Loading...
1
2// SPDX-License-Identifier: GPL-2.0-only
3/*
4 * linux/mm/memory.c
5 *
6 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 */
8
9/*
10 * demand-loading started 01.12.91 - seems it is high on the list of
11 * things wanted, and it should be easy to implement. - Linus
12 */
13
14/*
15 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
16 * pages started 02.12.91, seems to work. - Linus.
17 *
18 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
19 * would have taken more than the 6M I have free, but it worked well as
20 * far as I could see.
21 *
22 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
23 */
24
25/*
26 * Real VM (paging to/from disk) started 18.12.91. Much more work and
27 * thought has to go into this. Oh, well..
28 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
29 * Found it. Everything seems to work now.
30 * 20.12.91 - Ok, making the swap-device changeable like the root.
31 */
32
33/*
34 * 05.04.94 - Multi-page memory management added for v1.1.
35 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 *
37 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
38 * (Gerhard.Wichert@pdb.siemens.de)
39 *
40 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 */
42
43#include <linux/kernel_stat.h>
44#include <linux/mm.h>
45#include <linux/mm_inline.h>
46#include <linux/sched/mm.h>
47#include <linux/sched/coredump.h>
48#include <linux/sched/numa_balancing.h>
49#include <linux/sched/task.h>
50#include <linux/hugetlb.h>
51#include <linux/mman.h>
52#include <linux/swap.h>
53#include <linux/highmem.h>
54#include <linux/pagemap.h>
55#include <linux/memremap.h>
56#include <linux/kmsan.h>
57#include <linux/ksm.h>
58#include <linux/rmap.h>
59#include <linux/export.h>
60#include <linux/delayacct.h>
61#include <linux/init.h>
62#include <linux/pfn_t.h>
63#include <linux/writeback.h>
64#include <linux/memcontrol.h>
65#include <linux/mmu_notifier.h>
66#include <linux/swapops.h>
67#include <linux/elf.h>
68#include <linux/gfp.h>
69#include <linux/migrate.h>
70#include <linux/string.h>
71#include <linux/memory-tiers.h>
72#include <linux/debugfs.h>
73#include <linux/userfaultfd_k.h>
74#include <linux/dax.h>
75#include <linux/oom.h>
76#include <linux/numa.h>
77#include <linux/perf_event.h>
78#include <linux/ptrace.h>
79#include <linux/vmalloc.h>
80#include <linux/sched/sysctl.h>
81
82#include <trace/events/kmem.h>
83
84#include <asm/io.h>
85#include <asm/mmu_context.h>
86#include <asm/pgalloc.h>
87#include <linux/uaccess.h>
88#include <asm/tlb.h>
89#include <asm/tlbflush.h>
90
91#include "pgalloc-track.h"
92#include "internal.h"
93#include "swap.h"
94
95#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
96#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
97#endif
98
99#ifndef CONFIG_NUMA
100unsigned long max_mapnr;
101EXPORT_SYMBOL(max_mapnr);
102
103struct page *mem_map;
104EXPORT_SYMBOL(mem_map);
105#endif
106
107static vm_fault_t do_fault(struct vm_fault *vmf);
108static vm_fault_t do_anonymous_page(struct vm_fault *vmf);
109static bool vmf_pte_changed(struct vm_fault *vmf);
110
111/*
112 * Return true if the original pte was a uffd-wp pte marker (so the pte was
113 * wr-protected).
114 */
115static bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf)
116{
117 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
118 return false;
119
120 return pte_marker_uffd_wp(vmf->orig_pte);
121}
122
123/*
124 * A number of key systems in x86 including ioremap() rely on the assumption
125 * that high_memory defines the upper bound on direct map memory, then end
126 * of ZONE_NORMAL.
127 */
128void *high_memory;
129EXPORT_SYMBOL(high_memory);
130
131/*
132 * Randomize the address space (stacks, mmaps, brk, etc.).
133 *
134 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
135 * as ancient (libc5 based) binaries can segfault. )
136 */
137int randomize_va_space __read_mostly =
138#ifdef CONFIG_COMPAT_BRK
139 1;
140#else
141 2;
142#endif
143
144#ifndef arch_wants_old_prefaulted_pte
145static inline bool arch_wants_old_prefaulted_pte(void)
146{
147 /*
148 * Transitioning a PTE from 'old' to 'young' can be expensive on
149 * some architectures, even if it's performed in hardware. By
150 * default, "false" means prefaulted entries will be 'young'.
151 */
152 return false;
153}
154#endif
155
156static int __init disable_randmaps(char *s)
157{
158 randomize_va_space = 0;
159 return 1;
160}
161__setup("norandmaps", disable_randmaps);
162
163unsigned long zero_pfn __read_mostly;
164EXPORT_SYMBOL(zero_pfn);
165
166unsigned long highest_memmap_pfn __read_mostly;
167
168/*
169 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
170 */
171static int __init init_zero_pfn(void)
172{
173 zero_pfn = page_to_pfn(ZERO_PAGE(0));
174 return 0;
175}
176early_initcall(init_zero_pfn);
177
178void mm_trace_rss_stat(struct mm_struct *mm, int member)
179{
180 trace_rss_stat(mm, member);
181}
182
183/*
184 * Note: this doesn't free the actual pages themselves. That
185 * has been handled earlier when unmapping all the memory regions.
186 */
187static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
188 unsigned long addr)
189{
190 pgtable_t token = pmd_pgtable(*pmd);
191 pmd_clear(pmd);
192 pte_free_tlb(tlb, token, addr);
193 mm_dec_nr_ptes(tlb->mm);
194}
195
196static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
197 unsigned long addr, unsigned long end,
198 unsigned long floor, unsigned long ceiling)
199{
200 pmd_t *pmd;
201 unsigned long next;
202 unsigned long start;
203
204 start = addr;
205 pmd = pmd_offset(pud, addr);
206 do {
207 next = pmd_addr_end(addr, end);
208 if (pmd_none_or_clear_bad(pmd))
209 continue;
210 free_pte_range(tlb, pmd, addr);
211 } while (pmd++, addr = next, addr != end);
212
213 start &= PUD_MASK;
214 if (start < floor)
215 return;
216 if (ceiling) {
217 ceiling &= PUD_MASK;
218 if (!ceiling)
219 return;
220 }
221 if (end - 1 > ceiling - 1)
222 return;
223
224 pmd = pmd_offset(pud, start);
225 pud_clear(pud);
226 pmd_free_tlb(tlb, pmd, start);
227 mm_dec_nr_pmds(tlb->mm);
228}
229
230static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
231 unsigned long addr, unsigned long end,
232 unsigned long floor, unsigned long ceiling)
233{
234 pud_t *pud;
235 unsigned long next;
236 unsigned long start;
237
238 start = addr;
239 pud = pud_offset(p4d, addr);
240 do {
241 next = pud_addr_end(addr, end);
242 if (pud_none_or_clear_bad(pud))
243 continue;
244 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
245 } while (pud++, addr = next, addr != end);
246
247 start &= P4D_MASK;
248 if (start < floor)
249 return;
250 if (ceiling) {
251 ceiling &= P4D_MASK;
252 if (!ceiling)
253 return;
254 }
255 if (end - 1 > ceiling - 1)
256 return;
257
258 pud = pud_offset(p4d, start);
259 p4d_clear(p4d);
260 pud_free_tlb(tlb, pud, start);
261 mm_dec_nr_puds(tlb->mm);
262}
263
264static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
265 unsigned long addr, unsigned long end,
266 unsigned long floor, unsigned long ceiling)
267{
268 p4d_t *p4d;
269 unsigned long next;
270 unsigned long start;
271
272 start = addr;
273 p4d = p4d_offset(pgd, addr);
274 do {
275 next = p4d_addr_end(addr, end);
276 if (p4d_none_or_clear_bad(p4d))
277 continue;
278 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
279 } while (p4d++, addr = next, addr != end);
280
281 start &= PGDIR_MASK;
282 if (start < floor)
283 return;
284 if (ceiling) {
285 ceiling &= PGDIR_MASK;
286 if (!ceiling)
287 return;
288 }
289 if (end - 1 > ceiling - 1)
290 return;
291
292 p4d = p4d_offset(pgd, start);
293 pgd_clear(pgd);
294 p4d_free_tlb(tlb, p4d, start);
295}
296
297/*
298 * This function frees user-level page tables of a process.
299 */
300void free_pgd_range(struct mmu_gather *tlb,
301 unsigned long addr, unsigned long end,
302 unsigned long floor, unsigned long ceiling)
303{
304 pgd_t *pgd;
305 unsigned long next;
306
307 /*
308 * The next few lines have given us lots of grief...
309 *
310 * Why are we testing PMD* at this top level? Because often
311 * there will be no work to do at all, and we'd prefer not to
312 * go all the way down to the bottom just to discover that.
313 *
314 * Why all these "- 1"s? Because 0 represents both the bottom
315 * of the address space and the top of it (using -1 for the
316 * top wouldn't help much: the masks would do the wrong thing).
317 * The rule is that addr 0 and floor 0 refer to the bottom of
318 * the address space, but end 0 and ceiling 0 refer to the top
319 * Comparisons need to use "end - 1" and "ceiling - 1" (though
320 * that end 0 case should be mythical).
321 *
322 * Wherever addr is brought up or ceiling brought down, we must
323 * be careful to reject "the opposite 0" before it confuses the
324 * subsequent tests. But what about where end is brought down
325 * by PMD_SIZE below? no, end can't go down to 0 there.
326 *
327 * Whereas we round start (addr) and ceiling down, by different
328 * masks at different levels, in order to test whether a table
329 * now has no other vmas using it, so can be freed, we don't
330 * bother to round floor or end up - the tests don't need that.
331 */
332
333 addr &= PMD_MASK;
334 if (addr < floor) {
335 addr += PMD_SIZE;
336 if (!addr)
337 return;
338 }
339 if (ceiling) {
340 ceiling &= PMD_MASK;
341 if (!ceiling)
342 return;
343 }
344 if (end - 1 > ceiling - 1)
345 end -= PMD_SIZE;
346 if (addr > end - 1)
347 return;
348 /*
349 * We add page table cache pages with PAGE_SIZE,
350 * (see pte_free_tlb()), flush the tlb if we need
351 */
352 tlb_change_page_size(tlb, PAGE_SIZE);
353 pgd = pgd_offset(tlb->mm, addr);
354 do {
355 next = pgd_addr_end(addr, end);
356 if (pgd_none_or_clear_bad(pgd))
357 continue;
358 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
359 } while (pgd++, addr = next, addr != end);
360}
361
362void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
363 struct vm_area_struct *vma, unsigned long floor,
364 unsigned long ceiling, bool mm_wr_locked)
365{
366 do {
367 unsigned long addr = vma->vm_start;
368 struct vm_area_struct *next;
369
370 /*
371 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
372 * be 0. This will underflow and is okay.
373 */
374 next = mas_find(mas, ceiling - 1);
375 if (unlikely(xa_is_zero(next)))
376 next = NULL;
377
378 /*
379 * Hide vma from rmap and truncate_pagecache before freeing
380 * pgtables
381 */
382 if (mm_wr_locked)
383 vma_start_write(vma);
384 unlink_anon_vmas(vma);
385 unlink_file_vma(vma);
386
387 if (is_vm_hugetlb_page(vma)) {
388 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
389 floor, next ? next->vm_start : ceiling);
390 } else {
391 /*
392 * Optimization: gather nearby vmas into one call down
393 */
394 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
395 && !is_vm_hugetlb_page(next)) {
396 vma = next;
397 next = mas_find(mas, ceiling - 1);
398 if (unlikely(xa_is_zero(next)))
399 next = NULL;
400 if (mm_wr_locked)
401 vma_start_write(vma);
402 unlink_anon_vmas(vma);
403 unlink_file_vma(vma);
404 }
405 free_pgd_range(tlb, addr, vma->vm_end,
406 floor, next ? next->vm_start : ceiling);
407 }
408 vma = next;
409 } while (vma);
410}
411
412void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
413{
414 spinlock_t *ptl = pmd_lock(mm, pmd);
415
416 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
417 mm_inc_nr_ptes(mm);
418 /*
419 * Ensure all pte setup (eg. pte page lock and page clearing) are
420 * visible before the pte is made visible to other CPUs by being
421 * put into page tables.
422 *
423 * The other side of the story is the pointer chasing in the page
424 * table walking code (when walking the page table without locking;
425 * ie. most of the time). Fortunately, these data accesses consist
426 * of a chain of data-dependent loads, meaning most CPUs (alpha
427 * being the notable exception) will already guarantee loads are
428 * seen in-order. See the alpha page table accessors for the
429 * smp_rmb() barriers in page table walking code.
430 */
431 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
432 pmd_populate(mm, pmd, *pte);
433 *pte = NULL;
434 }
435 spin_unlock(ptl);
436}
437
438int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
439{
440 pgtable_t new = pte_alloc_one(mm);
441 if (!new)
442 return -ENOMEM;
443
444 pmd_install(mm, pmd, &new);
445 if (new)
446 pte_free(mm, new);
447 return 0;
448}
449
450int __pte_alloc_kernel(pmd_t *pmd)
451{
452 pte_t *new = pte_alloc_one_kernel(&init_mm);
453 if (!new)
454 return -ENOMEM;
455
456 spin_lock(&init_mm.page_table_lock);
457 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
458 smp_wmb(); /* See comment in pmd_install() */
459 pmd_populate_kernel(&init_mm, pmd, new);
460 new = NULL;
461 }
462 spin_unlock(&init_mm.page_table_lock);
463 if (new)
464 pte_free_kernel(&init_mm, new);
465 return 0;
466}
467
468static inline void init_rss_vec(int *rss)
469{
470 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
471}
472
473static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
474{
475 int i;
476
477 for (i = 0; i < NR_MM_COUNTERS; i++)
478 if (rss[i])
479 add_mm_counter(mm, i, rss[i]);
480}
481
482/*
483 * This function is called to print an error when a bad pte
484 * is found. For example, we might have a PFN-mapped pte in
485 * a region that doesn't allow it.
486 *
487 * The calling function must still handle the error.
488 */
489static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
490 pte_t pte, struct page *page)
491{
492 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
493 p4d_t *p4d = p4d_offset(pgd, addr);
494 pud_t *pud = pud_offset(p4d, addr);
495 pmd_t *pmd = pmd_offset(pud, addr);
496 struct address_space *mapping;
497 pgoff_t index;
498 static unsigned long resume;
499 static unsigned long nr_shown;
500 static unsigned long nr_unshown;
501
502 /*
503 * Allow a burst of 60 reports, then keep quiet for that minute;
504 * or allow a steady drip of one report per second.
505 */
506 if (nr_shown == 60) {
507 if (time_before(jiffies, resume)) {
508 nr_unshown++;
509 return;
510 }
511 if (nr_unshown) {
512 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
513 nr_unshown);
514 nr_unshown = 0;
515 }
516 nr_shown = 0;
517 }
518 if (nr_shown++ == 0)
519 resume = jiffies + 60 * HZ;
520
521 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
522 index = linear_page_index(vma, addr);
523
524 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
525 current->comm,
526 (long long)pte_val(pte), (long long)pmd_val(*pmd));
527 if (page)
528 dump_page(page, "bad pte");
529 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
530 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
531 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
532 vma->vm_file,
533 vma->vm_ops ? vma->vm_ops->fault : NULL,
534 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
535 mapping ? mapping->a_ops->read_folio : NULL);
536 dump_stack();
537 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
538}
539
540/*
541 * vm_normal_page -- This function gets the "struct page" associated with a pte.
542 *
543 * "Special" mappings do not wish to be associated with a "struct page" (either
544 * it doesn't exist, or it exists but they don't want to touch it). In this
545 * case, NULL is returned here. "Normal" mappings do have a struct page.
546 *
547 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
548 * pte bit, in which case this function is trivial. Secondly, an architecture
549 * may not have a spare pte bit, which requires a more complicated scheme,
550 * described below.
551 *
552 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
553 * special mapping (even if there are underlying and valid "struct pages").
554 * COWed pages of a VM_PFNMAP are always normal.
555 *
556 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
557 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
558 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
559 * mapping will always honor the rule
560 *
561 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
562 *
563 * And for normal mappings this is false.
564 *
565 * This restricts such mappings to be a linear translation from virtual address
566 * to pfn. To get around this restriction, we allow arbitrary mappings so long
567 * as the vma is not a COW mapping; in that case, we know that all ptes are
568 * special (because none can have been COWed).
569 *
570 *
571 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
572 *
573 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
574 * page" backing, however the difference is that _all_ pages with a struct
575 * page (that is, those where pfn_valid is true) are refcounted and considered
576 * normal pages by the VM. The disadvantage is that pages are refcounted
577 * (which can be slower and simply not an option for some PFNMAP users). The
578 * advantage is that we don't have to follow the strict linearity rule of
579 * PFNMAP mappings in order to support COWable mappings.
580 *
581 */
582struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
583 pte_t pte)
584{
585 unsigned long pfn = pte_pfn(pte);
586
587 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
588 if (likely(!pte_special(pte)))
589 goto check_pfn;
590 if (vma->vm_ops && vma->vm_ops->find_special_page)
591 return vma->vm_ops->find_special_page(vma, addr);
592 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
593 return NULL;
594 if (is_zero_pfn(pfn))
595 return NULL;
596 if (pte_devmap(pte))
597 /*
598 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
599 * and will have refcounts incremented on their struct pages
600 * when they are inserted into PTEs, thus they are safe to
601 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
602 * do not have refcounts. Example of legacy ZONE_DEVICE is
603 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
604 */
605 return NULL;
606
607 print_bad_pte(vma, addr, pte, NULL);
608 return NULL;
609 }
610
611 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
612
613 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
614 if (vma->vm_flags & VM_MIXEDMAP) {
615 if (!pfn_valid(pfn))
616 return NULL;
617 goto out;
618 } else {
619 unsigned long off;
620 off = (addr - vma->vm_start) >> PAGE_SHIFT;
621 if (pfn == vma->vm_pgoff + off)
622 return NULL;
623 if (!is_cow_mapping(vma->vm_flags))
624 return NULL;
625 }
626 }
627
628 if (is_zero_pfn(pfn))
629 return NULL;
630
631check_pfn:
632 if (unlikely(pfn > highest_memmap_pfn)) {
633 print_bad_pte(vma, addr, pte, NULL);
634 return NULL;
635 }
636
637 /*
638 * NOTE! We still have PageReserved() pages in the page tables.
639 * eg. VDSO mappings can cause them to exist.
640 */
641out:
642 return pfn_to_page(pfn);
643}
644
645struct folio *vm_normal_folio(struct vm_area_struct *vma, unsigned long addr,
646 pte_t pte)
647{
648 struct page *page = vm_normal_page(vma, addr, pte);
649
650 if (page)
651 return page_folio(page);
652 return NULL;
653}
654
655#ifdef CONFIG_TRANSPARENT_HUGEPAGE
656struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
657 pmd_t pmd)
658{
659 unsigned long pfn = pmd_pfn(pmd);
660
661 /*
662 * There is no pmd_special() but there may be special pmds, e.g.
663 * in a direct-access (dax) mapping, so let's just replicate the
664 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
665 */
666 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
667 if (vma->vm_flags & VM_MIXEDMAP) {
668 if (!pfn_valid(pfn))
669 return NULL;
670 goto out;
671 } else {
672 unsigned long off;
673 off = (addr - vma->vm_start) >> PAGE_SHIFT;
674 if (pfn == vma->vm_pgoff + off)
675 return NULL;
676 if (!is_cow_mapping(vma->vm_flags))
677 return NULL;
678 }
679 }
680
681 if (pmd_devmap(pmd))
682 return NULL;
683 if (is_huge_zero_pmd(pmd))
684 return NULL;
685 if (unlikely(pfn > highest_memmap_pfn))
686 return NULL;
687
688 /*
689 * NOTE! We still have PageReserved() pages in the page tables.
690 * eg. VDSO mappings can cause them to exist.
691 */
692out:
693 return pfn_to_page(pfn);
694}
695
696struct folio *vm_normal_folio_pmd(struct vm_area_struct *vma,
697 unsigned long addr, pmd_t pmd)
698{
699 struct page *page = vm_normal_page_pmd(vma, addr, pmd);
700
701 if (page)
702 return page_folio(page);
703 return NULL;
704}
705#endif
706
707static void restore_exclusive_pte(struct vm_area_struct *vma,
708 struct page *page, unsigned long address,
709 pte_t *ptep)
710{
711 struct folio *folio = page_folio(page);
712 pte_t orig_pte;
713 pte_t pte;
714 swp_entry_t entry;
715
716 orig_pte = ptep_get(ptep);
717 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
718 if (pte_swp_soft_dirty(orig_pte))
719 pte = pte_mksoft_dirty(pte);
720
721 entry = pte_to_swp_entry(orig_pte);
722 if (pte_swp_uffd_wp(orig_pte))
723 pte = pte_mkuffd_wp(pte);
724 else if (is_writable_device_exclusive_entry(entry))
725 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
726
727 VM_BUG_ON_FOLIO(pte_write(pte) && (!folio_test_anon(folio) &&
728 PageAnonExclusive(page)), folio);
729
730 /*
731 * No need to take a page reference as one was already
732 * created when the swap entry was made.
733 */
734 if (folio_test_anon(folio))
735 folio_add_anon_rmap_pte(folio, page, vma, address, RMAP_NONE);
736 else
737 /*
738 * Currently device exclusive access only supports anonymous
739 * memory so the entry shouldn't point to a filebacked page.
740 */
741 WARN_ON_ONCE(1);
742
743 set_pte_at(vma->vm_mm, address, ptep, pte);
744
745 /*
746 * No need to invalidate - it was non-present before. However
747 * secondary CPUs may have mappings that need invalidating.
748 */
749 update_mmu_cache(vma, address, ptep);
750}
751
752/*
753 * Tries to restore an exclusive pte if the page lock can be acquired without
754 * sleeping.
755 */
756static int
757try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
758 unsigned long addr)
759{
760 swp_entry_t entry = pte_to_swp_entry(ptep_get(src_pte));
761 struct page *page = pfn_swap_entry_to_page(entry);
762
763 if (trylock_page(page)) {
764 restore_exclusive_pte(vma, page, addr, src_pte);
765 unlock_page(page);
766 return 0;
767 }
768
769 return -EBUSY;
770}
771
772/*
773 * copy one vm_area from one task to the other. Assumes the page tables
774 * already present in the new task to be cleared in the whole range
775 * covered by this vma.
776 */
777
778static unsigned long
779copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
780 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
781 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
782{
783 unsigned long vm_flags = dst_vma->vm_flags;
784 pte_t orig_pte = ptep_get(src_pte);
785 pte_t pte = orig_pte;
786 struct folio *folio;
787 struct page *page;
788 swp_entry_t entry = pte_to_swp_entry(orig_pte);
789
790 if (likely(!non_swap_entry(entry))) {
791 if (swap_duplicate(entry) < 0)
792 return -EIO;
793
794 /* make sure dst_mm is on swapoff's mmlist. */
795 if (unlikely(list_empty(&dst_mm->mmlist))) {
796 spin_lock(&mmlist_lock);
797 if (list_empty(&dst_mm->mmlist))
798 list_add(&dst_mm->mmlist,
799 &src_mm->mmlist);
800 spin_unlock(&mmlist_lock);
801 }
802 /* Mark the swap entry as shared. */
803 if (pte_swp_exclusive(orig_pte)) {
804 pte = pte_swp_clear_exclusive(orig_pte);
805 set_pte_at(src_mm, addr, src_pte, pte);
806 }
807 rss[MM_SWAPENTS]++;
808 } else if (is_migration_entry(entry)) {
809 page = pfn_swap_entry_to_page(entry);
810
811 rss[mm_counter(page)]++;
812
813 if (!is_readable_migration_entry(entry) &&
814 is_cow_mapping(vm_flags)) {
815 /*
816 * COW mappings require pages in both parent and child
817 * to be set to read. A previously exclusive entry is
818 * now shared.
819 */
820 entry = make_readable_migration_entry(
821 swp_offset(entry));
822 pte = swp_entry_to_pte(entry);
823 if (pte_swp_soft_dirty(orig_pte))
824 pte = pte_swp_mksoft_dirty(pte);
825 if (pte_swp_uffd_wp(orig_pte))
826 pte = pte_swp_mkuffd_wp(pte);
827 set_pte_at(src_mm, addr, src_pte, pte);
828 }
829 } else if (is_device_private_entry(entry)) {
830 page = pfn_swap_entry_to_page(entry);
831 folio = page_folio(page);
832
833 /*
834 * Update rss count even for unaddressable pages, as
835 * they should treated just like normal pages in this
836 * respect.
837 *
838 * We will likely want to have some new rss counters
839 * for unaddressable pages, at some point. But for now
840 * keep things as they are.
841 */
842 folio_get(folio);
843 rss[mm_counter(page)]++;
844 /* Cannot fail as these pages cannot get pinned. */
845 folio_try_dup_anon_rmap_pte(folio, page, src_vma);
846
847 /*
848 * We do not preserve soft-dirty information, because so
849 * far, checkpoint/restore is the only feature that
850 * requires that. And checkpoint/restore does not work
851 * when a device driver is involved (you cannot easily
852 * save and restore device driver state).
853 */
854 if (is_writable_device_private_entry(entry) &&
855 is_cow_mapping(vm_flags)) {
856 entry = make_readable_device_private_entry(
857 swp_offset(entry));
858 pte = swp_entry_to_pte(entry);
859 if (pte_swp_uffd_wp(orig_pte))
860 pte = pte_swp_mkuffd_wp(pte);
861 set_pte_at(src_mm, addr, src_pte, pte);
862 }
863 } else if (is_device_exclusive_entry(entry)) {
864 /*
865 * Make device exclusive entries present by restoring the
866 * original entry then copying as for a present pte. Device
867 * exclusive entries currently only support private writable
868 * (ie. COW) mappings.
869 */
870 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
871 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
872 return -EBUSY;
873 return -ENOENT;
874 } else if (is_pte_marker_entry(entry)) {
875 pte_marker marker = copy_pte_marker(entry, dst_vma);
876
877 if (marker)
878 set_pte_at(dst_mm, addr, dst_pte,
879 make_pte_marker(marker));
880 return 0;
881 }
882 if (!userfaultfd_wp(dst_vma))
883 pte = pte_swp_clear_uffd_wp(pte);
884 set_pte_at(dst_mm, addr, dst_pte, pte);
885 return 0;
886}
887
888/*
889 * Copy a present and normal page.
890 *
891 * NOTE! The usual case is that this isn't required;
892 * instead, the caller can just increase the page refcount
893 * and re-use the pte the traditional way.
894 *
895 * And if we need a pre-allocated page but don't yet have
896 * one, return a negative error to let the preallocation
897 * code know so that it can do so outside the page table
898 * lock.
899 */
900static inline int
901copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
902 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
903 struct folio **prealloc, struct page *page)
904{
905 struct folio *new_folio;
906 pte_t pte;
907
908 new_folio = *prealloc;
909 if (!new_folio)
910 return -EAGAIN;
911
912 /*
913 * We have a prealloc page, all good! Take it
914 * over and copy the page & arm it.
915 */
916 *prealloc = NULL;
917 copy_user_highpage(&new_folio->page, page, addr, src_vma);
918 __folio_mark_uptodate(new_folio);
919 folio_add_new_anon_rmap(new_folio, dst_vma, addr);
920 folio_add_lru_vma(new_folio, dst_vma);
921 rss[MM_ANONPAGES]++;
922
923 /* All done, just insert the new page copy in the child */
924 pte = mk_pte(&new_folio->page, dst_vma->vm_page_prot);
925 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
926 if (userfaultfd_pte_wp(dst_vma, ptep_get(src_pte)))
927 /* Uffd-wp needs to be delivered to dest pte as well */
928 pte = pte_mkuffd_wp(pte);
929 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
930 return 0;
931}
932
933/*
934 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
935 * is required to copy this pte.
936 */
937static inline int
938copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
939 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
940 struct folio **prealloc)
941{
942 struct mm_struct *src_mm = src_vma->vm_mm;
943 unsigned long vm_flags = src_vma->vm_flags;
944 pte_t pte = ptep_get(src_pte);
945 struct page *page;
946 struct folio *folio;
947
948 page = vm_normal_page(src_vma, addr, pte);
949 if (page)
950 folio = page_folio(page);
951 if (page && folio_test_anon(folio)) {
952 /*
953 * If this page may have been pinned by the parent process,
954 * copy the page immediately for the child so that we'll always
955 * guarantee the pinned page won't be randomly replaced in the
956 * future.
957 */
958 folio_get(folio);
959 if (unlikely(folio_try_dup_anon_rmap_pte(folio, page, src_vma))) {
960 /* Page may be pinned, we have to copy. */
961 folio_put(folio);
962 return copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
963 addr, rss, prealloc, page);
964 }
965 rss[MM_ANONPAGES]++;
966 } else if (page) {
967 folio_get(folio);
968 folio_dup_file_rmap_pte(folio, page);
969 rss[mm_counter_file(page)]++;
970 }
971
972 /*
973 * If it's a COW mapping, write protect it both
974 * in the parent and the child
975 */
976 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
977 ptep_set_wrprotect(src_mm, addr, src_pte);
978 pte = pte_wrprotect(pte);
979 }
980 VM_BUG_ON(page && folio_test_anon(folio) && PageAnonExclusive(page));
981
982 /*
983 * If it's a shared mapping, mark it clean in
984 * the child
985 */
986 if (vm_flags & VM_SHARED)
987 pte = pte_mkclean(pte);
988 pte = pte_mkold(pte);
989
990 if (!userfaultfd_wp(dst_vma))
991 pte = pte_clear_uffd_wp(pte);
992
993 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
994 return 0;
995}
996
997static inline struct folio *folio_prealloc(struct mm_struct *src_mm,
998 struct vm_area_struct *vma, unsigned long addr, bool need_zero)
999{
1000 struct folio *new_folio;
1001
1002 if (need_zero)
1003 new_folio = vma_alloc_zeroed_movable_folio(vma, addr);
1004 else
1005 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma,
1006 addr, false);
1007
1008 if (!new_folio)
1009 return NULL;
1010
1011 if (mem_cgroup_charge(new_folio, src_mm, GFP_KERNEL)) {
1012 folio_put(new_folio);
1013 return NULL;
1014 }
1015 folio_throttle_swaprate(new_folio, GFP_KERNEL);
1016
1017 return new_folio;
1018}
1019
1020static int
1021copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1022 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1023 unsigned long end)
1024{
1025 struct mm_struct *dst_mm = dst_vma->vm_mm;
1026 struct mm_struct *src_mm = src_vma->vm_mm;
1027 pte_t *orig_src_pte, *orig_dst_pte;
1028 pte_t *src_pte, *dst_pte;
1029 pte_t ptent;
1030 spinlock_t *src_ptl, *dst_ptl;
1031 int progress, ret = 0;
1032 int rss[NR_MM_COUNTERS];
1033 swp_entry_t entry = (swp_entry_t){0};
1034 struct folio *prealloc = NULL;
1035
1036again:
1037 progress = 0;
1038 init_rss_vec(rss);
1039
1040 /*
1041 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1042 * error handling here, assume that exclusive mmap_lock on dst and src
1043 * protects anon from unexpected THP transitions; with shmem and file
1044 * protected by mmap_lock-less collapse skipping areas with anon_vma
1045 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1046 * can remove such assumptions later, but this is good enough for now.
1047 */
1048 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1049 if (!dst_pte) {
1050 ret = -ENOMEM;
1051 goto out;
1052 }
1053 src_pte = pte_offset_map_nolock(src_mm, src_pmd, addr, &src_ptl);
1054 if (!src_pte) {
1055 pte_unmap_unlock(dst_pte, dst_ptl);
1056 /* ret == 0 */
1057 goto out;
1058 }
1059 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1060 orig_src_pte = src_pte;
1061 orig_dst_pte = dst_pte;
1062 arch_enter_lazy_mmu_mode();
1063
1064 do {
1065 /*
1066 * We are holding two locks at this point - either of them
1067 * could generate latencies in another task on another CPU.
1068 */
1069 if (progress >= 32) {
1070 progress = 0;
1071 if (need_resched() ||
1072 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1073 break;
1074 }
1075 ptent = ptep_get(src_pte);
1076 if (pte_none(ptent)) {
1077 progress++;
1078 continue;
1079 }
1080 if (unlikely(!pte_present(ptent))) {
1081 ret = copy_nonpresent_pte(dst_mm, src_mm,
1082 dst_pte, src_pte,
1083 dst_vma, src_vma,
1084 addr, rss);
1085 if (ret == -EIO) {
1086 entry = pte_to_swp_entry(ptep_get(src_pte));
1087 break;
1088 } else if (ret == -EBUSY) {
1089 break;
1090 } else if (!ret) {
1091 progress += 8;
1092 continue;
1093 }
1094
1095 /*
1096 * Device exclusive entry restored, continue by copying
1097 * the now present pte.
1098 */
1099 WARN_ON_ONCE(ret != -ENOENT);
1100 }
1101 /* copy_present_pte() will clear `*prealloc' if consumed */
1102 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1103 addr, rss, &prealloc);
1104 /*
1105 * If we need a pre-allocated page for this pte, drop the
1106 * locks, allocate, and try again.
1107 */
1108 if (unlikely(ret == -EAGAIN))
1109 break;
1110 if (unlikely(prealloc)) {
1111 /*
1112 * pre-alloc page cannot be reused by next time so as
1113 * to strictly follow mempolicy (e.g., alloc_page_vma()
1114 * will allocate page according to address). This
1115 * could only happen if one pinned pte changed.
1116 */
1117 folio_put(prealloc);
1118 prealloc = NULL;
1119 }
1120 progress += 8;
1121 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1122
1123 arch_leave_lazy_mmu_mode();
1124 pte_unmap_unlock(orig_src_pte, src_ptl);
1125 add_mm_rss_vec(dst_mm, rss);
1126 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1127 cond_resched();
1128
1129 if (ret == -EIO) {
1130 VM_WARN_ON_ONCE(!entry.val);
1131 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1132 ret = -ENOMEM;
1133 goto out;
1134 }
1135 entry.val = 0;
1136 } else if (ret == -EBUSY) {
1137 goto out;
1138 } else if (ret == -EAGAIN) {
1139 prealloc = folio_prealloc(src_mm, src_vma, addr, false);
1140 if (!prealloc)
1141 return -ENOMEM;
1142 } else if (ret) {
1143 VM_WARN_ON_ONCE(1);
1144 }
1145
1146 /* We've captured and resolved the error. Reset, try again. */
1147 ret = 0;
1148
1149 if (addr != end)
1150 goto again;
1151out:
1152 if (unlikely(prealloc))
1153 folio_put(prealloc);
1154 return ret;
1155}
1156
1157static inline int
1158copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1159 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1160 unsigned long end)
1161{
1162 struct mm_struct *dst_mm = dst_vma->vm_mm;
1163 struct mm_struct *src_mm = src_vma->vm_mm;
1164 pmd_t *src_pmd, *dst_pmd;
1165 unsigned long next;
1166
1167 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1168 if (!dst_pmd)
1169 return -ENOMEM;
1170 src_pmd = pmd_offset(src_pud, addr);
1171 do {
1172 next = pmd_addr_end(addr, end);
1173 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1174 || pmd_devmap(*src_pmd)) {
1175 int err;
1176 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1177 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1178 addr, dst_vma, src_vma);
1179 if (err == -ENOMEM)
1180 return -ENOMEM;
1181 if (!err)
1182 continue;
1183 /* fall through */
1184 }
1185 if (pmd_none_or_clear_bad(src_pmd))
1186 continue;
1187 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1188 addr, next))
1189 return -ENOMEM;
1190 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1191 return 0;
1192}
1193
1194static inline int
1195copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1196 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1197 unsigned long end)
1198{
1199 struct mm_struct *dst_mm = dst_vma->vm_mm;
1200 struct mm_struct *src_mm = src_vma->vm_mm;
1201 pud_t *src_pud, *dst_pud;
1202 unsigned long next;
1203
1204 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1205 if (!dst_pud)
1206 return -ENOMEM;
1207 src_pud = pud_offset(src_p4d, addr);
1208 do {
1209 next = pud_addr_end(addr, end);
1210 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1211 int err;
1212
1213 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1214 err = copy_huge_pud(dst_mm, src_mm,
1215 dst_pud, src_pud, addr, src_vma);
1216 if (err == -ENOMEM)
1217 return -ENOMEM;
1218 if (!err)
1219 continue;
1220 /* fall through */
1221 }
1222 if (pud_none_or_clear_bad(src_pud))
1223 continue;
1224 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1225 addr, next))
1226 return -ENOMEM;
1227 } while (dst_pud++, src_pud++, addr = next, addr != end);
1228 return 0;
1229}
1230
1231static inline int
1232copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1233 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1234 unsigned long end)
1235{
1236 struct mm_struct *dst_mm = dst_vma->vm_mm;
1237 p4d_t *src_p4d, *dst_p4d;
1238 unsigned long next;
1239
1240 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1241 if (!dst_p4d)
1242 return -ENOMEM;
1243 src_p4d = p4d_offset(src_pgd, addr);
1244 do {
1245 next = p4d_addr_end(addr, end);
1246 if (p4d_none_or_clear_bad(src_p4d))
1247 continue;
1248 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1249 addr, next))
1250 return -ENOMEM;
1251 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1252 return 0;
1253}
1254
1255/*
1256 * Return true if the vma needs to copy the pgtable during this fork(). Return
1257 * false when we can speed up fork() by allowing lazy page faults later until
1258 * when the child accesses the memory range.
1259 */
1260static bool
1261vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1262{
1263 /*
1264 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1265 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1266 * contains uffd-wp protection information, that's something we can't
1267 * retrieve from page cache, and skip copying will lose those info.
1268 */
1269 if (userfaultfd_wp(dst_vma))
1270 return true;
1271
1272 if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
1273 return true;
1274
1275 if (src_vma->anon_vma)
1276 return true;
1277
1278 /*
1279 * Don't copy ptes where a page fault will fill them correctly. Fork
1280 * becomes much lighter when there are big shared or private readonly
1281 * mappings. The tradeoff is that copy_page_range is more efficient
1282 * than faulting.
1283 */
1284 return false;
1285}
1286
1287int
1288copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1289{
1290 pgd_t *src_pgd, *dst_pgd;
1291 unsigned long next;
1292 unsigned long addr = src_vma->vm_start;
1293 unsigned long end = src_vma->vm_end;
1294 struct mm_struct *dst_mm = dst_vma->vm_mm;
1295 struct mm_struct *src_mm = src_vma->vm_mm;
1296 struct mmu_notifier_range range;
1297 bool is_cow;
1298 int ret;
1299
1300 if (!vma_needs_copy(dst_vma, src_vma))
1301 return 0;
1302
1303 if (is_vm_hugetlb_page(src_vma))
1304 return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
1305
1306 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1307 /*
1308 * We do not free on error cases below as remove_vma
1309 * gets called on error from higher level routine
1310 */
1311 ret = track_pfn_copy(src_vma);
1312 if (ret)
1313 return ret;
1314 }
1315
1316 /*
1317 * We need to invalidate the secondary MMU mappings only when
1318 * there could be a permission downgrade on the ptes of the
1319 * parent mm. And a permission downgrade will only happen if
1320 * is_cow_mapping() returns true.
1321 */
1322 is_cow = is_cow_mapping(src_vma->vm_flags);
1323
1324 if (is_cow) {
1325 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1326 0, src_mm, addr, end);
1327 mmu_notifier_invalidate_range_start(&range);
1328 /*
1329 * Disabling preemption is not needed for the write side, as
1330 * the read side doesn't spin, but goes to the mmap_lock.
1331 *
1332 * Use the raw variant of the seqcount_t write API to avoid
1333 * lockdep complaining about preemptibility.
1334 */
1335 vma_assert_write_locked(src_vma);
1336 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1337 }
1338
1339 ret = 0;
1340 dst_pgd = pgd_offset(dst_mm, addr);
1341 src_pgd = pgd_offset(src_mm, addr);
1342 do {
1343 next = pgd_addr_end(addr, end);
1344 if (pgd_none_or_clear_bad(src_pgd))
1345 continue;
1346 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1347 addr, next))) {
1348 untrack_pfn_clear(dst_vma);
1349 ret = -ENOMEM;
1350 break;
1351 }
1352 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1353
1354 if (is_cow) {
1355 raw_write_seqcount_end(&src_mm->write_protect_seq);
1356 mmu_notifier_invalidate_range_end(&range);
1357 }
1358 return ret;
1359}
1360
1361/* Whether we should zap all COWed (private) pages too */
1362static inline bool should_zap_cows(struct zap_details *details)
1363{
1364 /* By default, zap all pages */
1365 if (!details)
1366 return true;
1367
1368 /* Or, we zap COWed pages only if the caller wants to */
1369 return details->even_cows;
1370}
1371
1372/* Decides whether we should zap this page with the page pointer specified */
1373static inline bool should_zap_page(struct zap_details *details, struct page *page)
1374{
1375 /* If we can make a decision without *page.. */
1376 if (should_zap_cows(details))
1377 return true;
1378
1379 /* E.g. the caller passes NULL for the case of a zero page */
1380 if (!page)
1381 return true;
1382
1383 /* Otherwise we should only zap non-anon pages */
1384 return !PageAnon(page);
1385}
1386
1387static inline bool zap_drop_file_uffd_wp(struct zap_details *details)
1388{
1389 if (!details)
1390 return false;
1391
1392 return details->zap_flags & ZAP_FLAG_DROP_MARKER;
1393}
1394
1395/*
1396 * This function makes sure that we'll replace the none pte with an uffd-wp
1397 * swap special pte marker when necessary. Must be with the pgtable lock held.
1398 */
1399static inline void
1400zap_install_uffd_wp_if_needed(struct vm_area_struct *vma,
1401 unsigned long addr, pte_t *pte,
1402 struct zap_details *details, pte_t pteval)
1403{
1404 /* Zap on anonymous always means dropping everything */
1405 if (vma_is_anonymous(vma))
1406 return;
1407
1408 if (zap_drop_file_uffd_wp(details))
1409 return;
1410
1411 pte_install_uffd_wp_if_needed(vma, addr, pte, pteval);
1412}
1413
1414static unsigned long zap_pte_range(struct mmu_gather *tlb,
1415 struct vm_area_struct *vma, pmd_t *pmd,
1416 unsigned long addr, unsigned long end,
1417 struct zap_details *details)
1418{
1419 struct mm_struct *mm = tlb->mm;
1420 int force_flush = 0;
1421 int rss[NR_MM_COUNTERS];
1422 spinlock_t *ptl;
1423 pte_t *start_pte;
1424 pte_t *pte;
1425 swp_entry_t entry;
1426
1427 tlb_change_page_size(tlb, PAGE_SIZE);
1428 init_rss_vec(rss);
1429 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1430 if (!pte)
1431 return addr;
1432
1433 flush_tlb_batched_pending(mm);
1434 arch_enter_lazy_mmu_mode();
1435 do {
1436 pte_t ptent = ptep_get(pte);
1437 struct folio *folio;
1438 struct page *page;
1439
1440 if (pte_none(ptent))
1441 continue;
1442
1443 if (need_resched())
1444 break;
1445
1446 if (pte_present(ptent)) {
1447 unsigned int delay_rmap;
1448
1449 page = vm_normal_page(vma, addr, ptent);
1450 if (unlikely(!should_zap_page(details, page)))
1451 continue;
1452 ptent = ptep_get_and_clear_full(mm, addr, pte,
1453 tlb->fullmm);
1454 arch_check_zapped_pte(vma, ptent);
1455 tlb_remove_tlb_entry(tlb, pte, addr);
1456 zap_install_uffd_wp_if_needed(vma, addr, pte, details,
1457 ptent);
1458 if (unlikely(!page)) {
1459 ksm_might_unmap_zero_page(mm, ptent);
1460 continue;
1461 }
1462
1463 folio = page_folio(page);
1464 delay_rmap = 0;
1465 if (!folio_test_anon(folio)) {
1466 if (pte_dirty(ptent)) {
1467 folio_mark_dirty(folio);
1468 if (tlb_delay_rmap(tlb)) {
1469 delay_rmap = 1;
1470 force_flush = 1;
1471 }
1472 }
1473 if (pte_young(ptent) && likely(vma_has_recency(vma)))
1474 folio_mark_accessed(folio);
1475 }
1476 rss[mm_counter(page)]--;
1477 if (!delay_rmap) {
1478 folio_remove_rmap_pte(folio, page, vma);
1479 if (unlikely(page_mapcount(page) < 0))
1480 print_bad_pte(vma, addr, ptent, page);
1481 }
1482 if (unlikely(__tlb_remove_page(tlb, page, delay_rmap))) {
1483 force_flush = 1;
1484 addr += PAGE_SIZE;
1485 break;
1486 }
1487 continue;
1488 }
1489
1490 entry = pte_to_swp_entry(ptent);
1491 if (is_device_private_entry(entry) ||
1492 is_device_exclusive_entry(entry)) {
1493 page = pfn_swap_entry_to_page(entry);
1494 folio = page_folio(page);
1495 if (unlikely(!should_zap_page(details, page)))
1496 continue;
1497 /*
1498 * Both device private/exclusive mappings should only
1499 * work with anonymous page so far, so we don't need to
1500 * consider uffd-wp bit when zap. For more information,
1501 * see zap_install_uffd_wp_if_needed().
1502 */
1503 WARN_ON_ONCE(!vma_is_anonymous(vma));
1504 rss[mm_counter(page)]--;
1505 if (is_device_private_entry(entry))
1506 folio_remove_rmap_pte(folio, page, vma);
1507 folio_put(folio);
1508 } else if (!non_swap_entry(entry)) {
1509 /* Genuine swap entry, hence a private anon page */
1510 if (!should_zap_cows(details))
1511 continue;
1512 rss[MM_SWAPENTS]--;
1513 if (unlikely(!free_swap_and_cache(entry)))
1514 print_bad_pte(vma, addr, ptent, NULL);
1515 } else if (is_migration_entry(entry)) {
1516 page = pfn_swap_entry_to_page(entry);
1517 if (!should_zap_page(details, page))
1518 continue;
1519 rss[mm_counter(page)]--;
1520 } else if (pte_marker_entry_uffd_wp(entry)) {
1521 /*
1522 * For anon: always drop the marker; for file: only
1523 * drop the marker if explicitly requested.
1524 */
1525 if (!vma_is_anonymous(vma) &&
1526 !zap_drop_file_uffd_wp(details))
1527 continue;
1528 } else if (is_hwpoison_entry(entry) ||
1529 is_poisoned_swp_entry(entry)) {
1530 if (!should_zap_cows(details))
1531 continue;
1532 } else {
1533 /* We should have covered all the swap entry types */
1534 pr_alert("unrecognized swap entry 0x%lx\n", entry.val);
1535 WARN_ON_ONCE(1);
1536 }
1537 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1538 zap_install_uffd_wp_if_needed(vma, addr, pte, details, ptent);
1539 } while (pte++, addr += PAGE_SIZE, addr != end);
1540
1541 add_mm_rss_vec(mm, rss);
1542 arch_leave_lazy_mmu_mode();
1543
1544 /* Do the actual TLB flush before dropping ptl */
1545 if (force_flush) {
1546 tlb_flush_mmu_tlbonly(tlb);
1547 tlb_flush_rmaps(tlb, vma);
1548 }
1549 pte_unmap_unlock(start_pte, ptl);
1550
1551 /*
1552 * If we forced a TLB flush (either due to running out of
1553 * batch buffers or because we needed to flush dirty TLB
1554 * entries before releasing the ptl), free the batched
1555 * memory too. Come back again if we didn't do everything.
1556 */
1557 if (force_flush)
1558 tlb_flush_mmu(tlb);
1559
1560 return addr;
1561}
1562
1563static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1564 struct vm_area_struct *vma, pud_t *pud,
1565 unsigned long addr, unsigned long end,
1566 struct zap_details *details)
1567{
1568 pmd_t *pmd;
1569 unsigned long next;
1570
1571 pmd = pmd_offset(pud, addr);
1572 do {
1573 next = pmd_addr_end(addr, end);
1574 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1575 if (next - addr != HPAGE_PMD_SIZE)
1576 __split_huge_pmd(vma, pmd, addr, false, NULL);
1577 else if (zap_huge_pmd(tlb, vma, pmd, addr)) {
1578 addr = next;
1579 continue;
1580 }
1581 /* fall through */
1582 } else if (details && details->single_folio &&
1583 folio_test_pmd_mappable(details->single_folio) &&
1584 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1585 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1586 /*
1587 * Take and drop THP pmd lock so that we cannot return
1588 * prematurely, while zap_huge_pmd() has cleared *pmd,
1589 * but not yet decremented compound_mapcount().
1590 */
1591 spin_unlock(ptl);
1592 }
1593 if (pmd_none(*pmd)) {
1594 addr = next;
1595 continue;
1596 }
1597 addr = zap_pte_range(tlb, vma, pmd, addr, next, details);
1598 if (addr != next)
1599 pmd--;
1600 } while (pmd++, cond_resched(), addr != end);
1601
1602 return addr;
1603}
1604
1605static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1606 struct vm_area_struct *vma, p4d_t *p4d,
1607 unsigned long addr, unsigned long end,
1608 struct zap_details *details)
1609{
1610 pud_t *pud;
1611 unsigned long next;
1612
1613 pud = pud_offset(p4d, addr);
1614 do {
1615 next = pud_addr_end(addr, end);
1616 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1617 if (next - addr != HPAGE_PUD_SIZE) {
1618 mmap_assert_locked(tlb->mm);
1619 split_huge_pud(vma, pud, addr);
1620 } else if (zap_huge_pud(tlb, vma, pud, addr))
1621 goto next;
1622 /* fall through */
1623 }
1624 if (pud_none_or_clear_bad(pud))
1625 continue;
1626 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1627next:
1628 cond_resched();
1629 } while (pud++, addr = next, addr != end);
1630
1631 return addr;
1632}
1633
1634static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1635 struct vm_area_struct *vma, pgd_t *pgd,
1636 unsigned long addr, unsigned long end,
1637 struct zap_details *details)
1638{
1639 p4d_t *p4d;
1640 unsigned long next;
1641
1642 p4d = p4d_offset(pgd, addr);
1643 do {
1644 next = p4d_addr_end(addr, end);
1645 if (p4d_none_or_clear_bad(p4d))
1646 continue;
1647 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1648 } while (p4d++, addr = next, addr != end);
1649
1650 return addr;
1651}
1652
1653void unmap_page_range(struct mmu_gather *tlb,
1654 struct vm_area_struct *vma,
1655 unsigned long addr, unsigned long end,
1656 struct zap_details *details)
1657{
1658 pgd_t *pgd;
1659 unsigned long next;
1660
1661 BUG_ON(addr >= end);
1662 tlb_start_vma(tlb, vma);
1663 pgd = pgd_offset(vma->vm_mm, addr);
1664 do {
1665 next = pgd_addr_end(addr, end);
1666 if (pgd_none_or_clear_bad(pgd))
1667 continue;
1668 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1669 } while (pgd++, addr = next, addr != end);
1670 tlb_end_vma(tlb, vma);
1671}
1672
1673
1674static void unmap_single_vma(struct mmu_gather *tlb,
1675 struct vm_area_struct *vma, unsigned long start_addr,
1676 unsigned long end_addr,
1677 struct zap_details *details, bool mm_wr_locked)
1678{
1679 unsigned long start = max(vma->vm_start, start_addr);
1680 unsigned long end;
1681
1682 if (start >= vma->vm_end)
1683 return;
1684 end = min(vma->vm_end, end_addr);
1685 if (end <= vma->vm_start)
1686 return;
1687
1688 if (vma->vm_file)
1689 uprobe_munmap(vma, start, end);
1690
1691 if (unlikely(vma->vm_flags & VM_PFNMAP))
1692 untrack_pfn(vma, 0, 0, mm_wr_locked);
1693
1694 if (start != end) {
1695 if (unlikely(is_vm_hugetlb_page(vma))) {
1696 /*
1697 * It is undesirable to test vma->vm_file as it
1698 * should be non-null for valid hugetlb area.
1699 * However, vm_file will be NULL in the error
1700 * cleanup path of mmap_region. When
1701 * hugetlbfs ->mmap method fails,
1702 * mmap_region() nullifies vma->vm_file
1703 * before calling this function to clean up.
1704 * Since no pte has actually been setup, it is
1705 * safe to do nothing in this case.
1706 */
1707 if (vma->vm_file) {
1708 zap_flags_t zap_flags = details ?
1709 details->zap_flags : 0;
1710 __unmap_hugepage_range(tlb, vma, start, end,
1711 NULL, zap_flags);
1712 }
1713 } else
1714 unmap_page_range(tlb, vma, start, end, details);
1715 }
1716}
1717
1718/**
1719 * unmap_vmas - unmap a range of memory covered by a list of vma's
1720 * @tlb: address of the caller's struct mmu_gather
1721 * @mas: the maple state
1722 * @vma: the starting vma
1723 * @start_addr: virtual address at which to start unmapping
1724 * @end_addr: virtual address at which to end unmapping
1725 * @tree_end: The maximum index to check
1726 * @mm_wr_locked: lock flag
1727 *
1728 * Unmap all pages in the vma list.
1729 *
1730 * Only addresses between `start' and `end' will be unmapped.
1731 *
1732 * The VMA list must be sorted in ascending virtual address order.
1733 *
1734 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1735 * range after unmap_vmas() returns. So the only responsibility here is to
1736 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1737 * drops the lock and schedules.
1738 */
1739void unmap_vmas(struct mmu_gather *tlb, struct ma_state *mas,
1740 struct vm_area_struct *vma, unsigned long start_addr,
1741 unsigned long end_addr, unsigned long tree_end,
1742 bool mm_wr_locked)
1743{
1744 struct mmu_notifier_range range;
1745 struct zap_details details = {
1746 .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
1747 /* Careful - we need to zap private pages too! */
1748 .even_cows = true,
1749 };
1750
1751 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm,
1752 start_addr, end_addr);
1753 mmu_notifier_invalidate_range_start(&range);
1754 do {
1755 unsigned long start = start_addr;
1756 unsigned long end = end_addr;
1757 hugetlb_zap_begin(vma, &start, &end);
1758 unmap_single_vma(tlb, vma, start, end, &details,
1759 mm_wr_locked);
1760 hugetlb_zap_end(vma, &details);
1761 vma = mas_find(mas, tree_end - 1);
1762 } while (vma && likely(!xa_is_zero(vma)));
1763 mmu_notifier_invalidate_range_end(&range);
1764}
1765
1766/**
1767 * zap_page_range_single - remove user pages in a given range
1768 * @vma: vm_area_struct holding the applicable pages
1769 * @address: starting address of pages to zap
1770 * @size: number of bytes to zap
1771 * @details: details of shared cache invalidation
1772 *
1773 * The range must fit into one VMA.
1774 */
1775void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1776 unsigned long size, struct zap_details *details)
1777{
1778 const unsigned long end = address + size;
1779 struct mmu_notifier_range range;
1780 struct mmu_gather tlb;
1781
1782 lru_add_drain();
1783 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1784 address, end);
1785 hugetlb_zap_begin(vma, &range.start, &range.end);
1786 tlb_gather_mmu(&tlb, vma->vm_mm);
1787 update_hiwater_rss(vma->vm_mm);
1788 mmu_notifier_invalidate_range_start(&range);
1789 /*
1790 * unmap 'address-end' not 'range.start-range.end' as range
1791 * could have been expanded for hugetlb pmd sharing.
1792 */
1793 unmap_single_vma(&tlb, vma, address, end, details, false);
1794 mmu_notifier_invalidate_range_end(&range);
1795 tlb_finish_mmu(&tlb);
1796 hugetlb_zap_end(vma, details);
1797}
1798
1799/**
1800 * zap_vma_ptes - remove ptes mapping the vma
1801 * @vma: vm_area_struct holding ptes to be zapped
1802 * @address: starting address of pages to zap
1803 * @size: number of bytes to zap
1804 *
1805 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1806 *
1807 * The entire address range must be fully contained within the vma.
1808 *
1809 */
1810void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1811 unsigned long size)
1812{
1813 if (!range_in_vma(vma, address, address + size) ||
1814 !(vma->vm_flags & VM_PFNMAP))
1815 return;
1816
1817 zap_page_range_single(vma, address, size, NULL);
1818}
1819EXPORT_SYMBOL_GPL(zap_vma_ptes);
1820
1821static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1822{
1823 pgd_t *pgd;
1824 p4d_t *p4d;
1825 pud_t *pud;
1826 pmd_t *pmd;
1827
1828 pgd = pgd_offset(mm, addr);
1829 p4d = p4d_alloc(mm, pgd, addr);
1830 if (!p4d)
1831 return NULL;
1832 pud = pud_alloc(mm, p4d, addr);
1833 if (!pud)
1834 return NULL;
1835 pmd = pmd_alloc(mm, pud, addr);
1836 if (!pmd)
1837 return NULL;
1838
1839 VM_BUG_ON(pmd_trans_huge(*pmd));
1840 return pmd;
1841}
1842
1843pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1844 spinlock_t **ptl)
1845{
1846 pmd_t *pmd = walk_to_pmd(mm, addr);
1847
1848 if (!pmd)
1849 return NULL;
1850 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1851}
1852
1853static int validate_page_before_insert(struct page *page)
1854{
1855 struct folio *folio = page_folio(page);
1856
1857 if (folio_test_anon(folio) || folio_test_slab(folio) ||
1858 page_has_type(page))
1859 return -EINVAL;
1860 flush_dcache_folio(folio);
1861 return 0;
1862}
1863
1864static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
1865 unsigned long addr, struct page *page, pgprot_t prot)
1866{
1867 struct folio *folio = page_folio(page);
1868
1869 if (!pte_none(ptep_get(pte)))
1870 return -EBUSY;
1871 /* Ok, finally just insert the thing.. */
1872 folio_get(folio);
1873 inc_mm_counter(vma->vm_mm, mm_counter_file(page));
1874 folio_add_file_rmap_pte(folio, page, vma);
1875 set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot));
1876 return 0;
1877}
1878
1879/*
1880 * This is the old fallback for page remapping.
1881 *
1882 * For historical reasons, it only allows reserved pages. Only
1883 * old drivers should use this, and they needed to mark their
1884 * pages reserved for the old functions anyway.
1885 */
1886static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1887 struct page *page, pgprot_t prot)
1888{
1889 int retval;
1890 pte_t *pte;
1891 spinlock_t *ptl;
1892
1893 retval = validate_page_before_insert(page);
1894 if (retval)
1895 goto out;
1896 retval = -ENOMEM;
1897 pte = get_locked_pte(vma->vm_mm, addr, &ptl);
1898 if (!pte)
1899 goto out;
1900 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot);
1901 pte_unmap_unlock(pte, ptl);
1902out:
1903 return retval;
1904}
1905
1906static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
1907 unsigned long addr, struct page *page, pgprot_t prot)
1908{
1909 int err;
1910
1911 if (!page_count(page))
1912 return -EINVAL;
1913 err = validate_page_before_insert(page);
1914 if (err)
1915 return err;
1916 return insert_page_into_pte_locked(vma, pte, addr, page, prot);
1917}
1918
1919/* insert_pages() amortizes the cost of spinlock operations
1920 * when inserting pages in a loop.
1921 */
1922static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1923 struct page **pages, unsigned long *num, pgprot_t prot)
1924{
1925 pmd_t *pmd = NULL;
1926 pte_t *start_pte, *pte;
1927 spinlock_t *pte_lock;
1928 struct mm_struct *const mm = vma->vm_mm;
1929 unsigned long curr_page_idx = 0;
1930 unsigned long remaining_pages_total = *num;
1931 unsigned long pages_to_write_in_pmd;
1932 int ret;
1933more:
1934 ret = -EFAULT;
1935 pmd = walk_to_pmd(mm, addr);
1936 if (!pmd)
1937 goto out;
1938
1939 pages_to_write_in_pmd = min_t(unsigned long,
1940 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1941
1942 /* Allocate the PTE if necessary; takes PMD lock once only. */
1943 ret = -ENOMEM;
1944 if (pte_alloc(mm, pmd))
1945 goto out;
1946
1947 while (pages_to_write_in_pmd) {
1948 int pte_idx = 0;
1949 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1950
1951 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1952 if (!start_pte) {
1953 ret = -EFAULT;
1954 goto out;
1955 }
1956 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1957 int err = insert_page_in_batch_locked(vma, pte,
1958 addr, pages[curr_page_idx], prot);
1959 if (unlikely(err)) {
1960 pte_unmap_unlock(start_pte, pte_lock);
1961 ret = err;
1962 remaining_pages_total -= pte_idx;
1963 goto out;
1964 }
1965 addr += PAGE_SIZE;
1966 ++curr_page_idx;
1967 }
1968 pte_unmap_unlock(start_pte, pte_lock);
1969 pages_to_write_in_pmd -= batch_size;
1970 remaining_pages_total -= batch_size;
1971 }
1972 if (remaining_pages_total)
1973 goto more;
1974 ret = 0;
1975out:
1976 *num = remaining_pages_total;
1977 return ret;
1978}
1979
1980/**
1981 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1982 * @vma: user vma to map to
1983 * @addr: target start user address of these pages
1984 * @pages: source kernel pages
1985 * @num: in: number of pages to map. out: number of pages that were *not*
1986 * mapped. (0 means all pages were successfully mapped).
1987 *
1988 * Preferred over vm_insert_page() when inserting multiple pages.
1989 *
1990 * In case of error, we may have mapped a subset of the provided
1991 * pages. It is the caller's responsibility to account for this case.
1992 *
1993 * The same restrictions apply as in vm_insert_page().
1994 */
1995int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1996 struct page **pages, unsigned long *num)
1997{
1998 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1999
2000 if (addr < vma->vm_start || end_addr >= vma->vm_end)
2001 return -EFAULT;
2002 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2003 BUG_ON(mmap_read_trylock(vma->vm_mm));
2004 BUG_ON(vma->vm_flags & VM_PFNMAP);
2005 vm_flags_set(vma, VM_MIXEDMAP);
2006 }
2007 /* Defer page refcount checking till we're about to map that page. */
2008 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
2009}
2010EXPORT_SYMBOL(vm_insert_pages);
2011
2012/**
2013 * vm_insert_page - insert single page into user vma
2014 * @vma: user vma to map to
2015 * @addr: target user address of this page
2016 * @page: source kernel page
2017 *
2018 * This allows drivers to insert individual pages they've allocated
2019 * into a user vma.
2020 *
2021 * The page has to be a nice clean _individual_ kernel allocation.
2022 * If you allocate a compound page, you need to have marked it as
2023 * such (__GFP_COMP), or manually just split the page up yourself
2024 * (see split_page()).
2025 *
2026 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2027 * took an arbitrary page protection parameter. This doesn't allow
2028 * that. Your vma protection will have to be set up correctly, which
2029 * means that if you want a shared writable mapping, you'd better
2030 * ask for a shared writable mapping!
2031 *
2032 * The page does not need to be reserved.
2033 *
2034 * Usually this function is called from f_op->mmap() handler
2035 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2036 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2037 * function from other places, for example from page-fault handler.
2038 *
2039 * Return: %0 on success, negative error code otherwise.
2040 */
2041int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2042 struct page *page)
2043{
2044 if (addr < vma->vm_start || addr >= vma->vm_end)
2045 return -EFAULT;
2046 if (!page_count(page))
2047 return -EINVAL;
2048 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2049 BUG_ON(mmap_read_trylock(vma->vm_mm));
2050 BUG_ON(vma->vm_flags & VM_PFNMAP);
2051 vm_flags_set(vma, VM_MIXEDMAP);
2052 }
2053 return insert_page(vma, addr, page, vma->vm_page_prot);
2054}
2055EXPORT_SYMBOL(vm_insert_page);
2056
2057/*
2058 * __vm_map_pages - maps range of kernel pages into user vma
2059 * @vma: user vma to map to
2060 * @pages: pointer to array of source kernel pages
2061 * @num: number of pages in page array
2062 * @offset: user's requested vm_pgoff
2063 *
2064 * This allows drivers to map range of kernel pages into a user vma.
2065 *
2066 * Return: 0 on success and error code otherwise.
2067 */
2068static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2069 unsigned long num, unsigned long offset)
2070{
2071 unsigned long count = vma_pages(vma);
2072 unsigned long uaddr = vma->vm_start;
2073 int ret, i;
2074
2075 /* Fail if the user requested offset is beyond the end of the object */
2076 if (offset >= num)
2077 return -ENXIO;
2078
2079 /* Fail if the user requested size exceeds available object size */
2080 if (count > num - offset)
2081 return -ENXIO;
2082
2083 for (i = 0; i < count; i++) {
2084 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2085 if (ret < 0)
2086 return ret;
2087 uaddr += PAGE_SIZE;
2088 }
2089
2090 return 0;
2091}
2092
2093/**
2094 * vm_map_pages - maps range of kernel pages starts with non zero offset
2095 * @vma: user vma to map to
2096 * @pages: pointer to array of source kernel pages
2097 * @num: number of pages in page array
2098 *
2099 * Maps an object consisting of @num pages, catering for the user's
2100 * requested vm_pgoff
2101 *
2102 * If we fail to insert any page into the vma, the function will return
2103 * immediately leaving any previously inserted pages present. Callers
2104 * from the mmap handler may immediately return the error as their caller
2105 * will destroy the vma, removing any successfully inserted pages. Other
2106 * callers should make their own arrangements for calling unmap_region().
2107 *
2108 * Context: Process context. Called by mmap handlers.
2109 * Return: 0 on success and error code otherwise.
2110 */
2111int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2112 unsigned long num)
2113{
2114 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2115}
2116EXPORT_SYMBOL(vm_map_pages);
2117
2118/**
2119 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2120 * @vma: user vma to map to
2121 * @pages: pointer to array of source kernel pages
2122 * @num: number of pages in page array
2123 *
2124 * Similar to vm_map_pages(), except that it explicitly sets the offset
2125 * to 0. This function is intended for the drivers that did not consider
2126 * vm_pgoff.
2127 *
2128 * Context: Process context. Called by mmap handlers.
2129 * Return: 0 on success and error code otherwise.
2130 */
2131int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2132 unsigned long num)
2133{
2134 return __vm_map_pages(vma, pages, num, 0);
2135}
2136EXPORT_SYMBOL(vm_map_pages_zero);
2137
2138static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2139 pfn_t pfn, pgprot_t prot, bool mkwrite)
2140{
2141 struct mm_struct *mm = vma->vm_mm;
2142 pte_t *pte, entry;
2143 spinlock_t *ptl;
2144
2145 pte = get_locked_pte(mm, addr, &ptl);
2146 if (!pte)
2147 return VM_FAULT_OOM;
2148 entry = ptep_get(pte);
2149 if (!pte_none(entry)) {
2150 if (mkwrite) {
2151 /*
2152 * For read faults on private mappings the PFN passed
2153 * in may not match the PFN we have mapped if the
2154 * mapped PFN is a writeable COW page. In the mkwrite
2155 * case we are creating a writable PTE for a shared
2156 * mapping and we expect the PFNs to match. If they
2157 * don't match, we are likely racing with block
2158 * allocation and mapping invalidation so just skip the
2159 * update.
2160 */
2161 if (pte_pfn(entry) != pfn_t_to_pfn(pfn)) {
2162 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry)));
2163 goto out_unlock;
2164 }
2165 entry = pte_mkyoung(entry);
2166 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2167 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2168 update_mmu_cache(vma, addr, pte);
2169 }
2170 goto out_unlock;
2171 }
2172
2173 /* Ok, finally just insert the thing.. */
2174 if (pfn_t_devmap(pfn))
2175 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2176 else
2177 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2178
2179 if (mkwrite) {
2180 entry = pte_mkyoung(entry);
2181 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2182 }
2183
2184 set_pte_at(mm, addr, pte, entry);
2185 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2186
2187out_unlock:
2188 pte_unmap_unlock(pte, ptl);
2189 return VM_FAULT_NOPAGE;
2190}
2191
2192/**
2193 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2194 * @vma: user vma to map to
2195 * @addr: target user address of this page
2196 * @pfn: source kernel pfn
2197 * @pgprot: pgprot flags for the inserted page
2198 *
2199 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2200 * to override pgprot on a per-page basis.
2201 *
2202 * This only makes sense for IO mappings, and it makes no sense for
2203 * COW mappings. In general, using multiple vmas is preferable;
2204 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2205 * impractical.
2206 *
2207 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2208 * caching- and encryption bits different than those of @vma->vm_page_prot,
2209 * because the caching- or encryption mode may not be known at mmap() time.
2210 *
2211 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2212 * to set caching and encryption bits for those vmas (except for COW pages).
2213 * This is ensured by core vm only modifying these page table entries using
2214 * functions that don't touch caching- or encryption bits, using pte_modify()
2215 * if needed. (See for example mprotect()).
2216 *
2217 * Also when new page-table entries are created, this is only done using the
2218 * fault() callback, and never using the value of vma->vm_page_prot,
2219 * except for page-table entries that point to anonymous pages as the result
2220 * of COW.
2221 *
2222 * Context: Process context. May allocate using %GFP_KERNEL.
2223 * Return: vm_fault_t value.
2224 */
2225vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2226 unsigned long pfn, pgprot_t pgprot)
2227{
2228 /*
2229 * Technically, architectures with pte_special can avoid all these
2230 * restrictions (same for remap_pfn_range). However we would like
2231 * consistency in testing and feature parity among all, so we should
2232 * try to keep these invariants in place for everybody.
2233 */
2234 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2235 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2236 (VM_PFNMAP|VM_MIXEDMAP));
2237 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2238 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2239
2240 if (addr < vma->vm_start || addr >= vma->vm_end)
2241 return VM_FAULT_SIGBUS;
2242
2243 if (!pfn_modify_allowed(pfn, pgprot))
2244 return VM_FAULT_SIGBUS;
2245
2246 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2247
2248 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2249 false);
2250}
2251EXPORT_SYMBOL(vmf_insert_pfn_prot);
2252
2253/**
2254 * vmf_insert_pfn - insert single pfn into user vma
2255 * @vma: user vma to map to
2256 * @addr: target user address of this page
2257 * @pfn: source kernel pfn
2258 *
2259 * Similar to vm_insert_page, this allows drivers to insert individual pages
2260 * they've allocated into a user vma. Same comments apply.
2261 *
2262 * This function should only be called from a vm_ops->fault handler, and
2263 * in that case the handler should return the result of this function.
2264 *
2265 * vma cannot be a COW mapping.
2266 *
2267 * As this is called only for pages that do not currently exist, we
2268 * do not need to flush old virtual caches or the TLB.
2269 *
2270 * Context: Process context. May allocate using %GFP_KERNEL.
2271 * Return: vm_fault_t value.
2272 */
2273vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2274 unsigned long pfn)
2275{
2276 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2277}
2278EXPORT_SYMBOL(vmf_insert_pfn);
2279
2280static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2281{
2282 /* these checks mirror the abort conditions in vm_normal_page */
2283 if (vma->vm_flags & VM_MIXEDMAP)
2284 return true;
2285 if (pfn_t_devmap(pfn))
2286 return true;
2287 if (pfn_t_special(pfn))
2288 return true;
2289 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2290 return true;
2291 return false;
2292}
2293
2294static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2295 unsigned long addr, pfn_t pfn, bool mkwrite)
2296{
2297 pgprot_t pgprot = vma->vm_page_prot;
2298 int err;
2299
2300 BUG_ON(!vm_mixed_ok(vma, pfn));
2301
2302 if (addr < vma->vm_start || addr >= vma->vm_end)
2303 return VM_FAULT_SIGBUS;
2304
2305 track_pfn_insert(vma, &pgprot, pfn);
2306
2307 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2308 return VM_FAULT_SIGBUS;
2309
2310 /*
2311 * If we don't have pte special, then we have to use the pfn_valid()
2312 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2313 * refcount the page if pfn_valid is true (hence insert_page rather
2314 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2315 * without pte special, it would there be refcounted as a normal page.
2316 */
2317 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2318 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2319 struct page *page;
2320
2321 /*
2322 * At this point we are committed to insert_page()
2323 * regardless of whether the caller specified flags that
2324 * result in pfn_t_has_page() == false.
2325 */
2326 page = pfn_to_page(pfn_t_to_pfn(pfn));
2327 err = insert_page(vma, addr, page, pgprot);
2328 } else {
2329 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2330 }
2331
2332 if (err == -ENOMEM)
2333 return VM_FAULT_OOM;
2334 if (err < 0 && err != -EBUSY)
2335 return VM_FAULT_SIGBUS;
2336
2337 return VM_FAULT_NOPAGE;
2338}
2339
2340vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2341 pfn_t pfn)
2342{
2343 return __vm_insert_mixed(vma, addr, pfn, false);
2344}
2345EXPORT_SYMBOL(vmf_insert_mixed);
2346
2347/*
2348 * If the insertion of PTE failed because someone else already added a
2349 * different entry in the mean time, we treat that as success as we assume
2350 * the same entry was actually inserted.
2351 */
2352vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2353 unsigned long addr, pfn_t pfn)
2354{
2355 return __vm_insert_mixed(vma, addr, pfn, true);
2356}
2357EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2358
2359/*
2360 * maps a range of physical memory into the requested pages. the old
2361 * mappings are removed. any references to nonexistent pages results
2362 * in null mappings (currently treated as "copy-on-access")
2363 */
2364static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2365 unsigned long addr, unsigned long end,
2366 unsigned long pfn, pgprot_t prot)
2367{
2368 pte_t *pte, *mapped_pte;
2369 spinlock_t *ptl;
2370 int err = 0;
2371
2372 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2373 if (!pte)
2374 return -ENOMEM;
2375 arch_enter_lazy_mmu_mode();
2376 do {
2377 BUG_ON(!pte_none(ptep_get(pte)));
2378 if (!pfn_modify_allowed(pfn, prot)) {
2379 err = -EACCES;
2380 break;
2381 }
2382 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2383 pfn++;
2384 } while (pte++, addr += PAGE_SIZE, addr != end);
2385 arch_leave_lazy_mmu_mode();
2386 pte_unmap_unlock(mapped_pte, ptl);
2387 return err;
2388}
2389
2390static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2391 unsigned long addr, unsigned long end,
2392 unsigned long pfn, pgprot_t prot)
2393{
2394 pmd_t *pmd;
2395 unsigned long next;
2396 int err;
2397
2398 pfn -= addr >> PAGE_SHIFT;
2399 pmd = pmd_alloc(mm, pud, addr);
2400 if (!pmd)
2401 return -ENOMEM;
2402 VM_BUG_ON(pmd_trans_huge(*pmd));
2403 do {
2404 next = pmd_addr_end(addr, end);
2405 err = remap_pte_range(mm, pmd, addr, next,
2406 pfn + (addr >> PAGE_SHIFT), prot);
2407 if (err)
2408 return err;
2409 } while (pmd++, addr = next, addr != end);
2410 return 0;
2411}
2412
2413static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2414 unsigned long addr, unsigned long end,
2415 unsigned long pfn, pgprot_t prot)
2416{
2417 pud_t *pud;
2418 unsigned long next;
2419 int err;
2420
2421 pfn -= addr >> PAGE_SHIFT;
2422 pud = pud_alloc(mm, p4d, addr);
2423 if (!pud)
2424 return -ENOMEM;
2425 do {
2426 next = pud_addr_end(addr, end);
2427 err = remap_pmd_range(mm, pud, addr, next,
2428 pfn + (addr >> PAGE_SHIFT), prot);
2429 if (err)
2430 return err;
2431 } while (pud++, addr = next, addr != end);
2432 return 0;
2433}
2434
2435static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2436 unsigned long addr, unsigned long end,
2437 unsigned long pfn, pgprot_t prot)
2438{
2439 p4d_t *p4d;
2440 unsigned long next;
2441 int err;
2442
2443 pfn -= addr >> PAGE_SHIFT;
2444 p4d = p4d_alloc(mm, pgd, addr);
2445 if (!p4d)
2446 return -ENOMEM;
2447 do {
2448 next = p4d_addr_end(addr, end);
2449 err = remap_pud_range(mm, p4d, addr, next,
2450 pfn + (addr >> PAGE_SHIFT), prot);
2451 if (err)
2452 return err;
2453 } while (p4d++, addr = next, addr != end);
2454 return 0;
2455}
2456
2457/*
2458 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2459 * must have pre-validated the caching bits of the pgprot_t.
2460 */
2461int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2462 unsigned long pfn, unsigned long size, pgprot_t prot)
2463{
2464 pgd_t *pgd;
2465 unsigned long next;
2466 unsigned long end = addr + PAGE_ALIGN(size);
2467 struct mm_struct *mm = vma->vm_mm;
2468 int err;
2469
2470 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2471 return -EINVAL;
2472
2473 /*
2474 * Physically remapped pages are special. Tell the
2475 * rest of the world about it:
2476 * VM_IO tells people not to look at these pages
2477 * (accesses can have side effects).
2478 * VM_PFNMAP tells the core MM that the base pages are just
2479 * raw PFN mappings, and do not have a "struct page" associated
2480 * with them.
2481 * VM_DONTEXPAND
2482 * Disable vma merging and expanding with mremap().
2483 * VM_DONTDUMP
2484 * Omit vma from core dump, even when VM_IO turned off.
2485 *
2486 * There's a horrible special case to handle copy-on-write
2487 * behaviour that some programs depend on. We mark the "original"
2488 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2489 * See vm_normal_page() for details.
2490 */
2491 if (is_cow_mapping(vma->vm_flags)) {
2492 if (addr != vma->vm_start || end != vma->vm_end)
2493 return -EINVAL;
2494 vma->vm_pgoff = pfn;
2495 }
2496
2497 vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
2498
2499 BUG_ON(addr >= end);
2500 pfn -= addr >> PAGE_SHIFT;
2501 pgd = pgd_offset(mm, addr);
2502 flush_cache_range(vma, addr, end);
2503 do {
2504 next = pgd_addr_end(addr, end);
2505 err = remap_p4d_range(mm, pgd, addr, next,
2506 pfn + (addr >> PAGE_SHIFT), prot);
2507 if (err)
2508 return err;
2509 } while (pgd++, addr = next, addr != end);
2510
2511 return 0;
2512}
2513
2514/**
2515 * remap_pfn_range - remap kernel memory to userspace
2516 * @vma: user vma to map to
2517 * @addr: target page aligned user address to start at
2518 * @pfn: page frame number of kernel physical memory address
2519 * @size: size of mapping area
2520 * @prot: page protection flags for this mapping
2521 *
2522 * Note: this is only safe if the mm semaphore is held when called.
2523 *
2524 * Return: %0 on success, negative error code otherwise.
2525 */
2526int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2527 unsigned long pfn, unsigned long size, pgprot_t prot)
2528{
2529 int err;
2530
2531 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2532 if (err)
2533 return -EINVAL;
2534
2535 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2536 if (err)
2537 untrack_pfn(vma, pfn, PAGE_ALIGN(size), true);
2538 return err;
2539}
2540EXPORT_SYMBOL(remap_pfn_range);
2541
2542/**
2543 * vm_iomap_memory - remap memory to userspace
2544 * @vma: user vma to map to
2545 * @start: start of the physical memory to be mapped
2546 * @len: size of area
2547 *
2548 * This is a simplified io_remap_pfn_range() for common driver use. The
2549 * driver just needs to give us the physical memory range to be mapped,
2550 * we'll figure out the rest from the vma information.
2551 *
2552 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2553 * whatever write-combining details or similar.
2554 *
2555 * Return: %0 on success, negative error code otherwise.
2556 */
2557int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2558{
2559 unsigned long vm_len, pfn, pages;
2560
2561 /* Check that the physical memory area passed in looks valid */
2562 if (start + len < start)
2563 return -EINVAL;
2564 /*
2565 * You *really* shouldn't map things that aren't page-aligned,
2566 * but we've historically allowed it because IO memory might
2567 * just have smaller alignment.
2568 */
2569 len += start & ~PAGE_MASK;
2570 pfn = start >> PAGE_SHIFT;
2571 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2572 if (pfn + pages < pfn)
2573 return -EINVAL;
2574
2575 /* We start the mapping 'vm_pgoff' pages into the area */
2576 if (vma->vm_pgoff > pages)
2577 return -EINVAL;
2578 pfn += vma->vm_pgoff;
2579 pages -= vma->vm_pgoff;
2580
2581 /* Can we fit all of the mapping? */
2582 vm_len = vma->vm_end - vma->vm_start;
2583 if (vm_len >> PAGE_SHIFT > pages)
2584 return -EINVAL;
2585
2586 /* Ok, let it rip */
2587 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2588}
2589EXPORT_SYMBOL(vm_iomap_memory);
2590
2591static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2592 unsigned long addr, unsigned long end,
2593 pte_fn_t fn, void *data, bool create,
2594 pgtbl_mod_mask *mask)
2595{
2596 pte_t *pte, *mapped_pte;
2597 int err = 0;
2598 spinlock_t *ptl;
2599
2600 if (create) {
2601 mapped_pte = pte = (mm == &init_mm) ?
2602 pte_alloc_kernel_track(pmd, addr, mask) :
2603 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2604 if (!pte)
2605 return -ENOMEM;
2606 } else {
2607 mapped_pte = pte = (mm == &init_mm) ?
2608 pte_offset_kernel(pmd, addr) :
2609 pte_offset_map_lock(mm, pmd, addr, &ptl);
2610 if (!pte)
2611 return -EINVAL;
2612 }
2613
2614 arch_enter_lazy_mmu_mode();
2615
2616 if (fn) {
2617 do {
2618 if (create || !pte_none(ptep_get(pte))) {
2619 err = fn(pte++, addr, data);
2620 if (err)
2621 break;
2622 }
2623 } while (addr += PAGE_SIZE, addr != end);
2624 }
2625 *mask |= PGTBL_PTE_MODIFIED;
2626
2627 arch_leave_lazy_mmu_mode();
2628
2629 if (mm != &init_mm)
2630 pte_unmap_unlock(mapped_pte, ptl);
2631 return err;
2632}
2633
2634static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2635 unsigned long addr, unsigned long end,
2636 pte_fn_t fn, void *data, bool create,
2637 pgtbl_mod_mask *mask)
2638{
2639 pmd_t *pmd;
2640 unsigned long next;
2641 int err = 0;
2642
2643 BUG_ON(pud_huge(*pud));
2644
2645 if (create) {
2646 pmd = pmd_alloc_track(mm, pud, addr, mask);
2647 if (!pmd)
2648 return -ENOMEM;
2649 } else {
2650 pmd = pmd_offset(pud, addr);
2651 }
2652 do {
2653 next = pmd_addr_end(addr, end);
2654 if (pmd_none(*pmd) && !create)
2655 continue;
2656 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2657 return -EINVAL;
2658 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2659 if (!create)
2660 continue;
2661 pmd_clear_bad(pmd);
2662 }
2663 err = apply_to_pte_range(mm, pmd, addr, next,
2664 fn, data, create, mask);
2665 if (err)
2666 break;
2667 } while (pmd++, addr = next, addr != end);
2668
2669 return err;
2670}
2671
2672static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2673 unsigned long addr, unsigned long end,
2674 pte_fn_t fn, void *data, bool create,
2675 pgtbl_mod_mask *mask)
2676{
2677 pud_t *pud;
2678 unsigned long next;
2679 int err = 0;
2680
2681 if (create) {
2682 pud = pud_alloc_track(mm, p4d, addr, mask);
2683 if (!pud)
2684 return -ENOMEM;
2685 } else {
2686 pud = pud_offset(p4d, addr);
2687 }
2688 do {
2689 next = pud_addr_end(addr, end);
2690 if (pud_none(*pud) && !create)
2691 continue;
2692 if (WARN_ON_ONCE(pud_leaf(*pud)))
2693 return -EINVAL;
2694 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2695 if (!create)
2696 continue;
2697 pud_clear_bad(pud);
2698 }
2699 err = apply_to_pmd_range(mm, pud, addr, next,
2700 fn, data, create, mask);
2701 if (err)
2702 break;
2703 } while (pud++, addr = next, addr != end);
2704
2705 return err;
2706}
2707
2708static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2709 unsigned long addr, unsigned long end,
2710 pte_fn_t fn, void *data, bool create,
2711 pgtbl_mod_mask *mask)
2712{
2713 p4d_t *p4d;
2714 unsigned long next;
2715 int err = 0;
2716
2717 if (create) {
2718 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2719 if (!p4d)
2720 return -ENOMEM;
2721 } else {
2722 p4d = p4d_offset(pgd, addr);
2723 }
2724 do {
2725 next = p4d_addr_end(addr, end);
2726 if (p4d_none(*p4d) && !create)
2727 continue;
2728 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2729 return -EINVAL;
2730 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2731 if (!create)
2732 continue;
2733 p4d_clear_bad(p4d);
2734 }
2735 err = apply_to_pud_range(mm, p4d, addr, next,
2736 fn, data, create, mask);
2737 if (err)
2738 break;
2739 } while (p4d++, addr = next, addr != end);
2740
2741 return err;
2742}
2743
2744static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2745 unsigned long size, pte_fn_t fn,
2746 void *data, bool create)
2747{
2748 pgd_t *pgd;
2749 unsigned long start = addr, next;
2750 unsigned long end = addr + size;
2751 pgtbl_mod_mask mask = 0;
2752 int err = 0;
2753
2754 if (WARN_ON(addr >= end))
2755 return -EINVAL;
2756
2757 pgd = pgd_offset(mm, addr);
2758 do {
2759 next = pgd_addr_end(addr, end);
2760 if (pgd_none(*pgd) && !create)
2761 continue;
2762 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2763 return -EINVAL;
2764 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2765 if (!create)
2766 continue;
2767 pgd_clear_bad(pgd);
2768 }
2769 err = apply_to_p4d_range(mm, pgd, addr, next,
2770 fn, data, create, &mask);
2771 if (err)
2772 break;
2773 } while (pgd++, addr = next, addr != end);
2774
2775 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2776 arch_sync_kernel_mappings(start, start + size);
2777
2778 return err;
2779}
2780
2781/*
2782 * Scan a region of virtual memory, filling in page tables as necessary
2783 * and calling a provided function on each leaf page table.
2784 */
2785int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2786 unsigned long size, pte_fn_t fn, void *data)
2787{
2788 return __apply_to_page_range(mm, addr, size, fn, data, true);
2789}
2790EXPORT_SYMBOL_GPL(apply_to_page_range);
2791
2792/*
2793 * Scan a region of virtual memory, calling a provided function on
2794 * each leaf page table where it exists.
2795 *
2796 * Unlike apply_to_page_range, this does _not_ fill in page tables
2797 * where they are absent.
2798 */
2799int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2800 unsigned long size, pte_fn_t fn, void *data)
2801{
2802 return __apply_to_page_range(mm, addr, size, fn, data, false);
2803}
2804EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2805
2806/*
2807 * handle_pte_fault chooses page fault handler according to an entry which was
2808 * read non-atomically. Before making any commitment, on those architectures
2809 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2810 * parts, do_swap_page must check under lock before unmapping the pte and
2811 * proceeding (but do_wp_page is only called after already making such a check;
2812 * and do_anonymous_page can safely check later on).
2813 */
2814static inline int pte_unmap_same(struct vm_fault *vmf)
2815{
2816 int same = 1;
2817#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2818 if (sizeof(pte_t) > sizeof(unsigned long)) {
2819 spin_lock(vmf->ptl);
2820 same = pte_same(ptep_get(vmf->pte), vmf->orig_pte);
2821 spin_unlock(vmf->ptl);
2822 }
2823#endif
2824 pte_unmap(vmf->pte);
2825 vmf->pte = NULL;
2826 return same;
2827}
2828
2829/*
2830 * Return:
2831 * 0: copied succeeded
2832 * -EHWPOISON: copy failed due to hwpoison in source page
2833 * -EAGAIN: copied failed (some other reason)
2834 */
2835static inline int __wp_page_copy_user(struct page *dst, struct page *src,
2836 struct vm_fault *vmf)
2837{
2838 int ret;
2839 void *kaddr;
2840 void __user *uaddr;
2841 struct vm_area_struct *vma = vmf->vma;
2842 struct mm_struct *mm = vma->vm_mm;
2843 unsigned long addr = vmf->address;
2844
2845 if (likely(src)) {
2846 if (copy_mc_user_highpage(dst, src, addr, vma)) {
2847 memory_failure_queue(page_to_pfn(src), 0);
2848 return -EHWPOISON;
2849 }
2850 return 0;
2851 }
2852
2853 /*
2854 * If the source page was a PFN mapping, we don't have
2855 * a "struct page" for it. We do a best-effort copy by
2856 * just copying from the original user address. If that
2857 * fails, we just zero-fill it. Live with it.
2858 */
2859 kaddr = kmap_local_page(dst);
2860 pagefault_disable();
2861 uaddr = (void __user *)(addr & PAGE_MASK);
2862
2863 /*
2864 * On architectures with software "accessed" bits, we would
2865 * take a double page fault, so mark it accessed here.
2866 */
2867 vmf->pte = NULL;
2868 if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) {
2869 pte_t entry;
2870
2871 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2872 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2873 /*
2874 * Other thread has already handled the fault
2875 * and update local tlb only
2876 */
2877 if (vmf->pte)
2878 update_mmu_tlb(vma, addr, vmf->pte);
2879 ret = -EAGAIN;
2880 goto pte_unlock;
2881 }
2882
2883 entry = pte_mkyoung(vmf->orig_pte);
2884 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2885 update_mmu_cache_range(vmf, vma, addr, vmf->pte, 1);
2886 }
2887
2888 /*
2889 * This really shouldn't fail, because the page is there
2890 * in the page tables. But it might just be unreadable,
2891 * in which case we just give up and fill the result with
2892 * zeroes.
2893 */
2894 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2895 if (vmf->pte)
2896 goto warn;
2897
2898 /* Re-validate under PTL if the page is still mapped */
2899 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2900 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2901 /* The PTE changed under us, update local tlb */
2902 if (vmf->pte)
2903 update_mmu_tlb(vma, addr, vmf->pte);
2904 ret = -EAGAIN;
2905 goto pte_unlock;
2906 }
2907
2908 /*
2909 * The same page can be mapped back since last copy attempt.
2910 * Try to copy again under PTL.
2911 */
2912 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2913 /*
2914 * Give a warn in case there can be some obscure
2915 * use-case
2916 */
2917warn:
2918 WARN_ON_ONCE(1);
2919 clear_page(kaddr);
2920 }
2921 }
2922
2923 ret = 0;
2924
2925pte_unlock:
2926 if (vmf->pte)
2927 pte_unmap_unlock(vmf->pte, vmf->ptl);
2928 pagefault_enable();
2929 kunmap_local(kaddr);
2930 flush_dcache_page(dst);
2931
2932 return ret;
2933}
2934
2935static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2936{
2937 struct file *vm_file = vma->vm_file;
2938
2939 if (vm_file)
2940 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2941
2942 /*
2943 * Special mappings (e.g. VDSO) do not have any file so fake
2944 * a default GFP_KERNEL for them.
2945 */
2946 return GFP_KERNEL;
2947}
2948
2949/*
2950 * Notify the address space that the page is about to become writable so that
2951 * it can prohibit this or wait for the page to get into an appropriate state.
2952 *
2953 * We do this without the lock held, so that it can sleep if it needs to.
2954 */
2955static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio)
2956{
2957 vm_fault_t ret;
2958 unsigned int old_flags = vmf->flags;
2959
2960 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2961
2962 if (vmf->vma->vm_file &&
2963 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2964 return VM_FAULT_SIGBUS;
2965
2966 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2967 /* Restore original flags so that caller is not surprised */
2968 vmf->flags = old_flags;
2969 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2970 return ret;
2971 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2972 folio_lock(folio);
2973 if (!folio->mapping) {
2974 folio_unlock(folio);
2975 return 0; /* retry */
2976 }
2977 ret |= VM_FAULT_LOCKED;
2978 } else
2979 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2980 return ret;
2981}
2982
2983/*
2984 * Handle dirtying of a page in shared file mapping on a write fault.
2985 *
2986 * The function expects the page to be locked and unlocks it.
2987 */
2988static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2989{
2990 struct vm_area_struct *vma = vmf->vma;
2991 struct address_space *mapping;
2992 struct folio *folio = page_folio(vmf->page);
2993 bool dirtied;
2994 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2995
2996 dirtied = folio_mark_dirty(folio);
2997 VM_BUG_ON_FOLIO(folio_test_anon(folio), folio);
2998 /*
2999 * Take a local copy of the address_space - folio.mapping may be zeroed
3000 * by truncate after folio_unlock(). The address_space itself remains
3001 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
3002 * release semantics to prevent the compiler from undoing this copying.
3003 */
3004 mapping = folio_raw_mapping(folio);
3005 folio_unlock(folio);
3006
3007 if (!page_mkwrite)
3008 file_update_time(vma->vm_file);
3009
3010 /*
3011 * Throttle page dirtying rate down to writeback speed.
3012 *
3013 * mapping may be NULL here because some device drivers do not
3014 * set page.mapping but still dirty their pages
3015 *
3016 * Drop the mmap_lock before waiting on IO, if we can. The file
3017 * is pinning the mapping, as per above.
3018 */
3019 if ((dirtied || page_mkwrite) && mapping) {
3020 struct file *fpin;
3021
3022 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
3023 balance_dirty_pages_ratelimited(mapping);
3024 if (fpin) {
3025 fput(fpin);
3026 return VM_FAULT_COMPLETED;
3027 }
3028 }
3029
3030 return 0;
3031}
3032
3033/*
3034 * Handle write page faults for pages that can be reused in the current vma
3035 *
3036 * This can happen either due to the mapping being with the VM_SHARED flag,
3037 * or due to us being the last reference standing to the page. In either
3038 * case, all we need to do here is to mark the page as writable and update
3039 * any related book-keeping.
3040 */
3041static inline void wp_page_reuse(struct vm_fault *vmf, struct folio *folio)
3042 __releases(vmf->ptl)
3043{
3044 struct vm_area_struct *vma = vmf->vma;
3045 pte_t entry;
3046
3047 VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE));
3048
3049 if (folio) {
3050 VM_BUG_ON(folio_test_anon(folio) &&
3051 !PageAnonExclusive(vmf->page));
3052 /*
3053 * Clear the folio's cpupid information as the existing
3054 * information potentially belongs to a now completely
3055 * unrelated process.
3056 */
3057 folio_xchg_last_cpupid(folio, (1 << LAST_CPUPID_SHIFT) - 1);
3058 }
3059
3060 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3061 entry = pte_mkyoung(vmf->orig_pte);
3062 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3063 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
3064 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3065 pte_unmap_unlock(vmf->pte, vmf->ptl);
3066 count_vm_event(PGREUSE);
3067}
3068
3069/*
3070 * We could add a bitflag somewhere, but for now, we know that all
3071 * vm_ops that have a ->map_pages have been audited and don't need
3072 * the mmap_lock to be held.
3073 */
3074static inline vm_fault_t vmf_can_call_fault(const struct vm_fault *vmf)
3075{
3076 struct vm_area_struct *vma = vmf->vma;
3077
3078 if (vma->vm_ops->map_pages || !(vmf->flags & FAULT_FLAG_VMA_LOCK))
3079 return 0;
3080 vma_end_read(vma);
3081 return VM_FAULT_RETRY;
3082}
3083
3084static vm_fault_t vmf_anon_prepare(struct vm_fault *vmf)
3085{
3086 struct vm_area_struct *vma = vmf->vma;
3087
3088 if (likely(vma->anon_vma))
3089 return 0;
3090 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3091 vma_end_read(vma);
3092 return VM_FAULT_RETRY;
3093 }
3094 if (__anon_vma_prepare(vma))
3095 return VM_FAULT_OOM;
3096 return 0;
3097}
3098
3099/*
3100 * Handle the case of a page which we actually need to copy to a new page,
3101 * either due to COW or unsharing.
3102 *
3103 * Called with mmap_lock locked and the old page referenced, but
3104 * without the ptl held.
3105 *
3106 * High level logic flow:
3107 *
3108 * - Allocate a page, copy the content of the old page to the new one.
3109 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3110 * - Take the PTL. If the pte changed, bail out and release the allocated page
3111 * - If the pte is still the way we remember it, update the page table and all
3112 * relevant references. This includes dropping the reference the page-table
3113 * held to the old page, as well as updating the rmap.
3114 * - In any case, unlock the PTL and drop the reference we took to the old page.
3115 */
3116static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3117{
3118 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3119 struct vm_area_struct *vma = vmf->vma;
3120 struct mm_struct *mm = vma->vm_mm;
3121 struct folio *old_folio = NULL;
3122 struct folio *new_folio = NULL;
3123 pte_t entry;
3124 int page_copied = 0;
3125 struct mmu_notifier_range range;
3126 vm_fault_t ret;
3127 bool pfn_is_zero;
3128
3129 delayacct_wpcopy_start();
3130
3131 if (vmf->page)
3132 old_folio = page_folio(vmf->page);
3133 ret = vmf_anon_prepare(vmf);
3134 if (unlikely(ret))
3135 goto out;
3136
3137 pfn_is_zero = is_zero_pfn(pte_pfn(vmf->orig_pte));
3138 new_folio = folio_prealloc(mm, vma, vmf->address, pfn_is_zero);
3139 if (!new_folio)
3140 goto oom;
3141
3142 if (!pfn_is_zero) {
3143 int err;
3144
3145 err = __wp_page_copy_user(&new_folio->page, vmf->page, vmf);
3146 if (err) {
3147 /*
3148 * COW failed, if the fault was solved by other,
3149 * it's fine. If not, userspace would re-fault on
3150 * the same address and we will handle the fault
3151 * from the second attempt.
3152 * The -EHWPOISON case will not be retried.
3153 */
3154 folio_put(new_folio);
3155 if (old_folio)
3156 folio_put(old_folio);
3157
3158 delayacct_wpcopy_end();
3159 return err == -EHWPOISON ? VM_FAULT_HWPOISON : 0;
3160 }
3161 kmsan_copy_page_meta(&new_folio->page, vmf->page);
3162 }
3163
3164 __folio_mark_uptodate(new_folio);
3165
3166 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
3167 vmf->address & PAGE_MASK,
3168 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3169 mmu_notifier_invalidate_range_start(&range);
3170
3171 /*
3172 * Re-check the pte - we dropped the lock
3173 */
3174 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3175 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3176 if (old_folio) {
3177 if (!folio_test_anon(old_folio)) {
3178 dec_mm_counter(mm, mm_counter_file(&old_folio->page));
3179 inc_mm_counter(mm, MM_ANONPAGES);
3180 }
3181 } else {
3182 ksm_might_unmap_zero_page(mm, vmf->orig_pte);
3183 inc_mm_counter(mm, MM_ANONPAGES);
3184 }
3185 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3186 entry = mk_pte(&new_folio->page, vma->vm_page_prot);
3187 entry = pte_sw_mkyoung(entry);
3188 if (unlikely(unshare)) {
3189 if (pte_soft_dirty(vmf->orig_pte))
3190 entry = pte_mksoft_dirty(entry);
3191 if (pte_uffd_wp(vmf->orig_pte))
3192 entry = pte_mkuffd_wp(entry);
3193 } else {
3194 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3195 }
3196
3197 /*
3198 * Clear the pte entry and flush it first, before updating the
3199 * pte with the new entry, to keep TLBs on different CPUs in
3200 * sync. This code used to set the new PTE then flush TLBs, but
3201 * that left a window where the new PTE could be loaded into
3202 * some TLBs while the old PTE remains in others.
3203 */
3204 ptep_clear_flush(vma, vmf->address, vmf->pte);
3205 folio_add_new_anon_rmap(new_folio, vma, vmf->address);
3206 folio_add_lru_vma(new_folio, vma);
3207 /*
3208 * We call the notify macro here because, when using secondary
3209 * mmu page tables (such as kvm shadow page tables), we want the
3210 * new page to be mapped directly into the secondary page table.
3211 */
3212 BUG_ON(unshare && pte_write(entry));
3213 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3214 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3215 if (old_folio) {
3216 /*
3217 * Only after switching the pte to the new page may
3218 * we remove the mapcount here. Otherwise another
3219 * process may come and find the rmap count decremented
3220 * before the pte is switched to the new page, and
3221 * "reuse" the old page writing into it while our pte
3222 * here still points into it and can be read by other
3223 * threads.
3224 *
3225 * The critical issue is to order this
3226 * folio_remove_rmap_pte() with the ptp_clear_flush
3227 * above. Those stores are ordered by (if nothing else,)
3228 * the barrier present in the atomic_add_negative
3229 * in folio_remove_rmap_pte();
3230 *
3231 * Then the TLB flush in ptep_clear_flush ensures that
3232 * no process can access the old page before the
3233 * decremented mapcount is visible. And the old page
3234 * cannot be reused until after the decremented
3235 * mapcount is visible. So transitively, TLBs to
3236 * old page will be flushed before it can be reused.
3237 */
3238 folio_remove_rmap_pte(old_folio, vmf->page, vma);
3239 }
3240
3241 /* Free the old page.. */
3242 new_folio = old_folio;
3243 page_copied = 1;
3244 pte_unmap_unlock(vmf->pte, vmf->ptl);
3245 } else if (vmf->pte) {
3246 update_mmu_tlb(vma, vmf->address, vmf->pte);
3247 pte_unmap_unlock(vmf->pte, vmf->ptl);
3248 }
3249
3250 mmu_notifier_invalidate_range_end(&range);
3251
3252 if (new_folio)
3253 folio_put(new_folio);
3254 if (old_folio) {
3255 if (page_copied)
3256 free_swap_cache(&old_folio->page);
3257 folio_put(old_folio);
3258 }
3259
3260 delayacct_wpcopy_end();
3261 return 0;
3262oom:
3263 ret = VM_FAULT_OOM;
3264out:
3265 if (old_folio)
3266 folio_put(old_folio);
3267
3268 delayacct_wpcopy_end();
3269 return ret;
3270}
3271
3272/**
3273 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3274 * writeable once the page is prepared
3275 *
3276 * @vmf: structure describing the fault
3277 * @folio: the folio of vmf->page
3278 *
3279 * This function handles all that is needed to finish a write page fault in a
3280 * shared mapping due to PTE being read-only once the mapped page is prepared.
3281 * It handles locking of PTE and modifying it.
3282 *
3283 * The function expects the page to be locked or other protection against
3284 * concurrent faults / writeback (such as DAX radix tree locks).
3285 *
3286 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3287 * we acquired PTE lock.
3288 */
3289static vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf, struct folio *folio)
3290{
3291 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3292 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3293 &vmf->ptl);
3294 if (!vmf->pte)
3295 return VM_FAULT_NOPAGE;
3296 /*
3297 * We might have raced with another page fault while we released the
3298 * pte_offset_map_lock.
3299 */
3300 if (!pte_same(ptep_get(vmf->pte), vmf->orig_pte)) {
3301 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3302 pte_unmap_unlock(vmf->pte, vmf->ptl);
3303 return VM_FAULT_NOPAGE;
3304 }
3305 wp_page_reuse(vmf, folio);
3306 return 0;
3307}
3308
3309/*
3310 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3311 * mapping
3312 */
3313static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3314{
3315 struct vm_area_struct *vma = vmf->vma;
3316
3317 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3318 vm_fault_t ret;
3319
3320 pte_unmap_unlock(vmf->pte, vmf->ptl);
3321 ret = vmf_can_call_fault(vmf);
3322 if (ret)
3323 return ret;
3324
3325 vmf->flags |= FAULT_FLAG_MKWRITE;
3326 ret = vma->vm_ops->pfn_mkwrite(vmf);
3327 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3328 return ret;
3329 return finish_mkwrite_fault(vmf, NULL);
3330 }
3331 wp_page_reuse(vmf, NULL);
3332 return 0;
3333}
3334
3335static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio)
3336 __releases(vmf->ptl)
3337{
3338 struct vm_area_struct *vma = vmf->vma;
3339 vm_fault_t ret = 0;
3340
3341 folio_get(folio);
3342
3343 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3344 vm_fault_t tmp;
3345
3346 pte_unmap_unlock(vmf->pte, vmf->ptl);
3347 tmp = vmf_can_call_fault(vmf);
3348 if (tmp) {
3349 folio_put(folio);
3350 return tmp;
3351 }
3352
3353 tmp = do_page_mkwrite(vmf, folio);
3354 if (unlikely(!tmp || (tmp &
3355 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3356 folio_put(folio);
3357 return tmp;
3358 }
3359 tmp = finish_mkwrite_fault(vmf, folio);
3360 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3361 folio_unlock(folio);
3362 folio_put(folio);
3363 return tmp;
3364 }
3365 } else {
3366 wp_page_reuse(vmf, folio);
3367 folio_lock(folio);
3368 }
3369 ret |= fault_dirty_shared_page(vmf);
3370 folio_put(folio);
3371
3372 return ret;
3373}
3374
3375static bool wp_can_reuse_anon_folio(struct folio *folio,
3376 struct vm_area_struct *vma)
3377{
3378 /*
3379 * We have to verify under folio lock: these early checks are
3380 * just an optimization to avoid locking the folio and freeing
3381 * the swapcache if there is little hope that we can reuse.
3382 *
3383 * KSM doesn't necessarily raise the folio refcount.
3384 */
3385 if (folio_test_ksm(folio) || folio_ref_count(folio) > 3)
3386 return false;
3387 if (!folio_test_lru(folio))
3388 /*
3389 * We cannot easily detect+handle references from
3390 * remote LRU caches or references to LRU folios.
3391 */
3392 lru_add_drain();
3393 if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio))
3394 return false;
3395 if (!folio_trylock(folio))
3396 return false;
3397 if (folio_test_swapcache(folio))
3398 folio_free_swap(folio);
3399 if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) {
3400 folio_unlock(folio);
3401 return false;
3402 }
3403 /*
3404 * Ok, we've got the only folio reference from our mapping
3405 * and the folio is locked, it's dark out, and we're wearing
3406 * sunglasses. Hit it.
3407 */
3408 folio_move_anon_rmap(folio, vma);
3409 folio_unlock(folio);
3410 return true;
3411}
3412
3413/*
3414 * This routine handles present pages, when
3415 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3416 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3417 * (FAULT_FLAG_UNSHARE)
3418 *
3419 * It is done by copying the page to a new address and decrementing the
3420 * shared-page counter for the old page.
3421 *
3422 * Note that this routine assumes that the protection checks have been
3423 * done by the caller (the low-level page fault routine in most cases).
3424 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3425 * done any necessary COW.
3426 *
3427 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3428 * though the page will change only once the write actually happens. This
3429 * avoids a few races, and potentially makes it more efficient.
3430 *
3431 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3432 * but allow concurrent faults), with pte both mapped and locked.
3433 * We return with mmap_lock still held, but pte unmapped and unlocked.
3434 */
3435static vm_fault_t do_wp_page(struct vm_fault *vmf)
3436 __releases(vmf->ptl)
3437{
3438 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3439 struct vm_area_struct *vma = vmf->vma;
3440 struct folio *folio = NULL;
3441 pte_t pte;
3442
3443 if (likely(!unshare)) {
3444 if (userfaultfd_pte_wp(vma, ptep_get(vmf->pte))) {
3445 if (!userfaultfd_wp_async(vma)) {
3446 pte_unmap_unlock(vmf->pte, vmf->ptl);
3447 return handle_userfault(vmf, VM_UFFD_WP);
3448 }
3449
3450 /*
3451 * Nothing needed (cache flush, TLB invalidations,
3452 * etc.) because we're only removing the uffd-wp bit,
3453 * which is completely invisible to the user.
3454 */
3455 pte = pte_clear_uffd_wp(ptep_get(vmf->pte));
3456
3457 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3458 /*
3459 * Update this to be prepared for following up CoW
3460 * handling
3461 */
3462 vmf->orig_pte = pte;
3463 }
3464
3465 /*
3466 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3467 * is flushed in this case before copying.
3468 */
3469 if (unlikely(userfaultfd_wp(vmf->vma) &&
3470 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3471 flush_tlb_page(vmf->vma, vmf->address);
3472 }
3473
3474 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3475
3476 if (vmf->page)
3477 folio = page_folio(vmf->page);
3478
3479 /*
3480 * Shared mapping: we are guaranteed to have VM_WRITE and
3481 * FAULT_FLAG_WRITE set at this point.
3482 */
3483 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
3484 /*
3485 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3486 * VM_PFNMAP VMA.
3487 *
3488 * We should not cow pages in a shared writeable mapping.
3489 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3490 */
3491 if (!vmf->page)
3492 return wp_pfn_shared(vmf);
3493 return wp_page_shared(vmf, folio);
3494 }
3495
3496 /*
3497 * Private mapping: create an exclusive anonymous page copy if reuse
3498 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3499 *
3500 * If we encounter a page that is marked exclusive, we must reuse
3501 * the page without further checks.
3502 */
3503 if (folio && folio_test_anon(folio) &&
3504 (PageAnonExclusive(vmf->page) || wp_can_reuse_anon_folio(folio, vma))) {
3505 if (!PageAnonExclusive(vmf->page))
3506 SetPageAnonExclusive(vmf->page);
3507 if (unlikely(unshare)) {
3508 pte_unmap_unlock(vmf->pte, vmf->ptl);
3509 return 0;
3510 }
3511 wp_page_reuse(vmf, folio);
3512 return 0;
3513 }
3514 /*
3515 * Ok, we need to copy. Oh, well..
3516 */
3517 if (folio)
3518 folio_get(folio);
3519
3520 pte_unmap_unlock(vmf->pte, vmf->ptl);
3521#ifdef CONFIG_KSM
3522 if (folio && folio_test_ksm(folio))
3523 count_vm_event(COW_KSM);
3524#endif
3525 return wp_page_copy(vmf);
3526}
3527
3528static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3529 unsigned long start_addr, unsigned long end_addr,
3530 struct zap_details *details)
3531{
3532 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3533}
3534
3535static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3536 pgoff_t first_index,
3537 pgoff_t last_index,
3538 struct zap_details *details)
3539{
3540 struct vm_area_struct *vma;
3541 pgoff_t vba, vea, zba, zea;
3542
3543 vma_interval_tree_foreach(vma, root, first_index, last_index) {
3544 vba = vma->vm_pgoff;
3545 vea = vba + vma_pages(vma) - 1;
3546 zba = max(first_index, vba);
3547 zea = min(last_index, vea);
3548
3549 unmap_mapping_range_vma(vma,
3550 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3551 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3552 details);
3553 }
3554}
3555
3556/**
3557 * unmap_mapping_folio() - Unmap single folio from processes.
3558 * @folio: The locked folio to be unmapped.
3559 *
3560 * Unmap this folio from any userspace process which still has it mmaped.
3561 * Typically, for efficiency, the range of nearby pages has already been
3562 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3563 * truncation or invalidation holds the lock on a folio, it may find that
3564 * the page has been remapped again: and then uses unmap_mapping_folio()
3565 * to unmap it finally.
3566 */
3567void unmap_mapping_folio(struct folio *folio)
3568{
3569 struct address_space *mapping = folio->mapping;
3570 struct zap_details details = { };
3571 pgoff_t first_index;
3572 pgoff_t last_index;
3573
3574 VM_BUG_ON(!folio_test_locked(folio));
3575
3576 first_index = folio->index;
3577 last_index = folio_next_index(folio) - 1;
3578
3579 details.even_cows = false;
3580 details.single_folio = folio;
3581 details.zap_flags = ZAP_FLAG_DROP_MARKER;
3582
3583 i_mmap_lock_read(mapping);
3584 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3585 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3586 last_index, &details);
3587 i_mmap_unlock_read(mapping);
3588}
3589
3590/**
3591 * unmap_mapping_pages() - Unmap pages from processes.
3592 * @mapping: The address space containing pages to be unmapped.
3593 * @start: Index of first page to be unmapped.
3594 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3595 * @even_cows: Whether to unmap even private COWed pages.
3596 *
3597 * Unmap the pages in this address space from any userspace process which
3598 * has them mmaped. Generally, you want to remove COWed pages as well when
3599 * a file is being truncated, but not when invalidating pages from the page
3600 * cache.
3601 */
3602void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3603 pgoff_t nr, bool even_cows)
3604{
3605 struct zap_details details = { };
3606 pgoff_t first_index = start;
3607 pgoff_t last_index = start + nr - 1;
3608
3609 details.even_cows = even_cows;
3610 if (last_index < first_index)
3611 last_index = ULONG_MAX;
3612
3613 i_mmap_lock_read(mapping);
3614 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3615 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3616 last_index, &details);
3617 i_mmap_unlock_read(mapping);
3618}
3619EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3620
3621/**
3622 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3623 * address_space corresponding to the specified byte range in the underlying
3624 * file.
3625 *
3626 * @mapping: the address space containing mmaps to be unmapped.
3627 * @holebegin: byte in first page to unmap, relative to the start of
3628 * the underlying file. This will be rounded down to a PAGE_SIZE
3629 * boundary. Note that this is different from truncate_pagecache(), which
3630 * must keep the partial page. In contrast, we must get rid of
3631 * partial pages.
3632 * @holelen: size of prospective hole in bytes. This will be rounded
3633 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3634 * end of the file.
3635 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3636 * but 0 when invalidating pagecache, don't throw away private data.
3637 */
3638void unmap_mapping_range(struct address_space *mapping,
3639 loff_t const holebegin, loff_t const holelen, int even_cows)
3640{
3641 pgoff_t hba = (pgoff_t)(holebegin) >> PAGE_SHIFT;
3642 pgoff_t hlen = ((pgoff_t)(holelen) + PAGE_SIZE - 1) >> PAGE_SHIFT;
3643
3644 /* Check for overflow. */
3645 if (sizeof(holelen) > sizeof(hlen)) {
3646 long long holeend =
3647 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3648 if (holeend & ~(long long)ULONG_MAX)
3649 hlen = ULONG_MAX - hba + 1;
3650 }
3651
3652 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3653}
3654EXPORT_SYMBOL(unmap_mapping_range);
3655
3656/*
3657 * Restore a potential device exclusive pte to a working pte entry
3658 */
3659static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3660{
3661 struct folio *folio = page_folio(vmf->page);
3662 struct vm_area_struct *vma = vmf->vma;
3663 struct mmu_notifier_range range;
3664 vm_fault_t ret;
3665
3666 /*
3667 * We need a reference to lock the folio because we don't hold
3668 * the PTL so a racing thread can remove the device-exclusive
3669 * entry and unmap it. If the folio is free the entry must
3670 * have been removed already. If it happens to have already
3671 * been re-allocated after being freed all we do is lock and
3672 * unlock it.
3673 */
3674 if (!folio_try_get(folio))
3675 return 0;
3676
3677 ret = folio_lock_or_retry(folio, vmf);
3678 if (ret) {
3679 folio_put(folio);
3680 return ret;
3681 }
3682 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
3683 vma->vm_mm, vmf->address & PAGE_MASK,
3684 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3685 mmu_notifier_invalidate_range_start(&range);
3686
3687 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3688 &vmf->ptl);
3689 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3690 restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte);
3691
3692 if (vmf->pte)
3693 pte_unmap_unlock(vmf->pte, vmf->ptl);
3694 folio_unlock(folio);
3695 folio_put(folio);
3696
3697 mmu_notifier_invalidate_range_end(&range);
3698 return 0;
3699}
3700
3701static inline bool should_try_to_free_swap(struct folio *folio,
3702 struct vm_area_struct *vma,
3703 unsigned int fault_flags)
3704{
3705 if (!folio_test_swapcache(folio))
3706 return false;
3707 if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) ||
3708 folio_test_mlocked(folio))
3709 return true;
3710 /*
3711 * If we want to map a page that's in the swapcache writable, we
3712 * have to detect via the refcount if we're really the exclusive
3713 * user. Try freeing the swapcache to get rid of the swapcache
3714 * reference only in case it's likely that we'll be the exlusive user.
3715 */
3716 return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
3717 folio_ref_count(folio) == 2;
3718}
3719
3720static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
3721{
3722 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
3723 vmf->address, &vmf->ptl);
3724 if (!vmf->pte)
3725 return 0;
3726 /*
3727 * Be careful so that we will only recover a special uffd-wp pte into a
3728 * none pte. Otherwise it means the pte could have changed, so retry.
3729 *
3730 * This should also cover the case where e.g. the pte changed
3731 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3732 * So is_pte_marker() check is not enough to safely drop the pte.
3733 */
3734 if (pte_same(vmf->orig_pte, ptep_get(vmf->pte)))
3735 pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte);
3736 pte_unmap_unlock(vmf->pte, vmf->ptl);
3737 return 0;
3738}
3739
3740static vm_fault_t do_pte_missing(struct vm_fault *vmf)
3741{
3742 if (vma_is_anonymous(vmf->vma))
3743 return do_anonymous_page(vmf);
3744 else
3745 return do_fault(vmf);
3746}
3747
3748/*
3749 * This is actually a page-missing access, but with uffd-wp special pte
3750 * installed. It means this pte was wr-protected before being unmapped.
3751 */
3752static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
3753{
3754 /*
3755 * Just in case there're leftover special ptes even after the region
3756 * got unregistered - we can simply clear them.
3757 */
3758 if (unlikely(!userfaultfd_wp(vmf->vma)))
3759 return pte_marker_clear(vmf);
3760
3761 return do_pte_missing(vmf);
3762}
3763
3764static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
3765{
3766 swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte);
3767 unsigned long marker = pte_marker_get(entry);
3768
3769 /*
3770 * PTE markers should never be empty. If anything weird happened,
3771 * the best thing to do is to kill the process along with its mm.
3772 */
3773 if (WARN_ON_ONCE(!marker))
3774 return VM_FAULT_SIGBUS;
3775
3776 /* Higher priority than uffd-wp when data corrupted */
3777 if (marker & PTE_MARKER_POISONED)
3778 return VM_FAULT_HWPOISON;
3779
3780 if (pte_marker_entry_uffd_wp(entry))
3781 return pte_marker_handle_uffd_wp(vmf);
3782
3783 /* This is an unknown pte marker */
3784 return VM_FAULT_SIGBUS;
3785}
3786
3787/*
3788 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3789 * but allow concurrent faults), and pte mapped but not yet locked.
3790 * We return with pte unmapped and unlocked.
3791 *
3792 * We return with the mmap_lock locked or unlocked in the same cases
3793 * as does filemap_fault().
3794 */
3795vm_fault_t do_swap_page(struct vm_fault *vmf)
3796{
3797 struct vm_area_struct *vma = vmf->vma;
3798 struct folio *swapcache, *folio = NULL;
3799 struct page *page;
3800 struct swap_info_struct *si = NULL;
3801 rmap_t rmap_flags = RMAP_NONE;
3802 bool need_clear_cache = false;
3803 bool exclusive = false;
3804 swp_entry_t entry;
3805 pte_t pte;
3806 vm_fault_t ret = 0;
3807 void *shadow = NULL;
3808
3809 if (!pte_unmap_same(vmf))
3810 goto out;
3811
3812 entry = pte_to_swp_entry(vmf->orig_pte);
3813 if (unlikely(non_swap_entry(entry))) {
3814 if (is_migration_entry(entry)) {
3815 migration_entry_wait(vma->vm_mm, vmf->pmd,
3816 vmf->address);
3817 } else if (is_device_exclusive_entry(entry)) {
3818 vmf->page = pfn_swap_entry_to_page(entry);
3819 ret = remove_device_exclusive_entry(vmf);
3820 } else if (is_device_private_entry(entry)) {
3821 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3822 /*
3823 * migrate_to_ram is not yet ready to operate
3824 * under VMA lock.
3825 */
3826 vma_end_read(vma);
3827 ret = VM_FAULT_RETRY;
3828 goto out;
3829 }
3830
3831 vmf->page = pfn_swap_entry_to_page(entry);
3832 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3833 vmf->address, &vmf->ptl);
3834 if (unlikely(!vmf->pte ||
3835 !pte_same(ptep_get(vmf->pte),
3836 vmf->orig_pte)))
3837 goto unlock;
3838
3839 /*
3840 * Get a page reference while we know the page can't be
3841 * freed.
3842 */
3843 get_page(vmf->page);
3844 pte_unmap_unlock(vmf->pte, vmf->ptl);
3845 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3846 put_page(vmf->page);
3847 } else if (is_hwpoison_entry(entry)) {
3848 ret = VM_FAULT_HWPOISON;
3849 } else if (is_pte_marker_entry(entry)) {
3850 ret = handle_pte_marker(vmf);
3851 } else {
3852 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3853 ret = VM_FAULT_SIGBUS;
3854 }
3855 goto out;
3856 }
3857
3858 /* Prevent swapoff from happening to us. */
3859 si = get_swap_device(entry);
3860 if (unlikely(!si))
3861 goto out;
3862
3863 folio = swap_cache_get_folio(entry, vma, vmf->address);
3864 if (folio)
3865 page = folio_file_page(folio, swp_offset(entry));
3866 swapcache = folio;
3867
3868 if (!folio) {
3869 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3870 __swap_count(entry) == 1) {
3871 /*
3872 * Prevent parallel swapin from proceeding with
3873 * the cache flag. Otherwise, another thread may
3874 * finish swapin first, free the entry, and swapout
3875 * reusing the same entry. It's undetectable as
3876 * pte_same() returns true due to entry reuse.
3877 */
3878 if (swapcache_prepare(entry)) {
3879 /* Relax a bit to prevent rapid repeated page faults */
3880 schedule_timeout_uninterruptible(1);
3881 goto out;
3882 }
3883 need_clear_cache = true;
3884
3885 /* skip swapcache */
3886 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0,
3887 vma, vmf->address, false);
3888 page = &folio->page;
3889 if (folio) {
3890 __folio_set_locked(folio);
3891 __folio_set_swapbacked(folio);
3892
3893 if (mem_cgroup_swapin_charge_folio(folio,
3894 vma->vm_mm, GFP_KERNEL,
3895 entry)) {
3896 ret = VM_FAULT_OOM;
3897 goto out_page;
3898 }
3899 mem_cgroup_swapin_uncharge_swap(entry);
3900
3901 shadow = get_shadow_from_swap_cache(entry);
3902 if (shadow)
3903 workingset_refault(folio, shadow);
3904
3905 folio_add_lru(folio);
3906
3907 /* To provide entry to swap_read_folio() */
3908 folio->swap = entry;
3909 swap_read_folio(folio, true, NULL);
3910 folio->private = NULL;
3911 }
3912 } else {
3913 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3914 vmf);
3915 if (page)
3916 folio = page_folio(page);
3917 swapcache = folio;
3918 }
3919
3920 if (!folio) {
3921 /*
3922 * Back out if somebody else faulted in this pte
3923 * while we released the pte lock.
3924 */
3925 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3926 vmf->address, &vmf->ptl);
3927 if (likely(vmf->pte &&
3928 pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3929 ret = VM_FAULT_OOM;
3930 goto unlock;
3931 }
3932
3933 /* Had to read the page from swap area: Major fault */
3934 ret = VM_FAULT_MAJOR;
3935 count_vm_event(PGMAJFAULT);
3936 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3937 } else if (PageHWPoison(page)) {
3938 /*
3939 * hwpoisoned dirty swapcache pages are kept for killing
3940 * owner processes (which may be unknown at hwpoison time)
3941 */
3942 ret = VM_FAULT_HWPOISON;
3943 goto out_release;
3944 }
3945
3946 ret |= folio_lock_or_retry(folio, vmf);
3947 if (ret & VM_FAULT_RETRY)
3948 goto out_release;
3949
3950 if (swapcache) {
3951 /*
3952 * Make sure folio_free_swap() or swapoff did not release the
3953 * swapcache from under us. The page pin, and pte_same test
3954 * below, are not enough to exclude that. Even if it is still
3955 * swapcache, we need to check that the page's swap has not
3956 * changed.
3957 */
3958 if (unlikely(!folio_test_swapcache(folio) ||
3959 page_swap_entry(page).val != entry.val))
3960 goto out_page;
3961
3962 /*
3963 * KSM sometimes has to copy on read faults, for example, if
3964 * page->index of !PageKSM() pages would be nonlinear inside the
3965 * anon VMA -- PageKSM() is lost on actual swapout.
3966 */
3967 folio = ksm_might_need_to_copy(folio, vma, vmf->address);
3968 if (unlikely(!folio)) {
3969 ret = VM_FAULT_OOM;
3970 folio = swapcache;
3971 goto out_page;
3972 } else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
3973 ret = VM_FAULT_HWPOISON;
3974 folio = swapcache;
3975 goto out_page;
3976 }
3977 if (folio != swapcache)
3978 page = folio_page(folio, 0);
3979
3980 /*
3981 * If we want to map a page that's in the swapcache writable, we
3982 * have to detect via the refcount if we're really the exclusive
3983 * owner. Try removing the extra reference from the local LRU
3984 * caches if required.
3985 */
3986 if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache &&
3987 !folio_test_ksm(folio) && !folio_test_lru(folio))
3988 lru_add_drain();
3989 }
3990
3991 folio_throttle_swaprate(folio, GFP_KERNEL);
3992
3993 /*
3994 * Back out if somebody else already faulted in this pte.
3995 */
3996 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3997 &vmf->ptl);
3998 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3999 goto out_nomap;
4000
4001 if (unlikely(!folio_test_uptodate(folio))) {
4002 ret = VM_FAULT_SIGBUS;
4003 goto out_nomap;
4004 }
4005
4006 /*
4007 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
4008 * must never point at an anonymous page in the swapcache that is
4009 * PG_anon_exclusive. Sanity check that this holds and especially, that
4010 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
4011 * check after taking the PT lock and making sure that nobody
4012 * concurrently faulted in this page and set PG_anon_exclusive.
4013 */
4014 BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio));
4015 BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page));
4016
4017 /*
4018 * Check under PT lock (to protect against concurrent fork() sharing
4019 * the swap entry concurrently) for certainly exclusive pages.
4020 */
4021 if (!folio_test_ksm(folio)) {
4022 exclusive = pte_swp_exclusive(vmf->orig_pte);
4023 if (folio != swapcache) {
4024 /*
4025 * We have a fresh page that is not exposed to the
4026 * swapcache -> certainly exclusive.
4027 */
4028 exclusive = true;
4029 } else if (exclusive && folio_test_writeback(folio) &&
4030 data_race(si->flags & SWP_STABLE_WRITES)) {
4031 /*
4032 * This is tricky: not all swap backends support
4033 * concurrent page modifications while under writeback.
4034 *
4035 * So if we stumble over such a page in the swapcache
4036 * we must not set the page exclusive, otherwise we can
4037 * map it writable without further checks and modify it
4038 * while still under writeback.
4039 *
4040 * For these problematic swap backends, simply drop the
4041 * exclusive marker: this is perfectly fine as we start
4042 * writeback only if we fully unmapped the page and
4043 * there are no unexpected references on the page after
4044 * unmapping succeeded. After fully unmapped, no
4045 * further GUP references (FOLL_GET and FOLL_PIN) can
4046 * appear, so dropping the exclusive marker and mapping
4047 * it only R/O is fine.
4048 */
4049 exclusive = false;
4050 }
4051 }
4052
4053 /*
4054 * Some architectures may have to restore extra metadata to the page
4055 * when reading from swap. This metadata may be indexed by swap entry
4056 * so this must be called before swap_free().
4057 */
4058 arch_swap_restore(entry, folio);
4059
4060 /*
4061 * Remove the swap entry and conditionally try to free up the swapcache.
4062 * We're already holding a reference on the page but haven't mapped it
4063 * yet.
4064 */
4065 swap_free(entry);
4066 if (should_try_to_free_swap(folio, vma, vmf->flags))
4067 folio_free_swap(folio);
4068
4069 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
4070 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
4071 pte = mk_pte(page, vma->vm_page_prot);
4072
4073 /*
4074 * Same logic as in do_wp_page(); however, optimize for pages that are
4075 * certainly not shared either because we just allocated them without
4076 * exposing them to the swapcache or because the swap entry indicates
4077 * exclusivity.
4078 */
4079 if (!folio_test_ksm(folio) &&
4080 (exclusive || folio_ref_count(folio) == 1)) {
4081 if (vmf->flags & FAULT_FLAG_WRITE) {
4082 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
4083 vmf->flags &= ~FAULT_FLAG_WRITE;
4084 }
4085 rmap_flags |= RMAP_EXCLUSIVE;
4086 }
4087 flush_icache_page(vma, page);
4088 if (pte_swp_soft_dirty(vmf->orig_pte))
4089 pte = pte_mksoft_dirty(pte);
4090 if (pte_swp_uffd_wp(vmf->orig_pte))
4091 pte = pte_mkuffd_wp(pte);
4092 vmf->orig_pte = pte;
4093
4094 /* ksm created a completely new copy */
4095 if (unlikely(folio != swapcache && swapcache)) {
4096 folio_add_new_anon_rmap(folio, vma, vmf->address);
4097 folio_add_lru_vma(folio, vma);
4098 } else {
4099 folio_add_anon_rmap_pte(folio, page, vma, vmf->address,
4100 rmap_flags);
4101 }
4102
4103 VM_BUG_ON(!folio_test_anon(folio) ||
4104 (pte_write(pte) && !PageAnonExclusive(page)));
4105 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
4106 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
4107
4108 folio_unlock(folio);
4109 if (folio != swapcache && swapcache) {
4110 /*
4111 * Hold the lock to avoid the swap entry to be reused
4112 * until we take the PT lock for the pte_same() check
4113 * (to avoid false positives from pte_same). For
4114 * further safety release the lock after the swap_free
4115 * so that the swap count won't change under a
4116 * parallel locked swapcache.
4117 */
4118 folio_unlock(swapcache);
4119 folio_put(swapcache);
4120 }
4121
4122 if (vmf->flags & FAULT_FLAG_WRITE) {
4123 ret |= do_wp_page(vmf);
4124 if (ret & VM_FAULT_ERROR)
4125 ret &= VM_FAULT_ERROR;
4126 goto out;
4127 }
4128
4129 /* No need to invalidate - it was non-present before */
4130 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
4131unlock:
4132 if (vmf->pte)
4133 pte_unmap_unlock(vmf->pte, vmf->ptl);
4134out:
4135 /* Clear the swap cache pin for direct swapin after PTL unlock */
4136 if (need_clear_cache)
4137 swapcache_clear(si, entry);
4138 if (si)
4139 put_swap_device(si);
4140 return ret;
4141out_nomap:
4142 if (vmf->pte)
4143 pte_unmap_unlock(vmf->pte, vmf->ptl);
4144out_page:
4145 folio_unlock(folio);
4146out_release:
4147 folio_put(folio);
4148 if (folio != swapcache && swapcache) {
4149 folio_unlock(swapcache);
4150 folio_put(swapcache);
4151 }
4152 if (need_clear_cache)
4153 swapcache_clear(si, entry);
4154 if (si)
4155 put_swap_device(si);
4156 return ret;
4157}
4158
4159static bool pte_range_none(pte_t *pte, int nr_pages)
4160{
4161 int i;
4162
4163 for (i = 0; i < nr_pages; i++) {
4164 if (!pte_none(ptep_get_lockless(pte + i)))
4165 return false;
4166 }
4167
4168 return true;
4169}
4170
4171static struct folio *alloc_anon_folio(struct vm_fault *vmf)
4172{
4173#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4174 struct vm_area_struct *vma = vmf->vma;
4175 unsigned long orders;
4176 struct folio *folio;
4177 unsigned long addr;
4178 pte_t *pte;
4179 gfp_t gfp;
4180 int order;
4181
4182 /*
4183 * If uffd is active for the vma we need per-page fault fidelity to
4184 * maintain the uffd semantics.
4185 */
4186 if (unlikely(userfaultfd_armed(vma)))
4187 goto fallback;
4188
4189 /*
4190 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4191 * for this vma. Then filter out the orders that can't be allocated over
4192 * the faulting address and still be fully contained in the vma.
4193 */
4194 orders = thp_vma_allowable_orders(vma, vma->vm_flags, false, true, true,
4195 BIT(PMD_ORDER) - 1);
4196 orders = thp_vma_suitable_orders(vma, vmf->address, orders);
4197
4198 if (!orders)
4199 goto fallback;
4200
4201 pte = pte_offset_map(vmf->pmd, vmf->address & PMD_MASK);
4202 if (!pte)
4203 return ERR_PTR(-EAGAIN);
4204
4205 /*
4206 * Find the highest order where the aligned range is completely
4207 * pte_none(). Note that all remaining orders will be completely
4208 * pte_none().
4209 */
4210 order = highest_order(orders);
4211 while (orders) {
4212 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4213 if (pte_range_none(pte + pte_index(addr), 1 << order))
4214 break;
4215 order = next_order(&orders, order);
4216 }
4217
4218 pte_unmap(pte);
4219
4220 /* Try allocating the highest of the remaining orders. */
4221 gfp = vma_thp_gfp_mask(vma);
4222 while (orders) {
4223 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4224 folio = vma_alloc_folio(gfp, order, vma, addr, true);
4225 if (folio) {
4226 clear_huge_page(&folio->page, vmf->address, 1 << order);
4227 return folio;
4228 }
4229 order = next_order(&orders, order);
4230 }
4231
4232fallback:
4233#endif
4234 return vma_alloc_zeroed_movable_folio(vmf->vma, vmf->address);
4235}
4236
4237/*
4238 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4239 * but allow concurrent faults), and pte mapped but not yet locked.
4240 * We return with mmap_lock still held, but pte unmapped and unlocked.
4241 */
4242static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
4243{
4244 bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4245 struct vm_area_struct *vma = vmf->vma;
4246 unsigned long addr = vmf->address;
4247 struct folio *folio;
4248 vm_fault_t ret = 0;
4249 int nr_pages = 1;
4250 pte_t entry;
4251 int i;
4252
4253 /* File mapping without ->vm_ops ? */
4254 if (vma->vm_flags & VM_SHARED)
4255 return VM_FAULT_SIGBUS;
4256
4257 /*
4258 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4259 * be distinguished from a transient failure of pte_offset_map().
4260 */
4261 if (pte_alloc(vma->vm_mm, vmf->pmd))
4262 return VM_FAULT_OOM;
4263
4264 /* Use the zero-page for reads */
4265 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
4266 !mm_forbids_zeropage(vma->vm_mm)) {
4267 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
4268 vma->vm_page_prot));
4269 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4270 vmf->address, &vmf->ptl);
4271 if (!vmf->pte)
4272 goto unlock;
4273 if (vmf_pte_changed(vmf)) {
4274 update_mmu_tlb(vma, vmf->address, vmf->pte);
4275 goto unlock;
4276 }
4277 ret = check_stable_address_space(vma->vm_mm);
4278 if (ret)
4279 goto unlock;
4280 /* Deliver the page fault to userland, check inside PT lock */
4281 if (userfaultfd_missing(vma)) {
4282 pte_unmap_unlock(vmf->pte, vmf->ptl);
4283 return handle_userfault(vmf, VM_UFFD_MISSING);
4284 }
4285 goto setpte;
4286 }
4287
4288 /* Allocate our own private page. */
4289 if (unlikely(anon_vma_prepare(vma)))
4290 goto oom;
4291 /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
4292 folio = alloc_anon_folio(vmf);
4293 if (IS_ERR(folio))
4294 return 0;
4295 if (!folio)
4296 goto oom;
4297
4298 nr_pages = folio_nr_pages(folio);
4299 addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE);
4300
4301 if (mem_cgroup_charge(folio, vma->vm_mm, GFP_KERNEL))
4302 goto oom_free_page;
4303 folio_throttle_swaprate(folio, GFP_KERNEL);
4304
4305 /*
4306 * The memory barrier inside __folio_mark_uptodate makes sure that
4307 * preceding stores to the page contents become visible before
4308 * the set_pte_at() write.
4309 */
4310 __folio_mark_uptodate(folio);
4311
4312 entry = mk_pte(&folio->page, vma->vm_page_prot);
4313 entry = pte_sw_mkyoung(entry);
4314 if (vma->vm_flags & VM_WRITE)
4315 entry = pte_mkwrite(pte_mkdirty(entry), vma);
4316
4317 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
4318 if (!vmf->pte)
4319 goto release;
4320 if (nr_pages == 1 && vmf_pte_changed(vmf)) {
4321 update_mmu_tlb(vma, addr, vmf->pte);
4322 goto release;
4323 } else if (nr_pages > 1 && !pte_range_none(vmf->pte, nr_pages)) {
4324 for (i = 0; i < nr_pages; i++)
4325 update_mmu_tlb(vma, addr + PAGE_SIZE * i, vmf->pte + i);
4326 goto release;
4327 }
4328
4329 ret = check_stable_address_space(vma->vm_mm);
4330 if (ret)
4331 goto release;
4332
4333 /* Deliver the page fault to userland, check inside PT lock */
4334 if (userfaultfd_missing(vma)) {
4335 pte_unmap_unlock(vmf->pte, vmf->ptl);
4336 folio_put(folio);
4337 return handle_userfault(vmf, VM_UFFD_MISSING);
4338 }
4339
4340 folio_ref_add(folio, nr_pages - 1);
4341 add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr_pages);
4342 folio_add_new_anon_rmap(folio, vma, addr);
4343 folio_add_lru_vma(folio, vma);
4344setpte:
4345 if (uffd_wp)
4346 entry = pte_mkuffd_wp(entry);
4347 set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr_pages);
4348
4349 /* No need to invalidate - it was non-present before */
4350 update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr_pages);
4351unlock:
4352 if (vmf->pte)
4353 pte_unmap_unlock(vmf->pte, vmf->ptl);
4354 return ret;
4355release:
4356 folio_put(folio);
4357 goto unlock;
4358oom_free_page:
4359 folio_put(folio);
4360oom:
4361 return VM_FAULT_OOM;
4362}
4363
4364/*
4365 * The mmap_lock must have been held on entry, and may have been
4366 * released depending on flags and vma->vm_ops->fault() return value.
4367 * See filemap_fault() and __lock_page_retry().
4368 */
4369static vm_fault_t __do_fault(struct vm_fault *vmf)
4370{
4371 struct vm_area_struct *vma = vmf->vma;
4372 struct folio *folio;
4373 vm_fault_t ret;
4374
4375 /*
4376 * Preallocate pte before we take page_lock because this might lead to
4377 * deadlocks for memcg reclaim which waits for pages under writeback:
4378 * lock_page(A)
4379 * SetPageWriteback(A)
4380 * unlock_page(A)
4381 * lock_page(B)
4382 * lock_page(B)
4383 * pte_alloc_one
4384 * shrink_page_list
4385 * wait_on_page_writeback(A)
4386 * SetPageWriteback(B)
4387 * unlock_page(B)
4388 * # flush A, B to clear the writeback
4389 */
4390 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
4391 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4392 if (!vmf->prealloc_pte)
4393 return VM_FAULT_OOM;
4394 }
4395
4396 ret = vma->vm_ops->fault(vmf);
4397 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
4398 VM_FAULT_DONE_COW)))
4399 return ret;
4400
4401 folio = page_folio(vmf->page);
4402 if (unlikely(PageHWPoison(vmf->page))) {
4403 vm_fault_t poisonret = VM_FAULT_HWPOISON;
4404 if (ret & VM_FAULT_LOCKED) {
4405 if (page_mapped(vmf->page))
4406 unmap_mapping_folio(folio);
4407 /* Retry if a clean folio was removed from the cache. */
4408 if (mapping_evict_folio(folio->mapping, folio))
4409 poisonret = VM_FAULT_NOPAGE;
4410 folio_unlock(folio);
4411 }
4412 folio_put(folio);
4413 vmf->page = NULL;
4414 return poisonret;
4415 }
4416
4417 if (unlikely(!(ret & VM_FAULT_LOCKED)))
4418 folio_lock(folio);
4419 else
4420 VM_BUG_ON_PAGE(!folio_test_locked(folio), vmf->page);
4421
4422 return ret;
4423}
4424
4425#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4426static void deposit_prealloc_pte(struct vm_fault *vmf)
4427{
4428 struct vm_area_struct *vma = vmf->vma;
4429
4430 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4431 /*
4432 * We are going to consume the prealloc table,
4433 * count that as nr_ptes.
4434 */
4435 mm_inc_nr_ptes(vma->vm_mm);
4436 vmf->prealloc_pte = NULL;
4437}
4438
4439vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4440{
4441 struct folio *folio = page_folio(page);
4442 struct vm_area_struct *vma = vmf->vma;
4443 bool write = vmf->flags & FAULT_FLAG_WRITE;
4444 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
4445 pmd_t entry;
4446 vm_fault_t ret = VM_FAULT_FALLBACK;
4447
4448 if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
4449 return ret;
4450
4451 if (page != &folio->page || folio_order(folio) != HPAGE_PMD_ORDER)
4452 return ret;
4453
4454 /*
4455 * Just backoff if any subpage of a THP is corrupted otherwise
4456 * the corrupted page may mapped by PMD silently to escape the
4457 * check. This kind of THP just can be PTE mapped. Access to
4458 * the corrupted subpage should trigger SIGBUS as expected.
4459 */
4460 if (unlikely(folio_test_has_hwpoisoned(folio)))
4461 return ret;
4462
4463 /*
4464 * Archs like ppc64 need additional space to store information
4465 * related to pte entry. Use the preallocated table for that.
4466 */
4467 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
4468 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4469 if (!vmf->prealloc_pte)
4470 return VM_FAULT_OOM;
4471 }
4472
4473 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4474 if (unlikely(!pmd_none(*vmf->pmd)))
4475 goto out;
4476
4477 flush_icache_pages(vma, page, HPAGE_PMD_NR);
4478
4479 entry = mk_huge_pmd(page, vma->vm_page_prot);
4480 if (write)
4481 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
4482
4483 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
4484 folio_add_file_rmap_pmd(folio, page, vma);
4485
4486 /*
4487 * deposit and withdraw with pmd lock held
4488 */
4489 if (arch_needs_pgtable_deposit())
4490 deposit_prealloc_pte(vmf);
4491
4492 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
4493
4494 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
4495
4496 /* fault is handled */
4497 ret = 0;
4498 count_vm_event(THP_FILE_MAPPED);
4499out:
4500 spin_unlock(vmf->ptl);
4501 return ret;
4502}
4503#else
4504vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4505{
4506 return VM_FAULT_FALLBACK;
4507}
4508#endif
4509
4510/**
4511 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4512 * @vmf: Fault decription.
4513 * @folio: The folio that contains @page.
4514 * @page: The first page to create a PTE for.
4515 * @nr: The number of PTEs to create.
4516 * @addr: The first address to create a PTE for.
4517 */
4518void set_pte_range(struct vm_fault *vmf, struct folio *folio,
4519 struct page *page, unsigned int nr, unsigned long addr)
4520{
4521 struct vm_area_struct *vma = vmf->vma;
4522 bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4523 bool write = vmf->flags & FAULT_FLAG_WRITE;
4524 bool prefault = in_range(vmf->address, addr, nr * PAGE_SIZE);
4525 pte_t entry;
4526
4527 flush_icache_pages(vma, page, nr);
4528 entry = mk_pte(page, vma->vm_page_prot);
4529
4530 if (prefault && arch_wants_old_prefaulted_pte())
4531 entry = pte_mkold(entry);
4532 else
4533 entry = pte_sw_mkyoung(entry);
4534
4535 if (write)
4536 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4537 if (unlikely(uffd_wp))
4538 entry = pte_mkuffd_wp(entry);
4539 /* copy-on-write page */
4540 if (write && !(vma->vm_flags & VM_SHARED)) {
4541 add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr);
4542 VM_BUG_ON_FOLIO(nr != 1, folio);
4543 folio_add_new_anon_rmap(folio, vma, addr);
4544 folio_add_lru_vma(folio, vma);
4545 } else {
4546 add_mm_counter(vma->vm_mm, mm_counter_file(page), nr);
4547 folio_add_file_rmap_ptes(folio, page, nr, vma);
4548 }
4549 set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr);
4550
4551 /* no need to invalidate: a not-present page won't be cached */
4552 update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr);
4553}
4554
4555static bool vmf_pte_changed(struct vm_fault *vmf)
4556{
4557 if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)
4558 return !pte_same(ptep_get(vmf->pte), vmf->orig_pte);
4559
4560 return !pte_none(ptep_get(vmf->pte));
4561}
4562
4563/**
4564 * finish_fault - finish page fault once we have prepared the page to fault
4565 *
4566 * @vmf: structure describing the fault
4567 *
4568 * This function handles all that is needed to finish a page fault once the
4569 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4570 * given page, adds reverse page mapping, handles memcg charges and LRU
4571 * addition.
4572 *
4573 * The function expects the page to be locked and on success it consumes a
4574 * reference of a page being mapped (for the PTE which maps it).
4575 *
4576 * Return: %0 on success, %VM_FAULT_ code in case of error.
4577 */
4578vm_fault_t finish_fault(struct vm_fault *vmf)
4579{
4580 struct vm_area_struct *vma = vmf->vma;
4581 struct page *page;
4582 vm_fault_t ret;
4583
4584 /* Did we COW the page? */
4585 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4586 page = vmf->cow_page;
4587 else
4588 page = vmf->page;
4589
4590 /*
4591 * check even for read faults because we might have lost our CoWed
4592 * page
4593 */
4594 if (!(vma->vm_flags & VM_SHARED)) {
4595 ret = check_stable_address_space(vma->vm_mm);
4596 if (ret)
4597 return ret;
4598 }
4599
4600 if (pmd_none(*vmf->pmd)) {
4601 if (PageTransCompound(page)) {
4602 ret = do_set_pmd(vmf, page);
4603 if (ret != VM_FAULT_FALLBACK)
4604 return ret;
4605 }
4606
4607 if (vmf->prealloc_pte)
4608 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte);
4609 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4610 return VM_FAULT_OOM;
4611 }
4612
4613 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4614 vmf->address, &vmf->ptl);
4615 if (!vmf->pte)
4616 return VM_FAULT_NOPAGE;
4617
4618 /* Re-check under ptl */
4619 if (likely(!vmf_pte_changed(vmf))) {
4620 struct folio *folio = page_folio(page);
4621
4622 set_pte_range(vmf, folio, page, 1, vmf->address);
4623 ret = 0;
4624 } else {
4625 update_mmu_tlb(vma, vmf->address, vmf->pte);
4626 ret = VM_FAULT_NOPAGE;
4627 }
4628
4629 pte_unmap_unlock(vmf->pte, vmf->ptl);
4630 return ret;
4631}
4632
4633static unsigned long fault_around_pages __read_mostly =
4634 65536 >> PAGE_SHIFT;
4635
4636#ifdef CONFIG_DEBUG_FS
4637static int fault_around_bytes_get(void *data, u64 *val)
4638{
4639 *val = fault_around_pages << PAGE_SHIFT;
4640 return 0;
4641}
4642
4643/*
4644 * fault_around_bytes must be rounded down to the nearest page order as it's
4645 * what do_fault_around() expects to see.
4646 */
4647static int fault_around_bytes_set(void *data, u64 val)
4648{
4649 if (val / PAGE_SIZE > PTRS_PER_PTE)
4650 return -EINVAL;
4651
4652 /*
4653 * The minimum value is 1 page, however this results in no fault-around
4654 * at all. See should_fault_around().
4655 */
4656 fault_around_pages = max(rounddown_pow_of_two(val) >> PAGE_SHIFT, 1UL);
4657
4658 return 0;
4659}
4660DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4661 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4662
4663static int __init fault_around_debugfs(void)
4664{
4665 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4666 &fault_around_bytes_fops);
4667 return 0;
4668}
4669late_initcall(fault_around_debugfs);
4670#endif
4671
4672/*
4673 * do_fault_around() tries to map few pages around the fault address. The hope
4674 * is that the pages will be needed soon and this will lower the number of
4675 * faults to handle.
4676 *
4677 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4678 * not ready to be mapped: not up-to-date, locked, etc.
4679 *
4680 * This function doesn't cross VMA or page table boundaries, in order to call
4681 * map_pages() and acquire a PTE lock only once.
4682 *
4683 * fault_around_pages defines how many pages we'll try to map.
4684 * do_fault_around() expects it to be set to a power of two less than or equal
4685 * to PTRS_PER_PTE.
4686 *
4687 * The virtual address of the area that we map is naturally aligned to
4688 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4689 * (and therefore to page order). This way it's easier to guarantee
4690 * that we don't cross page table boundaries.
4691 */
4692static vm_fault_t do_fault_around(struct vm_fault *vmf)
4693{
4694 pgoff_t nr_pages = READ_ONCE(fault_around_pages);
4695 pgoff_t pte_off = pte_index(vmf->address);
4696 /* The page offset of vmf->address within the VMA. */
4697 pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
4698 pgoff_t from_pte, to_pte;
4699 vm_fault_t ret;
4700
4701 /* The PTE offset of the start address, clamped to the VMA. */
4702 from_pte = max(ALIGN_DOWN(pte_off, nr_pages),
4703 pte_off - min(pte_off, vma_off));
4704
4705 /* The PTE offset of the end address, clamped to the VMA and PTE. */
4706 to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE,
4707 pte_off + vma_pages(vmf->vma) - vma_off) - 1;
4708
4709 if (pmd_none(*vmf->pmd)) {
4710 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4711 if (!vmf->prealloc_pte)
4712 return VM_FAULT_OOM;
4713 }
4714
4715 rcu_read_lock();
4716 ret = vmf->vma->vm_ops->map_pages(vmf,
4717 vmf->pgoff + from_pte - pte_off,
4718 vmf->pgoff + to_pte - pte_off);
4719 rcu_read_unlock();
4720
4721 return ret;
4722}
4723
4724/* Return true if we should do read fault-around, false otherwise */
4725static inline bool should_fault_around(struct vm_fault *vmf)
4726{
4727 /* No ->map_pages? No way to fault around... */
4728 if (!vmf->vma->vm_ops->map_pages)
4729 return false;
4730
4731 if (uffd_disable_fault_around(vmf->vma))
4732 return false;
4733
4734 /* A single page implies no faulting 'around' at all. */
4735 return fault_around_pages > 1;
4736}
4737
4738static vm_fault_t do_read_fault(struct vm_fault *vmf)
4739{
4740 vm_fault_t ret = 0;
4741 struct folio *folio;
4742
4743 /*
4744 * Let's call ->map_pages() first and use ->fault() as fallback
4745 * if page by the offset is not ready to be mapped (cold cache or
4746 * something).
4747 */
4748 if (should_fault_around(vmf)) {
4749 ret = do_fault_around(vmf);
4750 if (ret)
4751 return ret;
4752 }
4753
4754 ret = vmf_can_call_fault(vmf);
4755 if (ret)
4756 return ret;
4757
4758 ret = __do_fault(vmf);
4759 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4760 return ret;
4761
4762 ret |= finish_fault(vmf);
4763 folio = page_folio(vmf->page);
4764 folio_unlock(folio);
4765 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4766 folio_put(folio);
4767 return ret;
4768}
4769
4770static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4771{
4772 struct vm_area_struct *vma = vmf->vma;
4773 struct folio *folio;
4774 vm_fault_t ret;
4775
4776 ret = vmf_can_call_fault(vmf);
4777 if (!ret)
4778 ret = vmf_anon_prepare(vmf);
4779 if (ret)
4780 return ret;
4781
4782 folio = folio_prealloc(vma->vm_mm, vma, vmf->address, false);
4783 if (!folio)
4784 return VM_FAULT_OOM;
4785
4786 vmf->cow_page = &folio->page;
4787
4788 ret = __do_fault(vmf);
4789 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4790 goto uncharge_out;
4791 if (ret & VM_FAULT_DONE_COW)
4792 return ret;
4793
4794 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4795 __folio_mark_uptodate(folio);
4796
4797 ret |= finish_fault(vmf);
4798 unlock_page(vmf->page);
4799 put_page(vmf->page);
4800 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4801 goto uncharge_out;
4802 return ret;
4803uncharge_out:
4804 folio_put(folio);
4805 return ret;
4806}
4807
4808static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4809{
4810 struct vm_area_struct *vma = vmf->vma;
4811 vm_fault_t ret, tmp;
4812 struct folio *folio;
4813
4814 ret = vmf_can_call_fault(vmf);
4815 if (ret)
4816 return ret;
4817
4818 ret = __do_fault(vmf);
4819 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4820 return ret;
4821
4822 folio = page_folio(vmf->page);
4823
4824 /*
4825 * Check if the backing address space wants to know that the page is
4826 * about to become writable
4827 */
4828 if (vma->vm_ops->page_mkwrite) {
4829 folio_unlock(folio);
4830 tmp = do_page_mkwrite(vmf, folio);
4831 if (unlikely(!tmp ||
4832 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4833 folio_put(folio);
4834 return tmp;
4835 }
4836 }
4837
4838 ret |= finish_fault(vmf);
4839 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4840 VM_FAULT_RETRY))) {
4841 folio_unlock(folio);
4842 folio_put(folio);
4843 return ret;
4844 }
4845
4846 ret |= fault_dirty_shared_page(vmf);
4847 return ret;
4848}
4849
4850/*
4851 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4852 * but allow concurrent faults).
4853 * The mmap_lock may have been released depending on flags and our
4854 * return value. See filemap_fault() and __folio_lock_or_retry().
4855 * If mmap_lock is released, vma may become invalid (for example
4856 * by other thread calling munmap()).
4857 */
4858static vm_fault_t do_fault(struct vm_fault *vmf)
4859{
4860 struct vm_area_struct *vma = vmf->vma;
4861 struct mm_struct *vm_mm = vma->vm_mm;
4862 vm_fault_t ret;
4863
4864 /*
4865 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4866 */
4867 if (!vma->vm_ops->fault) {
4868 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
4869 vmf->address, &vmf->ptl);
4870 if (unlikely(!vmf->pte))
4871 ret = VM_FAULT_SIGBUS;
4872 else {
4873 /*
4874 * Make sure this is not a temporary clearing of pte
4875 * by holding ptl and checking again. A R/M/W update
4876 * of pte involves: take ptl, clearing the pte so that
4877 * we don't have concurrent modification by hardware
4878 * followed by an update.
4879 */
4880 if (unlikely(pte_none(ptep_get(vmf->pte))))
4881 ret = VM_FAULT_SIGBUS;
4882 else
4883 ret = VM_FAULT_NOPAGE;
4884
4885 pte_unmap_unlock(vmf->pte, vmf->ptl);
4886 }
4887 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4888 ret = do_read_fault(vmf);
4889 else if (!(vma->vm_flags & VM_SHARED))
4890 ret = do_cow_fault(vmf);
4891 else
4892 ret = do_shared_fault(vmf);
4893
4894 /* preallocated pagetable is unused: free it */
4895 if (vmf->prealloc_pte) {
4896 pte_free(vm_mm, vmf->prealloc_pte);
4897 vmf->prealloc_pte = NULL;
4898 }
4899 return ret;
4900}
4901
4902int numa_migrate_prep(struct folio *folio, struct vm_area_struct *vma,
4903 unsigned long addr, int page_nid, int *flags)
4904{
4905 folio_get(folio);
4906
4907 /* Record the current PID acceesing VMA */
4908 vma_set_access_pid_bit(vma);
4909
4910 count_vm_numa_event(NUMA_HINT_FAULTS);
4911 if (page_nid == numa_node_id()) {
4912 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4913 *flags |= TNF_FAULT_LOCAL;
4914 }
4915
4916 return mpol_misplaced(folio, vma, addr);
4917}
4918
4919static vm_fault_t do_numa_page(struct vm_fault *vmf)
4920{
4921 struct vm_area_struct *vma = vmf->vma;
4922 struct folio *folio = NULL;
4923 int nid = NUMA_NO_NODE;
4924 bool writable = false;
4925 int last_cpupid;
4926 int target_nid;
4927 pte_t pte, old_pte;
4928 int flags = 0;
4929
4930 /*
4931 * The "pte" at this point cannot be used safely without
4932 * validation through pte_unmap_same(). It's of NUMA type but
4933 * the pfn may be screwed if the read is non atomic.
4934 */
4935 spin_lock(vmf->ptl);
4936 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
4937 pte_unmap_unlock(vmf->pte, vmf->ptl);
4938 goto out;
4939 }
4940
4941 /* Get the normal PTE */
4942 old_pte = ptep_get(vmf->pte);
4943 pte = pte_modify(old_pte, vma->vm_page_prot);
4944
4945 /*
4946 * Detect now whether the PTE could be writable; this information
4947 * is only valid while holding the PT lock.
4948 */
4949 writable = pte_write(pte);
4950 if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
4951 can_change_pte_writable(vma, vmf->address, pte))
4952 writable = true;
4953
4954 folio = vm_normal_folio(vma, vmf->address, pte);
4955 if (!folio || folio_is_zone_device(folio))
4956 goto out_map;
4957
4958 /* TODO: handle PTE-mapped THP */
4959 if (folio_test_large(folio))
4960 goto out_map;
4961
4962 /*
4963 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4964 * much anyway since they can be in shared cache state. This misses
4965 * the case where a mapping is writable but the process never writes
4966 * to it but pte_write gets cleared during protection updates and
4967 * pte_dirty has unpredictable behaviour between PTE scan updates,
4968 * background writeback, dirty balancing and application behaviour.
4969 */
4970 if (!writable)
4971 flags |= TNF_NO_GROUP;
4972
4973 /*
4974 * Flag if the folio is shared between multiple address spaces. This
4975 * is later used when determining whether to group tasks together
4976 */
4977 if (folio_estimated_sharers(folio) > 1 && (vma->vm_flags & VM_SHARED))
4978 flags |= TNF_SHARED;
4979
4980 nid = folio_nid(folio);
4981 /*
4982 * For memory tiering mode, cpupid of slow memory page is used
4983 * to record page access time. So use default value.
4984 */
4985 if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4986 !node_is_toptier(nid))
4987 last_cpupid = (-1 & LAST_CPUPID_MASK);
4988 else
4989 last_cpupid = folio_last_cpupid(folio);
4990 target_nid = numa_migrate_prep(folio, vma, vmf->address, nid, &flags);
4991 if (target_nid == NUMA_NO_NODE) {
4992 folio_put(folio);
4993 goto out_map;
4994 }
4995 pte_unmap_unlock(vmf->pte, vmf->ptl);
4996 writable = false;
4997
4998 /* Migrate to the requested node */
4999 if (migrate_misplaced_folio(folio, vma, target_nid)) {
5000 nid = target_nid;
5001 flags |= TNF_MIGRATED;
5002 } else {
5003 flags |= TNF_MIGRATE_FAIL;
5004 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
5005 vmf->address, &vmf->ptl);
5006 if (unlikely(!vmf->pte))
5007 goto out;
5008 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
5009 pte_unmap_unlock(vmf->pte, vmf->ptl);
5010 goto out;
5011 }
5012 goto out_map;
5013 }
5014
5015out:
5016 if (nid != NUMA_NO_NODE)
5017 task_numa_fault(last_cpupid, nid, 1, flags);
5018 return 0;
5019out_map:
5020 /*
5021 * Make it present again, depending on how arch implements
5022 * non-accessible ptes, some can allow access by kernel mode.
5023 */
5024 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
5025 pte = pte_modify(old_pte, vma->vm_page_prot);
5026 pte = pte_mkyoung(pte);
5027 if (writable)
5028 pte = pte_mkwrite(pte, vma);
5029 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
5030 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
5031 pte_unmap_unlock(vmf->pte, vmf->ptl);
5032 goto out;
5033}
5034
5035static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
5036{
5037 struct vm_area_struct *vma = vmf->vma;
5038 if (vma_is_anonymous(vma))
5039 return do_huge_pmd_anonymous_page(vmf);
5040 if (vma->vm_ops->huge_fault)
5041 return vma->vm_ops->huge_fault(vmf, PMD_ORDER);
5042 return VM_FAULT_FALLBACK;
5043}
5044
5045/* `inline' is required to avoid gcc 4.1.2 build error */
5046static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
5047{
5048 struct vm_area_struct *vma = vmf->vma;
5049 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
5050 vm_fault_t ret;
5051
5052 if (vma_is_anonymous(vma)) {
5053 if (likely(!unshare) &&
5054 userfaultfd_huge_pmd_wp(vma, vmf->orig_pmd)) {
5055 if (userfaultfd_wp_async(vmf->vma))
5056 goto split;
5057 return handle_userfault(vmf, VM_UFFD_WP);
5058 }
5059 return do_huge_pmd_wp_page(vmf);
5060 }
5061
5062 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
5063 if (vma->vm_ops->huge_fault) {
5064 ret = vma->vm_ops->huge_fault(vmf, PMD_ORDER);
5065 if (!(ret & VM_FAULT_FALLBACK))
5066 return ret;
5067 }
5068 }
5069
5070split:
5071 /* COW or write-notify handled on pte level: split pmd. */
5072 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
5073
5074 return VM_FAULT_FALLBACK;
5075}
5076
5077static vm_fault_t create_huge_pud(struct vm_fault *vmf)
5078{
5079#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5080 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5081 struct vm_area_struct *vma = vmf->vma;
5082 /* No support for anonymous transparent PUD pages yet */
5083 if (vma_is_anonymous(vma))
5084 return VM_FAULT_FALLBACK;
5085 if (vma->vm_ops->huge_fault)
5086 return vma->vm_ops->huge_fault(vmf, PUD_ORDER);
5087#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5088 return VM_FAULT_FALLBACK;
5089}
5090
5091static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
5092{
5093#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5094 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5095 struct vm_area_struct *vma = vmf->vma;
5096 vm_fault_t ret;
5097
5098 /* No support for anonymous transparent PUD pages yet */
5099 if (vma_is_anonymous(vma))
5100 goto split;
5101 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
5102 if (vma->vm_ops->huge_fault) {
5103 ret = vma->vm_ops->huge_fault(vmf, PUD_ORDER);
5104 if (!(ret & VM_FAULT_FALLBACK))
5105 return ret;
5106 }
5107 }
5108split:
5109 /* COW or write-notify not handled on PUD level: split pud.*/
5110 __split_huge_pud(vma, vmf->pud, vmf->address);
5111#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
5112 return VM_FAULT_FALLBACK;
5113}
5114
5115/*
5116 * These routines also need to handle stuff like marking pages dirty
5117 * and/or accessed for architectures that don't do it in hardware (most
5118 * RISC architectures). The early dirtying is also good on the i386.
5119 *
5120 * There is also a hook called "update_mmu_cache()" that architectures
5121 * with external mmu caches can use to update those (ie the Sparc or
5122 * PowerPC hashed page tables that act as extended TLBs).
5123 *
5124 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
5125 * concurrent faults).
5126 *
5127 * The mmap_lock may have been released depending on flags and our return value.
5128 * See filemap_fault() and __folio_lock_or_retry().
5129 */
5130static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
5131{
5132 pte_t entry;
5133
5134 if (unlikely(pmd_none(*vmf->pmd))) {
5135 /*
5136 * Leave __pte_alloc() until later: because vm_ops->fault may
5137 * want to allocate huge page, and if we expose page table
5138 * for an instant, it will be difficult to retract from
5139 * concurrent faults and from rmap lookups.
5140 */
5141 vmf->pte = NULL;
5142 vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID;
5143 } else {
5144 /*
5145 * A regular pmd is established and it can't morph into a huge
5146 * pmd by anon khugepaged, since that takes mmap_lock in write
5147 * mode; but shmem or file collapse to THP could still morph
5148 * it into a huge pmd: just retry later if so.
5149 */
5150 vmf->pte = pte_offset_map_nolock(vmf->vma->vm_mm, vmf->pmd,
5151 vmf->address, &vmf->ptl);
5152 if (unlikely(!vmf->pte))
5153 return 0;
5154 vmf->orig_pte = ptep_get_lockless(vmf->pte);
5155 vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID;
5156
5157 if (pte_none(vmf->orig_pte)) {
5158 pte_unmap(vmf->pte);
5159 vmf->pte = NULL;
5160 }
5161 }
5162
5163 if (!vmf->pte)
5164 return do_pte_missing(vmf);
5165
5166 if (!pte_present(vmf->orig_pte))
5167 return do_swap_page(vmf);
5168
5169 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
5170 return do_numa_page(vmf);
5171
5172 spin_lock(vmf->ptl);
5173 entry = vmf->orig_pte;
5174 if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) {
5175 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
5176 goto unlock;
5177 }
5178 if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
5179 if (!pte_write(entry))
5180 return do_wp_page(vmf);
5181 else if (likely(vmf->flags & FAULT_FLAG_WRITE))
5182 entry = pte_mkdirty(entry);
5183 }
5184 entry = pte_mkyoung(entry);
5185 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
5186 vmf->flags & FAULT_FLAG_WRITE)) {
5187 update_mmu_cache_range(vmf, vmf->vma, vmf->address,
5188 vmf->pte, 1);
5189 } else {
5190 /* Skip spurious TLB flush for retried page fault */
5191 if (vmf->flags & FAULT_FLAG_TRIED)
5192 goto unlock;
5193 /*
5194 * This is needed only for protection faults but the arch code
5195 * is not yet telling us if this is a protection fault or not.
5196 * This still avoids useless tlb flushes for .text page faults
5197 * with threads.
5198 */
5199 if (vmf->flags & FAULT_FLAG_WRITE)
5200 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address,
5201 vmf->pte);
5202 }
5203unlock:
5204 pte_unmap_unlock(vmf->pte, vmf->ptl);
5205 return 0;
5206}
5207
5208/*
5209 * On entry, we hold either the VMA lock or the mmap_lock
5210 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5211 * the result, the mmap_lock is not held on exit. See filemap_fault()
5212 * and __folio_lock_or_retry().
5213 */
5214static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
5215 unsigned long address, unsigned int flags)
5216{
5217 struct vm_fault vmf = {
5218 .vma = vma,
5219 .address = address & PAGE_MASK,
5220 .real_address = address,
5221 .flags = flags,
5222 .pgoff = linear_page_index(vma, address),
5223 .gfp_mask = __get_fault_gfp_mask(vma),
5224 };
5225 struct mm_struct *mm = vma->vm_mm;
5226 unsigned long vm_flags = vma->vm_flags;
5227 pgd_t *pgd;
5228 p4d_t *p4d;
5229 vm_fault_t ret;
5230
5231 pgd = pgd_offset(mm, address);
5232 p4d = p4d_alloc(mm, pgd, address);
5233 if (!p4d)
5234 return VM_FAULT_OOM;
5235
5236 vmf.pud = pud_alloc(mm, p4d, address);
5237 if (!vmf.pud)
5238 return VM_FAULT_OOM;
5239retry_pud:
5240 if (pud_none(*vmf.pud) &&
5241 thp_vma_allowable_order(vma, vm_flags, false, true, true, PUD_ORDER)) {
5242 ret = create_huge_pud(&vmf);
5243 if (!(ret & VM_FAULT_FALLBACK))
5244 return ret;
5245 } else {
5246 pud_t orig_pud = *vmf.pud;
5247
5248 barrier();
5249 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
5250
5251 /*
5252 * TODO once we support anonymous PUDs: NUMA case and
5253 * FAULT_FLAG_UNSHARE handling.
5254 */
5255 if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) {
5256 ret = wp_huge_pud(&vmf, orig_pud);
5257 if (!(ret & VM_FAULT_FALLBACK))
5258 return ret;
5259 } else {
5260 huge_pud_set_accessed(&vmf, orig_pud);
5261 return 0;
5262 }
5263 }
5264 }
5265
5266 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
5267 if (!vmf.pmd)
5268 return VM_FAULT_OOM;
5269
5270 /* Huge pud page fault raced with pmd_alloc? */
5271 if (pud_trans_unstable(vmf.pud))
5272 goto retry_pud;
5273
5274 if (pmd_none(*vmf.pmd) &&
5275 thp_vma_allowable_order(vma, vm_flags, false, true, true, PMD_ORDER)) {
5276 ret = create_huge_pmd(&vmf);
5277 if (!(ret & VM_FAULT_FALLBACK))
5278 return ret;
5279 } else {
5280 vmf.orig_pmd = pmdp_get_lockless(vmf.pmd);
5281
5282 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
5283 VM_BUG_ON(thp_migration_supported() &&
5284 !is_pmd_migration_entry(vmf.orig_pmd));
5285 if (is_pmd_migration_entry(vmf.orig_pmd))
5286 pmd_migration_entry_wait(mm, vmf.pmd);
5287 return 0;
5288 }
5289 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
5290 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
5291 return do_huge_pmd_numa_page(&vmf);
5292
5293 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5294 !pmd_write(vmf.orig_pmd)) {
5295 ret = wp_huge_pmd(&vmf);
5296 if (!(ret & VM_FAULT_FALLBACK))
5297 return ret;
5298 } else {
5299 huge_pmd_set_accessed(&vmf);
5300 return 0;
5301 }
5302 }
5303 }
5304
5305 return handle_pte_fault(&vmf);
5306}
5307
5308/**
5309 * mm_account_fault - Do page fault accounting
5310 * @mm: mm from which memcg should be extracted. It can be NULL.
5311 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5312 * of perf event counters, but we'll still do the per-task accounting to
5313 * the task who triggered this page fault.
5314 * @address: the faulted address.
5315 * @flags: the fault flags.
5316 * @ret: the fault retcode.
5317 *
5318 * This will take care of most of the page fault accounting. Meanwhile, it
5319 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5320 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5321 * still be in per-arch page fault handlers at the entry of page fault.
5322 */
5323static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs,
5324 unsigned long address, unsigned int flags,
5325 vm_fault_t ret)
5326{
5327 bool major;
5328
5329 /* Incomplete faults will be accounted upon completion. */
5330 if (ret & VM_FAULT_RETRY)
5331 return;
5332
5333 /*
5334 * To preserve the behavior of older kernels, PGFAULT counters record
5335 * both successful and failed faults, as opposed to perf counters,
5336 * which ignore failed cases.
5337 */
5338 count_vm_event(PGFAULT);
5339 count_memcg_event_mm(mm, PGFAULT);
5340
5341 /*
5342 * Do not account for unsuccessful faults (e.g. when the address wasn't
5343 * valid). That includes arch_vma_access_permitted() failing before
5344 * reaching here. So this is not a "this many hardware page faults"
5345 * counter. We should use the hw profiling for that.
5346 */
5347 if (ret & VM_FAULT_ERROR)
5348 return;
5349
5350 /*
5351 * We define the fault as a major fault when the final successful fault
5352 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5353 * handle it immediately previously).
5354 */
5355 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
5356
5357 if (major)
5358 current->maj_flt++;
5359 else
5360 current->min_flt++;
5361
5362 /*
5363 * If the fault is done for GUP, regs will be NULL. We only do the
5364 * accounting for the per thread fault counters who triggered the
5365 * fault, and we skip the perf event updates.
5366 */
5367 if (!regs)
5368 return;
5369
5370 if (major)
5371 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
5372 else
5373 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
5374}
5375
5376#ifdef CONFIG_LRU_GEN
5377static void lru_gen_enter_fault(struct vm_area_struct *vma)
5378{
5379 /* the LRU algorithm only applies to accesses with recency */
5380 current->in_lru_fault = vma_has_recency(vma);
5381}
5382
5383static void lru_gen_exit_fault(void)
5384{
5385 current->in_lru_fault = false;
5386}
5387#else
5388static void lru_gen_enter_fault(struct vm_area_struct *vma)
5389{
5390}
5391
5392static void lru_gen_exit_fault(void)
5393{
5394}
5395#endif /* CONFIG_LRU_GEN */
5396
5397static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
5398 unsigned int *flags)
5399{
5400 if (unlikely(*flags & FAULT_FLAG_UNSHARE)) {
5401 if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE))
5402 return VM_FAULT_SIGSEGV;
5403 /*
5404 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5405 * just treat it like an ordinary read-fault otherwise.
5406 */
5407 if (!is_cow_mapping(vma->vm_flags))
5408 *flags &= ~FAULT_FLAG_UNSHARE;
5409 } else if (*flags & FAULT_FLAG_WRITE) {
5410 /* Write faults on read-only mappings are impossible ... */
5411 if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
5412 return VM_FAULT_SIGSEGV;
5413 /* ... and FOLL_FORCE only applies to COW mappings. */
5414 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
5415 !is_cow_mapping(vma->vm_flags)))
5416 return VM_FAULT_SIGSEGV;
5417 }
5418#ifdef CONFIG_PER_VMA_LOCK
5419 /*
5420 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5421 * the assumption that lock is dropped on VM_FAULT_RETRY.
5422 */
5423 if (WARN_ON_ONCE((*flags &
5424 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)) ==
5425 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)))
5426 return VM_FAULT_SIGSEGV;
5427#endif
5428
5429 return 0;
5430}
5431
5432/*
5433 * By the time we get here, we already hold the mm semaphore
5434 *
5435 * The mmap_lock may have been released depending on flags and our
5436 * return value. See filemap_fault() and __folio_lock_or_retry().
5437 */
5438vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
5439 unsigned int flags, struct pt_regs *regs)
5440{
5441 /* If the fault handler drops the mmap_lock, vma may be freed */
5442 struct mm_struct *mm = vma->vm_mm;
5443 vm_fault_t ret;
5444
5445 __set_current_state(TASK_RUNNING);
5446
5447 ret = sanitize_fault_flags(vma, &flags);
5448 if (ret)
5449 goto out;
5450
5451 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
5452 flags & FAULT_FLAG_INSTRUCTION,
5453 flags & FAULT_FLAG_REMOTE)) {
5454 ret = VM_FAULT_SIGSEGV;
5455 goto out;
5456 }
5457
5458 /*
5459 * Enable the memcg OOM handling for faults triggered in user
5460 * space. Kernel faults are handled more gracefully.
5461 */
5462 if (flags & FAULT_FLAG_USER)
5463 mem_cgroup_enter_user_fault();
5464
5465 lru_gen_enter_fault(vma);
5466
5467 if (unlikely(is_vm_hugetlb_page(vma)))
5468 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
5469 else
5470 ret = __handle_mm_fault(vma, address, flags);
5471
5472 lru_gen_exit_fault();
5473
5474 if (flags & FAULT_FLAG_USER) {
5475 mem_cgroup_exit_user_fault();
5476 /*
5477 * The task may have entered a memcg OOM situation but
5478 * if the allocation error was handled gracefully (no
5479 * VM_FAULT_OOM), there is no need to kill anything.
5480 * Just clean up the OOM state peacefully.
5481 */
5482 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
5483 mem_cgroup_oom_synchronize(false);
5484 }
5485out:
5486 mm_account_fault(mm, regs, address, flags, ret);
5487
5488 return ret;
5489}
5490EXPORT_SYMBOL_GPL(handle_mm_fault);
5491
5492#ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5493#include <linux/extable.h>
5494
5495static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5496{
5497 if (likely(mmap_read_trylock(mm)))
5498 return true;
5499
5500 if (regs && !user_mode(regs)) {
5501 unsigned long ip = exception_ip(regs);
5502 if (!search_exception_tables(ip))
5503 return false;
5504 }
5505
5506 return !mmap_read_lock_killable(mm);
5507}
5508
5509static inline bool mmap_upgrade_trylock(struct mm_struct *mm)
5510{
5511 /*
5512 * We don't have this operation yet.
5513 *
5514 * It should be easy enough to do: it's basically a
5515 * atomic_long_try_cmpxchg_acquire()
5516 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5517 * it also needs the proper lockdep magic etc.
5518 */
5519 return false;
5520}
5521
5522static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5523{
5524 mmap_read_unlock(mm);
5525 if (regs && !user_mode(regs)) {
5526 unsigned long ip = exception_ip(regs);
5527 if (!search_exception_tables(ip))
5528 return false;
5529 }
5530 return !mmap_write_lock_killable(mm);
5531}
5532
5533/*
5534 * Helper for page fault handling.
5535 *
5536 * This is kind of equivalend to "mmap_read_lock()" followed
5537 * by "find_extend_vma()", except it's a lot more careful about
5538 * the locking (and will drop the lock on failure).
5539 *
5540 * For example, if we have a kernel bug that causes a page
5541 * fault, we don't want to just use mmap_read_lock() to get
5542 * the mm lock, because that would deadlock if the bug were
5543 * to happen while we're holding the mm lock for writing.
5544 *
5545 * So this checks the exception tables on kernel faults in
5546 * order to only do this all for instructions that are actually
5547 * expected to fault.
5548 *
5549 * We can also actually take the mm lock for writing if we
5550 * need to extend the vma, which helps the VM layer a lot.
5551 */
5552struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm,
5553 unsigned long addr, struct pt_regs *regs)
5554{
5555 struct vm_area_struct *vma;
5556
5557 if (!get_mmap_lock_carefully(mm, regs))
5558 return NULL;
5559
5560 vma = find_vma(mm, addr);
5561 if (likely(vma && (vma->vm_start <= addr)))
5562 return vma;
5563
5564 /*
5565 * Well, dang. We might still be successful, but only
5566 * if we can extend a vma to do so.
5567 */
5568 if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) {
5569 mmap_read_unlock(mm);
5570 return NULL;
5571 }
5572
5573 /*
5574 * We can try to upgrade the mmap lock atomically,
5575 * in which case we can continue to use the vma
5576 * we already looked up.
5577 *
5578 * Otherwise we'll have to drop the mmap lock and
5579 * re-take it, and also look up the vma again,
5580 * re-checking it.
5581 */
5582 if (!mmap_upgrade_trylock(mm)) {
5583 if (!upgrade_mmap_lock_carefully(mm, regs))
5584 return NULL;
5585
5586 vma = find_vma(mm, addr);
5587 if (!vma)
5588 goto fail;
5589 if (vma->vm_start <= addr)
5590 goto success;
5591 if (!(vma->vm_flags & VM_GROWSDOWN))
5592 goto fail;
5593 }
5594
5595 if (expand_stack_locked(vma, addr))
5596 goto fail;
5597
5598success:
5599 mmap_write_downgrade(mm);
5600 return vma;
5601
5602fail:
5603 mmap_write_unlock(mm);
5604 return NULL;
5605}
5606#endif
5607
5608#ifdef CONFIG_PER_VMA_LOCK
5609/*
5610 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5611 * stable and not isolated. If the VMA is not found or is being modified the
5612 * function returns NULL.
5613 */
5614struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
5615 unsigned long address)
5616{
5617 MA_STATE(mas, &mm->mm_mt, address, address);
5618 struct vm_area_struct *vma;
5619
5620 rcu_read_lock();
5621retry:
5622 vma = mas_walk(&mas);
5623 if (!vma)
5624 goto inval;
5625
5626 if (!vma_start_read(vma))
5627 goto inval;
5628
5629 /*
5630 * find_mergeable_anon_vma uses adjacent vmas which are not locked.
5631 * This check must happen after vma_start_read(); otherwise, a
5632 * concurrent mremap() with MREMAP_DONTUNMAP could dissociate the VMA
5633 * from its anon_vma.
5634 */
5635 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma))
5636 goto inval_end_read;
5637
5638 /* Check since vm_start/vm_end might change before we lock the VMA */
5639 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
5640 goto inval_end_read;
5641
5642 /* Check if the VMA got isolated after we found it */
5643 if (vma->detached) {
5644 vma_end_read(vma);
5645 count_vm_vma_lock_event(VMA_LOCK_MISS);
5646 /* The area was replaced with another one */
5647 goto retry;
5648 }
5649
5650 rcu_read_unlock();
5651 return vma;
5652
5653inval_end_read:
5654 vma_end_read(vma);
5655inval:
5656 rcu_read_unlock();
5657 count_vm_vma_lock_event(VMA_LOCK_ABORT);
5658 return NULL;
5659}
5660#endif /* CONFIG_PER_VMA_LOCK */
5661
5662#ifndef __PAGETABLE_P4D_FOLDED
5663/*
5664 * Allocate p4d page table.
5665 * We've already handled the fast-path in-line.
5666 */
5667int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
5668{
5669 p4d_t *new = p4d_alloc_one(mm, address);
5670 if (!new)
5671 return -ENOMEM;
5672
5673 spin_lock(&mm->page_table_lock);
5674 if (pgd_present(*pgd)) { /* Another has populated it */
5675 p4d_free(mm, new);
5676 } else {
5677 smp_wmb(); /* See comment in pmd_install() */
5678 pgd_populate(mm, pgd, new);
5679 }
5680 spin_unlock(&mm->page_table_lock);
5681 return 0;
5682}
5683#endif /* __PAGETABLE_P4D_FOLDED */
5684
5685#ifndef __PAGETABLE_PUD_FOLDED
5686/*
5687 * Allocate page upper directory.
5688 * We've already handled the fast-path in-line.
5689 */
5690int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
5691{
5692 pud_t *new = pud_alloc_one(mm, address);
5693 if (!new)
5694 return -ENOMEM;
5695
5696 spin_lock(&mm->page_table_lock);
5697 if (!p4d_present(*p4d)) {
5698 mm_inc_nr_puds(mm);
5699 smp_wmb(); /* See comment in pmd_install() */
5700 p4d_populate(mm, p4d, new);
5701 } else /* Another has populated it */
5702 pud_free(mm, new);
5703 spin_unlock(&mm->page_table_lock);
5704 return 0;
5705}
5706#endif /* __PAGETABLE_PUD_FOLDED */
5707
5708#ifndef __PAGETABLE_PMD_FOLDED
5709/*
5710 * Allocate page middle directory.
5711 * We've already handled the fast-path in-line.
5712 */
5713int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
5714{
5715 spinlock_t *ptl;
5716 pmd_t *new = pmd_alloc_one(mm, address);
5717 if (!new)
5718 return -ENOMEM;
5719
5720 ptl = pud_lock(mm, pud);
5721 if (!pud_present(*pud)) {
5722 mm_inc_nr_pmds(mm);
5723 smp_wmb(); /* See comment in pmd_install() */
5724 pud_populate(mm, pud, new);
5725 } else { /* Another has populated it */
5726 pmd_free(mm, new);
5727 }
5728 spin_unlock(ptl);
5729 return 0;
5730}
5731#endif /* __PAGETABLE_PMD_FOLDED */
5732
5733/**
5734 * follow_pte - look up PTE at a user virtual address
5735 * @mm: the mm_struct of the target address space
5736 * @address: user virtual address
5737 * @ptepp: location to store found PTE
5738 * @ptlp: location to store the lock for the PTE
5739 *
5740 * On a successful return, the pointer to the PTE is stored in @ptepp;
5741 * the corresponding lock is taken and its location is stored in @ptlp.
5742 * The contents of the PTE are only stable until @ptlp is released;
5743 * any further use, if any, must be protected against invalidation
5744 * with MMU notifiers.
5745 *
5746 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5747 * should be taken for read.
5748 *
5749 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5750 * it is not a good general-purpose API.
5751 *
5752 * Return: zero on success, -ve otherwise.
5753 */
5754int follow_pte(struct mm_struct *mm, unsigned long address,
5755 pte_t **ptepp, spinlock_t **ptlp)
5756{
5757 pgd_t *pgd;
5758 p4d_t *p4d;
5759 pud_t *pud;
5760 pmd_t *pmd;
5761 pte_t *ptep;
5762
5763 pgd = pgd_offset(mm, address);
5764 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
5765 goto out;
5766
5767 p4d = p4d_offset(pgd, address);
5768 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
5769 goto out;
5770
5771 pud = pud_offset(p4d, address);
5772 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
5773 goto out;
5774
5775 pmd = pmd_offset(pud, address);
5776 VM_BUG_ON(pmd_trans_huge(*pmd));
5777
5778 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
5779 if (!ptep)
5780 goto out;
5781 if (!pte_present(ptep_get(ptep)))
5782 goto unlock;
5783 *ptepp = ptep;
5784 return 0;
5785unlock:
5786 pte_unmap_unlock(ptep, *ptlp);
5787out:
5788 return -EINVAL;
5789}
5790EXPORT_SYMBOL_GPL(follow_pte);
5791
5792/**
5793 * follow_pfn - look up PFN at a user virtual address
5794 * @vma: memory mapping
5795 * @address: user virtual address
5796 * @pfn: location to store found PFN
5797 *
5798 * Only IO mappings and raw PFN mappings are allowed.
5799 *
5800 * This function does not allow the caller to read the permissions
5801 * of the PTE. Do not use it.
5802 *
5803 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5804 */
5805int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5806 unsigned long *pfn)
5807{
5808 int ret = -EINVAL;
5809 spinlock_t *ptl;
5810 pte_t *ptep;
5811
5812 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5813 return ret;
5814
5815 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5816 if (ret)
5817 return ret;
5818 *pfn = pte_pfn(ptep_get(ptep));
5819 pte_unmap_unlock(ptep, ptl);
5820 return 0;
5821}
5822EXPORT_SYMBOL(follow_pfn);
5823
5824#ifdef CONFIG_HAVE_IOREMAP_PROT
5825int follow_phys(struct vm_area_struct *vma,
5826 unsigned long address, unsigned int flags,
5827 unsigned long *prot, resource_size_t *phys)
5828{
5829 int ret = -EINVAL;
5830 pte_t *ptep, pte;
5831 spinlock_t *ptl;
5832
5833 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5834 goto out;
5835
5836 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5837 goto out;
5838 pte = ptep_get(ptep);
5839
5840 if ((flags & FOLL_WRITE) && !pte_write(pte))
5841 goto unlock;
5842
5843 *prot = pgprot_val(pte_pgprot(pte));
5844 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5845
5846 ret = 0;
5847unlock:
5848 pte_unmap_unlock(ptep, ptl);
5849out:
5850 return ret;
5851}
5852
5853/**
5854 * generic_access_phys - generic implementation for iomem mmap access
5855 * @vma: the vma to access
5856 * @addr: userspace address, not relative offset within @vma
5857 * @buf: buffer to read/write
5858 * @len: length of transfer
5859 * @write: set to FOLL_WRITE when writing, otherwise reading
5860 *
5861 * This is a generic implementation for &vm_operations_struct.access for an
5862 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5863 * not page based.
5864 */
5865int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5866 void *buf, int len, int write)
5867{
5868 resource_size_t phys_addr;
5869 unsigned long prot = 0;
5870 void __iomem *maddr;
5871 pte_t *ptep, pte;
5872 spinlock_t *ptl;
5873 int offset = offset_in_page(addr);
5874 int ret = -EINVAL;
5875
5876 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5877 return -EINVAL;
5878
5879retry:
5880 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5881 return -EINVAL;
5882 pte = ptep_get(ptep);
5883 pte_unmap_unlock(ptep, ptl);
5884
5885 prot = pgprot_val(pte_pgprot(pte));
5886 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5887
5888 if ((write & FOLL_WRITE) && !pte_write(pte))
5889 return -EINVAL;
5890
5891 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5892 if (!maddr)
5893 return -ENOMEM;
5894
5895 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5896 goto out_unmap;
5897
5898 if (!pte_same(pte, ptep_get(ptep))) {
5899 pte_unmap_unlock(ptep, ptl);
5900 iounmap(maddr);
5901
5902 goto retry;
5903 }
5904
5905 if (write)
5906 memcpy_toio(maddr + offset, buf, len);
5907 else
5908 memcpy_fromio(buf, maddr + offset, len);
5909 ret = len;
5910 pte_unmap_unlock(ptep, ptl);
5911out_unmap:
5912 iounmap(maddr);
5913
5914 return ret;
5915}
5916EXPORT_SYMBOL_GPL(generic_access_phys);
5917#endif
5918
5919/*
5920 * Access another process' address space as given in mm.
5921 */
5922static int __access_remote_vm(struct mm_struct *mm, unsigned long addr,
5923 void *buf, int len, unsigned int gup_flags)
5924{
5925 void *old_buf = buf;
5926 int write = gup_flags & FOLL_WRITE;
5927
5928 if (mmap_read_lock_killable(mm))
5929 return 0;
5930
5931 /* Untag the address before looking up the VMA */
5932 addr = untagged_addr_remote(mm, addr);
5933
5934 /* Avoid triggering the temporary warning in __get_user_pages */
5935 if (!vma_lookup(mm, addr) && !expand_stack(mm, addr))
5936 return 0;
5937
5938 /* ignore errors, just check how much was successfully transferred */
5939 while (len) {
5940 int bytes, offset;
5941 void *maddr;
5942 struct vm_area_struct *vma = NULL;
5943 struct page *page = get_user_page_vma_remote(mm, addr,
5944 gup_flags, &vma);
5945
5946 if (IS_ERR(page)) {
5947 /* We might need to expand the stack to access it */
5948 vma = vma_lookup(mm, addr);
5949 if (!vma) {
5950 vma = expand_stack(mm, addr);
5951
5952 /* mmap_lock was dropped on failure */
5953 if (!vma)
5954 return buf - old_buf;
5955
5956 /* Try again if stack expansion worked */
5957 continue;
5958 }
5959
5960 /*
5961 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5962 * we can access using slightly different code.
5963 */
5964 bytes = 0;
5965#ifdef CONFIG_HAVE_IOREMAP_PROT
5966 if (vma->vm_ops && vma->vm_ops->access)
5967 bytes = vma->vm_ops->access(vma, addr, buf,
5968 len, write);
5969#endif
5970 if (bytes <= 0)
5971 break;
5972 } else {
5973 bytes = len;
5974 offset = addr & (PAGE_SIZE-1);
5975 if (bytes > PAGE_SIZE-offset)
5976 bytes = PAGE_SIZE-offset;
5977
5978 maddr = kmap_local_page(page);
5979 if (write) {
5980 copy_to_user_page(vma, page, addr,
5981 maddr + offset, buf, bytes);
5982 set_page_dirty_lock(page);
5983 } else {
5984 copy_from_user_page(vma, page, addr,
5985 buf, maddr + offset, bytes);
5986 }
5987 unmap_and_put_page(page, maddr);
5988 }
5989 len -= bytes;
5990 buf += bytes;
5991 addr += bytes;
5992 }
5993 mmap_read_unlock(mm);
5994
5995 return buf - old_buf;
5996}
5997
5998/**
5999 * access_remote_vm - access another process' address space
6000 * @mm: the mm_struct of the target address space
6001 * @addr: start address to access
6002 * @buf: source or destination buffer
6003 * @len: number of bytes to transfer
6004 * @gup_flags: flags modifying lookup behaviour
6005 *
6006 * The caller must hold a reference on @mm.
6007 *
6008 * Return: number of bytes copied from source to destination.
6009 */
6010int access_remote_vm(struct mm_struct *mm, unsigned long addr,
6011 void *buf, int len, unsigned int gup_flags)
6012{
6013 return __access_remote_vm(mm, addr, buf, len, gup_flags);
6014}
6015
6016/*
6017 * Access another process' address space.
6018 * Source/target buffer must be kernel space,
6019 * Do not walk the page table directly, use get_user_pages
6020 */
6021int access_process_vm(struct task_struct *tsk, unsigned long addr,
6022 void *buf, int len, unsigned int gup_flags)
6023{
6024 struct mm_struct *mm;
6025 int ret;
6026
6027 mm = get_task_mm(tsk);
6028 if (!mm)
6029 return 0;
6030
6031 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
6032
6033 mmput(mm);
6034
6035 return ret;
6036}
6037EXPORT_SYMBOL_GPL(access_process_vm);
6038
6039/*
6040 * Print the name of a VMA.
6041 */
6042void print_vma_addr(char *prefix, unsigned long ip)
6043{
6044 struct mm_struct *mm = current->mm;
6045 struct vm_area_struct *vma;
6046
6047 /*
6048 * we might be running from an atomic context so we cannot sleep
6049 */
6050 if (!mmap_read_trylock(mm))
6051 return;
6052
6053 vma = find_vma(mm, ip);
6054 if (vma && vma->vm_file) {
6055 struct file *f = vma->vm_file;
6056 char *buf = (char *)__get_free_page(GFP_NOWAIT);
6057 if (buf) {
6058 char *p;
6059
6060 p = file_path(f, buf, PAGE_SIZE);
6061 if (IS_ERR(p))
6062 p = "?";
6063 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
6064 vma->vm_start,
6065 vma->vm_end - vma->vm_start);
6066 free_page((unsigned long)buf);
6067 }
6068 }
6069 mmap_read_unlock(mm);
6070}
6071
6072#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6073void __might_fault(const char *file, int line)
6074{
6075 if (pagefault_disabled())
6076 return;
6077 __might_sleep(file, line);
6078#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6079 if (current->mm)
6080 might_lock_read(¤t->mm->mmap_lock);
6081#endif
6082}
6083EXPORT_SYMBOL(__might_fault);
6084#endif
6085
6086#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
6087/*
6088 * Process all subpages of the specified huge page with the specified
6089 * operation. The target subpage will be processed last to keep its
6090 * cache lines hot.
6091 */
6092static inline int process_huge_page(
6093 unsigned long addr_hint, unsigned int pages_per_huge_page,
6094 int (*process_subpage)(unsigned long addr, int idx, void *arg),
6095 void *arg)
6096{
6097 int i, n, base, l, ret;
6098 unsigned long addr = addr_hint &
6099 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6100
6101 /* Process target subpage last to keep its cache lines hot */
6102 might_sleep();
6103 n = (addr_hint - addr) / PAGE_SIZE;
6104 if (2 * n <= pages_per_huge_page) {
6105 /* If target subpage in first half of huge page */
6106 base = 0;
6107 l = n;
6108 /* Process subpages at the end of huge page */
6109 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
6110 cond_resched();
6111 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
6112 if (ret)
6113 return ret;
6114 }
6115 } else {
6116 /* If target subpage in second half of huge page */
6117 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
6118 l = pages_per_huge_page - n;
6119 /* Process subpages at the begin of huge page */
6120 for (i = 0; i < base; i++) {
6121 cond_resched();
6122 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
6123 if (ret)
6124 return ret;
6125 }
6126 }
6127 /*
6128 * Process remaining subpages in left-right-left-right pattern
6129 * towards the target subpage
6130 */
6131 for (i = 0; i < l; i++) {
6132 int left_idx = base + i;
6133 int right_idx = base + 2 * l - 1 - i;
6134
6135 cond_resched();
6136 ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
6137 if (ret)
6138 return ret;
6139 cond_resched();
6140 ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
6141 if (ret)
6142 return ret;
6143 }
6144 return 0;
6145}
6146
6147static void clear_gigantic_page(struct page *page,
6148 unsigned long addr,
6149 unsigned int pages_per_huge_page)
6150{
6151 int i;
6152 struct page *p;
6153
6154 might_sleep();
6155 for (i = 0; i < pages_per_huge_page; i++) {
6156 p = nth_page(page, i);
6157 cond_resched();
6158 clear_user_highpage(p, addr + i * PAGE_SIZE);
6159 }
6160}
6161
6162static int clear_subpage(unsigned long addr, int idx, void *arg)
6163{
6164 struct page *page = arg;
6165
6166 clear_user_highpage(page + idx, addr);
6167 return 0;
6168}
6169
6170void clear_huge_page(struct page *page,
6171 unsigned long addr_hint, unsigned int pages_per_huge_page)
6172{
6173 unsigned long addr = addr_hint &
6174 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6175
6176 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
6177 clear_gigantic_page(page, addr, pages_per_huge_page);
6178 return;
6179 }
6180
6181 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
6182}
6183
6184static int copy_user_gigantic_page(struct folio *dst, struct folio *src,
6185 unsigned long addr,
6186 struct vm_area_struct *vma,
6187 unsigned int pages_per_huge_page)
6188{
6189 int i;
6190 struct page *dst_page;
6191 struct page *src_page;
6192
6193 for (i = 0; i < pages_per_huge_page; i++) {
6194 dst_page = folio_page(dst, i);
6195 src_page = folio_page(src, i);
6196
6197 cond_resched();
6198 if (copy_mc_user_highpage(dst_page, src_page,
6199 addr + i*PAGE_SIZE, vma)) {
6200 memory_failure_queue(page_to_pfn(src_page), 0);
6201 return -EHWPOISON;
6202 }
6203 }
6204 return 0;
6205}
6206
6207struct copy_subpage_arg {
6208 struct page *dst;
6209 struct page *src;
6210 struct vm_area_struct *vma;
6211};
6212
6213static int copy_subpage(unsigned long addr, int idx, void *arg)
6214{
6215 struct copy_subpage_arg *copy_arg = arg;
6216
6217 if (copy_mc_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
6218 addr, copy_arg->vma)) {
6219 memory_failure_queue(page_to_pfn(copy_arg->src + idx), 0);
6220 return -EHWPOISON;
6221 }
6222 return 0;
6223}
6224
6225int copy_user_large_folio(struct folio *dst, struct folio *src,
6226 unsigned long addr_hint, struct vm_area_struct *vma)
6227{
6228 unsigned int pages_per_huge_page = folio_nr_pages(dst);
6229 unsigned long addr = addr_hint &
6230 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6231 struct copy_subpage_arg arg = {
6232 .dst = &dst->page,
6233 .src = &src->page,
6234 .vma = vma,
6235 };
6236
6237 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES))
6238 return copy_user_gigantic_page(dst, src, addr, vma,
6239 pages_per_huge_page);
6240
6241 return process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
6242}
6243
6244long copy_folio_from_user(struct folio *dst_folio,
6245 const void __user *usr_src,
6246 bool allow_pagefault)
6247{
6248 void *kaddr;
6249 unsigned long i, rc = 0;
6250 unsigned int nr_pages = folio_nr_pages(dst_folio);
6251 unsigned long ret_val = nr_pages * PAGE_SIZE;
6252 struct page *subpage;
6253
6254 for (i = 0; i < nr_pages; i++) {
6255 subpage = folio_page(dst_folio, i);
6256 kaddr = kmap_local_page(subpage);
6257 if (!allow_pagefault)
6258 pagefault_disable();
6259 rc = copy_from_user(kaddr, usr_src + i * PAGE_SIZE, PAGE_SIZE);
6260 if (!allow_pagefault)
6261 pagefault_enable();
6262 kunmap_local(kaddr);
6263
6264 ret_val -= (PAGE_SIZE - rc);
6265 if (rc)
6266 break;
6267
6268 flush_dcache_page(subpage);
6269
6270 cond_resched();
6271 }
6272 return ret_val;
6273}
6274#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6275
6276#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6277
6278static struct kmem_cache *page_ptl_cachep;
6279
6280void __init ptlock_cache_init(void)
6281{
6282 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
6283 SLAB_PANIC, NULL);
6284}
6285
6286bool ptlock_alloc(struct ptdesc *ptdesc)
6287{
6288 spinlock_t *ptl;
6289
6290 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
6291 if (!ptl)
6292 return false;
6293 ptdesc->ptl = ptl;
6294 return true;
6295}
6296
6297void ptlock_free(struct ptdesc *ptdesc)
6298{
6299 kmem_cache_free(page_ptl_cachep, ptdesc->ptl);
6300}
6301#endif
1/*
2 * linux/mm/memory.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
10 */
11
12/*
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
15 *
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
18 * far as I could see.
19 *
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21 */
22
23/*
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
29 */
30
31/*
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
34 *
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
37 *
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39 */
40
41#include <linux/kernel_stat.h>
42#include <linux/mm.h>
43#include <linux/hugetlb.h>
44#include <linux/mman.h>
45#include <linux/swap.h>
46#include <linux/highmem.h>
47#include <linux/pagemap.h>
48#include <linux/ksm.h>
49#include <linux/rmap.h>
50#include <linux/export.h>
51#include <linux/delayacct.h>
52#include <linux/init.h>
53#include <linux/writeback.h>
54#include <linux/memcontrol.h>
55#include <linux/mmu_notifier.h>
56#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
59#include <linux/gfp.h>
60#include <linux/migrate.h>
61#include <linux/string.h>
62#include <linux/dma-debug.h>
63#include <linux/debugfs.h>
64
65#include <asm/io.h>
66#include <asm/pgalloc.h>
67#include <asm/uaccess.h>
68#include <asm/tlb.h>
69#include <asm/tlbflush.h>
70#include <asm/pgtable.h>
71
72#include "internal.h"
73
74#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
76#endif
77
78#ifndef CONFIG_NEED_MULTIPLE_NODES
79/* use the per-pgdat data instead for discontigmem - mbligh */
80unsigned long max_mapnr;
81struct page *mem_map;
82
83EXPORT_SYMBOL(max_mapnr);
84EXPORT_SYMBOL(mem_map);
85#endif
86
87/*
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92 * and ZONE_HIGHMEM.
93 */
94void * high_memory;
95
96EXPORT_SYMBOL(high_memory);
97
98/*
99 * Randomize the address space (stacks, mmaps, brk, etc.).
100 *
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
103 */
104int randomize_va_space __read_mostly =
105#ifdef CONFIG_COMPAT_BRK
106 1;
107#else
108 2;
109#endif
110
111static int __init disable_randmaps(char *s)
112{
113 randomize_va_space = 0;
114 return 1;
115}
116__setup("norandmaps", disable_randmaps);
117
118unsigned long zero_pfn __read_mostly;
119unsigned long highest_memmap_pfn __read_mostly;
120
121/*
122 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123 */
124static int __init init_zero_pfn(void)
125{
126 zero_pfn = page_to_pfn(ZERO_PAGE(0));
127 return 0;
128}
129core_initcall(init_zero_pfn);
130
131
132#if defined(SPLIT_RSS_COUNTING)
133
134void sync_mm_rss(struct mm_struct *mm)
135{
136 int i;
137
138 for (i = 0; i < NR_MM_COUNTERS; i++) {
139 if (current->rss_stat.count[i]) {
140 add_mm_counter(mm, i, current->rss_stat.count[i]);
141 current->rss_stat.count[i] = 0;
142 }
143 }
144 current->rss_stat.events = 0;
145}
146
147static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148{
149 struct task_struct *task = current;
150
151 if (likely(task->mm == mm))
152 task->rss_stat.count[member] += val;
153 else
154 add_mm_counter(mm, member, val);
155}
156#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159/* sync counter once per 64 page faults */
160#define TASK_RSS_EVENTS_THRESH (64)
161static void check_sync_rss_stat(struct task_struct *task)
162{
163 if (unlikely(task != current))
164 return;
165 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
166 sync_mm_rss(task->mm);
167}
168#else /* SPLIT_RSS_COUNTING */
169
170#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173static void check_sync_rss_stat(struct task_struct *task)
174{
175}
176
177#endif /* SPLIT_RSS_COUNTING */
178
179#ifdef HAVE_GENERIC_MMU_GATHER
180
181static int tlb_next_batch(struct mmu_gather *tlb)
182{
183 struct mmu_gather_batch *batch;
184
185 batch = tlb->active;
186 if (batch->next) {
187 tlb->active = batch->next;
188 return 1;
189 }
190
191 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192 return 0;
193
194 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195 if (!batch)
196 return 0;
197
198 tlb->batch_count++;
199 batch->next = NULL;
200 batch->nr = 0;
201 batch->max = MAX_GATHER_BATCH;
202
203 tlb->active->next = batch;
204 tlb->active = batch;
205
206 return 1;
207}
208
209/* tlb_gather_mmu
210 * Called to initialize an (on-stack) mmu_gather structure for page-table
211 * tear-down from @mm. The @fullmm argument is used when @mm is without
212 * users and we're going to destroy the full address space (exit/execve).
213 */
214void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
215{
216 tlb->mm = mm;
217
218 /* Is it from 0 to ~0? */
219 tlb->fullmm = !(start | (end+1));
220 tlb->need_flush_all = 0;
221 tlb->start = start;
222 tlb->end = end;
223 tlb->need_flush = 0;
224 tlb->local.next = NULL;
225 tlb->local.nr = 0;
226 tlb->local.max = ARRAY_SIZE(tlb->__pages);
227 tlb->active = &tlb->local;
228 tlb->batch_count = 0;
229
230#ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 tlb->batch = NULL;
232#endif
233}
234
235static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
236{
237 tlb->need_flush = 0;
238 tlb_flush(tlb);
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb_table_flush(tlb);
241#endif
242}
243
244static void tlb_flush_mmu_free(struct mmu_gather *tlb)
245{
246 struct mmu_gather_batch *batch;
247
248 for (batch = &tlb->local; batch; batch = batch->next) {
249 free_pages_and_swap_cache(batch->pages, batch->nr);
250 batch->nr = 0;
251 }
252 tlb->active = &tlb->local;
253}
254
255void tlb_flush_mmu(struct mmu_gather *tlb)
256{
257 if (!tlb->need_flush)
258 return;
259 tlb_flush_mmu_tlbonly(tlb);
260 tlb_flush_mmu_free(tlb);
261}
262
263/* tlb_finish_mmu
264 * Called at the end of the shootdown operation to free up any resources
265 * that were required.
266 */
267void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
268{
269 struct mmu_gather_batch *batch, *next;
270
271 tlb_flush_mmu(tlb);
272
273 /* keep the page table cache within bounds */
274 check_pgt_cache();
275
276 for (batch = tlb->local.next; batch; batch = next) {
277 next = batch->next;
278 free_pages((unsigned long)batch, 0);
279 }
280 tlb->local.next = NULL;
281}
282
283/* __tlb_remove_page
284 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
285 * handling the additional races in SMP caused by other CPUs caching valid
286 * mappings in their TLBs. Returns the number of free page slots left.
287 * When out of page slots we must call tlb_flush_mmu().
288 */
289int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
290{
291 struct mmu_gather_batch *batch;
292
293 VM_BUG_ON(!tlb->need_flush);
294
295 batch = tlb->active;
296 batch->pages[batch->nr++] = page;
297 if (batch->nr == batch->max) {
298 if (!tlb_next_batch(tlb))
299 return 0;
300 batch = tlb->active;
301 }
302 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
303
304 return batch->max - batch->nr;
305}
306
307#endif /* HAVE_GENERIC_MMU_GATHER */
308
309#ifdef CONFIG_HAVE_RCU_TABLE_FREE
310
311/*
312 * See the comment near struct mmu_table_batch.
313 */
314
315static void tlb_remove_table_smp_sync(void *arg)
316{
317 /* Simply deliver the interrupt */
318}
319
320static void tlb_remove_table_one(void *table)
321{
322 /*
323 * This isn't an RCU grace period and hence the page-tables cannot be
324 * assumed to be actually RCU-freed.
325 *
326 * It is however sufficient for software page-table walkers that rely on
327 * IRQ disabling. See the comment near struct mmu_table_batch.
328 */
329 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
330 __tlb_remove_table(table);
331}
332
333static void tlb_remove_table_rcu(struct rcu_head *head)
334{
335 struct mmu_table_batch *batch;
336 int i;
337
338 batch = container_of(head, struct mmu_table_batch, rcu);
339
340 for (i = 0; i < batch->nr; i++)
341 __tlb_remove_table(batch->tables[i]);
342
343 free_page((unsigned long)batch);
344}
345
346void tlb_table_flush(struct mmu_gather *tlb)
347{
348 struct mmu_table_batch **batch = &tlb->batch;
349
350 if (*batch) {
351 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
352 *batch = NULL;
353 }
354}
355
356void tlb_remove_table(struct mmu_gather *tlb, void *table)
357{
358 struct mmu_table_batch **batch = &tlb->batch;
359
360 tlb->need_flush = 1;
361
362 /*
363 * When there's less then two users of this mm there cannot be a
364 * concurrent page-table walk.
365 */
366 if (atomic_read(&tlb->mm->mm_users) < 2) {
367 __tlb_remove_table(table);
368 return;
369 }
370
371 if (*batch == NULL) {
372 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
373 if (*batch == NULL) {
374 tlb_remove_table_one(table);
375 return;
376 }
377 (*batch)->nr = 0;
378 }
379 (*batch)->tables[(*batch)->nr++] = table;
380 if ((*batch)->nr == MAX_TABLE_BATCH)
381 tlb_table_flush(tlb);
382}
383
384#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
385
386/*
387 * Note: this doesn't free the actual pages themselves. That
388 * has been handled earlier when unmapping all the memory regions.
389 */
390static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
391 unsigned long addr)
392{
393 pgtable_t token = pmd_pgtable(*pmd);
394 pmd_clear(pmd);
395 pte_free_tlb(tlb, token, addr);
396 atomic_long_dec(&tlb->mm->nr_ptes);
397}
398
399static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
400 unsigned long addr, unsigned long end,
401 unsigned long floor, unsigned long ceiling)
402{
403 pmd_t *pmd;
404 unsigned long next;
405 unsigned long start;
406
407 start = addr;
408 pmd = pmd_offset(pud, addr);
409 do {
410 next = pmd_addr_end(addr, end);
411 if (pmd_none_or_clear_bad(pmd))
412 continue;
413 free_pte_range(tlb, pmd, addr);
414 } while (pmd++, addr = next, addr != end);
415
416 start &= PUD_MASK;
417 if (start < floor)
418 return;
419 if (ceiling) {
420 ceiling &= PUD_MASK;
421 if (!ceiling)
422 return;
423 }
424 if (end - 1 > ceiling - 1)
425 return;
426
427 pmd = pmd_offset(pud, start);
428 pud_clear(pud);
429 pmd_free_tlb(tlb, pmd, start);
430}
431
432static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
433 unsigned long addr, unsigned long end,
434 unsigned long floor, unsigned long ceiling)
435{
436 pud_t *pud;
437 unsigned long next;
438 unsigned long start;
439
440 start = addr;
441 pud = pud_offset(pgd, addr);
442 do {
443 next = pud_addr_end(addr, end);
444 if (pud_none_or_clear_bad(pud))
445 continue;
446 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
447 } while (pud++, addr = next, addr != end);
448
449 start &= PGDIR_MASK;
450 if (start < floor)
451 return;
452 if (ceiling) {
453 ceiling &= PGDIR_MASK;
454 if (!ceiling)
455 return;
456 }
457 if (end - 1 > ceiling - 1)
458 return;
459
460 pud = pud_offset(pgd, start);
461 pgd_clear(pgd);
462 pud_free_tlb(tlb, pud, start);
463}
464
465/*
466 * This function frees user-level page tables of a process.
467 */
468void free_pgd_range(struct mmu_gather *tlb,
469 unsigned long addr, unsigned long end,
470 unsigned long floor, unsigned long ceiling)
471{
472 pgd_t *pgd;
473 unsigned long next;
474
475 /*
476 * The next few lines have given us lots of grief...
477 *
478 * Why are we testing PMD* at this top level? Because often
479 * there will be no work to do at all, and we'd prefer not to
480 * go all the way down to the bottom just to discover that.
481 *
482 * Why all these "- 1"s? Because 0 represents both the bottom
483 * of the address space and the top of it (using -1 for the
484 * top wouldn't help much: the masks would do the wrong thing).
485 * The rule is that addr 0 and floor 0 refer to the bottom of
486 * the address space, but end 0 and ceiling 0 refer to the top
487 * Comparisons need to use "end - 1" and "ceiling - 1" (though
488 * that end 0 case should be mythical).
489 *
490 * Wherever addr is brought up or ceiling brought down, we must
491 * be careful to reject "the opposite 0" before it confuses the
492 * subsequent tests. But what about where end is brought down
493 * by PMD_SIZE below? no, end can't go down to 0 there.
494 *
495 * Whereas we round start (addr) and ceiling down, by different
496 * masks at different levels, in order to test whether a table
497 * now has no other vmas using it, so can be freed, we don't
498 * bother to round floor or end up - the tests don't need that.
499 */
500
501 addr &= PMD_MASK;
502 if (addr < floor) {
503 addr += PMD_SIZE;
504 if (!addr)
505 return;
506 }
507 if (ceiling) {
508 ceiling &= PMD_MASK;
509 if (!ceiling)
510 return;
511 }
512 if (end - 1 > ceiling - 1)
513 end -= PMD_SIZE;
514 if (addr > end - 1)
515 return;
516
517 pgd = pgd_offset(tlb->mm, addr);
518 do {
519 next = pgd_addr_end(addr, end);
520 if (pgd_none_or_clear_bad(pgd))
521 continue;
522 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
523 } while (pgd++, addr = next, addr != end);
524}
525
526void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
527 unsigned long floor, unsigned long ceiling)
528{
529 while (vma) {
530 struct vm_area_struct *next = vma->vm_next;
531 unsigned long addr = vma->vm_start;
532
533 /*
534 * Hide vma from rmap and truncate_pagecache before freeing
535 * pgtables
536 */
537 unlink_anon_vmas(vma);
538 unlink_file_vma(vma);
539
540 if (is_vm_hugetlb_page(vma)) {
541 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
542 floor, next? next->vm_start: ceiling);
543 } else {
544 /*
545 * Optimization: gather nearby vmas into one call down
546 */
547 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
548 && !is_vm_hugetlb_page(next)) {
549 vma = next;
550 next = vma->vm_next;
551 unlink_anon_vmas(vma);
552 unlink_file_vma(vma);
553 }
554 free_pgd_range(tlb, addr, vma->vm_end,
555 floor, next? next->vm_start: ceiling);
556 }
557 vma = next;
558 }
559}
560
561int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
562 pmd_t *pmd, unsigned long address)
563{
564 spinlock_t *ptl;
565 pgtable_t new = pte_alloc_one(mm, address);
566 int wait_split_huge_page;
567 if (!new)
568 return -ENOMEM;
569
570 /*
571 * Ensure all pte setup (eg. pte page lock and page clearing) are
572 * visible before the pte is made visible to other CPUs by being
573 * put into page tables.
574 *
575 * The other side of the story is the pointer chasing in the page
576 * table walking code (when walking the page table without locking;
577 * ie. most of the time). Fortunately, these data accesses consist
578 * of a chain of data-dependent loads, meaning most CPUs (alpha
579 * being the notable exception) will already guarantee loads are
580 * seen in-order. See the alpha page table accessors for the
581 * smp_read_barrier_depends() barriers in page table walking code.
582 */
583 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
584
585 ptl = pmd_lock(mm, pmd);
586 wait_split_huge_page = 0;
587 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
588 atomic_long_inc(&mm->nr_ptes);
589 pmd_populate(mm, pmd, new);
590 new = NULL;
591 } else if (unlikely(pmd_trans_splitting(*pmd)))
592 wait_split_huge_page = 1;
593 spin_unlock(ptl);
594 if (new)
595 pte_free(mm, new);
596 if (wait_split_huge_page)
597 wait_split_huge_page(vma->anon_vma, pmd);
598 return 0;
599}
600
601int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
602{
603 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
604 if (!new)
605 return -ENOMEM;
606
607 smp_wmb(); /* See comment in __pte_alloc */
608
609 spin_lock(&init_mm.page_table_lock);
610 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
611 pmd_populate_kernel(&init_mm, pmd, new);
612 new = NULL;
613 } else
614 VM_BUG_ON(pmd_trans_splitting(*pmd));
615 spin_unlock(&init_mm.page_table_lock);
616 if (new)
617 pte_free_kernel(&init_mm, new);
618 return 0;
619}
620
621static inline void init_rss_vec(int *rss)
622{
623 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
624}
625
626static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
627{
628 int i;
629
630 if (current->mm == mm)
631 sync_mm_rss(mm);
632 for (i = 0; i < NR_MM_COUNTERS; i++)
633 if (rss[i])
634 add_mm_counter(mm, i, rss[i]);
635}
636
637/*
638 * This function is called to print an error when a bad pte
639 * is found. For example, we might have a PFN-mapped pte in
640 * a region that doesn't allow it.
641 *
642 * The calling function must still handle the error.
643 */
644static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
645 pte_t pte, struct page *page)
646{
647 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
648 pud_t *pud = pud_offset(pgd, addr);
649 pmd_t *pmd = pmd_offset(pud, addr);
650 struct address_space *mapping;
651 pgoff_t index;
652 static unsigned long resume;
653 static unsigned long nr_shown;
654 static unsigned long nr_unshown;
655
656 /*
657 * Allow a burst of 60 reports, then keep quiet for that minute;
658 * or allow a steady drip of one report per second.
659 */
660 if (nr_shown == 60) {
661 if (time_before(jiffies, resume)) {
662 nr_unshown++;
663 return;
664 }
665 if (nr_unshown) {
666 printk(KERN_ALERT
667 "BUG: Bad page map: %lu messages suppressed\n",
668 nr_unshown);
669 nr_unshown = 0;
670 }
671 nr_shown = 0;
672 }
673 if (nr_shown++ == 0)
674 resume = jiffies + 60 * HZ;
675
676 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
677 index = linear_page_index(vma, addr);
678
679 printk(KERN_ALERT
680 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
681 current->comm,
682 (long long)pte_val(pte), (long long)pmd_val(*pmd));
683 if (page)
684 dump_page(page, "bad pte");
685 printk(KERN_ALERT
686 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
687 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
688 /*
689 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
690 */
691 if (vma->vm_ops)
692 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
693 vma->vm_ops->fault);
694 if (vma->vm_file)
695 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
696 vma->vm_file->f_op->mmap);
697 dump_stack();
698 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
699}
700
701static inline bool is_cow_mapping(vm_flags_t flags)
702{
703 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
704}
705
706/*
707 * vm_normal_page -- This function gets the "struct page" associated with a pte.
708 *
709 * "Special" mappings do not wish to be associated with a "struct page" (either
710 * it doesn't exist, or it exists but they don't want to touch it). In this
711 * case, NULL is returned here. "Normal" mappings do have a struct page.
712 *
713 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
714 * pte bit, in which case this function is trivial. Secondly, an architecture
715 * may not have a spare pte bit, which requires a more complicated scheme,
716 * described below.
717 *
718 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
719 * special mapping (even if there are underlying and valid "struct pages").
720 * COWed pages of a VM_PFNMAP are always normal.
721 *
722 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
723 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
724 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
725 * mapping will always honor the rule
726 *
727 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
728 *
729 * And for normal mappings this is false.
730 *
731 * This restricts such mappings to be a linear translation from virtual address
732 * to pfn. To get around this restriction, we allow arbitrary mappings so long
733 * as the vma is not a COW mapping; in that case, we know that all ptes are
734 * special (because none can have been COWed).
735 *
736 *
737 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
738 *
739 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
740 * page" backing, however the difference is that _all_ pages with a struct
741 * page (that is, those where pfn_valid is true) are refcounted and considered
742 * normal pages by the VM. The disadvantage is that pages are refcounted
743 * (which can be slower and simply not an option for some PFNMAP users). The
744 * advantage is that we don't have to follow the strict linearity rule of
745 * PFNMAP mappings in order to support COWable mappings.
746 *
747 */
748#ifdef __HAVE_ARCH_PTE_SPECIAL
749# define HAVE_PTE_SPECIAL 1
750#else
751# define HAVE_PTE_SPECIAL 0
752#endif
753struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
754 pte_t pte)
755{
756 unsigned long pfn = pte_pfn(pte);
757
758 if (HAVE_PTE_SPECIAL) {
759 if (likely(!pte_special(pte)))
760 goto check_pfn;
761 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
762 return NULL;
763 if (!is_zero_pfn(pfn))
764 print_bad_pte(vma, addr, pte, NULL);
765 return NULL;
766 }
767
768 /* !HAVE_PTE_SPECIAL case follows: */
769
770 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
771 if (vma->vm_flags & VM_MIXEDMAP) {
772 if (!pfn_valid(pfn))
773 return NULL;
774 goto out;
775 } else {
776 unsigned long off;
777 off = (addr - vma->vm_start) >> PAGE_SHIFT;
778 if (pfn == vma->vm_pgoff + off)
779 return NULL;
780 if (!is_cow_mapping(vma->vm_flags))
781 return NULL;
782 }
783 }
784
785 if (is_zero_pfn(pfn))
786 return NULL;
787check_pfn:
788 if (unlikely(pfn > highest_memmap_pfn)) {
789 print_bad_pte(vma, addr, pte, NULL);
790 return NULL;
791 }
792
793 /*
794 * NOTE! We still have PageReserved() pages in the page tables.
795 * eg. VDSO mappings can cause them to exist.
796 */
797out:
798 return pfn_to_page(pfn);
799}
800
801/*
802 * copy one vm_area from one task to the other. Assumes the page tables
803 * already present in the new task to be cleared in the whole range
804 * covered by this vma.
805 */
806
807static inline unsigned long
808copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
809 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
810 unsigned long addr, int *rss)
811{
812 unsigned long vm_flags = vma->vm_flags;
813 pte_t pte = *src_pte;
814 struct page *page;
815
816 /* pte contains position in swap or file, so copy. */
817 if (unlikely(!pte_present(pte))) {
818 if (!pte_file(pte)) {
819 swp_entry_t entry = pte_to_swp_entry(pte);
820
821 if (swap_duplicate(entry) < 0)
822 return entry.val;
823
824 /* make sure dst_mm is on swapoff's mmlist. */
825 if (unlikely(list_empty(&dst_mm->mmlist))) {
826 spin_lock(&mmlist_lock);
827 if (list_empty(&dst_mm->mmlist))
828 list_add(&dst_mm->mmlist,
829 &src_mm->mmlist);
830 spin_unlock(&mmlist_lock);
831 }
832 if (likely(!non_swap_entry(entry)))
833 rss[MM_SWAPENTS]++;
834 else if (is_migration_entry(entry)) {
835 page = migration_entry_to_page(entry);
836
837 if (PageAnon(page))
838 rss[MM_ANONPAGES]++;
839 else
840 rss[MM_FILEPAGES]++;
841
842 if (is_write_migration_entry(entry) &&
843 is_cow_mapping(vm_flags)) {
844 /*
845 * COW mappings require pages in both
846 * parent and child to be set to read.
847 */
848 make_migration_entry_read(&entry);
849 pte = swp_entry_to_pte(entry);
850 if (pte_swp_soft_dirty(*src_pte))
851 pte = pte_swp_mksoft_dirty(pte);
852 set_pte_at(src_mm, addr, src_pte, pte);
853 }
854 }
855 }
856 goto out_set_pte;
857 }
858
859 /*
860 * If it's a COW mapping, write protect it both
861 * in the parent and the child
862 */
863 if (is_cow_mapping(vm_flags)) {
864 ptep_set_wrprotect(src_mm, addr, src_pte);
865 pte = pte_wrprotect(pte);
866 }
867
868 /*
869 * If it's a shared mapping, mark it clean in
870 * the child
871 */
872 if (vm_flags & VM_SHARED)
873 pte = pte_mkclean(pte);
874 pte = pte_mkold(pte);
875
876 page = vm_normal_page(vma, addr, pte);
877 if (page) {
878 get_page(page);
879 page_dup_rmap(page);
880 if (PageAnon(page))
881 rss[MM_ANONPAGES]++;
882 else
883 rss[MM_FILEPAGES]++;
884 }
885
886out_set_pte:
887 set_pte_at(dst_mm, addr, dst_pte, pte);
888 return 0;
889}
890
891int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
892 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
893 unsigned long addr, unsigned long end)
894{
895 pte_t *orig_src_pte, *orig_dst_pte;
896 pte_t *src_pte, *dst_pte;
897 spinlock_t *src_ptl, *dst_ptl;
898 int progress = 0;
899 int rss[NR_MM_COUNTERS];
900 swp_entry_t entry = (swp_entry_t){0};
901
902again:
903 init_rss_vec(rss);
904
905 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
906 if (!dst_pte)
907 return -ENOMEM;
908 src_pte = pte_offset_map(src_pmd, addr);
909 src_ptl = pte_lockptr(src_mm, src_pmd);
910 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
911 orig_src_pte = src_pte;
912 orig_dst_pte = dst_pte;
913 arch_enter_lazy_mmu_mode();
914
915 do {
916 /*
917 * We are holding two locks at this point - either of them
918 * could generate latencies in another task on another CPU.
919 */
920 if (progress >= 32) {
921 progress = 0;
922 if (need_resched() ||
923 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
924 break;
925 }
926 if (pte_none(*src_pte)) {
927 progress++;
928 continue;
929 }
930 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
931 vma, addr, rss);
932 if (entry.val)
933 break;
934 progress += 8;
935 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
936
937 arch_leave_lazy_mmu_mode();
938 spin_unlock(src_ptl);
939 pte_unmap(orig_src_pte);
940 add_mm_rss_vec(dst_mm, rss);
941 pte_unmap_unlock(orig_dst_pte, dst_ptl);
942 cond_resched();
943
944 if (entry.val) {
945 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
946 return -ENOMEM;
947 progress = 0;
948 }
949 if (addr != end)
950 goto again;
951 return 0;
952}
953
954static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
955 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
956 unsigned long addr, unsigned long end)
957{
958 pmd_t *src_pmd, *dst_pmd;
959 unsigned long next;
960
961 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
962 if (!dst_pmd)
963 return -ENOMEM;
964 src_pmd = pmd_offset(src_pud, addr);
965 do {
966 next = pmd_addr_end(addr, end);
967 if (pmd_trans_huge(*src_pmd)) {
968 int err;
969 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
970 err = copy_huge_pmd(dst_mm, src_mm,
971 dst_pmd, src_pmd, addr, vma);
972 if (err == -ENOMEM)
973 return -ENOMEM;
974 if (!err)
975 continue;
976 /* fall through */
977 }
978 if (pmd_none_or_clear_bad(src_pmd))
979 continue;
980 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
981 vma, addr, next))
982 return -ENOMEM;
983 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
984 return 0;
985}
986
987static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
988 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
989 unsigned long addr, unsigned long end)
990{
991 pud_t *src_pud, *dst_pud;
992 unsigned long next;
993
994 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
995 if (!dst_pud)
996 return -ENOMEM;
997 src_pud = pud_offset(src_pgd, addr);
998 do {
999 next = pud_addr_end(addr, end);
1000 if (pud_none_or_clear_bad(src_pud))
1001 continue;
1002 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1003 vma, addr, next))
1004 return -ENOMEM;
1005 } while (dst_pud++, src_pud++, addr = next, addr != end);
1006 return 0;
1007}
1008
1009int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1010 struct vm_area_struct *vma)
1011{
1012 pgd_t *src_pgd, *dst_pgd;
1013 unsigned long next;
1014 unsigned long addr = vma->vm_start;
1015 unsigned long end = vma->vm_end;
1016 unsigned long mmun_start; /* For mmu_notifiers */
1017 unsigned long mmun_end; /* For mmu_notifiers */
1018 bool is_cow;
1019 int ret;
1020
1021 /*
1022 * Don't copy ptes where a page fault will fill them correctly.
1023 * Fork becomes much lighter when there are big shared or private
1024 * readonly mappings. The tradeoff is that copy_page_range is more
1025 * efficient than faulting.
1026 */
1027 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1028 VM_PFNMAP | VM_MIXEDMAP))) {
1029 if (!vma->anon_vma)
1030 return 0;
1031 }
1032
1033 if (is_vm_hugetlb_page(vma))
1034 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1035
1036 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1037 /*
1038 * We do not free on error cases below as remove_vma
1039 * gets called on error from higher level routine
1040 */
1041 ret = track_pfn_copy(vma);
1042 if (ret)
1043 return ret;
1044 }
1045
1046 /*
1047 * We need to invalidate the secondary MMU mappings only when
1048 * there could be a permission downgrade on the ptes of the
1049 * parent mm. And a permission downgrade will only happen if
1050 * is_cow_mapping() returns true.
1051 */
1052 is_cow = is_cow_mapping(vma->vm_flags);
1053 mmun_start = addr;
1054 mmun_end = end;
1055 if (is_cow)
1056 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1057 mmun_end);
1058
1059 ret = 0;
1060 dst_pgd = pgd_offset(dst_mm, addr);
1061 src_pgd = pgd_offset(src_mm, addr);
1062 do {
1063 next = pgd_addr_end(addr, end);
1064 if (pgd_none_or_clear_bad(src_pgd))
1065 continue;
1066 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1067 vma, addr, next))) {
1068 ret = -ENOMEM;
1069 break;
1070 }
1071 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1072
1073 if (is_cow)
1074 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1075 return ret;
1076}
1077
1078static unsigned long zap_pte_range(struct mmu_gather *tlb,
1079 struct vm_area_struct *vma, pmd_t *pmd,
1080 unsigned long addr, unsigned long end,
1081 struct zap_details *details)
1082{
1083 struct mm_struct *mm = tlb->mm;
1084 int force_flush = 0;
1085 int rss[NR_MM_COUNTERS];
1086 spinlock_t *ptl;
1087 pte_t *start_pte;
1088 pte_t *pte;
1089
1090again:
1091 init_rss_vec(rss);
1092 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1093 pte = start_pte;
1094 arch_enter_lazy_mmu_mode();
1095 do {
1096 pte_t ptent = *pte;
1097 if (pte_none(ptent)) {
1098 continue;
1099 }
1100
1101 if (pte_present(ptent)) {
1102 struct page *page;
1103
1104 page = vm_normal_page(vma, addr, ptent);
1105 if (unlikely(details) && page) {
1106 /*
1107 * unmap_shared_mapping_pages() wants to
1108 * invalidate cache without truncating:
1109 * unmap shared but keep private pages.
1110 */
1111 if (details->check_mapping &&
1112 details->check_mapping != page->mapping)
1113 continue;
1114 /*
1115 * Each page->index must be checked when
1116 * invalidating or truncating nonlinear.
1117 */
1118 if (details->nonlinear_vma &&
1119 (page->index < details->first_index ||
1120 page->index > details->last_index))
1121 continue;
1122 }
1123 ptent = ptep_get_and_clear_full(mm, addr, pte,
1124 tlb->fullmm);
1125 tlb_remove_tlb_entry(tlb, pte, addr);
1126 if (unlikely(!page))
1127 continue;
1128 if (unlikely(details) && details->nonlinear_vma
1129 && linear_page_index(details->nonlinear_vma,
1130 addr) != page->index) {
1131 pte_t ptfile = pgoff_to_pte(page->index);
1132 if (pte_soft_dirty(ptent))
1133 pte_file_mksoft_dirty(ptfile);
1134 set_pte_at(mm, addr, pte, ptfile);
1135 }
1136 if (PageAnon(page))
1137 rss[MM_ANONPAGES]--;
1138 else {
1139 if (pte_dirty(ptent)) {
1140 force_flush = 1;
1141 set_page_dirty(page);
1142 }
1143 if (pte_young(ptent) &&
1144 likely(!(vma->vm_flags & VM_SEQ_READ)))
1145 mark_page_accessed(page);
1146 rss[MM_FILEPAGES]--;
1147 }
1148 page_remove_rmap(page);
1149 if (unlikely(page_mapcount(page) < 0))
1150 print_bad_pte(vma, addr, ptent, page);
1151 if (unlikely(!__tlb_remove_page(tlb, page))) {
1152 force_flush = 1;
1153 break;
1154 }
1155 continue;
1156 }
1157 /*
1158 * If details->check_mapping, we leave swap entries;
1159 * if details->nonlinear_vma, we leave file entries.
1160 */
1161 if (unlikely(details))
1162 continue;
1163 if (pte_file(ptent)) {
1164 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1165 print_bad_pte(vma, addr, ptent, NULL);
1166 } else {
1167 swp_entry_t entry = pte_to_swp_entry(ptent);
1168
1169 if (!non_swap_entry(entry))
1170 rss[MM_SWAPENTS]--;
1171 else if (is_migration_entry(entry)) {
1172 struct page *page;
1173
1174 page = migration_entry_to_page(entry);
1175
1176 if (PageAnon(page))
1177 rss[MM_ANONPAGES]--;
1178 else
1179 rss[MM_FILEPAGES]--;
1180 }
1181 if (unlikely(!free_swap_and_cache(entry)))
1182 print_bad_pte(vma, addr, ptent, NULL);
1183 }
1184 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1185 } while (pte++, addr += PAGE_SIZE, addr != end);
1186
1187 add_mm_rss_vec(mm, rss);
1188 arch_leave_lazy_mmu_mode();
1189
1190 /* Do the actual TLB flush before dropping ptl */
1191 if (force_flush) {
1192 unsigned long old_end;
1193
1194 /*
1195 * Flush the TLB just for the previous segment,
1196 * then update the range to be the remaining
1197 * TLB range.
1198 */
1199 old_end = tlb->end;
1200 tlb->end = addr;
1201 tlb_flush_mmu_tlbonly(tlb);
1202 tlb->start = addr;
1203 tlb->end = old_end;
1204 }
1205 pte_unmap_unlock(start_pte, ptl);
1206
1207 /*
1208 * If we forced a TLB flush (either due to running out of
1209 * batch buffers or because we needed to flush dirty TLB
1210 * entries before releasing the ptl), free the batched
1211 * memory too. Restart if we didn't do everything.
1212 */
1213 if (force_flush) {
1214 force_flush = 0;
1215 tlb_flush_mmu_free(tlb);
1216
1217 if (addr != end)
1218 goto again;
1219 }
1220
1221 return addr;
1222}
1223
1224static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225 struct vm_area_struct *vma, pud_t *pud,
1226 unsigned long addr, unsigned long end,
1227 struct zap_details *details)
1228{
1229 pmd_t *pmd;
1230 unsigned long next;
1231
1232 pmd = pmd_offset(pud, addr);
1233 do {
1234 next = pmd_addr_end(addr, end);
1235 if (pmd_trans_huge(*pmd)) {
1236 if (next - addr != HPAGE_PMD_SIZE) {
1237#ifdef CONFIG_DEBUG_VM
1238 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1239 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240 __func__, addr, end,
1241 vma->vm_start,
1242 vma->vm_end);
1243 BUG();
1244 }
1245#endif
1246 split_huge_page_pmd(vma, addr, pmd);
1247 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1248 goto next;
1249 /* fall through */
1250 }
1251 /*
1252 * Here there can be other concurrent MADV_DONTNEED or
1253 * trans huge page faults running, and if the pmd is
1254 * none or trans huge it can change under us. This is
1255 * because MADV_DONTNEED holds the mmap_sem in read
1256 * mode.
1257 */
1258 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1259 goto next;
1260 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1261next:
1262 cond_resched();
1263 } while (pmd++, addr = next, addr != end);
1264
1265 return addr;
1266}
1267
1268static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1269 struct vm_area_struct *vma, pgd_t *pgd,
1270 unsigned long addr, unsigned long end,
1271 struct zap_details *details)
1272{
1273 pud_t *pud;
1274 unsigned long next;
1275
1276 pud = pud_offset(pgd, addr);
1277 do {
1278 next = pud_addr_end(addr, end);
1279 if (pud_none_or_clear_bad(pud))
1280 continue;
1281 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1282 } while (pud++, addr = next, addr != end);
1283
1284 return addr;
1285}
1286
1287static void unmap_page_range(struct mmu_gather *tlb,
1288 struct vm_area_struct *vma,
1289 unsigned long addr, unsigned long end,
1290 struct zap_details *details)
1291{
1292 pgd_t *pgd;
1293 unsigned long next;
1294
1295 if (details && !details->check_mapping && !details->nonlinear_vma)
1296 details = NULL;
1297
1298 BUG_ON(addr >= end);
1299 mem_cgroup_uncharge_start();
1300 tlb_start_vma(tlb, vma);
1301 pgd = pgd_offset(vma->vm_mm, addr);
1302 do {
1303 next = pgd_addr_end(addr, end);
1304 if (pgd_none_or_clear_bad(pgd))
1305 continue;
1306 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307 } while (pgd++, addr = next, addr != end);
1308 tlb_end_vma(tlb, vma);
1309 mem_cgroup_uncharge_end();
1310}
1311
1312
1313static void unmap_single_vma(struct mmu_gather *tlb,
1314 struct vm_area_struct *vma, unsigned long start_addr,
1315 unsigned long end_addr,
1316 struct zap_details *details)
1317{
1318 unsigned long start = max(vma->vm_start, start_addr);
1319 unsigned long end;
1320
1321 if (start >= vma->vm_end)
1322 return;
1323 end = min(vma->vm_end, end_addr);
1324 if (end <= vma->vm_start)
1325 return;
1326
1327 if (vma->vm_file)
1328 uprobe_munmap(vma, start, end);
1329
1330 if (unlikely(vma->vm_flags & VM_PFNMAP))
1331 untrack_pfn(vma, 0, 0);
1332
1333 if (start != end) {
1334 if (unlikely(is_vm_hugetlb_page(vma))) {
1335 /*
1336 * It is undesirable to test vma->vm_file as it
1337 * should be non-null for valid hugetlb area.
1338 * However, vm_file will be NULL in the error
1339 * cleanup path of mmap_region. When
1340 * hugetlbfs ->mmap method fails,
1341 * mmap_region() nullifies vma->vm_file
1342 * before calling this function to clean up.
1343 * Since no pte has actually been setup, it is
1344 * safe to do nothing in this case.
1345 */
1346 if (vma->vm_file) {
1347 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1348 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1349 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1350 }
1351 } else
1352 unmap_page_range(tlb, vma, start, end, details);
1353 }
1354}
1355
1356/**
1357 * unmap_vmas - unmap a range of memory covered by a list of vma's
1358 * @tlb: address of the caller's struct mmu_gather
1359 * @vma: the starting vma
1360 * @start_addr: virtual address at which to start unmapping
1361 * @end_addr: virtual address at which to end unmapping
1362 *
1363 * Unmap all pages in the vma list.
1364 *
1365 * Only addresses between `start' and `end' will be unmapped.
1366 *
1367 * The VMA list must be sorted in ascending virtual address order.
1368 *
1369 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370 * range after unmap_vmas() returns. So the only responsibility here is to
1371 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372 * drops the lock and schedules.
1373 */
1374void unmap_vmas(struct mmu_gather *tlb,
1375 struct vm_area_struct *vma, unsigned long start_addr,
1376 unsigned long end_addr)
1377{
1378 struct mm_struct *mm = vma->vm_mm;
1379
1380 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1381 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1382 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1383 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1384}
1385
1386/**
1387 * zap_page_range - remove user pages in a given range
1388 * @vma: vm_area_struct holding the applicable pages
1389 * @start: starting address of pages to zap
1390 * @size: number of bytes to zap
1391 * @details: details of nonlinear truncation or shared cache invalidation
1392 *
1393 * Caller must protect the VMA list
1394 */
1395void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1396 unsigned long size, struct zap_details *details)
1397{
1398 struct mm_struct *mm = vma->vm_mm;
1399 struct mmu_gather tlb;
1400 unsigned long end = start + size;
1401
1402 lru_add_drain();
1403 tlb_gather_mmu(&tlb, mm, start, end);
1404 update_hiwater_rss(mm);
1405 mmu_notifier_invalidate_range_start(mm, start, end);
1406 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1407 unmap_single_vma(&tlb, vma, start, end, details);
1408 mmu_notifier_invalidate_range_end(mm, start, end);
1409 tlb_finish_mmu(&tlb, start, end);
1410}
1411
1412/**
1413 * zap_page_range_single - remove user pages in a given range
1414 * @vma: vm_area_struct holding the applicable pages
1415 * @address: starting address of pages to zap
1416 * @size: number of bytes to zap
1417 * @details: details of nonlinear truncation or shared cache invalidation
1418 *
1419 * The range must fit into one VMA.
1420 */
1421static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1422 unsigned long size, struct zap_details *details)
1423{
1424 struct mm_struct *mm = vma->vm_mm;
1425 struct mmu_gather tlb;
1426 unsigned long end = address + size;
1427
1428 lru_add_drain();
1429 tlb_gather_mmu(&tlb, mm, address, end);
1430 update_hiwater_rss(mm);
1431 mmu_notifier_invalidate_range_start(mm, address, end);
1432 unmap_single_vma(&tlb, vma, address, end, details);
1433 mmu_notifier_invalidate_range_end(mm, address, end);
1434 tlb_finish_mmu(&tlb, address, end);
1435}
1436
1437/**
1438 * zap_vma_ptes - remove ptes mapping the vma
1439 * @vma: vm_area_struct holding ptes to be zapped
1440 * @address: starting address of pages to zap
1441 * @size: number of bytes to zap
1442 *
1443 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444 *
1445 * The entire address range must be fully contained within the vma.
1446 *
1447 * Returns 0 if successful.
1448 */
1449int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1450 unsigned long size)
1451{
1452 if (address < vma->vm_start || address + size > vma->vm_end ||
1453 !(vma->vm_flags & VM_PFNMAP))
1454 return -1;
1455 zap_page_range_single(vma, address, size, NULL);
1456 return 0;
1457}
1458EXPORT_SYMBOL_GPL(zap_vma_ptes);
1459
1460/**
1461 * follow_page_mask - look up a page descriptor from a user-virtual address
1462 * @vma: vm_area_struct mapping @address
1463 * @address: virtual address to look up
1464 * @flags: flags modifying lookup behaviour
1465 * @page_mask: on output, *page_mask is set according to the size of the page
1466 *
1467 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1468 *
1469 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1470 * an error pointer if there is a mapping to something not represented
1471 * by a page descriptor (see also vm_normal_page()).
1472 */
1473struct page *follow_page_mask(struct vm_area_struct *vma,
1474 unsigned long address, unsigned int flags,
1475 unsigned int *page_mask)
1476{
1477 pgd_t *pgd;
1478 pud_t *pud;
1479 pmd_t *pmd;
1480 pte_t *ptep, pte;
1481 spinlock_t *ptl;
1482 struct page *page;
1483 struct mm_struct *mm = vma->vm_mm;
1484
1485 *page_mask = 0;
1486
1487 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1488 if (!IS_ERR(page)) {
1489 BUG_ON(flags & FOLL_GET);
1490 goto out;
1491 }
1492
1493 page = NULL;
1494 pgd = pgd_offset(mm, address);
1495 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1496 goto no_page_table;
1497
1498 pud = pud_offset(pgd, address);
1499 if (pud_none(*pud))
1500 goto no_page_table;
1501 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1502 if (flags & FOLL_GET)
1503 goto out;
1504 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1505 goto out;
1506 }
1507 if (unlikely(pud_bad(*pud)))
1508 goto no_page_table;
1509
1510 pmd = pmd_offset(pud, address);
1511 if (pmd_none(*pmd))
1512 goto no_page_table;
1513 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1514 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1515 if (flags & FOLL_GET) {
1516 /*
1517 * Refcount on tail pages are not well-defined and
1518 * shouldn't be taken. The caller should handle a NULL
1519 * return when trying to follow tail pages.
1520 */
1521 if (PageHead(page))
1522 get_page(page);
1523 else {
1524 page = NULL;
1525 goto out;
1526 }
1527 }
1528 goto out;
1529 }
1530 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1531 goto no_page_table;
1532 if (pmd_trans_huge(*pmd)) {
1533 if (flags & FOLL_SPLIT) {
1534 split_huge_page_pmd(vma, address, pmd);
1535 goto split_fallthrough;
1536 }
1537 ptl = pmd_lock(mm, pmd);
1538 if (likely(pmd_trans_huge(*pmd))) {
1539 if (unlikely(pmd_trans_splitting(*pmd))) {
1540 spin_unlock(ptl);
1541 wait_split_huge_page(vma->anon_vma, pmd);
1542 } else {
1543 page = follow_trans_huge_pmd(vma, address,
1544 pmd, flags);
1545 spin_unlock(ptl);
1546 *page_mask = HPAGE_PMD_NR - 1;
1547 goto out;
1548 }
1549 } else
1550 spin_unlock(ptl);
1551 /* fall through */
1552 }
1553split_fallthrough:
1554 if (unlikely(pmd_bad(*pmd)))
1555 goto no_page_table;
1556
1557 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1558
1559 pte = *ptep;
1560 if (!pte_present(pte)) {
1561 swp_entry_t entry;
1562 /*
1563 * KSM's break_ksm() relies upon recognizing a ksm page
1564 * even while it is being migrated, so for that case we
1565 * need migration_entry_wait().
1566 */
1567 if (likely(!(flags & FOLL_MIGRATION)))
1568 goto no_page;
1569 if (pte_none(pte) || pte_file(pte))
1570 goto no_page;
1571 entry = pte_to_swp_entry(pte);
1572 if (!is_migration_entry(entry))
1573 goto no_page;
1574 pte_unmap_unlock(ptep, ptl);
1575 migration_entry_wait(mm, pmd, address);
1576 goto split_fallthrough;
1577 }
1578 if ((flags & FOLL_NUMA) && pte_numa(pte))
1579 goto no_page;
1580 if ((flags & FOLL_WRITE) && !pte_write(pte))
1581 goto unlock;
1582
1583 page = vm_normal_page(vma, address, pte);
1584 if (unlikely(!page)) {
1585 if ((flags & FOLL_DUMP) ||
1586 !is_zero_pfn(pte_pfn(pte)))
1587 goto bad_page;
1588 page = pte_page(pte);
1589 }
1590
1591 if (flags & FOLL_GET)
1592 get_page_foll(page);
1593 if (flags & FOLL_TOUCH) {
1594 if ((flags & FOLL_WRITE) &&
1595 !pte_dirty(pte) && !PageDirty(page))
1596 set_page_dirty(page);
1597 /*
1598 * pte_mkyoung() would be more correct here, but atomic care
1599 * is needed to avoid losing the dirty bit: it is easier to use
1600 * mark_page_accessed().
1601 */
1602 mark_page_accessed(page);
1603 }
1604 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1605 /*
1606 * The preliminary mapping check is mainly to avoid the
1607 * pointless overhead of lock_page on the ZERO_PAGE
1608 * which might bounce very badly if there is contention.
1609 *
1610 * If the page is already locked, we don't need to
1611 * handle it now - vmscan will handle it later if and
1612 * when it attempts to reclaim the page.
1613 */
1614 if (page->mapping && trylock_page(page)) {
1615 lru_add_drain(); /* push cached pages to LRU */
1616 /*
1617 * Because we lock page here, and migration is
1618 * blocked by the pte's page reference, and we
1619 * know the page is still mapped, we don't even
1620 * need to check for file-cache page truncation.
1621 */
1622 mlock_vma_page(page);
1623 unlock_page(page);
1624 }
1625 }
1626unlock:
1627 pte_unmap_unlock(ptep, ptl);
1628out:
1629 return page;
1630
1631bad_page:
1632 pte_unmap_unlock(ptep, ptl);
1633 return ERR_PTR(-EFAULT);
1634
1635no_page:
1636 pte_unmap_unlock(ptep, ptl);
1637 if (!pte_none(pte))
1638 return page;
1639
1640no_page_table:
1641 /*
1642 * When core dumping an enormous anonymous area that nobody
1643 * has touched so far, we don't want to allocate unnecessary pages or
1644 * page tables. Return error instead of NULL to skip handle_mm_fault,
1645 * then get_dump_page() will return NULL to leave a hole in the dump.
1646 * But we can only make this optimization where a hole would surely
1647 * be zero-filled if handle_mm_fault() actually did handle it.
1648 */
1649 if ((flags & FOLL_DUMP) &&
1650 (!vma->vm_ops || !vma->vm_ops->fault))
1651 return ERR_PTR(-EFAULT);
1652 return page;
1653}
1654
1655static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1656{
1657 return stack_guard_page_start(vma, addr) ||
1658 stack_guard_page_end(vma, addr+PAGE_SIZE);
1659}
1660
1661/**
1662 * __get_user_pages() - pin user pages in memory
1663 * @tsk: task_struct of target task
1664 * @mm: mm_struct of target mm
1665 * @start: starting user address
1666 * @nr_pages: number of pages from start to pin
1667 * @gup_flags: flags modifying pin behaviour
1668 * @pages: array that receives pointers to the pages pinned.
1669 * Should be at least nr_pages long. Or NULL, if caller
1670 * only intends to ensure the pages are faulted in.
1671 * @vmas: array of pointers to vmas corresponding to each page.
1672 * Or NULL if the caller does not require them.
1673 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1674 *
1675 * Returns number of pages pinned. This may be fewer than the number
1676 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1677 * were pinned, returns -errno. Each page returned must be released
1678 * with a put_page() call when it is finished with. vmas will only
1679 * remain valid while mmap_sem is held.
1680 *
1681 * Must be called with mmap_sem held for read or write.
1682 *
1683 * __get_user_pages walks a process's page tables and takes a reference to
1684 * each struct page that each user address corresponds to at a given
1685 * instant. That is, it takes the page that would be accessed if a user
1686 * thread accesses the given user virtual address at that instant.
1687 *
1688 * This does not guarantee that the page exists in the user mappings when
1689 * __get_user_pages returns, and there may even be a completely different
1690 * page there in some cases (eg. if mmapped pagecache has been invalidated
1691 * and subsequently re faulted). However it does guarantee that the page
1692 * won't be freed completely. And mostly callers simply care that the page
1693 * contains data that was valid *at some point in time*. Typically, an IO
1694 * or similar operation cannot guarantee anything stronger anyway because
1695 * locks can't be held over the syscall boundary.
1696 *
1697 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1698 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1699 * appropriate) must be called after the page is finished with, and
1700 * before put_page is called.
1701 *
1702 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1703 * or mmap_sem contention, and if waiting is needed to pin all pages,
1704 * *@nonblocking will be set to 0.
1705 *
1706 * In most cases, get_user_pages or get_user_pages_fast should be used
1707 * instead of __get_user_pages. __get_user_pages should be used only if
1708 * you need some special @gup_flags.
1709 */
1710long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1711 unsigned long start, unsigned long nr_pages,
1712 unsigned int gup_flags, struct page **pages,
1713 struct vm_area_struct **vmas, int *nonblocking)
1714{
1715 long i;
1716 unsigned long vm_flags;
1717 unsigned int page_mask;
1718
1719 if (!nr_pages)
1720 return 0;
1721
1722 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1723
1724 /*
1725 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1726 * would be called on PROT_NONE ranges. We must never invoke
1727 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1728 * page faults would unprotect the PROT_NONE ranges if
1729 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1730 * bitflag. So to avoid that, don't set FOLL_NUMA if
1731 * FOLL_FORCE is set.
1732 */
1733 if (!(gup_flags & FOLL_FORCE))
1734 gup_flags |= FOLL_NUMA;
1735
1736 i = 0;
1737
1738 do {
1739 struct vm_area_struct *vma;
1740
1741 vma = find_extend_vma(mm, start);
1742 if (!vma && in_gate_area(mm, start)) {
1743 unsigned long pg = start & PAGE_MASK;
1744 pgd_t *pgd;
1745 pud_t *pud;
1746 pmd_t *pmd;
1747 pte_t *pte;
1748
1749 /* user gate pages are read-only */
1750 if (gup_flags & FOLL_WRITE)
1751 goto efault;
1752 if (pg > TASK_SIZE)
1753 pgd = pgd_offset_k(pg);
1754 else
1755 pgd = pgd_offset_gate(mm, pg);
1756 BUG_ON(pgd_none(*pgd));
1757 pud = pud_offset(pgd, pg);
1758 BUG_ON(pud_none(*pud));
1759 pmd = pmd_offset(pud, pg);
1760 if (pmd_none(*pmd))
1761 goto efault;
1762 VM_BUG_ON(pmd_trans_huge(*pmd));
1763 pte = pte_offset_map(pmd, pg);
1764 if (pte_none(*pte)) {
1765 pte_unmap(pte);
1766 goto efault;
1767 }
1768 vma = get_gate_vma(mm);
1769 if (pages) {
1770 struct page *page;
1771
1772 page = vm_normal_page(vma, start, *pte);
1773 if (!page) {
1774 if (!(gup_flags & FOLL_DUMP) &&
1775 is_zero_pfn(pte_pfn(*pte)))
1776 page = pte_page(*pte);
1777 else {
1778 pte_unmap(pte);
1779 goto efault;
1780 }
1781 }
1782 pages[i] = page;
1783 get_page(page);
1784 }
1785 pte_unmap(pte);
1786 page_mask = 0;
1787 goto next_page;
1788 }
1789
1790 if (!vma)
1791 goto efault;
1792 vm_flags = vma->vm_flags;
1793 if (vm_flags & (VM_IO | VM_PFNMAP))
1794 goto efault;
1795
1796 if (gup_flags & FOLL_WRITE) {
1797 if (!(vm_flags & VM_WRITE)) {
1798 if (!(gup_flags & FOLL_FORCE))
1799 goto efault;
1800 /*
1801 * We used to let the write,force case do COW
1802 * in a VM_MAYWRITE VM_SHARED !VM_WRITE vma, so
1803 * ptrace could set a breakpoint in a read-only
1804 * mapping of an executable, without corrupting
1805 * the file (yet only when that file had been
1806 * opened for writing!). Anon pages in shared
1807 * mappings are surprising: now just reject it.
1808 */
1809 if (!is_cow_mapping(vm_flags)) {
1810 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
1811 goto efault;
1812 }
1813 }
1814 } else {
1815 if (!(vm_flags & VM_READ)) {
1816 if (!(gup_flags & FOLL_FORCE))
1817 goto efault;
1818 /*
1819 * Is there actually any vma we can reach here
1820 * which does not have VM_MAYREAD set?
1821 */
1822 if (!(vm_flags & VM_MAYREAD))
1823 goto efault;
1824 }
1825 }
1826
1827 if (is_vm_hugetlb_page(vma)) {
1828 i = follow_hugetlb_page(mm, vma, pages, vmas,
1829 &start, &nr_pages, i, gup_flags);
1830 continue;
1831 }
1832
1833 do {
1834 struct page *page;
1835 unsigned int foll_flags = gup_flags;
1836 unsigned int page_increm;
1837
1838 /*
1839 * If we have a pending SIGKILL, don't keep faulting
1840 * pages and potentially allocating memory.
1841 */
1842 if (unlikely(fatal_signal_pending(current)))
1843 return i ? i : -ERESTARTSYS;
1844
1845 cond_resched();
1846 while (!(page = follow_page_mask(vma, start,
1847 foll_flags, &page_mask))) {
1848 int ret;
1849 unsigned int fault_flags = 0;
1850
1851 /* For mlock, just skip the stack guard page. */
1852 if (foll_flags & FOLL_MLOCK) {
1853 if (stack_guard_page(vma, start))
1854 goto next_page;
1855 }
1856 if (foll_flags & FOLL_WRITE)
1857 fault_flags |= FAULT_FLAG_WRITE;
1858 if (nonblocking)
1859 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1860 if (foll_flags & FOLL_NOWAIT)
1861 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1862
1863 ret = handle_mm_fault(mm, vma, start,
1864 fault_flags);
1865
1866 if (ret & VM_FAULT_ERROR) {
1867 if (ret & VM_FAULT_OOM)
1868 return i ? i : -ENOMEM;
1869 if (ret & (VM_FAULT_HWPOISON |
1870 VM_FAULT_HWPOISON_LARGE)) {
1871 if (i)
1872 return i;
1873 else if (gup_flags & FOLL_HWPOISON)
1874 return -EHWPOISON;
1875 else
1876 return -EFAULT;
1877 }
1878 if (ret & VM_FAULT_SIGBUS)
1879 goto efault;
1880 BUG();
1881 }
1882
1883 if (tsk) {
1884 if (ret & VM_FAULT_MAJOR)
1885 tsk->maj_flt++;
1886 else
1887 tsk->min_flt++;
1888 }
1889
1890 if (ret & VM_FAULT_RETRY) {
1891 if (nonblocking)
1892 *nonblocking = 0;
1893 return i;
1894 }
1895
1896 /*
1897 * The VM_FAULT_WRITE bit tells us that
1898 * do_wp_page has broken COW when necessary,
1899 * even if maybe_mkwrite decided not to set
1900 * pte_write. We can thus safely do subsequent
1901 * page lookups as if they were reads. But only
1902 * do so when looping for pte_write is futile:
1903 * in some cases userspace may also be wanting
1904 * to write to the gotten user page, which a
1905 * read fault here might prevent (a readonly
1906 * page might get reCOWed by userspace write).
1907 */
1908 if ((ret & VM_FAULT_WRITE) &&
1909 !(vma->vm_flags & VM_WRITE))
1910 foll_flags &= ~FOLL_WRITE;
1911
1912 cond_resched();
1913 }
1914 if (IS_ERR(page))
1915 return i ? i : PTR_ERR(page);
1916 if (pages) {
1917 pages[i] = page;
1918
1919 flush_anon_page(vma, page, start);
1920 flush_dcache_page(page);
1921 page_mask = 0;
1922 }
1923next_page:
1924 if (vmas) {
1925 vmas[i] = vma;
1926 page_mask = 0;
1927 }
1928 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1929 if (page_increm > nr_pages)
1930 page_increm = nr_pages;
1931 i += page_increm;
1932 start += page_increm * PAGE_SIZE;
1933 nr_pages -= page_increm;
1934 } while (nr_pages && start < vma->vm_end);
1935 } while (nr_pages);
1936 return i;
1937efault:
1938 return i ? : -EFAULT;
1939}
1940EXPORT_SYMBOL(__get_user_pages);
1941
1942/*
1943 * fixup_user_fault() - manually resolve a user page fault
1944 * @tsk: the task_struct to use for page fault accounting, or
1945 * NULL if faults are not to be recorded.
1946 * @mm: mm_struct of target mm
1947 * @address: user address
1948 * @fault_flags:flags to pass down to handle_mm_fault()
1949 *
1950 * This is meant to be called in the specific scenario where for locking reasons
1951 * we try to access user memory in atomic context (within a pagefault_disable()
1952 * section), this returns -EFAULT, and we want to resolve the user fault before
1953 * trying again.
1954 *
1955 * Typically this is meant to be used by the futex code.
1956 *
1957 * The main difference with get_user_pages() is that this function will
1958 * unconditionally call handle_mm_fault() which will in turn perform all the
1959 * necessary SW fixup of the dirty and young bits in the PTE, while
1960 * handle_mm_fault() only guarantees to update these in the struct page.
1961 *
1962 * This is important for some architectures where those bits also gate the
1963 * access permission to the page because they are maintained in software. On
1964 * such architectures, gup() will not be enough to make a subsequent access
1965 * succeed.
1966 *
1967 * This should be called with the mm_sem held for read.
1968 */
1969int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1970 unsigned long address, unsigned int fault_flags)
1971{
1972 struct vm_area_struct *vma;
1973 vm_flags_t vm_flags;
1974 int ret;
1975
1976 vma = find_extend_vma(mm, address);
1977 if (!vma || address < vma->vm_start)
1978 return -EFAULT;
1979
1980 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
1981 if (!(vm_flags & vma->vm_flags))
1982 return -EFAULT;
1983
1984 ret = handle_mm_fault(mm, vma, address, fault_flags);
1985 if (ret & VM_FAULT_ERROR) {
1986 if (ret & VM_FAULT_OOM)
1987 return -ENOMEM;
1988 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1989 return -EHWPOISON;
1990 if (ret & VM_FAULT_SIGBUS)
1991 return -EFAULT;
1992 BUG();
1993 }
1994 if (tsk) {
1995 if (ret & VM_FAULT_MAJOR)
1996 tsk->maj_flt++;
1997 else
1998 tsk->min_flt++;
1999 }
2000 return 0;
2001}
2002
2003/*
2004 * get_user_pages() - pin user pages in memory
2005 * @tsk: the task_struct to use for page fault accounting, or
2006 * NULL if faults are not to be recorded.
2007 * @mm: mm_struct of target mm
2008 * @start: starting user address
2009 * @nr_pages: number of pages from start to pin
2010 * @write: whether pages will be written to by the caller
2011 * @force: whether to force access even when user mapping is currently
2012 * protected (but never forces write access to shared mapping).
2013 * @pages: array that receives pointers to the pages pinned.
2014 * Should be at least nr_pages long. Or NULL, if caller
2015 * only intends to ensure the pages are faulted in.
2016 * @vmas: array of pointers to vmas corresponding to each page.
2017 * Or NULL if the caller does not require them.
2018 *
2019 * Returns number of pages pinned. This may be fewer than the number
2020 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2021 * were pinned, returns -errno. Each page returned must be released
2022 * with a put_page() call when it is finished with. vmas will only
2023 * remain valid while mmap_sem is held.
2024 *
2025 * Must be called with mmap_sem held for read or write.
2026 *
2027 * get_user_pages walks a process's page tables and takes a reference to
2028 * each struct page that each user address corresponds to at a given
2029 * instant. That is, it takes the page that would be accessed if a user
2030 * thread accesses the given user virtual address at that instant.
2031 *
2032 * This does not guarantee that the page exists in the user mappings when
2033 * get_user_pages returns, and there may even be a completely different
2034 * page there in some cases (eg. if mmapped pagecache has been invalidated
2035 * and subsequently re faulted). However it does guarantee that the page
2036 * won't be freed completely. And mostly callers simply care that the page
2037 * contains data that was valid *at some point in time*. Typically, an IO
2038 * or similar operation cannot guarantee anything stronger anyway because
2039 * locks can't be held over the syscall boundary.
2040 *
2041 * If write=0, the page must not be written to. If the page is written to,
2042 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2043 * after the page is finished with, and before put_page is called.
2044 *
2045 * get_user_pages is typically used for fewer-copy IO operations, to get a
2046 * handle on the memory by some means other than accesses via the user virtual
2047 * addresses. The pages may be submitted for DMA to devices or accessed via
2048 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2049 * use the correct cache flushing APIs.
2050 *
2051 * See also get_user_pages_fast, for performance critical applications.
2052 */
2053long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2054 unsigned long start, unsigned long nr_pages, int write,
2055 int force, struct page **pages, struct vm_area_struct **vmas)
2056{
2057 int flags = FOLL_TOUCH;
2058
2059 if (pages)
2060 flags |= FOLL_GET;
2061 if (write)
2062 flags |= FOLL_WRITE;
2063 if (force)
2064 flags |= FOLL_FORCE;
2065
2066 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2067 NULL);
2068}
2069EXPORT_SYMBOL(get_user_pages);
2070
2071/**
2072 * get_dump_page() - pin user page in memory while writing it to core dump
2073 * @addr: user address
2074 *
2075 * Returns struct page pointer of user page pinned for dump,
2076 * to be freed afterwards by page_cache_release() or put_page().
2077 *
2078 * Returns NULL on any kind of failure - a hole must then be inserted into
2079 * the corefile, to preserve alignment with its headers; and also returns
2080 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2081 * allowing a hole to be left in the corefile to save diskspace.
2082 *
2083 * Called without mmap_sem, but after all other threads have been killed.
2084 */
2085#ifdef CONFIG_ELF_CORE
2086struct page *get_dump_page(unsigned long addr)
2087{
2088 struct vm_area_struct *vma;
2089 struct page *page;
2090
2091 if (__get_user_pages(current, current->mm, addr, 1,
2092 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2093 NULL) < 1)
2094 return NULL;
2095 flush_cache_page(vma, addr, page_to_pfn(page));
2096 return page;
2097}
2098#endif /* CONFIG_ELF_CORE */
2099
2100pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2101 spinlock_t **ptl)
2102{
2103 pgd_t * pgd = pgd_offset(mm, addr);
2104 pud_t * pud = pud_alloc(mm, pgd, addr);
2105 if (pud) {
2106 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2107 if (pmd) {
2108 VM_BUG_ON(pmd_trans_huge(*pmd));
2109 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2110 }
2111 }
2112 return NULL;
2113}
2114
2115/*
2116 * This is the old fallback for page remapping.
2117 *
2118 * For historical reasons, it only allows reserved pages. Only
2119 * old drivers should use this, and they needed to mark their
2120 * pages reserved for the old functions anyway.
2121 */
2122static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2123 struct page *page, pgprot_t prot)
2124{
2125 struct mm_struct *mm = vma->vm_mm;
2126 int retval;
2127 pte_t *pte;
2128 spinlock_t *ptl;
2129
2130 retval = -EINVAL;
2131 if (PageAnon(page))
2132 goto out;
2133 retval = -ENOMEM;
2134 flush_dcache_page(page);
2135 pte = get_locked_pte(mm, addr, &ptl);
2136 if (!pte)
2137 goto out;
2138 retval = -EBUSY;
2139 if (!pte_none(*pte))
2140 goto out_unlock;
2141
2142 /* Ok, finally just insert the thing.. */
2143 get_page(page);
2144 inc_mm_counter_fast(mm, MM_FILEPAGES);
2145 page_add_file_rmap(page);
2146 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2147
2148 retval = 0;
2149 pte_unmap_unlock(pte, ptl);
2150 return retval;
2151out_unlock:
2152 pte_unmap_unlock(pte, ptl);
2153out:
2154 return retval;
2155}
2156
2157/**
2158 * vm_insert_page - insert single page into user vma
2159 * @vma: user vma to map to
2160 * @addr: target user address of this page
2161 * @page: source kernel page
2162 *
2163 * This allows drivers to insert individual pages they've allocated
2164 * into a user vma.
2165 *
2166 * The page has to be a nice clean _individual_ kernel allocation.
2167 * If you allocate a compound page, you need to have marked it as
2168 * such (__GFP_COMP), or manually just split the page up yourself
2169 * (see split_page()).
2170 *
2171 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2172 * took an arbitrary page protection parameter. This doesn't allow
2173 * that. Your vma protection will have to be set up correctly, which
2174 * means that if you want a shared writable mapping, you'd better
2175 * ask for a shared writable mapping!
2176 *
2177 * The page does not need to be reserved.
2178 *
2179 * Usually this function is called from f_op->mmap() handler
2180 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2181 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2182 * function from other places, for example from page-fault handler.
2183 */
2184int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2185 struct page *page)
2186{
2187 if (addr < vma->vm_start || addr >= vma->vm_end)
2188 return -EFAULT;
2189 if (!page_count(page))
2190 return -EINVAL;
2191 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2192 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2193 BUG_ON(vma->vm_flags & VM_PFNMAP);
2194 vma->vm_flags |= VM_MIXEDMAP;
2195 }
2196 return insert_page(vma, addr, page, vma->vm_page_prot);
2197}
2198EXPORT_SYMBOL(vm_insert_page);
2199
2200static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2201 unsigned long pfn, pgprot_t prot)
2202{
2203 struct mm_struct *mm = vma->vm_mm;
2204 int retval;
2205 pte_t *pte, entry;
2206 spinlock_t *ptl;
2207
2208 retval = -ENOMEM;
2209 pte = get_locked_pte(mm, addr, &ptl);
2210 if (!pte)
2211 goto out;
2212 retval = -EBUSY;
2213 if (!pte_none(*pte))
2214 goto out_unlock;
2215
2216 /* Ok, finally just insert the thing.. */
2217 entry = pte_mkspecial(pfn_pte(pfn, prot));
2218 set_pte_at(mm, addr, pte, entry);
2219 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2220
2221 retval = 0;
2222out_unlock:
2223 pte_unmap_unlock(pte, ptl);
2224out:
2225 return retval;
2226}
2227
2228/**
2229 * vm_insert_pfn - insert single pfn into user vma
2230 * @vma: user vma to map to
2231 * @addr: target user address of this page
2232 * @pfn: source kernel pfn
2233 *
2234 * Similar to vm_insert_page, this allows drivers to insert individual pages
2235 * they've allocated into a user vma. Same comments apply.
2236 *
2237 * This function should only be called from a vm_ops->fault handler, and
2238 * in that case the handler should return NULL.
2239 *
2240 * vma cannot be a COW mapping.
2241 *
2242 * As this is called only for pages that do not currently exist, we
2243 * do not need to flush old virtual caches or the TLB.
2244 */
2245int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2246 unsigned long pfn)
2247{
2248 int ret;
2249 pgprot_t pgprot = vma->vm_page_prot;
2250 /*
2251 * Technically, architectures with pte_special can avoid all these
2252 * restrictions (same for remap_pfn_range). However we would like
2253 * consistency in testing and feature parity among all, so we should
2254 * try to keep these invariants in place for everybody.
2255 */
2256 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2257 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2258 (VM_PFNMAP|VM_MIXEDMAP));
2259 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2260 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2261
2262 if (addr < vma->vm_start || addr >= vma->vm_end)
2263 return -EFAULT;
2264 if (track_pfn_insert(vma, &pgprot, pfn))
2265 return -EINVAL;
2266
2267 ret = insert_pfn(vma, addr, pfn, pgprot);
2268
2269 return ret;
2270}
2271EXPORT_SYMBOL(vm_insert_pfn);
2272
2273int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2274 unsigned long pfn)
2275{
2276 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2277
2278 if (addr < vma->vm_start || addr >= vma->vm_end)
2279 return -EFAULT;
2280
2281 /*
2282 * If we don't have pte special, then we have to use the pfn_valid()
2283 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2284 * refcount the page if pfn_valid is true (hence insert_page rather
2285 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2286 * without pte special, it would there be refcounted as a normal page.
2287 */
2288 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2289 struct page *page;
2290
2291 page = pfn_to_page(pfn);
2292 return insert_page(vma, addr, page, vma->vm_page_prot);
2293 }
2294 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2295}
2296EXPORT_SYMBOL(vm_insert_mixed);
2297
2298/*
2299 * maps a range of physical memory into the requested pages. the old
2300 * mappings are removed. any references to nonexistent pages results
2301 * in null mappings (currently treated as "copy-on-access")
2302 */
2303static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2304 unsigned long addr, unsigned long end,
2305 unsigned long pfn, pgprot_t prot)
2306{
2307 pte_t *pte;
2308 spinlock_t *ptl;
2309
2310 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2311 if (!pte)
2312 return -ENOMEM;
2313 arch_enter_lazy_mmu_mode();
2314 do {
2315 BUG_ON(!pte_none(*pte));
2316 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2317 pfn++;
2318 } while (pte++, addr += PAGE_SIZE, addr != end);
2319 arch_leave_lazy_mmu_mode();
2320 pte_unmap_unlock(pte - 1, ptl);
2321 return 0;
2322}
2323
2324static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2325 unsigned long addr, unsigned long end,
2326 unsigned long pfn, pgprot_t prot)
2327{
2328 pmd_t *pmd;
2329 unsigned long next;
2330
2331 pfn -= addr >> PAGE_SHIFT;
2332 pmd = pmd_alloc(mm, pud, addr);
2333 if (!pmd)
2334 return -ENOMEM;
2335 VM_BUG_ON(pmd_trans_huge(*pmd));
2336 do {
2337 next = pmd_addr_end(addr, end);
2338 if (remap_pte_range(mm, pmd, addr, next,
2339 pfn + (addr >> PAGE_SHIFT), prot))
2340 return -ENOMEM;
2341 } while (pmd++, addr = next, addr != end);
2342 return 0;
2343}
2344
2345static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2346 unsigned long addr, unsigned long end,
2347 unsigned long pfn, pgprot_t prot)
2348{
2349 pud_t *pud;
2350 unsigned long next;
2351
2352 pfn -= addr >> PAGE_SHIFT;
2353 pud = pud_alloc(mm, pgd, addr);
2354 if (!pud)
2355 return -ENOMEM;
2356 do {
2357 next = pud_addr_end(addr, end);
2358 if (remap_pmd_range(mm, pud, addr, next,
2359 pfn + (addr >> PAGE_SHIFT), prot))
2360 return -ENOMEM;
2361 } while (pud++, addr = next, addr != end);
2362 return 0;
2363}
2364
2365/**
2366 * remap_pfn_range - remap kernel memory to userspace
2367 * @vma: user vma to map to
2368 * @addr: target user address to start at
2369 * @pfn: physical address of kernel memory
2370 * @size: size of map area
2371 * @prot: page protection flags for this mapping
2372 *
2373 * Note: this is only safe if the mm semaphore is held when called.
2374 */
2375int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2376 unsigned long pfn, unsigned long size, pgprot_t prot)
2377{
2378 pgd_t *pgd;
2379 unsigned long next;
2380 unsigned long end = addr + PAGE_ALIGN(size);
2381 struct mm_struct *mm = vma->vm_mm;
2382 int err;
2383
2384 /*
2385 * Physically remapped pages are special. Tell the
2386 * rest of the world about it:
2387 * VM_IO tells people not to look at these pages
2388 * (accesses can have side effects).
2389 * VM_PFNMAP tells the core MM that the base pages are just
2390 * raw PFN mappings, and do not have a "struct page" associated
2391 * with them.
2392 * VM_DONTEXPAND
2393 * Disable vma merging and expanding with mremap().
2394 * VM_DONTDUMP
2395 * Omit vma from core dump, even when VM_IO turned off.
2396 *
2397 * There's a horrible special case to handle copy-on-write
2398 * behaviour that some programs depend on. We mark the "original"
2399 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2400 * See vm_normal_page() for details.
2401 */
2402 if (is_cow_mapping(vma->vm_flags)) {
2403 if (addr != vma->vm_start || end != vma->vm_end)
2404 return -EINVAL;
2405 vma->vm_pgoff = pfn;
2406 }
2407
2408 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2409 if (err)
2410 return -EINVAL;
2411
2412 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2413
2414 BUG_ON(addr >= end);
2415 pfn -= addr >> PAGE_SHIFT;
2416 pgd = pgd_offset(mm, addr);
2417 flush_cache_range(vma, addr, end);
2418 do {
2419 next = pgd_addr_end(addr, end);
2420 err = remap_pud_range(mm, pgd, addr, next,
2421 pfn + (addr >> PAGE_SHIFT), prot);
2422 if (err)
2423 break;
2424 } while (pgd++, addr = next, addr != end);
2425
2426 if (err)
2427 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2428
2429 return err;
2430}
2431EXPORT_SYMBOL(remap_pfn_range);
2432
2433/**
2434 * vm_iomap_memory - remap memory to userspace
2435 * @vma: user vma to map to
2436 * @start: start of area
2437 * @len: size of area
2438 *
2439 * This is a simplified io_remap_pfn_range() for common driver use. The
2440 * driver just needs to give us the physical memory range to be mapped,
2441 * we'll figure out the rest from the vma information.
2442 *
2443 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2444 * whatever write-combining details or similar.
2445 */
2446int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2447{
2448 unsigned long vm_len, pfn, pages;
2449
2450 /* Check that the physical memory area passed in looks valid */
2451 if (start + len < start)
2452 return -EINVAL;
2453 /*
2454 * You *really* shouldn't map things that aren't page-aligned,
2455 * but we've historically allowed it because IO memory might
2456 * just have smaller alignment.
2457 */
2458 len += start & ~PAGE_MASK;
2459 pfn = start >> PAGE_SHIFT;
2460 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2461 if (pfn + pages < pfn)
2462 return -EINVAL;
2463
2464 /* We start the mapping 'vm_pgoff' pages into the area */
2465 if (vma->vm_pgoff > pages)
2466 return -EINVAL;
2467 pfn += vma->vm_pgoff;
2468 pages -= vma->vm_pgoff;
2469
2470 /* Can we fit all of the mapping? */
2471 vm_len = vma->vm_end - vma->vm_start;
2472 if (vm_len >> PAGE_SHIFT > pages)
2473 return -EINVAL;
2474
2475 /* Ok, let it rip */
2476 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2477}
2478EXPORT_SYMBOL(vm_iomap_memory);
2479
2480static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2481 unsigned long addr, unsigned long end,
2482 pte_fn_t fn, void *data)
2483{
2484 pte_t *pte;
2485 int err;
2486 pgtable_t token;
2487 spinlock_t *uninitialized_var(ptl);
2488
2489 pte = (mm == &init_mm) ?
2490 pte_alloc_kernel(pmd, addr) :
2491 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2492 if (!pte)
2493 return -ENOMEM;
2494
2495 BUG_ON(pmd_huge(*pmd));
2496
2497 arch_enter_lazy_mmu_mode();
2498
2499 token = pmd_pgtable(*pmd);
2500
2501 do {
2502 err = fn(pte++, token, addr, data);
2503 if (err)
2504 break;
2505 } while (addr += PAGE_SIZE, addr != end);
2506
2507 arch_leave_lazy_mmu_mode();
2508
2509 if (mm != &init_mm)
2510 pte_unmap_unlock(pte-1, ptl);
2511 return err;
2512}
2513
2514static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2515 unsigned long addr, unsigned long end,
2516 pte_fn_t fn, void *data)
2517{
2518 pmd_t *pmd;
2519 unsigned long next;
2520 int err;
2521
2522 BUG_ON(pud_huge(*pud));
2523
2524 pmd = pmd_alloc(mm, pud, addr);
2525 if (!pmd)
2526 return -ENOMEM;
2527 do {
2528 next = pmd_addr_end(addr, end);
2529 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2530 if (err)
2531 break;
2532 } while (pmd++, addr = next, addr != end);
2533 return err;
2534}
2535
2536static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2537 unsigned long addr, unsigned long end,
2538 pte_fn_t fn, void *data)
2539{
2540 pud_t *pud;
2541 unsigned long next;
2542 int err;
2543
2544 pud = pud_alloc(mm, pgd, addr);
2545 if (!pud)
2546 return -ENOMEM;
2547 do {
2548 next = pud_addr_end(addr, end);
2549 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2550 if (err)
2551 break;
2552 } while (pud++, addr = next, addr != end);
2553 return err;
2554}
2555
2556/*
2557 * Scan a region of virtual memory, filling in page tables as necessary
2558 * and calling a provided function on each leaf page table.
2559 */
2560int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2561 unsigned long size, pte_fn_t fn, void *data)
2562{
2563 pgd_t *pgd;
2564 unsigned long next;
2565 unsigned long end = addr + size;
2566 int err;
2567
2568 BUG_ON(addr >= end);
2569 pgd = pgd_offset(mm, addr);
2570 do {
2571 next = pgd_addr_end(addr, end);
2572 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2573 if (err)
2574 break;
2575 } while (pgd++, addr = next, addr != end);
2576
2577 return err;
2578}
2579EXPORT_SYMBOL_GPL(apply_to_page_range);
2580
2581/*
2582 * handle_pte_fault chooses page fault handler according to an entry
2583 * which was read non-atomically. Before making any commitment, on
2584 * those architectures or configurations (e.g. i386 with PAE) which
2585 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2586 * must check under lock before unmapping the pte and proceeding
2587 * (but do_wp_page is only called after already making such a check;
2588 * and do_anonymous_page can safely check later on).
2589 */
2590static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2591 pte_t *page_table, pte_t orig_pte)
2592{
2593 int same = 1;
2594#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2595 if (sizeof(pte_t) > sizeof(unsigned long)) {
2596 spinlock_t *ptl = pte_lockptr(mm, pmd);
2597 spin_lock(ptl);
2598 same = pte_same(*page_table, orig_pte);
2599 spin_unlock(ptl);
2600 }
2601#endif
2602 pte_unmap(page_table);
2603 return same;
2604}
2605
2606static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2607{
2608 debug_dma_assert_idle(src);
2609
2610 /*
2611 * If the source page was a PFN mapping, we don't have
2612 * a "struct page" for it. We do a best-effort copy by
2613 * just copying from the original user address. If that
2614 * fails, we just zero-fill it. Live with it.
2615 */
2616 if (unlikely(!src)) {
2617 void *kaddr = kmap_atomic(dst);
2618 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2619
2620 /*
2621 * This really shouldn't fail, because the page is there
2622 * in the page tables. But it might just be unreadable,
2623 * in which case we just give up and fill the result with
2624 * zeroes.
2625 */
2626 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2627 clear_page(kaddr);
2628 kunmap_atomic(kaddr);
2629 flush_dcache_page(dst);
2630 } else
2631 copy_user_highpage(dst, src, va, vma);
2632}
2633
2634/*
2635 * Notify the address space that the page is about to become writable so that
2636 * it can prohibit this or wait for the page to get into an appropriate state.
2637 *
2638 * We do this without the lock held, so that it can sleep if it needs to.
2639 */
2640static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2641 unsigned long address)
2642{
2643 struct vm_fault vmf;
2644 int ret;
2645
2646 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2647 vmf.pgoff = page->index;
2648 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2649 vmf.page = page;
2650
2651 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2652 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2653 return ret;
2654 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2655 lock_page(page);
2656 if (!page->mapping) {
2657 unlock_page(page);
2658 return 0; /* retry */
2659 }
2660 ret |= VM_FAULT_LOCKED;
2661 } else
2662 VM_BUG_ON_PAGE(!PageLocked(page), page);
2663 return ret;
2664}
2665
2666/*
2667 * This routine handles present pages, when users try to write
2668 * to a shared page. It is done by copying the page to a new address
2669 * and decrementing the shared-page counter for the old page.
2670 *
2671 * Note that this routine assumes that the protection checks have been
2672 * done by the caller (the low-level page fault routine in most cases).
2673 * Thus we can safely just mark it writable once we've done any necessary
2674 * COW.
2675 *
2676 * We also mark the page dirty at this point even though the page will
2677 * change only once the write actually happens. This avoids a few races,
2678 * and potentially makes it more efficient.
2679 *
2680 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2681 * but allow concurrent faults), with pte both mapped and locked.
2682 * We return with mmap_sem still held, but pte unmapped and unlocked.
2683 */
2684static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2685 unsigned long address, pte_t *page_table, pmd_t *pmd,
2686 spinlock_t *ptl, pte_t orig_pte)
2687 __releases(ptl)
2688{
2689 struct page *old_page, *new_page = NULL;
2690 pte_t entry;
2691 int ret = 0;
2692 int page_mkwrite = 0;
2693 struct page *dirty_page = NULL;
2694 unsigned long mmun_start = 0; /* For mmu_notifiers */
2695 unsigned long mmun_end = 0; /* For mmu_notifiers */
2696
2697 old_page = vm_normal_page(vma, address, orig_pte);
2698 if (!old_page) {
2699 /*
2700 * VM_MIXEDMAP !pfn_valid() case
2701 *
2702 * We should not cow pages in a shared writeable mapping.
2703 * Just mark the pages writable as we can't do any dirty
2704 * accounting on raw pfn maps.
2705 */
2706 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2707 (VM_WRITE|VM_SHARED))
2708 goto reuse;
2709 goto gotten;
2710 }
2711
2712 /*
2713 * Take out anonymous pages first, anonymous shared vmas are
2714 * not dirty accountable.
2715 */
2716 if (PageAnon(old_page) && !PageKsm(old_page)) {
2717 if (!trylock_page(old_page)) {
2718 page_cache_get(old_page);
2719 pte_unmap_unlock(page_table, ptl);
2720 lock_page(old_page);
2721 page_table = pte_offset_map_lock(mm, pmd, address,
2722 &ptl);
2723 if (!pte_same(*page_table, orig_pte)) {
2724 unlock_page(old_page);
2725 goto unlock;
2726 }
2727 page_cache_release(old_page);
2728 }
2729 if (reuse_swap_page(old_page)) {
2730 /*
2731 * The page is all ours. Move it to our anon_vma so
2732 * the rmap code will not search our parent or siblings.
2733 * Protected against the rmap code by the page lock.
2734 */
2735 page_move_anon_rmap(old_page, vma, address);
2736 unlock_page(old_page);
2737 goto reuse;
2738 }
2739 unlock_page(old_page);
2740 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2741 (VM_WRITE|VM_SHARED))) {
2742 /*
2743 * Only catch write-faults on shared writable pages,
2744 * read-only shared pages can get COWed by
2745 * get_user_pages(.write=1, .force=1).
2746 */
2747 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2748 int tmp;
2749 page_cache_get(old_page);
2750 pte_unmap_unlock(page_table, ptl);
2751 tmp = do_page_mkwrite(vma, old_page, address);
2752 if (unlikely(!tmp || (tmp &
2753 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2754 page_cache_release(old_page);
2755 return tmp;
2756 }
2757 /*
2758 * Since we dropped the lock we need to revalidate
2759 * the PTE as someone else may have changed it. If
2760 * they did, we just return, as we can count on the
2761 * MMU to tell us if they didn't also make it writable.
2762 */
2763 page_table = pte_offset_map_lock(mm, pmd, address,
2764 &ptl);
2765 if (!pte_same(*page_table, orig_pte)) {
2766 unlock_page(old_page);
2767 goto unlock;
2768 }
2769
2770 page_mkwrite = 1;
2771 }
2772 dirty_page = old_page;
2773 get_page(dirty_page);
2774
2775reuse:
2776 /*
2777 * Clear the pages cpupid information as the existing
2778 * information potentially belongs to a now completely
2779 * unrelated process.
2780 */
2781 if (old_page)
2782 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2783
2784 flush_cache_page(vma, address, pte_pfn(orig_pte));
2785 entry = pte_mkyoung(orig_pte);
2786 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2787 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2788 update_mmu_cache(vma, address, page_table);
2789 pte_unmap_unlock(page_table, ptl);
2790 ret |= VM_FAULT_WRITE;
2791
2792 if (!dirty_page)
2793 return ret;
2794
2795 /*
2796 * Yes, Virginia, this is actually required to prevent a race
2797 * with clear_page_dirty_for_io() from clearing the page dirty
2798 * bit after it clear all dirty ptes, but before a racing
2799 * do_wp_page installs a dirty pte.
2800 *
2801 * do_shared_fault is protected similarly.
2802 */
2803 if (!page_mkwrite) {
2804 wait_on_page_locked(dirty_page);
2805 set_page_dirty_balance(dirty_page);
2806 /* file_update_time outside page_lock */
2807 if (vma->vm_file)
2808 file_update_time(vma->vm_file);
2809 }
2810 put_page(dirty_page);
2811 if (page_mkwrite) {
2812 struct address_space *mapping = dirty_page->mapping;
2813
2814 set_page_dirty(dirty_page);
2815 unlock_page(dirty_page);
2816 page_cache_release(dirty_page);
2817 if (mapping) {
2818 /*
2819 * Some device drivers do not set page.mapping
2820 * but still dirty their pages
2821 */
2822 balance_dirty_pages_ratelimited(mapping);
2823 }
2824 }
2825
2826 return ret;
2827 }
2828
2829 /*
2830 * Ok, we need to copy. Oh, well..
2831 */
2832 page_cache_get(old_page);
2833gotten:
2834 pte_unmap_unlock(page_table, ptl);
2835
2836 if (unlikely(anon_vma_prepare(vma)))
2837 goto oom;
2838
2839 if (is_zero_pfn(pte_pfn(orig_pte))) {
2840 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2841 if (!new_page)
2842 goto oom;
2843 } else {
2844 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2845 if (!new_page)
2846 goto oom;
2847 cow_user_page(new_page, old_page, address, vma);
2848 }
2849 __SetPageUptodate(new_page);
2850
2851 if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))
2852 goto oom_free_new;
2853
2854 mmun_start = address & PAGE_MASK;
2855 mmun_end = mmun_start + PAGE_SIZE;
2856 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2857
2858 /*
2859 * Re-check the pte - we dropped the lock
2860 */
2861 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2862 if (likely(pte_same(*page_table, orig_pte))) {
2863 if (old_page) {
2864 if (!PageAnon(old_page)) {
2865 dec_mm_counter_fast(mm, MM_FILEPAGES);
2866 inc_mm_counter_fast(mm, MM_ANONPAGES);
2867 }
2868 } else
2869 inc_mm_counter_fast(mm, MM_ANONPAGES);
2870 flush_cache_page(vma, address, pte_pfn(orig_pte));
2871 entry = mk_pte(new_page, vma->vm_page_prot);
2872 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2873 /*
2874 * Clear the pte entry and flush it first, before updating the
2875 * pte with the new entry. This will avoid a race condition
2876 * seen in the presence of one thread doing SMC and another
2877 * thread doing COW.
2878 */
2879 ptep_clear_flush(vma, address, page_table);
2880 page_add_new_anon_rmap(new_page, vma, address);
2881 /*
2882 * We call the notify macro here because, when using secondary
2883 * mmu page tables (such as kvm shadow page tables), we want the
2884 * new page to be mapped directly into the secondary page table.
2885 */
2886 set_pte_at_notify(mm, address, page_table, entry);
2887 update_mmu_cache(vma, address, page_table);
2888 if (old_page) {
2889 /*
2890 * Only after switching the pte to the new page may
2891 * we remove the mapcount here. Otherwise another
2892 * process may come and find the rmap count decremented
2893 * before the pte is switched to the new page, and
2894 * "reuse" the old page writing into it while our pte
2895 * here still points into it and can be read by other
2896 * threads.
2897 *
2898 * The critical issue is to order this
2899 * page_remove_rmap with the ptp_clear_flush above.
2900 * Those stores are ordered by (if nothing else,)
2901 * the barrier present in the atomic_add_negative
2902 * in page_remove_rmap.
2903 *
2904 * Then the TLB flush in ptep_clear_flush ensures that
2905 * no process can access the old page before the
2906 * decremented mapcount is visible. And the old page
2907 * cannot be reused until after the decremented
2908 * mapcount is visible. So transitively, TLBs to
2909 * old page will be flushed before it can be reused.
2910 */
2911 page_remove_rmap(old_page);
2912 }
2913
2914 /* Free the old page.. */
2915 new_page = old_page;
2916 ret |= VM_FAULT_WRITE;
2917 } else
2918 mem_cgroup_uncharge_page(new_page);
2919
2920 if (new_page)
2921 page_cache_release(new_page);
2922unlock:
2923 pte_unmap_unlock(page_table, ptl);
2924 if (mmun_end > mmun_start)
2925 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2926 if (old_page) {
2927 /*
2928 * Don't let another task, with possibly unlocked vma,
2929 * keep the mlocked page.
2930 */
2931 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2932 lock_page(old_page); /* LRU manipulation */
2933 munlock_vma_page(old_page);
2934 unlock_page(old_page);
2935 }
2936 page_cache_release(old_page);
2937 }
2938 return ret;
2939oom_free_new:
2940 page_cache_release(new_page);
2941oom:
2942 if (old_page)
2943 page_cache_release(old_page);
2944 return VM_FAULT_OOM;
2945}
2946
2947static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2948 unsigned long start_addr, unsigned long end_addr,
2949 struct zap_details *details)
2950{
2951 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2952}
2953
2954static inline void unmap_mapping_range_tree(struct rb_root *root,
2955 struct zap_details *details)
2956{
2957 struct vm_area_struct *vma;
2958 pgoff_t vba, vea, zba, zea;
2959
2960 vma_interval_tree_foreach(vma, root,
2961 details->first_index, details->last_index) {
2962
2963 vba = vma->vm_pgoff;
2964 vea = vba + vma_pages(vma) - 1;
2965 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2966 zba = details->first_index;
2967 if (zba < vba)
2968 zba = vba;
2969 zea = details->last_index;
2970 if (zea > vea)
2971 zea = vea;
2972
2973 unmap_mapping_range_vma(vma,
2974 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2975 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2976 details);
2977 }
2978}
2979
2980static inline void unmap_mapping_range_list(struct list_head *head,
2981 struct zap_details *details)
2982{
2983 struct vm_area_struct *vma;
2984
2985 /*
2986 * In nonlinear VMAs there is no correspondence between virtual address
2987 * offset and file offset. So we must perform an exhaustive search
2988 * across *all* the pages in each nonlinear VMA, not just the pages
2989 * whose virtual address lies outside the file truncation point.
2990 */
2991 list_for_each_entry(vma, head, shared.nonlinear) {
2992 details->nonlinear_vma = vma;
2993 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2994 }
2995}
2996
2997/**
2998 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2999 * @mapping: the address space containing mmaps to be unmapped.
3000 * @holebegin: byte in first page to unmap, relative to the start of
3001 * the underlying file. This will be rounded down to a PAGE_SIZE
3002 * boundary. Note that this is different from truncate_pagecache(), which
3003 * must keep the partial page. In contrast, we must get rid of
3004 * partial pages.
3005 * @holelen: size of prospective hole in bytes. This will be rounded
3006 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3007 * end of the file.
3008 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3009 * but 0 when invalidating pagecache, don't throw away private data.
3010 */
3011void unmap_mapping_range(struct address_space *mapping,
3012 loff_t const holebegin, loff_t const holelen, int even_cows)
3013{
3014 struct zap_details details;
3015 pgoff_t hba = holebegin >> PAGE_SHIFT;
3016 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3017
3018 /* Check for overflow. */
3019 if (sizeof(holelen) > sizeof(hlen)) {
3020 long long holeend =
3021 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3022 if (holeend & ~(long long)ULONG_MAX)
3023 hlen = ULONG_MAX - hba + 1;
3024 }
3025
3026 details.check_mapping = even_cows? NULL: mapping;
3027 details.nonlinear_vma = NULL;
3028 details.first_index = hba;
3029 details.last_index = hba + hlen - 1;
3030 if (details.last_index < details.first_index)
3031 details.last_index = ULONG_MAX;
3032
3033
3034 mutex_lock(&mapping->i_mmap_mutex);
3035 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
3036 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3037 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
3038 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
3039 mutex_unlock(&mapping->i_mmap_mutex);
3040}
3041EXPORT_SYMBOL(unmap_mapping_range);
3042
3043/*
3044 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3045 * but allow concurrent faults), and pte mapped but not yet locked.
3046 * We return with mmap_sem still held, but pte unmapped and unlocked.
3047 */
3048static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3049 unsigned long address, pte_t *page_table, pmd_t *pmd,
3050 unsigned int flags, pte_t orig_pte)
3051{
3052 spinlock_t *ptl;
3053 struct page *page, *swapcache;
3054 swp_entry_t entry;
3055 pte_t pte;
3056 int locked;
3057 struct mem_cgroup *ptr;
3058 int exclusive = 0;
3059 int ret = 0;
3060
3061 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3062 goto out;
3063
3064 entry = pte_to_swp_entry(orig_pte);
3065 if (unlikely(non_swap_entry(entry))) {
3066 if (is_migration_entry(entry)) {
3067 migration_entry_wait(mm, pmd, address);
3068 } else if (is_hwpoison_entry(entry)) {
3069 ret = VM_FAULT_HWPOISON;
3070 } else {
3071 print_bad_pte(vma, address, orig_pte, NULL);
3072 ret = VM_FAULT_SIGBUS;
3073 }
3074 goto out;
3075 }
3076 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3077 page = lookup_swap_cache(entry);
3078 if (!page) {
3079 page = swapin_readahead(entry,
3080 GFP_HIGHUSER_MOVABLE, vma, address);
3081 if (!page) {
3082 /*
3083 * Back out if somebody else faulted in this pte
3084 * while we released the pte lock.
3085 */
3086 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3087 if (likely(pte_same(*page_table, orig_pte)))
3088 ret = VM_FAULT_OOM;
3089 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3090 goto unlock;
3091 }
3092
3093 /* Had to read the page from swap area: Major fault */
3094 ret = VM_FAULT_MAJOR;
3095 count_vm_event(PGMAJFAULT);
3096 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3097 } else if (PageHWPoison(page)) {
3098 /*
3099 * hwpoisoned dirty swapcache pages are kept for killing
3100 * owner processes (which may be unknown at hwpoison time)
3101 */
3102 ret = VM_FAULT_HWPOISON;
3103 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3104 swapcache = page;
3105 goto out_release;
3106 }
3107
3108 swapcache = page;
3109 locked = lock_page_or_retry(page, mm, flags);
3110
3111 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3112 if (!locked) {
3113 ret |= VM_FAULT_RETRY;
3114 goto out_release;
3115 }
3116
3117 /*
3118 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3119 * release the swapcache from under us. The page pin, and pte_same
3120 * test below, are not enough to exclude that. Even if it is still
3121 * swapcache, we need to check that the page's swap has not changed.
3122 */
3123 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3124 goto out_page;
3125
3126 page = ksm_might_need_to_copy(page, vma, address);
3127 if (unlikely(!page)) {
3128 ret = VM_FAULT_OOM;
3129 page = swapcache;
3130 goto out_page;
3131 }
3132
3133 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3134 ret = VM_FAULT_OOM;
3135 goto out_page;
3136 }
3137
3138 /*
3139 * Back out if somebody else already faulted in this pte.
3140 */
3141 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3142 if (unlikely(!pte_same(*page_table, orig_pte)))
3143 goto out_nomap;
3144
3145 if (unlikely(!PageUptodate(page))) {
3146 ret = VM_FAULT_SIGBUS;
3147 goto out_nomap;
3148 }
3149
3150 /*
3151 * The page isn't present yet, go ahead with the fault.
3152 *
3153 * Be careful about the sequence of operations here.
3154 * To get its accounting right, reuse_swap_page() must be called
3155 * while the page is counted on swap but not yet in mapcount i.e.
3156 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3157 * must be called after the swap_free(), or it will never succeed.
3158 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3159 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3160 * in page->private. In this case, a record in swap_cgroup is silently
3161 * discarded at swap_free().
3162 */
3163
3164 inc_mm_counter_fast(mm, MM_ANONPAGES);
3165 dec_mm_counter_fast(mm, MM_SWAPENTS);
3166 pte = mk_pte(page, vma->vm_page_prot);
3167 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3168 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3169 flags &= ~FAULT_FLAG_WRITE;
3170 ret |= VM_FAULT_WRITE;
3171 exclusive = 1;
3172 }
3173 flush_icache_page(vma, page);
3174 if (pte_swp_soft_dirty(orig_pte))
3175 pte = pte_mksoft_dirty(pte);
3176 set_pte_at(mm, address, page_table, pte);
3177 if (page == swapcache)
3178 do_page_add_anon_rmap(page, vma, address, exclusive);
3179 else /* ksm created a completely new copy */
3180 page_add_new_anon_rmap(page, vma, address);
3181 /* It's better to call commit-charge after rmap is established */
3182 mem_cgroup_commit_charge_swapin(page, ptr);
3183
3184 swap_free(entry);
3185 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3186 try_to_free_swap(page);
3187 unlock_page(page);
3188 if (page != swapcache) {
3189 /*
3190 * Hold the lock to avoid the swap entry to be reused
3191 * until we take the PT lock for the pte_same() check
3192 * (to avoid false positives from pte_same). For
3193 * further safety release the lock after the swap_free
3194 * so that the swap count won't change under a
3195 * parallel locked swapcache.
3196 */
3197 unlock_page(swapcache);
3198 page_cache_release(swapcache);
3199 }
3200
3201 if (flags & FAULT_FLAG_WRITE) {
3202 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3203 if (ret & VM_FAULT_ERROR)
3204 ret &= VM_FAULT_ERROR;
3205 goto out;
3206 }
3207
3208 /* No need to invalidate - it was non-present before */
3209 update_mmu_cache(vma, address, page_table);
3210unlock:
3211 pte_unmap_unlock(page_table, ptl);
3212out:
3213 return ret;
3214out_nomap:
3215 mem_cgroup_cancel_charge_swapin(ptr);
3216 pte_unmap_unlock(page_table, ptl);
3217out_page:
3218 unlock_page(page);
3219out_release:
3220 page_cache_release(page);
3221 if (page != swapcache) {
3222 unlock_page(swapcache);
3223 page_cache_release(swapcache);
3224 }
3225 return ret;
3226}
3227
3228/*
3229 * This is like a special single-page "expand_{down|up}wards()",
3230 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3231 * doesn't hit another vma.
3232 */
3233static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3234{
3235 address &= PAGE_MASK;
3236 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3237 struct vm_area_struct *prev = vma->vm_prev;
3238
3239 /*
3240 * Is there a mapping abutting this one below?
3241 *
3242 * That's only ok if it's the same stack mapping
3243 * that has gotten split..
3244 */
3245 if (prev && prev->vm_end == address)
3246 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3247
3248 expand_downwards(vma, address - PAGE_SIZE);
3249 }
3250 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3251 struct vm_area_struct *next = vma->vm_next;
3252
3253 /* As VM_GROWSDOWN but s/below/above/ */
3254 if (next && next->vm_start == address + PAGE_SIZE)
3255 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3256
3257 expand_upwards(vma, address + PAGE_SIZE);
3258 }
3259 return 0;
3260}
3261
3262/*
3263 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3264 * but allow concurrent faults), and pte mapped but not yet locked.
3265 * We return with mmap_sem still held, but pte unmapped and unlocked.
3266 */
3267static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3268 unsigned long address, pte_t *page_table, pmd_t *pmd,
3269 unsigned int flags)
3270{
3271 struct page *page;
3272 spinlock_t *ptl;
3273 pte_t entry;
3274
3275 pte_unmap(page_table);
3276
3277 /* Check if we need to add a guard page to the stack */
3278 if (check_stack_guard_page(vma, address) < 0)
3279 return VM_FAULT_SIGBUS;
3280
3281 /* Use the zero-page for reads */
3282 if (!(flags & FAULT_FLAG_WRITE)) {
3283 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3284 vma->vm_page_prot));
3285 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3286 if (!pte_none(*page_table))
3287 goto unlock;
3288 goto setpte;
3289 }
3290
3291 /* Allocate our own private page. */
3292 if (unlikely(anon_vma_prepare(vma)))
3293 goto oom;
3294 page = alloc_zeroed_user_highpage_movable(vma, address);
3295 if (!page)
3296 goto oom;
3297 /*
3298 * The memory barrier inside __SetPageUptodate makes sure that
3299 * preceeding stores to the page contents become visible before
3300 * the set_pte_at() write.
3301 */
3302 __SetPageUptodate(page);
3303
3304 if (mem_cgroup_charge_anon(page, mm, GFP_KERNEL))
3305 goto oom_free_page;
3306
3307 entry = mk_pte(page, vma->vm_page_prot);
3308 if (vma->vm_flags & VM_WRITE)
3309 entry = pte_mkwrite(pte_mkdirty(entry));
3310
3311 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3312 if (!pte_none(*page_table))
3313 goto release;
3314
3315 inc_mm_counter_fast(mm, MM_ANONPAGES);
3316 page_add_new_anon_rmap(page, vma, address);
3317setpte:
3318 set_pte_at(mm, address, page_table, entry);
3319
3320 /* No need to invalidate - it was non-present before */
3321 update_mmu_cache(vma, address, page_table);
3322unlock:
3323 pte_unmap_unlock(page_table, ptl);
3324 return 0;
3325release:
3326 mem_cgroup_uncharge_page(page);
3327 page_cache_release(page);
3328 goto unlock;
3329oom_free_page:
3330 page_cache_release(page);
3331oom:
3332 return VM_FAULT_OOM;
3333}
3334
3335static int __do_fault(struct vm_area_struct *vma, unsigned long address,
3336 pgoff_t pgoff, unsigned int flags, struct page **page)
3337{
3338 struct vm_fault vmf;
3339 int ret;
3340
3341 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3342 vmf.pgoff = pgoff;
3343 vmf.flags = flags;
3344 vmf.page = NULL;
3345
3346 ret = vma->vm_ops->fault(vma, &vmf);
3347 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3348 return ret;
3349
3350 if (unlikely(PageHWPoison(vmf.page))) {
3351 if (ret & VM_FAULT_LOCKED)
3352 unlock_page(vmf.page);
3353 page_cache_release(vmf.page);
3354 return VM_FAULT_HWPOISON;
3355 }
3356
3357 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3358 lock_page(vmf.page);
3359 else
3360 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
3361
3362 *page = vmf.page;
3363 return ret;
3364}
3365
3366/**
3367 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
3368 *
3369 * @vma: virtual memory area
3370 * @address: user virtual address
3371 * @page: page to map
3372 * @pte: pointer to target page table entry
3373 * @write: true, if new entry is writable
3374 * @anon: true, if it's anonymous page
3375 *
3376 * Caller must hold page table lock relevant for @pte.
3377 *
3378 * Target users are page handler itself and implementations of
3379 * vm_ops->map_pages.
3380 */
3381void do_set_pte(struct vm_area_struct *vma, unsigned long address,
3382 struct page *page, pte_t *pte, bool write, bool anon)
3383{
3384 pte_t entry;
3385
3386 flush_icache_page(vma, page);
3387 entry = mk_pte(page, vma->vm_page_prot);
3388 if (write)
3389 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3390 else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
3391 pte_mksoft_dirty(entry);
3392 if (anon) {
3393 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3394 page_add_new_anon_rmap(page, vma, address);
3395 } else {
3396 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
3397 page_add_file_rmap(page);
3398 }
3399 set_pte_at(vma->vm_mm, address, pte, entry);
3400
3401 /* no need to invalidate: a not-present page won't be cached */
3402 update_mmu_cache(vma, address, pte);
3403}
3404
3405#define FAULT_AROUND_ORDER 4
3406
3407#ifdef CONFIG_DEBUG_FS
3408static unsigned int fault_around_order = FAULT_AROUND_ORDER;
3409
3410static int fault_around_order_get(void *data, u64 *val)
3411{
3412 *val = fault_around_order;
3413 return 0;
3414}
3415
3416static int fault_around_order_set(void *data, u64 val)
3417{
3418 BUILD_BUG_ON((1UL << FAULT_AROUND_ORDER) > PTRS_PER_PTE);
3419 if (1UL << val > PTRS_PER_PTE)
3420 return -EINVAL;
3421 fault_around_order = val;
3422 return 0;
3423}
3424DEFINE_SIMPLE_ATTRIBUTE(fault_around_order_fops,
3425 fault_around_order_get, fault_around_order_set, "%llu\n");
3426
3427static int __init fault_around_debugfs(void)
3428{
3429 void *ret;
3430
3431 ret = debugfs_create_file("fault_around_order", 0644, NULL, NULL,
3432 &fault_around_order_fops);
3433 if (!ret)
3434 pr_warn("Failed to create fault_around_order in debugfs");
3435 return 0;
3436}
3437late_initcall(fault_around_debugfs);
3438
3439static inline unsigned long fault_around_pages(void)
3440{
3441 return 1UL << fault_around_order;
3442}
3443
3444static inline unsigned long fault_around_mask(void)
3445{
3446 return ~((1UL << (PAGE_SHIFT + fault_around_order)) - 1);
3447}
3448#else
3449static inline unsigned long fault_around_pages(void)
3450{
3451 unsigned long nr_pages;
3452
3453 nr_pages = 1UL << FAULT_AROUND_ORDER;
3454 BUILD_BUG_ON(nr_pages > PTRS_PER_PTE);
3455 return nr_pages;
3456}
3457
3458static inline unsigned long fault_around_mask(void)
3459{
3460 return ~((1UL << (PAGE_SHIFT + FAULT_AROUND_ORDER)) - 1);
3461}
3462#endif
3463
3464static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
3465 pte_t *pte, pgoff_t pgoff, unsigned int flags)
3466{
3467 unsigned long start_addr;
3468 pgoff_t max_pgoff;
3469 struct vm_fault vmf;
3470 int off;
3471
3472 start_addr = max(address & fault_around_mask(), vma->vm_start);
3473 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3474 pte -= off;
3475 pgoff -= off;
3476
3477 /*
3478 * max_pgoff is either end of page table or end of vma
3479 * or fault_around_pages() from pgoff, depending what is neast.
3480 */
3481 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3482 PTRS_PER_PTE - 1;
3483 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
3484 pgoff + fault_around_pages() - 1);
3485
3486 /* Check if it makes any sense to call ->map_pages */
3487 while (!pte_none(*pte)) {
3488 if (++pgoff > max_pgoff)
3489 return;
3490 start_addr += PAGE_SIZE;
3491 if (start_addr >= vma->vm_end)
3492 return;
3493 pte++;
3494 }
3495
3496 vmf.virtual_address = (void __user *) start_addr;
3497 vmf.pte = pte;
3498 vmf.pgoff = pgoff;
3499 vmf.max_pgoff = max_pgoff;
3500 vmf.flags = flags;
3501 vma->vm_ops->map_pages(vma, &vmf);
3502}
3503
3504static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3505 unsigned long address, pmd_t *pmd,
3506 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3507{
3508 struct page *fault_page;
3509 spinlock_t *ptl;
3510 pte_t *pte;
3511 int ret = 0;
3512
3513 /*
3514 * Let's call ->map_pages() first and use ->fault() as fallback
3515 * if page by the offset is not ready to be mapped (cold cache or
3516 * something).
3517 */
3518 if (vma->vm_ops->map_pages) {
3519 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3520 do_fault_around(vma, address, pte, pgoff, flags);
3521 if (!pte_same(*pte, orig_pte))
3522 goto unlock_out;
3523 pte_unmap_unlock(pte, ptl);
3524 }
3525
3526 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3527 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3528 return ret;
3529
3530 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3531 if (unlikely(!pte_same(*pte, orig_pte))) {
3532 pte_unmap_unlock(pte, ptl);
3533 unlock_page(fault_page);
3534 page_cache_release(fault_page);
3535 return ret;
3536 }
3537 do_set_pte(vma, address, fault_page, pte, false, false);
3538 unlock_page(fault_page);
3539unlock_out:
3540 pte_unmap_unlock(pte, ptl);
3541 return ret;
3542}
3543
3544static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3545 unsigned long address, pmd_t *pmd,
3546 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3547{
3548 struct page *fault_page, *new_page;
3549 spinlock_t *ptl;
3550 pte_t *pte;
3551 int ret;
3552
3553 if (unlikely(anon_vma_prepare(vma)))
3554 return VM_FAULT_OOM;
3555
3556 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3557 if (!new_page)
3558 return VM_FAULT_OOM;
3559
3560 if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)) {
3561 page_cache_release(new_page);
3562 return VM_FAULT_OOM;
3563 }
3564
3565 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3566 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3567 goto uncharge_out;
3568
3569 copy_user_highpage(new_page, fault_page, address, vma);
3570 __SetPageUptodate(new_page);
3571
3572 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3573 if (unlikely(!pte_same(*pte, orig_pte))) {
3574 pte_unmap_unlock(pte, ptl);
3575 unlock_page(fault_page);
3576 page_cache_release(fault_page);
3577 goto uncharge_out;
3578 }
3579 do_set_pte(vma, address, new_page, pte, true, true);
3580 pte_unmap_unlock(pte, ptl);
3581 unlock_page(fault_page);
3582 page_cache_release(fault_page);
3583 return ret;
3584uncharge_out:
3585 mem_cgroup_uncharge_page(new_page);
3586 page_cache_release(new_page);
3587 return ret;
3588}
3589
3590static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3591 unsigned long address, pmd_t *pmd,
3592 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3593{
3594 struct page *fault_page;
3595 struct address_space *mapping;
3596 spinlock_t *ptl;
3597 pte_t *pte;
3598 int dirtied = 0;
3599 int ret, tmp;
3600
3601 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3602 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3603 return ret;
3604
3605 /*
3606 * Check if the backing address space wants to know that the page is
3607 * about to become writable
3608 */
3609 if (vma->vm_ops->page_mkwrite) {
3610 unlock_page(fault_page);
3611 tmp = do_page_mkwrite(vma, fault_page, address);
3612 if (unlikely(!tmp ||
3613 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3614 page_cache_release(fault_page);
3615 return tmp;
3616 }
3617 }
3618
3619 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3620 if (unlikely(!pte_same(*pte, orig_pte))) {
3621 pte_unmap_unlock(pte, ptl);
3622 unlock_page(fault_page);
3623 page_cache_release(fault_page);
3624 return ret;
3625 }
3626 do_set_pte(vma, address, fault_page, pte, true, false);
3627 pte_unmap_unlock(pte, ptl);
3628
3629 if (set_page_dirty(fault_page))
3630 dirtied = 1;
3631 mapping = fault_page->mapping;
3632 unlock_page(fault_page);
3633 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3634 /*
3635 * Some device drivers do not set page.mapping but still
3636 * dirty their pages
3637 */
3638 balance_dirty_pages_ratelimited(mapping);
3639 }
3640
3641 /* file_update_time outside page_lock */
3642 if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3643 file_update_time(vma->vm_file);
3644
3645 return ret;
3646}
3647
3648static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3649 unsigned long address, pte_t *page_table, pmd_t *pmd,
3650 unsigned int flags, pte_t orig_pte)
3651{
3652 pgoff_t pgoff = (((address & PAGE_MASK)
3653 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3654
3655 pte_unmap(page_table);
3656 if (!(flags & FAULT_FLAG_WRITE))
3657 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3658 orig_pte);
3659 if (!(vma->vm_flags & VM_SHARED))
3660 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3661 orig_pte);
3662 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3663}
3664
3665/*
3666 * Fault of a previously existing named mapping. Repopulate the pte
3667 * from the encoded file_pte if possible. This enables swappable
3668 * nonlinear vmas.
3669 *
3670 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3671 * but allow concurrent faults), and pte mapped but not yet locked.
3672 * We return with mmap_sem still held, but pte unmapped and unlocked.
3673 */
3674static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3675 unsigned long address, pte_t *page_table, pmd_t *pmd,
3676 unsigned int flags, pte_t orig_pte)
3677{
3678 pgoff_t pgoff;
3679
3680 flags |= FAULT_FLAG_NONLINEAR;
3681
3682 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3683 return 0;
3684
3685 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3686 /*
3687 * Page table corrupted: show pte and kill process.
3688 */
3689 print_bad_pte(vma, address, orig_pte, NULL);
3690 return VM_FAULT_SIGBUS;
3691 }
3692
3693 pgoff = pte_to_pgoff(orig_pte);
3694 if (!(flags & FAULT_FLAG_WRITE))
3695 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3696 orig_pte);
3697 if (!(vma->vm_flags & VM_SHARED))
3698 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3699 orig_pte);
3700 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3701}
3702
3703static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3704 unsigned long addr, int page_nid,
3705 int *flags)
3706{
3707 get_page(page);
3708
3709 count_vm_numa_event(NUMA_HINT_FAULTS);
3710 if (page_nid == numa_node_id()) {
3711 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3712 *flags |= TNF_FAULT_LOCAL;
3713 }
3714
3715 return mpol_misplaced(page, vma, addr);
3716}
3717
3718static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3719 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3720{
3721 struct page *page = NULL;
3722 spinlock_t *ptl;
3723 int page_nid = -1;
3724 int last_cpupid;
3725 int target_nid;
3726 bool migrated = false;
3727 int flags = 0;
3728
3729 /*
3730 * The "pte" at this point cannot be used safely without
3731 * validation through pte_unmap_same(). It's of NUMA type but
3732 * the pfn may be screwed if the read is non atomic.
3733 *
3734 * ptep_modify_prot_start is not called as this is clearing
3735 * the _PAGE_NUMA bit and it is not really expected that there
3736 * would be concurrent hardware modifications to the PTE.
3737 */
3738 ptl = pte_lockptr(mm, pmd);
3739 spin_lock(ptl);
3740 if (unlikely(!pte_same(*ptep, pte))) {
3741 pte_unmap_unlock(ptep, ptl);
3742 goto out;
3743 }
3744
3745 pte = pte_mknonnuma(pte);
3746 set_pte_at(mm, addr, ptep, pte);
3747 update_mmu_cache(vma, addr, ptep);
3748
3749 page = vm_normal_page(vma, addr, pte);
3750 if (!page) {
3751 pte_unmap_unlock(ptep, ptl);
3752 return 0;
3753 }
3754 BUG_ON(is_zero_pfn(page_to_pfn(page)));
3755
3756 /*
3757 * Avoid grouping on DSO/COW pages in specific and RO pages
3758 * in general, RO pages shouldn't hurt as much anyway since
3759 * they can be in shared cache state.
3760 */
3761 if (!pte_write(pte))
3762 flags |= TNF_NO_GROUP;
3763
3764 /*
3765 * Flag if the page is shared between multiple address spaces. This
3766 * is later used when determining whether to group tasks together
3767 */
3768 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3769 flags |= TNF_SHARED;
3770
3771 last_cpupid = page_cpupid_last(page);
3772 page_nid = page_to_nid(page);
3773 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3774 pte_unmap_unlock(ptep, ptl);
3775 if (target_nid == -1) {
3776 put_page(page);
3777 goto out;
3778 }
3779
3780 /* Migrate to the requested node */
3781 migrated = migrate_misplaced_page(page, vma, target_nid);
3782 if (migrated) {
3783 page_nid = target_nid;
3784 flags |= TNF_MIGRATED;
3785 }
3786
3787out:
3788 if (page_nid != -1)
3789 task_numa_fault(last_cpupid, page_nid, 1, flags);
3790 return 0;
3791}
3792
3793/*
3794 * These routines also need to handle stuff like marking pages dirty
3795 * and/or accessed for architectures that don't do it in hardware (most
3796 * RISC architectures). The early dirtying is also good on the i386.
3797 *
3798 * There is also a hook called "update_mmu_cache()" that architectures
3799 * with external mmu caches can use to update those (ie the Sparc or
3800 * PowerPC hashed page tables that act as extended TLBs).
3801 *
3802 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3803 * but allow concurrent faults), and pte mapped but not yet locked.
3804 * We return with mmap_sem still held, but pte unmapped and unlocked.
3805 */
3806static int handle_pte_fault(struct mm_struct *mm,
3807 struct vm_area_struct *vma, unsigned long address,
3808 pte_t *pte, pmd_t *pmd, unsigned int flags)
3809{
3810 pte_t entry;
3811 spinlock_t *ptl;
3812
3813 entry = *pte;
3814 if (!pte_present(entry)) {
3815 if (pte_none(entry)) {
3816 if (vma->vm_ops) {
3817 if (likely(vma->vm_ops->fault))
3818 return do_linear_fault(mm, vma, address,
3819 pte, pmd, flags, entry);
3820 }
3821 return do_anonymous_page(mm, vma, address,
3822 pte, pmd, flags);
3823 }
3824 if (pte_file(entry))
3825 return do_nonlinear_fault(mm, vma, address,
3826 pte, pmd, flags, entry);
3827 return do_swap_page(mm, vma, address,
3828 pte, pmd, flags, entry);
3829 }
3830
3831 if (pte_numa(entry))
3832 return do_numa_page(mm, vma, address, entry, pte, pmd);
3833
3834 ptl = pte_lockptr(mm, pmd);
3835 spin_lock(ptl);
3836 if (unlikely(!pte_same(*pte, entry)))
3837 goto unlock;
3838 if (flags & FAULT_FLAG_WRITE) {
3839 if (!pte_write(entry))
3840 return do_wp_page(mm, vma, address,
3841 pte, pmd, ptl, entry);
3842 entry = pte_mkdirty(entry);
3843 }
3844 entry = pte_mkyoung(entry);
3845 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3846 update_mmu_cache(vma, address, pte);
3847 } else {
3848 /*
3849 * This is needed only for protection faults but the arch code
3850 * is not yet telling us if this is a protection fault or not.
3851 * This still avoids useless tlb flushes for .text page faults
3852 * with threads.
3853 */
3854 if (flags & FAULT_FLAG_WRITE)
3855 flush_tlb_fix_spurious_fault(vma, address);
3856 }
3857unlock:
3858 pte_unmap_unlock(pte, ptl);
3859 return 0;
3860}
3861
3862/*
3863 * By the time we get here, we already hold the mm semaphore
3864 */
3865static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3866 unsigned long address, unsigned int flags)
3867{
3868 pgd_t *pgd;
3869 pud_t *pud;
3870 pmd_t *pmd;
3871 pte_t *pte;
3872
3873 if (unlikely(is_vm_hugetlb_page(vma)))
3874 return hugetlb_fault(mm, vma, address, flags);
3875
3876 pgd = pgd_offset(mm, address);
3877 pud = pud_alloc(mm, pgd, address);
3878 if (!pud)
3879 return VM_FAULT_OOM;
3880 pmd = pmd_alloc(mm, pud, address);
3881 if (!pmd)
3882 return VM_FAULT_OOM;
3883 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3884 int ret = VM_FAULT_FALLBACK;
3885 if (!vma->vm_ops)
3886 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3887 pmd, flags);
3888 if (!(ret & VM_FAULT_FALLBACK))
3889 return ret;
3890 } else {
3891 pmd_t orig_pmd = *pmd;
3892 int ret;
3893
3894 barrier();
3895 if (pmd_trans_huge(orig_pmd)) {
3896 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3897
3898 /*
3899 * If the pmd is splitting, return and retry the
3900 * the fault. Alternative: wait until the split
3901 * is done, and goto retry.
3902 */
3903 if (pmd_trans_splitting(orig_pmd))
3904 return 0;
3905
3906 if (pmd_numa(orig_pmd))
3907 return do_huge_pmd_numa_page(mm, vma, address,
3908 orig_pmd, pmd);
3909
3910 if (dirty && !pmd_write(orig_pmd)) {
3911 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3912 orig_pmd);
3913 if (!(ret & VM_FAULT_FALLBACK))
3914 return ret;
3915 } else {
3916 huge_pmd_set_accessed(mm, vma, address, pmd,
3917 orig_pmd, dirty);
3918 return 0;
3919 }
3920 }
3921 }
3922
3923 /* THP should already have been handled */
3924 BUG_ON(pmd_numa(*pmd));
3925
3926 /*
3927 * Use __pte_alloc instead of pte_alloc_map, because we can't
3928 * run pte_offset_map on the pmd, if an huge pmd could
3929 * materialize from under us from a different thread.
3930 */
3931 if (unlikely(pmd_none(*pmd)) &&
3932 unlikely(__pte_alloc(mm, vma, pmd, address)))
3933 return VM_FAULT_OOM;
3934 /* if an huge pmd materialized from under us just retry later */
3935 if (unlikely(pmd_trans_huge(*pmd)))
3936 return 0;
3937 /*
3938 * A regular pmd is established and it can't morph into a huge pmd
3939 * from under us anymore at this point because we hold the mmap_sem
3940 * read mode and khugepaged takes it in write mode. So now it's
3941 * safe to run pte_offset_map().
3942 */
3943 pte = pte_offset_map(pmd, address);
3944
3945 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3946}
3947
3948int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3949 unsigned long address, unsigned int flags)
3950{
3951 int ret;
3952
3953 __set_current_state(TASK_RUNNING);
3954
3955 count_vm_event(PGFAULT);
3956 mem_cgroup_count_vm_event(mm, PGFAULT);
3957
3958 /* do counter updates before entering really critical section. */
3959 check_sync_rss_stat(current);
3960
3961 /*
3962 * Enable the memcg OOM handling for faults triggered in user
3963 * space. Kernel faults are handled more gracefully.
3964 */
3965 if (flags & FAULT_FLAG_USER)
3966 mem_cgroup_oom_enable();
3967
3968 ret = __handle_mm_fault(mm, vma, address, flags);
3969
3970 if (flags & FAULT_FLAG_USER) {
3971 mem_cgroup_oom_disable();
3972 /*
3973 * The task may have entered a memcg OOM situation but
3974 * if the allocation error was handled gracefully (no
3975 * VM_FAULT_OOM), there is no need to kill anything.
3976 * Just clean up the OOM state peacefully.
3977 */
3978 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3979 mem_cgroup_oom_synchronize(false);
3980 }
3981
3982 return ret;
3983}
3984
3985#ifndef __PAGETABLE_PUD_FOLDED
3986/*
3987 * Allocate page upper directory.
3988 * We've already handled the fast-path in-line.
3989 */
3990int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3991{
3992 pud_t *new = pud_alloc_one(mm, address);
3993 if (!new)
3994 return -ENOMEM;
3995
3996 smp_wmb(); /* See comment in __pte_alloc */
3997
3998 spin_lock(&mm->page_table_lock);
3999 if (pgd_present(*pgd)) /* Another has populated it */
4000 pud_free(mm, new);
4001 else
4002 pgd_populate(mm, pgd, new);
4003 spin_unlock(&mm->page_table_lock);
4004 return 0;
4005}
4006#endif /* __PAGETABLE_PUD_FOLDED */
4007
4008#ifndef __PAGETABLE_PMD_FOLDED
4009/*
4010 * Allocate page middle directory.
4011 * We've already handled the fast-path in-line.
4012 */
4013int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4014{
4015 pmd_t *new = pmd_alloc_one(mm, address);
4016 if (!new)
4017 return -ENOMEM;
4018
4019 smp_wmb(); /* See comment in __pte_alloc */
4020
4021 spin_lock(&mm->page_table_lock);
4022#ifndef __ARCH_HAS_4LEVEL_HACK
4023 if (pud_present(*pud)) /* Another has populated it */
4024 pmd_free(mm, new);
4025 else
4026 pud_populate(mm, pud, new);
4027#else
4028 if (pgd_present(*pud)) /* Another has populated it */
4029 pmd_free(mm, new);
4030 else
4031 pgd_populate(mm, pud, new);
4032#endif /* __ARCH_HAS_4LEVEL_HACK */
4033 spin_unlock(&mm->page_table_lock);
4034 return 0;
4035}
4036#endif /* __PAGETABLE_PMD_FOLDED */
4037
4038#if !defined(__HAVE_ARCH_GATE_AREA)
4039
4040#if defined(AT_SYSINFO_EHDR)
4041static struct vm_area_struct gate_vma;
4042
4043static int __init gate_vma_init(void)
4044{
4045 gate_vma.vm_mm = NULL;
4046 gate_vma.vm_start = FIXADDR_USER_START;
4047 gate_vma.vm_end = FIXADDR_USER_END;
4048 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
4049 gate_vma.vm_page_prot = __P101;
4050
4051 return 0;
4052}
4053__initcall(gate_vma_init);
4054#endif
4055
4056struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
4057{
4058#ifdef AT_SYSINFO_EHDR
4059 return &gate_vma;
4060#else
4061 return NULL;
4062#endif
4063}
4064
4065int in_gate_area_no_mm(unsigned long addr)
4066{
4067#ifdef AT_SYSINFO_EHDR
4068 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
4069 return 1;
4070#endif
4071 return 0;
4072}
4073
4074#endif /* __HAVE_ARCH_GATE_AREA */
4075
4076static int __follow_pte(struct mm_struct *mm, unsigned long address,
4077 pte_t **ptepp, spinlock_t **ptlp)
4078{
4079 pgd_t *pgd;
4080 pud_t *pud;
4081 pmd_t *pmd;
4082 pte_t *ptep;
4083
4084 pgd = pgd_offset(mm, address);
4085 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4086 goto out;
4087
4088 pud = pud_offset(pgd, address);
4089 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4090 goto out;
4091
4092 pmd = pmd_offset(pud, address);
4093 VM_BUG_ON(pmd_trans_huge(*pmd));
4094 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4095 goto out;
4096
4097 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
4098 if (pmd_huge(*pmd))
4099 goto out;
4100
4101 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4102 if (!ptep)
4103 goto out;
4104 if (!pte_present(*ptep))
4105 goto unlock;
4106 *ptepp = ptep;
4107 return 0;
4108unlock:
4109 pte_unmap_unlock(ptep, *ptlp);
4110out:
4111 return -EINVAL;
4112}
4113
4114static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4115 pte_t **ptepp, spinlock_t **ptlp)
4116{
4117 int res;
4118
4119 /* (void) is needed to make gcc happy */
4120 (void) __cond_lock(*ptlp,
4121 !(res = __follow_pte(mm, address, ptepp, ptlp)));
4122 return res;
4123}
4124
4125/**
4126 * follow_pfn - look up PFN at a user virtual address
4127 * @vma: memory mapping
4128 * @address: user virtual address
4129 * @pfn: location to store found PFN
4130 *
4131 * Only IO mappings and raw PFN mappings are allowed.
4132 *
4133 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4134 */
4135int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4136 unsigned long *pfn)
4137{
4138 int ret = -EINVAL;
4139 spinlock_t *ptl;
4140 pte_t *ptep;
4141
4142 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4143 return ret;
4144
4145 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4146 if (ret)
4147 return ret;
4148 *pfn = pte_pfn(*ptep);
4149 pte_unmap_unlock(ptep, ptl);
4150 return 0;
4151}
4152EXPORT_SYMBOL(follow_pfn);
4153
4154#ifdef CONFIG_HAVE_IOREMAP_PROT
4155int follow_phys(struct vm_area_struct *vma,
4156 unsigned long address, unsigned int flags,
4157 unsigned long *prot, resource_size_t *phys)
4158{
4159 int ret = -EINVAL;
4160 pte_t *ptep, pte;
4161 spinlock_t *ptl;
4162
4163 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4164 goto out;
4165
4166 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4167 goto out;
4168 pte = *ptep;
4169
4170 if ((flags & FOLL_WRITE) && !pte_write(pte))
4171 goto unlock;
4172
4173 *prot = pgprot_val(pte_pgprot(pte));
4174 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4175
4176 ret = 0;
4177unlock:
4178 pte_unmap_unlock(ptep, ptl);
4179out:
4180 return ret;
4181}
4182
4183int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4184 void *buf, int len, int write)
4185{
4186 resource_size_t phys_addr;
4187 unsigned long prot = 0;
4188 void __iomem *maddr;
4189 int offset = addr & (PAGE_SIZE-1);
4190
4191 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4192 return -EINVAL;
4193
4194 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4195 if (write)
4196 memcpy_toio(maddr + offset, buf, len);
4197 else
4198 memcpy_fromio(buf, maddr + offset, len);
4199 iounmap(maddr);
4200
4201 return len;
4202}
4203EXPORT_SYMBOL_GPL(generic_access_phys);
4204#endif
4205
4206/*
4207 * Access another process' address space as given in mm. If non-NULL, use the
4208 * given task for page fault accounting.
4209 */
4210static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4211 unsigned long addr, void *buf, int len, int write)
4212{
4213 struct vm_area_struct *vma;
4214 void *old_buf = buf;
4215
4216 down_read(&mm->mmap_sem);
4217 /* ignore errors, just check how much was successfully transferred */
4218 while (len) {
4219 int bytes, ret, offset;
4220 void *maddr;
4221 struct page *page = NULL;
4222
4223 ret = get_user_pages(tsk, mm, addr, 1,
4224 write, 1, &page, &vma);
4225 if (ret <= 0) {
4226 /*
4227 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4228 * we can access using slightly different code.
4229 */
4230#ifdef CONFIG_HAVE_IOREMAP_PROT
4231 vma = find_vma(mm, addr);
4232 if (!vma || vma->vm_start > addr)
4233 break;
4234 if (vma->vm_ops && vma->vm_ops->access)
4235 ret = vma->vm_ops->access(vma, addr, buf,
4236 len, write);
4237 if (ret <= 0)
4238#endif
4239 break;
4240 bytes = ret;
4241 } else {
4242 bytes = len;
4243 offset = addr & (PAGE_SIZE-1);
4244 if (bytes > PAGE_SIZE-offset)
4245 bytes = PAGE_SIZE-offset;
4246
4247 maddr = kmap(page);
4248 if (write) {
4249 copy_to_user_page(vma, page, addr,
4250 maddr + offset, buf, bytes);
4251 set_page_dirty_lock(page);
4252 } else {
4253 copy_from_user_page(vma, page, addr,
4254 buf, maddr + offset, bytes);
4255 }
4256 kunmap(page);
4257 page_cache_release(page);
4258 }
4259 len -= bytes;
4260 buf += bytes;
4261 addr += bytes;
4262 }
4263 up_read(&mm->mmap_sem);
4264
4265 return buf - old_buf;
4266}
4267
4268/**
4269 * access_remote_vm - access another process' address space
4270 * @mm: the mm_struct of the target address space
4271 * @addr: start address to access
4272 * @buf: source or destination buffer
4273 * @len: number of bytes to transfer
4274 * @write: whether the access is a write
4275 *
4276 * The caller must hold a reference on @mm.
4277 */
4278int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4279 void *buf, int len, int write)
4280{
4281 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4282}
4283
4284/*
4285 * Access another process' address space.
4286 * Source/target buffer must be kernel space,
4287 * Do not walk the page table directly, use get_user_pages
4288 */
4289int access_process_vm(struct task_struct *tsk, unsigned long addr,
4290 void *buf, int len, int write)
4291{
4292 struct mm_struct *mm;
4293 int ret;
4294
4295 mm = get_task_mm(tsk);
4296 if (!mm)
4297 return 0;
4298
4299 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4300 mmput(mm);
4301
4302 return ret;
4303}
4304
4305/*
4306 * Print the name of a VMA.
4307 */
4308void print_vma_addr(char *prefix, unsigned long ip)
4309{
4310 struct mm_struct *mm = current->mm;
4311 struct vm_area_struct *vma;
4312
4313 /*
4314 * Do not print if we are in atomic
4315 * contexts (in exception stacks, etc.):
4316 */
4317 if (preempt_count())
4318 return;
4319
4320 down_read(&mm->mmap_sem);
4321 vma = find_vma(mm, ip);
4322 if (vma && vma->vm_file) {
4323 struct file *f = vma->vm_file;
4324 char *buf = (char *)__get_free_page(GFP_KERNEL);
4325 if (buf) {
4326 char *p;
4327
4328 p = d_path(&f->f_path, buf, PAGE_SIZE);
4329 if (IS_ERR(p))
4330 p = "?";
4331 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4332 vma->vm_start,
4333 vma->vm_end - vma->vm_start);
4334 free_page((unsigned long)buf);
4335 }
4336 }
4337 up_read(&mm->mmap_sem);
4338}
4339
4340#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4341void might_fault(void)
4342{
4343 /*
4344 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4345 * holding the mmap_sem, this is safe because kernel memory doesn't
4346 * get paged out, therefore we'll never actually fault, and the
4347 * below annotations will generate false positives.
4348 */
4349 if (segment_eq(get_fs(), KERNEL_DS))
4350 return;
4351
4352 /*
4353 * it would be nicer only to annotate paths which are not under
4354 * pagefault_disable, however that requires a larger audit and
4355 * providing helpers like get_user_atomic.
4356 */
4357 if (in_atomic())
4358 return;
4359
4360 __might_sleep(__FILE__, __LINE__, 0);
4361
4362 if (current->mm)
4363 might_lock_read(¤t->mm->mmap_sem);
4364}
4365EXPORT_SYMBOL(might_fault);
4366#endif
4367
4368#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4369static void clear_gigantic_page(struct page *page,
4370 unsigned long addr,
4371 unsigned int pages_per_huge_page)
4372{
4373 int i;
4374 struct page *p = page;
4375
4376 might_sleep();
4377 for (i = 0; i < pages_per_huge_page;
4378 i++, p = mem_map_next(p, page, i)) {
4379 cond_resched();
4380 clear_user_highpage(p, addr + i * PAGE_SIZE);
4381 }
4382}
4383void clear_huge_page(struct page *page,
4384 unsigned long addr, unsigned int pages_per_huge_page)
4385{
4386 int i;
4387
4388 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4389 clear_gigantic_page(page, addr, pages_per_huge_page);
4390 return;
4391 }
4392
4393 might_sleep();
4394 for (i = 0; i < pages_per_huge_page; i++) {
4395 cond_resched();
4396 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4397 }
4398}
4399
4400static void copy_user_gigantic_page(struct page *dst, struct page *src,
4401 unsigned long addr,
4402 struct vm_area_struct *vma,
4403 unsigned int pages_per_huge_page)
4404{
4405 int i;
4406 struct page *dst_base = dst;
4407 struct page *src_base = src;
4408
4409 for (i = 0; i < pages_per_huge_page; ) {
4410 cond_resched();
4411 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4412
4413 i++;
4414 dst = mem_map_next(dst, dst_base, i);
4415 src = mem_map_next(src, src_base, i);
4416 }
4417}
4418
4419void copy_user_huge_page(struct page *dst, struct page *src,
4420 unsigned long addr, struct vm_area_struct *vma,
4421 unsigned int pages_per_huge_page)
4422{
4423 int i;
4424
4425 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4426 copy_user_gigantic_page(dst, src, addr, vma,
4427 pages_per_huge_page);
4428 return;
4429 }
4430
4431 might_sleep();
4432 for (i = 0; i < pages_per_huge_page; i++) {
4433 cond_resched();
4434 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4435 }
4436}
4437#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4438
4439#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4440
4441static struct kmem_cache *page_ptl_cachep;
4442
4443void __init ptlock_cache_init(void)
4444{
4445 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4446 SLAB_PANIC, NULL);
4447}
4448
4449bool ptlock_alloc(struct page *page)
4450{
4451 spinlock_t *ptl;
4452
4453 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4454 if (!ptl)
4455 return false;
4456 page->ptl = ptl;
4457 return true;
4458}
4459
4460void ptlock_free(struct page *page)
4461{
4462 kmem_cache_free(page_ptl_cachep, page->ptl);
4463}
4464#endif