Merge branch 'for-2.6.29' of git://linux-nfs.org/~bfields/linux
[linux-2.6] / 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         else if (is_cow_mapping(vma->vm_flags))
1671                 return -EINVAL;
1672
1673         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1674
1675         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1676         if (err) {
1677                 /*
1678                  * To indicate that track_pfn related cleanup is not
1679                  * needed from higher level routine calling unmap_vmas
1680                  */
1681                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1682                 return -EINVAL;
1683         }
1684
1685         BUG_ON(addr >= end);
1686         pfn -= addr >> PAGE_SHIFT;
1687         pgd = pgd_offset(mm, addr);
1688         flush_cache_range(vma, addr, end);
1689         do {
1690                 next = pgd_addr_end(addr, end);
1691                 err = remap_pud_range(mm, pgd, addr, next,
1692                                 pfn + (addr >> PAGE_SHIFT), prot);
1693                 if (err)
1694                         break;
1695         } while (pgd++, addr = next, addr != end);
1696
1697         if (err)
1698                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1699
1700         return err;
1701 }
1702 EXPORT_SYMBOL(remap_pfn_range);
1703
1704 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1705                                      unsigned long addr, unsigned long end,
1706                                      pte_fn_t fn, void *data)
1707 {
1708         pte_t *pte;
1709         int err;
1710         pgtable_t token;
1711         spinlock_t *uninitialized_var(ptl);
1712
1713         pte = (mm == &init_mm) ?
1714                 pte_alloc_kernel(pmd, addr) :
1715                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1716         if (!pte)
1717                 return -ENOMEM;
1718
1719         BUG_ON(pmd_huge(*pmd));
1720
1721         arch_enter_lazy_mmu_mode();
1722
1723         token = pmd_pgtable(*pmd);
1724
1725         do {
1726                 err = fn(pte, token, addr, data);
1727                 if (err)
1728                         break;
1729         } while (pte++, addr += PAGE_SIZE, addr != end);
1730
1731         arch_leave_lazy_mmu_mode();
1732
1733         if (mm != &init_mm)
1734                 pte_unmap_unlock(pte-1, ptl);
1735         return err;
1736 }
1737
1738 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1739                                      unsigned long addr, unsigned long end,
1740                                      pte_fn_t fn, void *data)
1741 {
1742         pmd_t *pmd;
1743         unsigned long next;
1744         int err;
1745
1746         BUG_ON(pud_huge(*pud));
1747
1748         pmd = pmd_alloc(mm, pud, addr);
1749         if (!pmd)
1750                 return -ENOMEM;
1751         do {
1752                 next = pmd_addr_end(addr, end);
1753                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1754                 if (err)
1755                         break;
1756         } while (pmd++, addr = next, addr != end);
1757         return err;
1758 }
1759
1760 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1761                                      unsigned long addr, unsigned long end,
1762                                      pte_fn_t fn, void *data)
1763 {
1764         pud_t *pud;
1765         unsigned long next;
1766         int err;
1767
1768         pud = pud_alloc(mm, pgd, addr);
1769         if (!pud)
1770                 return -ENOMEM;
1771         do {
1772                 next = pud_addr_end(addr, end);
1773                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1774                 if (err)
1775                         break;
1776         } while (pud++, addr = next, addr != end);
1777         return err;
1778 }
1779
1780 /*
1781  * Scan a region of virtual memory, filling in page tables as necessary
1782  * and calling a provided function on each leaf page table.
1783  */
1784 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1785                         unsigned long size, pte_fn_t fn, void *data)
1786 {
1787         pgd_t *pgd;
1788         unsigned long next;
1789         unsigned long start = addr, end = addr + size;
1790         int err;
1791
1792         BUG_ON(addr >= end);
1793         mmu_notifier_invalidate_range_start(mm, start, end);
1794         pgd = pgd_offset(mm, addr);
1795         do {
1796                 next = pgd_addr_end(addr, end);
1797                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1798                 if (err)
1799                         break;
1800         } while (pgd++, addr = next, addr != end);
1801         mmu_notifier_invalidate_range_end(mm, start, end);
1802         return err;
1803 }
1804 EXPORT_SYMBOL_GPL(apply_to_page_range);
1805
1806 /*
1807  * handle_pte_fault chooses page fault handler according to an entry
1808  * which was read non-atomically.  Before making any commitment, on
1809  * those architectures or configurations (e.g. i386 with PAE) which
1810  * might give a mix of unmatched parts, do_swap_page and do_file_page
1811  * must check under lock before unmapping the pte and proceeding
1812  * (but do_wp_page is only called after already making such a check;
1813  * and do_anonymous_page and do_no_page can safely check later on).
1814  */
1815 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1816                                 pte_t *page_table, pte_t orig_pte)
1817 {
1818         int same = 1;
1819 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1820         if (sizeof(pte_t) > sizeof(unsigned long)) {
1821                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1822                 spin_lock(ptl);
1823                 same = pte_same(*page_table, orig_pte);
1824                 spin_unlock(ptl);
1825         }
1826 #endif
1827         pte_unmap(page_table);
1828         return same;
1829 }
1830
1831 /*
1832  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1833  * servicing faults for write access.  In the normal case, do always want
1834  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1835  * that do not have writing enabled, when used by access_process_vm.
1836  */
1837 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1838 {
1839         if (likely(vma->vm_flags & VM_WRITE))
1840                 pte = pte_mkwrite(pte);
1841         return pte;
1842 }
1843
1844 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1845 {
1846         /*
1847          * If the source page was a PFN mapping, we don't have
1848          * a "struct page" for it. We do a best-effort copy by
1849          * just copying from the original user address. If that
1850          * fails, we just zero-fill it. Live with it.
1851          */
1852         if (unlikely(!src)) {
1853                 void *kaddr = kmap_atomic(dst, KM_USER0);
1854                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1855
1856                 /*
1857                  * This really shouldn't fail, because the page is there
1858                  * in the page tables. But it might just be unreadable,
1859                  * in which case we just give up and fill the result with
1860                  * zeroes.
1861                  */
1862                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1863                         memset(kaddr, 0, PAGE_SIZE);
1864                 kunmap_atomic(kaddr, KM_USER0);
1865                 flush_dcache_page(dst);
1866         } else
1867                 copy_user_highpage(dst, src, va, vma);
1868 }
1869
1870 /*
1871  * This routine handles present pages, when users try to write
1872  * to a shared page. It is done by copying the page to a new address
1873  * and decrementing the shared-page counter for the old page.
1874  *
1875  * Note that this routine assumes that the protection checks have been
1876  * done by the caller (the low-level page fault routine in most cases).
1877  * Thus we can safely just mark it writable once we've done any necessary
1878  * COW.
1879  *
1880  * We also mark the page dirty at this point even though the page will
1881  * change only once the write actually happens. This avoids a few races,
1882  * and potentially makes it more efficient.
1883  *
1884  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1885  * but allow concurrent faults), with pte both mapped and locked.
1886  * We return with mmap_sem still held, but pte unmapped and unlocked.
1887  */
1888 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1889                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1890                 spinlock_t *ptl, pte_t orig_pte)
1891 {
1892         struct page *old_page, *new_page;
1893         pte_t entry;
1894         int reuse = 0, ret = 0;
1895         int page_mkwrite = 0;
1896         struct page *dirty_page = NULL;
1897
1898         old_page = vm_normal_page(vma, address, orig_pte);
1899         if (!old_page) {
1900                 /*
1901                  * VM_MIXEDMAP !pfn_valid() case
1902                  *
1903                  * We should not cow pages in a shared writeable mapping.
1904                  * Just mark the pages writable as we can't do any dirty
1905                  * accounting on raw pfn maps.
1906                  */
1907                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1908                                      (VM_WRITE|VM_SHARED))
1909                         goto reuse;
1910                 goto gotten;
1911         }
1912
1913         /*
1914          * Take out anonymous pages first, anonymous shared vmas are
1915          * not dirty accountable.
1916          */
1917         if (PageAnon(old_page)) {
1918                 if (!trylock_page(old_page)) {
1919                         page_cache_get(old_page);
1920                         pte_unmap_unlock(page_table, ptl);
1921                         lock_page(old_page);
1922                         page_table = pte_offset_map_lock(mm, pmd, address,
1923                                                          &ptl);
1924                         if (!pte_same(*page_table, orig_pte)) {
1925                                 unlock_page(old_page);
1926                                 page_cache_release(old_page);
1927                                 goto unlock;
1928                         }
1929                         page_cache_release(old_page);
1930                 }
1931                 reuse = reuse_swap_page(old_page);
1932                 unlock_page(old_page);
1933         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1934                                         (VM_WRITE|VM_SHARED))) {
1935                 /*
1936                  * Only catch write-faults on shared writable pages,
1937                  * read-only shared pages can get COWed by
1938                  * get_user_pages(.write=1, .force=1).
1939                  */
1940                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1941                         /*
1942                          * Notify the address space that the page is about to
1943                          * become writable so that it can prohibit this or wait
1944                          * for the page to get into an appropriate state.
1945                          *
1946                          * We do this without the lock held, so that it can
1947                          * sleep if it needs to.
1948                          */
1949                         page_cache_get(old_page);
1950                         pte_unmap_unlock(page_table, ptl);
1951
1952                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1953                                 goto unwritable_page;
1954
1955                         /*
1956                          * Since we dropped the lock we need to revalidate
1957                          * the PTE as someone else may have changed it.  If
1958                          * they did, we just return, as we can count on the
1959                          * MMU to tell us if they didn't also make it writable.
1960                          */
1961                         page_table = pte_offset_map_lock(mm, pmd, address,
1962                                                          &ptl);
1963                         page_cache_release(old_page);
1964                         if (!pte_same(*page_table, orig_pte))
1965                                 goto unlock;
1966
1967                         page_mkwrite = 1;
1968                 }
1969                 dirty_page = old_page;
1970                 get_page(dirty_page);
1971                 reuse = 1;
1972         }
1973
1974         if (reuse) {
1975 reuse:
1976                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1977                 entry = pte_mkyoung(orig_pte);
1978                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1979                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1980                         update_mmu_cache(vma, address, entry);
1981                 ret |= VM_FAULT_WRITE;
1982                 goto unlock;
1983         }
1984
1985         /*
1986          * Ok, we need to copy. Oh, well..
1987          */
1988         page_cache_get(old_page);
1989 gotten:
1990         pte_unmap_unlock(page_table, ptl);
1991
1992         if (unlikely(anon_vma_prepare(vma)))
1993                 goto oom;
1994         VM_BUG_ON(old_page == ZERO_PAGE(0));
1995         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1996         if (!new_page)
1997                 goto oom;
1998         /*
1999          * Don't let another task, with possibly unlocked vma,
2000          * keep the mlocked page.
2001          */
2002         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2003                 lock_page(old_page);    /* for LRU manipulation */
2004                 clear_page_mlock(old_page);
2005                 unlock_page(old_page);
2006         }
2007         cow_user_page(new_page, old_page, address, vma);
2008         __SetPageUptodate(new_page);
2009
2010         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2011                 goto oom_free_new;
2012
2013         /*
2014          * Re-check the pte - we dropped the lock
2015          */
2016         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2017         if (likely(pte_same(*page_table, orig_pte))) {
2018                 if (old_page) {
2019                         if (!PageAnon(old_page)) {
2020                                 dec_mm_counter(mm, file_rss);
2021                                 inc_mm_counter(mm, anon_rss);
2022                         }
2023                 } else
2024                         inc_mm_counter(mm, anon_rss);
2025                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2026                 entry = mk_pte(new_page, vma->vm_page_prot);
2027                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2028                 /*
2029                  * Clear the pte entry and flush it first, before updating the
2030                  * pte with the new entry. This will avoid a race condition
2031                  * seen in the presence of one thread doing SMC and another
2032                  * thread doing COW.
2033                  */
2034                 ptep_clear_flush_notify(vma, address, page_table);
2035                 page_add_new_anon_rmap(new_page, vma, address);
2036                 set_pte_at(mm, address, page_table, entry);
2037                 update_mmu_cache(vma, address, entry);
2038                 if (old_page) {
2039                         /*
2040                          * Only after switching the pte to the new page may
2041                          * we remove the mapcount here. Otherwise another
2042                          * process may come and find the rmap count decremented
2043                          * before the pte is switched to the new page, and
2044                          * "reuse" the old page writing into it while our pte
2045                          * here still points into it and can be read by other
2046                          * threads.
2047                          *
2048                          * The critical issue is to order this
2049                          * page_remove_rmap with the ptp_clear_flush above.
2050                          * Those stores are ordered by (if nothing else,)
2051                          * the barrier present in the atomic_add_negative
2052                          * in page_remove_rmap.
2053                          *
2054                          * Then the TLB flush in ptep_clear_flush ensures that
2055                          * no process can access the old page before the
2056                          * decremented mapcount is visible. And the old page
2057                          * cannot be reused until after the decremented
2058                          * mapcount is visible. So transitively, TLBs to
2059                          * old page will be flushed before it can be reused.
2060                          */
2061                         page_remove_rmap(old_page);
2062                 }
2063
2064                 /* Free the old page.. */
2065                 new_page = old_page;
2066                 ret |= VM_FAULT_WRITE;
2067         } else
2068                 mem_cgroup_uncharge_page(new_page);
2069
2070         if (new_page)
2071                 page_cache_release(new_page);
2072         if (old_page)
2073                 page_cache_release(old_page);
2074 unlock:
2075         pte_unmap_unlock(page_table, ptl);
2076         if (dirty_page) {
2077                 if (vma->vm_file)
2078                         file_update_time(vma->vm_file);
2079
2080                 /*
2081                  * Yes, Virginia, this is actually required to prevent a race
2082                  * with clear_page_dirty_for_io() from clearing the page dirty
2083                  * bit after it clear all dirty ptes, but before a racing
2084                  * do_wp_page installs a dirty pte.
2085                  *
2086                  * do_no_page is protected similarly.
2087                  */
2088                 wait_on_page_locked(dirty_page);
2089                 set_page_dirty_balance(dirty_page, page_mkwrite);
2090                 put_page(dirty_page);
2091         }
2092         return ret;
2093 oom_free_new:
2094         page_cache_release(new_page);
2095 oom:
2096         if (old_page)
2097                 page_cache_release(old_page);
2098         return VM_FAULT_OOM;
2099
2100 unwritable_page:
2101         page_cache_release(old_page);
2102         return VM_FAULT_SIGBUS;
2103 }
2104
2105 /*
2106  * Helper functions for unmap_mapping_range().
2107  *
2108  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2109  *
2110  * We have to restart searching the prio_tree whenever we drop the lock,
2111  * since the iterator is only valid while the lock is held, and anyway
2112  * a later vma might be split and reinserted earlier while lock dropped.
2113  *
2114  * The list of nonlinear vmas could be handled more efficiently, using
2115  * a placeholder, but handle it in the same way until a need is shown.
2116  * It is important to search the prio_tree before nonlinear list: a vma
2117  * may become nonlinear and be shifted from prio_tree to nonlinear list
2118  * while the lock is dropped; but never shifted from list to prio_tree.
2119  *
2120  * In order to make forward progress despite restarting the search,
2121  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2122  * quickly skip it next time around.  Since the prio_tree search only
2123  * shows us those vmas affected by unmapping the range in question, we
2124  * can't efficiently keep all vmas in step with mapping->truncate_count:
2125  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2126  * mapping->truncate_count and vma->vm_truncate_count are protected by
2127  * i_mmap_lock.
2128  *
2129  * In order to make forward progress despite repeatedly restarting some
2130  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2131  * and restart from that address when we reach that vma again.  It might
2132  * have been split or merged, shrunk or extended, but never shifted: so
2133  * restart_addr remains valid so long as it remains in the vma's range.
2134  * unmap_mapping_range forces truncate_count to leap over page-aligned
2135  * values so we can save vma's restart_addr in its truncate_count field.
2136  */
2137 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2138
2139 static void reset_vma_truncate_counts(struct address_space *mapping)
2140 {
2141         struct vm_area_struct *vma;
2142         struct prio_tree_iter iter;
2143
2144         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2145                 vma->vm_truncate_count = 0;
2146         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2147                 vma->vm_truncate_count = 0;
2148 }
2149
2150 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2151                 unsigned long start_addr, unsigned long end_addr,
2152                 struct zap_details *details)
2153 {
2154         unsigned long restart_addr;
2155         int need_break;
2156
2157         /*
2158          * files that support invalidating or truncating portions of the
2159          * file from under mmaped areas must have their ->fault function
2160          * return a locked page (and set VM_FAULT_LOCKED in the return).
2161          * This provides synchronisation against concurrent unmapping here.
2162          */
2163
2164 again:
2165         restart_addr = vma->vm_truncate_count;
2166         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2167                 start_addr = restart_addr;
2168                 if (start_addr >= end_addr) {
2169                         /* Top of vma has been split off since last time */
2170                         vma->vm_truncate_count = details->truncate_count;
2171                         return 0;
2172                 }
2173         }
2174
2175         restart_addr = zap_page_range(vma, start_addr,
2176                                         end_addr - start_addr, details);
2177         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2178
2179         if (restart_addr >= end_addr) {
2180                 /* We have now completed this vma: mark it so */
2181                 vma->vm_truncate_count = details->truncate_count;
2182                 if (!need_break)
2183                         return 0;
2184         } else {
2185                 /* Note restart_addr in vma's truncate_count field */
2186                 vma->vm_truncate_count = restart_addr;
2187                 if (!need_break)
2188                         goto again;
2189         }
2190
2191         spin_unlock(details->i_mmap_lock);
2192         cond_resched();
2193         spin_lock(details->i_mmap_lock);
2194         return -EINTR;
2195 }
2196
2197 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2198                                             struct zap_details *details)
2199 {
2200         struct vm_area_struct *vma;
2201         struct prio_tree_iter iter;
2202         pgoff_t vba, vea, zba, zea;
2203
2204 restart:
2205         vma_prio_tree_foreach(vma, &iter, root,
2206                         details->first_index, details->last_index) {
2207                 /* Skip quickly over those we have already dealt with */
2208                 if (vma->vm_truncate_count == details->truncate_count)
2209                         continue;
2210
2211                 vba = vma->vm_pgoff;
2212                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2213                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2214                 zba = details->first_index;
2215                 if (zba < vba)
2216                         zba = vba;
2217                 zea = details->last_index;
2218                 if (zea > vea)
2219                         zea = vea;
2220
2221                 if (unmap_mapping_range_vma(vma,
2222                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2223                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2224                                 details) < 0)
2225                         goto restart;
2226         }
2227 }
2228
2229 static inline void unmap_mapping_range_list(struct list_head *head,
2230                                             struct zap_details *details)
2231 {
2232         struct vm_area_struct *vma;
2233
2234         /*
2235          * In nonlinear VMAs there is no correspondence between virtual address
2236          * offset and file offset.  So we must perform an exhaustive search
2237          * across *all* the pages in each nonlinear VMA, not just the pages
2238          * whose virtual address lies outside the file truncation point.
2239          */
2240 restart:
2241         list_for_each_entry(vma, head, shared.vm_set.list) {
2242                 /* Skip quickly over those we have already dealt with */
2243                 if (vma->vm_truncate_count == details->truncate_count)
2244                         continue;
2245                 details->nonlinear_vma = vma;
2246                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2247                                         vma->vm_end, details) < 0)
2248                         goto restart;
2249         }
2250 }
2251
2252 /**
2253  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2254  * @mapping: the address space containing mmaps to be unmapped.
2255  * @holebegin: byte in first page to unmap, relative to the start of
2256  * the underlying file.  This will be rounded down to a PAGE_SIZE
2257  * boundary.  Note that this is different from vmtruncate(), which
2258  * must keep the partial page.  In contrast, we must get rid of
2259  * partial pages.
2260  * @holelen: size of prospective hole in bytes.  This will be rounded
2261  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2262  * end of the file.
2263  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2264  * but 0 when invalidating pagecache, don't throw away private data.
2265  */
2266 void unmap_mapping_range(struct address_space *mapping,
2267                 loff_t const holebegin, loff_t const holelen, int even_cows)
2268 {
2269         struct zap_details details;
2270         pgoff_t hba = holebegin >> PAGE_SHIFT;
2271         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2272
2273         /* Check for overflow. */
2274         if (sizeof(holelen) > sizeof(hlen)) {
2275                 long long holeend =
2276                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2277                 if (holeend & ~(long long)ULONG_MAX)
2278                         hlen = ULONG_MAX - hba + 1;
2279         }
2280
2281         details.check_mapping = even_cows? NULL: mapping;
2282         details.nonlinear_vma = NULL;
2283         details.first_index = hba;
2284         details.last_index = hba + hlen - 1;
2285         if (details.last_index < details.first_index)
2286                 details.last_index = ULONG_MAX;
2287         details.i_mmap_lock = &mapping->i_mmap_lock;
2288
2289         spin_lock(&mapping->i_mmap_lock);
2290
2291         /* Protect against endless unmapping loops */
2292         mapping->truncate_count++;
2293         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2294                 if (mapping->truncate_count == 0)
2295                         reset_vma_truncate_counts(mapping);
2296                 mapping->truncate_count++;
2297         }
2298         details.truncate_count = mapping->truncate_count;
2299
2300         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2301                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2302         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2303                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2304         spin_unlock(&mapping->i_mmap_lock);
2305 }
2306 EXPORT_SYMBOL(unmap_mapping_range);
2307
2308 /**
2309  * vmtruncate - unmap mappings "freed" by truncate() syscall
2310  * @inode: inode of the file used
2311  * @offset: file offset to start truncating
2312  *
2313  * NOTE! We have to be ready to update the memory sharing
2314  * between the file and the memory map for a potential last
2315  * incomplete page.  Ugly, but necessary.
2316  */
2317 int vmtruncate(struct inode * inode, loff_t offset)
2318 {
2319         if (inode->i_size < offset) {
2320                 unsigned long limit;
2321
2322                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2323                 if (limit != RLIM_INFINITY && offset > limit)
2324                         goto out_sig;
2325                 if (offset > inode->i_sb->s_maxbytes)
2326                         goto out_big;
2327                 i_size_write(inode, offset);
2328         } else {
2329                 struct address_space *mapping = inode->i_mapping;
2330
2331                 /*
2332                  * truncation of in-use swapfiles is disallowed - it would
2333                  * cause subsequent swapout to scribble on the now-freed
2334                  * blocks.
2335                  */
2336                 if (IS_SWAPFILE(inode))
2337                         return -ETXTBSY;
2338                 i_size_write(inode, offset);
2339
2340                 /*
2341                  * unmap_mapping_range is called twice, first simply for
2342                  * efficiency so that truncate_inode_pages does fewer
2343                  * single-page unmaps.  However after this first call, and
2344                  * before truncate_inode_pages finishes, it is possible for
2345                  * private pages to be COWed, which remain after
2346                  * truncate_inode_pages finishes, hence the second
2347                  * unmap_mapping_range call must be made for correctness.
2348                  */
2349                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2350                 truncate_inode_pages(mapping, offset);
2351                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2352         }
2353
2354         if (inode->i_op->truncate)
2355                 inode->i_op->truncate(inode);
2356         return 0;
2357
2358 out_sig:
2359         send_sig(SIGXFSZ, current, 0);
2360 out_big:
2361         return -EFBIG;
2362 }
2363 EXPORT_SYMBOL(vmtruncate);
2364
2365 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2366 {
2367         struct address_space *mapping = inode->i_mapping;
2368
2369         /*
2370          * If the underlying filesystem is not going to provide
2371          * a way to truncate a range of blocks (punch a hole) -
2372          * we should return failure right now.
2373          */
2374         if (!inode->i_op->truncate_range)
2375                 return -ENOSYS;
2376
2377         mutex_lock(&inode->i_mutex);
2378         down_write(&inode->i_alloc_sem);
2379         unmap_mapping_range(mapping, offset, (end - offset), 1);
2380         truncate_inode_pages_range(mapping, offset, end);
2381         unmap_mapping_range(mapping, offset, (end - offset), 1);
2382         inode->i_op->truncate_range(inode, offset, end);
2383         up_write(&inode->i_alloc_sem);
2384         mutex_unlock(&inode->i_mutex);
2385
2386         return 0;
2387 }
2388
2389 /*
2390  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2391  * but allow concurrent faults), and pte mapped but not yet locked.
2392  * We return with mmap_sem still held, but pte unmapped and unlocked.
2393  */
2394 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2395                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2396                 int write_access, pte_t orig_pte)
2397 {
2398         spinlock_t *ptl;
2399         struct page *page;
2400         swp_entry_t entry;
2401         pte_t pte;
2402         struct mem_cgroup *ptr = NULL;
2403         int ret = 0;
2404
2405         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2406                 goto out;
2407
2408         entry = pte_to_swp_entry(orig_pte);
2409         if (is_migration_entry(entry)) {
2410                 migration_entry_wait(mm, pmd, address);
2411                 goto out;
2412         }
2413         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2414         page = lookup_swap_cache(entry);
2415         if (!page) {
2416                 grab_swap_token(); /* Contend for token _before_ read-in */
2417                 page = swapin_readahead(entry,
2418                                         GFP_HIGHUSER_MOVABLE, vma, address);
2419                 if (!page) {
2420                         /*
2421                          * Back out if somebody else faulted in this pte
2422                          * while we released the pte lock.
2423                          */
2424                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2425                         if (likely(pte_same(*page_table, orig_pte)))
2426                                 ret = VM_FAULT_OOM;
2427                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2428                         goto unlock;
2429                 }
2430
2431                 /* Had to read the page from swap area: Major fault */
2432                 ret = VM_FAULT_MAJOR;
2433                 count_vm_event(PGMAJFAULT);
2434         }
2435
2436         mark_page_accessed(page);
2437
2438         lock_page(page);
2439         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2440
2441         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2442                 ret = VM_FAULT_OOM;
2443                 unlock_page(page);
2444                 goto out;
2445         }
2446
2447         /*
2448          * Back out if somebody else already faulted in this pte.
2449          */
2450         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2451         if (unlikely(!pte_same(*page_table, orig_pte)))
2452                 goto out_nomap;
2453
2454         if (unlikely(!PageUptodate(page))) {
2455                 ret = VM_FAULT_SIGBUS;
2456                 goto out_nomap;
2457         }
2458
2459         /*
2460          * The page isn't present yet, go ahead with the fault.
2461          *
2462          * Be careful about the sequence of operations here.
2463          * To get its accounting right, reuse_swap_page() must be called
2464          * while the page is counted on swap but not yet in mapcount i.e.
2465          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2466          * must be called after the swap_free(), or it will never succeed.
2467          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2468          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2469          * in page->private. In this case, a record in swap_cgroup  is silently
2470          * discarded at swap_free().
2471          */
2472
2473         inc_mm_counter(mm, anon_rss);
2474         pte = mk_pte(page, vma->vm_page_prot);
2475         if (write_access && reuse_swap_page(page)) {
2476                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2477                 write_access = 0;
2478         }
2479         flush_icache_page(vma, page);
2480         set_pte_at(mm, address, page_table, pte);
2481         page_add_anon_rmap(page, vma, address);
2482         /* It's better to call commit-charge after rmap is established */
2483         mem_cgroup_commit_charge_swapin(page, ptr);
2484
2485         swap_free(entry);
2486         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2487                 try_to_free_swap(page);
2488         unlock_page(page);
2489
2490         if (write_access) {
2491                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2492                 if (ret & VM_FAULT_ERROR)
2493                         ret &= VM_FAULT_ERROR;
2494                 goto out;
2495         }
2496
2497         /* No need to invalidate - it was non-present before */
2498         update_mmu_cache(vma, address, pte);
2499 unlock:
2500         pte_unmap_unlock(page_table, ptl);
2501 out:
2502         return ret;
2503 out_nomap:
2504         mem_cgroup_cancel_charge_swapin(ptr);
2505         pte_unmap_unlock(page_table, ptl);
2506         unlock_page(page);
2507         page_cache_release(page);
2508         return ret;
2509 }
2510
2511 /*
2512  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2513  * but allow concurrent faults), and pte mapped but not yet locked.
2514  * We return with mmap_sem still held, but pte unmapped and unlocked.
2515  */
2516 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2517                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2518                 int write_access)
2519 {
2520         struct page *page;
2521         spinlock_t *ptl;
2522         pte_t entry;
2523
2524         /* Allocate our own private page. */
2525         pte_unmap(page_table);
2526
2527         if (unlikely(anon_vma_prepare(vma)))
2528                 goto oom;
2529         page = alloc_zeroed_user_highpage_movable(vma, address);
2530         if (!page)
2531                 goto oom;
2532         __SetPageUptodate(page);
2533
2534         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2535                 goto oom_free_page;
2536
2537         entry = mk_pte(page, vma->vm_page_prot);
2538         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2539
2540         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2541         if (!pte_none(*page_table))
2542                 goto release;
2543         inc_mm_counter(mm, anon_rss);
2544         page_add_new_anon_rmap(page, vma, address);
2545         set_pte_at(mm, address, page_table, entry);
2546
2547         /* No need to invalidate - it was non-present before */
2548         update_mmu_cache(vma, address, entry);
2549 unlock:
2550         pte_unmap_unlock(page_table, ptl);
2551         return 0;
2552 release:
2553         mem_cgroup_uncharge_page(page);
2554         page_cache_release(page);
2555         goto unlock;
2556 oom_free_page:
2557         page_cache_release(page);
2558 oom:
2559         return VM_FAULT_OOM;
2560 }
2561
2562 /*
2563  * __do_fault() tries to create a new page mapping. It aggressively
2564  * tries to share with existing pages, but makes a separate copy if
2565  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2566  * the next page fault.
2567  *
2568  * As this is called only for pages that do not currently exist, we
2569  * do not need to flush old virtual caches or the TLB.
2570  *
2571  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2572  * but allow concurrent faults), and pte neither mapped nor locked.
2573  * We return with mmap_sem still held, but pte unmapped and unlocked.
2574  */
2575 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2576                 unsigned long address, pmd_t *pmd,
2577                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2578 {
2579         pte_t *page_table;
2580         spinlock_t *ptl;
2581         struct page *page;
2582         pte_t entry;
2583         int anon = 0;
2584         int charged = 0;
2585         struct page *dirty_page = NULL;
2586         struct vm_fault vmf;
2587         int ret;
2588         int page_mkwrite = 0;
2589
2590         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2591         vmf.pgoff = pgoff;
2592         vmf.flags = flags;
2593         vmf.page = NULL;
2594
2595         ret = vma->vm_ops->fault(vma, &vmf);
2596         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2597                 return ret;
2598
2599         /*
2600          * For consistency in subsequent calls, make the faulted page always
2601          * locked.
2602          */
2603         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2604                 lock_page(vmf.page);
2605         else
2606                 VM_BUG_ON(!PageLocked(vmf.page));
2607
2608         /*
2609          * Should we do an early C-O-W break?
2610          */
2611         page = vmf.page;
2612         if (flags & FAULT_FLAG_WRITE) {
2613                 if (!(vma->vm_flags & VM_SHARED)) {
2614                         anon = 1;
2615                         if (unlikely(anon_vma_prepare(vma))) {
2616                                 ret = VM_FAULT_OOM;
2617                                 goto out;
2618                         }
2619                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2620                                                 vma, address);
2621                         if (!page) {
2622                                 ret = VM_FAULT_OOM;
2623                                 goto out;
2624                         }
2625                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2626                                 ret = VM_FAULT_OOM;
2627                                 page_cache_release(page);
2628                                 goto out;
2629                         }
2630                         charged = 1;
2631                         /*
2632                          * Don't let another task, with possibly unlocked vma,
2633                          * keep the mlocked page.
2634                          */
2635                         if (vma->vm_flags & VM_LOCKED)
2636                                 clear_page_mlock(vmf.page);
2637                         copy_user_highpage(page, vmf.page, address, vma);
2638                         __SetPageUptodate(page);
2639                 } else {
2640                         /*
2641                          * If the page will be shareable, see if the backing
2642                          * address space wants to know that the page is about
2643                          * to become writable
2644                          */
2645                         if (vma->vm_ops->page_mkwrite) {
2646                                 unlock_page(page);
2647                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2648                                         ret = VM_FAULT_SIGBUS;
2649                                         anon = 1; /* no anon but release vmf.page */
2650                                         goto out_unlocked;
2651                                 }
2652                                 lock_page(page);
2653                                 /*
2654                                  * XXX: this is not quite right (racy vs
2655                                  * invalidate) to unlock and relock the page
2656                                  * like this, however a better fix requires
2657                                  * reworking page_mkwrite locking API, which
2658                                  * is better done later.
2659                                  */
2660                                 if (!page->mapping) {
2661                                         ret = 0;
2662                                         anon = 1; /* no anon but release vmf.page */
2663                                         goto out;
2664                                 }
2665                                 page_mkwrite = 1;
2666                         }
2667                 }
2668
2669         }
2670
2671         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2672
2673         /*
2674          * This silly early PAGE_DIRTY setting removes a race
2675          * due to the bad i386 page protection. But it's valid
2676          * for other architectures too.
2677          *
2678          * Note that if write_access is true, we either now have
2679          * an exclusive copy of the page, or this is a shared mapping,
2680          * so we can make it writable and dirty to avoid having to
2681          * handle that later.
2682          */
2683         /* Only go through if we didn't race with anybody else... */
2684         if (likely(pte_same(*page_table, orig_pte))) {
2685                 flush_icache_page(vma, page);
2686                 entry = mk_pte(page, vma->vm_page_prot);
2687                 if (flags & FAULT_FLAG_WRITE)
2688                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2689                 if (anon) {
2690                         inc_mm_counter(mm, anon_rss);
2691                         page_add_new_anon_rmap(page, vma, address);
2692                 } else {
2693                         inc_mm_counter(mm, file_rss);
2694                         page_add_file_rmap(page);
2695                         if (flags & FAULT_FLAG_WRITE) {
2696                                 dirty_page = page;
2697                                 get_page(dirty_page);
2698                         }
2699                 }
2700                 set_pte_at(mm, address, page_table, entry);
2701
2702                 /* no need to invalidate: a not-present page won't be cached */
2703                 update_mmu_cache(vma, address, entry);
2704         } else {
2705                 if (charged)
2706                         mem_cgroup_uncharge_page(page);
2707                 if (anon)
2708                         page_cache_release(page);
2709                 else
2710                         anon = 1; /* no anon but release faulted_page */
2711         }
2712
2713         pte_unmap_unlock(page_table, ptl);
2714
2715 out:
2716         unlock_page(vmf.page);
2717 out_unlocked:
2718         if (anon)
2719                 page_cache_release(vmf.page);
2720         else if (dirty_page) {
2721                 if (vma->vm_file)
2722                         file_update_time(vma->vm_file);
2723
2724                 set_page_dirty_balance(dirty_page, page_mkwrite);
2725                 put_page(dirty_page);
2726         }
2727
2728         return ret;
2729 }
2730
2731 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2732                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2733                 int write_access, pte_t orig_pte)
2734 {
2735         pgoff_t pgoff = (((address & PAGE_MASK)
2736                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2737         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2738
2739         pte_unmap(page_table);
2740         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2741 }
2742
2743 /*
2744  * Fault of a previously existing named mapping. Repopulate the pte
2745  * from the encoded file_pte if possible. This enables swappable
2746  * nonlinear vmas.
2747  *
2748  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2749  * but allow concurrent faults), and pte mapped but not yet locked.
2750  * We return with mmap_sem still held, but pte unmapped and unlocked.
2751  */
2752 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2753                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2754                 int write_access, pte_t orig_pte)
2755 {
2756         unsigned int flags = FAULT_FLAG_NONLINEAR |
2757                                 (write_access ? FAULT_FLAG_WRITE : 0);
2758         pgoff_t pgoff;
2759
2760         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2761                 return 0;
2762
2763         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2764                 /*
2765                  * Page table corrupted: show pte and kill process.
2766                  */
2767                 print_bad_pte(vma, address, orig_pte, NULL);
2768                 return VM_FAULT_OOM;
2769         }
2770
2771         pgoff = pte_to_pgoff(orig_pte);
2772         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2773 }
2774
2775 /*
2776  * These routines also need to handle stuff like marking pages dirty
2777  * and/or accessed for architectures that don't do it in hardware (most
2778  * RISC architectures).  The early dirtying is also good on the i386.
2779  *
2780  * There is also a hook called "update_mmu_cache()" that architectures
2781  * with external mmu caches can use to update those (ie the Sparc or
2782  * PowerPC hashed page tables that act as extended TLBs).
2783  *
2784  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2785  * but allow concurrent faults), and pte mapped but not yet locked.
2786  * We return with mmap_sem still held, but pte unmapped and unlocked.
2787  */
2788 static inline int handle_pte_fault(struct mm_struct *mm,
2789                 struct vm_area_struct *vma, unsigned long address,
2790                 pte_t *pte, pmd_t *pmd, int write_access)
2791 {
2792         pte_t entry;
2793         spinlock_t *ptl;
2794
2795         entry = *pte;
2796         if (!pte_present(entry)) {
2797                 if (pte_none(entry)) {
2798                         if (vma->vm_ops) {
2799                                 if (likely(vma->vm_ops->fault))
2800                                         return do_linear_fault(mm, vma, address,
2801                                                 pte, pmd, write_access, entry);
2802                         }
2803                         return do_anonymous_page(mm, vma, address,
2804                                                  pte, pmd, write_access);
2805                 }
2806                 if (pte_file(entry))
2807                         return do_nonlinear_fault(mm, vma, address,
2808                                         pte, pmd, write_access, entry);
2809                 return do_swap_page(mm, vma, address,
2810                                         pte, pmd, write_access, entry);
2811         }
2812
2813         ptl = pte_lockptr(mm, pmd);
2814         spin_lock(ptl);
2815         if (unlikely(!pte_same(*pte, entry)))
2816                 goto unlock;
2817         if (write_access) {
2818                 if (!pte_write(entry))
2819                         return do_wp_page(mm, vma, address,
2820                                         pte, pmd, ptl, entry);
2821                 entry = pte_mkdirty(entry);
2822         }
2823         entry = pte_mkyoung(entry);
2824         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2825                 update_mmu_cache(vma, address, entry);
2826         } else {
2827                 /*
2828                  * This is needed only for protection faults but the arch code
2829                  * is not yet telling us if this is a protection fault or not.
2830                  * This still avoids useless tlb flushes for .text page faults
2831                  * with threads.
2832                  */
2833                 if (write_access)
2834                         flush_tlb_page(vma, address);
2835         }
2836 unlock:
2837         pte_unmap_unlock(pte, ptl);
2838         return 0;
2839 }
2840
2841 /*
2842  * By the time we get here, we already hold the mm semaphore
2843  */
2844 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2845                 unsigned long address, int write_access)
2846 {
2847         pgd_t *pgd;
2848         pud_t *pud;
2849         pmd_t *pmd;
2850         pte_t *pte;
2851
2852         __set_current_state(TASK_RUNNING);
2853
2854         count_vm_event(PGFAULT);
2855
2856         if (unlikely(is_vm_hugetlb_page(vma)))
2857                 return hugetlb_fault(mm, vma, address, write_access);
2858
2859         pgd = pgd_offset(mm, address);
2860         pud = pud_alloc(mm, pgd, address);
2861         if (!pud)
2862                 return VM_FAULT_OOM;
2863         pmd = pmd_alloc(mm, pud, address);
2864         if (!pmd)
2865                 return VM_FAULT_OOM;
2866         pte = pte_alloc_map(mm, pmd, address);
2867         if (!pte)
2868                 return VM_FAULT_OOM;
2869
2870         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2871 }
2872
2873 #ifndef __PAGETABLE_PUD_FOLDED
2874 /*
2875  * Allocate page upper directory.
2876  * We've already handled the fast-path in-line.
2877  */
2878 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2879 {
2880         pud_t *new = pud_alloc_one(mm, address);
2881         if (!new)
2882                 return -ENOMEM;
2883
2884         smp_wmb(); /* See comment in __pte_alloc */
2885
2886         spin_lock(&mm->page_table_lock);
2887         if (pgd_present(*pgd))          /* Another has populated it */
2888                 pud_free(mm, new);
2889         else
2890                 pgd_populate(mm, pgd, new);
2891         spin_unlock(&mm->page_table_lock);
2892         return 0;
2893 }
2894 #endif /* __PAGETABLE_PUD_FOLDED */
2895
2896 #ifndef __PAGETABLE_PMD_FOLDED
2897 /*
2898  * Allocate page middle directory.
2899  * We've already handled the fast-path in-line.
2900  */
2901 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2902 {
2903         pmd_t *new = pmd_alloc_one(mm, address);
2904         if (!new)
2905                 return -ENOMEM;
2906
2907         smp_wmb(); /* See comment in __pte_alloc */
2908
2909         spin_lock(&mm->page_table_lock);
2910 #ifndef __ARCH_HAS_4LEVEL_HACK
2911         if (pud_present(*pud))          /* Another has populated it */
2912                 pmd_free(mm, new);
2913         else
2914                 pud_populate(mm, pud, new);
2915 #else
2916         if (pgd_present(*pud))          /* Another has populated it */
2917                 pmd_free(mm, new);
2918         else
2919                 pgd_populate(mm, pud, new);
2920 #endif /* __ARCH_HAS_4LEVEL_HACK */
2921         spin_unlock(&mm->page_table_lock);
2922         return 0;
2923 }
2924 #endif /* __PAGETABLE_PMD_FOLDED */
2925
2926 int make_pages_present(unsigned long addr, unsigned long end)
2927 {
2928         int ret, len, write;
2929         struct vm_area_struct * vma;
2930
2931         vma = find_vma(current->mm, addr);
2932         if (!vma)
2933                 return -ENOMEM;
2934         write = (vma->vm_flags & VM_WRITE) != 0;
2935         BUG_ON(addr >= end);
2936         BUG_ON(end > vma->vm_end);
2937         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2938         ret = get_user_pages(current, current->mm, addr,
2939                         len, write, 0, NULL, NULL);
2940         if (ret < 0)
2941                 return ret;
2942         return ret == len ? 0 : -EFAULT;
2943 }
2944
2945 #if !defined(__HAVE_ARCH_GATE_AREA)
2946
2947 #if defined(AT_SYSINFO_EHDR)
2948 static struct vm_area_struct gate_vma;
2949
2950 static int __init gate_vma_init(void)
2951 {
2952         gate_vma.vm_mm = NULL;
2953         gate_vma.vm_start = FIXADDR_USER_START;
2954         gate_vma.vm_end = FIXADDR_USER_END;
2955         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2956         gate_vma.vm_page_prot = __P101;
2957         /*
2958          * Make sure the vDSO gets into every core dump.
2959          * Dumping its contents makes post-mortem fully interpretable later
2960          * without matching up the same kernel and hardware config to see
2961          * what PC values meant.
2962          */
2963         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2964         return 0;
2965 }
2966 __initcall(gate_vma_init);
2967 #endif
2968
2969 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2970 {
2971 #ifdef AT_SYSINFO_EHDR
2972         return &gate_vma;
2973 #else
2974         return NULL;
2975 #endif
2976 }
2977
2978 int in_gate_area_no_task(unsigned long addr)
2979 {
2980 #ifdef AT_SYSINFO_EHDR
2981         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2982                 return 1;
2983 #endif
2984         return 0;
2985 }
2986
2987 #endif  /* __HAVE_ARCH_GATE_AREA */
2988
2989 #ifdef CONFIG_HAVE_IOREMAP_PROT
2990 int follow_phys(struct vm_area_struct *vma,
2991                 unsigned long address, unsigned int flags,
2992                 unsigned long *prot, resource_size_t *phys)
2993 {
2994         pgd_t *pgd;
2995         pud_t *pud;
2996         pmd_t *pmd;
2997         pte_t *ptep, pte;
2998         spinlock_t *ptl;
2999         resource_size_t phys_addr = 0;
3000         struct mm_struct *mm = vma->vm_mm;
3001         int ret = -EINVAL;
3002
3003         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3004                 goto out;
3005
3006         pgd = pgd_offset(mm, address);
3007         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3008                 goto out;
3009
3010         pud = pud_offset(pgd, address);
3011         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3012                 goto out;
3013
3014         pmd = pmd_offset(pud, address);
3015         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3016                 goto out;
3017
3018         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3019         if (pmd_huge(*pmd))
3020                 goto out;
3021
3022         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
3023         if (!ptep)
3024                 goto out;
3025
3026         pte = *ptep;
3027         if (!pte_present(pte))
3028                 goto unlock;
3029         if ((flags & FOLL_WRITE) && !pte_write(pte))
3030                 goto unlock;
3031         phys_addr = pte_pfn(pte);
3032         phys_addr <<= PAGE_SHIFT; /* Shift here to avoid overflow on PAE */
3033
3034         *prot = pgprot_val(pte_pgprot(pte));
3035         *phys = phys_addr;
3036         ret = 0;
3037
3038 unlock:
3039         pte_unmap_unlock(ptep, ptl);
3040 out:
3041         return ret;
3042 }
3043
3044 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3045                         void *buf, int len, int write)
3046 {
3047         resource_size_t phys_addr;
3048         unsigned long prot = 0;
3049         void __iomem *maddr;
3050         int offset = addr & (PAGE_SIZE-1);
3051
3052         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3053                 return -EINVAL;
3054
3055         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3056         if (write)
3057                 memcpy_toio(maddr + offset, buf, len);
3058         else
3059                 memcpy_fromio(buf, maddr + offset, len);
3060         iounmap(maddr);
3061
3062         return len;
3063 }
3064 #endif
3065
3066 /*
3067  * Access another process' address space.
3068  * Source/target buffer must be kernel space,
3069  * Do not walk the page table directly, use get_user_pages
3070  */
3071 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3072 {
3073         struct mm_struct *mm;
3074         struct vm_area_struct *vma;
3075         void *old_buf = buf;
3076
3077         mm = get_task_mm(tsk);
3078         if (!mm)
3079                 return 0;
3080
3081         down_read(&mm->mmap_sem);
3082         /* ignore errors, just check how much was successfully transferred */
3083         while (len) {
3084                 int bytes, ret, offset;
3085                 void *maddr;
3086                 struct page *page = NULL;
3087
3088                 ret = get_user_pages(tsk, mm, addr, 1,
3089                                 write, 1, &page, &vma);
3090                 if (ret <= 0) {
3091                         /*
3092                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3093                          * we can access using slightly different code.
3094                          */
3095 #ifdef CONFIG_HAVE_IOREMAP_PROT
3096                         vma = find_vma(mm, addr);
3097                         if (!vma)
3098                                 break;
3099                         if (vma->vm_ops && vma->vm_ops->access)
3100                                 ret = vma->vm_ops->access(vma, addr, buf,
3101                                                           len, write);
3102                         if (ret <= 0)
3103 #endif
3104                                 break;
3105                         bytes = ret;
3106                 } else {
3107                         bytes = len;
3108                         offset = addr & (PAGE_SIZE-1);
3109                         if (bytes > PAGE_SIZE-offset)
3110                                 bytes = PAGE_SIZE-offset;
3111
3112                         maddr = kmap(page);
3113                         if (write) {
3114                                 copy_to_user_page(vma, page, addr,
3115                                                   maddr + offset, buf, bytes);
3116                                 set_page_dirty_lock(page);
3117                         } else {
3118                                 copy_from_user_page(vma, page, addr,
3119                                                     buf, maddr + offset, bytes);
3120                         }
3121                         kunmap(page);
3122                         page_cache_release(page);
3123                 }
3124                 len -= bytes;
3125                 buf += bytes;
3126                 addr += bytes;
3127         }
3128         up_read(&mm->mmap_sem);
3129         mmput(mm);
3130
3131         return buf - old_buf;
3132 }
3133
3134 /*
3135  * Print the name of a VMA.
3136  */
3137 void print_vma_addr(char *prefix, unsigned long ip)
3138 {
3139         struct mm_struct *mm = current->mm;
3140         struct vm_area_struct *vma;
3141
3142         /*
3143          * Do not print if we are in atomic
3144          * contexts (in exception stacks, etc.):
3145          */
3146         if (preempt_count())
3147                 return;
3148
3149         down_read(&mm->mmap_sem);
3150         vma = find_vma(mm, ip);
3151         if (vma && vma->vm_file) {
3152                 struct file *f = vma->vm_file;
3153                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3154                 if (buf) {
3155                         char *p, *s;
3156
3157                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3158                         if (IS_ERR(p))
3159                                 p = "?";
3160                         s = strrchr(p, '/');
3161                         if (s)
3162                                 p = s+1;
3163                         printk("%s%s[%lx+%lx]", prefix, p,
3164                                         vma->vm_start,
3165                                         vma->vm_end - vma->vm_start);
3166                         free_page((unsigned long)buf);
3167                 }
3168         }
3169         up_read(&current->mm->mmap_sem);
3170 }
3171
3172 #ifdef CONFIG_PROVE_LOCKING
3173 void might_fault(void)
3174 {
3175         /*
3176          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3177          * holding the mmap_sem, this is safe because kernel memory doesn't
3178          * get paged out, therefore we'll never actually fault, and the
3179          * below annotations will generate false positives.
3180          */
3181         if (segment_eq(get_fs(), KERNEL_DS))
3182                 return;
3183
3184         might_sleep();
3185         /*
3186          * it would be nicer only to annotate paths which are not under
3187          * pagefault_disable, however that requires a larger audit and
3188          * providing helpers like get_user_atomic.
3189          */
3190         if (!in_atomic() && current->mm)
3191                 might_lock_read(&current->mm->mmap_sem);
3192 }
3193 EXPORT_SYMBOL(might_fault);
3194 #endif