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/debugfs.h>
69#include <linux/userfaultfd_k.h>
70#include <linux/dax.h>
71#include <linux/oom.h>
72#include <linux/numa.h>
73#include <linux/perf_event.h>
74#include <linux/ptrace.h>
75#include <linux/vmalloc.h>
76
77#include <trace/events/kmem.h>
78
79#include <asm/io.h>
80#include <asm/mmu_context.h>
81#include <asm/pgalloc.h>
82#include <linux/uaccess.h>
83#include <asm/tlb.h>
84#include <asm/tlbflush.h>
85
86#include "pgalloc-track.h"
87#include "internal.h"
88
89#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
90#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
91#endif
92
93#ifndef CONFIG_NUMA
94unsigned long max_mapnr;
95EXPORT_SYMBOL(max_mapnr);
96
97struct page *mem_map;
98EXPORT_SYMBOL(mem_map);
99#endif
100
101/*
102 * A number of key systems in x86 including ioremap() rely on the assumption
103 * that high_memory defines the upper bound on direct map memory, then end
104 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
105 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
106 * and ZONE_HIGHMEM.
107 */
108void *high_memory;
109EXPORT_SYMBOL(high_memory);
110
111/*
112 * Randomize the address space (stacks, mmaps, brk, etc.).
113 *
114 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
115 * as ancient (libc5 based) binaries can segfault. )
116 */
117int randomize_va_space __read_mostly =
118#ifdef CONFIG_COMPAT_BRK
119 1;
120#else
121 2;
122#endif
123
124#ifndef arch_faults_on_old_pte
125static inline bool arch_faults_on_old_pte(void)
126{
127 /*
128 * Those arches which don't have hw access flag feature need to
129 * implement their own helper. By default, "true" means pagefault
130 * will be hit on old pte.
131 */
132 return true;
133}
134#endif
135
136#ifndef arch_wants_old_prefaulted_pte
137static inline bool arch_wants_old_prefaulted_pte(void)
138{
139 /*
140 * Transitioning a PTE from 'old' to 'young' can be expensive on
141 * some architectures, even if it's performed in hardware. By
142 * default, "false" means prefaulted entries will be 'young'.
143 */
144 return false;
145}
146#endif
147
148static int __init disable_randmaps(char *s)
149{
150 randomize_va_space = 0;
151 return 1;
152}
153__setup("norandmaps", disable_randmaps);
154
155unsigned long zero_pfn __read_mostly;
156EXPORT_SYMBOL(zero_pfn);
157
158unsigned long highest_memmap_pfn __read_mostly;
159
160/*
161 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
162 */
163static int __init init_zero_pfn(void)
164{
165 zero_pfn = page_to_pfn(ZERO_PAGE(0));
166 return 0;
167}
168early_initcall(init_zero_pfn);
169
170void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
171{
172 trace_rss_stat(mm, member, count);
173}
174
175#if defined(SPLIT_RSS_COUNTING)
176
177void sync_mm_rss(struct mm_struct *mm)
178{
179 int i;
180
181 for (i = 0; i < NR_MM_COUNTERS; i++) {
182 if (current->rss_stat.count[i]) {
183 add_mm_counter(mm, i, current->rss_stat.count[i]);
184 current->rss_stat.count[i] = 0;
185 }
186 }
187 current->rss_stat.events = 0;
188}
189
190static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
191{
192 struct task_struct *task = current;
193
194 if (likely(task->mm == mm))
195 task->rss_stat.count[member] += val;
196 else
197 add_mm_counter(mm, member, val);
198}
199#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
200#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
201
202/* sync counter once per 64 page faults */
203#define TASK_RSS_EVENTS_THRESH (64)
204static void check_sync_rss_stat(struct task_struct *task)
205{
206 if (unlikely(task != current))
207 return;
208 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
209 sync_mm_rss(task->mm);
210}
211#else /* SPLIT_RSS_COUNTING */
212
213#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
214#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
215
216static void check_sync_rss_stat(struct task_struct *task)
217{
218}
219
220#endif /* SPLIT_RSS_COUNTING */
221
222/*
223 * Note: this doesn't free the actual pages themselves. That
224 * has been handled earlier when unmapping all the memory regions.
225 */
226static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
227 unsigned long addr)
228{
229 pgtable_t token = pmd_pgtable(*pmd);
230 pmd_clear(pmd);
231 pte_free_tlb(tlb, token, addr);
232 mm_dec_nr_ptes(tlb->mm);
233}
234
235static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
236 unsigned long addr, unsigned long end,
237 unsigned long floor, unsigned long ceiling)
238{
239 pmd_t *pmd;
240 unsigned long next;
241 unsigned long start;
242
243 start = addr;
244 pmd = pmd_offset(pud, addr);
245 do {
246 next = pmd_addr_end(addr, end);
247 if (pmd_none_or_clear_bad(pmd))
248 continue;
249 free_pte_range(tlb, pmd, addr);
250 } while (pmd++, addr = next, addr != end);
251
252 start &= PUD_MASK;
253 if (start < floor)
254 return;
255 if (ceiling) {
256 ceiling &= PUD_MASK;
257 if (!ceiling)
258 return;
259 }
260 if (end - 1 > ceiling - 1)
261 return;
262
263 pmd = pmd_offset(pud, start);
264 pud_clear(pud);
265 pmd_free_tlb(tlb, pmd, start);
266 mm_dec_nr_pmds(tlb->mm);
267}
268
269static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
270 unsigned long addr, unsigned long end,
271 unsigned long floor, unsigned long ceiling)
272{
273 pud_t *pud;
274 unsigned long next;
275 unsigned long start;
276
277 start = addr;
278 pud = pud_offset(p4d, addr);
279 do {
280 next = pud_addr_end(addr, end);
281 if (pud_none_or_clear_bad(pud))
282 continue;
283 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
284 } while (pud++, addr = next, addr != end);
285
286 start &= P4D_MASK;
287 if (start < floor)
288 return;
289 if (ceiling) {
290 ceiling &= P4D_MASK;
291 if (!ceiling)
292 return;
293 }
294 if (end - 1 > ceiling - 1)
295 return;
296
297 pud = pud_offset(p4d, start);
298 p4d_clear(p4d);
299 pud_free_tlb(tlb, pud, start);
300 mm_dec_nr_puds(tlb->mm);
301}
302
303static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
304 unsigned long addr, unsigned long end,
305 unsigned long floor, unsigned long ceiling)
306{
307 p4d_t *p4d;
308 unsigned long next;
309 unsigned long start;
310
311 start = addr;
312 p4d = p4d_offset(pgd, addr);
313 do {
314 next = p4d_addr_end(addr, end);
315 if (p4d_none_or_clear_bad(p4d))
316 continue;
317 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
318 } while (p4d++, addr = next, addr != end);
319
320 start &= PGDIR_MASK;
321 if (start < floor)
322 return;
323 if (ceiling) {
324 ceiling &= PGDIR_MASK;
325 if (!ceiling)
326 return;
327 }
328 if (end - 1 > ceiling - 1)
329 return;
330
331 p4d = p4d_offset(pgd, start);
332 pgd_clear(pgd);
333 p4d_free_tlb(tlb, p4d, start);
334}
335
336/*
337 * This function frees user-level page tables of a process.
338 */
339void free_pgd_range(struct mmu_gather *tlb,
340 unsigned long addr, unsigned long end,
341 unsigned long floor, unsigned long ceiling)
342{
343 pgd_t *pgd;
344 unsigned long next;
345
346 /*
347 * The next few lines have given us lots of grief...
348 *
349 * Why are we testing PMD* at this top level? Because often
350 * there will be no work to do at all, and we'd prefer not to
351 * go all the way down to the bottom just to discover that.
352 *
353 * Why all these "- 1"s? Because 0 represents both the bottom
354 * of the address space and the top of it (using -1 for the
355 * top wouldn't help much: the masks would do the wrong thing).
356 * The rule is that addr 0 and floor 0 refer to the bottom of
357 * the address space, but end 0 and ceiling 0 refer to the top
358 * Comparisons need to use "end - 1" and "ceiling - 1" (though
359 * that end 0 case should be mythical).
360 *
361 * Wherever addr is brought up or ceiling brought down, we must
362 * be careful to reject "the opposite 0" before it confuses the
363 * subsequent tests. But what about where end is brought down
364 * by PMD_SIZE below? no, end can't go down to 0 there.
365 *
366 * Whereas we round start (addr) and ceiling down, by different
367 * masks at different levels, in order to test whether a table
368 * now has no other vmas using it, so can be freed, we don't
369 * bother to round floor or end up - the tests don't need that.
370 */
371
372 addr &= PMD_MASK;
373 if (addr < floor) {
374 addr += PMD_SIZE;
375 if (!addr)
376 return;
377 }
378 if (ceiling) {
379 ceiling &= PMD_MASK;
380 if (!ceiling)
381 return;
382 }
383 if (end - 1 > ceiling - 1)
384 end -= PMD_SIZE;
385 if (addr > end - 1)
386 return;
387 /*
388 * We add page table cache pages with PAGE_SIZE,
389 * (see pte_free_tlb()), flush the tlb if we need
390 */
391 tlb_change_page_size(tlb, PAGE_SIZE);
392 pgd = pgd_offset(tlb->mm, addr);
393 do {
394 next = pgd_addr_end(addr, end);
395 if (pgd_none_or_clear_bad(pgd))
396 continue;
397 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
398 } while (pgd++, addr = next, addr != end);
399}
400
401void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
402 unsigned long floor, unsigned long ceiling)
403{
404 while (vma) {
405 struct vm_area_struct *next = vma->vm_next;
406 unsigned long addr = vma->vm_start;
407
408 /*
409 * Hide vma from rmap and truncate_pagecache before freeing
410 * pgtables
411 */
412 unlink_anon_vmas(vma);
413 unlink_file_vma(vma);
414
415 if (is_vm_hugetlb_page(vma)) {
416 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
417 floor, next ? next->vm_start : ceiling);
418 } else {
419 /*
420 * Optimization: gather nearby vmas into one call down
421 */
422 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
423 && !is_vm_hugetlb_page(next)) {
424 vma = next;
425 next = vma->vm_next;
426 unlink_anon_vmas(vma);
427 unlink_file_vma(vma);
428 }
429 free_pgd_range(tlb, addr, vma->vm_end,
430 floor, next ? next->vm_start : ceiling);
431 }
432 vma = next;
433 }
434}
435
436int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
437{
438 spinlock_t *ptl;
439 pgtable_t new = pte_alloc_one(mm);
440 if (!new)
441 return -ENOMEM;
442
443 /*
444 * Ensure all pte setup (eg. pte page lock and page clearing) are
445 * visible before the pte is made visible to other CPUs by being
446 * put into page tables.
447 *
448 * The other side of the story is the pointer chasing in the page
449 * table walking code (when walking the page table without locking;
450 * ie. most of the time). Fortunately, these data accesses consist
451 * of a chain of data-dependent loads, meaning most CPUs (alpha
452 * being the notable exception) will already guarantee loads are
453 * seen in-order. See the alpha page table accessors for the
454 * smp_rmb() barriers in page table walking code.
455 */
456 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
457
458 ptl = pmd_lock(mm, pmd);
459 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
460 mm_inc_nr_ptes(mm);
461 pmd_populate(mm, pmd, new);
462 new = NULL;
463 }
464 spin_unlock(ptl);
465 if (new)
466 pte_free(mm, new);
467 return 0;
468}
469
470int __pte_alloc_kernel(pmd_t *pmd)
471{
472 pte_t *new = pte_alloc_one_kernel(&init_mm);
473 if (!new)
474 return -ENOMEM;
475
476 smp_wmb(); /* See comment in __pte_alloc */
477
478 spin_lock(&init_mm.page_table_lock);
479 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
480 pmd_populate_kernel(&init_mm, pmd, new);
481 new = NULL;
482 }
483 spin_unlock(&init_mm.page_table_lock);
484 if (new)
485 pte_free_kernel(&init_mm, new);
486 return 0;
487}
488
489static inline void init_rss_vec(int *rss)
490{
491 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
492}
493
494static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
495{
496 int i;
497
498 if (current->mm == mm)
499 sync_mm_rss(mm);
500 for (i = 0; i < NR_MM_COUNTERS; i++)
501 if (rss[i])
502 add_mm_counter(mm, i, rss[i]);
503}
504
505/*
506 * This function is called to print an error when a bad pte
507 * is found. For example, we might have a PFN-mapped pte in
508 * a region that doesn't allow it.
509 *
510 * The calling function must still handle the error.
511 */
512static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
513 pte_t pte, struct page *page)
514{
515 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
516 p4d_t *p4d = p4d_offset(pgd, addr);
517 pud_t *pud = pud_offset(p4d, addr);
518 pmd_t *pmd = pmd_offset(pud, addr);
519 struct address_space *mapping;
520 pgoff_t index;
521 static unsigned long resume;
522 static unsigned long nr_shown;
523 static unsigned long nr_unshown;
524
525 /*
526 * Allow a burst of 60 reports, then keep quiet for that minute;
527 * or allow a steady drip of one report per second.
528 */
529 if (nr_shown == 60) {
530 if (time_before(jiffies, resume)) {
531 nr_unshown++;
532 return;
533 }
534 if (nr_unshown) {
535 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
536 nr_unshown);
537 nr_unshown = 0;
538 }
539 nr_shown = 0;
540 }
541 if (nr_shown++ == 0)
542 resume = jiffies + 60 * HZ;
543
544 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
545 index = linear_page_index(vma, addr);
546
547 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
548 current->comm,
549 (long long)pte_val(pte), (long long)pmd_val(*pmd));
550 if (page)
551 dump_page(page, "bad pte");
552 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
553 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
554 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
555 vma->vm_file,
556 vma->vm_ops ? vma->vm_ops->fault : NULL,
557 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
558 mapping ? mapping->a_ops->readpage : NULL);
559 dump_stack();
560 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
561}
562
563/*
564 * vm_normal_page -- This function gets the "struct page" associated with a pte.
565 *
566 * "Special" mappings do not wish to be associated with a "struct page" (either
567 * it doesn't exist, or it exists but they don't want to touch it). In this
568 * case, NULL is returned here. "Normal" mappings do have a struct page.
569 *
570 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
571 * pte bit, in which case this function is trivial. Secondly, an architecture
572 * may not have a spare pte bit, which requires a more complicated scheme,
573 * described below.
574 *
575 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
576 * special mapping (even if there are underlying and valid "struct pages").
577 * COWed pages of a VM_PFNMAP are always normal.
578 *
579 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
580 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
581 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
582 * mapping will always honor the rule
583 *
584 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
585 *
586 * And for normal mappings this is false.
587 *
588 * This restricts such mappings to be a linear translation from virtual address
589 * to pfn. To get around this restriction, we allow arbitrary mappings so long
590 * as the vma is not a COW mapping; in that case, we know that all ptes are
591 * special (because none can have been COWed).
592 *
593 *
594 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
595 *
596 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
597 * page" backing, however the difference is that _all_ pages with a struct
598 * page (that is, those where pfn_valid is true) are refcounted and considered
599 * normal pages by the VM. The disadvantage is that pages are refcounted
600 * (which can be slower and simply not an option for some PFNMAP users). The
601 * advantage is that we don't have to follow the strict linearity rule of
602 * PFNMAP mappings in order to support COWable mappings.
603 *
604 */
605struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
606 pte_t pte)
607{
608 unsigned long pfn = pte_pfn(pte);
609
610 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
611 if (likely(!pte_special(pte)))
612 goto check_pfn;
613 if (vma->vm_ops && vma->vm_ops->find_special_page)
614 return vma->vm_ops->find_special_page(vma, addr);
615 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
616 return NULL;
617 if (is_zero_pfn(pfn))
618 return NULL;
619 if (pte_devmap(pte))
620 return NULL;
621
622 print_bad_pte(vma, addr, pte, NULL);
623 return NULL;
624 }
625
626 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
627
628 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
629 if (vma->vm_flags & VM_MIXEDMAP) {
630 if (!pfn_valid(pfn))
631 return NULL;
632 goto out;
633 } else {
634 unsigned long off;
635 off = (addr - vma->vm_start) >> PAGE_SHIFT;
636 if (pfn == vma->vm_pgoff + off)
637 return NULL;
638 if (!is_cow_mapping(vma->vm_flags))
639 return NULL;
640 }
641 }
642
643 if (is_zero_pfn(pfn))
644 return NULL;
645
646check_pfn:
647 if (unlikely(pfn > highest_memmap_pfn)) {
648 print_bad_pte(vma, addr, pte, NULL);
649 return NULL;
650 }
651
652 /*
653 * NOTE! We still have PageReserved() pages in the page tables.
654 * eg. VDSO mappings can cause them to exist.
655 */
656out:
657 return pfn_to_page(pfn);
658}
659
660#ifdef CONFIG_TRANSPARENT_HUGEPAGE
661struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
662 pmd_t pmd)
663{
664 unsigned long pfn = pmd_pfn(pmd);
665
666 /*
667 * There is no pmd_special() but there may be special pmds, e.g.
668 * in a direct-access (dax) mapping, so let's just replicate the
669 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
670 */
671 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
672 if (vma->vm_flags & VM_MIXEDMAP) {
673 if (!pfn_valid(pfn))
674 return NULL;
675 goto out;
676 } else {
677 unsigned long off;
678 off = (addr - vma->vm_start) >> PAGE_SHIFT;
679 if (pfn == vma->vm_pgoff + off)
680 return NULL;
681 if (!is_cow_mapping(vma->vm_flags))
682 return NULL;
683 }
684 }
685
686 if (pmd_devmap(pmd))
687 return NULL;
688 if (is_huge_zero_pmd(pmd))
689 return NULL;
690 if (unlikely(pfn > highest_memmap_pfn))
691 return NULL;
692
693 /*
694 * NOTE! We still have PageReserved() pages in the page tables.
695 * eg. VDSO mappings can cause them to exist.
696 */
697out:
698 return pfn_to_page(pfn);
699}
700#endif
701
702static void restore_exclusive_pte(struct vm_area_struct *vma,
703 struct page *page, unsigned long address,
704 pte_t *ptep)
705{
706 pte_t pte;
707 swp_entry_t entry;
708
709 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
710 if (pte_swp_soft_dirty(*ptep))
711 pte = pte_mksoft_dirty(pte);
712
713 entry = pte_to_swp_entry(*ptep);
714 if (pte_swp_uffd_wp(*ptep))
715 pte = pte_mkuffd_wp(pte);
716 else if (is_writable_device_exclusive_entry(entry))
717 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
718
719 set_pte_at(vma->vm_mm, address, ptep, pte);
720
721 /*
722 * No need to take a page reference as one was already
723 * created when the swap entry was made.
724 */
725 if (PageAnon(page))
726 page_add_anon_rmap(page, vma, address, false);
727 else
728 /*
729 * Currently device exclusive access only supports anonymous
730 * memory so the entry shouldn't point to a filebacked page.
731 */
732 WARN_ON_ONCE(!PageAnon(page));
733
734 if (vma->vm_flags & VM_LOCKED)
735 mlock_vma_page(page);
736
737 /*
738 * No need to invalidate - it was non-present before. However
739 * secondary CPUs may have mappings that need invalidating.
740 */
741 update_mmu_cache(vma, address, ptep);
742}
743
744/*
745 * Tries to restore an exclusive pte if the page lock can be acquired without
746 * sleeping.
747 */
748static int
749try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
750 unsigned long addr)
751{
752 swp_entry_t entry = pte_to_swp_entry(*src_pte);
753 struct page *page = pfn_swap_entry_to_page(entry);
754
755 if (trylock_page(page)) {
756 restore_exclusive_pte(vma, page, addr, src_pte);
757 unlock_page(page);
758 return 0;
759 }
760
761 return -EBUSY;
762}
763
764/*
765 * copy one vm_area from one task to the other. Assumes the page tables
766 * already present in the new task to be cleared in the whole range
767 * covered by this vma.
768 */
769
770static unsigned long
771copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
772 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
773 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
774{
775 unsigned long vm_flags = dst_vma->vm_flags;
776 pte_t pte = *src_pte;
777 struct page *page;
778 swp_entry_t entry = pte_to_swp_entry(pte);
779
780 if (likely(!non_swap_entry(entry))) {
781 if (swap_duplicate(entry) < 0)
782 return -EIO;
783
784 /* make sure dst_mm is on swapoff's mmlist. */
785 if (unlikely(list_empty(&dst_mm->mmlist))) {
786 spin_lock(&mmlist_lock);
787 if (list_empty(&dst_mm->mmlist))
788 list_add(&dst_mm->mmlist,
789 &src_mm->mmlist);
790 spin_unlock(&mmlist_lock);
791 }
792 rss[MM_SWAPENTS]++;
793 } else if (is_migration_entry(entry)) {
794 page = pfn_swap_entry_to_page(entry);
795
796 rss[mm_counter(page)]++;
797
798 if (is_writable_migration_entry(entry) &&
799 is_cow_mapping(vm_flags)) {
800 /*
801 * COW mappings require pages in both
802 * parent and child to be set to read.
803 */
804 entry = make_readable_migration_entry(
805 swp_offset(entry));
806 pte = swp_entry_to_pte(entry);
807 if (pte_swp_soft_dirty(*src_pte))
808 pte = pte_swp_mksoft_dirty(pte);
809 if (pte_swp_uffd_wp(*src_pte))
810 pte = pte_swp_mkuffd_wp(pte);
811 set_pte_at(src_mm, addr, src_pte, pte);
812 }
813 } else if (is_device_private_entry(entry)) {
814 page = pfn_swap_entry_to_page(entry);
815
816 /*
817 * Update rss count even for unaddressable pages, as
818 * they should treated just like normal pages in this
819 * respect.
820 *
821 * We will likely want to have some new rss counters
822 * for unaddressable pages, at some point. But for now
823 * keep things as they are.
824 */
825 get_page(page);
826 rss[mm_counter(page)]++;
827 page_dup_rmap(page, false);
828
829 /*
830 * We do not preserve soft-dirty information, because so
831 * far, checkpoint/restore is the only feature that
832 * requires that. And checkpoint/restore does not work
833 * when a device driver is involved (you cannot easily
834 * save and restore device driver state).
835 */
836 if (is_writable_device_private_entry(entry) &&
837 is_cow_mapping(vm_flags)) {
838 entry = make_readable_device_private_entry(
839 swp_offset(entry));
840 pte = swp_entry_to_pte(entry);
841 if (pte_swp_uffd_wp(*src_pte))
842 pte = pte_swp_mkuffd_wp(pte);
843 set_pte_at(src_mm, addr, src_pte, pte);
844 }
845 } else if (is_device_exclusive_entry(entry)) {
846 /*
847 * Make device exclusive entries present by restoring the
848 * original entry then copying as for a present pte. Device
849 * exclusive entries currently only support private writable
850 * (ie. COW) mappings.
851 */
852 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
853 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
854 return -EBUSY;
855 return -ENOENT;
856 }
857 if (!userfaultfd_wp(dst_vma))
858 pte = pte_swp_clear_uffd_wp(pte);
859 set_pte_at(dst_mm, addr, dst_pte, pte);
860 return 0;
861}
862
863/*
864 * Copy a present and normal page if necessary.
865 *
866 * NOTE! The usual case is that this doesn't need to do
867 * anything, and can just return a positive value. That
868 * will let the caller know that it can just increase
869 * the page refcount and re-use the pte the traditional
870 * way.
871 *
872 * But _if_ we need to copy it because it needs to be
873 * pinned in the parent (and the child should get its own
874 * copy rather than just a reference to the same page),
875 * we'll do that here and return zero to let the caller
876 * know we're done.
877 *
878 * And if we need a pre-allocated page but don't yet have
879 * one, return a negative error to let the preallocation
880 * code know so that it can do so outside the page table
881 * lock.
882 */
883static inline int
884copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
885 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
886 struct page **prealloc, pte_t pte, struct page *page)
887{
888 struct page *new_page;
889
890 /*
891 * What we want to do is to check whether this page may
892 * have been pinned by the parent process. If so,
893 * instead of wrprotect the pte on both sides, we copy
894 * the page immediately so that we'll always guarantee
895 * the pinned page won't be randomly replaced in the
896 * future.
897 *
898 * The page pinning checks are just "has this mm ever
899 * seen pinning", along with the (inexact) check of
900 * the page count. That might give false positives for
901 * for pinning, but it will work correctly.
902 */
903 if (likely(!page_needs_cow_for_dma(src_vma, page)))
904 return 1;
905
906 new_page = *prealloc;
907 if (!new_page)
908 return -EAGAIN;
909
910 /*
911 * We have a prealloc page, all good! Take it
912 * over and copy the page & arm it.
913 */
914 *prealloc = NULL;
915 copy_user_highpage(new_page, page, addr, src_vma);
916 __SetPageUptodate(new_page);
917 page_add_new_anon_rmap(new_page, dst_vma, addr, false);
918 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
919 rss[mm_counter(new_page)]++;
920
921 /* All done, just insert the new page copy in the child */
922 pte = mk_pte(new_page, dst_vma->vm_page_prot);
923 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
924 if (userfaultfd_pte_wp(dst_vma, *src_pte))
925 /* Uffd-wp needs to be delivered to dest pte as well */
926 pte = pte_wrprotect(pte_mkuffd_wp(pte));
927 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
928 return 0;
929}
930
931/*
932 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
933 * is required to copy this pte.
934 */
935static inline int
936copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
937 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
938 struct page **prealloc)
939{
940 struct mm_struct *src_mm = src_vma->vm_mm;
941 unsigned long vm_flags = src_vma->vm_flags;
942 pte_t pte = *src_pte;
943 struct page *page;
944
945 page = vm_normal_page(src_vma, addr, pte);
946 if (page) {
947 int retval;
948
949 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
950 addr, rss, prealloc, pte, page);
951 if (retval <= 0)
952 return retval;
953
954 get_page(page);
955 page_dup_rmap(page, false);
956 rss[mm_counter(page)]++;
957 }
958
959 /*
960 * If it's a COW mapping, write protect it both
961 * in the parent and the child
962 */
963 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
964 ptep_set_wrprotect(src_mm, addr, src_pte);
965 pte = pte_wrprotect(pte);
966 }
967
968 /*
969 * If it's a shared mapping, mark it clean in
970 * the child
971 */
972 if (vm_flags & VM_SHARED)
973 pte = pte_mkclean(pte);
974 pte = pte_mkold(pte);
975
976 if (!userfaultfd_wp(dst_vma))
977 pte = pte_clear_uffd_wp(pte);
978
979 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
980 return 0;
981}
982
983static inline struct page *
984page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
985 unsigned long addr)
986{
987 struct page *new_page;
988
989 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
990 if (!new_page)
991 return NULL;
992
993 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
994 put_page(new_page);
995 return NULL;
996 }
997 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
998
999 return new_page;
1000}
1001
1002static int
1003copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1004 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1005 unsigned long end)
1006{
1007 struct mm_struct *dst_mm = dst_vma->vm_mm;
1008 struct mm_struct *src_mm = src_vma->vm_mm;
1009 pte_t *orig_src_pte, *orig_dst_pte;
1010 pte_t *src_pte, *dst_pte;
1011 spinlock_t *src_ptl, *dst_ptl;
1012 int progress, ret = 0;
1013 int rss[NR_MM_COUNTERS];
1014 swp_entry_t entry = (swp_entry_t){0};
1015 struct page *prealloc = NULL;
1016
1017again:
1018 progress = 0;
1019 init_rss_vec(rss);
1020
1021 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1022 if (!dst_pte) {
1023 ret = -ENOMEM;
1024 goto out;
1025 }
1026 src_pte = pte_offset_map(src_pmd, addr);
1027 src_ptl = pte_lockptr(src_mm, src_pmd);
1028 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1029 orig_src_pte = src_pte;
1030 orig_dst_pte = dst_pte;
1031 arch_enter_lazy_mmu_mode();
1032
1033 do {
1034 /*
1035 * We are holding two locks at this point - either of them
1036 * could generate latencies in another task on another CPU.
1037 */
1038 if (progress >= 32) {
1039 progress = 0;
1040 if (need_resched() ||
1041 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1042 break;
1043 }
1044 if (pte_none(*src_pte)) {
1045 progress++;
1046 continue;
1047 }
1048 if (unlikely(!pte_present(*src_pte))) {
1049 ret = copy_nonpresent_pte(dst_mm, src_mm,
1050 dst_pte, src_pte,
1051 dst_vma, src_vma,
1052 addr, rss);
1053 if (ret == -EIO) {
1054 entry = pte_to_swp_entry(*src_pte);
1055 break;
1056 } else if (ret == -EBUSY) {
1057 break;
1058 } else if (!ret) {
1059 progress += 8;
1060 continue;
1061 }
1062
1063 /*
1064 * Device exclusive entry restored, continue by copying
1065 * the now present pte.
1066 */
1067 WARN_ON_ONCE(ret != -ENOENT);
1068 }
1069 /* copy_present_pte() will clear `*prealloc' if consumed */
1070 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1071 addr, rss, &prealloc);
1072 /*
1073 * If we need a pre-allocated page for this pte, drop the
1074 * locks, allocate, and try again.
1075 */
1076 if (unlikely(ret == -EAGAIN))
1077 break;
1078 if (unlikely(prealloc)) {
1079 /*
1080 * pre-alloc page cannot be reused by next time so as
1081 * to strictly follow mempolicy (e.g., alloc_page_vma()
1082 * will allocate page according to address). This
1083 * could only happen if one pinned pte changed.
1084 */
1085 put_page(prealloc);
1086 prealloc = NULL;
1087 }
1088 progress += 8;
1089 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1090
1091 arch_leave_lazy_mmu_mode();
1092 spin_unlock(src_ptl);
1093 pte_unmap(orig_src_pte);
1094 add_mm_rss_vec(dst_mm, rss);
1095 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1096 cond_resched();
1097
1098 if (ret == -EIO) {
1099 VM_WARN_ON_ONCE(!entry.val);
1100 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1101 ret = -ENOMEM;
1102 goto out;
1103 }
1104 entry.val = 0;
1105 } else if (ret == -EBUSY) {
1106 goto out;
1107 } else if (ret == -EAGAIN) {
1108 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1109 if (!prealloc)
1110 return -ENOMEM;
1111 } else if (ret) {
1112 VM_WARN_ON_ONCE(1);
1113 }
1114
1115 /* We've captured and resolved the error. Reset, try again. */
1116 ret = 0;
1117
1118 if (addr != end)
1119 goto again;
1120out:
1121 if (unlikely(prealloc))
1122 put_page(prealloc);
1123 return ret;
1124}
1125
1126static inline int
1127copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1128 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1129 unsigned long end)
1130{
1131 struct mm_struct *dst_mm = dst_vma->vm_mm;
1132 struct mm_struct *src_mm = src_vma->vm_mm;
1133 pmd_t *src_pmd, *dst_pmd;
1134 unsigned long next;
1135
1136 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1137 if (!dst_pmd)
1138 return -ENOMEM;
1139 src_pmd = pmd_offset(src_pud, addr);
1140 do {
1141 next = pmd_addr_end(addr, end);
1142 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1143 || pmd_devmap(*src_pmd)) {
1144 int err;
1145 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1146 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1147 addr, dst_vma, src_vma);
1148 if (err == -ENOMEM)
1149 return -ENOMEM;
1150 if (!err)
1151 continue;
1152 /* fall through */
1153 }
1154 if (pmd_none_or_clear_bad(src_pmd))
1155 continue;
1156 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1157 addr, next))
1158 return -ENOMEM;
1159 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1160 return 0;
1161}
1162
1163static inline int
1164copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1165 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1166 unsigned long end)
1167{
1168 struct mm_struct *dst_mm = dst_vma->vm_mm;
1169 struct mm_struct *src_mm = src_vma->vm_mm;
1170 pud_t *src_pud, *dst_pud;
1171 unsigned long next;
1172
1173 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1174 if (!dst_pud)
1175 return -ENOMEM;
1176 src_pud = pud_offset(src_p4d, addr);
1177 do {
1178 next = pud_addr_end(addr, end);
1179 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1180 int err;
1181
1182 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1183 err = copy_huge_pud(dst_mm, src_mm,
1184 dst_pud, src_pud, addr, src_vma);
1185 if (err == -ENOMEM)
1186 return -ENOMEM;
1187 if (!err)
1188 continue;
1189 /* fall through */
1190 }
1191 if (pud_none_or_clear_bad(src_pud))
1192 continue;
1193 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1194 addr, next))
1195 return -ENOMEM;
1196 } while (dst_pud++, src_pud++, addr = next, addr != end);
1197 return 0;
1198}
1199
1200static inline int
1201copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1202 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1203 unsigned long end)
1204{
1205 struct mm_struct *dst_mm = dst_vma->vm_mm;
1206 p4d_t *src_p4d, *dst_p4d;
1207 unsigned long next;
1208
1209 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1210 if (!dst_p4d)
1211 return -ENOMEM;
1212 src_p4d = p4d_offset(src_pgd, addr);
1213 do {
1214 next = p4d_addr_end(addr, end);
1215 if (p4d_none_or_clear_bad(src_p4d))
1216 continue;
1217 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1218 addr, next))
1219 return -ENOMEM;
1220 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1221 return 0;
1222}
1223
1224int
1225copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1226{
1227 pgd_t *src_pgd, *dst_pgd;
1228 unsigned long next;
1229 unsigned long addr = src_vma->vm_start;
1230 unsigned long end = src_vma->vm_end;
1231 struct mm_struct *dst_mm = dst_vma->vm_mm;
1232 struct mm_struct *src_mm = src_vma->vm_mm;
1233 struct mmu_notifier_range range;
1234 bool is_cow;
1235 int ret;
1236
1237 /*
1238 * Don't copy ptes where a page fault will fill them correctly.
1239 * Fork becomes much lighter when there are big shared or private
1240 * readonly mappings. The tradeoff is that copy_page_range is more
1241 * efficient than faulting.
1242 */
1243 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1244 !src_vma->anon_vma)
1245 return 0;
1246
1247 if (is_vm_hugetlb_page(src_vma))
1248 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1249
1250 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1251 /*
1252 * We do not free on error cases below as remove_vma
1253 * gets called on error from higher level routine
1254 */
1255 ret = track_pfn_copy(src_vma);
1256 if (ret)
1257 return ret;
1258 }
1259
1260 /*
1261 * We need to invalidate the secondary MMU mappings only when
1262 * there could be a permission downgrade on the ptes of the
1263 * parent mm. And a permission downgrade will only happen if
1264 * is_cow_mapping() returns true.
1265 */
1266 is_cow = is_cow_mapping(src_vma->vm_flags);
1267
1268 if (is_cow) {
1269 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1270 0, src_vma, src_mm, addr, end);
1271 mmu_notifier_invalidate_range_start(&range);
1272 /*
1273 * Disabling preemption is not needed for the write side, as
1274 * the read side doesn't spin, but goes to the mmap_lock.
1275 *
1276 * Use the raw variant of the seqcount_t write API to avoid
1277 * lockdep complaining about preemptibility.
1278 */
1279 mmap_assert_write_locked(src_mm);
1280 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1281 }
1282
1283 ret = 0;
1284 dst_pgd = pgd_offset(dst_mm, addr);
1285 src_pgd = pgd_offset(src_mm, addr);
1286 do {
1287 next = pgd_addr_end(addr, end);
1288 if (pgd_none_or_clear_bad(src_pgd))
1289 continue;
1290 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1291 addr, next))) {
1292 ret = -ENOMEM;
1293 break;
1294 }
1295 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1296
1297 if (is_cow) {
1298 raw_write_seqcount_end(&src_mm->write_protect_seq);
1299 mmu_notifier_invalidate_range_end(&range);
1300 }
1301 return ret;
1302}
1303
1304static unsigned long zap_pte_range(struct mmu_gather *tlb,
1305 struct vm_area_struct *vma, pmd_t *pmd,
1306 unsigned long addr, unsigned long end,
1307 struct zap_details *details)
1308{
1309 struct mm_struct *mm = tlb->mm;
1310 int force_flush = 0;
1311 int rss[NR_MM_COUNTERS];
1312 spinlock_t *ptl;
1313 pte_t *start_pte;
1314 pte_t *pte;
1315 swp_entry_t entry;
1316
1317 tlb_change_page_size(tlb, PAGE_SIZE);
1318again:
1319 init_rss_vec(rss);
1320 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1321 pte = start_pte;
1322 flush_tlb_batched_pending(mm);
1323 arch_enter_lazy_mmu_mode();
1324 do {
1325 pte_t ptent = *pte;
1326 if (pte_none(ptent))
1327 continue;
1328
1329 if (need_resched())
1330 break;
1331
1332 if (pte_present(ptent)) {
1333 struct page *page;
1334
1335 page = vm_normal_page(vma, addr, ptent);
1336 if (unlikely(details) && page) {
1337 /*
1338 * unmap_shared_mapping_pages() wants to
1339 * invalidate cache without truncating:
1340 * unmap shared but keep private pages.
1341 */
1342 if (details->check_mapping &&
1343 details->check_mapping != page_rmapping(page))
1344 continue;
1345 }
1346 ptent = ptep_get_and_clear_full(mm, addr, pte,
1347 tlb->fullmm);
1348 tlb_remove_tlb_entry(tlb, pte, addr);
1349 if (unlikely(!page))
1350 continue;
1351
1352 if (!PageAnon(page)) {
1353 if (pte_dirty(ptent)) {
1354 force_flush = 1;
1355 set_page_dirty(page);
1356 }
1357 if (pte_young(ptent) &&
1358 likely(!(vma->vm_flags & VM_SEQ_READ)))
1359 mark_page_accessed(page);
1360 }
1361 rss[mm_counter(page)]--;
1362 page_remove_rmap(page, false);
1363 if (unlikely(page_mapcount(page) < 0))
1364 print_bad_pte(vma, addr, ptent, page);
1365 if (unlikely(__tlb_remove_page(tlb, page))) {
1366 force_flush = 1;
1367 addr += PAGE_SIZE;
1368 break;
1369 }
1370 continue;
1371 }
1372
1373 entry = pte_to_swp_entry(ptent);
1374 if (is_device_private_entry(entry) ||
1375 is_device_exclusive_entry(entry)) {
1376 struct page *page = pfn_swap_entry_to_page(entry);
1377
1378 if (unlikely(details && details->check_mapping)) {
1379 /*
1380 * unmap_shared_mapping_pages() wants to
1381 * invalidate cache without truncating:
1382 * unmap shared but keep private pages.
1383 */
1384 if (details->check_mapping !=
1385 page_rmapping(page))
1386 continue;
1387 }
1388
1389 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1390 rss[mm_counter(page)]--;
1391
1392 if (is_device_private_entry(entry))
1393 page_remove_rmap(page, false);
1394
1395 put_page(page);
1396 continue;
1397 }
1398
1399 /* If details->check_mapping, we leave swap entries. */
1400 if (unlikely(details))
1401 continue;
1402
1403 if (!non_swap_entry(entry))
1404 rss[MM_SWAPENTS]--;
1405 else if (is_migration_entry(entry)) {
1406 struct page *page;
1407
1408 page = pfn_swap_entry_to_page(entry);
1409 rss[mm_counter(page)]--;
1410 }
1411 if (unlikely(!free_swap_and_cache(entry)))
1412 print_bad_pte(vma, addr, ptent, NULL);
1413 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1414 } while (pte++, addr += PAGE_SIZE, addr != end);
1415
1416 add_mm_rss_vec(mm, rss);
1417 arch_leave_lazy_mmu_mode();
1418
1419 /* Do the actual TLB flush before dropping ptl */
1420 if (force_flush)
1421 tlb_flush_mmu_tlbonly(tlb);
1422 pte_unmap_unlock(start_pte, ptl);
1423
1424 /*
1425 * If we forced a TLB flush (either due to running out of
1426 * batch buffers or because we needed to flush dirty TLB
1427 * entries before releasing the ptl), free the batched
1428 * memory too. Restart if we didn't do everything.
1429 */
1430 if (force_flush) {
1431 force_flush = 0;
1432 tlb_flush_mmu(tlb);
1433 }
1434
1435 if (addr != end) {
1436 cond_resched();
1437 goto again;
1438 }
1439
1440 return addr;
1441}
1442
1443static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1444 struct vm_area_struct *vma, pud_t *pud,
1445 unsigned long addr, unsigned long end,
1446 struct zap_details *details)
1447{
1448 pmd_t *pmd;
1449 unsigned long next;
1450
1451 pmd = pmd_offset(pud, addr);
1452 do {
1453 next = pmd_addr_end(addr, end);
1454 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1455 if (next - addr != HPAGE_PMD_SIZE)
1456 __split_huge_pmd(vma, pmd, addr, false, NULL);
1457 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1458 goto next;
1459 /* fall through */
1460 } else if (details && details->single_page &&
1461 PageTransCompound(details->single_page) &&
1462 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1463 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1464 /*
1465 * Take and drop THP pmd lock so that we cannot return
1466 * prematurely, while zap_huge_pmd() has cleared *pmd,
1467 * but not yet decremented compound_mapcount().
1468 */
1469 spin_unlock(ptl);
1470 }
1471
1472 /*
1473 * Here there can be other concurrent MADV_DONTNEED or
1474 * trans huge page faults running, and if the pmd is
1475 * none or trans huge it can change under us. This is
1476 * because MADV_DONTNEED holds the mmap_lock in read
1477 * mode.
1478 */
1479 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1480 goto next;
1481 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1482next:
1483 cond_resched();
1484 } while (pmd++, addr = next, addr != end);
1485
1486 return addr;
1487}
1488
1489static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1490 struct vm_area_struct *vma, p4d_t *p4d,
1491 unsigned long addr, unsigned long end,
1492 struct zap_details *details)
1493{
1494 pud_t *pud;
1495 unsigned long next;
1496
1497 pud = pud_offset(p4d, addr);
1498 do {
1499 next = pud_addr_end(addr, end);
1500 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1501 if (next - addr != HPAGE_PUD_SIZE) {
1502 mmap_assert_locked(tlb->mm);
1503 split_huge_pud(vma, pud, addr);
1504 } else if (zap_huge_pud(tlb, vma, pud, addr))
1505 goto next;
1506 /* fall through */
1507 }
1508 if (pud_none_or_clear_bad(pud))
1509 continue;
1510 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1511next:
1512 cond_resched();
1513 } while (pud++, addr = next, addr != end);
1514
1515 return addr;
1516}
1517
1518static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1519 struct vm_area_struct *vma, pgd_t *pgd,
1520 unsigned long addr, unsigned long end,
1521 struct zap_details *details)
1522{
1523 p4d_t *p4d;
1524 unsigned long next;
1525
1526 p4d = p4d_offset(pgd, addr);
1527 do {
1528 next = p4d_addr_end(addr, end);
1529 if (p4d_none_or_clear_bad(p4d))
1530 continue;
1531 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1532 } while (p4d++, addr = next, addr != end);
1533
1534 return addr;
1535}
1536
1537void unmap_page_range(struct mmu_gather *tlb,
1538 struct vm_area_struct *vma,
1539 unsigned long addr, unsigned long end,
1540 struct zap_details *details)
1541{
1542 pgd_t *pgd;
1543 unsigned long next;
1544
1545 BUG_ON(addr >= end);
1546 tlb_start_vma(tlb, vma);
1547 pgd = pgd_offset(vma->vm_mm, addr);
1548 do {
1549 next = pgd_addr_end(addr, end);
1550 if (pgd_none_or_clear_bad(pgd))
1551 continue;
1552 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1553 } while (pgd++, addr = next, addr != end);
1554 tlb_end_vma(tlb, vma);
1555}
1556
1557
1558static void unmap_single_vma(struct mmu_gather *tlb,
1559 struct vm_area_struct *vma, unsigned long start_addr,
1560 unsigned long end_addr,
1561 struct zap_details *details)
1562{
1563 unsigned long start = max(vma->vm_start, start_addr);
1564 unsigned long end;
1565
1566 if (start >= vma->vm_end)
1567 return;
1568 end = min(vma->vm_end, end_addr);
1569 if (end <= vma->vm_start)
1570 return;
1571
1572 if (vma->vm_file)
1573 uprobe_munmap(vma, start, end);
1574
1575 if (unlikely(vma->vm_flags & VM_PFNMAP))
1576 untrack_pfn(vma, 0, 0);
1577
1578 if (start != end) {
1579 if (unlikely(is_vm_hugetlb_page(vma))) {
1580 /*
1581 * It is undesirable to test vma->vm_file as it
1582 * should be non-null for valid hugetlb area.
1583 * However, vm_file will be NULL in the error
1584 * cleanup path of mmap_region. When
1585 * hugetlbfs ->mmap method fails,
1586 * mmap_region() nullifies vma->vm_file
1587 * before calling this function to clean up.
1588 * Since no pte has actually been setup, it is
1589 * safe to do nothing in this case.
1590 */
1591 if (vma->vm_file) {
1592 i_mmap_lock_write(vma->vm_file->f_mapping);
1593 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1594 i_mmap_unlock_write(vma->vm_file->f_mapping);
1595 }
1596 } else
1597 unmap_page_range(tlb, vma, start, end, details);
1598 }
1599}
1600
1601/**
1602 * unmap_vmas - unmap a range of memory covered by a list of vma's
1603 * @tlb: address of the caller's struct mmu_gather
1604 * @vma: the starting vma
1605 * @start_addr: virtual address at which to start unmapping
1606 * @end_addr: virtual address at which to end unmapping
1607 *
1608 * Unmap all pages in the vma list.
1609 *
1610 * Only addresses between `start' and `end' will be unmapped.
1611 *
1612 * The VMA list must be sorted in ascending virtual address order.
1613 *
1614 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1615 * range after unmap_vmas() returns. So the only responsibility here is to
1616 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1617 * drops the lock and schedules.
1618 */
1619void unmap_vmas(struct mmu_gather *tlb,
1620 struct vm_area_struct *vma, unsigned long start_addr,
1621 unsigned long end_addr)
1622{
1623 struct mmu_notifier_range range;
1624
1625 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1626 start_addr, end_addr);
1627 mmu_notifier_invalidate_range_start(&range);
1628 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1629 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1630 mmu_notifier_invalidate_range_end(&range);
1631}
1632
1633/**
1634 * zap_page_range - remove user pages in a given range
1635 * @vma: vm_area_struct holding the applicable pages
1636 * @start: starting address of pages to zap
1637 * @size: number of bytes to zap
1638 *
1639 * Caller must protect the VMA list
1640 */
1641void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1642 unsigned long size)
1643{
1644 struct mmu_notifier_range range;
1645 struct mmu_gather tlb;
1646
1647 lru_add_drain();
1648 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1649 start, start + size);
1650 tlb_gather_mmu(&tlb, vma->vm_mm);
1651 update_hiwater_rss(vma->vm_mm);
1652 mmu_notifier_invalidate_range_start(&range);
1653 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1654 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1655 mmu_notifier_invalidate_range_end(&range);
1656 tlb_finish_mmu(&tlb);
1657}
1658
1659/**
1660 * zap_page_range_single - remove user pages in a given range
1661 * @vma: vm_area_struct holding the applicable pages
1662 * @address: starting address of pages to zap
1663 * @size: number of bytes to zap
1664 * @details: details of shared cache invalidation
1665 *
1666 * The range must fit into one VMA.
1667 */
1668static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1669 unsigned long size, struct zap_details *details)
1670{
1671 struct mmu_notifier_range range;
1672 struct mmu_gather tlb;
1673
1674 lru_add_drain();
1675 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1676 address, address + size);
1677 tlb_gather_mmu(&tlb, vma->vm_mm);
1678 update_hiwater_rss(vma->vm_mm);
1679 mmu_notifier_invalidate_range_start(&range);
1680 unmap_single_vma(&tlb, vma, address, range.end, details);
1681 mmu_notifier_invalidate_range_end(&range);
1682 tlb_finish_mmu(&tlb);
1683}
1684
1685/**
1686 * zap_vma_ptes - remove ptes mapping the vma
1687 * @vma: vm_area_struct holding ptes to be zapped
1688 * @address: starting address of pages to zap
1689 * @size: number of bytes to zap
1690 *
1691 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1692 *
1693 * The entire address range must be fully contained within the vma.
1694 *
1695 */
1696void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1697 unsigned long size)
1698{
1699 if (address < vma->vm_start || address + size > vma->vm_end ||
1700 !(vma->vm_flags & VM_PFNMAP))
1701 return;
1702
1703 zap_page_range_single(vma, address, size, NULL);
1704}
1705EXPORT_SYMBOL_GPL(zap_vma_ptes);
1706
1707static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1708{
1709 pgd_t *pgd;
1710 p4d_t *p4d;
1711 pud_t *pud;
1712 pmd_t *pmd;
1713
1714 pgd = pgd_offset(mm, addr);
1715 p4d = p4d_alloc(mm, pgd, addr);
1716 if (!p4d)
1717 return NULL;
1718 pud = pud_alloc(mm, p4d, addr);
1719 if (!pud)
1720 return NULL;
1721 pmd = pmd_alloc(mm, pud, addr);
1722 if (!pmd)
1723 return NULL;
1724
1725 VM_BUG_ON(pmd_trans_huge(*pmd));
1726 return pmd;
1727}
1728
1729pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1730 spinlock_t **ptl)
1731{
1732 pmd_t *pmd = walk_to_pmd(mm, addr);
1733
1734 if (!pmd)
1735 return NULL;
1736 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1737}
1738
1739static int validate_page_before_insert(struct page *page)
1740{
1741 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1742 return -EINVAL;
1743 flush_dcache_page(page);
1744 return 0;
1745}
1746
1747static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1748 unsigned long addr, struct page *page, pgprot_t prot)
1749{
1750 if (!pte_none(*pte))
1751 return -EBUSY;
1752 /* Ok, finally just insert the thing.. */
1753 get_page(page);
1754 inc_mm_counter_fast(mm, mm_counter_file(page));
1755 page_add_file_rmap(page, false);
1756 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1757 return 0;
1758}
1759
1760/*
1761 * This is the old fallback for page remapping.
1762 *
1763 * For historical reasons, it only allows reserved pages. Only
1764 * old drivers should use this, and they needed to mark their
1765 * pages reserved for the old functions anyway.
1766 */
1767static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1768 struct page *page, pgprot_t prot)
1769{
1770 struct mm_struct *mm = vma->vm_mm;
1771 int retval;
1772 pte_t *pte;
1773 spinlock_t *ptl;
1774
1775 retval = validate_page_before_insert(page);
1776 if (retval)
1777 goto out;
1778 retval = -ENOMEM;
1779 pte = get_locked_pte(mm, addr, &ptl);
1780 if (!pte)
1781 goto out;
1782 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1783 pte_unmap_unlock(pte, ptl);
1784out:
1785 return retval;
1786}
1787
1788#ifdef pte_index
1789static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1790 unsigned long addr, struct page *page, pgprot_t prot)
1791{
1792 int err;
1793
1794 if (!page_count(page))
1795 return -EINVAL;
1796 err = validate_page_before_insert(page);
1797 if (err)
1798 return err;
1799 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1800}
1801
1802/* insert_pages() amortizes the cost of spinlock operations
1803 * when inserting pages in a loop. Arch *must* define pte_index.
1804 */
1805static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1806 struct page **pages, unsigned long *num, pgprot_t prot)
1807{
1808 pmd_t *pmd = NULL;
1809 pte_t *start_pte, *pte;
1810 spinlock_t *pte_lock;
1811 struct mm_struct *const mm = vma->vm_mm;
1812 unsigned long curr_page_idx = 0;
1813 unsigned long remaining_pages_total = *num;
1814 unsigned long pages_to_write_in_pmd;
1815 int ret;
1816more:
1817 ret = -EFAULT;
1818 pmd = walk_to_pmd(mm, addr);
1819 if (!pmd)
1820 goto out;
1821
1822 pages_to_write_in_pmd = min_t(unsigned long,
1823 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1824
1825 /* Allocate the PTE if necessary; takes PMD lock once only. */
1826 ret = -ENOMEM;
1827 if (pte_alloc(mm, pmd))
1828 goto out;
1829
1830 while (pages_to_write_in_pmd) {
1831 int pte_idx = 0;
1832 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1833
1834 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1835 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1836 int err = insert_page_in_batch_locked(mm, pte,
1837 addr, pages[curr_page_idx], prot);
1838 if (unlikely(err)) {
1839 pte_unmap_unlock(start_pte, pte_lock);
1840 ret = err;
1841 remaining_pages_total -= pte_idx;
1842 goto out;
1843 }
1844 addr += PAGE_SIZE;
1845 ++curr_page_idx;
1846 }
1847 pte_unmap_unlock(start_pte, pte_lock);
1848 pages_to_write_in_pmd -= batch_size;
1849 remaining_pages_total -= batch_size;
1850 }
1851 if (remaining_pages_total)
1852 goto more;
1853 ret = 0;
1854out:
1855 *num = remaining_pages_total;
1856 return ret;
1857}
1858#endif /* ifdef pte_index */
1859
1860/**
1861 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1862 * @vma: user vma to map to
1863 * @addr: target start user address of these pages
1864 * @pages: source kernel pages
1865 * @num: in: number of pages to map. out: number of pages that were *not*
1866 * mapped. (0 means all pages were successfully mapped).
1867 *
1868 * Preferred over vm_insert_page() when inserting multiple pages.
1869 *
1870 * In case of error, we may have mapped a subset of the provided
1871 * pages. It is the caller's responsibility to account for this case.
1872 *
1873 * The same restrictions apply as in vm_insert_page().
1874 */
1875int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1876 struct page **pages, unsigned long *num)
1877{
1878#ifdef pte_index
1879 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1880
1881 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1882 return -EFAULT;
1883 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1884 BUG_ON(mmap_read_trylock(vma->vm_mm));
1885 BUG_ON(vma->vm_flags & VM_PFNMAP);
1886 vma->vm_flags |= VM_MIXEDMAP;
1887 }
1888 /* Defer page refcount checking till we're about to map that page. */
1889 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1890#else
1891 unsigned long idx = 0, pgcount = *num;
1892 int err = -EINVAL;
1893
1894 for (; idx < pgcount; ++idx) {
1895 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1896 if (err)
1897 break;
1898 }
1899 *num = pgcount - idx;
1900 return err;
1901#endif /* ifdef pte_index */
1902}
1903EXPORT_SYMBOL(vm_insert_pages);
1904
1905/**
1906 * vm_insert_page - insert single page into user vma
1907 * @vma: user vma to map to
1908 * @addr: target user address of this page
1909 * @page: source kernel page
1910 *
1911 * This allows drivers to insert individual pages they've allocated
1912 * into a user vma.
1913 *
1914 * The page has to be a nice clean _individual_ kernel allocation.
1915 * If you allocate a compound page, you need to have marked it as
1916 * such (__GFP_COMP), or manually just split the page up yourself
1917 * (see split_page()).
1918 *
1919 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1920 * took an arbitrary page protection parameter. This doesn't allow
1921 * that. Your vma protection will have to be set up correctly, which
1922 * means that if you want a shared writable mapping, you'd better
1923 * ask for a shared writable mapping!
1924 *
1925 * The page does not need to be reserved.
1926 *
1927 * Usually this function is called from f_op->mmap() handler
1928 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1929 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1930 * function from other places, for example from page-fault handler.
1931 *
1932 * Return: %0 on success, negative error code otherwise.
1933 */
1934int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1935 struct page *page)
1936{
1937 if (addr < vma->vm_start || addr >= vma->vm_end)
1938 return -EFAULT;
1939 if (!page_count(page))
1940 return -EINVAL;
1941 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1942 BUG_ON(mmap_read_trylock(vma->vm_mm));
1943 BUG_ON(vma->vm_flags & VM_PFNMAP);
1944 vma->vm_flags |= VM_MIXEDMAP;
1945 }
1946 return insert_page(vma, addr, page, vma->vm_page_prot);
1947}
1948EXPORT_SYMBOL(vm_insert_page);
1949
1950/*
1951 * __vm_map_pages - maps range of kernel pages into user vma
1952 * @vma: user vma to map to
1953 * @pages: pointer to array of source kernel pages
1954 * @num: number of pages in page array
1955 * @offset: user's requested vm_pgoff
1956 *
1957 * This allows drivers to map range of kernel pages into a user vma.
1958 *
1959 * Return: 0 on success and error code otherwise.
1960 */
1961static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1962 unsigned long num, unsigned long offset)
1963{
1964 unsigned long count = vma_pages(vma);
1965 unsigned long uaddr = vma->vm_start;
1966 int ret, i;
1967
1968 /* Fail if the user requested offset is beyond the end of the object */
1969 if (offset >= num)
1970 return -ENXIO;
1971
1972 /* Fail if the user requested size exceeds available object size */
1973 if (count > num - offset)
1974 return -ENXIO;
1975
1976 for (i = 0; i < count; i++) {
1977 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1978 if (ret < 0)
1979 return ret;
1980 uaddr += PAGE_SIZE;
1981 }
1982
1983 return 0;
1984}
1985
1986/**
1987 * vm_map_pages - maps range of kernel pages starts with non zero offset
1988 * @vma: user vma to map to
1989 * @pages: pointer to array of source kernel pages
1990 * @num: number of pages in page array
1991 *
1992 * Maps an object consisting of @num pages, catering for the user's
1993 * requested vm_pgoff
1994 *
1995 * If we fail to insert any page into the vma, the function will return
1996 * immediately leaving any previously inserted pages present. Callers
1997 * from the mmap handler may immediately return the error as their caller
1998 * will destroy the vma, removing any successfully inserted pages. Other
1999 * callers should make their own arrangements for calling unmap_region().
2000 *
2001 * Context: Process context. Called by mmap handlers.
2002 * Return: 0 on success and error code otherwise.
2003 */
2004int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2005 unsigned long num)
2006{
2007 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2008}
2009EXPORT_SYMBOL(vm_map_pages);
2010
2011/**
2012 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2013 * @vma: user vma to map to
2014 * @pages: pointer to array of source kernel pages
2015 * @num: number of pages in page array
2016 *
2017 * Similar to vm_map_pages(), except that it explicitly sets the offset
2018 * to 0. This function is intended for the drivers that did not consider
2019 * vm_pgoff.
2020 *
2021 * Context: Process context. Called by mmap handlers.
2022 * Return: 0 on success and error code otherwise.
2023 */
2024int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2025 unsigned long num)
2026{
2027 return __vm_map_pages(vma, pages, num, 0);
2028}
2029EXPORT_SYMBOL(vm_map_pages_zero);
2030
2031static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2032 pfn_t pfn, pgprot_t prot, bool mkwrite)
2033{
2034 struct mm_struct *mm = vma->vm_mm;
2035 pte_t *pte, entry;
2036 spinlock_t *ptl;
2037
2038 pte = get_locked_pte(mm, addr, &ptl);
2039 if (!pte)
2040 return VM_FAULT_OOM;
2041 if (!pte_none(*pte)) {
2042 if (mkwrite) {
2043 /*
2044 * For read faults on private mappings the PFN passed
2045 * in may not match the PFN we have mapped if the
2046 * mapped PFN is a writeable COW page. In the mkwrite
2047 * case we are creating a writable PTE for a shared
2048 * mapping and we expect the PFNs to match. If they
2049 * don't match, we are likely racing with block
2050 * allocation and mapping invalidation so just skip the
2051 * update.
2052 */
2053 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
2054 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
2055 goto out_unlock;
2056 }
2057 entry = pte_mkyoung(*pte);
2058 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2059 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2060 update_mmu_cache(vma, addr, pte);
2061 }
2062 goto out_unlock;
2063 }
2064
2065 /* Ok, finally just insert the thing.. */
2066 if (pfn_t_devmap(pfn))
2067 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2068 else
2069 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2070
2071 if (mkwrite) {
2072 entry = pte_mkyoung(entry);
2073 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2074 }
2075
2076 set_pte_at(mm, addr, pte, entry);
2077 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2078
2079out_unlock:
2080 pte_unmap_unlock(pte, ptl);
2081 return VM_FAULT_NOPAGE;
2082}
2083
2084/**
2085 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2086 * @vma: user vma to map to
2087 * @addr: target user address of this page
2088 * @pfn: source kernel pfn
2089 * @pgprot: pgprot flags for the inserted page
2090 *
2091 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2092 * to override pgprot on a per-page basis.
2093 *
2094 * This only makes sense for IO mappings, and it makes no sense for
2095 * COW mappings. In general, using multiple vmas is preferable;
2096 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2097 * impractical.
2098 *
2099 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2100 * a value of @pgprot different from that of @vma->vm_page_prot.
2101 *
2102 * Context: Process context. May allocate using %GFP_KERNEL.
2103 * Return: vm_fault_t value.
2104 */
2105vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2106 unsigned long pfn, pgprot_t pgprot)
2107{
2108 /*
2109 * Technically, architectures with pte_special can avoid all these
2110 * restrictions (same for remap_pfn_range). However we would like
2111 * consistency in testing and feature parity among all, so we should
2112 * try to keep these invariants in place for everybody.
2113 */
2114 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2115 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2116 (VM_PFNMAP|VM_MIXEDMAP));
2117 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2118 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2119
2120 if (addr < vma->vm_start || addr >= vma->vm_end)
2121 return VM_FAULT_SIGBUS;
2122
2123 if (!pfn_modify_allowed(pfn, pgprot))
2124 return VM_FAULT_SIGBUS;
2125
2126 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2127
2128 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2129 false);
2130}
2131EXPORT_SYMBOL(vmf_insert_pfn_prot);
2132
2133/**
2134 * vmf_insert_pfn - insert single pfn into user vma
2135 * @vma: user vma to map to
2136 * @addr: target user address of this page
2137 * @pfn: source kernel pfn
2138 *
2139 * Similar to vm_insert_page, this allows drivers to insert individual pages
2140 * they've allocated into a user vma. Same comments apply.
2141 *
2142 * This function should only be called from a vm_ops->fault handler, and
2143 * in that case the handler should return the result of this function.
2144 *
2145 * vma cannot be a COW mapping.
2146 *
2147 * As this is called only for pages that do not currently exist, we
2148 * do not need to flush old virtual caches or the TLB.
2149 *
2150 * Context: Process context. May allocate using %GFP_KERNEL.
2151 * Return: vm_fault_t value.
2152 */
2153vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2154 unsigned long pfn)
2155{
2156 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2157}
2158EXPORT_SYMBOL(vmf_insert_pfn);
2159
2160static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2161{
2162 /* these checks mirror the abort conditions in vm_normal_page */
2163 if (vma->vm_flags & VM_MIXEDMAP)
2164 return true;
2165 if (pfn_t_devmap(pfn))
2166 return true;
2167 if (pfn_t_special(pfn))
2168 return true;
2169 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2170 return true;
2171 return false;
2172}
2173
2174static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2175 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2176 bool mkwrite)
2177{
2178 int err;
2179
2180 BUG_ON(!vm_mixed_ok(vma, pfn));
2181
2182 if (addr < vma->vm_start || addr >= vma->vm_end)
2183 return VM_FAULT_SIGBUS;
2184
2185 track_pfn_insert(vma, &pgprot, pfn);
2186
2187 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2188 return VM_FAULT_SIGBUS;
2189
2190 /*
2191 * If we don't have pte special, then we have to use the pfn_valid()
2192 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2193 * refcount the page if pfn_valid is true (hence insert_page rather
2194 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2195 * without pte special, it would there be refcounted as a normal page.
2196 */
2197 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2198 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2199 struct page *page;
2200
2201 /*
2202 * At this point we are committed to insert_page()
2203 * regardless of whether the caller specified flags that
2204 * result in pfn_t_has_page() == false.
2205 */
2206 page = pfn_to_page(pfn_t_to_pfn(pfn));
2207 err = insert_page(vma, addr, page, pgprot);
2208 } else {
2209 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2210 }
2211
2212 if (err == -ENOMEM)
2213 return VM_FAULT_OOM;
2214 if (err < 0 && err != -EBUSY)
2215 return VM_FAULT_SIGBUS;
2216
2217 return VM_FAULT_NOPAGE;
2218}
2219
2220/**
2221 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2222 * @vma: user vma to map to
2223 * @addr: target user address of this page
2224 * @pfn: source kernel pfn
2225 * @pgprot: pgprot flags for the inserted page
2226 *
2227 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2228 * to override pgprot on a per-page basis.
2229 *
2230 * Typically this function should be used by drivers to set caching- and
2231 * encryption bits different than those of @vma->vm_page_prot, because
2232 * the caching- or encryption mode may not be known at mmap() time.
2233 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2234 * to set caching and encryption bits for those vmas (except for COW pages).
2235 * This is ensured by core vm only modifying these page table entries using
2236 * functions that don't touch caching- or encryption bits, using pte_modify()
2237 * if needed. (See for example mprotect()).
2238 * Also when new page-table entries are created, this is only done using the
2239 * fault() callback, and never using the value of vma->vm_page_prot,
2240 * except for page-table entries that point to anonymous pages as the result
2241 * of COW.
2242 *
2243 * Context: Process context. May allocate using %GFP_KERNEL.
2244 * Return: vm_fault_t value.
2245 */
2246vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2247 pfn_t pfn, pgprot_t pgprot)
2248{
2249 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2250}
2251EXPORT_SYMBOL(vmf_insert_mixed_prot);
2252
2253vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2254 pfn_t pfn)
2255{
2256 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2257}
2258EXPORT_SYMBOL(vmf_insert_mixed);
2259
2260/*
2261 * If the insertion of PTE failed because someone else already added a
2262 * different entry in the mean time, we treat that as success as we assume
2263 * the same entry was actually inserted.
2264 */
2265vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2266 unsigned long addr, pfn_t pfn)
2267{
2268 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2269}
2270EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2271
2272/*
2273 * maps a range of physical memory into the requested pages. the old
2274 * mappings are removed. any references to nonexistent pages results
2275 * in null mappings (currently treated as "copy-on-access")
2276 */
2277static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2278 unsigned long addr, unsigned long end,
2279 unsigned long pfn, pgprot_t prot)
2280{
2281 pte_t *pte, *mapped_pte;
2282 spinlock_t *ptl;
2283 int err = 0;
2284
2285 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2286 if (!pte)
2287 return -ENOMEM;
2288 arch_enter_lazy_mmu_mode();
2289 do {
2290 BUG_ON(!pte_none(*pte));
2291 if (!pfn_modify_allowed(pfn, prot)) {
2292 err = -EACCES;
2293 break;
2294 }
2295 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2296 pfn++;
2297 } while (pte++, addr += PAGE_SIZE, addr != end);
2298 arch_leave_lazy_mmu_mode();
2299 pte_unmap_unlock(mapped_pte, ptl);
2300 return err;
2301}
2302
2303static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2304 unsigned long addr, unsigned long end,
2305 unsigned long pfn, pgprot_t prot)
2306{
2307 pmd_t *pmd;
2308 unsigned long next;
2309 int err;
2310
2311 pfn -= addr >> PAGE_SHIFT;
2312 pmd = pmd_alloc(mm, pud, addr);
2313 if (!pmd)
2314 return -ENOMEM;
2315 VM_BUG_ON(pmd_trans_huge(*pmd));
2316 do {
2317 next = pmd_addr_end(addr, end);
2318 err = remap_pte_range(mm, pmd, addr, next,
2319 pfn + (addr >> PAGE_SHIFT), prot);
2320 if (err)
2321 return err;
2322 } while (pmd++, addr = next, addr != end);
2323 return 0;
2324}
2325
2326static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2327 unsigned long addr, unsigned long end,
2328 unsigned long pfn, pgprot_t prot)
2329{
2330 pud_t *pud;
2331 unsigned long next;
2332 int err;
2333
2334 pfn -= addr >> PAGE_SHIFT;
2335 pud = pud_alloc(mm, p4d, addr);
2336 if (!pud)
2337 return -ENOMEM;
2338 do {
2339 next = pud_addr_end(addr, end);
2340 err = remap_pmd_range(mm, pud, addr, next,
2341 pfn + (addr >> PAGE_SHIFT), prot);
2342 if (err)
2343 return err;
2344 } while (pud++, addr = next, addr != end);
2345 return 0;
2346}
2347
2348static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2349 unsigned long addr, unsigned long end,
2350 unsigned long pfn, pgprot_t prot)
2351{
2352 p4d_t *p4d;
2353 unsigned long next;
2354 int err;
2355
2356 pfn -= addr >> PAGE_SHIFT;
2357 p4d = p4d_alloc(mm, pgd, addr);
2358 if (!p4d)
2359 return -ENOMEM;
2360 do {
2361 next = p4d_addr_end(addr, end);
2362 err = remap_pud_range(mm, p4d, addr, next,
2363 pfn + (addr >> PAGE_SHIFT), prot);
2364 if (err)
2365 return err;
2366 } while (p4d++, addr = next, addr != end);
2367 return 0;
2368}
2369
2370/*
2371 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2372 * must have pre-validated the caching bits of the pgprot_t.
2373 */
2374int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2375 unsigned long pfn, unsigned long size, pgprot_t prot)
2376{
2377 pgd_t *pgd;
2378 unsigned long next;
2379 unsigned long end = addr + PAGE_ALIGN(size);
2380 struct mm_struct *mm = vma->vm_mm;
2381 int err;
2382
2383 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2384 return -EINVAL;
2385
2386 /*
2387 * Physically remapped pages are special. Tell the
2388 * rest of the world about it:
2389 * VM_IO tells people not to look at these pages
2390 * (accesses can have side effects).
2391 * VM_PFNMAP tells the core MM that the base pages are just
2392 * raw PFN mappings, and do not have a "struct page" associated
2393 * with them.
2394 * VM_DONTEXPAND
2395 * Disable vma merging and expanding with mremap().
2396 * VM_DONTDUMP
2397 * Omit vma from core dump, even when VM_IO turned off.
2398 *
2399 * There's a horrible special case to handle copy-on-write
2400 * behaviour that some programs depend on. We mark the "original"
2401 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2402 * See vm_normal_page() for details.
2403 */
2404 if (is_cow_mapping(vma->vm_flags)) {
2405 if (addr != vma->vm_start || end != vma->vm_end)
2406 return -EINVAL;
2407 vma->vm_pgoff = pfn;
2408 }
2409
2410 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2411
2412 BUG_ON(addr >= end);
2413 pfn -= addr >> PAGE_SHIFT;
2414 pgd = pgd_offset(mm, addr);
2415 flush_cache_range(vma, addr, end);
2416 do {
2417 next = pgd_addr_end(addr, end);
2418 err = remap_p4d_range(mm, pgd, addr, next,
2419 pfn + (addr >> PAGE_SHIFT), prot);
2420 if (err)
2421 return err;
2422 } while (pgd++, addr = next, addr != end);
2423
2424 return 0;
2425}
2426
2427/**
2428 * remap_pfn_range - remap kernel memory to userspace
2429 * @vma: user vma to map to
2430 * @addr: target page aligned user address to start at
2431 * @pfn: page frame number of kernel physical memory address
2432 * @size: size of mapping area
2433 * @prot: page protection flags for this mapping
2434 *
2435 * Note: this is only safe if the mm semaphore is held when called.
2436 *
2437 * Return: %0 on success, negative error code otherwise.
2438 */
2439int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2440 unsigned long pfn, unsigned long size, pgprot_t prot)
2441{
2442 int err;
2443
2444 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2445 if (err)
2446 return -EINVAL;
2447
2448 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2449 if (err)
2450 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2451 return err;
2452}
2453EXPORT_SYMBOL(remap_pfn_range);
2454
2455/**
2456 * vm_iomap_memory - remap memory to userspace
2457 * @vma: user vma to map to
2458 * @start: start of the physical memory to be mapped
2459 * @len: size of area
2460 *
2461 * This is a simplified io_remap_pfn_range() for common driver use. The
2462 * driver just needs to give us the physical memory range to be mapped,
2463 * we'll figure out the rest from the vma information.
2464 *
2465 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2466 * whatever write-combining details or similar.
2467 *
2468 * Return: %0 on success, negative error code otherwise.
2469 */
2470int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2471{
2472 unsigned long vm_len, pfn, pages;
2473
2474 /* Check that the physical memory area passed in looks valid */
2475 if (start + len < start)
2476 return -EINVAL;
2477 /*
2478 * You *really* shouldn't map things that aren't page-aligned,
2479 * but we've historically allowed it because IO memory might
2480 * just have smaller alignment.
2481 */
2482 len += start & ~PAGE_MASK;
2483 pfn = start >> PAGE_SHIFT;
2484 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2485 if (pfn + pages < pfn)
2486 return -EINVAL;
2487
2488 /* We start the mapping 'vm_pgoff' pages into the area */
2489 if (vma->vm_pgoff > pages)
2490 return -EINVAL;
2491 pfn += vma->vm_pgoff;
2492 pages -= vma->vm_pgoff;
2493
2494 /* Can we fit all of the mapping? */
2495 vm_len = vma->vm_end - vma->vm_start;
2496 if (vm_len >> PAGE_SHIFT > pages)
2497 return -EINVAL;
2498
2499 /* Ok, let it rip */
2500 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2501}
2502EXPORT_SYMBOL(vm_iomap_memory);
2503
2504static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2505 unsigned long addr, unsigned long end,
2506 pte_fn_t fn, void *data, bool create,
2507 pgtbl_mod_mask *mask)
2508{
2509 pte_t *pte, *mapped_pte;
2510 int err = 0;
2511 spinlock_t *ptl;
2512
2513 if (create) {
2514 mapped_pte = pte = (mm == &init_mm) ?
2515 pte_alloc_kernel_track(pmd, addr, mask) :
2516 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2517 if (!pte)
2518 return -ENOMEM;
2519 } else {
2520 mapped_pte = pte = (mm == &init_mm) ?
2521 pte_offset_kernel(pmd, addr) :
2522 pte_offset_map_lock(mm, pmd, addr, &ptl);
2523 }
2524
2525 BUG_ON(pmd_huge(*pmd));
2526
2527 arch_enter_lazy_mmu_mode();
2528
2529 if (fn) {
2530 do {
2531 if (create || !pte_none(*pte)) {
2532 err = fn(pte++, addr, data);
2533 if (err)
2534 break;
2535 }
2536 } while (addr += PAGE_SIZE, addr != end);
2537 }
2538 *mask |= PGTBL_PTE_MODIFIED;
2539
2540 arch_leave_lazy_mmu_mode();
2541
2542 if (mm != &init_mm)
2543 pte_unmap_unlock(mapped_pte, ptl);
2544 return err;
2545}
2546
2547static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2548 unsigned long addr, unsigned long end,
2549 pte_fn_t fn, void *data, bool create,
2550 pgtbl_mod_mask *mask)
2551{
2552 pmd_t *pmd;
2553 unsigned long next;
2554 int err = 0;
2555
2556 BUG_ON(pud_huge(*pud));
2557
2558 if (create) {
2559 pmd = pmd_alloc_track(mm, pud, addr, mask);
2560 if (!pmd)
2561 return -ENOMEM;
2562 } else {
2563 pmd = pmd_offset(pud, addr);
2564 }
2565 do {
2566 next = pmd_addr_end(addr, end);
2567 if (pmd_none(*pmd) && !create)
2568 continue;
2569 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2570 return -EINVAL;
2571 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2572 if (!create)
2573 continue;
2574 pmd_clear_bad(pmd);
2575 }
2576 err = apply_to_pte_range(mm, pmd, addr, next,
2577 fn, data, create, mask);
2578 if (err)
2579 break;
2580 } while (pmd++, addr = next, addr != end);
2581
2582 return err;
2583}
2584
2585static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2586 unsigned long addr, unsigned long end,
2587 pte_fn_t fn, void *data, bool create,
2588 pgtbl_mod_mask *mask)
2589{
2590 pud_t *pud;
2591 unsigned long next;
2592 int err = 0;
2593
2594 if (create) {
2595 pud = pud_alloc_track(mm, p4d, addr, mask);
2596 if (!pud)
2597 return -ENOMEM;
2598 } else {
2599 pud = pud_offset(p4d, addr);
2600 }
2601 do {
2602 next = pud_addr_end(addr, end);
2603 if (pud_none(*pud) && !create)
2604 continue;
2605 if (WARN_ON_ONCE(pud_leaf(*pud)))
2606 return -EINVAL;
2607 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2608 if (!create)
2609 continue;
2610 pud_clear_bad(pud);
2611 }
2612 err = apply_to_pmd_range(mm, pud, addr, next,
2613 fn, data, create, mask);
2614 if (err)
2615 break;
2616 } while (pud++, addr = next, addr != end);
2617
2618 return err;
2619}
2620
2621static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2622 unsigned long addr, unsigned long end,
2623 pte_fn_t fn, void *data, bool create,
2624 pgtbl_mod_mask *mask)
2625{
2626 p4d_t *p4d;
2627 unsigned long next;
2628 int err = 0;
2629
2630 if (create) {
2631 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2632 if (!p4d)
2633 return -ENOMEM;
2634 } else {
2635 p4d = p4d_offset(pgd, addr);
2636 }
2637 do {
2638 next = p4d_addr_end(addr, end);
2639 if (p4d_none(*p4d) && !create)
2640 continue;
2641 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2642 return -EINVAL;
2643 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2644 if (!create)
2645 continue;
2646 p4d_clear_bad(p4d);
2647 }
2648 err = apply_to_pud_range(mm, p4d, addr, next,
2649 fn, data, create, mask);
2650 if (err)
2651 break;
2652 } while (p4d++, addr = next, addr != end);
2653
2654 return err;
2655}
2656
2657static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2658 unsigned long size, pte_fn_t fn,
2659 void *data, bool create)
2660{
2661 pgd_t *pgd;
2662 unsigned long start = addr, next;
2663 unsigned long end = addr + size;
2664 pgtbl_mod_mask mask = 0;
2665 int err = 0;
2666
2667 if (WARN_ON(addr >= end))
2668 return -EINVAL;
2669
2670 pgd = pgd_offset(mm, addr);
2671 do {
2672 next = pgd_addr_end(addr, end);
2673 if (pgd_none(*pgd) && !create)
2674 continue;
2675 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2676 return -EINVAL;
2677 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2678 if (!create)
2679 continue;
2680 pgd_clear_bad(pgd);
2681 }
2682 err = apply_to_p4d_range(mm, pgd, addr, next,
2683 fn, data, create, &mask);
2684 if (err)
2685 break;
2686 } while (pgd++, addr = next, addr != end);
2687
2688 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2689 arch_sync_kernel_mappings(start, start + size);
2690
2691 return err;
2692}
2693
2694/*
2695 * Scan a region of virtual memory, filling in page tables as necessary
2696 * and calling a provided function on each leaf page table.
2697 */
2698int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2699 unsigned long size, pte_fn_t fn, void *data)
2700{
2701 return __apply_to_page_range(mm, addr, size, fn, data, true);
2702}
2703EXPORT_SYMBOL_GPL(apply_to_page_range);
2704
2705/*
2706 * Scan a region of virtual memory, calling a provided function on
2707 * each leaf page table where it exists.
2708 *
2709 * Unlike apply_to_page_range, this does _not_ fill in page tables
2710 * where they are absent.
2711 */
2712int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2713 unsigned long size, pte_fn_t fn, void *data)
2714{
2715 return __apply_to_page_range(mm, addr, size, fn, data, false);
2716}
2717EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2718
2719/*
2720 * handle_pte_fault chooses page fault handler according to an entry which was
2721 * read non-atomically. Before making any commitment, on those architectures
2722 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2723 * parts, do_swap_page must check under lock before unmapping the pte and
2724 * proceeding (but do_wp_page is only called after already making such a check;
2725 * and do_anonymous_page can safely check later on).
2726 */
2727static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2728 pte_t *page_table, pte_t orig_pte)
2729{
2730 int same = 1;
2731#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2732 if (sizeof(pte_t) > sizeof(unsigned long)) {
2733 spinlock_t *ptl = pte_lockptr(mm, pmd);
2734 spin_lock(ptl);
2735 same = pte_same(*page_table, orig_pte);
2736 spin_unlock(ptl);
2737 }
2738#endif
2739 pte_unmap(page_table);
2740 return same;
2741}
2742
2743static inline bool cow_user_page(struct page *dst, struct page *src,
2744 struct vm_fault *vmf)
2745{
2746 bool ret;
2747 void *kaddr;
2748 void __user *uaddr;
2749 bool locked = false;
2750 struct vm_area_struct *vma = vmf->vma;
2751 struct mm_struct *mm = vma->vm_mm;
2752 unsigned long addr = vmf->address;
2753
2754 if (likely(src)) {
2755 copy_user_highpage(dst, src, addr, vma);
2756 return true;
2757 }
2758
2759 /*
2760 * If the source page was a PFN mapping, we don't have
2761 * a "struct page" for it. We do a best-effort copy by
2762 * just copying from the original user address. If that
2763 * fails, we just zero-fill it. Live with it.
2764 */
2765 kaddr = kmap_atomic(dst);
2766 uaddr = (void __user *)(addr & PAGE_MASK);
2767
2768 /*
2769 * On architectures with software "accessed" bits, we would
2770 * take a double page fault, so mark it accessed here.
2771 */
2772 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2773 pte_t entry;
2774
2775 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2776 locked = true;
2777 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2778 /*
2779 * Other thread has already handled the fault
2780 * and update local tlb only
2781 */
2782 update_mmu_tlb(vma, addr, vmf->pte);
2783 ret = false;
2784 goto pte_unlock;
2785 }
2786
2787 entry = pte_mkyoung(vmf->orig_pte);
2788 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2789 update_mmu_cache(vma, addr, vmf->pte);
2790 }
2791
2792 /*
2793 * This really shouldn't fail, because the page is there
2794 * in the page tables. But it might just be unreadable,
2795 * in which case we just give up and fill the result with
2796 * zeroes.
2797 */
2798 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2799 if (locked)
2800 goto warn;
2801
2802 /* Re-validate under PTL if the page is still mapped */
2803 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2804 locked = true;
2805 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2806 /* The PTE changed under us, update local tlb */
2807 update_mmu_tlb(vma, addr, vmf->pte);
2808 ret = false;
2809 goto pte_unlock;
2810 }
2811
2812 /*
2813 * The same page can be mapped back since last copy attempt.
2814 * Try to copy again under PTL.
2815 */
2816 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2817 /*
2818 * Give a warn in case there can be some obscure
2819 * use-case
2820 */
2821warn:
2822 WARN_ON_ONCE(1);
2823 clear_page(kaddr);
2824 }
2825 }
2826
2827 ret = true;
2828
2829pte_unlock:
2830 if (locked)
2831 pte_unmap_unlock(vmf->pte, vmf->ptl);
2832 kunmap_atomic(kaddr);
2833 flush_dcache_page(dst);
2834
2835 return ret;
2836}
2837
2838static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2839{
2840 struct file *vm_file = vma->vm_file;
2841
2842 if (vm_file)
2843 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2844
2845 /*
2846 * Special mappings (e.g. VDSO) do not have any file so fake
2847 * a default GFP_KERNEL for them.
2848 */
2849 return GFP_KERNEL;
2850}
2851
2852/*
2853 * Notify the address space that the page is about to become writable so that
2854 * it can prohibit this or wait for the page to get into an appropriate state.
2855 *
2856 * We do this without the lock held, so that it can sleep if it needs to.
2857 */
2858static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2859{
2860 vm_fault_t ret;
2861 struct page *page = vmf->page;
2862 unsigned int old_flags = vmf->flags;
2863
2864 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2865
2866 if (vmf->vma->vm_file &&
2867 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2868 return VM_FAULT_SIGBUS;
2869
2870 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2871 /* Restore original flags so that caller is not surprised */
2872 vmf->flags = old_flags;
2873 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2874 return ret;
2875 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2876 lock_page(page);
2877 if (!page->mapping) {
2878 unlock_page(page);
2879 return 0; /* retry */
2880 }
2881 ret |= VM_FAULT_LOCKED;
2882 } else
2883 VM_BUG_ON_PAGE(!PageLocked(page), page);
2884 return ret;
2885}
2886
2887/*
2888 * Handle dirtying of a page in shared file mapping on a write fault.
2889 *
2890 * The function expects the page to be locked and unlocks it.
2891 */
2892static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2893{
2894 struct vm_area_struct *vma = vmf->vma;
2895 struct address_space *mapping;
2896 struct page *page = vmf->page;
2897 bool dirtied;
2898 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2899
2900 dirtied = set_page_dirty(page);
2901 VM_BUG_ON_PAGE(PageAnon(page), page);
2902 /*
2903 * Take a local copy of the address_space - page.mapping may be zeroed
2904 * by truncate after unlock_page(). The address_space itself remains
2905 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2906 * release semantics to prevent the compiler from undoing this copying.
2907 */
2908 mapping = page_rmapping(page);
2909 unlock_page(page);
2910
2911 if (!page_mkwrite)
2912 file_update_time(vma->vm_file);
2913
2914 /*
2915 * Throttle page dirtying rate down to writeback speed.
2916 *
2917 * mapping may be NULL here because some device drivers do not
2918 * set page.mapping but still dirty their pages
2919 *
2920 * Drop the mmap_lock before waiting on IO, if we can. The file
2921 * is pinning the mapping, as per above.
2922 */
2923 if ((dirtied || page_mkwrite) && mapping) {
2924 struct file *fpin;
2925
2926 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2927 balance_dirty_pages_ratelimited(mapping);
2928 if (fpin) {
2929 fput(fpin);
2930 return VM_FAULT_RETRY;
2931 }
2932 }
2933
2934 return 0;
2935}
2936
2937/*
2938 * Handle write page faults for pages that can be reused in the current vma
2939 *
2940 * This can happen either due to the mapping being with the VM_SHARED flag,
2941 * or due to us being the last reference standing to the page. In either
2942 * case, all we need to do here is to mark the page as writable and update
2943 * any related book-keeping.
2944 */
2945static inline void wp_page_reuse(struct vm_fault *vmf)
2946 __releases(vmf->ptl)
2947{
2948 struct vm_area_struct *vma = vmf->vma;
2949 struct page *page = vmf->page;
2950 pte_t entry;
2951 /*
2952 * Clear the pages cpupid information as the existing
2953 * information potentially belongs to a now completely
2954 * unrelated process.
2955 */
2956 if (page)
2957 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2958
2959 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2960 entry = pte_mkyoung(vmf->orig_pte);
2961 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2962 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2963 update_mmu_cache(vma, vmf->address, vmf->pte);
2964 pte_unmap_unlock(vmf->pte, vmf->ptl);
2965 count_vm_event(PGREUSE);
2966}
2967
2968/*
2969 * Handle the case of a page which we actually need to copy to a new page.
2970 *
2971 * Called with mmap_lock locked and the old page referenced, but
2972 * without the ptl held.
2973 *
2974 * High level logic flow:
2975 *
2976 * - Allocate a page, copy the content of the old page to the new one.
2977 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2978 * - Take the PTL. If the pte changed, bail out and release the allocated page
2979 * - If the pte is still the way we remember it, update the page table and all
2980 * relevant references. This includes dropping the reference the page-table
2981 * held to the old page, as well as updating the rmap.
2982 * - In any case, unlock the PTL and drop the reference we took to the old page.
2983 */
2984static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2985{
2986 struct vm_area_struct *vma = vmf->vma;
2987 struct mm_struct *mm = vma->vm_mm;
2988 struct page *old_page = vmf->page;
2989 struct page *new_page = NULL;
2990 pte_t entry;
2991 int page_copied = 0;
2992 struct mmu_notifier_range range;
2993
2994 if (unlikely(anon_vma_prepare(vma)))
2995 goto oom;
2996
2997 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2998 new_page = alloc_zeroed_user_highpage_movable(vma,
2999 vmf->address);
3000 if (!new_page)
3001 goto oom;
3002 } else {
3003 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3004 vmf->address);
3005 if (!new_page)
3006 goto oom;
3007
3008 if (!cow_user_page(new_page, old_page, vmf)) {
3009 /*
3010 * COW failed, if the fault was solved by other,
3011 * it's fine. If not, userspace would re-fault on
3012 * the same address and we will handle the fault
3013 * from the second attempt.
3014 */
3015 put_page(new_page);
3016 if (old_page)
3017 put_page(old_page);
3018 return 0;
3019 }
3020 }
3021
3022 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
3023 goto oom_free_new;
3024 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
3025
3026 __SetPageUptodate(new_page);
3027
3028 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
3029 vmf->address & PAGE_MASK,
3030 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3031 mmu_notifier_invalidate_range_start(&range);
3032
3033 /*
3034 * Re-check the pte - we dropped the lock
3035 */
3036 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3037 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
3038 if (old_page) {
3039 if (!PageAnon(old_page)) {
3040 dec_mm_counter_fast(mm,
3041 mm_counter_file(old_page));
3042 inc_mm_counter_fast(mm, MM_ANONPAGES);
3043 }
3044 } else {
3045 inc_mm_counter_fast(mm, MM_ANONPAGES);
3046 }
3047 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3048 entry = mk_pte(new_page, vma->vm_page_prot);
3049 entry = pte_sw_mkyoung(entry);
3050 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3051
3052 /*
3053 * Clear the pte entry and flush it first, before updating the
3054 * pte with the new entry, to keep TLBs on different CPUs in
3055 * sync. This code used to set the new PTE then flush TLBs, but
3056 * that left a window where the new PTE could be loaded into
3057 * some TLBs while the old PTE remains in others.
3058 */
3059 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
3060 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
3061 lru_cache_add_inactive_or_unevictable(new_page, vma);
3062 /*
3063 * We call the notify macro here because, when using secondary
3064 * mmu page tables (such as kvm shadow page tables), we want the
3065 * new page to be mapped directly into the secondary page table.
3066 */
3067 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3068 update_mmu_cache(vma, vmf->address, vmf->pte);
3069 if (old_page) {
3070 /*
3071 * Only after switching the pte to the new page may
3072 * we remove the mapcount here. Otherwise another
3073 * process may come and find the rmap count decremented
3074 * before the pte is switched to the new page, and
3075 * "reuse" the old page writing into it while our pte
3076 * here still points into it and can be read by other
3077 * threads.
3078 *
3079 * The critical issue is to order this
3080 * page_remove_rmap with the ptp_clear_flush above.
3081 * Those stores are ordered by (if nothing else,)
3082 * the barrier present in the atomic_add_negative
3083 * in page_remove_rmap.
3084 *
3085 * Then the TLB flush in ptep_clear_flush ensures that
3086 * no process can access the old page before the
3087 * decremented mapcount is visible. And the old page
3088 * cannot be reused until after the decremented
3089 * mapcount is visible. So transitively, TLBs to
3090 * old page will be flushed before it can be reused.
3091 */
3092 page_remove_rmap(old_page, false);
3093 }
3094
3095 /* Free the old page.. */
3096 new_page = old_page;
3097 page_copied = 1;
3098 } else {
3099 update_mmu_tlb(vma, vmf->address, vmf->pte);
3100 }
3101
3102 if (new_page)
3103 put_page(new_page);
3104
3105 pte_unmap_unlock(vmf->pte, vmf->ptl);
3106 /*
3107 * No need to double call mmu_notifier->invalidate_range() callback as
3108 * the above ptep_clear_flush_notify() did already call it.
3109 */
3110 mmu_notifier_invalidate_range_only_end(&range);
3111 if (old_page) {
3112 /*
3113 * Don't let another task, with possibly unlocked vma,
3114 * keep the mlocked page.
3115 */
3116 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
3117 lock_page(old_page); /* LRU manipulation */
3118 if (PageMlocked(old_page))
3119 munlock_vma_page(old_page);
3120 unlock_page(old_page);
3121 }
3122 if (page_copied)
3123 free_swap_cache(old_page);
3124 put_page(old_page);
3125 }
3126 return page_copied ? VM_FAULT_WRITE : 0;
3127oom_free_new:
3128 put_page(new_page);
3129oom:
3130 if (old_page)
3131 put_page(old_page);
3132 return VM_FAULT_OOM;
3133}
3134
3135/**
3136 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3137 * writeable once the page is prepared
3138 *
3139 * @vmf: structure describing the fault
3140 *
3141 * This function handles all that is needed to finish a write page fault in a
3142 * shared mapping due to PTE being read-only once the mapped page is prepared.
3143 * It handles locking of PTE and modifying it.
3144 *
3145 * The function expects the page to be locked or other protection against
3146 * concurrent faults / writeback (such as DAX radix tree locks).
3147 *
3148 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3149 * we acquired PTE lock.
3150 */
3151vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3152{
3153 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3154 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3155 &vmf->ptl);
3156 /*
3157 * We might have raced with another page fault while we released the
3158 * pte_offset_map_lock.
3159 */
3160 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3161 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3162 pte_unmap_unlock(vmf->pte, vmf->ptl);
3163 return VM_FAULT_NOPAGE;
3164 }
3165 wp_page_reuse(vmf);
3166 return 0;
3167}
3168
3169/*
3170 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3171 * mapping
3172 */
3173static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3174{
3175 struct vm_area_struct *vma = vmf->vma;
3176
3177 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3178 vm_fault_t ret;
3179
3180 pte_unmap_unlock(vmf->pte, vmf->ptl);
3181 vmf->flags |= FAULT_FLAG_MKWRITE;
3182 ret = vma->vm_ops->pfn_mkwrite(vmf);
3183 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3184 return ret;
3185 return finish_mkwrite_fault(vmf);
3186 }
3187 wp_page_reuse(vmf);
3188 return VM_FAULT_WRITE;
3189}
3190
3191static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3192 __releases(vmf->ptl)
3193{
3194 struct vm_area_struct *vma = vmf->vma;
3195 vm_fault_t ret = VM_FAULT_WRITE;
3196
3197 get_page(vmf->page);
3198
3199 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3200 vm_fault_t tmp;
3201
3202 pte_unmap_unlock(vmf->pte, vmf->ptl);
3203 tmp = do_page_mkwrite(vmf);
3204 if (unlikely(!tmp || (tmp &
3205 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3206 put_page(vmf->page);
3207 return tmp;
3208 }
3209 tmp = finish_mkwrite_fault(vmf);
3210 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3211 unlock_page(vmf->page);
3212 put_page(vmf->page);
3213 return tmp;
3214 }
3215 } else {
3216 wp_page_reuse(vmf);
3217 lock_page(vmf->page);
3218 }
3219 ret |= fault_dirty_shared_page(vmf);
3220 put_page(vmf->page);
3221
3222 return ret;
3223}
3224
3225/*
3226 * This routine handles present pages, when users try to write
3227 * to a shared page. It is done by copying the page to a new address
3228 * and decrementing the shared-page counter for the old page.
3229 *
3230 * Note that this routine assumes that the protection checks have been
3231 * done by the caller (the low-level page fault routine in most cases).
3232 * Thus we can safely just mark it writable once we've done any necessary
3233 * COW.
3234 *
3235 * We also mark the page dirty at this point even though the page will
3236 * change only once the write actually happens. This avoids a few races,
3237 * and potentially makes it more efficient.
3238 *
3239 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3240 * but allow concurrent faults), with pte both mapped and locked.
3241 * We return with mmap_lock still held, but pte unmapped and unlocked.
3242 */
3243static vm_fault_t do_wp_page(struct vm_fault *vmf)
3244 __releases(vmf->ptl)
3245{
3246 struct vm_area_struct *vma = vmf->vma;
3247
3248 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3249 pte_unmap_unlock(vmf->pte, vmf->ptl);
3250 return handle_userfault(vmf, VM_UFFD_WP);
3251 }
3252
3253 /*
3254 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3255 * is flushed in this case before copying.
3256 */
3257 if (unlikely(userfaultfd_wp(vmf->vma) &&
3258 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3259 flush_tlb_page(vmf->vma, vmf->address);
3260
3261 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3262 if (!vmf->page) {
3263 /*
3264 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3265 * VM_PFNMAP VMA.
3266 *
3267 * We should not cow pages in a shared writeable mapping.
3268 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3269 */
3270 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3271 (VM_WRITE|VM_SHARED))
3272 return wp_pfn_shared(vmf);
3273
3274 pte_unmap_unlock(vmf->pte, vmf->ptl);
3275 return wp_page_copy(vmf);
3276 }
3277
3278 /*
3279 * Take out anonymous pages first, anonymous shared vmas are
3280 * not dirty accountable.
3281 */
3282 if (PageAnon(vmf->page)) {
3283 struct page *page = vmf->page;
3284
3285 /* PageKsm() doesn't necessarily raise the page refcount */
3286 if (PageKsm(page) || page_count(page) != 1)
3287 goto copy;
3288 if (!trylock_page(page))
3289 goto copy;
3290 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3291 unlock_page(page);
3292 goto copy;
3293 }
3294 /*
3295 * Ok, we've got the only map reference, and the only
3296 * page count reference, and the page is locked,
3297 * it's dark out, and we're wearing sunglasses. Hit it.
3298 */
3299 unlock_page(page);
3300 wp_page_reuse(vmf);
3301 return VM_FAULT_WRITE;
3302 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3303 (VM_WRITE|VM_SHARED))) {
3304 return wp_page_shared(vmf);
3305 }
3306copy:
3307 /*
3308 * Ok, we need to copy. Oh, well..
3309 */
3310 get_page(vmf->page);
3311
3312 pte_unmap_unlock(vmf->pte, vmf->ptl);
3313 return wp_page_copy(vmf);
3314}
3315
3316static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3317 unsigned long start_addr, unsigned long end_addr,
3318 struct zap_details *details)
3319{
3320 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3321}
3322
3323static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3324 struct zap_details *details)
3325{
3326 struct vm_area_struct *vma;
3327 pgoff_t vba, vea, zba, zea;
3328
3329 vma_interval_tree_foreach(vma, root,
3330 details->first_index, details->last_index) {
3331
3332 vba = vma->vm_pgoff;
3333 vea = vba + vma_pages(vma) - 1;
3334 zba = details->first_index;
3335 if (zba < vba)
3336 zba = vba;
3337 zea = details->last_index;
3338 if (zea > vea)
3339 zea = vea;
3340
3341 unmap_mapping_range_vma(vma,
3342 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3343 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3344 details);
3345 }
3346}
3347
3348/**
3349 * unmap_mapping_page() - Unmap single page from processes.
3350 * @page: The locked page to be unmapped.
3351 *
3352 * Unmap this page from any userspace process which still has it mmaped.
3353 * Typically, for efficiency, the range of nearby pages has already been
3354 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3355 * truncation or invalidation holds the lock on a page, it may find that
3356 * the page has been remapped again: and then uses unmap_mapping_page()
3357 * to unmap it finally.
3358 */
3359void unmap_mapping_page(struct page *page)
3360{
3361 struct address_space *mapping = page->mapping;
3362 struct zap_details details = { };
3363
3364 VM_BUG_ON(!PageLocked(page));
3365 VM_BUG_ON(PageTail(page));
3366
3367 details.check_mapping = mapping;
3368 details.first_index = page->index;
3369 details.last_index = page->index + thp_nr_pages(page) - 1;
3370 details.single_page = page;
3371
3372 i_mmap_lock_write(mapping);
3373 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3374 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3375 i_mmap_unlock_write(mapping);
3376}
3377
3378/**
3379 * unmap_mapping_pages() - Unmap pages from processes.
3380 * @mapping: The address space containing pages to be unmapped.
3381 * @start: Index of first page to be unmapped.
3382 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3383 * @even_cows: Whether to unmap even private COWed pages.
3384 *
3385 * Unmap the pages in this address space from any userspace process which
3386 * has them mmaped. Generally, you want to remove COWed pages as well when
3387 * a file is being truncated, but not when invalidating pages from the page
3388 * cache.
3389 */
3390void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3391 pgoff_t nr, bool even_cows)
3392{
3393 struct zap_details details = { };
3394
3395 details.check_mapping = even_cows ? NULL : mapping;
3396 details.first_index = start;
3397 details.last_index = start + nr - 1;
3398 if (details.last_index < details.first_index)
3399 details.last_index = ULONG_MAX;
3400
3401 i_mmap_lock_write(mapping);
3402 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3403 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3404 i_mmap_unlock_write(mapping);
3405}
3406
3407/**
3408 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3409 * address_space corresponding to the specified byte range in the underlying
3410 * file.
3411 *
3412 * @mapping: the address space containing mmaps to be unmapped.
3413 * @holebegin: byte in first page to unmap, relative to the start of
3414 * the underlying file. This will be rounded down to a PAGE_SIZE
3415 * boundary. Note that this is different from truncate_pagecache(), which
3416 * must keep the partial page. In contrast, we must get rid of
3417 * partial pages.
3418 * @holelen: size of prospective hole in bytes. This will be rounded
3419 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3420 * end of the file.
3421 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3422 * but 0 when invalidating pagecache, don't throw away private data.
3423 */
3424void unmap_mapping_range(struct address_space *mapping,
3425 loff_t const holebegin, loff_t const holelen, int even_cows)
3426{
3427 pgoff_t hba = holebegin >> PAGE_SHIFT;
3428 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3429
3430 /* Check for overflow. */
3431 if (sizeof(holelen) > sizeof(hlen)) {
3432 long long holeend =
3433 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3434 if (holeend & ~(long long)ULONG_MAX)
3435 hlen = ULONG_MAX - hba + 1;
3436 }
3437
3438 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3439}
3440EXPORT_SYMBOL(unmap_mapping_range);
3441
3442/*
3443 * Restore a potential device exclusive pte to a working pte entry
3444 */
3445static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3446{
3447 struct page *page = vmf->page;
3448 struct vm_area_struct *vma = vmf->vma;
3449 struct mmu_notifier_range range;
3450
3451 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags))
3452 return VM_FAULT_RETRY;
3453 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
3454 vma->vm_mm, vmf->address & PAGE_MASK,
3455 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3456 mmu_notifier_invalidate_range_start(&range);
3457
3458 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3459 &vmf->ptl);
3460 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3461 restore_exclusive_pte(vma, page, vmf->address, vmf->pte);
3462
3463 pte_unmap_unlock(vmf->pte, vmf->ptl);
3464 unlock_page(page);
3465
3466 mmu_notifier_invalidate_range_end(&range);
3467 return 0;
3468}
3469
3470/*
3471 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3472 * but allow concurrent faults), and pte mapped but not yet locked.
3473 * We return with pte unmapped and unlocked.
3474 *
3475 * We return with the mmap_lock locked or unlocked in the same cases
3476 * as does filemap_fault().
3477 */
3478vm_fault_t do_swap_page(struct vm_fault *vmf)
3479{
3480 struct vm_area_struct *vma = vmf->vma;
3481 struct page *page = NULL, *swapcache;
3482 struct swap_info_struct *si = NULL;
3483 swp_entry_t entry;
3484 pte_t pte;
3485 int locked;
3486 int exclusive = 0;
3487 vm_fault_t ret = 0;
3488 void *shadow = NULL;
3489
3490 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3491 goto out;
3492
3493 entry = pte_to_swp_entry(vmf->orig_pte);
3494 if (unlikely(non_swap_entry(entry))) {
3495 if (is_migration_entry(entry)) {
3496 migration_entry_wait(vma->vm_mm, vmf->pmd,
3497 vmf->address);
3498 } else if (is_device_exclusive_entry(entry)) {
3499 vmf->page = pfn_swap_entry_to_page(entry);
3500 ret = remove_device_exclusive_entry(vmf);
3501 } else if (is_device_private_entry(entry)) {
3502 vmf->page = pfn_swap_entry_to_page(entry);
3503 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3504 } else if (is_hwpoison_entry(entry)) {
3505 ret = VM_FAULT_HWPOISON;
3506 } else {
3507 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3508 ret = VM_FAULT_SIGBUS;
3509 }
3510 goto out;
3511 }
3512
3513 /* Prevent swapoff from happening to us. */
3514 si = get_swap_device(entry);
3515 if (unlikely(!si))
3516 goto out;
3517
3518 delayacct_set_flag(current, DELAYACCT_PF_SWAPIN);
3519 page = lookup_swap_cache(entry, vma, vmf->address);
3520 swapcache = page;
3521
3522 if (!page) {
3523 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3524 __swap_count(entry) == 1) {
3525 /* skip swapcache */
3526 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3527 vmf->address);
3528 if (page) {
3529 __SetPageLocked(page);
3530 __SetPageSwapBacked(page);
3531
3532 if (mem_cgroup_swapin_charge_page(page,
3533 vma->vm_mm, GFP_KERNEL, entry)) {
3534 ret = VM_FAULT_OOM;
3535 goto out_page;
3536 }
3537 mem_cgroup_swapin_uncharge_swap(entry);
3538
3539 shadow = get_shadow_from_swap_cache(entry);
3540 if (shadow)
3541 workingset_refault(page, shadow);
3542
3543 lru_cache_add(page);
3544
3545 /* To provide entry to swap_readpage() */
3546 set_page_private(page, entry.val);
3547 swap_readpage(page, true);
3548 set_page_private(page, 0);
3549 }
3550 } else {
3551 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3552 vmf);
3553 swapcache = page;
3554 }
3555
3556 if (!page) {
3557 /*
3558 * Back out if somebody else faulted in this pte
3559 * while we released the pte lock.
3560 */
3561 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3562 vmf->address, &vmf->ptl);
3563 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3564 ret = VM_FAULT_OOM;
3565 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3566 goto unlock;
3567 }
3568
3569 /* Had to read the page from swap area: Major fault */
3570 ret = VM_FAULT_MAJOR;
3571 count_vm_event(PGMAJFAULT);
3572 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3573 } else if (PageHWPoison(page)) {
3574 /*
3575 * hwpoisoned dirty swapcache pages are kept for killing
3576 * owner processes (which may be unknown at hwpoison time)
3577 */
3578 ret = VM_FAULT_HWPOISON;
3579 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3580 goto out_release;
3581 }
3582
3583 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3584
3585 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3586 if (!locked) {
3587 ret |= VM_FAULT_RETRY;
3588 goto out_release;
3589 }
3590
3591 /*
3592 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3593 * release the swapcache from under us. The page pin, and pte_same
3594 * test below, are not enough to exclude that. Even if it is still
3595 * swapcache, we need to check that the page's swap has not changed.
3596 */
3597 if (unlikely((!PageSwapCache(page) ||
3598 page_private(page) != entry.val)) && swapcache)
3599 goto out_page;
3600
3601 page = ksm_might_need_to_copy(page, vma, vmf->address);
3602 if (unlikely(!page)) {
3603 ret = VM_FAULT_OOM;
3604 page = swapcache;
3605 goto out_page;
3606 }
3607
3608 cgroup_throttle_swaprate(page, GFP_KERNEL);
3609
3610 /*
3611 * Back out if somebody else already faulted in this pte.
3612 */
3613 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3614 &vmf->ptl);
3615 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3616 goto out_nomap;
3617
3618 if (unlikely(!PageUptodate(page))) {
3619 ret = VM_FAULT_SIGBUS;
3620 goto out_nomap;
3621 }
3622
3623 /*
3624 * The page isn't present yet, go ahead with the fault.
3625 *
3626 * Be careful about the sequence of operations here.
3627 * To get its accounting right, reuse_swap_page() must be called
3628 * while the page is counted on swap but not yet in mapcount i.e.
3629 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3630 * must be called after the swap_free(), or it will never succeed.
3631 */
3632
3633 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3634 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3635 pte = mk_pte(page, vma->vm_page_prot);
3636 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3637 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3638 vmf->flags &= ~FAULT_FLAG_WRITE;
3639 ret |= VM_FAULT_WRITE;
3640 exclusive = RMAP_EXCLUSIVE;
3641 }
3642 flush_icache_page(vma, page);
3643 if (pte_swp_soft_dirty(vmf->orig_pte))
3644 pte = pte_mksoft_dirty(pte);
3645 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3646 pte = pte_mkuffd_wp(pte);
3647 pte = pte_wrprotect(pte);
3648 }
3649 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3650 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3651 vmf->orig_pte = pte;
3652
3653 /* ksm created a completely new copy */
3654 if (unlikely(page != swapcache && swapcache)) {
3655 page_add_new_anon_rmap(page, vma, vmf->address, false);
3656 lru_cache_add_inactive_or_unevictable(page, vma);
3657 } else {
3658 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3659 }
3660
3661 swap_free(entry);
3662 if (mem_cgroup_swap_full(page) ||
3663 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3664 try_to_free_swap(page);
3665 unlock_page(page);
3666 if (page != swapcache && swapcache) {
3667 /*
3668 * Hold the lock to avoid the swap entry to be reused
3669 * until we take the PT lock for the pte_same() check
3670 * (to avoid false positives from pte_same). For
3671 * further safety release the lock after the swap_free
3672 * so that the swap count won't change under a
3673 * parallel locked swapcache.
3674 */
3675 unlock_page(swapcache);
3676 put_page(swapcache);
3677 }
3678
3679 if (vmf->flags & FAULT_FLAG_WRITE) {
3680 ret |= do_wp_page(vmf);
3681 if (ret & VM_FAULT_ERROR)
3682 ret &= VM_FAULT_ERROR;
3683 goto out;
3684 }
3685
3686 /* No need to invalidate - it was non-present before */
3687 update_mmu_cache(vma, vmf->address, vmf->pte);
3688unlock:
3689 pte_unmap_unlock(vmf->pte, vmf->ptl);
3690out:
3691 if (si)
3692 put_swap_device(si);
3693 return ret;
3694out_nomap:
3695 pte_unmap_unlock(vmf->pte, vmf->ptl);
3696out_page:
3697 unlock_page(page);
3698out_release:
3699 put_page(page);
3700 if (page != swapcache && swapcache) {
3701 unlock_page(swapcache);
3702 put_page(swapcache);
3703 }
3704 if (si)
3705 put_swap_device(si);
3706 return ret;
3707}
3708
3709/*
3710 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3711 * but allow concurrent faults), and pte mapped but not yet locked.
3712 * We return with mmap_lock still held, but pte unmapped and unlocked.
3713 */
3714static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3715{
3716 struct vm_area_struct *vma = vmf->vma;
3717 struct page *page;
3718 vm_fault_t ret = 0;
3719 pte_t entry;
3720
3721 /* File mapping without ->vm_ops ? */
3722 if (vma->vm_flags & VM_SHARED)
3723 return VM_FAULT_SIGBUS;
3724
3725 /*
3726 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3727 * pte_offset_map() on pmds where a huge pmd might be created
3728 * from a different thread.
3729 *
3730 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3731 * parallel threads are excluded by other means.
3732 *
3733 * Here we only have mmap_read_lock(mm).
3734 */
3735 if (pte_alloc(vma->vm_mm, vmf->pmd))
3736 return VM_FAULT_OOM;
3737
3738 /* See comment in handle_pte_fault() */
3739 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3740 return 0;
3741
3742 /* Use the zero-page for reads */
3743 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3744 !mm_forbids_zeropage(vma->vm_mm)) {
3745 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3746 vma->vm_page_prot));
3747 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3748 vmf->address, &vmf->ptl);
3749 if (!pte_none(*vmf->pte)) {
3750 update_mmu_tlb(vma, vmf->address, vmf->pte);
3751 goto unlock;
3752 }
3753 ret = check_stable_address_space(vma->vm_mm);
3754 if (ret)
3755 goto unlock;
3756 /* Deliver the page fault to userland, check inside PT lock */
3757 if (userfaultfd_missing(vma)) {
3758 pte_unmap_unlock(vmf->pte, vmf->ptl);
3759 return handle_userfault(vmf, VM_UFFD_MISSING);
3760 }
3761 goto setpte;
3762 }
3763
3764 /* Allocate our own private page. */
3765 if (unlikely(anon_vma_prepare(vma)))
3766 goto oom;
3767 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3768 if (!page)
3769 goto oom;
3770
3771 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3772 goto oom_free_page;
3773 cgroup_throttle_swaprate(page, GFP_KERNEL);
3774
3775 /*
3776 * The memory barrier inside __SetPageUptodate makes sure that
3777 * preceding stores to the page contents become visible before
3778 * the set_pte_at() write.
3779 */
3780 __SetPageUptodate(page);
3781
3782 entry = mk_pte(page, vma->vm_page_prot);
3783 entry = pte_sw_mkyoung(entry);
3784 if (vma->vm_flags & VM_WRITE)
3785 entry = pte_mkwrite(pte_mkdirty(entry));
3786
3787 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3788 &vmf->ptl);
3789 if (!pte_none(*vmf->pte)) {
3790 update_mmu_cache(vma, vmf->address, vmf->pte);
3791 goto release;
3792 }
3793
3794 ret = check_stable_address_space(vma->vm_mm);
3795 if (ret)
3796 goto release;
3797
3798 /* Deliver the page fault to userland, check inside PT lock */
3799 if (userfaultfd_missing(vma)) {
3800 pte_unmap_unlock(vmf->pte, vmf->ptl);
3801 put_page(page);
3802 return handle_userfault(vmf, VM_UFFD_MISSING);
3803 }
3804
3805 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3806 page_add_new_anon_rmap(page, vma, vmf->address, false);
3807 lru_cache_add_inactive_or_unevictable(page, vma);
3808setpte:
3809 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3810
3811 /* No need to invalidate - it was non-present before */
3812 update_mmu_cache(vma, vmf->address, vmf->pte);
3813unlock:
3814 pte_unmap_unlock(vmf->pte, vmf->ptl);
3815 return ret;
3816release:
3817 put_page(page);
3818 goto unlock;
3819oom_free_page:
3820 put_page(page);
3821oom:
3822 return VM_FAULT_OOM;
3823}
3824
3825/*
3826 * The mmap_lock must have been held on entry, and may have been
3827 * released depending on flags and vma->vm_ops->fault() return value.
3828 * See filemap_fault() and __lock_page_retry().
3829 */
3830static vm_fault_t __do_fault(struct vm_fault *vmf)
3831{
3832 struct vm_area_struct *vma = vmf->vma;
3833 vm_fault_t ret;
3834
3835 /*
3836 * Preallocate pte before we take page_lock because this might lead to
3837 * deadlocks for memcg reclaim which waits for pages under writeback:
3838 * lock_page(A)
3839 * SetPageWriteback(A)
3840 * unlock_page(A)
3841 * lock_page(B)
3842 * lock_page(B)
3843 * pte_alloc_one
3844 * shrink_page_list
3845 * wait_on_page_writeback(A)
3846 * SetPageWriteback(B)
3847 * unlock_page(B)
3848 * # flush A, B to clear the writeback
3849 */
3850 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3851 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3852 if (!vmf->prealloc_pte)
3853 return VM_FAULT_OOM;
3854 smp_wmb(); /* See comment in __pte_alloc() */
3855 }
3856
3857 ret = vma->vm_ops->fault(vmf);
3858 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3859 VM_FAULT_DONE_COW)))
3860 return ret;
3861
3862 if (unlikely(PageHWPoison(vmf->page))) {
3863 if (ret & VM_FAULT_LOCKED)
3864 unlock_page(vmf->page);
3865 put_page(vmf->page);
3866 vmf->page = NULL;
3867 return VM_FAULT_HWPOISON;
3868 }
3869
3870 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3871 lock_page(vmf->page);
3872 else
3873 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3874
3875 return ret;
3876}
3877
3878#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3879static void deposit_prealloc_pte(struct vm_fault *vmf)
3880{
3881 struct vm_area_struct *vma = vmf->vma;
3882
3883 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3884 /*
3885 * We are going to consume the prealloc table,
3886 * count that as nr_ptes.
3887 */
3888 mm_inc_nr_ptes(vma->vm_mm);
3889 vmf->prealloc_pte = NULL;
3890}
3891
3892vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3893{
3894 struct vm_area_struct *vma = vmf->vma;
3895 bool write = vmf->flags & FAULT_FLAG_WRITE;
3896 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3897 pmd_t entry;
3898 int i;
3899 vm_fault_t ret = VM_FAULT_FALLBACK;
3900
3901 if (!transhuge_vma_suitable(vma, haddr))
3902 return ret;
3903
3904 page = compound_head(page);
3905 if (compound_order(page) != HPAGE_PMD_ORDER)
3906 return ret;
3907
3908 /*
3909 * Archs like ppc64 need additional space to store information
3910 * related to pte entry. Use the preallocated table for that.
3911 */
3912 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3913 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3914 if (!vmf->prealloc_pte)
3915 return VM_FAULT_OOM;
3916 smp_wmb(); /* See comment in __pte_alloc() */
3917 }
3918
3919 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3920 if (unlikely(!pmd_none(*vmf->pmd)))
3921 goto out;
3922
3923 for (i = 0; i < HPAGE_PMD_NR; i++)
3924 flush_icache_page(vma, page + i);
3925
3926 entry = mk_huge_pmd(page, vma->vm_page_prot);
3927 if (write)
3928 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3929
3930 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3931 page_add_file_rmap(page, true);
3932 /*
3933 * deposit and withdraw with pmd lock held
3934 */
3935 if (arch_needs_pgtable_deposit())
3936 deposit_prealloc_pte(vmf);
3937
3938 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3939
3940 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3941
3942 /* fault is handled */
3943 ret = 0;
3944 count_vm_event(THP_FILE_MAPPED);
3945out:
3946 spin_unlock(vmf->ptl);
3947 return ret;
3948}
3949#else
3950vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3951{
3952 return VM_FAULT_FALLBACK;
3953}
3954#endif
3955
3956void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
3957{
3958 struct vm_area_struct *vma = vmf->vma;
3959 bool write = vmf->flags & FAULT_FLAG_WRITE;
3960 bool prefault = vmf->address != addr;
3961 pte_t entry;
3962
3963 flush_icache_page(vma, page);
3964 entry = mk_pte(page, vma->vm_page_prot);
3965
3966 if (prefault && arch_wants_old_prefaulted_pte())
3967 entry = pte_mkold(entry);
3968 else
3969 entry = pte_sw_mkyoung(entry);
3970
3971 if (write)
3972 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3973 /* copy-on-write page */
3974 if (write && !(vma->vm_flags & VM_SHARED)) {
3975 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3976 page_add_new_anon_rmap(page, vma, addr, false);
3977 lru_cache_add_inactive_or_unevictable(page, vma);
3978 } else {
3979 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3980 page_add_file_rmap(page, false);
3981 }
3982 set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
3983}
3984
3985/**
3986 * finish_fault - finish page fault once we have prepared the page to fault
3987 *
3988 * @vmf: structure describing the fault
3989 *
3990 * This function handles all that is needed to finish a page fault once the
3991 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3992 * given page, adds reverse page mapping, handles memcg charges and LRU
3993 * addition.
3994 *
3995 * The function expects the page to be locked and on success it consumes a
3996 * reference of a page being mapped (for the PTE which maps it).
3997 *
3998 * Return: %0 on success, %VM_FAULT_ code in case of error.
3999 */
4000vm_fault_t finish_fault(struct vm_fault *vmf)
4001{
4002 struct vm_area_struct *vma = vmf->vma;
4003 struct page *page;
4004 vm_fault_t ret;
4005
4006 /* Did we COW the page? */
4007 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4008 page = vmf->cow_page;
4009 else
4010 page = vmf->page;
4011
4012 /*
4013 * check even for read faults because we might have lost our CoWed
4014 * page
4015 */
4016 if (!(vma->vm_flags & VM_SHARED)) {
4017 ret = check_stable_address_space(vma->vm_mm);
4018 if (ret)
4019 return ret;
4020 }
4021
4022 if (pmd_none(*vmf->pmd)) {
4023 if (PageTransCompound(page)) {
4024 ret = do_set_pmd(vmf, page);
4025 if (ret != VM_FAULT_FALLBACK)
4026 return ret;
4027 }
4028
4029 if (vmf->prealloc_pte) {
4030 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4031 if (likely(pmd_none(*vmf->pmd))) {
4032 mm_inc_nr_ptes(vma->vm_mm);
4033 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4034 vmf->prealloc_pte = NULL;
4035 }
4036 spin_unlock(vmf->ptl);
4037 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
4038 return VM_FAULT_OOM;
4039 }
4040 }
4041
4042 /* See comment in handle_pte_fault() */
4043 if (pmd_devmap_trans_unstable(vmf->pmd))
4044 return 0;
4045
4046 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4047 vmf->address, &vmf->ptl);
4048 ret = 0;
4049 /* Re-check under ptl */
4050 if (likely(pte_none(*vmf->pte)))
4051 do_set_pte(vmf, page, vmf->address);
4052 else
4053 ret = VM_FAULT_NOPAGE;
4054
4055 update_mmu_tlb(vma, vmf->address, vmf->pte);
4056 pte_unmap_unlock(vmf->pte, vmf->ptl);
4057 return ret;
4058}
4059
4060static unsigned long fault_around_bytes __read_mostly =
4061 rounddown_pow_of_two(65536);
4062
4063#ifdef CONFIG_DEBUG_FS
4064static int fault_around_bytes_get(void *data, u64 *val)
4065{
4066 *val = fault_around_bytes;
4067 return 0;
4068}
4069
4070/*
4071 * fault_around_bytes must be rounded down to the nearest page order as it's
4072 * what do_fault_around() expects to see.
4073 */
4074static int fault_around_bytes_set(void *data, u64 val)
4075{
4076 if (val / PAGE_SIZE > PTRS_PER_PTE)
4077 return -EINVAL;
4078 if (val > PAGE_SIZE)
4079 fault_around_bytes = rounddown_pow_of_two(val);
4080 else
4081 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
4082 return 0;
4083}
4084DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4085 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4086
4087static int __init fault_around_debugfs(void)
4088{
4089 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4090 &fault_around_bytes_fops);
4091 return 0;
4092}
4093late_initcall(fault_around_debugfs);
4094#endif
4095
4096/*
4097 * do_fault_around() tries to map few pages around the fault address. The hope
4098 * is that the pages will be needed soon and this will lower the number of
4099 * faults to handle.
4100 *
4101 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4102 * not ready to be mapped: not up-to-date, locked, etc.
4103 *
4104 * This function is called with the page table lock taken. In the split ptlock
4105 * case the page table lock only protects only those entries which belong to
4106 * the page table corresponding to the fault address.
4107 *
4108 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4109 * only once.
4110 *
4111 * fault_around_bytes defines how many bytes we'll try to map.
4112 * do_fault_around() expects it to be set to a power of two less than or equal
4113 * to PTRS_PER_PTE.
4114 *
4115 * The virtual address of the area that we map is naturally aligned to
4116 * fault_around_bytes rounded down to the machine page size
4117 * (and therefore to page order). This way it's easier to guarantee
4118 * that we don't cross page table boundaries.
4119 */
4120static vm_fault_t do_fault_around(struct vm_fault *vmf)
4121{
4122 unsigned long address = vmf->address, nr_pages, mask;
4123 pgoff_t start_pgoff = vmf->pgoff;
4124 pgoff_t end_pgoff;
4125 int off;
4126
4127 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
4128 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
4129
4130 address = max(address & mask, vmf->vma->vm_start);
4131 off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
4132 start_pgoff -= off;
4133
4134 /*
4135 * end_pgoff is either the end of the page table, the end of
4136 * the vma or nr_pages from start_pgoff, depending what is nearest.
4137 */
4138 end_pgoff = start_pgoff -
4139 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
4140 PTRS_PER_PTE - 1;
4141 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
4142 start_pgoff + nr_pages - 1);
4143
4144 if (pmd_none(*vmf->pmd)) {
4145 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4146 if (!vmf->prealloc_pte)
4147 return VM_FAULT_OOM;
4148 smp_wmb(); /* See comment in __pte_alloc() */
4149 }
4150
4151 return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
4152}
4153
4154static vm_fault_t do_read_fault(struct vm_fault *vmf)
4155{
4156 struct vm_area_struct *vma = vmf->vma;
4157 vm_fault_t ret = 0;
4158
4159 /*
4160 * Let's call ->map_pages() first and use ->fault() as fallback
4161 * if page by the offset is not ready to be mapped (cold cache or
4162 * something).
4163 */
4164 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
4165 if (likely(!userfaultfd_minor(vmf->vma))) {
4166 ret = do_fault_around(vmf);
4167 if (ret)
4168 return ret;
4169 }
4170 }
4171
4172 ret = __do_fault(vmf);
4173 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4174 return ret;
4175
4176 ret |= finish_fault(vmf);
4177 unlock_page(vmf->page);
4178 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4179 put_page(vmf->page);
4180 return ret;
4181}
4182
4183static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4184{
4185 struct vm_area_struct *vma = vmf->vma;
4186 vm_fault_t ret;
4187
4188 if (unlikely(anon_vma_prepare(vma)))
4189 return VM_FAULT_OOM;
4190
4191 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4192 if (!vmf->cow_page)
4193 return VM_FAULT_OOM;
4194
4195 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4196 put_page(vmf->cow_page);
4197 return VM_FAULT_OOM;
4198 }
4199 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4200
4201 ret = __do_fault(vmf);
4202 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4203 goto uncharge_out;
4204 if (ret & VM_FAULT_DONE_COW)
4205 return ret;
4206
4207 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4208 __SetPageUptodate(vmf->cow_page);
4209
4210 ret |= finish_fault(vmf);
4211 unlock_page(vmf->page);
4212 put_page(vmf->page);
4213 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4214 goto uncharge_out;
4215 return ret;
4216uncharge_out:
4217 put_page(vmf->cow_page);
4218 return ret;
4219}
4220
4221static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4222{
4223 struct vm_area_struct *vma = vmf->vma;
4224 vm_fault_t ret, tmp;
4225
4226 ret = __do_fault(vmf);
4227 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4228 return ret;
4229
4230 /*
4231 * Check if the backing address space wants to know that the page is
4232 * about to become writable
4233 */
4234 if (vma->vm_ops->page_mkwrite) {
4235 unlock_page(vmf->page);
4236 tmp = do_page_mkwrite(vmf);
4237 if (unlikely(!tmp ||
4238 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4239 put_page(vmf->page);
4240 return tmp;
4241 }
4242 }
4243
4244 ret |= finish_fault(vmf);
4245 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4246 VM_FAULT_RETRY))) {
4247 unlock_page(vmf->page);
4248 put_page(vmf->page);
4249 return ret;
4250 }
4251
4252 ret |= fault_dirty_shared_page(vmf);
4253 return ret;
4254}
4255
4256/*
4257 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4258 * but allow concurrent faults).
4259 * The mmap_lock may have been released depending on flags and our
4260 * return value. See filemap_fault() and __lock_page_or_retry().
4261 * If mmap_lock is released, vma may become invalid (for example
4262 * by other thread calling munmap()).
4263 */
4264static vm_fault_t do_fault(struct vm_fault *vmf)
4265{
4266 struct vm_area_struct *vma = vmf->vma;
4267 struct mm_struct *vm_mm = vma->vm_mm;
4268 vm_fault_t ret;
4269
4270 /*
4271 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4272 */
4273 if (!vma->vm_ops->fault) {
4274 /*
4275 * If we find a migration pmd entry or a none pmd entry, which
4276 * should never happen, return SIGBUS
4277 */
4278 if (unlikely(!pmd_present(*vmf->pmd)))
4279 ret = VM_FAULT_SIGBUS;
4280 else {
4281 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4282 vmf->pmd,
4283 vmf->address,
4284 &vmf->ptl);
4285 /*
4286 * Make sure this is not a temporary clearing of pte
4287 * by holding ptl and checking again. A R/M/W update
4288 * of pte involves: take ptl, clearing the pte so that
4289 * we don't have concurrent modification by hardware
4290 * followed by an update.
4291 */
4292 if (unlikely(pte_none(*vmf->pte)))
4293 ret = VM_FAULT_SIGBUS;
4294 else
4295 ret = VM_FAULT_NOPAGE;
4296
4297 pte_unmap_unlock(vmf->pte, vmf->ptl);
4298 }
4299 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4300 ret = do_read_fault(vmf);
4301 else if (!(vma->vm_flags & VM_SHARED))
4302 ret = do_cow_fault(vmf);
4303 else
4304 ret = do_shared_fault(vmf);
4305
4306 /* preallocated pagetable is unused: free it */
4307 if (vmf->prealloc_pte) {
4308 pte_free(vm_mm, vmf->prealloc_pte);
4309 vmf->prealloc_pte = NULL;
4310 }
4311 return ret;
4312}
4313
4314int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4315 unsigned long addr, int page_nid, int *flags)
4316{
4317 get_page(page);
4318
4319 count_vm_numa_event(NUMA_HINT_FAULTS);
4320 if (page_nid == numa_node_id()) {
4321 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4322 *flags |= TNF_FAULT_LOCAL;
4323 }
4324
4325 return mpol_misplaced(page, vma, addr);
4326}
4327
4328static vm_fault_t do_numa_page(struct vm_fault *vmf)
4329{
4330 struct vm_area_struct *vma = vmf->vma;
4331 struct page *page = NULL;
4332 int page_nid = NUMA_NO_NODE;
4333 int last_cpupid;
4334 int target_nid;
4335 pte_t pte, old_pte;
4336 bool was_writable = pte_savedwrite(vmf->orig_pte);
4337 int flags = 0;
4338
4339 /*
4340 * The "pte" at this point cannot be used safely without
4341 * validation through pte_unmap_same(). It's of NUMA type but
4342 * the pfn may be screwed if the read is non atomic.
4343 */
4344 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4345 spin_lock(vmf->ptl);
4346 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4347 pte_unmap_unlock(vmf->pte, vmf->ptl);
4348 goto out;
4349 }
4350
4351 /* Get the normal PTE */
4352 old_pte = ptep_get(vmf->pte);
4353 pte = pte_modify(old_pte, vma->vm_page_prot);
4354
4355 page = vm_normal_page(vma, vmf->address, pte);
4356 if (!page)
4357 goto out_map;
4358
4359 /* TODO: handle PTE-mapped THP */
4360 if (PageCompound(page))
4361 goto out_map;
4362
4363 /*
4364 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4365 * much anyway since they can be in shared cache state. This misses
4366 * the case where a mapping is writable but the process never writes
4367 * to it but pte_write gets cleared during protection updates and
4368 * pte_dirty has unpredictable behaviour between PTE scan updates,
4369 * background writeback, dirty balancing and application behaviour.
4370 */
4371 if (!was_writable)
4372 flags |= TNF_NO_GROUP;
4373
4374 /*
4375 * Flag if the page is shared between multiple address spaces. This
4376 * is later used when determining whether to group tasks together
4377 */
4378 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4379 flags |= TNF_SHARED;
4380
4381 last_cpupid = page_cpupid_last(page);
4382 page_nid = page_to_nid(page);
4383 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4384 &flags);
4385 if (target_nid == NUMA_NO_NODE) {
4386 put_page(page);
4387 goto out_map;
4388 }
4389 pte_unmap_unlock(vmf->pte, vmf->ptl);
4390
4391 /* Migrate to the requested node */
4392 if (migrate_misplaced_page(page, vma, target_nid)) {
4393 page_nid = target_nid;
4394 flags |= TNF_MIGRATED;
4395 } else {
4396 flags |= TNF_MIGRATE_FAIL;
4397 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4398 spin_lock(vmf->ptl);
4399 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4400 pte_unmap_unlock(vmf->pte, vmf->ptl);
4401 goto out;
4402 }
4403 goto out_map;
4404 }
4405
4406out:
4407 if (page_nid != NUMA_NO_NODE)
4408 task_numa_fault(last_cpupid, page_nid, 1, flags);
4409 return 0;
4410out_map:
4411 /*
4412 * Make it present again, depending on how arch implements
4413 * non-accessible ptes, some can allow access by kernel mode.
4414 */
4415 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4416 pte = pte_modify(old_pte, vma->vm_page_prot);
4417 pte = pte_mkyoung(pte);
4418 if (was_writable)
4419 pte = pte_mkwrite(pte);
4420 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4421 update_mmu_cache(vma, vmf->address, vmf->pte);
4422 pte_unmap_unlock(vmf->pte, vmf->ptl);
4423 goto out;
4424}
4425
4426static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4427{
4428 if (vma_is_anonymous(vmf->vma))
4429 return do_huge_pmd_anonymous_page(vmf);
4430 if (vmf->vma->vm_ops->huge_fault)
4431 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4432 return VM_FAULT_FALLBACK;
4433}
4434
4435/* `inline' is required to avoid gcc 4.1.2 build error */
4436static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
4437{
4438 if (vma_is_anonymous(vmf->vma)) {
4439 if (userfaultfd_huge_pmd_wp(vmf->vma, vmf->orig_pmd))
4440 return handle_userfault(vmf, VM_UFFD_WP);
4441 return do_huge_pmd_wp_page(vmf);
4442 }
4443 if (vmf->vma->vm_ops->huge_fault) {
4444 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4445
4446 if (!(ret & VM_FAULT_FALLBACK))
4447 return ret;
4448 }
4449
4450 /* COW or write-notify handled on pte level: split pmd. */
4451 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4452
4453 return VM_FAULT_FALLBACK;
4454}
4455
4456static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4457{
4458#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4459 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4460 /* No support for anonymous transparent PUD pages yet */
4461 if (vma_is_anonymous(vmf->vma))
4462 goto split;
4463 if (vmf->vma->vm_ops->huge_fault) {
4464 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4465
4466 if (!(ret & VM_FAULT_FALLBACK))
4467 return ret;
4468 }
4469split:
4470 /* COW or write-notify not handled on PUD level: split pud.*/
4471 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4472#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4473 return VM_FAULT_FALLBACK;
4474}
4475
4476static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4477{
4478#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4479 /* No support for anonymous transparent PUD pages yet */
4480 if (vma_is_anonymous(vmf->vma))
4481 return VM_FAULT_FALLBACK;
4482 if (vmf->vma->vm_ops->huge_fault)
4483 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4484#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4485 return VM_FAULT_FALLBACK;
4486}
4487
4488/*
4489 * These routines also need to handle stuff like marking pages dirty
4490 * and/or accessed for architectures that don't do it in hardware (most
4491 * RISC architectures). The early dirtying is also good on the i386.
4492 *
4493 * There is also a hook called "update_mmu_cache()" that architectures
4494 * with external mmu caches can use to update those (ie the Sparc or
4495 * PowerPC hashed page tables that act as extended TLBs).
4496 *
4497 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4498 * concurrent faults).
4499 *
4500 * The mmap_lock may have been released depending on flags and our return value.
4501 * See filemap_fault() and __lock_page_or_retry().
4502 */
4503static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4504{
4505 pte_t entry;
4506
4507 if (unlikely(pmd_none(*vmf->pmd))) {
4508 /*
4509 * Leave __pte_alloc() until later: because vm_ops->fault may
4510 * want to allocate huge page, and if we expose page table
4511 * for an instant, it will be difficult to retract from
4512 * concurrent faults and from rmap lookups.
4513 */
4514 vmf->pte = NULL;
4515 } else {
4516 /*
4517 * If a huge pmd materialized under us just retry later. Use
4518 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4519 * of pmd_trans_huge() to ensure the pmd didn't become
4520 * pmd_trans_huge under us and then back to pmd_none, as a
4521 * result of MADV_DONTNEED running immediately after a huge pmd
4522 * fault in a different thread of this mm, in turn leading to a
4523 * misleading pmd_trans_huge() retval. All we have to ensure is
4524 * that it is a regular pmd that we can walk with
4525 * pte_offset_map() and we can do that through an atomic read
4526 * in C, which is what pmd_trans_unstable() provides.
4527 */
4528 if (pmd_devmap_trans_unstable(vmf->pmd))
4529 return 0;
4530 /*
4531 * A regular pmd is established and it can't morph into a huge
4532 * pmd from under us anymore at this point because we hold the
4533 * mmap_lock read mode and khugepaged takes it in write mode.
4534 * So now it's safe to run pte_offset_map().
4535 */
4536 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4537 vmf->orig_pte = *vmf->pte;
4538
4539 /*
4540 * some architectures can have larger ptes than wordsize,
4541 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4542 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4543 * accesses. The code below just needs a consistent view
4544 * for the ifs and we later double check anyway with the
4545 * ptl lock held. So here a barrier will do.
4546 */
4547 barrier();
4548 if (pte_none(vmf->orig_pte)) {
4549 pte_unmap(vmf->pte);
4550 vmf->pte = NULL;
4551 }
4552 }
4553
4554 if (!vmf->pte) {
4555 if (vma_is_anonymous(vmf->vma))
4556 return do_anonymous_page(vmf);
4557 else
4558 return do_fault(vmf);
4559 }
4560
4561 if (!pte_present(vmf->orig_pte))
4562 return do_swap_page(vmf);
4563
4564 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4565 return do_numa_page(vmf);
4566
4567 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4568 spin_lock(vmf->ptl);
4569 entry = vmf->orig_pte;
4570 if (unlikely(!pte_same(*vmf->pte, entry))) {
4571 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4572 goto unlock;
4573 }
4574 if (vmf->flags & FAULT_FLAG_WRITE) {
4575 if (!pte_write(entry))
4576 return do_wp_page(vmf);
4577 entry = pte_mkdirty(entry);
4578 }
4579 entry = pte_mkyoung(entry);
4580 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4581 vmf->flags & FAULT_FLAG_WRITE)) {
4582 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4583 } else {
4584 /* Skip spurious TLB flush for retried page fault */
4585 if (vmf->flags & FAULT_FLAG_TRIED)
4586 goto unlock;
4587 /*
4588 * This is needed only for protection faults but the arch code
4589 * is not yet telling us if this is a protection fault or not.
4590 * This still avoids useless tlb flushes for .text page faults
4591 * with threads.
4592 */
4593 if (vmf->flags & FAULT_FLAG_WRITE)
4594 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4595 }
4596unlock:
4597 pte_unmap_unlock(vmf->pte, vmf->ptl);
4598 return 0;
4599}
4600
4601/*
4602 * By the time we get here, we already hold the mm semaphore
4603 *
4604 * The mmap_lock may have been released depending on flags and our
4605 * return value. See filemap_fault() and __lock_page_or_retry().
4606 */
4607static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4608 unsigned long address, unsigned int flags)
4609{
4610 struct vm_fault vmf = {
4611 .vma = vma,
4612 .address = address & PAGE_MASK,
4613 .flags = flags,
4614 .pgoff = linear_page_index(vma, address),
4615 .gfp_mask = __get_fault_gfp_mask(vma),
4616 };
4617 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4618 struct mm_struct *mm = vma->vm_mm;
4619 pgd_t *pgd;
4620 p4d_t *p4d;
4621 vm_fault_t ret;
4622
4623 pgd = pgd_offset(mm, address);
4624 p4d = p4d_alloc(mm, pgd, address);
4625 if (!p4d)
4626 return VM_FAULT_OOM;
4627
4628 vmf.pud = pud_alloc(mm, p4d, address);
4629 if (!vmf.pud)
4630 return VM_FAULT_OOM;
4631retry_pud:
4632 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4633 ret = create_huge_pud(&vmf);
4634 if (!(ret & VM_FAULT_FALLBACK))
4635 return ret;
4636 } else {
4637 pud_t orig_pud = *vmf.pud;
4638
4639 barrier();
4640 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4641
4642 /* NUMA case for anonymous PUDs would go here */
4643
4644 if (dirty && !pud_write(orig_pud)) {
4645 ret = wp_huge_pud(&vmf, orig_pud);
4646 if (!(ret & VM_FAULT_FALLBACK))
4647 return ret;
4648 } else {
4649 huge_pud_set_accessed(&vmf, orig_pud);
4650 return 0;
4651 }
4652 }
4653 }
4654
4655 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4656 if (!vmf.pmd)
4657 return VM_FAULT_OOM;
4658
4659 /* Huge pud page fault raced with pmd_alloc? */
4660 if (pud_trans_unstable(vmf.pud))
4661 goto retry_pud;
4662
4663 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4664 ret = create_huge_pmd(&vmf);
4665 if (!(ret & VM_FAULT_FALLBACK))
4666 return ret;
4667 } else {
4668 vmf.orig_pmd = *vmf.pmd;
4669
4670 barrier();
4671 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
4672 VM_BUG_ON(thp_migration_supported() &&
4673 !is_pmd_migration_entry(vmf.orig_pmd));
4674 if (is_pmd_migration_entry(vmf.orig_pmd))
4675 pmd_migration_entry_wait(mm, vmf.pmd);
4676 return 0;
4677 }
4678 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
4679 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
4680 return do_huge_pmd_numa_page(&vmf);
4681
4682 if (dirty && !pmd_write(vmf.orig_pmd)) {
4683 ret = wp_huge_pmd(&vmf);
4684 if (!(ret & VM_FAULT_FALLBACK))
4685 return ret;
4686 } else {
4687 huge_pmd_set_accessed(&vmf);
4688 return 0;
4689 }
4690 }
4691 }
4692
4693 return handle_pte_fault(&vmf);
4694}
4695
4696/**
4697 * mm_account_fault - Do page fault accounting
4698 *
4699 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4700 * of perf event counters, but we'll still do the per-task accounting to
4701 * the task who triggered this page fault.
4702 * @address: the faulted address.
4703 * @flags: the fault flags.
4704 * @ret: the fault retcode.
4705 *
4706 * This will take care of most of the page fault accounting. Meanwhile, it
4707 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4708 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4709 * still be in per-arch page fault handlers at the entry of page fault.
4710 */
4711static inline void mm_account_fault(struct pt_regs *regs,
4712 unsigned long address, unsigned int flags,
4713 vm_fault_t ret)
4714{
4715 bool major;
4716
4717 /*
4718 * We don't do accounting for some specific faults:
4719 *
4720 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4721 * includes arch_vma_access_permitted() failing before reaching here.
4722 * So this is not a "this many hardware page faults" counter. We
4723 * should use the hw profiling for that.
4724 *
4725 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4726 * once they're completed.
4727 */
4728 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4729 return;
4730
4731 /*
4732 * We define the fault as a major fault when the final successful fault
4733 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4734 * handle it immediately previously).
4735 */
4736 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4737
4738 if (major)
4739 current->maj_flt++;
4740 else
4741 current->min_flt++;
4742
4743 /*
4744 * If the fault is done for GUP, regs will be NULL. We only do the
4745 * accounting for the per thread fault counters who triggered the
4746 * fault, and we skip the perf event updates.
4747 */
4748 if (!regs)
4749 return;
4750
4751 if (major)
4752 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4753 else
4754 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4755}
4756
4757/*
4758 * By the time we get here, we already hold the mm semaphore
4759 *
4760 * The mmap_lock may have been released depending on flags and our
4761 * return value. See filemap_fault() and __lock_page_or_retry().
4762 */
4763vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4764 unsigned int flags, struct pt_regs *regs)
4765{
4766 vm_fault_t ret;
4767
4768 __set_current_state(TASK_RUNNING);
4769
4770 count_vm_event(PGFAULT);
4771 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4772
4773 /* do counter updates before entering really critical section. */
4774 check_sync_rss_stat(current);
4775
4776 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4777 flags & FAULT_FLAG_INSTRUCTION,
4778 flags & FAULT_FLAG_REMOTE))
4779 return VM_FAULT_SIGSEGV;
4780
4781 /*
4782 * Enable the memcg OOM handling for faults triggered in user
4783 * space. Kernel faults are handled more gracefully.
4784 */
4785 if (flags & FAULT_FLAG_USER)
4786 mem_cgroup_enter_user_fault();
4787
4788 if (unlikely(is_vm_hugetlb_page(vma)))
4789 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4790 else
4791 ret = __handle_mm_fault(vma, address, flags);
4792
4793 if (flags & FAULT_FLAG_USER) {
4794 mem_cgroup_exit_user_fault();
4795 /*
4796 * The task may have entered a memcg OOM situation but
4797 * if the allocation error was handled gracefully (no
4798 * VM_FAULT_OOM), there is no need to kill anything.
4799 * Just clean up the OOM state peacefully.
4800 */
4801 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4802 mem_cgroup_oom_synchronize(false);
4803 }
4804
4805 mm_account_fault(regs, address, flags, ret);
4806
4807 return ret;
4808}
4809EXPORT_SYMBOL_GPL(handle_mm_fault);
4810
4811#ifndef __PAGETABLE_P4D_FOLDED
4812/*
4813 * Allocate p4d page table.
4814 * We've already handled the fast-path in-line.
4815 */
4816int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4817{
4818 p4d_t *new = p4d_alloc_one(mm, address);
4819 if (!new)
4820 return -ENOMEM;
4821
4822 smp_wmb(); /* See comment in __pte_alloc */
4823
4824 spin_lock(&mm->page_table_lock);
4825 if (pgd_present(*pgd)) /* Another has populated it */
4826 p4d_free(mm, new);
4827 else
4828 pgd_populate(mm, pgd, new);
4829 spin_unlock(&mm->page_table_lock);
4830 return 0;
4831}
4832#endif /* __PAGETABLE_P4D_FOLDED */
4833
4834#ifndef __PAGETABLE_PUD_FOLDED
4835/*
4836 * Allocate page upper directory.
4837 * We've already handled the fast-path in-line.
4838 */
4839int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4840{
4841 pud_t *new = pud_alloc_one(mm, address);
4842 if (!new)
4843 return -ENOMEM;
4844
4845 smp_wmb(); /* See comment in __pte_alloc */
4846
4847 spin_lock(&mm->page_table_lock);
4848 if (!p4d_present(*p4d)) {
4849 mm_inc_nr_puds(mm);
4850 p4d_populate(mm, p4d, new);
4851 } else /* Another has populated it */
4852 pud_free(mm, new);
4853 spin_unlock(&mm->page_table_lock);
4854 return 0;
4855}
4856#endif /* __PAGETABLE_PUD_FOLDED */
4857
4858#ifndef __PAGETABLE_PMD_FOLDED
4859/*
4860 * Allocate page middle directory.
4861 * We've already handled the fast-path in-line.
4862 */
4863int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4864{
4865 spinlock_t *ptl;
4866 pmd_t *new = pmd_alloc_one(mm, address);
4867 if (!new)
4868 return -ENOMEM;
4869
4870 smp_wmb(); /* See comment in __pte_alloc */
4871
4872 ptl = pud_lock(mm, pud);
4873 if (!pud_present(*pud)) {
4874 mm_inc_nr_pmds(mm);
4875 pud_populate(mm, pud, new);
4876 } else /* Another has populated it */
4877 pmd_free(mm, new);
4878 spin_unlock(ptl);
4879 return 0;
4880}
4881#endif /* __PAGETABLE_PMD_FOLDED */
4882
4883int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4884 struct mmu_notifier_range *range, pte_t **ptepp,
4885 pmd_t **pmdpp, spinlock_t **ptlp)
4886{
4887 pgd_t *pgd;
4888 p4d_t *p4d;
4889 pud_t *pud;
4890 pmd_t *pmd;
4891 pte_t *ptep;
4892
4893 pgd = pgd_offset(mm, address);
4894 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4895 goto out;
4896
4897 p4d = p4d_offset(pgd, address);
4898 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4899 goto out;
4900
4901 pud = pud_offset(p4d, address);
4902 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4903 goto out;
4904
4905 pmd = pmd_offset(pud, address);
4906 VM_BUG_ON(pmd_trans_huge(*pmd));
4907
4908 if (pmd_huge(*pmd)) {
4909 if (!pmdpp)
4910 goto out;
4911
4912 if (range) {
4913 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4914 NULL, mm, address & PMD_MASK,
4915 (address & PMD_MASK) + PMD_SIZE);
4916 mmu_notifier_invalidate_range_start(range);
4917 }
4918 *ptlp = pmd_lock(mm, pmd);
4919 if (pmd_huge(*pmd)) {
4920 *pmdpp = pmd;
4921 return 0;
4922 }
4923 spin_unlock(*ptlp);
4924 if (range)
4925 mmu_notifier_invalidate_range_end(range);
4926 }
4927
4928 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4929 goto out;
4930
4931 if (range) {
4932 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4933 address & PAGE_MASK,
4934 (address & PAGE_MASK) + PAGE_SIZE);
4935 mmu_notifier_invalidate_range_start(range);
4936 }
4937 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4938 if (!pte_present(*ptep))
4939 goto unlock;
4940 *ptepp = ptep;
4941 return 0;
4942unlock:
4943 pte_unmap_unlock(ptep, *ptlp);
4944 if (range)
4945 mmu_notifier_invalidate_range_end(range);
4946out:
4947 return -EINVAL;
4948}
4949
4950/**
4951 * follow_pte - look up PTE at a user virtual address
4952 * @mm: the mm_struct of the target address space
4953 * @address: user virtual address
4954 * @ptepp: location to store found PTE
4955 * @ptlp: location to store the lock for the PTE
4956 *
4957 * On a successful return, the pointer to the PTE is stored in @ptepp;
4958 * the corresponding lock is taken and its location is stored in @ptlp.
4959 * The contents of the PTE are only stable until @ptlp is released;
4960 * any further use, if any, must be protected against invalidation
4961 * with MMU notifiers.
4962 *
4963 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4964 * should be taken for read.
4965 *
4966 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
4967 * it is not a good general-purpose API.
4968 *
4969 * Return: zero on success, -ve otherwise.
4970 */
4971int follow_pte(struct mm_struct *mm, unsigned long address,
4972 pte_t **ptepp, spinlock_t **ptlp)
4973{
4974 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
4975}
4976EXPORT_SYMBOL_GPL(follow_pte);
4977
4978/**
4979 * follow_pfn - look up PFN at a user virtual address
4980 * @vma: memory mapping
4981 * @address: user virtual address
4982 * @pfn: location to store found PFN
4983 *
4984 * Only IO mappings and raw PFN mappings are allowed.
4985 *
4986 * This function does not allow the caller to read the permissions
4987 * of the PTE. Do not use it.
4988 *
4989 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4990 */
4991int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4992 unsigned long *pfn)
4993{
4994 int ret = -EINVAL;
4995 spinlock_t *ptl;
4996 pte_t *ptep;
4997
4998 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4999 return ret;
5000
5001 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5002 if (ret)
5003 return ret;
5004 *pfn = pte_pfn(*ptep);
5005 pte_unmap_unlock(ptep, ptl);
5006 return 0;
5007}
5008EXPORT_SYMBOL(follow_pfn);
5009
5010#ifdef CONFIG_HAVE_IOREMAP_PROT
5011int follow_phys(struct vm_area_struct *vma,
5012 unsigned long address, unsigned int flags,
5013 unsigned long *prot, resource_size_t *phys)
5014{
5015 int ret = -EINVAL;
5016 pte_t *ptep, pte;
5017 spinlock_t *ptl;
5018
5019 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5020 goto out;
5021
5022 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5023 goto out;
5024 pte = *ptep;
5025
5026 if ((flags & FOLL_WRITE) && !pte_write(pte))
5027 goto unlock;
5028
5029 *prot = pgprot_val(pte_pgprot(pte));
5030 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5031
5032 ret = 0;
5033unlock:
5034 pte_unmap_unlock(ptep, ptl);
5035out:
5036 return ret;
5037}
5038
5039/**
5040 * generic_access_phys - generic implementation for iomem mmap access
5041 * @vma: the vma to access
5042 * @addr: userspace address, not relative offset within @vma
5043 * @buf: buffer to read/write
5044 * @len: length of transfer
5045 * @write: set to FOLL_WRITE when writing, otherwise reading
5046 *
5047 * This is a generic implementation for &vm_operations_struct.access for an
5048 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5049 * not page based.
5050 */
5051int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5052 void *buf, int len, int write)
5053{
5054 resource_size_t phys_addr;
5055 unsigned long prot = 0;
5056 void __iomem *maddr;
5057 pte_t *ptep, pte;
5058 spinlock_t *ptl;
5059 int offset = offset_in_page(addr);
5060 int ret = -EINVAL;
5061
5062 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5063 return -EINVAL;
5064
5065retry:
5066 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5067 return -EINVAL;
5068 pte = *ptep;
5069 pte_unmap_unlock(ptep, ptl);
5070
5071 prot = pgprot_val(pte_pgprot(pte));
5072 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5073
5074 if ((write & FOLL_WRITE) && !pte_write(pte))
5075 return -EINVAL;
5076
5077 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5078 if (!maddr)
5079 return -ENOMEM;
5080
5081 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5082 goto out_unmap;
5083
5084 if (!pte_same(pte, *ptep)) {
5085 pte_unmap_unlock(ptep, ptl);
5086 iounmap(maddr);
5087
5088 goto retry;
5089 }
5090
5091 if (write)
5092 memcpy_toio(maddr + offset, buf, len);
5093 else
5094 memcpy_fromio(buf, maddr + offset, len);
5095 ret = len;
5096 pte_unmap_unlock(ptep, ptl);
5097out_unmap:
5098 iounmap(maddr);
5099
5100 return ret;
5101}
5102EXPORT_SYMBOL_GPL(generic_access_phys);
5103#endif
5104
5105/*
5106 * Access another process' address space as given in mm.
5107 */
5108int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
5109 int len, unsigned int gup_flags)
5110{
5111 struct vm_area_struct *vma;
5112 void *old_buf = buf;
5113 int write = gup_flags & FOLL_WRITE;
5114
5115 if (mmap_read_lock_killable(mm))
5116 return 0;
5117
5118 /* ignore errors, just check how much was successfully transferred */
5119 while (len) {
5120 int bytes, ret, offset;
5121 void *maddr;
5122 struct page *page = NULL;
5123
5124 ret = get_user_pages_remote(mm, addr, 1,
5125 gup_flags, &page, &vma, NULL);
5126 if (ret <= 0) {
5127#ifndef CONFIG_HAVE_IOREMAP_PROT
5128 break;
5129#else
5130 /*
5131 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5132 * we can access using slightly different code.
5133 */
5134 vma = vma_lookup(mm, addr);
5135 if (!vma)
5136 break;
5137 if (vma->vm_ops && vma->vm_ops->access)
5138 ret = vma->vm_ops->access(vma, addr, buf,
5139 len, write);
5140 if (ret <= 0)
5141 break;
5142 bytes = ret;
5143#endif
5144 } else {
5145 bytes = len;
5146 offset = addr & (PAGE_SIZE-1);
5147 if (bytes > PAGE_SIZE-offset)
5148 bytes = PAGE_SIZE-offset;
5149
5150 maddr = kmap(page);
5151 if (write) {
5152 copy_to_user_page(vma, page, addr,
5153 maddr + offset, buf, bytes);
5154 set_page_dirty_lock(page);
5155 } else {
5156 copy_from_user_page(vma, page, addr,
5157 buf, maddr + offset, bytes);
5158 }
5159 kunmap(page);
5160 put_page(page);
5161 }
5162 len -= bytes;
5163 buf += bytes;
5164 addr += bytes;
5165 }
5166 mmap_read_unlock(mm);
5167
5168 return buf - old_buf;
5169}
5170
5171/**
5172 * access_remote_vm - access another process' address space
5173 * @mm: the mm_struct of the target address space
5174 * @addr: start address to access
5175 * @buf: source or destination buffer
5176 * @len: number of bytes to transfer
5177 * @gup_flags: flags modifying lookup behaviour
5178 *
5179 * The caller must hold a reference on @mm.
5180 *
5181 * Return: number of bytes copied from source to destination.
5182 */
5183int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5184 void *buf, int len, unsigned int gup_flags)
5185{
5186 return __access_remote_vm(mm, addr, buf, len, gup_flags);
5187}
5188
5189/*
5190 * Access another process' address space.
5191 * Source/target buffer must be kernel space,
5192 * Do not walk the page table directly, use get_user_pages
5193 */
5194int access_process_vm(struct task_struct *tsk, unsigned long addr,
5195 void *buf, int len, unsigned int gup_flags)
5196{
5197 struct mm_struct *mm;
5198 int ret;
5199
5200 mm = get_task_mm(tsk);
5201 if (!mm)
5202 return 0;
5203
5204 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5205
5206 mmput(mm);
5207
5208 return ret;
5209}
5210EXPORT_SYMBOL_GPL(access_process_vm);
5211
5212/*
5213 * Print the name of a VMA.
5214 */
5215void print_vma_addr(char *prefix, unsigned long ip)
5216{
5217 struct mm_struct *mm = current->mm;
5218 struct vm_area_struct *vma;
5219
5220 /*
5221 * we might be running from an atomic context so we cannot sleep
5222 */
5223 if (!mmap_read_trylock(mm))
5224 return;
5225
5226 vma = find_vma(mm, ip);
5227 if (vma && vma->vm_file) {
5228 struct file *f = vma->vm_file;
5229 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5230 if (buf) {
5231 char *p;
5232
5233 p = file_path(f, buf, PAGE_SIZE);
5234 if (IS_ERR(p))
5235 p = "?";
5236 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5237 vma->vm_start,
5238 vma->vm_end - vma->vm_start);
5239 free_page((unsigned long)buf);
5240 }
5241 }
5242 mmap_read_unlock(mm);
5243}
5244
5245#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5246void __might_fault(const char *file, int line)
5247{
5248 /*
5249 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5250 * holding the mmap_lock, this is safe because kernel memory doesn't
5251 * get paged out, therefore we'll never actually fault, and the
5252 * below annotations will generate false positives.
5253 */
5254 if (uaccess_kernel())
5255 return;
5256 if (pagefault_disabled())
5257 return;
5258 __might_sleep(file, line, 0);
5259#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5260 if (current->mm)
5261 might_lock_read(¤t->mm->mmap_lock);
5262#endif
5263}
5264EXPORT_SYMBOL(__might_fault);
5265#endif
5266
5267#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5268/*
5269 * Process all subpages of the specified huge page with the specified
5270 * operation. The target subpage will be processed last to keep its
5271 * cache lines hot.
5272 */
5273static inline void process_huge_page(
5274 unsigned long addr_hint, unsigned int pages_per_huge_page,
5275 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5276 void *arg)
5277{
5278 int i, n, base, l;
5279 unsigned long addr = addr_hint &
5280 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5281
5282 /* Process target subpage last to keep its cache lines hot */
5283 might_sleep();
5284 n = (addr_hint - addr) / PAGE_SIZE;
5285 if (2 * n <= pages_per_huge_page) {
5286 /* If target subpage in first half of huge page */
5287 base = 0;
5288 l = n;
5289 /* Process subpages at the end of huge page */
5290 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5291 cond_resched();
5292 process_subpage(addr + i * PAGE_SIZE, i, arg);
5293 }
5294 } else {
5295 /* If target subpage in second half of huge page */
5296 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5297 l = pages_per_huge_page - n;
5298 /* Process subpages at the begin of huge page */
5299 for (i = 0; i < base; i++) {
5300 cond_resched();
5301 process_subpage(addr + i * PAGE_SIZE, i, arg);
5302 }
5303 }
5304 /*
5305 * Process remaining subpages in left-right-left-right pattern
5306 * towards the target subpage
5307 */
5308 for (i = 0; i < l; i++) {
5309 int left_idx = base + i;
5310 int right_idx = base + 2 * l - 1 - i;
5311
5312 cond_resched();
5313 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5314 cond_resched();
5315 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5316 }
5317}
5318
5319static void clear_gigantic_page(struct page *page,
5320 unsigned long addr,
5321 unsigned int pages_per_huge_page)
5322{
5323 int i;
5324 struct page *p = page;
5325
5326 might_sleep();
5327 for (i = 0; i < pages_per_huge_page;
5328 i++, p = mem_map_next(p, page, i)) {
5329 cond_resched();
5330 clear_user_highpage(p, addr + i * PAGE_SIZE);
5331 }
5332}
5333
5334static void clear_subpage(unsigned long addr, int idx, void *arg)
5335{
5336 struct page *page = arg;
5337
5338 clear_user_highpage(page + idx, addr);
5339}
5340
5341void clear_huge_page(struct page *page,
5342 unsigned long addr_hint, unsigned int pages_per_huge_page)
5343{
5344 unsigned long addr = addr_hint &
5345 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5346
5347 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5348 clear_gigantic_page(page, addr, pages_per_huge_page);
5349 return;
5350 }
5351
5352 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5353}
5354
5355static void copy_user_gigantic_page(struct page *dst, struct page *src,
5356 unsigned long addr,
5357 struct vm_area_struct *vma,
5358 unsigned int pages_per_huge_page)
5359{
5360 int i;
5361 struct page *dst_base = dst;
5362 struct page *src_base = src;
5363
5364 for (i = 0; i < pages_per_huge_page; ) {
5365 cond_resched();
5366 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5367
5368 i++;
5369 dst = mem_map_next(dst, dst_base, i);
5370 src = mem_map_next(src, src_base, i);
5371 }
5372}
5373
5374struct copy_subpage_arg {
5375 struct page *dst;
5376 struct page *src;
5377 struct vm_area_struct *vma;
5378};
5379
5380static void copy_subpage(unsigned long addr, int idx, void *arg)
5381{
5382 struct copy_subpage_arg *copy_arg = arg;
5383
5384 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5385 addr, copy_arg->vma);
5386}
5387
5388void copy_user_huge_page(struct page *dst, struct page *src,
5389 unsigned long addr_hint, struct vm_area_struct *vma,
5390 unsigned int pages_per_huge_page)
5391{
5392 unsigned long addr = addr_hint &
5393 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5394 struct copy_subpage_arg arg = {
5395 .dst = dst,
5396 .src = src,
5397 .vma = vma,
5398 };
5399
5400 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5401 copy_user_gigantic_page(dst, src, addr, vma,
5402 pages_per_huge_page);
5403 return;
5404 }
5405
5406 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5407}
5408
5409long copy_huge_page_from_user(struct page *dst_page,
5410 const void __user *usr_src,
5411 unsigned int pages_per_huge_page,
5412 bool allow_pagefault)
5413{
5414 void *src = (void *)usr_src;
5415 void *page_kaddr;
5416 unsigned long i, rc = 0;
5417 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5418 struct page *subpage = dst_page;
5419
5420 for (i = 0; i < pages_per_huge_page;
5421 i++, subpage = mem_map_next(subpage, dst_page, i)) {
5422 if (allow_pagefault)
5423 page_kaddr = kmap(subpage);
5424 else
5425 page_kaddr = kmap_atomic(subpage);
5426 rc = copy_from_user(page_kaddr,
5427 (const void __user *)(src + i * PAGE_SIZE),
5428 PAGE_SIZE);
5429 if (allow_pagefault)
5430 kunmap(subpage);
5431 else
5432 kunmap_atomic(page_kaddr);
5433
5434 ret_val -= (PAGE_SIZE - rc);
5435 if (rc)
5436 break;
5437
5438 cond_resched();
5439 }
5440 return ret_val;
5441}
5442#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5443
5444#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5445
5446static struct kmem_cache *page_ptl_cachep;
5447
5448void __init ptlock_cache_init(void)
5449{
5450 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5451 SLAB_PANIC, NULL);
5452}
5453
5454bool ptlock_alloc(struct page *page)
5455{
5456 spinlock_t *ptl;
5457
5458 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5459 if (!ptl)
5460 return false;
5461 page->ptl = ptl;
5462 return true;
5463}
5464
5465void ptlock_free(struct page *page)
5466{
5467 kmem_cache_free(page_ptl_cachep, page->ptl);
5468}
5469#endif
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