Merge branch 'akpm' (patches from Andrew)
[muen/linux.git] / mm / memory.c
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/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72 #include <linux/numa.h>
73
74 #include <asm/io.h>
75 #include <asm/mmu_context.h>
76 #include <asm/pgalloc.h>
77 #include <linux/uaccess.h>
78 #include <asm/tlb.h>
79 #include <asm/tlbflush.h>
80 #include <asm/pgtable.h>
81
82 #include "internal.h"
83
84 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
85 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
86 #endif
87
88 #ifndef CONFIG_NEED_MULTIPLE_NODES
89 /* use the per-pgdat data instead for discontigmem - mbligh */
90 unsigned long max_mapnr;
91 EXPORT_SYMBOL(max_mapnr);
92
93 struct page *mem_map;
94 EXPORT_SYMBOL(mem_map);
95 #endif
96
97 /*
98  * A number of key systems in x86 including ioremap() rely on the assumption
99  * that high_memory defines the upper bound on direct map memory, then end
100  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
101  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
102  * and ZONE_HIGHMEM.
103  */
104 void *high_memory;
105 EXPORT_SYMBOL(high_memory);
106
107 /*
108  * Randomize the address space (stacks, mmaps, brk, etc.).
109  *
110  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
111  *   as ancient (libc5 based) binaries can segfault. )
112  */
113 int randomize_va_space __read_mostly =
114 #ifdef CONFIG_COMPAT_BRK
115                                         1;
116 #else
117                                         2;
118 #endif
119
120 static int __init disable_randmaps(char *s)
121 {
122         randomize_va_space = 0;
123         return 1;
124 }
125 __setup("norandmaps", disable_randmaps);
126
127 unsigned long zero_pfn __read_mostly;
128 EXPORT_SYMBOL(zero_pfn);
129
130 unsigned long highest_memmap_pfn __read_mostly;
131
132 /*
133  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
134  */
135 static int __init init_zero_pfn(void)
136 {
137         zero_pfn = page_to_pfn(ZERO_PAGE(0));
138         return 0;
139 }
140 core_initcall(init_zero_pfn);
141
142
143 #if defined(SPLIT_RSS_COUNTING)
144
145 void sync_mm_rss(struct mm_struct *mm)
146 {
147         int i;
148
149         for (i = 0; i < NR_MM_COUNTERS; i++) {
150                 if (current->rss_stat.count[i]) {
151                         add_mm_counter(mm, i, current->rss_stat.count[i]);
152                         current->rss_stat.count[i] = 0;
153                 }
154         }
155         current->rss_stat.events = 0;
156 }
157
158 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
159 {
160         struct task_struct *task = current;
161
162         if (likely(task->mm == mm))
163                 task->rss_stat.count[member] += val;
164         else
165                 add_mm_counter(mm, member, val);
166 }
167 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
168 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
169
170 /* sync counter once per 64 page faults */
171 #define TASK_RSS_EVENTS_THRESH  (64)
172 static void check_sync_rss_stat(struct task_struct *task)
173 {
174         if (unlikely(task != current))
175                 return;
176         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
177                 sync_mm_rss(task->mm);
178 }
179 #else /* SPLIT_RSS_COUNTING */
180
181 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
182 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
183
184 static void check_sync_rss_stat(struct task_struct *task)
185 {
186 }
187
188 #endif /* SPLIT_RSS_COUNTING */
189
190 /*
191  * Note: this doesn't free the actual pages themselves. That
192  * has been handled earlier when unmapping all the memory regions.
193  */
194 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
195                            unsigned long addr)
196 {
197         pgtable_t token = pmd_pgtable(*pmd);
198         pmd_clear(pmd);
199         pte_free_tlb(tlb, token, addr);
200         mm_dec_nr_ptes(tlb->mm);
201 }
202
203 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
204                                 unsigned long addr, unsigned long end,
205                                 unsigned long floor, unsigned long ceiling)
206 {
207         pmd_t *pmd;
208         unsigned long next;
209         unsigned long start;
210
211         start = addr;
212         pmd = pmd_offset(pud, addr);
213         do {
214                 next = pmd_addr_end(addr, end);
215                 if (pmd_none_or_clear_bad(pmd))
216                         continue;
217                 free_pte_range(tlb, pmd, addr);
218         } while (pmd++, addr = next, addr != end);
219
220         start &= PUD_MASK;
221         if (start < floor)
222                 return;
223         if (ceiling) {
224                 ceiling &= PUD_MASK;
225                 if (!ceiling)
226                         return;
227         }
228         if (end - 1 > ceiling - 1)
229                 return;
230
231         pmd = pmd_offset(pud, start);
232         pud_clear(pud);
233         pmd_free_tlb(tlb, pmd, start);
234         mm_dec_nr_pmds(tlb->mm);
235 }
236
237 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
238                                 unsigned long addr, unsigned long end,
239                                 unsigned long floor, unsigned long ceiling)
240 {
241         pud_t *pud;
242         unsigned long next;
243         unsigned long start;
244
245         start = addr;
246         pud = pud_offset(p4d, addr);
247         do {
248                 next = pud_addr_end(addr, end);
249                 if (pud_none_or_clear_bad(pud))
250                         continue;
251                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
252         } while (pud++, addr = next, addr != end);
253
254         start &= P4D_MASK;
255         if (start < floor)
256                 return;
257         if (ceiling) {
258                 ceiling &= P4D_MASK;
259                 if (!ceiling)
260                         return;
261         }
262         if (end - 1 > ceiling - 1)
263                 return;
264
265         pud = pud_offset(p4d, start);
266         p4d_clear(p4d);
267         pud_free_tlb(tlb, pud, start);
268         mm_dec_nr_puds(tlb->mm);
269 }
270
271 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
272                                 unsigned long addr, unsigned long end,
273                                 unsigned long floor, unsigned long ceiling)
274 {
275         p4d_t *p4d;
276         unsigned long next;
277         unsigned long start;
278
279         start = addr;
280         p4d = p4d_offset(pgd, addr);
281         do {
282                 next = p4d_addr_end(addr, end);
283                 if (p4d_none_or_clear_bad(p4d))
284                         continue;
285                 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
286         } while (p4d++, addr = next, addr != end);
287
288         start &= PGDIR_MASK;
289         if (start < floor)
290                 return;
291         if (ceiling) {
292                 ceiling &= PGDIR_MASK;
293                 if (!ceiling)
294                         return;
295         }
296         if (end - 1 > ceiling - 1)
297                 return;
298
299         p4d = p4d_offset(pgd, start);
300         pgd_clear(pgd);
301         p4d_free_tlb(tlb, p4d, start);
302 }
303
304 /*
305  * This function frees user-level page tables of a process.
306  */
307 void free_pgd_range(struct mmu_gather *tlb,
308                         unsigned long addr, unsigned long end,
309                         unsigned long floor, unsigned long ceiling)
310 {
311         pgd_t *pgd;
312         unsigned long next;
313
314         /*
315          * The next few lines have given us lots of grief...
316          *
317          * Why are we testing PMD* at this top level?  Because often
318          * there will be no work to do at all, and we'd prefer not to
319          * go all the way down to the bottom just to discover that.
320          *
321          * Why all these "- 1"s?  Because 0 represents both the bottom
322          * of the address space and the top of it (using -1 for the
323          * top wouldn't help much: the masks would do the wrong thing).
324          * The rule is that addr 0 and floor 0 refer to the bottom of
325          * the address space, but end 0 and ceiling 0 refer to the top
326          * Comparisons need to use "end - 1" and "ceiling - 1" (though
327          * that end 0 case should be mythical).
328          *
329          * Wherever addr is brought up or ceiling brought down, we must
330          * be careful to reject "the opposite 0" before it confuses the
331          * subsequent tests.  But what about where end is brought down
332          * by PMD_SIZE below? no, end can't go down to 0 there.
333          *
334          * Whereas we round start (addr) and ceiling down, by different
335          * masks at different levels, in order to test whether a table
336          * now has no other vmas using it, so can be freed, we don't
337          * bother to round floor or end up - the tests don't need that.
338          */
339
340         addr &= PMD_MASK;
341         if (addr < floor) {
342                 addr += PMD_SIZE;
343                 if (!addr)
344                         return;
345         }
346         if (ceiling) {
347                 ceiling &= PMD_MASK;
348                 if (!ceiling)
349                         return;
350         }
351         if (end - 1 > ceiling - 1)
352                 end -= PMD_SIZE;
353         if (addr > end - 1)
354                 return;
355         /*
356          * We add page table cache pages with PAGE_SIZE,
357          * (see pte_free_tlb()), flush the tlb if we need
358          */
359         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
360         pgd = pgd_offset(tlb->mm, addr);
361         do {
362                 next = pgd_addr_end(addr, end);
363                 if (pgd_none_or_clear_bad(pgd))
364                         continue;
365                 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
366         } while (pgd++, addr = next, addr != end);
367 }
368
369 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
370                 unsigned long floor, unsigned long ceiling)
371 {
372         while (vma) {
373                 struct vm_area_struct *next = vma->vm_next;
374                 unsigned long addr = vma->vm_start;
375
376                 /*
377                  * Hide vma from rmap and truncate_pagecache before freeing
378                  * pgtables
379                  */
380                 unlink_anon_vmas(vma);
381                 unlink_file_vma(vma);
382
383                 if (is_vm_hugetlb_page(vma)) {
384                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
385                                 floor, next ? next->vm_start : ceiling);
386                 } else {
387                         /*
388                          * Optimization: gather nearby vmas into one call down
389                          */
390                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
391                                && !is_vm_hugetlb_page(next)) {
392                                 vma = next;
393                                 next = vma->vm_next;
394                                 unlink_anon_vmas(vma);
395                                 unlink_file_vma(vma);
396                         }
397                         free_pgd_range(tlb, addr, vma->vm_end,
398                                 floor, next ? next->vm_start : ceiling);
399                 }
400                 vma = next;
401         }
402 }
403
404 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
405 {
406         spinlock_t *ptl;
407         pgtable_t new = pte_alloc_one(mm);
408         if (!new)
409                 return -ENOMEM;
410
411         /*
412          * Ensure all pte setup (eg. pte page lock and page clearing) are
413          * visible before the pte is made visible to other CPUs by being
414          * put into page tables.
415          *
416          * The other side of the story is the pointer chasing in the page
417          * table walking code (when walking the page table without locking;
418          * ie. most of the time). Fortunately, these data accesses consist
419          * of a chain of data-dependent loads, meaning most CPUs (alpha
420          * being the notable exception) will already guarantee loads are
421          * seen in-order. See the alpha page table accessors for the
422          * smp_read_barrier_depends() barriers in page table walking code.
423          */
424         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
425
426         ptl = pmd_lock(mm, pmd);
427         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
428                 mm_inc_nr_ptes(mm);
429                 pmd_populate(mm, pmd, new);
430                 new = NULL;
431         }
432         spin_unlock(ptl);
433         if (new)
434                 pte_free(mm, new);
435         return 0;
436 }
437
438 int __pte_alloc_kernel(pmd_t *pmd)
439 {
440         pte_t *new = pte_alloc_one_kernel(&init_mm);
441         if (!new)
442                 return -ENOMEM;
443
444         smp_wmb(); /* See comment in __pte_alloc */
445
446         spin_lock(&init_mm.page_table_lock);
447         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
448                 pmd_populate_kernel(&init_mm, pmd, new);
449                 new = NULL;
450         }
451         spin_unlock(&init_mm.page_table_lock);
452         if (new)
453                 pte_free_kernel(&init_mm, new);
454         return 0;
455 }
456
457 static inline void init_rss_vec(int *rss)
458 {
459         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
460 }
461
462 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
463 {
464         int i;
465
466         if (current->mm == mm)
467                 sync_mm_rss(mm);
468         for (i = 0; i < NR_MM_COUNTERS; i++)
469                 if (rss[i])
470                         add_mm_counter(mm, i, rss[i]);
471 }
472
473 /*
474  * This function is called to print an error when a bad pte
475  * is found. For example, we might have a PFN-mapped pte in
476  * a region that doesn't allow it.
477  *
478  * The calling function must still handle the error.
479  */
480 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
481                           pte_t pte, struct page *page)
482 {
483         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
484         p4d_t *p4d = p4d_offset(pgd, addr);
485         pud_t *pud = pud_offset(p4d, addr);
486         pmd_t *pmd = pmd_offset(pud, addr);
487         struct address_space *mapping;
488         pgoff_t index;
489         static unsigned long resume;
490         static unsigned long nr_shown;
491         static unsigned long nr_unshown;
492
493         /*
494          * Allow a burst of 60 reports, then keep quiet for that minute;
495          * or allow a steady drip of one report per second.
496          */
497         if (nr_shown == 60) {
498                 if (time_before(jiffies, resume)) {
499                         nr_unshown++;
500                         return;
501                 }
502                 if (nr_unshown) {
503                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
504                                  nr_unshown);
505                         nr_unshown = 0;
506                 }
507                 nr_shown = 0;
508         }
509         if (nr_shown++ == 0)
510                 resume = jiffies + 60 * HZ;
511
512         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
513         index = linear_page_index(vma, addr);
514
515         pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
516                  current->comm,
517                  (long long)pte_val(pte), (long long)pmd_val(*pmd));
518         if (page)
519                 dump_page(page, "bad pte");
520         pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
521                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
522         pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
523                  vma->vm_file,
524                  vma->vm_ops ? vma->vm_ops->fault : NULL,
525                  vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
526                  mapping ? mapping->a_ops->readpage : NULL);
527         dump_stack();
528         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
529 }
530
531 /*
532  * vm_normal_page -- This function gets the "struct page" associated with a pte.
533  *
534  * "Special" mappings do not wish to be associated with a "struct page" (either
535  * it doesn't exist, or it exists but they don't want to touch it). In this
536  * case, NULL is returned here. "Normal" mappings do have a struct page.
537  *
538  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
539  * pte bit, in which case this function is trivial. Secondly, an architecture
540  * may not have a spare pte bit, which requires a more complicated scheme,
541  * described below.
542  *
543  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
544  * special mapping (even if there are underlying and valid "struct pages").
545  * COWed pages of a VM_PFNMAP are always normal.
546  *
547  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
548  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
549  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
550  * mapping will always honor the rule
551  *
552  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
553  *
554  * And for normal mappings this is false.
555  *
556  * This restricts such mappings to be a linear translation from virtual address
557  * to pfn. To get around this restriction, we allow arbitrary mappings so long
558  * as the vma is not a COW mapping; in that case, we know that all ptes are
559  * special (because none can have been COWed).
560  *
561  *
562  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
563  *
564  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
565  * page" backing, however the difference is that _all_ pages with a struct
566  * page (that is, those where pfn_valid is true) are refcounted and considered
567  * normal pages by the VM. The disadvantage is that pages are refcounted
568  * (which can be slower and simply not an option for some PFNMAP users). The
569  * advantage is that we don't have to follow the strict linearity rule of
570  * PFNMAP mappings in order to support COWable mappings.
571  *
572  */
573 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
574                              pte_t pte, bool with_public_device)
575 {
576         unsigned long pfn = pte_pfn(pte);
577
578         if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
579                 if (likely(!pte_special(pte)))
580                         goto check_pfn;
581                 if (vma->vm_ops && vma->vm_ops->find_special_page)
582                         return vma->vm_ops->find_special_page(vma, addr);
583                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
584                         return NULL;
585                 if (is_zero_pfn(pfn))
586                         return NULL;
587
588                 /*
589                  * Device public pages are special pages (they are ZONE_DEVICE
590                  * pages but different from persistent memory). They behave
591                  * allmost like normal pages. The difference is that they are
592                  * not on the lru and thus should never be involve with any-
593                  * thing that involve lru manipulation (mlock, numa balancing,
594                  * ...).
595                  *
596                  * This is why we still want to return NULL for such page from
597                  * vm_normal_page() so that we do not have to special case all
598                  * call site of vm_normal_page().
599                  */
600                 if (likely(pfn <= highest_memmap_pfn)) {
601                         struct page *page = pfn_to_page(pfn);
602
603                         if (is_device_public_page(page)) {
604                                 if (with_public_device)
605                                         return page;
606                                 return NULL;
607                         }
608                 }
609
610                 if (pte_devmap(pte))
611                         return NULL;
612
613                 print_bad_pte(vma, addr, pte, NULL);
614                 return NULL;
615         }
616
617         /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
618
619         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
620                 if (vma->vm_flags & VM_MIXEDMAP) {
621                         if (!pfn_valid(pfn))
622                                 return NULL;
623                         goto out;
624                 } else {
625                         unsigned long off;
626                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
627                         if (pfn == vma->vm_pgoff + off)
628                                 return NULL;
629                         if (!is_cow_mapping(vma->vm_flags))
630                                 return NULL;
631                 }
632         }
633
634         if (is_zero_pfn(pfn))
635                 return NULL;
636
637 check_pfn:
638         if (unlikely(pfn > highest_memmap_pfn)) {
639                 print_bad_pte(vma, addr, pte, NULL);
640                 return NULL;
641         }
642
643         /*
644          * NOTE! We still have PageReserved() pages in the page tables.
645          * eg. VDSO mappings can cause them to exist.
646          */
647 out:
648         return pfn_to_page(pfn);
649 }
650
651 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
652 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
653                                 pmd_t pmd)
654 {
655         unsigned long pfn = pmd_pfn(pmd);
656
657         /*
658          * There is no pmd_special() but there may be special pmds, e.g.
659          * in a direct-access (dax) mapping, so let's just replicate the
660          * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
661          */
662         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
663                 if (vma->vm_flags & VM_MIXEDMAP) {
664                         if (!pfn_valid(pfn))
665                                 return NULL;
666                         goto out;
667                 } else {
668                         unsigned long off;
669                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
670                         if (pfn == vma->vm_pgoff + off)
671                                 return NULL;
672                         if (!is_cow_mapping(vma->vm_flags))
673                                 return NULL;
674                 }
675         }
676
677         if (pmd_devmap(pmd))
678                 return NULL;
679         if (is_zero_pfn(pfn))
680                 return NULL;
681         if (unlikely(pfn > highest_memmap_pfn))
682                 return NULL;
683
684         /*
685          * NOTE! We still have PageReserved() pages in the page tables.
686          * eg. VDSO mappings can cause them to exist.
687          */
688 out:
689         return pfn_to_page(pfn);
690 }
691 #endif
692
693 /*
694  * copy one vm_area from one task to the other. Assumes the page tables
695  * already present in the new task to be cleared in the whole range
696  * covered by this vma.
697  */
698
699 static inline unsigned long
700 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
701                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
702                 unsigned long addr, int *rss)
703 {
704         unsigned long vm_flags = vma->vm_flags;
705         pte_t pte = *src_pte;
706         struct page *page;
707
708         /* pte contains position in swap or file, so copy. */
709         if (unlikely(!pte_present(pte))) {
710                 swp_entry_t entry = pte_to_swp_entry(pte);
711
712                 if (likely(!non_swap_entry(entry))) {
713                         if (swap_duplicate(entry) < 0)
714                                 return entry.val;
715
716                         /* make sure dst_mm is on swapoff's mmlist. */
717                         if (unlikely(list_empty(&dst_mm->mmlist))) {
718                                 spin_lock(&mmlist_lock);
719                                 if (list_empty(&dst_mm->mmlist))
720                                         list_add(&dst_mm->mmlist,
721                                                         &src_mm->mmlist);
722                                 spin_unlock(&mmlist_lock);
723                         }
724                         rss[MM_SWAPENTS]++;
725                 } else if (is_migration_entry(entry)) {
726                         page = migration_entry_to_page(entry);
727
728                         rss[mm_counter(page)]++;
729
730                         if (is_write_migration_entry(entry) &&
731                                         is_cow_mapping(vm_flags)) {
732                                 /*
733                                  * COW mappings require pages in both
734                                  * parent and child to be set to read.
735                                  */
736                                 make_migration_entry_read(&entry);
737                                 pte = swp_entry_to_pte(entry);
738                                 if (pte_swp_soft_dirty(*src_pte))
739                                         pte = pte_swp_mksoft_dirty(pte);
740                                 set_pte_at(src_mm, addr, src_pte, pte);
741                         }
742                 } else if (is_device_private_entry(entry)) {
743                         page = device_private_entry_to_page(entry);
744
745                         /*
746                          * Update rss count even for unaddressable pages, as
747                          * they should treated just like normal pages in this
748                          * respect.
749                          *
750                          * We will likely want to have some new rss counters
751                          * for unaddressable pages, at some point. But for now
752                          * keep things as they are.
753                          */
754                         get_page(page);
755                         rss[mm_counter(page)]++;
756                         page_dup_rmap(page, false);
757
758                         /*
759                          * We do not preserve soft-dirty information, because so
760                          * far, checkpoint/restore is the only feature that
761                          * requires that. And checkpoint/restore does not work
762                          * when a device driver is involved (you cannot easily
763                          * save and restore device driver state).
764                          */
765                         if (is_write_device_private_entry(entry) &&
766                             is_cow_mapping(vm_flags)) {
767                                 make_device_private_entry_read(&entry);
768                                 pte = swp_entry_to_pte(entry);
769                                 set_pte_at(src_mm, addr, src_pte, pte);
770                         }
771                 }
772                 goto out_set_pte;
773         }
774
775         /*
776          * If it's a COW mapping, write protect it both
777          * in the parent and the child
778          */
779         if (is_cow_mapping(vm_flags) && pte_write(pte)) {
780                 ptep_set_wrprotect(src_mm, addr, src_pte);
781                 pte = pte_wrprotect(pte);
782         }
783
784         /*
785          * If it's a shared mapping, mark it clean in
786          * the child
787          */
788         if (vm_flags & VM_SHARED)
789                 pte = pte_mkclean(pte);
790         pte = pte_mkold(pte);
791
792         page = vm_normal_page(vma, addr, pte);
793         if (page) {
794                 get_page(page);
795                 page_dup_rmap(page, false);
796                 rss[mm_counter(page)]++;
797         } else if (pte_devmap(pte)) {
798                 page = pte_page(pte);
799
800                 /*
801                  * Cache coherent device memory behave like regular page and
802                  * not like persistent memory page. For more informations see
803                  * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
804                  */
805                 if (is_device_public_page(page)) {
806                         get_page(page);
807                         page_dup_rmap(page, false);
808                         rss[mm_counter(page)]++;
809                 }
810         }
811
812 out_set_pte:
813         set_pte_at(dst_mm, addr, dst_pte, pte);
814         return 0;
815 }
816
817 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
818                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
819                    unsigned long addr, unsigned long end)
820 {
821         pte_t *orig_src_pte, *orig_dst_pte;
822         pte_t *src_pte, *dst_pte;
823         spinlock_t *src_ptl, *dst_ptl;
824         int progress = 0;
825         int rss[NR_MM_COUNTERS];
826         swp_entry_t entry = (swp_entry_t){0};
827
828 again:
829         init_rss_vec(rss);
830
831         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
832         if (!dst_pte)
833                 return -ENOMEM;
834         src_pte = pte_offset_map(src_pmd, addr);
835         src_ptl = pte_lockptr(src_mm, src_pmd);
836         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
837         orig_src_pte = src_pte;
838         orig_dst_pte = dst_pte;
839         arch_enter_lazy_mmu_mode();
840
841         do {
842                 /*
843                  * We are holding two locks at this point - either of them
844                  * could generate latencies in another task on another CPU.
845                  */
846                 if (progress >= 32) {
847                         progress = 0;
848                         if (need_resched() ||
849                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
850                                 break;
851                 }
852                 if (pte_none(*src_pte)) {
853                         progress++;
854                         continue;
855                 }
856                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
857                                                         vma, addr, rss);
858                 if (entry.val)
859                         break;
860                 progress += 8;
861         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
862
863         arch_leave_lazy_mmu_mode();
864         spin_unlock(src_ptl);
865         pte_unmap(orig_src_pte);
866         add_mm_rss_vec(dst_mm, rss);
867         pte_unmap_unlock(orig_dst_pte, dst_ptl);
868         cond_resched();
869
870         if (entry.val) {
871                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
872                         return -ENOMEM;
873                 progress = 0;
874         }
875         if (addr != end)
876                 goto again;
877         return 0;
878 }
879
880 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
881                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
882                 unsigned long addr, unsigned long end)
883 {
884         pmd_t *src_pmd, *dst_pmd;
885         unsigned long next;
886
887         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
888         if (!dst_pmd)
889                 return -ENOMEM;
890         src_pmd = pmd_offset(src_pud, addr);
891         do {
892                 next = pmd_addr_end(addr, end);
893                 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
894                         || pmd_devmap(*src_pmd)) {
895                         int err;
896                         VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
897                         err = copy_huge_pmd(dst_mm, src_mm,
898                                             dst_pmd, src_pmd, addr, vma);
899                         if (err == -ENOMEM)
900                                 return -ENOMEM;
901                         if (!err)
902                                 continue;
903                         /* fall through */
904                 }
905                 if (pmd_none_or_clear_bad(src_pmd))
906                         continue;
907                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
908                                                 vma, addr, next))
909                         return -ENOMEM;
910         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
911         return 0;
912 }
913
914 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
915                 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
916                 unsigned long addr, unsigned long end)
917 {
918         pud_t *src_pud, *dst_pud;
919         unsigned long next;
920
921         dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
922         if (!dst_pud)
923                 return -ENOMEM;
924         src_pud = pud_offset(src_p4d, addr);
925         do {
926                 next = pud_addr_end(addr, end);
927                 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
928                         int err;
929
930                         VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
931                         err = copy_huge_pud(dst_mm, src_mm,
932                                             dst_pud, src_pud, addr, vma);
933                         if (err == -ENOMEM)
934                                 return -ENOMEM;
935                         if (!err)
936                                 continue;
937                         /* fall through */
938                 }
939                 if (pud_none_or_clear_bad(src_pud))
940                         continue;
941                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
942                                                 vma, addr, next))
943                         return -ENOMEM;
944         } while (dst_pud++, src_pud++, addr = next, addr != end);
945         return 0;
946 }
947
948 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
949                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
950                 unsigned long addr, unsigned long end)
951 {
952         p4d_t *src_p4d, *dst_p4d;
953         unsigned long next;
954
955         dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
956         if (!dst_p4d)
957                 return -ENOMEM;
958         src_p4d = p4d_offset(src_pgd, addr);
959         do {
960                 next = p4d_addr_end(addr, end);
961                 if (p4d_none_or_clear_bad(src_p4d))
962                         continue;
963                 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
964                                                 vma, addr, next))
965                         return -ENOMEM;
966         } while (dst_p4d++, src_p4d++, addr = next, addr != end);
967         return 0;
968 }
969
970 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
971                 struct vm_area_struct *vma)
972 {
973         pgd_t *src_pgd, *dst_pgd;
974         unsigned long next;
975         unsigned long addr = vma->vm_start;
976         unsigned long end = vma->vm_end;
977         struct mmu_notifier_range range;
978         bool is_cow;
979         int ret;
980
981         /*
982          * Don't copy ptes where a page fault will fill them correctly.
983          * Fork becomes much lighter when there are big shared or private
984          * readonly mappings. The tradeoff is that copy_page_range is more
985          * efficient than faulting.
986          */
987         if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
988                         !vma->anon_vma)
989                 return 0;
990
991         if (is_vm_hugetlb_page(vma))
992                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
993
994         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
995                 /*
996                  * We do not free on error cases below as remove_vma
997                  * gets called on error from higher level routine
998                  */
999                 ret = track_pfn_copy(vma);
1000                 if (ret)
1001                         return ret;
1002         }
1003
1004         /*
1005          * We need to invalidate the secondary MMU mappings only when
1006          * there could be a permission downgrade on the ptes of the
1007          * parent mm. And a permission downgrade will only happen if
1008          * is_cow_mapping() returns true.
1009          */
1010         is_cow = is_cow_mapping(vma->vm_flags);
1011
1012         if (is_cow) {
1013                 mmu_notifier_range_init(&range, src_mm, addr, end);
1014                 mmu_notifier_invalidate_range_start(&range);
1015         }
1016
1017         ret = 0;
1018         dst_pgd = pgd_offset(dst_mm, addr);
1019         src_pgd = pgd_offset(src_mm, addr);
1020         do {
1021                 next = pgd_addr_end(addr, end);
1022                 if (pgd_none_or_clear_bad(src_pgd))
1023                         continue;
1024                 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1025                                             vma, addr, next))) {
1026                         ret = -ENOMEM;
1027                         break;
1028                 }
1029         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1030
1031         if (is_cow)
1032                 mmu_notifier_invalidate_range_end(&range);
1033         return ret;
1034 }
1035
1036 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1037                                 struct vm_area_struct *vma, pmd_t *pmd,
1038                                 unsigned long addr, unsigned long end,
1039                                 struct zap_details *details)
1040 {
1041         struct mm_struct *mm = tlb->mm;
1042         int force_flush = 0;
1043         int rss[NR_MM_COUNTERS];
1044         spinlock_t *ptl;
1045         pte_t *start_pte;
1046         pte_t *pte;
1047         swp_entry_t entry;
1048
1049         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1050 again:
1051         init_rss_vec(rss);
1052         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1053         pte = start_pte;
1054         flush_tlb_batched_pending(mm);
1055         arch_enter_lazy_mmu_mode();
1056         do {
1057                 pte_t ptent = *pte;
1058                 if (pte_none(ptent))
1059                         continue;
1060
1061                 if (pte_present(ptent)) {
1062                         struct page *page;
1063
1064                         page = _vm_normal_page(vma, addr, ptent, true);
1065                         if (unlikely(details) && page) {
1066                                 /*
1067                                  * unmap_shared_mapping_pages() wants to
1068                                  * invalidate cache without truncating:
1069                                  * unmap shared but keep private pages.
1070                                  */
1071                                 if (details->check_mapping &&
1072                                     details->check_mapping != page_rmapping(page))
1073                                         continue;
1074                         }
1075                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1076                                                         tlb->fullmm);
1077                         tlb_remove_tlb_entry(tlb, pte, addr);
1078                         if (unlikely(!page))
1079                                 continue;
1080
1081                         if (!PageAnon(page)) {
1082                                 if (pte_dirty(ptent)) {
1083                                         force_flush = 1;
1084                                         set_page_dirty(page);
1085                                 }
1086                                 if (pte_young(ptent) &&
1087                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1088                                         mark_page_accessed(page);
1089                         }
1090                         rss[mm_counter(page)]--;
1091                         page_remove_rmap(page, false);
1092                         if (unlikely(page_mapcount(page) < 0))
1093                                 print_bad_pte(vma, addr, ptent, page);
1094                         if (unlikely(__tlb_remove_page(tlb, page))) {
1095                                 force_flush = 1;
1096                                 addr += PAGE_SIZE;
1097                                 break;
1098                         }
1099                         continue;
1100                 }
1101
1102                 entry = pte_to_swp_entry(ptent);
1103                 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1104                         struct page *page = device_private_entry_to_page(entry);
1105
1106                         if (unlikely(details && details->check_mapping)) {
1107                                 /*
1108                                  * unmap_shared_mapping_pages() wants to
1109                                  * invalidate cache without truncating:
1110                                  * unmap shared but keep private pages.
1111                                  */
1112                                 if (details->check_mapping !=
1113                                     page_rmapping(page))
1114                                         continue;
1115                         }
1116
1117                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1118                         rss[mm_counter(page)]--;
1119                         page_remove_rmap(page, false);
1120                         put_page(page);
1121                         continue;
1122                 }
1123
1124                 /* If details->check_mapping, we leave swap entries. */
1125                 if (unlikely(details))
1126                         continue;
1127
1128                 entry = pte_to_swp_entry(ptent);
1129                 if (!non_swap_entry(entry))
1130                         rss[MM_SWAPENTS]--;
1131                 else if (is_migration_entry(entry)) {
1132                         struct page *page;
1133
1134                         page = migration_entry_to_page(entry);
1135                         rss[mm_counter(page)]--;
1136                 }
1137                 if (unlikely(!free_swap_and_cache(entry)))
1138                         print_bad_pte(vma, addr, ptent, NULL);
1139                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1140         } while (pte++, addr += PAGE_SIZE, addr != end);
1141
1142         add_mm_rss_vec(mm, rss);
1143         arch_leave_lazy_mmu_mode();
1144
1145         /* Do the actual TLB flush before dropping ptl */
1146         if (force_flush)
1147                 tlb_flush_mmu_tlbonly(tlb);
1148         pte_unmap_unlock(start_pte, ptl);
1149
1150         /*
1151          * If we forced a TLB flush (either due to running out of
1152          * batch buffers or because we needed to flush dirty TLB
1153          * entries before releasing the ptl), free the batched
1154          * memory too. Restart if we didn't do everything.
1155          */
1156         if (force_flush) {
1157                 force_flush = 0;
1158                 tlb_flush_mmu_free(tlb);
1159                 if (addr != end)
1160                         goto again;
1161         }
1162
1163         return addr;
1164 }
1165
1166 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1167                                 struct vm_area_struct *vma, pud_t *pud,
1168                                 unsigned long addr, unsigned long end,
1169                                 struct zap_details *details)
1170 {
1171         pmd_t *pmd;
1172         unsigned long next;
1173
1174         pmd = pmd_offset(pud, addr);
1175         do {
1176                 next = pmd_addr_end(addr, end);
1177                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1178                         if (next - addr != HPAGE_PMD_SIZE)
1179                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1180                         else if (zap_huge_pmd(tlb, vma, pmd, addr))
1181                                 goto next;
1182                         /* fall through */
1183                 }
1184                 /*
1185                  * Here there can be other concurrent MADV_DONTNEED or
1186                  * trans huge page faults running, and if the pmd is
1187                  * none or trans huge it can change under us. This is
1188                  * because MADV_DONTNEED holds the mmap_sem in read
1189                  * mode.
1190                  */
1191                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1192                         goto next;
1193                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1194 next:
1195                 cond_resched();
1196         } while (pmd++, addr = next, addr != end);
1197
1198         return addr;
1199 }
1200
1201 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1202                                 struct vm_area_struct *vma, p4d_t *p4d,
1203                                 unsigned long addr, unsigned long end,
1204                                 struct zap_details *details)
1205 {
1206         pud_t *pud;
1207         unsigned long next;
1208
1209         pud = pud_offset(p4d, addr);
1210         do {
1211                 next = pud_addr_end(addr, end);
1212                 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1213                         if (next - addr != HPAGE_PUD_SIZE) {
1214                                 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1215                                 split_huge_pud(vma, pud, addr);
1216                         } else if (zap_huge_pud(tlb, vma, pud, addr))
1217                                 goto next;
1218                         /* fall through */
1219                 }
1220                 if (pud_none_or_clear_bad(pud))
1221                         continue;
1222                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1223 next:
1224                 cond_resched();
1225         } while (pud++, addr = next, addr != end);
1226
1227         return addr;
1228 }
1229
1230 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1231                                 struct vm_area_struct *vma, pgd_t *pgd,
1232                                 unsigned long addr, unsigned long end,
1233                                 struct zap_details *details)
1234 {
1235         p4d_t *p4d;
1236         unsigned long next;
1237
1238         p4d = p4d_offset(pgd, addr);
1239         do {
1240                 next = p4d_addr_end(addr, end);
1241                 if (p4d_none_or_clear_bad(p4d))
1242                         continue;
1243                 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1244         } while (p4d++, addr = next, addr != end);
1245
1246         return addr;
1247 }
1248
1249 void unmap_page_range(struct mmu_gather *tlb,
1250                              struct vm_area_struct *vma,
1251                              unsigned long addr, unsigned long end,
1252                              struct zap_details *details)
1253 {
1254         pgd_t *pgd;
1255         unsigned long next;
1256
1257         BUG_ON(addr >= end);
1258         tlb_start_vma(tlb, vma);
1259         pgd = pgd_offset(vma->vm_mm, addr);
1260         do {
1261                 next = pgd_addr_end(addr, end);
1262                 if (pgd_none_or_clear_bad(pgd))
1263                         continue;
1264                 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1265         } while (pgd++, addr = next, addr != end);
1266         tlb_end_vma(tlb, vma);
1267 }
1268
1269
1270 static void unmap_single_vma(struct mmu_gather *tlb,
1271                 struct vm_area_struct *vma, unsigned long start_addr,
1272                 unsigned long end_addr,
1273                 struct zap_details *details)
1274 {
1275         unsigned long start = max(vma->vm_start, start_addr);
1276         unsigned long end;
1277
1278         if (start >= vma->vm_end)
1279                 return;
1280         end = min(vma->vm_end, end_addr);
1281         if (end <= vma->vm_start)
1282                 return;
1283
1284         if (vma->vm_file)
1285                 uprobe_munmap(vma, start, end);
1286
1287         if (unlikely(vma->vm_flags & VM_PFNMAP))
1288                 untrack_pfn(vma, 0, 0);
1289
1290         if (start != end) {
1291                 if (unlikely(is_vm_hugetlb_page(vma))) {
1292                         /*
1293                          * It is undesirable to test vma->vm_file as it
1294                          * should be non-null for valid hugetlb area.
1295                          * However, vm_file will be NULL in the error
1296                          * cleanup path of mmap_region. When
1297                          * hugetlbfs ->mmap method fails,
1298                          * mmap_region() nullifies vma->vm_file
1299                          * before calling this function to clean up.
1300                          * Since no pte has actually been setup, it is
1301                          * safe to do nothing in this case.
1302                          */
1303                         if (vma->vm_file) {
1304                                 i_mmap_lock_write(vma->vm_file->f_mapping);
1305                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1306                                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1307                         }
1308                 } else
1309                         unmap_page_range(tlb, vma, start, end, details);
1310         }
1311 }
1312
1313 /**
1314  * unmap_vmas - unmap a range of memory covered by a list of vma's
1315  * @tlb: address of the caller's struct mmu_gather
1316  * @vma: the starting vma
1317  * @start_addr: virtual address at which to start unmapping
1318  * @end_addr: virtual address at which to end unmapping
1319  *
1320  * Unmap all pages in the vma list.
1321  *
1322  * Only addresses between `start' and `end' will be unmapped.
1323  *
1324  * The VMA list must be sorted in ascending virtual address order.
1325  *
1326  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1327  * range after unmap_vmas() returns.  So the only responsibility here is to
1328  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1329  * drops the lock and schedules.
1330  */
1331 void unmap_vmas(struct mmu_gather *tlb,
1332                 struct vm_area_struct *vma, unsigned long start_addr,
1333                 unsigned long end_addr)
1334 {
1335         struct mmu_notifier_range range;
1336
1337         mmu_notifier_range_init(&range, vma->vm_mm, start_addr, end_addr);
1338         mmu_notifier_invalidate_range_start(&range);
1339         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1340                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1341         mmu_notifier_invalidate_range_end(&range);
1342 }
1343
1344 /**
1345  * zap_page_range - remove user pages in a given range
1346  * @vma: vm_area_struct holding the applicable pages
1347  * @start: starting address of pages to zap
1348  * @size: number of bytes to zap
1349  *
1350  * Caller must protect the VMA list
1351  */
1352 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1353                 unsigned long size)
1354 {
1355         struct mmu_notifier_range range;
1356         struct mmu_gather tlb;
1357
1358         lru_add_drain();
1359         mmu_notifier_range_init(&range, vma->vm_mm, start, start + size);
1360         tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1361         update_hiwater_rss(vma->vm_mm);
1362         mmu_notifier_invalidate_range_start(&range);
1363         for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1364                 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1365         mmu_notifier_invalidate_range_end(&range);
1366         tlb_finish_mmu(&tlb, start, range.end);
1367 }
1368
1369 /**
1370  * zap_page_range_single - remove user pages in a given range
1371  * @vma: vm_area_struct holding the applicable pages
1372  * @address: starting address of pages to zap
1373  * @size: number of bytes to zap
1374  * @details: details of shared cache invalidation
1375  *
1376  * The range must fit into one VMA.
1377  */
1378 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1379                 unsigned long size, struct zap_details *details)
1380 {
1381         struct mmu_notifier_range range;
1382         struct mmu_gather tlb;
1383
1384         lru_add_drain();
1385         mmu_notifier_range_init(&range, vma->vm_mm, address, address + size);
1386         tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1387         update_hiwater_rss(vma->vm_mm);
1388         mmu_notifier_invalidate_range_start(&range);
1389         unmap_single_vma(&tlb, vma, address, range.end, details);
1390         mmu_notifier_invalidate_range_end(&range);
1391         tlb_finish_mmu(&tlb, address, range.end);
1392 }
1393
1394 /**
1395  * zap_vma_ptes - remove ptes mapping the vma
1396  * @vma: vm_area_struct holding ptes to be zapped
1397  * @address: starting address of pages to zap
1398  * @size: number of bytes to zap
1399  *
1400  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1401  *
1402  * The entire address range must be fully contained within the vma.
1403  *
1404  */
1405 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1406                 unsigned long size)
1407 {
1408         if (address < vma->vm_start || address + size > vma->vm_end ||
1409                         !(vma->vm_flags & VM_PFNMAP))
1410                 return;
1411
1412         zap_page_range_single(vma, address, size, NULL);
1413 }
1414 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1415
1416 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1417                         spinlock_t **ptl)
1418 {
1419         pgd_t *pgd;
1420         p4d_t *p4d;
1421         pud_t *pud;
1422         pmd_t *pmd;
1423
1424         pgd = pgd_offset(mm, addr);
1425         p4d = p4d_alloc(mm, pgd, addr);
1426         if (!p4d)
1427                 return NULL;
1428         pud = pud_alloc(mm, p4d, addr);
1429         if (!pud)
1430                 return NULL;
1431         pmd = pmd_alloc(mm, pud, addr);
1432         if (!pmd)
1433                 return NULL;
1434
1435         VM_BUG_ON(pmd_trans_huge(*pmd));
1436         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1437 }
1438
1439 /*
1440  * This is the old fallback for page remapping.
1441  *
1442  * For historical reasons, it only allows reserved pages. Only
1443  * old drivers should use this, and they needed to mark their
1444  * pages reserved for the old functions anyway.
1445  */
1446 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1447                         struct page *page, pgprot_t prot)
1448 {
1449         struct mm_struct *mm = vma->vm_mm;
1450         int retval;
1451         pte_t *pte;
1452         spinlock_t *ptl;
1453
1454         retval = -EINVAL;
1455         if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1456                 goto out;
1457         retval = -ENOMEM;
1458         flush_dcache_page(page);
1459         pte = get_locked_pte(mm, addr, &ptl);
1460         if (!pte)
1461                 goto out;
1462         retval = -EBUSY;
1463         if (!pte_none(*pte))
1464                 goto out_unlock;
1465
1466         /* Ok, finally just insert the thing.. */
1467         get_page(page);
1468         inc_mm_counter_fast(mm, mm_counter_file(page));
1469         page_add_file_rmap(page, false);
1470         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1471
1472         retval = 0;
1473         pte_unmap_unlock(pte, ptl);
1474         return retval;
1475 out_unlock:
1476         pte_unmap_unlock(pte, ptl);
1477 out:
1478         return retval;
1479 }
1480
1481 /**
1482  * vm_insert_page - insert single page into user vma
1483  * @vma: user vma to map to
1484  * @addr: target user address of this page
1485  * @page: source kernel page
1486  *
1487  * This allows drivers to insert individual pages they've allocated
1488  * into a user vma.
1489  *
1490  * The page has to be a nice clean _individual_ kernel allocation.
1491  * If you allocate a compound page, you need to have marked it as
1492  * such (__GFP_COMP), or manually just split the page up yourself
1493  * (see split_page()).
1494  *
1495  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1496  * took an arbitrary page protection parameter. This doesn't allow
1497  * that. Your vma protection will have to be set up correctly, which
1498  * means that if you want a shared writable mapping, you'd better
1499  * ask for a shared writable mapping!
1500  *
1501  * The page does not need to be reserved.
1502  *
1503  * Usually this function is called from f_op->mmap() handler
1504  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1505  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1506  * function from other places, for example from page-fault handler.
1507  *
1508  * Return: %0 on success, negative error code otherwise.
1509  */
1510 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1511                         struct page *page)
1512 {
1513         if (addr < vma->vm_start || addr >= vma->vm_end)
1514                 return -EFAULT;
1515         if (!page_count(page))
1516                 return -EINVAL;
1517         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1518                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1519                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1520                 vma->vm_flags |= VM_MIXEDMAP;
1521         }
1522         return insert_page(vma, addr, page, vma->vm_page_prot);
1523 }
1524 EXPORT_SYMBOL(vm_insert_page);
1525
1526 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1527                         pfn_t pfn, pgprot_t prot, bool mkwrite)
1528 {
1529         struct mm_struct *mm = vma->vm_mm;
1530         pte_t *pte, entry;
1531         spinlock_t *ptl;
1532
1533         pte = get_locked_pte(mm, addr, &ptl);
1534         if (!pte)
1535                 return VM_FAULT_OOM;
1536         if (!pte_none(*pte)) {
1537                 if (mkwrite) {
1538                         /*
1539                          * For read faults on private mappings the PFN passed
1540                          * in may not match the PFN we have mapped if the
1541                          * mapped PFN is a writeable COW page.  In the mkwrite
1542                          * case we are creating a writable PTE for a shared
1543                          * mapping and we expect the PFNs to match. If they
1544                          * don't match, we are likely racing with block
1545                          * allocation and mapping invalidation so just skip the
1546                          * update.
1547                          */
1548                         if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1549                                 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1550                                 goto out_unlock;
1551                         }
1552                         entry = pte_mkyoung(*pte);
1553                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1554                         if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1555                                 update_mmu_cache(vma, addr, pte);
1556                 }
1557                 goto out_unlock;
1558         }
1559
1560         /* Ok, finally just insert the thing.. */
1561         if (pfn_t_devmap(pfn))
1562                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1563         else
1564                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1565
1566         if (mkwrite) {
1567                 entry = pte_mkyoung(entry);
1568                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1569         }
1570
1571         set_pte_at(mm, addr, pte, entry);
1572         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1573
1574 out_unlock:
1575         pte_unmap_unlock(pte, ptl);
1576         return VM_FAULT_NOPAGE;
1577 }
1578
1579 /**
1580  * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1581  * @vma: user vma to map to
1582  * @addr: target user address of this page
1583  * @pfn: source kernel pfn
1584  * @pgprot: pgprot flags for the inserted page
1585  *
1586  * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1587  * to override pgprot on a per-page basis.
1588  *
1589  * This only makes sense for IO mappings, and it makes no sense for
1590  * COW mappings.  In general, using multiple vmas is preferable;
1591  * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1592  * impractical.
1593  *
1594  * Context: Process context.  May allocate using %GFP_KERNEL.
1595  * Return: vm_fault_t value.
1596  */
1597 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1598                         unsigned long pfn, pgprot_t pgprot)
1599 {
1600         /*
1601          * Technically, architectures with pte_special can avoid all these
1602          * restrictions (same for remap_pfn_range).  However we would like
1603          * consistency in testing and feature parity among all, so we should
1604          * try to keep these invariants in place for everybody.
1605          */
1606         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1607         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1608                                                 (VM_PFNMAP|VM_MIXEDMAP));
1609         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1610         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1611
1612         if (addr < vma->vm_start || addr >= vma->vm_end)
1613                 return VM_FAULT_SIGBUS;
1614
1615         if (!pfn_modify_allowed(pfn, pgprot))
1616                 return VM_FAULT_SIGBUS;
1617
1618         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1619
1620         return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1621                         false);
1622 }
1623 EXPORT_SYMBOL(vmf_insert_pfn_prot);
1624
1625 /**
1626  * vmf_insert_pfn - insert single pfn into user vma
1627  * @vma: user vma to map to
1628  * @addr: target user address of this page
1629  * @pfn: source kernel pfn
1630  *
1631  * Similar to vm_insert_page, this allows drivers to insert individual pages
1632  * they've allocated into a user vma. Same comments apply.
1633  *
1634  * This function should only be called from a vm_ops->fault handler, and
1635  * in that case the handler should return the result of this function.
1636  *
1637  * vma cannot be a COW mapping.
1638  *
1639  * As this is called only for pages that do not currently exist, we
1640  * do not need to flush old virtual caches or the TLB.
1641  *
1642  * Context: Process context.  May allocate using %GFP_KERNEL.
1643  * Return: vm_fault_t value.
1644  */
1645 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1646                         unsigned long pfn)
1647 {
1648         return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1649 }
1650 EXPORT_SYMBOL(vmf_insert_pfn);
1651
1652 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1653 {
1654         /* these checks mirror the abort conditions in vm_normal_page */
1655         if (vma->vm_flags & VM_MIXEDMAP)
1656                 return true;
1657         if (pfn_t_devmap(pfn))
1658                 return true;
1659         if (pfn_t_special(pfn))
1660                 return true;
1661         if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1662                 return true;
1663         return false;
1664 }
1665
1666 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1667                 unsigned long addr, pfn_t pfn, bool mkwrite)
1668 {
1669         pgprot_t pgprot = vma->vm_page_prot;
1670         int err;
1671
1672         BUG_ON(!vm_mixed_ok(vma, pfn));
1673
1674         if (addr < vma->vm_start || addr >= vma->vm_end)
1675                 return VM_FAULT_SIGBUS;
1676
1677         track_pfn_insert(vma, &pgprot, pfn);
1678
1679         if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1680                 return VM_FAULT_SIGBUS;
1681
1682         /*
1683          * If we don't have pte special, then we have to use the pfn_valid()
1684          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1685          * refcount the page if pfn_valid is true (hence insert_page rather
1686          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1687          * without pte special, it would there be refcounted as a normal page.
1688          */
1689         if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1690             !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1691                 struct page *page;
1692
1693                 /*
1694                  * At this point we are committed to insert_page()
1695                  * regardless of whether the caller specified flags that
1696                  * result in pfn_t_has_page() == false.
1697                  */
1698                 page = pfn_to_page(pfn_t_to_pfn(pfn));
1699                 err = insert_page(vma, addr, page, pgprot);
1700         } else {
1701                 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1702         }
1703
1704         if (err == -ENOMEM)
1705                 return VM_FAULT_OOM;
1706         if (err < 0 && err != -EBUSY)
1707                 return VM_FAULT_SIGBUS;
1708
1709         return VM_FAULT_NOPAGE;
1710 }
1711
1712 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1713                 pfn_t pfn)
1714 {
1715         return __vm_insert_mixed(vma, addr, pfn, false);
1716 }
1717 EXPORT_SYMBOL(vmf_insert_mixed);
1718
1719 /*
1720  *  If the insertion of PTE failed because someone else already added a
1721  *  different entry in the mean time, we treat that as success as we assume
1722  *  the same entry was actually inserted.
1723  */
1724 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1725                 unsigned long addr, pfn_t pfn)
1726 {
1727         return __vm_insert_mixed(vma, addr, pfn, true);
1728 }
1729 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1730
1731 /*
1732  * maps a range of physical memory into the requested pages. the old
1733  * mappings are removed. any references to nonexistent pages results
1734  * in null mappings (currently treated as "copy-on-access")
1735  */
1736 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1737                         unsigned long addr, unsigned long end,
1738                         unsigned long pfn, pgprot_t prot)
1739 {
1740         pte_t *pte;
1741         spinlock_t *ptl;
1742         int err = 0;
1743
1744         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1745         if (!pte)
1746                 return -ENOMEM;
1747         arch_enter_lazy_mmu_mode();
1748         do {
1749                 BUG_ON(!pte_none(*pte));
1750                 if (!pfn_modify_allowed(pfn, prot)) {
1751                         err = -EACCES;
1752                         break;
1753                 }
1754                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1755                 pfn++;
1756         } while (pte++, addr += PAGE_SIZE, addr != end);
1757         arch_leave_lazy_mmu_mode();
1758         pte_unmap_unlock(pte - 1, ptl);
1759         return err;
1760 }
1761
1762 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1763                         unsigned long addr, unsigned long end,
1764                         unsigned long pfn, pgprot_t prot)
1765 {
1766         pmd_t *pmd;
1767         unsigned long next;
1768         int err;
1769
1770         pfn -= addr >> PAGE_SHIFT;
1771         pmd = pmd_alloc(mm, pud, addr);
1772         if (!pmd)
1773                 return -ENOMEM;
1774         VM_BUG_ON(pmd_trans_huge(*pmd));
1775         do {
1776                 next = pmd_addr_end(addr, end);
1777                 err = remap_pte_range(mm, pmd, addr, next,
1778                                 pfn + (addr >> PAGE_SHIFT), prot);
1779                 if (err)
1780                         return err;
1781         } while (pmd++, addr = next, addr != end);
1782         return 0;
1783 }
1784
1785 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1786                         unsigned long addr, unsigned long end,
1787                         unsigned long pfn, pgprot_t prot)
1788 {
1789         pud_t *pud;
1790         unsigned long next;
1791         int err;
1792
1793         pfn -= addr >> PAGE_SHIFT;
1794         pud = pud_alloc(mm, p4d, addr);
1795         if (!pud)
1796                 return -ENOMEM;
1797         do {
1798                 next = pud_addr_end(addr, end);
1799                 err = remap_pmd_range(mm, pud, addr, next,
1800                                 pfn + (addr >> PAGE_SHIFT), prot);
1801                 if (err)
1802                         return err;
1803         } while (pud++, addr = next, addr != end);
1804         return 0;
1805 }
1806
1807 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1808                         unsigned long addr, unsigned long end,
1809                         unsigned long pfn, pgprot_t prot)
1810 {
1811         p4d_t *p4d;
1812         unsigned long next;
1813         int err;
1814
1815         pfn -= addr >> PAGE_SHIFT;
1816         p4d = p4d_alloc(mm, pgd, addr);
1817         if (!p4d)
1818                 return -ENOMEM;
1819         do {
1820                 next = p4d_addr_end(addr, end);
1821                 err = remap_pud_range(mm, p4d, addr, next,
1822                                 pfn + (addr >> PAGE_SHIFT), prot);
1823                 if (err)
1824                         return err;
1825         } while (p4d++, addr = next, addr != end);
1826         return 0;
1827 }
1828
1829 /**
1830  * remap_pfn_range - remap kernel memory to userspace
1831  * @vma: user vma to map to
1832  * @addr: target user address to start at
1833  * @pfn: physical address of kernel memory
1834  * @size: size of map area
1835  * @prot: page protection flags for this mapping
1836  *
1837  * Note: this is only safe if the mm semaphore is held when called.
1838  *
1839  * Return: %0 on success, negative error code otherwise.
1840  */
1841 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1842                     unsigned long pfn, unsigned long size, pgprot_t prot)
1843 {
1844         pgd_t *pgd;
1845         unsigned long next;
1846         unsigned long end = addr + PAGE_ALIGN(size);
1847         struct mm_struct *mm = vma->vm_mm;
1848         unsigned long remap_pfn = pfn;
1849         int err;
1850
1851         /*
1852          * Physically remapped pages are special. Tell the
1853          * rest of the world about it:
1854          *   VM_IO tells people not to look at these pages
1855          *      (accesses can have side effects).
1856          *   VM_PFNMAP tells the core MM that the base pages are just
1857          *      raw PFN mappings, and do not have a "struct page" associated
1858          *      with them.
1859          *   VM_DONTEXPAND
1860          *      Disable vma merging and expanding with mremap().
1861          *   VM_DONTDUMP
1862          *      Omit vma from core dump, even when VM_IO turned off.
1863          *
1864          * There's a horrible special case to handle copy-on-write
1865          * behaviour that some programs depend on. We mark the "original"
1866          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1867          * See vm_normal_page() for details.
1868          */
1869         if (is_cow_mapping(vma->vm_flags)) {
1870                 if (addr != vma->vm_start || end != vma->vm_end)
1871                         return -EINVAL;
1872                 vma->vm_pgoff = pfn;
1873         }
1874
1875         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1876         if (err)
1877                 return -EINVAL;
1878
1879         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1880
1881         BUG_ON(addr >= end);
1882         pfn -= addr >> PAGE_SHIFT;
1883         pgd = pgd_offset(mm, addr);
1884         flush_cache_range(vma, addr, end);
1885         do {
1886                 next = pgd_addr_end(addr, end);
1887                 err = remap_p4d_range(mm, pgd, addr, next,
1888                                 pfn + (addr >> PAGE_SHIFT), prot);
1889                 if (err)
1890                         break;
1891         } while (pgd++, addr = next, addr != end);
1892
1893         if (err)
1894                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1895
1896         return err;
1897 }
1898 EXPORT_SYMBOL(remap_pfn_range);
1899
1900 /**
1901  * vm_iomap_memory - remap memory to userspace
1902  * @vma: user vma to map to
1903  * @start: start of area
1904  * @len: size of area
1905  *
1906  * This is a simplified io_remap_pfn_range() for common driver use. The
1907  * driver just needs to give us the physical memory range to be mapped,
1908  * we'll figure out the rest from the vma information.
1909  *
1910  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1911  * whatever write-combining details or similar.
1912  *
1913  * Return: %0 on success, negative error code otherwise.
1914  */
1915 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1916 {
1917         unsigned long vm_len, pfn, pages;
1918
1919         /* Check that the physical memory area passed in looks valid */
1920         if (start + len < start)
1921                 return -EINVAL;
1922         /*
1923          * You *really* shouldn't map things that aren't page-aligned,
1924          * but we've historically allowed it because IO memory might
1925          * just have smaller alignment.
1926          */
1927         len += start & ~PAGE_MASK;
1928         pfn = start >> PAGE_SHIFT;
1929         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1930         if (pfn + pages < pfn)
1931                 return -EINVAL;
1932
1933         /* We start the mapping 'vm_pgoff' pages into the area */
1934         if (vma->vm_pgoff > pages)
1935                 return -EINVAL;
1936         pfn += vma->vm_pgoff;
1937         pages -= vma->vm_pgoff;
1938
1939         /* Can we fit all of the mapping? */
1940         vm_len = vma->vm_end - vma->vm_start;
1941         if (vm_len >> PAGE_SHIFT > pages)
1942                 return -EINVAL;
1943
1944         /* Ok, let it rip */
1945         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1946 }
1947 EXPORT_SYMBOL(vm_iomap_memory);
1948
1949 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1950                                      unsigned long addr, unsigned long end,
1951                                      pte_fn_t fn, void *data)
1952 {
1953         pte_t *pte;
1954         int err;
1955         pgtable_t token;
1956         spinlock_t *uninitialized_var(ptl);
1957
1958         pte = (mm == &init_mm) ?
1959                 pte_alloc_kernel(pmd, addr) :
1960                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1961         if (!pte)
1962                 return -ENOMEM;
1963
1964         BUG_ON(pmd_huge(*pmd));
1965
1966         arch_enter_lazy_mmu_mode();
1967
1968         token = pmd_pgtable(*pmd);
1969
1970         do {
1971                 err = fn(pte++, token, addr, data);
1972                 if (err)
1973                         break;
1974         } while (addr += PAGE_SIZE, addr != end);
1975
1976         arch_leave_lazy_mmu_mode();
1977
1978         if (mm != &init_mm)
1979                 pte_unmap_unlock(pte-1, ptl);
1980         return err;
1981 }
1982
1983 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1984                                      unsigned long addr, unsigned long end,
1985                                      pte_fn_t fn, void *data)
1986 {
1987         pmd_t *pmd;
1988         unsigned long next;
1989         int err;
1990
1991         BUG_ON(pud_huge(*pud));
1992
1993         pmd = pmd_alloc(mm, pud, addr);
1994         if (!pmd)
1995                 return -ENOMEM;
1996         do {
1997                 next = pmd_addr_end(addr, end);
1998                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1999                 if (err)
2000                         break;
2001         } while (pmd++, addr = next, addr != end);
2002         return err;
2003 }
2004
2005 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2006                                      unsigned long addr, unsigned long end,
2007                                      pte_fn_t fn, void *data)
2008 {
2009         pud_t *pud;
2010         unsigned long next;
2011         int err;
2012
2013         pud = pud_alloc(mm, p4d, addr);
2014         if (!pud)
2015                 return -ENOMEM;
2016         do {
2017                 next = pud_addr_end(addr, end);
2018                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2019                 if (err)
2020                         break;
2021         } while (pud++, addr = next, addr != end);
2022         return err;
2023 }
2024
2025 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2026                                      unsigned long addr, unsigned long end,
2027                                      pte_fn_t fn, void *data)
2028 {
2029         p4d_t *p4d;
2030         unsigned long next;
2031         int err;
2032
2033         p4d = p4d_alloc(mm, pgd, addr);
2034         if (!p4d)
2035                 return -ENOMEM;
2036         do {
2037                 next = p4d_addr_end(addr, end);
2038                 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2039                 if (err)
2040                         break;
2041         } while (p4d++, addr = next, addr != end);
2042         return err;
2043 }
2044
2045 /*
2046  * Scan a region of virtual memory, filling in page tables as necessary
2047  * and calling a provided function on each leaf page table.
2048  */
2049 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2050                         unsigned long size, pte_fn_t fn, void *data)
2051 {
2052         pgd_t *pgd;
2053         unsigned long next;
2054         unsigned long end = addr + size;
2055         int err;
2056
2057         if (WARN_ON(addr >= end))
2058                 return -EINVAL;
2059
2060         pgd = pgd_offset(mm, addr);
2061         do {
2062                 next = pgd_addr_end(addr, end);
2063                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2064                 if (err)
2065                         break;
2066         } while (pgd++, addr = next, addr != end);
2067
2068         return err;
2069 }
2070 EXPORT_SYMBOL_GPL(apply_to_page_range);
2071
2072 /*
2073  * handle_pte_fault chooses page fault handler according to an entry which was
2074  * read non-atomically.  Before making any commitment, on those architectures
2075  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2076  * parts, do_swap_page must check under lock before unmapping the pte and
2077  * proceeding (but do_wp_page is only called after already making such a check;
2078  * and do_anonymous_page can safely check later on).
2079  */
2080 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2081                                 pte_t *page_table, pte_t orig_pte)
2082 {
2083         int same = 1;
2084 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2085         if (sizeof(pte_t) > sizeof(unsigned long)) {
2086                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2087                 spin_lock(ptl);
2088                 same = pte_same(*page_table, orig_pte);
2089                 spin_unlock(ptl);
2090         }
2091 #endif
2092         pte_unmap(page_table);
2093         return same;
2094 }
2095
2096 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2097 {
2098         debug_dma_assert_idle(src);
2099
2100         /*
2101          * If the source page was a PFN mapping, we don't have
2102          * a "struct page" for it. We do a best-effort copy by
2103          * just copying from the original user address. If that
2104          * fails, we just zero-fill it. Live with it.
2105          */
2106         if (unlikely(!src)) {
2107                 void *kaddr = kmap_atomic(dst);
2108                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2109
2110                 /*
2111                  * This really shouldn't fail, because the page is there
2112                  * in the page tables. But it might just be unreadable,
2113                  * in which case we just give up and fill the result with
2114                  * zeroes.
2115                  */
2116                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2117                         clear_page(kaddr);
2118                 kunmap_atomic(kaddr);
2119                 flush_dcache_page(dst);
2120         } else
2121                 copy_user_highpage(dst, src, va, vma);
2122 }
2123
2124 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2125 {
2126         struct file *vm_file = vma->vm_file;
2127
2128         if (vm_file)
2129                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2130
2131         /*
2132          * Special mappings (e.g. VDSO) do not have any file so fake
2133          * a default GFP_KERNEL for them.
2134          */
2135         return GFP_KERNEL;
2136 }
2137
2138 /*
2139  * Notify the address space that the page is about to become writable so that
2140  * it can prohibit this or wait for the page to get into an appropriate state.
2141  *
2142  * We do this without the lock held, so that it can sleep if it needs to.
2143  */
2144 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2145 {
2146         vm_fault_t ret;
2147         struct page *page = vmf->page;
2148         unsigned int old_flags = vmf->flags;
2149
2150         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2151
2152         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2153         /* Restore original flags so that caller is not surprised */
2154         vmf->flags = old_flags;
2155         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2156                 return ret;
2157         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2158                 lock_page(page);
2159                 if (!page->mapping) {
2160                         unlock_page(page);
2161                         return 0; /* retry */
2162                 }
2163                 ret |= VM_FAULT_LOCKED;
2164         } else
2165                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2166         return ret;
2167 }
2168
2169 /*
2170  * Handle dirtying of a page in shared file mapping on a write fault.
2171  *
2172  * The function expects the page to be locked and unlocks it.
2173  */
2174 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2175                                     struct page *page)
2176 {
2177         struct address_space *mapping;
2178         bool dirtied;
2179         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2180
2181         dirtied = set_page_dirty(page);
2182         VM_BUG_ON_PAGE(PageAnon(page), page);
2183         /*
2184          * Take a local copy of the address_space - page.mapping may be zeroed
2185          * by truncate after unlock_page().   The address_space itself remains
2186          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2187          * release semantics to prevent the compiler from undoing this copying.
2188          */
2189         mapping = page_rmapping(page);
2190         unlock_page(page);
2191
2192         if ((dirtied || page_mkwrite) && mapping) {
2193                 /*
2194                  * Some device drivers do not set page.mapping
2195                  * but still dirty their pages
2196                  */
2197                 balance_dirty_pages_ratelimited(mapping);
2198         }
2199
2200         if (!page_mkwrite)
2201                 file_update_time(vma->vm_file);
2202 }
2203
2204 /*
2205  * Handle write page faults for pages that can be reused in the current vma
2206  *
2207  * This can happen either due to the mapping being with the VM_SHARED flag,
2208  * or due to us being the last reference standing to the page. In either
2209  * case, all we need to do here is to mark the page as writable and update
2210  * any related book-keeping.
2211  */
2212 static inline void wp_page_reuse(struct vm_fault *vmf)
2213         __releases(vmf->ptl)
2214 {
2215         struct vm_area_struct *vma = vmf->vma;
2216         struct page *page = vmf->page;
2217         pte_t entry;
2218         /*
2219          * Clear the pages cpupid information as the existing
2220          * information potentially belongs to a now completely
2221          * unrelated process.
2222          */
2223         if (page)
2224                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2225
2226         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2227         entry = pte_mkyoung(vmf->orig_pte);
2228         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2229         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2230                 update_mmu_cache(vma, vmf->address, vmf->pte);
2231         pte_unmap_unlock(vmf->pte, vmf->ptl);
2232 }
2233
2234 /*
2235  * Handle the case of a page which we actually need to copy to a new page.
2236  *
2237  * Called with mmap_sem locked and the old page referenced, but
2238  * without the ptl held.
2239  *
2240  * High level logic flow:
2241  *
2242  * - Allocate a page, copy the content of the old page to the new one.
2243  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2244  * - Take the PTL. If the pte changed, bail out and release the allocated page
2245  * - If the pte is still the way we remember it, update the page table and all
2246  *   relevant references. This includes dropping the reference the page-table
2247  *   held to the old page, as well as updating the rmap.
2248  * - In any case, unlock the PTL and drop the reference we took to the old page.
2249  */
2250 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2251 {
2252         struct vm_area_struct *vma = vmf->vma;
2253         struct mm_struct *mm = vma->vm_mm;
2254         struct page *old_page = vmf->page;
2255         struct page *new_page = NULL;
2256         pte_t entry;
2257         int page_copied = 0;
2258         struct mem_cgroup *memcg;
2259         struct mmu_notifier_range range;
2260
2261         if (unlikely(anon_vma_prepare(vma)))
2262                 goto oom;
2263
2264         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2265                 new_page = alloc_zeroed_user_highpage_movable(vma,
2266                                                               vmf->address);
2267                 if (!new_page)
2268                         goto oom;
2269         } else {
2270                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2271                                 vmf->address);
2272                 if (!new_page)
2273                         goto oom;
2274                 cow_user_page(new_page, old_page, vmf->address, vma);
2275         }
2276
2277         if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2278                 goto oom_free_new;
2279
2280         __SetPageUptodate(new_page);
2281
2282         mmu_notifier_range_init(&range, mm, vmf->address & PAGE_MASK,
2283                                 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2284         mmu_notifier_invalidate_range_start(&range);
2285
2286         /*
2287          * Re-check the pte - we dropped the lock
2288          */
2289         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2290         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2291                 if (old_page) {
2292                         if (!PageAnon(old_page)) {
2293                                 dec_mm_counter_fast(mm,
2294                                                 mm_counter_file(old_page));
2295                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2296                         }
2297                 } else {
2298                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2299                 }
2300                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2301                 entry = mk_pte(new_page, vma->vm_page_prot);
2302                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2303                 /*
2304                  * Clear the pte entry and flush it first, before updating the
2305                  * pte with the new entry. This will avoid a race condition
2306                  * seen in the presence of one thread doing SMC and another
2307                  * thread doing COW.
2308                  */
2309                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2310                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2311                 mem_cgroup_commit_charge(new_page, memcg, false, false);
2312                 lru_cache_add_active_or_unevictable(new_page, vma);
2313                 /*
2314                  * We call the notify macro here because, when using secondary
2315                  * mmu page tables (such as kvm shadow page tables), we want the
2316                  * new page to be mapped directly into the secondary page table.
2317                  */
2318                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2319                 update_mmu_cache(vma, vmf->address, vmf->pte);
2320                 if (old_page) {
2321                         /*
2322                          * Only after switching the pte to the new page may
2323                          * we remove the mapcount here. Otherwise another
2324                          * process may come and find the rmap count decremented
2325                          * before the pte is switched to the new page, and
2326                          * "reuse" the old page writing into it while our pte
2327                          * here still points into it and can be read by other
2328                          * threads.
2329                          *
2330                          * The critical issue is to order this
2331                          * page_remove_rmap with the ptp_clear_flush above.
2332                          * Those stores are ordered by (if nothing else,)
2333                          * the barrier present in the atomic_add_negative
2334                          * in page_remove_rmap.
2335                          *
2336                          * Then the TLB flush in ptep_clear_flush ensures that
2337                          * no process can access the old page before the
2338                          * decremented mapcount is visible. And the old page
2339                          * cannot be reused until after the decremented
2340                          * mapcount is visible. So transitively, TLBs to
2341                          * old page will be flushed before it can be reused.
2342                          */
2343                         page_remove_rmap(old_page, false);
2344                 }
2345
2346                 /* Free the old page.. */
2347                 new_page = old_page;
2348                 page_copied = 1;
2349         } else {
2350                 mem_cgroup_cancel_charge(new_page, memcg, false);
2351         }
2352
2353         if (new_page)
2354                 put_page(new_page);
2355
2356         pte_unmap_unlock(vmf->pte, vmf->ptl);
2357         /*
2358          * No need to double call mmu_notifier->invalidate_range() callback as
2359          * the above ptep_clear_flush_notify() did already call it.
2360          */
2361         mmu_notifier_invalidate_range_only_end(&range);
2362         if (old_page) {
2363                 /*
2364                  * Don't let another task, with possibly unlocked vma,
2365                  * keep the mlocked page.
2366                  */
2367                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2368                         lock_page(old_page);    /* LRU manipulation */
2369                         if (PageMlocked(old_page))
2370                                 munlock_vma_page(old_page);
2371                         unlock_page(old_page);
2372                 }
2373                 put_page(old_page);
2374         }
2375         return page_copied ? VM_FAULT_WRITE : 0;
2376 oom_free_new:
2377         put_page(new_page);
2378 oom:
2379         if (old_page)
2380                 put_page(old_page);
2381         return VM_FAULT_OOM;
2382 }
2383
2384 /**
2385  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2386  *                        writeable once the page is prepared
2387  *
2388  * @vmf: structure describing the fault
2389  *
2390  * This function handles all that is needed to finish a write page fault in a
2391  * shared mapping due to PTE being read-only once the mapped page is prepared.
2392  * It handles locking of PTE and modifying it.
2393  *
2394  * The function expects the page to be locked or other protection against
2395  * concurrent faults / writeback (such as DAX radix tree locks).
2396  *
2397  * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2398  * we acquired PTE lock.
2399  */
2400 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2401 {
2402         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2403         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2404                                        &vmf->ptl);
2405         /*
2406          * We might have raced with another page fault while we released the
2407          * pte_offset_map_lock.
2408          */
2409         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2410                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2411                 return VM_FAULT_NOPAGE;
2412         }
2413         wp_page_reuse(vmf);
2414         return 0;
2415 }
2416
2417 /*
2418  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2419  * mapping
2420  */
2421 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2422 {
2423         struct vm_area_struct *vma = vmf->vma;
2424
2425         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2426                 vm_fault_t ret;
2427
2428                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2429                 vmf->flags |= FAULT_FLAG_MKWRITE;
2430                 ret = vma->vm_ops->pfn_mkwrite(vmf);
2431                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2432                         return ret;
2433                 return finish_mkwrite_fault(vmf);
2434         }
2435         wp_page_reuse(vmf);
2436         return VM_FAULT_WRITE;
2437 }
2438
2439 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2440         __releases(vmf->ptl)
2441 {
2442         struct vm_area_struct *vma = vmf->vma;
2443
2444         get_page(vmf->page);
2445
2446         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2447                 vm_fault_t tmp;
2448
2449                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2450                 tmp = do_page_mkwrite(vmf);
2451                 if (unlikely(!tmp || (tmp &
2452                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2453                         put_page(vmf->page);
2454                         return tmp;
2455                 }
2456                 tmp = finish_mkwrite_fault(vmf);
2457                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2458                         unlock_page(vmf->page);
2459                         put_page(vmf->page);
2460                         return tmp;
2461                 }
2462         } else {
2463                 wp_page_reuse(vmf);
2464                 lock_page(vmf->page);
2465         }
2466         fault_dirty_shared_page(vma, vmf->page);
2467         put_page(vmf->page);
2468
2469         return VM_FAULT_WRITE;
2470 }
2471
2472 /*
2473  * This routine handles present pages, when users try to write
2474  * to a shared page. It is done by copying the page to a new address
2475  * and decrementing the shared-page counter for the old page.
2476  *
2477  * Note that this routine assumes that the protection checks have been
2478  * done by the caller (the low-level page fault routine in most cases).
2479  * Thus we can safely just mark it writable once we've done any necessary
2480  * COW.
2481  *
2482  * We also mark the page dirty at this point even though the page will
2483  * change only once the write actually happens. This avoids a few races,
2484  * and potentially makes it more efficient.
2485  *
2486  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2487  * but allow concurrent faults), with pte both mapped and locked.
2488  * We return with mmap_sem still held, but pte unmapped and unlocked.
2489  */
2490 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2491         __releases(vmf->ptl)
2492 {
2493         struct vm_area_struct *vma = vmf->vma;
2494
2495         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2496         if (!vmf->page) {
2497                 /*
2498                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2499                  * VM_PFNMAP VMA.
2500                  *
2501                  * We should not cow pages in a shared writeable mapping.
2502                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
2503                  */
2504                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2505                                      (VM_WRITE|VM_SHARED))
2506                         return wp_pfn_shared(vmf);
2507
2508                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2509                 return wp_page_copy(vmf);
2510         }
2511
2512         /*
2513          * Take out anonymous pages first, anonymous shared vmas are
2514          * not dirty accountable.
2515          */
2516         if (PageAnon(vmf->page)) {
2517                 int total_map_swapcount;
2518                 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2519                                            page_count(vmf->page) != 1))
2520                         goto copy;
2521                 if (!trylock_page(vmf->page)) {
2522                         get_page(vmf->page);
2523                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2524                         lock_page(vmf->page);
2525                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2526                                         vmf->address, &vmf->ptl);
2527                         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2528                                 unlock_page(vmf->page);
2529                                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2530                                 put_page(vmf->page);
2531                                 return 0;
2532                         }
2533                         put_page(vmf->page);
2534                 }
2535                 if (PageKsm(vmf->page)) {
2536                         bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2537                                                      vmf->address);
2538                         unlock_page(vmf->page);
2539                         if (!reused)
2540                                 goto copy;
2541                         wp_page_reuse(vmf);
2542                         return VM_FAULT_WRITE;
2543                 }
2544                 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2545                         if (total_map_swapcount == 1) {
2546                                 /*
2547                                  * The page is all ours. Move it to
2548                                  * our anon_vma so the rmap code will
2549                                  * not search our parent or siblings.
2550                                  * Protected against the rmap code by
2551                                  * the page lock.
2552                                  */
2553                                 page_move_anon_rmap(vmf->page, vma);
2554                         }
2555                         unlock_page(vmf->page);
2556                         wp_page_reuse(vmf);
2557                         return VM_FAULT_WRITE;
2558                 }
2559                 unlock_page(vmf->page);
2560         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2561                                         (VM_WRITE|VM_SHARED))) {
2562                 return wp_page_shared(vmf);
2563         }
2564 copy:
2565         /*
2566          * Ok, we need to copy. Oh, well..
2567          */
2568         get_page(vmf->page);
2569
2570         pte_unmap_unlock(vmf->pte, vmf->ptl);
2571         return wp_page_copy(vmf);
2572 }
2573
2574 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2575                 unsigned long start_addr, unsigned long end_addr,
2576                 struct zap_details *details)
2577 {
2578         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2579 }
2580
2581 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2582                                             struct zap_details *details)
2583 {
2584         struct vm_area_struct *vma;
2585         pgoff_t vba, vea, zba, zea;
2586
2587         vma_interval_tree_foreach(vma, root,
2588                         details->first_index, details->last_index) {
2589
2590                 vba = vma->vm_pgoff;
2591                 vea = vba + vma_pages(vma) - 1;
2592                 zba = details->first_index;
2593                 if (zba < vba)
2594                         zba = vba;
2595                 zea = details->last_index;
2596                 if (zea > vea)
2597                         zea = vea;
2598
2599                 unmap_mapping_range_vma(vma,
2600                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2601                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2602                                 details);
2603         }
2604 }
2605
2606 /**
2607  * unmap_mapping_pages() - Unmap pages from processes.
2608  * @mapping: The address space containing pages to be unmapped.
2609  * @start: Index of first page to be unmapped.
2610  * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
2611  * @even_cows: Whether to unmap even private COWed pages.
2612  *
2613  * Unmap the pages in this address space from any userspace process which
2614  * has them mmaped.  Generally, you want to remove COWed pages as well when
2615  * a file is being truncated, but not when invalidating pages from the page
2616  * cache.
2617  */
2618 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2619                 pgoff_t nr, bool even_cows)
2620 {
2621         struct zap_details details = { };
2622
2623         details.check_mapping = even_cows ? NULL : mapping;
2624         details.first_index = start;
2625         details.last_index = start + nr - 1;
2626         if (details.last_index < details.first_index)
2627                 details.last_index = ULONG_MAX;
2628
2629         i_mmap_lock_write(mapping);
2630         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2631                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2632         i_mmap_unlock_write(mapping);
2633 }
2634
2635 /**
2636  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2637  * address_space corresponding to the specified byte range in the underlying
2638  * file.
2639  *
2640  * @mapping: the address space containing mmaps to be unmapped.
2641  * @holebegin: byte in first page to unmap, relative to the start of
2642  * the underlying file.  This will be rounded down to a PAGE_SIZE
2643  * boundary.  Note that this is different from truncate_pagecache(), which
2644  * must keep the partial page.  In contrast, we must get rid of
2645  * partial pages.
2646  * @holelen: size of prospective hole in bytes.  This will be rounded
2647  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2648  * end of the file.
2649  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2650  * but 0 when invalidating pagecache, don't throw away private data.
2651  */
2652 void unmap_mapping_range(struct address_space *mapping,
2653                 loff_t const holebegin, loff_t const holelen, int even_cows)
2654 {
2655         pgoff_t hba = holebegin >> PAGE_SHIFT;
2656         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2657
2658         /* Check for overflow. */
2659         if (sizeof(holelen) > sizeof(hlen)) {
2660                 long long holeend =
2661                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2662                 if (holeend & ~(long long)ULONG_MAX)
2663                         hlen = ULONG_MAX - hba + 1;
2664         }
2665
2666         unmap_mapping_pages(mapping, hba, hlen, even_cows);
2667 }
2668 EXPORT_SYMBOL(unmap_mapping_range);
2669
2670 /*
2671  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2672  * but allow concurrent faults), and pte mapped but not yet locked.
2673  * We return with pte unmapped and unlocked.
2674  *
2675  * We return with the mmap_sem locked or unlocked in the same cases
2676  * as does filemap_fault().
2677  */
2678 vm_fault_t do_swap_page(struct vm_fault *vmf)
2679 {
2680         struct vm_area_struct *vma = vmf->vma;
2681         struct page *page = NULL, *swapcache;
2682         struct mem_cgroup *memcg;
2683         swp_entry_t entry;
2684         pte_t pte;
2685         int locked;
2686         int exclusive = 0;
2687         vm_fault_t ret = 0;
2688
2689         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2690                 goto out;
2691
2692         entry = pte_to_swp_entry(vmf->orig_pte);
2693         if (unlikely(non_swap_entry(entry))) {
2694                 if (is_migration_entry(entry)) {
2695                         migration_entry_wait(vma->vm_mm, vmf->pmd,
2696                                              vmf->address);
2697                 } else if (is_device_private_entry(entry)) {
2698                         /*
2699                          * For un-addressable device memory we call the pgmap
2700                          * fault handler callback. The callback must migrate
2701                          * the page back to some CPU accessible page.
2702                          */
2703                         ret = device_private_entry_fault(vma, vmf->address, entry,
2704                                                  vmf->flags, vmf->pmd);
2705                 } else if (is_hwpoison_entry(entry)) {
2706                         ret = VM_FAULT_HWPOISON;
2707                 } else {
2708                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2709                         ret = VM_FAULT_SIGBUS;
2710                 }
2711                 goto out;
2712         }
2713
2714
2715         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2716         page = lookup_swap_cache(entry, vma, vmf->address);
2717         swapcache = page;
2718
2719         if (!page) {
2720                 struct swap_info_struct *si = swp_swap_info(entry);
2721
2722                 if (si->flags & SWP_SYNCHRONOUS_IO &&
2723                                 __swap_count(si, entry) == 1) {
2724                         /* skip swapcache */
2725                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2726                                                         vmf->address);
2727                         if (page) {
2728                                 __SetPageLocked(page);
2729                                 __SetPageSwapBacked(page);
2730                                 set_page_private(page, entry.val);
2731                                 lru_cache_add_anon(page);
2732                                 swap_readpage(page, true);
2733                         }
2734                 } else {
2735                         page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2736                                                 vmf);
2737                         swapcache = page;
2738                 }
2739
2740                 if (!page) {
2741                         /*
2742                          * Back out if somebody else faulted in this pte
2743                          * while we released the pte lock.
2744                          */
2745                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2746                                         vmf->address, &vmf->ptl);
2747                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2748                                 ret = VM_FAULT_OOM;
2749                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2750                         goto unlock;
2751                 }
2752
2753                 /* Had to read the page from swap area: Major fault */
2754                 ret = VM_FAULT_MAJOR;
2755                 count_vm_event(PGMAJFAULT);
2756                 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2757         } else if (PageHWPoison(page)) {
2758                 /*
2759                  * hwpoisoned dirty swapcache pages are kept for killing
2760                  * owner processes (which may be unknown at hwpoison time)
2761                  */
2762                 ret = VM_FAULT_HWPOISON;
2763                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2764                 goto out_release;
2765         }
2766
2767         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2768
2769         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2770         if (!locked) {
2771                 ret |= VM_FAULT_RETRY;
2772                 goto out_release;
2773         }
2774
2775         /*
2776          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2777          * release the swapcache from under us.  The page pin, and pte_same
2778          * test below, are not enough to exclude that.  Even if it is still
2779          * swapcache, we need to check that the page's swap has not changed.
2780          */
2781         if (unlikely((!PageSwapCache(page) ||
2782                         page_private(page) != entry.val)) && swapcache)
2783                 goto out_page;
2784
2785         page = ksm_might_need_to_copy(page, vma, vmf->address);
2786         if (unlikely(!page)) {
2787                 ret = VM_FAULT_OOM;
2788                 page = swapcache;
2789                 goto out_page;
2790         }
2791
2792         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2793                                         &memcg, false)) {
2794                 ret = VM_FAULT_OOM;
2795                 goto out_page;
2796         }
2797
2798         /*
2799          * Back out if somebody else already faulted in this pte.
2800          */
2801         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2802                         &vmf->ptl);
2803         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2804                 goto out_nomap;
2805
2806         if (unlikely(!PageUptodate(page))) {
2807                 ret = VM_FAULT_SIGBUS;
2808                 goto out_nomap;
2809         }
2810
2811         /*
2812          * The page isn't present yet, go ahead with the fault.
2813          *
2814          * Be careful about the sequence of operations here.
2815          * To get its accounting right, reuse_swap_page() must be called
2816          * while the page is counted on swap but not yet in mapcount i.e.
2817          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2818          * must be called after the swap_free(), or it will never succeed.
2819          */
2820
2821         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2822         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2823         pte = mk_pte(page, vma->vm_page_prot);
2824         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2825                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2826                 vmf->flags &= ~FAULT_FLAG_WRITE;
2827                 ret |= VM_FAULT_WRITE;
2828                 exclusive = RMAP_EXCLUSIVE;
2829         }
2830         flush_icache_page(vma, page);
2831         if (pte_swp_soft_dirty(vmf->orig_pte))
2832                 pte = pte_mksoft_dirty(pte);
2833         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2834         arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2835         vmf->orig_pte = pte;
2836
2837         /* ksm created a completely new copy */
2838         if (unlikely(page != swapcache && swapcache)) {
2839                 page_add_new_anon_rmap(page, vma, vmf->address, false);
2840                 mem_cgroup_commit_charge(page, memcg, false, false);
2841                 lru_cache_add_active_or_unevictable(page, vma);
2842         } else {
2843                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2844                 mem_cgroup_commit_charge(page, memcg, true, false);
2845                 activate_page(page);
2846         }
2847
2848         swap_free(entry);
2849         if (mem_cgroup_swap_full(page) ||
2850             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2851                 try_to_free_swap(page);
2852         unlock_page(page);
2853         if (page != swapcache && swapcache) {
2854                 /*
2855                  * Hold the lock to avoid the swap entry to be reused
2856                  * until we take the PT lock for the pte_same() check
2857                  * (to avoid false positives from pte_same). For
2858                  * further safety release the lock after the swap_free
2859                  * so that the swap count won't change under a
2860                  * parallel locked swapcache.
2861                  */
2862                 unlock_page(swapcache);
2863                 put_page(swapcache);
2864         }
2865
2866         if (vmf->flags & FAULT_FLAG_WRITE) {
2867                 ret |= do_wp_page(vmf);
2868                 if (ret & VM_FAULT_ERROR)
2869                         ret &= VM_FAULT_ERROR;
2870                 goto out;
2871         }
2872
2873         /* No need to invalidate - it was non-present before */
2874         update_mmu_cache(vma, vmf->address, vmf->pte);
2875 unlock:
2876         pte_unmap_unlock(vmf->pte, vmf->ptl);
2877 out:
2878         return ret;
2879 out_nomap:
2880         mem_cgroup_cancel_charge(page, memcg, false);
2881         pte_unmap_unlock(vmf->pte, vmf->ptl);
2882 out_page:
2883         unlock_page(page);
2884 out_release:
2885         put_page(page);
2886         if (page != swapcache && swapcache) {
2887                 unlock_page(swapcache);
2888                 put_page(swapcache);
2889         }
2890         return ret;
2891 }
2892
2893 /*
2894  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2895  * but allow concurrent faults), and pte mapped but not yet locked.
2896  * We return with mmap_sem still held, but pte unmapped and unlocked.
2897  */
2898 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
2899 {
2900         struct vm_area_struct *vma = vmf->vma;
2901         struct mem_cgroup *memcg;
2902         struct page *page;
2903         vm_fault_t ret = 0;
2904         pte_t entry;
2905
2906         /* File mapping without ->vm_ops ? */
2907         if (vma->vm_flags & VM_SHARED)
2908                 return VM_FAULT_SIGBUS;
2909
2910         /*
2911          * Use pte_alloc() instead of pte_alloc_map().  We can't run
2912          * pte_offset_map() on pmds where a huge pmd might be created
2913          * from a different thread.
2914          *
2915          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2916          * parallel threads are excluded by other means.
2917          *
2918          * Here we only have down_read(mmap_sem).
2919          */
2920         if (pte_alloc(vma->vm_mm, vmf->pmd))
2921                 return VM_FAULT_OOM;
2922
2923         /* See the comment in pte_alloc_one_map() */
2924         if (unlikely(pmd_trans_unstable(vmf->pmd)))
2925                 return 0;
2926
2927         /* Use the zero-page for reads */
2928         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2929                         !mm_forbids_zeropage(vma->vm_mm)) {
2930                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2931                                                 vma->vm_page_prot));
2932                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2933                                 vmf->address, &vmf->ptl);
2934                 if (!pte_none(*vmf->pte))
2935                         goto unlock;
2936                 ret = check_stable_address_space(vma->vm_mm);
2937                 if (ret)
2938                         goto unlock;
2939                 /* Deliver the page fault to userland, check inside PT lock */
2940                 if (userfaultfd_missing(vma)) {
2941                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2942                         return handle_userfault(vmf, VM_UFFD_MISSING);
2943                 }
2944                 goto setpte;
2945         }
2946
2947         /* Allocate our own private page. */
2948         if (unlikely(anon_vma_prepare(vma)))
2949                 goto oom;
2950         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2951         if (!page)
2952                 goto oom;
2953
2954         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
2955                                         false))
2956                 goto oom_free_page;
2957
2958         /*
2959          * The memory barrier inside __SetPageUptodate makes sure that
2960          * preceeding stores to the page contents become visible before
2961          * the set_pte_at() write.
2962          */
2963         __SetPageUptodate(page);
2964
2965         entry = mk_pte(page, vma->vm_page_prot);
2966         if (vma->vm_flags & VM_WRITE)
2967                 entry = pte_mkwrite(pte_mkdirty(entry));
2968
2969         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2970                         &vmf->ptl);
2971         if (!pte_none(*vmf->pte))
2972                 goto release;
2973
2974         ret = check_stable_address_space(vma->vm_mm);
2975         if (ret)
2976                 goto release;
2977
2978         /* Deliver the page fault to userland, check inside PT lock */
2979         if (userfaultfd_missing(vma)) {
2980                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2981                 mem_cgroup_cancel_charge(page, memcg, false);
2982                 put_page(page);
2983                 return handle_userfault(vmf, VM_UFFD_MISSING);
2984         }
2985
2986         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2987         page_add_new_anon_rmap(page, vma, vmf->address, false);
2988         mem_cgroup_commit_charge(page, memcg, false, false);
2989         lru_cache_add_active_or_unevictable(page, vma);
2990 setpte:
2991         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2992
2993         /* No need to invalidate - it was non-present before */
2994         update_mmu_cache(vma, vmf->address, vmf->pte);
2995 unlock:
2996         pte_unmap_unlock(vmf->pte, vmf->ptl);
2997         return ret;
2998 release:
2999         mem_cgroup_cancel_charge(page, memcg, false);
3000         put_page(page);
3001         goto unlock;
3002 oom_free_page:
3003         put_page(page);
3004 oom:
3005         return VM_FAULT_OOM;
3006 }
3007
3008 /*
3009  * The mmap_sem must have been held on entry, and may have been
3010  * released depending on flags and vma->vm_ops->fault() return value.
3011  * See filemap_fault() and __lock_page_retry().
3012  */
3013 static vm_fault_t __do_fault(struct vm_fault *vmf)
3014 {
3015         struct vm_area_struct *vma = vmf->vma;
3016         vm_fault_t ret;
3017
3018         /*
3019          * Preallocate pte before we take page_lock because this might lead to
3020          * deadlocks for memcg reclaim which waits for pages under writeback:
3021          *                              lock_page(A)
3022          *                              SetPageWriteback(A)
3023          *                              unlock_page(A)
3024          * lock_page(B)
3025          *                              lock_page(B)
3026          * pte_alloc_pne
3027          *   shrink_page_list
3028          *     wait_on_page_writeback(A)
3029          *                              SetPageWriteback(B)
3030          *                              unlock_page(B)
3031          *                              # flush A, B to clear the writeback
3032          */
3033         if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3034                 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3035                 if (!vmf->prealloc_pte)
3036                         return VM_FAULT_OOM;
3037                 smp_wmb(); /* See comment in __pte_alloc() */
3038         }
3039
3040         ret = vma->vm_ops->fault(vmf);
3041         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3042                             VM_FAULT_DONE_COW)))
3043                 return ret;
3044
3045         if (unlikely(PageHWPoison(vmf->page))) {
3046                 if (ret & VM_FAULT_LOCKED)
3047                         unlock_page(vmf->page);
3048                 put_page(vmf->page);
3049                 vmf->page = NULL;
3050                 return VM_FAULT_HWPOISON;
3051         }
3052
3053         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3054                 lock_page(vmf->page);
3055         else
3056                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3057
3058         return ret;
3059 }
3060
3061 /*
3062  * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3063  * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3064  * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3065  * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3066  */
3067 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3068 {
3069         return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3070 }
3071
3072 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3073 {
3074         struct vm_area_struct *vma = vmf->vma;
3075
3076         if (!pmd_none(*vmf->pmd))
3077                 goto map_pte;
3078         if (vmf->prealloc_pte) {
3079                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3080                 if (unlikely(!pmd_none(*vmf->pmd))) {
3081                         spin_unlock(vmf->ptl);
3082                         goto map_pte;
3083                 }
3084
3085                 mm_inc_nr_ptes(vma->vm_mm);
3086                 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3087                 spin_unlock(vmf->ptl);
3088                 vmf->prealloc_pte = NULL;
3089         } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3090                 return VM_FAULT_OOM;
3091         }
3092 map_pte:
3093         /*
3094          * If a huge pmd materialized under us just retry later.  Use
3095          * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3096          * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3097          * under us and then back to pmd_none, as a result of MADV_DONTNEED
3098          * running immediately after a huge pmd fault in a different thread of
3099          * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3100          * All we have to ensure is that it is a regular pmd that we can walk
3101          * with pte_offset_map() and we can do that through an atomic read in
3102          * C, which is what pmd_trans_unstable() provides.
3103          */
3104         if (pmd_devmap_trans_unstable(vmf->pmd))
3105                 return VM_FAULT_NOPAGE;
3106
3107         /*
3108          * At this point we know that our vmf->pmd points to a page of ptes
3109          * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3110          * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3111          * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3112          * be valid and we will re-check to make sure the vmf->pte isn't
3113          * pte_none() under vmf->ptl protection when we return to
3114          * alloc_set_pte().
3115          */
3116         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3117                         &vmf->ptl);
3118         return 0;
3119 }
3120
3121 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3122
3123 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3124 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3125                 unsigned long haddr)
3126 {
3127         if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3128                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3129                 return false;
3130         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3131                 return false;
3132         return true;
3133 }
3134
3135 static void deposit_prealloc_pte(struct vm_fault *vmf)
3136 {
3137         struct vm_area_struct *vma = vmf->vma;
3138
3139         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3140         /*
3141          * We are going to consume the prealloc table,
3142          * count that as nr_ptes.
3143          */
3144         mm_inc_nr_ptes(vma->vm_mm);
3145         vmf->prealloc_pte = NULL;
3146 }
3147
3148 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3149 {
3150         struct vm_area_struct *vma = vmf->vma;
3151         bool write = vmf->flags & FAULT_FLAG_WRITE;
3152         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3153         pmd_t entry;
3154         int i;
3155         vm_fault_t ret;
3156
3157         if (!transhuge_vma_suitable(vma, haddr))
3158                 return VM_FAULT_FALLBACK;
3159
3160         ret = VM_FAULT_FALLBACK;
3161         page = compound_head(page);
3162
3163         /*
3164          * Archs like ppc64 need additonal space to store information
3165          * related to pte entry. Use the preallocated table for that.
3166          */
3167         if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3168                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3169                 if (!vmf->prealloc_pte)
3170                         return VM_FAULT_OOM;
3171                 smp_wmb(); /* See comment in __pte_alloc() */
3172         }
3173
3174         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3175         if (unlikely(!pmd_none(*vmf->pmd)))
3176                 goto out;
3177
3178         for (i = 0; i < HPAGE_PMD_NR; i++)
3179                 flush_icache_page(vma, page + i);
3180
3181         entry = mk_huge_pmd(page, vma->vm_page_prot);
3182         if (write)
3183                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3184
3185         add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3186         page_add_file_rmap(page, true);
3187         /*
3188          * deposit and withdraw with pmd lock held
3189          */
3190         if (arch_needs_pgtable_deposit())
3191                 deposit_prealloc_pte(vmf);
3192
3193         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3194
3195         update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3196
3197         /* fault is handled */
3198         ret = 0;
3199         count_vm_event(THP_FILE_MAPPED);
3200 out:
3201         spin_unlock(vmf->ptl);
3202         return ret;
3203 }
3204 #else
3205 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3206 {
3207         BUILD_BUG();
3208         return 0;
3209 }
3210 #endif
3211
3212 /**
3213  * alloc_set_pte - setup new PTE entry for given page and add reverse page
3214  * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3215  *
3216  * @vmf: fault environment
3217  * @memcg: memcg to charge page (only for private mappings)
3218  * @page: page to map
3219  *
3220  * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3221  * return.
3222  *
3223  * Target users are page handler itself and implementations of
3224  * vm_ops->map_pages.
3225  *
3226  * Return: %0 on success, %VM_FAULT_ code in case of error.
3227  */
3228 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3229                 struct page *page)
3230 {
3231         struct vm_area_struct *vma = vmf->vma;
3232         bool write = vmf->flags & FAULT_FLAG_WRITE;
3233         pte_t entry;
3234         vm_fault_t ret;
3235
3236         if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3237                         IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3238                 /* THP on COW? */
3239                 VM_BUG_ON_PAGE(memcg, page);
3240
3241                 ret = do_set_pmd(vmf, page);
3242                 if (ret != VM_FAULT_FALLBACK)
3243                         return ret;
3244         }
3245
3246         if (!vmf->pte) {
3247                 ret = pte_alloc_one_map(vmf);
3248                 if (ret)
3249                         return ret;
3250         }
3251
3252         /* Re-check under ptl */
3253         if (unlikely(!pte_none(*vmf->pte)))
3254                 return VM_FAULT_NOPAGE;
3255
3256         flush_icache_page(vma, page);
3257         entry = mk_pte(page, vma->vm_page_prot);
3258         if (write)
3259                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3260         /* copy-on-write page */
3261         if (write && !(vma->vm_flags & VM_SHARED)) {
3262                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3263                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3264                 mem_cgroup_commit_charge(page, memcg, false, false);
3265                 lru_cache_add_active_or_unevictable(page, vma);
3266         } else {
3267                 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3268                 page_add_file_rmap(page, false);
3269         }
3270         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3271
3272         /* no need to invalidate: a not-present page won't be cached */
3273         update_mmu_cache(vma, vmf->address, vmf->pte);
3274
3275         return 0;
3276 }
3277
3278
3279 /**
3280  * finish_fault - finish page fault once we have prepared the page to fault
3281  *
3282  * @vmf: structure describing the fault
3283  *
3284  * This function handles all that is needed to finish a page fault once the
3285  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3286  * given page, adds reverse page mapping, handles memcg charges and LRU
3287  * addition.
3288  *
3289  * The function expects the page to be locked and on success it consumes a
3290  * reference of a page being mapped (for the PTE which maps it).
3291  *
3292  * Return: %0 on success, %VM_FAULT_ code in case of error.
3293  */
3294 vm_fault_t finish_fault(struct vm_fault *vmf)
3295 {
3296         struct page *page;
3297         vm_fault_t ret = 0;
3298
3299         /* Did we COW the page? */
3300         if ((vmf->flags & FAULT_FLAG_WRITE) &&
3301             !(vmf->vma->vm_flags & VM_SHARED))
3302                 page = vmf->cow_page;
3303         else
3304                 page = vmf->page;
3305
3306         /*
3307          * check even for read faults because we might have lost our CoWed
3308          * page
3309          */
3310         if (!(vmf->vma->vm_flags & VM_SHARED))
3311                 ret = check_stable_address_space(vmf->vma->vm_mm);
3312         if (!ret)
3313                 ret = alloc_set_pte(vmf, vmf->memcg, page);
3314         if (vmf->pte)
3315                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3316         return ret;
3317 }
3318
3319 static unsigned long fault_around_bytes __read_mostly =
3320         rounddown_pow_of_two(65536);
3321
3322 #ifdef CONFIG_DEBUG_FS
3323 static int fault_around_bytes_get(void *data, u64 *val)
3324 {
3325         *val = fault_around_bytes;
3326         return 0;
3327 }
3328
3329 /*
3330  * fault_around_bytes must be rounded down to the nearest page order as it's
3331  * what do_fault_around() expects to see.
3332  */
3333 static int fault_around_bytes_set(void *data, u64 val)
3334 {
3335         if (val / PAGE_SIZE > PTRS_PER_PTE)
3336                 return -EINVAL;
3337         if (val > PAGE_SIZE)
3338                 fault_around_bytes = rounddown_pow_of_two(val);
3339         else
3340                 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3341         return 0;
3342 }
3343 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3344                 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3345
3346 static int __init fault_around_debugfs(void)
3347 {
3348         debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3349                                    &fault_around_bytes_fops);
3350         return 0;
3351 }
3352 late_initcall(fault_around_debugfs);
3353 #endif
3354
3355 /*
3356  * do_fault_around() tries to map few pages around the fault address. The hope
3357  * is that the pages will be needed soon and this will lower the number of
3358  * faults to handle.
3359  *
3360  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3361  * not ready to be mapped: not up-to-date, locked, etc.
3362  *
3363  * This function is called with the page table lock taken. In the split ptlock
3364  * case the page table lock only protects only those entries which belong to
3365  * the page table corresponding to the fault address.
3366  *
3367  * This function doesn't cross the VMA boundaries, in order to call map_pages()
3368  * only once.
3369  *
3370  * fault_around_bytes defines how many bytes we'll try to map.
3371  * do_fault_around() expects it to be set to a power of two less than or equal
3372  * to PTRS_PER_PTE.
3373  *
3374  * The virtual address of the area that we map is naturally aligned to
3375  * fault_around_bytes rounded down to the machine page size
3376  * (and therefore to page order).  This way it's easier to guarantee
3377  * that we don't cross page table boundaries.
3378  */
3379 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3380 {
3381         unsigned long address = vmf->address, nr_pages, mask;
3382         pgoff_t start_pgoff = vmf->pgoff;
3383         pgoff_t end_pgoff;
3384         int off;
3385         vm_fault_t ret = 0;
3386
3387         nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3388         mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3389
3390         vmf->address = max(address & mask, vmf->vma->vm_start);
3391         off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3392         start_pgoff -= off;
3393
3394         /*
3395          *  end_pgoff is either the end of the page table, the end of
3396          *  the vma or nr_pages from start_pgoff, depending what is nearest.
3397          */
3398         end_pgoff = start_pgoff -
3399                 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3400                 PTRS_PER_PTE - 1;
3401         end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3402                         start_pgoff + nr_pages - 1);
3403
3404         if (pmd_none(*vmf->pmd)) {
3405                 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3406                 if (!vmf->prealloc_pte)
3407                         goto out;
3408                 smp_wmb(); /* See comment in __pte_alloc() */
3409         }
3410
3411         vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3412
3413         /* Huge page is mapped? Page fault is solved */
3414         if (pmd_trans_huge(*vmf->pmd)) {
3415                 ret = VM_FAULT_NOPAGE;
3416                 goto out;
3417         }
3418
3419         /* ->map_pages() haven't done anything useful. Cold page cache? */
3420         if (!vmf->pte)
3421                 goto out;
3422
3423         /* check if the page fault is solved */
3424         vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3425         if (!pte_none(*vmf->pte))
3426                 ret = VM_FAULT_NOPAGE;
3427         pte_unmap_unlock(vmf->pte, vmf->ptl);
3428 out:
3429         vmf->address = address;
3430         vmf->pte = NULL;
3431         return ret;
3432 }
3433
3434 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3435 {
3436         struct vm_area_struct *vma = vmf->vma;
3437         vm_fault_t ret = 0;
3438
3439         /*
3440          * Let's call ->map_pages() first and use ->fault() as fallback
3441          * if page by the offset is not ready to be mapped (cold cache or
3442          * something).
3443          */
3444         if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3445                 ret = do_fault_around(vmf);
3446                 if (ret)
3447                         return ret;
3448         }
3449
3450         ret = __do_fault(vmf);
3451         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3452                 return ret;
3453
3454         ret |= finish_fault(vmf);
3455         unlock_page(vmf->page);
3456         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3457                 put_page(vmf->page);
3458         return ret;
3459 }
3460
3461 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3462 {
3463         struct vm_area_struct *vma = vmf->vma;
3464         vm_fault_t ret;
3465
3466         if (unlikely(anon_vma_prepare(vma)))
3467                 return VM_FAULT_OOM;
3468
3469         vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3470         if (!vmf->