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#include <linux/perf_event.h>
75#include <linux/ptrace.h>
76#include <linux/vmalloc.h>
77
78#include <trace/events/kmem.h>
79
80#include <asm/io.h>
81#include <asm/mmu_context.h>
82#include <asm/pgalloc.h>
83#include <linux/uaccess.h>
84#include <asm/tlb.h>
85#include <asm/tlbflush.h>
86
87#include "pgalloc-track.h"
88#include "internal.h"
89
90#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
91#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
92#endif
93
94#ifndef CONFIG_NEED_MULTIPLE_NODES
95/* use the per-pgdat data instead for discontigmem - mbligh */
96unsigned long max_mapnr;
97EXPORT_SYMBOL(max_mapnr);
98
99struct page *mem_map;
100EXPORT_SYMBOL(mem_map);
101#endif
102
103/*
104 * A number of key systems in x86 including ioremap() rely on the assumption
105 * that high_memory defines the upper bound on direct map memory, then end
106 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
107 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
108 * and ZONE_HIGHMEM.
109 */
110void *high_memory;
111EXPORT_SYMBOL(high_memory);
112
113/*
114 * Randomize the address space (stacks, mmaps, brk, etc.).
115 *
116 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
117 * as ancient (libc5 based) binaries can segfault. )
118 */
119int randomize_va_space __read_mostly =
120#ifdef CONFIG_COMPAT_BRK
121 1;
122#else
123 2;
124#endif
125
126#ifndef arch_faults_on_old_pte
127static inline bool arch_faults_on_old_pte(void)
128{
129 /*
130 * Those arches which don't have hw access flag feature need to
131 * implement their own helper. By default, "true" means pagefault
132 * will be hit on old pte.
133 */
134 return true;
135}
136#endif
137
138static int __init disable_randmaps(char *s)
139{
140 randomize_va_space = 0;
141 return 1;
142}
143__setup("norandmaps", disable_randmaps);
144
145unsigned long zero_pfn __read_mostly;
146EXPORT_SYMBOL(zero_pfn);
147
148unsigned long highest_memmap_pfn __read_mostly;
149
150/*
151 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
152 */
153static int __init init_zero_pfn(void)
154{
155 zero_pfn = page_to_pfn(ZERO_PAGE(0));
156 return 0;
157}
158core_initcall(init_zero_pfn);
159
160void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
161{
162 trace_rss_stat(mm, member, count);
163}
164
165#if defined(SPLIT_RSS_COUNTING)
166
167void sync_mm_rss(struct mm_struct *mm)
168{
169 int i;
170
171 for (i = 0; i < NR_MM_COUNTERS; i++) {
172 if (current->rss_stat.count[i]) {
173 add_mm_counter(mm, i, current->rss_stat.count[i]);
174 current->rss_stat.count[i] = 0;
175 }
176 }
177 current->rss_stat.events = 0;
178}
179
180static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
181{
182 struct task_struct *task = current;
183
184 if (likely(task->mm == mm))
185 task->rss_stat.count[member] += val;
186 else
187 add_mm_counter(mm, member, val);
188}
189#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
190#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
191
192/* sync counter once per 64 page faults */
193#define TASK_RSS_EVENTS_THRESH (64)
194static void check_sync_rss_stat(struct task_struct *task)
195{
196 if (unlikely(task != current))
197 return;
198 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
199 sync_mm_rss(task->mm);
200}
201#else /* SPLIT_RSS_COUNTING */
202
203#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
204#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
205
206static void check_sync_rss_stat(struct task_struct *task)
207{
208}
209
210#endif /* SPLIT_RSS_COUNTING */
211
212/*
213 * Note: this doesn't free the actual pages themselves. That
214 * has been handled earlier when unmapping all the memory regions.
215 */
216static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
217 unsigned long addr)
218{
219 pgtable_t token = pmd_pgtable(*pmd);
220 pmd_clear(pmd);
221 pte_free_tlb(tlb, token, addr);
222 mm_dec_nr_ptes(tlb->mm);
223}
224
225static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
226 unsigned long addr, unsigned long end,
227 unsigned long floor, unsigned long ceiling)
228{
229 pmd_t *pmd;
230 unsigned long next;
231 unsigned long start;
232
233 start = addr;
234 pmd = pmd_offset(pud, addr);
235 do {
236 next = pmd_addr_end(addr, end);
237 if (pmd_none_or_clear_bad(pmd))
238 continue;
239 free_pte_range(tlb, pmd, addr);
240 } while (pmd++, addr = next, addr != end);
241
242 start &= PUD_MASK;
243 if (start < floor)
244 return;
245 if (ceiling) {
246 ceiling &= PUD_MASK;
247 if (!ceiling)
248 return;
249 }
250 if (end - 1 > ceiling - 1)
251 return;
252
253 pmd = pmd_offset(pud, start);
254 pud_clear(pud);
255 pmd_free_tlb(tlb, pmd, start);
256 mm_dec_nr_pmds(tlb->mm);
257}
258
259static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
260 unsigned long addr, unsigned long end,
261 unsigned long floor, unsigned long ceiling)
262{
263 pud_t *pud;
264 unsigned long next;
265 unsigned long start;
266
267 start = addr;
268 pud = pud_offset(p4d, addr);
269 do {
270 next = pud_addr_end(addr, end);
271 if (pud_none_or_clear_bad(pud))
272 continue;
273 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
274 } while (pud++, addr = next, addr != end);
275
276 start &= P4D_MASK;
277 if (start < floor)
278 return;
279 if (ceiling) {
280 ceiling &= P4D_MASK;
281 if (!ceiling)
282 return;
283 }
284 if (end - 1 > ceiling - 1)
285 return;
286
287 pud = pud_offset(p4d, start);
288 p4d_clear(p4d);
289 pud_free_tlb(tlb, pud, start);
290 mm_dec_nr_puds(tlb->mm);
291}
292
293static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
294 unsigned long addr, unsigned long end,
295 unsigned long floor, unsigned long ceiling)
296{
297 p4d_t *p4d;
298 unsigned long next;
299 unsigned long start;
300
301 start = addr;
302 p4d = p4d_offset(pgd, addr);
303 do {
304 next = p4d_addr_end(addr, end);
305 if (p4d_none_or_clear_bad(p4d))
306 continue;
307 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
308 } while (p4d++, addr = next, addr != end);
309
310 start &= PGDIR_MASK;
311 if (start < floor)
312 return;
313 if (ceiling) {
314 ceiling &= PGDIR_MASK;
315 if (!ceiling)
316 return;
317 }
318 if (end - 1 > ceiling - 1)
319 return;
320
321 p4d = p4d_offset(pgd, start);
322 pgd_clear(pgd);
323 p4d_free_tlb(tlb, p4d, start);
324}
325
326/*
327 * This function frees user-level page tables of a process.
328 */
329void free_pgd_range(struct mmu_gather *tlb,
330 unsigned long addr, unsigned long end,
331 unsigned long floor, unsigned long ceiling)
332{
333 pgd_t *pgd;
334 unsigned long next;
335
336 /*
337 * The next few lines have given us lots of grief...
338 *
339 * Why are we testing PMD* at this top level? Because often
340 * there will be no work to do at all, and we'd prefer not to
341 * go all the way down to the bottom just to discover that.
342 *
343 * Why all these "- 1"s? Because 0 represents both the bottom
344 * of the address space and the top of it (using -1 for the
345 * top wouldn't help much: the masks would do the wrong thing).
346 * The rule is that addr 0 and floor 0 refer to the bottom of
347 * the address space, but end 0 and ceiling 0 refer to the top
348 * Comparisons need to use "end - 1" and "ceiling - 1" (though
349 * that end 0 case should be mythical).
350 *
351 * Wherever addr is brought up or ceiling brought down, we must
352 * be careful to reject "the opposite 0" before it confuses the
353 * subsequent tests. But what about where end is brought down
354 * by PMD_SIZE below? no, end can't go down to 0 there.
355 *
356 * Whereas we round start (addr) and ceiling down, by different
357 * masks at different levels, in order to test whether a table
358 * now has no other vmas using it, so can be freed, we don't
359 * bother to round floor or end up - the tests don't need that.
360 */
361
362 addr &= PMD_MASK;
363 if (addr < floor) {
364 addr += PMD_SIZE;
365 if (!addr)
366 return;
367 }
368 if (ceiling) {
369 ceiling &= PMD_MASK;
370 if (!ceiling)
371 return;
372 }
373 if (end - 1 > ceiling - 1)
374 end -= PMD_SIZE;
375 if (addr > end - 1)
376 return;
377 /*
378 * We add page table cache pages with PAGE_SIZE,
379 * (see pte_free_tlb()), flush the tlb if we need
380 */
381 tlb_change_page_size(tlb, PAGE_SIZE);
382 pgd = pgd_offset(tlb->mm, addr);
383 do {
384 next = pgd_addr_end(addr, end);
385 if (pgd_none_or_clear_bad(pgd))
386 continue;
387 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
388 } while (pgd++, addr = next, addr != end);
389}
390
391void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
392 unsigned long floor, unsigned long ceiling)
393{
394 while (vma) {
395 struct vm_area_struct *next = vma->vm_next;
396 unsigned long addr = vma->vm_start;
397
398 /*
399 * Hide vma from rmap and truncate_pagecache before freeing
400 * pgtables
401 */
402 unlink_anon_vmas(vma);
403 unlink_file_vma(vma);
404
405 if (is_vm_hugetlb_page(vma)) {
406 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
407 floor, next ? next->vm_start : ceiling);
408 } else {
409 /*
410 * Optimization: gather nearby vmas into one call down
411 */
412 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
413 && !is_vm_hugetlb_page(next)) {
414 vma = next;
415 next = vma->vm_next;
416 unlink_anon_vmas(vma);
417 unlink_file_vma(vma);
418 }
419 free_pgd_range(tlb, addr, vma->vm_end,
420 floor, next ? next->vm_start : ceiling);
421 }
422 vma = next;
423 }
424}
425
426int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
427{
428 spinlock_t *ptl;
429 pgtable_t new = pte_alloc_one(mm);
430 if (!new)
431 return -ENOMEM;
432
433 /*
434 * Ensure all pte setup (eg. pte page lock and page clearing) are
435 * visible before the pte is made visible to other CPUs by being
436 * put into page tables.
437 *
438 * The other side of the story is the pointer chasing in the page
439 * table walking code (when walking the page table without locking;
440 * ie. most of the time). Fortunately, these data accesses consist
441 * of a chain of data-dependent loads, meaning most CPUs (alpha
442 * being the notable exception) will already guarantee loads are
443 * seen in-order. See the alpha page table accessors for the
444 * smp_rmb() barriers in page table walking code.
445 */
446 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
447
448 ptl = pmd_lock(mm, pmd);
449 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
450 mm_inc_nr_ptes(mm);
451 pmd_populate(mm, pmd, new);
452 new = NULL;
453 }
454 spin_unlock(ptl);
455 if (new)
456 pte_free(mm, new);
457 return 0;
458}
459
460int __pte_alloc_kernel(pmd_t *pmd)
461{
462 pte_t *new = pte_alloc_one_kernel(&init_mm);
463 if (!new)
464 return -ENOMEM;
465
466 smp_wmb(); /* See comment in __pte_alloc */
467
468 spin_lock(&init_mm.page_table_lock);
469 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
470 pmd_populate_kernel(&init_mm, pmd, new);
471 new = NULL;
472 }
473 spin_unlock(&init_mm.page_table_lock);
474 if (new)
475 pte_free_kernel(&init_mm, new);
476 return 0;
477}
478
479static inline void init_rss_vec(int *rss)
480{
481 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
482}
483
484static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
485{
486 int i;
487
488 if (current->mm == mm)
489 sync_mm_rss(mm);
490 for (i = 0; i < NR_MM_COUNTERS; i++)
491 if (rss[i])
492 add_mm_counter(mm, i, rss[i]);
493}
494
495/*
496 * This function is called to print an error when a bad pte
497 * is found. For example, we might have a PFN-mapped pte in
498 * a region that doesn't allow it.
499 *
500 * The calling function must still handle the error.
501 */
502static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
503 pte_t pte, struct page *page)
504{
505 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
506 p4d_t *p4d = p4d_offset(pgd, addr);
507 pud_t *pud = pud_offset(p4d, addr);
508 pmd_t *pmd = pmd_offset(pud, addr);
509 struct address_space *mapping;
510 pgoff_t index;
511 static unsigned long resume;
512 static unsigned long nr_shown;
513 static unsigned long nr_unshown;
514
515 /*
516 * Allow a burst of 60 reports, then keep quiet for that minute;
517 * or allow a steady drip of one report per second.
518 */
519 if (nr_shown == 60) {
520 if (time_before(jiffies, resume)) {
521 nr_unshown++;
522 return;
523 }
524 if (nr_unshown) {
525 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
526 nr_unshown);
527 nr_unshown = 0;
528 }
529 nr_shown = 0;
530 }
531 if (nr_shown++ == 0)
532 resume = jiffies + 60 * HZ;
533
534 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
535 index = linear_page_index(vma, addr);
536
537 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
538 current->comm,
539 (long long)pte_val(pte), (long long)pmd_val(*pmd));
540 if (page)
541 dump_page(page, "bad pte");
542 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
543 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
544 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
545 vma->vm_file,
546 vma->vm_ops ? vma->vm_ops->fault : NULL,
547 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
548 mapping ? mapping->a_ops->readpage : NULL);
549 dump_stack();
550 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
551}
552
553/*
554 * vm_normal_page -- This function gets the "struct page" associated with a pte.
555 *
556 * "Special" mappings do not wish to be associated with a "struct page" (either
557 * it doesn't exist, or it exists but they don't want to touch it). In this
558 * case, NULL is returned here. "Normal" mappings do have a struct page.
559 *
560 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
561 * pte bit, in which case this function is trivial. Secondly, an architecture
562 * may not have a spare pte bit, which requires a more complicated scheme,
563 * described below.
564 *
565 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
566 * special mapping (even if there are underlying and valid "struct pages").
567 * COWed pages of a VM_PFNMAP are always normal.
568 *
569 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
570 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
571 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
572 * mapping will always honor the rule
573 *
574 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
575 *
576 * And for normal mappings this is false.
577 *
578 * This restricts such mappings to be a linear translation from virtual address
579 * to pfn. To get around this restriction, we allow arbitrary mappings so long
580 * as the vma is not a COW mapping; in that case, we know that all ptes are
581 * special (because none can have been COWed).
582 *
583 *
584 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
585 *
586 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
587 * page" backing, however the difference is that _all_ pages with a struct
588 * page (that is, those where pfn_valid is true) are refcounted and considered
589 * normal pages by the VM. The disadvantage is that pages are refcounted
590 * (which can be slower and simply not an option for some PFNMAP users). The
591 * advantage is that we don't have to follow the strict linearity rule of
592 * PFNMAP mappings in order to support COWable mappings.
593 *
594 */
595struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
596 pte_t pte)
597{
598 unsigned long pfn = pte_pfn(pte);
599
600 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
601 if (likely(!pte_special(pte)))
602 goto check_pfn;
603 if (vma->vm_ops && vma->vm_ops->find_special_page)
604 return vma->vm_ops->find_special_page(vma, addr);
605 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
606 return NULL;
607 if (is_zero_pfn(pfn))
608 return NULL;
609 if (pte_devmap(pte))
610 return NULL;
611
612 print_bad_pte(vma, addr, pte, NULL);
613 return NULL;
614 }
615
616 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
617
618 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
619 if (vma->vm_flags & VM_MIXEDMAP) {
620 if (!pfn_valid(pfn))
621 return NULL;
622 goto out;
623 } else {
624 unsigned long off;
625 off = (addr - vma->vm_start) >> PAGE_SHIFT;
626 if (pfn == vma->vm_pgoff + off)
627 return NULL;
628 if (!is_cow_mapping(vma->vm_flags))
629 return NULL;
630 }
631 }
632
633 if (is_zero_pfn(pfn))
634 return NULL;
635
636check_pfn:
637 if (unlikely(pfn > highest_memmap_pfn)) {
638 print_bad_pte(vma, addr, pte, NULL);
639 return NULL;
640 }
641
642 /*
643 * NOTE! We still have PageReserved() pages in the page tables.
644 * eg. VDSO mappings can cause them to exist.
645 */
646out:
647 return pfn_to_page(pfn);
648}
649
650#ifdef CONFIG_TRANSPARENT_HUGEPAGE
651struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
652 pmd_t pmd)
653{
654 unsigned long pfn = pmd_pfn(pmd);
655
656 /*
657 * There is no pmd_special() but there may be special pmds, e.g.
658 * in a direct-access (dax) mapping, so let's just replicate the
659 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
660 */
661 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
662 if (vma->vm_flags & VM_MIXEDMAP) {
663 if (!pfn_valid(pfn))
664 return NULL;
665 goto out;
666 } else {
667 unsigned long off;
668 off = (addr - vma->vm_start) >> PAGE_SHIFT;
669 if (pfn == vma->vm_pgoff + off)
670 return NULL;
671 if (!is_cow_mapping(vma->vm_flags))
672 return NULL;
673 }
674 }
675
676 if (pmd_devmap(pmd))
677 return NULL;
678 if (is_huge_zero_pmd(pmd))
679 return NULL;
680 if (unlikely(pfn > highest_memmap_pfn))
681 return NULL;
682
683 /*
684 * NOTE! We still have PageReserved() pages in the page tables.
685 * eg. VDSO mappings can cause them to exist.
686 */
687out:
688 return pfn_to_page(pfn);
689}
690#endif
691
692/*
693 * copy one vm_area from one task to the other. Assumes the page tables
694 * already present in the new task to be cleared in the whole range
695 * covered by this vma.
696 */
697
698static unsigned long
699copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
700 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
701 unsigned long addr, int *rss)
702{
703 unsigned long vm_flags = vma->vm_flags;
704 pte_t pte = *src_pte;
705 struct page *page;
706 swp_entry_t entry = pte_to_swp_entry(pte);
707
708 if (likely(!non_swap_entry(entry))) {
709 if (swap_duplicate(entry) < 0)
710 return entry.val;
711
712 /* make sure dst_mm is on swapoff's mmlist. */
713 if (unlikely(list_empty(&dst_mm->mmlist))) {
714 spin_lock(&mmlist_lock);
715 if (list_empty(&dst_mm->mmlist))
716 list_add(&dst_mm->mmlist,
717 &src_mm->mmlist);
718 spin_unlock(&mmlist_lock);
719 }
720 rss[MM_SWAPENTS]++;
721 } else if (is_migration_entry(entry)) {
722 page = migration_entry_to_page(entry);
723
724 rss[mm_counter(page)]++;
725
726 if (is_write_migration_entry(entry) &&
727 is_cow_mapping(vm_flags)) {
728 /*
729 * COW mappings require pages in both
730 * parent and child to be set to read.
731 */
732 make_migration_entry_read(&entry);
733 pte = swp_entry_to_pte(entry);
734 if (pte_swp_soft_dirty(*src_pte))
735 pte = pte_swp_mksoft_dirty(pte);
736 if (pte_swp_uffd_wp(*src_pte))
737 pte = pte_swp_mkuffd_wp(pte);
738 set_pte_at(src_mm, addr, src_pte, pte);
739 }
740 } else if (is_device_private_entry(entry)) {
741 page = device_private_entry_to_page(entry);
742
743 /*
744 * Update rss count even for unaddressable pages, as
745 * they should treated just like normal pages in this
746 * respect.
747 *
748 * We will likely want to have some new rss counters
749 * for unaddressable pages, at some point. But for now
750 * keep things as they are.
751 */
752 get_page(page);
753 rss[mm_counter(page)]++;
754 page_dup_rmap(page, false);
755
756 /*
757 * We do not preserve soft-dirty information, because so
758 * far, checkpoint/restore is the only feature that
759 * requires that. And checkpoint/restore does not work
760 * when a device driver is involved (you cannot easily
761 * save and restore device driver state).
762 */
763 if (is_write_device_private_entry(entry) &&
764 is_cow_mapping(vm_flags)) {
765 make_device_private_entry_read(&entry);
766 pte = swp_entry_to_pte(entry);
767 if (pte_swp_uffd_wp(*src_pte))
768 pte = pte_swp_mkuffd_wp(pte);
769 set_pte_at(src_mm, addr, src_pte, pte);
770 }
771 }
772 set_pte_at(dst_mm, addr, dst_pte, pte);
773 return 0;
774}
775
776/*
777 * Copy a present and normal page if necessary.
778 *
779 * NOTE! The usual case is that this doesn't need to do
780 * anything, and can just return a positive value. That
781 * will let the caller know that it can just increase
782 * the page refcount and re-use the pte the traditional
783 * way.
784 *
785 * But _if_ we need to copy it because it needs to be
786 * pinned in the parent (and the child should get its own
787 * copy rather than just a reference to the same page),
788 * we'll do that here and return zero to let the caller
789 * know we're done.
790 *
791 * And if we need a pre-allocated page but don't yet have
792 * one, return a negative error to let the preallocation
793 * code know so that it can do so outside the page table
794 * lock.
795 */
796static inline int
797copy_present_page(struct mm_struct *dst_mm, struct mm_struct *src_mm,
798 pte_t *dst_pte, pte_t *src_pte,
799 struct vm_area_struct *vma, struct vm_area_struct *new,
800 unsigned long addr, int *rss, struct page **prealloc,
801 pte_t pte, struct page *page)
802{
803 struct page *new_page;
804
805 if (!is_cow_mapping(vma->vm_flags))
806 return 1;
807
808 /*
809 * What we want to do is to check whether this page may
810 * have been pinned by the parent process. If so,
811 * instead of wrprotect the pte on both sides, we copy
812 * the page immediately so that we'll always guarantee
813 * the pinned page won't be randomly replaced in the
814 * future.
815 *
816 * The page pinning checks are just "has this mm ever
817 * seen pinning", along with the (inexact) check of
818 * the page count. That might give false positives for
819 * for pinning, but it will work correctly.
820 */
821 if (likely(!atomic_read(&src_mm->has_pinned)))
822 return 1;
823 if (likely(!page_maybe_dma_pinned(page)))
824 return 1;
825
826 new_page = *prealloc;
827 if (!new_page)
828 return -EAGAIN;
829
830 /*
831 * We have a prealloc page, all good! Take it
832 * over and copy the page & arm it.
833 */
834 *prealloc = NULL;
835 copy_user_highpage(new_page, page, addr, vma);
836 __SetPageUptodate(new_page);
837 page_add_new_anon_rmap(new_page, new, addr, false);
838 lru_cache_add_inactive_or_unevictable(new_page, new);
839 rss[mm_counter(new_page)]++;
840
841 /* All done, just insert the new page copy in the child */
842 pte = mk_pte(new_page, new->vm_page_prot);
843 pte = maybe_mkwrite(pte_mkdirty(pte), new);
844 set_pte_at(dst_mm, addr, dst_pte, pte);
845 return 0;
846}
847
848/*
849 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
850 * is required to copy this pte.
851 */
852static inline int
853copy_present_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
854 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
855 struct vm_area_struct *new,
856 unsigned long addr, int *rss, struct page **prealloc)
857{
858 unsigned long vm_flags = vma->vm_flags;
859 pte_t pte = *src_pte;
860 struct page *page;
861
862 page = vm_normal_page(vma, addr, pte);
863 if (page) {
864 int retval;
865
866 retval = copy_present_page(dst_mm, src_mm,
867 dst_pte, src_pte,
868 vma, new,
869 addr, rss, prealloc,
870 pte, page);
871 if (retval <= 0)
872 return retval;
873
874 get_page(page);
875 page_dup_rmap(page, false);
876 rss[mm_counter(page)]++;
877 }
878
879 /*
880 * If it's a COW mapping, write protect it both
881 * in the parent and the child
882 */
883 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
884 ptep_set_wrprotect(src_mm, addr, src_pte);
885 pte = pte_wrprotect(pte);
886 }
887
888 /*
889 * If it's a shared mapping, mark it clean in
890 * the child
891 */
892 if (vm_flags & VM_SHARED)
893 pte = pte_mkclean(pte);
894 pte = pte_mkold(pte);
895
896 /*
897 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
898 * does not have the VM_UFFD_WP, which means that the uffd
899 * fork event is not enabled.
900 */
901 if (!(vm_flags & VM_UFFD_WP))
902 pte = pte_clear_uffd_wp(pte);
903
904 set_pte_at(dst_mm, addr, dst_pte, pte);
905 return 0;
906}
907
908static inline struct page *
909page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
910 unsigned long addr)
911{
912 struct page *new_page;
913
914 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
915 if (!new_page)
916 return NULL;
917
918 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
919 put_page(new_page);
920 return NULL;
921 }
922 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
923
924 return new_page;
925}
926
927static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
928 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
929 struct vm_area_struct *new,
930 unsigned long addr, unsigned long end)
931{
932 pte_t *orig_src_pte, *orig_dst_pte;
933 pte_t *src_pte, *dst_pte;
934 spinlock_t *src_ptl, *dst_ptl;
935 int progress, ret = 0;
936 int rss[NR_MM_COUNTERS];
937 swp_entry_t entry = (swp_entry_t){0};
938 struct page *prealloc = NULL;
939
940again:
941 progress = 0;
942 init_rss_vec(rss);
943
944 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
945 if (!dst_pte) {
946 ret = -ENOMEM;
947 goto out;
948 }
949 src_pte = pte_offset_map(src_pmd, addr);
950 src_ptl = pte_lockptr(src_mm, src_pmd);
951 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
952 orig_src_pte = src_pte;
953 orig_dst_pte = dst_pte;
954 arch_enter_lazy_mmu_mode();
955
956 do {
957 /*
958 * We are holding two locks at this point - either of them
959 * could generate latencies in another task on another CPU.
960 */
961 if (progress >= 32) {
962 progress = 0;
963 if (need_resched() ||
964 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
965 break;
966 }
967 if (pte_none(*src_pte)) {
968 progress++;
969 continue;
970 }
971 if (unlikely(!pte_present(*src_pte))) {
972 entry.val = copy_nonpresent_pte(dst_mm, src_mm,
973 dst_pte, src_pte,
974 vma, addr, rss);
975 if (entry.val)
976 break;
977 progress += 8;
978 continue;
979 }
980 /* copy_present_pte() will clear `*prealloc' if consumed */
981 ret = copy_present_pte(dst_mm, src_mm, dst_pte, src_pte,
982 vma, new, addr, rss, &prealloc);
983 /*
984 * If we need a pre-allocated page for this pte, drop the
985 * locks, allocate, and try again.
986 */
987 if (unlikely(ret == -EAGAIN))
988 break;
989 if (unlikely(prealloc)) {
990 /*
991 * pre-alloc page cannot be reused by next time so as
992 * to strictly follow mempolicy (e.g., alloc_page_vma()
993 * will allocate page according to address). This
994 * could only happen if one pinned pte changed.
995 */
996 put_page(prealloc);
997 prealloc = NULL;
998 }
999 progress += 8;
1000 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1001
1002 arch_leave_lazy_mmu_mode();
1003 spin_unlock(src_ptl);
1004 pte_unmap(orig_src_pte);
1005 add_mm_rss_vec(dst_mm, rss);
1006 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1007 cond_resched();
1008
1009 if (entry.val) {
1010 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1011 ret = -ENOMEM;
1012 goto out;
1013 }
1014 entry.val = 0;
1015 } else if (ret) {
1016 WARN_ON_ONCE(ret != -EAGAIN);
1017 prealloc = page_copy_prealloc(src_mm, vma, addr);
1018 if (!prealloc)
1019 return -ENOMEM;
1020 /* We've captured and resolved the error. Reset, try again. */
1021 ret = 0;
1022 }
1023 if (addr != end)
1024 goto again;
1025out:
1026 if (unlikely(prealloc))
1027 put_page(prealloc);
1028 return ret;
1029}
1030
1031static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1032 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1033 struct vm_area_struct *new,
1034 unsigned long addr, unsigned long end)
1035{
1036 pmd_t *src_pmd, *dst_pmd;
1037 unsigned long next;
1038
1039 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1040 if (!dst_pmd)
1041 return -ENOMEM;
1042 src_pmd = pmd_offset(src_pud, addr);
1043 do {
1044 next = pmd_addr_end(addr, end);
1045 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1046 || pmd_devmap(*src_pmd)) {
1047 int err;
1048 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1049 err = copy_huge_pmd(dst_mm, src_mm,
1050 dst_pmd, src_pmd, addr, vma);
1051 if (err == -ENOMEM)
1052 return -ENOMEM;
1053 if (!err)
1054 continue;
1055 /* fall through */
1056 }
1057 if (pmd_none_or_clear_bad(src_pmd))
1058 continue;
1059 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1060 vma, new, addr, next))
1061 return -ENOMEM;
1062 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1063 return 0;
1064}
1065
1066static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1067 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1068 struct vm_area_struct *new,
1069 unsigned long addr, unsigned long end)
1070{
1071 pud_t *src_pud, *dst_pud;
1072 unsigned long next;
1073
1074 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1075 if (!dst_pud)
1076 return -ENOMEM;
1077 src_pud = pud_offset(src_p4d, addr);
1078 do {
1079 next = pud_addr_end(addr, end);
1080 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1081 int err;
1082
1083 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1084 err = copy_huge_pud(dst_mm, src_mm,
1085 dst_pud, src_pud, addr, vma);
1086 if (err == -ENOMEM)
1087 return -ENOMEM;
1088 if (!err)
1089 continue;
1090 /* fall through */
1091 }
1092 if (pud_none_or_clear_bad(src_pud))
1093 continue;
1094 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1095 vma, new, addr, next))
1096 return -ENOMEM;
1097 } while (dst_pud++, src_pud++, addr = next, addr != end);
1098 return 0;
1099}
1100
1101static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1102 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1103 struct vm_area_struct *new,
1104 unsigned long addr, unsigned long end)
1105{
1106 p4d_t *src_p4d, *dst_p4d;
1107 unsigned long next;
1108
1109 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1110 if (!dst_p4d)
1111 return -ENOMEM;
1112 src_p4d = p4d_offset(src_pgd, addr);
1113 do {
1114 next = p4d_addr_end(addr, end);
1115 if (p4d_none_or_clear_bad(src_p4d))
1116 continue;
1117 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1118 vma, new, addr, next))
1119 return -ENOMEM;
1120 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1121 return 0;
1122}
1123
1124int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1125 struct vm_area_struct *vma, struct vm_area_struct *new)
1126{
1127 pgd_t *src_pgd, *dst_pgd;
1128 unsigned long next;
1129 unsigned long addr = vma->vm_start;
1130 unsigned long end = vma->vm_end;
1131 struct mmu_notifier_range range;
1132 bool is_cow;
1133 int ret;
1134
1135 /*
1136 * Don't copy ptes where a page fault will fill them correctly.
1137 * Fork becomes much lighter when there are big shared or private
1138 * readonly mappings. The tradeoff is that copy_page_range is more
1139 * efficient than faulting.
1140 */
1141 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1142 !vma->anon_vma)
1143 return 0;
1144
1145 if (is_vm_hugetlb_page(vma))
1146 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1147
1148 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1149 /*
1150 * We do not free on error cases below as remove_vma
1151 * gets called on error from higher level routine
1152 */
1153 ret = track_pfn_copy(vma);
1154 if (ret)
1155 return ret;
1156 }
1157
1158 /*
1159 * We need to invalidate the secondary MMU mappings only when
1160 * there could be a permission downgrade on the ptes of the
1161 * parent mm. And a permission downgrade will only happen if
1162 * is_cow_mapping() returns true.
1163 */
1164 is_cow = is_cow_mapping(vma->vm_flags);
1165
1166 if (is_cow) {
1167 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1168 0, vma, src_mm, addr, end);
1169 mmu_notifier_invalidate_range_start(&range);
1170 }
1171
1172 ret = 0;
1173 dst_pgd = pgd_offset(dst_mm, addr);
1174 src_pgd = pgd_offset(src_mm, addr);
1175 do {
1176 next = pgd_addr_end(addr, end);
1177 if (pgd_none_or_clear_bad(src_pgd))
1178 continue;
1179 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1180 vma, new, addr, next))) {
1181 ret = -ENOMEM;
1182 break;
1183 }
1184 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1185
1186 if (is_cow)
1187 mmu_notifier_invalidate_range_end(&range);
1188 return ret;
1189}
1190
1191static unsigned long zap_pte_range(struct mmu_gather *tlb,
1192 struct vm_area_struct *vma, pmd_t *pmd,
1193 unsigned long addr, unsigned long end,
1194 struct zap_details *details)
1195{
1196 struct mm_struct *mm = tlb->mm;
1197 int force_flush = 0;
1198 int rss[NR_MM_COUNTERS];
1199 spinlock_t *ptl;
1200 pte_t *start_pte;
1201 pte_t *pte;
1202 swp_entry_t entry;
1203
1204 tlb_change_page_size(tlb, PAGE_SIZE);
1205again:
1206 init_rss_vec(rss);
1207 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1208 pte = start_pte;
1209 flush_tlb_batched_pending(mm);
1210 arch_enter_lazy_mmu_mode();
1211 do {
1212 pte_t ptent = *pte;
1213 if (pte_none(ptent))
1214 continue;
1215
1216 if (need_resched())
1217 break;
1218
1219 if (pte_present(ptent)) {
1220 struct page *page;
1221
1222 page = vm_normal_page(vma, addr, ptent);
1223 if (unlikely(details) && page) {
1224 /*
1225 * unmap_shared_mapping_pages() wants to
1226 * invalidate cache without truncating:
1227 * unmap shared but keep private pages.
1228 */
1229 if (details->check_mapping &&
1230 details->check_mapping != page_rmapping(page))
1231 continue;
1232 }
1233 ptent = ptep_get_and_clear_full(mm, addr, pte,
1234 tlb->fullmm);
1235 tlb_remove_tlb_entry(tlb, pte, addr);
1236 if (unlikely(!page))
1237 continue;
1238
1239 if (!PageAnon(page)) {
1240 if (pte_dirty(ptent)) {
1241 force_flush = 1;
1242 set_page_dirty(page);
1243 }
1244 if (pte_young(ptent) &&
1245 likely(!(vma->vm_flags & VM_SEQ_READ)))
1246 mark_page_accessed(page);
1247 }
1248 rss[mm_counter(page)]--;
1249 page_remove_rmap(page, false);
1250 if (unlikely(page_mapcount(page) < 0))
1251 print_bad_pte(vma, addr, ptent, page);
1252 if (unlikely(__tlb_remove_page(tlb, page))) {
1253 force_flush = 1;
1254 addr += PAGE_SIZE;
1255 break;
1256 }
1257 continue;
1258 }
1259
1260 entry = pte_to_swp_entry(ptent);
1261 if (is_device_private_entry(entry)) {
1262 struct page *page = device_private_entry_to_page(entry);
1263
1264 if (unlikely(details && details->check_mapping)) {
1265 /*
1266 * unmap_shared_mapping_pages() wants to
1267 * invalidate cache without truncating:
1268 * unmap shared but keep private pages.
1269 */
1270 if (details->check_mapping !=
1271 page_rmapping(page))
1272 continue;
1273 }
1274
1275 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1276 rss[mm_counter(page)]--;
1277 page_remove_rmap(page, false);
1278 put_page(page);
1279 continue;
1280 }
1281
1282 /* If details->check_mapping, we leave swap entries. */
1283 if (unlikely(details))
1284 continue;
1285
1286 if (!non_swap_entry(entry))
1287 rss[MM_SWAPENTS]--;
1288 else if (is_migration_entry(entry)) {
1289 struct page *page;
1290
1291 page = migration_entry_to_page(entry);
1292 rss[mm_counter(page)]--;
1293 }
1294 if (unlikely(!free_swap_and_cache(entry)))
1295 print_bad_pte(vma, addr, ptent, NULL);
1296 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1297 } while (pte++, addr += PAGE_SIZE, addr != end);
1298
1299 add_mm_rss_vec(mm, rss);
1300 arch_leave_lazy_mmu_mode();
1301
1302 /* Do the actual TLB flush before dropping ptl */
1303 if (force_flush)
1304 tlb_flush_mmu_tlbonly(tlb);
1305 pte_unmap_unlock(start_pte, ptl);
1306
1307 /*
1308 * If we forced a TLB flush (either due to running out of
1309 * batch buffers or because we needed to flush dirty TLB
1310 * entries before releasing the ptl), free the batched
1311 * memory too. Restart if we didn't do everything.
1312 */
1313 if (force_flush) {
1314 force_flush = 0;
1315 tlb_flush_mmu(tlb);
1316 }
1317
1318 if (addr != end) {
1319 cond_resched();
1320 goto again;
1321 }
1322
1323 return addr;
1324}
1325
1326static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1327 struct vm_area_struct *vma, pud_t *pud,
1328 unsigned long addr, unsigned long end,
1329 struct zap_details *details)
1330{
1331 pmd_t *pmd;
1332 unsigned long next;
1333
1334 pmd = pmd_offset(pud, addr);
1335 do {
1336 next = pmd_addr_end(addr, end);
1337 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1338 if (next - addr != HPAGE_PMD_SIZE)
1339 __split_huge_pmd(vma, pmd, addr, false, NULL);
1340 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1341 goto next;
1342 /* fall through */
1343 }
1344 /*
1345 * Here there can be other concurrent MADV_DONTNEED or
1346 * trans huge page faults running, and if the pmd is
1347 * none or trans huge it can change under us. This is
1348 * because MADV_DONTNEED holds the mmap_lock in read
1349 * mode.
1350 */
1351 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1352 goto next;
1353 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1354next:
1355 cond_resched();
1356 } while (pmd++, addr = next, addr != end);
1357
1358 return addr;
1359}
1360
1361static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1362 struct vm_area_struct *vma, p4d_t *p4d,
1363 unsigned long addr, unsigned long end,
1364 struct zap_details *details)
1365{
1366 pud_t *pud;
1367 unsigned long next;
1368
1369 pud = pud_offset(p4d, addr);
1370 do {
1371 next = pud_addr_end(addr, end);
1372 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1373 if (next - addr != HPAGE_PUD_SIZE) {
1374 mmap_assert_locked(tlb->mm);
1375 split_huge_pud(vma, pud, addr);
1376 } else if (zap_huge_pud(tlb, vma, pud, addr))
1377 goto next;
1378 /* fall through */
1379 }
1380 if (pud_none_or_clear_bad(pud))
1381 continue;
1382 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1383next:
1384 cond_resched();
1385 } while (pud++, addr = next, addr != end);
1386
1387 return addr;
1388}
1389
1390static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1391 struct vm_area_struct *vma, pgd_t *pgd,
1392 unsigned long addr, unsigned long end,
1393 struct zap_details *details)
1394{
1395 p4d_t *p4d;
1396 unsigned long next;
1397
1398 p4d = p4d_offset(pgd, addr);
1399 do {
1400 next = p4d_addr_end(addr, end);
1401 if (p4d_none_or_clear_bad(p4d))
1402 continue;
1403 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1404 } while (p4d++, addr = next, addr != end);
1405
1406 return addr;
1407}
1408
1409void unmap_page_range(struct mmu_gather *tlb,
1410 struct vm_area_struct *vma,
1411 unsigned long addr, unsigned long end,
1412 struct zap_details *details)
1413{
1414 pgd_t *pgd;
1415 unsigned long next;
1416
1417 BUG_ON(addr >= end);
1418 tlb_start_vma(tlb, vma);
1419 pgd = pgd_offset(vma->vm_mm, addr);
1420 do {
1421 next = pgd_addr_end(addr, end);
1422 if (pgd_none_or_clear_bad(pgd))
1423 continue;
1424 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1425 } while (pgd++, addr = next, addr != end);
1426 tlb_end_vma(tlb, vma);
1427}
1428
1429
1430static void unmap_single_vma(struct mmu_gather *tlb,
1431 struct vm_area_struct *vma, unsigned long start_addr,
1432 unsigned long end_addr,
1433 struct zap_details *details)
1434{
1435 unsigned long start = max(vma->vm_start, start_addr);
1436 unsigned long end;
1437
1438 if (start >= vma->vm_end)
1439 return;
1440 end = min(vma->vm_end, end_addr);
1441 if (end <= vma->vm_start)
1442 return;
1443
1444 if (vma->vm_file)
1445 uprobe_munmap(vma, start, end);
1446
1447 if (unlikely(vma->vm_flags & VM_PFNMAP))
1448 untrack_pfn(vma, 0, 0);
1449
1450 if (start != end) {
1451 if (unlikely(is_vm_hugetlb_page(vma))) {
1452 /*
1453 * It is undesirable to test vma->vm_file as it
1454 * should be non-null for valid hugetlb area.
1455 * However, vm_file will be NULL in the error
1456 * cleanup path of mmap_region. When
1457 * hugetlbfs ->mmap method fails,
1458 * mmap_region() nullifies vma->vm_file
1459 * before calling this function to clean up.
1460 * Since no pte has actually been setup, it is
1461 * safe to do nothing in this case.
1462 */
1463 if (vma->vm_file) {
1464 i_mmap_lock_write(vma->vm_file->f_mapping);
1465 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1466 i_mmap_unlock_write(vma->vm_file->f_mapping);
1467 }
1468 } else
1469 unmap_page_range(tlb, vma, start, end, details);
1470 }
1471}
1472
1473/**
1474 * unmap_vmas - unmap a range of memory covered by a list of vma's
1475 * @tlb: address of the caller's struct mmu_gather
1476 * @vma: the starting vma
1477 * @start_addr: virtual address at which to start unmapping
1478 * @end_addr: virtual address at which to end unmapping
1479 *
1480 * Unmap all pages in the vma list.
1481 *
1482 * Only addresses between `start' and `end' will be unmapped.
1483 *
1484 * The VMA list must be sorted in ascending virtual address order.
1485 *
1486 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1487 * range after unmap_vmas() returns. So the only responsibility here is to
1488 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1489 * drops the lock and schedules.
1490 */
1491void unmap_vmas(struct mmu_gather *tlb,
1492 struct vm_area_struct *vma, unsigned long start_addr,
1493 unsigned long end_addr)
1494{
1495 struct mmu_notifier_range range;
1496
1497 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1498 start_addr, end_addr);
1499 mmu_notifier_invalidate_range_start(&range);
1500 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1501 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1502 mmu_notifier_invalidate_range_end(&range);
1503}
1504
1505/**
1506 * zap_page_range - remove user pages in a given range
1507 * @vma: vm_area_struct holding the applicable pages
1508 * @start: starting address of pages to zap
1509 * @size: number of bytes to zap
1510 *
1511 * Caller must protect the VMA list
1512 */
1513void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1514 unsigned long size)
1515{
1516 struct mmu_notifier_range range;
1517 struct mmu_gather tlb;
1518
1519 lru_add_drain();
1520 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1521 start, start + size);
1522 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1523 update_hiwater_rss(vma->vm_mm);
1524 mmu_notifier_invalidate_range_start(&range);
1525 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1526 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1527 mmu_notifier_invalidate_range_end(&range);
1528 tlb_finish_mmu(&tlb, start, range.end);
1529}
1530
1531/**
1532 * zap_page_range_single - remove user pages in a given range
1533 * @vma: vm_area_struct holding the applicable pages
1534 * @address: starting address of pages to zap
1535 * @size: number of bytes to zap
1536 * @details: details of shared cache invalidation
1537 *
1538 * The range must fit into one VMA.
1539 */
1540static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1541 unsigned long size, struct zap_details *details)
1542{
1543 struct mmu_notifier_range range;
1544 struct mmu_gather tlb;
1545
1546 lru_add_drain();
1547 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1548 address, address + size);
1549 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1550 update_hiwater_rss(vma->vm_mm);
1551 mmu_notifier_invalidate_range_start(&range);
1552 unmap_single_vma(&tlb, vma, address, range.end, details);
1553 mmu_notifier_invalidate_range_end(&range);
1554 tlb_finish_mmu(&tlb, address, range.end);
1555}
1556
1557/**
1558 * zap_vma_ptes - remove ptes mapping the vma
1559 * @vma: vm_area_struct holding ptes to be zapped
1560 * @address: starting address of pages to zap
1561 * @size: number of bytes to zap
1562 *
1563 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1564 *
1565 * The entire address range must be fully contained within the vma.
1566 *
1567 */
1568void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1569 unsigned long size)
1570{
1571 if (address < vma->vm_start || address + size > vma->vm_end ||
1572 !(vma->vm_flags & VM_PFNMAP))
1573 return;
1574
1575 zap_page_range_single(vma, address, size, NULL);
1576}
1577EXPORT_SYMBOL_GPL(zap_vma_ptes);
1578
1579static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1580{
1581 pgd_t *pgd;
1582 p4d_t *p4d;
1583 pud_t *pud;
1584 pmd_t *pmd;
1585
1586 pgd = pgd_offset(mm, addr);
1587 p4d = p4d_alloc(mm, pgd, addr);
1588 if (!p4d)
1589 return NULL;
1590 pud = pud_alloc(mm, p4d, addr);
1591 if (!pud)
1592 return NULL;
1593 pmd = pmd_alloc(mm, pud, addr);
1594 if (!pmd)
1595 return NULL;
1596
1597 VM_BUG_ON(pmd_trans_huge(*pmd));
1598 return pmd;
1599}
1600
1601pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1602 spinlock_t **ptl)
1603{
1604 pmd_t *pmd = walk_to_pmd(mm, addr);
1605
1606 if (!pmd)
1607 return NULL;
1608 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1609}
1610
1611static int validate_page_before_insert(struct page *page)
1612{
1613 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1614 return -EINVAL;
1615 flush_dcache_page(page);
1616 return 0;
1617}
1618
1619static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1620 unsigned long addr, struct page *page, pgprot_t prot)
1621{
1622 if (!pte_none(*pte))
1623 return -EBUSY;
1624 /* Ok, finally just insert the thing.. */
1625 get_page(page);
1626 inc_mm_counter_fast(mm, mm_counter_file(page));
1627 page_add_file_rmap(page, false);
1628 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1629 return 0;
1630}
1631
1632/*
1633 * This is the old fallback for page remapping.
1634 *
1635 * For historical reasons, it only allows reserved pages. Only
1636 * old drivers should use this, and they needed to mark their
1637 * pages reserved for the old functions anyway.
1638 */
1639static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1640 struct page *page, pgprot_t prot)
1641{
1642 struct mm_struct *mm = vma->vm_mm;
1643 int retval;
1644 pte_t *pte;
1645 spinlock_t *ptl;
1646
1647 retval = validate_page_before_insert(page);
1648 if (retval)
1649 goto out;
1650 retval = -ENOMEM;
1651 pte = get_locked_pte(mm, addr, &ptl);
1652 if (!pte)
1653 goto out;
1654 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1655 pte_unmap_unlock(pte, ptl);
1656out:
1657 return retval;
1658}
1659
1660#ifdef pte_index
1661static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1662 unsigned long addr, struct page *page, pgprot_t prot)
1663{
1664 int err;
1665
1666 if (!page_count(page))
1667 return -EINVAL;
1668 err = validate_page_before_insert(page);
1669 if (err)
1670 return err;
1671 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1672}
1673
1674/* insert_pages() amortizes the cost of spinlock operations
1675 * when inserting pages in a loop. Arch *must* define pte_index.
1676 */
1677static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1678 struct page **pages, unsigned long *num, pgprot_t prot)
1679{
1680 pmd_t *pmd = NULL;
1681 pte_t *start_pte, *pte;
1682 spinlock_t *pte_lock;
1683 struct mm_struct *const mm = vma->vm_mm;
1684 unsigned long curr_page_idx = 0;
1685 unsigned long remaining_pages_total = *num;
1686 unsigned long pages_to_write_in_pmd;
1687 int ret;
1688more:
1689 ret = -EFAULT;
1690 pmd = walk_to_pmd(mm, addr);
1691 if (!pmd)
1692 goto out;
1693
1694 pages_to_write_in_pmd = min_t(unsigned long,
1695 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1696
1697 /* Allocate the PTE if necessary; takes PMD lock once only. */
1698 ret = -ENOMEM;
1699 if (pte_alloc(mm, pmd))
1700 goto out;
1701
1702 while (pages_to_write_in_pmd) {
1703 int pte_idx = 0;
1704 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1705
1706 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1707 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1708 int err = insert_page_in_batch_locked(mm, pte,
1709 addr, pages[curr_page_idx], prot);
1710 if (unlikely(err)) {
1711 pte_unmap_unlock(start_pte, pte_lock);
1712 ret = err;
1713 remaining_pages_total -= pte_idx;
1714 goto out;
1715 }
1716 addr += PAGE_SIZE;
1717 ++curr_page_idx;
1718 }
1719 pte_unmap_unlock(start_pte, pte_lock);
1720 pages_to_write_in_pmd -= batch_size;
1721 remaining_pages_total -= batch_size;
1722 }
1723 if (remaining_pages_total)
1724 goto more;
1725 ret = 0;
1726out:
1727 *num = remaining_pages_total;
1728 return ret;
1729}
1730#endif /* ifdef pte_index */
1731
1732/**
1733 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1734 * @vma: user vma to map to
1735 * @addr: target start user address of these pages
1736 * @pages: source kernel pages
1737 * @num: in: number of pages to map. out: number of pages that were *not*
1738 * mapped. (0 means all pages were successfully mapped).
1739 *
1740 * Preferred over vm_insert_page() when inserting multiple pages.
1741 *
1742 * In case of error, we may have mapped a subset of the provided
1743 * pages. It is the caller's responsibility to account for this case.
1744 *
1745 * The same restrictions apply as in vm_insert_page().
1746 */
1747int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1748 struct page **pages, unsigned long *num)
1749{
1750#ifdef pte_index
1751 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1752
1753 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1754 return -EFAULT;
1755 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1756 BUG_ON(mmap_read_trylock(vma->vm_mm));
1757 BUG_ON(vma->vm_flags & VM_PFNMAP);
1758 vma->vm_flags |= VM_MIXEDMAP;
1759 }
1760 /* Defer page refcount checking till we're about to map that page. */
1761 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1762#else
1763 unsigned long idx = 0, pgcount = *num;
1764 int err = -EINVAL;
1765
1766 for (; idx < pgcount; ++idx) {
1767 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1768 if (err)
1769 break;
1770 }
1771 *num = pgcount - idx;
1772 return err;
1773#endif /* ifdef pte_index */
1774}
1775EXPORT_SYMBOL(vm_insert_pages);
1776
1777/**
1778 * vm_insert_page - insert single page into user vma
1779 * @vma: user vma to map to
1780 * @addr: target user address of this page
1781 * @page: source kernel page
1782 *
1783 * This allows drivers to insert individual pages they've allocated
1784 * into a user vma.
1785 *
1786 * The page has to be a nice clean _individual_ kernel allocation.
1787 * If you allocate a compound page, you need to have marked it as
1788 * such (__GFP_COMP), or manually just split the page up yourself
1789 * (see split_page()).
1790 *
1791 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1792 * took an arbitrary page protection parameter. This doesn't allow
1793 * that. Your vma protection will have to be set up correctly, which
1794 * means that if you want a shared writable mapping, you'd better
1795 * ask for a shared writable mapping!
1796 *
1797 * The page does not need to be reserved.
1798 *
1799 * Usually this function is called from f_op->mmap() handler
1800 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1801 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1802 * function from other places, for example from page-fault handler.
1803 *
1804 * Return: %0 on success, negative error code otherwise.
1805 */
1806int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1807 struct page *page)
1808{
1809 if (addr < vma->vm_start || addr >= vma->vm_end)
1810 return -EFAULT;
1811 if (!page_count(page))
1812 return -EINVAL;
1813 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1814 BUG_ON(mmap_read_trylock(vma->vm_mm));
1815 BUG_ON(vma->vm_flags & VM_PFNMAP);
1816 vma->vm_flags |= VM_MIXEDMAP;
1817 }
1818 return insert_page(vma, addr, page, vma->vm_page_prot);
1819}
1820EXPORT_SYMBOL(vm_insert_page);
1821
1822/*
1823 * __vm_map_pages - maps range of kernel pages into user vma
1824 * @vma: user vma to map to
1825 * @pages: pointer to array of source kernel pages
1826 * @num: number of pages in page array
1827 * @offset: user's requested vm_pgoff
1828 *
1829 * This allows drivers to map range of kernel pages into a user vma.
1830 *
1831 * Return: 0 on success and error code otherwise.
1832 */
1833static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1834 unsigned long num, unsigned long offset)
1835{
1836 unsigned long count = vma_pages(vma);
1837 unsigned long uaddr = vma->vm_start;
1838 int ret, i;
1839
1840 /* Fail if the user requested offset is beyond the end of the object */
1841 if (offset >= num)
1842 return -ENXIO;
1843
1844 /* Fail if the user requested size exceeds available object size */
1845 if (count > num - offset)
1846 return -ENXIO;
1847
1848 for (i = 0; i < count; i++) {
1849 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1850 if (ret < 0)
1851 return ret;
1852 uaddr += PAGE_SIZE;
1853 }
1854
1855 return 0;
1856}
1857
1858/**
1859 * vm_map_pages - maps range of kernel pages starts with non zero offset
1860 * @vma: user vma to map to
1861 * @pages: pointer to array of source kernel pages
1862 * @num: number of pages in page array
1863 *
1864 * Maps an object consisting of @num pages, catering for the user's
1865 * requested vm_pgoff
1866 *
1867 * If we fail to insert any page into the vma, the function will return
1868 * immediately leaving any previously inserted pages present. Callers
1869 * from the mmap handler may immediately return the error as their caller
1870 * will destroy the vma, removing any successfully inserted pages. Other
1871 * callers should make their own arrangements for calling unmap_region().
1872 *
1873 * Context: Process context. Called by mmap handlers.
1874 * Return: 0 on success and error code otherwise.
1875 */
1876int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1877 unsigned long num)
1878{
1879 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1880}
1881EXPORT_SYMBOL(vm_map_pages);
1882
1883/**
1884 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1885 * @vma: user vma to map to
1886 * @pages: pointer to array of source kernel pages
1887 * @num: number of pages in page array
1888 *
1889 * Similar to vm_map_pages(), except that it explicitly sets the offset
1890 * to 0. This function is intended for the drivers that did not consider
1891 * vm_pgoff.
1892 *
1893 * Context: Process context. Called by mmap handlers.
1894 * Return: 0 on success and error code otherwise.
1895 */
1896int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1897 unsigned long num)
1898{
1899 return __vm_map_pages(vma, pages, num, 0);
1900}
1901EXPORT_SYMBOL(vm_map_pages_zero);
1902
1903static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1904 pfn_t pfn, pgprot_t prot, bool mkwrite)
1905{
1906 struct mm_struct *mm = vma->vm_mm;
1907 pte_t *pte, entry;
1908 spinlock_t *ptl;
1909
1910 pte = get_locked_pte(mm, addr, &ptl);
1911 if (!pte)
1912 return VM_FAULT_OOM;
1913 if (!pte_none(*pte)) {
1914 if (mkwrite) {
1915 /*
1916 * For read faults on private mappings the PFN passed
1917 * in may not match the PFN we have mapped if the
1918 * mapped PFN is a writeable COW page. In the mkwrite
1919 * case we are creating a writable PTE for a shared
1920 * mapping and we expect the PFNs to match. If they
1921 * don't match, we are likely racing with block
1922 * allocation and mapping invalidation so just skip the
1923 * update.
1924 */
1925 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1926 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1927 goto out_unlock;
1928 }
1929 entry = pte_mkyoung(*pte);
1930 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1931 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1932 update_mmu_cache(vma, addr, pte);
1933 }
1934 goto out_unlock;
1935 }
1936
1937 /* Ok, finally just insert the thing.. */
1938 if (pfn_t_devmap(pfn))
1939 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1940 else
1941 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1942
1943 if (mkwrite) {
1944 entry = pte_mkyoung(entry);
1945 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1946 }
1947
1948 set_pte_at(mm, addr, pte, entry);
1949 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1950
1951out_unlock:
1952 pte_unmap_unlock(pte, ptl);
1953 return VM_FAULT_NOPAGE;
1954}
1955
1956/**
1957 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1958 * @vma: user vma to map to
1959 * @addr: target user address of this page
1960 * @pfn: source kernel pfn
1961 * @pgprot: pgprot flags for the inserted page
1962 *
1963 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1964 * to override pgprot on a per-page basis.
1965 *
1966 * This only makes sense for IO mappings, and it makes no sense for
1967 * COW mappings. In general, using multiple vmas is preferable;
1968 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1969 * impractical.
1970 *
1971 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1972 * a value of @pgprot different from that of @vma->vm_page_prot.
1973 *
1974 * Context: Process context. May allocate using %GFP_KERNEL.
1975 * Return: vm_fault_t value.
1976 */
1977vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1978 unsigned long pfn, pgprot_t pgprot)
1979{
1980 /*
1981 * Technically, architectures with pte_special can avoid all these
1982 * restrictions (same for remap_pfn_range). However we would like
1983 * consistency in testing and feature parity among all, so we should
1984 * try to keep these invariants in place for everybody.
1985 */
1986 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1987 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1988 (VM_PFNMAP|VM_MIXEDMAP));
1989 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1990 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1991
1992 if (addr < vma->vm_start || addr >= vma->vm_end)
1993 return VM_FAULT_SIGBUS;
1994
1995 if (!pfn_modify_allowed(pfn, pgprot))
1996 return VM_FAULT_SIGBUS;
1997
1998 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1999
2000 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2001 false);
2002}
2003EXPORT_SYMBOL(vmf_insert_pfn_prot);
2004
2005/**
2006 * vmf_insert_pfn - insert single pfn into user vma
2007 * @vma: user vma to map to
2008 * @addr: target user address of this page
2009 * @pfn: source kernel pfn
2010 *
2011 * Similar to vm_insert_page, this allows drivers to insert individual pages
2012 * they've allocated into a user vma. Same comments apply.
2013 *
2014 * This function should only be called from a vm_ops->fault handler, and
2015 * in that case the handler should return the result of this function.
2016 *
2017 * vma cannot be a COW mapping.
2018 *
2019 * As this is called only for pages that do not currently exist, we
2020 * do not need to flush old virtual caches or the TLB.
2021 *
2022 * Context: Process context. May allocate using %GFP_KERNEL.
2023 * Return: vm_fault_t value.
2024 */
2025vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2026 unsigned long pfn)
2027{
2028 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2029}
2030EXPORT_SYMBOL(vmf_insert_pfn);
2031
2032static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2033{
2034 /* these checks mirror the abort conditions in vm_normal_page */
2035 if (vma->vm_flags & VM_MIXEDMAP)
2036 return true;
2037 if (pfn_t_devmap(pfn))
2038 return true;
2039 if (pfn_t_special(pfn))
2040 return true;
2041 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2042 return true;
2043 return false;
2044}
2045
2046static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2047 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2048 bool mkwrite)
2049{
2050 int err;
2051
2052 BUG_ON(!vm_mixed_ok(vma, pfn));
2053
2054 if (addr < vma->vm_start || addr >= vma->vm_end)
2055 return VM_FAULT_SIGBUS;
2056
2057 track_pfn_insert(vma, &pgprot, pfn);
2058
2059 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2060 return VM_FAULT_SIGBUS;
2061
2062 /*
2063 * If we don't have pte special, then we have to use the pfn_valid()
2064 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2065 * refcount the page if pfn_valid is true (hence insert_page rather
2066 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2067 * without pte special, it would there be refcounted as a normal page.
2068 */
2069 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2070 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2071 struct page *page;
2072
2073 /*
2074 * At this point we are committed to insert_page()
2075 * regardless of whether the caller specified flags that
2076 * result in pfn_t_has_page() == false.
2077 */
2078 page = pfn_to_page(pfn_t_to_pfn(pfn));
2079 err = insert_page(vma, addr, page, pgprot);
2080 } else {
2081 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2082 }
2083
2084 if (err == -ENOMEM)
2085 return VM_FAULT_OOM;
2086 if (err < 0 && err != -EBUSY)
2087 return VM_FAULT_SIGBUS;
2088
2089 return VM_FAULT_NOPAGE;
2090}
2091
2092/**
2093 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2094 * @vma: user vma to map to
2095 * @addr: target user address of this page
2096 * @pfn: source kernel pfn
2097 * @pgprot: pgprot flags for the inserted page
2098 *
2099 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2100 * to override pgprot on a per-page basis.
2101 *
2102 * Typically this function should be used by drivers to set caching- and
2103 * encryption bits different than those of @vma->vm_page_prot, because
2104 * the caching- or encryption mode may not be known at mmap() time.
2105 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2106 * to set caching and encryption bits for those vmas (except for COW pages).
2107 * This is ensured by core vm only modifying these page table entries using
2108 * functions that don't touch caching- or encryption bits, using pte_modify()
2109 * if needed. (See for example mprotect()).
2110 * Also when new page-table entries are created, this is only done using the
2111 * fault() callback, and never using the value of vma->vm_page_prot,
2112 * except for page-table entries that point to anonymous pages as the result
2113 * of COW.
2114 *
2115 * Context: Process context. May allocate using %GFP_KERNEL.
2116 * Return: vm_fault_t value.
2117 */
2118vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2119 pfn_t pfn, pgprot_t pgprot)
2120{
2121 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2122}
2123EXPORT_SYMBOL(vmf_insert_mixed_prot);
2124
2125vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2126 pfn_t pfn)
2127{
2128 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2129}
2130EXPORT_SYMBOL(vmf_insert_mixed);
2131
2132/*
2133 * If the insertion of PTE failed because someone else already added a
2134 * different entry in the mean time, we treat that as success as we assume
2135 * the same entry was actually inserted.
2136 */
2137vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2138 unsigned long addr, pfn_t pfn)
2139{
2140 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2141}
2142EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2143
2144/*
2145 * maps a range of physical memory into the requested pages. the old
2146 * mappings are removed. any references to nonexistent pages results
2147 * in null mappings (currently treated as "copy-on-access")
2148 */
2149static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2150 unsigned long addr, unsigned long end,
2151 unsigned long pfn, pgprot_t prot)
2152{
2153 pte_t *pte;
2154 spinlock_t *ptl;
2155 int err = 0;
2156
2157 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2158 if (!pte)
2159 return -ENOMEM;
2160 arch_enter_lazy_mmu_mode();
2161 do {
2162 BUG_ON(!pte_none(*pte));
2163 if (!pfn_modify_allowed(pfn, prot)) {
2164 err = -EACCES;
2165 break;
2166 }
2167 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2168 pfn++;
2169 } while (pte++, addr += PAGE_SIZE, addr != end);
2170 arch_leave_lazy_mmu_mode();
2171 pte_unmap_unlock(pte - 1, ptl);
2172 return err;
2173}
2174
2175static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2176 unsigned long addr, unsigned long end,
2177 unsigned long pfn, pgprot_t prot)
2178{
2179 pmd_t *pmd;
2180 unsigned long next;
2181 int err;
2182
2183 pfn -= addr >> PAGE_SHIFT;
2184 pmd = pmd_alloc(mm, pud, addr);
2185 if (!pmd)
2186 return -ENOMEM;
2187 VM_BUG_ON(pmd_trans_huge(*pmd));
2188 do {
2189 next = pmd_addr_end(addr, end);
2190 err = remap_pte_range(mm, pmd, addr, next,
2191 pfn + (addr >> PAGE_SHIFT), prot);
2192 if (err)
2193 return err;
2194 } while (pmd++, addr = next, addr != end);
2195 return 0;
2196}
2197
2198static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2199 unsigned long addr, unsigned long end,
2200 unsigned long pfn, pgprot_t prot)
2201{
2202 pud_t *pud;
2203 unsigned long next;
2204 int err;
2205
2206 pfn -= addr >> PAGE_SHIFT;
2207 pud = pud_alloc(mm, p4d, addr);
2208 if (!pud)
2209 return -ENOMEM;
2210 do {
2211 next = pud_addr_end(addr, end);
2212 err = remap_pmd_range(mm, pud, addr, next,
2213 pfn + (addr >> PAGE_SHIFT), prot);
2214 if (err)
2215 return err;
2216 } while (pud++, addr = next, addr != end);
2217 return 0;
2218}
2219
2220static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2221 unsigned long addr, unsigned long end,
2222 unsigned long pfn, pgprot_t prot)
2223{
2224 p4d_t *p4d;
2225 unsigned long next;
2226 int err;
2227
2228 pfn -= addr >> PAGE_SHIFT;
2229 p4d = p4d_alloc(mm, pgd, addr);
2230 if (!p4d)
2231 return -ENOMEM;
2232 do {
2233 next = p4d_addr_end(addr, end);
2234 err = remap_pud_range(mm, p4d, addr, next,
2235 pfn + (addr >> PAGE_SHIFT), prot);
2236 if (err)
2237 return err;
2238 } while (p4d++, addr = next, addr != end);
2239 return 0;
2240}
2241
2242/**
2243 * remap_pfn_range - remap kernel memory to userspace
2244 * @vma: user vma to map to
2245 * @addr: target page aligned user address to start at
2246 * @pfn: page frame number of kernel physical memory address
2247 * @size: size of mapping area
2248 * @prot: page protection flags for this mapping
2249 *
2250 * Note: this is only safe if the mm semaphore is held when called.
2251 *
2252 * Return: %0 on success, negative error code otherwise.
2253 */
2254int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2255 unsigned long pfn, unsigned long size, pgprot_t prot)
2256{
2257 pgd_t *pgd;
2258 unsigned long next;
2259 unsigned long end = addr + PAGE_ALIGN(size);
2260 struct mm_struct *mm = vma->vm_mm;
2261 unsigned long remap_pfn = pfn;
2262 int err;
2263
2264 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2265 return -EINVAL;
2266
2267 /*
2268 * Physically remapped pages are special. Tell the
2269 * rest of the world about it:
2270 * VM_IO tells people not to look at these pages
2271 * (accesses can have side effects).
2272 * VM_PFNMAP tells the core MM that the base pages are just
2273 * raw PFN mappings, and do not have a "struct page" associated
2274 * with them.
2275 * VM_DONTEXPAND
2276 * Disable vma merging and expanding with mremap().
2277 * VM_DONTDUMP
2278 * Omit vma from core dump, even when VM_IO turned off.
2279 *
2280 * There's a horrible special case to handle copy-on-write
2281 * behaviour that some programs depend on. We mark the "original"
2282 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2283 * See vm_normal_page() for details.
2284 */
2285 if (is_cow_mapping(vma->vm_flags)) {
2286 if (addr != vma->vm_start || end != vma->vm_end)
2287 return -EINVAL;
2288 vma->vm_pgoff = pfn;
2289 }
2290
2291 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2292 if (err)
2293 return -EINVAL;
2294
2295 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2296
2297 BUG_ON(addr >= end);
2298 pfn -= addr >> PAGE_SHIFT;
2299 pgd = pgd_offset(mm, addr);
2300 flush_cache_range(vma, addr, end);
2301 do {
2302 next = pgd_addr_end(addr, end);
2303 err = remap_p4d_range(mm, pgd, addr, next,
2304 pfn + (addr >> PAGE_SHIFT), prot);
2305 if (err)
2306 break;
2307 } while (pgd++, addr = next, addr != end);
2308
2309 if (err)
2310 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2311
2312 return err;
2313}
2314EXPORT_SYMBOL(remap_pfn_range);
2315
2316/**
2317 * vm_iomap_memory - remap memory to userspace
2318 * @vma: user vma to map to
2319 * @start: start of the physical memory to be mapped
2320 * @len: size of area
2321 *
2322 * This is a simplified io_remap_pfn_range() for common driver use. The
2323 * driver just needs to give us the physical memory range to be mapped,
2324 * we'll figure out the rest from the vma information.
2325 *
2326 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2327 * whatever write-combining details or similar.
2328 *
2329 * Return: %0 on success, negative error code otherwise.
2330 */
2331int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2332{
2333 unsigned long vm_len, pfn, pages;
2334
2335 /* Check that the physical memory area passed in looks valid */
2336 if (start + len < start)
2337 return -EINVAL;
2338 /*
2339 * You *really* shouldn't map things that aren't page-aligned,
2340 * but we've historically allowed it because IO memory might
2341 * just have smaller alignment.
2342 */
2343 len += start & ~PAGE_MASK;
2344 pfn = start >> PAGE_SHIFT;
2345 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2346 if (pfn + pages < pfn)
2347 return -EINVAL;
2348
2349 /* We start the mapping 'vm_pgoff' pages into the area */
2350 if (vma->vm_pgoff > pages)
2351 return -EINVAL;
2352 pfn += vma->vm_pgoff;
2353 pages -= vma->vm_pgoff;
2354
2355 /* Can we fit all of the mapping? */
2356 vm_len = vma->vm_end - vma->vm_start;
2357 if (vm_len >> PAGE_SHIFT > pages)
2358 return -EINVAL;
2359
2360 /* Ok, let it rip */
2361 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2362}
2363EXPORT_SYMBOL(vm_iomap_memory);
2364
2365static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2366 unsigned long addr, unsigned long end,
2367 pte_fn_t fn, void *data, bool create,
2368 pgtbl_mod_mask *mask)
2369{
2370 pte_t *pte;
2371 int err = 0;
2372 spinlock_t *ptl;
2373
2374 if (create) {
2375 pte = (mm == &init_mm) ?
2376 pte_alloc_kernel_track(pmd, addr, mask) :
2377 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2378 if (!pte)
2379 return -ENOMEM;
2380 } else {
2381 pte = (mm == &init_mm) ?
2382 pte_offset_kernel(pmd, addr) :
2383 pte_offset_map_lock(mm, pmd, addr, &ptl);
2384 }
2385
2386 BUG_ON(pmd_huge(*pmd));
2387
2388 arch_enter_lazy_mmu_mode();
2389
2390 do {
2391 if (create || !pte_none(*pte)) {
2392 err = fn(pte++, addr, data);
2393 if (err)
2394 break;
2395 }
2396 } while (addr += PAGE_SIZE, addr != end);
2397 *mask |= PGTBL_PTE_MODIFIED;
2398
2399 arch_leave_lazy_mmu_mode();
2400
2401 if (mm != &init_mm)
2402 pte_unmap_unlock(pte-1, ptl);
2403 return err;
2404}
2405
2406static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2407 unsigned long addr, unsigned long end,
2408 pte_fn_t fn, void *data, bool create,
2409 pgtbl_mod_mask *mask)
2410{
2411 pmd_t *pmd;
2412 unsigned long next;
2413 int err = 0;
2414
2415 BUG_ON(pud_huge(*pud));
2416
2417 if (create) {
2418 pmd = pmd_alloc_track(mm, pud, addr, mask);
2419 if (!pmd)
2420 return -ENOMEM;
2421 } else {
2422 pmd = pmd_offset(pud, addr);
2423 }
2424 do {
2425 next = pmd_addr_end(addr, end);
2426 if (create || !pmd_none_or_clear_bad(pmd)) {
2427 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2428 create, mask);
2429 if (err)
2430 break;
2431 }
2432 } while (pmd++, addr = next, addr != end);
2433 return err;
2434}
2435
2436static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2437 unsigned long addr, unsigned long end,
2438 pte_fn_t fn, void *data, bool create,
2439 pgtbl_mod_mask *mask)
2440{
2441 pud_t *pud;
2442 unsigned long next;
2443 int err = 0;
2444
2445 if (create) {
2446 pud = pud_alloc_track(mm, p4d, addr, mask);
2447 if (!pud)
2448 return -ENOMEM;
2449 } else {
2450 pud = pud_offset(p4d, addr);
2451 }
2452 do {
2453 next = pud_addr_end(addr, end);
2454 if (create || !pud_none_or_clear_bad(pud)) {
2455 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2456 create, mask);
2457 if (err)
2458 break;
2459 }
2460 } while (pud++, addr = next, addr != end);
2461 return err;
2462}
2463
2464static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2465 unsigned long addr, unsigned long end,
2466 pte_fn_t fn, void *data, bool create,
2467 pgtbl_mod_mask *mask)
2468{
2469 p4d_t *p4d;
2470 unsigned long next;
2471 int err = 0;
2472
2473 if (create) {
2474 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2475 if (!p4d)
2476 return -ENOMEM;
2477 } else {
2478 p4d = p4d_offset(pgd, addr);
2479 }
2480 do {
2481 next = p4d_addr_end(addr, end);
2482 if (create || !p4d_none_or_clear_bad(p4d)) {
2483 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2484 create, mask);
2485 if (err)
2486 break;
2487 }
2488 } while (p4d++, addr = next, addr != end);
2489 return err;
2490}
2491
2492static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2493 unsigned long size, pte_fn_t fn,
2494 void *data, bool create)
2495{
2496 pgd_t *pgd;
2497 unsigned long start = addr, next;
2498 unsigned long end = addr + size;
2499 pgtbl_mod_mask mask = 0;
2500 int err = 0;
2501
2502 if (WARN_ON(addr >= end))
2503 return -EINVAL;
2504
2505 pgd = pgd_offset(mm, addr);
2506 do {
2507 next = pgd_addr_end(addr, end);
2508 if (!create && pgd_none_or_clear_bad(pgd))
2509 continue;
2510 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask);
2511 if (err)
2512 break;
2513 } while (pgd++, addr = next, addr != end);
2514
2515 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2516 arch_sync_kernel_mappings(start, start + size);
2517
2518 return err;
2519}
2520
2521/*
2522 * Scan a region of virtual memory, filling in page tables as necessary
2523 * and calling a provided function on each leaf page table.
2524 */
2525int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2526 unsigned long size, pte_fn_t fn, void *data)
2527{
2528 return __apply_to_page_range(mm, addr, size, fn, data, true);
2529}
2530EXPORT_SYMBOL_GPL(apply_to_page_range);
2531
2532/*
2533 * Scan a region of virtual memory, calling a provided function on
2534 * each leaf page table where it exists.
2535 *
2536 * Unlike apply_to_page_range, this does _not_ fill in page tables
2537 * where they are absent.
2538 */
2539int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2540 unsigned long size, pte_fn_t fn, void *data)
2541{
2542 return __apply_to_page_range(mm, addr, size, fn, data, false);
2543}
2544EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2545
2546/*
2547 * handle_pte_fault chooses page fault handler according to an entry which was
2548 * read non-atomically. Before making any commitment, on those architectures
2549 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2550 * parts, do_swap_page must check under lock before unmapping the pte and
2551 * proceeding (but do_wp_page is only called after already making such a check;
2552 * and do_anonymous_page can safely check later on).
2553 */
2554static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2555 pte_t *page_table, pte_t orig_pte)
2556{
2557 int same = 1;
2558#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2559 if (sizeof(pte_t) > sizeof(unsigned long)) {
2560 spinlock_t *ptl = pte_lockptr(mm, pmd);
2561 spin_lock(ptl);
2562 same = pte_same(*page_table, orig_pte);
2563 spin_unlock(ptl);
2564 }
2565#endif
2566 pte_unmap(page_table);
2567 return same;
2568}
2569
2570static inline bool cow_user_page(struct page *dst, struct page *src,
2571 struct vm_fault *vmf)
2572{
2573 bool ret;
2574 void *kaddr;
2575 void __user *uaddr;
2576 bool locked = false;
2577 struct vm_area_struct *vma = vmf->vma;
2578 struct mm_struct *mm = vma->vm_mm;
2579 unsigned long addr = vmf->address;
2580
2581 if (likely(src)) {
2582 copy_user_highpage(dst, src, addr, vma);
2583 return true;
2584 }
2585
2586 /*
2587 * If the source page was a PFN mapping, we don't have
2588 * a "struct page" for it. We do a best-effort copy by
2589 * just copying from the original user address. If that
2590 * fails, we just zero-fill it. Live with it.
2591 */
2592 kaddr = kmap_atomic(dst);
2593 uaddr = (void __user *)(addr & PAGE_MASK);
2594
2595 /*
2596 * On architectures with software "accessed" bits, we would
2597 * take a double page fault, so mark it accessed here.
2598 */
2599 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2600 pte_t entry;
2601
2602 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2603 locked = true;
2604 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2605 /*
2606 * Other thread has already handled the fault
2607 * and update local tlb only
2608 */
2609 update_mmu_tlb(vma, addr, vmf->pte);
2610 ret = false;
2611 goto pte_unlock;
2612 }
2613
2614 entry = pte_mkyoung(vmf->orig_pte);
2615 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2616 update_mmu_cache(vma, addr, vmf->pte);
2617 }
2618
2619 /*
2620 * This really shouldn't fail, because the page is there
2621 * in the page tables. But it might just be unreadable,
2622 * in which case we just give up and fill the result with
2623 * zeroes.
2624 */
2625 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2626 if (locked)
2627 goto warn;
2628
2629 /* Re-validate under PTL if the page is still mapped */
2630 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2631 locked = true;
2632 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2633 /* The PTE changed under us, update local tlb */
2634 update_mmu_tlb(vma, addr, vmf->pte);
2635 ret = false;
2636 goto pte_unlock;
2637 }
2638
2639 /*
2640 * The same page can be mapped back since last copy attempt.
2641 * Try to copy again under PTL.
2642 */
2643 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2644 /*
2645 * Give a warn in case there can be some obscure
2646 * use-case
2647 */
2648warn:
2649 WARN_ON_ONCE(1);
2650 clear_page(kaddr);
2651 }
2652 }
2653
2654 ret = true;
2655
2656pte_unlock:
2657 if (locked)
2658 pte_unmap_unlock(vmf->pte, vmf->ptl);
2659 kunmap_atomic(kaddr);
2660 flush_dcache_page(dst);
2661
2662 return ret;
2663}
2664
2665static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2666{
2667 struct file *vm_file = vma->vm_file;
2668
2669 if (vm_file)
2670 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2671
2672 /*
2673 * Special mappings (e.g. VDSO) do not have any file so fake
2674 * a default GFP_KERNEL for them.
2675 */
2676 return GFP_KERNEL;
2677}
2678
2679/*
2680 * Notify the address space that the page is about to become writable so that
2681 * it can prohibit this or wait for the page to get into an appropriate state.
2682 *
2683 * We do this without the lock held, so that it can sleep if it needs to.
2684 */
2685static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2686{
2687 vm_fault_t ret;
2688 struct page *page = vmf->page;
2689 unsigned int old_flags = vmf->flags;
2690
2691 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2692
2693 if (vmf->vma->vm_file &&
2694 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2695 return VM_FAULT_SIGBUS;
2696
2697 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2698 /* Restore original flags so that caller is not surprised */
2699 vmf->flags = old_flags;
2700 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2701 return ret;
2702 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2703 lock_page(page);
2704 if (!page->mapping) {
2705 unlock_page(page);
2706 return 0; /* retry */
2707 }
2708 ret |= VM_FAULT_LOCKED;
2709 } else
2710 VM_BUG_ON_PAGE(!PageLocked(page), page);
2711 return ret;
2712}
2713
2714/*
2715 * Handle dirtying of a page in shared file mapping on a write fault.
2716 *
2717 * The function expects the page to be locked and unlocks it.
2718 */
2719static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2720{
2721 struct vm_area_struct *vma = vmf->vma;
2722 struct address_space *mapping;
2723 struct page *page = vmf->page;
2724 bool dirtied;
2725 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2726
2727 dirtied = set_page_dirty(page);
2728 VM_BUG_ON_PAGE(PageAnon(page), page);
2729 /*
2730 * Take a local copy of the address_space - page.mapping may be zeroed
2731 * by truncate after unlock_page(). The address_space itself remains
2732 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2733 * release semantics to prevent the compiler from undoing this copying.
2734 */
2735 mapping = page_rmapping(page);
2736 unlock_page(page);
2737
2738 if (!page_mkwrite)
2739 file_update_time(vma->vm_file);
2740
2741 /*
2742 * Throttle page dirtying rate down to writeback speed.
2743 *
2744 * mapping may be NULL here because some device drivers do not
2745 * set page.mapping but still dirty their pages
2746 *
2747 * Drop the mmap_lock before waiting on IO, if we can. The file
2748 * is pinning the mapping, as per above.
2749 */
2750 if ((dirtied || page_mkwrite) && mapping) {
2751 struct file *fpin;
2752
2753 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2754 balance_dirty_pages_ratelimited(mapping);
2755 if (fpin) {
2756 fput(fpin);
2757 return VM_FAULT_RETRY;
2758 }
2759 }
2760
2761 return 0;
2762}
2763
2764/*
2765 * Handle write page faults for pages that can be reused in the current vma
2766 *
2767 * This can happen either due to the mapping being with the VM_SHARED flag,
2768 * or due to us being the last reference standing to the page. In either
2769 * case, all we need to do here is to mark the page as writable and update
2770 * any related book-keeping.
2771 */
2772static inline void wp_page_reuse(struct vm_fault *vmf)
2773 __releases(vmf->ptl)
2774{
2775 struct vm_area_struct *vma = vmf->vma;
2776 struct page *page = vmf->page;
2777 pte_t entry;
2778 /*
2779 * Clear the pages cpupid information as the existing
2780 * information potentially belongs to a now completely
2781 * unrelated process.
2782 */
2783 if (page)
2784 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2785
2786 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2787 entry = pte_mkyoung(vmf->orig_pte);
2788 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2789 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2790 update_mmu_cache(vma, vmf->address, vmf->pte);
2791 pte_unmap_unlock(vmf->pte, vmf->ptl);
2792 count_vm_event(PGREUSE);
2793}
2794
2795/*
2796 * Handle the case of a page which we actually need to copy to a new page.
2797 *
2798 * Called with mmap_lock locked and the old page referenced, but
2799 * without the ptl held.
2800 *
2801 * High level logic flow:
2802 *
2803 * - Allocate a page, copy the content of the old page to the new one.
2804 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2805 * - Take the PTL. If the pte changed, bail out and release the allocated page
2806 * - If the pte is still the way we remember it, update the page table and all
2807 * relevant references. This includes dropping the reference the page-table
2808 * held to the old page, as well as updating the rmap.
2809 * - In any case, unlock the PTL and drop the reference we took to the old page.
2810 */
2811static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2812{
2813 struct vm_area_struct *vma = vmf->vma;
2814 struct mm_struct *mm = vma->vm_mm;
2815 struct page *old_page = vmf->page;
2816 struct page *new_page = NULL;
2817 pte_t entry;
2818 int page_copied = 0;
2819 struct mmu_notifier_range range;
2820
2821 if (unlikely(anon_vma_prepare(vma)))
2822 goto oom;
2823
2824 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2825 new_page = alloc_zeroed_user_highpage_movable(vma,
2826 vmf->address);
2827 if (!new_page)
2828 goto oom;
2829 } else {
2830 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2831 vmf->address);
2832 if (!new_page)
2833 goto oom;
2834
2835 if (!cow_user_page(new_page, old_page, vmf)) {
2836 /*
2837 * COW failed, if the fault was solved by other,
2838 * it's fine. If not, userspace would re-fault on
2839 * the same address and we will handle the fault
2840 * from the second attempt.
2841 */
2842 put_page(new_page);
2843 if (old_page)
2844 put_page(old_page);
2845 return 0;
2846 }
2847 }
2848
2849 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2850 goto oom_free_new;
2851 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2852
2853 __SetPageUptodate(new_page);
2854
2855 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2856 vmf->address & PAGE_MASK,
2857 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2858 mmu_notifier_invalidate_range_start(&range);
2859
2860 /*
2861 * Re-check the pte - we dropped the lock
2862 */
2863 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2864 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2865 if (old_page) {
2866 if (!PageAnon(old_page)) {
2867 dec_mm_counter_fast(mm,
2868 mm_counter_file(old_page));
2869 inc_mm_counter_fast(mm, MM_ANONPAGES);
2870 }
2871 } else {
2872 inc_mm_counter_fast(mm, MM_ANONPAGES);
2873 }
2874 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2875 entry = mk_pte(new_page, vma->vm_page_prot);
2876 entry = pte_sw_mkyoung(entry);
2877 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2878 /*
2879 * Clear the pte entry and flush it first, before updating the
2880 * pte with the new entry. This will avoid a race condition
2881 * seen in the presence of one thread doing SMC and another
2882 * thread doing COW.
2883 */
2884 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2885 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2886 lru_cache_add_inactive_or_unevictable(new_page, vma);
2887 /*
2888 * We call the notify macro here because, when using secondary
2889 * mmu page tables (such as kvm shadow page tables), we want the
2890 * new page to be mapped directly into the secondary page table.
2891 */
2892 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2893 update_mmu_cache(vma, vmf->address, vmf->pte);
2894 if (old_page) {
2895 /*
2896 * Only after switching the pte to the new page may
2897 * we remove the mapcount here. Otherwise another
2898 * process may come and find the rmap count decremented
2899 * before the pte is switched to the new page, and
2900 * "reuse" the old page writing into it while our pte
2901 * here still points into it and can be read by other
2902 * threads.
2903 *
2904 * The critical issue is to order this
2905 * page_remove_rmap with the ptp_clear_flush above.
2906 * Those stores are ordered by (if nothing else,)
2907 * the barrier present in the atomic_add_negative
2908 * in page_remove_rmap.
2909 *
2910 * Then the TLB flush in ptep_clear_flush ensures that
2911 * no process can access the old page before the
2912 * decremented mapcount is visible. And the old page
2913 * cannot be reused until after the decremented
2914 * mapcount is visible. So transitively, TLBs to
2915 * old page will be flushed before it can be reused.
2916 */
2917 page_remove_rmap(old_page, false);
2918 }
2919
2920 /* Free the old page.. */
2921 new_page = old_page;
2922 page_copied = 1;
2923 } else {
2924 update_mmu_tlb(vma, vmf->address, vmf->pte);
2925 }
2926
2927 if (new_page)
2928 put_page(new_page);
2929
2930 pte_unmap_unlock(vmf->pte, vmf->ptl);
2931 /*
2932 * No need to double call mmu_notifier->invalidate_range() callback as
2933 * the above ptep_clear_flush_notify() did already call it.
2934 */
2935 mmu_notifier_invalidate_range_only_end(&range);
2936 if (old_page) {
2937 /*
2938 * Don't let another task, with possibly unlocked vma,
2939 * keep the mlocked page.
2940 */
2941 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2942 lock_page(old_page); /* LRU manipulation */
2943 if (PageMlocked(old_page))
2944 munlock_vma_page(old_page);
2945 unlock_page(old_page);
2946 }
2947 put_page(old_page);
2948 }
2949 return page_copied ? VM_FAULT_WRITE : 0;
2950oom_free_new:
2951 put_page(new_page);
2952oom:
2953 if (old_page)
2954 put_page(old_page);
2955 return VM_FAULT_OOM;
2956}
2957
2958/**
2959 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2960 * writeable once the page is prepared
2961 *
2962 * @vmf: structure describing the fault
2963 *
2964 * This function handles all that is needed to finish a write page fault in a
2965 * shared mapping due to PTE being read-only once the mapped page is prepared.
2966 * It handles locking of PTE and modifying it.
2967 *
2968 * The function expects the page to be locked or other protection against
2969 * concurrent faults / writeback (such as DAX radix tree locks).
2970 *
2971 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2972 * we acquired PTE lock.
2973 */
2974vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2975{
2976 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2977 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2978 &vmf->ptl);
2979 /*
2980 * We might have raced with another page fault while we released the
2981 * pte_offset_map_lock.
2982 */
2983 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2984 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
2985 pte_unmap_unlock(vmf->pte, vmf->ptl);
2986 return VM_FAULT_NOPAGE;
2987 }
2988 wp_page_reuse(vmf);
2989 return 0;
2990}
2991
2992/*
2993 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2994 * mapping
2995 */
2996static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2997{
2998 struct vm_area_struct *vma = vmf->vma;
2999
3000 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3001 vm_fault_t ret;
3002
3003 pte_unmap_unlock(vmf->pte, vmf->ptl);
3004 vmf->flags |= FAULT_FLAG_MKWRITE;
3005 ret = vma->vm_ops->pfn_mkwrite(vmf);
3006 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3007 return ret;
3008 return finish_mkwrite_fault(vmf);
3009 }
3010 wp_page_reuse(vmf);
3011 return VM_FAULT_WRITE;
3012}
3013
3014static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3015 __releases(vmf->ptl)
3016{
3017 struct vm_area_struct *vma = vmf->vma;
3018 vm_fault_t ret = VM_FAULT_WRITE;
3019
3020 get_page(vmf->page);
3021
3022 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3023 vm_fault_t tmp;
3024
3025 pte_unmap_unlock(vmf->pte, vmf->ptl);
3026 tmp = do_page_mkwrite(vmf);
3027 if (unlikely(!tmp || (tmp &
3028 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3029 put_page(vmf->page);
3030 return tmp;
3031 }
3032 tmp = finish_mkwrite_fault(vmf);
3033 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3034 unlock_page(vmf->page);
3035 put_page(vmf->page);
3036 return tmp;
3037 }
3038 } else {
3039 wp_page_reuse(vmf);
3040 lock_page(vmf->page);
3041 }
3042 ret |= fault_dirty_shared_page(vmf);
3043 put_page(vmf->page);
3044
3045 return ret;
3046}
3047
3048/*
3049 * This routine handles present pages, when users try to write
3050 * to a shared page. It is done by copying the page to a new address
3051 * and decrementing the shared-page counter for the old page.
3052 *
3053 * Note that this routine assumes that the protection checks have been
3054 * done by the caller (the low-level page fault routine in most cases).
3055 * Thus we can safely just mark it writable once we've done any necessary
3056 * COW.
3057 *
3058 * We also mark the page dirty at this point even though the page will
3059 * change only once the write actually happens. This avoids a few races,
3060 * and potentially makes it more efficient.
3061 *
3062 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3063 * but allow concurrent faults), with pte both mapped and locked.
3064 * We return with mmap_lock still held, but pte unmapped and unlocked.
3065 */
3066static vm_fault_t do_wp_page(struct vm_fault *vmf)
3067 __releases(vmf->ptl)
3068{
3069 struct vm_area_struct *vma = vmf->vma;
3070
3071 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3072 pte_unmap_unlock(vmf->pte, vmf->ptl);
3073 return handle_userfault(vmf, VM_UFFD_WP);
3074 }
3075
3076 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3077 if (!vmf->page) {
3078 /*
3079 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3080 * VM_PFNMAP VMA.
3081 *
3082 * We should not cow pages in a shared writeable mapping.
3083 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3084 */
3085 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3086 (VM_WRITE|VM_SHARED))
3087 return wp_pfn_shared(vmf);
3088
3089 pte_unmap_unlock(vmf->pte, vmf->ptl);
3090 return wp_page_copy(vmf);
3091 }
3092
3093 /*
3094 * Take out anonymous pages first, anonymous shared vmas are
3095 * not dirty accountable.
3096 */
3097 if (PageAnon(vmf->page)) {
3098 struct page *page = vmf->page;
3099
3100 /* PageKsm() doesn't necessarily raise the page refcount */
3101 if (PageKsm(page) || page_count(page) != 1)
3102 goto copy;
3103 if (!trylock_page(page))
3104 goto copy;
3105 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3106 unlock_page(page);
3107 goto copy;
3108 }
3109 /*
3110 * Ok, we've got the only map reference, and the only
3111 * page count reference, and the page is locked,
3112 * it's dark out, and we're wearing sunglasses. Hit it.
3113 */
3114 unlock_page(page);
3115 wp_page_reuse(vmf);
3116 return VM_FAULT_WRITE;
3117 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3118 (VM_WRITE|VM_SHARED))) {
3119 return wp_page_shared(vmf);
3120 }
3121copy:
3122 /*
3123 * Ok, we need to copy. Oh, well..
3124 */
3125 get_page(vmf->page);
3126
3127 pte_unmap_unlock(vmf->pte, vmf->ptl);
3128 return wp_page_copy(vmf);
3129}
3130
3131static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3132 unsigned long start_addr, unsigned long end_addr,
3133 struct zap_details *details)
3134{
3135 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3136}
3137
3138static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3139 struct zap_details *details)
3140{
3141 struct vm_area_struct *vma;
3142 pgoff_t vba, vea, zba, zea;
3143
3144 vma_interval_tree_foreach(vma, root,
3145 details->first_index, details->last_index) {
3146
3147 vba = vma->vm_pgoff;
3148 vea = vba + vma_pages(vma) - 1;
3149 zba = details->first_index;
3150 if (zba < vba)
3151 zba = vba;
3152 zea = details->last_index;
3153 if (zea > vea)
3154 zea = vea;
3155
3156 unmap_mapping_range_vma(vma,
3157 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3158 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3159 details);
3160 }
3161}
3162
3163/**
3164 * unmap_mapping_pages() - Unmap pages from processes.
3165 * @mapping: The address space containing pages to be unmapped.
3166 * @start: Index of first page to be unmapped.
3167 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3168 * @even_cows: Whether to unmap even private COWed pages.
3169 *
3170 * Unmap the pages in this address space from any userspace process which
3171 * has them mmaped. Generally, you want to remove COWed pages as well when
3172 * a file is being truncated, but not when invalidating pages from the page
3173 * cache.
3174 */
3175void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3176 pgoff_t nr, bool even_cows)
3177{
3178 struct zap_details details = { };
3179
3180 details.check_mapping = even_cows ? NULL : mapping;
3181 details.first_index = start;
3182 details.last_index = start + nr - 1;
3183 if (details.last_index < details.first_index)
3184 details.last_index = ULONG_MAX;
3185
3186 i_mmap_lock_write(mapping);
3187 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3188 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3189 i_mmap_unlock_write(mapping);
3190}
3191
3192/**
3193 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3194 * address_space corresponding to the specified byte range in the underlying
3195 * file.
3196 *
3197 * @mapping: the address space containing mmaps to be unmapped.
3198 * @holebegin: byte in first page to unmap, relative to the start of
3199 * the underlying file. This will be rounded down to a PAGE_SIZE
3200 * boundary. Note that this is different from truncate_pagecache(), which
3201 * must keep the partial page. In contrast, we must get rid of
3202 * partial pages.
3203 * @holelen: size of prospective hole in bytes. This will be rounded
3204 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3205 * end of the file.
3206 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3207 * but 0 when invalidating pagecache, don't throw away private data.
3208 */
3209void unmap_mapping_range(struct address_space *mapping,
3210 loff_t const holebegin, loff_t const holelen, int even_cows)
3211{
3212 pgoff_t hba = holebegin >> PAGE_SHIFT;
3213 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3214
3215 /* Check for overflow. */
3216 if (sizeof(holelen) > sizeof(hlen)) {
3217 long long holeend =
3218 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3219 if (holeend & ~(long long)ULONG_MAX)
3220 hlen = ULONG_MAX - hba + 1;
3221 }
3222
3223 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3224}
3225EXPORT_SYMBOL(unmap_mapping_range);
3226
3227/*
3228 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3229 * but allow concurrent faults), and pte mapped but not yet locked.
3230 * We return with pte unmapped and unlocked.
3231 *
3232 * We return with the mmap_lock locked or unlocked in the same cases
3233 * as does filemap_fault().
3234 */
3235vm_fault_t do_swap_page(struct vm_fault *vmf)
3236{
3237 struct vm_area_struct *vma = vmf->vma;
3238 struct page *page = NULL, *swapcache;
3239 swp_entry_t entry;
3240 pte_t pte;
3241 int locked;
3242 int exclusive = 0;
3243 vm_fault_t ret = 0;
3244 void *shadow = NULL;
3245
3246 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3247 goto out;
3248
3249 entry = pte_to_swp_entry(vmf->orig_pte);
3250 if (unlikely(non_swap_entry(entry))) {
3251 if (is_migration_entry(entry)) {
3252 migration_entry_wait(vma->vm_mm, vmf->pmd,
3253 vmf->address);
3254 } else if (is_device_private_entry(entry)) {
3255 vmf->page = device_private_entry_to_page(entry);
3256 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3257 } else if (is_hwpoison_entry(entry)) {
3258 ret = VM_FAULT_HWPOISON;
3259 } else {
3260 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3261 ret = VM_FAULT_SIGBUS;
3262 }
3263 goto out;
3264 }
3265
3266
3267 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3268 page = lookup_swap_cache(entry, vma, vmf->address);
3269 swapcache = page;
3270
3271 if (!page) {
3272 struct swap_info_struct *si = swp_swap_info(entry);
3273
3274 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3275 __swap_count(entry) == 1) {
3276 /* skip swapcache */
3277 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3278 vmf->address);
3279 if (page) {
3280 int err;
3281
3282 __SetPageLocked(page);
3283 __SetPageSwapBacked(page);
3284 set_page_private(page, entry.val);
3285
3286 /* Tell memcg to use swap ownership records */
3287 SetPageSwapCache(page);
3288 err = mem_cgroup_charge(page, vma->vm_mm,
3289 GFP_KERNEL);
3290 ClearPageSwapCache(page);
3291 if (err) {
3292 ret = VM_FAULT_OOM;
3293 goto out_page;
3294 }
3295
3296 shadow = get_shadow_from_swap_cache(entry);
3297 if (shadow)
3298 workingset_refault(page, shadow);
3299
3300 lru_cache_add(page);
3301 swap_readpage(page, true);
3302 }
3303 } else {
3304 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3305 vmf);
3306 swapcache = page;
3307 }
3308
3309 if (!page) {
3310 /*
3311 * Back out if somebody else faulted in this pte
3312 * while we released the pte lock.
3313 */
3314 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3315 vmf->address, &vmf->ptl);
3316 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3317 ret = VM_FAULT_OOM;
3318 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3319 goto unlock;
3320 }
3321
3322 /* Had to read the page from swap area: Major fault */
3323 ret = VM_FAULT_MAJOR;
3324 count_vm_event(PGMAJFAULT);
3325 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3326 } else if (PageHWPoison(page)) {
3327 /*
3328 * hwpoisoned dirty swapcache pages are kept for killing
3329 * owner processes (which may be unknown at hwpoison time)
3330 */
3331 ret = VM_FAULT_HWPOISON;
3332 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3333 goto out_release;
3334 }
3335
3336 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3337
3338 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3339 if (!locked) {
3340 ret |= VM_FAULT_RETRY;
3341 goto out_release;
3342 }
3343
3344 /*
3345 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3346 * release the swapcache from under us. The page pin, and pte_same
3347 * test below, are not enough to exclude that. Even if it is still
3348 * swapcache, we need to check that the page's swap has not changed.
3349 */
3350 if (unlikely((!PageSwapCache(page) ||
3351 page_private(page) != entry.val)) && swapcache)
3352 goto out_page;
3353
3354 page = ksm_might_need_to_copy(page, vma, vmf->address);
3355 if (unlikely(!page)) {
3356 ret = VM_FAULT_OOM;
3357 page = swapcache;
3358 goto out_page;
3359 }
3360
3361 cgroup_throttle_swaprate(page, GFP_KERNEL);
3362
3363 /*
3364 * Back out if somebody else already faulted in this pte.
3365 */
3366 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3367 &vmf->ptl);
3368 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3369 goto out_nomap;
3370
3371 if (unlikely(!PageUptodate(page))) {
3372 ret = VM_FAULT_SIGBUS;
3373 goto out_nomap;
3374 }
3375
3376 /*
3377 * The page isn't present yet, go ahead with the fault.
3378 *
3379 * Be careful about the sequence of operations here.
3380 * To get its accounting right, reuse_swap_page() must be called
3381 * while the page is counted on swap but not yet in mapcount i.e.
3382 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3383 * must be called after the swap_free(), or it will never succeed.
3384 */
3385
3386 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3387 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3388 pte = mk_pte(page, vma->vm_page_prot);
3389 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3390 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3391 vmf->flags &= ~FAULT_FLAG_WRITE;
3392 ret |= VM_FAULT_WRITE;
3393 exclusive = RMAP_EXCLUSIVE;
3394 }
3395 flush_icache_page(vma, page);
3396 if (pte_swp_soft_dirty(vmf->orig_pte))
3397 pte = pte_mksoft_dirty(pte);
3398 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3399 pte = pte_mkuffd_wp(pte);
3400 pte = pte_wrprotect(pte);
3401 }
3402 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3403 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3404 vmf->orig_pte = pte;
3405
3406 /* ksm created a completely new copy */
3407 if (unlikely(page != swapcache && swapcache)) {
3408 page_add_new_anon_rmap(page, vma, vmf->address, false);
3409 lru_cache_add_inactive_or_unevictable(page, vma);
3410 } else {
3411 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3412 }
3413
3414 swap_free(entry);
3415 if (mem_cgroup_swap_full(page) ||
3416 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3417 try_to_free_swap(page);
3418 unlock_page(page);
3419 if (page != swapcache && swapcache) {
3420 /*
3421 * Hold the lock to avoid the swap entry to be reused
3422 * until we take the PT lock for the pte_same() check
3423 * (to avoid false positives from pte_same). For
3424 * further safety release the lock after the swap_free
3425 * so that the swap count won't change under a
3426 * parallel locked swapcache.
3427 */
3428 unlock_page(swapcache);
3429 put_page(swapcache);
3430 }
3431
3432 if (vmf->flags & FAULT_FLAG_WRITE) {
3433 ret |= do_wp_page(vmf);
3434 if (ret & VM_FAULT_ERROR)
3435 ret &= VM_FAULT_ERROR;
3436 goto out;
3437 }
3438
3439 /* No need to invalidate - it was non-present before */
3440 update_mmu_cache(vma, vmf->address, vmf->pte);
3441unlock:
3442 pte_unmap_unlock(vmf->pte, vmf->ptl);
3443out:
3444 return ret;
3445out_nomap:
3446 pte_unmap_unlock(vmf->pte, vmf->ptl);
3447out_page:
3448 unlock_page(page);
3449out_release:
3450 put_page(page);
3451 if (page != swapcache && swapcache) {
3452 unlock_page(swapcache);
3453 put_page(swapcache);
3454 }
3455 return ret;
3456}
3457
3458/*
3459 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3460 * but allow concurrent faults), and pte mapped but not yet locked.
3461 * We return with mmap_lock still held, but pte unmapped and unlocked.
3462 */
3463static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3464{
3465 struct vm_area_struct *vma = vmf->vma;
3466 struct page *page;
3467 vm_fault_t ret = 0;
3468 pte_t entry;
3469
3470 /* File mapping without ->vm_ops ? */
3471 if (vma->vm_flags & VM_SHARED)
3472 return VM_FAULT_SIGBUS;
3473
3474 /*
3475 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3476 * pte_offset_map() on pmds where a huge pmd might be created
3477 * from a different thread.
3478 *
3479 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3480 * parallel threads are excluded by other means.
3481 *
3482 * Here we only have mmap_read_lock(mm).
3483 */
3484 if (pte_alloc(vma->vm_mm, vmf->pmd))
3485 return VM_FAULT_OOM;
3486
3487 /* See the comment in pte_alloc_one_map() */
3488 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3489 return 0;
3490
3491 /* Use the zero-page for reads */
3492 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3493 !mm_forbids_zeropage(vma->vm_mm)) {
3494 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3495 vma->vm_page_prot));
3496 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3497 vmf->address, &vmf->ptl);
3498 if (!pte_none(*vmf->pte)) {
3499 update_mmu_tlb(vma, vmf->address, vmf->pte);
3500 goto unlock;
3501 }
3502 ret = check_stable_address_space(vma->vm_mm);
3503 if (ret)
3504 goto unlock;
3505 /* Deliver the page fault to userland, check inside PT lock */
3506 if (userfaultfd_missing(vma)) {
3507 pte_unmap_unlock(vmf->pte, vmf->ptl);
3508 return handle_userfault(vmf, VM_UFFD_MISSING);
3509 }
3510 goto setpte;
3511 }
3512
3513 /* Allocate our own private page. */
3514 if (unlikely(anon_vma_prepare(vma)))
3515 goto oom;
3516 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3517 if (!page)
3518 goto oom;
3519
3520 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3521 goto oom_free_page;
3522 cgroup_throttle_swaprate(page, GFP_KERNEL);
3523
3524 /*
3525 * The memory barrier inside __SetPageUptodate makes sure that
3526 * preceding stores to the page contents become visible before
3527 * the set_pte_at() write.
3528 */
3529 __SetPageUptodate(page);
3530
3531 entry = mk_pte(page, vma->vm_page_prot);
3532 entry = pte_sw_mkyoung(entry);
3533 if (vma->vm_flags & VM_WRITE)
3534 entry = pte_mkwrite(pte_mkdirty(entry));
3535
3536 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3537 &vmf->ptl);
3538 if (!pte_none(*vmf->pte)) {
3539 update_mmu_cache(vma, vmf->address, vmf->pte);
3540 goto release;
3541 }
3542
3543 ret = check_stable_address_space(vma->vm_mm);
3544 if (ret)
3545 goto release;
3546
3547 /* Deliver the page fault to userland, check inside PT lock */
3548 if (userfaultfd_missing(vma)) {
3549 pte_unmap_unlock(vmf->pte, vmf->ptl);
3550 put_page(page);
3551 return handle_userfault(vmf, VM_UFFD_MISSING);
3552 }
3553
3554 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3555 page_add_new_anon_rmap(page, vma, vmf->address, false);
3556 lru_cache_add_inactive_or_unevictable(page, vma);
3557setpte:
3558 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3559
3560 /* No need to invalidate - it was non-present before */
3561 update_mmu_cache(vma, vmf->address, vmf->pte);
3562unlock:
3563 pte_unmap_unlock(vmf->pte, vmf->ptl);
3564 return ret;
3565release:
3566 put_page(page);
3567 goto unlock;
3568oom_free_page:
3569 put_page(page);
3570oom:
3571 return VM_FAULT_OOM;
3572}
3573
3574/*
3575 * The mmap_lock must have been held on entry, and may have been
3576 * released depending on flags and vma->vm_ops->fault() return value.
3577 * See filemap_fault() and __lock_page_retry().
3578 */
3579static vm_fault_t __do_fault(struct vm_fault *vmf)
3580{
3581 struct vm_area_struct *vma = vmf->vma;
3582 vm_fault_t ret;
3583
3584 /*
3585 * Preallocate pte before we take page_lock because this might lead to
3586 * deadlocks for memcg reclaim which waits for pages under writeback:
3587 * lock_page(A)
3588 * SetPageWriteback(A)
3589 * unlock_page(A)
3590 * lock_page(B)
3591 * lock_page(B)
3592 * pte_alloc_pne
3593 * shrink_page_list
3594 * wait_on_page_writeback(A)
3595 * SetPageWriteback(B)
3596 * unlock_page(B)
3597 * # flush A, B to clear the writeback
3598 */
3599 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3600 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3601 if (!vmf->prealloc_pte)
3602 return VM_FAULT_OOM;
3603 smp_wmb(); /* See comment in __pte_alloc() */
3604 }
3605
3606 ret = vma->vm_ops->fault(vmf);
3607 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3608 VM_FAULT_DONE_COW)))
3609 return ret;
3610
3611 if (unlikely(PageHWPoison(vmf->page))) {
3612 if (ret & VM_FAULT_LOCKED)
3613 unlock_page(vmf->page);
3614 put_page(vmf->page);
3615 vmf->page = NULL;
3616 return VM_FAULT_HWPOISON;
3617 }
3618
3619 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3620 lock_page(vmf->page);
3621 else
3622 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3623
3624 return ret;
3625}
3626
3627/*
3628 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3629 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3630 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3631 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3632 */
3633static int pmd_devmap_trans_unstable(pmd_t *pmd)
3634{
3635 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3636}
3637
3638static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3639{
3640 struct vm_area_struct *vma = vmf->vma;
3641
3642 if (!pmd_none(*vmf->pmd))
3643 goto map_pte;
3644 if (vmf->prealloc_pte) {
3645 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3646 if (unlikely(!pmd_none(*vmf->pmd))) {
3647 spin_unlock(vmf->ptl);
3648 goto map_pte;
3649 }
3650
3651 mm_inc_nr_ptes(vma->vm_mm);
3652 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3653 spin_unlock(vmf->ptl);
3654 vmf->prealloc_pte = NULL;
3655 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3656 return VM_FAULT_OOM;
3657 }
3658map_pte:
3659 /*
3660 * If a huge pmd materialized under us just retry later. Use
3661 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3662 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3663 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3664 * running immediately after a huge pmd fault in a different thread of
3665 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3666 * All we have to ensure is that it is a regular pmd that we can walk
3667 * with pte_offset_map() and we can do that through an atomic read in
3668 * C, which is what pmd_trans_unstable() provides.
3669 */
3670 if (pmd_devmap_trans_unstable(vmf->pmd))
3671 return VM_FAULT_NOPAGE;
3672
3673 /*
3674 * At this point we know that our vmf->pmd points to a page of ptes
3675 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3676 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3677 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3678 * be valid and we will re-check to make sure the vmf->pte isn't
3679 * pte_none() under vmf->ptl protection when we return to
3680 * alloc_set_pte().
3681 */
3682 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3683 &vmf->ptl);
3684 return 0;
3685}
3686
3687#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3688static void deposit_prealloc_pte(struct vm_fault *vmf)
3689{
3690 struct vm_area_struct *vma = vmf->vma;
3691
3692 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3693 /*
3694 * We are going to consume the prealloc table,
3695 * count that as nr_ptes.
3696 */
3697 mm_inc_nr_ptes(vma->vm_mm);
3698 vmf->prealloc_pte = NULL;
3699}
3700
3701static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3702{
3703 struct vm_area_struct *vma = vmf->vma;
3704 bool write = vmf->flags & FAULT_FLAG_WRITE;
3705 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3706 pmd_t entry;
3707 int i;
3708 vm_fault_t ret;
3709
3710 if (!transhuge_vma_suitable(vma, haddr))
3711 return VM_FAULT_FALLBACK;
3712
3713 ret = VM_FAULT_FALLBACK;
3714 page = compound_head(page);
3715
3716 /*
3717 * Archs like ppc64 need additonal space to store information
3718 * related to pte entry. Use the preallocated table for that.
3719 */
3720 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3721 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3722 if (!vmf->prealloc_pte)
3723 return VM_FAULT_OOM;
3724 smp_wmb(); /* See comment in __pte_alloc() */
3725 }
3726
3727 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3728 if (unlikely(!pmd_none(*vmf->pmd)))
3729 goto out;
3730
3731 for (i = 0; i < HPAGE_PMD_NR; i++)
3732 flush_icache_page(vma, page + i);
3733
3734 entry = mk_huge_pmd(page, vma->vm_page_prot);
3735 if (write)
3736 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3737
3738 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3739 page_add_file_rmap(page, true);
3740 /*
3741 * deposit and withdraw with pmd lock held
3742 */
3743 if (arch_needs_pgtable_deposit())
3744 deposit_prealloc_pte(vmf);
3745
3746 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3747
3748 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3749
3750 /* fault is handled */
3751 ret = 0;
3752 count_vm_event(THP_FILE_MAPPED);
3753out:
3754 spin_unlock(vmf->ptl);
3755 return ret;
3756}
3757#else
3758static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3759{
3760 BUILD_BUG();
3761 return 0;
3762}
3763#endif
3764
3765/**
3766 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3767 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3768 *
3769 * @vmf: fault environment
3770 * @page: page to map
3771 *
3772 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3773 * return.
3774 *
3775 * Target users are page handler itself and implementations of
3776 * vm_ops->map_pages.
3777 *
3778 * Return: %0 on success, %VM_FAULT_ code in case of error.
3779 */
3780vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct page *page)
3781{
3782 struct vm_area_struct *vma = vmf->vma;
3783 bool write = vmf->flags & FAULT_FLAG_WRITE;
3784 pte_t entry;
3785 vm_fault_t ret;
3786
3787 if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
3788 ret = do_set_pmd(vmf, page);
3789 if (ret != VM_FAULT_FALLBACK)
3790 return ret;
3791 }
3792
3793 if (!vmf->pte) {
3794 ret = pte_alloc_one_map(vmf);
3795 if (ret)
3796 return ret;
3797 }
3798
3799 /* Re-check under ptl */
3800 if (unlikely(!pte_none(*vmf->pte))) {
3801 update_mmu_tlb(vma, vmf->address, vmf->pte);
3802 return VM_FAULT_NOPAGE;
3803 }
3804
3805 flush_icache_page(vma, page);
3806 entry = mk_pte(page, vma->vm_page_prot);
3807 entry = pte_sw_mkyoung(entry);
3808 if (write)
3809 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3810 /* copy-on-write page */
3811 if (write && !(vma->vm_flags & VM_SHARED)) {
3812 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3813 page_add_new_anon_rmap(page, vma, vmf->address, false);
3814 lru_cache_add_inactive_or_unevictable(page, vma);
3815 } else {
3816 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3817 page_add_file_rmap(page, false);
3818 }
3819 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3820
3821 /* no need to invalidate: a not-present page won't be cached */
3822 update_mmu_cache(vma, vmf->address, vmf->pte);
3823
3824 return 0;
3825}
3826
3827
3828/**
3829 * finish_fault - finish page fault once we have prepared the page to fault
3830 *
3831 * @vmf: structure describing the fault
3832 *
3833 * This function handles all that is needed to finish a page fault once the
3834 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3835 * given page, adds reverse page mapping, handles memcg charges and LRU
3836 * addition.
3837 *
3838 * The function expects the page to be locked and on success it consumes a
3839 * reference of a page being mapped (for the PTE which maps it).
3840 *
3841 * Return: %0 on success, %VM_FAULT_ code in case of error.
3842 */
3843vm_fault_t finish_fault(struct vm_fault *vmf)
3844{
3845 struct page *page;
3846 vm_fault_t ret = 0;
3847
3848 /* Did we COW the page? */
3849 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3850 !(vmf->vma->vm_flags & VM_SHARED))
3851 page = vmf->cow_page;
3852 else
3853 page = vmf->page;
3854
3855 /*
3856 * check even for read faults because we might have lost our CoWed
3857 * page
3858 */
3859 if (!(vmf->vma->vm_flags & VM_SHARED))
3860 ret = check_stable_address_space(vmf->vma->vm_mm);
3861 if (!ret)
3862 ret = alloc_set_pte(vmf, page);
3863 if (vmf->pte)
3864 pte_unmap_unlock(vmf->pte, vmf->ptl);
3865 return ret;
3866}
3867
3868static unsigned long fault_around_bytes __read_mostly =
3869 rounddown_pow_of_two(65536);
3870
3871#ifdef CONFIG_DEBUG_FS
3872static int fault_around_bytes_get(void *data, u64 *val)
3873{
3874 *val = fault_around_bytes;
3875 return 0;
3876}
3877
3878/*
3879 * fault_around_bytes must be rounded down to the nearest page order as it's
3880 * what do_fault_around() expects to see.
3881 */
3882static int fault_around_bytes_set(void *data, u64 val)
3883{
3884 if (val / PAGE_SIZE > PTRS_PER_PTE)
3885 return -EINVAL;
3886 if (val > PAGE_SIZE)
3887 fault_around_bytes = rounddown_pow_of_two(val);
3888 else
3889 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3890 return 0;
3891}
3892DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3893 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3894
3895static int __init fault_around_debugfs(void)
3896{
3897 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3898 &fault_around_bytes_fops);
3899 return 0;
3900}
3901late_initcall(fault_around_debugfs);
3902#endif
3903
3904/*
3905 * do_fault_around() tries to map few pages around the fault address. The hope
3906 * is that the pages will be needed soon and this will lower the number of
3907 * faults to handle.
3908 *
3909 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3910 * not ready to be mapped: not up-to-date, locked, etc.
3911 *
3912 * This function is called with the page table lock taken. In the split ptlock
3913 * case the page table lock only protects only those entries which belong to
3914 * the page table corresponding to the fault address.
3915 *
3916 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3917 * only once.
3918 *
3919 * fault_around_bytes defines how many bytes we'll try to map.
3920 * do_fault_around() expects it to be set to a power of two less than or equal
3921 * to PTRS_PER_PTE.
3922 *
3923 * The virtual address of the area that we map is naturally aligned to
3924 * fault_around_bytes rounded down to the machine page size
3925 * (and therefore to page order). This way it's easier to guarantee
3926 * that we don't cross page table boundaries.
3927 */
3928static vm_fault_t do_fault_around(struct vm_fault *vmf)
3929{
3930 unsigned long address = vmf->address, nr_pages, mask;
3931 pgoff_t start_pgoff = vmf->pgoff;
3932 pgoff_t end_pgoff;
3933 int off;
3934 vm_fault_t ret = 0;
3935
3936 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3937 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3938
3939 vmf->address = max(address & mask, vmf->vma->vm_start);
3940 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3941 start_pgoff -= off;
3942
3943 /*
3944 * end_pgoff is either the end of the page table, the end of
3945 * the vma or nr_pages from start_pgoff, depending what is nearest.
3946 */
3947 end_pgoff = start_pgoff -
3948 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3949 PTRS_PER_PTE - 1;
3950 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3951 start_pgoff + nr_pages - 1);
3952
3953 if (pmd_none(*vmf->pmd)) {
3954 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3955 if (!vmf->prealloc_pte)
3956 goto out;
3957 smp_wmb(); /* See comment in __pte_alloc() */
3958 }
3959
3960 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3961
3962 /* Huge page is mapped? Page fault is solved */
3963 if (pmd_trans_huge(*vmf->pmd)) {
3964 ret = VM_FAULT_NOPAGE;
3965 goto out;
3966 }
3967
3968 /* ->map_pages() haven't done anything useful. Cold page cache? */
3969 if (!vmf->pte)
3970 goto out;
3971
3972 /* check if the page fault is solved */
3973 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3974 if (!pte_none(*vmf->pte))
3975 ret = VM_FAULT_NOPAGE;
3976 pte_unmap_unlock(vmf->pte, vmf->ptl);
3977out:
3978 vmf->address = address;
3979 vmf->pte = NULL;
3980 return ret;
3981}
3982
3983static vm_fault_t do_read_fault(struct vm_fault *vmf)
3984{
3985 struct vm_area_struct *vma = vmf->vma;
3986 vm_fault_t ret = 0;
3987
3988 /*
3989 * Let's call ->map_pages() first and use ->fault() as fallback
3990 * if page by the offset is not ready to be mapped (cold cache or
3991 * something).
3992 */
3993 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3994 ret = do_fault_around(vmf);
3995 if (ret)
3996 return ret;
3997 }
3998
3999 ret = __do_fault(vmf);
4000 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4001 return ret;
4002
4003 ret |= finish_fault(vmf);
4004 unlock_page(vmf->page);
4005 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4006 put_page(vmf->page);
4007 return ret;
4008}
4009
4010static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4011{
4012 struct vm_area_struct *vma = vmf->vma;
4013 vm_fault_t ret;
4014
4015 if (unlikely(anon_vma_prepare(vma)))
4016 return VM_FAULT_OOM;
4017
4018 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4019 if (!vmf->cow_page)
4020 return VM_FAULT_OOM;
4021
4022 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4023 put_page(vmf->cow_page);
4024 return VM_FAULT_OOM;
4025 }
4026 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4027
4028 ret = __do_fault(vmf);
4029 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4030 goto uncharge_out;
4031 if (ret & VM_FAULT_DONE_COW)
4032 return ret;
4033
4034 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4035 __SetPageUptodate(vmf->cow_page);
4036
4037 ret |= finish_fault(vmf);
4038 unlock_page(vmf->page);
4039 put_page(vmf->page);
4040 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4041 goto uncharge_out;
4042 return ret;
4043uncharge_out:
4044 put_page(vmf->cow_page);
4045 return ret;
4046}
4047
4048static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4049{
4050 struct vm_area_struct *vma = vmf->vma;
4051 vm_fault_t ret, tmp;
4052
4053 ret = __do_fault(vmf);
4054 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4055 return ret;
4056
4057 /*
4058 * Check if the backing address space wants to know that the page is
4059 * about to become writable
4060 */
4061 if (vma->vm_ops->page_mkwrite) {
4062 unlock_page(vmf->page);
4063 tmp = do_page_mkwrite(vmf);
4064 if (unlikely(!tmp ||
4065 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4066 put_page(vmf->page);
4067 return tmp;
4068 }
4069 }
4070
4071 ret |= finish_fault(vmf);
4072 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4073 VM_FAULT_RETRY))) {
4074 unlock_page(vmf->page);
4075 put_page(vmf->page);
4076 return ret;
4077 }
4078
4079 ret |= fault_dirty_shared_page(vmf);
4080 return ret;
4081}
4082
4083/*
4084 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4085 * but allow concurrent faults).
4086 * The mmap_lock may have been released depending on flags and our
4087 * return value. See filemap_fault() and __lock_page_or_retry().
4088 * If mmap_lock is released, vma may become invalid (for example
4089 * by other thread calling munmap()).
4090 */
4091static vm_fault_t do_fault(struct vm_fault *vmf)
4092{
4093 struct vm_area_struct *vma = vmf->vma;
4094 struct mm_struct *vm_mm = vma->vm_mm;
4095 vm_fault_t ret;
4096
4097 /*
4098 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4099 */
4100 if (!vma->vm_ops->fault) {
4101 /*
4102 * If we find a migration pmd entry or a none pmd entry, which
4103 * should never happen, return SIGBUS
4104 */
4105 if (unlikely(!pmd_present(*vmf->pmd)))
4106 ret = VM_FAULT_SIGBUS;
4107 else {
4108 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4109 vmf->pmd,
4110 vmf->address,
4111 &vmf->ptl);
4112 /*
4113 * Make sure this is not a temporary clearing of pte
4114 * by holding ptl and checking again. A R/M/W update
4115 * of pte involves: take ptl, clearing the pte so that
4116 * we don't have concurrent modification by hardware
4117 * followed by an update.
4118 */
4119 if (unlikely(pte_none(*vmf->pte)))
4120 ret = VM_FAULT_SIGBUS;
4121 else
4122 ret = VM_FAULT_NOPAGE;
4123
4124 pte_unmap_unlock(vmf->pte, vmf->ptl);
4125 }
4126 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4127 ret = do_read_fault(vmf);
4128 else if (!(vma->vm_flags & VM_SHARED))
4129 ret = do_cow_fault(vmf);
4130 else
4131 ret = do_shared_fault(vmf);
4132
4133 /* preallocated pagetable is unused: free it */
4134 if (vmf->prealloc_pte) {
4135 pte_free(vm_mm, vmf->prealloc_pte);
4136 vmf->prealloc_pte = NULL;
4137 }
4138 return ret;
4139}
4140
4141static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4142 unsigned long addr, int page_nid,
4143 int *flags)
4144{
4145 get_page(page);
4146
4147 count_vm_numa_event(NUMA_HINT_FAULTS);
4148 if (page_nid == numa_node_id()) {
4149 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4150 *flags |= TNF_FAULT_LOCAL;
4151 }
4152
4153 return mpol_misplaced(page, vma, addr);
4154}
4155
4156static vm_fault_t do_numa_page(struct vm_fault *vmf)
4157{
4158 struct vm_area_struct *vma = vmf->vma;
4159 struct page *page = NULL;
4160 int page_nid = NUMA_NO_NODE;
4161 int last_cpupid;
4162 int target_nid;
4163 bool migrated = false;
4164 pte_t pte, old_pte;
4165 bool was_writable = pte_savedwrite(vmf->orig_pte);
4166 int flags = 0;
4167
4168 /*
4169 * The "pte" at this point cannot be used safely without
4170 * validation through pte_unmap_same(). It's of NUMA type but
4171 * the pfn may be screwed if the read is non atomic.
4172 */
4173 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4174 spin_lock(vmf->ptl);
4175 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4176 pte_unmap_unlock(vmf->pte, vmf->ptl);
4177 goto out;
4178 }
4179
4180 /*
4181 * Make it present again, Depending on how arch implementes non
4182 * accessible ptes, some can allow access by kernel mode.
4183 */
4184 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4185 pte = pte_modify(old_pte, vma->vm_page_prot);
4186 pte = pte_mkyoung(pte);
4187 if (was_writable)
4188 pte = pte_mkwrite(pte);
4189 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4190 update_mmu_cache(vma, vmf->address, vmf->pte);
4191
4192 page = vm_normal_page(vma, vmf->address, pte);
4193 if (!page) {
4194 pte_unmap_unlock(vmf->pte, vmf->ptl);
4195 return 0;
4196 }
4197
4198 /* TODO: handle PTE-mapped THP */
4199 if (PageCompound(page)) {
4200 pte_unmap_unlock(vmf->pte, vmf->ptl);
4201 return 0;
4202 }
4203
4204 /*
4205 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4206 * much anyway since they can be in shared cache state. This misses
4207 * the case where a mapping is writable but the process never writes
4208 * to it but pte_write gets cleared during protection updates and
4209 * pte_dirty has unpredictable behaviour between PTE scan updates,
4210 * background writeback, dirty balancing and application behaviour.
4211 */
4212 if (!pte_write(pte))
4213 flags |= TNF_NO_GROUP;
4214
4215 /*
4216 * Flag if the page is shared between multiple address spaces. This
4217 * is later used when determining whether to group tasks together
4218 */
4219 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4220 flags |= TNF_SHARED;
4221
4222 last_cpupid = page_cpupid_last(page);
4223 page_nid = page_to_nid(page);
4224 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4225 &flags);
4226 pte_unmap_unlock(vmf->pte, vmf->ptl);
4227 if (target_nid == NUMA_NO_NODE) {
4228 put_page(page);
4229 goto out;
4230 }
4231
4232 /* Migrate to the requested node */
4233 migrated = migrate_misplaced_page(page, vma, target_nid);
4234 if (migrated) {
4235 page_nid = target_nid;
4236 flags |= TNF_MIGRATED;
4237 } else
4238 flags |= TNF_MIGRATE_FAIL;
4239
4240out:
4241 if (page_nid != NUMA_NO_NODE)
4242 task_numa_fault(last_cpupid, page_nid, 1, flags);
4243 return 0;
4244}
4245
4246static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4247{
4248 if (vma_is_anonymous(vmf->vma))
4249 return do_huge_pmd_anonymous_page(vmf);
4250 if (vmf->vma->vm_ops->huge_fault)
4251 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4252 return VM_FAULT_FALLBACK;
4253}
4254
4255/* `inline' is required to avoid gcc 4.1.2 build error */
4256static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4257{
4258 if (vma_is_anonymous(vmf->vma)) {
4259 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4260 return handle_userfault(vmf, VM_UFFD_WP);
4261 return do_huge_pmd_wp_page(vmf, orig_pmd);
4262 }
4263 if (vmf->vma->vm_ops->huge_fault) {
4264 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4265
4266 if (!(ret & VM_FAULT_FALLBACK))
4267 return ret;
4268 }
4269
4270 /* COW or write-notify handled on pte level: split pmd. */
4271 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4272
4273 return VM_FAULT_FALLBACK;
4274}
4275
4276static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4277{
4278#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4279 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4280 /* No support for anonymous transparent PUD pages yet */
4281 if (vma_is_anonymous(vmf->vma))
4282 goto split;
4283 if (vmf->vma->vm_ops->huge_fault) {
4284 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4285
4286 if (!(ret & VM_FAULT_FALLBACK))
4287 return ret;
4288 }
4289split:
4290 /* COW or write-notify not handled on PUD level: split pud.*/
4291 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4292#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4293 return VM_FAULT_FALLBACK;
4294}
4295
4296static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4297{
4298#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4299 /* No support for anonymous transparent PUD pages yet */
4300 if (vma_is_anonymous(vmf->vma))
4301 return VM_FAULT_FALLBACK;
4302 if (vmf->vma->vm_ops->huge_fault)
4303 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4304#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4305 return VM_FAULT_FALLBACK;
4306}
4307
4308/*
4309 * These routines also need to handle stuff like marking pages dirty
4310 * and/or accessed for architectures that don't do it in hardware (most
4311 * RISC architectures). The early dirtying is also good on the i386.
4312 *
4313 * There is also a hook called "update_mmu_cache()" that architectures
4314 * with external mmu caches can use to update those (ie the Sparc or
4315 * PowerPC hashed page tables that act as extended TLBs).
4316 *
4317 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4318 * concurrent faults).
4319 *
4320 * The mmap_lock may have been released depending on flags and our return value.
4321 * See filemap_fault() and __lock_page_or_retry().
4322 */
4323static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4324{
4325 pte_t entry;
4326
4327 if (unlikely(pmd_none(*vmf->pmd))) {
4328 /*
4329 * Leave __pte_alloc() until later: because vm_ops->fault may
4330 * want to allocate huge page, and if we expose page table
4331 * for an instant, it will be difficult to retract from
4332 * concurrent faults and from rmap lookups.
4333 */
4334 vmf->pte = NULL;
4335 } else {
4336 /* See comment in pte_alloc_one_map() */
4337 if (pmd_devmap_trans_unstable(vmf->pmd))
4338 return 0;
4339 /*
4340 * A regular pmd is established and it can't morph into a huge
4341 * pmd from under us anymore at this point because we hold the
4342 * mmap_lock read mode and khugepaged takes it in write mode.
4343 * So now it's safe to run pte_offset_map().
4344 */
4345 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4346 vmf->orig_pte = *vmf->pte;
4347
4348 /*
4349 * some architectures can have larger ptes than wordsize,
4350 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4351 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4352 * accesses. The code below just needs a consistent view
4353 * for the ifs and we later double check anyway with the
4354 * ptl lock held. So here a barrier will do.
4355 */
4356 barrier();
4357 if (pte_none(vmf->orig_pte)) {
4358 pte_unmap(vmf->pte);
4359 vmf->pte = NULL;
4360 }
4361 }
4362
4363 if (!vmf->pte) {
4364 if (vma_is_anonymous(vmf->vma))
4365 return do_anonymous_page(vmf);
4366 else
4367 return do_fault(vmf);
4368 }
4369
4370 if (!pte_present(vmf->orig_pte))
4371 return do_swap_page(vmf);
4372
4373 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4374 return do_numa_page(vmf);
4375
4376 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4377 spin_lock(vmf->ptl);
4378 entry = vmf->orig_pte;
4379 if (unlikely(!pte_same(*vmf->pte, entry))) {
4380 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4381 goto unlock;
4382 }
4383 if (vmf->flags & FAULT_FLAG_WRITE) {
4384 if (!pte_write(entry))
4385 return do_wp_page(vmf);
4386 entry = pte_mkdirty(entry);
4387 }
4388 entry = pte_mkyoung(entry);
4389 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4390 vmf->flags & FAULT_FLAG_WRITE)) {
4391 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4392 } else {
4393 /* Skip spurious TLB flush for retried page fault */
4394 if (vmf->flags & FAULT_FLAG_TRIED)
4395 goto unlock;
4396 /*
4397 * This is needed only for protection faults but the arch code
4398 * is not yet telling us if this is a protection fault or not.
4399 * This still avoids useless tlb flushes for .text page faults
4400 * with threads.
4401 */
4402 if (vmf->flags & FAULT_FLAG_WRITE)
4403 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4404 }
4405unlock:
4406 pte_unmap_unlock(vmf->pte, vmf->ptl);
4407 return 0;
4408}
4409
4410/*
4411 * By the time we get here, we already hold the mm semaphore
4412 *
4413 * The mmap_lock may have been released depending on flags and our
4414 * return value. See filemap_fault() and __lock_page_or_retry().
4415 */
4416static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4417 unsigned long address, unsigned int flags)
4418{
4419 struct vm_fault vmf = {
4420 .vma = vma,
4421 .address = address & PAGE_MASK,
4422 .flags = flags,
4423 .pgoff = linear_page_index(vma, address),
4424 .gfp_mask = __get_fault_gfp_mask(vma),
4425 };
4426 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4427 struct mm_struct *mm = vma->vm_mm;
4428 pgd_t *pgd;
4429 p4d_t *p4d;
4430 vm_fault_t ret;
4431
4432 pgd = pgd_offset(mm, address);
4433 p4d = p4d_alloc(mm, pgd, address);
4434 if (!p4d)
4435 return VM_FAULT_OOM;
4436
4437 vmf.pud = pud_alloc(mm, p4d, address);
4438 if (!vmf.pud)
4439 return VM_FAULT_OOM;
4440retry_pud:
4441 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4442 ret = create_huge_pud(&vmf);
4443 if (!(ret & VM_FAULT_FALLBACK))
4444 return ret;
4445 } else {
4446 pud_t orig_pud = *vmf.pud;
4447
4448 barrier();
4449 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4450
4451 /* NUMA case for anonymous PUDs would go here */
4452
4453 if (dirty && !pud_write(orig_pud)) {
4454 ret = wp_huge_pud(&vmf, orig_pud);
4455 if (!(ret & VM_FAULT_FALLBACK))
4456 return ret;
4457 } else {
4458 huge_pud_set_accessed(&vmf, orig_pud);
4459 return 0;
4460 }
4461 }
4462 }
4463
4464 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4465 if (!vmf.pmd)
4466 return VM_FAULT_OOM;
4467
4468 /* Huge pud page fault raced with pmd_alloc? */
4469 if (pud_trans_unstable(vmf.pud))
4470 goto retry_pud;
4471
4472 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4473 ret = create_huge_pmd(&vmf);
4474 if (!(ret & VM_FAULT_FALLBACK))
4475 return ret;
4476 } else {
4477 pmd_t orig_pmd = *vmf.pmd;
4478
4479 barrier();
4480 if (unlikely(is_swap_pmd(orig_pmd))) {
4481 VM_BUG_ON(thp_migration_supported() &&
4482 !is_pmd_migration_entry(orig_pmd));
4483 if (is_pmd_migration_entry(orig_pmd))
4484 pmd_migration_entry_wait(mm, vmf.pmd);
4485 return 0;
4486 }
4487 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4488 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4489 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4490
4491 if (dirty && !pmd_write(orig_pmd)) {
4492 ret = wp_huge_pmd(&vmf, orig_pmd);
4493 if (!(ret & VM_FAULT_FALLBACK))
4494 return ret;
4495 } else {
4496 huge_pmd_set_accessed(&vmf, orig_pmd);
4497 return 0;
4498 }
4499 }
4500 }
4501
4502 return handle_pte_fault(&vmf);
4503}
4504
4505/**
4506 * mm_account_fault - Do page fault accountings
4507 *
4508 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4509 * of perf event counters, but we'll still do the per-task accounting to
4510 * the task who triggered this page fault.
4511 * @address: the faulted address.
4512 * @flags: the fault flags.
4513 * @ret: the fault retcode.
4514 *
4515 * This will take care of most of the page fault accountings. Meanwhile, it
4516 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4517 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4518 * still be in per-arch page fault handlers at the entry of page fault.
4519 */
4520static inline void mm_account_fault(struct pt_regs *regs,
4521 unsigned long address, unsigned int flags,
4522 vm_fault_t ret)
4523{
4524 bool major;
4525
4526 /*
4527 * We don't do accounting for some specific faults:
4528 *
4529 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4530 * includes arch_vma_access_permitted() failing before reaching here.
4531 * So this is not a "this many hardware page faults" counter. We
4532 * should use the hw profiling for that.
4533 *
4534 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4535 * once they're completed.
4536 */
4537 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4538 return;
4539
4540 /*
4541 * We define the fault as a major fault when the final successful fault
4542 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4543 * handle it immediately previously).
4544 */
4545 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4546
4547 if (major)
4548 current->maj_flt++;
4549 else
4550 current->min_flt++;
4551
4552 /*
4553 * If the fault is done for GUP, regs will be NULL. We only do the
4554 * accounting for the per thread fault counters who triggered the
4555 * fault, and we skip the perf event updates.
4556 */
4557 if (!regs)
4558 return;
4559
4560 if (major)
4561 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4562 else
4563 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4564}
4565
4566/*
4567 * By the time we get here, we already hold the mm semaphore
4568 *
4569 * The mmap_lock may have been released depending on flags and our
4570 * return value. See filemap_fault() and __lock_page_or_retry().
4571 */
4572vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4573 unsigned int flags, struct pt_regs *regs)
4574{
4575 vm_fault_t ret;
4576
4577 __set_current_state(TASK_RUNNING);
4578
4579 count_vm_event(PGFAULT);
4580 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4581
4582 /* do counter updates before entering really critical section. */
4583 check_sync_rss_stat(current);
4584
4585 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4586 flags & FAULT_FLAG_INSTRUCTION,
4587 flags & FAULT_FLAG_REMOTE))
4588 return VM_FAULT_SIGSEGV;
4589
4590 /*
4591 * Enable the memcg OOM handling for faults triggered in user
4592 * space. Kernel faults are handled more gracefully.
4593 */
4594 if (flags & FAULT_FLAG_USER)
4595 mem_cgroup_enter_user_fault();
4596
4597 if (unlikely(is_vm_hugetlb_page(vma)))
4598 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4599 else
4600 ret = __handle_mm_fault(vma, address, flags);
4601
4602 if (flags & FAULT_FLAG_USER) {
4603 mem_cgroup_exit_user_fault();
4604 /*
4605 * The task may have entered a memcg OOM situation but
4606 * if the allocation error was handled gracefully (no
4607 * VM_FAULT_OOM), there is no need to kill anything.
4608 * Just clean up the OOM state peacefully.
4609 */
4610 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4611 mem_cgroup_oom_synchronize(false);
4612 }
4613
4614 mm_account_fault(regs, address, flags, ret);
4615
4616 return ret;
4617}
4618EXPORT_SYMBOL_GPL(handle_mm_fault);
4619
4620#ifndef __PAGETABLE_P4D_FOLDED
4621/*
4622 * Allocate p4d page table.
4623 * We've already handled the fast-path in-line.
4624 */
4625int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4626{
4627 p4d_t *new = p4d_alloc_one(mm, address);
4628 if (!new)
4629 return -ENOMEM;
4630
4631 smp_wmb(); /* See comment in __pte_alloc */
4632
4633 spin_lock(&mm->page_table_lock);
4634 if (pgd_present(*pgd)) /* Another has populated it */
4635 p4d_free(mm, new);
4636 else
4637 pgd_populate(mm, pgd, new);
4638 spin_unlock(&mm->page_table_lock);
4639 return 0;
4640}
4641#endif /* __PAGETABLE_P4D_FOLDED */
4642
4643#ifndef __PAGETABLE_PUD_FOLDED
4644/*
4645 * Allocate page upper directory.
4646 * We've already handled the fast-path in-line.
4647 */
4648int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4649{
4650 pud_t *new = pud_alloc_one(mm, address);
4651 if (!new)
4652 return -ENOMEM;
4653
4654 smp_wmb(); /* See comment in __pte_alloc */
4655
4656 spin_lock(&mm->page_table_lock);
4657 if (!p4d_present(*p4d)) {
4658 mm_inc_nr_puds(mm);
4659 p4d_populate(mm, p4d, new);
4660 } else /* Another has populated it */
4661 pud_free(mm, new);
4662 spin_unlock(&mm->page_table_lock);
4663 return 0;
4664}
4665#endif /* __PAGETABLE_PUD_FOLDED */
4666
4667#ifndef __PAGETABLE_PMD_FOLDED
4668/*
4669 * Allocate page middle directory.
4670 * We've already handled the fast-path in-line.
4671 */
4672int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4673{
4674 spinlock_t *ptl;
4675 pmd_t *new = pmd_alloc_one(mm, address);
4676 if (!new)
4677 return -ENOMEM;
4678
4679 smp_wmb(); /* See comment in __pte_alloc */
4680
4681 ptl = pud_lock(mm, pud);
4682 if (!pud_present(*pud)) {
4683 mm_inc_nr_pmds(mm);
4684 pud_populate(mm, pud, new);
4685 } else /* Another has populated it */
4686 pmd_free(mm, new);
4687 spin_unlock(ptl);
4688 return 0;
4689}
4690#endif /* __PAGETABLE_PMD_FOLDED */
4691
4692static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4693 struct mmu_notifier_range *range,
4694 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4695{
4696 pgd_t *pgd;
4697 p4d_t *p4d;
4698 pud_t *pud;
4699 pmd_t *pmd;
4700 pte_t *ptep;
4701
4702 pgd = pgd_offset(mm, address);
4703 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4704 goto out;
4705
4706 p4d = p4d_offset(pgd, address);
4707 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4708 goto out;
4709
4710 pud = pud_offset(p4d, address);
4711 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4712 goto out;
4713
4714 pmd = pmd_offset(pud, address);
4715 VM_BUG_ON(pmd_trans_huge(*pmd));
4716
4717 if (pmd_huge(*pmd)) {
4718 if (!pmdpp)
4719 goto out;
4720
4721 if (range) {
4722 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4723 NULL, mm, address & PMD_MASK,
4724 (address & PMD_MASK) + PMD_SIZE);
4725 mmu_notifier_invalidate_range_start(range);
4726 }
4727 *ptlp = pmd_lock(mm, pmd);
4728 if (pmd_huge(*pmd)) {
4729 *pmdpp = pmd;
4730 return 0;
4731 }
4732 spin_unlock(*ptlp);
4733 if (range)
4734 mmu_notifier_invalidate_range_end(range);
4735 }
4736
4737 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4738 goto out;
4739
4740 if (range) {
4741 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4742 address & PAGE_MASK,
4743 (address & PAGE_MASK) + PAGE_SIZE);
4744 mmu_notifier_invalidate_range_start(range);
4745 }
4746 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4747 if (!pte_present(*ptep))
4748 goto unlock;
4749 *ptepp = ptep;
4750 return 0;
4751unlock:
4752 pte_unmap_unlock(ptep, *ptlp);
4753 if (range)
4754 mmu_notifier_invalidate_range_end(range);
4755out:
4756 return -EINVAL;
4757}
4758
4759static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4760 pte_t **ptepp, spinlock_t **ptlp)
4761{
4762 int res;
4763
4764 /* (void) is needed to make gcc happy */
4765 (void) __cond_lock(*ptlp,
4766 !(res = __follow_pte_pmd(mm, address, NULL,
4767 ptepp, NULL, ptlp)));
4768 return res;
4769}
4770
4771int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4772 struct mmu_notifier_range *range,
4773 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4774{
4775 int res;
4776
4777 /* (void) is needed to make gcc happy */
4778 (void) __cond_lock(*ptlp,
4779 !(res = __follow_pte_pmd(mm, address, range,
4780 ptepp, pmdpp, ptlp)));
4781 return res;
4782}
4783EXPORT_SYMBOL(follow_pte_pmd);
4784
4785/**
4786 * follow_pfn - look up PFN at a user virtual address
4787 * @vma: memory mapping
4788 * @address: user virtual address
4789 * @pfn: location to store found PFN
4790 *
4791 * Only IO mappings and raw PFN mappings are allowed.
4792 *
4793 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4794 */
4795int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4796 unsigned long *pfn)
4797{
4798 int ret = -EINVAL;
4799 spinlock_t *ptl;
4800 pte_t *ptep;
4801
4802 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4803 return ret;
4804
4805 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4806 if (ret)
4807 return ret;
4808 *pfn = pte_pfn(*ptep);
4809 pte_unmap_unlock(ptep, ptl);
4810 return 0;
4811}
4812EXPORT_SYMBOL(follow_pfn);
4813
4814#ifdef CONFIG_HAVE_IOREMAP_PROT
4815int follow_phys(struct vm_area_struct *vma,
4816 unsigned long address, unsigned int flags,
4817 unsigned long *prot, resource_size_t *phys)
4818{
4819 int ret = -EINVAL;
4820 pte_t *ptep, pte;
4821 spinlock_t *ptl;
4822
4823 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4824 goto out;
4825
4826 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4827 goto out;
4828 pte = *ptep;
4829
4830 if ((flags & FOLL_WRITE) && !pte_write(pte))
4831 goto unlock;
4832
4833 *prot = pgprot_val(pte_pgprot(pte));
4834 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4835
4836 ret = 0;
4837unlock:
4838 pte_unmap_unlock(ptep, ptl);
4839out:
4840 return ret;
4841}
4842
4843int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4844 void *buf, int len, int write)
4845{
4846 resource_size_t phys_addr;
4847 unsigned long prot = 0;
4848 void __iomem *maddr;
4849 int offset = addr & (PAGE_SIZE-1);
4850
4851 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4852 return -EINVAL;
4853
4854 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4855 if (!maddr)
4856 return -ENOMEM;
4857
4858 if (write)
4859 memcpy_toio(maddr + offset, buf, len);
4860 else
4861 memcpy_fromio(buf, maddr + offset, len);
4862 iounmap(maddr);
4863
4864 return len;
4865}
4866EXPORT_SYMBOL_GPL(generic_access_phys);
4867#endif
4868
4869/*
4870 * Access another process' address space as given in mm. If non-NULL, use the
4871 * given task for page fault accounting.
4872 */
4873int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4874 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4875{
4876 struct vm_area_struct *vma;
4877 void *old_buf = buf;
4878 int write = gup_flags & FOLL_WRITE;
4879
4880 if (mmap_read_lock_killable(mm))
4881 return 0;
4882
4883 /* ignore errors, just check how much was successfully transferred */
4884 while (len) {
4885 int bytes, ret, offset;
4886 void *maddr;
4887 struct page *page = NULL;
4888
4889 ret = get_user_pages_remote(mm, addr, 1,
4890 gup_flags, &page, &vma, NULL);
4891 if (ret <= 0) {
4892#ifndef CONFIG_HAVE_IOREMAP_PROT
4893 break;
4894#else
4895 /*
4896 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4897 * we can access using slightly different code.
4898 */
4899 vma = find_vma(mm, addr);
4900 if (!vma || vma->vm_start > addr)
4901 break;
4902 if (vma->vm_ops && vma->vm_ops->access)
4903 ret = vma->vm_ops->access(vma, addr, buf,
4904 len, write);
4905 if (ret <= 0)
4906 break;
4907 bytes = ret;
4908#endif
4909 } else {
4910 bytes = len;
4911 offset = addr & (PAGE_SIZE-1);
4912 if (bytes > PAGE_SIZE-offset)
4913 bytes = PAGE_SIZE-offset;
4914
4915 maddr = kmap(page);
4916 if (write) {
4917 copy_to_user_page(vma, page, addr,
4918 maddr + offset, buf, bytes);
4919 set_page_dirty_lock(page);
4920 } else {
4921 copy_from_user_page(vma, page, addr,
4922 buf, maddr + offset, bytes);
4923 }
4924 kunmap(page);
4925 put_page(page);
4926 }
4927 len -= bytes;
4928 buf += bytes;
4929 addr += bytes;
4930 }
4931 mmap_read_unlock(mm);
4932
4933 return buf - old_buf;
4934}
4935
4936/**
4937 * access_remote_vm - access another process' address space
4938 * @mm: the mm_struct of the target address space
4939 * @addr: start address to access
4940 * @buf: source or destination buffer
4941 * @len: number of bytes to transfer
4942 * @gup_flags: flags modifying lookup behaviour
4943 *
4944 * The caller must hold a reference on @mm.
4945 *
4946 * Return: number of bytes copied from source to destination.
4947 */
4948int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4949 void *buf, int len, unsigned int gup_flags)
4950{
4951 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4952}
4953
4954/*
4955 * Access another process' address space.
4956 * Source/target buffer must be kernel space,
4957 * Do not walk the page table directly, use get_user_pages
4958 */
4959int access_process_vm(struct task_struct *tsk, unsigned long addr,
4960 void *buf, int len, unsigned int gup_flags)
4961{
4962 struct mm_struct *mm;
4963 int ret;
4964
4965 mm = get_task_mm(tsk);
4966 if (!mm)
4967 return 0;
4968
4969 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4970
4971 mmput(mm);
4972
4973 return ret;
4974}
4975EXPORT_SYMBOL_GPL(access_process_vm);
4976
4977/*
4978 * Print the name of a VMA.
4979 */
4980void print_vma_addr(char *prefix, unsigned long ip)
4981{
4982 struct mm_struct *mm = current->mm;
4983 struct vm_area_struct *vma;
4984
4985 /*
4986 * we might be running from an atomic context so we cannot sleep
4987 */
4988 if (!mmap_read_trylock(mm))
4989 return;
4990
4991 vma = find_vma(mm, ip);
4992 if (vma && vma->vm_file) {
4993 struct file *f = vma->vm_file;
4994 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4995 if (buf) {
4996 char *p;
4997
4998 p = file_path(f, buf, PAGE_SIZE);
4999 if (IS_ERR(p))
5000 p = "?";
5001 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5002 vma->vm_start,
5003 vma->vm_end - vma->vm_start);
5004 free_page((unsigned long)buf);
5005 }
5006 }
5007 mmap_read_unlock(mm);
5008}
5009
5010#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5011void __might_fault(const char *file, int line)
5012{
5013 /*
5014 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5015 * holding the mmap_lock, this is safe because kernel memory doesn't
5016 * get paged out, therefore we'll never actually fault, and the
5017 * below annotations will generate false positives.
5018 */
5019 if (uaccess_kernel())
5020 return;
5021 if (pagefault_disabled())
5022 return;
5023 __might_sleep(file, line, 0);
5024#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5025 if (current->mm)
5026 might_lock_read(¤t->mm->mmap_lock);
5027#endif
5028}
5029EXPORT_SYMBOL(__might_fault);
5030#endif
5031
5032#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5033/*
5034 * Process all subpages of the specified huge page with the specified
5035 * operation. The target subpage will be processed last to keep its
5036 * cache lines hot.
5037 */
5038static inline void process_huge_page(
5039 unsigned long addr_hint, unsigned int pages_per_huge_page,
5040 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5041 void *arg)
5042{
5043 int i, n, base, l;
5044 unsigned long addr = addr_hint &
5045 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5046
5047 /* Process target subpage last to keep its cache lines hot */
5048 might_sleep();
5049 n = (addr_hint - addr) / PAGE_SIZE;
5050 if (2 * n <= pages_per_huge_page) {
5051 /* If target subpage in first half of huge page */
5052 base = 0;
5053 l = n;
5054 /* Process subpages at the end of huge page */
5055 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5056 cond_resched();
5057 process_subpage(addr + i * PAGE_SIZE, i, arg);
5058 }
5059 } else {
5060 /* If target subpage in second half of huge page */
5061 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5062 l = pages_per_huge_page - n;
5063 /* Process subpages at the begin of huge page */
5064 for (i = 0; i < base; i++) {
5065 cond_resched();
5066 process_subpage(addr + i * PAGE_SIZE, i, arg);
5067 }
5068 }
5069 /*
5070 * Process remaining subpages in left-right-left-right pattern
5071 * towards the target subpage
5072 */
5073 for (i = 0; i < l; i++) {
5074 int left_idx = base + i;
5075 int right_idx = base + 2 * l - 1 - i;
5076
5077 cond_resched();
5078 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5079 cond_resched();
5080 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5081 }
5082}
5083
5084static void clear_gigantic_page(struct page *page,
5085 unsigned long addr,
5086 unsigned int pages_per_huge_page)
5087{
5088 int i;
5089 struct page *p = page;
5090
5091 might_sleep();
5092 for (i = 0; i < pages_per_huge_page;
5093 i++, p = mem_map_next(p, page, i)) {
5094 cond_resched();
5095 clear_user_highpage(p, addr + i * PAGE_SIZE);
5096 }
5097}
5098
5099static void clear_subpage(unsigned long addr, int idx, void *arg)
5100{
5101 struct page *page = arg;
5102
5103 clear_user_highpage(page + idx, addr);
5104}
5105
5106void clear_huge_page(struct page *page,
5107 unsigned long addr_hint, unsigned int pages_per_huge_page)
5108{
5109 unsigned long addr = addr_hint &
5110 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5111
5112 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5113 clear_gigantic_page(page, addr, pages_per_huge_page);
5114 return;
5115 }
5116
5117 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5118}
5119
5120static void copy_user_gigantic_page(struct page *dst, struct page *src,
5121 unsigned long addr,
5122 struct vm_area_struct *vma,
5123 unsigned int pages_per_huge_page)
5124{
5125 int i;
5126 struct page *dst_base = dst;
5127 struct page *src_base = src;
5128
5129 for (i = 0; i < pages_per_huge_page; ) {
5130 cond_resched();
5131 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5132
5133 i++;
5134 dst = mem_map_next(dst, dst_base, i);
5135 src = mem_map_next(src, src_base, i);
5136 }
5137}
5138
5139struct copy_subpage_arg {
5140 struct page *dst;
5141 struct page *src;
5142 struct vm_area_struct *vma;
5143};
5144
5145static void copy_subpage(unsigned long addr, int idx, void *arg)
5146{
5147 struct copy_subpage_arg *copy_arg = arg;
5148
5149 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5150 addr, copy_arg->vma);
5151}
5152
5153void copy_user_huge_page(struct page *dst, struct page *src,
5154 unsigned long addr_hint, struct vm_area_struct *vma,
5155 unsigned int pages_per_huge_page)
5156{
5157 unsigned long addr = addr_hint &
5158 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5159 struct copy_subpage_arg arg = {
5160 .dst = dst,
5161 .src = src,
5162 .vma = vma,
5163 };
5164
5165 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5166 copy_user_gigantic_page(dst, src, addr, vma,
5167 pages_per_huge_page);
5168 return;
5169 }
5170
5171 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5172}
5173
5174long copy_huge_page_from_user(struct page *dst_page,
5175 const void __user *usr_src,
5176 unsigned int pages_per_huge_page,
5177 bool allow_pagefault)
5178{
5179 void *src = (void *)usr_src;
5180 void *page_kaddr;
5181 unsigned long i, rc = 0;
5182 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5183
5184 for (i = 0; i < pages_per_huge_page; i++) {
5185 if (allow_pagefault)
5186 page_kaddr = kmap(dst_page + i);
5187 else
5188 page_kaddr = kmap_atomic(dst_page + i);
5189 rc = copy_from_user(page_kaddr,
5190 (const void __user *)(src + i * PAGE_SIZE),
5191 PAGE_SIZE);
5192 if (allow_pagefault)
5193 kunmap(dst_page + i);
5194 else
5195 kunmap_atomic(page_kaddr);
5196
5197 ret_val -= (PAGE_SIZE - rc);
5198 if (rc)
5199 break;
5200
5201 cond_resched();
5202 }
5203 return ret_val;
5204}
5205#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5206
5207#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5208
5209static struct kmem_cache *page_ptl_cachep;
5210
5211void __init ptlock_cache_init(void)
5212{
5213 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5214 SLAB_PANIC, NULL);
5215}
5216
5217bool ptlock_alloc(struct page *page)
5218{
5219 spinlock_t *ptl;
5220
5221 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5222 if (!ptl)
5223 return false;
5224 page->ptl = ptl;
5225 return true;
5226}
5227
5228void ptlock_free(struct page *page)
5229{
5230 kmem_cache_free(page_ptl_cachep, page->ptl);
5231}
5232#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#include <linux/migrate.h>
61#include <linux/string.h>
62#include <linux/dma-debug.h>
63#include <linux/debugfs.h>
64
65#include <asm/io.h>
66#include <asm/pgalloc.h>
67#include <asm/uaccess.h>
68#include <asm/tlb.h>
69#include <asm/tlbflush.h>
70#include <asm/pgtable.h>
71
72#include "internal.h"
73
74#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
76#endif
77
78#ifndef CONFIG_NEED_MULTIPLE_NODES
79/* use the per-pgdat data instead for discontigmem - mbligh */
80unsigned long max_mapnr;
81struct page *mem_map;
82
83EXPORT_SYMBOL(max_mapnr);
84EXPORT_SYMBOL(mem_map);
85#endif
86
87/*
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92 * and ZONE_HIGHMEM.
93 */
94void * high_memory;
95
96EXPORT_SYMBOL(high_memory);
97
98/*
99 * Randomize the address space (stacks, mmaps, brk, etc.).
100 *
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
103 */
104int randomize_va_space __read_mostly =
105#ifdef CONFIG_COMPAT_BRK
106 1;
107#else
108 2;
109#endif
110
111static int __init disable_randmaps(char *s)
112{
113 randomize_va_space = 0;
114 return 1;
115}
116__setup("norandmaps", disable_randmaps);
117
118unsigned long zero_pfn __read_mostly;
119unsigned long highest_memmap_pfn __read_mostly;
120
121/*
122 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123 */
124static int __init init_zero_pfn(void)
125{
126 zero_pfn = page_to_pfn(ZERO_PAGE(0));
127 return 0;
128}
129core_initcall(init_zero_pfn);
130
131
132#if defined(SPLIT_RSS_COUNTING)
133
134void sync_mm_rss(struct mm_struct *mm)
135{
136 int i;
137
138 for (i = 0; i < NR_MM_COUNTERS; i++) {
139 if (current->rss_stat.count[i]) {
140 add_mm_counter(mm, i, current->rss_stat.count[i]);
141 current->rss_stat.count[i] = 0;
142 }
143 }
144 current->rss_stat.events = 0;
145}
146
147static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148{
149 struct task_struct *task = current;
150
151 if (likely(task->mm == mm))
152 task->rss_stat.count[member] += val;
153 else
154 add_mm_counter(mm, member, val);
155}
156#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159/* sync counter once per 64 page faults */
160#define TASK_RSS_EVENTS_THRESH (64)
161static void check_sync_rss_stat(struct task_struct *task)
162{
163 if (unlikely(task != current))
164 return;
165 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
166 sync_mm_rss(task->mm);
167}
168#else /* SPLIT_RSS_COUNTING */
169
170#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173static void check_sync_rss_stat(struct task_struct *task)
174{
175}
176
177#endif /* SPLIT_RSS_COUNTING */
178
179#ifdef HAVE_GENERIC_MMU_GATHER
180
181static int tlb_next_batch(struct mmu_gather *tlb)
182{
183 struct mmu_gather_batch *batch;
184
185 batch = tlb->active;
186 if (batch->next) {
187 tlb->active = batch->next;
188 return 1;
189 }
190
191 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192 return 0;
193
194 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195 if (!batch)
196 return 0;
197
198 tlb->batch_count++;
199 batch->next = NULL;
200 batch->nr = 0;
201 batch->max = MAX_GATHER_BATCH;
202
203 tlb->active->next = batch;
204 tlb->active = batch;
205
206 return 1;
207}
208
209/* tlb_gather_mmu
210 * Called to initialize an (on-stack) mmu_gather structure for page-table
211 * tear-down from @mm. The @fullmm argument is used when @mm is without
212 * users and we're going to destroy the full address space (exit/execve).
213 */
214void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
215{
216 tlb->mm = mm;
217
218 /* Is it from 0 to ~0? */
219 tlb->fullmm = !(start | (end+1));
220 tlb->need_flush_all = 0;
221 tlb->start = start;
222 tlb->end = end;
223 tlb->need_flush = 0;
224 tlb->local.next = NULL;
225 tlb->local.nr = 0;
226 tlb->local.max = ARRAY_SIZE(tlb->__pages);
227 tlb->active = &tlb->local;
228 tlb->batch_count = 0;
229
230#ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 tlb->batch = NULL;
232#endif
233}
234
235static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
236{
237 tlb->need_flush = 0;
238 tlb_flush(tlb);
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb_table_flush(tlb);
241#endif
242}
243
244static void tlb_flush_mmu_free(struct mmu_gather *tlb)
245{
246 struct mmu_gather_batch *batch;
247
248 for (batch = &tlb->local; batch; batch = batch->next) {
249 free_pages_and_swap_cache(batch->pages, batch->nr);
250 batch->nr = 0;
251 }
252 tlb->active = &tlb->local;
253}
254
255void tlb_flush_mmu(struct mmu_gather *tlb)
256{
257 if (!tlb->need_flush)
258 return;
259 tlb_flush_mmu_tlbonly(tlb);
260 tlb_flush_mmu_free(tlb);
261}
262
263/* tlb_finish_mmu
264 * Called at the end of the shootdown operation to free up any resources
265 * that were required.
266 */
267void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
268{
269 struct mmu_gather_batch *batch, *next;
270
271 tlb_flush_mmu(tlb);
272
273 /* keep the page table cache within bounds */
274 check_pgt_cache();
275
276 for (batch = tlb->local.next; batch; batch = next) {
277 next = batch->next;
278 free_pages((unsigned long)batch, 0);
279 }
280 tlb->local.next = NULL;
281}
282
283/* __tlb_remove_page
284 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
285 * handling the additional races in SMP caused by other CPUs caching valid
286 * mappings in their TLBs. Returns the number of free page slots left.
287 * When out of page slots we must call tlb_flush_mmu().
288 */
289int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
290{
291 struct mmu_gather_batch *batch;
292
293 VM_BUG_ON(!tlb->need_flush);
294
295 batch = tlb->active;
296 batch->pages[batch->nr++] = page;
297 if (batch->nr == batch->max) {
298 if (!tlb_next_batch(tlb))
299 return 0;
300 batch = tlb->active;
301 }
302 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
303
304 return batch->max - batch->nr;
305}
306
307#endif /* HAVE_GENERIC_MMU_GATHER */
308
309#ifdef CONFIG_HAVE_RCU_TABLE_FREE
310
311/*
312 * See the comment near struct mmu_table_batch.
313 */
314
315static void tlb_remove_table_smp_sync(void *arg)
316{
317 /* Simply deliver the interrupt */
318}
319
320static void tlb_remove_table_one(void *table)
321{
322 /*
323 * This isn't an RCU grace period and hence the page-tables cannot be
324 * assumed to be actually RCU-freed.
325 *
326 * It is however sufficient for software page-table walkers that rely on
327 * IRQ disabling. See the comment near struct mmu_table_batch.
328 */
329 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
330 __tlb_remove_table(table);
331}
332
333static void tlb_remove_table_rcu(struct rcu_head *head)
334{
335 struct mmu_table_batch *batch;
336 int i;
337
338 batch = container_of(head, struct mmu_table_batch, rcu);
339
340 for (i = 0; i < batch->nr; i++)
341 __tlb_remove_table(batch->tables[i]);
342
343 free_page((unsigned long)batch);
344}
345
346void tlb_table_flush(struct mmu_gather *tlb)
347{
348 struct mmu_table_batch **batch = &tlb->batch;
349
350 if (*batch) {
351 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
352 *batch = NULL;
353 }
354}
355
356void tlb_remove_table(struct mmu_gather *tlb, void *table)
357{
358 struct mmu_table_batch **batch = &tlb->batch;
359
360 tlb->need_flush = 1;
361
362 /*
363 * When there's less then two users of this mm there cannot be a
364 * concurrent page-table walk.
365 */
366 if (atomic_read(&tlb->mm->mm_users) < 2) {
367 __tlb_remove_table(table);
368 return;
369 }
370
371 if (*batch == NULL) {
372 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
373 if (*batch == NULL) {
374 tlb_remove_table_one(table);
375 return;
376 }
377 (*batch)->nr = 0;
378 }
379 (*batch)->tables[(*batch)->nr++] = table;
380 if ((*batch)->nr == MAX_TABLE_BATCH)
381 tlb_table_flush(tlb);
382}
383
384#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
385
386/*
387 * Note: this doesn't free the actual pages themselves. That
388 * has been handled earlier when unmapping all the memory regions.
389 */
390static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
391 unsigned long addr)
392{
393 pgtable_t token = pmd_pgtable(*pmd);
394 pmd_clear(pmd);
395 pte_free_tlb(tlb, token, addr);
396 atomic_long_dec(&tlb->mm->nr_ptes);
397}
398
399static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
400 unsigned long addr, unsigned long end,
401 unsigned long floor, unsigned long ceiling)
402{
403 pmd_t *pmd;
404 unsigned long next;
405 unsigned long start;
406
407 start = addr;
408 pmd = pmd_offset(pud, addr);
409 do {
410 next = pmd_addr_end(addr, end);
411 if (pmd_none_or_clear_bad(pmd))
412 continue;
413 free_pte_range(tlb, pmd, addr);
414 } while (pmd++, addr = next, addr != end);
415
416 start &= PUD_MASK;
417 if (start < floor)
418 return;
419 if (ceiling) {
420 ceiling &= PUD_MASK;
421 if (!ceiling)
422 return;
423 }
424 if (end - 1 > ceiling - 1)
425 return;
426
427 pmd = pmd_offset(pud, start);
428 pud_clear(pud);
429 pmd_free_tlb(tlb, pmd, start);
430}
431
432static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
433 unsigned long addr, unsigned long end,
434 unsigned long floor, unsigned long ceiling)
435{
436 pud_t *pud;
437 unsigned long next;
438 unsigned long start;
439
440 start = addr;
441 pud = pud_offset(pgd, addr);
442 do {
443 next = pud_addr_end(addr, end);
444 if (pud_none_or_clear_bad(pud))
445 continue;
446 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
447 } while (pud++, addr = next, addr != end);
448
449 start &= PGDIR_MASK;
450 if (start < floor)
451 return;
452 if (ceiling) {
453 ceiling &= PGDIR_MASK;
454 if (!ceiling)
455 return;
456 }
457 if (end - 1 > ceiling - 1)
458 return;
459
460 pud = pud_offset(pgd, start);
461 pgd_clear(pgd);
462 pud_free_tlb(tlb, pud, start);
463}
464
465/*
466 * This function frees user-level page tables of a process.
467 */
468void free_pgd_range(struct mmu_gather *tlb,
469 unsigned long addr, unsigned long end,
470 unsigned long floor, unsigned long ceiling)
471{
472 pgd_t *pgd;
473 unsigned long next;
474
475 /*
476 * The next few lines have given us lots of grief...
477 *
478 * Why are we testing PMD* at this top level? Because often
479 * there will be no work to do at all, and we'd prefer not to
480 * go all the way down to the bottom just to discover that.
481 *
482 * Why all these "- 1"s? Because 0 represents both the bottom
483 * of the address space and the top of it (using -1 for the
484 * top wouldn't help much: the masks would do the wrong thing).
485 * The rule is that addr 0 and floor 0 refer to the bottom of
486 * the address space, but end 0 and ceiling 0 refer to the top
487 * Comparisons need to use "end - 1" and "ceiling - 1" (though
488 * that end 0 case should be mythical).
489 *
490 * Wherever addr is brought up or ceiling brought down, we must
491 * be careful to reject "the opposite 0" before it confuses the
492 * subsequent tests. But what about where end is brought down
493 * by PMD_SIZE below? no, end can't go down to 0 there.
494 *
495 * Whereas we round start (addr) and ceiling down, by different
496 * masks at different levels, in order to test whether a table
497 * now has no other vmas using it, so can be freed, we don't
498 * bother to round floor or end up - the tests don't need that.
499 */
500
501 addr &= PMD_MASK;
502 if (addr < floor) {
503 addr += PMD_SIZE;
504 if (!addr)
505 return;
506 }
507 if (ceiling) {
508 ceiling &= PMD_MASK;
509 if (!ceiling)
510 return;
511 }
512 if (end - 1 > ceiling - 1)
513 end -= PMD_SIZE;
514 if (addr > end - 1)
515 return;
516
517 pgd = pgd_offset(tlb->mm, addr);
518 do {
519 next = pgd_addr_end(addr, end);
520 if (pgd_none_or_clear_bad(pgd))
521 continue;
522 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
523 } while (pgd++, addr = next, addr != end);
524}
525
526void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
527 unsigned long floor, unsigned long ceiling)
528{
529 while (vma) {
530 struct vm_area_struct *next = vma->vm_next;
531 unsigned long addr = vma->vm_start;
532
533 /*
534 * Hide vma from rmap and truncate_pagecache before freeing
535 * pgtables
536 */
537 unlink_anon_vmas(vma);
538 unlink_file_vma(vma);
539
540 if (is_vm_hugetlb_page(vma)) {
541 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
542 floor, next? next->vm_start: ceiling);
543 } else {
544 /*
545 * Optimization: gather nearby vmas into one call down
546 */
547 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
548 && !is_vm_hugetlb_page(next)) {
549 vma = next;
550 next = vma->vm_next;
551 unlink_anon_vmas(vma);
552 unlink_file_vma(vma);
553 }
554 free_pgd_range(tlb, addr, vma->vm_end,
555 floor, next? next->vm_start: ceiling);
556 }
557 vma = next;
558 }
559}
560
561int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
562 pmd_t *pmd, unsigned long address)
563{
564 spinlock_t *ptl;
565 pgtable_t new = pte_alloc_one(mm, address);
566 int wait_split_huge_page;
567 if (!new)
568 return -ENOMEM;
569
570 /*
571 * Ensure all pte setup (eg. pte page lock and page clearing) are
572 * visible before the pte is made visible to other CPUs by being
573 * put into page tables.
574 *
575 * The other side of the story is the pointer chasing in the page
576 * table walking code (when walking the page table without locking;
577 * ie. most of the time). Fortunately, these data accesses consist
578 * of a chain of data-dependent loads, meaning most CPUs (alpha
579 * being the notable exception) will already guarantee loads are
580 * seen in-order. See the alpha page table accessors for the
581 * smp_read_barrier_depends() barriers in page table walking code.
582 */
583 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
584
585 ptl = pmd_lock(mm, pmd);
586 wait_split_huge_page = 0;
587 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
588 atomic_long_inc(&mm->nr_ptes);
589 pmd_populate(mm, pmd, new);
590 new = NULL;
591 } else if (unlikely(pmd_trans_splitting(*pmd)))
592 wait_split_huge_page = 1;
593 spin_unlock(ptl);
594 if (new)
595 pte_free(mm, new);
596 if (wait_split_huge_page)
597 wait_split_huge_page(vma->anon_vma, pmd);
598 return 0;
599}
600
601int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
602{
603 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
604 if (!new)
605 return -ENOMEM;
606
607 smp_wmb(); /* See comment in __pte_alloc */
608
609 spin_lock(&init_mm.page_table_lock);
610 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
611 pmd_populate_kernel(&init_mm, pmd, new);
612 new = NULL;
613 } else
614 VM_BUG_ON(pmd_trans_splitting(*pmd));
615 spin_unlock(&init_mm.page_table_lock);
616 if (new)
617 pte_free_kernel(&init_mm, new);
618 return 0;
619}
620
621static inline void init_rss_vec(int *rss)
622{
623 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
624}
625
626static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
627{
628 int i;
629
630 if (current->mm == mm)
631 sync_mm_rss(mm);
632 for (i = 0; i < NR_MM_COUNTERS; i++)
633 if (rss[i])
634 add_mm_counter(mm, i, rss[i]);
635}
636
637/*
638 * This function is called to print an error when a bad pte
639 * is found. For example, we might have a PFN-mapped pte in
640 * a region that doesn't allow it.
641 *
642 * The calling function must still handle the error.
643 */
644static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
645 pte_t pte, struct page *page)
646{
647 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
648 pud_t *pud = pud_offset(pgd, addr);
649 pmd_t *pmd = pmd_offset(pud, addr);
650 struct address_space *mapping;
651 pgoff_t index;
652 static unsigned long resume;
653 static unsigned long nr_shown;
654 static unsigned long nr_unshown;
655
656 /*
657 * Allow a burst of 60 reports, then keep quiet for that minute;
658 * or allow a steady drip of one report per second.
659 */
660 if (nr_shown == 60) {
661 if (time_before(jiffies, resume)) {
662 nr_unshown++;
663 return;
664 }
665 if (nr_unshown) {
666 printk(KERN_ALERT
667 "BUG: Bad page map: %lu messages suppressed\n",
668 nr_unshown);
669 nr_unshown = 0;
670 }
671 nr_shown = 0;
672 }
673 if (nr_shown++ == 0)
674 resume = jiffies + 60 * HZ;
675
676 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
677 index = linear_page_index(vma, addr);
678
679 printk(KERN_ALERT
680 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
681 current->comm,
682 (long long)pte_val(pte), (long long)pmd_val(*pmd));
683 if (page)
684 dump_page(page, "bad pte");
685 printk(KERN_ALERT
686 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
687 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
688 /*
689 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
690 */
691 if (vma->vm_ops)
692 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
693 vma->vm_ops->fault);
694 if (vma->vm_file)
695 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
696 vma->vm_file->f_op->mmap);
697 dump_stack();
698 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
699}
700
701static inline bool is_cow_mapping(vm_flags_t flags)
702{
703 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
704}
705
706/*
707 * vm_normal_page -- This function gets the "struct page" associated with a pte.
708 *
709 * "Special" mappings do not wish to be associated with a "struct page" (either
710 * it doesn't exist, or it exists but they don't want to touch it). In this
711 * case, NULL is returned here. "Normal" mappings do have a struct page.
712 *
713 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
714 * pte bit, in which case this function is trivial. Secondly, an architecture
715 * may not have a spare pte bit, which requires a more complicated scheme,
716 * described below.
717 *
718 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
719 * special mapping (even if there are underlying and valid "struct pages").
720 * COWed pages of a VM_PFNMAP are always normal.
721 *
722 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
723 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
724 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
725 * mapping will always honor the rule
726 *
727 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
728 *
729 * And for normal mappings this is false.
730 *
731 * This restricts such mappings to be a linear translation from virtual address
732 * to pfn. To get around this restriction, we allow arbitrary mappings so long
733 * as the vma is not a COW mapping; in that case, we know that all ptes are
734 * special (because none can have been COWed).
735 *
736 *
737 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
738 *
739 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
740 * page" backing, however the difference is that _all_ pages with a struct
741 * page (that is, those where pfn_valid is true) are refcounted and considered
742 * normal pages by the VM. The disadvantage is that pages are refcounted
743 * (which can be slower and simply not an option for some PFNMAP users). The
744 * advantage is that we don't have to follow the strict linearity rule of
745 * PFNMAP mappings in order to support COWable mappings.
746 *
747 */
748#ifdef __HAVE_ARCH_PTE_SPECIAL
749# define HAVE_PTE_SPECIAL 1
750#else
751# define HAVE_PTE_SPECIAL 0
752#endif
753struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
754 pte_t pte)
755{
756 unsigned long pfn = pte_pfn(pte);
757
758 if (HAVE_PTE_SPECIAL) {
759 if (likely(!pte_special(pte)))
760 goto check_pfn;
761 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
762 return NULL;
763 if (!is_zero_pfn(pfn))
764 print_bad_pte(vma, addr, pte, NULL);
765 return NULL;
766 }
767
768 /* !HAVE_PTE_SPECIAL case follows: */
769
770 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
771 if (vma->vm_flags & VM_MIXEDMAP) {
772 if (!pfn_valid(pfn))
773 return NULL;
774 goto out;
775 } else {
776 unsigned long off;
777 off = (addr - vma->vm_start) >> PAGE_SHIFT;
778 if (pfn == vma->vm_pgoff + off)
779 return NULL;
780 if (!is_cow_mapping(vma->vm_flags))
781 return NULL;
782 }
783 }
784
785 if (is_zero_pfn(pfn))
786 return NULL;
787check_pfn:
788 if (unlikely(pfn > highest_memmap_pfn)) {
789 print_bad_pte(vma, addr, pte, NULL);
790 return NULL;
791 }
792
793 /*
794 * NOTE! We still have PageReserved() pages in the page tables.
795 * eg. VDSO mappings can cause them to exist.
796 */
797out:
798 return pfn_to_page(pfn);
799}
800
801/*
802 * copy one vm_area from one task to the other. Assumes the page tables
803 * already present in the new task to be cleared in the whole range
804 * covered by this vma.
805 */
806
807static inline unsigned long
808copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
809 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
810 unsigned long addr, int *rss)
811{
812 unsigned long vm_flags = vma->vm_flags;
813 pte_t pte = *src_pte;
814 struct page *page;
815
816 /* pte contains position in swap or file, so copy. */
817 if (unlikely(!pte_present(pte))) {
818 if (!pte_file(pte)) {
819 swp_entry_t entry = pte_to_swp_entry(pte);
820
821 if (swap_duplicate(entry) < 0)
822 return entry.val;
823
824 /* make sure dst_mm is on swapoff's mmlist. */
825 if (unlikely(list_empty(&dst_mm->mmlist))) {
826 spin_lock(&mmlist_lock);
827 if (list_empty(&dst_mm->mmlist))
828 list_add(&dst_mm->mmlist,
829 &src_mm->mmlist);
830 spin_unlock(&mmlist_lock);
831 }
832 if (likely(!non_swap_entry(entry)))
833 rss[MM_SWAPENTS]++;
834 else if (is_migration_entry(entry)) {
835 page = migration_entry_to_page(entry);
836
837 if (PageAnon(page))
838 rss[MM_ANONPAGES]++;
839 else
840 rss[MM_FILEPAGES]++;
841
842 if (is_write_migration_entry(entry) &&
843 is_cow_mapping(vm_flags)) {
844 /*
845 * COW mappings require pages in both
846 * parent and child to be set to read.
847 */
848 make_migration_entry_read(&entry);
849 pte = swp_entry_to_pte(entry);
850 if (pte_swp_soft_dirty(*src_pte))
851 pte = pte_swp_mksoft_dirty(pte);
852 set_pte_at(src_mm, addr, src_pte, pte);
853 }
854 }
855 }
856 goto out_set_pte;
857 }
858
859 /*
860 * If it's a COW mapping, write protect it both
861 * in the parent and the child
862 */
863 if (is_cow_mapping(vm_flags)) {
864 ptep_set_wrprotect(src_mm, addr, src_pte);
865 pte = pte_wrprotect(pte);
866 }
867
868 /*
869 * If it's a shared mapping, mark it clean in
870 * the child
871 */
872 if (vm_flags & VM_SHARED)
873 pte = pte_mkclean(pte);
874 pte = pte_mkold(pte);
875
876 page = vm_normal_page(vma, addr, pte);
877 if (page) {
878 get_page(page);
879 page_dup_rmap(page);
880 if (PageAnon(page))
881 rss[MM_ANONPAGES]++;
882 else
883 rss[MM_FILEPAGES]++;
884 }
885
886out_set_pte:
887 set_pte_at(dst_mm, addr, dst_pte, pte);
888 return 0;
889}
890
891int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
892 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
893 unsigned long addr, unsigned long end)
894{
895 pte_t *orig_src_pte, *orig_dst_pte;
896 pte_t *src_pte, *dst_pte;
897 spinlock_t *src_ptl, *dst_ptl;
898 int progress = 0;
899 int rss[NR_MM_COUNTERS];
900 swp_entry_t entry = (swp_entry_t){0};
901
902again:
903 init_rss_vec(rss);
904
905 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
906 if (!dst_pte)
907 return -ENOMEM;
908 src_pte = pte_offset_map(src_pmd, addr);
909 src_ptl = pte_lockptr(src_mm, src_pmd);
910 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
911 orig_src_pte = src_pte;
912 orig_dst_pte = dst_pte;
913 arch_enter_lazy_mmu_mode();
914
915 do {
916 /*
917 * We are holding two locks at this point - either of them
918 * could generate latencies in another task on another CPU.
919 */
920 if (progress >= 32) {
921 progress = 0;
922 if (need_resched() ||
923 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
924 break;
925 }
926 if (pte_none(*src_pte)) {
927 progress++;
928 continue;
929 }
930 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
931 vma, addr, rss);
932 if (entry.val)
933 break;
934 progress += 8;
935 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
936
937 arch_leave_lazy_mmu_mode();
938 spin_unlock(src_ptl);
939 pte_unmap(orig_src_pte);
940 add_mm_rss_vec(dst_mm, rss);
941 pte_unmap_unlock(orig_dst_pte, dst_ptl);
942 cond_resched();
943
944 if (entry.val) {
945 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
946 return -ENOMEM;
947 progress = 0;
948 }
949 if (addr != end)
950 goto again;
951 return 0;
952}
953
954static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
955 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
956 unsigned long addr, unsigned long end)
957{
958 pmd_t *src_pmd, *dst_pmd;
959 unsigned long next;
960
961 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
962 if (!dst_pmd)
963 return -ENOMEM;
964 src_pmd = pmd_offset(src_pud, addr);
965 do {
966 next = pmd_addr_end(addr, end);
967 if (pmd_trans_huge(*src_pmd)) {
968 int err;
969 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
970 err = copy_huge_pmd(dst_mm, src_mm,
971 dst_pmd, src_pmd, addr, vma);
972 if (err == -ENOMEM)
973 return -ENOMEM;
974 if (!err)
975 continue;
976 /* fall through */
977 }
978 if (pmd_none_or_clear_bad(src_pmd))
979 continue;
980 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
981 vma, addr, next))
982 return -ENOMEM;
983 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
984 return 0;
985}
986
987static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
988 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
989 unsigned long addr, unsigned long end)
990{
991 pud_t *src_pud, *dst_pud;
992 unsigned long next;
993
994 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
995 if (!dst_pud)
996 return -ENOMEM;
997 src_pud = pud_offset(src_pgd, addr);
998 do {
999 next = pud_addr_end(addr, end);
1000 if (pud_none_or_clear_bad(src_pud))
1001 continue;
1002 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1003 vma, addr, next))
1004 return -ENOMEM;
1005 } while (dst_pud++, src_pud++, addr = next, addr != end);
1006 return 0;
1007}
1008
1009int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1010 struct vm_area_struct *vma)
1011{
1012 pgd_t *src_pgd, *dst_pgd;
1013 unsigned long next;
1014 unsigned long addr = vma->vm_start;
1015 unsigned long end = vma->vm_end;
1016 unsigned long mmun_start; /* For mmu_notifiers */
1017 unsigned long mmun_end; /* For mmu_notifiers */
1018 bool is_cow;
1019 int ret;
1020
1021 /*
1022 * Don't copy ptes where a page fault will fill them correctly.
1023 * Fork becomes much lighter when there are big shared or private
1024 * readonly mappings. The tradeoff is that copy_page_range is more
1025 * efficient than faulting.
1026 */
1027 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1028 VM_PFNMAP | VM_MIXEDMAP))) {
1029 if (!vma->anon_vma)
1030 return 0;
1031 }
1032
1033 if (is_vm_hugetlb_page(vma))
1034 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1035
1036 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1037 /*
1038 * We do not free on error cases below as remove_vma
1039 * gets called on error from higher level routine
1040 */
1041 ret = track_pfn_copy(vma);
1042 if (ret)
1043 return ret;
1044 }
1045
1046 /*
1047 * We need to invalidate the secondary MMU mappings only when
1048 * there could be a permission downgrade on the ptes of the
1049 * parent mm. And a permission downgrade will only happen if
1050 * is_cow_mapping() returns true.
1051 */
1052 is_cow = is_cow_mapping(vma->vm_flags);
1053 mmun_start = addr;
1054 mmun_end = end;
1055 if (is_cow)
1056 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1057 mmun_end);
1058
1059 ret = 0;
1060 dst_pgd = pgd_offset(dst_mm, addr);
1061 src_pgd = pgd_offset(src_mm, addr);
1062 do {
1063 next = pgd_addr_end(addr, end);
1064 if (pgd_none_or_clear_bad(src_pgd))
1065 continue;
1066 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1067 vma, addr, next))) {
1068 ret = -ENOMEM;
1069 break;
1070 }
1071 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1072
1073 if (is_cow)
1074 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1075 return ret;
1076}
1077
1078static unsigned long zap_pte_range(struct mmu_gather *tlb,
1079 struct vm_area_struct *vma, pmd_t *pmd,
1080 unsigned long addr, unsigned long end,
1081 struct zap_details *details)
1082{
1083 struct mm_struct *mm = tlb->mm;
1084 int force_flush = 0;
1085 int rss[NR_MM_COUNTERS];
1086 spinlock_t *ptl;
1087 pte_t *start_pte;
1088 pte_t *pte;
1089
1090again:
1091 init_rss_vec(rss);
1092 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1093 pte = start_pte;
1094 arch_enter_lazy_mmu_mode();
1095 do {
1096 pte_t ptent = *pte;
1097 if (pte_none(ptent)) {
1098 continue;
1099 }
1100
1101 if (pte_present(ptent)) {
1102 struct page *page;
1103
1104 page = vm_normal_page(vma, addr, ptent);
1105 if (unlikely(details) && page) {
1106 /*
1107 * unmap_shared_mapping_pages() wants to
1108 * invalidate cache without truncating:
1109 * unmap shared but keep private pages.
1110 */
1111 if (details->check_mapping &&
1112 details->check_mapping != page->mapping)
1113 continue;
1114 /*
1115 * Each page->index must be checked when
1116 * invalidating or truncating nonlinear.
1117 */
1118 if (details->nonlinear_vma &&
1119 (page->index < details->first_index ||
1120 page->index > details->last_index))
1121 continue;
1122 }
1123 ptent = ptep_get_and_clear_full(mm, addr, pte,
1124 tlb->fullmm);
1125 tlb_remove_tlb_entry(tlb, pte, addr);
1126 if (unlikely(!page))
1127 continue;
1128 if (unlikely(details) && details->nonlinear_vma
1129 && linear_page_index(details->nonlinear_vma,
1130 addr) != page->index) {
1131 pte_t ptfile = pgoff_to_pte(page->index);
1132 if (pte_soft_dirty(ptent))
1133 pte_file_mksoft_dirty(ptfile);
1134 set_pte_at(mm, addr, pte, ptfile);
1135 }
1136 if (PageAnon(page))
1137 rss[MM_ANONPAGES]--;
1138 else {
1139 if (pte_dirty(ptent)) {
1140 force_flush = 1;
1141 set_page_dirty(page);
1142 }
1143 if (pte_young(ptent) &&
1144 likely(!(vma->vm_flags & VM_SEQ_READ)))
1145 mark_page_accessed(page);
1146 rss[MM_FILEPAGES]--;
1147 }
1148 page_remove_rmap(page);
1149 if (unlikely(page_mapcount(page) < 0))
1150 print_bad_pte(vma, addr, ptent, page);
1151 if (unlikely(!__tlb_remove_page(tlb, page))) {
1152 force_flush = 1;
1153 break;
1154 }
1155 continue;
1156 }
1157 /*
1158 * If details->check_mapping, we leave swap entries;
1159 * if details->nonlinear_vma, we leave file entries.
1160 */
1161 if (unlikely(details))
1162 continue;
1163 if (pte_file(ptent)) {
1164 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1165 print_bad_pte(vma, addr, ptent, NULL);
1166 } else {
1167 swp_entry_t entry = pte_to_swp_entry(ptent);
1168
1169 if (!non_swap_entry(entry))
1170 rss[MM_SWAPENTS]--;
1171 else if (is_migration_entry(entry)) {
1172 struct page *page;
1173
1174 page = migration_entry_to_page(entry);
1175
1176 if (PageAnon(page))
1177 rss[MM_ANONPAGES]--;
1178 else
1179 rss[MM_FILEPAGES]--;
1180 }
1181 if (unlikely(!free_swap_and_cache(entry)))
1182 print_bad_pte(vma, addr, ptent, NULL);
1183 }
1184 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1185 } while (pte++, addr += PAGE_SIZE, addr != end);
1186
1187 add_mm_rss_vec(mm, rss);
1188 arch_leave_lazy_mmu_mode();
1189
1190 /* Do the actual TLB flush before dropping ptl */
1191 if (force_flush) {
1192 unsigned long old_end;
1193
1194 /*
1195 * Flush the TLB just for the previous segment,
1196 * then update the range to be the remaining
1197 * TLB range.
1198 */
1199 old_end = tlb->end;
1200 tlb->end = addr;
1201 tlb_flush_mmu_tlbonly(tlb);
1202 tlb->start = addr;
1203 tlb->end = old_end;
1204 }
1205 pte_unmap_unlock(start_pte, ptl);
1206
1207 /*
1208 * If we forced a TLB flush (either due to running out of
1209 * batch buffers or because we needed to flush dirty TLB
1210 * entries before releasing the ptl), free the batched
1211 * memory too. Restart if we didn't do everything.
1212 */
1213 if (force_flush) {
1214 force_flush = 0;
1215 tlb_flush_mmu_free(tlb);
1216
1217 if (addr != end)
1218 goto again;
1219 }
1220
1221 return addr;
1222}
1223
1224static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225 struct vm_area_struct *vma, pud_t *pud,
1226 unsigned long addr, unsigned long end,
1227 struct zap_details *details)
1228{
1229 pmd_t *pmd;
1230 unsigned long next;
1231
1232 pmd = pmd_offset(pud, addr);
1233 do {
1234 next = pmd_addr_end(addr, end);
1235 if (pmd_trans_huge(*pmd)) {
1236 if (next - addr != HPAGE_PMD_SIZE) {
1237#ifdef CONFIG_DEBUG_VM
1238 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1239 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240 __func__, addr, end,
1241 vma->vm_start,
1242 vma->vm_end);
1243 BUG();
1244 }
1245#endif
1246 split_huge_page_pmd(vma, addr, pmd);
1247 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1248 goto next;
1249 /* fall through */
1250 }
1251 /*
1252 * Here there can be other concurrent MADV_DONTNEED or
1253 * trans huge page faults running, and if the pmd is
1254 * none or trans huge it can change under us. This is
1255 * because MADV_DONTNEED holds the mmap_sem in read
1256 * mode.
1257 */
1258 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1259 goto next;
1260 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1261next:
1262 cond_resched();
1263 } while (pmd++, addr = next, addr != end);
1264
1265 return addr;
1266}
1267
1268static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1269 struct vm_area_struct *vma, pgd_t *pgd,
1270 unsigned long addr, unsigned long end,
1271 struct zap_details *details)
1272{
1273 pud_t *pud;
1274 unsigned long next;
1275
1276 pud = pud_offset(pgd, addr);
1277 do {
1278 next = pud_addr_end(addr, end);
1279 if (pud_none_or_clear_bad(pud))
1280 continue;
1281 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1282 } while (pud++, addr = next, addr != end);
1283
1284 return addr;
1285}
1286
1287static void unmap_page_range(struct mmu_gather *tlb,
1288 struct vm_area_struct *vma,
1289 unsigned long addr, unsigned long end,
1290 struct zap_details *details)
1291{
1292 pgd_t *pgd;
1293 unsigned long next;
1294
1295 if (details && !details->check_mapping && !details->nonlinear_vma)
1296 details = NULL;
1297
1298 BUG_ON(addr >= end);
1299 mem_cgroup_uncharge_start();
1300 tlb_start_vma(tlb, vma);
1301 pgd = pgd_offset(vma->vm_mm, addr);
1302 do {
1303 next = pgd_addr_end(addr, end);
1304 if (pgd_none_or_clear_bad(pgd))
1305 continue;
1306 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307 } while (pgd++, addr = next, addr != end);
1308 tlb_end_vma(tlb, vma);
1309 mem_cgroup_uncharge_end();
1310}
1311
1312
1313static void unmap_single_vma(struct mmu_gather *tlb,
1314 struct vm_area_struct *vma, unsigned long start_addr,
1315 unsigned long end_addr,
1316 struct zap_details *details)
1317{
1318 unsigned long start = max(vma->vm_start, start_addr);
1319 unsigned long end;
1320
1321 if (start >= vma->vm_end)
1322 return;
1323 end = min(vma->vm_end, end_addr);
1324 if (end <= vma->vm_start)
1325 return;
1326
1327 if (vma->vm_file)
1328 uprobe_munmap(vma, start, end);
1329
1330 if (unlikely(vma->vm_flags & VM_PFNMAP))
1331 untrack_pfn(vma, 0, 0);
1332
1333 if (start != end) {
1334 if (unlikely(is_vm_hugetlb_page(vma))) {
1335 /*
1336 * It is undesirable to test vma->vm_file as it
1337 * should be non-null for valid hugetlb area.
1338 * However, vm_file will be NULL in the error
1339 * cleanup path of mmap_region. When
1340 * hugetlbfs ->mmap method fails,
1341 * mmap_region() nullifies vma->vm_file
1342 * before calling this function to clean up.
1343 * Since no pte has actually been setup, it is
1344 * safe to do nothing in this case.
1345 */
1346 if (vma->vm_file) {
1347 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1348 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1349 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1350 }
1351 } else
1352 unmap_page_range(tlb, vma, start, end, details);
1353 }
1354}
1355
1356/**
1357 * unmap_vmas - unmap a range of memory covered by a list of vma's
1358 * @tlb: address of the caller's struct mmu_gather
1359 * @vma: the starting vma
1360 * @start_addr: virtual address at which to start unmapping
1361 * @end_addr: virtual address at which to end unmapping
1362 *
1363 * Unmap all pages in the vma list.
1364 *
1365 * Only addresses between `start' and `end' will be unmapped.
1366 *
1367 * The VMA list must be sorted in ascending virtual address order.
1368 *
1369 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370 * range after unmap_vmas() returns. So the only responsibility here is to
1371 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372 * drops the lock and schedules.
1373 */
1374void unmap_vmas(struct mmu_gather *tlb,
1375 struct vm_area_struct *vma, unsigned long start_addr,
1376 unsigned long end_addr)
1377{
1378 struct mm_struct *mm = vma->vm_mm;
1379
1380 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1381 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1382 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1383 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1384}
1385
1386/**
1387 * zap_page_range - remove user pages in a given range
1388 * @vma: vm_area_struct holding the applicable pages
1389 * @start: starting address of pages to zap
1390 * @size: number of bytes to zap
1391 * @details: details of nonlinear truncation or shared cache invalidation
1392 *
1393 * Caller must protect the VMA list
1394 */
1395void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1396 unsigned long size, struct zap_details *details)
1397{
1398 struct mm_struct *mm = vma->vm_mm;
1399 struct mmu_gather tlb;
1400 unsigned long end = start + size;
1401
1402 lru_add_drain();
1403 tlb_gather_mmu(&tlb, mm, start, end);
1404 update_hiwater_rss(mm);
1405 mmu_notifier_invalidate_range_start(mm, start, end);
1406 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1407 unmap_single_vma(&tlb, vma, start, end, details);
1408 mmu_notifier_invalidate_range_end(mm, start, end);
1409 tlb_finish_mmu(&tlb, start, end);
1410}
1411
1412/**
1413 * zap_page_range_single - remove user pages in a given range
1414 * @vma: vm_area_struct holding the applicable pages
1415 * @address: starting address of pages to zap
1416 * @size: number of bytes to zap
1417 * @details: details of nonlinear truncation or shared cache invalidation
1418 *
1419 * The range must fit into one VMA.
1420 */
1421static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1422 unsigned long size, struct zap_details *details)
1423{
1424 struct mm_struct *mm = vma->vm_mm;
1425 struct mmu_gather tlb;
1426 unsigned long end = address + size;
1427
1428 lru_add_drain();
1429 tlb_gather_mmu(&tlb, mm, address, end);
1430 update_hiwater_rss(mm);
1431 mmu_notifier_invalidate_range_start(mm, address, end);
1432 unmap_single_vma(&tlb, vma, address, end, details);
1433 mmu_notifier_invalidate_range_end(mm, address, end);
1434 tlb_finish_mmu(&tlb, address, end);
1435}
1436
1437/**
1438 * zap_vma_ptes - remove ptes mapping the vma
1439 * @vma: vm_area_struct holding ptes to be zapped
1440 * @address: starting address of pages to zap
1441 * @size: number of bytes to zap
1442 *
1443 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444 *
1445 * The entire address range must be fully contained within the vma.
1446 *
1447 * Returns 0 if successful.
1448 */
1449int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1450 unsigned long size)
1451{
1452 if (address < vma->vm_start || address + size > vma->vm_end ||
1453 !(vma->vm_flags & VM_PFNMAP))
1454 return -1;
1455 zap_page_range_single(vma, address, size, NULL);
1456 return 0;
1457}
1458EXPORT_SYMBOL_GPL(zap_vma_ptes);
1459
1460/**
1461 * follow_page_mask - look up a page descriptor from a user-virtual address
1462 * @vma: vm_area_struct mapping @address
1463 * @address: virtual address to look up
1464 * @flags: flags modifying lookup behaviour
1465 * @page_mask: on output, *page_mask is set according to the size of the page
1466 *
1467 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1468 *
1469 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1470 * an error pointer if there is a mapping to something not represented
1471 * by a page descriptor (see also vm_normal_page()).
1472 */
1473struct page *follow_page_mask(struct vm_area_struct *vma,
1474 unsigned long address, unsigned int flags,
1475 unsigned int *page_mask)
1476{
1477 pgd_t *pgd;
1478 pud_t *pud;
1479 pmd_t *pmd;
1480 pte_t *ptep, pte;
1481 spinlock_t *ptl;
1482 struct page *page;
1483 struct mm_struct *mm = vma->vm_mm;
1484
1485 *page_mask = 0;
1486
1487 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1488 if (!IS_ERR(page)) {
1489 BUG_ON(flags & FOLL_GET);
1490 goto out;
1491 }
1492
1493 page = NULL;
1494 pgd = pgd_offset(mm, address);
1495 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1496 goto no_page_table;
1497
1498 pud = pud_offset(pgd, address);
1499 if (pud_none(*pud))
1500 goto no_page_table;
1501 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1502 if (flags & FOLL_GET)
1503 goto out;
1504 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1505 goto out;
1506 }
1507 if (unlikely(pud_bad(*pud)))
1508 goto no_page_table;
1509
1510 pmd = pmd_offset(pud, address);
1511 if (pmd_none(*pmd))
1512 goto no_page_table;
1513 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1514 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1515 if (flags & FOLL_GET) {
1516 /*
1517 * Refcount on tail pages are not well-defined and
1518 * shouldn't be taken. The caller should handle a NULL
1519 * return when trying to follow tail pages.
1520 */
1521 if (PageHead(page))
1522 get_page(page);
1523 else {
1524 page = NULL;
1525 goto out;
1526 }
1527 }
1528 goto out;
1529 }
1530 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1531 goto no_page_table;
1532 if (pmd_trans_huge(*pmd)) {
1533 if (flags & FOLL_SPLIT) {
1534 split_huge_page_pmd(vma, address, pmd);
1535 goto split_fallthrough;
1536 }
1537 ptl = pmd_lock(mm, pmd);
1538 if (likely(pmd_trans_huge(*pmd))) {
1539 if (unlikely(pmd_trans_splitting(*pmd))) {
1540 spin_unlock(ptl);
1541 wait_split_huge_page(vma->anon_vma, pmd);
1542 } else {
1543 page = follow_trans_huge_pmd(vma, address,
1544 pmd, flags);
1545 spin_unlock(ptl);
1546 *page_mask = HPAGE_PMD_NR - 1;
1547 goto out;
1548 }
1549 } else
1550 spin_unlock(ptl);
1551 /* fall through */
1552 }
1553split_fallthrough:
1554 if (unlikely(pmd_bad(*pmd)))
1555 goto no_page_table;
1556
1557 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1558
1559 pte = *ptep;
1560 if (!pte_present(pte)) {
1561 swp_entry_t entry;
1562 /*
1563 * KSM's break_ksm() relies upon recognizing a ksm page
1564 * even while it is being migrated, so for that case we
1565 * need migration_entry_wait().
1566 */
1567 if (likely(!(flags & FOLL_MIGRATION)))
1568 goto no_page;
1569 if (pte_none(pte) || pte_file(pte))
1570 goto no_page;
1571 entry = pte_to_swp_entry(pte);
1572 if (!is_migration_entry(entry))
1573 goto no_page;
1574 pte_unmap_unlock(ptep, ptl);
1575 migration_entry_wait(mm, pmd, address);
1576 goto split_fallthrough;
1577 }
1578 if ((flags & FOLL_NUMA) && pte_numa(pte))
1579 goto no_page;
1580 if ((flags & FOLL_WRITE) && !pte_write(pte))
1581 goto unlock;
1582
1583 page = vm_normal_page(vma, address, pte);
1584 if (unlikely(!page)) {
1585 if ((flags & FOLL_DUMP) ||
1586 !is_zero_pfn(pte_pfn(pte)))
1587 goto bad_page;
1588 page = pte_page(pte);
1589 }
1590
1591 if (flags & FOLL_GET)
1592 get_page_foll(page);
1593 if (flags & FOLL_TOUCH) {
1594 if ((flags & FOLL_WRITE) &&
1595 !pte_dirty(pte) && !PageDirty(page))
1596 set_page_dirty(page);
1597 /*
1598 * pte_mkyoung() would be more correct here, but atomic care
1599 * is needed to avoid losing the dirty bit: it is easier to use
1600 * mark_page_accessed().
1601 */
1602 mark_page_accessed(page);
1603 }
1604 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1605 /*
1606 * The preliminary mapping check is mainly to avoid the
1607 * pointless overhead of lock_page on the ZERO_PAGE
1608 * which might bounce very badly if there is contention.
1609 *
1610 * If the page is already locked, we don't need to
1611 * handle it now - vmscan will handle it later if and
1612 * when it attempts to reclaim the page.
1613 */
1614 if (page->mapping && trylock_page(page)) {
1615 lru_add_drain(); /* push cached pages to LRU */
1616 /*
1617 * Because we lock page here, and migration is
1618 * blocked by the pte's page reference, and we
1619 * know the page is still mapped, we don't even
1620 * need to check for file-cache page truncation.
1621 */
1622 mlock_vma_page(page);
1623 unlock_page(page);
1624 }
1625 }
1626unlock:
1627 pte_unmap_unlock(ptep, ptl);
1628out:
1629 return page;
1630
1631bad_page:
1632 pte_unmap_unlock(ptep, ptl);
1633 return ERR_PTR(-EFAULT);
1634
1635no_page:
1636 pte_unmap_unlock(ptep, ptl);
1637 if (!pte_none(pte))
1638 return page;
1639
1640no_page_table:
1641 /*
1642 * When core dumping an enormous anonymous area that nobody
1643 * has touched so far, we don't want to allocate unnecessary pages or
1644 * page tables. Return error instead of NULL to skip handle_mm_fault,
1645 * then get_dump_page() will return NULL to leave a hole in the dump.
1646 * But we can only make this optimization where a hole would surely
1647 * be zero-filled if handle_mm_fault() actually did handle it.
1648 */
1649 if ((flags & FOLL_DUMP) &&
1650 (!vma->vm_ops || !vma->vm_ops->fault))
1651 return ERR_PTR(-EFAULT);
1652 return page;
1653}
1654
1655static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1656{
1657 return stack_guard_page_start(vma, addr) ||
1658 stack_guard_page_end(vma, addr+PAGE_SIZE);
1659}
1660
1661/**
1662 * __get_user_pages() - pin user pages in memory
1663 * @tsk: task_struct of target task
1664 * @mm: mm_struct of target mm
1665 * @start: starting user address
1666 * @nr_pages: number of pages from start to pin
1667 * @gup_flags: flags modifying pin behaviour
1668 * @pages: array that receives pointers to the pages pinned.
1669 * Should be at least nr_pages long. Or NULL, if caller
1670 * only intends to ensure the pages are faulted in.
1671 * @vmas: array of pointers to vmas corresponding to each page.
1672 * Or NULL if the caller does not require them.
1673 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1674 *
1675 * Returns number of pages pinned. This may be fewer than the number
1676 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1677 * were pinned, returns -errno. Each page returned must be released
1678 * with a put_page() call when it is finished with. vmas will only
1679 * remain valid while mmap_sem is held.
1680 *
1681 * Must be called with mmap_sem held for read or write.
1682 *
1683 * __get_user_pages walks a process's page tables and takes a reference to
1684 * each struct page that each user address corresponds to at a given
1685 * instant. That is, it takes the page that would be accessed if a user
1686 * thread accesses the given user virtual address at that instant.
1687 *
1688 * This does not guarantee that the page exists in the user mappings when
1689 * __get_user_pages returns, and there may even be a completely different
1690 * page there in some cases (eg. if mmapped pagecache has been invalidated
1691 * and subsequently re faulted). However it does guarantee that the page
1692 * won't be freed completely. And mostly callers simply care that the page
1693 * contains data that was valid *at some point in time*. Typically, an IO
1694 * or similar operation cannot guarantee anything stronger anyway because
1695 * locks can't be held over the syscall boundary.
1696 *
1697 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1698 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1699 * appropriate) must be called after the page is finished with, and
1700 * before put_page is called.
1701 *
1702 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1703 * or mmap_sem contention, and if waiting is needed to pin all pages,
1704 * *@nonblocking will be set to 0.
1705 *
1706 * In most cases, get_user_pages or get_user_pages_fast should be used
1707 * instead of __get_user_pages. __get_user_pages should be used only if
1708 * you need some special @gup_flags.
1709 */
1710long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1711 unsigned long start, unsigned long nr_pages,
1712 unsigned int gup_flags, struct page **pages,
1713 struct vm_area_struct **vmas, int *nonblocking)
1714{
1715 long i;
1716 unsigned long vm_flags;
1717 unsigned int page_mask;
1718
1719 if (!nr_pages)
1720 return 0;
1721
1722 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1723
1724 /*
1725 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1726 * would be called on PROT_NONE ranges. We must never invoke
1727 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1728 * page faults would unprotect the PROT_NONE ranges if
1729 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1730 * bitflag. So to avoid that, don't set FOLL_NUMA if
1731 * FOLL_FORCE is set.
1732 */
1733 if (!(gup_flags & FOLL_FORCE))
1734 gup_flags |= FOLL_NUMA;
1735
1736 i = 0;
1737
1738 do {
1739 struct vm_area_struct *vma;
1740
1741 vma = find_extend_vma(mm, start);
1742 if (!vma && in_gate_area(mm, start)) {
1743 unsigned long pg = start & PAGE_MASK;
1744 pgd_t *pgd;
1745 pud_t *pud;
1746 pmd_t *pmd;
1747 pte_t *pte;
1748
1749 /* user gate pages are read-only */
1750 if (gup_flags & FOLL_WRITE)
1751 goto efault;
1752 if (pg > TASK_SIZE)
1753 pgd = pgd_offset_k(pg);
1754 else
1755 pgd = pgd_offset_gate(mm, pg);
1756 BUG_ON(pgd_none(*pgd));
1757 pud = pud_offset(pgd, pg);
1758 BUG_ON(pud_none(*pud));
1759 pmd = pmd_offset(pud, pg);
1760 if (pmd_none(*pmd))
1761 goto efault;
1762 VM_BUG_ON(pmd_trans_huge(*pmd));
1763 pte = pte_offset_map(pmd, pg);
1764 if (pte_none(*pte)) {
1765 pte_unmap(pte);
1766 goto efault;
1767 }
1768 vma = get_gate_vma(mm);
1769 if (pages) {
1770 struct page *page;
1771
1772 page = vm_normal_page(vma, start, *pte);
1773 if (!page) {
1774 if (!(gup_flags & FOLL_DUMP) &&
1775 is_zero_pfn(pte_pfn(*pte)))
1776 page = pte_page(*pte);
1777 else {
1778 pte_unmap(pte);
1779 goto efault;
1780 }
1781 }
1782 pages[i] = page;
1783 get_page(page);
1784 }
1785 pte_unmap(pte);
1786 page_mask = 0;
1787 goto next_page;
1788 }
1789
1790 if (!vma)
1791 goto efault;
1792 vm_flags = vma->vm_flags;
1793 if (vm_flags & (VM_IO | VM_PFNMAP))
1794 goto efault;
1795
1796 if (gup_flags & FOLL_WRITE) {
1797 if (!(vm_flags & VM_WRITE)) {
1798 if (!(gup_flags & FOLL_FORCE))
1799 goto efault;
1800 /*
1801 * We used to let the write,force case do COW
1802 * in a VM_MAYWRITE VM_SHARED !VM_WRITE vma, so
1803 * ptrace could set a breakpoint in a read-only
1804 * mapping of an executable, without corrupting
1805 * the file (yet only when that file had been
1806 * opened for writing!). Anon pages in shared
1807 * mappings are surprising: now just reject it.
1808 */
1809 if (!is_cow_mapping(vm_flags)) {
1810 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
1811 goto efault;
1812 }
1813 }
1814 } else {
1815 if (!(vm_flags & VM_READ)) {
1816 if (!(gup_flags & FOLL_FORCE))
1817 goto efault;
1818 /*
1819 * Is there actually any vma we can reach here
1820 * which does not have VM_MAYREAD set?
1821 */
1822 if (!(vm_flags & VM_MAYREAD))
1823 goto efault;
1824 }
1825 }
1826
1827 if (is_vm_hugetlb_page(vma)) {
1828 i = follow_hugetlb_page(mm, vma, pages, vmas,
1829 &start, &nr_pages, i, gup_flags);
1830 continue;
1831 }
1832
1833 do {
1834 struct page *page;
1835 unsigned int foll_flags = gup_flags;
1836 unsigned int page_increm;
1837
1838 /*
1839 * If we have a pending SIGKILL, don't keep faulting
1840 * pages and potentially allocating memory.
1841 */
1842 if (unlikely(fatal_signal_pending(current)))
1843 return i ? i : -ERESTARTSYS;
1844
1845 cond_resched();
1846 while (!(page = follow_page_mask(vma, start,
1847 foll_flags, &page_mask))) {
1848 int ret;
1849 unsigned int fault_flags = 0;
1850
1851 /* For mlock, just skip the stack guard page. */
1852 if (foll_flags & FOLL_MLOCK) {
1853 if (stack_guard_page(vma, start))
1854 goto next_page;
1855 }
1856 if (foll_flags & FOLL_WRITE)
1857 fault_flags |= FAULT_FLAG_WRITE;
1858 if (nonblocking)
1859 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1860 if (foll_flags & FOLL_NOWAIT)
1861 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1862
1863 ret = handle_mm_fault(mm, vma, start,
1864 fault_flags);
1865
1866 if (ret & VM_FAULT_ERROR) {
1867 if (ret & VM_FAULT_OOM)
1868 return i ? i : -ENOMEM;
1869 if (ret & (VM_FAULT_HWPOISON |
1870 VM_FAULT_HWPOISON_LARGE)) {
1871 if (i)
1872 return i;
1873 else if (gup_flags & FOLL_HWPOISON)
1874 return -EHWPOISON;
1875 else
1876 return -EFAULT;
1877 }
1878 if (ret & VM_FAULT_SIGBUS)
1879 goto efault;
1880 BUG();
1881 }
1882
1883 if (tsk) {
1884 if (ret & VM_FAULT_MAJOR)
1885 tsk->maj_flt++;
1886 else
1887 tsk->min_flt++;
1888 }
1889
1890 if (ret & VM_FAULT_RETRY) {
1891 if (nonblocking)
1892 *nonblocking = 0;
1893 return i;
1894 }
1895
1896 /*
1897 * The VM_FAULT_WRITE bit tells us that
1898 * do_wp_page has broken COW when necessary,
1899 * even if maybe_mkwrite decided not to set
1900 * pte_write. We can thus safely do subsequent
1901 * page lookups as if they were reads. But only
1902 * do so when looping for pte_write is futile:
1903 * in some cases userspace may also be wanting
1904 * to write to the gotten user page, which a
1905 * read fault here might prevent (a readonly
1906 * page might get reCOWed by userspace write).
1907 */
1908 if ((ret & VM_FAULT_WRITE) &&
1909 !(vma->vm_flags & VM_WRITE))
1910 foll_flags &= ~FOLL_WRITE;
1911
1912 cond_resched();
1913 }
1914 if (IS_ERR(page))
1915 return i ? i : PTR_ERR(page);
1916 if (pages) {
1917 pages[i] = page;
1918
1919 flush_anon_page(vma, page, start);
1920 flush_dcache_page(page);
1921 page_mask = 0;
1922 }
1923next_page:
1924 if (vmas) {
1925 vmas[i] = vma;
1926 page_mask = 0;
1927 }
1928 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1929 if (page_increm > nr_pages)
1930 page_increm = nr_pages;
1931 i += page_increm;
1932 start += page_increm * PAGE_SIZE;
1933 nr_pages -= page_increm;
1934 } while (nr_pages && start < vma->vm_end);
1935 } while (nr_pages);
1936 return i;
1937efault:
1938 return i ? : -EFAULT;
1939}
1940EXPORT_SYMBOL(__get_user_pages);
1941
1942/*
1943 * fixup_user_fault() - manually resolve a user page fault
1944 * @tsk: the task_struct to use for page fault accounting, or
1945 * NULL if faults are not to be recorded.
1946 * @mm: mm_struct of target mm
1947 * @address: user address
1948 * @fault_flags:flags to pass down to handle_mm_fault()
1949 *
1950 * This is meant to be called in the specific scenario where for locking reasons
1951 * we try to access user memory in atomic context (within a pagefault_disable()
1952 * section), this returns -EFAULT, and we want to resolve the user fault before
1953 * trying again.
1954 *
1955 * Typically this is meant to be used by the futex code.
1956 *
1957 * The main difference with get_user_pages() is that this function will
1958 * unconditionally call handle_mm_fault() which will in turn perform all the
1959 * necessary SW fixup of the dirty and young bits in the PTE, while
1960 * handle_mm_fault() only guarantees to update these in the struct page.
1961 *
1962 * This is important for some architectures where those bits also gate the
1963 * access permission to the page because they are maintained in software. On
1964 * such architectures, gup() will not be enough to make a subsequent access
1965 * succeed.
1966 *
1967 * This should be called with the mm_sem held for read.
1968 */
1969int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1970 unsigned long address, unsigned int fault_flags)
1971{
1972 struct vm_area_struct *vma;
1973 vm_flags_t vm_flags;
1974 int ret;
1975
1976 vma = find_extend_vma(mm, address);
1977 if (!vma || address < vma->vm_start)
1978 return -EFAULT;
1979
1980 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
1981 if (!(vm_flags & vma->vm_flags))
1982 return -EFAULT;
1983
1984 ret = handle_mm_fault(mm, vma, address, fault_flags);
1985 if (ret & VM_FAULT_ERROR) {
1986 if (ret & VM_FAULT_OOM)
1987 return -ENOMEM;
1988 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1989 return -EHWPOISON;
1990 if (ret & VM_FAULT_SIGBUS)
1991 return -EFAULT;
1992 BUG();
1993 }
1994 if (tsk) {
1995 if (ret & VM_FAULT_MAJOR)
1996 tsk->maj_flt++;
1997 else
1998 tsk->min_flt++;
1999 }
2000 return 0;
2001}
2002
2003/*
2004 * get_user_pages() - pin user pages in memory
2005 * @tsk: the task_struct to use for page fault accounting, or
2006 * NULL if faults are not to be recorded.
2007 * @mm: mm_struct of target mm
2008 * @start: starting user address
2009 * @nr_pages: number of pages from start to pin
2010 * @write: whether pages will be written to by the caller
2011 * @force: whether to force access even when user mapping is currently
2012 * protected (but never forces write access to shared mapping).
2013 * @pages: array that receives pointers to the pages pinned.
2014 * Should be at least nr_pages long. Or NULL, if caller
2015 * only intends to ensure the pages are faulted in.
2016 * @vmas: array of pointers to vmas corresponding to each page.
2017 * Or NULL if the caller does not require them.
2018 *
2019 * Returns number of pages pinned. This may be fewer than the number
2020 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2021 * were pinned, returns -errno. Each page returned must be released
2022 * with a put_page() call when it is finished with. vmas will only
2023 * remain valid while mmap_sem is held.
2024 *
2025 * Must be called with mmap_sem held for read or write.
2026 *
2027 * get_user_pages walks a process's page tables and takes a reference to
2028 * each struct page that each user address corresponds to at a given
2029 * instant. That is, it takes the page that would be accessed if a user
2030 * thread accesses the given user virtual address at that instant.
2031 *
2032 * This does not guarantee that the page exists in the user mappings when
2033 * get_user_pages returns, and there may even be a completely different
2034 * page there in some cases (eg. if mmapped pagecache has been invalidated
2035 * and subsequently re faulted). However it does guarantee that the page
2036 * won't be freed completely. And mostly callers simply care that the page
2037 * contains data that was valid *at some point in time*. Typically, an IO
2038 * or similar operation cannot guarantee anything stronger anyway because
2039 * locks can't be held over the syscall boundary.
2040 *
2041 * If write=0, the page must not be written to. If the page is written to,
2042 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2043 * after the page is finished with, and before put_page is called.
2044 *
2045 * get_user_pages is typically used for fewer-copy IO operations, to get a
2046 * handle on the memory by some means other than accesses via the user virtual
2047 * addresses. The pages may be submitted for DMA to devices or accessed via
2048 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2049 * use the correct cache flushing APIs.
2050 *
2051 * See also get_user_pages_fast, for performance critical applications.
2052 */
2053long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2054 unsigned long start, unsigned long nr_pages, int write,
2055 int force, struct page **pages, struct vm_area_struct **vmas)
2056{
2057 int flags = FOLL_TOUCH;
2058
2059 if (pages)
2060 flags |= FOLL_GET;
2061 if (write)
2062 flags |= FOLL_WRITE;
2063 if (force)
2064 flags |= FOLL_FORCE;
2065
2066 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2067 NULL);
2068}
2069EXPORT_SYMBOL(get_user_pages);
2070
2071/**
2072 * get_dump_page() - pin user page in memory while writing it to core dump
2073 * @addr: user address
2074 *
2075 * Returns struct page pointer of user page pinned for dump,
2076 * to be freed afterwards by page_cache_release() or put_page().
2077 *
2078 * Returns NULL on any kind of failure - a hole must then be inserted into
2079 * the corefile, to preserve alignment with its headers; and also returns
2080 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2081 * allowing a hole to be left in the corefile to save diskspace.
2082 *
2083 * Called without mmap_sem, but after all other threads have been killed.
2084 */
2085#ifdef CONFIG_ELF_CORE
2086struct page *get_dump_page(unsigned long addr)
2087{
2088 struct vm_area_struct *vma;
2089 struct page *page;
2090
2091 if (__get_user_pages(current, current->mm, addr, 1,
2092 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2093 NULL) < 1)
2094 return NULL;
2095 flush_cache_page(vma, addr, page_to_pfn(page));
2096 return page;
2097}
2098#endif /* CONFIG_ELF_CORE */
2099
2100pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2101 spinlock_t **ptl)
2102{
2103 pgd_t * pgd = pgd_offset(mm, addr);
2104 pud_t * pud = pud_alloc(mm, pgd, addr);
2105 if (pud) {
2106 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2107 if (pmd) {
2108 VM_BUG_ON(pmd_trans_huge(*pmd));
2109 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2110 }
2111 }
2112 return NULL;
2113}
2114
2115/*
2116 * This is the old fallback for page remapping.
2117 *
2118 * For historical reasons, it only allows reserved pages. Only
2119 * old drivers should use this, and they needed to mark their
2120 * pages reserved for the old functions anyway.
2121 */
2122static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2123 struct page *page, pgprot_t prot)
2124{
2125 struct mm_struct *mm = vma->vm_mm;
2126 int retval;
2127 pte_t *pte;
2128 spinlock_t *ptl;
2129
2130 retval = -EINVAL;
2131 if (PageAnon(page))
2132 goto out;
2133 retval = -ENOMEM;
2134 flush_dcache_page(page);
2135 pte = get_locked_pte(mm, addr, &ptl);
2136 if (!pte)
2137 goto out;
2138 retval = -EBUSY;
2139 if (!pte_none(*pte))
2140 goto out_unlock;
2141
2142 /* Ok, finally just insert the thing.. */
2143 get_page(page);
2144 inc_mm_counter_fast(mm, MM_FILEPAGES);
2145 page_add_file_rmap(page);
2146 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2147
2148 retval = 0;
2149 pte_unmap_unlock(pte, ptl);
2150 return retval;
2151out_unlock:
2152 pte_unmap_unlock(pte, ptl);
2153out:
2154 return retval;
2155}
2156
2157/**
2158 * vm_insert_page - insert single page into user vma
2159 * @vma: user vma to map to
2160 * @addr: target user address of this page
2161 * @page: source kernel page
2162 *
2163 * This allows drivers to insert individual pages they've allocated
2164 * into a user vma.
2165 *
2166 * The page has to be a nice clean _individual_ kernel allocation.
2167 * If you allocate a compound page, you need to have marked it as
2168 * such (__GFP_COMP), or manually just split the page up yourself
2169 * (see split_page()).
2170 *
2171 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2172 * took an arbitrary page protection parameter. This doesn't allow
2173 * that. Your vma protection will have to be set up correctly, which
2174 * means that if you want a shared writable mapping, you'd better
2175 * ask for a shared writable mapping!
2176 *
2177 * The page does not need to be reserved.
2178 *
2179 * Usually this function is called from f_op->mmap() handler
2180 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2181 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2182 * function from other places, for example from page-fault handler.
2183 */
2184int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2185 struct page *page)
2186{
2187 if (addr < vma->vm_start || addr >= vma->vm_end)
2188 return -EFAULT;
2189 if (!page_count(page))
2190 return -EINVAL;
2191 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2192 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2193 BUG_ON(vma->vm_flags & VM_PFNMAP);
2194 vma->vm_flags |= VM_MIXEDMAP;
2195 }
2196 return insert_page(vma, addr, page, vma->vm_page_prot);
2197}
2198EXPORT_SYMBOL(vm_insert_page);
2199
2200static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2201 unsigned long pfn, pgprot_t prot)
2202{
2203 struct mm_struct *mm = vma->vm_mm;
2204 int retval;
2205 pte_t *pte, entry;
2206 spinlock_t *ptl;
2207
2208 retval = -ENOMEM;
2209 pte = get_locked_pte(mm, addr, &ptl);
2210 if (!pte)
2211 goto out;
2212 retval = -EBUSY;
2213 if (!pte_none(*pte))
2214 goto out_unlock;
2215
2216 /* Ok, finally just insert the thing.. */
2217 entry = pte_mkspecial(pfn_pte(pfn, prot));
2218 set_pte_at(mm, addr, pte, entry);
2219 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2220
2221 retval = 0;
2222out_unlock:
2223 pte_unmap_unlock(pte, ptl);
2224out:
2225 return retval;
2226}
2227
2228/**
2229 * vm_insert_pfn - insert single pfn into user vma
2230 * @vma: user vma to map to
2231 * @addr: target user address of this page
2232 * @pfn: source kernel pfn
2233 *
2234 * Similar to vm_insert_page, this allows drivers to insert individual pages
2235 * they've allocated into a user vma. Same comments apply.
2236 *
2237 * This function should only be called from a vm_ops->fault handler, and
2238 * in that case the handler should return NULL.
2239 *
2240 * vma cannot be a COW mapping.
2241 *
2242 * As this is called only for pages that do not currently exist, we
2243 * do not need to flush old virtual caches or the TLB.
2244 */
2245int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2246 unsigned long pfn)
2247{
2248 int ret;
2249 pgprot_t pgprot = vma->vm_page_prot;
2250 /*
2251 * Technically, architectures with pte_special can avoid all these
2252 * restrictions (same for remap_pfn_range). However we would like
2253 * consistency in testing and feature parity among all, so we should
2254 * try to keep these invariants in place for everybody.
2255 */
2256 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2257 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2258 (VM_PFNMAP|VM_MIXEDMAP));
2259 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2260 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2261
2262 if (addr < vma->vm_start || addr >= vma->vm_end)
2263 return -EFAULT;
2264 if (track_pfn_insert(vma, &pgprot, pfn))
2265 return -EINVAL;
2266
2267 ret = insert_pfn(vma, addr, pfn, pgprot);
2268
2269 return ret;
2270}
2271EXPORT_SYMBOL(vm_insert_pfn);
2272
2273int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2274 unsigned long pfn)
2275{
2276 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2277
2278 if (addr < vma->vm_start || addr >= vma->vm_end)
2279 return -EFAULT;
2280
2281 /*
2282 * If we don't have pte special, then we have to use the pfn_valid()
2283 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2284 * refcount the page if pfn_valid is true (hence insert_page rather
2285 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2286 * without pte special, it would there be refcounted as a normal page.
2287 */
2288 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2289 struct page *page;
2290
2291 page = pfn_to_page(pfn);
2292 return insert_page(vma, addr, page, vma->vm_page_prot);
2293 }
2294 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2295}
2296EXPORT_SYMBOL(vm_insert_mixed);
2297
2298/*
2299 * maps a range of physical memory into the requested pages. the old
2300 * mappings are removed. any references to nonexistent pages results
2301 * in null mappings (currently treated as "copy-on-access")
2302 */
2303static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2304 unsigned long addr, unsigned long end,
2305 unsigned long pfn, pgprot_t prot)
2306{
2307 pte_t *pte;
2308 spinlock_t *ptl;
2309
2310 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2311 if (!pte)
2312 return -ENOMEM;
2313 arch_enter_lazy_mmu_mode();
2314 do {
2315 BUG_ON(!pte_none(*pte));
2316 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2317 pfn++;
2318 } while (pte++, addr += PAGE_SIZE, addr != end);
2319 arch_leave_lazy_mmu_mode();
2320 pte_unmap_unlock(pte - 1, ptl);
2321 return 0;
2322}
2323
2324static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2325 unsigned long addr, unsigned long end,
2326 unsigned long pfn, pgprot_t prot)
2327{
2328 pmd_t *pmd;
2329 unsigned long next;
2330
2331 pfn -= addr >> PAGE_SHIFT;
2332 pmd = pmd_alloc(mm, pud, addr);
2333 if (!pmd)
2334 return -ENOMEM;
2335 VM_BUG_ON(pmd_trans_huge(*pmd));
2336 do {
2337 next = pmd_addr_end(addr, end);
2338 if (remap_pte_range(mm, pmd, addr, next,
2339 pfn + (addr >> PAGE_SHIFT), prot))
2340 return -ENOMEM;
2341 } while (pmd++, addr = next, addr != end);
2342 return 0;
2343}
2344
2345static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2346 unsigned long addr, unsigned long end,
2347 unsigned long pfn, pgprot_t prot)
2348{
2349 pud_t *pud;
2350 unsigned long next;
2351
2352 pfn -= addr >> PAGE_SHIFT;
2353 pud = pud_alloc(mm, pgd, addr);
2354 if (!pud)
2355 return -ENOMEM;
2356 do {
2357 next = pud_addr_end(addr, end);
2358 if (remap_pmd_range(mm, pud, addr, next,
2359 pfn + (addr >> PAGE_SHIFT), prot))
2360 return -ENOMEM;
2361 } while (pud++, addr = next, addr != end);
2362 return 0;
2363}
2364
2365/**
2366 * remap_pfn_range - remap kernel memory to userspace
2367 * @vma: user vma to map to
2368 * @addr: target user address to start at
2369 * @pfn: physical address of kernel memory
2370 * @size: size of map area
2371 * @prot: page protection flags for this mapping
2372 *
2373 * Note: this is only safe if the mm semaphore is held when called.
2374 */
2375int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2376 unsigned long pfn, unsigned long size, pgprot_t prot)
2377{
2378 pgd_t *pgd;
2379 unsigned long next;
2380 unsigned long end = addr + PAGE_ALIGN(size);
2381 struct mm_struct *mm = vma->vm_mm;
2382 int err;
2383
2384 /*
2385 * Physically remapped pages are special. Tell the
2386 * rest of the world about it:
2387 * VM_IO tells people not to look at these pages
2388 * (accesses can have side effects).
2389 * VM_PFNMAP tells the core MM that the base pages are just
2390 * raw PFN mappings, and do not have a "struct page" associated
2391 * with them.
2392 * VM_DONTEXPAND
2393 * Disable vma merging and expanding with mremap().
2394 * VM_DONTDUMP
2395 * Omit vma from core dump, even when VM_IO turned off.
2396 *
2397 * There's a horrible special case to handle copy-on-write
2398 * behaviour that some programs depend on. We mark the "original"
2399 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2400 * See vm_normal_page() for details.
2401 */
2402 if (is_cow_mapping(vma->vm_flags)) {
2403 if (addr != vma->vm_start || end != vma->vm_end)
2404 return -EINVAL;
2405 vma->vm_pgoff = pfn;
2406 }
2407
2408 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2409 if (err)
2410 return -EINVAL;
2411
2412 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2413
2414 BUG_ON(addr >= end);
2415 pfn -= addr >> PAGE_SHIFT;
2416 pgd = pgd_offset(mm, addr);
2417 flush_cache_range(vma, addr, end);
2418 do {
2419 next = pgd_addr_end(addr, end);
2420 err = remap_pud_range(mm, pgd, addr, next,
2421 pfn + (addr >> PAGE_SHIFT), prot);
2422 if (err)
2423 break;
2424 } while (pgd++, addr = next, addr != end);
2425
2426 if (err)
2427 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2428
2429 return err;
2430}
2431EXPORT_SYMBOL(remap_pfn_range);
2432
2433/**
2434 * vm_iomap_memory - remap memory to userspace
2435 * @vma: user vma to map to
2436 * @start: start of area
2437 * @len: size of area
2438 *
2439 * This is a simplified io_remap_pfn_range() for common driver use. The
2440 * driver just needs to give us the physical memory range to be mapped,
2441 * we'll figure out the rest from the vma information.
2442 *
2443 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2444 * whatever write-combining details or similar.
2445 */
2446int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2447{
2448 unsigned long vm_len, pfn, pages;
2449
2450 /* Check that the physical memory area passed in looks valid */
2451 if (start + len < start)
2452 return -EINVAL;
2453 /*
2454 * You *really* shouldn't map things that aren't page-aligned,
2455 * but we've historically allowed it because IO memory might
2456 * just have smaller alignment.
2457 */
2458 len += start & ~PAGE_MASK;
2459 pfn = start >> PAGE_SHIFT;
2460 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2461 if (pfn + pages < pfn)
2462 return -EINVAL;
2463
2464 /* We start the mapping 'vm_pgoff' pages into the area */
2465 if (vma->vm_pgoff > pages)
2466 return -EINVAL;
2467 pfn += vma->vm_pgoff;
2468 pages -= vma->vm_pgoff;
2469
2470 /* Can we fit all of the mapping? */
2471 vm_len = vma->vm_end - vma->vm_start;
2472 if (vm_len >> PAGE_SHIFT > pages)
2473 return -EINVAL;
2474
2475 /* Ok, let it rip */
2476 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2477}
2478EXPORT_SYMBOL(vm_iomap_memory);
2479
2480static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2481 unsigned long addr, unsigned long end,
2482 pte_fn_t fn, void *data)
2483{
2484 pte_t *pte;
2485 int err;
2486 pgtable_t token;
2487 spinlock_t *uninitialized_var(ptl);
2488
2489 pte = (mm == &init_mm) ?
2490 pte_alloc_kernel(pmd, addr) :
2491 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2492 if (!pte)
2493 return -ENOMEM;
2494
2495 BUG_ON(pmd_huge(*pmd));
2496
2497 arch_enter_lazy_mmu_mode();
2498
2499 token = pmd_pgtable(*pmd);
2500
2501 do {
2502 err = fn(pte++, token, addr, data);
2503 if (err)
2504 break;
2505 } while (addr += PAGE_SIZE, addr != end);
2506
2507 arch_leave_lazy_mmu_mode();
2508
2509 if (mm != &init_mm)
2510 pte_unmap_unlock(pte-1, ptl);
2511 return err;
2512}
2513
2514static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2515 unsigned long addr, unsigned long end,
2516 pte_fn_t fn, void *data)
2517{
2518 pmd_t *pmd;
2519 unsigned long next;
2520 int err;
2521
2522 BUG_ON(pud_huge(*pud));
2523
2524 pmd = pmd_alloc(mm, pud, addr);
2525 if (!pmd)
2526 return -ENOMEM;
2527 do {
2528 next = pmd_addr_end(addr, end);
2529 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2530 if (err)
2531 break;
2532 } while (pmd++, addr = next, addr != end);
2533 return err;
2534}
2535
2536static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2537 unsigned long addr, unsigned long end,
2538 pte_fn_t fn, void *data)
2539{
2540 pud_t *pud;
2541 unsigned long next;
2542 int err;
2543
2544 pud = pud_alloc(mm, pgd, addr);
2545 if (!pud)
2546 return -ENOMEM;
2547 do {
2548 next = pud_addr_end(addr, end);
2549 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2550 if (err)
2551 break;
2552 } while (pud++, addr = next, addr != end);
2553 return err;
2554}
2555
2556/*
2557 * Scan a region of virtual memory, filling in page tables as necessary
2558 * and calling a provided function on each leaf page table.
2559 */
2560int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2561 unsigned long size, pte_fn_t fn, void *data)
2562{
2563 pgd_t *pgd;
2564 unsigned long next;
2565 unsigned long end = addr + size;
2566 int err;
2567
2568 BUG_ON(addr >= end);
2569 pgd = pgd_offset(mm, addr);
2570 do {
2571 next = pgd_addr_end(addr, end);
2572 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2573 if (err)
2574 break;
2575 } while (pgd++, addr = next, addr != end);
2576
2577 return err;
2578}
2579EXPORT_SYMBOL_GPL(apply_to_page_range);
2580
2581/*
2582 * handle_pte_fault chooses page fault handler according to an entry
2583 * which was read non-atomically. Before making any commitment, on
2584 * those architectures or configurations (e.g. i386 with PAE) which
2585 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2586 * must check under lock before unmapping the pte and proceeding
2587 * (but do_wp_page is only called after already making such a check;
2588 * and do_anonymous_page can safely check later on).
2589 */
2590static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2591 pte_t *page_table, pte_t orig_pte)
2592{
2593 int same = 1;
2594#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2595 if (sizeof(pte_t) > sizeof(unsigned long)) {
2596 spinlock_t *ptl = pte_lockptr(mm, pmd);
2597 spin_lock(ptl);
2598 same = pte_same(*page_table, orig_pte);
2599 spin_unlock(ptl);
2600 }
2601#endif
2602 pte_unmap(page_table);
2603 return same;
2604}
2605
2606static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2607{
2608 debug_dma_assert_idle(src);
2609
2610 /*
2611 * If the source page was a PFN mapping, we don't have
2612 * a "struct page" for it. We do a best-effort copy by
2613 * just copying from the original user address. If that
2614 * fails, we just zero-fill it. Live with it.
2615 */
2616 if (unlikely(!src)) {
2617 void *kaddr = kmap_atomic(dst);
2618 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2619
2620 /*
2621 * This really shouldn't fail, because the page is there
2622 * in the page tables. But it might just be unreadable,
2623 * in which case we just give up and fill the result with
2624 * zeroes.
2625 */
2626 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2627 clear_page(kaddr);
2628 kunmap_atomic(kaddr);
2629 flush_dcache_page(dst);
2630 } else
2631 copy_user_highpage(dst, src, va, vma);
2632}
2633
2634/*
2635 * Notify the address space that the page is about to become writable so that
2636 * it can prohibit this or wait for the page to get into an appropriate state.
2637 *
2638 * We do this without the lock held, so that it can sleep if it needs to.
2639 */
2640static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2641 unsigned long address)
2642{
2643 struct vm_fault vmf;
2644 int ret;
2645
2646 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2647 vmf.pgoff = page->index;
2648 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2649 vmf.page = page;
2650
2651 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2652 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2653 return ret;
2654 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2655 lock_page(page);
2656 if (!page->mapping) {
2657 unlock_page(page);
2658 return 0; /* retry */
2659 }
2660 ret |= VM_FAULT_LOCKED;
2661 } else
2662 VM_BUG_ON_PAGE(!PageLocked(page), page);
2663 return ret;
2664}
2665
2666/*
2667 * This routine handles present pages, when users try to write
2668 * to a shared page. It is done by copying the page to a new address
2669 * and decrementing the shared-page counter for the old page.
2670 *
2671 * Note that this routine assumes that the protection checks have been
2672 * done by the caller (the low-level page fault routine in most cases).
2673 * Thus we can safely just mark it writable once we've done any necessary
2674 * COW.
2675 *
2676 * We also mark the page dirty at this point even though the page will
2677 * change only once the write actually happens. This avoids a few races,
2678 * and potentially makes it more efficient.
2679 *
2680 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2681 * but allow concurrent faults), with pte both mapped and locked.
2682 * We return with mmap_sem still held, but pte unmapped and unlocked.
2683 */
2684static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2685 unsigned long address, pte_t *page_table, pmd_t *pmd,
2686 spinlock_t *ptl, pte_t orig_pte)
2687 __releases(ptl)
2688{
2689 struct page *old_page, *new_page = NULL;
2690 pte_t entry;
2691 int ret = 0;
2692 int page_mkwrite = 0;
2693 struct page *dirty_page = NULL;
2694 unsigned long mmun_start = 0; /* For mmu_notifiers */
2695 unsigned long mmun_end = 0; /* For mmu_notifiers */
2696
2697 old_page = vm_normal_page(vma, address, orig_pte);
2698 if (!old_page) {
2699 /*
2700 * VM_MIXEDMAP !pfn_valid() case
2701 *
2702 * We should not cow pages in a shared writeable mapping.
2703 * Just mark the pages writable as we can't do any dirty
2704 * accounting on raw pfn maps.
2705 */
2706 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2707 (VM_WRITE|VM_SHARED))
2708 goto reuse;
2709 goto gotten;
2710 }
2711
2712 /*
2713 * Take out anonymous pages first, anonymous shared vmas are
2714 * not dirty accountable.
2715 */
2716 if (PageAnon(old_page) && !PageKsm(old_page)) {
2717 if (!trylock_page(old_page)) {
2718 page_cache_get(old_page);
2719 pte_unmap_unlock(page_table, ptl);
2720 lock_page(old_page);
2721 page_table = pte_offset_map_lock(mm, pmd, address,
2722 &ptl);
2723 if (!pte_same(*page_table, orig_pte)) {
2724 unlock_page(old_page);
2725 goto unlock;
2726 }
2727 page_cache_release(old_page);
2728 }
2729 if (reuse_swap_page(old_page)) {
2730 /*
2731 * The page is all ours. Move it to our anon_vma so
2732 * the rmap code will not search our parent or siblings.
2733 * Protected against the rmap code by the page lock.
2734 */
2735 page_move_anon_rmap(old_page, vma, address);
2736 unlock_page(old_page);
2737 goto reuse;
2738 }
2739 unlock_page(old_page);
2740 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2741 (VM_WRITE|VM_SHARED))) {
2742 /*
2743 * Only catch write-faults on shared writable pages,
2744 * read-only shared pages can get COWed by
2745 * get_user_pages(.write=1, .force=1).
2746 */
2747 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2748 int tmp;
2749 page_cache_get(old_page);
2750 pte_unmap_unlock(page_table, ptl);
2751 tmp = do_page_mkwrite(vma, old_page, address);
2752 if (unlikely(!tmp || (tmp &
2753 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2754 page_cache_release(old_page);
2755 return tmp;
2756 }
2757 /*
2758 * Since we dropped the lock we need to revalidate
2759 * the PTE as someone else may have changed it. If
2760 * they did, we just return, as we can count on the
2761 * MMU to tell us if they didn't also make it writable.
2762 */
2763 page_table = pte_offset_map_lock(mm, pmd, address,
2764 &ptl);
2765 if (!pte_same(*page_table, orig_pte)) {
2766 unlock_page(old_page);
2767 goto unlock;
2768 }
2769
2770 page_mkwrite = 1;
2771 }
2772 dirty_page = old_page;
2773 get_page(dirty_page);
2774
2775reuse:
2776 /*
2777 * Clear the pages cpupid information as the existing
2778 * information potentially belongs to a now completely
2779 * unrelated process.
2780 */
2781 if (old_page)
2782 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2783
2784 flush_cache_page(vma, address, pte_pfn(orig_pte));
2785 entry = pte_mkyoung(orig_pte);
2786 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2787 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2788 update_mmu_cache(vma, address, page_table);
2789 pte_unmap_unlock(page_table, ptl);
2790 ret |= VM_FAULT_WRITE;
2791
2792 if (!dirty_page)
2793 return ret;
2794
2795 /*
2796 * Yes, Virginia, this is actually required to prevent a race
2797 * with clear_page_dirty_for_io() from clearing the page dirty
2798 * bit after it clear all dirty ptes, but before a racing
2799 * do_wp_page installs a dirty pte.
2800 *
2801 * do_shared_fault is protected similarly.
2802 */
2803 if (!page_mkwrite) {
2804 wait_on_page_locked(dirty_page);
2805 set_page_dirty_balance(dirty_page);
2806 /* file_update_time outside page_lock */
2807 if (vma->vm_file)
2808 file_update_time(vma->vm_file);
2809 }
2810 put_page(dirty_page);
2811 if (page_mkwrite) {
2812 struct address_space *mapping = dirty_page->mapping;
2813
2814 set_page_dirty(dirty_page);
2815 unlock_page(dirty_page);
2816 page_cache_release(dirty_page);
2817 if (mapping) {
2818 /*
2819 * Some device drivers do not set page.mapping
2820 * but still dirty their pages
2821 */
2822 balance_dirty_pages_ratelimited(mapping);
2823 }
2824 }
2825
2826 return ret;
2827 }
2828
2829 /*
2830 * Ok, we need to copy. Oh, well..
2831 */
2832 page_cache_get(old_page);
2833gotten:
2834 pte_unmap_unlock(page_table, ptl);
2835
2836 if (unlikely(anon_vma_prepare(vma)))
2837 goto oom;
2838
2839 if (is_zero_pfn(pte_pfn(orig_pte))) {
2840 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2841 if (!new_page)
2842 goto oom;
2843 } else {
2844 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2845 if (!new_page)
2846 goto oom;
2847 cow_user_page(new_page, old_page, address, vma);
2848 }
2849 __SetPageUptodate(new_page);
2850
2851 if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))
2852 goto oom_free_new;
2853
2854 mmun_start = address & PAGE_MASK;
2855 mmun_end = mmun_start + PAGE_SIZE;
2856 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2857
2858 /*
2859 * Re-check the pte - we dropped the lock
2860 */
2861 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2862 if (likely(pte_same(*page_table, orig_pte))) {
2863 if (old_page) {
2864 if (!PageAnon(old_page)) {
2865 dec_mm_counter_fast(mm, MM_FILEPAGES);
2866 inc_mm_counter_fast(mm, MM_ANONPAGES);
2867 }
2868 } else
2869 inc_mm_counter_fast(mm, MM_ANONPAGES);
2870 flush_cache_page(vma, address, pte_pfn(orig_pte));
2871 entry = mk_pte(new_page, vma->vm_page_prot);
2872 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2873 /*
2874 * Clear the pte entry and flush it first, before updating the
2875 * pte with the new entry. This will avoid a race condition
2876 * seen in the presence of one thread doing SMC and another
2877 * thread doing COW.
2878 */
2879 ptep_clear_flush(vma, address, page_table);
2880 page_add_new_anon_rmap(new_page, vma, address);
2881 /*
2882 * We call the notify macro here because, when using secondary
2883 * mmu page tables (such as kvm shadow page tables), we want the
2884 * new page to be mapped directly into the secondary page table.
2885 */
2886 set_pte_at_notify(mm, address, page_table, entry);
2887 update_mmu_cache(vma, address, page_table);
2888 if (old_page) {
2889 /*
2890 * Only after switching the pte to the new page may
2891 * we remove the mapcount here. Otherwise another
2892 * process may come and find the rmap count decremented
2893 * before the pte is switched to the new page, and
2894 * "reuse" the old page writing into it while our pte
2895 * here still points into it and can be read by other
2896 * threads.
2897 *
2898 * The critical issue is to order this
2899 * page_remove_rmap with the ptp_clear_flush above.
2900 * Those stores are ordered by (if nothing else,)
2901 * the barrier present in the atomic_add_negative
2902 * in page_remove_rmap.
2903 *
2904 * Then the TLB flush in ptep_clear_flush ensures that
2905 * no process can access the old page before the
2906 * decremented mapcount is visible. And the old page
2907 * cannot be reused until after the decremented
2908 * mapcount is visible. So transitively, TLBs to
2909 * old page will be flushed before it can be reused.
2910 */
2911 page_remove_rmap(old_page);
2912 }
2913
2914 /* Free the old page.. */
2915 new_page = old_page;
2916 ret |= VM_FAULT_WRITE;
2917 } else
2918 mem_cgroup_uncharge_page(new_page);
2919
2920 if (new_page)
2921 page_cache_release(new_page);
2922unlock:
2923 pte_unmap_unlock(page_table, ptl);
2924 if (mmun_end > mmun_start)
2925 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2926 if (old_page) {
2927 /*
2928 * Don't let another task, with possibly unlocked vma,
2929 * keep the mlocked page.
2930 */
2931 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2932 lock_page(old_page); /* LRU manipulation */
2933 munlock_vma_page(old_page);
2934 unlock_page(old_page);
2935 }
2936 page_cache_release(old_page);
2937 }
2938 return ret;
2939oom_free_new:
2940 page_cache_release(new_page);
2941oom:
2942 if (old_page)
2943 page_cache_release(old_page);
2944 return VM_FAULT_OOM;
2945}
2946
2947static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2948 unsigned long start_addr, unsigned long end_addr,
2949 struct zap_details *details)
2950{
2951 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2952}
2953
2954static inline void unmap_mapping_range_tree(struct rb_root *root,
2955 struct zap_details *details)
2956{
2957 struct vm_area_struct *vma;
2958 pgoff_t vba, vea, zba, zea;
2959
2960 vma_interval_tree_foreach(vma, root,
2961 details->first_index, details->last_index) {
2962
2963 vba = vma->vm_pgoff;
2964 vea = vba + vma_pages(vma) - 1;
2965 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2966 zba = details->first_index;
2967 if (zba < vba)
2968 zba = vba;
2969 zea = details->last_index;
2970 if (zea > vea)
2971 zea = vea;
2972
2973 unmap_mapping_range_vma(vma,
2974 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2975 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2976 details);
2977 }
2978}
2979
2980static inline void unmap_mapping_range_list(struct list_head *head,
2981 struct zap_details *details)
2982{
2983 struct vm_area_struct *vma;
2984
2985 /*
2986 * In nonlinear VMAs there is no correspondence between virtual address
2987 * offset and file offset. So we must perform an exhaustive search
2988 * across *all* the pages in each nonlinear VMA, not just the pages
2989 * whose virtual address lies outside the file truncation point.
2990 */
2991 list_for_each_entry(vma, head, shared.nonlinear) {
2992 details->nonlinear_vma = vma;
2993 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2994 }
2995}
2996
2997/**
2998 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2999 * @mapping: the address space containing mmaps to be unmapped.
3000 * @holebegin: byte in first page to unmap, relative to the start of
3001 * the underlying file. This will be rounded down to a PAGE_SIZE
3002 * boundary. Note that this is different from truncate_pagecache(), which
3003 * must keep the partial page. In contrast, we must get rid of
3004 * partial pages.
3005 * @holelen: size of prospective hole in bytes. This will be rounded
3006 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3007 * end of the file.
3008 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3009 * but 0 when invalidating pagecache, don't throw away private data.
3010 */
3011void unmap_mapping_range(struct address_space *mapping,
3012 loff_t const holebegin, loff_t const holelen, int even_cows)
3013{
3014 struct zap_details details;
3015 pgoff_t hba = holebegin >> PAGE_SHIFT;
3016 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3017
3018 /* Check for overflow. */
3019 if (sizeof(holelen) > sizeof(hlen)) {
3020 long long holeend =
3021 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3022 if (holeend & ~(long long)ULONG_MAX)
3023 hlen = ULONG_MAX - hba + 1;
3024 }
3025
3026 details.check_mapping = even_cows? NULL: mapping;
3027 details.nonlinear_vma = NULL;
3028 details.first_index = hba;
3029 details.last_index = hba + hlen - 1;
3030 if (details.last_index < details.first_index)
3031 details.last_index = ULONG_MAX;
3032
3033
3034 mutex_lock(&mapping->i_mmap_mutex);
3035 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
3036 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3037 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
3038 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
3039 mutex_unlock(&mapping->i_mmap_mutex);
3040}
3041EXPORT_SYMBOL(unmap_mapping_range);
3042
3043/*
3044 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3045 * but allow concurrent faults), and pte mapped but not yet locked.
3046 * We return with mmap_sem still held, but pte unmapped and unlocked.
3047 */
3048static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3049 unsigned long address, pte_t *page_table, pmd_t *pmd,
3050 unsigned int flags, pte_t orig_pte)
3051{
3052 spinlock_t *ptl;
3053 struct page *page, *swapcache;
3054 swp_entry_t entry;
3055 pte_t pte;
3056 int locked;
3057 struct mem_cgroup *ptr;
3058 int exclusive = 0;
3059 int ret = 0;
3060
3061 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3062 goto out;
3063
3064 entry = pte_to_swp_entry(orig_pte);
3065 if (unlikely(non_swap_entry(entry))) {
3066 if (is_migration_entry(entry)) {
3067 migration_entry_wait(mm, pmd, address);
3068 } else if (is_hwpoison_entry(entry)) {
3069 ret = VM_FAULT_HWPOISON;
3070 } else {
3071 print_bad_pte(vma, address, orig_pte, NULL);
3072 ret = VM_FAULT_SIGBUS;
3073 }
3074 goto out;
3075 }
3076 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3077 page = lookup_swap_cache(entry);
3078 if (!page) {
3079 page = swapin_readahead(entry,
3080 GFP_HIGHUSER_MOVABLE, vma, address);
3081 if (!page) {
3082 /*
3083 * Back out if somebody else faulted in this pte
3084 * while we released the pte lock.
3085 */
3086 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3087 if (likely(pte_same(*page_table, orig_pte)))
3088 ret = VM_FAULT_OOM;
3089 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3090 goto unlock;
3091 }
3092
3093 /* Had to read the page from swap area: Major fault */
3094 ret = VM_FAULT_MAJOR;
3095 count_vm_event(PGMAJFAULT);
3096 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3097 } else if (PageHWPoison(page)) {
3098 /*
3099 * hwpoisoned dirty swapcache pages are kept for killing
3100 * owner processes (which may be unknown at hwpoison time)
3101 */
3102 ret = VM_FAULT_HWPOISON;
3103 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3104 swapcache = page;
3105 goto out_release;
3106 }
3107
3108 swapcache = page;
3109 locked = lock_page_or_retry(page, mm, flags);
3110
3111 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3112 if (!locked) {
3113 ret |= VM_FAULT_RETRY;
3114 goto out_release;
3115 }
3116
3117 /*
3118 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3119 * release the swapcache from under us. The page pin, and pte_same
3120 * test below, are not enough to exclude that. Even if it is still
3121 * swapcache, we need to check that the page's swap has not changed.
3122 */
3123 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3124 goto out_page;
3125
3126 page = ksm_might_need_to_copy(page, vma, address);
3127 if (unlikely(!page)) {
3128 ret = VM_FAULT_OOM;
3129 page = swapcache;
3130 goto out_page;
3131 }
3132
3133 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3134 ret = VM_FAULT_OOM;
3135 goto out_page;
3136 }
3137
3138 /*
3139 * Back out if somebody else already faulted in this pte.
3140 */
3141 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3142 if (unlikely(!pte_same(*page_table, orig_pte)))
3143 goto out_nomap;
3144
3145 if (unlikely(!PageUptodate(page))) {
3146 ret = VM_FAULT_SIGBUS;
3147 goto out_nomap;
3148 }
3149
3150 /*
3151 * The page isn't present yet, go ahead with the fault.
3152 *
3153 * Be careful about the sequence of operations here.
3154 * To get its accounting right, reuse_swap_page() must be called
3155 * while the page is counted on swap but not yet in mapcount i.e.
3156 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3157 * must be called after the swap_free(), or it will never succeed.
3158 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3159 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3160 * in page->private. In this case, a record in swap_cgroup is silently
3161 * discarded at swap_free().
3162 */
3163
3164 inc_mm_counter_fast(mm, MM_ANONPAGES);
3165 dec_mm_counter_fast(mm, MM_SWAPENTS);
3166 pte = mk_pte(page, vma->vm_page_prot);
3167 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3168 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3169 flags &= ~FAULT_FLAG_WRITE;
3170 ret |= VM_FAULT_WRITE;
3171 exclusive = 1;
3172 }
3173 flush_icache_page(vma, page);
3174 if (pte_swp_soft_dirty(orig_pte))
3175 pte = pte_mksoft_dirty(pte);
3176 set_pte_at(mm, address, page_table, pte);
3177 if (page == swapcache)
3178 do_page_add_anon_rmap(page, vma, address, exclusive);
3179 else /* ksm created a completely new copy */
3180 page_add_new_anon_rmap(page, vma, address);
3181 /* It's better to call commit-charge after rmap is established */
3182 mem_cgroup_commit_charge_swapin(page, ptr);
3183
3184 swap_free(entry);
3185 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3186 try_to_free_swap(page);
3187 unlock_page(page);
3188 if (page != swapcache) {
3189 /*
3190 * Hold the lock to avoid the swap entry to be reused
3191 * until we take the PT lock for the pte_same() check
3192 * (to avoid false positives from pte_same). For
3193 * further safety release the lock after the swap_free
3194 * so that the swap count won't change under a
3195 * parallel locked swapcache.
3196 */
3197 unlock_page(swapcache);
3198 page_cache_release(swapcache);
3199 }
3200
3201 if (flags & FAULT_FLAG_WRITE) {
3202 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3203 if (ret & VM_FAULT_ERROR)
3204 ret &= VM_FAULT_ERROR;
3205 goto out;
3206 }
3207
3208 /* No need to invalidate - it was non-present before */
3209 update_mmu_cache(vma, address, page_table);
3210unlock:
3211 pte_unmap_unlock(page_table, ptl);
3212out:
3213 return ret;
3214out_nomap:
3215 mem_cgroup_cancel_charge_swapin(ptr);
3216 pte_unmap_unlock(page_table, ptl);
3217out_page:
3218 unlock_page(page);
3219out_release:
3220 page_cache_release(page);
3221 if (page != swapcache) {
3222 unlock_page(swapcache);
3223 page_cache_release(swapcache);
3224 }
3225 return ret;
3226}
3227
3228/*
3229 * This is like a special single-page "expand_{down|up}wards()",
3230 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3231 * doesn't hit another vma.
3232 */
3233static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3234{
3235 address &= PAGE_MASK;
3236 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3237 struct vm_area_struct *prev = vma->vm_prev;
3238
3239 /*
3240 * Is there a mapping abutting this one below?
3241 *
3242 * That's only ok if it's the same stack mapping
3243 * that has gotten split..
3244 */
3245 if (prev && prev->vm_end == address)
3246 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3247
3248 expand_downwards(vma, address - PAGE_SIZE);
3249 }
3250 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3251 struct vm_area_struct *next = vma->vm_next;
3252
3253 /* As VM_GROWSDOWN but s/below/above/ */
3254 if (next && next->vm_start == address + PAGE_SIZE)
3255 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3256
3257 expand_upwards(vma, address + PAGE_SIZE);
3258 }
3259 return 0;
3260}
3261
3262/*
3263 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3264 * but allow concurrent faults), and pte mapped but not yet locked.
3265 * We return with mmap_sem still held, but pte unmapped and unlocked.
3266 */
3267static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3268 unsigned long address, pte_t *page_table, pmd_t *pmd,
3269 unsigned int flags)
3270{
3271 struct page *page;
3272 spinlock_t *ptl;
3273 pte_t entry;
3274
3275 pte_unmap(page_table);
3276
3277 /* Check if we need to add a guard page to the stack */
3278 if (check_stack_guard_page(vma, address) < 0)
3279 return VM_FAULT_SIGBUS;
3280
3281 /* Use the zero-page for reads */
3282 if (!(flags & FAULT_FLAG_WRITE)) {
3283 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3284 vma->vm_page_prot));
3285 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3286 if (!pte_none(*page_table))
3287 goto unlock;
3288 goto setpte;
3289 }
3290
3291 /* Allocate our own private page. */
3292 if (unlikely(anon_vma_prepare(vma)))
3293 goto oom;
3294 page = alloc_zeroed_user_highpage_movable(vma, address);
3295 if (!page)
3296 goto oom;
3297 /*
3298 * The memory barrier inside __SetPageUptodate makes sure that
3299 * preceeding stores to the page contents become visible before
3300 * the set_pte_at() write.
3301 */
3302 __SetPageUptodate(page);
3303
3304 if (mem_cgroup_charge_anon(page, mm, GFP_KERNEL))
3305 goto oom_free_page;
3306
3307 entry = mk_pte(page, vma->vm_page_prot);
3308 if (vma->vm_flags & VM_WRITE)
3309 entry = pte_mkwrite(pte_mkdirty(entry));
3310
3311 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3312 if (!pte_none(*page_table))
3313 goto release;
3314
3315 inc_mm_counter_fast(mm, MM_ANONPAGES);
3316 page_add_new_anon_rmap(page, vma, address);
3317setpte:
3318 set_pte_at(mm, address, page_table, entry);
3319
3320 /* No need to invalidate - it was non-present before */
3321 update_mmu_cache(vma, address, page_table);
3322unlock:
3323 pte_unmap_unlock(page_table, ptl);
3324 return 0;
3325release:
3326 mem_cgroup_uncharge_page(page);
3327 page_cache_release(page);
3328 goto unlock;
3329oom_free_page:
3330 page_cache_release(page);
3331oom:
3332 return VM_FAULT_OOM;
3333}
3334
3335static int __do_fault(struct vm_area_struct *vma, unsigned long address,
3336 pgoff_t pgoff, unsigned int flags, struct page **page)
3337{
3338 struct vm_fault vmf;
3339 int ret;
3340
3341 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3342 vmf.pgoff = pgoff;
3343 vmf.flags = flags;
3344 vmf.page = NULL;
3345
3346 ret = vma->vm_ops->fault(vma, &vmf);
3347 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3348 return ret;
3349
3350 if (unlikely(PageHWPoison(vmf.page))) {
3351 if (ret & VM_FAULT_LOCKED)
3352 unlock_page(vmf.page);
3353 page_cache_release(vmf.page);
3354 return VM_FAULT_HWPOISON;
3355 }
3356
3357 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3358 lock_page(vmf.page);
3359 else
3360 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
3361
3362 *page = vmf.page;
3363 return ret;
3364}
3365
3366/**
3367 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
3368 *
3369 * @vma: virtual memory area
3370 * @address: user virtual address
3371 * @page: page to map
3372 * @pte: pointer to target page table entry
3373 * @write: true, if new entry is writable
3374 * @anon: true, if it's anonymous page
3375 *
3376 * Caller must hold page table lock relevant for @pte.
3377 *
3378 * Target users are page handler itself and implementations of
3379 * vm_ops->map_pages.
3380 */
3381void do_set_pte(struct vm_area_struct *vma, unsigned long address,
3382 struct page *page, pte_t *pte, bool write, bool anon)
3383{
3384 pte_t entry;
3385
3386 flush_icache_page(vma, page);
3387 entry = mk_pte(page, vma->vm_page_prot);
3388 if (write)
3389 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3390 else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
3391 pte_mksoft_dirty(entry);
3392 if (anon) {
3393 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3394 page_add_new_anon_rmap(page, vma, address);
3395 } else {
3396 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
3397 page_add_file_rmap(page);
3398 }
3399 set_pte_at(vma->vm_mm, address, pte, entry);
3400
3401 /* no need to invalidate: a not-present page won't be cached */
3402 update_mmu_cache(vma, address, pte);
3403}
3404
3405#define FAULT_AROUND_ORDER 4
3406
3407#ifdef CONFIG_DEBUG_FS
3408static unsigned int fault_around_order = FAULT_AROUND_ORDER;
3409
3410static int fault_around_order_get(void *data, u64 *val)
3411{
3412 *val = fault_around_order;
3413 return 0;
3414}
3415
3416static int fault_around_order_set(void *data, u64 val)
3417{
3418 BUILD_BUG_ON((1UL << FAULT_AROUND_ORDER) > PTRS_PER_PTE);
3419 if (1UL << val > PTRS_PER_PTE)
3420 return -EINVAL;
3421 fault_around_order = val;
3422 return 0;
3423}
3424DEFINE_SIMPLE_ATTRIBUTE(fault_around_order_fops,
3425 fault_around_order_get, fault_around_order_set, "%llu\n");
3426
3427static int __init fault_around_debugfs(void)
3428{
3429 void *ret;
3430
3431 ret = debugfs_create_file("fault_around_order", 0644, NULL, NULL,
3432 &fault_around_order_fops);
3433 if (!ret)
3434 pr_warn("Failed to create fault_around_order in debugfs");
3435 return 0;
3436}
3437late_initcall(fault_around_debugfs);
3438
3439static inline unsigned long fault_around_pages(void)
3440{
3441 return 1UL << fault_around_order;
3442}
3443
3444static inline unsigned long fault_around_mask(void)
3445{
3446 return ~((1UL << (PAGE_SHIFT + fault_around_order)) - 1);
3447}
3448#else
3449static inline unsigned long fault_around_pages(void)
3450{
3451 unsigned long nr_pages;
3452
3453 nr_pages = 1UL << FAULT_AROUND_ORDER;
3454 BUILD_BUG_ON(nr_pages > PTRS_PER_PTE);
3455 return nr_pages;
3456}
3457
3458static inline unsigned long fault_around_mask(void)
3459{
3460 return ~((1UL << (PAGE_SHIFT + FAULT_AROUND_ORDER)) - 1);
3461}
3462#endif
3463
3464static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
3465 pte_t *pte, pgoff_t pgoff, unsigned int flags)
3466{
3467 unsigned long start_addr;
3468 pgoff_t max_pgoff;
3469 struct vm_fault vmf;
3470 int off;
3471
3472 start_addr = max(address & fault_around_mask(), vma->vm_start);
3473 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3474 pte -= off;
3475 pgoff -= off;
3476
3477 /*
3478 * max_pgoff is either end of page table or end of vma
3479 * or fault_around_pages() from pgoff, depending what is neast.
3480 */
3481 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3482 PTRS_PER_PTE - 1;
3483 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
3484 pgoff + fault_around_pages() - 1);
3485
3486 /* Check if it makes any sense to call ->map_pages */
3487 while (!pte_none(*pte)) {
3488 if (++pgoff > max_pgoff)
3489 return;
3490 start_addr += PAGE_SIZE;
3491 if (start_addr >= vma->vm_end)
3492 return;
3493 pte++;
3494 }
3495
3496 vmf.virtual_address = (void __user *) start_addr;
3497 vmf.pte = pte;
3498 vmf.pgoff = pgoff;
3499 vmf.max_pgoff = max_pgoff;
3500 vmf.flags = flags;
3501 vma->vm_ops->map_pages(vma, &vmf);
3502}
3503
3504static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3505 unsigned long address, pmd_t *pmd,
3506 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3507{
3508 struct page *fault_page;
3509 spinlock_t *ptl;
3510 pte_t *pte;
3511 int ret = 0;
3512
3513 /*
3514 * Let's call ->map_pages() first and use ->fault() as fallback
3515 * if page by the offset is not ready to be mapped (cold cache or
3516 * something).
3517 */
3518 if (vma->vm_ops->map_pages) {
3519 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3520 do_fault_around(vma, address, pte, pgoff, flags);
3521 if (!pte_same(*pte, orig_pte))
3522 goto unlock_out;
3523 pte_unmap_unlock(pte, ptl);
3524 }
3525
3526 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3527 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3528 return ret;
3529
3530 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3531 if (unlikely(!pte_same(*pte, orig_pte))) {
3532 pte_unmap_unlock(pte, ptl);
3533 unlock_page(fault_page);
3534 page_cache_release(fault_page);
3535 return ret;
3536 }
3537 do_set_pte(vma, address, fault_page, pte, false, false);
3538 unlock_page(fault_page);
3539unlock_out:
3540 pte_unmap_unlock(pte, ptl);
3541 return ret;
3542}
3543
3544static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3545 unsigned long address, pmd_t *pmd,
3546 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3547{
3548 struct page *fault_page, *new_page;
3549 spinlock_t *ptl;
3550 pte_t *pte;
3551 int ret;
3552
3553 if (unlikely(anon_vma_prepare(vma)))
3554 return VM_FAULT_OOM;
3555
3556 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3557 if (!new_page)
3558 return VM_FAULT_OOM;
3559
3560 if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)) {
3561 page_cache_release(new_page);
3562 return VM_FAULT_OOM;
3563 }
3564
3565 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3566 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3567 goto uncharge_out;
3568
3569 copy_user_highpage(new_page, fault_page, address, vma);
3570 __SetPageUptodate(new_page);
3571
3572 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3573 if (unlikely(!pte_same(*pte, orig_pte))) {
3574 pte_unmap_unlock(pte, ptl);
3575 unlock_page(fault_page);
3576 page_cache_release(fault_page);
3577 goto uncharge_out;
3578 }
3579 do_set_pte(vma, address, new_page, pte, true, true);
3580 pte_unmap_unlock(pte, ptl);
3581 unlock_page(fault_page);
3582 page_cache_release(fault_page);
3583 return ret;
3584uncharge_out:
3585 mem_cgroup_uncharge_page(new_page);
3586 page_cache_release(new_page);
3587 return ret;
3588}
3589
3590static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3591 unsigned long address, pmd_t *pmd,
3592 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3593{
3594 struct page *fault_page;
3595 struct address_space *mapping;
3596 spinlock_t *ptl;
3597 pte_t *pte;
3598 int dirtied = 0;
3599 int ret, tmp;
3600
3601 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3602 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3603 return ret;
3604
3605 /*
3606 * Check if the backing address space wants to know that the page is
3607 * about to become writable
3608 */
3609 if (vma->vm_ops->page_mkwrite) {
3610 unlock_page(fault_page);
3611 tmp = do_page_mkwrite(vma, fault_page, address);
3612 if (unlikely(!tmp ||
3613 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3614 page_cache_release(fault_page);
3615 return tmp;
3616 }
3617 }
3618
3619 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3620 if (unlikely(!pte_same(*pte, orig_pte))) {
3621 pte_unmap_unlock(pte, ptl);
3622 unlock_page(fault_page);
3623 page_cache_release(fault_page);
3624 return ret;
3625 }
3626 do_set_pte(vma, address, fault_page, pte, true, false);
3627 pte_unmap_unlock(pte, ptl);
3628
3629 if (set_page_dirty(fault_page))
3630 dirtied = 1;
3631 mapping = fault_page->mapping;
3632 unlock_page(fault_page);
3633 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3634 /*
3635 * Some device drivers do not set page.mapping but still
3636 * dirty their pages
3637 */
3638 balance_dirty_pages_ratelimited(mapping);
3639 }
3640
3641 /* file_update_time outside page_lock */
3642 if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3643 file_update_time(vma->vm_file);
3644
3645 return ret;
3646}
3647
3648static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3649 unsigned long address, pte_t *page_table, pmd_t *pmd,
3650 unsigned int flags, pte_t orig_pte)
3651{
3652 pgoff_t pgoff = (((address & PAGE_MASK)
3653 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3654
3655 pte_unmap(page_table);
3656 if (!(flags & FAULT_FLAG_WRITE))
3657 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3658 orig_pte);
3659 if (!(vma->vm_flags & VM_SHARED))
3660 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3661 orig_pte);
3662 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3663}
3664
3665/*
3666 * Fault of a previously existing named mapping. Repopulate the pte
3667 * from the encoded file_pte if possible. This enables swappable
3668 * nonlinear vmas.
3669 *
3670 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3671 * but allow concurrent faults), and pte mapped but not yet locked.
3672 * We return with mmap_sem still held, but pte unmapped and unlocked.
3673 */
3674static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3675 unsigned long address, pte_t *page_table, pmd_t *pmd,
3676 unsigned int flags, pte_t orig_pte)
3677{
3678 pgoff_t pgoff;
3679
3680 flags |= FAULT_FLAG_NONLINEAR;
3681
3682 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3683 return 0;
3684
3685 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3686 /*
3687 * Page table corrupted: show pte and kill process.
3688 */
3689 print_bad_pte(vma, address, orig_pte, NULL);
3690 return VM_FAULT_SIGBUS;
3691 }
3692
3693 pgoff = pte_to_pgoff(orig_pte);
3694 if (!(flags & FAULT_FLAG_WRITE))
3695 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3696 orig_pte);
3697 if (!(vma->vm_flags & VM_SHARED))
3698 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3699 orig_pte);
3700 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3701}
3702
3703static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3704 unsigned long addr, int page_nid,
3705 int *flags)
3706{
3707 get_page(page);
3708
3709 count_vm_numa_event(NUMA_HINT_FAULTS);
3710 if (page_nid == numa_node_id()) {
3711 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3712 *flags |= TNF_FAULT_LOCAL;
3713 }
3714
3715 return mpol_misplaced(page, vma, addr);
3716}
3717
3718static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3719 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3720{
3721 struct page *page = NULL;
3722 spinlock_t *ptl;
3723 int page_nid = -1;
3724 int last_cpupid;
3725 int target_nid;
3726 bool migrated = false;
3727 int flags = 0;
3728
3729 /*
3730 * The "pte" at this point cannot be used safely without
3731 * validation through pte_unmap_same(). It's of NUMA type but
3732 * the pfn may be screwed if the read is non atomic.
3733 *
3734 * ptep_modify_prot_start is not called as this is clearing
3735 * the _PAGE_NUMA bit and it is not really expected that there
3736 * would be concurrent hardware modifications to the PTE.
3737 */
3738 ptl = pte_lockptr(mm, pmd);
3739 spin_lock(ptl);
3740 if (unlikely(!pte_same(*ptep, pte))) {
3741 pte_unmap_unlock(ptep, ptl);
3742 goto out;
3743 }
3744
3745 pte = pte_mknonnuma(pte);
3746 set_pte_at(mm, addr, ptep, pte);
3747 update_mmu_cache(vma, addr, ptep);
3748
3749 page = vm_normal_page(vma, addr, pte);
3750 if (!page) {
3751 pte_unmap_unlock(ptep, ptl);
3752 return 0;
3753 }
3754 BUG_ON(is_zero_pfn(page_to_pfn(page)));
3755
3756 /*
3757 * Avoid grouping on DSO/COW pages in specific and RO pages
3758 * in general, RO pages shouldn't hurt as much anyway since
3759 * they can be in shared cache state.
3760 */
3761 if (!pte_write(pte))
3762 flags |= TNF_NO_GROUP;
3763
3764 /*
3765 * Flag if the page is shared between multiple address spaces. This
3766 * is later used when determining whether to group tasks together
3767 */
3768 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3769 flags |= TNF_SHARED;
3770
3771 last_cpupid = page_cpupid_last(page);
3772 page_nid = page_to_nid(page);
3773 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3774 pte_unmap_unlock(ptep, ptl);
3775 if (target_nid == -1) {
3776 put_page(page);
3777 goto out;
3778 }
3779
3780 /* Migrate to the requested node */
3781 migrated = migrate_misplaced_page(page, vma, target_nid);
3782 if (migrated) {
3783 page_nid = target_nid;
3784 flags |= TNF_MIGRATED;
3785 }
3786
3787out:
3788 if (page_nid != -1)
3789 task_numa_fault(last_cpupid, page_nid, 1, flags);
3790 return 0;
3791}
3792
3793/*
3794 * These routines also need to handle stuff like marking pages dirty
3795 * and/or accessed for architectures that don't do it in hardware (most
3796 * RISC architectures). The early dirtying is also good on the i386.
3797 *
3798 * There is also a hook called "update_mmu_cache()" that architectures
3799 * with external mmu caches can use to update those (ie the Sparc or
3800 * PowerPC hashed page tables that act as extended TLBs).
3801 *
3802 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3803 * but allow concurrent faults), and pte mapped but not yet locked.
3804 * We return with mmap_sem still held, but pte unmapped and unlocked.
3805 */
3806static int handle_pte_fault(struct mm_struct *mm,
3807 struct vm_area_struct *vma, unsigned long address,
3808 pte_t *pte, pmd_t *pmd, unsigned int flags)
3809{
3810 pte_t entry;
3811 spinlock_t *ptl;
3812
3813 entry = *pte;
3814 if (!pte_present(entry)) {
3815 if (pte_none(entry)) {
3816 if (vma->vm_ops) {
3817 if (likely(vma->vm_ops->fault))
3818 return do_linear_fault(mm, vma, address,
3819 pte, pmd, flags, entry);
3820 }
3821 return do_anonymous_page(mm, vma, address,
3822 pte, pmd, flags);
3823 }
3824 if (pte_file(entry))
3825 return do_nonlinear_fault(mm, vma, address,
3826 pte, pmd, flags, entry);
3827 return do_swap_page(mm, vma, address,
3828 pte, pmd, flags, entry);
3829 }
3830
3831 if (pte_numa(entry))
3832 return do_numa_page(mm, vma, address, entry, pte, pmd);
3833
3834 ptl = pte_lockptr(mm, pmd);
3835 spin_lock(ptl);
3836 if (unlikely(!pte_same(*pte, entry)))
3837 goto unlock;
3838 if (flags & FAULT_FLAG_WRITE) {
3839 if (!pte_write(entry))
3840 return do_wp_page(mm, vma, address,
3841 pte, pmd, ptl, entry);
3842 entry = pte_mkdirty(entry);
3843 }
3844 entry = pte_mkyoung(entry);
3845 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3846 update_mmu_cache(vma, address, pte);
3847 } else {
3848 /*
3849 * This is needed only for protection faults but the arch code
3850 * is not yet telling us if this is a protection fault or not.
3851 * This still avoids useless tlb flushes for .text page faults
3852 * with threads.
3853 */
3854 if (flags & FAULT_FLAG_WRITE)
3855 flush_tlb_fix_spurious_fault(vma, address);
3856 }
3857unlock:
3858 pte_unmap_unlock(pte, ptl);
3859 return 0;
3860}
3861
3862/*
3863 * By the time we get here, we already hold the mm semaphore
3864 */
3865static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3866 unsigned long address, unsigned int flags)
3867{
3868 pgd_t *pgd;
3869 pud_t *pud;
3870 pmd_t *pmd;
3871 pte_t *pte;
3872
3873 if (unlikely(is_vm_hugetlb_page(vma)))
3874 return hugetlb_fault(mm, vma, address, flags);
3875
3876 pgd = pgd_offset(mm, address);
3877 pud = pud_alloc(mm, pgd, address);
3878 if (!pud)
3879 return VM_FAULT_OOM;
3880 pmd = pmd_alloc(mm, pud, address);
3881 if (!pmd)
3882 return VM_FAULT_OOM;
3883 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3884 int ret = VM_FAULT_FALLBACK;
3885 if (!vma->vm_ops)
3886 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3887 pmd, flags);
3888 if (!(ret & VM_FAULT_FALLBACK))
3889 return ret;
3890 } else {
3891 pmd_t orig_pmd = *pmd;
3892 int ret;
3893
3894 barrier();
3895 if (pmd_trans_huge(orig_pmd)) {
3896 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3897
3898 /*
3899 * If the pmd is splitting, return and retry the
3900 * the fault. Alternative: wait until the split
3901 * is done, and goto retry.
3902 */
3903 if (pmd_trans_splitting(orig_pmd))
3904 return 0;
3905
3906 if (pmd_numa(orig_pmd))
3907 return do_huge_pmd_numa_page(mm, vma, address,
3908 orig_pmd, pmd);
3909
3910 if (dirty && !pmd_write(orig_pmd)) {
3911 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3912 orig_pmd);
3913 if (!(ret & VM_FAULT_FALLBACK))
3914 return ret;
3915 } else {
3916 huge_pmd_set_accessed(mm, vma, address, pmd,
3917 orig_pmd, dirty);
3918 return 0;
3919 }
3920 }
3921 }
3922
3923 /* THP should already have been handled */
3924 BUG_ON(pmd_numa(*pmd));
3925
3926 /*
3927 * Use __pte_alloc instead of pte_alloc_map, because we can't
3928 * run pte_offset_map on the pmd, if an huge pmd could
3929 * materialize from under us from a different thread.
3930 */
3931 if (unlikely(pmd_none(*pmd)) &&
3932 unlikely(__pte_alloc(mm, vma, pmd, address)))
3933 return VM_FAULT_OOM;
3934 /* if an huge pmd materialized from under us just retry later */
3935 if (unlikely(pmd_trans_huge(*pmd)))
3936 return 0;
3937 /*
3938 * A regular pmd is established and it can't morph into a huge pmd
3939 * from under us anymore at this point because we hold the mmap_sem
3940 * read mode and khugepaged takes it in write mode. So now it's
3941 * safe to run pte_offset_map().
3942 */
3943 pte = pte_offset_map(pmd, address);
3944
3945 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3946}
3947
3948int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3949 unsigned long address, unsigned int flags)
3950{
3951 int ret;
3952
3953 __set_current_state(TASK_RUNNING);
3954
3955 count_vm_event(PGFAULT);
3956 mem_cgroup_count_vm_event(mm, PGFAULT);
3957
3958 /* do counter updates before entering really critical section. */
3959 check_sync_rss_stat(current);
3960
3961 /*
3962 * Enable the memcg OOM handling for faults triggered in user
3963 * space. Kernel faults are handled more gracefully.
3964 */
3965 if (flags & FAULT_FLAG_USER)
3966 mem_cgroup_oom_enable();
3967
3968 ret = __handle_mm_fault(mm, vma, address, flags);
3969
3970 if (flags & FAULT_FLAG_USER) {
3971 mem_cgroup_oom_disable();
3972 /*
3973 * The task may have entered a memcg OOM situation but
3974 * if the allocation error was handled gracefully (no
3975 * VM_FAULT_OOM), there is no need to kill anything.
3976 * Just clean up the OOM state peacefully.
3977 */
3978 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3979 mem_cgroup_oom_synchronize(false);
3980 }
3981
3982 return ret;
3983}
3984
3985#ifndef __PAGETABLE_PUD_FOLDED
3986/*
3987 * Allocate page upper directory.
3988 * We've already handled the fast-path in-line.
3989 */
3990int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3991{
3992 pud_t *new = pud_alloc_one(mm, address);
3993 if (!new)
3994 return -ENOMEM;
3995
3996 smp_wmb(); /* See comment in __pte_alloc */
3997
3998 spin_lock(&mm->page_table_lock);
3999 if (pgd_present(*pgd)) /* Another has populated it */
4000 pud_free(mm, new);
4001 else
4002 pgd_populate(mm, pgd, new);
4003 spin_unlock(&mm->page_table_lock);
4004 return 0;
4005}
4006#endif /* __PAGETABLE_PUD_FOLDED */
4007
4008#ifndef __PAGETABLE_PMD_FOLDED
4009/*
4010 * Allocate page middle directory.
4011 * We've already handled the fast-path in-line.
4012 */
4013int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4014{
4015 pmd_t *new = pmd_alloc_one(mm, address);
4016 if (!new)
4017 return -ENOMEM;
4018
4019 smp_wmb(); /* See comment in __pte_alloc */
4020
4021 spin_lock(&mm->page_table_lock);
4022#ifndef __ARCH_HAS_4LEVEL_HACK
4023 if (pud_present(*pud)) /* Another has populated it */
4024 pmd_free(mm, new);
4025 else
4026 pud_populate(mm, pud, new);
4027#else
4028 if (pgd_present(*pud)) /* Another has populated it */
4029 pmd_free(mm, new);
4030 else
4031 pgd_populate(mm, pud, new);
4032#endif /* __ARCH_HAS_4LEVEL_HACK */
4033 spin_unlock(&mm->page_table_lock);
4034 return 0;
4035}
4036#endif /* __PAGETABLE_PMD_FOLDED */
4037
4038#if !defined(__HAVE_ARCH_GATE_AREA)
4039
4040#if defined(AT_SYSINFO_EHDR)
4041static struct vm_area_struct gate_vma;
4042
4043static int __init gate_vma_init(void)
4044{
4045 gate_vma.vm_mm = NULL;
4046 gate_vma.vm_start = FIXADDR_USER_START;
4047 gate_vma.vm_end = FIXADDR_USER_END;
4048 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
4049 gate_vma.vm_page_prot = __P101;
4050
4051 return 0;
4052}
4053__initcall(gate_vma_init);
4054#endif
4055
4056struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
4057{
4058#ifdef AT_SYSINFO_EHDR
4059 return &gate_vma;
4060#else
4061 return NULL;
4062#endif
4063}
4064
4065int in_gate_area_no_mm(unsigned long addr)
4066{
4067#ifdef AT_SYSINFO_EHDR
4068 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
4069 return 1;
4070#endif
4071 return 0;
4072}
4073
4074#endif /* __HAVE_ARCH_GATE_AREA */
4075
4076static int __follow_pte(struct mm_struct *mm, unsigned long address,
4077 pte_t **ptepp, spinlock_t **ptlp)
4078{
4079 pgd_t *pgd;
4080 pud_t *pud;
4081 pmd_t *pmd;
4082 pte_t *ptep;
4083
4084 pgd = pgd_offset(mm, address);
4085 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4086 goto out;
4087
4088 pud = pud_offset(pgd, address);
4089 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4090 goto out;
4091
4092 pmd = pmd_offset(pud, address);
4093 VM_BUG_ON(pmd_trans_huge(*pmd));
4094 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4095 goto out;
4096
4097 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
4098 if (pmd_huge(*pmd))
4099 goto out;
4100
4101 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4102 if (!ptep)
4103 goto out;
4104 if (!pte_present(*ptep))
4105 goto unlock;
4106 *ptepp = ptep;
4107 return 0;
4108unlock:
4109 pte_unmap_unlock(ptep, *ptlp);
4110out:
4111 return -EINVAL;
4112}
4113
4114static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4115 pte_t **ptepp, spinlock_t **ptlp)
4116{
4117 int res;
4118
4119 /* (void) is needed to make gcc happy */
4120 (void) __cond_lock(*ptlp,
4121 !(res = __follow_pte(mm, address, ptepp, ptlp)));
4122 return res;
4123}
4124
4125/**
4126 * follow_pfn - look up PFN at a user virtual address
4127 * @vma: memory mapping
4128 * @address: user virtual address
4129 * @pfn: location to store found PFN
4130 *
4131 * Only IO mappings and raw PFN mappings are allowed.
4132 *
4133 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4134 */
4135int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4136 unsigned long *pfn)
4137{
4138 int ret = -EINVAL;
4139 spinlock_t *ptl;
4140 pte_t *ptep;
4141
4142 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4143 return ret;
4144
4145 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4146 if (ret)
4147 return ret;
4148 *pfn = pte_pfn(*ptep);
4149 pte_unmap_unlock(ptep, ptl);
4150 return 0;
4151}
4152EXPORT_SYMBOL(follow_pfn);
4153
4154#ifdef CONFIG_HAVE_IOREMAP_PROT
4155int follow_phys(struct vm_area_struct *vma,
4156 unsigned long address, unsigned int flags,
4157 unsigned long *prot, resource_size_t *phys)
4158{
4159 int ret = -EINVAL;
4160 pte_t *ptep, pte;
4161 spinlock_t *ptl;
4162
4163 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4164 goto out;
4165
4166 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4167 goto out;
4168 pte = *ptep;
4169
4170 if ((flags & FOLL_WRITE) && !pte_write(pte))
4171 goto unlock;
4172
4173 *prot = pgprot_val(pte_pgprot(pte));
4174 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4175
4176 ret = 0;
4177unlock:
4178 pte_unmap_unlock(ptep, ptl);
4179out:
4180 return ret;
4181}
4182
4183int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4184 void *buf, int len, int write)
4185{
4186 resource_size_t phys_addr;
4187 unsigned long prot = 0;
4188 void __iomem *maddr;
4189 int offset = addr & (PAGE_SIZE-1);
4190
4191 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4192 return -EINVAL;
4193
4194 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4195 if (write)
4196 memcpy_toio(maddr + offset, buf, len);
4197 else
4198 memcpy_fromio(buf, maddr + offset, len);
4199 iounmap(maddr);
4200
4201 return len;
4202}
4203EXPORT_SYMBOL_GPL(generic_access_phys);
4204#endif
4205
4206/*
4207 * Access another process' address space as given in mm. If non-NULL, use the
4208 * given task for page fault accounting.
4209 */
4210static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4211 unsigned long addr, void *buf, int len, int write)
4212{
4213 struct vm_area_struct *vma;
4214 void *old_buf = buf;
4215
4216 down_read(&mm->mmap_sem);
4217 /* ignore errors, just check how much was successfully transferred */
4218 while (len) {
4219 int bytes, ret, offset;
4220 void *maddr;
4221 struct page *page = NULL;
4222
4223 ret = get_user_pages(tsk, mm, addr, 1,
4224 write, 1, &page, &vma);
4225 if (ret <= 0) {
4226 /*
4227 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4228 * we can access using slightly different code.
4229 */
4230#ifdef CONFIG_HAVE_IOREMAP_PROT
4231 vma = find_vma(mm, addr);
4232 if (!vma || vma->vm_start > addr)
4233 break;
4234 if (vma->vm_ops && vma->vm_ops->access)
4235 ret = vma->vm_ops->access(vma, addr, buf,
4236 len, write);
4237 if (ret <= 0)
4238#endif
4239 break;
4240 bytes = ret;
4241 } else {
4242 bytes = len;
4243 offset = addr & (PAGE_SIZE-1);
4244 if (bytes > PAGE_SIZE-offset)
4245 bytes = PAGE_SIZE-offset;
4246
4247 maddr = kmap(page);
4248 if (write) {
4249 copy_to_user_page(vma, page, addr,
4250 maddr + offset, buf, bytes);
4251 set_page_dirty_lock(page);
4252 } else {
4253 copy_from_user_page(vma, page, addr,
4254 buf, maddr + offset, bytes);
4255 }
4256 kunmap(page);
4257 page_cache_release(page);
4258 }
4259 len -= bytes;
4260 buf += bytes;
4261 addr += bytes;
4262 }
4263 up_read(&mm->mmap_sem);
4264
4265 return buf - old_buf;
4266}
4267
4268/**
4269 * access_remote_vm - access another process' address space
4270 * @mm: the mm_struct of the target address space
4271 * @addr: start address to access
4272 * @buf: source or destination buffer
4273 * @len: number of bytes to transfer
4274 * @write: whether the access is a write
4275 *
4276 * The caller must hold a reference on @mm.
4277 */
4278int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4279 void *buf, int len, int write)
4280{
4281 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4282}
4283
4284/*
4285 * Access another process' address space.
4286 * Source/target buffer must be kernel space,
4287 * Do not walk the page table directly, use get_user_pages
4288 */
4289int access_process_vm(struct task_struct *tsk, unsigned long addr,
4290 void *buf, int len, int write)
4291{
4292 struct mm_struct *mm;
4293 int ret;
4294
4295 mm = get_task_mm(tsk);
4296 if (!mm)
4297 return 0;
4298
4299 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4300 mmput(mm);
4301
4302 return ret;
4303}
4304
4305/*
4306 * Print the name of a VMA.
4307 */
4308void print_vma_addr(char *prefix, unsigned long ip)
4309{
4310 struct mm_struct *mm = current->mm;
4311 struct vm_area_struct *vma;
4312
4313 /*
4314 * Do not print if we are in atomic
4315 * contexts (in exception stacks, etc.):
4316 */
4317 if (preempt_count())
4318 return;
4319
4320 down_read(&mm->mmap_sem);
4321 vma = find_vma(mm, ip);
4322 if (vma && vma->vm_file) {
4323 struct file *f = vma->vm_file;
4324 char *buf = (char *)__get_free_page(GFP_KERNEL);
4325 if (buf) {
4326 char *p;
4327
4328 p = d_path(&f->f_path, buf, PAGE_SIZE);
4329 if (IS_ERR(p))
4330 p = "?";
4331 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4332 vma->vm_start,
4333 vma->vm_end - vma->vm_start);
4334 free_page((unsigned long)buf);
4335 }
4336 }
4337 up_read(&mm->mmap_sem);
4338}
4339
4340#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4341void might_fault(void)
4342{
4343 /*
4344 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4345 * holding the mmap_sem, this is safe because kernel memory doesn't
4346 * get paged out, therefore we'll never actually fault, and the
4347 * below annotations will generate false positives.
4348 */
4349 if (segment_eq(get_fs(), KERNEL_DS))
4350 return;
4351
4352 /*
4353 * it would be nicer only to annotate paths which are not under
4354 * pagefault_disable, however that requires a larger audit and
4355 * providing helpers like get_user_atomic.
4356 */
4357 if (in_atomic())
4358 return;
4359
4360 __might_sleep(__FILE__, __LINE__, 0);
4361
4362 if (current->mm)
4363 might_lock_read(¤t->mm->mmap_sem);
4364}
4365EXPORT_SYMBOL(might_fault);
4366#endif
4367
4368#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4369static void clear_gigantic_page(struct page *page,
4370 unsigned long addr,
4371 unsigned int pages_per_huge_page)
4372{
4373 int i;
4374 struct page *p = page;
4375
4376 might_sleep();
4377 for (i = 0; i < pages_per_huge_page;
4378 i++, p = mem_map_next(p, page, i)) {
4379 cond_resched();
4380 clear_user_highpage(p, addr + i * PAGE_SIZE);
4381 }
4382}
4383void clear_huge_page(struct page *page,
4384 unsigned long addr, unsigned int pages_per_huge_page)
4385{
4386 int i;
4387
4388 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4389 clear_gigantic_page(page, addr, pages_per_huge_page);
4390 return;
4391 }
4392
4393 might_sleep();
4394 for (i = 0; i < pages_per_huge_page; i++) {
4395 cond_resched();
4396 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4397 }
4398}
4399
4400static void copy_user_gigantic_page(struct page *dst, struct page *src,
4401 unsigned long addr,
4402 struct vm_area_struct *vma,
4403 unsigned int pages_per_huge_page)
4404{
4405 int i;
4406 struct page *dst_base = dst;
4407 struct page *src_base = src;
4408
4409 for (i = 0; i < pages_per_huge_page; ) {
4410 cond_resched();
4411 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4412
4413 i++;
4414 dst = mem_map_next(dst, dst_base, i);
4415 src = mem_map_next(src, src_base, i);
4416 }
4417}
4418
4419void copy_user_huge_page(struct page *dst, struct page *src,
4420 unsigned long addr, struct vm_area_struct *vma,
4421 unsigned int pages_per_huge_page)
4422{
4423 int i;
4424
4425 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4426 copy_user_gigantic_page(dst, src, addr, vma,
4427 pages_per_huge_page);
4428 return;
4429 }
4430
4431 might_sleep();
4432 for (i = 0; i < pages_per_huge_page; i++) {
4433 cond_resched();
4434 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4435 }
4436}
4437#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4438
4439#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4440
4441static struct kmem_cache *page_ptl_cachep;
4442
4443void __init ptlock_cache_init(void)
4444{
4445 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4446 SLAB_PANIC, NULL);
4447}
4448
4449bool ptlock_alloc(struct page *page)
4450{
4451 spinlock_t *ptl;
4452
4453 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4454 if (!ptl)
4455 return false;
4456 page->ptl = ptl;
4457 return true;
4458}
4459
4460void ptlock_free(struct page *page)
4461{
4462 kmem_cache_free(page_ptl_cachep, page->ptl);
4463}
4464#endif