
96#endif
97
98/*
99 * A number of key systems in x86 including ioremap() rely on the assumption
100 * that high_memory defines the upper bound on direct map memory, then end
101 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
102 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
103 * and ZONE_HIGHMEM.
104 */
105void *high_memory;
106EXPORT_SYMBOL(high_memory);
107
108/*
109 * Randomize the address space (stacks, mmaps, brk, etc.).
110 *
111 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
112 * as ancient (libc5 based) binaries can segfault. )
113 */
114int randomize_va_space __read_mostly =
115#ifdef CONFIG_COMPAT_BRK
116 1;
117#else
118 2;
119#endif
120
121static int __init disable_randmaps(char *s)
122{
123 randomize_va_space = 0;
124 return 1;
125}
126__setup("norandmaps", disable_randmaps);
127
128unsigned long zero_pfn __read_mostly;
129EXPORT_SYMBOL(zero_pfn);
130
131unsigned long highest_memmap_pfn __read_mostly;
132
133/*
134 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
135 */
136static int __init init_zero_pfn(void)
137{
138 zero_pfn = page_to_pfn(ZERO_PAGE(0));
139 return 0;
140}
141core_initcall(init_zero_pfn);
142
143
144#if defined(SPLIT_RSS_COUNTING)
145
146void sync_mm_rss(struct mm_struct *mm)
147{
148 int i;
149
150 for (i = 0; i < NR_MM_COUNTERS; i++) {
151 if (current->rss_stat.count[i]) {
152 add_mm_counter(mm, i, current->rss_stat.count[i]);
153 current->rss_stat.count[i] = 0;
154 }
155 }
156 current->rss_stat.events = 0;
157}
158
159static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
160{
161 struct task_struct *task = current;
162
163 if (likely(task->mm == mm))
164 task->rss_stat.count[member] += val;
165 else
166 add_mm_counter(mm, member, val);
167}
168#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
169#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
170
171/* sync counter once per 64 page faults */
172#define TASK_RSS_EVENTS_THRESH (64)
173static void check_sync_rss_stat(struct task_struct *task)
174{
175 if (unlikely(task != current))
176 return;
177 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
178 sync_mm_rss(task->mm);
179}
180#else /* SPLIT_RSS_COUNTING */
181
182#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
183#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
184
185static void check_sync_rss_stat(struct task_struct *task)
186{
187}
188
189#endif /* SPLIT_RSS_COUNTING */
190
191/*
192 * Note: this doesn't free the actual pages themselves. That
193 * has been handled earlier when unmapping all the memory regions.
194 */
195static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
196 unsigned long addr)
197{
198 pgtable_t token = pmd_pgtable(*pmd);
199 pmd_clear(pmd);
200 pte_free_tlb(tlb, token, addr);
201 mm_dec_nr_ptes(tlb->mm);
202}
203
204static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
205 unsigned long addr, unsigned long end,
206 unsigned long floor, unsigned long ceiling)
207{
208 pmd_t *pmd;
209 unsigned long next;
210 unsigned long start;
211
212 start = addr;
213 pmd = pmd_offset(pud, addr);
214 do {
215 next = pmd_addr_end(addr, end);
216 if (pmd_none_or_clear_bad(pmd))
217 continue;
218 free_pte_range(tlb, pmd, addr);
219 } while (pmd++, addr = next, addr != end);
220
221 start &= PUD_MASK;
222 if (start < floor)
223 return;
224 if (ceiling) {
225 ceiling &= PUD_MASK;
226 if (!ceiling)
227 return;
228 }
229 if (end - 1 > ceiling - 1)
230 return;
231
232 pmd = pmd_offset(pud, start);
233 pud_clear(pud);
234 pmd_free_tlb(tlb, pmd, start);
235 mm_dec_nr_pmds(tlb->mm);
236}
237
238static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
239 unsigned long addr, unsigned long end,
240 unsigned long floor, unsigned long ceiling)
241{
242 pud_t *pud;
243 unsigned long next;
244 unsigned long start;
245
246 start = addr;
247 pud = pud_offset(p4d, addr);
248 do {
249 next = pud_addr_end(addr, end);
250 if (pud_none_or_clear_bad(pud))
251 continue;
252 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
253 } while (pud++, addr = next, addr != end);
254
255 start &= P4D_MASK;
256 if (start < floor)
257 return;
258 if (ceiling) {
259 ceiling &= P4D_MASK;
260 if (!ceiling)
261 return;
262 }
263 if (end - 1 > ceiling - 1)
264 return;
265
266 pud = pud_offset(p4d, start);
267 p4d_clear(p4d);
268 pud_free_tlb(tlb, pud, start);
269 mm_dec_nr_puds(tlb->mm);
270}
271
272static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
273 unsigned long addr, unsigned long end,
274 unsigned long floor, unsigned long ceiling)
275{
276 p4d_t *p4d;
277 unsigned long next;
278 unsigned long start;
279
280 start = addr;
281 p4d = p4d_offset(pgd, addr);
282 do {
283 next = p4d_addr_end(addr, end);
284 if (p4d_none_or_clear_bad(p4d))
285 continue;
286 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
287 } while (p4d++, addr = next, addr != end);
288
289 start &= PGDIR_MASK;
290 if (start < floor)
291 return;
292 if (ceiling) {
293 ceiling &= PGDIR_MASK;
294 if (!ceiling)
295 return;
296 }
297 if (end - 1 > ceiling - 1)
298 return;
299
300 p4d = p4d_offset(pgd, start);
301 pgd_clear(pgd);
302 p4d_free_tlb(tlb, p4d, start);
303}
304
305/*
306 * This function frees user-level page tables of a process.
307 */
308void free_pgd_range(struct mmu_gather *tlb,
309 unsigned long addr, unsigned long end,
310 unsigned long floor, unsigned long ceiling)
311{
312 pgd_t *pgd;
313 unsigned long next;
314
315 /*
316 * The next few lines have given us lots of grief...
317 *
318 * Why are we testing PMD* at this top level? Because often
319 * there will be no work to do at all, and we'd prefer not to
320 * go all the way down to the bottom just to discover that.
321 *
322 * Why all these "- 1"s? Because 0 represents both the bottom
323 * of the address space and the top of it (using -1 for the
324 * top wouldn't help much: the masks would do the wrong thing).
325 * The rule is that addr 0 and floor 0 refer to the bottom of
326 * the address space, but end 0 and ceiling 0 refer to the top
327 * Comparisons need to use "end - 1" and "ceiling - 1" (though
328 * that end 0 case should be mythical).
329 *
330 * Wherever addr is brought up or ceiling brought down, we must
331 * be careful to reject "the opposite 0" before it confuses the
332 * subsequent tests. But what about where end is brought down
333 * by PMD_SIZE below? no, end can't go down to 0 there.
334 *
335 * Whereas we round start (addr) and ceiling down, by different
336 * masks at different levels, in order to test whether a table
337 * now has no other vmas using it, so can be freed, we don't
338 * bother to round floor or end up - the tests don't need that.
339 */
340
341 addr &= PMD_MASK;
342 if (addr < floor) {
343 addr += PMD_SIZE;
344 if (!addr)
345 return;
346 }
347 if (ceiling) {
348 ceiling &= PMD_MASK;
349 if (!ceiling)
350 return;
351 }
352 if (end - 1 > ceiling - 1)
353 end -= PMD_SIZE;
354 if (addr > end - 1)
355 return;
356 /*
357 * We add page table cache pages with PAGE_SIZE,
358 * (see pte_free_tlb()), flush the tlb if we need
359 */
360 tlb_change_page_size(tlb, PAGE_SIZE);
361 pgd = pgd_offset(tlb->mm, addr);
362 do {
363 next = pgd_addr_end(addr, end);
364 if (pgd_none_or_clear_bad(pgd))
365 continue;
366 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
367 } while (pgd++, addr = next, addr != end);
368}
369
370void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
371 unsigned long floor, unsigned long ceiling)
372{
373 while (vma) {
374 struct vm_area_struct *next = vma->vm_next;
375 unsigned long addr = vma->vm_start;
376
377 /*
378 * Hide vma from rmap and truncate_pagecache before freeing
379 * pgtables
380 */
381 unlink_anon_vmas(vma);
382 unlink_file_vma(vma);
383
384 if (is_vm_hugetlb_page(vma)) {
385 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
386 floor, next ? next->vm_start : ceiling);
387 } else {
388 /*
389 * Optimization: gather nearby vmas into one call down
390 */
391 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
392 && !is_vm_hugetlb_page(next)) {
393 vma = next;
394 next = vma->vm_next;
395 unlink_anon_vmas(vma);
396 unlink_file_vma(vma);
397 }
398 free_pgd_range(tlb, addr, vma->vm_end,
399 floor, next ? next->vm_start : ceiling);
400 }
401 vma = next;
402 }
403}
404
405int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
406{
407 spinlock_t *ptl;
408 pgtable_t new = pte_alloc_one(mm);
409 if (!new)
410 return -ENOMEM;
411
412 /*
413 * Ensure all pte setup (eg. pte page lock and page clearing) are
414 * visible before the pte is made visible to other CPUs by being
415 * put into page tables.
416 *
417 * The other side of the story is the pointer chasing in the page
418 * table walking code (when walking the page table without locking;
419 * ie. most of the time). Fortunately, these data accesses consist
420 * of a chain of data-dependent loads, meaning most CPUs (alpha
421 * being the notable exception) will already guarantee loads are
422 * seen in-order. See the alpha page table accessors for the
423 * smp_read_barrier_depends() barriers in page table walking code.
424 */
425 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
426
427 ptl = pmd_lock(mm, pmd);
428 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
429 mm_inc_nr_ptes(mm);
430 pmd_populate(mm, pmd, new);
431 new = NULL;
432 }
433 spin_unlock(ptl);
434 if (new)
435 pte_free(mm, new);
436 return 0;
437}
438
439int __pte_alloc_kernel(pmd_t *pmd)
440{
441 pte_t *new = pte_alloc_one_kernel(&init_mm);
442 if (!new)
443 return -ENOMEM;
444
445 smp_wmb(); /* See comment in __pte_alloc */
446
447 spin_lock(&init_mm.page_table_lock);
448 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
449 pmd_populate_kernel(&init_mm, pmd, new);
450 new = NULL;
451 }
452 spin_unlock(&init_mm.page_table_lock);
453 if (new)
454 pte_free_kernel(&init_mm, new);
455 return 0;
456}
457
458static inline void init_rss_vec(int *rss)
459{
460 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
461}
462
463static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
464{
465 int i;
466
467 if (current->mm == mm)
468 sync_mm_rss(mm);
469 for (i = 0; i < NR_MM_COUNTERS; i++)
470 if (rss[i])
471 add_mm_counter(mm, i, rss[i]);
472}
473
474/*
475 * This function is called to print an error when a bad pte
476 * is found. For example, we might have a PFN-mapped pte in
477 * a region that doesn't allow it.
478 *
479 * The calling function must still handle the error.
480 */
481static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
482 pte_t pte, struct page *page)
483{
484 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
485 p4d_t *p4d = p4d_offset(pgd, addr);
486 pud_t *pud = pud_offset(p4d, addr);
487 pmd_t *pmd = pmd_offset(pud, addr);
488 struct address_space *mapping;
489 pgoff_t index;
490 static unsigned long resume;
491 static unsigned long nr_shown;
492 static unsigned long nr_unshown;
493
494 /*
495 * Allow a burst of 60 reports, then keep quiet for that minute;
496 * or allow a steady drip of one report per second.
497 */
498 if (nr_shown == 60) {
499 if (time_before(jiffies, resume)) {
500 nr_unshown++;
501 return;
502 }
503 if (nr_unshown) {
504 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
505 nr_unshown);
506 nr_unshown = 0;
507 }
508 nr_shown = 0;
509 }
510 if (nr_shown++ == 0)
511 resume = jiffies + 60 * HZ;
512
513 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
514 index = linear_page_index(vma, addr);
515
516 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
517 current->comm,
518 (long long)pte_val(pte), (long long)pmd_val(*pmd));
519 if (page)
520 dump_page(page, "bad pte");
521 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
522 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
523 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
524 vma->vm_file,
525 vma->vm_ops ? vma->vm_ops->fault : NULL,
526 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
527 mapping ? mapping->a_ops->readpage : NULL);
528 dump_stack();
529 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
530}
531
532/*
533 * vm_normal_page -- This function gets the "struct page" associated with a pte.
534 *
535 * "Special" mappings do not wish to be associated with a "struct page" (either
536 * it doesn't exist, or it exists but they don't want to touch it). In this
537 * case, NULL is returned here. "Normal" mappings do have a struct page.
538 *
539 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
540 * pte bit, in which case this function is trivial. Secondly, an architecture
541 * may not have a spare pte bit, which requires a more complicated scheme,
542 * described below.
543 *
544 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
545 * special mapping (even if there are underlying and valid "struct pages").
546 * COWed pages of a VM_PFNMAP are always normal.
547 *
548 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
549 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
550 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
551 * mapping will always honor the rule
552 *
553 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
554 *
555 * And for normal mappings this is false.
556 *
557 * This restricts such mappings to be a linear translation from virtual address
558 * to pfn. To get around this restriction, we allow arbitrary mappings so long
559 * as the vma is not a COW mapping; in that case, we know that all ptes are
560 * special (because none can have been COWed).
561 *
562 *
563 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
564 *
565 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
566 * page" backing, however the difference is that _all_ pages with a struct
567 * page (that is, those where pfn_valid is true) are refcounted and considered
568 * normal pages by the VM. The disadvantage is that pages are refcounted
569 * (which can be slower and simply not an option for some PFNMAP users). The
570 * advantage is that we don't have to follow the strict linearity rule of
571 * PFNMAP mappings in order to support COWable mappings.
572 *
573 */
574struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
575 pte_t pte)
576{
577 unsigned long pfn = pte_pfn(pte);
578
579 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
580 if (likely(!pte_special(pte)))
581 goto check_pfn;
582 if (vma->vm_ops && vma->vm_ops->find_special_page)
583 return vma->vm_ops->find_special_page(vma, addr);
584 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
585 return NULL;
586 if (is_zero_pfn(pfn))
587 return NULL;
588 if (pte_devmap(pte))
589 return NULL;
590
591 print_bad_pte(vma, addr, pte, NULL);
592 return NULL;
593 }
594
595 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
596
597 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
598 if (vma->vm_flags & VM_MIXEDMAP) {
599 if (!pfn_valid(pfn))
600 return NULL;
601 goto out;
602 } else {
603 unsigned long off;
604 off = (addr - vma->vm_start) >> PAGE_SHIFT;
605 if (pfn == vma->vm_pgoff + off)
606 return NULL;
607 if (!is_cow_mapping(vma->vm_flags))
608 return NULL;
609 }
610 }
611
612 if (is_zero_pfn(pfn))
613 return NULL;
614
615check_pfn:
616 if (unlikely(pfn > highest_memmap_pfn)) {
617 print_bad_pte(vma, addr, pte, NULL);
618 return NULL;
619 }
620
621 /*
622 * NOTE! We still have PageReserved() pages in the page tables.
623 * eg. VDSO mappings can cause them to exist.
624 */
625out:
626 return pfn_to_page(pfn);
627}
628
629#ifdef CONFIG_TRANSPARENT_HUGEPAGE
630struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
631 pmd_t pmd)
632{
633 unsigned long pfn = pmd_pfn(pmd);
634
635 /*
636 * There is no pmd_special() but there may be special pmds, e.g.
637 * in a direct-access (dax) mapping, so let's just replicate the
638 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
639 */
640 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
641 if (vma->vm_flags & VM_MIXEDMAP) {
642 if (!pfn_valid(pfn))
643 return NULL;
644 goto out;
645 } else {
646 unsigned long off;
647 off = (addr - vma->vm_start) >> PAGE_SHIFT;
648 if (pfn == vma->vm_pgoff + off)
649 return NULL;
650 if (!is_cow_mapping(vma->vm_flags))
651 return NULL;
652 }
653 }
654
655 if (pmd_devmap(pmd))
656 return NULL;
657 if (is_zero_pfn(pfn))
658 return NULL;
659 if (unlikely(pfn > highest_memmap_pfn))
660 return NULL;
661
662 /*
663 * NOTE! We still have PageReserved() pages in the page tables.
664 * eg. VDSO mappings can cause them to exist.
665 */
666out:
667 return pfn_to_page(pfn);
668}
669#endif
670
671/*
672 * copy one vm_area from one task to the other. Assumes the page tables
673 * already present in the new task to be cleared in the whole range
674 * covered by this vma.
675 */
676
677static inline unsigned long
678copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
679 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
680 unsigned long addr, int *rss)
681{
682 unsigned long vm_flags = vma->vm_flags;
683 pte_t pte = *src_pte;
684 struct page *page;
685
686 /* pte contains position in swap or file, so copy. */
687 if (unlikely(!pte_present(pte))) {
688 swp_entry_t entry = pte_to_swp_entry(pte);
689
690 if (likely(!non_swap_entry(entry))) {
691 if (swap_duplicate(entry) < 0)
692 return entry.val;
693
694 /* make sure dst_mm is on swapoff's mmlist. */
695 if (unlikely(list_empty(&dst_mm->mmlist))) {
696 spin_lock(&mmlist_lock);
697 if (list_empty(&dst_mm->mmlist))
698 list_add(&dst_mm->mmlist,
699 &src_mm->mmlist);
700 spin_unlock(&mmlist_lock);
701 }
702 rss[MM_SWAPENTS]++;
703 } else if (is_migration_entry(entry)) {
704 page = migration_entry_to_page(entry);
705
706 rss[mm_counter(page)]++;
707
708 if (is_write_migration_entry(entry) &&
709 is_cow_mapping(vm_flags)) {
710 /*
711 * COW mappings require pages in both
712 * parent and child to be set to read.
713 */
714 make_migration_entry_read(&entry);
715 pte = swp_entry_to_pte(entry);
716 if (pte_swp_soft_dirty(*src_pte))
717 pte = pte_swp_mksoft_dirty(pte);
718 set_pte_at(src_mm, addr, src_pte, pte);
719 }
720 } else if (is_device_private_entry(entry)) {
721 page = device_private_entry_to_page(entry);
722
723 /*
724 * Update rss count even for unaddressable pages, as
725 * they should treated just like normal pages in this
726 * respect.
727 *
728 * We will likely want to have some new rss counters
729 * for unaddressable pages, at some point. But for now
730 * keep things as they are.
731 */
732 get_page(page);
733 rss[mm_counter(page)]++;
734 page_dup_rmap(page, false);
735
736 /*
737 * We do not preserve soft-dirty information, because so
738 * far, checkpoint/restore is the only feature that
739 * requires that. And checkpoint/restore does not work
740 * when a device driver is involved (you cannot easily
741 * save and restore device driver state).
742 */
743 if (is_write_device_private_entry(entry) &&
744 is_cow_mapping(vm_flags)) {
745 make_device_private_entry_read(&entry);
746 pte = swp_entry_to_pte(entry);
747 set_pte_at(src_mm, addr, src_pte, pte);
748 }
749 }
750 goto out_set_pte;
751 }
752
753 /*
754 * If it's a COW mapping, write protect it both
755 * in the parent and the child
756 */
757 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
758 ptep_set_wrprotect(src_mm, addr, src_pte);
759 pte = pte_wrprotect(pte);
760 }
761
762 /*
763 * If it's a shared mapping, mark it clean in
764 * the child
765 */
766 if (vm_flags & VM_SHARED)
767 pte = pte_mkclean(pte);
768 pte = pte_mkold(pte);
769
770 page = vm_normal_page(vma, addr, pte);
771 if (page) {
772 get_page(page);
773 page_dup_rmap(page, false);
774 rss[mm_counter(page)]++;
775 } else if (pte_devmap(pte)) {
776 page = pte_page(pte);
777 }
778
779out_set_pte:
780 set_pte_at(dst_mm, addr, dst_pte, pte);
781 return 0;
782}
783
784static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
785 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
786 unsigned long addr, unsigned long end)
787{
788 pte_t *orig_src_pte, *orig_dst_pte;
789 pte_t *src_pte, *dst_pte;
790 spinlock_t *src_ptl, *dst_ptl;
791 int progress = 0;
792 int rss[NR_MM_COUNTERS];
793 swp_entry_t entry = (swp_entry_t){0};
794
795again:
796 init_rss_vec(rss);
797
798 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
799 if (!dst_pte)
800 return -ENOMEM;
801 src_pte = pte_offset_map(src_pmd, addr);
802 src_ptl = pte_lockptr(src_mm, src_pmd);
803 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
804 orig_src_pte = src_pte;
805 orig_dst_pte = dst_pte;
806 arch_enter_lazy_mmu_mode();
807
808 do {
809 /*
810 * We are holding two locks at this point - either of them
811 * could generate latencies in another task on another CPU.
812 */
813 if (progress >= 32) {
814 progress = 0;
815 if (need_resched() ||
816 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
817 break;
818 }
819 if (pte_none(*src_pte)) {
820 progress++;
821 continue;
822 }
823 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
824 vma, addr, rss);
825 if (entry.val)
826 break;
827 progress += 8;
828 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
829
830 arch_leave_lazy_mmu_mode();
831 spin_unlock(src_ptl);
832 pte_unmap(orig_src_pte);
833 add_mm_rss_vec(dst_mm, rss);
834 pte_unmap_unlock(orig_dst_pte, dst_ptl);
835 cond_resched();
836
837 if (entry.val) {
838 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
839 return -ENOMEM;
840 progress = 0;
841 }
842 if (addr != end)
843 goto again;
844 return 0;
845}
846
847static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
848 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
849 unsigned long addr, unsigned long end)
850{
851 pmd_t *src_pmd, *dst_pmd;
852 unsigned long next;
853
854 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
855 if (!dst_pmd)
856 return -ENOMEM;
857 src_pmd = pmd_offset(src_pud, addr);
858 do {
859 next = pmd_addr_end(addr, end);
860 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
861 || pmd_devmap(*src_pmd)) {
862 int err;
863 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
864 err = copy_huge_pmd(dst_mm, src_mm,
865 dst_pmd, src_pmd, addr, vma);
866 if (err == -ENOMEM)
867 return -ENOMEM;
868 if (!err)
869 continue;
870 /* fall through */
871 }
872 if (pmd_none_or_clear_bad(src_pmd))
873 continue;
874 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
875 vma, addr, next))
876 return -ENOMEM;
877 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
878 return 0;
879}
880
881static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
882 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
883 unsigned long addr, unsigned long end)
884{
885 pud_t *src_pud, *dst_pud;
886 unsigned long next;
887
888 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
889 if (!dst_pud)
890 return -ENOMEM;
891 src_pud = pud_offset(src_p4d, addr);
892 do {
893 next = pud_addr_end(addr, end);
894 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
895 int err;
896
897 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
898 err = copy_huge_pud(dst_mm, src_mm,
899 dst_pud, src_pud, addr, vma);
900 if (err == -ENOMEM)
901 return -ENOMEM;
902 if (!err)
903 continue;
904 /* fall through */
905 }
906 if (pud_none_or_clear_bad(src_pud))
907 continue;
908 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
909 vma, addr, next))
910 return -ENOMEM;
911 } while (dst_pud++, src_pud++, addr = next, addr != end);
912 return 0;
913}
914
915static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
916 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
917 unsigned long addr, unsigned long end)
918{
919 p4d_t *src_p4d, *dst_p4d;
920 unsigned long next;
921
922 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
923 if (!dst_p4d)
924 return -ENOMEM;
925 src_p4d = p4d_offset(src_pgd, addr);
926 do {
927 next = p4d_addr_end(addr, end);
928 if (p4d_none_or_clear_bad(src_p4d))
929 continue;
930 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
931 vma, addr, next))
932 return -ENOMEM;
933 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
934 return 0;
935}
936
937int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
938 struct vm_area_struct *vma)
939{
940 pgd_t *src_pgd, *dst_pgd;
941 unsigned long next;
942 unsigned long addr = vma->vm_start;
943 unsigned long end = vma->vm_end;
944 struct mmu_notifier_range range;
945 bool is_cow;
946 int ret;
947
948 /*
949 * Don't copy ptes where a page fault will fill them correctly.
950 * Fork becomes much lighter when there are big shared or private
951 * readonly mappings. The tradeoff is that copy_page_range is more
952 * efficient than faulting.
953 */
954 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
955 !vma->anon_vma)
956 return 0;
957
958 if (is_vm_hugetlb_page(vma))
959 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
960
961 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
962 /*
963 * We do not free on error cases below as remove_vma
964 * gets called on error from higher level routine
965 */
966 ret = track_pfn_copy(vma);
967 if (ret)
968 return ret;
969 }
970
971 /*
972 * We need to invalidate the secondary MMU mappings only when
973 * there could be a permission downgrade on the ptes of the
974 * parent mm. And a permission downgrade will only happen if
975 * is_cow_mapping() returns true.
976 */
977 is_cow = is_cow_mapping(vma->vm_flags);
978
979 if (is_cow) {
980 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
981 0, vma, src_mm, addr, end);
982 mmu_notifier_invalidate_range_start(&range);
983 }
984
985 ret = 0;
986 dst_pgd = pgd_offset(dst_mm, addr);
987 src_pgd = pgd_offset(src_mm, addr);
988 do {
989 next = pgd_addr_end(addr, end);
990 if (pgd_none_or_clear_bad(src_pgd))
991 continue;
992 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
993 vma, addr, next))) {
994 ret = -ENOMEM;
995 break;
996 }
997 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
998
999 if (is_cow)
1000 mmu_notifier_invalidate_range_end(&range);
1001 return ret;
1002}
1003
1004static unsigned long zap_pte_range(struct mmu_gather *tlb,
1005 struct vm_area_struct *vma, pmd_t *pmd,
1006 unsigned long addr, unsigned long end,
1007 struct zap_details *details)
1008{
1009 struct mm_struct *mm = tlb->mm;
1010 int force_flush = 0;
1011 int rss[NR_MM_COUNTERS];
1012 spinlock_t *ptl;
1013 pte_t *start_pte;
1014 pte_t *pte;
1015 swp_entry_t entry;
1016
1017 tlb_change_page_size(tlb, PAGE_SIZE);
1018again:
1019 init_rss_vec(rss);
1020 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1021 pte = start_pte;
1022 flush_tlb_batched_pending(mm);
1023 arch_enter_lazy_mmu_mode();
1024 do {
1025 pte_t ptent = *pte;
1026 if (pte_none(ptent))
1027 continue;
1028
1029 if (need_resched())
1030 break;
1031
1032 if (pte_present(ptent)) {
1033 struct page *page;
1034
1035 page = vm_normal_page(vma, addr, ptent);
1036 if (unlikely(details) && page) {
1037 /*
1038 * unmap_shared_mapping_pages() wants to
1039 * invalidate cache without truncating:
1040 * unmap shared but keep private pages.
1041 */
1042 if (details->check_mapping &&
1043 details->check_mapping != page_rmapping(page))
1044 continue;
1045 }
1046 ptent = ptep_get_and_clear_full(mm, addr, pte,
1047 tlb->fullmm);
1048 tlb_remove_tlb_entry(tlb, pte, addr);
1049 if (unlikely(!page))
1050 continue;
1051
1052 if (!PageAnon(page)) {
1053 if (pte_dirty(ptent)) {
1054 force_flush = 1;
1055 set_page_dirty(page);
1056 }
1057 if (pte_young(ptent) &&
1058 likely(!(vma->vm_flags & VM_SEQ_READ)))
1059 mark_page_accessed(page);
1060 }
1061 rss[mm_counter(page)]--;
1062 page_remove_rmap(page, false);
1063 if (unlikely(page_mapcount(page) < 0))
1064 print_bad_pte(vma, addr, ptent, page);
1065 if (unlikely(__tlb_remove_page(tlb, page))) {
1066 force_flush = 1;
1067 addr += PAGE_SIZE;
1068 break;
1069 }
1070 continue;
1071 }
1072
1073 entry = pte_to_swp_entry(ptent);
1074 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1075 struct page *page = device_private_entry_to_page(entry);
1076
1077 if (unlikely(details && details->check_mapping)) {
1078 /*
1079 * unmap_shared_mapping_pages() wants to
1080 * invalidate cache without truncating:
1081 * unmap shared but keep private pages.
1082 */
1083 if (details->check_mapping !=
1084 page_rmapping(page))
1085 continue;
1086 }
1087
1088 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1089 rss[mm_counter(page)]--;
1090 page_remove_rmap(page, false);
1091 put_page(page);
1092 continue;
1093 }
1094
1095 /* If details->check_mapping, we leave swap entries. */
1096 if (unlikely(details))
1097 continue;
1098
1099 if (!non_swap_entry(entry))
1100 rss[MM_SWAPENTS]--;
1101 else if (is_migration_entry(entry)) {
1102 struct page *page;
1103
1104 page = migration_entry_to_page(entry);
1105 rss[mm_counter(page)]--;
1106 }
1107 if (unlikely(!free_swap_and_cache(entry)))
1108 print_bad_pte(vma, addr, ptent, NULL);
1109 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1110 } while (pte++, addr += PAGE_SIZE, addr != end);
1111
1112 add_mm_rss_vec(mm, rss);
1113 arch_leave_lazy_mmu_mode();
1114
1115 /* Do the actual TLB flush before dropping ptl */
1116 if (force_flush)
1117 tlb_flush_mmu_tlbonly(tlb);
1118 pte_unmap_unlock(start_pte, ptl);
1119
1120 /*
1121 * If we forced a TLB flush (either due to running out of
1122 * batch buffers or because we needed to flush dirty TLB
1123 * entries before releasing the ptl), free the batched
1124 * memory too. Restart if we didn't do everything.
1125 */
1126 if (force_flush) {
1127 force_flush = 0;
1128 tlb_flush_mmu(tlb);
1129 }
1130
1131 if (addr != end) {
1132 cond_resched();
1133 goto again;
1134 }
1135
1136 return addr;
1137}
1138
1139static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1140 struct vm_area_struct *vma, pud_t *pud,
1141 unsigned long addr, unsigned long end,
1142 struct zap_details *details)
1143{
1144 pmd_t *pmd;
1145 unsigned long next;
1146
1147 pmd = pmd_offset(pud, addr);
1148 do {
1149 next = pmd_addr_end(addr, end);
1150 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1151 if (next - addr != HPAGE_PMD_SIZE)
1152 __split_huge_pmd(vma, pmd, addr, false, NULL);
1153 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1154 goto next;
1155 /* fall through */
1156 }
1157 /*
1158 * Here there can be other concurrent MADV_DONTNEED or
1159 * trans huge page faults running, and if the pmd is
1160 * none or trans huge it can change under us. This is
1161 * because MADV_DONTNEED holds the mmap_sem in read
1162 * mode.
1163 */
1164 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1165 goto next;
1166 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1167next:
1168 cond_resched();
1169 } while (pmd++, addr = next, addr != end);
1170
1171 return addr;
1172}
1173
1174static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1175 struct vm_area_struct *vma, p4d_t *p4d,
1176 unsigned long addr, unsigned long end,
1177 struct zap_details *details)
1178{
1179 pud_t *pud;
1180 unsigned long next;
1181
1182 pud = pud_offset(p4d, addr);
1183 do {
1184 next = pud_addr_end(addr, end);
1185 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1186 if (next - addr != HPAGE_PUD_SIZE) {
1187 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1188 split_huge_pud(vma, pud, addr);
1189 } else if (zap_huge_pud(tlb, vma, pud, addr))
1190 goto next;
1191 /* fall through */
1192 }
1193 if (pud_none_or_clear_bad(pud))
1194 continue;
1195 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1196next:
1197 cond_resched();
1198 } while (pud++, addr = next, addr != end);
1199
1200 return addr;
1201}
1202
1203static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1204 struct vm_area_struct *vma, pgd_t *pgd,
1205 unsigned long addr, unsigned long end,
1206 struct zap_details *details)
1207{
1208 p4d_t *p4d;
1209 unsigned long next;
1210
1211 p4d = p4d_offset(pgd, addr);
1212 do {
1213 next = p4d_addr_end(addr, end);
1214 if (p4d_none_or_clear_bad(p4d))
1215 continue;
1216 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1217 } while (p4d++, addr = next, addr != end);
1218
1219 return addr;
1220}
1221
1222void unmap_page_range(struct mmu_gather *tlb,
1223 struct vm_area_struct *vma,
1224 unsigned long addr, unsigned long end,
1225 struct zap_details *details)
1226{
1227 pgd_t *pgd;
1228 unsigned long next;
1229
1230 BUG_ON(addr >= end);
1231 tlb_start_vma(tlb, vma);
1232 pgd = pgd_offset(vma->vm_mm, addr);
1233 do {
1234 next = pgd_addr_end(addr, end);
1235 if (pgd_none_or_clear_bad(pgd))
1236 continue;
1237 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1238 } while (pgd++, addr = next, addr != end);
1239 tlb_end_vma(tlb, vma);
1240}
1241
1242
1243static void unmap_single_vma(struct mmu_gather *tlb,
1244 struct vm_area_struct *vma, unsigned long start_addr,
1245 unsigned long end_addr,
1246 struct zap_details *details)
1247{
1248 unsigned long start = max(vma->vm_start, start_addr);
1249 unsigned long end;
1250
1251 if (start >= vma->vm_end)
1252 return;
1253 end = min(vma->vm_end, end_addr);
1254 if (end <= vma->vm_start)
1255 return;
1256
1257 if (vma->vm_file)
1258 uprobe_munmap(vma, start, end);
1259
1260 if (unlikely(vma->vm_flags & VM_PFNMAP))
1261 untrack_pfn(vma, 0, 0);
1262
1263 if (start != end) {
1264 if (unlikely(is_vm_hugetlb_page(vma))) {
1265 /*
1266 * It is undesirable to test vma->vm_file as it
1267 * should be non-null for valid hugetlb area.
1268 * However, vm_file will be NULL in the error
1269 * cleanup path of mmap_region. When
1270 * hugetlbfs ->mmap method fails,
1271 * mmap_region() nullifies vma->vm_file
1272 * before calling this function to clean up.
1273 * Since no pte has actually been setup, it is
1274 * safe to do nothing in this case.
1275 */
1276 if (vma->vm_file) {
1277 i_mmap_lock_write(vma->vm_file->f_mapping);
1278 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1279 i_mmap_unlock_write(vma->vm_file->f_mapping);
1280 }
1281 } else
1282 unmap_page_range(tlb, vma, start, end, details);
1283 }
1284}
1285
1286/**
1287 * unmap_vmas - unmap a range of memory covered by a list of vma's
1288 * @tlb: address of the caller's struct mmu_gather
1289 * @vma: the starting vma
1290 * @start_addr: virtual address at which to start unmapping
1291 * @end_addr: virtual address at which to end unmapping
1292 *
1293 * Unmap all pages in the vma list.
1294 *
1295 * Only addresses between `start' and `end' will be unmapped.
1296 *
1297 * The VMA list must be sorted in ascending virtual address order.
1298 *
1299 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1300 * range after unmap_vmas() returns. So the only responsibility here is to
1301 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1302 * drops the lock and schedules.
1303 */
1304void unmap_vmas(struct mmu_gather *tlb,
1305 struct vm_area_struct *vma, unsigned long start_addr,
1306 unsigned long end_addr)
1307{
1308 struct mmu_notifier_range range;
1309
1310 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1311 start_addr, end_addr);
1312 mmu_notifier_invalidate_range_start(&range);
1313 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1314 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1315 mmu_notifier_invalidate_range_end(&range);
1316}
1317
1318/**
1319 * zap_page_range - remove user pages in a given range
1320 * @vma: vm_area_struct holding the applicable pages
1321 * @start: starting address of pages to zap
1322 * @size: number of bytes to zap
1323 *
1324 * Caller must protect the VMA list
1325 */
1326void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1327 unsigned long size)
1328{
1329 struct mmu_notifier_range range;
1330 struct mmu_gather tlb;
1331
1332 lru_add_drain();
1333 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1334 start, start + size);
1335 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1336 update_hiwater_rss(vma->vm_mm);
1337 mmu_notifier_invalidate_range_start(&range);
1338 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1339 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1340 mmu_notifier_invalidate_range_end(&range);
1341 tlb_finish_mmu(&tlb, start, range.end);
1342}
1343
1344/**
1345 * zap_page_range_single - remove user pages in a given range
1346 * @vma: vm_area_struct holding the applicable pages
1347 * @address: starting address of pages to zap
1348 * @size: number of bytes to zap
1349 * @details: details of shared cache invalidation
1350 *
1351 * The range must fit into one VMA.
1352 */
1353static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1354 unsigned long size, struct zap_details *details)
1355{
1356 struct mmu_notifier_range range;
1357 struct mmu_gather tlb;
1358
1359 lru_add_drain();
1360 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1361 address, address + size);
1362 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1363 update_hiwater_rss(vma->vm_mm);
1364 mmu_notifier_invalidate_range_start(&range);
1365 unmap_single_vma(&tlb, vma, address, range.end, details);
1366 mmu_notifier_invalidate_range_end(&range);
1367 tlb_finish_mmu(&tlb, address, range.end);
1368}
1369
1370/**
1371 * zap_vma_ptes - remove ptes mapping the vma
1372 * @vma: vm_area_struct holding ptes to be zapped
1373 * @address: starting address of pages to zap
1374 * @size: number of bytes to zap
1375 *
1376 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1377 *
1378 * The entire address range must be fully contained within the vma.
1379 *
1380 */
1381void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1382 unsigned long size)
1383{
1384 if (address < vma->vm_start || address + size > vma->vm_end ||
1385 !(vma->vm_flags & VM_PFNMAP))
1386 return;
1387
1388 zap_page_range_single(vma, address, size, NULL);
1389}
1390EXPORT_SYMBOL_GPL(zap_vma_ptes);
1391
1392pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1393 spinlock_t **ptl)
1394{
1395 pgd_t *pgd;
1396 p4d_t *p4d;
1397 pud_t *pud;
1398 pmd_t *pmd;
1399
1400 pgd = pgd_offset(mm, addr);
1401 p4d = p4d_alloc(mm, pgd, addr);
1402 if (!p4d)
1403 return NULL;
1404 pud = pud_alloc(mm, p4d, addr);
1405 if (!pud)
1406 return NULL;
1407 pmd = pmd_alloc(mm, pud, addr);
1408 if (!pmd)
1409 return NULL;
1410
1411 VM_BUG_ON(pmd_trans_huge(*pmd));
1412 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1413}
1414
1415/*
1416 * This is the old fallback for page remapping.
1417 *
1418 * For historical reasons, it only allows reserved pages. Only
1419 * old drivers should use this, and they needed to mark their
1420 * pages reserved for the old functions anyway.
1421 */
1422static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1423 struct page *page, pgprot_t prot)
1424{
1425 struct mm_struct *mm = vma->vm_mm;
1426 int retval;
1427 pte_t *pte;
1428 spinlock_t *ptl;
1429
1430 retval = -EINVAL;
1431 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1432 goto out;
1433 retval = -ENOMEM;
1434 flush_dcache_page(page);
1435 pte = get_locked_pte(mm, addr, &ptl);
1436 if (!pte)
1437 goto out;
1438 retval = -EBUSY;
1439 if (!pte_none(*pte))
1440 goto out_unlock;
1441
1442 /* Ok, finally just insert the thing.. */
1443 get_page(page);
1444 inc_mm_counter_fast(mm, mm_counter_file(page));
1445 page_add_file_rmap(page, false);
1446 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1447
1448 retval = 0;
1449out_unlock:
1450 pte_unmap_unlock(pte, ptl);
1451out:
1452 return retval;
1453}
1454
1455/**
1456 * vm_insert_page - insert single page into user vma
1457 * @vma: user vma to map to
1458 * @addr: target user address of this page
1459 * @page: source kernel page
1460 *
1461 * This allows drivers to insert individual pages they've allocated
1462 * into a user vma.
1463 *
1464 * The page has to be a nice clean _individual_ kernel allocation.
1465 * If you allocate a compound page, you need to have marked it as
1466 * such (__GFP_COMP), or manually just split the page up yourself
1467 * (see split_page()).
1468 *
1469 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1470 * took an arbitrary page protection parameter. This doesn't allow
1471 * that. Your vma protection will have to be set up correctly, which
1472 * means that if you want a shared writable mapping, you'd better
1473 * ask for a shared writable mapping!
1474 *
1475 * The page does not need to be reserved.
1476 *
1477 * Usually this function is called from f_op->mmap() handler
1478 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1479 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1480 * function from other places, for example from page-fault handler.
1481 *
1482 * Return: %0 on success, negative error code otherwise.
1483 */
1484int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1485 struct page *page)
1486{
1487 if (addr < vma->vm_start || addr >= vma->vm_end)
1488 return -EFAULT;
1489 if (!page_count(page))
1490 return -EINVAL;
1491 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1492 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1493 BUG_ON(vma->vm_flags & VM_PFNMAP);
1494 vma->vm_flags |= VM_MIXEDMAP;
1495 }
1496 return insert_page(vma, addr, page, vma->vm_page_prot);
1497}
1498EXPORT_SYMBOL(vm_insert_page);
1499
1500/*
1501 * __vm_map_pages - maps range of kernel pages into user vma
1502 * @vma: user vma to map to
1503 * @pages: pointer to array of source kernel pages
1504 * @num: number of pages in page array
1505 * @offset: user's requested vm_pgoff
1506 *
1507 * This allows drivers to map range of kernel pages into a user vma.
1508 *
1509 * Return: 0 on success and error code otherwise.
1510 */
1511static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1512 unsigned long num, unsigned long offset)
1513{
1514 unsigned long count = vma_pages(vma);
1515 unsigned long uaddr = vma->vm_start;
1516 int ret, i;
1517
1518 /* Fail if the user requested offset is beyond the end of the object */
1519 if (offset >= num)
1520 return -ENXIO;
1521
1522 /* Fail if the user requested size exceeds available object size */
1523 if (count > num - offset)
1524 return -ENXIO;
1525
1526 for (i = 0; i < count; i++) {
1527 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1528 if (ret < 0)
1529 return ret;
1530 uaddr += PAGE_SIZE;
1531 }
1532
1533 return 0;
1534}
1535
1536/**
1537 * vm_map_pages - maps range of kernel pages starts with non zero offset
1538 * @vma: user vma to map to
1539 * @pages: pointer to array of source kernel pages
1540 * @num: number of pages in page array
1541 *
1542 * Maps an object consisting of @num pages, catering for the user's
1543 * requested vm_pgoff
1544 *
1545 * If we fail to insert any page into the vma, the function will return
1546 * immediately leaving any previously inserted pages present. Callers
1547 * from the mmap handler may immediately return the error as their caller
1548 * will destroy the vma, removing any successfully inserted pages. Other
1549 * callers should make their own arrangements for calling unmap_region().
1550 *
1551 * Context: Process context. Called by mmap handlers.
1552 * Return: 0 on success and error code otherwise.
1553 */
1554int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1555 unsigned long num)
1556{
1557 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1558}
1559EXPORT_SYMBOL(vm_map_pages);
1560
1561/**
1562 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1563 * @vma: user vma to map to
1564 * @pages: pointer to array of source kernel pages
1565 * @num: number of pages in page array
1566 *
1567 * Similar to vm_map_pages(), except that it explicitly sets the offset
1568 * to 0. This function is intended for the drivers that did not consider
1569 * vm_pgoff.
1570 *
1571 * Context: Process context. Called by mmap handlers.
1572 * Return: 0 on success and error code otherwise.
1573 */
1574int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1575 unsigned long num)
1576{
1577 return __vm_map_pages(vma, pages, num, 0);
1578}
1579EXPORT_SYMBOL(vm_map_pages_zero);
1580
1581static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1582 pfn_t pfn, pgprot_t prot, bool mkwrite)
1583{
1584 struct mm_struct *mm = vma->vm_mm;
1585 pte_t *pte, entry;
1586 spinlock_t *ptl;
1587
1588 pte = get_locked_pte(mm, addr, &ptl);
1589 if (!pte)
1590 return VM_FAULT_OOM;
1591 if (!pte_none(*pte)) {
1592 if (mkwrite) {
1593 /*
1594 * For read faults on private mappings the PFN passed
1595 * in may not match the PFN we have mapped if the
1596 * mapped PFN is a writeable COW page. In the mkwrite
1597 * case we are creating a writable PTE for a shared
1598 * mapping and we expect the PFNs to match. If they
1599 * don't match, we are likely racing with block
1600 * allocation and mapping invalidation so just skip the
1601 * update.
1602 */
1603 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1604 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1605 goto out_unlock;
1606 }
1607 entry = pte_mkyoung(*pte);
1608 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1609 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1610 update_mmu_cache(vma, addr, pte);
1611 }
1612 goto out_unlock;
1613 }
1614
1615 /* Ok, finally just insert the thing.. */
1616 if (pfn_t_devmap(pfn))
1617 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1618 else
1619 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1620
1621 if (mkwrite) {
1622 entry = pte_mkyoung(entry);
1623 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1624 }
1625
1626 set_pte_at(mm, addr, pte, entry);
1627 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1628
1629out_unlock:
1630 pte_unmap_unlock(pte, ptl);
1631 return VM_FAULT_NOPAGE;
1632}
1633
1634/**
1635 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1636 * @vma: user vma to map to
1637 * @addr: target user address of this page
1638 * @pfn: source kernel pfn
1639 * @pgprot: pgprot flags for the inserted page
1640 *
1641 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1642 * to override pgprot on a per-page basis.
1643 *
1644 * This only makes sense for IO mappings, and it makes no sense for
1645 * COW mappings. In general, using multiple vmas is preferable;
1646 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1647 * impractical.
1648 *
1649 * Context: Process context. May allocate using %GFP_KERNEL.
1650 * Return: vm_fault_t value.
1651 */
1652vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1653 unsigned long pfn, pgprot_t pgprot)
1654{
1655 /*
1656 * Technically, architectures with pte_special can avoid all these
1657 * restrictions (same for remap_pfn_range). However we would like
1658 * consistency in testing and feature parity among all, so we should
1659 * try to keep these invariants in place for everybody.
1660 */
1661 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1662 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1663 (VM_PFNMAP|VM_MIXEDMAP));
1664 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1665 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1666
1667 if (addr < vma->vm_start || addr >= vma->vm_end)
1668 return VM_FAULT_SIGBUS;
1669
1670 if (!pfn_modify_allowed(pfn, pgprot))
1671 return VM_FAULT_SIGBUS;
1672
1673 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1674
1675 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1676 false);
1677}
1678EXPORT_SYMBOL(vmf_insert_pfn_prot);
1679
1680/**
1681 * vmf_insert_pfn - insert single pfn into user vma
1682 * @vma: user vma to map to
1683 * @addr: target user address of this page
1684 * @pfn: source kernel pfn
1685 *
1686 * Similar to vm_insert_page, this allows drivers to insert individual pages
1687 * they've allocated into a user vma. Same comments apply.
1688 *
1689 * This function should only be called from a vm_ops->fault handler, and
1690 * in that case the handler should return the result of this function.
1691 *
1692 * vma cannot be a COW mapping.
1693 *
1694 * As this is called only for pages that do not currently exist, we
1695 * do not need to flush old virtual caches or the TLB.
1696 *
1697 * Context: Process context. May allocate using %GFP_KERNEL.
1698 * Return: vm_fault_t value.
1699 */
1700vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1701 unsigned long pfn)
1702{
1703 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1704}
1705EXPORT_SYMBOL(vmf_insert_pfn);
1706
1707static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1708{
1709 /* these checks mirror the abort conditions in vm_normal_page */
1710 if (vma->vm_flags & VM_MIXEDMAP)
1711 return true;
1712 if (pfn_t_devmap(pfn))
1713 return true;
1714 if (pfn_t_special(pfn))
1715 return true;
1716 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1717 return true;
1718 return false;
1719}
1720
1721static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1722 unsigned long addr, pfn_t pfn, bool mkwrite)
1723{
1724 pgprot_t pgprot = vma->vm_page_prot;
1725 int err;
1726
1727 BUG_ON(!vm_mixed_ok(vma, pfn));
1728
1729 if (addr < vma->vm_start || addr >= vma->vm_end)
1730 return VM_FAULT_SIGBUS;
1731
1732 track_pfn_insert(vma, &pgprot, pfn);
1733
1734 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1735 return VM_FAULT_SIGBUS;
1736
1737 /*
1738 * If we don't have pte special, then we have to use the pfn_valid()
1739 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1740 * refcount the page if pfn_valid is true (hence insert_page rather
1741 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1742 * without pte special, it would there be refcounted as a normal page.
1743 */
1744 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1745 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1746 struct page *page;
1747
1748 /*
1749 * At this point we are committed to insert_page()
1750 * regardless of whether the caller specified flags that
1751 * result in pfn_t_has_page() == false.
1752 */
1753 page = pfn_to_page(pfn_t_to_pfn(pfn));
1754 err = insert_page(vma, addr, page, pgprot);
1755 } else {
1756 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1757 }
1758
1759 if (err == -ENOMEM)
1760 return VM_FAULT_OOM;
1761 if (err < 0 && err != -EBUSY)
1762 return VM_FAULT_SIGBUS;
1763
1764 return VM_FAULT_NOPAGE;
1765}
1766
1767vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1768 pfn_t pfn)
1769{
1770 return __vm_insert_mixed(vma, addr, pfn, false);
1771}
1772EXPORT_SYMBOL(vmf_insert_mixed);
1773
1774/*
1775 * If the insertion of PTE failed because someone else already added a
1776 * different entry in the mean time, we treat that as success as we assume
1777 * the same entry was actually inserted.
1778 */
1779vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1780 unsigned long addr, pfn_t pfn)
1781{
1782 return __vm_insert_mixed(vma, addr, pfn, true);
1783}
1784EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1785
1786/*
1787 * maps a range of physical memory into the requested pages. the old
1788 * mappings are removed. any references to nonexistent pages results
1789 * in null mappings (currently treated as "copy-on-access")
1790 */
1791static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1792 unsigned long addr, unsigned long end,
1793 unsigned long pfn, pgprot_t prot)
1794{
1795 pte_t *pte;
1796 spinlock_t *ptl;
1797 int err = 0;
1798
1799 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1800 if (!pte)
1801 return -ENOMEM;
1802 arch_enter_lazy_mmu_mode();
1803 do {
1804 BUG_ON(!pte_none(*pte));
1805 if (!pfn_modify_allowed(pfn, prot)) {
1806 err = -EACCES;
1807 break;
1808 }
1809 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1810 pfn++;
1811 } while (pte++, addr += PAGE_SIZE, addr != end);
1812 arch_leave_lazy_mmu_mode();
1813 pte_unmap_unlock(pte - 1, ptl);
1814 return err;
1815}
1816
1817static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1818 unsigned long addr, unsigned long end,
1819 unsigned long pfn, pgprot_t prot)
1820{
1821 pmd_t *pmd;
1822 unsigned long next;
1823 int err;
1824
1825 pfn -= addr >> PAGE_SHIFT;
1826 pmd = pmd_alloc(mm, pud, addr);
1827 if (!pmd)
1828 return -ENOMEM;
1829 VM_BUG_ON(pmd_trans_huge(*pmd));
1830 do {
1831 next = pmd_addr_end(addr, end);
1832 err = remap_pte_range(mm, pmd, addr, next,
1833 pfn + (addr >> PAGE_SHIFT), prot);
1834 if (err)
1835 return err;
1836 } while (pmd++, addr = next, addr != end);
1837 return 0;
1838}
1839
1840static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1841 unsigned long addr, unsigned long end,
1842 unsigned long pfn, pgprot_t prot)
1843{
1844 pud_t *pud;
1845 unsigned long next;
1846 int err;
1847
1848 pfn -= addr >> PAGE_SHIFT;
1849 pud = pud_alloc(mm, p4d, addr);
1850 if (!pud)
1851 return -ENOMEM;
1852 do {
1853 next = pud_addr_end(addr, end);
1854 err = remap_pmd_range(mm, pud, addr, next,
1855 pfn + (addr >> PAGE_SHIFT), prot);
1856 if (err)
1857 return err;
1858 } while (pud++, addr = next, addr != end);
1859 return 0;
1860}
1861
1862static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1863 unsigned long addr, unsigned long end,
1864 unsigned long pfn, pgprot_t prot)
1865{
1866 p4d_t *p4d;
1867 unsigned long next;
1868 int err;
1869
1870 pfn -= addr >> PAGE_SHIFT;
1871 p4d = p4d_alloc(mm, pgd, addr);
1872 if (!p4d)
1873 return -ENOMEM;
1874 do {
1875 next = p4d_addr_end(addr, end);
1876 err = remap_pud_range(mm, p4d, addr, next,
1877 pfn + (addr >> PAGE_SHIFT), prot);
1878 if (err)
1879 return err;
1880 } while (p4d++, addr = next, addr != end);
1881 return 0;
1882}
1883
1884/**
1885 * remap_pfn_range - remap kernel memory to userspace
1886 * @vma: user vma to map to
1887 * @addr: target user address to start at
1888 * @pfn: physical address of kernel memory
1889 * @size: size of map area
1890 * @prot: page protection flags for this mapping
1891 *
1892 * Note: this is only safe if the mm semaphore is held when called.
1893 *
1894 * Return: %0 on success, negative error code otherwise.
1895 */
1896int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1897 unsigned long pfn, unsigned long size, pgprot_t prot)
1898{
1899 pgd_t *pgd;
1900 unsigned long next;
1901 unsigned long end = addr + PAGE_ALIGN(size);
1902 struct mm_struct *mm = vma->vm_mm;
1903 unsigned long remap_pfn = pfn;
1904 int err;
1905
1906 /*
1907 * Physically remapped pages are special. Tell the
1908 * rest of the world about it:
1909 * VM_IO tells people not to look at these pages
1910 * (accesses can have side effects).
1911 * VM_PFNMAP tells the core MM that the base pages are just
1912 * raw PFN mappings, and do not have a "struct page" associated
1913 * with them.
1914 * VM_DONTEXPAND
1915 * Disable vma merging and expanding with mremap().
1916 * VM_DONTDUMP
1917 * Omit vma from core dump, even when VM_IO turned off.
1918 *
1919 * There's a horrible special case to handle copy-on-write
1920 * behaviour that some programs depend on. We mark the "original"
1921 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1922 * See vm_normal_page() for details.
1923 */
1924 if (is_cow_mapping(vma->vm_flags)) {
1925 if (addr != vma->vm_start || end != vma->vm_end)
1926 return -EINVAL;
1927 vma->vm_pgoff = pfn;
1928 }
1929
1930 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1931 if (err)
1932 return -EINVAL;
1933
1934 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1935
1936 BUG_ON(addr >= end);
1937 pfn -= addr >> PAGE_SHIFT;
1938 pgd = pgd_offset(mm, addr);
1939 flush_cache_range(vma, addr, end);
1940 do {
1941 next = pgd_addr_end(addr, end);
1942 err = remap_p4d_range(mm, pgd, addr, next,
1943 pfn + (addr >> PAGE_SHIFT), prot);
1944 if (err)
1945 break;
1946 } while (pgd++, addr = next, addr != end);
1947
1948 if (err)
1949 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1950
1951 return err;
1952}
1953EXPORT_SYMBOL(remap_pfn_range);
1954
1955/**
1956 * vm_iomap_memory - remap memory to userspace
1957 * @vma: user vma to map to
1958 * @start: start of area
1959 * @len: size of area
1960 *
1961 * This is a simplified io_remap_pfn_range() for common driver use. The
1962 * driver just needs to give us the physical memory range to be mapped,
1963 * we'll figure out the rest from the vma information.
1964 *
1965 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1966 * whatever write-combining details or similar.
1967 *
1968 * Return: %0 on success, negative error code otherwise.
1969 */
1970int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1971{
1972 unsigned long vm_len, pfn, pages;
1973
1974 /* Check that the physical memory area passed in looks valid */
1975 if (start + len < start)
1976 return -EINVAL;
1977 /*
1978 * You *really* shouldn't map things that aren't page-aligned,
1979 * but we've historically allowed it because IO memory might
1980 * just have smaller alignment.
1981 */
1982 len += start & ~PAGE_MASK;
1983 pfn = start >> PAGE_SHIFT;
1984 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1985 if (pfn + pages < pfn)
1986 return -EINVAL;
1987
1988 /* We start the mapping 'vm_pgoff' pages into the area */
1989 if (vma->vm_pgoff > pages)
1990 return -EINVAL;
1991 pfn += vma->vm_pgoff;
1992 pages -= vma->vm_pgoff;
1993
1994 /* Can we fit all of the mapping? */
1995 vm_len = vma->vm_end - vma->vm_start;
1996 if (vm_len >> PAGE_SHIFT > pages)
1997 return -EINVAL;
1998
1999 /* Ok, let it rip */
2000 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2001}
2002EXPORT_SYMBOL(vm_iomap_memory);
2003
2004static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2005 unsigned long addr, unsigned long end,
2006 pte_fn_t fn, void *data)
2007{
2008 pte_t *pte;
2009 int err;
2010 spinlock_t *uninitialized_var(ptl);
2011
2012 pte = (mm == &init_mm) ?
2013 pte_alloc_kernel(pmd, addr) :
2014 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2015 if (!pte)
2016 return -ENOMEM;
2017
2018 BUG_ON(pmd_huge(*pmd));
2019
2020 arch_enter_lazy_mmu_mode();
2021
2022 do {
2023 err = fn(pte++, addr, data);
2024 if (err)
2025 break;
2026 } while (addr += PAGE_SIZE, addr != end);
2027
2028 arch_leave_lazy_mmu_mode();
2029
2030 if (mm != &init_mm)
2031 pte_unmap_unlock(pte-1, ptl);
2032 return err;
2033}
2034
2035static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2036 unsigned long addr, unsigned long end,
2037 pte_fn_t fn, void *data)
2038{
2039 pmd_t *pmd;
2040 unsigned long next;
2041 int err;
2042
2043 BUG_ON(pud_huge(*pud));
2044
2045 pmd = pmd_alloc(mm, pud, addr);
2046 if (!pmd)
2047 return -ENOMEM;
2048 do {
2049 next = pmd_addr_end(addr, end);
2050 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2051 if (err)
2052 break;
2053 } while (pmd++, addr = next, addr != end);
2054 return err;
2055}
2056
2057static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2058 unsigned long addr, unsigned long end,
2059 pte_fn_t fn, void *data)
2060{
2061 pud_t *pud;
2062 unsigned long next;
2063 int err;
2064
2065 pud = pud_alloc(mm, p4d, addr);
2066 if (!pud)
2067 return -ENOMEM;
2068 do {
2069 next = pud_addr_end(addr, end);
2070 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2071 if (err)
2072 break;
2073 } while (pud++, addr = next, addr != end);
2074 return err;
2075}
2076
2077static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2078 unsigned long addr, unsigned long end,
2079 pte_fn_t fn, void *data)
2080{
2081 p4d_t *p4d;
2082 unsigned long next;
2083 int err;
2084
2085 p4d = p4d_alloc(mm, pgd, addr);
2086 if (!p4d)
2087 return -ENOMEM;
2088 do {
2089 next = p4d_addr_end(addr, end);
2090 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2091 if (err)
2092 break;
2093 } while (p4d++, addr = next, addr != end);
2094 return err;
2095}
2096
2097/*
2098 * Scan a region of virtual memory, filling in page tables as necessary
2099 * and calling a provided function on each leaf page table.
2100 */
2101int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2102 unsigned long size, pte_fn_t fn, void *data)
2103{
2104 pgd_t *pgd;
2105 unsigned long next;
2106 unsigned long end = addr + size;
2107 int err;
2108
2109 if (WARN_ON(addr >= end))
2110 return -EINVAL;
2111
2112 pgd = pgd_offset(mm, addr);
2113 do {
2114 next = pgd_addr_end(addr, end);
2115 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2116 if (err)
2117 break;
2118 } while (pgd++, addr = next, addr != end);
2119
2120 return err;
2121}
2122EXPORT_SYMBOL_GPL(apply_to_page_range);
2123
2124/*
2125 * handle_pte_fault chooses page fault handler according to an entry which was
2126 * read non-atomically. Before making any commitment, on those architectures
2127 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2128 * parts, do_swap_page must check under lock before unmapping the pte and
2129 * proceeding (but do_wp_page is only called after already making such a check;
2130 * and do_anonymous_page can safely check later on).
2131 */
2132static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2133 pte_t *page_table, pte_t orig_pte)
2134{
2135 int same = 1;
2136#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2137 if (sizeof(pte_t) > sizeof(unsigned long)) {
2138 spinlock_t *ptl = pte_lockptr(mm, pmd);
2139 spin_lock(ptl);
2140 same = pte_same(*page_table, orig_pte);
2141 spin_unlock(ptl);
2142 }
2143#endif
2144 pte_unmap(page_table);
2145 return same;
2146}
2147
2148static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2149{
2150 debug_dma_assert_idle(src);
2151
2152 /*
2153 * If the source page was a PFN mapping, we don't have
2154 * a "struct page" for it. We do a best-effort copy by
2155 * just copying from the original user address. If that
2156 * fails, we just zero-fill it. Live with it.
2157 */
2158 if (unlikely(!src)) {
2159 void *kaddr = kmap_atomic(dst);
2160 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2161
2162 /*
2163 * This really shouldn't fail, because the page is there
2164 * in the page tables. But it might just be unreadable,
2165 * in which case we just give up and fill the result with
2166 * zeroes.
2167 */
2168 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2169 clear_page(kaddr);
2170 kunmap_atomic(kaddr);
2171 flush_dcache_page(dst);
2172 } else
2173 copy_user_highpage(dst, src, va, vma);
2174}
2175
2176static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2177{
2178 struct file *vm_file = vma->vm_file;
2179
2180 if (vm_file)
2181 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2182
2183 /*
2184 * Special mappings (e.g. VDSO) do not have any file so fake
2185 * a default GFP_KERNEL for them.
2186 */
2187 return GFP_KERNEL;
2188}
2189
2190/*
2191 * Notify the address space that the page is about to become writable so that
2192 * it can prohibit this or wait for the page to get into an appropriate state.
2193 *
2194 * We do this without the lock held, so that it can sleep if it needs to.
2195 */
2196static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2197{
2198 vm_fault_t ret;
2199 struct page *page = vmf->page;
2200 unsigned int old_flags = vmf->flags;
2201
2202 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2203
2204 if (vmf->vma->vm_file &&
2205 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2206 return VM_FAULT_SIGBUS;
2207
2208 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2209 /* Restore original flags so that caller is not surprised */
2210 vmf->flags = old_flags;
2211 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2212 return ret;
2213 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2214 lock_page(page);
2215 if (!page->mapping) {
2216 unlock_page(page);
2217 return 0; /* retry */
2218 }
2219 ret |= VM_FAULT_LOCKED;
2220 } else
2221 VM_BUG_ON_PAGE(!PageLocked(page), page);
2222 return ret;
2223}
2224
2225/*
2226 * Handle dirtying of a page in shared file mapping on a write fault.
2227 *
2228 * The function expects the page to be locked and unlocks it.
2229 */
2230static void fault_dirty_shared_page(struct vm_area_struct *vma,
2231 struct page *page)
2232{
2233 struct address_space *mapping;
2234 bool dirtied;
2235 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2236
2237 dirtied = set_page_dirty(page);
2238 VM_BUG_ON_PAGE(PageAnon(page), page);
2239 /*
2240 * Take a local copy of the address_space - page.mapping may be zeroed
2241 * by truncate after unlock_page(). The address_space itself remains
2242 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2243 * release semantics to prevent the compiler from undoing this copying.
2244 */
2245 mapping = page_rmapping(page);
2246 unlock_page(page);
2247
2248 if ((dirtied || page_mkwrite) && mapping) {
2249 /*
2250 * Some device drivers do not set page.mapping
2251 * but still dirty their pages
2252 */
2253 balance_dirty_pages_ratelimited(mapping);
2254 }
2255
2256 if (!page_mkwrite)
2257 file_update_time(vma->vm_file);
2258}
2259
2260/*
2261 * Handle write page faults for pages that can be reused in the current vma
2262 *
2263 * This can happen either due to the mapping being with the VM_SHARED flag,
2264 * or due to us being the last reference standing to the page. In either
2265 * case, all we need to do here is to mark the page as writable and update
2266 * any related book-keeping.
2267 */
2268static inline void wp_page_reuse(struct vm_fault *vmf)
2269 __releases(vmf->ptl)
2270{
2271 struct vm_area_struct *vma = vmf->vma;
2272 struct page *page = vmf->page;
2273 pte_t entry;
2274 /*
2275 * Clear the pages cpupid information as the existing
2276 * information potentially belongs to a now completely
2277 * unrelated process.
2278 */
2279 if (page)
2280 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2281
2282 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2283 entry = pte_mkyoung(vmf->orig_pte);
2284 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2285 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2286 update_mmu_cache(vma, vmf->address, vmf->pte);
2287 pte_unmap_unlock(vmf->pte, vmf->ptl);
2288}
2289
2290/*
2291 * Handle the case of a page which we actually need to copy to a new page.
2292 *
2293 * Called with mmap_sem locked and the old page referenced, but
2294 * without the ptl held.
2295 *
2296 * High level logic flow:
2297 *
2298 * - Allocate a page, copy the content of the old page to the new one.
2299 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2300 * - Take the PTL. If the pte changed, bail out and release the allocated page
2301 * - If the pte is still the way we remember it, update the page table and all
2302 * relevant references. This includes dropping the reference the page-table
2303 * held to the old page, as well as updating the rmap.
2304 * - In any case, unlock the PTL and drop the reference we took to the old page.
2305 */
2306static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2307{
2308 struct vm_area_struct *vma = vmf->vma;
2309 struct mm_struct *mm = vma->vm_mm;
2310 struct page *old_page = vmf->page;
2311 struct page *new_page = NULL;
2312 pte_t entry;
2313 int page_copied = 0;
2314 struct mem_cgroup *memcg;
2315 struct mmu_notifier_range range;
2316
2317 if (unlikely(anon_vma_prepare(vma)))
2318 goto oom;
2319
2320 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2321 new_page = alloc_zeroed_user_highpage_movable(vma,
2322 vmf->address);
2323 if (!new_page)
2324 goto oom;
2325 } else {
2326 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2327 vmf->address);
2328 if (!new_page)
2329 goto oom;
2330 cow_user_page(new_page, old_page, vmf->address, vma);
2331 }
2332
2333 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2334 goto oom_free_new;
2335
2336 __SetPageUptodate(new_page);
2337
2338 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2339 vmf->address & PAGE_MASK,
2340 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2341 mmu_notifier_invalidate_range_start(&range);
2342
2343 /*
2344 * Re-check the pte - we dropped the lock
2345 */
2346 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2347 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2348 if (old_page) {
2349 if (!PageAnon(old_page)) {
2350 dec_mm_counter_fast(mm,
2351 mm_counter_file(old_page));
2352 inc_mm_counter_fast(mm, MM_ANONPAGES);
2353 }
2354 } else {
2355 inc_mm_counter_fast(mm, MM_ANONPAGES);
2356 }
2357 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2358 entry = mk_pte(new_page, vma->vm_page_prot);
2359 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2360 /*
2361 * Clear the pte entry and flush it first, before updating the
2362 * pte with the new entry. This will avoid a race condition
2363 * seen in the presence of one thread doing SMC and another
2364 * thread doing COW.
2365 */
2366 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2367 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2368 mem_cgroup_commit_charge(new_page, memcg, false, false);
2369 lru_cache_add_active_or_unevictable(new_page, vma);
2370 /*
2371 * We call the notify macro here because, when using secondary
2372 * mmu page tables (such as kvm shadow page tables), we want the
2373 * new page to be mapped directly into the secondary page table.
2374 */
2375 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2376 update_mmu_cache(vma, vmf->address, vmf->pte);
2377 if (old_page) {
2378 /*
2379 * Only after switching the pte to the new page may
2380 * we remove the mapcount here. Otherwise another
2381 * process may come and find the rmap count decremented
2382 * before the pte is switched to the new page, and
2383 * "reuse" the old page writing into it while our pte
2384 * here still points into it and can be read by other
2385 * threads.
2386 *
2387 * The critical issue is to order this
2388 * page_remove_rmap with the ptp_clear_flush above.
2389 * Those stores are ordered by (if nothing else,)
2390 * the barrier present in the atomic_add_negative
2391 * in page_remove_rmap.
2392 *
2393 * Then the TLB flush in ptep_clear_flush ensures that
2394 * no process can access the old page before the
2395 * decremented mapcount is visible. And the old page
2396 * cannot be reused until after the decremented
2397 * mapcount is visible. So transitively, TLBs to
2398 * old page will be flushed before it can be reused.
2399 */
2400 page_remove_rmap(old_page, false);
2401 }
2402
2403 /* Free the old page.. */
2404 new_page = old_page;
2405 page_copied = 1;
2406 } else {
2407 mem_cgroup_cancel_charge(new_page, memcg, false);
2408 }
2409
2410 if (new_page)
2411 put_page(new_page);
2412
2413 pte_unmap_unlock(vmf->pte, vmf->ptl);
2414 /*
2415 * No need to double call mmu_notifier->invalidate_range() callback as
2416 * the above ptep_clear_flush_notify() did already call it.
2417 */
2418 mmu_notifier_invalidate_range_only_end(&range);
2419 if (old_page) {
2420 /*
2421 * Don't let another task, with possibly unlocked vma,
2422 * keep the mlocked page.
2423 */
2424 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2425 lock_page(old_page); /* LRU manipulation */
2426 if (PageMlocked(old_page))
2427 munlock_vma_page(old_page);
2428 unlock_page(old_page);
2429 }
2430 put_page(old_page);
2431 }
2432 return page_copied ? VM_FAULT_WRITE : 0;
2433oom_free_new:
2434 put_page(new_page);
2435oom:
2436 if (old_page)
2437 put_page(old_page);
2438 return VM_FAULT_OOM;
2439}
2440
2441/**
2442 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2443 * writeable once the page is prepared
2444 *
2445 * @vmf: structure describing the fault
2446 *
2447 * This function handles all that is needed to finish a write page fault in a
2448 * shared mapping due to PTE being read-only once the mapped page is prepared.
2449 * It handles locking of PTE and modifying it.
2450 *
2451 * The function expects the page to be locked or other protection against
2452 * concurrent faults / writeback (such as DAX radix tree locks).
2453 *
2454 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2455 * we acquired PTE lock.
2456 */
2457vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2458{
2459 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2460 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2461 &vmf->ptl);
2462 /*
2463 * We might have raced with another page fault while we released the
2464 * pte_offset_map_lock.
2465 */
2466 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2467 pte_unmap_unlock(vmf->pte, vmf->ptl);
2468 return VM_FAULT_NOPAGE;
2469 }
2470 wp_page_reuse(vmf);
2471 return 0;
2472}
2473
2474/*
2475 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2476 * mapping
2477 */
2478static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2479{
2480 struct vm_area_struct *vma = vmf->vma;
2481
2482 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2483 vm_fault_t ret;
2484
2485 pte_unmap_unlock(vmf->pte, vmf->ptl);
2486 vmf->flags |= FAULT_FLAG_MKWRITE;
2487 ret = vma->vm_ops->pfn_mkwrite(vmf);
2488 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2489 return ret;
2490 return finish_mkwrite_fault(vmf);
2491 }
2492 wp_page_reuse(vmf);
2493 return VM_FAULT_WRITE;
2494}
2495
2496static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2497 __releases(vmf->ptl)
2498{
2499 struct vm_area_struct *vma = vmf->vma;
2500
2501 get_page(vmf->page);
2502
2503 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2504 vm_fault_t tmp;
2505
2506 pte_unmap_unlock(vmf->pte, vmf->ptl);
2507 tmp = do_page_mkwrite(vmf);
2508 if (unlikely(!tmp || (tmp &
2509 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2510 put_page(vmf->page);
2511 return tmp;
2512 }
2513 tmp = finish_mkwrite_fault(vmf);
2514 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2515 unlock_page(vmf->page);
2516 put_page(vmf->page);
2517 return tmp;
2518 }
2519 } else {
2520 wp_page_reuse(vmf);
2521 lock_page(vmf->page);
2522 }
2523 fault_dirty_shared_page(vma, vmf->page);
2524 put_page(vmf->page);
2525
2526 return VM_FAULT_WRITE;
2527}
2528
2529/*
2530 * This routine handles present pages, when users try to write
2531 * to a shared page. It is done by copying the page to a new address
2532 * and decrementing the shared-page counter for the old page.
2533 *
2534 * Note that this routine assumes that the protection checks have been
2535 * done by the caller (the low-level page fault routine in most cases).
2536 * Thus we can safely just mark it writable once we've done any necessary
2537 * COW.
2538 *
2539 * We also mark the page dirty at this point even though the page will
2540 * change only once the write actually happens. This avoids a few races,
2541 * and potentially makes it more efficient.
2542 *
2543 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2544 * but allow concurrent faults), with pte both mapped and locked.
2545 * We return with mmap_sem still held, but pte unmapped and unlocked.
2546 */
2547static vm_fault_t do_wp_page(struct vm_fault *vmf)
2548 __releases(vmf->ptl)
2549{
2550 struct vm_area_struct *vma = vmf->vma;
2551
2552 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2553 if (!vmf->page) {
2554 /*
2555 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2556 * VM_PFNMAP VMA.
2557 *
2558 * We should not cow pages in a shared writeable mapping.
2559 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2560 */
2561 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2562 (VM_WRITE|VM_SHARED))
2563 return wp_pfn_shared(vmf);
2564
2565 pte_unmap_unlock(vmf->pte, vmf->ptl);
2566 return wp_page_copy(vmf);
2567 }
2568
2569 /*
2570 * Take out anonymous pages first, anonymous shared vmas are
2571 * not dirty accountable.
2572 */
2573 if (PageAnon(vmf->page)) {
2574 int total_map_swapcount;
2575 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2576 page_count(vmf->page) != 1))
2577 goto copy;
2578 if (!trylock_page(vmf->page)) {
2579 get_page(vmf->page);
2580 pte_unmap_unlock(vmf->pte, vmf->ptl);
2581 lock_page(vmf->page);
2582 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2583 vmf->address, &vmf->ptl);
2584 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2585 unlock_page(vmf->page);
2586 pte_unmap_unlock(vmf->pte, vmf->ptl);
2587 put_page(vmf->page);
2588 return 0;
2589 }
2590 put_page(vmf->page);
2591 }
2592 if (PageKsm(vmf->page)) {
2593 bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2594 vmf->address);
2595 unlock_page(vmf->page);
2596 if (!reused)
2597 goto copy;
2598 wp_page_reuse(vmf);
2599 return VM_FAULT_WRITE;
2600 }
2601 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2602 if (total_map_swapcount == 1) {
2603 /*
2604 * The page is all ours. Move it to
2605 * our anon_vma so the rmap code will
2606 * not search our parent or siblings.
2607 * Protected against the rmap code by
2608 * the page lock.
2609 */
2610 page_move_anon_rmap(vmf->page, vma);
2611 }
2612 unlock_page(vmf->page);
2613 wp_page_reuse(vmf);
2614 return VM_FAULT_WRITE;
2615 }
2616 unlock_page(vmf->page);
2617 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2618 (VM_WRITE|VM_SHARED))) {
2619 return wp_page_shared(vmf);
2620 }
2621copy:
2622 /*
2623 * Ok, we need to copy. Oh, well..
2624 */
2625 get_page(vmf->page);
2626
2627 pte_unmap_unlock(vmf->pte, vmf->ptl);
2628 return wp_page_copy(vmf);
2629}
2630
2631static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2632 unsigned long start_addr, unsigned long end_addr,
2633 struct zap_details *details)
2634{
2635 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2636}
2637
2638static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2639 struct zap_details *details)
2640{
2641 struct vm_area_struct *vma;
2642 pgoff_t vba, vea, zba, zea;
2643
2644 vma_interval_tree_foreach(vma, root,
2645 details->first_index, details->last_index) {
2646
2647 vba = vma->vm_pgoff;
2648 vea = vba + vma_pages(vma) - 1;
2649 zba = details->first_index;
2650 if (zba < vba)
2651 zba = vba;
2652 zea = details->last_index;
2653 if (zea > vea)
2654 zea = vea;
2655
2656 unmap_mapping_range_vma(vma,
2657 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2658 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2659 details);
2660 }
2661}
2662
2663/**
2664 * unmap_mapping_pages() - Unmap pages from processes.
2665 * @mapping: The address space containing pages to be unmapped.
2666 * @start: Index of first page to be unmapped.
2667 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2668 * @even_cows: Whether to unmap even private COWed pages.
2669 *
2670 * Unmap the pages in this address space from any userspace process which
2671 * has them mmaped. Generally, you want to remove COWed pages as well when
2672 * a file is being truncated, but not when invalidating pages from the page
2673 * cache.
2674 */
2675void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2676 pgoff_t nr, bool even_cows)
2677{
2678 struct zap_details details = { };
2679
2680 details.check_mapping = even_cows ? NULL : mapping;
2681 details.first_index = start;
2682 details.last_index = start + nr - 1;
2683 if (details.last_index < details.first_index)
2684 details.last_index = ULONG_MAX;
2685
2686 i_mmap_lock_write(mapping);
2687 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2688 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2689 i_mmap_unlock_write(mapping);
2690}
2691
2692/**
2693 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2694 * address_space corresponding to the specified byte range in the underlying
2695 * file.
2696 *
2697 * @mapping: the address space containing mmaps to be unmapped.
2698 * @holebegin: byte in first page to unmap, relative to the start of
2699 * the underlying file. This will be rounded down to a PAGE_SIZE
2700 * boundary. Note that this is different from truncate_pagecache(), which
2701 * must keep the partial page. In contrast, we must get rid of
2702 * partial pages.
2703 * @holelen: size of prospective hole in bytes. This will be rounded
2704 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2705 * end of the file.
2706 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2707 * but 0 when invalidating pagecache, don't throw away private data.
2708 */
2709void unmap_mapping_range(struct address_space *mapping,
2710 loff_t const holebegin, loff_t const holelen, int even_cows)
2711{
2712 pgoff_t hba = holebegin >> PAGE_SHIFT;
2713 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2714
2715 /* Check for overflow. */
2716 if (sizeof(holelen) > sizeof(hlen)) {
2717 long long holeend =
2718 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2719 if (holeend & ~(long long)ULONG_MAX)
2720 hlen = ULONG_MAX - hba + 1;
2721 }
2722
2723 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2724}
2725EXPORT_SYMBOL(unmap_mapping_range);
2726
2727/*
2728 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2729 * but allow concurrent faults), and pte mapped but not yet locked.
2730 * We return with pte unmapped and unlocked.
2731 *
2732 * We return with the mmap_sem locked or unlocked in the same cases
2733 * as does filemap_fault().
2734 */
2735vm_fault_t do_swap_page(struct vm_fault *vmf)
2736{
2737 struct vm_area_struct *vma = vmf->vma;
2738 struct page *page = NULL, *swapcache;
2739 struct mem_cgroup *memcg;
2740 swp_entry_t entry;
2741 pte_t pte;
2742 int locked;
2743 int exclusive = 0;
2744 vm_fault_t ret = 0;
2745
2746 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2747 goto out;
2748
2749 entry = pte_to_swp_entry(vmf->orig_pte);
2750 if (unlikely(non_swap_entry(entry))) {
2751 if (is_migration_entry(entry)) {
2752 migration_entry_wait(vma->vm_mm, vmf->pmd,
2753 vmf->address);
2754 } else if (is_device_private_entry(entry)) {
2755 vmf->page = device_private_entry_to_page(entry);
2756 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
2757 } else if (is_hwpoison_entry(entry)) {
2758 ret = VM_FAULT_HWPOISON;
2759 } else {
2760 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2761 ret = VM_FAULT_SIGBUS;
2762 }
2763 goto out;
2764 }
2765
2766
2767 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2768 page = lookup_swap_cache(entry, vma, vmf->address);
2769 swapcache = page;
2770
2771 if (!page) {
2772 struct swap_info_struct *si = swp_swap_info(entry);
2773
2774 if (si->flags & SWP_SYNCHRONOUS_IO &&
2775 __swap_count(entry) == 1) {
2776 /* skip swapcache */
2777 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2778 vmf->address);
2779 if (page) {
2780 __SetPageLocked(page);
2781 __SetPageSwapBacked(page);
2782 set_page_private(page, entry.val);
2783 lru_cache_add_anon(page);
2784 swap_readpage(page, true);
2785 }
2786 } else {
2787 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2788 vmf);
2789 swapcache = page;
2790 }
2791
2792 if (!page) {
2793 /*
2794 * Back out if somebody else faulted in this pte
2795 * while we released the pte lock.
2796 */
2797 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2798 vmf->address, &vmf->ptl);
2799 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2800 ret = VM_FAULT_OOM;
2801 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2802 goto unlock;
2803 }
2804
2805 /* Had to read the page from swap area: Major fault */
2806 ret = VM_FAULT_MAJOR;
2807 count_vm_event(PGMAJFAULT);
2808 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2809 } else if (PageHWPoison(page)) {
2810 /*
2811 * hwpoisoned dirty swapcache pages are kept for killing
2812 * owner processes (which may be unknown at hwpoison time)
2813 */
2814 ret = VM_FAULT_HWPOISON;
2815 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2816 goto out_release;
2817 }
2818
2819 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2820
2821 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2822 if (!locked) {
2823 ret |= VM_FAULT_RETRY;
2824 goto out_release;
2825 }
2826
2827 /*
2828 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2829 * release the swapcache from under us. The page pin, and pte_same
2830 * test below, are not enough to exclude that. Even if it is still
2831 * swapcache, we need to check that the page's swap has not changed.
2832 */
2833 if (unlikely((!PageSwapCache(page) ||
2834 page_private(page) != entry.val)) && swapcache)
2835 goto out_page;
2836
2837 page = ksm_might_need_to_copy(page, vma, vmf->address);
2838 if (unlikely(!page)) {
2839 ret = VM_FAULT_OOM;
2840 page = swapcache;
2841 goto out_page;
2842 }
2843
2844 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2845 &memcg, false)) {
2846 ret = VM_FAULT_OOM;
2847 goto out_page;
2848 }
2849
2850 /*
2851 * Back out if somebody else already faulted in this pte.
2852 */
2853 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2854 &vmf->ptl);
2855 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2856 goto out_nomap;
2857
2858 if (unlikely(!PageUptodate(page))) {
2859 ret = VM_FAULT_SIGBUS;
2860 goto out_nomap;
2861 }
2862
2863 /*
2864 * The page isn't present yet, go ahead with the fault.
2865 *
2866 * Be careful about the sequence of operations here.
2867 * To get its accounting right, reuse_swap_page() must be called
2868 * while the page is counted on swap but not yet in mapcount i.e.
2869 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2870 * must be called after the swap_free(), or it will never succeed.
2871 */
2872
2873 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2874 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2875 pte = mk_pte(page, vma->vm_page_prot);
2876 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2877 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2878 vmf->flags &= ~FAULT_FLAG_WRITE;
2879 ret |= VM_FAULT_WRITE;
2880 exclusive = RMAP_EXCLUSIVE;
2881 }
2882 flush_icache_page(vma, page);
2883 if (pte_swp_soft_dirty(vmf->orig_pte))
2884 pte = pte_mksoft_dirty(pte);
2885 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2886 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2887 vmf->orig_pte = pte;
2888
2889 /* ksm created a completely new copy */
2890 if (unlikely(page != swapcache && swapcache)) {
2891 page_add_new_anon_rmap(page, vma, vmf->address, false);
2892 mem_cgroup_commit_charge(page, memcg, false, false);
2893 lru_cache_add_active_or_unevictable(page, vma);
2894 } else {
2895 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2896 mem_cgroup_commit_charge(page, memcg, true, false);
2897 activate_page(page);
2898 }
2899
2900 swap_free(entry);
2901 if (mem_cgroup_swap_full(page) ||
2902 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2903 try_to_free_swap(page);
2904 unlock_page(page);
2905 if (page != swapcache && swapcache) {
2906 /*
2907 * Hold the lock to avoid the swap entry to be reused
2908 * until we take the PT lock for the pte_same() check
2909 * (to avoid false positives from pte_same). For
2910 * further safety release the lock after the swap_free
2911 * so that the swap count won't change under a
2912 * parallel locked swapcache.
2913 */
2914 unlock_page(swapcache);
2915 put_page(swapcache);
2916 }
2917
2918 if (vmf->flags & FAULT_FLAG_WRITE) {
2919 ret |= do_wp_page(vmf);
2920 if (ret & VM_FAULT_ERROR)
2921 ret &= VM_FAULT_ERROR;
2922 goto out;
2923 }
2924
2925 /* No need to invalidate - it was non-present before */
2926 update_mmu_cache(vma, vmf->address, vmf->pte);
2927unlock:
2928 pte_unmap_unlock(vmf->pte, vmf->ptl);
2929out:
2930 return ret;
2931out_nomap:
2932 mem_cgroup_cancel_charge(page, memcg, false);
2933 pte_unmap_unlock(vmf->pte, vmf->ptl);
2934out_page:
2935 unlock_page(page);
2936out_release:
2937 put_page(page);
2938 if (page != swapcache && swapcache) {
2939 unlock_page(swapcache);
2940 put_page(swapcache);
2941 }
2942 return ret;
2943}
2944
2945/*
2946 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2947 * but allow concurrent faults), and pte mapped but not yet locked.
2948 * We return with mmap_sem still held, but pte unmapped and unlocked.
2949 */
2950static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
2951{
2952 struct vm_area_struct *vma = vmf->vma;
2953 struct mem_cgroup *memcg;
2954 struct page *page;
2955 vm_fault_t ret = 0;
2956 pte_t entry;
2957
2958 /* File mapping without ->vm_ops ? */
2959 if (vma->vm_flags & VM_SHARED)
2960 return VM_FAULT_SIGBUS;
2961
2962 /*
2963 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2964 * pte_offset_map() on pmds where a huge pmd might be created
2965 * from a different thread.
2966 *
2967 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2968 * parallel threads are excluded by other means.
2969 *
2970 * Here we only have down_read(mmap_sem).
2971 */
2972 if (pte_alloc(vma->vm_mm, vmf->pmd))
2973 return VM_FAULT_OOM;
2974
2975 /* See the comment in pte_alloc_one_map() */
2976 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2977 return 0;
2978
2979 /* Use the zero-page for reads */
2980 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2981 !mm_forbids_zeropage(vma->vm_mm)) {
2982 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2983 vma->vm_page_prot));
2984 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2985 vmf->address, &vmf->ptl);
2986 if (!pte_none(*vmf->pte))
2987 goto unlock;
2988 ret = check_stable_address_space(vma->vm_mm);
2989 if (ret)
2990 goto unlock;
2991 /* Deliver the page fault to userland, check inside PT lock */
2992 if (userfaultfd_missing(vma)) {
2993 pte_unmap_unlock(vmf->pte, vmf->ptl);
2994 return handle_userfault(vmf, VM_UFFD_MISSING);
2995 }
2996 goto setpte;
2997 }
2998
2999 /* Allocate our own private page. */
3000 if (unlikely(anon_vma_prepare(vma)))
3001 goto oom;
3002 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3003 if (!page)
3004 goto oom;
3005
3006 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3007 false))
3008 goto oom_free_page;
3009
3010 /*
3011 * The memory barrier inside __SetPageUptodate makes sure that
3012 * preceeding stores to the page contents become visible before
3013 * the set_pte_at() write.
3014 */
3015 __SetPageUptodate(page);
3016
3017 entry = mk_pte(page, vma->vm_page_prot);
3018 if (vma->vm_flags & VM_WRITE)
3019 entry = pte_mkwrite(pte_mkdirty(entry));
3020
3021 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3022 &vmf->ptl);
3023 if (!pte_none(*vmf->pte))
3024 goto release;
3025
3026 ret = check_stable_address_space(vma->vm_mm);
3027 if (ret)
3028 goto release;
3029
3030 /* Deliver the page fault to userland, check inside PT lock */
3031 if (userfaultfd_missing(vma)) {
3032 pte_unmap_unlock(vmf->pte, vmf->ptl);
3033 mem_cgroup_cancel_charge(page, memcg, false);
3034 put_page(page);
3035 return handle_userfault(vmf, VM_UFFD_MISSING);
3036 }
3037
3038 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3039 page_add_new_anon_rmap(page, vma, vmf->address, false);
3040 mem_cgroup_commit_charge(page, memcg, false, false);
3041 lru_cache_add_active_or_unevictable(page, vma);
3042setpte:
3043 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3044
3045 /* No need to invalidate - it was non-present before */
3046 update_mmu_cache(vma, vmf->address, vmf->pte);
3047unlock:
3048 pte_unmap_unlock(vmf->pte, vmf->ptl);
3049 return ret;
3050release:
3051 mem_cgroup_cancel_charge(page, memcg, false);
3052 put_page(page);
3053 goto unlock;
3054oom_free_page:
3055 put_page(page);
3056oom:
3057 return VM_FAULT_OOM;
3058}
3059
3060/*
3061 * The mmap_sem must have been held on entry, and may have been
3062 * released depending on flags and vma->vm_ops->fault() return value.
3063 * See filemap_fault() and __lock_page_retry().
3064 */
3065static vm_fault_t __do_fault(struct vm_fault *vmf)
3066{
3067 struct vm_area_struct *vma = vmf->vma;
3068 vm_fault_t ret;
3069
3070 /*
3071 * Preallocate pte before we take page_lock because this might lead to
3072 * deadlocks for memcg reclaim which waits for pages under writeback:
3073 * lock_page(A)
3074 * SetPageWriteback(A)
3075 * unlock_page(A)
3076 * lock_page(B)
3077 * lock_page(B)
3078 * pte_alloc_pne
3079 * shrink_page_list
3080 * wait_on_page_writeback(A)
3081 * SetPageWriteback(B)
3082 * unlock_page(B)
3083 * # flush A, B to clear the writeback
3084 */
3085 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3086 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3087 if (!vmf->prealloc_pte)
3088 return VM_FAULT_OOM;
3089 smp_wmb(); /* See comment in __pte_alloc() */
3090 }
3091
3092 ret = vma->vm_ops->fault(vmf);
3093 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3094 VM_FAULT_DONE_COW)))
3095 return ret;
3096
3097 if (unlikely(PageHWPoison(vmf->page))) {
3098 if (ret & VM_FAULT_LOCKED)
3099 unlock_page(vmf->page);
3100 put_page(vmf->page);
3101 vmf->page = NULL;
3102 return VM_FAULT_HWPOISON;
3103 }
3104
3105 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3106 lock_page(vmf->page);
3107 else
3108 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3109
3110 return ret;
3111}
3112
3113/*
3114 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3115 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3116 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3117 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3118 */
3119static int pmd_devmap_trans_unstable(pmd_t *pmd)
3120{
3121 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3122}
3123
3124static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3125{
3126 struct vm_area_struct *vma = vmf->vma;
3127
3128 if (!pmd_none(*vmf->pmd))
3129 goto map_pte;
3130 if (vmf->prealloc_pte) {
3131 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3132 if (unlikely(!pmd_none(*vmf->pmd))) {
3133 spin_unlock(vmf->ptl);
3134 goto map_pte;
3135 }
3136
3137 mm_inc_nr_ptes(vma->vm_mm);
3138 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3139 spin_unlock(vmf->ptl);
3140 vmf->prealloc_pte = NULL;
3141 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3142 return VM_FAULT_OOM;
3143 }
3144map_pte:
3145 /*
3146 * If a huge pmd materialized under us just retry later. Use
3147 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3148 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3149 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3150 * running immediately after a huge pmd fault in a different thread of
3151 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3152 * All we have to ensure is that it is a regular pmd that we can walk
3153 * with pte_offset_map() and we can do that through an atomic read in
3154 * C, which is what pmd_trans_unstable() provides.
3155 */
3156 if (pmd_devmap_trans_unstable(vmf->pmd))
3157 return VM_FAULT_NOPAGE;
3158
3159 /*
3160 * At this point we know that our vmf->pmd points to a page of ptes
3161 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3162 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3163 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3164 * be valid and we will re-check to make sure the vmf->pte isn't
3165 * pte_none() under vmf->ptl protection when we return to
3166 * alloc_set_pte().
3167 */
3168 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3169 &vmf->ptl);
3170 return 0;
3171}
3172
3173#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3174static void deposit_prealloc_pte(struct vm_fault *vmf)
3175{
3176 struct vm_area_struct *vma = vmf->vma;
3177
3178 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3179 /*
3180 * We are going to consume the prealloc table,
3181 * count that as nr_ptes.
3182 */
3183 mm_inc_nr_ptes(vma->vm_mm);
3184 vmf->prealloc_pte = NULL;
3185}
3186
3187static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3188{
3189 struct vm_area_struct *vma = vmf->vma;
3190 bool write = vmf->flags & FAULT_FLAG_WRITE;
3191 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3192 pmd_t entry;
3193 int i;
3194 vm_fault_t ret;
3195
3196 if (!transhuge_vma_suitable(vma, haddr))
3197 return VM_FAULT_FALLBACK;
3198
3199 ret = VM_FAULT_FALLBACK;
3200 page = compound_head(page);
3201
3202 /*
3203 * Archs like ppc64 need additonal space to store information
3204 * related to pte entry. Use the preallocated table for that.
3205 */
3206 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3207 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3208 if (!vmf->prealloc_pte)
3209 return VM_FAULT_OOM;
3210 smp_wmb(); /* See comment in __pte_alloc() */
3211 }
3212
3213 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3214 if (unlikely(!pmd_none(*vmf->pmd)))
3215 goto out;
3216
3217 for (i = 0; i < HPAGE_PMD_NR; i++)
3218 flush_icache_page(vma, page + i);
3219
3220 entry = mk_huge_pmd(page, vma->vm_page_prot);
3221 if (write)
3222 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3223
3224 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3225 page_add_file_rmap(page, true);
3226 /*
3227 * deposit and withdraw with pmd lock held
3228 */
3229 if (arch_needs_pgtable_deposit())
3230 deposit_prealloc_pte(vmf);
3231
3232 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3233
3234 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3235
3236 /* fault is handled */
3237 ret = 0;
3238 count_vm_event(THP_FILE_MAPPED);
3239out:
3240 spin_unlock(vmf->ptl);
3241 return ret;
3242}
3243#else
3244static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3245{
3246 BUILD_BUG();
3247 return 0;
3248}
3249#endif
3250
3251/**
3252 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3253 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3254 *
3255 * @vmf: fault environment
3256 * @memcg: memcg to charge page (only for private mappings)
3257 * @page: page to map
3258 *
3259 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3260 * return.
3261 *
3262 * Target users are page handler itself and implementations of
3263 * vm_ops->map_pages.
3264 *
3265 * Return: %0 on success, %VM_FAULT_ code in case of error.
3266 */
3267vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3268 struct page *page)
3269{
3270 struct vm_area_struct *vma = vmf->vma;
3271 bool write = vmf->flags & FAULT_FLAG_WRITE;
3272 pte_t entry;
3273 vm_fault_t ret;
3274
3275 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3276 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3277 /* THP on COW? */
3278 VM_BUG_ON_PAGE(memcg, page);
3279
3280 ret = do_set_pmd(vmf, page);
3281 if (ret != VM_FAULT_FALLBACK)
3282 return ret;
3283 }
3284
3285 if (!vmf->pte) {
3286 ret = pte_alloc_one_map(vmf);
3287 if (ret)
3288 return ret;
3289 }
3290
3291 /* Re-check under ptl */
3292 if (unlikely(!pte_none(*vmf->pte)))
3293 return VM_FAULT_NOPAGE;
3294
3295 flush_icache_page(vma, page);
3296 entry = mk_pte(page, vma->vm_page_prot);
3297 if (write)
3298 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3299 /* copy-on-write page */
3300 if (write && !(vma->vm_flags & VM_SHARED)) {
3301 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3302 page_add_new_anon_rmap(page, vma, vmf->address, false);
3303 mem_cgroup_commit_charge(page, memcg, false, false);
3304 lru_cache_add_active_or_unevictable(page, vma);
3305 } else {
3306 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3307 page_add_file_rmap(page, false);
3308 }
3309 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3310
3311 /* no need to invalidate: a not-present page won't be cached */
3312 update_mmu_cache(vma, vmf->address, vmf->pte);
3313
3314 return 0;
3315}
3316
3317
3318/**
3319 * finish_fault - finish page fault once we have prepared the page to fault
3320 *
3321 * @vmf: structure describing the fault
3322 *
3323 * This function handles all that is needed to finish a page fault once the
3324 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3325 * given page, adds reverse page mapping, handles memcg charges and LRU
3326 * addition.
3327 *
3328 * The function expects the page to be locked and on success it consumes a
3329 * reference of a page being mapped (for the PTE which maps it).
3330 *
3331 * Return: %0 on success, %VM_FAULT_ code in case of error.
3332 */
3333vm_fault_t finish_fault(struct vm_fault *vmf)
3334{
3335 struct page *page;
3336 vm_fault_t ret = 0;
3337
3338 /* Did we COW the page? */
3339 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3340 !(vmf->vma->vm_flags & VM_SHARED))
3341 page = vmf->cow_page;
3342 else
3343 page = vmf->page;
3344
3345 /*
3346 * check even for read faults because we might have lost our CoWed
3347 * page
3348 */
3349 if (!(vmf->vma->vm_flags & VM_SHARED))
3350 ret = check_stable_address_space(vmf->vma->vm_mm);
3351 if (!ret)
3352 ret = alloc_set_pte(vmf, vmf->memcg, page);
3353 if (vmf->pte)
3354 pte_unmap_unlock(vmf->pte, vmf->ptl);
3355 return ret;
3356}
3357
3358static unsigned long fault_around_bytes __read_mostly =
3359 rounddown_pow_of_two(65536);
3360
3361#ifdef CONFIG_DEBUG_FS
3362static int fault_around_bytes_get(void *data, u64 *val)
3363{
3364 *val = fault_around_bytes;
3365 return 0;
3366}
3367
3368/*
3369 * fault_around_bytes must be rounded down to the nearest page order as it's
3370 * what do_fault_around() expects to see.
3371 */
3372static int fault_around_bytes_set(void *data, u64 val)
3373{
3374 if (val / PAGE_SIZE > PTRS_PER_PTE)
3375 return -EINVAL;
3376 if (val > PAGE_SIZE)
3377 fault_around_bytes = rounddown_pow_of_two(val);
3378 else
3379 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3380 return 0;
3381}
3382DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3383 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3384
3385static int __init fault_around_debugfs(void)
3386{
3387 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3388 &fault_around_bytes_fops);
3389 return 0;
3390}
3391late_initcall(fault_around_debugfs);
3392#endif
3393
3394/*
3395 * do_fault_around() tries to map few pages around the fault address. The hope
3396 * is that the pages will be needed soon and this will lower the number of
3397 * faults to handle.
3398 *
3399 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3400 * not ready to be mapped: not up-to-date, locked, etc.
3401 *
3402 * This function is called with the page table lock taken. In the split ptlock
3403 * case the page table lock only protects only those entries which belong to
3404 * the page table corresponding to the fault address.
3405 *
3406 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3407 * only once.
3408 *
3409 * fault_around_bytes defines how many bytes we'll try to map.
3410 * do_fault_around() expects it to be set to a power of two less than or equal
3411 * to PTRS_PER_PTE.
3412 *
3413 * The virtual address of the area that we map is naturally aligned to
3414 * fault_around_bytes rounded down to the machine page size
3415 * (and therefore to page order). This way it's easier to guarantee
3416 * that we don't cross page table boundaries.
3417 */
3418static vm_fault_t do_fault_around(struct vm_fault *vmf)
3419{
3420 unsigned long address = vmf->address, nr_pages, mask;
3421 pgoff_t start_pgoff = vmf->pgoff;
3422 pgoff_t end_pgoff;
3423 int off;
3424 vm_fault_t ret = 0;
3425
3426 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3427 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3428
3429 vmf->address = max(address & mask, vmf->vma->vm_start);
3430 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3431 start_pgoff -= off;
3432
3433 /*
3434 * end_pgoff is either the end of the page table, the end of
3435 * the vma or nr_pages from start_pgoff, depending what is nearest.
3436 */
3437 end_pgoff = start_pgoff -
3438 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3439 PTRS_PER_PTE - 1;
3440 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3441 start_pgoff + nr_pages - 1);
3442
3443 if (pmd_none(*vmf->pmd)) {
3444 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3445 if (!vmf->prealloc_pte)
3446 goto out;
3447 smp_wmb(); /* See comment in __pte_alloc() */
3448 }
3449
3450 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3451
3452 /* Huge page is mapped? Page fault is solved */
3453 if (pmd_trans_huge(*vmf->pmd)) {
3454 ret = VM_FAULT_NOPAGE;
3455 goto out;
3456 }
3457
3458 /* ->map_pages() haven't done anything useful. Cold page cache? */
3459 if (!vmf->pte)
3460 goto out;
3461
3462 /* check if the page fault is solved */
3463 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3464 if (!pte_none(*vmf->pte))
3465 ret = VM_FAULT_NOPAGE;
3466 pte_unmap_unlock(vmf->pte, vmf->ptl);
3467out:
3468 vmf->address = address;
3469 vmf->pte = NULL;
3470 return ret;
3471}
3472
3473static vm_fault_t do_read_fault(struct vm_fault *vmf)
3474{
3475 struct vm_area_struct *vma = vmf->vma;
3476 vm_fault_t ret = 0;
3477
3478 /*
3479 * Let's call ->map_pages() first and use ->fault() as fallback
3480 * if page by the offset is not ready to be mapped (cold cache or
3481 * something).
3482 */
3483 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3484 ret = do_fault_around(vmf);
3485 if (ret)
3486 return ret;
3487 }
3488
3489 ret = __do_fault(vmf);
3490 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3491 return ret;
3492
3493 ret |= finish_fault(vmf);
3494 unlock_page(vmf->page);
3495 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3496 put_page(vmf->page);
3497 return ret;
3498}
3499
3500static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3501{
3502 struct vm_area_struct *vma = vmf->vma;
3503 vm_fault_t ret;
3504
3505 if (unlikely(anon_vma_prepare(vma)))
3506 return VM_FAULT_OOM;
3507
3508 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3509 if (!vmf->cow_page)
3510 return VM_FAULT_OOM;
3511
3512 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3513 &vmf->memcg, false)) {
3514 put_page(vmf->cow_page);
3515 return VM_FAULT_OOM;
3516 }
3517
3518 ret = __do_fault(vmf);
3519 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3520 goto uncharge_out;
3521 if (ret & VM_FAULT_DONE_COW)
3522 return ret;
3523
3524 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3525 __SetPageUptodate(vmf->cow_page);
3526
3527 ret |= finish_fault(vmf);
3528 unlock_page(vmf->page);
3529 put_page(vmf->page);
3530 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3531 goto uncharge_out;
3532 return ret;
3533uncharge_out:
3534 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3535 put_page(vmf->cow_page);
3536 return ret;
3537}
3538
3539static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3540{
3541 struct vm_area_struct *vma = vmf->vma;
3542 vm_fault_t ret, tmp;
3543
3544 ret = __do_fault(vmf);
3545 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3546 return ret;
3547
3548 /*
3549 * Check if the backing address space wants to know that the page is
3550 * about to become writable
3551 */
3552 if (vma->vm_ops->page_mkwrite) {
3553 unlock_page(vmf->page);
3554 tmp = do_page_mkwrite(vmf);
3555 if (unlikely(!tmp ||
3556 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3557 put_page(vmf->page);
3558 return tmp;
3559 }
3560 }
3561
3562 ret |= finish_fault(vmf);
3563 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3564 VM_FAULT_RETRY))) {
3565 unlock_page(vmf->page);
3566 put_page(vmf->page);
3567 return ret;
3568 }
3569
3570 fault_dirty_shared_page(vma, vmf->page);
3571 return ret;
3572}
3573
3574/*
3575 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3576 * but allow concurrent faults).
3577 * The mmap_sem may have been released depending on flags and our
3578 * return value. See filemap_fault() and __lock_page_or_retry().
3579 * If mmap_sem is released, vma may become invalid (for example
3580 * by other thread calling munmap()).
3581 */
3582static vm_fault_t do_fault(struct vm_fault *vmf)
3583{
3584 struct vm_area_struct *vma = vmf->vma;
3585 struct mm_struct *vm_mm = vma->vm_mm;
3586 vm_fault_t ret;
3587
3588 /*
3589 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3590 */
3591 if (!vma->vm_ops->fault) {
3592 /*
3593 * If we find a migration pmd entry or a none pmd entry, which
3594 * should never happen, return SIGBUS
3595 */
3596 if (unlikely(!pmd_present(*vmf->pmd)))
3597 ret = VM_FAULT_SIGBUS;
3598 else {
3599 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3600 vmf->pmd,
3601 vmf->address,
3602 &vmf->ptl);
3603 /*
3604 * Make sure this is not a temporary clearing of pte
3605 * by holding ptl and checking again. A R/M/W update
3606 * of pte involves: take ptl, clearing the pte so that
3607 * we don't have concurrent modification by hardware
3608 * followed by an update.
3609 */
3610 if (unlikely(pte_none(*vmf->pte)))
3611 ret = VM_FAULT_SIGBUS;
3612 else
3613 ret = VM_FAULT_NOPAGE;
3614
3615 pte_unmap_unlock(vmf->pte, vmf->ptl);
3616 }
3617 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3618 ret = do_read_fault(vmf);
3619 else if (!(vma->vm_flags & VM_SHARED))
3620 ret = do_cow_fault(vmf);
3621 else
3622 ret = do_shared_fault(vmf);
3623
3624 /* preallocated pagetable is unused: free it */
3625 if (vmf->prealloc_pte) {
3626 pte_free(vm_mm, vmf->prealloc_pte);
3627 vmf->prealloc_pte = NULL;
3628 }
3629 return ret;
3630}
3631
3632static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3633 unsigned long addr, int page_nid,
3634 int *flags)
3635{
3636 get_page(page);
3637
3638 count_vm_numa_event(NUMA_HINT_FAULTS);
3639 if (page_nid == numa_node_id()) {
3640 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3641 *flags |= TNF_FAULT_LOCAL;
3642 }
3643
3644 return mpol_misplaced(page, vma, addr);
3645}
3646
3647static vm_fault_t do_numa_page(struct vm_fault *vmf)
3648{
3649 struct vm_area_struct *vma = vmf->vma;
3650 struct page *page = NULL;
3651 int page_nid = NUMA_NO_NODE;
3652 int last_cpupid;
3653 int target_nid;
3654 bool migrated = false;
3655 pte_t pte, old_pte;
3656 bool was_writable = pte_savedwrite(vmf->orig_pte);
3657 int flags = 0;
3658
3659 /*
3660 * The "pte" at this point cannot be used safely without
3661 * validation through pte_unmap_same(). It's of NUMA type but
3662 * the pfn may be screwed if the read is non atomic.
3663 */
3664 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3665 spin_lock(vmf->ptl);
3666 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3667 pte_unmap_unlock(vmf->pte, vmf->ptl);
3668 goto out;
3669 }
3670
3671 /*
3672 * Make it present again, Depending on how arch implementes non
3673 * accessible ptes, some can allow access by kernel mode.
3674 */
3675 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
3676 pte = pte_modify(old_pte, vma->vm_page_prot);
3677 pte = pte_mkyoung(pte);
3678 if (was_writable)
3679 pte = pte_mkwrite(pte);
3680 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
3681 update_mmu_cache(vma, vmf->address, vmf->pte);
3682
3683 page = vm_normal_page(vma, vmf->address, pte);
3684 if (!page) {
3685 pte_unmap_unlock(vmf->pte, vmf->ptl);
3686 return 0;
3687 }
3688
3689 /* TODO: handle PTE-mapped THP */
3690 if (PageCompound(page)) {
3691 pte_unmap_unlock(vmf->pte, vmf->ptl);
3692 return 0;
3693 }
3694
3695 /*
3696 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3697 * much anyway since they can be in shared cache state. This misses
3698 * the case where a mapping is writable but the process never writes
3699 * to it but pte_write gets cleared during protection updates and
3700 * pte_dirty has unpredictable behaviour between PTE scan updates,
3701 * background writeback, dirty balancing and application behaviour.
3702 */
3703 if (!pte_write(pte))
3704 flags |= TNF_NO_GROUP;
3705
3706 /*
3707 * Flag if the page is shared between multiple address spaces. This
3708 * is later used when determining whether to group tasks together
3709 */
3710 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3711 flags |= TNF_SHARED;
3712
3713 last_cpupid = page_cpupid_last(page);
3714 page_nid = page_to_nid(page);
3715 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3716 &flags);
3717 pte_unmap_unlock(vmf->pte, vmf->ptl);
3718 if (target_nid == NUMA_NO_NODE) {
3719 put_page(page);
3720 goto out;
3721 }
3722
3723 /* Migrate to the requested node */
3724 migrated = migrate_misplaced_page(page, vma, target_nid);
3725 if (migrated) {
3726 page_nid = target_nid;
3727 flags |= TNF_MIGRATED;
3728 } else
3729 flags |= TNF_MIGRATE_FAIL;
3730
3731out:
3732 if (page_nid != NUMA_NO_NODE)
3733 task_numa_fault(last_cpupid, page_nid, 1, flags);
3734 return 0;
3735}
3736
3737static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
3738{
3739 if (vma_is_anonymous(vmf->vma))
3740 return do_huge_pmd_anonymous_page(vmf);
3741 if (vmf->vma->vm_ops->huge_fault)
3742 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3743 return VM_FAULT_FALLBACK;
3744}
3745
3746/* `inline' is required to avoid gcc 4.1.2 build error */
3747static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3748{
3749 if (vma_is_anonymous(vmf->vma))
3750 return do_huge_pmd_wp_page(vmf, orig_pmd);
3751 if (vmf->vma->vm_ops->huge_fault)
3752 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3753
3754 /* COW handled on pte level: split pmd */
3755 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3756 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3757
3758 return VM_FAULT_FALLBACK;
3759}
3760
3761static inline bool vma_is_accessible(struct vm_area_struct *vma)
3762{
3763 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3764}
3765
3766static vm_fault_t create_huge_pud(struct vm_fault *vmf)
3767{
3768#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3769 /* No support for anonymous transparent PUD pages yet */
3770 if (vma_is_anonymous(vmf->vma))
3771 return VM_FAULT_FALLBACK;
3772 if (vmf->vma->vm_ops->huge_fault)
3773 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3774#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3775 return VM_FAULT_FALLBACK;
3776}
3777
3778static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3779{
3780#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3781 /* No support for anonymous transparent PUD pages yet */
3782 if (vma_is_anonymous(vmf->vma))
3783 return VM_FAULT_FALLBACK;
3784 if (vmf->vma->vm_ops->huge_fault)
3785 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3786#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3787 return VM_FAULT_FALLBACK;
3788}
3789
3790/*
3791 * These routines also need to handle stuff like marking pages dirty
3792 * and/or accessed for architectures that don't do it in hardware (most
3793 * RISC architectures). The early dirtying is also good on the i386.
3794 *
3795 * There is also a hook called "update_mmu_cache()" that architectures
3796 * with external mmu caches can use to update those (ie the Sparc or
3797 * PowerPC hashed page tables that act as extended TLBs).
3798 *
3799 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3800 * concurrent faults).
3801 *
3802 * The mmap_sem may have been released depending on flags and our return value.
3803 * See filemap_fault() and __lock_page_or_retry().
3804 */
3805static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
3806{
3807 pte_t entry;
3808
3809 if (unlikely(pmd_none(*vmf->pmd))) {
3810 /*
3811 * Leave __pte_alloc() until later: because vm_ops->fault may
3812 * want to allocate huge page, and if we expose page table
3813 * for an instant, it will be difficult to retract from
3814 * concurrent faults and from rmap lookups.
3815 */
3816 vmf->pte = NULL;
3817 } else {
3818 /* See comment in pte_alloc_one_map() */
3819 if (pmd_devmap_trans_unstable(vmf->pmd))
3820 return 0;
3821 /*
3822 * A regular pmd is established and it can't morph into a huge
3823 * pmd from under us anymore at this point because we hold the
3824 * mmap_sem read mode and khugepaged takes it in write mode.
3825 * So now it's safe to run pte_offset_map().
3826 */
3827 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3828 vmf->orig_pte = *vmf->pte;
3829
3830 /*
3831 * some architectures can have larger ptes than wordsize,
3832 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3833 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3834 * accesses. The code below just needs a consistent view
3835 * for the ifs and we later double check anyway with the
3836 * ptl lock held. So here a barrier will do.
3837 */
3838 barrier();
3839 if (pte_none(vmf->orig_pte)) {
3840 pte_unmap(vmf->pte);
3841 vmf->pte = NULL;
3842 }
3843 }
3844
3845 if (!vmf->pte) {
3846 if (vma_is_anonymous(vmf->vma))
3847 return do_anonymous_page(vmf);
3848 else
3849 return do_fault(vmf);
3850 }
3851
3852 if (!pte_present(vmf->orig_pte))
3853 return do_swap_page(vmf);
3854
3855 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3856 return do_numa_page(vmf);
3857
3858 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3859 spin_lock(vmf->ptl);
3860 entry = vmf->orig_pte;
3861 if (unlikely(!pte_same(*vmf->pte, entry)))
3862 goto unlock;
3863 if (vmf->flags & FAULT_FLAG_WRITE) {
3864 if (!pte_write(entry))
3865 return do_wp_page(vmf);
3866 entry = pte_mkdirty(entry);
3867 }
3868 entry = pte_mkyoung(entry);
3869 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3870 vmf->flags & FAULT_FLAG_WRITE)) {
3871 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3872 } else {
3873 /*
3874 * This is needed only for protection faults but the arch code
3875 * is not yet telling us if this is a protection fault or not.
3876 * This still avoids useless tlb flushes for .text page faults
3877 * with threads.
3878 */
3879 if (vmf->flags & FAULT_FLAG_WRITE)
3880 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3881 }
3882unlock:
3883 pte_unmap_unlock(vmf->pte, vmf->ptl);
3884 return 0;
3885}
3886
3887/*
3888 * By the time we get here, we already hold the mm semaphore
3889 *
3890 * The mmap_sem may have been released depending on flags and our
3891 * return value. See filemap_fault() and __lock_page_or_retry().
3892 */
3893static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
3894 unsigned long address, unsigned int flags)
3895{
3896 struct vm_fault vmf = {
3897 .vma = vma,
3898 .address = address & PAGE_MASK,
3899 .flags = flags,
3900 .pgoff = linear_page_index(vma, address),
3901 .gfp_mask = __get_fault_gfp_mask(vma),
3902 };
3903 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3904 struct mm_struct *mm = vma->vm_mm;
3905 pgd_t *pgd;
3906 p4d_t *p4d;
3907 vm_fault_t ret;
3908
3909 pgd = pgd_offset(mm, address);
3910 p4d = p4d_alloc(mm, pgd, address);
3911 if (!p4d)
3912 return VM_FAULT_OOM;
3913
3914 vmf.pud = pud_alloc(mm, p4d, address);
3915 if (!vmf.pud)
3916 return VM_FAULT_OOM;
3917 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
3918 ret = create_huge_pud(&vmf);
3919 if (!(ret & VM_FAULT_FALLBACK))
3920 return ret;
3921 } else {
3922 pud_t orig_pud = *vmf.pud;
3923
3924 barrier();
3925 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3926
3927 /* NUMA case for anonymous PUDs would go here */
3928
3929 if (dirty && !pud_write(orig_pud)) {
3930 ret = wp_huge_pud(&vmf, orig_pud);
3931 if (!(ret & VM_FAULT_FALLBACK))
3932 return ret;
3933 } else {
3934 huge_pud_set_accessed(&vmf, orig_pud);
3935 return 0;
3936 }
3937 }
3938 }
3939
3940 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3941 if (!vmf.pmd)
3942 return VM_FAULT_OOM;
3943 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
3944 ret = create_huge_pmd(&vmf);
3945 if (!(ret & VM_FAULT_FALLBACK))
3946 return ret;
3947 } else {
3948 pmd_t orig_pmd = *vmf.pmd;
3949
3950 barrier();
3951 if (unlikely(is_swap_pmd(orig_pmd))) {
3952 VM_BUG_ON(thp_migration_supported() &&
3953 !is_pmd_migration_entry(orig_pmd));
3954 if (is_pmd_migration_entry(orig_pmd))
3955 pmd_migration_entry_wait(mm, vmf.pmd);
3956 return 0;
3957 }
3958 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3959 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3960 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3961
3962 if (dirty && !pmd_write(orig_pmd)) {
3963 ret = wp_huge_pmd(&vmf, orig_pmd);
3964 if (!(ret & VM_FAULT_FALLBACK))
3965 return ret;
3966 } else {
3967 huge_pmd_set_accessed(&vmf, orig_pmd);
3968 return 0;
3969 }
3970 }
3971 }
3972
3973 return handle_pte_fault(&vmf);
3974}
3975
3976/*
3977 * By the time we get here, we already hold the mm semaphore
3978 *
3979 * The mmap_sem may have been released depending on flags and our
3980 * return value. See filemap_fault() and __lock_page_or_retry().
3981 */
3982vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3983 unsigned int flags)
3984{
3985 vm_fault_t ret;
3986
3987 __set_current_state(TASK_RUNNING);
3988
3989 count_vm_event(PGFAULT);
3990 count_memcg_event_mm(vma->vm_mm, PGFAULT);
3991
3992 /* do counter updates before entering really critical section. */
3993 check_sync_rss_stat(current);
3994
3995 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3996 flags & FAULT_FLAG_INSTRUCTION,
3997 flags & FAULT_FLAG_REMOTE))
3998 return VM_FAULT_SIGSEGV;
3999
4000 /*
4001 * Enable the memcg OOM handling for faults triggered in user
4002 * space. Kernel faults are handled more gracefully.
4003 */
4004 if (flags & FAULT_FLAG_USER)
4005 mem_cgroup_enter_user_fault();
4006
4007 if (unlikely(is_vm_hugetlb_page(vma)))
4008 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4009 else
4010 ret = __handle_mm_fault(vma, address, flags);
4011
4012 if (flags & FAULT_FLAG_USER) {
4013 mem_cgroup_exit_user_fault();
4014 /*
4015 * The task may have entered a memcg OOM situation but
4016 * if the allocation error was handled gracefully (no
4017 * VM_FAULT_OOM), there is no need to kill anything.
4018 * Just clean up the OOM state peacefully.
4019 */
4020 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4021 mem_cgroup_oom_synchronize(false);
4022 }
4023
4024 return ret;
4025}
4026EXPORT_SYMBOL_GPL(handle_mm_fault);
4027
4028#ifndef __PAGETABLE_P4D_FOLDED
4029/*
4030 * Allocate p4d page table.
4031 * We've already handled the fast-path in-line.
4032 */
4033int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4034{
4035 p4d_t *new = p4d_alloc_one(mm, address);
4036 if (!new)
4037 return -ENOMEM;
4038
4039 smp_wmb(); /* See comment in __pte_alloc */
4040
4041 spin_lock(&mm->page_table_lock);
4042 if (pgd_present(*pgd)) /* Another has populated it */
4043 p4d_free(mm, new);
4044 else
4045 pgd_populate(mm, pgd, new);
4046 spin_unlock(&mm->page_table_lock);
4047 return 0;
4048}
4049#endif /* __PAGETABLE_P4D_FOLDED */
4050
4051#ifndef __PAGETABLE_PUD_FOLDED
4052/*
4053 * Allocate page upper directory.
4054 * We've already handled the fast-path in-line.
4055 */
4056int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4057{
4058 pud_t *new = pud_alloc_one(mm, address);
4059 if (!new)
4060 return -ENOMEM;
4061
4062 smp_wmb(); /* See comment in __pte_alloc */
4063
4064 spin_lock(&mm->page_table_lock);
4065#ifndef __ARCH_HAS_5LEVEL_HACK
4066 if (!p4d_present(*p4d)) {
4067 mm_inc_nr_puds(mm);
4068 p4d_populate(mm, p4d, new);
4069 } else /* Another has populated it */
4070 pud_free(mm, new);
4071#else
4072 if (!pgd_present(*p4d)) {
4073 mm_inc_nr_puds(mm);
4074 pgd_populate(mm, p4d, new);
4075 } else /* Another has populated it */
4076 pud_free(mm, new);
4077#endif /* __ARCH_HAS_5LEVEL_HACK */
4078 spin_unlock(&mm->page_table_lock);
4079 return 0;
4080}
4081#endif /* __PAGETABLE_PUD_FOLDED */
4082
4083#ifndef __PAGETABLE_PMD_FOLDED
4084/*
4085 * Allocate page middle directory.
4086 * We've already handled the fast-path in-line.
4087 */
4088int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4089{
4090 spinlock_t *ptl;
4091 pmd_t *new = pmd_alloc_one(mm, address);
4092 if (!new)
4093 return -ENOMEM;
4094
4095 smp_wmb(); /* See comment in __pte_alloc */
4096
4097 ptl = pud_lock(mm, pud);
4098#ifndef __ARCH_HAS_4LEVEL_HACK
4099 if (!pud_present(*pud)) {
4100 mm_inc_nr_pmds(mm);
4101 pud_populate(mm, pud, new);
4102 } else /* Another has populated it */
4103 pmd_free(mm, new);
4104#else
4105 if (!pgd_present(*pud)) {
4106 mm_inc_nr_pmds(mm);
4107 pgd_populate(mm, pud, new);
4108 } else /* Another has populated it */
4109 pmd_free(mm, new);
4110#endif /* __ARCH_HAS_4LEVEL_HACK */
4111 spin_unlock(ptl);
4112 return 0;
4113}
4114#endif /* __PAGETABLE_PMD_FOLDED */
4115
4116static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4117 struct mmu_notifier_range *range,
4118 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4119{
4120 pgd_t *pgd;
4121 p4d_t *p4d;
4122 pud_t *pud;
4123 pmd_t *pmd;
4124 pte_t *ptep;
4125
4126 pgd = pgd_offset(mm, address);
4127 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4128 goto out;
4129
4130 p4d = p4d_offset(pgd, address);
4131 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4132 goto out;
4133
4134 pud = pud_offset(p4d, address);
4135 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4136 goto out;
4137
4138 pmd = pmd_offset(pud, address);
4139 VM_BUG_ON(pmd_trans_huge(*pmd));
4140
4141 if (pmd_huge(*pmd)) {
4142 if (!pmdpp)
4143 goto out;
4144
4145 if (range) {
4146 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4147 NULL, mm, address & PMD_MASK,
4148 (address & PMD_MASK) + PMD_SIZE);
4149 mmu_notifier_invalidate_range_start(range);
4150 }
4151 *ptlp = pmd_lock(mm, pmd);
4152 if (pmd_huge(*pmd)) {
4153 *pmdpp = pmd;
4154 return 0;
4155 }
4156 spin_unlock(*ptlp);
4157 if (range)
4158 mmu_notifier_invalidate_range_end(range);
4159 }
4160
4161 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4162 goto out;
4163
4164 if (range) {
4165 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4166 address & PAGE_MASK,
4167 (address & PAGE_MASK) + PAGE_SIZE);
4168 mmu_notifier_invalidate_range_start(range);
4169 }
4170 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4171 if (!pte_present(*ptep))
4172 goto unlock;
4173 *ptepp = ptep;
4174 return 0;
4175unlock:
4176 pte_unmap_unlock(ptep, *ptlp);
4177 if (range)
4178 mmu_notifier_invalidate_range_end(range);
4179out:
4180 return -EINVAL;
4181}
4182
4183static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4184 pte_t **ptepp, spinlock_t **ptlp)
4185{
4186 int res;
4187
4188 /* (void) is needed to make gcc happy */
4189 (void) __cond_lock(*ptlp,
4190 !(res = __follow_pte_pmd(mm, address, NULL,
4191 ptepp, NULL, ptlp)));
4192 return res;
4193}
4194
4195int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4196 struct mmu_notifier_range *range,
4197 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4198{
4199 int res;
4200
4201 /* (void) is needed to make gcc happy */
4202 (void) __cond_lock(*ptlp,
4203 !(res = __follow_pte_pmd(mm, address, range,
4204 ptepp, pmdpp, ptlp)));
4205 return res;
4206}
4207EXPORT_SYMBOL(follow_pte_pmd);
4208
4209/**
4210 * follow_pfn - look up PFN at a user virtual address
4211 * @vma: memory mapping
4212 * @address: user virtual address
4213 * @pfn: location to store found PFN
4214 *
4215 * Only IO mappings and raw PFN mappings are allowed.
4216 *
4217 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4218 */
4219int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4220 unsigned long *pfn)
4221{
4222 int ret = -EINVAL;
4223 spinlock_t *ptl;
4224 pte_t *ptep;
4225
4226 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4227 return ret;
4228
4229 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4230 if (ret)
4231 return ret;
4232 *pfn = pte_pfn(*ptep);
4233 pte_unmap_unlock(ptep, ptl);
4234 return 0;
4235}
4236EXPORT_SYMBOL(follow_pfn);
4237
4238#ifdef CONFIG_HAVE_IOREMAP_PROT
4239int follow_phys(struct vm_area_struct *vma,
4240 unsigned long address, unsigned int flags,
4241 unsigned long *prot, resource_size_t *phys)
4242{
4243 int ret = -EINVAL;
4244 pte_t *ptep, pte;
4245 spinlock_t *ptl;
4246
4247 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4248 goto out;
4249
4250 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4251 goto out;
4252 pte = *ptep;
4253
4254 if ((flags & FOLL_WRITE) && !pte_write(pte))
4255 goto unlock;
4256
4257 *prot = pgprot_val(pte_pgprot(pte));
4258 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4259
4260 ret = 0;
4261unlock:
4262 pte_unmap_unlock(ptep, ptl);
4263out:
4264 return ret;
4265}
4266
4267int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4268 void *buf, int len, int write)
4269{
4270 resource_size_t phys_addr;
4271 unsigned long prot = 0;
4272 void __iomem *maddr;
4273 int offset = addr & (PAGE_SIZE-1);
4274
4275 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4276 return -EINVAL;
4277
4278 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4279 if (!maddr)
4280 return -ENOMEM;
4281
4282 if (write)
4283 memcpy_toio(maddr + offset, buf, len);
4284 else
4285 memcpy_fromio(buf, maddr + offset, len);
4286 iounmap(maddr);
4287
4288 return len;
4289}
4290EXPORT_SYMBOL_GPL(generic_access_phys);
4291#endif
4292
4293/*
4294 * Access another process' address space as given in mm. If non-NULL, use the
4295 * given task for page fault accounting.
4296 */
4297int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4298 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4299{
4300 struct vm_area_struct *vma;
4301 void *old_buf = buf;
4302 int write = gup_flags & FOLL_WRITE;
4303
4304 if (down_read_killable(&mm->mmap_sem))
4305 return 0;
4306
4307 /* ignore errors, just check how much was successfully transferred */
4308 while (len) {
4309 int bytes, ret, offset;
4310 void *maddr;
4311 struct page *page = NULL;
4312
4313 ret = get_user_pages_remote(tsk, mm, addr, 1,
4314 gup_flags, &page, &vma, NULL);
4315 if (ret <= 0) {
4316#ifndef CONFIG_HAVE_IOREMAP_PROT
4317 break;
4318#else
4319 /*
4320 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4321 * we can access using slightly different code.
4322 */
4323 vma = find_vma(mm, addr);
4324 if (!vma || vma->vm_start > addr)
4325 break;
4326 if (vma->vm_ops && vma->vm_ops->access)
4327 ret = vma->vm_ops->access(vma, addr, buf,
4328 len, write);
4329 if (ret <= 0)
4330 break;
4331 bytes = ret;
4332#endif
4333 } else {
4334 bytes = len;
4335 offset = addr & (PAGE_SIZE-1);
4336 if (bytes > PAGE_SIZE-offset)
4337 bytes = PAGE_SIZE-offset;
4338
4339 maddr = kmap(page);
4340 if (write) {
4341 copy_to_user_page(vma, page, addr,
4342 maddr + offset, buf, bytes);
4343 set_page_dirty_lock(page);
4344 } else {
4345 copy_from_user_page(vma, page, addr,
4346 buf, maddr + offset, bytes);
4347 }
4348 kunmap(page);
4349 put_page(page);
4350 }
4351 len -= bytes;
4352 buf += bytes;
4353 addr += bytes;
4354 }
4355 up_read(&mm->mmap_sem);
4356
4357 return buf - old_buf;
4358}
4359
4360/**
4361 * access_remote_vm - access another process' address space
4362 * @mm: the mm_struct of the target address space
4363 * @addr: start address to access
4364 * @buf: source or destination buffer
4365 * @len: number of bytes to transfer
4366 * @gup_flags: flags modifying lookup behaviour
4367 *
4368 * The caller must hold a reference on @mm.
4369 *
4370 * Return: number of bytes copied from source to destination.
4371 */
4372int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4373 void *buf, int len, unsigned int gup_flags)
4374{
4375 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4376}
4377
4378/*
4379 * Access another process' address space.
4380 * Source/target buffer must be kernel space,
4381 * Do not walk the page table directly, use get_user_pages
4382 */
4383int access_process_vm(struct task_struct *tsk, unsigned long addr,
4384 void *buf, int len, unsigned int gup_flags)
4385{
4386 struct mm_struct *mm;
4387 int ret;
4388
4389 mm = get_task_mm(tsk);
4390 if (!mm)
4391 return 0;
4392
4393 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4394
4395 mmput(mm);
4396
4397 return ret;
4398}
4399EXPORT_SYMBOL_GPL(access_process_vm);
4400
4401/*
4402 * Print the name of a VMA.
4403 */
4404void print_vma_addr(char *prefix, unsigned long ip)
4405{
4406 struct mm_struct *mm = current->mm;
4407 struct vm_area_struct *vma;
4408
4409 /*
4410 * we might be running from an atomic context so we cannot sleep
4411 */
4412 if (!down_read_trylock(&mm->mmap_sem))
4413 return;
4414
4415 vma = find_vma(mm, ip);
4416 if (vma && vma->vm_file) {
4417 struct file *f = vma->vm_file;
4418 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4419 if (buf) {
4420 char *p;
4421
4422 p = file_path(f, buf, PAGE_SIZE);
4423 if (IS_ERR(p))
4424 p = "?";
4425 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4426 vma->vm_start,
4427 vma->vm_end - vma->vm_start);
4428 free_page((unsigned long)buf);
4429 }
4430 }
4431 up_read(&mm->mmap_sem);
4432}
4433
4434#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4435void __might_fault(const char *file, int line)
4436{
4437 /*
4438 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4439 * holding the mmap_sem, this is safe because kernel memory doesn't
4440 * get paged out, therefore we'll never actually fault, and the
4441 * below annotations will generate false positives.
4442 */
4443 if (uaccess_kernel())
4444 return;
4445 if (pagefault_disabled())
4446 return;
4447 __might_sleep(file, line, 0);
4448#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4449 if (current->mm)
4450 might_lock_read(¤t->mm->mmap_sem);
4451#endif
4452}
4453EXPORT_SYMBOL(__might_fault);
4454#endif
4455
4456#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4457/*
4458 * Process all subpages of the specified huge page with the specified
4459 * operation. The target subpage will be processed last to keep its
4460 * cache lines hot.
4461 */
4462static inline void process_huge_page(
4463 unsigned long addr_hint, unsigned int pages_per_huge_page,
4464 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4465 void *arg)
4466{
4467 int i, n, base, l;
4468 unsigned long addr = addr_hint &
4469 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4470
4471 /* Process target subpage last to keep its cache lines hot */
4472 might_sleep();
4473 n = (addr_hint - addr) / PAGE_SIZE;
4474 if (2 * n <= pages_per_huge_page) {
4475 /* If target subpage in first half of huge page */
4476 base = 0;
4477 l = n;
4478 /* Process subpages at the end of huge page */
4479 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4480 cond_resched();
4481 process_subpage(addr + i * PAGE_SIZE, i, arg);
4482 }
4483 } else {
4484 /* If target subpage in second half of huge page */
4485 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4486 l = pages_per_huge_page - n;
4487 /* Process subpages at the begin of huge page */
4488 for (i = 0; i < base; i++) {
4489 cond_resched();
4490 process_subpage(addr + i * PAGE_SIZE, i, arg);
4491 }
4492 }
4493 /*
4494 * Process remaining subpages in left-right-left-right pattern
4495 * towards the target subpage
4496 */
4497 for (i = 0; i < l; i++) {
4498 int left_idx = base + i;
4499 int right_idx = base + 2 * l - 1 - i;
4500
4501 cond_resched();
4502 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4503 cond_resched();
4504 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4505 }
4506}
4507
4508static void clear_gigantic_page(struct page *page,
4509 unsigned long addr,
4510 unsigned int pages_per_huge_page)
4511{
4512 int i;
4513 struct page *p = page;
4514
4515 might_sleep();
4516 for (i = 0; i < pages_per_huge_page;
4517 i++, p = mem_map_next(p, page, i)) {
4518 cond_resched();
4519 clear_user_highpage(p, addr + i * PAGE_SIZE);
4520 }
4521}
4522
4523static void clear_subpage(unsigned long addr, int idx, void *arg)
4524{
4525 struct page *page = arg;
4526
4527 clear_user_highpage(page + idx, addr);
4528}
4529
4530void clear_huge_page(struct page *page,
4531 unsigned long addr_hint, unsigned int pages_per_huge_page)
4532{
4533 unsigned long addr = addr_hint &
4534 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4535
4536 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4537 clear_gigantic_page(page, addr, pages_per_huge_page);
4538 return;
4539 }
4540
4541 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4542}
4543
4544static void copy_user_gigantic_page(struct page *dst, struct page *src,
4545 unsigned long addr,
4546 struct vm_area_struct *vma,
4547 unsigned int pages_per_huge_page)
4548{
4549 int i;
4550 struct page *dst_base = dst;
4551 struct page *src_base = src;
4552
4553 for (i = 0; i < pages_per_huge_page; ) {
4554 cond_resched();
4555 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4556
4557 i++;
4558 dst = mem_map_next(dst, dst_base, i);
4559 src = mem_map_next(src, src_base, i);
4560 }
4561}
4562
4563struct copy_subpage_arg {
4564 struct page *dst;
4565 struct page *src;
4566 struct vm_area_struct *vma;
4567};
4568
4569static void copy_subpage(unsigned long addr, int idx, void *arg)
4570{
4571 struct copy_subpage_arg *copy_arg = arg;
4572
4573 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4574 addr, copy_arg->vma);
4575}
4576
4577void copy_user_huge_page(struct page *dst, struct page *src,
4578 unsigned long addr_hint, struct vm_area_struct *vma,
4579 unsigned int pages_per_huge_page)
4580{
4581 unsigned long addr = addr_hint &
4582 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4583 struct copy_subpage_arg arg = {
4584 .dst = dst,
4585 .src = src,
4586 .vma = vma,
4587 };
4588
4589 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4590 copy_user_gigantic_page(dst, src, addr, vma,
4591 pages_per_huge_page);
4592 return;
4593 }
4594
4595 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4596}
4597
4598long copy_huge_page_from_user(struct page *dst_page,
4599 const void __user *usr_src,
4600 unsigned int pages_per_huge_page,
4601 bool allow_pagefault)
4602{
4603 void *src = (void *)usr_src;
4604 void *page_kaddr;
4605 unsigned long i, rc = 0;
4606 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4607
4608 for (i = 0; i < pages_per_huge_page; i++) {
4609 if (allow_pagefault)
4610 page_kaddr = kmap(dst_page + i);
4611 else
4612 page_kaddr = kmap_atomic(dst_page + i);
4613 rc = copy_from_user(page_kaddr,
4614 (const void __user *)(src + i * PAGE_SIZE),
4615 PAGE_SIZE);
4616 if (allow_pagefault)
4617 kunmap(dst_page + i);
4618 else
4619 kunmap_atomic(page_kaddr);
4620
4621 ret_val -= (PAGE_SIZE - rc);
4622 if (rc)
4623 break;
4624
4625 cond_resched();
4626 }
4627 return ret_val;
4628}
4629#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4630
4631#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4632
4633static struct kmem_cache *page_ptl_cachep;
4634
4635void __init ptlock_cache_init(void)
4636{
4637 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4638 SLAB_PANIC, NULL);
4639}
4640
4641bool ptlock_alloc(struct page *page)
4642{
4643 spinlock_t *ptl;
4644
4645 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4646 if (!ptl)
4647 return false;
4648 page->ptl = ptl;
4649 return true;
4650}
4651
4652void ptlock_free(struct page *page)
4653{
4654 kmem_cache_free(page_ptl_cachep, page->ptl);
4655}
4656#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 */