VM, x86, PAT: Change is_linear_pfn_mapping to not use vm_pgoff
[linux-2.6.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/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/delayacct.h>
51 #include <linux/init.h>
52 #include <linux/writeback.h>
53 #include <linux/memcontrol.h>
54 #include <linux/mmu_notifier.h>
55 #include <linux/kallsyms.h>
56 #include <linux/swapops.h>
57 #include <linux/elf.h>
58
59 #include <asm/pgalloc.h>
60 #include <asm/uaccess.h>
61 #include <asm/tlb.h>
62 #include <asm/tlbflush.h>
63 #include <asm/pgtable.h>
64
65 #include "internal.h"
66
67 #ifndef CONFIG_NEED_MULTIPLE_NODES
68 /* use the per-pgdat data instead for discontigmem - mbligh */
69 unsigned long max_mapnr;
70 struct page *mem_map;
71
72 EXPORT_SYMBOL(max_mapnr);
73 EXPORT_SYMBOL(mem_map);
74 #endif
75
76 unsigned long num_physpages;
77 /*
78  * A number of key systems in x86 including ioremap() rely on the assumption
79  * that high_memory defines the upper bound on direct map memory, then end
80  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
81  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
82  * and ZONE_HIGHMEM.
83  */
84 void * high_memory;
85
86 EXPORT_SYMBOL(num_physpages);
87 EXPORT_SYMBOL(high_memory);
88
89 /*
90  * Randomize the address space (stacks, mmaps, brk, etc.).
91  *
92  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
93  *   as ancient (libc5 based) binaries can segfault. )
94  */
95 int randomize_va_space __read_mostly =
96 #ifdef CONFIG_COMPAT_BRK
97                                         1;
98 #else
99                                         2;
100 #endif
101
102 static int __init disable_randmaps(char *s)
103 {
104         randomize_va_space = 0;
105         return 1;
106 }
107 __setup("norandmaps", disable_randmaps);
108
109
110 /*
111  * If a p?d_bad entry is found while walking page tables, report
112  * the error, before resetting entry to p?d_none.  Usually (but
113  * very seldom) called out from the p?d_none_or_clear_bad macros.
114  */
115
116 void pgd_clear_bad(pgd_t *pgd)
117 {
118         pgd_ERROR(*pgd);
119         pgd_clear(pgd);
120 }
121
122 void pud_clear_bad(pud_t *pud)
123 {
124         pud_ERROR(*pud);
125         pud_clear(pud);
126 }
127
128 void pmd_clear_bad(pmd_t *pmd)
129 {
130         pmd_ERROR(*pmd);
131         pmd_clear(pmd);
132 }
133
134 /*
135  * Note: this doesn't free the actual pages themselves. That
136  * has been handled earlier when unmapping all the memory regions.
137  */
138 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
139 {
140         pgtable_t token = pmd_pgtable(*pmd);
141         pmd_clear(pmd);
142         pte_free_tlb(tlb, token);
143         tlb->mm->nr_ptes--;
144 }
145
146 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
147                                 unsigned long addr, unsigned long end,
148                                 unsigned long floor, unsigned long ceiling)
149 {
150         pmd_t *pmd;
151         unsigned long next;
152         unsigned long start;
153
154         start = addr;
155         pmd = pmd_offset(pud, addr);
156         do {
157                 next = pmd_addr_end(addr, end);
158                 if (pmd_none_or_clear_bad(pmd))
159                         continue;
160                 free_pte_range(tlb, pmd);
161         } while (pmd++, addr = next, addr != end);
162
163         start &= PUD_MASK;
164         if (start < floor)
165                 return;
166         if (ceiling) {
167                 ceiling &= PUD_MASK;
168                 if (!ceiling)
169                         return;
170         }
171         if (end - 1 > ceiling - 1)
172                 return;
173
174         pmd = pmd_offset(pud, start);
175         pud_clear(pud);
176         pmd_free_tlb(tlb, pmd);
177 }
178
179 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
180                                 unsigned long addr, unsigned long end,
181                                 unsigned long floor, unsigned long ceiling)
182 {
183         pud_t *pud;
184         unsigned long next;
185         unsigned long start;
186
187         start = addr;
188         pud = pud_offset(pgd, addr);
189         do {
190                 next = pud_addr_end(addr, end);
191                 if (pud_none_or_clear_bad(pud))
192                         continue;
193                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
194         } while (pud++, addr = next, addr != end);
195
196         start &= PGDIR_MASK;
197         if (start < floor)
198                 return;
199         if (ceiling) {
200                 ceiling &= PGDIR_MASK;
201                 if (!ceiling)
202                         return;
203         }
204         if (end - 1 > ceiling - 1)
205                 return;
206
207         pud = pud_offset(pgd, start);
208         pgd_clear(pgd);
209         pud_free_tlb(tlb, pud);
210 }
211
212 /*
213  * This function frees user-level page tables of a process.
214  *
215  * Must be called with pagetable lock held.
216  */
217 void free_pgd_range(struct mmu_gather *tlb,
218                         unsigned long addr, unsigned long end,
219                         unsigned long floor, unsigned long ceiling)
220 {
221         pgd_t *pgd;
222         unsigned long next;
223         unsigned long start;
224
225         /*
226          * The next few lines have given us lots of grief...
227          *
228          * Why are we testing PMD* at this top level?  Because often
229          * there will be no work to do at all, and we'd prefer not to
230          * go all the way down to the bottom just to discover that.
231          *
232          * Why all these "- 1"s?  Because 0 represents both the bottom
233          * of the address space and the top of it (using -1 for the
234          * top wouldn't help much: the masks would do the wrong thing).
235          * The rule is that addr 0 and floor 0 refer to the bottom of
236          * the address space, but end 0 and ceiling 0 refer to the top
237          * Comparisons need to use "end - 1" and "ceiling - 1" (though
238          * that end 0 case should be mythical).
239          *
240          * Wherever addr is brought up or ceiling brought down, we must
241          * be careful to reject "the opposite 0" before it confuses the
242          * subsequent tests.  But what about where end is brought down
243          * by PMD_SIZE below? no, end can't go down to 0 there.
244          *
245          * Whereas we round start (addr) and ceiling down, by different
246          * masks at different levels, in order to test whether a table
247          * now has no other vmas using it, so can be freed, we don't
248          * bother to round floor or end up - the tests don't need that.
249          */
250
251         addr &= PMD_MASK;
252         if (addr < floor) {
253                 addr += PMD_SIZE;
254                 if (!addr)
255                         return;
256         }
257         if (ceiling) {
258                 ceiling &= PMD_MASK;
259                 if (!ceiling)
260                         return;
261         }
262         if (end - 1 > ceiling - 1)
263                 end -= PMD_SIZE;
264         if (addr > end - 1)
265                 return;
266
267         start = addr;
268         pgd = pgd_offset(tlb->mm, addr);
269         do {
270                 next = pgd_addr_end(addr, end);
271                 if (pgd_none_or_clear_bad(pgd))
272                         continue;
273                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
274         } while (pgd++, addr = next, addr != end);
275 }
276
277 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
278                 unsigned long floor, unsigned long ceiling)
279 {
280         while (vma) {
281                 struct vm_area_struct *next = vma->vm_next;
282                 unsigned long addr = vma->vm_start;
283
284                 /*
285                  * Hide vma from rmap and vmtruncate before freeing pgtables
286                  */
287                 anon_vma_unlink(vma);
288                 unlink_file_vma(vma);
289
290                 if (is_vm_hugetlb_page(vma)) {
291                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
292                                 floor, next? next->vm_start: ceiling);
293                 } else {
294                         /*
295                          * Optimization: gather nearby vmas into one call down
296                          */
297                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
298                                && !is_vm_hugetlb_page(next)) {
299                                 vma = next;
300                                 next = vma->vm_next;
301                                 anon_vma_unlink(vma);
302                                 unlink_file_vma(vma);
303                         }
304                         free_pgd_range(tlb, addr, vma->vm_end,
305                                 floor, next? next->vm_start: ceiling);
306                 }
307                 vma = next;
308         }
309 }
310
311 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
312 {
313         pgtable_t new = pte_alloc_one(mm, address);
314         if (!new)
315                 return -ENOMEM;
316
317         /*
318          * Ensure all pte setup (eg. pte page lock and page clearing) are
319          * visible before the pte is made visible to other CPUs by being
320          * put into page tables.
321          *
322          * The other side of the story is the pointer chasing in the page
323          * table walking code (when walking the page table without locking;
324          * ie. most of the time). Fortunately, these data accesses consist
325          * of a chain of data-dependent loads, meaning most CPUs (alpha
326          * being the notable exception) will already guarantee loads are
327          * seen in-order. See the alpha page table accessors for the
328          * smp_read_barrier_depends() barriers in page table walking code.
329          */
330         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
331
332         spin_lock(&mm->page_table_lock);
333         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
334                 mm->nr_ptes++;
335                 pmd_populate(mm, pmd, new);
336                 new = NULL;
337         }
338         spin_unlock(&mm->page_table_lock);
339         if (new)
340                 pte_free(mm, new);
341         return 0;
342 }
343
344 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
345 {
346         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
347         if (!new)
348                 return -ENOMEM;
349
350         smp_wmb(); /* See comment in __pte_alloc */
351
352         spin_lock(&init_mm.page_table_lock);
353         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
354                 pmd_populate_kernel(&init_mm, pmd, new);
355                 new = NULL;
356         }
357         spin_unlock(&init_mm.page_table_lock);
358         if (new)
359                 pte_free_kernel(&init_mm, new);
360         return 0;
361 }
362
363 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
364 {
365         if (file_rss)
366                 add_mm_counter(mm, file_rss, file_rss);
367         if (anon_rss)
368                 add_mm_counter(mm, anon_rss, anon_rss);
369 }
370
371 /*
372  * This function is called to print an error when a bad pte
373  * is found. For example, we might have a PFN-mapped pte in
374  * a region that doesn't allow it.
375  *
376  * The calling function must still handle the error.
377  */
378 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
379                           pte_t pte, struct page *page)
380 {
381         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
382         pud_t *pud = pud_offset(pgd, addr);
383         pmd_t *pmd = pmd_offset(pud, addr);
384         struct address_space *mapping;
385         pgoff_t index;
386         static unsigned long resume;
387         static unsigned long nr_shown;
388         static unsigned long nr_unshown;
389
390         /*
391          * Allow a burst of 60 reports, then keep quiet for that minute;
392          * or allow a steady drip of one report per second.
393          */
394         if (nr_shown == 60) {
395                 if (time_before(jiffies, resume)) {
396                         nr_unshown++;
397                         return;
398                 }
399                 if (nr_unshown) {
400                         printk(KERN_ALERT
401                                 "BUG: Bad page map: %lu messages suppressed\n",
402                                 nr_unshown);
403                         nr_unshown = 0;
404                 }
405                 nr_shown = 0;
406         }
407         if (nr_shown++ == 0)
408                 resume = jiffies + 60 * HZ;
409
410         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
411         index = linear_page_index(vma, addr);
412
413         printk(KERN_ALERT
414                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
415                 current->comm,
416                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
417         if (page) {
418                 printk(KERN_ALERT
419                 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
420                 page, (void *)page->flags, page_count(page),
421                 page_mapcount(page), page->mapping, page->index);
422         }
423         printk(KERN_ALERT
424                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
425                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
426         /*
427          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
428          */
429         if (vma->vm_ops)
430                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
431                                 (unsigned long)vma->vm_ops->fault);
432         if (vma->vm_file && vma->vm_file->f_op)
433                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
434                                 (unsigned long)vma->vm_file->f_op->mmap);
435         dump_stack();
436         add_taint(TAINT_BAD_PAGE);
437 }
438
439 static inline int is_cow_mapping(unsigned int flags)
440 {
441         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
442 }
443
444 /*
445  * vm_normal_page -- This function gets the "struct page" associated with a pte.
446  *
447  * "Special" mappings do not wish to be associated with a "struct page" (either
448  * it doesn't exist, or it exists but they don't want to touch it). In this
449  * case, NULL is returned here. "Normal" mappings do have a struct page.
450  *
451  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
452  * pte bit, in which case this function is trivial. Secondly, an architecture
453  * may not have a spare pte bit, which requires a more complicated scheme,
454  * described below.
455  *
456  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
457  * special mapping (even if there are underlying and valid "struct pages").
458  * COWed pages of a VM_PFNMAP are always normal.
459  *
460  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
461  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
462  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
463  * mapping will always honor the rule
464  *
465  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
466  *
467  * And for normal mappings this is false.
468  *
469  * This restricts such mappings to be a linear translation from virtual address
470  * to pfn. To get around this restriction, we allow arbitrary mappings so long
471  * as the vma is not a COW mapping; in that case, we know that all ptes are
472  * special (because none can have been COWed).
473  *
474  *
475  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
476  *
477  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
478  * page" backing, however the difference is that _all_ pages with a struct
479  * page (that is, those where pfn_valid is true) are refcounted and considered
480  * normal pages by the VM. The disadvantage is that pages are refcounted
481  * (which can be slower and simply not an option for some PFNMAP users). The
482  * advantage is that we don't have to follow the strict linearity rule of
483  * PFNMAP mappings in order to support COWable mappings.
484  *
485  */
486 #ifdef __HAVE_ARCH_PTE_SPECIAL
487 # define HAVE_PTE_SPECIAL 1
488 #else
489 # define HAVE_PTE_SPECIAL 0
490 #endif
491 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
492                                 pte_t pte)
493 {
494         unsigned long pfn = pte_pfn(pte);
495
496         if (HAVE_PTE_SPECIAL) {
497                 if (likely(!pte_special(pte)))
498                         goto check_pfn;
499                 if (!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)))
500                         print_bad_pte(vma, addr, pte, NULL);
501                 return NULL;
502         }
503
504         /* !HAVE_PTE_SPECIAL case follows: */
505
506         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
507                 if (vma->vm_flags & VM_MIXEDMAP) {
508                         if (!pfn_valid(pfn))
509                                 return NULL;
510                         goto out;
511                 } else {
512                         unsigned long off;
513                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
514                         if (pfn == vma->vm_pgoff + off)
515                                 return NULL;
516                         if (!is_cow_mapping(vma->vm_flags))
517                                 return NULL;
518                 }
519         }
520
521 check_pfn:
522         if (unlikely(pfn > highest_memmap_pfn)) {
523                 print_bad_pte(vma, addr, pte, NULL);
524                 return NULL;
525         }
526
527         /*
528          * NOTE! We still have PageReserved() pages in the page tables.
529          * eg. VDSO mappings can cause them to exist.
530          */
531 out:
532         return pfn_to_page(pfn);
533 }
534
535 /*
536  * copy one vm_area from one task to the other. Assumes the page tables
537  * already present in the new task to be cleared in the whole range
538  * covered by this vma.
539  */
540
541 static inline void
542 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
543                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
544                 unsigned long addr, int *rss)
545 {
546         unsigned long vm_flags = vma->vm_flags;
547         pte_t pte = *src_pte;
548         struct page *page;
549
550         /* pte contains position in swap or file, so copy. */
551         if (unlikely(!pte_present(pte))) {
552                 if (!pte_file(pte)) {
553                         swp_entry_t entry = pte_to_swp_entry(pte);
554
555                         swap_duplicate(entry);
556                         /* make sure dst_mm is on swapoff's mmlist. */
557                         if (unlikely(list_empty(&dst_mm->mmlist))) {
558                                 spin_lock(&mmlist_lock);
559                                 if (list_empty(&dst_mm->mmlist))
560                                         list_add(&dst_mm->mmlist,
561                                                  &src_mm->mmlist);
562                                 spin_unlock(&mmlist_lock);
563                         }
564                         if (is_write_migration_entry(entry) &&
565                                         is_cow_mapping(vm_flags)) {
566                                 /*
567                                  * COW mappings require pages in both parent
568                                  * and child to be set to read.
569                                  */
570                                 make_migration_entry_read(&entry);
571                                 pte = swp_entry_to_pte(entry);
572                                 set_pte_at(src_mm, addr, src_pte, pte);
573                         }
574                 }
575                 goto out_set_pte;
576         }
577
578         /*
579          * If it's a COW mapping, write protect it both
580          * in the parent and the child
581          */
582         if (is_cow_mapping(vm_flags)) {
583                 ptep_set_wrprotect(src_mm, addr, src_pte);
584                 pte = pte_wrprotect(pte);
585         }
586
587         /*
588          * If it's a shared mapping, mark it clean in
589          * the child
590          */
591         if (vm_flags & VM_SHARED)
592                 pte = pte_mkclean(pte);
593         pte = pte_mkold(pte);
594
595         page = vm_normal_page(vma, addr, pte);
596         if (page) {
597                 get_page(page);
598                 page_dup_rmap(page, vma, addr);
599                 rss[!!PageAnon(page)]++;
600         }
601
602 out_set_pte:
603         set_pte_at(dst_mm, addr, dst_pte, pte);
604 }
605
606 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
607                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
608                 unsigned long addr, unsigned long end)
609 {
610         pte_t *src_pte, *dst_pte;
611         spinlock_t *src_ptl, *dst_ptl;
612         int progress = 0;
613         int rss[2];
614
615 again:
616         rss[1] = rss[0] = 0;
617         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
618         if (!dst_pte)
619                 return -ENOMEM;
620         src_pte = pte_offset_map_nested(src_pmd, addr);
621         src_ptl = pte_lockptr(src_mm, src_pmd);
622         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
623         arch_enter_lazy_mmu_mode();
624
625         do {
626                 /*
627                  * We are holding two locks at this point - either of them
628                  * could generate latencies in another task on another CPU.
629                  */
630                 if (progress >= 32) {
631                         progress = 0;
632                         if (need_resched() ||
633                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
634                                 break;
635                 }
636                 if (pte_none(*src_pte)) {
637                         progress++;
638                         continue;
639                 }
640                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
641                 progress += 8;
642         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
643
644         arch_leave_lazy_mmu_mode();
645         spin_unlock(src_ptl);
646         pte_unmap_nested(src_pte - 1);
647         add_mm_rss(dst_mm, rss[0], rss[1]);
648         pte_unmap_unlock(dst_pte - 1, dst_ptl);
649         cond_resched();
650         if (addr != end)
651                 goto again;
652         return 0;
653 }
654
655 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
656                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
657                 unsigned long addr, unsigned long end)
658 {
659         pmd_t *src_pmd, *dst_pmd;
660         unsigned long next;
661
662         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
663         if (!dst_pmd)
664                 return -ENOMEM;
665         src_pmd = pmd_offset(src_pud, addr);
666         do {
667                 next = pmd_addr_end(addr, end);
668                 if (pmd_none_or_clear_bad(src_pmd))
669                         continue;
670                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
671                                                 vma, addr, next))
672                         return -ENOMEM;
673         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
674         return 0;
675 }
676
677 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
678                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
679                 unsigned long addr, unsigned long end)
680 {
681         pud_t *src_pud, *dst_pud;
682         unsigned long next;
683
684         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
685         if (!dst_pud)
686                 return -ENOMEM;
687         src_pud = pud_offset(src_pgd, addr);
688         do {
689                 next = pud_addr_end(addr, end);
690                 if (pud_none_or_clear_bad(src_pud))
691                         continue;
692                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
693                                                 vma, addr, next))
694                         return -ENOMEM;
695         } while (dst_pud++, src_pud++, addr = next, addr != end);
696         return 0;
697 }
698
699 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
700                 struct vm_area_struct *vma)
701 {
702         pgd_t *src_pgd, *dst_pgd;
703         unsigned long next;
704         unsigned long addr = vma->vm_start;
705         unsigned long end = vma->vm_end;
706         int ret;
707
708         /*
709          * Don't copy ptes where a page fault will fill them correctly.
710          * Fork becomes much lighter when there are big shared or private
711          * readonly mappings. The tradeoff is that copy_page_range is more
712          * efficient than faulting.
713          */
714         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
715                 if (!vma->anon_vma)
716                         return 0;
717         }
718
719         if (is_vm_hugetlb_page(vma))
720                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
721
722         if (unlikely(is_pfn_mapping(vma))) {
723                 /*
724                  * We do not free on error cases below as remove_vma
725                  * gets called on error from higher level routine
726                  */
727                 ret = track_pfn_vma_copy(vma);
728                 if (ret)
729                         return ret;
730         }
731
732         /*
733          * We need to invalidate the secondary MMU mappings only when
734          * there could be a permission downgrade on the ptes of the
735          * parent mm. And a permission downgrade will only happen if
736          * is_cow_mapping() returns true.
737          */
738         if (is_cow_mapping(vma->vm_flags))
739                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
740
741         ret = 0;
742         dst_pgd = pgd_offset(dst_mm, addr);
743         src_pgd = pgd_offset(src_mm, addr);
744         do {
745                 next = pgd_addr_end(addr, end);
746                 if (pgd_none_or_clear_bad(src_pgd))
747                         continue;
748                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
749                                             vma, addr, next))) {
750                         ret = -ENOMEM;
751                         break;
752                 }
753         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
754
755         if (is_cow_mapping(vma->vm_flags))
756                 mmu_notifier_invalidate_range_end(src_mm,
757                                                   vma->vm_start, end);
758         return ret;
759 }
760
761 static unsigned long zap_pte_range(struct mmu_gather *tlb,
762                                 struct vm_area_struct *vma, pmd_t *pmd,
763                                 unsigned long addr, unsigned long end,
764                                 long *zap_work, struct zap_details *details)
765 {
766         struct mm_struct *mm = tlb->mm;
767         pte_t *pte;
768         spinlock_t *ptl;
769         int file_rss = 0;
770         int anon_rss = 0;
771
772         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
773         arch_enter_lazy_mmu_mode();
774         do {
775                 pte_t ptent = *pte;
776                 if (pte_none(ptent)) {
777                         (*zap_work)--;
778                         continue;
779                 }
780
781                 (*zap_work) -= PAGE_SIZE;
782
783                 if (pte_present(ptent)) {
784                         struct page *page;
785
786                         page = vm_normal_page(vma, addr, ptent);
787                         if (unlikely(details) && page) {
788                                 /*
789                                  * unmap_shared_mapping_pages() wants to
790                                  * invalidate cache without truncating:
791                                  * unmap shared but keep private pages.
792                                  */
793                                 if (details->check_mapping &&
794                                     details->check_mapping != page->mapping)
795                                         continue;
796                                 /*
797                                  * Each page->index must be checked when
798                                  * invalidating or truncating nonlinear.
799                                  */
800                                 if (details->nonlinear_vma &&
801                                     (page->index < details->first_index ||
802                                      page->index > details->last_index))
803                                         continue;
804                         }
805                         ptent = ptep_get_and_clear_full(mm, addr, pte,
806                                                         tlb->fullmm);
807                         tlb_remove_tlb_entry(tlb, pte, addr);
808                         if (unlikely(!page))
809                                 continue;
810                         if (unlikely(details) && details->nonlinear_vma
811                             && linear_page_index(details->nonlinear_vma,
812                                                 addr) != page->index)
813                                 set_pte_at(mm, addr, pte,
814                                            pgoff_to_pte(page->index));
815                         if (PageAnon(page))
816                                 anon_rss--;
817                         else {
818                                 if (pte_dirty(ptent))
819                                         set_page_dirty(page);
820                                 if (pte_young(ptent) &&
821                                     likely(!VM_SequentialReadHint(vma)))
822                                         mark_page_accessed(page);
823                                 file_rss--;
824                         }
825                         page_remove_rmap(page);
826                         if (unlikely(page_mapcount(page) < 0))
827                                 print_bad_pte(vma, addr, ptent, page);
828                         tlb_remove_page(tlb, page);
829                         continue;
830                 }
831                 /*
832                  * If details->check_mapping, we leave swap entries;
833                  * if details->nonlinear_vma, we leave file entries.
834                  */
835                 if (unlikely(details))
836                         continue;
837                 if (pte_file(ptent)) {
838                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
839                                 print_bad_pte(vma, addr, ptent, NULL);
840                 } else if
841                   (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
842                         print_bad_pte(vma, addr, ptent, NULL);
843                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
844         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
845
846         add_mm_rss(mm, file_rss, anon_rss);
847         arch_leave_lazy_mmu_mode();
848         pte_unmap_unlock(pte - 1, ptl);
849
850         return addr;
851 }
852
853 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
854                                 struct vm_area_struct *vma, pud_t *pud,
855                                 unsigned long addr, unsigned long end,
856                                 long *zap_work, struct zap_details *details)
857 {
858         pmd_t *pmd;
859         unsigned long next;
860
861         pmd = pmd_offset(pud, addr);
862         do {
863                 next = pmd_addr_end(addr, end);
864                 if (pmd_none_or_clear_bad(pmd)) {
865                         (*zap_work)--;
866                         continue;
867                 }
868                 next = zap_pte_range(tlb, vma, pmd, addr, next,
869                                                 zap_work, details);
870         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
871
872         return addr;
873 }
874
875 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
876                                 struct vm_area_struct *vma, pgd_t *pgd,
877                                 unsigned long addr, unsigned long end,
878                                 long *zap_work, struct zap_details *details)
879 {
880         pud_t *pud;
881         unsigned long next;
882
883         pud = pud_offset(pgd, addr);
884         do {
885                 next = pud_addr_end(addr, end);
886                 if (pud_none_or_clear_bad(pud)) {
887                         (*zap_work)--;
888                         continue;
889                 }
890                 next = zap_pmd_range(tlb, vma, pud, addr, next,
891                                                 zap_work, details);
892         } while (pud++, addr = next, (addr != end && *zap_work > 0));
893
894         return addr;
895 }
896
897 static unsigned long unmap_page_range(struct mmu_gather *tlb,
898                                 struct vm_area_struct *vma,
899                                 unsigned long addr, unsigned long end,
900                                 long *zap_work, struct zap_details *details)
901 {
902         pgd_t *pgd;
903         unsigned long next;
904
905         if (details && !details->check_mapping && !details->nonlinear_vma)
906                 details = NULL;
907
908         BUG_ON(addr >= end);
909         tlb_start_vma(tlb, vma);
910         pgd = pgd_offset(vma->vm_mm, addr);
911         do {
912                 next = pgd_addr_end(addr, end);
913                 if (pgd_none_or_clear_bad(pgd)) {
914                         (*zap_work)--;
915                         continue;
916                 }
917                 next = zap_pud_range(tlb, vma, pgd, addr, next,
918                                                 zap_work, details);
919         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
920         tlb_end_vma(tlb, vma);
921
922         return addr;
923 }
924
925 #ifdef CONFIG_PREEMPT
926 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
927 #else
928 /* No preempt: go for improved straight-line efficiency */
929 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
930 #endif
931
932 /**
933  * unmap_vmas - unmap a range of memory covered by a list of vma's
934  * @tlbp: address of the caller's struct mmu_gather
935  * @vma: the starting vma
936  * @start_addr: virtual address at which to start unmapping
937  * @end_addr: virtual address at which to end unmapping
938  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
939  * @details: details of nonlinear truncation or shared cache invalidation
940  *
941  * Returns the end address of the unmapping (restart addr if interrupted).
942  *
943  * Unmap all pages in the vma list.
944  *
945  * We aim to not hold locks for too long (for scheduling latency reasons).
946  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
947  * return the ending mmu_gather to the caller.
948  *
949  * Only addresses between `start' and `end' will be unmapped.
950  *
951  * The VMA list must be sorted in ascending virtual address order.
952  *
953  * unmap_vmas() assumes that the caller will flush the whole unmapped address
954  * range after unmap_vmas() returns.  So the only responsibility here is to
955  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
956  * drops the lock and schedules.
957  */
958 unsigned long unmap_vmas(struct mmu_gather **tlbp,
959                 struct vm_area_struct *vma, unsigned long start_addr,
960                 unsigned long end_addr, unsigned long *nr_accounted,
961                 struct zap_details *details)
962 {
963         long zap_work = ZAP_BLOCK_SIZE;
964         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
965         int tlb_start_valid = 0;
966         unsigned long start = start_addr;
967         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
968         int fullmm = (*tlbp)->fullmm;
969         struct mm_struct *mm = vma->vm_mm;
970
971         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
972         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
973                 unsigned long end;
974
975                 start = max(vma->vm_start, start_addr);
976                 if (start >= vma->vm_end)
977                         continue;
978                 end = min(vma->vm_end, end_addr);
979                 if (end <= vma->vm_start)
980                         continue;
981
982                 if (vma->vm_flags & VM_ACCOUNT)
983                         *nr_accounted += (end - start) >> PAGE_SHIFT;
984
985                 if (unlikely(is_pfn_mapping(vma)))
986                         untrack_pfn_vma(vma, 0, 0);
987
988                 while (start != end) {
989                         if (!tlb_start_valid) {
990                                 tlb_start = start;
991                                 tlb_start_valid = 1;
992                         }
993
994                         if (unlikely(is_vm_hugetlb_page(vma))) {
995                                 /*
996                                  * It is undesirable to test vma->vm_file as it
997                                  * should be non-null for valid hugetlb area.
998                                  * However, vm_file will be NULL in the error
999                                  * cleanup path of do_mmap_pgoff. When
1000                                  * hugetlbfs ->mmap method fails,
1001                                  * do_mmap_pgoff() nullifies vma->vm_file
1002                                  * before calling this function to clean up.
1003                                  * Since no pte has actually been setup, it is
1004                                  * safe to do nothing in this case.
1005                                  */
1006                                 if (vma->vm_file) {
1007                                         unmap_hugepage_range(vma, start, end, NULL);
1008                                         zap_work -= (end - start) /
1009                                         pages_per_huge_page(hstate_vma(vma));
1010                                 }
1011
1012                                 start = end;
1013                         } else
1014                                 start = unmap_page_range(*tlbp, vma,
1015                                                 start, end, &zap_work, details);
1016
1017                         if (zap_work > 0) {
1018                                 BUG_ON(start != end);
1019                                 break;
1020                         }
1021
1022                         tlb_finish_mmu(*tlbp, tlb_start, start);
1023
1024                         if (need_resched() ||
1025                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1026                                 if (i_mmap_lock) {
1027                                         *tlbp = NULL;
1028                                         goto out;
1029                                 }
1030                                 cond_resched();
1031                         }
1032
1033                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1034                         tlb_start_valid = 0;
1035                         zap_work = ZAP_BLOCK_SIZE;
1036                 }
1037         }
1038 out:
1039         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1040         return start;   /* which is now the end (or restart) address */
1041 }
1042
1043 /**
1044  * zap_page_range - remove user pages in a given range
1045  * @vma: vm_area_struct holding the applicable pages
1046  * @address: starting address of pages to zap
1047  * @size: number of bytes to zap
1048  * @details: details of nonlinear truncation or shared cache invalidation
1049  */
1050 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1051                 unsigned long size, struct zap_details *details)
1052 {
1053         struct mm_struct *mm = vma->vm_mm;
1054         struct mmu_gather *tlb;
1055         unsigned long end = address + size;
1056         unsigned long nr_accounted = 0;
1057
1058         lru_add_drain();
1059         tlb = tlb_gather_mmu(mm, 0);
1060         update_hiwater_rss(mm);
1061         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1062         if (tlb)
1063                 tlb_finish_mmu(tlb, address, end);
1064         return end;
1065 }
1066
1067 /**
1068  * zap_vma_ptes - remove ptes mapping the vma
1069  * @vma: vm_area_struct holding ptes to be zapped
1070  * @address: starting address of pages to zap
1071  * @size: number of bytes to zap
1072  *
1073  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1074  *
1075  * The entire address range must be fully contained within the vma.
1076  *
1077  * Returns 0 if successful.
1078  */
1079 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1080                 unsigned long size)
1081 {
1082         if (address < vma->vm_start || address + size > vma->vm_end ||
1083                         !(vma->vm_flags & VM_PFNMAP))
1084                 return -1;
1085         zap_page_range(vma, address, size, NULL);
1086         return 0;
1087 }
1088 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1089
1090 /*
1091  * Do a quick page-table lookup for a single page.
1092  */
1093 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1094                         unsigned int flags)
1095 {
1096         pgd_t *pgd;
1097         pud_t *pud;
1098         pmd_t *pmd;
1099         pte_t *ptep, pte;
1100         spinlock_t *ptl;
1101         struct page *page;
1102         struct mm_struct *mm = vma->vm_mm;
1103
1104         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1105         if (!IS_ERR(page)) {
1106                 BUG_ON(flags & FOLL_GET);
1107                 goto out;
1108         }
1109
1110         page = NULL;
1111         pgd = pgd_offset(mm, address);
1112         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1113                 goto no_page_table;
1114
1115         pud = pud_offset(pgd, address);
1116         if (pud_none(*pud))
1117                 goto no_page_table;
1118         if (pud_huge(*pud)) {
1119                 BUG_ON(flags & FOLL_GET);
1120                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1121                 goto out;
1122         }
1123         if (unlikely(pud_bad(*pud)))
1124                 goto no_page_table;
1125
1126         pmd = pmd_offset(pud, address);
1127         if (pmd_none(*pmd))
1128                 goto no_page_table;
1129         if (pmd_huge(*pmd)) {
1130                 BUG_ON(flags & FOLL_GET);
1131                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1132                 goto out;
1133         }
1134         if (unlikely(pmd_bad(*pmd)))
1135                 goto no_page_table;
1136
1137         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1138
1139         pte = *ptep;
1140         if (!pte_present(pte))
1141                 goto no_page;
1142         if ((flags & FOLL_WRITE) && !pte_write(pte))
1143                 goto unlock;
1144         page = vm_normal_page(vma, address, pte);
1145         if (unlikely(!page))
1146                 goto bad_page;
1147
1148         if (flags & FOLL_GET)
1149                 get_page(page);
1150         if (flags & FOLL_TOUCH) {
1151                 if ((flags & FOLL_WRITE) &&
1152                     !pte_dirty(pte) && !PageDirty(page))
1153                         set_page_dirty(page);
1154                 mark_page_accessed(page);
1155         }
1156 unlock:
1157         pte_unmap_unlock(ptep, ptl);
1158 out:
1159         return page;
1160
1161 bad_page:
1162         pte_unmap_unlock(ptep, ptl);
1163         return ERR_PTR(-EFAULT);
1164
1165 no_page:
1166         pte_unmap_unlock(ptep, ptl);
1167         if (!pte_none(pte))
1168                 return page;
1169         /* Fall through to ZERO_PAGE handling */
1170 no_page_table:
1171         /*
1172          * When core dumping an enormous anonymous area that nobody
1173          * has touched so far, we don't want to allocate page tables.
1174          */
1175         if (flags & FOLL_ANON) {
1176                 page = ZERO_PAGE(0);
1177                 if (flags & FOLL_GET)
1178                         get_page(page);
1179                 BUG_ON(flags & FOLL_WRITE);
1180         }
1181         return page;
1182 }
1183
1184 /* Can we do the FOLL_ANON optimization? */
1185 static inline int use_zero_page(struct vm_area_struct *vma)
1186 {
1187         /*
1188          * We don't want to optimize FOLL_ANON for make_pages_present()
1189          * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1190          * we want to get the page from the page tables to make sure
1191          * that we serialize and update with any other user of that
1192          * mapping.
1193          */
1194         if (vma->vm_flags & (VM_LOCKED | VM_SHARED))
1195                 return 0;
1196         /*
1197          * And if we have a fault routine, it's not an anonymous region.
1198          */
1199         return !vma->vm_ops || !vma->vm_ops->fault;
1200 }
1201
1202
1203
1204 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1205                      unsigned long start, int len, int flags,
1206                 struct page **pages, struct vm_area_struct **vmas)
1207 {
1208         int i;
1209         unsigned int vm_flags = 0;
1210         int write = !!(flags & GUP_FLAGS_WRITE);
1211         int force = !!(flags & GUP_FLAGS_FORCE);
1212         int ignore = !!(flags & GUP_FLAGS_IGNORE_VMA_PERMISSIONS);
1213         int ignore_sigkill = !!(flags & GUP_FLAGS_IGNORE_SIGKILL);
1214
1215         if (len <= 0)
1216                 return 0;
1217         /* 
1218          * Require read or write permissions.
1219          * If 'force' is set, we only require the "MAY" flags.
1220          */
1221         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1222         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1223         i = 0;
1224
1225         do {
1226                 struct vm_area_struct *vma;
1227                 unsigned int foll_flags;
1228
1229                 vma = find_extend_vma(mm, start);
1230                 if (!vma && in_gate_area(tsk, start)) {
1231                         unsigned long pg = start & PAGE_MASK;
1232                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1233                         pgd_t *pgd;
1234                         pud_t *pud;
1235                         pmd_t *pmd;
1236                         pte_t *pte;
1237
1238                         /* user gate pages are read-only */
1239                         if (!ignore && write)
1240                                 return i ? : -EFAULT;
1241                         if (pg > TASK_SIZE)
1242                                 pgd = pgd_offset_k(pg);
1243                         else
1244                                 pgd = pgd_offset_gate(mm, pg);
1245                         BUG_ON(pgd_none(*pgd));
1246                         pud = pud_offset(pgd, pg);
1247                         BUG_ON(pud_none(*pud));
1248                         pmd = pmd_offset(pud, pg);
1249                         if (pmd_none(*pmd))
1250                                 return i ? : -EFAULT;
1251                         pte = pte_offset_map(pmd, pg);
1252                         if (pte_none(*pte)) {
1253                                 pte_unmap(pte);
1254                                 return i ? : -EFAULT;
1255                         }
1256                         if (pages) {
1257                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1258                                 pages[i] = page;
1259                                 if (page)
1260                                         get_page(page);
1261                         }
1262                         pte_unmap(pte);
1263                         if (vmas)
1264                                 vmas[i] = gate_vma;
1265                         i++;
1266                         start += PAGE_SIZE;
1267                         len--;
1268                         continue;
1269                 }
1270
1271                 if (!vma ||
1272                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1273                     (!ignore && !(vm_flags & vma->vm_flags)))
1274                         return i ? : -EFAULT;
1275
1276                 if (is_vm_hugetlb_page(vma)) {
1277                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1278                                                 &start, &len, i, write);
1279                         continue;
1280                 }
1281
1282                 foll_flags = FOLL_TOUCH;
1283                 if (pages)
1284                         foll_flags |= FOLL_GET;
1285                 if (!write && use_zero_page(vma))
1286                         foll_flags |= FOLL_ANON;
1287
1288                 do {
1289                         struct page *page;
1290
1291                         /*
1292                          * If we have a pending SIGKILL, don't keep faulting
1293                          * pages and potentially allocating memory, unless
1294                          * current is handling munlock--e.g., on exit. In
1295                          * that case, we are not allocating memory.  Rather,
1296                          * we're only unlocking already resident/mapped pages.
1297                          */
1298                         if (unlikely(!ignore_sigkill &&
1299                                         fatal_signal_pending(current)))
1300                                 return i ? i : -ERESTARTSYS;
1301
1302                         if (write)
1303                                 foll_flags |= FOLL_WRITE;
1304
1305                         cond_resched();
1306                         while (!(page = follow_page(vma, start, foll_flags))) {
1307                                 int ret;
1308                                 ret = handle_mm_fault(mm, vma, start,
1309                                                 foll_flags & FOLL_WRITE);
1310                                 if (ret & VM_FAULT_ERROR) {
1311                                         if (ret & VM_FAULT_OOM)
1312                                                 return i ? i : -ENOMEM;
1313                                         else if (ret & VM_FAULT_SIGBUS)
1314                                                 return i ? i : -EFAULT;
1315                                         BUG();
1316                                 }
1317                                 if (ret & VM_FAULT_MAJOR)
1318                                         tsk->maj_flt++;
1319                                 else
1320                                         tsk->min_flt++;
1321
1322                                 /*
1323                                  * The VM_FAULT_WRITE bit tells us that
1324                                  * do_wp_page has broken COW when necessary,
1325                                  * even if maybe_mkwrite decided not to set
1326                                  * pte_write. We can thus safely do subsequent
1327                                  * page lookups as if they were reads. But only
1328                                  * do so when looping for pte_write is futile:
1329                                  * in some cases userspace may also be wanting
1330                                  * to write to the gotten user page, which a
1331                                  * read fault here might prevent (a readonly
1332                                  * page might get reCOWed by userspace write).
1333                                  */
1334                                 if ((ret & VM_FAULT_WRITE) &&
1335                                     !(vma->vm_flags & VM_WRITE))
1336                                         foll_flags &= ~FOLL_WRITE;
1337
1338                                 cond_resched();
1339                         }
1340                         if (IS_ERR(page))
1341                                 return i ? i : PTR_ERR(page);
1342                         if (pages) {
1343                                 pages[i] = page;
1344
1345                                 flush_anon_page(vma, page, start);
1346                                 flush_dcache_page(page);
1347                         }
1348                         if (vmas)
1349                                 vmas[i] = vma;
1350                         i++;
1351                         start += PAGE_SIZE;
1352                         len--;
1353                 } while (len && start < vma->vm_end);
1354         } while (len);
1355         return i;
1356 }
1357
1358 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1359                 unsigned long start, int len, int write, int force,
1360                 struct page **pages, struct vm_area_struct **vmas)
1361 {
1362         int flags = 0;
1363
1364         if (write)
1365                 flags |= GUP_FLAGS_WRITE;
1366         if (force)
1367                 flags |= GUP_FLAGS_FORCE;
1368
1369         return __get_user_pages(tsk, mm,
1370                                 start, len, flags,
1371                                 pages, vmas);
1372 }
1373
1374 EXPORT_SYMBOL(get_user_pages);
1375
1376 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1377                         spinlock_t **ptl)
1378 {
1379         pgd_t * pgd = pgd_offset(mm, addr);
1380         pud_t * pud = pud_alloc(mm, pgd, addr);
1381         if (pud) {
1382                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1383                 if (pmd)
1384                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1385         }
1386         return NULL;
1387 }
1388
1389 /*
1390  * This is the old fallback for page remapping.
1391  *
1392  * For historical reasons, it only allows reserved pages. Only
1393  * old drivers should use this, and they needed to mark their
1394  * pages reserved for the old functions anyway.
1395  */
1396 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1397                         struct page *page, pgprot_t prot)
1398 {
1399         struct mm_struct *mm = vma->vm_mm;
1400         int retval;
1401         pte_t *pte;
1402         spinlock_t *ptl;
1403
1404         retval = -EINVAL;
1405         if (PageAnon(page))
1406                 goto out;
1407         retval = -ENOMEM;
1408         flush_dcache_page(page);
1409         pte = get_locked_pte(mm, addr, &ptl);
1410         if (!pte)
1411                 goto out;
1412         retval = -EBUSY;
1413         if (!pte_none(*pte))
1414                 goto out_unlock;
1415
1416         /* Ok, finally just insert the thing.. */
1417         get_page(page);
1418         inc_mm_counter(mm, file_rss);
1419         page_add_file_rmap(page);
1420         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1421
1422         retval = 0;
1423         pte_unmap_unlock(pte, ptl);
1424         return retval;
1425 out_unlock:
1426         pte_unmap_unlock(pte, ptl);
1427 out:
1428         return retval;
1429 }
1430
1431 /**
1432  * vm_insert_page - insert single page into user vma
1433  * @vma: user vma to map to
1434  * @addr: target user address of this page
1435  * @page: source kernel page
1436  *
1437  * This allows drivers to insert individual pages they've allocated
1438  * into a user vma.
1439  *
1440  * The page has to be a nice clean _individual_ kernel allocation.
1441  * If you allocate a compound page, you need to have marked it as
1442  * such (__GFP_COMP), or manually just split the page up yourself
1443  * (see split_page()).
1444  *
1445  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1446  * took an arbitrary page protection parameter. This doesn't allow
1447  * that. Your vma protection will have to be set up correctly, which
1448  * means that if you want a shared writable mapping, you'd better
1449  * ask for a shared writable mapping!
1450  *
1451  * The page does not need to be reserved.
1452  */
1453 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1454                         struct page *page)
1455 {
1456         if (addr < vma->vm_start || addr >= vma->vm_end)
1457                 return -EFAULT;
1458         if (!page_count(page))
1459                 return -EINVAL;
1460         vma->vm_flags |= VM_INSERTPAGE;
1461         return insert_page(vma, addr, page, vma->vm_page_prot);
1462 }
1463 EXPORT_SYMBOL(vm_insert_page);
1464
1465 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1466                         unsigned long pfn, pgprot_t prot)
1467 {
1468         struct mm_struct *mm = vma->vm_mm;
1469         int retval;
1470         pte_t *pte, entry;
1471         spinlock_t *ptl;
1472
1473         retval = -ENOMEM;
1474         pte = get_locked_pte(mm, addr, &ptl);
1475         if (!pte)
1476                 goto out;
1477         retval = -EBUSY;
1478         if (!pte_none(*pte))
1479                 goto out_unlock;
1480
1481         /* Ok, finally just insert the thing.. */
1482         entry = pte_mkspecial(pfn_pte(pfn, prot));
1483         set_pte_at(mm, addr, pte, entry);
1484         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1485
1486         retval = 0;
1487 out_unlock:
1488         pte_unmap_unlock(pte, ptl);
1489 out:
1490         return retval;
1491 }
1492
1493 /**
1494  * vm_insert_pfn - insert single pfn into user vma
1495  * @vma: user vma to map to
1496  * @addr: target user address of this page
1497  * @pfn: source kernel pfn
1498  *
1499  * Similar to vm_inert_page, this allows drivers to insert individual pages
1500  * they've allocated into a user vma. Same comments apply.
1501  *
1502  * This function should only be called from a vm_ops->fault handler, and
1503  * in that case the handler should return NULL.
1504  *
1505  * vma cannot be a COW mapping.
1506  *
1507  * As this is called only for pages that do not currently exist, we
1508  * do not need to flush old virtual caches or the TLB.
1509  */
1510 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1511                         unsigned long pfn)
1512 {
1513         int ret;
1514         pgprot_t pgprot = vma->vm_page_prot;
1515         /*
1516          * Technically, architectures with pte_special can avoid all these
1517          * restrictions (same for remap_pfn_range).  However we would like
1518          * consistency in testing and feature parity among all, so we should
1519          * try to keep these invariants in place for everybody.
1520          */
1521         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1522         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1523                                                 (VM_PFNMAP|VM_MIXEDMAP));
1524         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1525         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1526
1527         if (addr < vma->vm_start || addr >= vma->vm_end)
1528                 return -EFAULT;
1529         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1530                 return -EINVAL;
1531
1532         ret = insert_pfn(vma, addr, pfn, pgprot);
1533
1534         if (ret)
1535                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1536
1537         return ret;
1538 }
1539 EXPORT_SYMBOL(vm_insert_pfn);
1540
1541 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1542                         unsigned long pfn)
1543 {
1544         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1545
1546         if (addr < vma->vm_start || addr >= vma->vm_end)
1547                 return -EFAULT;
1548
1549         /*
1550          * If we don't have pte special, then we have to use the pfn_valid()
1551          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1552          * refcount the page if pfn_valid is true (hence insert_page rather
1553          * than insert_pfn).
1554          */
1555         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1556                 struct page *page;
1557
1558                 page = pfn_to_page(pfn);
1559                 return insert_page(vma, addr, page, vma->vm_page_prot);
1560         }
1561         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1562 }
1563 EXPORT_SYMBOL(vm_insert_mixed);
1564
1565 /*
1566  * maps a range of physical memory into the requested pages. the old
1567  * mappings are removed. any references to nonexistent pages results
1568  * in null mappings (currently treated as "copy-on-access")
1569  */
1570 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1571                         unsigned long addr, unsigned long end,
1572                         unsigned long pfn, pgprot_t prot)
1573 {
1574         pte_t *pte;
1575         spinlock_t *ptl;
1576
1577         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1578         if (!pte)
1579                 return -ENOMEM;
1580         arch_enter_lazy_mmu_mode();
1581         do {
1582                 BUG_ON(!pte_none(*pte));
1583                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1584                 pfn++;
1585         } while (pte++, addr += PAGE_SIZE, addr != end);
1586         arch_leave_lazy_mmu_mode();
1587         pte_unmap_unlock(pte - 1, ptl);
1588         return 0;
1589 }
1590
1591 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1592                         unsigned long addr, unsigned long end,
1593                         unsigned long pfn, pgprot_t prot)
1594 {
1595         pmd_t *pmd;
1596         unsigned long next;
1597
1598         pfn -= addr >> PAGE_SHIFT;
1599         pmd = pmd_alloc(mm, pud, addr);
1600         if (!pmd)
1601                 return -ENOMEM;
1602         do {
1603                 next = pmd_addr_end(addr, end);
1604                 if (remap_pte_range(mm, pmd, addr, next,
1605                                 pfn + (addr >> PAGE_SHIFT), prot))
1606                         return -ENOMEM;
1607         } while (pmd++, addr = next, addr != end);
1608         return 0;
1609 }
1610
1611 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1612                         unsigned long addr, unsigned long end,
1613                         unsigned long pfn, pgprot_t prot)
1614 {
1615         pud_t *pud;
1616         unsigned long next;
1617
1618         pfn -= addr >> PAGE_SHIFT;
1619         pud = pud_alloc(mm, pgd, addr);
1620         if (!pud)
1621                 return -ENOMEM;
1622         do {
1623                 next = pud_addr_end(addr, end);
1624                 if (remap_pmd_range(mm, pud, addr, next,
1625                                 pfn + (addr >> PAGE_SHIFT), prot))
1626                         return -ENOMEM;
1627         } while (pud++, addr = next, addr != end);
1628         return 0;
1629 }
1630
1631 /**
1632  * remap_pfn_range - remap kernel memory to userspace
1633  * @vma: user vma to map to
1634  * @addr: target user address to start at
1635  * @pfn: physical address of kernel memory
1636  * @size: size of map area
1637  * @prot: page protection flags for this mapping
1638  *
1639  *  Note: this is only safe if the mm semaphore is held when called.
1640  */
1641 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1642                     unsigned long pfn, unsigned long size, pgprot_t prot)
1643 {
1644         pgd_t *pgd;
1645         unsigned long next;
1646         unsigned long end = addr + PAGE_ALIGN(size);
1647         struct mm_struct *mm = vma->vm_mm;
1648         int err;
1649
1650         /*
1651          * Physically remapped pages are special. Tell the
1652          * rest of the world about it:
1653          *   VM_IO tells people not to look at these pages
1654          *      (accesses can have side effects).
1655          *   VM_RESERVED is specified all over the place, because
1656          *      in 2.4 it kept swapout's vma scan off this vma; but
1657          *      in 2.6 the LRU scan won't even find its pages, so this
1658          *      flag means no more than count its pages in reserved_vm,
1659          *      and omit it from core dump, even when VM_IO turned off.
1660          *   VM_PFNMAP tells the core MM that the base pages are just
1661          *      raw PFN mappings, and do not have a "struct page" associated
1662          *      with them.
1663          *
1664          * There's a horrible special case to handle copy-on-write
1665          * behaviour that some programs depend on. We mark the "original"
1666          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1667          */
1668         if (addr == vma->vm_start && end == vma->vm_end) {
1669                 vma->vm_pgoff = pfn;
1670                 vma->vm_flags |= VM_PFNMAP_AT_MMAP;
1671         } else if (is_cow_mapping(vma->vm_flags))
1672                 return -EINVAL;
1673
1674         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1675
1676         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1677         if (err) {
1678                 /*
1679                  * To indicate that track_pfn related cleanup is not
1680                  * needed from higher level routine calling unmap_vmas
1681                  */
1682                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1683                 vma->vm_flags &= ~VM_PFNMAP_AT_MMAP;
1684                 return -EINVAL;
1685         }
1686
1687         BUG_ON(addr >= end);
1688         pfn -= addr >> PAGE_SHIFT;
1689         pgd = pgd_offset(mm, addr);
1690         flush_cache_range(vma, addr, end);
1691         do {
1692                 next = pgd_addr_end(addr, end);
1693                 err = remap_pud_range(mm, pgd, addr, next,
1694                                 pfn + (addr >> PAGE_SHIFT), prot);
1695                 if (err)
1696                         break;
1697         } while (pgd++, addr = next, addr != end);
1698
1699         if (err)
1700                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1701
1702         return err;
1703 }
1704 EXPORT_SYMBOL(remap_pfn_range);
1705
1706 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1707                                      unsigned long addr, unsigned long end,
1708                                      pte_fn_t fn, void *data)
1709 {
1710         pte_t *pte;
1711         int err;
1712         pgtable_t token;
1713         spinlock_t *uninitialized_var(ptl);
1714
1715         pte = (mm == &init_mm) ?
1716                 pte_alloc_kernel(pmd, addr) :
1717                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1718         if (!pte)
1719                 return -ENOMEM;
1720
1721         BUG_ON(pmd_huge(*pmd));
1722
1723         arch_enter_lazy_mmu_mode();
1724
1725         token = pmd_pgtable(*pmd);
1726
1727         do {
1728                 err = fn(pte, token, addr, data);
1729                 if (err)
1730                         break;
1731         } while (pte++, addr += PAGE_SIZE, addr != end);
1732
1733         arch_leave_lazy_mmu_mode();
1734
1735         if (mm != &init_mm)
1736                 pte_unmap_unlock(pte-1, ptl);
1737         return err;
1738 }
1739
1740 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1741                                      unsigned long addr, unsigned long end,
1742                                      pte_fn_t fn, void *data)
1743 {
1744         pmd_t *pmd;
1745         unsigned long next;
1746         int err;
1747
1748         BUG_ON(pud_huge(*pud));
1749
1750         pmd = pmd_alloc(mm, pud, addr);
1751         if (!pmd)
1752                 return -ENOMEM;
1753         do {
1754                 next = pmd_addr_end(addr, end);
1755                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1756                 if (err)
1757                         break;
1758         } while (pmd++, addr = next, addr != end);
1759         return err;
1760 }
1761
1762 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1763                                      unsigned long addr, unsigned long end,
1764                                      pte_fn_t fn, void *data)
1765 {
1766         pud_t *pud;
1767         unsigned long next;
1768         int err;
1769
1770         pud = pud_alloc(mm, pgd, addr);
1771         if (!pud)
1772                 return -ENOMEM;
1773         do {
1774                 next = pud_addr_end(addr, end);
1775                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1776                 if (err)
1777                         break;
1778         } while (pud++, addr = next, addr != end);
1779         return err;
1780 }
1781
1782 /*
1783  * Scan a region of virtual memory, filling in page tables as necessary
1784  * and calling a provided function on each leaf page table.
1785  */
1786 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1787                         unsigned long size, pte_fn_t fn, void *data)
1788 {
1789         pgd_t *pgd;
1790         unsigned long next;
1791         unsigned long start = addr, end = addr + size;
1792         int err;
1793
1794         BUG_ON(addr >= end);
1795         mmu_notifier_invalidate_range_start(mm, start, end);
1796         pgd = pgd_offset(mm, addr);
1797         do {
1798                 next = pgd_addr_end(addr, end);
1799                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1800                 if (err)
1801                         break;
1802         } while (pgd++, addr = next, addr != end);
1803         mmu_notifier_invalidate_range_end(mm, start, end);
1804         return err;
1805 }
1806 EXPORT_SYMBOL_GPL(apply_to_page_range);
1807
1808 /*
1809  * handle_pte_fault chooses page fault handler according to an entry
1810  * which was read non-atomically.  Before making any commitment, on
1811  * those architectures or configurations (e.g. i386 with PAE) which
1812  * might give a mix of unmatched parts, do_swap_page and do_file_page
1813  * must check under lock before unmapping the pte and proceeding
1814  * (but do_wp_page is only called after already making such a check;
1815  * and do_anonymous_page and do_no_page can safely check later on).
1816  */
1817 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1818                                 pte_t *page_table, pte_t orig_pte)
1819 {
1820         int same = 1;
1821 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1822         if (sizeof(pte_t) > sizeof(unsigned long)) {
1823                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1824                 spin_lock(ptl);
1825                 same = pte_same(*page_table, orig_pte);
1826                 spin_unlock(ptl);
1827         }
1828 #endif
1829         pte_unmap(page_table);
1830         return same;
1831 }
1832
1833 /*
1834  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1835  * servicing faults for write access.  In the normal case, do always want
1836  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1837  * that do not have writing enabled, when used by access_process_vm.
1838  */
1839 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1840 {
1841         if (likely(vma->vm_flags & VM_WRITE))
1842                 pte = pte_mkwrite(pte);
1843         return pte;
1844 }
1845
1846 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1847 {
1848         /*
1849          * If the source page was a PFN mapping, we don't have
1850          * a "struct page" for it. We do a best-effort copy by
1851          * just copying from the original user address. If that
1852          * fails, we just zero-fill it. Live with it.
1853          */
1854         if (unlikely(!src)) {
1855                 void *kaddr = kmap_atomic(dst, KM_USER0);
1856                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1857
1858                 /*
1859                  * This really shouldn't fail, because the page is there
1860                  * in the page tables. But it might just be unreadable,
1861                  * in which case we just give up and fill the result with
1862                  * zeroes.
1863                  */
1864                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1865                         memset(kaddr, 0, PAGE_SIZE);
1866                 kunmap_atomic(kaddr, KM_USER0);
1867                 flush_dcache_page(dst);
1868         } else
1869                 copy_user_highpage(dst, src, va, vma);
1870 }
1871
1872 /*
1873  * This routine handles present pages, when users try to write
1874  * to a shared page. It is done by copying the page to a new address
1875  * and decrementing the shared-page counter for the old page.
1876  *
1877  * Note that this routine assumes that the protection checks have been
1878  * done by the caller (the low-level page fault routine in most cases).
1879  * Thus we can safely just mark it writable once we've done any necessary
1880  * COW.
1881  *
1882  * We also mark the page dirty at this point even though the page will
1883  * change only once the write actually happens. This avoids a few races,
1884  * and potentially makes it more efficient.
1885  *
1886  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1887  * but allow concurrent faults), with pte both mapped and locked.
1888  * We return with mmap_sem still held, but pte unmapped and unlocked.
1889  */
1890 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1891                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1892                 spinlock_t *ptl, pte_t orig_pte)
1893 {
1894         struct page *old_page, *new_page;
1895         pte_t entry;
1896         int reuse = 0, ret = 0;
1897         int page_mkwrite = 0;
1898         struct page *dirty_page = NULL;
1899
1900         old_page = vm_normal_page(vma, address, orig_pte);
1901         if (!old_page) {
1902                 /*
1903                  * VM_MIXEDMAP !pfn_valid() case
1904                  *
1905                  * We should not cow pages in a shared writeable mapping.
1906                  * Just mark the pages writable as we can't do any dirty
1907                  * accounting on raw pfn maps.
1908                  */
1909                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1910                                      (VM_WRITE|VM_SHARED))
1911                         goto reuse;
1912                 goto gotten;
1913         }
1914
1915         /*
1916          * Take out anonymous pages first, anonymous shared vmas are
1917          * not dirty accountable.
1918          */
1919         if (PageAnon(old_page)) {
1920                 if (!trylock_page(old_page)) {
1921                         page_cache_get(old_page);
1922                         pte_unmap_unlock(page_table, ptl);
1923                         lock_page(old_page);
1924                         page_table = pte_offset_map_lock(mm, pmd, address,
1925                                                          &ptl);
1926                         if (!pte_same(*page_table, orig_pte)) {
1927                                 unlock_page(old_page);
1928                                 page_cache_release(old_page);
1929                                 goto unlock;
1930                         }
1931                         page_cache_release(old_page);
1932                 }
1933                 reuse = reuse_swap_page(old_page);
1934                 unlock_page(old_page);
1935         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1936                                         (VM_WRITE|VM_SHARED))) {
1937                 /*
1938                  * Only catch write-faults on shared writable pages,
1939                  * read-only shared pages can get COWed by
1940                  * get_user_pages(.write=1, .force=1).
1941                  */
1942                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1943                         /*
1944                          * Notify the address space that the page is about to
1945                          * become writable so that it can prohibit this or wait
1946                          * for the page to get into an appropriate state.
1947                          *
1948                          * We do this without the lock held, so that it can
1949                          * sleep if it needs to.
1950                          */
1951                         page_cache_get(old_page);
1952                         pte_unmap_unlock(page_table, ptl);
1953
1954                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1955                                 goto unwritable_page;
1956
1957                         /*
1958                          * Since we dropped the lock we need to revalidate
1959                          * the PTE as someone else may have changed it.  If
1960                          * they did, we just return, as we can count on the
1961                          * MMU to tell us if they didn't also make it writable.
1962                          */
1963                         page_table = pte_offset_map_lock(mm, pmd, address,
1964                                                          &ptl);
1965                         page_cache_release(old_page);
1966                         if (!pte_same(*page_table, orig_pte))
1967                                 goto unlock;
1968
1969                         page_mkwrite = 1;
1970                 }
1971                 dirty_page = old_page;
1972                 get_page(dirty_page);
1973                 reuse = 1;
1974         }
1975
1976         if (reuse) {
1977 reuse:
1978                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1979                 entry = pte_mkyoung(orig_pte);
1980                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1981                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1982                         update_mmu_cache(vma, address, entry);
1983                 ret |= VM_FAULT_WRITE;
1984                 goto unlock;
1985         }
1986
1987         /*
1988          * Ok, we need to copy. Oh, well..
1989          */
1990         page_cache_get(old_page);
1991 gotten:
1992         pte_unmap_unlock(page_table, ptl);
1993
1994         if (unlikely(anon_vma_prepare(vma)))
1995                 goto oom;
1996         VM_BUG_ON(old_page == ZERO_PAGE(0));
1997         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1998         if (!new_page)
1999                 goto oom;
2000         /*
2001          * Don't let another task, with possibly unlocked vma,
2002          * keep the mlocked page.
2003          */
2004         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2005                 lock_page(old_page);    /* for LRU manipulation */
2006                 clear_page_mlock(old_page);
2007                 unlock_page(old_page);
2008         }
2009         cow_user_page(new_page, old_page, address, vma);
2010         __SetPageUptodate(new_page);
2011
2012         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2013                 goto oom_free_new;
2014
2015         /*
2016          * Re-check the pte - we dropped the lock
2017          */
2018         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2019         if (likely(pte_same(*page_table, orig_pte))) {
2020                 if (old_page) {
2021                         if (!PageAnon(old_page)) {
2022                                 dec_mm_counter(mm, file_rss);
2023                                 inc_mm_counter(mm, anon_rss);
2024                         }
2025                 } else
2026                         inc_mm_counter(mm, anon_rss);
2027                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2028                 entry = mk_pte(new_page, vma->vm_page_prot);
2029                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2030                 /*
2031                  * Clear the pte entry and flush it first, before updating the
2032                  * pte with the new entry. This will avoid a race condition
2033                  * seen in the presence of one thread doing SMC and another
2034                  * thread doing COW.
2035                  */
2036                 ptep_clear_flush_notify(vma, address, page_table);
2037                 page_add_new_anon_rmap(new_page, vma, address);
2038                 set_pte_at(mm, address, page_table, entry);
2039                 update_mmu_cache(vma, address, entry);
2040                 if (old_page) {
2041                         /*
2042                          * Only after switching the pte to the new page may
2043                          * we remove the mapcount here. Otherwise another
2044                          * process may come and find the rmap count decremented
2045                          * before the pte is switched to the new page, and
2046                          * "reuse" the old page writing into it while our pte
2047                          * here still points into it and can be read by other
2048                          * threads.
2049                          *
2050                          * The critical issue is to order this
2051                          * page_remove_rmap with the ptp_clear_flush above.
2052                          * Those stores are ordered by (if nothing else,)
2053                          * the barrier present in the atomic_add_negative
2054                          * in page_remove_rmap.
2055                          *
2056                          * Then the TLB flush in ptep_clear_flush ensures that
2057                          * no process can access the old page before the
2058                          * decremented mapcount is visible. And the old page
2059                          * cannot be reused until after the decremented
2060                          * mapcount is visible. So transitively, TLBs to
2061                          * old page will be flushed before it can be reused.
2062                          */
2063                         page_remove_rmap(old_page);
2064                 }
2065
2066                 /* Free the old page.. */
2067                 new_page = old_page;
2068                 ret |= VM_FAULT_WRITE;
2069         } else
2070                 mem_cgroup_uncharge_page(new_page);
2071
2072         if (new_page)
2073                 page_cache_release(new_page);
2074         if (old_page)
2075                 page_cache_release(old_page);
2076 unlock:
2077         pte_unmap_unlock(page_table, ptl);
2078         if (dirty_page) {
2079                 if (vma->vm_file)
2080                         file_update_time(vma->vm_file);
2081
2082                 /*
2083                  * Yes, Virginia, this is actually required to prevent a race
2084                  * with clear_page_dirty_for_io() from clearing the page dirty
2085                  * bit after it clear all dirty ptes, but before a racing
2086                  * do_wp_page installs a dirty pte.
2087                  *
2088                  * do_no_page is protected similarly.
2089                  */
2090                 wait_on_page_locked(dirty_page);
2091                 set_page_dirty_balance(dirty_page, page_mkwrite);
2092                 put_page(dirty_page);
2093         }
2094         return ret;
2095 oom_free_new:
2096         page_cache_release(new_page);
2097 oom:
2098         if (old_page)
2099                 page_cache_release(old_page);
2100         return VM_FAULT_OOM;
2101
2102 unwritable_page:
2103         page_cache_release(old_page);
2104         return VM_FAULT_SIGBUS;
2105 }
2106
2107 /*
2108  * Helper functions for unmap_mapping_range().
2109  *
2110  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2111  *
2112  * We have to restart searching the prio_tree whenever we drop the lock,
2113  * since the iterator is only valid while the lock is held, and anyway
2114  * a later vma might be split and reinserted earlier while lock dropped.
2115  *
2116  * The list of nonlinear vmas could be handled more efficiently, using
2117  * a placeholder, but handle it in the same way until a need is shown.
2118  * It is important to search the prio_tree before nonlinear list: a vma
2119  * may become nonlinear and be shifted from prio_tree to nonlinear list
2120  * while the lock is dropped; but never shifted from list to prio_tree.
2121  *
2122  * In order to make forward progress despite restarting the search,
2123  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2124  * quickly skip it next time around.  Since the prio_tree search only
2125  * shows us those vmas affected by unmapping the range in question, we
2126  * can't efficiently keep all vmas in step with mapping->truncate_count:
2127  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2128  * mapping->truncate_count and vma->vm_truncate_count are protected by
2129  * i_mmap_lock.
2130  *
2131  * In order to make forward progress despite repeatedly restarting some
2132  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2133  * and restart from that address when we reach that vma again.  It might
2134  * have been split or merged, shrunk or extended, but never shifted: so
2135  * restart_addr remains valid so long as it remains in the vma's range.
2136  * unmap_mapping_range forces truncate_count to leap over page-aligned
2137  * values so we can save vma's restart_addr in its truncate_count field.
2138  */
2139 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2140
2141 static void reset_vma_truncate_counts(struct address_space *mapping)
2142 {
2143         struct vm_area_struct *vma;
2144         struct prio_tree_iter iter;
2145
2146         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2147                 vma->vm_truncate_count = 0;
2148         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2149                 vma->vm_truncate_count = 0;
2150 }
2151
2152 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2153                 unsigned long start_addr, unsigned long end_addr,
2154                 struct zap_details *details)
2155 {
2156         unsigned long restart_addr;
2157         int need_break;
2158
2159         /*
2160          * files that support invalidating or truncating portions of the
2161          * file from under mmaped areas must have their ->fault function
2162          * return a locked page (and set VM_FAULT_LOCKED in the return).
2163          * This provides synchronisation against concurrent unmapping here.
2164          */
2165
2166 again:
2167         restart_addr = vma->vm_truncate_count;
2168         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2169                 start_addr = restart_addr;
2170                 if (start_addr >= end_addr) {
2171                         /* Top of vma has been split off since last time */
2172                         vma->vm_truncate_count = details->truncate_count;
2173                         return 0;
2174                 }
2175         }
2176
2177         restart_addr = zap_page_range(vma, start_addr,
2178                                         end_addr - start_addr, details);
2179         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2180
2181         if (restart_addr >= end_addr) {
2182                 /* We have now completed this vma: mark it so */
2183                 vma->vm_truncate_count = details->truncate_count;
2184                 if (!need_break)
2185                         return 0;
2186         } else {
2187                 /* Note restart_addr in vma's truncate_count field */
2188                 vma->vm_truncate_count = restart_addr;
2189                 if (!need_break)
2190                         goto again;
2191         }
2192
2193         spin_unlock(details->i_mmap_lock);
2194         cond_resched();
2195         spin_lock(details->i_mmap_lock);
2196         return -EINTR;
2197 }
2198
2199 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2200                                             struct zap_details *details)
2201 {
2202         struct vm_area_struct *vma;
2203         struct prio_tree_iter iter;
2204         pgoff_t vba, vea, zba, zea;
2205
2206 restart:
2207         vma_prio_tree_foreach(vma, &iter, root,
2208                         details->first_index, details->last_index) {
2209                 /* Skip quickly over those we have already dealt with */
2210                 if (vma->vm_truncate_count == details->truncate_count)
2211                         continue;
2212
2213                 vba = vma->vm_pgoff;
2214                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2215                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2216                 zba = details->first_index;
2217                 if (zba < vba)
2218                         zba = vba;
2219                 zea = details->last_index;
2220                 if (zea > vea)
2221                         zea = vea;
2222
2223                 if (unmap_mapping_range_vma(vma,
2224                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2225                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2226                                 details) < 0)
2227                         goto restart;
2228         }
2229 }
2230
2231 static inline void unmap_mapping_range_list(struct list_head *head,
2232                                             struct zap_details *details)
2233 {
2234         struct vm_area_struct *vma;
2235
2236         /*
2237          * In nonlinear VMAs there is no correspondence between virtual address
2238          * offset and file offset.  So we must perform an exhaustive search
2239          * across *all* the pages in each nonlinear VMA, not just the pages
2240          * whose virtual address lies outside the file truncation point.
2241          */
2242 restart:
2243         list_for_each_entry(vma, head, shared.vm_set.list) {
2244                 /* Skip quickly over those we have already dealt with */
2245                 if (vma->vm_truncate_count == details->truncate_count)
2246                         continue;
2247                 details->nonlinear_vma = vma;
2248                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2249                                         vma->vm_end, details) < 0)
2250                         goto restart;
2251         }
2252 }
2253
2254 /**
2255  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2256  * @mapping: the address space containing mmaps to be unmapped.
2257  * @holebegin: byte in first page to unmap, relative to the start of
2258  * the underlying file.  This will be rounded down to a PAGE_SIZE
2259  * boundary.  Note that this is different from vmtruncate(), which
2260  * must keep the partial page.  In contrast, we must get rid of
2261  * partial pages.
2262  * @holelen: size of prospective hole in bytes.  This will be rounded
2263  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2264  * end of the file.
2265  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2266  * but 0 when invalidating pagecache, don't throw away private data.
2267  */
2268 void unmap_mapping_range(struct address_space *mapping,
2269                 loff_t const holebegin, loff_t const holelen, int even_cows)
2270 {
2271         struct zap_details details;
2272         pgoff_t hba = holebegin >> PAGE_SHIFT;
2273         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2274
2275         /* Check for overflow. */
2276         if (sizeof(holelen) > sizeof(hlen)) {
2277                 long long holeend =
2278                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2279                 if (holeend & ~(long long)ULONG_MAX)
2280                         hlen = ULONG_MAX - hba + 1;
2281         }
2282
2283         details.check_mapping = even_cows? NULL: mapping;
2284         details.nonlinear_vma = NULL;
2285         details.first_index = hba;
2286         details.last_index = hba + hlen - 1;
2287         if (details.last_index < details.first_index)
2288                 details.last_index = ULONG_MAX;
2289         details.i_mmap_lock = &mapping->i_mmap_lock;
2290
2291         spin_lock(&mapping->i_mmap_lock);
2292
2293         /* Protect against endless unmapping loops */
2294         mapping->truncate_count++;
2295         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2296                 if (mapping->truncate_count == 0)
2297                         reset_vma_truncate_counts(mapping);
2298                 mapping->truncate_count++;
2299         }
2300         details.truncate_count = mapping->truncate_count;
2301
2302         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2303                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2304         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2305                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2306         spin_unlock(&mapping->i_mmap_lock);
2307 }
2308 EXPORT_SYMBOL(unmap_mapping_range);
2309
2310 /**
2311  * vmtruncate - unmap mappings "freed" by truncate() syscall
2312  * @inode: inode of the file used
2313  * @offset: file offset to start truncating
2314  *
2315  * NOTE! We have to be ready to update the memory sharing
2316  * between the file and the memory map for a potential last
2317  * incomplete page.  Ugly, but necessary.
2318  */
2319 int vmtruncate(struct inode * inode, loff_t offset)
2320 {
2321         if (inode->i_size < offset) {
2322                 unsigned long limit;
2323
2324                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2325                 if (limit != RLIM_INFINITY && offset > limit)
2326                         goto out_sig;
2327                 if (offset > inode->i_sb->s_maxbytes)
2328                         goto out_big;
2329                 i_size_write(inode, offset);
2330         } else {
2331                 struct address_space *mapping = inode->i_mapping;
2332
2333                 /*
2334                  * truncation of in-use swapfiles is disallowed - it would
2335                  * cause subsequent swapout to scribble on the now-freed
2336                  * blocks.
2337                  */
2338                 if (IS_SWAPFILE(inode))
2339                         return -ETXTBSY;
2340                 i_size_write(inode, offset);
2341
2342                 /*
2343                  * unmap_mapping_range is called twice, first simply for
2344                  * efficiency so that truncate_inode_pages does fewer
2345                  * single-page unmaps.  However after this first call, and
2346                  * before truncate_inode_pages finishes, it is possible for
2347                  * private pages to be COWed, which remain after
2348                  * truncate_inode_pages finishes, hence the second
2349                  * unmap_mapping_range call must be made for correctness.
2350                  */
2351                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2352                 truncate_inode_pages(mapping, offset);
2353                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2354         }
2355
2356         if (inode->i_op->truncate)
2357                 inode->i_op->truncate(inode);
2358         return 0;
2359
2360 out_sig:
2361         send_sig(SIGXFSZ, current, 0);
2362 out_big:
2363         return -EFBIG;
2364 }
2365 EXPORT_SYMBOL(vmtruncate);
2366
2367 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2368 {
2369         struct address_space *mapping = inode->i_mapping;
2370
2371         /*
2372          * If the underlying filesystem is not going to provide
2373          * a way to truncate a range of blocks (punch a hole) -
2374          * we should return failure right now.
2375          */
2376         if (!inode->i_op->truncate_range)
2377                 return -ENOSYS;
2378
2379         mutex_lock(&inode->i_mutex);
2380         down_write(&inode->i_alloc_sem);
2381         unmap_mapping_range(mapping, offset, (end - offset), 1);
2382         truncate_inode_pages_range(mapping, offset, end);
2383         unmap_mapping_range(mapping, offset, (end - offset), 1);
2384         inode->i_op->truncate_range(inode, offset, end);
2385         up_write(&inode->i_alloc_sem);
2386         mutex_unlock(&inode->i_mutex);
2387
2388         return 0;
2389 }
2390
2391 /*
2392  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2393  * but allow concurrent faults), and pte mapped but not yet locked.
2394  * We return with mmap_sem still held, but pte unmapped and unlocked.
2395  */
2396 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2397                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2398                 int write_access, pte_t orig_pte)
2399 {
2400         spinlock_t *ptl;
2401         struct page *page;
2402         swp_entry_t entry;
2403         pte_t pte;
2404         struct mem_cgroup *ptr = NULL;
2405         int ret = 0;
2406
2407         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2408                 goto out;
2409
2410         entry = pte_to_swp_entry(orig_pte);
2411         if (is_migration_entry(entry)) {
2412                 migration_entry_wait(mm, pmd, address);
2413                 goto out;
2414         }
2415         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2416         page = lookup_swap_cache(entry);
2417         if (!page) {
2418                 grab_swap_token(); /* Contend for token _before_ read-in */
2419                 page = swapin_readahead(entry,
2420                                         GFP_HIGHUSER_MOVABLE, vma, address);
2421                 if (!page) {
2422                         /*
2423                          * Back out if somebody else faulted in this pte
2424                          * while we released the pte lock.
2425                          */
2426                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2427                         if (likely(pte_same(*page_table, orig_pte)))
2428                                 ret = VM_FAULT_OOM;
2429                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2430                         goto unlock;
2431                 }
2432
2433                 /* Had to read the page from swap area: Major fault */
2434                 ret = VM_FAULT_MAJOR;
2435                 count_vm_event(PGMAJFAULT);
2436         }
2437
2438         mark_page_accessed(page);
2439
2440         lock_page(page);
2441         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2442
2443         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2444                 ret = VM_FAULT_OOM;
2445                 unlock_page(page);
2446                 goto out;
2447         }
2448
2449         /*
2450          * Back out if somebody else already faulted in this pte.
2451          */
2452         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2453         if (unlikely(!pte_same(*page_table, orig_pte)))
2454                 goto out_nomap;
2455
2456         if (unlikely(!PageUptodate(page))) {
2457                 ret = VM_FAULT_SIGBUS;
2458                 goto out_nomap;
2459         }
2460
2461         /*
2462          * The page isn't present yet, go ahead with the fault.
2463          *
2464          * Be careful about the sequence of operations here.
2465          * To get its accounting right, reuse_swap_page() must be called
2466          * while the page is counted on swap but not yet in mapcount i.e.
2467          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2468          * must be called after the swap_free(), or it will never succeed.
2469          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2470          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2471          * in page->private. In this case, a record in swap_cgroup  is silently
2472          * discarded at swap_free().
2473          */
2474
2475         inc_mm_counter(mm, anon_rss);
2476         pte = mk_pte(page, vma->vm_page_prot);
2477         if (write_access && reuse_swap_page(page)) {
2478                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2479                 write_access = 0;
2480         }
2481         flush_icache_page(vma, page);
2482         set_pte_at(mm, address, page_table, pte);
2483         page_add_anon_rmap(page, vma, address);
2484         /* It's better to call commit-charge after rmap is established */
2485         mem_cgroup_commit_charge_swapin(page, ptr);
2486
2487         swap_free(entry);
2488         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2489                 try_to_free_swap(page);
2490         unlock_page(page);
2491
2492         if (write_access) {
2493                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2494                 if (ret & VM_FAULT_ERROR)
2495                         ret &= VM_FAULT_ERROR;
2496                 goto out;
2497         }
2498
2499         /* No need to invalidate - it was non-present before */
2500         update_mmu_cache(vma, address, pte);
2501 unlock:
2502         pte_unmap_unlock(page_table, ptl);
2503 out:
2504         return ret;
2505 out_nomap:
2506         mem_cgroup_cancel_charge_swapin(ptr);
2507         pte_unmap_unlock(page_table, ptl);
2508         unlock_page(page);
2509         page_cache_release(page);
2510         return ret;
2511 }
2512
2513 /*
2514  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2515  * but allow concurrent faults), and pte mapped but not yet locked.
2516  * We return with mmap_sem still held, but pte unmapped and unlocked.
2517  */
2518 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2519                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2520                 int write_access)
2521 {
2522         struct page *page;
2523         spinlock_t *ptl;
2524         pte_t entry;
2525
2526         /* Allocate our own private page. */
2527         pte_unmap(page_table);
2528
2529         if (unlikely(anon_vma_prepare(vma)))
2530                 goto oom;
2531         page = alloc_zeroed_user_highpage_movable(vma, address);
2532         if (!page)
2533                 goto oom;
2534         __SetPageUptodate(page);
2535
2536         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2537                 goto oom_free_page;
2538
2539         entry = mk_pte(page, vma->vm_page_prot);
2540         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2541
2542         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2543         if (!pte_none(*page_table))
2544                 goto release;
2545         inc_mm_counter(mm, anon_rss);
2546         page_add_new_anon_rmap(page, vma, address);
2547         set_pte_at(mm, address, page_table, entry);
2548
2549         /* No need to invalidate - it was non-present before */
2550         update_mmu_cache(vma, address, entry);
2551 unlock:
2552         pte_unmap_unlock(page_table, ptl);
2553         return 0;
2554 release:
2555         mem_cgroup_uncharge_page(page);
2556         page_cache_release(page);
2557         goto unlock;
2558 oom_free_page:
2559         page_cache_release(page);
2560 oom:
2561         return VM_FAULT_OOM;
2562 }
2563
2564 /*
2565  * __do_fault() tries to create a new page mapping. It aggressively
2566  * tries to share with existing pages, but makes a separate copy if
2567  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2568  * the next page fault.
2569  *
2570  * As this is called only for pages that do not currently exist, we
2571  * do not need to flush old virtual caches or the TLB.
2572  *
2573  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2574  * but allow concurrent faults), and pte neither mapped nor locked.
2575  * We return with mmap_sem still held, but pte unmapped and unlocked.
2576  */
2577 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2578                 unsigned long address, pmd_t *pmd,
2579                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2580 {
2581         pte_t *page_table;
2582         spinlock_t *ptl;
2583         struct page *page;
2584         pte_t entry;
2585         int anon = 0;
2586         int charged = 0;
2587         struct page *dirty_page = NULL;
2588         struct vm_fault vmf;
2589         int ret;
2590         int page_mkwrite = 0;
2591
2592         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2593         vmf.pgoff = pgoff;
2594         vmf.flags = flags;
2595         vmf.page = NULL;
2596
2597         ret = vma->vm_ops->fault(vma, &vmf);
2598         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2599                 return ret;
2600
2601         /*
2602          * For consistency in subsequent calls, make the faulted page always
2603          * locked.
2604          */
2605         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2606                 lock_page(vmf.page);
2607         else
2608                 VM_BUG_ON(!PageLocked(vmf.page));
2609
2610         /*
2611          * Should we do an early C-O-W break?
2612          */
2613         page = vmf.page;
2614         if (flags & FAULT_FLAG_WRITE) {
2615                 if (!(vma->vm_flags & VM_SHARED)) {
2616                         anon = 1;
2617                         if (unlikely(anon_vma_prepare(vma))) {
2618                                 ret = VM_FAULT_OOM;
2619                                 goto out;
2620                         }
2621                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2622                                                 vma, address);
2623                         if (!page) {
2624                                 ret = VM_FAULT_OOM;
2625                                 goto out;
2626                         }
2627                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2628                                 ret = VM_FAULT_OOM;
2629                                 page_cache_release(page);
2630                                 goto out;
2631                         }
2632                         charged = 1;
2633                         /*
2634                          * Don't let another task, with possibly unlocked vma,
2635                          * keep the mlocked page.
2636                          */
2637                         if (vma->vm_flags & VM_LOCKED)
2638                                 clear_page_mlock(vmf.page);
2639                         copy_user_highpage(page, vmf.page, address, vma);
2640                         __SetPageUptodate(page);
2641                 } else {
2642                         /*
2643                          * If the page will be shareable, see if the backing
2644                          * address space wants to know that the page is about
2645                          * to become writable
2646                          */
2647                         if (vma->vm_ops->page_mkwrite) {
2648                                 unlock_page(page);
2649                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2650                                         ret = VM_FAULT_SIGBUS;
2651                                         anon = 1; /* no anon but release vmf.page */
2652                                         goto out_unlocked;
2653                                 }
2654                                 lock_page(page);
2655                                 /*
2656                                  * XXX: this is not quite right (racy vs
2657                                  * invalidate) to unlock and relock the page
2658                                  * like this, however a better fix requires
2659                                  * reworking page_mkwrite locking API, which
2660                                  * is better done later.
2661                                  */
2662                                 if (!page->mapping) {
2663                                         ret = 0;
2664                                         anon = 1; /* no anon but release vmf.page */
2665                                         goto out;
2666                                 }
2667                                 page_mkwrite = 1;
2668                         }
2669                 }
2670
2671         }
2672
2673         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2674
2675         /*
2676          * This silly early PAGE_DIRTY setting removes a race
2677          * due to the bad i386 page protection. But it's valid
2678          * for other architectures too.
2679          *
2680          * Note that if write_access is true, we either now have
2681          * an exclusive copy of the page, or this is a shared mapping,
2682          * so we can make it writable and dirty to avoid having to
2683          * handle that later.
2684          */
2685         /* Only go through if we didn't race with anybody else... */
2686         if (likely(pte_same(*page_table, orig_pte))) {
2687                 flush_icache_page(vma, page);
2688                 entry = mk_pte(page, vma->vm_page_prot);
2689                 if (flags & FAULT_FLAG_WRITE)
2690                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2691                 if (anon) {
2692                         inc_mm_counter(mm, anon_rss);
2693                         page_add_new_anon_rmap(page, vma, address);
2694                 } else {
2695                         inc_mm_counter(mm, file_rss);
2696                         page_add_file_rmap(page);
2697                         if (flags & FAULT_FLAG_WRITE) {
2698                                 dirty_page = page;
2699                                 get_page(dirty_page);
2700                         }
2701                 }
2702                 set_pte_at(mm, address, page_table, entry);
2703
2704                 /* no need to invalidate: a not-present page won't be cached */
2705                 update_mmu_cache(vma, address, entry);
2706         } else {
2707                 if (charged)
2708                         mem_cgroup_uncharge_page(page);
2709                 if (anon)
2710                         page_cache_release(page);
2711                 else
2712                         anon = 1; /* no anon but release faulted_page */
2713         }
2714
2715         pte_unmap_unlock(page_table, ptl);
2716
2717 out:
2718         unlock_page(vmf.page);
2719 out_unlocked:
2720         if (anon)
2721                 page_cache_release(vmf.page);
2722         else if (dirty_page) {
2723                 if (vma->vm_file)
2724                         file_update_time(vma->vm_file);
2725
2726                 set_page_dirty_balance(dirty_page, page_mkwrite);
2727                 put_page(dirty_page);
2728         }
2729
2730         return ret;
2731 }
2732
2733 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2734                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2735                 int write_access, pte_t orig_pte)
2736 {
2737         pgoff_t pgoff = (((address & PAGE_MASK)
2738                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2739         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2740
2741         pte_unmap(page_table);
2742         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2743 }
2744
2745 /*
2746  * Fault of a previously existing named mapping. Repopulate the pte
2747  * from the encoded file_pte if possible. This enables swappable
2748  * nonlinear vmas.
2749  *
2750  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2751  * but allow concurrent faults), and pte mapped but not yet locked.
2752  * We return with mmap_sem still held, but pte unmapped and unlocked.
2753  */
2754 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2755                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2756                 int write_access, pte_t orig_pte)
2757 {
2758         unsigned int flags = FAULT_FLAG_NONLINEAR |
2759                                 (write_access ? FAULT_FLAG_WRITE : 0);
2760         pgoff_t pgoff;
2761
2762         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2763                 return 0;
2764
2765         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2766                 /*
2767                  * Page table corrupted: show pte and kill process.
2768                  */
2769                 print_bad_pte(vma, address, orig_pte, NULL);
2770                 return VM_FAULT_OOM;
2771         }
2772
2773         pgoff = pte_to_pgoff(orig_pte);
2774         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2775 }
2776
2777 /*
2778  * These routines also need to handle stuff like marking pages dirty
2779  * and/or accessed for architectures that don't do it in hardware (most
2780  * RISC architectures).  The early dirtying is also good on the i386.
2781  *
2782  * There is also a hook called "update_mmu_cache()" that architectures
2783  * with external mmu caches can use to update those (ie the Sparc or
2784  * PowerPC hashed page tables that act as extended TLBs).
2785  *
2786  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2787  * but allow concurrent faults), and pte mapped but not yet locked.
2788  * We return with mmap_sem still held, but pte unmapped and unlocked.
2789  */
2790 static inline int handle_pte_fault(struct mm_struct *mm,
2791                 struct vm_area_struct *vma, unsigned long address,
2792                 pte_t *pte, pmd_t *pmd, int write_access)
2793 {
2794         pte_t entry;
2795         spinlock_t *ptl;
2796
2797         entry = *pte;
2798         if (!pte_present(entry)) {
2799                 if (pte_none(entry)) {
2800                         if (vma->vm_ops) {
2801                                 if (likely(vma->vm_ops->fault))
2802                                         return do_linear_fault(mm, vma, address,
2803                                                 pte, pmd, write_access, entry);
2804                         }
2805                         return do_anonymous_page(mm, vma, address,
2806                                                  pte, pmd, write_access);
2807                 }
2808                 if (pte_file(entry))
2809                         return do_nonlinear_fault(mm, vma, address,
2810                                         pte, pmd, write_access, entry);
2811                 return do_swap_page(mm, vma, address,
2812                                         pte, pmd, write_access, entry);
2813         }
2814
2815         ptl = pte_lockptr(mm, pmd);
2816         spin_lock(ptl);
2817         if (unlikely(!pte_same(*pte, entry)))
2818                 goto unlock;
2819         if (write_access) {
2820                 if (!pte_write(entry))
2821                         return do_wp_page(mm, vma, address,
2822                                         pte, pmd, ptl, entry);
2823                 entry = pte_mkdirty(entry);
2824         }
2825         entry = pte_mkyoung(entry);
2826         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2827                 update_mmu_cache(vma, address, entry);
2828         } else {
2829                 /*
2830                  * This is needed only for protection faults but the arch code
2831                  * is not yet telling us if this is a protection fault or not.
2832                  * This still avoids useless tlb flushes for .text page faults
2833                  * with threads.
2834                  */
2835                 if (write_access)
2836                         flush_tlb_page(vma, address);
2837         }
2838 unlock:
2839         pte_unmap_unlock(pte, ptl);
2840         return 0;
2841 }
2842
2843 /*
2844  * By the time we get here, we already hold the mm semaphore
2845  */
2846 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2847                 unsigned long address, int write_access)
2848 {
2849         pgd_t *pgd;
2850         pud_t *pud;
2851         pmd_t *pmd;
2852         pte_t *pte;
2853
2854         __set_current_state(TASK_RUNNING);
2855
2856         count_vm_event(PGFAULT);
2857
2858         if (unlikely(is_vm_hugetlb_page(vma)))
2859                 return hugetlb_fault(mm, vma, address, write_access);
2860
2861         pgd = pgd_offset(mm, address);
2862         pud = pud_alloc(mm, pgd, address);
2863         if (!pud)
2864                 return VM_FAULT_OOM;
2865         pmd = pmd_alloc(mm, pud, address);
2866         if (!pmd)
2867                 return VM_FAULT_OOM;
2868         pte = pte_alloc_map(mm, pmd, address);
2869         if (!pte)
2870                 return VM_FAULT_OOM;
2871
2872         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2873 }
2874
2875 #ifndef __PAGETABLE_PUD_FOLDED
2876 /*
2877  * Allocate page upper directory.
2878  * We've already handled the fast-path in-line.
2879  */
2880 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2881 {
2882         pud_t *new = pud_alloc_one(mm, address);
2883         if (!new)
2884                 return -ENOMEM;
2885
2886         smp_wmb(); /* See comment in __pte_alloc */
2887
2888         spin_lock(&mm->page_table_lock);
2889         if (pgd_present(*pgd))          /* Another has populated it */
2890                 pud_free(mm, new);
2891         else
2892                 pgd_populate(mm, pgd, new);
2893         spin_unlock(&mm->page_table_lock);
2894         return 0;
2895 }
2896 #endif /* __PAGETABLE_PUD_FOLDED */
2897
2898 #ifndef __PAGETABLE_PMD_FOLDED
2899 /*
2900  * Allocate page middle directory.
2901  * We've already handled the fast-path in-line.
2902  */
2903 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2904 {
2905         pmd_t *new = pmd_alloc_one(mm, address);
2906         if (!new)
2907                 return -ENOMEM;
2908
2909         smp_wmb(); /* See comment in __pte_alloc */
2910
2911         spin_lock(&mm->page_table_lock);
2912 #ifndef __ARCH_HAS_4LEVEL_HACK
2913         if (pud_present(*pud))          /* Another has populated it */
2914                 pmd_free(mm, new);
2915         else
2916                 pud_populate(mm, pud, new);
2917 #else
2918         if (pgd_present(*pud))          /* Another has populated it */
2919                 pmd_free(mm, new);
2920         else
2921                 pgd_populate(mm, pud, new);
2922 #endif /* __ARCH_HAS_4LEVEL_HACK */
2923         spin_unlock(&mm->page_table_lock);
2924         return 0;
2925 }
2926 #endif /* __PAGETABLE_PMD_FOLDED */
2927
2928 int make_pages_present(unsigned long addr, unsigned long end)
2929 {
2930         int ret, len, write;
2931         struct vm_area_struct * vma;
2932
2933         vma = find_vma(current->mm, addr);
2934         if (!vma)
2935                 return -ENOMEM;
2936         write = (vma->vm_flags & VM_WRITE) != 0;
2937         BUG_ON(addr >= end);
2938         BUG_ON(end > vma->vm_end);
2939         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2940         ret = get_user_pages(current, current->mm, addr,
2941                         len, write, 0, NULL, NULL);
2942         if (ret < 0)
2943                 return ret;
2944         return ret == len ? 0 : -EFAULT;
2945 }
2946
2947 #if !defined(__HAVE_ARCH_GATE_AREA)
2948
2949 #if defined(AT_SYSINFO_EHDR)
2950 static struct vm_area_struct gate_vma;
2951
2952 static int __init gate_vma_init(void)
2953 {
2954         gate_vma.vm_mm = NULL;
2955         gate_vma.vm_start = FIXADDR_USER_START;
2956         gate_vma.vm_end = FIXADDR_USER_END;
2957         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2958         gate_vma.vm_page_prot = __P101;
2959         /*
2960          * Make sure the vDSO gets into every core dump.
2961          * Dumping its contents makes post-mortem fully interpretable later
2962          * without matching up the same kernel and hardware config to see
2963          * what PC values meant.
2964          */
2965         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2966         return 0;
2967 }
2968 __initcall(gate_vma_init);
2969 #endif
2970
2971 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2972 {
2973 #ifdef AT_SYSINFO_EHDR
2974         return &gate_vma;
2975 #else
2976         return NULL;
2977 #endif
2978 }
2979
2980 int in_gate_area_no_task(unsigned long addr)
2981 {
2982 #ifdef AT_SYSINFO_EHDR
2983         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2984                 return 1;
2985 #endif
2986         return 0;
2987 }
2988
2989 #endif  /* __HAVE_ARCH_GATE_AREA */
2990
2991 #ifdef CONFIG_HAVE_IOREMAP_PROT
2992 int follow_phys(struct vm_area_struct *vma,
2993                 unsigned long address, unsigned int flags,
2994                 unsigned long *prot, resource_size_t *phys)
2995 {
2996         pgd_t *pgd;
2997         pud_t *pud;
2998         pmd_t *pmd;
2999         pte_t *ptep, pte;
3000         spinlock_t *ptl;
3001         resource_size_t phys_addr = 0;
3002         struct mm_struct *mm = vma->vm_mm;
3003         int ret = -EINVAL;
3004
3005         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3006                 goto out;
3007
3008         pgd = pgd_offset(mm, address);
3009         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3010                 goto out;
3011
3012         pud = pud_offset(pgd, address);
3013         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3014                 goto out;
3015
3016         pmd = pmd_offset(pud, address);
3017         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3018                 goto out;
3019
3020         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3021         if (pmd_huge(*pmd))
3022                 goto out;
3023
3024         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
3025         if (!ptep)
3026                 goto out;
3027
3028         pte = *ptep;
3029         if (!pte_present(pte))
3030                 goto unlock;
3031         if ((flags & FOLL_WRITE) && !pte_write(pte))
3032                 goto unlock;
3033         phys_addr = pte_pfn(pte);
3034         phys_addr <<= PAGE_SHIFT; /* Shift here to avoid overflow on PAE */
3035
3036         *prot = pgprot_val(pte_pgprot(pte));
3037         *phys = phys_addr;
3038         ret = 0;
3039
3040 unlock:
3041         pte_unmap_unlock(ptep, ptl);
3042 out:
3043         return ret;
3044 }
3045
3046 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3047                         void *buf, int len, int write)
3048 {
3049         resource_size_t phys_addr;
3050         unsigned long prot = 0;
3051         void __iomem *maddr;
3052         int offset = addr & (PAGE_SIZE-1);
3053
3054         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3055                 return -EINVAL;
3056
3057         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3058         if (write)
3059                 memcpy_toio(maddr + offset, buf, len);
3060         else
3061                 memcpy_fromio(buf, maddr + offset, len);
3062         iounmap(maddr);
3063
3064         return len;
3065 }
3066 #endif
3067
3068 /*
3069  * Access another process' address space.
3070  * Source/target buffer must be kernel space,
3071  * Do not walk the page table directly, use get_user_pages
3072  */
3073 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3074 {
3075         struct mm_struct *mm;
3076         struct vm_area_struct *vma;
3077         void *old_buf = buf;
3078
3079         mm = get_task_mm(tsk);
3080         if (!mm)
3081                 return 0;
3082
3083         down_read(&mm->mmap_sem);
3084         /* ignore errors, just check how much was successfully transferred */
3085         while (len) {
3086                 int bytes, ret, offset;
3087                 void *maddr;
3088                 struct page *page = NULL;
3089
3090                 ret = get_user_pages(tsk, mm, addr, 1,
3091                                 write, 1, &page, &vma);
3092                 if (ret <= 0) {
3093                         /*
3094                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3095                          * we can access using slightly different code.
3096                          */
3097 #ifdef CONFIG_HAVE_IOREMAP_PROT
3098                         vma = find_vma(mm, addr);
3099                         if (!vma)
3100                                 break;
3101                         if (vma->vm_ops && vma->vm_ops->access)
3102                                 ret = vma->vm_ops->access(vma, addr, buf,
3103                                                           len, write);
3104                         if (ret <= 0)
3105 #endif
3106                                 break;
3107                         bytes = ret;
3108                 } else {
3109                         bytes = len;
3110                         offset = addr & (PAGE_SIZE-1);
3111                         if (bytes > PAGE_SIZE-offset)
3112                                 bytes = PAGE_SIZE-offset;
3113
3114                         maddr = kmap(page);
3115                         if (write) {
3116                                 copy_to_user_page(vma, page, addr,
3117                                                   maddr + offset, buf, bytes);
3118                                 set_page_dirty_lock(page);
3119                         } else {
3120                                 copy_from_user_page(vma, page, addr,
3121                                                     buf, maddr + offset, bytes);
3122                         }
3123                         kunmap(page);
3124                         page_cache_release(page);
3125                 }
3126                 len -= bytes;
3127                 buf += bytes;
3128                 addr += bytes;
3129         }
3130         up_read(&mm->mmap_sem);
3131         mmput(mm);
3132
3133         return buf - old_buf;
3134 }
3135
3136 /*
3137  * Print the name of a VMA.
3138  */
3139 void print_vma_addr(char *prefix, unsigned long ip)
3140 {
3141         struct mm_struct *mm = current->mm;
3142         struct vm_area_struct *vma;
3143
3144         /*
3145          * Do not print if we are in atomic
3146          * contexts (in exception stacks, etc.):
3147          */
3148         if (preempt_count())
3149                 return;
3150
3151         down_read(&mm->mmap_sem);
3152         vma = find_vma(mm, ip);
3153         if (vma && vma->vm_file) {
3154                 struct file *f = vma->vm_file;
3155                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3156                 if (buf) {
3157                         char *p, *s;
3158
3159                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3160                         if (IS_ERR(p))
3161                                 p = "?";
3162                         s = strrchr(p, '/');
3163                         if (s)
3164                                 p = s+1;
3165                         printk("%s%s[%lx+%lx]", prefix, p,
3166                                         vma->vm_start,
3167                                         vma->vm_end - vma->vm_start);
3168                         free_page((unsigned long)buf);
3169                 }
3170         }
3171         up_read(&current->mm->mmap_sem);
3172 }
3173
3174 #ifdef CONFIG_PROVE_LOCKING
3175 void might_fault(void)
3176 {
3177         /*
3178          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3179          * holding the mmap_sem, this is safe because kernel memory doesn't
3180          * get paged out, therefore we'll never actually fault, and the
3181          * below annotations will generate false positives.
3182          */
3183         if (segment_eq(get_fs(), KERNEL_DS))
3184                 return;
3185
3186         might_sleep();
3187         /*
3188          * it would be nicer only to annotate paths which are not under
3189          * pagefault_disable, however that requires a larger audit and
3190          * providing helpers like get_user_atomic.
3191          */
3192         if (!in_atomic() && current->mm)
3193                 might_lock_read(&current->mm->mmap_sem);
3194 }
3195 EXPORT_SYMBOL(might_fault);
3196 #endif