hugetlb: introduce pud_huge
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/delayacct.h>
51 #include <linux/init.h>
52 #include <linux/writeback.h>
53 #include <linux/memcontrol.h>
54
55 #include <asm/pgalloc.h>
56 #include <asm/uaccess.h>
57 #include <asm/tlb.h>
58 #include <asm/tlbflush.h>
59 #include <asm/pgtable.h>
60
61 #include <linux/swapops.h>
62 #include <linux/elf.h>
63
64 #include "internal.h"
65
66 #ifndef CONFIG_NEED_MULTIPLE_NODES
67 /* use the per-pgdat data instead for discontigmem - mbligh */
68 unsigned long max_mapnr;
69 struct page *mem_map;
70
71 EXPORT_SYMBOL(max_mapnr);
72 EXPORT_SYMBOL(mem_map);
73 #endif
74
75 unsigned long num_physpages;
76 /*
77  * A number of key systems in x86 including ioremap() rely on the assumption
78  * that high_memory defines the upper bound on direct map memory, then end
79  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
80  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
81  * and ZONE_HIGHMEM.
82  */
83 void * high_memory;
84
85 EXPORT_SYMBOL(num_physpages);
86 EXPORT_SYMBOL(high_memory);
87
88 /*
89  * Randomize the address space (stacks, mmaps, brk, etc.).
90  *
91  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
92  *   as ancient (libc5 based) binaries can segfault. )
93  */
94 int randomize_va_space __read_mostly =
95 #ifdef CONFIG_COMPAT_BRK
96                                         1;
97 #else
98                                         2;
99 #endif
100
101 static int __init disable_randmaps(char *s)
102 {
103         randomize_va_space = 0;
104         return 1;
105 }
106 __setup("norandmaps", disable_randmaps);
107
108
109 /*
110  * If a p?d_bad entry is found while walking page tables, report
111  * the error, before resetting entry to p?d_none.  Usually (but
112  * very seldom) called out from the p?d_none_or_clear_bad macros.
113  */
114
115 void pgd_clear_bad(pgd_t *pgd)
116 {
117         pgd_ERROR(*pgd);
118         pgd_clear(pgd);
119 }
120
121 void pud_clear_bad(pud_t *pud)
122 {
123         pud_ERROR(*pud);
124         pud_clear(pud);
125 }
126
127 void pmd_clear_bad(pmd_t *pmd)
128 {
129         pmd_ERROR(*pmd);
130         pmd_clear(pmd);
131 }
132
133 /*
134  * Note: this doesn't free the actual pages themselves. That
135  * has been handled earlier when unmapping all the memory regions.
136  */
137 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
138 {
139         pgtable_t token = pmd_pgtable(*pmd);
140         pmd_clear(pmd);
141         pte_free_tlb(tlb, token);
142         tlb->mm->nr_ptes--;
143 }
144
145 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
146                                 unsigned long addr, unsigned long end,
147                                 unsigned long floor, unsigned long ceiling)
148 {
149         pmd_t *pmd;
150         unsigned long next;
151         unsigned long start;
152
153         start = addr;
154         pmd = pmd_offset(pud, addr);
155         do {
156                 next = pmd_addr_end(addr, end);
157                 if (pmd_none_or_clear_bad(pmd))
158                         continue;
159                 free_pte_range(tlb, pmd);
160         } while (pmd++, addr = next, addr != end);
161
162         start &= PUD_MASK;
163         if (start < floor)
164                 return;
165         if (ceiling) {
166                 ceiling &= PUD_MASK;
167                 if (!ceiling)
168                         return;
169         }
170         if (end - 1 > ceiling - 1)
171                 return;
172
173         pmd = pmd_offset(pud, start);
174         pud_clear(pud);
175         pmd_free_tlb(tlb, pmd);
176 }
177
178 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
179                                 unsigned long addr, unsigned long end,
180                                 unsigned long floor, unsigned long ceiling)
181 {
182         pud_t *pud;
183         unsigned long next;
184         unsigned long start;
185
186         start = addr;
187         pud = pud_offset(pgd, addr);
188         do {
189                 next = pud_addr_end(addr, end);
190                 if (pud_none_or_clear_bad(pud))
191                         continue;
192                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
193         } while (pud++, addr = next, addr != end);
194
195         start &= PGDIR_MASK;
196         if (start < floor)
197                 return;
198         if (ceiling) {
199                 ceiling &= PGDIR_MASK;
200                 if (!ceiling)
201                         return;
202         }
203         if (end - 1 > ceiling - 1)
204                 return;
205
206         pud = pud_offset(pgd, start);
207         pgd_clear(pgd);
208         pud_free_tlb(tlb, pud);
209 }
210
211 /*
212  * This function frees user-level page tables of a process.
213  *
214  * Must be called with pagetable lock held.
215  */
216 void free_pgd_range(struct mmu_gather *tlb,
217                         unsigned long addr, unsigned long end,
218                         unsigned long floor, unsigned long ceiling)
219 {
220         pgd_t *pgd;
221         unsigned long next;
222         unsigned long start;
223
224         /*
225          * The next few lines have given us lots of grief...
226          *
227          * Why are we testing PMD* at this top level?  Because often
228          * there will be no work to do at all, and we'd prefer not to
229          * go all the way down to the bottom just to discover that.
230          *
231          * Why all these "- 1"s?  Because 0 represents both the bottom
232          * of the address space and the top of it (using -1 for the
233          * top wouldn't help much: the masks would do the wrong thing).
234          * The rule is that addr 0 and floor 0 refer to the bottom of
235          * the address space, but end 0 and ceiling 0 refer to the top
236          * Comparisons need to use "end - 1" and "ceiling - 1" (though
237          * that end 0 case should be mythical).
238          *
239          * Wherever addr is brought up or ceiling brought down, we must
240          * be careful to reject "the opposite 0" before it confuses the
241          * subsequent tests.  But what about where end is brought down
242          * by PMD_SIZE below? no, end can't go down to 0 there.
243          *
244          * Whereas we round start (addr) and ceiling down, by different
245          * masks at different levels, in order to test whether a table
246          * now has no other vmas using it, so can be freed, we don't
247          * bother to round floor or end up - the tests don't need that.
248          */
249
250         addr &= PMD_MASK;
251         if (addr < floor) {
252                 addr += PMD_SIZE;
253                 if (!addr)
254                         return;
255         }
256         if (ceiling) {
257                 ceiling &= PMD_MASK;
258                 if (!ceiling)
259                         return;
260         }
261         if (end - 1 > ceiling - 1)
262                 end -= PMD_SIZE;
263         if (addr > end - 1)
264                 return;
265
266         start = addr;
267         pgd = pgd_offset(tlb->mm, addr);
268         do {
269                 next = pgd_addr_end(addr, end);
270                 if (pgd_none_or_clear_bad(pgd))
271                         continue;
272                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
273         } while (pgd++, addr = next, addr != end);
274 }
275
276 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
277                 unsigned long floor, unsigned long ceiling)
278 {
279         while (vma) {
280                 struct vm_area_struct *next = vma->vm_next;
281                 unsigned long addr = vma->vm_start;
282
283                 /*
284                  * Hide vma from rmap and vmtruncate before freeing pgtables
285                  */
286                 anon_vma_unlink(vma);
287                 unlink_file_vma(vma);
288
289                 if (is_vm_hugetlb_page(vma)) {
290                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
291                                 floor, next? next->vm_start: ceiling);
292                 } else {
293                         /*
294                          * Optimization: gather nearby vmas into one call down
295                          */
296                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
297                                && !is_vm_hugetlb_page(next)) {
298                                 vma = next;
299                                 next = vma->vm_next;
300                                 anon_vma_unlink(vma);
301                                 unlink_file_vma(vma);
302                         }
303                         free_pgd_range(tlb, addr, vma->vm_end,
304                                 floor, next? next->vm_start: ceiling);
305                 }
306                 vma = next;
307         }
308 }
309
310 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
311 {
312         pgtable_t new = pte_alloc_one(mm, address);
313         if (!new)
314                 return -ENOMEM;
315
316         /*
317          * Ensure all pte setup (eg. pte page lock and page clearing) are
318          * visible before the pte is made visible to other CPUs by being
319          * put into page tables.
320          *
321          * The other side of the story is the pointer chasing in the page
322          * table walking code (when walking the page table without locking;
323          * ie. most of the time). Fortunately, these data accesses consist
324          * of a chain of data-dependent loads, meaning most CPUs (alpha
325          * being the notable exception) will already guarantee loads are
326          * seen in-order. See the alpha page table accessors for the
327          * smp_read_barrier_depends() barriers in page table walking code.
328          */
329         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
330
331         spin_lock(&mm->page_table_lock);
332         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
333                 mm->nr_ptes++;
334                 pmd_populate(mm, pmd, new);
335                 new = NULL;
336         }
337         spin_unlock(&mm->page_table_lock);
338         if (new)
339                 pte_free(mm, new);
340         return 0;
341 }
342
343 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
344 {
345         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
346         if (!new)
347                 return -ENOMEM;
348
349         smp_wmb(); /* See comment in __pte_alloc */
350
351         spin_lock(&init_mm.page_table_lock);
352         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
353                 pmd_populate_kernel(&init_mm, pmd, new);
354                 new = NULL;
355         }
356         spin_unlock(&init_mm.page_table_lock);
357         if (new)
358                 pte_free_kernel(&init_mm, new);
359         return 0;
360 }
361
362 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
363 {
364         if (file_rss)
365                 add_mm_counter(mm, file_rss, file_rss);
366         if (anon_rss)
367                 add_mm_counter(mm, anon_rss, anon_rss);
368 }
369
370 /*
371  * This function is called to print an error when a bad pte
372  * is found. For example, we might have a PFN-mapped pte in
373  * a region that doesn't allow it.
374  *
375  * The calling function must still handle the error.
376  */
377 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
378 {
379         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
380                         "vm_flags = %lx, vaddr = %lx\n",
381                 (long long)pte_val(pte),
382                 (vma->vm_mm == current->mm ? current->comm : "???"),
383                 vma->vm_flags, vaddr);
384         dump_stack();
385 }
386
387 static inline int is_cow_mapping(unsigned int flags)
388 {
389         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
390 }
391
392 /*
393  * vm_normal_page -- This function gets the "struct page" associated with a pte.
394  *
395  * "Special" mappings do not wish to be associated with a "struct page" (either
396  * it doesn't exist, or it exists but they don't want to touch it). In this
397  * case, NULL is returned here. "Normal" mappings do have a struct page.
398  *
399  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
400  * pte bit, in which case this function is trivial. Secondly, an architecture
401  * may not have a spare pte bit, which requires a more complicated scheme,
402  * described below.
403  *
404  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
405  * special mapping (even if there are underlying and valid "struct pages").
406  * COWed pages of a VM_PFNMAP are always normal.
407  *
408  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
409  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
410  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
411  * mapping will always honor the rule
412  *
413  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
414  *
415  * And for normal mappings this is false.
416  *
417  * This restricts such mappings to be a linear translation from virtual address
418  * to pfn. To get around this restriction, we allow arbitrary mappings so long
419  * as the vma is not a COW mapping; in that case, we know that all ptes are
420  * special (because none can have been COWed).
421  *
422  *
423  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
424  *
425  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
426  * page" backing, however the difference is that _all_ pages with a struct
427  * page (that is, those where pfn_valid is true) are refcounted and considered
428  * normal pages by the VM. The disadvantage is that pages are refcounted
429  * (which can be slower and simply not an option for some PFNMAP users). The
430  * advantage is that we don't have to follow the strict linearity rule of
431  * PFNMAP mappings in order to support COWable mappings.
432  *
433  */
434 #ifdef __HAVE_ARCH_PTE_SPECIAL
435 # define HAVE_PTE_SPECIAL 1
436 #else
437 # define HAVE_PTE_SPECIAL 0
438 #endif
439 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
440                                 pte_t pte)
441 {
442         unsigned long pfn;
443
444         if (HAVE_PTE_SPECIAL) {
445                 if (likely(!pte_special(pte))) {
446                         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
447                         return pte_page(pte);
448                 }
449                 VM_BUG_ON(!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)));
450                 return NULL;
451         }
452
453         /* !HAVE_PTE_SPECIAL case follows: */
454
455         pfn = pte_pfn(pte);
456
457         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
458                 if (vma->vm_flags & VM_MIXEDMAP) {
459                         if (!pfn_valid(pfn))
460                                 return NULL;
461                         goto out;
462                 } else {
463                         unsigned long off;
464                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
465                         if (pfn == vma->vm_pgoff + off)
466                                 return NULL;
467                         if (!is_cow_mapping(vma->vm_flags))
468                                 return NULL;
469                 }
470         }
471
472         VM_BUG_ON(!pfn_valid(pfn));
473
474         /*
475          * NOTE! We still have PageReserved() pages in the page tables.
476          *
477          * eg. VDSO mappings can cause them to exist.
478          */
479 out:
480         return pfn_to_page(pfn);
481 }
482
483 /*
484  * copy one vm_area from one task to the other. Assumes the page tables
485  * already present in the new task to be cleared in the whole range
486  * covered by this vma.
487  */
488
489 static inline void
490 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
491                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
492                 unsigned long addr, int *rss)
493 {
494         unsigned long vm_flags = vma->vm_flags;
495         pte_t pte = *src_pte;
496         struct page *page;
497
498         /* pte contains position in swap or file, so copy. */
499         if (unlikely(!pte_present(pte))) {
500                 if (!pte_file(pte)) {
501                         swp_entry_t entry = pte_to_swp_entry(pte);
502
503                         swap_duplicate(entry);
504                         /* make sure dst_mm is on swapoff's mmlist. */
505                         if (unlikely(list_empty(&dst_mm->mmlist))) {
506                                 spin_lock(&mmlist_lock);
507                                 if (list_empty(&dst_mm->mmlist))
508                                         list_add(&dst_mm->mmlist,
509                                                  &src_mm->mmlist);
510                                 spin_unlock(&mmlist_lock);
511                         }
512                         if (is_write_migration_entry(entry) &&
513                                         is_cow_mapping(vm_flags)) {
514                                 /*
515                                  * COW mappings require pages in both parent
516                                  * and child to be set to read.
517                                  */
518                                 make_migration_entry_read(&entry);
519                                 pte = swp_entry_to_pte(entry);
520                                 set_pte_at(src_mm, addr, src_pte, pte);
521                         }
522                 }
523                 goto out_set_pte;
524         }
525
526         /*
527          * If it's a COW mapping, write protect it both
528          * in the parent and the child
529          */
530         if (is_cow_mapping(vm_flags)) {
531                 ptep_set_wrprotect(src_mm, addr, src_pte);
532                 pte = pte_wrprotect(pte);
533         }
534
535         /*
536          * If it's a shared mapping, mark it clean in
537          * the child
538          */
539         if (vm_flags & VM_SHARED)
540                 pte = pte_mkclean(pte);
541         pte = pte_mkold(pte);
542
543         page = vm_normal_page(vma, addr, pte);
544         if (page) {
545                 get_page(page);
546                 page_dup_rmap(page, vma, addr);
547                 rss[!!PageAnon(page)]++;
548         }
549
550 out_set_pte:
551         set_pte_at(dst_mm, addr, dst_pte, pte);
552 }
553
554 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
555                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
556                 unsigned long addr, unsigned long end)
557 {
558         pte_t *src_pte, *dst_pte;
559         spinlock_t *src_ptl, *dst_ptl;
560         int progress = 0;
561         int rss[2];
562
563 again:
564         rss[1] = rss[0] = 0;
565         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
566         if (!dst_pte)
567                 return -ENOMEM;
568         src_pte = pte_offset_map_nested(src_pmd, addr);
569         src_ptl = pte_lockptr(src_mm, src_pmd);
570         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
571         arch_enter_lazy_mmu_mode();
572
573         do {
574                 /*
575                  * We are holding two locks at this point - either of them
576                  * could generate latencies in another task on another CPU.
577                  */
578                 if (progress >= 32) {
579                         progress = 0;
580                         if (need_resched() ||
581                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
582                                 break;
583                 }
584                 if (pte_none(*src_pte)) {
585                         progress++;
586                         continue;
587                 }
588                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
589                 progress += 8;
590         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
591
592         arch_leave_lazy_mmu_mode();
593         spin_unlock(src_ptl);
594         pte_unmap_nested(src_pte - 1);
595         add_mm_rss(dst_mm, rss[0], rss[1]);
596         pte_unmap_unlock(dst_pte - 1, dst_ptl);
597         cond_resched();
598         if (addr != end)
599                 goto again;
600         return 0;
601 }
602
603 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
604                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
605                 unsigned long addr, unsigned long end)
606 {
607         pmd_t *src_pmd, *dst_pmd;
608         unsigned long next;
609
610         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
611         if (!dst_pmd)
612                 return -ENOMEM;
613         src_pmd = pmd_offset(src_pud, addr);
614         do {
615                 next = pmd_addr_end(addr, end);
616                 if (pmd_none_or_clear_bad(src_pmd))
617                         continue;
618                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
619                                                 vma, addr, next))
620                         return -ENOMEM;
621         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
622         return 0;
623 }
624
625 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
626                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
627                 unsigned long addr, unsigned long end)
628 {
629         pud_t *src_pud, *dst_pud;
630         unsigned long next;
631
632         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
633         if (!dst_pud)
634                 return -ENOMEM;
635         src_pud = pud_offset(src_pgd, addr);
636         do {
637                 next = pud_addr_end(addr, end);
638                 if (pud_none_or_clear_bad(src_pud))
639                         continue;
640                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
641                                                 vma, addr, next))
642                         return -ENOMEM;
643         } while (dst_pud++, src_pud++, addr = next, addr != end);
644         return 0;
645 }
646
647 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
648                 struct vm_area_struct *vma)
649 {
650         pgd_t *src_pgd, *dst_pgd;
651         unsigned long next;
652         unsigned long addr = vma->vm_start;
653         unsigned long end = vma->vm_end;
654
655         /*
656          * Don't copy ptes where a page fault will fill them correctly.
657          * Fork becomes much lighter when there are big shared or private
658          * readonly mappings. The tradeoff is that copy_page_range is more
659          * efficient than faulting.
660          */
661         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
662                 if (!vma->anon_vma)
663                         return 0;
664         }
665
666         if (is_vm_hugetlb_page(vma))
667                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
668
669         dst_pgd = pgd_offset(dst_mm, addr);
670         src_pgd = pgd_offset(src_mm, addr);
671         do {
672                 next = pgd_addr_end(addr, end);
673                 if (pgd_none_or_clear_bad(src_pgd))
674                         continue;
675                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
676                                                 vma, addr, next))
677                         return -ENOMEM;
678         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
679         return 0;
680 }
681
682 static unsigned long zap_pte_range(struct mmu_gather *tlb,
683                                 struct vm_area_struct *vma, pmd_t *pmd,
684                                 unsigned long addr, unsigned long end,
685                                 long *zap_work, struct zap_details *details)
686 {
687         struct mm_struct *mm = tlb->mm;
688         pte_t *pte;
689         spinlock_t *ptl;
690         int file_rss = 0;
691         int anon_rss = 0;
692
693         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
694         arch_enter_lazy_mmu_mode();
695         do {
696                 pte_t ptent = *pte;
697                 if (pte_none(ptent)) {
698                         (*zap_work)--;
699                         continue;
700                 }
701
702                 (*zap_work) -= PAGE_SIZE;
703
704                 if (pte_present(ptent)) {
705                         struct page *page;
706
707                         page = vm_normal_page(vma, addr, ptent);
708                         if (unlikely(details) && page) {
709                                 /*
710                                  * unmap_shared_mapping_pages() wants to
711                                  * invalidate cache without truncating:
712                                  * unmap shared but keep private pages.
713                                  */
714                                 if (details->check_mapping &&
715                                     details->check_mapping != page->mapping)
716                                         continue;
717                                 /*
718                                  * Each page->index must be checked when
719                                  * invalidating or truncating nonlinear.
720                                  */
721                                 if (details->nonlinear_vma &&
722                                     (page->index < details->first_index ||
723                                      page->index > details->last_index))
724                                         continue;
725                         }
726                         ptent = ptep_get_and_clear_full(mm, addr, pte,
727                                                         tlb->fullmm);
728                         tlb_remove_tlb_entry(tlb, pte, addr);
729                         if (unlikely(!page))
730                                 continue;
731                         if (unlikely(details) && details->nonlinear_vma
732                             && linear_page_index(details->nonlinear_vma,
733                                                 addr) != page->index)
734                                 set_pte_at(mm, addr, pte,
735                                            pgoff_to_pte(page->index));
736                         if (PageAnon(page))
737                                 anon_rss--;
738                         else {
739                                 if (pte_dirty(ptent))
740                                         set_page_dirty(page);
741                                 if (pte_young(ptent))
742                                         SetPageReferenced(page);
743                                 file_rss--;
744                         }
745                         page_remove_rmap(page, vma);
746                         tlb_remove_page(tlb, page);
747                         continue;
748                 }
749                 /*
750                  * If details->check_mapping, we leave swap entries;
751                  * if details->nonlinear_vma, we leave file entries.
752                  */
753                 if (unlikely(details))
754                         continue;
755                 if (!pte_file(ptent))
756                         free_swap_and_cache(pte_to_swp_entry(ptent));
757                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
758         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
759
760         add_mm_rss(mm, file_rss, anon_rss);
761         arch_leave_lazy_mmu_mode();
762         pte_unmap_unlock(pte - 1, ptl);
763
764         return addr;
765 }
766
767 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
768                                 struct vm_area_struct *vma, pud_t *pud,
769                                 unsigned long addr, unsigned long end,
770                                 long *zap_work, struct zap_details *details)
771 {
772         pmd_t *pmd;
773         unsigned long next;
774
775         pmd = pmd_offset(pud, addr);
776         do {
777                 next = pmd_addr_end(addr, end);
778                 if (pmd_none_or_clear_bad(pmd)) {
779                         (*zap_work)--;
780                         continue;
781                 }
782                 next = zap_pte_range(tlb, vma, pmd, addr, next,
783                                                 zap_work, details);
784         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
785
786         return addr;
787 }
788
789 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
790                                 struct vm_area_struct *vma, pgd_t *pgd,
791                                 unsigned long addr, unsigned long end,
792                                 long *zap_work, struct zap_details *details)
793 {
794         pud_t *pud;
795         unsigned long next;
796
797         pud = pud_offset(pgd, addr);
798         do {
799                 next = pud_addr_end(addr, end);
800                 if (pud_none_or_clear_bad(pud)) {
801                         (*zap_work)--;
802                         continue;
803                 }
804                 next = zap_pmd_range(tlb, vma, pud, addr, next,
805                                                 zap_work, details);
806         } while (pud++, addr = next, (addr != end && *zap_work > 0));
807
808         return addr;
809 }
810
811 static unsigned long unmap_page_range(struct mmu_gather *tlb,
812                                 struct vm_area_struct *vma,
813                                 unsigned long addr, unsigned long end,
814                                 long *zap_work, struct zap_details *details)
815 {
816         pgd_t *pgd;
817         unsigned long next;
818
819         if (details && !details->check_mapping && !details->nonlinear_vma)
820                 details = NULL;
821
822         BUG_ON(addr >= end);
823         tlb_start_vma(tlb, vma);
824         pgd = pgd_offset(vma->vm_mm, addr);
825         do {
826                 next = pgd_addr_end(addr, end);
827                 if (pgd_none_or_clear_bad(pgd)) {
828                         (*zap_work)--;
829                         continue;
830                 }
831                 next = zap_pud_range(tlb, vma, pgd, addr, next,
832                                                 zap_work, details);
833         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
834         tlb_end_vma(tlb, vma);
835
836         return addr;
837 }
838
839 #ifdef CONFIG_PREEMPT
840 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
841 #else
842 /* No preempt: go for improved straight-line efficiency */
843 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
844 #endif
845
846 /**
847  * unmap_vmas - unmap a range of memory covered by a list of vma's
848  * @tlbp: address of the caller's struct mmu_gather
849  * @vma: the starting vma
850  * @start_addr: virtual address at which to start unmapping
851  * @end_addr: virtual address at which to end unmapping
852  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
853  * @details: details of nonlinear truncation or shared cache invalidation
854  *
855  * Returns the end address of the unmapping (restart addr if interrupted).
856  *
857  * Unmap all pages in the vma list.
858  *
859  * We aim to not hold locks for too long (for scheduling latency reasons).
860  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
861  * return the ending mmu_gather to the caller.
862  *
863  * Only addresses between `start' and `end' will be unmapped.
864  *
865  * The VMA list must be sorted in ascending virtual address order.
866  *
867  * unmap_vmas() assumes that the caller will flush the whole unmapped address
868  * range after unmap_vmas() returns.  So the only responsibility here is to
869  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
870  * drops the lock and schedules.
871  */
872 unsigned long unmap_vmas(struct mmu_gather **tlbp,
873                 struct vm_area_struct *vma, unsigned long start_addr,
874                 unsigned long end_addr, unsigned long *nr_accounted,
875                 struct zap_details *details)
876 {
877         long zap_work = ZAP_BLOCK_SIZE;
878         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
879         int tlb_start_valid = 0;
880         unsigned long start = start_addr;
881         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
882         int fullmm = (*tlbp)->fullmm;
883
884         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
885                 unsigned long end;
886
887                 start = max(vma->vm_start, start_addr);
888                 if (start >= vma->vm_end)
889                         continue;
890                 end = min(vma->vm_end, end_addr);
891                 if (end <= vma->vm_start)
892                         continue;
893
894                 if (vma->vm_flags & VM_ACCOUNT)
895                         *nr_accounted += (end - start) >> PAGE_SHIFT;
896
897                 while (start != end) {
898                         if (!tlb_start_valid) {
899                                 tlb_start = start;
900                                 tlb_start_valid = 1;
901                         }
902
903                         if (unlikely(is_vm_hugetlb_page(vma))) {
904                                 /*
905                                  * It is undesirable to test vma->vm_file as it
906                                  * should be non-null for valid hugetlb area.
907                                  * However, vm_file will be NULL in the error
908                                  * cleanup path of do_mmap_pgoff. When
909                                  * hugetlbfs ->mmap method fails,
910                                  * do_mmap_pgoff() nullifies vma->vm_file
911                                  * before calling this function to clean up.
912                                  * Since no pte has actually been setup, it is
913                                  * safe to do nothing in this case.
914                                  */
915                                 if (vma->vm_file) {
916                                         unmap_hugepage_range(vma, start, end, NULL);
917                                         zap_work -= (end - start) /
918                                         pages_per_huge_page(hstate_vma(vma));
919                                 }
920
921                                 start = end;
922                         } else
923                                 start = unmap_page_range(*tlbp, vma,
924                                                 start, end, &zap_work, details);
925
926                         if (zap_work > 0) {
927                                 BUG_ON(start != end);
928                                 break;
929                         }
930
931                         tlb_finish_mmu(*tlbp, tlb_start, start);
932
933                         if (need_resched() ||
934                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
935                                 if (i_mmap_lock) {
936                                         *tlbp = NULL;
937                                         goto out;
938                                 }
939                                 cond_resched();
940                         }
941
942                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
943                         tlb_start_valid = 0;
944                         zap_work = ZAP_BLOCK_SIZE;
945                 }
946         }
947 out:
948         return start;   /* which is now the end (or restart) address */
949 }
950
951 /**
952  * zap_page_range - remove user pages in a given range
953  * @vma: vm_area_struct holding the applicable pages
954  * @address: starting address of pages to zap
955  * @size: number of bytes to zap
956  * @details: details of nonlinear truncation or shared cache invalidation
957  */
958 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
959                 unsigned long size, struct zap_details *details)
960 {
961         struct mm_struct *mm = vma->vm_mm;
962         struct mmu_gather *tlb;
963         unsigned long end = address + size;
964         unsigned long nr_accounted = 0;
965
966         lru_add_drain();
967         tlb = tlb_gather_mmu(mm, 0);
968         update_hiwater_rss(mm);
969         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
970         if (tlb)
971                 tlb_finish_mmu(tlb, address, end);
972         return end;
973 }
974
975 /*
976  * Do a quick page-table lookup for a single page.
977  */
978 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
979                         unsigned int flags)
980 {
981         pgd_t *pgd;
982         pud_t *pud;
983         pmd_t *pmd;
984         pte_t *ptep, pte;
985         spinlock_t *ptl;
986         struct page *page;
987         struct mm_struct *mm = vma->vm_mm;
988
989         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
990         if (!IS_ERR(page)) {
991                 BUG_ON(flags & FOLL_GET);
992                 goto out;
993         }
994
995         page = NULL;
996         pgd = pgd_offset(mm, address);
997         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
998                 goto no_page_table;
999
1000         pud = pud_offset(pgd, address);
1001         if (pud_none(*pud))
1002                 goto no_page_table;
1003         if (pud_huge(*pud)) {
1004                 BUG_ON(flags & FOLL_GET);
1005                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1006                 goto out;
1007         }
1008         if (unlikely(pud_bad(*pud)))
1009                 goto no_page_table;
1010
1011         pmd = pmd_offset(pud, address);
1012         if (pmd_none(*pmd))
1013                 goto no_page_table;
1014         if (pmd_huge(*pmd)) {
1015                 BUG_ON(flags & FOLL_GET);
1016                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1017                 goto out;
1018         }
1019         if (unlikely(pmd_bad(*pmd)))
1020                 goto no_page_table;
1021
1022         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1023
1024         pte = *ptep;
1025         if (!pte_present(pte))
1026                 goto no_page;
1027         if ((flags & FOLL_WRITE) && !pte_write(pte))
1028                 goto unlock;
1029         page = vm_normal_page(vma, address, pte);
1030         if (unlikely(!page))
1031                 goto bad_page;
1032
1033         if (flags & FOLL_GET)
1034                 get_page(page);
1035         if (flags & FOLL_TOUCH) {
1036                 if ((flags & FOLL_WRITE) &&
1037                     !pte_dirty(pte) && !PageDirty(page))
1038                         set_page_dirty(page);
1039                 mark_page_accessed(page);
1040         }
1041 unlock:
1042         pte_unmap_unlock(ptep, ptl);
1043 out:
1044         return page;
1045
1046 bad_page:
1047         pte_unmap_unlock(ptep, ptl);
1048         return ERR_PTR(-EFAULT);
1049
1050 no_page:
1051         pte_unmap_unlock(ptep, ptl);
1052         if (!pte_none(pte))
1053                 return page;
1054         /* Fall through to ZERO_PAGE handling */
1055 no_page_table:
1056         /*
1057          * When core dumping an enormous anonymous area that nobody
1058          * has touched so far, we don't want to allocate page tables.
1059          */
1060         if (flags & FOLL_ANON) {
1061                 page = ZERO_PAGE(0);
1062                 if (flags & FOLL_GET)
1063                         get_page(page);
1064                 BUG_ON(flags & FOLL_WRITE);
1065         }
1066         return page;
1067 }
1068
1069 /* Can we do the FOLL_ANON optimization? */
1070 static inline int use_zero_page(struct vm_area_struct *vma)
1071 {
1072         /*
1073          * We don't want to optimize FOLL_ANON for make_pages_present()
1074          * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1075          * we want to get the page from the page tables to make sure
1076          * that we serialize and update with any other user of that
1077          * mapping.
1078          */
1079         if (vma->vm_flags & (VM_LOCKED | VM_SHARED))
1080                 return 0;
1081         /*
1082          * And if we have a fault routine, it's not an anonymous region.
1083          */
1084         return !vma->vm_ops || !vma->vm_ops->fault;
1085 }
1086
1087 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1088                 unsigned long start, int len, int write, int force,
1089                 struct page **pages, struct vm_area_struct **vmas)
1090 {
1091         int i;
1092         unsigned int vm_flags;
1093
1094         if (len <= 0)
1095                 return 0;
1096         /* 
1097          * Require read or write permissions.
1098          * If 'force' is set, we only require the "MAY" flags.
1099          */
1100         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1101         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1102         i = 0;
1103
1104         do {
1105                 struct vm_area_struct *vma;
1106                 unsigned int foll_flags;
1107
1108                 vma = find_extend_vma(mm, start);
1109                 if (!vma && in_gate_area(tsk, start)) {
1110                         unsigned long pg = start & PAGE_MASK;
1111                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1112                         pgd_t *pgd;
1113                         pud_t *pud;
1114                         pmd_t *pmd;
1115                         pte_t *pte;
1116                         if (write) /* user gate pages are read-only */
1117                                 return i ? : -EFAULT;
1118                         if (pg > TASK_SIZE)
1119                                 pgd = pgd_offset_k(pg);
1120                         else
1121                                 pgd = pgd_offset_gate(mm, pg);
1122                         BUG_ON(pgd_none(*pgd));
1123                         pud = pud_offset(pgd, pg);
1124                         BUG_ON(pud_none(*pud));
1125                         pmd = pmd_offset(pud, pg);
1126                         if (pmd_none(*pmd))
1127                                 return i ? : -EFAULT;
1128                         pte = pte_offset_map(pmd, pg);
1129                         if (pte_none(*pte)) {
1130                                 pte_unmap(pte);
1131                                 return i ? : -EFAULT;
1132                         }
1133                         if (pages) {
1134                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1135                                 pages[i] = page;
1136                                 if (page)
1137                                         get_page(page);
1138                         }
1139                         pte_unmap(pte);
1140                         if (vmas)
1141                                 vmas[i] = gate_vma;
1142                         i++;
1143                         start += PAGE_SIZE;
1144                         len--;
1145                         continue;
1146                 }
1147
1148                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1149                                 || !(vm_flags & vma->vm_flags))
1150                         return i ? : -EFAULT;
1151
1152                 if (is_vm_hugetlb_page(vma)) {
1153                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1154                                                 &start, &len, i, write);
1155                         continue;
1156                 }
1157
1158                 foll_flags = FOLL_TOUCH;
1159                 if (pages)
1160                         foll_flags |= FOLL_GET;
1161                 if (!write && use_zero_page(vma))
1162                         foll_flags |= FOLL_ANON;
1163
1164                 do {
1165                         struct page *page;
1166
1167                         /*
1168                          * If tsk is ooming, cut off its access to large memory
1169                          * allocations. It has a pending SIGKILL, but it can't
1170                          * be processed until returning to user space.
1171                          */
1172                         if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1173                                 return i ? i : -ENOMEM;
1174
1175                         if (write)
1176                                 foll_flags |= FOLL_WRITE;
1177
1178                         cond_resched();
1179                         while (!(page = follow_page(vma, start, foll_flags))) {
1180                                 int ret;
1181                                 ret = handle_mm_fault(mm, vma, start,
1182                                                 foll_flags & FOLL_WRITE);
1183                                 if (ret & VM_FAULT_ERROR) {
1184                                         if (ret & VM_FAULT_OOM)
1185                                                 return i ? i : -ENOMEM;
1186                                         else if (ret & VM_FAULT_SIGBUS)
1187                                                 return i ? i : -EFAULT;
1188                                         BUG();
1189                                 }
1190                                 if (ret & VM_FAULT_MAJOR)
1191                                         tsk->maj_flt++;
1192                                 else
1193                                         tsk->min_flt++;
1194
1195                                 /*
1196                                  * The VM_FAULT_WRITE bit tells us that
1197                                  * do_wp_page has broken COW when necessary,
1198                                  * even if maybe_mkwrite decided not to set
1199                                  * pte_write. We can thus safely do subsequent
1200                                  * page lookups as if they were reads.
1201                                  */
1202                                 if (ret & VM_FAULT_WRITE)
1203                                         foll_flags &= ~FOLL_WRITE;
1204
1205                                 cond_resched();
1206                         }
1207                         if (IS_ERR(page))
1208                                 return i ? i : PTR_ERR(page);
1209                         if (pages) {
1210                                 pages[i] = page;
1211
1212                                 flush_anon_page(vma, page, start);
1213                                 flush_dcache_page(page);
1214                         }
1215                         if (vmas)
1216                                 vmas[i] = vma;
1217                         i++;
1218                         start += PAGE_SIZE;
1219                         len--;
1220                 } while (len && start < vma->vm_end);
1221         } while (len);
1222         return i;
1223 }
1224 EXPORT_SYMBOL(get_user_pages);
1225
1226 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1227                         spinlock_t **ptl)
1228 {
1229         pgd_t * pgd = pgd_offset(mm, addr);
1230         pud_t * pud = pud_alloc(mm, pgd, addr);
1231         if (pud) {
1232                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1233                 if (pmd)
1234                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1235         }
1236         return NULL;
1237 }
1238
1239 /*
1240  * This is the old fallback for page remapping.
1241  *
1242  * For historical reasons, it only allows reserved pages. Only
1243  * old drivers should use this, and they needed to mark their
1244  * pages reserved for the old functions anyway.
1245  */
1246 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1247                         struct page *page, pgprot_t prot)
1248 {
1249         struct mm_struct *mm = vma->vm_mm;
1250         int retval;
1251         pte_t *pte;
1252         spinlock_t *ptl;
1253
1254         retval = mem_cgroup_charge(page, mm, GFP_KERNEL);
1255         if (retval)
1256                 goto out;
1257
1258         retval = -EINVAL;
1259         if (PageAnon(page))
1260                 goto out_uncharge;
1261         retval = -ENOMEM;
1262         flush_dcache_page(page);
1263         pte = get_locked_pte(mm, addr, &ptl);
1264         if (!pte)
1265                 goto out_uncharge;
1266         retval = -EBUSY;
1267         if (!pte_none(*pte))
1268                 goto out_unlock;
1269
1270         /* Ok, finally just insert the thing.. */
1271         get_page(page);
1272         inc_mm_counter(mm, file_rss);
1273         page_add_file_rmap(page);
1274         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1275
1276         retval = 0;
1277         pte_unmap_unlock(pte, ptl);
1278         return retval;
1279 out_unlock:
1280         pte_unmap_unlock(pte, ptl);
1281 out_uncharge:
1282         mem_cgroup_uncharge_page(page);
1283 out:
1284         return retval;
1285 }
1286
1287 /**
1288  * vm_insert_page - insert single page into user vma
1289  * @vma: user vma to map to
1290  * @addr: target user address of this page
1291  * @page: source kernel page
1292  *
1293  * This allows drivers to insert individual pages they've allocated
1294  * into a user vma.
1295  *
1296  * The page has to be a nice clean _individual_ kernel allocation.
1297  * If you allocate a compound page, you need to have marked it as
1298  * such (__GFP_COMP), or manually just split the page up yourself
1299  * (see split_page()).
1300  *
1301  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1302  * took an arbitrary page protection parameter. This doesn't allow
1303  * that. Your vma protection will have to be set up correctly, which
1304  * means that if you want a shared writable mapping, you'd better
1305  * ask for a shared writable mapping!
1306  *
1307  * The page does not need to be reserved.
1308  */
1309 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1310                         struct page *page)
1311 {
1312         if (addr < vma->vm_start || addr >= vma->vm_end)
1313                 return -EFAULT;
1314         if (!page_count(page))
1315                 return -EINVAL;
1316         vma->vm_flags |= VM_INSERTPAGE;
1317         return insert_page(vma, addr, page, vma->vm_page_prot);
1318 }
1319 EXPORT_SYMBOL(vm_insert_page);
1320
1321 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1322                         unsigned long pfn, pgprot_t prot)
1323 {
1324         struct mm_struct *mm = vma->vm_mm;
1325         int retval;
1326         pte_t *pte, entry;
1327         spinlock_t *ptl;
1328
1329         retval = -ENOMEM;
1330         pte = get_locked_pte(mm, addr, &ptl);
1331         if (!pte)
1332                 goto out;
1333         retval = -EBUSY;
1334         if (!pte_none(*pte))
1335                 goto out_unlock;
1336
1337         /* Ok, finally just insert the thing.. */
1338         entry = pte_mkspecial(pfn_pte(pfn, prot));
1339         set_pte_at(mm, addr, pte, entry);
1340         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1341
1342         retval = 0;
1343 out_unlock:
1344         pte_unmap_unlock(pte, ptl);
1345 out:
1346         return retval;
1347 }
1348
1349 /**
1350  * vm_insert_pfn - insert single pfn into user vma
1351  * @vma: user vma to map to
1352  * @addr: target user address of this page
1353  * @pfn: source kernel pfn
1354  *
1355  * Similar to vm_inert_page, this allows drivers to insert individual pages
1356  * they've allocated into a user vma. Same comments apply.
1357  *
1358  * This function should only be called from a vm_ops->fault handler, and
1359  * in that case the handler should return NULL.
1360  *
1361  * vma cannot be a COW mapping.
1362  *
1363  * As this is called only for pages that do not currently exist, we
1364  * do not need to flush old virtual caches or the TLB.
1365  */
1366 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1367                         unsigned long pfn)
1368 {
1369         /*
1370          * Technically, architectures with pte_special can avoid all these
1371          * restrictions (same for remap_pfn_range).  However we would like
1372          * consistency in testing and feature parity among all, so we should
1373          * try to keep these invariants in place for everybody.
1374          */
1375         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1376         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1377                                                 (VM_PFNMAP|VM_MIXEDMAP));
1378         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1379         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1380
1381         if (addr < vma->vm_start || addr >= vma->vm_end)
1382                 return -EFAULT;
1383         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1384 }
1385 EXPORT_SYMBOL(vm_insert_pfn);
1386
1387 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1388                         unsigned long pfn)
1389 {
1390         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1391
1392         if (addr < vma->vm_start || addr >= vma->vm_end)
1393                 return -EFAULT;
1394
1395         /*
1396          * If we don't have pte special, then we have to use the pfn_valid()
1397          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1398          * refcount the page if pfn_valid is true (hence insert_page rather
1399          * than insert_pfn).
1400          */
1401         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1402                 struct page *page;
1403
1404                 page = pfn_to_page(pfn);
1405                 return insert_page(vma, addr, page, vma->vm_page_prot);
1406         }
1407         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1408 }
1409 EXPORT_SYMBOL(vm_insert_mixed);
1410
1411 /*
1412  * maps a range of physical memory into the requested pages. the old
1413  * mappings are removed. any references to nonexistent pages results
1414  * in null mappings (currently treated as "copy-on-access")
1415  */
1416 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1417                         unsigned long addr, unsigned long end,
1418                         unsigned long pfn, pgprot_t prot)
1419 {
1420         pte_t *pte;
1421         spinlock_t *ptl;
1422
1423         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1424         if (!pte)
1425                 return -ENOMEM;
1426         arch_enter_lazy_mmu_mode();
1427         do {
1428                 BUG_ON(!pte_none(*pte));
1429                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1430                 pfn++;
1431         } while (pte++, addr += PAGE_SIZE, addr != end);
1432         arch_leave_lazy_mmu_mode();
1433         pte_unmap_unlock(pte - 1, ptl);
1434         return 0;
1435 }
1436
1437 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1438                         unsigned long addr, unsigned long end,
1439                         unsigned long pfn, pgprot_t prot)
1440 {
1441         pmd_t *pmd;
1442         unsigned long next;
1443
1444         pfn -= addr >> PAGE_SHIFT;
1445         pmd = pmd_alloc(mm, pud, addr);
1446         if (!pmd)
1447                 return -ENOMEM;
1448         do {
1449                 next = pmd_addr_end(addr, end);
1450                 if (remap_pte_range(mm, pmd, addr, next,
1451                                 pfn + (addr >> PAGE_SHIFT), prot))
1452                         return -ENOMEM;
1453         } while (pmd++, addr = next, addr != end);
1454         return 0;
1455 }
1456
1457 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1458                         unsigned long addr, unsigned long end,
1459                         unsigned long pfn, pgprot_t prot)
1460 {
1461         pud_t *pud;
1462         unsigned long next;
1463
1464         pfn -= addr >> PAGE_SHIFT;
1465         pud = pud_alloc(mm, pgd, addr);
1466         if (!pud)
1467                 return -ENOMEM;
1468         do {
1469                 next = pud_addr_end(addr, end);
1470                 if (remap_pmd_range(mm, pud, addr, next,
1471                                 pfn + (addr >> PAGE_SHIFT), prot))
1472                         return -ENOMEM;
1473         } while (pud++, addr = next, addr != end);
1474         return 0;
1475 }
1476
1477 /**
1478  * remap_pfn_range - remap kernel memory to userspace
1479  * @vma: user vma to map to
1480  * @addr: target user address to start at
1481  * @pfn: physical address of kernel memory
1482  * @size: size of map area
1483  * @prot: page protection flags for this mapping
1484  *
1485  *  Note: this is only safe if the mm semaphore is held when called.
1486  */
1487 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1488                     unsigned long pfn, unsigned long size, pgprot_t prot)
1489 {
1490         pgd_t *pgd;
1491         unsigned long next;
1492         unsigned long end = addr + PAGE_ALIGN(size);
1493         struct mm_struct *mm = vma->vm_mm;
1494         int err;
1495
1496         /*
1497          * Physically remapped pages are special. Tell the
1498          * rest of the world about it:
1499          *   VM_IO tells people not to look at these pages
1500          *      (accesses can have side effects).
1501          *   VM_RESERVED is specified all over the place, because
1502          *      in 2.4 it kept swapout's vma scan off this vma; but
1503          *      in 2.6 the LRU scan won't even find its pages, so this
1504          *      flag means no more than count its pages in reserved_vm,
1505          *      and omit it from core dump, even when VM_IO turned off.
1506          *   VM_PFNMAP tells the core MM that the base pages are just
1507          *      raw PFN mappings, and do not have a "struct page" associated
1508          *      with them.
1509          *
1510          * There's a horrible special case to handle copy-on-write
1511          * behaviour that some programs depend on. We mark the "original"
1512          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1513          */
1514         if (is_cow_mapping(vma->vm_flags)) {
1515                 if (addr != vma->vm_start || end != vma->vm_end)
1516                         return -EINVAL;
1517                 vma->vm_pgoff = pfn;
1518         }
1519
1520         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1521
1522         BUG_ON(addr >= end);
1523         pfn -= addr >> PAGE_SHIFT;
1524         pgd = pgd_offset(mm, addr);
1525         flush_cache_range(vma, addr, end);
1526         do {
1527                 next = pgd_addr_end(addr, end);
1528                 err = remap_pud_range(mm, pgd, addr, next,
1529                                 pfn + (addr >> PAGE_SHIFT), prot);
1530                 if (err)
1531                         break;
1532         } while (pgd++, addr = next, addr != end);
1533         return err;
1534 }
1535 EXPORT_SYMBOL(remap_pfn_range);
1536
1537 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1538                                      unsigned long addr, unsigned long end,
1539                                      pte_fn_t fn, void *data)
1540 {
1541         pte_t *pte;
1542         int err;
1543         pgtable_t token;
1544         spinlock_t *uninitialized_var(ptl);
1545
1546         pte = (mm == &init_mm) ?
1547                 pte_alloc_kernel(pmd, addr) :
1548                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1549         if (!pte)
1550                 return -ENOMEM;
1551
1552         BUG_ON(pmd_huge(*pmd));
1553
1554         token = pmd_pgtable(*pmd);
1555
1556         do {
1557                 err = fn(pte, token, addr, data);
1558                 if (err)
1559                         break;
1560         } while (pte++, addr += PAGE_SIZE, addr != end);
1561
1562         if (mm != &init_mm)
1563                 pte_unmap_unlock(pte-1, ptl);
1564         return err;
1565 }
1566
1567 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1568                                      unsigned long addr, unsigned long end,
1569                                      pte_fn_t fn, void *data)
1570 {
1571         pmd_t *pmd;
1572         unsigned long next;
1573         int err;
1574
1575         BUG_ON(pud_huge(*pud));
1576
1577         pmd = pmd_alloc(mm, pud, addr);
1578         if (!pmd)
1579                 return -ENOMEM;
1580         do {
1581                 next = pmd_addr_end(addr, end);
1582                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1583                 if (err)
1584                         break;
1585         } while (pmd++, addr = next, addr != end);
1586         return err;
1587 }
1588
1589 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1590                                      unsigned long addr, unsigned long end,
1591                                      pte_fn_t fn, void *data)
1592 {
1593         pud_t *pud;
1594         unsigned long next;
1595         int err;
1596
1597         pud = pud_alloc(mm, pgd, addr);
1598         if (!pud)
1599                 return -ENOMEM;
1600         do {
1601                 next = pud_addr_end(addr, end);
1602                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1603                 if (err)
1604                         break;
1605         } while (pud++, addr = next, addr != end);
1606         return err;
1607 }
1608
1609 /*
1610  * Scan a region of virtual memory, filling in page tables as necessary
1611  * and calling a provided function on each leaf page table.
1612  */
1613 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1614                         unsigned long size, pte_fn_t fn, void *data)
1615 {
1616         pgd_t *pgd;
1617         unsigned long next;
1618         unsigned long end = addr + size;
1619         int err;
1620
1621         BUG_ON(addr >= end);
1622         pgd = pgd_offset(mm, addr);
1623         do {
1624                 next = pgd_addr_end(addr, end);
1625                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1626                 if (err)
1627                         break;
1628         } while (pgd++, addr = next, addr != end);
1629         return err;
1630 }
1631 EXPORT_SYMBOL_GPL(apply_to_page_range);
1632
1633 /*
1634  * handle_pte_fault chooses page fault handler according to an entry
1635  * which was read non-atomically.  Before making any commitment, on
1636  * those architectures or configurations (e.g. i386 with PAE) which
1637  * might give a mix of unmatched parts, do_swap_page and do_file_page
1638  * must check under lock before unmapping the pte and proceeding
1639  * (but do_wp_page is only called after already making such a check;
1640  * and do_anonymous_page and do_no_page can safely check later on).
1641  */
1642 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1643                                 pte_t *page_table, pte_t orig_pte)
1644 {
1645         int same = 1;
1646 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1647         if (sizeof(pte_t) > sizeof(unsigned long)) {
1648                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1649                 spin_lock(ptl);
1650                 same = pte_same(*page_table, orig_pte);
1651                 spin_unlock(ptl);
1652         }
1653 #endif
1654         pte_unmap(page_table);
1655         return same;
1656 }
1657
1658 /*
1659  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1660  * servicing faults for write access.  In the normal case, do always want
1661  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1662  * that do not have writing enabled, when used by access_process_vm.
1663  */
1664 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1665 {
1666         if (likely(vma->vm_flags & VM_WRITE))
1667                 pte = pte_mkwrite(pte);
1668         return pte;
1669 }
1670
1671 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1672 {
1673         /*
1674          * If the source page was a PFN mapping, we don't have
1675          * a "struct page" for it. We do a best-effort copy by
1676          * just copying from the original user address. If that
1677          * fails, we just zero-fill it. Live with it.
1678          */
1679         if (unlikely(!src)) {
1680                 void *kaddr = kmap_atomic(dst, KM_USER0);
1681                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1682
1683                 /*
1684                  * This really shouldn't fail, because the page is there
1685                  * in the page tables. But it might just be unreadable,
1686                  * in which case we just give up and fill the result with
1687                  * zeroes.
1688                  */
1689                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1690                         memset(kaddr, 0, PAGE_SIZE);
1691                 kunmap_atomic(kaddr, KM_USER0);
1692                 flush_dcache_page(dst);
1693         } else
1694                 copy_user_highpage(dst, src, va, vma);
1695 }
1696
1697 /*
1698  * This routine handles present pages, when users try to write
1699  * to a shared page. It is done by copying the page to a new address
1700  * and decrementing the shared-page counter for the old page.
1701  *
1702  * Note that this routine assumes that the protection checks have been
1703  * done by the caller (the low-level page fault routine in most cases).
1704  * Thus we can safely just mark it writable once we've done any necessary
1705  * COW.
1706  *
1707  * We also mark the page dirty at this point even though the page will
1708  * change only once the write actually happens. This avoids a few races,
1709  * and potentially makes it more efficient.
1710  *
1711  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1712  * but allow concurrent faults), with pte both mapped and locked.
1713  * We return with mmap_sem still held, but pte unmapped and unlocked.
1714  */
1715 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1716                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1717                 spinlock_t *ptl, pte_t orig_pte)
1718 {
1719         struct page *old_page, *new_page;
1720         pte_t entry;
1721         int reuse = 0, ret = 0;
1722         int page_mkwrite = 0;
1723         struct page *dirty_page = NULL;
1724
1725         old_page = vm_normal_page(vma, address, orig_pte);
1726         if (!old_page) {
1727                 /*
1728                  * VM_MIXEDMAP !pfn_valid() case
1729                  *
1730                  * We should not cow pages in a shared writeable mapping.
1731                  * Just mark the pages writable as we can't do any dirty
1732                  * accounting on raw pfn maps.
1733                  */
1734                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1735                                      (VM_WRITE|VM_SHARED))
1736                         goto reuse;
1737                 goto gotten;
1738         }
1739
1740         /*
1741          * Take out anonymous pages first, anonymous shared vmas are
1742          * not dirty accountable.
1743          */
1744         if (PageAnon(old_page)) {
1745                 if (!TestSetPageLocked(old_page)) {
1746                         reuse = can_share_swap_page(old_page);
1747                         unlock_page(old_page);
1748                 }
1749         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1750                                         (VM_WRITE|VM_SHARED))) {
1751                 /*
1752                  * Only catch write-faults on shared writable pages,
1753                  * read-only shared pages can get COWed by
1754                  * get_user_pages(.write=1, .force=1).
1755                  */
1756                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1757                         /*
1758                          * Notify the address space that the page is about to
1759                          * become writable so that it can prohibit this or wait
1760                          * for the page to get into an appropriate state.
1761                          *
1762                          * We do this without the lock held, so that it can
1763                          * sleep if it needs to.
1764                          */
1765                         page_cache_get(old_page);
1766                         pte_unmap_unlock(page_table, ptl);
1767
1768                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1769                                 goto unwritable_page;
1770
1771                         /*
1772                          * Since we dropped the lock we need to revalidate
1773                          * the PTE as someone else may have changed it.  If
1774                          * they did, we just return, as we can count on the
1775                          * MMU to tell us if they didn't also make it writable.
1776                          */
1777                         page_table = pte_offset_map_lock(mm, pmd, address,
1778                                                          &ptl);
1779                         page_cache_release(old_page);
1780                         if (!pte_same(*page_table, orig_pte))
1781                                 goto unlock;
1782
1783                         page_mkwrite = 1;
1784                 }
1785                 dirty_page = old_page;
1786                 get_page(dirty_page);
1787                 reuse = 1;
1788         }
1789
1790         if (reuse) {
1791 reuse:
1792                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1793                 entry = pte_mkyoung(orig_pte);
1794                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1795                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1796                         update_mmu_cache(vma, address, entry);
1797                 ret |= VM_FAULT_WRITE;
1798                 goto unlock;
1799         }
1800
1801         /*
1802          * Ok, we need to copy. Oh, well..
1803          */
1804         page_cache_get(old_page);
1805 gotten:
1806         pte_unmap_unlock(page_table, ptl);
1807
1808         if (unlikely(anon_vma_prepare(vma)))
1809                 goto oom;
1810         VM_BUG_ON(old_page == ZERO_PAGE(0));
1811         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1812         if (!new_page)
1813                 goto oom;
1814         cow_user_page(new_page, old_page, address, vma);
1815         __SetPageUptodate(new_page);
1816
1817         if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
1818                 goto oom_free_new;
1819
1820         /*
1821          * Re-check the pte - we dropped the lock
1822          */
1823         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1824         if (likely(pte_same(*page_table, orig_pte))) {
1825                 if (old_page) {
1826                         if (!PageAnon(old_page)) {
1827                                 dec_mm_counter(mm, file_rss);
1828                                 inc_mm_counter(mm, anon_rss);
1829                         }
1830                 } else
1831                         inc_mm_counter(mm, anon_rss);
1832                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1833                 entry = mk_pte(new_page, vma->vm_page_prot);
1834                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1835                 /*
1836                  * Clear the pte entry and flush it first, before updating the
1837                  * pte with the new entry. This will avoid a race condition
1838                  * seen in the presence of one thread doing SMC and another
1839                  * thread doing COW.
1840                  */
1841                 ptep_clear_flush(vma, address, page_table);
1842                 set_pte_at(mm, address, page_table, entry);
1843                 update_mmu_cache(vma, address, entry);
1844                 lru_cache_add_active(new_page);
1845                 page_add_new_anon_rmap(new_page, vma, address);
1846
1847                 if (old_page) {
1848                         /*
1849                          * Only after switching the pte to the new page may
1850                          * we remove the mapcount here. Otherwise another
1851                          * process may come and find the rmap count decremented
1852                          * before the pte is switched to the new page, and
1853                          * "reuse" the old page writing into it while our pte
1854                          * here still points into it and can be read by other
1855                          * threads.
1856                          *
1857                          * The critical issue is to order this
1858                          * page_remove_rmap with the ptp_clear_flush above.
1859                          * Those stores are ordered by (if nothing else,)
1860                          * the barrier present in the atomic_add_negative
1861                          * in page_remove_rmap.
1862                          *
1863                          * Then the TLB flush in ptep_clear_flush ensures that
1864                          * no process can access the old page before the
1865                          * decremented mapcount is visible. And the old page
1866                          * cannot be reused until after the decremented
1867                          * mapcount is visible. So transitively, TLBs to
1868                          * old page will be flushed before it can be reused.
1869                          */
1870                         page_remove_rmap(old_page, vma);
1871                 }
1872
1873                 /* Free the old page.. */
1874                 new_page = old_page;
1875                 ret |= VM_FAULT_WRITE;
1876         } else
1877                 mem_cgroup_uncharge_page(new_page);
1878
1879         if (new_page)
1880                 page_cache_release(new_page);
1881         if (old_page)
1882                 page_cache_release(old_page);
1883 unlock:
1884         pte_unmap_unlock(page_table, ptl);
1885         if (dirty_page) {
1886                 if (vma->vm_file)
1887                         file_update_time(vma->vm_file);
1888
1889                 /*
1890                  * Yes, Virginia, this is actually required to prevent a race
1891                  * with clear_page_dirty_for_io() from clearing the page dirty
1892                  * bit after it clear all dirty ptes, but before a racing
1893                  * do_wp_page installs a dirty pte.
1894                  *
1895                  * do_no_page is protected similarly.
1896                  */
1897                 wait_on_page_locked(dirty_page);
1898                 set_page_dirty_balance(dirty_page, page_mkwrite);
1899                 put_page(dirty_page);
1900         }
1901         return ret;
1902 oom_free_new:
1903         page_cache_release(new_page);
1904 oom:
1905         if (old_page)
1906                 page_cache_release(old_page);
1907         return VM_FAULT_OOM;
1908
1909 unwritable_page:
1910         page_cache_release(old_page);
1911         return VM_FAULT_SIGBUS;
1912 }
1913
1914 /*
1915  * Helper functions for unmap_mapping_range().
1916  *
1917  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1918  *
1919  * We have to restart searching the prio_tree whenever we drop the lock,
1920  * since the iterator is only valid while the lock is held, and anyway
1921  * a later vma might be split and reinserted earlier while lock dropped.
1922  *
1923  * The list of nonlinear vmas could be handled more efficiently, using
1924  * a placeholder, but handle it in the same way until a need is shown.
1925  * It is important to search the prio_tree before nonlinear list: a vma
1926  * may become nonlinear and be shifted from prio_tree to nonlinear list
1927  * while the lock is dropped; but never shifted from list to prio_tree.
1928  *
1929  * In order to make forward progress despite restarting the search,
1930  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1931  * quickly skip it next time around.  Since the prio_tree search only
1932  * shows us those vmas affected by unmapping the range in question, we
1933  * can't efficiently keep all vmas in step with mapping->truncate_count:
1934  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1935  * mapping->truncate_count and vma->vm_truncate_count are protected by
1936  * i_mmap_lock.
1937  *
1938  * In order to make forward progress despite repeatedly restarting some
1939  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1940  * and restart from that address when we reach that vma again.  It might
1941  * have been split or merged, shrunk or extended, but never shifted: so
1942  * restart_addr remains valid so long as it remains in the vma's range.
1943  * unmap_mapping_range forces truncate_count to leap over page-aligned
1944  * values so we can save vma's restart_addr in its truncate_count field.
1945  */
1946 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1947
1948 static void reset_vma_truncate_counts(struct address_space *mapping)
1949 {
1950         struct vm_area_struct *vma;
1951         struct prio_tree_iter iter;
1952
1953         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1954                 vma->vm_truncate_count = 0;
1955         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1956                 vma->vm_truncate_count = 0;
1957 }
1958
1959 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1960                 unsigned long start_addr, unsigned long end_addr,
1961                 struct zap_details *details)
1962 {
1963         unsigned long restart_addr;
1964         int need_break;
1965
1966         /*
1967          * files that support invalidating or truncating portions of the
1968          * file from under mmaped areas must have their ->fault function
1969          * return a locked page (and set VM_FAULT_LOCKED in the return).
1970          * This provides synchronisation against concurrent unmapping here.
1971          */
1972
1973 again:
1974         restart_addr = vma->vm_truncate_count;
1975         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1976                 start_addr = restart_addr;
1977                 if (start_addr >= end_addr) {
1978                         /* Top of vma has been split off since last time */
1979                         vma->vm_truncate_count = details->truncate_count;
1980                         return 0;
1981                 }
1982         }
1983
1984         restart_addr = zap_page_range(vma, start_addr,
1985                                         end_addr - start_addr, details);
1986         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
1987
1988         if (restart_addr >= end_addr) {
1989                 /* We have now completed this vma: mark it so */
1990                 vma->vm_truncate_count = details->truncate_count;
1991                 if (!need_break)
1992                         return 0;
1993         } else {
1994                 /* Note restart_addr in vma's truncate_count field */
1995                 vma->vm_truncate_count = restart_addr;
1996                 if (!need_break)
1997                         goto again;
1998         }
1999
2000         spin_unlock(details->i_mmap_lock);
2001         cond_resched();
2002         spin_lock(details->i_mmap_lock);
2003         return -EINTR;
2004 }
2005
2006 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2007                                             struct zap_details *details)
2008 {
2009         struct vm_area_struct *vma;
2010         struct prio_tree_iter iter;
2011         pgoff_t vba, vea, zba, zea;
2012
2013 restart:
2014         vma_prio_tree_foreach(vma, &iter, root,
2015                         details->first_index, details->last_index) {
2016                 /* Skip quickly over those we have already dealt with */
2017                 if (vma->vm_truncate_count == details->truncate_count)
2018                         continue;
2019
2020                 vba = vma->vm_pgoff;
2021                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2022                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2023                 zba = details->first_index;
2024                 if (zba < vba)
2025                         zba = vba;
2026                 zea = details->last_index;
2027                 if (zea > vea)
2028                         zea = vea;
2029
2030                 if (unmap_mapping_range_vma(vma,
2031                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2032                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2033                                 details) < 0)
2034                         goto restart;
2035         }
2036 }
2037
2038 static inline void unmap_mapping_range_list(struct list_head *head,
2039                                             struct zap_details *details)
2040 {
2041         struct vm_area_struct *vma;
2042
2043         /*
2044          * In nonlinear VMAs there is no correspondence between virtual address
2045          * offset and file offset.  So we must perform an exhaustive search
2046          * across *all* the pages in each nonlinear VMA, not just the pages
2047          * whose virtual address lies outside the file truncation point.
2048          */
2049 restart:
2050         list_for_each_entry(vma, head, shared.vm_set.list) {
2051                 /* Skip quickly over those we have already dealt with */
2052                 if (vma->vm_truncate_count == details->truncate_count)
2053                         continue;
2054                 details->nonlinear_vma = vma;
2055                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2056                                         vma->vm_end, details) < 0)
2057                         goto restart;
2058         }
2059 }
2060
2061 /**
2062  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2063  * @mapping: the address space containing mmaps to be unmapped.
2064  * @holebegin: byte in first page to unmap, relative to the start of
2065  * the underlying file.  This will be rounded down to a PAGE_SIZE
2066  * boundary.  Note that this is different from vmtruncate(), which
2067  * must keep the partial page.  In contrast, we must get rid of
2068  * partial pages.
2069  * @holelen: size of prospective hole in bytes.  This will be rounded
2070  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2071  * end of the file.
2072  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2073  * but 0 when invalidating pagecache, don't throw away private data.
2074  */
2075 void unmap_mapping_range(struct address_space *mapping,
2076                 loff_t const holebegin, loff_t const holelen, int even_cows)
2077 {
2078         struct zap_details details;
2079         pgoff_t hba = holebegin >> PAGE_SHIFT;
2080         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2081
2082         /* Check for overflow. */
2083         if (sizeof(holelen) > sizeof(hlen)) {
2084                 long long holeend =
2085                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2086                 if (holeend & ~(long long)ULONG_MAX)
2087                         hlen = ULONG_MAX - hba + 1;
2088         }
2089
2090         details.check_mapping = even_cows? NULL: mapping;
2091         details.nonlinear_vma = NULL;
2092         details.first_index = hba;
2093         details.last_index = hba + hlen - 1;
2094         if (details.last_index < details.first_index)
2095                 details.last_index = ULONG_MAX;
2096         details.i_mmap_lock = &mapping->i_mmap_lock;
2097
2098         spin_lock(&mapping->i_mmap_lock);
2099
2100         /* Protect against endless unmapping loops */
2101         mapping->truncate_count++;
2102         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2103                 if (mapping->truncate_count == 0)
2104                         reset_vma_truncate_counts(mapping);
2105                 mapping->truncate_count++;
2106         }
2107         details.truncate_count = mapping->truncate_count;
2108
2109         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2110                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2111         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2112                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2113         spin_unlock(&mapping->i_mmap_lock);
2114 }
2115 EXPORT_SYMBOL(unmap_mapping_range);
2116
2117 /**
2118  * vmtruncate - unmap mappings "freed" by truncate() syscall
2119  * @inode: inode of the file used
2120  * @offset: file offset to start truncating
2121  *
2122  * NOTE! We have to be ready to update the memory sharing
2123  * between the file and the memory map for a potential last
2124  * incomplete page.  Ugly, but necessary.
2125  */
2126 int vmtruncate(struct inode * inode, loff_t offset)
2127 {
2128         if (inode->i_size < offset) {
2129                 unsigned long limit;
2130
2131                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2132                 if (limit != RLIM_INFINITY && offset > limit)
2133                         goto out_sig;
2134                 if (offset > inode->i_sb->s_maxbytes)
2135                         goto out_big;
2136                 i_size_write(inode, offset);
2137         } else {
2138                 struct address_space *mapping = inode->i_mapping;
2139
2140                 /*
2141                  * truncation of in-use swapfiles is disallowed - it would
2142                  * cause subsequent swapout to scribble on the now-freed
2143                  * blocks.
2144                  */
2145                 if (IS_SWAPFILE(inode))
2146                         return -ETXTBSY;
2147                 i_size_write(inode, offset);
2148
2149                 /*
2150                  * unmap_mapping_range is called twice, first simply for
2151                  * efficiency so that truncate_inode_pages does fewer
2152                  * single-page unmaps.  However after this first call, and
2153                  * before truncate_inode_pages finishes, it is possible for
2154                  * private pages to be COWed, which remain after
2155                  * truncate_inode_pages finishes, hence the second
2156                  * unmap_mapping_range call must be made for correctness.
2157                  */
2158                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2159                 truncate_inode_pages(mapping, offset);
2160                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2161         }
2162
2163         if (inode->i_op && inode->i_op->truncate)
2164                 inode->i_op->truncate(inode);
2165         return 0;
2166
2167 out_sig:
2168         send_sig(SIGXFSZ, current, 0);
2169 out_big:
2170         return -EFBIG;
2171 }
2172 EXPORT_SYMBOL(vmtruncate);
2173
2174 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2175 {
2176         struct address_space *mapping = inode->i_mapping;
2177
2178         /*
2179          * If the underlying filesystem is not going to provide
2180          * a way to truncate a range of blocks (punch a hole) -
2181          * we should return failure right now.
2182          */
2183         if (!inode->i_op || !inode->i_op->truncate_range)
2184                 return -ENOSYS;
2185
2186         mutex_lock(&inode->i_mutex);
2187         down_write(&inode->i_alloc_sem);
2188         unmap_mapping_range(mapping, offset, (end - offset), 1);
2189         truncate_inode_pages_range(mapping, offset, end);
2190         unmap_mapping_range(mapping, offset, (end - offset), 1);
2191         inode->i_op->truncate_range(inode, offset, end);
2192         up_write(&inode->i_alloc_sem);
2193         mutex_unlock(&inode->i_mutex);
2194
2195         return 0;
2196 }
2197
2198 /*
2199  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2200  * but allow concurrent faults), and pte mapped but not yet locked.
2201  * We return with mmap_sem still held, but pte unmapped and unlocked.
2202  */
2203 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2204                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2205                 int write_access, pte_t orig_pte)
2206 {
2207         spinlock_t *ptl;
2208         struct page *page;
2209         swp_entry_t entry;
2210         pte_t pte;
2211         int ret = 0;
2212
2213         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2214                 goto out;
2215
2216         entry = pte_to_swp_entry(orig_pte);
2217         if (is_migration_entry(entry)) {
2218                 migration_entry_wait(mm, pmd, address);
2219                 goto out;
2220         }
2221         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2222         page = lookup_swap_cache(entry);
2223         if (!page) {
2224                 grab_swap_token(); /* Contend for token _before_ read-in */
2225                 page = swapin_readahead(entry,
2226                                         GFP_HIGHUSER_MOVABLE, vma, address);
2227                 if (!page) {
2228                         /*
2229                          * Back out if somebody else faulted in this pte
2230                          * while we released the pte lock.
2231                          */
2232                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2233                         if (likely(pte_same(*page_table, orig_pte)))
2234                                 ret = VM_FAULT_OOM;
2235                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2236                         goto unlock;
2237                 }
2238
2239                 /* Had to read the page from swap area: Major fault */
2240                 ret = VM_FAULT_MAJOR;
2241                 count_vm_event(PGMAJFAULT);
2242         }
2243
2244         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2245                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2246                 ret = VM_FAULT_OOM;
2247                 goto out;
2248         }
2249
2250         mark_page_accessed(page);
2251         lock_page(page);
2252         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2253
2254         /*
2255          * Back out if somebody else already faulted in this pte.
2256          */
2257         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2258         if (unlikely(!pte_same(*page_table, orig_pte)))
2259                 goto out_nomap;
2260
2261         if (unlikely(!PageUptodate(page))) {
2262                 ret = VM_FAULT_SIGBUS;
2263                 goto out_nomap;
2264         }
2265
2266         /* The page isn't present yet, go ahead with the fault. */
2267
2268         inc_mm_counter(mm, anon_rss);
2269         pte = mk_pte(page, vma->vm_page_prot);
2270         if (write_access && can_share_swap_page(page)) {
2271                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2272                 write_access = 0;
2273         }
2274
2275         flush_icache_page(vma, page);
2276         set_pte_at(mm, address, page_table, pte);
2277         page_add_anon_rmap(page, vma, address);
2278
2279         swap_free(entry);
2280         if (vm_swap_full())
2281                 remove_exclusive_swap_page(page);
2282         unlock_page(page);
2283
2284         if (write_access) {
2285                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2286                 if (ret & VM_FAULT_ERROR)
2287                         ret &= VM_FAULT_ERROR;
2288                 goto out;
2289         }
2290
2291         /* No need to invalidate - it was non-present before */
2292         update_mmu_cache(vma, address, pte);
2293 unlock:
2294         pte_unmap_unlock(page_table, ptl);
2295 out:
2296         return ret;
2297 out_nomap:
2298         mem_cgroup_uncharge_page(page);
2299         pte_unmap_unlock(page_table, ptl);
2300         unlock_page(page);
2301         page_cache_release(page);
2302         return ret;
2303 }
2304
2305 /*
2306  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2307  * but allow concurrent faults), and pte mapped but not yet locked.
2308  * We return with mmap_sem still held, but pte unmapped and unlocked.
2309  */
2310 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2311                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2312                 int write_access)
2313 {
2314         struct page *page;
2315         spinlock_t *ptl;
2316         pte_t entry;
2317
2318         /* Allocate our own private page. */
2319         pte_unmap(page_table);
2320
2321         if (unlikely(anon_vma_prepare(vma)))
2322                 goto oom;
2323         page = alloc_zeroed_user_highpage_movable(vma, address);
2324         if (!page)
2325                 goto oom;
2326         __SetPageUptodate(page);
2327
2328         if (mem_cgroup_charge(page, mm, GFP_KERNEL))
2329                 goto oom_free_page;
2330
2331         entry = mk_pte(page, vma->vm_page_prot);
2332         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2333
2334         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2335         if (!pte_none(*page_table))
2336                 goto release;
2337         inc_mm_counter(mm, anon_rss);
2338         lru_cache_add_active(page);
2339         page_add_new_anon_rmap(page, vma, address);
2340         set_pte_at(mm, address, page_table, entry);
2341
2342         /* No need to invalidate - it was non-present before */
2343         update_mmu_cache(vma, address, entry);
2344 unlock:
2345         pte_unmap_unlock(page_table, ptl);
2346         return 0;
2347 release:
2348         mem_cgroup_uncharge_page(page);
2349         page_cache_release(page);
2350         goto unlock;
2351 oom_free_page:
2352         page_cache_release(page);
2353 oom:
2354         return VM_FAULT_OOM;
2355 }
2356
2357 /*
2358  * __do_fault() tries to create a new page mapping. It aggressively
2359  * tries to share with existing pages, but makes a separate copy if
2360  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2361  * the next page fault.
2362  *
2363  * As this is called only for pages that do not currently exist, we
2364  * do not need to flush old virtual caches or the TLB.
2365  *
2366  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2367  * but allow concurrent faults), and pte neither mapped nor locked.
2368  * We return with mmap_sem still held, but pte unmapped and unlocked.
2369  */
2370 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2371                 unsigned long address, pmd_t *pmd,
2372                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2373 {
2374         pte_t *page_table;
2375         spinlock_t *ptl;
2376         struct page *page;
2377         pte_t entry;
2378         int anon = 0;
2379         struct page *dirty_page = NULL;
2380         struct vm_fault vmf;
2381         int ret;
2382         int page_mkwrite = 0;
2383
2384         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2385         vmf.pgoff = pgoff;
2386         vmf.flags = flags;
2387         vmf.page = NULL;
2388
2389         ret = vma->vm_ops->fault(vma, &vmf);
2390         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2391                 return ret;
2392
2393         /*
2394          * For consistency in subsequent calls, make the faulted page always
2395          * locked.
2396          */
2397         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2398                 lock_page(vmf.page);
2399         else
2400                 VM_BUG_ON(!PageLocked(vmf.page));
2401
2402         /*
2403          * Should we do an early C-O-W break?
2404          */
2405         page = vmf.page;
2406         if (flags & FAULT_FLAG_WRITE) {
2407                 if (!(vma->vm_flags & VM_SHARED)) {
2408                         anon = 1;
2409                         if (unlikely(anon_vma_prepare(vma))) {
2410                                 ret = VM_FAULT_OOM;
2411                                 goto out;
2412                         }
2413                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2414                                                 vma, address);
2415                         if (!page) {
2416                                 ret = VM_FAULT_OOM;
2417                                 goto out;
2418                         }
2419                         copy_user_highpage(page, vmf.page, address, vma);
2420                         __SetPageUptodate(page);
2421                 } else {
2422                         /*
2423                          * If the page will be shareable, see if the backing
2424                          * address space wants to know that the page is about
2425                          * to become writable
2426                          */
2427                         if (vma->vm_ops->page_mkwrite) {
2428                                 unlock_page(page);
2429                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2430                                         ret = VM_FAULT_SIGBUS;
2431                                         anon = 1; /* no anon but release vmf.page */
2432                                         goto out_unlocked;
2433                                 }
2434                                 lock_page(page);
2435                                 /*
2436                                  * XXX: this is not quite right (racy vs
2437                                  * invalidate) to unlock and relock the page
2438                                  * like this, however a better fix requires
2439                                  * reworking page_mkwrite locking API, which
2440                                  * is better done later.
2441                                  */
2442                                 if (!page->mapping) {
2443                                         ret = 0;
2444                                         anon = 1; /* no anon but release vmf.page */
2445                                         goto out;
2446                                 }
2447                                 page_mkwrite = 1;
2448                         }
2449                 }
2450
2451         }
2452
2453         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2454                 ret = VM_FAULT_OOM;
2455                 goto out;
2456         }
2457
2458         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2459
2460         /*
2461          * This silly early PAGE_DIRTY setting removes a race
2462          * due to the bad i386 page protection. But it's valid
2463          * for other architectures too.
2464          *
2465          * Note that if write_access is true, we either now have
2466          * an exclusive copy of the page, or this is a shared mapping,
2467          * so we can make it writable and dirty to avoid having to
2468          * handle that later.
2469          */
2470         /* Only go through if we didn't race with anybody else... */
2471         if (likely(pte_same(*page_table, orig_pte))) {
2472                 flush_icache_page(vma, page);
2473                 entry = mk_pte(page, vma->vm_page_prot);
2474                 if (flags & FAULT_FLAG_WRITE)
2475                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2476                 set_pte_at(mm, address, page_table, entry);
2477                 if (anon) {
2478                         inc_mm_counter(mm, anon_rss);
2479                         lru_cache_add_active(page);
2480                         page_add_new_anon_rmap(page, vma, address);
2481                 } else {
2482                         inc_mm_counter(mm, file_rss);
2483                         page_add_file_rmap(page);
2484                         if (flags & FAULT_FLAG_WRITE) {
2485                                 dirty_page = page;
2486                                 get_page(dirty_page);
2487                         }
2488                 }
2489
2490                 /* no need to invalidate: a not-present page won't be cached */
2491                 update_mmu_cache(vma, address, entry);
2492         } else {
2493                 mem_cgroup_uncharge_page(page);
2494                 if (anon)
2495                         page_cache_release(page);
2496                 else
2497                         anon = 1; /* no anon but release faulted_page */
2498         }
2499
2500         pte_unmap_unlock(page_table, ptl);
2501
2502 out:
2503         unlock_page(vmf.page);
2504 out_unlocked:
2505         if (anon)
2506                 page_cache_release(vmf.page);
2507         else if (dirty_page) {
2508                 if (vma->vm_file)
2509                         file_update_time(vma->vm_file);
2510
2511                 set_page_dirty_balance(dirty_page, page_mkwrite);
2512                 put_page(dirty_page);
2513         }
2514
2515         return ret;
2516 }
2517
2518 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2519                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2520                 int write_access, pte_t orig_pte)
2521 {
2522         pgoff_t pgoff = (((address & PAGE_MASK)
2523                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2524         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2525
2526         pte_unmap(page_table);
2527         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2528 }
2529
2530 /*
2531  * Fault of a previously existing named mapping. Repopulate the pte
2532  * from the encoded file_pte if possible. This enables swappable
2533  * nonlinear vmas.
2534  *
2535  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2536  * but allow concurrent faults), and pte mapped but not yet locked.
2537  * We return with mmap_sem still held, but pte unmapped and unlocked.
2538  */
2539 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2540                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2541                 int write_access, pte_t orig_pte)
2542 {
2543         unsigned int flags = FAULT_FLAG_NONLINEAR |
2544                                 (write_access ? FAULT_FLAG_WRITE : 0);
2545         pgoff_t pgoff;
2546
2547         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2548                 return 0;
2549
2550         if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2551                         !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2552                 /*
2553                  * Page table corrupted: show pte and kill process.
2554                  */
2555                 print_bad_pte(vma, orig_pte, address);
2556                 return VM_FAULT_OOM;
2557         }
2558
2559         pgoff = pte_to_pgoff(orig_pte);
2560         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2561 }
2562
2563 /*
2564  * These routines also need to handle stuff like marking pages dirty
2565  * and/or accessed for architectures that don't do it in hardware (most
2566  * RISC architectures).  The early dirtying is also good on the i386.
2567  *
2568  * There is also a hook called "update_mmu_cache()" that architectures
2569  * with external mmu caches can use to update those (ie the Sparc or
2570  * PowerPC hashed page tables that act as extended TLBs).
2571  *
2572  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2573  * but allow concurrent faults), and pte mapped but not yet locked.
2574  * We return with mmap_sem still held, but pte unmapped and unlocked.
2575  */
2576 static inline int handle_pte_fault(struct mm_struct *mm,
2577                 struct vm_area_struct *vma, unsigned long address,
2578                 pte_t *pte, pmd_t *pmd, int write_access)
2579 {
2580         pte_t entry;
2581         spinlock_t *ptl;
2582
2583         entry = *pte;
2584         if (!pte_present(entry)) {
2585                 if (pte_none(entry)) {
2586                         if (vma->vm_ops) {
2587                                 if (likely(vma->vm_ops->fault))
2588                                         return do_linear_fault(mm, vma, address,
2589                                                 pte, pmd, write_access, entry);
2590                         }
2591                         return do_anonymous_page(mm, vma, address,
2592                                                  pte, pmd, write_access);
2593                 }
2594                 if (pte_file(entry))
2595                         return do_nonlinear_fault(mm, vma, address,
2596                                         pte, pmd, write_access, entry);
2597                 return do_swap_page(mm, vma, address,
2598                                         pte, pmd, write_access, entry);
2599         }
2600
2601         ptl = pte_lockptr(mm, pmd);
2602         spin_lock(ptl);
2603         if (unlikely(!pte_same(*pte, entry)))
2604                 goto unlock;
2605         if (write_access) {
2606                 if (!pte_write(entry))
2607                         return do_wp_page(mm, vma, address,
2608                                         pte, pmd, ptl, entry);
2609                 entry = pte_mkdirty(entry);
2610         }
2611         entry = pte_mkyoung(entry);
2612         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2613                 update_mmu_cache(vma, address, entry);
2614         } else {
2615                 /*
2616                  * This is needed only for protection faults but the arch code
2617                  * is not yet telling us if this is a protection fault or not.
2618                  * This still avoids useless tlb flushes for .text page faults
2619                  * with threads.
2620                  */
2621                 if (write_access)
2622                         flush_tlb_page(vma, address);
2623         }
2624 unlock:
2625         pte_unmap_unlock(pte, ptl);
2626         return 0;
2627 }
2628
2629 /*
2630  * By the time we get here, we already hold the mm semaphore
2631  */
2632 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2633                 unsigned long address, int write_access)
2634 {
2635         pgd_t *pgd;
2636         pud_t *pud;
2637         pmd_t *pmd;
2638         pte_t *pte;
2639
2640         __set_current_state(TASK_RUNNING);
2641
2642         count_vm_event(PGFAULT);
2643
2644         if (unlikely(is_vm_hugetlb_page(vma)))
2645                 return hugetlb_fault(mm, vma, address, write_access);
2646
2647         pgd = pgd_offset(mm, address);
2648         pud = pud_alloc(mm, pgd, address);
2649         if (!pud)
2650                 return VM_FAULT_OOM;
2651         pmd = pmd_alloc(mm, pud, address);
2652         if (!pmd)
2653                 return VM_FAULT_OOM;
2654         pte = pte_alloc_map(mm, pmd, address);
2655         if (!pte)
2656                 return VM_FAULT_OOM;
2657
2658         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2659 }
2660
2661 #ifndef __PAGETABLE_PUD_FOLDED
2662 /*
2663  * Allocate page upper directory.
2664  * We've already handled the fast-path in-line.
2665  */
2666 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2667 {
2668         pud_t *new = pud_alloc_one(mm, address);
2669         if (!new)
2670                 return -ENOMEM;
2671
2672         smp_wmb(); /* See comment in __pte_alloc */
2673
2674         spin_lock(&mm->page_table_lock);
2675         if (pgd_present(*pgd))          /* Another has populated it */
2676                 pud_free(mm, new);
2677         else
2678                 pgd_populate(mm, pgd, new);
2679         spin_unlock(&mm->page_table_lock);
2680         return 0;
2681 }
2682 #endif /* __PAGETABLE_PUD_FOLDED */
2683
2684 #ifndef __PAGETABLE_PMD_FOLDED
2685 /*
2686  * Allocate page middle directory.
2687  * We've already handled the fast-path in-line.
2688  */
2689 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2690 {
2691         pmd_t *new = pmd_alloc_one(mm, address);
2692         if (!new)
2693                 return -ENOMEM;
2694
2695         smp_wmb(); /* See comment in __pte_alloc */
2696
2697         spin_lock(&mm->page_table_lock);
2698 #ifndef __ARCH_HAS_4LEVEL_HACK
2699         if (pud_present(*pud))          /* Another has populated it */
2700                 pmd_free(mm, new);
2701         else
2702                 pud_populate(mm, pud, new);
2703 #else
2704         if (pgd_present(*pud))          /* Another has populated it */
2705                 pmd_free(mm, new);
2706         else
2707                 pgd_populate(mm, pud, new);
2708 #endif /* __ARCH_HAS_4LEVEL_HACK */
2709         spin_unlock(&mm->page_table_lock);
2710         return 0;
2711 }
2712 #endif /* __PAGETABLE_PMD_FOLDED */
2713
2714 int make_pages_present(unsigned long addr, unsigned long end)
2715 {
2716         int ret, len, write;
2717         struct vm_area_struct * vma;
2718
2719         vma = find_vma(current->mm, addr);
2720         if (!vma)
2721                 return -1;
2722         write = (vma->vm_flags & VM_WRITE) != 0;
2723         BUG_ON(addr >= end);
2724         BUG_ON(end > vma->vm_end);
2725         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2726         ret = get_user_pages(current, current->mm, addr,
2727                         len, write, 0, NULL, NULL);
2728         if (ret < 0)
2729                 return ret;
2730         return ret == len ? 0 : -1;
2731 }
2732
2733 #if !defined(__HAVE_ARCH_GATE_AREA)
2734
2735 #if defined(AT_SYSINFO_EHDR)
2736 static struct vm_area_struct gate_vma;
2737
2738 static int __init gate_vma_init(void)
2739 {
2740         gate_vma.vm_mm = NULL;
2741         gate_vma.vm_start = FIXADDR_USER_START;
2742         gate_vma.vm_end = FIXADDR_USER_END;
2743         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2744         gate_vma.vm_page_prot = __P101;
2745         /*
2746          * Make sure the vDSO gets into every core dump.
2747          * Dumping its contents makes post-mortem fully interpretable later
2748          * without matching up the same kernel and hardware config to see
2749          * what PC values meant.
2750          */
2751         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2752         return 0;
2753 }
2754 __initcall(gate_vma_init);
2755 #endif
2756
2757 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2758 {
2759 #ifdef AT_SYSINFO_EHDR
2760         return &gate_vma;
2761 #else
2762         return NULL;
2763 #endif
2764 }
2765
2766 int in_gate_area_no_task(unsigned long addr)
2767 {
2768 #ifdef AT_SYSINFO_EHDR
2769         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2770                 return 1;
2771 #endif
2772         return 0;
2773 }
2774
2775 #endif  /* __HAVE_ARCH_GATE_AREA */
2776
2777 #ifdef CONFIG_HAVE_IOREMAP_PROT
2778 static resource_size_t follow_phys(struct vm_area_struct *vma,
2779                         unsigned long address, unsigned int flags,
2780                         unsigned long *prot)
2781 {
2782         pgd_t *pgd;
2783         pud_t *pud;
2784         pmd_t *pmd;
2785         pte_t *ptep, pte;
2786         spinlock_t *ptl;
2787         resource_size_t phys_addr = 0;
2788         struct mm_struct *mm = vma->vm_mm;
2789
2790         VM_BUG_ON(!(vma->vm_flags & (VM_IO | VM_PFNMAP)));
2791
2792         pgd = pgd_offset(mm, address);
2793         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
2794                 goto no_page_table;
2795
2796         pud = pud_offset(pgd, address);
2797         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
2798                 goto no_page_table;
2799
2800         pmd = pmd_offset(pud, address);
2801         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
2802                 goto no_page_table;
2803
2804         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
2805         if (pmd_huge(*pmd))
2806                 goto no_page_table;
2807
2808         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
2809         if (!ptep)
2810                 goto out;
2811
2812         pte = *ptep;
2813         if (!pte_present(pte))
2814                 goto unlock;
2815         if ((flags & FOLL_WRITE) && !pte_write(pte))
2816                 goto unlock;
2817         phys_addr = pte_pfn(pte);
2818         phys_addr <<= PAGE_SHIFT; /* Shift here to avoid overflow on PAE */
2819
2820         *prot = pgprot_val(pte_pgprot(pte));
2821
2822 unlock:
2823         pte_unmap_unlock(ptep, ptl);
2824 out:
2825         return phys_addr;
2826 no_page_table:
2827         return 0;
2828 }
2829
2830 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
2831                         void *buf, int len, int write)
2832 {
2833         resource_size_t phys_addr;
2834         unsigned long prot = 0;
2835         void *maddr;
2836         int offset = addr & (PAGE_SIZE-1);
2837
2838         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
2839                 return -EINVAL;
2840
2841         phys_addr = follow_phys(vma, addr, write, &prot);
2842
2843         if (!phys_addr)
2844                 return -EINVAL;
2845
2846         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
2847         if (write)
2848                 memcpy_toio(maddr + offset, buf, len);
2849         else
2850                 memcpy_fromio(buf, maddr + offset, len);
2851         iounmap(maddr);
2852
2853         return len;
2854 }
2855 #endif
2856
2857 /*
2858  * Access another process' address space.
2859  * Source/target buffer must be kernel space,
2860  * Do not walk the page table directly, use get_user_pages
2861  */
2862 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2863 {
2864         struct mm_struct *mm;
2865         struct vm_area_struct *vma;
2866         void *old_buf = buf;
2867
2868         mm = get_task_mm(tsk);
2869         if (!mm)
2870                 return 0;
2871
2872         down_read(&mm->mmap_sem);
2873         /* ignore errors, just check how much was successfully transferred */
2874         while (len) {
2875                 int bytes, ret, offset;
2876                 void *maddr;
2877                 struct page *page = NULL;
2878
2879                 ret = get_user_pages(tsk, mm, addr, 1,
2880                                 write, 1, &page, &vma);
2881                 if (ret <= 0) {
2882                         /*
2883                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
2884                          * we can access using slightly different code.
2885                          */
2886 #ifdef CONFIG_HAVE_IOREMAP_PROT
2887                         vma = find_vma(mm, addr);
2888                         if (!vma)
2889                                 break;
2890                         if (vma->vm_ops && vma->vm_ops->access)
2891                                 ret = vma->vm_ops->access(vma, addr, buf,
2892                                                           len, write);
2893                         if (ret <= 0)
2894 #endif
2895                                 break;
2896                         bytes = ret;
2897                 } else {
2898                         bytes = len;
2899                         offset = addr & (PAGE_SIZE-1);
2900                         if (bytes > PAGE_SIZE-offset)
2901                                 bytes = PAGE_SIZE-offset;
2902
2903                         maddr = kmap(page);
2904                         if (write) {
2905                                 copy_to_user_page(vma, page, addr,
2906                                                   maddr + offset, buf, bytes);
2907                                 set_page_dirty_lock(page);
2908                         } else {
2909                                 copy_from_user_page(vma, page, addr,
2910                                                     buf, maddr + offset, bytes);
2911                         }
2912                         kunmap(page);
2913                         page_cache_release(page);
2914                 }
2915                 len -= bytes;
2916                 buf += bytes;
2917                 addr += bytes;
2918         }
2919         up_read(&mm->mmap_sem);
2920         mmput(mm);
2921
2922         return buf - old_buf;
2923 }
2924
2925 /*
2926  * Print the name of a VMA.
2927  */
2928 void print_vma_addr(char *prefix, unsigned long ip)
2929 {
2930         struct mm_struct *mm = current->mm;
2931         struct vm_area_struct *vma;
2932
2933         /*
2934          * Do not print if we are in atomic
2935          * contexts (in exception stacks, etc.):
2936          */
2937         if (preempt_count())
2938                 return;
2939
2940         down_read(&mm->mmap_sem);
2941         vma = find_vma(mm, ip);
2942         if (vma && vma->vm_file) {
2943                 struct file *f = vma->vm_file;
2944                 char *buf = (char *)__get_free_page(GFP_KERNEL);
2945                 if (buf) {
2946                         char *p, *s;
2947
2948                         p = d_path(&f->f_path, buf, PAGE_SIZE);
2949                         if (IS_ERR(p))
2950                                 p = "?";
2951                         s = strrchr(p, '/');
2952                         if (s)
2953                                 p = s+1;
2954                         printk("%s%s[%lx+%lx]", prefix, p,
2955                                         vma->vm_start,
2956                                         vma->vm_end - vma->vm_start);
2957                         free_page((unsigned long)buf);
2958                 }
2959         }
2960         up_read(&current->mm->mmap_sem);
2961 }