Version:  2.0.40 2.2.26 2.4.37 3.0 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16

Linux/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/ksm.h>
 49 #include <linux/rmap.h>
 50 #include <linux/module.h>
 51 #include <linux/delayacct.h>
 52 #include <linux/init.h>
 53 #include <linux/writeback.h>
 54 #include <linux/memcontrol.h>
 55 #include <linux/mmu_notifier.h>
 56 #include <linux/kallsyms.h>
 57 #include <linux/swapops.h>
 58 #include <linux/elf.h>
 59 #include <linux/gfp.h>
 60 
 61 #include <asm/io.h>
 62 #include <asm/pgalloc.h>
 63 #include <asm/uaccess.h>
 64 #include <asm/tlb.h>
 65 #include <asm/tlbflush.h>
 66 #include <asm/pgtable.h>
 67 
 68 #include "internal.h"
 69 
 70 #ifndef CONFIG_NEED_MULTIPLE_NODES
 71 /* use the per-pgdat data instead for discontigmem - mbligh */
 72 unsigned long max_mapnr;
 73 struct page *mem_map;
 74 
 75 EXPORT_SYMBOL(max_mapnr);
 76 EXPORT_SYMBOL(mem_map);
 77 #endif
 78 
 79 unsigned long num_physpages;
 80 /*
 81  * A number of key systems in x86 including ioremap() rely on the assumption
 82  * that high_memory defines the upper bound on direct map memory, then end
 83  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
 84  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
 85  * and ZONE_HIGHMEM.
 86  */
 87 void * high_memory;
 88 
 89 EXPORT_SYMBOL(num_physpages);
 90 EXPORT_SYMBOL(high_memory);
 91 
 92 /*
 93  * Randomize the address space (stacks, mmaps, brk, etc.).
 94  *
 95  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
 96  *   as ancient (libc5 based) binaries can segfault. )
 97  */
 98 int randomize_va_space __read_mostly =
 99 #ifdef CONFIG_COMPAT_BRK
100                                         1;
101 #else
102                                         2;
103 #endif
104 
105 static int __init disable_randmaps(char *s)
106 {
107         randomize_va_space = 0;
108         return 1;
109 }
110 __setup("norandmaps", disable_randmaps);
111 
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
114 
115 /*
116  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117  */
118 static int __init init_zero_pfn(void)
119 {
120         zero_pfn = page_to_pfn(ZERO_PAGE(0));
121         return 0;
122 }
123 core_initcall(init_zero_pfn);
124 
125 
126 #if defined(SPLIT_RSS_COUNTING)
127 
128 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
129 {
130         int i;
131 
132         for (i = 0; i < NR_MM_COUNTERS; i++) {
133                 if (task->rss_stat.count[i]) {
134                         add_mm_counter(mm, i, task->rss_stat.count[i]);
135                         task->rss_stat.count[i] = 0;
136                 }
137         }
138         task->rss_stat.events = 0;
139 }
140 
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142 {
143         struct task_struct *task = current;
144 
145         if (likely(task->mm == mm))
146                 task->rss_stat.count[member] += val;
147         else
148                 add_mm_counter(mm, member, val);
149 }
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
152 
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH  (64)
155 static void check_sync_rss_stat(struct task_struct *task)
156 {
157         if (unlikely(task != current))
158                 return;
159         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160                 __sync_task_rss_stat(task, task->mm);
161 }
162 
163 unsigned long get_mm_counter(struct mm_struct *mm, int member)
164 {
165         long val = 0;
166 
167         /*
168          * Don't use task->mm here...for avoiding to use task_get_mm()..
169          * The caller must guarantee task->mm is not invalid.
170          */
171         val = atomic_long_read(&mm->rss_stat.count[member]);
172         /*
173          * counter is updated in asynchronous manner and may go to minus.
174          * But it's never be expected number for users.
175          */
176         if (val < 0)
177                 return 0;
178         return (unsigned long)val;
179 }
180 
181 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182 {
183         __sync_task_rss_stat(task, mm);
184 }
185 #else
186 
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
189 
190 static void check_sync_rss_stat(struct task_struct *task)
191 {
192 }
193 
194 #endif
195 
196 /*
197  * If a p?d_bad entry is found while walking page tables, report
198  * the error, before resetting entry to p?d_none.  Usually (but
199  * very seldom) called out from the p?d_none_or_clear_bad macros.
200  */
201 
202 void pgd_clear_bad(pgd_t *pgd)
203 {
204         pgd_ERROR(*pgd);
205         pgd_clear(pgd);
206 }
207 
208 void pud_clear_bad(pud_t *pud)
209 {
210         pud_ERROR(*pud);
211         pud_clear(pud);
212 }
213 
214 void pmd_clear_bad(pmd_t *pmd)
215 {
216         pmd_ERROR(*pmd);
217         pmd_clear(pmd);
218 }
219 
220 /*
221  * Note: this doesn't free the actual pages themselves. That
222  * has been handled earlier when unmapping all the memory regions.
223  */
224 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
225                            unsigned long addr)
226 {
227         pgtable_t token = pmd_pgtable(*pmd);
228         pmd_clear(pmd);
229         pte_free_tlb(tlb, token, addr);
230         tlb->mm->nr_ptes--;
231 }
232 
233 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
234                                 unsigned long addr, unsigned long end,
235                                 unsigned long floor, unsigned long ceiling)
236 {
237         pmd_t *pmd;
238         unsigned long next;
239         unsigned long start;
240 
241         start = addr;
242         pmd = pmd_offset(pud, addr);
243         do {
244                 next = pmd_addr_end(addr, end);
245                 if (pmd_none_or_clear_bad(pmd))
246                         continue;
247                 free_pte_range(tlb, pmd, addr);
248         } while (pmd++, addr = next, addr != end);
249 
250         start &= PUD_MASK;
251         if (start < floor)
252                 return;
253         if (ceiling) {
254                 ceiling &= PUD_MASK;
255                 if (!ceiling)
256                         return;
257         }
258         if (end - 1 > ceiling - 1)
259                 return;
260 
261         pmd = pmd_offset(pud, start);
262         pud_clear(pud);
263         pmd_free_tlb(tlb, pmd, start);
264 }
265 
266 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
267                                 unsigned long addr, unsigned long end,
268                                 unsigned long floor, unsigned long ceiling)
269 {
270         pud_t *pud;
271         unsigned long next;
272         unsigned long start;
273 
274         start = addr;
275         pud = pud_offset(pgd, addr);
276         do {
277                 next = pud_addr_end(addr, end);
278                 if (pud_none_or_clear_bad(pud))
279                         continue;
280                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
281         } while (pud++, addr = next, addr != end);
282 
283         start &= PGDIR_MASK;
284         if (start < floor)
285                 return;
286         if (ceiling) {
287                 ceiling &= PGDIR_MASK;
288                 if (!ceiling)
289                         return;
290         }
291         if (end - 1 > ceiling - 1)
292                 return;
293 
294         pud = pud_offset(pgd, start);
295         pgd_clear(pgd);
296         pud_free_tlb(tlb, pud, start);
297 }
298 
299 /*
300  * This function frees user-level page tables of a process.
301  *
302  * Must be called with pagetable lock held.
303  */
304 void free_pgd_range(struct mmu_gather *tlb,
305                         unsigned long addr, unsigned long end,
306                         unsigned long floor, unsigned long ceiling)
307 {
308         pgd_t *pgd;
309         unsigned long next;
310 
311         /*
312          * The next few lines have given us lots of grief...
313          *
314          * Why are we testing PMD* at this top level?  Because often
315          * there will be no work to do at all, and we'd prefer not to
316          * go all the way down to the bottom just to discover that.
317          *
318          * Why all these "- 1"s?  Because 0 represents both the bottom
319          * of the address space and the top of it (using -1 for the
320          * top wouldn't help much: the masks would do the wrong thing).
321          * The rule is that addr 0 and floor 0 refer to the bottom of
322          * the address space, but end 0 and ceiling 0 refer to the top
323          * Comparisons need to use "end - 1" and "ceiling - 1" (though
324          * that end 0 case should be mythical).
325          *
326          * Wherever addr is brought up or ceiling brought down, we must
327          * be careful to reject "the opposite 0" before it confuses the
328          * subsequent tests.  But what about where end is brought down
329          * by PMD_SIZE below? no, end can't go down to 0 there.
330          *
331          * Whereas we round start (addr) and ceiling down, by different
332          * masks at different levels, in order to test whether a table
333          * now has no other vmas using it, so can be freed, we don't
334          * bother to round floor or end up - the tests don't need that.
335          */
336 
337         addr &= PMD_MASK;
338         if (addr < floor) {
339                 addr += PMD_SIZE;
340                 if (!addr)
341                         return;
342         }
343         if (ceiling) {
344                 ceiling &= PMD_MASK;
345                 if (!ceiling)
346                         return;
347         }
348         if (end - 1 > ceiling - 1)
349                 end -= PMD_SIZE;
350         if (addr > end - 1)
351                 return;
352 
353         pgd = pgd_offset(tlb->mm, addr);
354         do {
355                 next = pgd_addr_end(addr, end);
356                 if (pgd_none_or_clear_bad(pgd))
357                         continue;
358                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
359         } while (pgd++, addr = next, addr != end);
360 }
361 
362 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
363                 unsigned long floor, unsigned long ceiling)
364 {
365         while (vma) {
366                 struct vm_area_struct *next = vma->vm_next;
367                 unsigned long addr = vma->vm_start;
368 
369                 /*
370                  * Hide vma from rmap and truncate_pagecache before freeing
371                  * pgtables
372                  */
373                 unlink_anon_vmas(vma);
374                 unlink_file_vma(vma);
375 
376                 if (is_vm_hugetlb_page(vma)) {
377                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
378                                 floor, next? next->vm_start: ceiling);
379                 } else {
380                         /*
381                          * Optimization: gather nearby vmas into one call down
382                          */
383                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
384                                && !is_vm_hugetlb_page(next)) {
385                                 vma = next;
386                                 next = vma->vm_next;
387                                 unlink_anon_vmas(vma);
388                                 unlink_file_vma(vma);
389                         }
390                         free_pgd_range(tlb, addr, vma->vm_end,
391                                 floor, next? next->vm_start: ceiling);
392                 }
393                 vma = next;
394         }
395 }
396 
397 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
398                 pmd_t *pmd, unsigned long address)
399 {
400         pgtable_t new = pte_alloc_one(mm, address);
401         int wait_split_huge_page;
402         if (!new)
403                 return -ENOMEM;
404 
405         /*
406          * Ensure all pte setup (eg. pte page lock and page clearing) are
407          * visible before the pte is made visible to other CPUs by being
408          * put into page tables.
409          *
410          * The other side of the story is the pointer chasing in the page
411          * table walking code (when walking the page table without locking;
412          * ie. most of the time). Fortunately, these data accesses consist
413          * of a chain of data-dependent loads, meaning most CPUs (alpha
414          * being the notable exception) will already guarantee loads are
415          * seen in-order. See the alpha page table accessors for the
416          * smp_read_barrier_depends() barriers in page table walking code.
417          */
418         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
419 
420         spin_lock(&mm->page_table_lock);
421         wait_split_huge_page = 0;
422         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
423                 mm->nr_ptes++;
424                 pmd_populate(mm, pmd, new);
425                 new = NULL;
426         } else if (unlikely(pmd_trans_splitting(*pmd)))
427                 wait_split_huge_page = 1;
428         spin_unlock(&mm->page_table_lock);
429         if (new)
430                 pte_free(mm, new);
431         if (wait_split_huge_page)
432                 wait_split_huge_page(vma->anon_vma, pmd);
433         return 0;
434 }
435 
436 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
437 {
438         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
439         if (!new)
440                 return -ENOMEM;
441 
442         smp_wmb(); /* See comment in __pte_alloc */
443 
444         spin_lock(&init_mm.page_table_lock);
445         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
446                 pmd_populate_kernel(&init_mm, pmd, new);
447                 new = NULL;
448         } else
449                 VM_BUG_ON(pmd_trans_splitting(*pmd));
450         spin_unlock(&init_mm.page_table_lock);
451         if (new)
452                 pte_free_kernel(&init_mm, new);
453         return 0;
454 }
455 
456 static inline void init_rss_vec(int *rss)
457 {
458         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
459 }
460 
461 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
462 {
463         int i;
464 
465         if (current->mm == mm)
466                 sync_mm_rss(current, mm);
467         for (i = 0; i < NR_MM_COUNTERS; i++)
468                 if (rss[i])
469                         add_mm_counter(mm, i, rss[i]);
470 }
471 
472 /*
473  * This function is called to print an error when a bad pte
474  * is found. For example, we might have a PFN-mapped pte in
475  * a region that doesn't allow it.
476  *
477  * The calling function must still handle the error.
478  */
479 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
480                           pte_t pte, struct page *page)
481 {
482         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
483         pud_t *pud = pud_offset(pgd, addr);
484         pmd_t *pmd = pmd_offset(pud, addr);
485         struct address_space *mapping;
486         pgoff_t index;
487         static unsigned long resume;
488         static unsigned long nr_shown;
489         static unsigned long nr_unshown;
490 
491         /*
492          * Allow a burst of 60 reports, then keep quiet for that minute;
493          * or allow a steady drip of one report per second.
494          */
495         if (nr_shown == 60) {
496                 if (time_before(jiffies, resume)) {
497                         nr_unshown++;
498                         return;
499                 }
500                 if (nr_unshown) {
501                         printk(KERN_ALERT
502                                 "BUG: Bad page map: %lu messages suppressed\n",
503                                 nr_unshown);
504                         nr_unshown = 0;
505                 }
506                 nr_shown = 0;
507         }
508         if (nr_shown++ == 0)
509                 resume = jiffies + 60 * HZ;
510 
511         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
512         index = linear_page_index(vma, addr);
513 
514         printk(KERN_ALERT
515                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
516                 current->comm,
517                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
518         if (page)
519                 dump_page(page);
520         printk(KERN_ALERT
521                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
522                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
523         /*
524          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
525          */
526         if (vma->vm_ops)
527                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
528                                 (unsigned long)vma->vm_ops->fault);
529         if (vma->vm_file && vma->vm_file->f_op)
530                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
531                                 (unsigned long)vma->vm_file->f_op->mmap);
532         dump_stack();
533         add_taint(TAINT_BAD_PAGE);
534 }
535 
536 static inline int is_cow_mapping(unsigned int flags)
537 {
538         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
539 }
540 
541 #ifndef is_zero_pfn
542 static inline int is_zero_pfn(unsigned long pfn)
543 {
544         return pfn == zero_pfn;
545 }
546 #endif
547 
548 #ifndef my_zero_pfn
549 static inline unsigned long my_zero_pfn(unsigned long addr)
550 {
551         return zero_pfn;
552 }
553 #endif
554 
555 /*
556  * vm_normal_page -- This function gets the "struct page" associated with a pte.
557  *
558  * "Special" mappings do not wish to be associated with a "struct page" (either
559  * it doesn't exist, or it exists but they don't want to touch it). In this
560  * case, NULL is returned here. "Normal" mappings do have a struct page.
561  *
562  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
563  * pte bit, in which case this function is trivial. Secondly, an architecture
564  * may not have a spare pte bit, which requires a more complicated scheme,
565  * described below.
566  *
567  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
568  * special mapping (even if there are underlying and valid "struct pages").
569  * COWed pages of a VM_PFNMAP are always normal.
570  *
571  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
572  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
573  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
574  * mapping will always honor the rule
575  *
576  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
577  *
578  * And for normal mappings this is false.
579  *
580  * This restricts such mappings to be a linear translation from virtual address
581  * to pfn. To get around this restriction, we allow arbitrary mappings so long
582  * as the vma is not a COW mapping; in that case, we know that all ptes are
583  * special (because none can have been COWed).
584  *
585  *
586  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
587  *
588  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
589  * page" backing, however the difference is that _all_ pages with a struct
590  * page (that is, those where pfn_valid is true) are refcounted and considered
591  * normal pages by the VM. The disadvantage is that pages are refcounted
592  * (which can be slower and simply not an option for some PFNMAP users). The
593  * advantage is that we don't have to follow the strict linearity rule of
594  * PFNMAP mappings in order to support COWable mappings.
595  *
596  */
597 #ifdef __HAVE_ARCH_PTE_SPECIAL
598 # define HAVE_PTE_SPECIAL 1
599 #else
600 # define HAVE_PTE_SPECIAL 0
601 #endif
602 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
603                                 pte_t pte)
604 {
605         unsigned long pfn = pte_pfn(pte);
606 
607         if (HAVE_PTE_SPECIAL) {
608                 if (likely(!pte_special(pte)))
609                         goto check_pfn;
610                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
611                         return NULL;
612                 if (!is_zero_pfn(pfn))
613                         print_bad_pte(vma, addr, pte, NULL);
614                 return NULL;
615         }
616 
617         /* !HAVE_PTE_SPECIAL case follows: */
618 
619         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
620                 if (vma->vm_flags & VM_MIXEDMAP) {
621                         if (!pfn_valid(pfn))
622                                 return NULL;
623                         goto out;
624                 } else {
625                         unsigned long off;
626                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
627                         if (pfn == vma->vm_pgoff + off)
628                                 return NULL;
629                         if (!is_cow_mapping(vma->vm_flags))
630                                 return NULL;
631                 }
632         }
633 
634         if (is_zero_pfn(pfn))
635                 return NULL;
636 check_pfn:
637         if (unlikely(pfn > highest_memmap_pfn)) {
638                 print_bad_pte(vma, addr, pte, NULL);
639                 return NULL;
640         }
641 
642         /*
643          * NOTE! We still have PageReserved() pages in the page tables.
644          * eg. VDSO mappings can cause them to exist.
645          */
646 out:
647         return pfn_to_page(pfn);
648 }
649 
650 /*
651  * copy one vm_area from one task to the other. Assumes the page tables
652  * already present in the new task to be cleared in the whole range
653  * covered by this vma.
654  */
655 
656 static inline unsigned long
657 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
658                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
659                 unsigned long addr, int *rss)
660 {
661         unsigned long vm_flags = vma->vm_flags;
662         pte_t pte = *src_pte;
663         struct page *page;
664 
665         /* pte contains position in swap or file, so copy. */
666         if (unlikely(!pte_present(pte))) {
667                 if (!pte_file(pte)) {
668                         swp_entry_t entry = pte_to_swp_entry(pte);
669 
670                         if (swap_duplicate(entry) < 0)
671                                 return entry.val;
672 
673                         /* make sure dst_mm is on swapoff's mmlist. */
674                         if (unlikely(list_empty(&dst_mm->mmlist))) {
675                                 spin_lock(&mmlist_lock);
676                                 if (list_empty(&dst_mm->mmlist))
677                                         list_add(&dst_mm->mmlist,
678                                                  &src_mm->mmlist);
679                                 spin_unlock(&mmlist_lock);
680                         }
681                         if (likely(!non_swap_entry(entry)))
682                                 rss[MM_SWAPENTS]++;
683                         else if (is_write_migration_entry(entry) &&
684                                         is_cow_mapping(vm_flags)) {
685                                 /*
686                                  * COW mappings require pages in both parent
687                                  * and child to be set to read.
688                                  */
689                                 make_migration_entry_read(&entry);
690                                 pte = swp_entry_to_pte(entry);
691                                 set_pte_at(src_mm, addr, src_pte, pte);
692                         }
693                 }
694                 goto out_set_pte;
695         }
696 
697         /*
698          * If it's a COW mapping, write protect it both
699          * in the parent and the child
700          */
701         if (is_cow_mapping(vm_flags)) {
702                 ptep_set_wrprotect(src_mm, addr, src_pte);
703                 pte = pte_wrprotect(pte);
704         }
705 
706         /*
707          * If it's a shared mapping, mark it clean in
708          * the child
709          */
710         if (vm_flags & VM_SHARED)
711                 pte = pte_mkclean(pte);
712         pte = pte_mkold(pte);
713 
714         page = vm_normal_page(vma, addr, pte);
715         if (page) {
716                 get_page(page);
717                 page_dup_rmap(page);
718                 if (PageAnon(page))
719                         rss[MM_ANONPAGES]++;
720                 else
721                         rss[MM_FILEPAGES]++;
722         }
723 
724 out_set_pte:
725         set_pte_at(dst_mm, addr, dst_pte, pte);
726         return 0;
727 }
728 
729 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
730                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
731                    unsigned long addr, unsigned long end)
732 {
733         pte_t *orig_src_pte, *orig_dst_pte;
734         pte_t *src_pte, *dst_pte;
735         spinlock_t *src_ptl, *dst_ptl;
736         int progress = 0;
737         int rss[NR_MM_COUNTERS];
738         swp_entry_t entry = (swp_entry_t){0};
739 
740 again:
741         init_rss_vec(rss);
742 
743         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
744         if (!dst_pte)
745                 return -ENOMEM;
746         src_pte = pte_offset_map(src_pmd, addr);
747         src_ptl = pte_lockptr(src_mm, src_pmd);
748         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
749         orig_src_pte = src_pte;
750         orig_dst_pte = dst_pte;
751         arch_enter_lazy_mmu_mode();
752 
753         do {
754                 /*
755                  * We are holding two locks at this point - either of them
756                  * could generate latencies in another task on another CPU.
757                  */
758                 if (progress >= 32) {
759                         progress = 0;
760                         if (need_resched() ||
761                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
762                                 break;
763                 }
764                 if (pte_none(*src_pte)) {
765                         progress++;
766                         continue;
767                 }
768                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
769                                                         vma, addr, rss);
770                 if (entry.val)
771                         break;
772                 progress += 8;
773         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
774 
775         arch_leave_lazy_mmu_mode();
776         spin_unlock(src_ptl);
777         pte_unmap(orig_src_pte);
778         add_mm_rss_vec(dst_mm, rss);
779         pte_unmap_unlock(orig_dst_pte, dst_ptl);
780         cond_resched();
781 
782         if (entry.val) {
783                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
784                         return -ENOMEM;
785                 progress = 0;
786         }
787         if (addr != end)
788                 goto again;
789         return 0;
790 }
791 
792 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
793                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
794                 unsigned long addr, unsigned long end)
795 {
796         pmd_t *src_pmd, *dst_pmd;
797         unsigned long next;
798 
799         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
800         if (!dst_pmd)
801                 return -ENOMEM;
802         src_pmd = pmd_offset(src_pud, addr);
803         do {
804                 next = pmd_addr_end(addr, end);
805                 if (pmd_trans_huge(*src_pmd)) {
806                         int err;
807                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
808                         err = copy_huge_pmd(dst_mm, src_mm,
809                                             dst_pmd, src_pmd, addr, vma);
810                         if (err == -ENOMEM)
811                                 return -ENOMEM;
812                         if (!err)
813                                 continue;
814                         /* fall through */
815                 }
816                 if (pmd_none_or_clear_bad(src_pmd))
817                         continue;
818                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
819                                                 vma, addr, next))
820                         return -ENOMEM;
821         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
822         return 0;
823 }
824 
825 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
826                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
827                 unsigned long addr, unsigned long end)
828 {
829         pud_t *src_pud, *dst_pud;
830         unsigned long next;
831 
832         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
833         if (!dst_pud)
834                 return -ENOMEM;
835         src_pud = pud_offset(src_pgd, addr);
836         do {
837                 next = pud_addr_end(addr, end);
838                 if (pud_none_or_clear_bad(src_pud))
839                         continue;
840                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
841                                                 vma, addr, next))
842                         return -ENOMEM;
843         } while (dst_pud++, src_pud++, addr = next, addr != end);
844         return 0;
845 }
846 
847 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
848                 struct vm_area_struct *vma)
849 {
850         pgd_t *src_pgd, *dst_pgd;
851         unsigned long next;
852         unsigned long addr = vma->vm_start;
853         unsigned long end = vma->vm_end;
854         int ret;
855 
856         /*
857          * Don't copy ptes where a page fault will fill them correctly.
858          * Fork becomes much lighter when there are big shared or private
859          * readonly mappings. The tradeoff is that copy_page_range is more
860          * efficient than faulting.
861          */
862         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
863                 if (!vma->anon_vma)
864                         return 0;
865         }
866 
867         if (is_vm_hugetlb_page(vma))
868                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
869 
870         if (unlikely(is_pfn_mapping(vma))) {
871                 /*
872                  * We do not free on error cases below as remove_vma
873                  * gets called on error from higher level routine
874                  */
875                 ret = track_pfn_vma_copy(vma);
876                 if (ret)
877                         return ret;
878         }
879 
880         /*
881          * We need to invalidate the secondary MMU mappings only when
882          * there could be a permission downgrade on the ptes of the
883          * parent mm. And a permission downgrade will only happen if
884          * is_cow_mapping() returns true.
885          */
886         if (is_cow_mapping(vma->vm_flags))
887                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
888 
889         ret = 0;
890         dst_pgd = pgd_offset(dst_mm, addr);
891         src_pgd = pgd_offset(src_mm, addr);
892         do {
893                 next = pgd_addr_end(addr, end);
894                 if (pgd_none_or_clear_bad(src_pgd))
895                         continue;
896                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
897                                             vma, addr, next))) {
898                         ret = -ENOMEM;
899                         break;
900                 }
901         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
902 
903         if (is_cow_mapping(vma->vm_flags))
904                 mmu_notifier_invalidate_range_end(src_mm,
905                                                   vma->vm_start, end);
906         return ret;
907 }
908 
909 static unsigned long zap_pte_range(struct mmu_gather *tlb,
910                                 struct vm_area_struct *vma, pmd_t *pmd,
911                                 unsigned long addr, unsigned long end,
912                                 long *zap_work, struct zap_details *details)
913 {
914         struct mm_struct *mm = tlb->mm;
915         pte_t *pte;
916         spinlock_t *ptl;
917         int rss[NR_MM_COUNTERS];
918 
919         init_rss_vec(rss);
920 
921         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
922         arch_enter_lazy_mmu_mode();
923         do {
924                 pte_t ptent = *pte;
925                 if (pte_none(ptent)) {
926                         (*zap_work)--;
927                         continue;
928                 }
929 
930                 (*zap_work) -= PAGE_SIZE;
931 
932                 if (pte_present(ptent)) {
933                         struct page *page;
934 
935                         page = vm_normal_page(vma, addr, ptent);
936                         if (unlikely(details) && page) {
937                                 /*
938                                  * unmap_shared_mapping_pages() wants to
939                                  * invalidate cache without truncating:
940                                  * unmap shared but keep private pages.
941                                  */
942                                 if (details->check_mapping &&
943                                     details->check_mapping != page->mapping)
944                                         continue;
945                                 /*
946                                  * Each page->index must be checked when
947                                  * invalidating or truncating nonlinear.
948                                  */
949                                 if (details->nonlinear_vma &&
950                                     (page->index < details->first_index ||
951                                      page->index > details->last_index))
952                                         continue;
953                         }
954                         ptent = ptep_get_and_clear_full(mm, addr, pte,
955                                                         tlb->fullmm);
956                         tlb_remove_tlb_entry(tlb, pte, addr);
957                         if (unlikely(!page))
958                                 continue;
959                         if (unlikely(details) && details->nonlinear_vma
960                             && linear_page_index(details->nonlinear_vma,
961                                                 addr) != page->index)
962                                 set_pte_at(mm, addr, pte,
963                                            pgoff_to_pte(page->index));
964                         if (PageAnon(page))
965                                 rss[MM_ANONPAGES]--;
966                         else {
967                                 if (pte_dirty(ptent))
968                                         set_page_dirty(page);
969                                 if (pte_young(ptent) &&
970                                     likely(!VM_SequentialReadHint(vma)))
971                                         mark_page_accessed(page);
972                                 rss[MM_FILEPAGES]--;
973                         }
974                         page_remove_rmap(page);
975                         if (unlikely(page_mapcount(page) < 0))
976                                 print_bad_pte(vma, addr, ptent, page);
977                         tlb_remove_page(tlb, page);
978                         continue;
979                 }
980                 /*
981                  * If details->check_mapping, we leave swap entries;
982                  * if details->nonlinear_vma, we leave file entries.
983                  */
984                 if (unlikely(details))
985                         continue;
986                 if (pte_file(ptent)) {
987                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
988                                 print_bad_pte(vma, addr, ptent, NULL);
989                 } else {
990                         swp_entry_t entry = pte_to_swp_entry(ptent);
991 
992                         if (!non_swap_entry(entry))
993                                 rss[MM_SWAPENTS]--;
994                         if (unlikely(!free_swap_and_cache(entry)))
995                                 print_bad_pte(vma, addr, ptent, NULL);
996                 }
997                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
998         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
999 
1000         add_mm_rss_vec(mm, rss);
1001         arch_leave_lazy_mmu_mode();
1002         pte_unmap_unlock(pte - 1, ptl);
1003 
1004         return addr;
1005 }
1006 
1007 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1008                                 struct vm_area_struct *vma, pud_t *pud,
1009                                 unsigned long addr, unsigned long end,
1010                                 long *zap_work, struct zap_details *details)
1011 {
1012         pmd_t *pmd;
1013         unsigned long next;
1014 
1015         pmd = pmd_offset(pud, addr);
1016         do {
1017                 next = pmd_addr_end(addr, end);
1018                 if (pmd_trans_huge(*pmd)) {
1019                         if (next-addr != HPAGE_PMD_SIZE) {
1020                                 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1021                                 split_huge_page_pmd(vma->vm_mm, pmd);
1022                         } else if (zap_huge_pmd(tlb, vma, pmd)) {
1023                                 (*zap_work)--;
1024                                 continue;
1025                         }
1026                         /* fall through */
1027                 }
1028                 if (pmd_none_or_clear_bad(pmd)) {
1029                         (*zap_work)--;
1030                         continue;
1031                 }
1032                 next = zap_pte_range(tlb, vma, pmd, addr, next,
1033                                                 zap_work, details);
1034         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
1035 
1036         return addr;
1037 }
1038 
1039 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1040                                 struct vm_area_struct *vma, pgd_t *pgd,
1041                                 unsigned long addr, unsigned long end,
1042                                 long *zap_work, struct zap_details *details)
1043 {
1044         pud_t *pud;
1045         unsigned long next;
1046 
1047         pud = pud_offset(pgd, addr);
1048         do {
1049                 next = pud_addr_end(addr, end);
1050                 if (pud_none_or_clear_bad(pud)) {
1051                         (*zap_work)--;
1052                         continue;
1053                 }
1054                 next = zap_pmd_range(tlb, vma, pud, addr, next,
1055                                                 zap_work, details);
1056         } while (pud++, addr = next, (addr != end && *zap_work > 0));
1057 
1058         return addr;
1059 }
1060 
1061 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1062                                 struct vm_area_struct *vma,
1063                                 unsigned long addr, unsigned long end,
1064                                 long *zap_work, struct zap_details *details)
1065 {
1066         pgd_t *pgd;
1067         unsigned long next;
1068 
1069         if (details && !details->check_mapping && !details->nonlinear_vma)
1070                 details = NULL;
1071 
1072         BUG_ON(addr >= end);
1073         mem_cgroup_uncharge_start();
1074         tlb_start_vma(tlb, vma);
1075         pgd = pgd_offset(vma->vm_mm, addr);
1076         do {
1077                 next = pgd_addr_end(addr, end);
1078                 if (pgd_none_or_clear_bad(pgd)) {
1079                         (*zap_work)--;
1080                         continue;
1081                 }
1082                 next = zap_pud_range(tlb, vma, pgd, addr, next,
1083                                                 zap_work, details);
1084         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
1085         tlb_end_vma(tlb, vma);
1086         mem_cgroup_uncharge_end();
1087 
1088         return addr;
1089 }
1090 
1091 #ifdef CONFIG_PREEMPT
1092 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1093 #else
1094 /* No preempt: go for improved straight-line efficiency */
1095 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1096 #endif
1097 
1098 /**
1099  * unmap_vmas - unmap a range of memory covered by a list of vma's
1100  * @tlbp: address of the caller's struct mmu_gather
1101  * @vma: the starting vma
1102  * @start_addr: virtual address at which to start unmapping
1103  * @end_addr: virtual address at which to end unmapping
1104  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1105  * @details: details of nonlinear truncation or shared cache invalidation
1106  *
1107  * Returns the end address of the unmapping (restart addr if interrupted).
1108  *
1109  * Unmap all pages in the vma list.
1110  *
1111  * We aim to not hold locks for too long (for scheduling latency reasons).
1112  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
1113  * return the ending mmu_gather to the caller.
1114  *
1115  * Only addresses between `start' and `end' will be unmapped.
1116  *
1117  * The VMA list must be sorted in ascending virtual address order.
1118  *
1119  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1120  * range after unmap_vmas() returns.  So the only responsibility here is to
1121  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1122  * drops the lock and schedules.
1123  */
1124 unsigned long unmap_vmas(struct mmu_gather **tlbp,
1125                 struct vm_area_struct *vma, unsigned long start_addr,
1126                 unsigned long end_addr, unsigned long *nr_accounted,
1127                 struct zap_details *details)
1128 {
1129         long zap_work = ZAP_BLOCK_SIZE;
1130         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
1131         int tlb_start_valid = 0;
1132         unsigned long start = start_addr;
1133         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1134         int fullmm = (*tlbp)->fullmm;
1135         struct mm_struct *mm = vma->vm_mm;
1136 
1137         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1138         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1139                 unsigned long end;
1140 
1141                 start = max(vma->vm_start, start_addr);
1142                 if (start >= vma->vm_end)
1143                         continue;
1144                 end = min(vma->vm_end, end_addr);
1145                 if (end <= vma->vm_start)
1146                         continue;
1147 
1148                 if (vma->vm_flags & VM_ACCOUNT)
1149                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1150 
1151                 if (unlikely(is_pfn_mapping(vma)))
1152                         untrack_pfn_vma(vma, 0, 0);
1153 
1154                 while (start != end) {
1155                         if (!tlb_start_valid) {
1156                                 tlb_start = start;
1157                                 tlb_start_valid = 1;
1158                         }
1159 
1160                         if (unlikely(is_vm_hugetlb_page(vma))) {
1161                                 /*
1162                                  * It is undesirable to test vma->vm_file as it
1163                                  * should be non-null for valid hugetlb area.
1164                                  * However, vm_file will be NULL in the error
1165                                  * cleanup path of do_mmap_pgoff. When
1166                                  * hugetlbfs ->mmap method fails,
1167                                  * do_mmap_pgoff() nullifies vma->vm_file
1168                                  * before calling this function to clean up.
1169                                  * Since no pte has actually been setup, it is
1170                                  * safe to do nothing in this case.
1171                                  */
1172                                 if (vma->vm_file) {
1173                                         unmap_hugepage_range(vma, start, end, NULL);
1174                                         zap_work -= (end - start) /
1175                                         pages_per_huge_page(hstate_vma(vma));
1176                                 }
1177 
1178                                 start = end;
1179                         } else
1180                                 start = unmap_page_range(*tlbp, vma,
1181                                                 start, end, &zap_work, details);
1182 
1183                         if (zap_work > 0) {
1184                                 BUG_ON(start != end);
1185                                 break;
1186                         }
1187 
1188                         tlb_finish_mmu(*tlbp, tlb_start, start);
1189 
1190                         if (need_resched() ||
1191                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1192                                 if (i_mmap_lock) {
1193                                         *tlbp = NULL;
1194                                         goto out;
1195                                 }
1196                                 cond_resched();
1197                         }
1198 
1199                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1200                         tlb_start_valid = 0;
1201                         zap_work = ZAP_BLOCK_SIZE;
1202                 }
1203         }
1204 out:
1205         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1206         return start;   /* which is now the end (or restart) address */
1207 }
1208 
1209 /**
1210  * zap_page_range - remove user pages in a given range
1211  * @vma: vm_area_struct holding the applicable pages
1212  * @address: starting address of pages to zap
1213  * @size: number of bytes to zap
1214  * @details: details of nonlinear truncation or shared cache invalidation
1215  */
1216 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1217                 unsigned long size, struct zap_details *details)
1218 {
1219         struct mm_struct *mm = vma->vm_mm;
1220         struct mmu_gather *tlb;
1221         unsigned long end = address + size;
1222         unsigned long nr_accounted = 0;
1223 
1224         lru_add_drain();
1225         tlb = tlb_gather_mmu(mm, 0);
1226         update_hiwater_rss(mm);
1227         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1228         if (tlb)
1229                 tlb_finish_mmu(tlb, address, end);
1230         return end;
1231 }
1232 
1233 /**
1234  * zap_vma_ptes - remove ptes mapping the vma
1235  * @vma: vm_area_struct holding ptes to be zapped
1236  * @address: starting address of pages to zap
1237  * @size: number of bytes to zap
1238  *
1239  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1240  *
1241  * The entire address range must be fully contained within the vma.
1242  *
1243  * Returns 0 if successful.
1244  */
1245 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1246                 unsigned long size)
1247 {
1248         if (address < vma->vm_start || address + size > vma->vm_end ||
1249                         !(vma->vm_flags & VM_PFNMAP))
1250                 return -1;
1251         zap_page_range(vma, address, size, NULL);
1252         return 0;
1253 }
1254 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1255 
1256 /**
1257  * follow_page - look up a page descriptor from a user-virtual address
1258  * @vma: vm_area_struct mapping @address
1259  * @address: virtual address to look up
1260  * @flags: flags modifying lookup behaviour
1261  *
1262  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1263  *
1264  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1265  * an error pointer if there is a mapping to something not represented
1266  * by a page descriptor (see also vm_normal_page()).
1267  */
1268 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1269                         unsigned int flags)
1270 {
1271         pgd_t *pgd;
1272         pud_t *pud;
1273         pmd_t *pmd;
1274         pte_t *ptep, pte;
1275         spinlock_t *ptl;
1276         struct page *page;
1277         struct mm_struct *mm = vma->vm_mm;
1278 
1279         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1280         if (!IS_ERR(page)) {
1281                 BUG_ON(flags & FOLL_GET);
1282                 goto out;
1283         }
1284 
1285         page = NULL;
1286         pgd = pgd_offset(mm, address);
1287         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1288                 goto no_page_table;
1289 
1290         pud = pud_offset(pgd, address);
1291         if (pud_none(*pud))
1292                 goto no_page_table;
1293         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1294                 BUG_ON(flags & FOLL_GET);
1295                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1296                 goto out;
1297         }
1298         if (unlikely(pud_bad(*pud)))
1299                 goto no_page_table;
1300 
1301         pmd = pmd_offset(pud, address);
1302         if (pmd_none(*pmd))
1303                 goto no_page_table;
1304         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1305                 BUG_ON(flags & FOLL_GET);
1306                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1307                 goto out;
1308         }
1309         if (pmd_trans_huge(*pmd)) {
1310                 if (flags & FOLL_SPLIT) {
1311                         split_huge_page_pmd(mm, pmd);
1312                         goto split_fallthrough;
1313                 }
1314                 spin_lock(&mm->page_table_lock);
1315                 if (likely(pmd_trans_huge(*pmd))) {
1316                         if (unlikely(pmd_trans_splitting(*pmd))) {
1317                                 spin_unlock(&mm->page_table_lock);
1318                                 wait_split_huge_page(vma->anon_vma, pmd);
1319                         } else {
1320                                 page = follow_trans_huge_pmd(mm, address,
1321                                                              pmd, flags);
1322                                 spin_unlock(&mm->page_table_lock);
1323                                 goto out;
1324                         }
1325                 } else
1326                         spin_unlock(&mm->page_table_lock);
1327                 /* fall through */
1328         }
1329 split_fallthrough:
1330         if (unlikely(pmd_bad(*pmd)))
1331                 goto no_page_table;
1332 
1333         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1334 
1335         pte = *ptep;
1336         if (!pte_present(pte))
1337                 goto no_page;
1338         if ((flags & FOLL_WRITE) && !pte_write(pte))
1339                 goto unlock;
1340 
1341         page = vm_normal_page(vma, address, pte);
1342         if (unlikely(!page)) {
1343                 if ((flags & FOLL_DUMP) ||
1344                     !is_zero_pfn(pte_pfn(pte)))
1345                         goto bad_page;
1346                 page = pte_page(pte);
1347         }
1348 
1349         if (flags & FOLL_GET)
1350                 get_page(page);
1351         if (flags & FOLL_TOUCH) {
1352                 if ((flags & FOLL_WRITE) &&
1353                     !pte_dirty(pte) && !PageDirty(page))
1354                         set_page_dirty(page);
1355                 /*
1356                  * pte_mkyoung() would be more correct here, but atomic care
1357                  * is needed to avoid losing the dirty bit: it is easier to use
1358                  * mark_page_accessed().
1359                  */
1360                 mark_page_accessed(page);
1361         }
1362         if (flags & FOLL_MLOCK) {
1363                 /*
1364                  * The preliminary mapping check is mainly to avoid the
1365                  * pointless overhead of lock_page on the ZERO_PAGE
1366                  * which might bounce very badly if there is contention.
1367                  *
1368                  * If the page is already locked, we don't need to
1369                  * handle it now - vmscan will handle it later if and
1370                  * when it attempts to reclaim the page.
1371                  */
1372                 if (page->mapping && trylock_page(page)) {
1373                         lru_add_drain();  /* push cached pages to LRU */
1374                         /*
1375                          * Because we lock page here and migration is
1376                          * blocked by the pte's page reference, we need
1377                          * only check for file-cache page truncation.
1378                          */
1379                         if (page->mapping)
1380                                 mlock_vma_page(page);
1381                         unlock_page(page);
1382                 }
1383         }
1384 unlock:
1385         pte_unmap_unlock(ptep, ptl);
1386 out:
1387         return page;
1388 
1389 bad_page:
1390         pte_unmap_unlock(ptep, ptl);
1391         return ERR_PTR(-EFAULT);
1392 
1393 no_page:
1394         pte_unmap_unlock(ptep, ptl);
1395         if (!pte_none(pte))
1396                 return page;
1397 
1398 no_page_table:
1399         /*
1400          * When core dumping an enormous anonymous area that nobody
1401          * has touched so far, we don't want to allocate unnecessary pages or
1402          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1403          * then get_dump_page() will return NULL to leave a hole in the dump.
1404          * But we can only make this optimization where a hole would surely
1405          * be zero-filled if handle_mm_fault() actually did handle it.
1406          */
1407         if ((flags & FOLL_DUMP) &&
1408             (!vma->vm_ops || !vma->vm_ops->fault))
1409                 return ERR_PTR(-EFAULT);
1410         return page;
1411 }
1412 
1413 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1414                      unsigned long start, int nr_pages, unsigned int gup_flags,
1415                      struct page **pages, struct vm_area_struct **vmas,
1416                      int *nonblocking)
1417 {
1418         int i;
1419         unsigned long vm_flags;
1420 
1421         if (nr_pages <= 0)
1422                 return 0;
1423 
1424         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1425 
1426         /* 
1427          * Require read or write permissions.
1428          * If FOLL_FORCE is set, we only require the "MAY" flags.
1429          */
1430         vm_flags  = (gup_flags & FOLL_WRITE) ?
1431                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1432         vm_flags &= (gup_flags & FOLL_FORCE) ?
1433                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1434         i = 0;
1435 
1436         do {
1437                 struct vm_area_struct *vma;
1438 
1439                 vma = find_extend_vma(mm, start);
1440                 if (!vma && in_gate_area(tsk, start)) {
1441                         unsigned long pg = start & PAGE_MASK;
1442                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1443                         pgd_t *pgd;
1444                         pud_t *pud;
1445                         pmd_t *pmd;
1446                         pte_t *pte;
1447 
1448                         /* user gate pages are read-only */
1449                         if (gup_flags & FOLL_WRITE)
1450                                 return i ? : -EFAULT;
1451                         if (pg > TASK_SIZE)
1452                                 pgd = pgd_offset_k(pg);
1453                         else
1454                                 pgd = pgd_offset_gate(mm, pg);
1455                         BUG_ON(pgd_none(*pgd));
1456                         pud = pud_offset(pgd, pg);
1457                         BUG_ON(pud_none(*pud));
1458                         pmd = pmd_offset(pud, pg);
1459                         if (pmd_none(*pmd))
1460                                 return i ? : -EFAULT;
1461                         VM_BUG_ON(pmd_trans_huge(*pmd));
1462                         pte = pte_offset_map(pmd, pg);
1463                         if (pte_none(*pte)) {
1464                                 pte_unmap(pte);
1465                                 return i ? : -EFAULT;
1466                         }
1467                         if (pages) {
1468                                 struct page *page;
1469 
1470                                 page = vm_normal_page(gate_vma, start, *pte);
1471                                 if (!page) {
1472                                         if (!(gup_flags & FOLL_DUMP) &&
1473                                              is_zero_pfn(pte_pfn(*pte)))
1474                                                 page = pte_page(*pte);
1475                                         else {
1476                                                 pte_unmap(pte);
1477                                                 return i ? : -EFAULT;
1478                                         }
1479                                 }
1480                                 pages[i] = page;
1481                                 get_page(page);
1482                         }
1483                         pte_unmap(pte);
1484                         if (vmas)
1485                                 vmas[i] = gate_vma;
1486                         i++;
1487                         start += PAGE_SIZE;
1488                         nr_pages--;
1489                         continue;
1490                 }
1491 
1492                 if (!vma ||
1493                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1494                     !(vm_flags & vma->vm_flags))
1495                         return i ? : -EFAULT;
1496 
1497                 if (is_vm_hugetlb_page(vma)) {
1498                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1499                                         &start, &nr_pages, i, gup_flags);
1500                         continue;
1501                 }
1502 
1503                 do {
1504                         struct page *page;
1505                         unsigned int foll_flags = gup_flags;
1506 
1507                         /*
1508                          * If we have a pending SIGKILL, don't keep faulting
1509                          * pages and potentially allocating memory.
1510                          */
1511                         if (unlikely(fatal_signal_pending(current)))
1512                                 return i ? i : -ERESTARTSYS;
1513 
1514                         cond_resched();
1515                         while (!(page = follow_page(vma, start, foll_flags))) {
1516                                 int ret;
1517                                 unsigned int fault_flags = 0;
1518 
1519                                 if (foll_flags & FOLL_WRITE)
1520                                         fault_flags |= FAULT_FLAG_WRITE;
1521                                 if (nonblocking)
1522                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1523 
1524                                 ret = handle_mm_fault(mm, vma, start,
1525                                                         fault_flags);
1526 
1527                                 if (ret & VM_FAULT_ERROR) {
1528                                         if (ret & VM_FAULT_OOM)
1529                                                 return i ? i : -ENOMEM;
1530                                         if (ret &
1531                                             (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE|
1532                                              VM_FAULT_SIGBUS))
1533                                                 return i ? i : -EFAULT;
1534                                         BUG();
1535                                 }
1536                                 if (ret & VM_FAULT_MAJOR)
1537                                         tsk->maj_flt++;
1538                                 else
1539                                         tsk->min_flt++;
1540 
1541                                 if (ret & VM_FAULT_RETRY) {
1542                                         *nonblocking = 0;
1543                                         return i;
1544                                 }
1545 
1546                                 /*
1547                                  * The VM_FAULT_WRITE bit tells us that
1548                                  * do_wp_page has broken COW when necessary,
1549                                  * even if maybe_mkwrite decided not to set
1550                                  * pte_write. We can thus safely do subsequent
1551                                  * page lookups as if they were reads. But only
1552                                  * do so when looping for pte_write is futile:
1553                                  * in some cases userspace may also be wanting
1554                                  * to write to the gotten user page, which a
1555                                  * read fault here might prevent (a readonly
1556                                  * page might get reCOWed by userspace write).
1557                                  */
1558                                 if ((ret & VM_FAULT_WRITE) &&
1559                                     !(vma->vm_flags & VM_WRITE))
1560                                         foll_flags &= ~FOLL_WRITE;
1561 
1562                                 cond_resched();
1563                         }
1564                         if (IS_ERR(page))
1565                                 return i ? i : PTR_ERR(page);
1566                         if (pages) {
1567                                 pages[i] = page;
1568 
1569                                 flush_anon_page(vma, page, start);
1570                                 flush_dcache_page(page);
1571                         }
1572                         if (vmas)
1573                                 vmas[i] = vma;
1574                         i++;
1575                         start += PAGE_SIZE;
1576                         nr_pages--;
1577                 } while (nr_pages && start < vma->vm_end);
1578         } while (nr_pages);
1579         return i;
1580 }
1581 
1582 /**
1583  * get_user_pages() - pin user pages in memory
1584  * @tsk:        task_struct of target task
1585  * @mm:         mm_struct of target mm
1586  * @start:      starting user address
1587  * @nr_pages:   number of pages from start to pin
1588  * @write:      whether pages will be written to by the caller
1589  * @force:      whether to force write access even if user mapping is
1590  *              readonly. This will result in the page being COWed even
1591  *              in MAP_SHARED mappings. You do not want this.
1592  * @pages:      array that receives pointers to the pages pinned.
1593  *              Should be at least nr_pages long. Or NULL, if caller
1594  *              only intends to ensure the pages are faulted in.
1595  * @vmas:       array of pointers to vmas corresponding to each page.
1596  *              Or NULL if the caller does not require them.
1597  *
1598  * Returns number of pages pinned. This may be fewer than the number
1599  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1600  * were pinned, returns -errno. Each page returned must be released
1601  * with a put_page() call when it is finished with. vmas will only
1602  * remain valid while mmap_sem is held.
1603  *
1604  * Must be called with mmap_sem held for read or write.
1605  *
1606  * get_user_pages walks a process's page tables and takes a reference to
1607  * each struct page that each user address corresponds to at a given
1608  * instant. That is, it takes the page that would be accessed if a user
1609  * thread accesses the given user virtual address at that instant.
1610  *
1611  * This does not guarantee that the page exists in the user mappings when
1612  * get_user_pages returns, and there may even be a completely different
1613  * page there in some cases (eg. if mmapped pagecache has been invalidated
1614  * and subsequently re faulted). However it does guarantee that the page
1615  * won't be freed completely. And mostly callers simply care that the page
1616  * contains data that was valid *at some point in time*. Typically, an IO
1617  * or similar operation cannot guarantee anything stronger anyway because
1618  * locks can't be held over the syscall boundary.
1619  *
1620  * If write=0, the page must not be written to. If the page is written to,
1621  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1622  * after the page is finished with, and before put_page is called.
1623  *
1624  * get_user_pages is typically used for fewer-copy IO operations, to get a
1625  * handle on the memory by some means other than accesses via the user virtual
1626  * addresses. The pages may be submitted for DMA to devices or accessed via
1627  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1628  * use the correct cache flushing APIs.
1629  *
1630  * See also get_user_pages_fast, for performance critical applications.
1631  */
1632 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1633                 unsigned long start, int nr_pages, int write, int force,
1634                 struct page **pages, struct vm_area_struct **vmas)
1635 {
1636         int flags = FOLL_TOUCH;
1637 
1638         if (pages)
1639                 flags |= FOLL_GET;
1640         if (write)
1641                 flags |= FOLL_WRITE;
1642         if (force)
1643                 flags |= FOLL_FORCE;
1644 
1645         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1646                                 NULL);
1647 }
1648 EXPORT_SYMBOL(get_user_pages);
1649 
1650 /**
1651  * get_dump_page() - pin user page in memory while writing it to core dump
1652  * @addr: user address
1653  *
1654  * Returns struct page pointer of user page pinned for dump,
1655  * to be freed afterwards by page_cache_release() or put_page().
1656  *
1657  * Returns NULL on any kind of failure - a hole must then be inserted into
1658  * the corefile, to preserve alignment with its headers; and also returns
1659  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1660  * allowing a hole to be left in the corefile to save diskspace.
1661  *
1662  * Called without mmap_sem, but after all other threads have been killed.
1663  */
1664 #ifdef CONFIG_ELF_CORE
1665 struct page *get_dump_page(unsigned long addr)
1666 {
1667         struct vm_area_struct *vma;
1668         struct page *page;
1669 
1670         if (__get_user_pages(current, current->mm, addr, 1,
1671                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1672                              NULL) < 1)
1673                 return NULL;
1674         flush_cache_page(vma, addr, page_to_pfn(page));
1675         return page;
1676 }
1677 #endif /* CONFIG_ELF_CORE */
1678 
1679 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1680                         spinlock_t **ptl)
1681 {
1682         pgd_t * pgd = pgd_offset(mm, addr);
1683         pud_t * pud = pud_alloc(mm, pgd, addr);
1684         if (pud) {
1685                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1686                 if (pmd) {
1687                         VM_BUG_ON(pmd_trans_huge(*pmd));
1688                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1689                 }
1690         }
1691         return NULL;
1692 }
1693 
1694 /*
1695  * This is the old fallback for page remapping.
1696  *
1697  * For historical reasons, it only allows reserved pages. Only
1698  * old drivers should use this, and they needed to mark their
1699  * pages reserved for the old functions anyway.
1700  */
1701 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1702                         struct page *page, pgprot_t prot)
1703 {
1704         struct mm_struct *mm = vma->vm_mm;
1705         int retval;
1706         pte_t *pte;
1707         spinlock_t *ptl;
1708 
1709         retval = -EINVAL;
1710         if (PageAnon(page))
1711                 goto out;
1712         retval = -ENOMEM;
1713         flush_dcache_page(page);
1714         pte = get_locked_pte(mm, addr, &ptl);
1715         if (!pte)
1716                 goto out;
1717         retval = -EBUSY;
1718         if (!pte_none(*pte))
1719                 goto out_unlock;
1720 
1721         /* Ok, finally just insert the thing.. */
1722         get_page(page);
1723         inc_mm_counter_fast(mm, MM_FILEPAGES);
1724         page_add_file_rmap(page);
1725         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1726 
1727         retval = 0;
1728         pte_unmap_unlock(pte, ptl);
1729         return retval;
1730 out_unlock:
1731         pte_unmap_unlock(pte, ptl);
1732 out:
1733         return retval;
1734 }
1735 
1736 /**
1737  * vm_insert_page - insert single page into user vma
1738  * @vma: user vma to map to
1739  * @addr: target user address of this page
1740  * @page: source kernel page
1741  *
1742  * This allows drivers to insert individual pages they've allocated
1743  * into a user vma.
1744  *
1745  * The page has to be a nice clean _individual_ kernel allocation.
1746  * If you allocate a compound page, you need to have marked it as
1747  * such (__GFP_COMP), or manually just split the page up yourself
1748  * (see split_page()).
1749  *
1750  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1751  * took an arbitrary page protection parameter. This doesn't allow
1752  * that. Your vma protection will have to be set up correctly, which
1753  * means that if you want a shared writable mapping, you'd better
1754  * ask for a shared writable mapping!
1755  *
1756  * The page does not need to be reserved.
1757  */
1758 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1759                         struct page *page)
1760 {
1761         if (addr < vma->vm_start || addr >= vma->vm_end)
1762                 return -EFAULT;
1763         if (!page_count(page))
1764                 return -EINVAL;
1765         vma->vm_flags |= VM_INSERTPAGE;
1766         return insert_page(vma, addr, page, vma->vm_page_prot);
1767 }
1768 EXPORT_SYMBOL(vm_insert_page);
1769 
1770 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1771                         unsigned long pfn, pgprot_t prot)
1772 {
1773         struct mm_struct *mm = vma->vm_mm;
1774         int retval;
1775         pte_t *pte, entry;
1776         spinlock_t *ptl;
1777 
1778         retval = -ENOMEM;
1779         pte = get_locked_pte(mm, addr, &ptl);
1780         if (!pte)
1781                 goto out;
1782         retval = -EBUSY;
1783         if (!pte_none(*pte))
1784                 goto out_unlock;
1785 
1786         /* Ok, finally just insert the thing.. */
1787         entry = pte_mkspecial(pfn_pte(pfn, prot));
1788         set_pte_at(mm, addr, pte, entry);
1789         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1790 
1791         retval = 0;
1792 out_unlock:
1793         pte_unmap_unlock(pte, ptl);
1794 out:
1795         return retval;
1796 }
1797 
1798 /**
1799  * vm_insert_pfn - insert single pfn into user vma
1800  * @vma: user vma to map to
1801  * @addr: target user address of this page
1802  * @pfn: source kernel pfn
1803  *
1804  * Similar to vm_inert_page, this allows drivers to insert individual pages
1805  * they've allocated into a user vma. Same comments apply.
1806  *
1807  * This function should only be called from a vm_ops->fault handler, and
1808  * in that case the handler should return NULL.
1809  *
1810  * vma cannot be a COW mapping.
1811  *
1812  * As this is called only for pages that do not currently exist, we
1813  * do not need to flush old virtual caches or the TLB.
1814  */
1815 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1816                         unsigned long pfn)
1817 {
1818         int ret;
1819         pgprot_t pgprot = vma->vm_page_prot;
1820         /*
1821          * Technically, architectures with pte_special can avoid all these
1822          * restrictions (same for remap_pfn_range).  However we would like
1823          * consistency in testing and feature parity among all, so we should
1824          * try to keep these invariants in place for everybody.
1825          */
1826         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1827         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1828                                                 (VM_PFNMAP|VM_MIXEDMAP));
1829         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1830         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1831 
1832         if (addr < vma->vm_start || addr >= vma->vm_end)
1833                 return -EFAULT;
1834         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1835                 return -EINVAL;
1836 
1837         ret = insert_pfn(vma, addr, pfn, pgprot);
1838 
1839         if (ret)
1840                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1841 
1842         return ret;
1843 }
1844 EXPORT_SYMBOL(vm_insert_pfn);
1845 
1846 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1847                         unsigned long pfn)
1848 {
1849         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1850 
1851         if (addr < vma->vm_start || addr >= vma->vm_end)
1852                 return -EFAULT;
1853 
1854         /*
1855          * If we don't have pte special, then we have to use the pfn_valid()
1856          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1857          * refcount the page if pfn_valid is true (hence insert_page rather
1858          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1859          * without pte special, it would there be refcounted as a normal page.
1860          */
1861         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1862                 struct page *page;
1863 
1864                 page = pfn_to_page(pfn);
1865                 return insert_page(vma, addr, page, vma->vm_page_prot);
1866         }
1867         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1868 }
1869 EXPORT_SYMBOL(vm_insert_mixed);
1870 
1871 /*
1872  * maps a range of physical memory into the requested pages. the old
1873  * mappings are removed. any references to nonexistent pages results
1874  * in null mappings (currently treated as "copy-on-access")
1875  */
1876 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1877                         unsigned long addr, unsigned long end,
1878                         unsigned long pfn, pgprot_t prot)
1879 {
1880         pte_t *pte;
1881         spinlock_t *ptl;
1882 
1883         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1884         if (!pte)
1885                 return -ENOMEM;
1886         arch_enter_lazy_mmu_mode();
1887         do {
1888                 BUG_ON(!pte_none(*pte));
1889                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1890                 pfn++;
1891         } while (pte++, addr += PAGE_SIZE, addr != end);
1892         arch_leave_lazy_mmu_mode();
1893         pte_unmap_unlock(pte - 1, ptl);
1894         return 0;
1895 }
1896 
1897 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1898                         unsigned long addr, unsigned long end,
1899                         unsigned long pfn, pgprot_t prot)
1900 {
1901         pmd_t *pmd;
1902         unsigned long next;
1903 
1904         pfn -= addr >> PAGE_SHIFT;
1905         pmd = pmd_alloc(mm, pud, addr);
1906         if (!pmd)
1907                 return -ENOMEM;
1908         VM_BUG_ON(pmd_trans_huge(*pmd));
1909         do {
1910                 next = pmd_addr_end(addr, end);
1911                 if (remap_pte_range(mm, pmd, addr, next,
1912                                 pfn + (addr >> PAGE_SHIFT), prot))
1913                         return -ENOMEM;
1914         } while (pmd++, addr = next, addr != end);
1915         return 0;
1916 }
1917 
1918 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1919                         unsigned long addr, unsigned long end,
1920                         unsigned long pfn, pgprot_t prot)
1921 {
1922         pud_t *pud;
1923         unsigned long next;
1924 
1925         pfn -= addr >> PAGE_SHIFT;
1926         pud = pud_alloc(mm, pgd, addr);
1927         if (!pud)
1928                 return -ENOMEM;
1929         do {
1930                 next = pud_addr_end(addr, end);
1931                 if (remap_pmd_range(mm, pud, addr, next,
1932                                 pfn + (addr >> PAGE_SHIFT), prot))
1933                         return -ENOMEM;
1934         } while (pud++, addr = next, addr != end);
1935         return 0;
1936 }
1937 
1938 /**
1939  * remap_pfn_range - remap kernel memory to userspace
1940  * @vma: user vma to map to
1941  * @addr: target user address to start at
1942  * @pfn: physical address of kernel memory
1943  * @size: size of map area
1944  * @prot: page protection flags for this mapping
1945  *
1946  *  Note: this is only safe if the mm semaphore is held when called.
1947  */
1948 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1949                     unsigned long pfn, unsigned long size, pgprot_t prot)
1950 {
1951         pgd_t *pgd;
1952         unsigned long next;
1953         unsigned long end = addr + PAGE_ALIGN(size);
1954         struct mm_struct *mm = vma->vm_mm;
1955         int err;
1956 
1957         /*
1958          * Physically remapped pages are special. Tell the
1959          * rest of the world about it:
1960          *   VM_IO tells people not to look at these pages
1961          *      (accesses can have side effects).
1962          *   VM_RESERVED is specified all over the place, because
1963          *      in 2.4 it kept swapout's vma scan off this vma; but
1964          *      in 2.6 the LRU scan won't even find its pages, so this
1965          *      flag means no more than count its pages in reserved_vm,
1966          *      and omit it from core dump, even when VM_IO turned off.
1967          *   VM_PFNMAP tells the core MM that the base pages are just
1968          *      raw PFN mappings, and do not have a "struct page" associated
1969          *      with them.
1970          *
1971          * There's a horrible special case to handle copy-on-write
1972          * behaviour that some programs depend on. We mark the "original"
1973          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1974          */
1975         if (addr == vma->vm_start && end == vma->vm_end) {
1976                 vma->vm_pgoff = pfn;
1977                 vma->vm_flags |= VM_PFN_AT_MMAP;
1978         } else if (is_cow_mapping(vma->vm_flags))
1979                 return -EINVAL;
1980 
1981         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1982 
1983         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1984         if (err) {
1985                 /*
1986                  * To indicate that track_pfn related cleanup is not
1987                  * needed from higher level routine calling unmap_vmas
1988                  */
1989                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1990                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1991                 return -EINVAL;
1992         }
1993 
1994         BUG_ON(addr >= end);
1995         pfn -= addr >> PAGE_SHIFT;
1996         pgd = pgd_offset(mm, addr);
1997         flush_cache_range(vma, addr, end);
1998         do {
1999                 next = pgd_addr_end(addr, end);
2000                 err = remap_pud_range(mm, pgd, addr, next,
2001                                 pfn + (addr >> PAGE_SHIFT), prot);
2002                 if (err)
2003                         break;
2004         } while (pgd++, addr = next, addr != end);
2005 
2006         if (err)
2007                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2008 
2009         return err;
2010 }
2011 EXPORT_SYMBOL(remap_pfn_range);
2012 
2013 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2014                                      unsigned long addr, unsigned long end,
2015                                      pte_fn_t fn, void *data)
2016 {
2017         pte_t *pte;
2018         int err;
2019         pgtable_t token;
2020         spinlock_t *uninitialized_var(ptl);
2021 
2022         pte = (mm == &init_mm) ?
2023                 pte_alloc_kernel(pmd, addr) :
2024                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2025         if (!pte)
2026                 return -ENOMEM;
2027 
2028         BUG_ON(pmd_huge(*pmd));
2029 
2030         arch_enter_lazy_mmu_mode();
2031 
2032         token = pmd_pgtable(*pmd);
2033 
2034         do {
2035                 err = fn(pte++, token, addr, data);
2036                 if (err)
2037                         break;
2038         } while (addr += PAGE_SIZE, addr != end);
2039 
2040         arch_leave_lazy_mmu_mode();
2041 
2042         if (mm != &init_mm)
2043                 pte_unmap_unlock(pte-1, ptl);
2044         return err;
2045 }
2046 
2047 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2048                                      unsigned long addr, unsigned long end,
2049                                      pte_fn_t fn, void *data)
2050 {
2051         pmd_t *pmd;
2052         unsigned long next;
2053         int err;
2054 
2055         BUG_ON(pud_huge(*pud));
2056 
2057         pmd = pmd_alloc(mm, pud, addr);
2058         if (!pmd)
2059                 return -ENOMEM;
2060         do {
2061                 next = pmd_addr_end(addr, end);
2062                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2063                 if (err)
2064                         break;
2065         } while (pmd++, addr = next, addr != end);
2066         return err;
2067 }
2068 
2069 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2070                                      unsigned long addr, unsigned long end,
2071                                      pte_fn_t fn, void *data)
2072 {
2073         pud_t *pud;
2074         unsigned long next;
2075         int err;
2076 
2077         pud = pud_alloc(mm, pgd, addr);
2078         if (!pud)
2079                 return -ENOMEM;
2080         do {
2081                 next = pud_addr_end(addr, end);
2082                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2083                 if (err)
2084                         break;
2085         } while (pud++, addr = next, addr != end);
2086         return err;
2087 }
2088 
2089 /*
2090  * Scan a region of virtual memory, filling in page tables as necessary
2091  * and calling a provided function on each leaf page table.
2092  */
2093 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2094                         unsigned long size, pte_fn_t fn, void *data)
2095 {
2096         pgd_t *pgd;
2097         unsigned long next;
2098         unsigned long end = addr + size;
2099         int err;
2100 
2101         BUG_ON(addr >= end);
2102         pgd = pgd_offset(mm, addr);
2103         do {
2104                 next = pgd_addr_end(addr, end);
2105                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2106                 if (err)
2107                         break;
2108         } while (pgd++, addr = next, addr != end);
2109 
2110         return err;
2111 }
2112 EXPORT_SYMBOL_GPL(apply_to_page_range);
2113 
2114 /*
2115  * handle_pte_fault chooses page fault handler according to an entry
2116  * which was read non-atomically.  Before making any commitment, on
2117  * those architectures or configurations (e.g. i386 with PAE) which
2118  * might give a mix of unmatched parts, do_swap_page and do_file_page
2119  * must check under lock before unmapping the pte and proceeding
2120  * (but do_wp_page is only called after already making such a check;
2121  * and do_anonymous_page and do_no_page can safely check later on).
2122  */
2123 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2124                                 pte_t *page_table, pte_t orig_pte)
2125 {
2126         int same = 1;
2127 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2128         if (sizeof(pte_t) > sizeof(unsigned long)) {
2129                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2130                 spin_lock(ptl);
2131                 same = pte_same(*page_table, orig_pte);
2132                 spin_unlock(ptl);
2133         }
2134 #endif
2135         pte_unmap(page_table);
2136         return same;
2137 }
2138 
2139 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2140 {
2141         /*
2142          * If the source page was a PFN mapping, we don't have
2143          * a "struct page" for it. We do a best-effort copy by
2144          * just copying from the original user address. If that
2145          * fails, we just zero-fill it. Live with it.
2146          */
2147         if (unlikely(!src)) {
2148                 void *kaddr = kmap_atomic(dst, KM_USER0);
2149                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2150 
2151                 /*
2152                  * This really shouldn't fail, because the page is there
2153                  * in the page tables. But it might just be unreadable,
2154                  * in which case we just give up and fill the result with
2155                  * zeroes.
2156                  */
2157                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2158                         clear_page(kaddr);
2159                 kunmap_atomic(kaddr, KM_USER0);
2160                 flush_dcache_page(dst);
2161         } else
2162                 copy_user_highpage(dst, src, va, vma);
2163 }
2164 
2165 /*
2166  * This routine handles present pages, when users try to write
2167  * to a shared page. It is done by copying the page to a new address
2168  * and decrementing the shared-page counter for the old page.
2169  *
2170  * Note that this routine assumes that the protection checks have been
2171  * done by the caller (the low-level page fault routine in most cases).
2172  * Thus we can safely just mark it writable once we've done any necessary
2173  * COW.
2174  *
2175  * We also mark the page dirty at this point even though the page will
2176  * change only once the write actually happens. This avoids a few races,
2177  * and potentially makes it more efficient.
2178  *
2179  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2180  * but allow concurrent faults), with pte both mapped and locked.
2181  * We return with mmap_sem still held, but pte unmapped and unlocked.
2182  */
2183 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2184                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2185                 spinlock_t *ptl, pte_t orig_pte)
2186         __releases(ptl)
2187 {
2188         struct page *old_page, *new_page;
2189         pte_t entry;
2190         int ret = 0;
2191         int page_mkwrite = 0;
2192         struct page *dirty_page = NULL;
2193 
2194         old_page = vm_normal_page(vma, address, orig_pte);
2195         if (!old_page) {
2196                 /*
2197                  * VM_MIXEDMAP !pfn_valid() case
2198                  *
2199                  * We should not cow pages in a shared writeable mapping.
2200                  * Just mark the pages writable as we can't do any dirty
2201                  * accounting on raw pfn maps.
2202                  */
2203                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2204                                      (VM_WRITE|VM_SHARED))
2205                         goto reuse;
2206                 goto gotten;
2207         }
2208 
2209         /*
2210          * Take out anonymous pages first, anonymous shared vmas are
2211          * not dirty accountable.
2212          */
2213         if (PageAnon(old_page) && !PageKsm(old_page)) {
2214                 if (!trylock_page(old_page)) {
2215                         page_cache_get(old_page);
2216                         pte_unmap_unlock(page_table, ptl);
2217                         lock_page(old_page);
2218                         page_table = pte_offset_map_lock(mm, pmd, address,
2219                                                          &ptl);
2220                         if (!pte_same(*page_table, orig_pte)) {
2221                                 unlock_page(old_page);
2222                                 goto unlock;
2223                         }
2224                         page_cache_release(old_page);
2225                 }
2226                 if (reuse_swap_page(old_page)) {
2227                         /*
2228                          * The page is all ours.  Move it to our anon_vma so
2229                          * the rmap code will not search our parent or siblings.
2230                          * Protected against the rmap code by the page lock.
2231                          */
2232                         page_move_anon_rmap(old_page, vma, address);
2233                         unlock_page(old_page);
2234                         goto reuse;
2235                 }
2236                 unlock_page(old_page);
2237         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2238                                         (VM_WRITE|VM_SHARED))) {
2239                 /*
2240                  * Only catch write-faults on shared writable pages,
2241                  * read-only shared pages can get COWed by
2242                  * get_user_pages(.write=1, .force=1).
2243                  */
2244                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2245                         struct vm_fault vmf;
2246                         int tmp;
2247 
2248                         vmf.virtual_address = (void __user *)(address &
2249                                                                 PAGE_MASK);
2250                         vmf.pgoff = old_page->index;
2251                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2252                         vmf.page = old_page;
2253 
2254                         /*
2255                          * Notify the address space that the page is about to
2256                          * become writable so that it can prohibit this or wait
2257                          * for the page to get into an appropriate state.
2258                          *
2259                          * We do this without the lock held, so that it can
2260                          * sleep if it needs to.
2261                          */
2262                         page_cache_get(old_page);
2263                         pte_unmap_unlock(page_table, ptl);
2264 
2265                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2266                         if (unlikely(tmp &
2267                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2268                                 ret = tmp;
2269                                 goto unwritable_page;
2270                         }
2271                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2272                                 lock_page(old_page);
2273                                 if (!old_page->mapping) {
2274                                         ret = 0; /* retry the fault */
2275                                         unlock_page(old_page);
2276                                         goto unwritable_page;
2277                                 }
2278                         } else
2279                                 VM_BUG_ON(!PageLocked(old_page));
2280 
2281                         /*
2282                          * Since we dropped the lock we need to revalidate
2283                          * the PTE as someone else may have changed it.  If
2284                          * they did, we just return, as we can count on the
2285                          * MMU to tell us if they didn't also make it writable.
2286                          */
2287                         page_table = pte_offset_map_lock(mm, pmd, address,
2288                                                          &ptl);
2289                         if (!pte_same(*page_table, orig_pte)) {
2290                                 unlock_page(old_page);
2291                                 goto unlock;
2292                         }
2293 
2294                         page_mkwrite = 1;
2295                 }
2296                 dirty_page = old_page;
2297                 get_page(dirty_page);
2298 
2299 reuse:
2300                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2301                 entry = pte_mkyoung(orig_pte);
2302                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2303                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2304                         update_mmu_cache(vma, address, page_table);
2305                 pte_unmap_unlock(page_table, ptl);
2306                 ret |= VM_FAULT_WRITE;
2307 
2308                 if (!dirty_page)
2309                         return ret;
2310 
2311                 /*
2312                  * Yes, Virginia, this is actually required to prevent a race
2313                  * with clear_page_dirty_for_io() from clearing the page dirty
2314                  * bit after it clear all dirty ptes, but before a racing
2315                  * do_wp_page installs a dirty pte.
2316                  *
2317                  * do_no_page is protected similarly.
2318                  */
2319                 if (!page_mkwrite) {
2320                         wait_on_page_locked(dirty_page);
2321                         set_page_dirty_balance(dirty_page, page_mkwrite);
2322                 }
2323                 put_page(dirty_page);
2324                 if (page_mkwrite) {
2325                         struct address_space *mapping = dirty_page->mapping;
2326 
2327                         set_page_dirty(dirty_page);
2328                         unlock_page(dirty_page);
2329                         page_cache_release(dirty_page);
2330                         if (mapping)    {
2331                                 /*
2332                                  * Some device drivers do not set page.mapping
2333                                  * but still dirty their pages
2334                                  */
2335                                 balance_dirty_pages_ratelimited(mapping);
2336                         }
2337                 }
2338 
2339                 /* file_update_time outside page_lock */
2340                 if (vma->vm_file)
2341                         file_update_time(vma->vm_file);
2342 
2343                 return ret;
2344         }
2345 
2346         /*
2347          * Ok, we need to copy. Oh, well..
2348          */
2349         page_cache_get(old_page);
2350 gotten:
2351         pte_unmap_unlock(page_table, ptl);
2352 
2353         if (unlikely(anon_vma_prepare(vma)))
2354                 goto oom;
2355 
2356         if (is_zero_pfn(pte_pfn(orig_pte))) {
2357                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2358                 if (!new_page)
2359                         goto oom;
2360         } else {
2361                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2362                 if (!new_page)
2363                         goto oom;
2364                 cow_user_page(new_page, old_page, address, vma);
2365         }
2366         __SetPageUptodate(new_page);
2367 
2368         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2369                 goto oom_free_new;
2370 
2371         /*
2372          * Re-check the pte - we dropped the lock
2373          */
2374         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2375         if (likely(pte_same(*page_table, orig_pte))) {
2376                 if (old_page) {
2377                         if (!PageAnon(old_page)) {
2378                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2379                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2380                         }
2381                 } else
2382                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2383                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2384                 entry = mk_pte(new_page, vma->vm_page_prot);
2385                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2386                 /*
2387                  * Clear the pte entry and flush it first, before updating the
2388                  * pte with the new entry. This will avoid a race condition
2389                  * seen in the presence of one thread doing SMC and another
2390                  * thread doing COW.
2391                  */
2392                 ptep_clear_flush(vma, address, page_table);
2393                 page_add_new_anon_rmap(new_page, vma, address);
2394                 /*
2395                  * We call the notify macro here because, when using secondary
2396                  * mmu page tables (such as kvm shadow page tables), we want the
2397                  * new page to be mapped directly into the secondary page table.
2398                  */
2399                 set_pte_at_notify(mm, address, page_table, entry);
2400                 update_mmu_cache(vma, address, page_table);
2401                 if (old_page) {
2402                         /*
2403                          * Only after switching the pte to the new page may
2404                          * we remove the mapcount here. Otherwise another
2405                          * process may come and find the rmap count decremented
2406                          * before the pte is switched to the new page, and
2407                          * "reuse" the old page writing into it while our pte
2408                          * here still points into it and can be read by other
2409                          * threads.
2410                          *
2411                          * The critical issue is to order this
2412                          * page_remove_rmap with the ptp_clear_flush above.
2413                          * Those stores are ordered by (if nothing else,)
2414                          * the barrier present in the atomic_add_negative
2415                          * in page_remove_rmap.
2416                          *
2417                          * Then the TLB flush in ptep_clear_flush ensures that
2418                          * no process can access the old page before the
2419                          * decremented mapcount is visible. And the old page
2420                          * cannot be reused until after the decremented
2421                          * mapcount is visible. So transitively, TLBs to
2422                          * old page will be flushed before it can be reused.
2423                          */
2424                         page_remove_rmap(old_page);
2425                 }
2426 
2427                 /* Free the old page.. */
2428                 new_page = old_page;
2429                 ret |= VM_FAULT_WRITE;
2430         } else
2431                 mem_cgroup_uncharge_page(new_page);
2432 
2433         if (new_page)
2434                 page_cache_release(new_page);
2435 unlock:
2436         pte_unmap_unlock(page_table, ptl);
2437         if (old_page) {
2438                 /*
2439                  * Don't let another task, with possibly unlocked vma,
2440                  * keep the mlocked page.
2441                  */
2442                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2443                         lock_page(old_page);    /* LRU manipulation */
2444                         munlock_vma_page(old_page);
2445                         unlock_page(old_page);
2446                 }
2447                 page_cache_release(old_page);
2448         }
2449         return ret;
2450 oom_free_new:
2451         page_cache_release(new_page);
2452 oom:
2453         if (old_page) {
2454                 if (page_mkwrite) {
2455                         unlock_page(old_page);
2456                         page_cache_release(old_page);
2457                 }
2458                 page_cache_release(old_page);
2459         }
2460         return VM_FAULT_OOM;
2461 
2462 unwritable_page:
2463         page_cache_release(old_page);
2464         return ret;
2465 }
2466 
2467 /*
2468  * Helper functions for unmap_mapping_range().
2469  *
2470  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2471  *
2472  * We have to restart searching the prio_tree whenever we drop the lock,
2473  * since the iterator is only valid while the lock is held, and anyway
2474  * a later vma might be split and reinserted earlier while lock dropped.
2475  *
2476  * The list of nonlinear vmas could be handled more efficiently, using
2477  * a placeholder, but handle it in the same way until a need is shown.
2478  * It is important to search the prio_tree before nonlinear list: a vma
2479  * may become nonlinear and be shifted from prio_tree to nonlinear list
2480  * while the lock is dropped; but never shifted from list to prio_tree.
2481  *
2482  * In order to make forward progress despite restarting the search,
2483  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2484  * quickly skip it next time around.  Since the prio_tree search only
2485  * shows us those vmas affected by unmapping the range in question, we
2486  * can't efficiently keep all vmas in step with mapping->truncate_count:
2487  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2488  * mapping->truncate_count and vma->vm_truncate_count are protected by
2489  * i_mmap_lock.
2490  *
2491  * In order to make forward progress despite repeatedly restarting some
2492  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2493  * and restart from that address when we reach that vma again.  It might
2494  * have been split or merged, shrunk or extended, but never shifted: so
2495  * restart_addr remains valid so long as it remains in the vma's range.
2496  * unmap_mapping_range forces truncate_count to leap over page-aligned
2497  * values so we can save vma's restart_addr in its truncate_count field.
2498  */
2499 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2500 
2501 static void reset_vma_truncate_counts(struct address_space *mapping)
2502 {
2503         struct vm_area_struct *vma;
2504         struct prio_tree_iter iter;
2505 
2506         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2507                 vma->vm_truncate_count = 0;
2508         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2509                 vma->vm_truncate_count = 0;
2510 }
2511 
2512 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2513                 unsigned long start_addr, unsigned long end_addr,
2514                 struct zap_details *details)
2515 {
2516         unsigned long restart_addr;
2517         int need_break;
2518 
2519         /*
2520          * files that support invalidating or truncating portions of the
2521          * file from under mmaped areas must have their ->fault function
2522          * return a locked page (and set VM_FAULT_LOCKED in the return).
2523          * This provides synchronisation against concurrent unmapping here.
2524          */
2525 
2526 again:
2527         restart_addr = vma->vm_truncate_count;
2528         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2529                 start_addr = restart_addr;
2530                 if (start_addr >= end_addr) {
2531                         /* Top of vma has been split off since last time */
2532                         vma->vm_truncate_count = details->truncate_count;
2533                         return 0;
2534                 }
2535         }
2536 
2537         restart_addr = zap_page_range(vma, start_addr,
2538                                         end_addr - start_addr, details);
2539         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2540 
2541         if (restart_addr >= end_addr) {
2542                 /* We have now completed this vma: mark it so */
2543                 vma->vm_truncate_count = details->truncate_count;
2544                 if (!need_break)
2545                         return 0;
2546         } else {
2547                 /* Note restart_addr in vma's truncate_count field */
2548                 vma->vm_truncate_count = restart_addr;
2549                 if (!need_break)
2550                         goto again;
2551         }
2552 
2553         spin_unlock(details->i_mmap_lock);
2554         cond_resched();
2555         spin_lock(details->i_mmap_lock);
2556         return -EINTR;
2557 }
2558 
2559 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2560                                             struct zap_details *details)
2561 {
2562         struct vm_area_struct *vma;
2563         struct prio_tree_iter iter;
2564         pgoff_t vba, vea, zba, zea;
2565 
2566 restart:
2567         vma_prio_tree_foreach(vma, &iter, root,
2568                         details->first_index, details->last_index) {
2569                 /* Skip quickly over those we have already dealt with */
2570                 if (vma->vm_truncate_count == details->truncate_count)
2571                         continue;
2572 
2573                 vba = vma->vm_pgoff;
2574                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2575                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2576                 zba = details->first_index;
2577                 if (zba < vba)
2578                         zba = vba;
2579                 zea = details->last_index;
2580                 if (zea > vea)
2581                         zea = vea;
2582 
2583                 if (unmap_mapping_range_vma(vma,
2584                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2585                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2586                                 details) < 0)
2587                         goto restart;
2588         }
2589 }
2590 
2591 static inline void unmap_mapping_range_list(struct list_head *head,
2592                                             struct zap_details *details)
2593 {
2594         struct vm_area_struct *vma;
2595 
2596         /*
2597          * In nonlinear VMAs there is no correspondence between virtual address
2598          * offset and file offset.  So we must perform an exhaustive search
2599          * across *all* the pages in each nonlinear VMA, not just the pages
2600          * whose virtual address lies outside the file truncation point.
2601          */
2602 restart:
2603         list_for_each_entry(vma, head, shared.vm_set.list) {
2604                 /* Skip quickly over those we have already dealt with */
2605                 if (vma->vm_truncate_count == details->truncate_count)
2606                         continue;
2607                 details->nonlinear_vma = vma;
2608                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2609                                         vma->vm_end, details) < 0)
2610                         goto restart;
2611         }
2612 }
2613 
2614 /**
2615  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2616  * @mapping: the address space containing mmaps to be unmapped.
2617  * @holebegin: byte in first page to unmap, relative to the start of
2618  * the underlying file.  This will be rounded down to a PAGE_SIZE
2619  * boundary.  Note that this is different from truncate_pagecache(), which
2620  * must keep the partial page.  In contrast, we must get rid of
2621  * partial pages.
2622  * @holelen: size of prospective hole in bytes.  This will be rounded
2623  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2624  * end of the file.
2625  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2626  * but 0 when invalidating pagecache, don't throw away private data.
2627  */
2628 void unmap_mapping_range(struct address_space *mapping,
2629                 loff_t const holebegin, loff_t const holelen, int even_cows)
2630 {
2631         struct zap_details details;
2632         pgoff_t hba = holebegin >> PAGE_SHIFT;
2633         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2634 
2635         /* Check for overflow. */
2636         if (sizeof(holelen) > sizeof(hlen)) {
2637                 long long holeend =
2638                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2639                 if (holeend & ~(long long)ULONG_MAX)
2640                         hlen = ULONG_MAX - hba + 1;
2641         }
2642 
2643         details.check_mapping = even_cows? NULL: mapping;
2644         details.nonlinear_vma = NULL;
2645         details.first_index = hba;
2646         details.last_index = hba + hlen - 1;
2647         if (details.last_index < details.first_index)
2648                 details.last_index = ULONG_MAX;
2649         details.i_mmap_lock = &mapping->i_mmap_lock;
2650 
2651         mutex_lock(&mapping->unmap_mutex);
2652         spin_lock(&mapping->i_mmap_lock);
2653 
2654         /* Protect against endless unmapping loops */
2655         mapping->truncate_count++;
2656         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2657                 if (mapping->truncate_count == 0)
2658                         reset_vma_truncate_counts(mapping);
2659                 mapping->truncate_count++;
2660         }
2661         details.truncate_count = mapping->truncate_count;
2662 
2663         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2664                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2665         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2666                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2667         spin_unlock(&mapping->i_mmap_lock);
2668         mutex_unlock(&mapping->unmap_mutex);
2669 }
2670 EXPORT_SYMBOL(unmap_mapping_range);
2671 
2672 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2673 {
2674         struct address_space *mapping = inode->i_mapping;
2675 
2676         /*
2677          * If the underlying filesystem is not going to provide
2678          * a way to truncate a range of blocks (punch a hole) -
2679          * we should return failure right now.
2680          */
2681         if (!inode->i_op->truncate_range)
2682                 return -ENOSYS;
2683 
2684         mutex_lock(&inode->i_mutex);
2685         down_write(&inode->i_alloc_sem);
2686         unmap_mapping_range(mapping, offset, (end - offset), 1);
2687         truncate_inode_pages_range(mapping, offset, end);
2688         unmap_mapping_range(mapping, offset, (end - offset), 1);
2689         inode->i_op->truncate_range(inode, offset, end);
2690         up_write(&inode->i_alloc_sem);
2691         mutex_unlock(&inode->i_mutex);
2692 
2693         return 0;
2694 }
2695 
2696 /*
2697  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2698  * but allow concurrent faults), and pte mapped but not yet locked.
2699  * We return with mmap_sem still held, but pte unmapped and unlocked.
2700  */
2701 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2702                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2703                 unsigned int flags, pte_t orig_pte)
2704 {
2705         spinlock_t *ptl;
2706         struct page *page, *swapcache = NULL;
2707         swp_entry_t entry;
2708         pte_t pte;
2709         int locked;
2710         struct mem_cgroup *ptr = NULL;
2711         int exclusive = 0;
2712         int ret = 0;
2713 
2714         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2715                 goto out;
2716 
2717         entry = pte_to_swp_entry(orig_pte);
2718         if (unlikely(non_swap_entry(entry))) {
2719                 if (is_migration_entry(entry)) {
2720                         migration_entry_wait(mm, pmd, address);
2721                 } else if (is_hwpoison_entry(entry)) {
2722                         ret = VM_FAULT_HWPOISON;
2723                 } else {
2724                         print_bad_pte(vma, address, orig_pte, NULL);
2725                         ret = VM_FAULT_SIGBUS;
2726                 }
2727                 goto out;
2728         }
2729         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2730         page = lookup_swap_cache(entry);
2731         if (!page) {
2732                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2733                 page = swapin_readahead(entry,
2734                                         GFP_HIGHUSER_MOVABLE, vma, address);
2735                 if (!page) {
2736                         /*
2737                          * Back out if somebody else faulted in this pte
2738                          * while we released the pte lock.
2739                          */
2740                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2741                         if (likely(pte_same(*page_table, orig_pte)))
2742                                 ret = VM_FAULT_OOM;
2743                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2744                         goto unlock;
2745                 }
2746 
2747                 /* Had to read the page from swap area: Major fault */
2748                 ret = VM_FAULT_MAJOR;
2749                 count_vm_event(PGMAJFAULT);
2750         } else if (PageHWPoison(page)) {
2751                 /*
2752                  * hwpoisoned dirty swapcache pages are kept for killing
2753                  * owner processes (which may be unknown at hwpoison time)
2754                  */
2755                 ret = VM_FAULT_HWPOISON;
2756                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2757                 goto out_release;
2758         }
2759 
2760         locked = lock_page_or_retry(page, mm, flags);
2761         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2762         if (!locked) {
2763                 ret |= VM_FAULT_RETRY;
2764                 goto out_release;
2765         }
2766 
2767         /*
2768          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2769          * release the swapcache from under us.  The page pin, and pte_same
2770          * test below, are not enough to exclude that.  Even if it is still
2771          * swapcache, we need to check that the page's swap has not changed.
2772          */
2773         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2774                 goto out_page;
2775 
2776         if (ksm_might_need_to_copy(page, vma, address)) {
2777                 swapcache = page;
2778                 page = ksm_does_need_to_copy(page, vma, address);
2779 
2780                 if (unlikely(!page)) {
2781                         ret = VM_FAULT_OOM;
2782                         page = swapcache;
2783                         swapcache = NULL;
2784                         goto out_page;
2785                 }
2786         }
2787 
2788         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2789                 ret = VM_FAULT_OOM;
2790                 goto out_page;
2791         }
2792 
2793         /*
2794          * Back out if somebody else already faulted in this pte.
2795          */
2796         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2797         if (unlikely(!pte_same(*page_table, orig_pte)))
2798                 goto out_nomap;
2799 
2800         if (unlikely(!PageUptodate(page))) {
2801                 ret = VM_FAULT_SIGBUS;
2802                 goto out_nomap;
2803         }
2804 
2805         /*
2806          * The page isn't present yet, go ahead with the fault.
2807          *
2808          * Be careful about the sequence of operations here.
2809          * To get its accounting right, reuse_swap_page() must be called
2810          * while the page is counted on swap but not yet in mapcount i.e.
2811          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2812          * must be called after the swap_free(), or it will never succeed.
2813          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2814          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2815          * in page->private. In this case, a record in swap_cgroup  is silently
2816          * discarded at swap_free().
2817          */
2818 
2819         inc_mm_counter_fast(mm, MM_ANONPAGES);
2820         dec_mm_counter_fast(mm, MM_SWAPENTS);
2821         pte = mk_pte(page, vma->vm_page_prot);
2822         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2823                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2824                 flags &= ~FAULT_FLAG_WRITE;
2825                 ret |= VM_FAULT_WRITE;
2826                 exclusive = 1;
2827         }
2828         flush_icache_page(vma, page);
2829         set_pte_at(mm, address, page_table, pte);
2830         do_page_add_anon_rmap(page, vma, address, exclusive);
2831         /* It's better to call commit-charge after rmap is established */
2832         mem_cgroup_commit_charge_swapin(page, ptr);
2833 
2834         swap_free(entry);
2835         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2836                 try_to_free_swap(page);
2837         unlock_page(page);
2838         if (swapcache) {
2839                 /*
2840                  * Hold the lock to avoid the swap entry to be reused
2841                  * until we take the PT lock for the pte_same() check
2842                  * (to avoid false positives from pte_same). For
2843                  * further safety release the lock after the swap_free
2844                  * so that the swap count won't change under a
2845                  * parallel locked swapcache.
2846                  */
2847                 unlock_page(swapcache);
2848                 page_cache_release(swapcache);
2849         }
2850 
2851         if (flags & FAULT_FLAG_WRITE) {
2852                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2853                 if (ret & VM_FAULT_ERROR)
2854                         ret &= VM_FAULT_ERROR;
2855                 goto out;
2856         }
2857 
2858         /* No need to invalidate - it was non-present before */
2859         update_mmu_cache(vma, address, page_table);
2860 unlock:
2861         pte_unmap_unlock(page_table, ptl);
2862 out:
2863         return ret;
2864 out_nomap:
2865         mem_cgroup_cancel_charge_swapin(ptr);
2866         pte_unmap_unlock(page_table, ptl);
2867 out_page:
2868         unlock_page(page);
2869 out_release:
2870         page_cache_release(page);
2871         if (swapcache) {
2872                 unlock_page(swapcache);
2873                 page_cache_release(swapcache);
2874         }
2875         return ret;
2876 }
2877 
2878 /*
2879  * This is like a special single-page "expand_{down|up}wards()",
2880  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2881  * doesn't hit another vma.
2882  */
2883 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2884 {
2885         address &= PAGE_MASK;
2886         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2887                 struct vm_area_struct *prev = vma->vm_prev;
2888 
2889                 /*
2890                  * Is there a mapping abutting this one below?
2891                  *
2892                  * That's only ok if it's the same stack mapping
2893                  * that has gotten split..
2894                  */
2895                 if (prev && prev->vm_end == address)
2896                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2897 
2898                 expand_stack(vma, address - PAGE_SIZE);
2899         }
2900         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2901                 struct vm_area_struct *next = vma->vm_next;
2902 
2903                 /* As VM_GROWSDOWN but s/below/above/ */
2904                 if (next && next->vm_start == address + PAGE_SIZE)
2905                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2906 
2907                 expand_upwards(vma, address + PAGE_SIZE);
2908         }
2909         return 0;
2910 }
2911 
2912 /*
2913  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2914  * but allow concurrent faults), and pte mapped but not yet locked.
2915  * We return with mmap_sem still held, but pte unmapped and unlocked.
2916  */
2917 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2918                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2919                 unsigned int flags)
2920 {
2921         struct page *page;
2922         spinlock_t *ptl;
2923         pte_t entry;
2924 
2925         pte_unmap(page_table);
2926 
2927         /* Check if we need to add a guard page to the stack */
2928         if (check_stack_guard_page(vma, address) < 0)
2929                 return VM_FAULT_SIGBUS;
2930 
2931         /* Use the zero-page for reads */
2932         if (!(flags & FAULT_FLAG_WRITE)) {
2933                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2934                                                 vma->vm_page_prot));
2935                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2936                 if (!pte_none(*page_table))
2937                         goto unlock;
2938                 goto setpte;
2939         }
2940 
2941         /* Allocate our own private page. */
2942         if (unlikely(anon_vma_prepare(vma)))
2943                 goto oom;
2944         page = alloc_zeroed_user_highpage_movable(vma, address);
2945         if (!page)
2946                 goto oom;
2947         __SetPageUptodate(page);
2948 
2949         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2950                 goto oom_free_page;
2951 
2952         entry = mk_pte(page, vma->vm_page_prot);
2953         if (vma->vm_flags & VM_WRITE)
2954                 entry = pte_mkwrite(pte_mkdirty(entry));
2955 
2956         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2957         if (!pte_none(*page_table))
2958                 goto release;
2959 
2960         inc_mm_counter_fast(mm, MM_ANONPAGES);
2961         page_add_new_anon_rmap(page, vma, address);
2962 setpte:
2963         set_pte_at(mm, address, page_table, entry);
2964 
2965         /* No need to invalidate - it was non-present before */
2966         update_mmu_cache(vma, address, page_table);
2967 unlock:
2968         pte_unmap_unlock(page_table, ptl);
2969         return 0;
2970 release:
2971         mem_cgroup_uncharge_page(page);
2972         page_cache_release(page);
2973         goto unlock;
2974 oom_free_page:
2975         page_cache_release(page);
2976 oom:
2977         return VM_FAULT_OOM;
2978 }
2979 
2980 /*
2981  * __do_fault() tries to create a new page mapping. It aggressively
2982  * tries to share with existing pages, but makes a separate copy if
2983  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2984  * the next page fault.
2985  *
2986  * As this is called only for pages that do not currently exist, we
2987  * do not need to flush old virtual caches or the TLB.
2988  *
2989  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2990  * but allow concurrent faults), and pte neither mapped nor locked.
2991  * We return with mmap_sem still held, but pte unmapped and unlocked.
2992  */
2993 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2994                 unsigned long address, pmd_t *pmd,
2995                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2996 {
2997         pte_t *page_table;
2998         spinlock_t *ptl;
2999         struct page *page;
3000         pte_t entry;
3001         int anon = 0;
3002         int charged = 0;
3003         struct page *dirty_page = NULL;
3004         struct vm_fault vmf;
3005         int ret;
3006         int page_mkwrite = 0;
3007 
3008         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3009         vmf.pgoff = pgoff;
3010         vmf.flags = flags;
3011         vmf.page = NULL;
3012 
3013         ret = vma->vm_ops->fault(vma, &vmf);
3014         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3015                             VM_FAULT_RETRY)))
3016                 return ret;
3017 
3018         if (unlikely(PageHWPoison(vmf.page))) {
3019                 if (ret & VM_FAULT_LOCKED)
3020                         unlock_page(vmf.page);
3021                 return VM_FAULT_HWPOISON;
3022         }
3023 
3024         /*
3025          * For consistency in subsequent calls, make the faulted page always
3026          * locked.
3027          */
3028         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3029                 lock_page(vmf.page);
3030         else
3031                 VM_BUG_ON(!PageLocked(vmf.page));
3032 
3033         /*
3034          * Should we do an early C-O-W break?
3035          */
3036         page = vmf.page;
3037         if (flags & FAULT_FLAG_WRITE) {
3038                 if (!(vma->vm_flags & VM_SHARED)) {
3039                         anon = 1;
3040                         if (unlikely(anon_vma_prepare(vma))) {
3041                                 ret = VM_FAULT_OOM;
3042                                 goto out;
3043                         }
3044                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3045                                                 vma, address);
3046                         if (!page) {
3047                                 ret = VM_FAULT_OOM;
3048                                 goto out;
3049                         }
3050                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3051                                 ret = VM_FAULT_OOM;
3052                                 page_cache_release(page);
3053                                 goto out;
3054                         }
3055                         charged = 1;
3056                         copy_user_highpage(page, vmf.page, address, vma);
3057                         __SetPageUptodate(page);
3058                 } else {
3059                         /*
3060                          * If the page will be shareable, see if the backing
3061                          * address space wants to know that the page is about
3062                          * to become writable
3063                          */
3064                         if (vma->vm_ops->page_mkwrite) {
3065                                 int tmp;
3066 
3067                                 unlock_page(page);
3068                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3069                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3070                                 if (unlikely(tmp &
3071                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3072                                         ret = tmp;
3073                                         goto unwritable_page;
3074                                 }
3075                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3076                                         lock_page(page);
3077                                         if (!page->mapping) {
3078                                                 ret = 0; /* retry the fault */
3079                                                 unlock_page(page);
3080                                                 goto unwritable_page;
3081                                         }
3082                                 } else
3083                                         VM_BUG_ON(!PageLocked(page));
3084                                 page_mkwrite = 1;
3085                         }
3086                 }
3087 
3088         }
3089 
3090         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3091 
3092         /*
3093          * This silly early PAGE_DIRTY setting removes a race
3094          * due to the bad i386 page protection. But it's valid
3095          * for other architectures too.
3096          *
3097          * Note that if FAULT_FLAG_WRITE is set, we either now have
3098          * an exclusive copy of the page, or this is a shared mapping,
3099          * so we can make it writable and dirty to avoid having to
3100          * handle that later.
3101          */
3102         /* Only go through if we didn't race with anybody else... */
3103         if (likely(pte_same(*page_table, orig_pte))) {
3104                 flush_icache_page(vma, page);
3105                 entry = mk_pte(page, vma->vm_page_prot);
3106                 if (flags & FAULT_FLAG_WRITE)
3107                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3108                 if (anon) {
3109                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3110                         page_add_new_anon_rmap(page, vma, address);
3111                 } else {
3112                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3113                         page_add_file_rmap(page);
3114                         if (flags & FAULT_FLAG_WRITE) {
3115                                 dirty_page = page;
3116                                 get_page(dirty_page);
3117                         }
3118                 }
3119                 set_pte_at(mm, address, page_table, entry);
3120 
3121                 /* no need to invalidate: a not-present page won't be cached */
3122                 update_mmu_cache(vma, address, page_table);
3123         } else {
3124                 if (charged)
3125                         mem_cgroup_uncharge_page(page);
3126                 if (anon)
3127                         page_cache_release(page);
3128                 else
3129                         anon = 1; /* no anon but release faulted_page */
3130         }
3131 
3132         pte_unmap_unlock(page_table, ptl);
3133 
3134 out:
3135         if (dirty_page) {
3136                 struct address_space *mapping = page->mapping;
3137 
3138                 if (set_page_dirty(dirty_page))
3139                         page_mkwrite = 1;
3140                 unlock_page(dirty_page);
3141                 put_page(dirty_page);
3142                 if (page_mkwrite && mapping) {
3143                         /*
3144                          * Some device drivers do not set page.mapping but still
3145                          * dirty their pages
3146                          */
3147                         balance_dirty_pages_ratelimited(mapping);
3148                 }
3149 
3150                 /* file_update_time outside page_lock */
3151                 if (vma->vm_file)
3152                         file_update_time(vma->vm_file);
3153         } else {
3154                 unlock_page(vmf.page);
3155                 if (anon)
3156                         page_cache_release(vmf.page);
3157         }
3158 
3159         return ret;
3160 
3161 unwritable_page:
3162         page_cache_release(page);
3163         return ret;
3164 }
3165 
3166 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3167                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3168                 unsigned int flags, pte_t orig_pte)
3169 {
3170         pgoff_t pgoff = (((address & PAGE_MASK)
3171                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3172 
3173         pte_unmap(page_table);
3174         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3175 }
3176 
3177 /*
3178  * Fault of a previously existing named mapping. Repopulate the pte
3179  * from the encoded file_pte if possible. This enables swappable
3180  * nonlinear vmas.
3181  *
3182  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3183  * but allow concurrent faults), and pte mapped but not yet locked.
3184  * We return with mmap_sem still held, but pte unmapped and unlocked.
3185  */
3186 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3187                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3188                 unsigned int flags, pte_t orig_pte)
3189 {
3190         pgoff_t pgoff;
3191 
3192         flags |= FAULT_FLAG_NONLINEAR;
3193 
3194         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3195                 return 0;
3196 
3197         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3198                 /*
3199                  * Page table corrupted: show pte and kill process.
3200                  */
3201                 print_bad_pte(vma, address, orig_pte, NULL);
3202                 return VM_FAULT_SIGBUS;
3203         }
3204 
3205         pgoff = pte_to_pgoff(orig_pte);
3206         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3207 }
3208 
3209 /*
3210  * These routines also need to handle stuff like marking pages dirty
3211  * and/or accessed for architectures that don't do it in hardware (most
3212  * RISC architectures).  The early dirtying is also good on the i386.
3213  *
3214  * There is also a hook called "update_mmu_cache()" that architectures
3215  * with external mmu caches can use to update those (ie the Sparc or
3216  * PowerPC hashed page tables that act as extended TLBs).
3217  *
3218  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3219  * but allow concurrent faults), and pte mapped but not yet locked.
3220  * We return with mmap_sem still held, but pte unmapped and unlocked.
3221  */
3222 int handle_pte_fault(struct mm_struct *mm,
3223                      struct vm_area_struct *vma, unsigned long address,
3224                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3225 {
3226         pte_t entry;
3227         spinlock_t *ptl;
3228 
3229         entry = *pte;
3230         if (!pte_present(entry)) {
3231                 if (pte_none(entry)) {
3232                         if (vma->vm_ops) {
3233                                 if (likely(vma->vm_ops->fault))
3234                                         return do_linear_fault(mm, vma, address,
3235                                                 pte, pmd, flags, entry);
3236                         }
3237                         return do_anonymous_page(mm, vma, address,
3238                                                  pte, pmd, flags);
3239                 }
3240                 if (pte_file(entry))
3241                         return do_nonlinear_fault(mm, vma, address,
3242                                         pte, pmd, flags, entry);
3243                 return do_swap_page(mm, vma, address,
3244                                         pte, pmd, flags, entry);
3245         }
3246 
3247         ptl = pte_lockptr(mm, pmd);
3248         spin_lock(ptl);
3249         if (unlikely(!pte_same(*pte, entry)))
3250                 goto unlock;
3251         if (flags & FAULT_FLAG_WRITE) {
3252                 if (!pte_write(entry))
3253                         return do_wp_page(mm, vma, address,
3254                                         pte, pmd, ptl, entry);
3255                 entry = pte_mkdirty(entry);
3256         }
3257         entry = pte_mkyoung(entry);
3258         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3259                 update_mmu_cache(vma, address, pte);
3260         } else {
3261                 /*
3262                  * This is needed only for protection faults but the arch code
3263                  * is not yet telling us if this is a protection fault or not.
3264                  * This still avoids useless tlb flushes for .text page faults
3265                  * with threads.
3266                  */
3267                 if (flags & FAULT_FLAG_WRITE)
3268                         flush_tlb_fix_spurious_fault(vma, address);
3269         }
3270 unlock:
3271         pte_unmap_unlock(pte, ptl);
3272         return 0;
3273 }
3274 
3275 /*
3276  * By the time we get here, we already hold the mm semaphore
3277  */
3278 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3279                 unsigned long address, unsigned int flags)
3280 {
3281         pgd_t *pgd;
3282         pud_t *pud;
3283         pmd_t *pmd;
3284         pte_t *pte;
3285 
3286         __set_current_state(TASK_RUNNING);
3287 
3288         count_vm_event(PGFAULT);
3289 
3290         /* do counter updates before entering really critical section. */
3291         check_sync_rss_stat(current);
3292 
3293         if (unlikely(is_vm_hugetlb_page(vma)))
3294                 return hugetlb_fault(mm, vma, address, flags);
3295 
3296         pgd = pgd_offset(mm, address);
3297         pud = pud_alloc(mm, pgd, address);
3298         if (!pud)
3299                 return VM_FAULT_OOM;
3300         pmd = pmd_alloc(mm, pud, address);
3301         if (!pmd)
3302                 return VM_FAULT_OOM;
3303         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3304                 if (!vma->vm_ops)
3305                         return do_huge_pmd_anonymous_page(mm, vma, address,
3306                                                           pmd, flags);
3307         } else {
3308                 pmd_t orig_pmd = *pmd;
3309                 barrier();
3310                 if (pmd_trans_huge(orig_pmd)) {
3311                         if (flags & FAULT_FLAG_WRITE &&
3312                             !pmd_write(orig_pmd) &&
3313                             !pmd_trans_splitting(orig_pmd))
3314                                 return do_huge_pmd_wp_page(mm, vma, address,
3315                                                            pmd, orig_pmd);
3316                         return 0;
3317                 }
3318         }
3319 
3320         /*
3321          * Use __pte_alloc instead of pte_alloc_map, because we can't
3322          * run pte_offset_map on the pmd, if an huge pmd could
3323          * materialize from under us from a different thread.
3324          */
3325         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
3326                 return VM_FAULT_OOM;
3327         /* if an huge pmd materialized from under us just retry later */
3328         if (unlikely(pmd_trans_huge(*pmd)))
3329                 return 0;
3330         /*
3331          * A regular pmd is established and it can't morph into a huge pmd
3332          * from under us anymore at this point because we hold the mmap_sem
3333          * read mode and khugepaged takes it in write mode. So now it's
3334          * safe to run pte_offset_map().
3335          */
3336         pte = pte_offset_map(pmd, address);
3337 
3338         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3339 }
3340 
3341 #ifndef __PAGETABLE_PUD_FOLDED
3342 /*
3343  * Allocate page upper directory.
3344  * We've already handled the fast-path in-line.
3345  */
3346 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3347 {
3348         pud_t *new = pud_alloc_one(mm, address);
3349         if (!new)
3350                 return -ENOMEM;
3351 
3352         smp_wmb(); /* See comment in __pte_alloc */
3353 
3354         spin_lock(&mm->page_table_lock);
3355         if (pgd_present(*pgd))          /* Another has populated it */
3356                 pud_free(mm, new);
3357         else
3358                 pgd_populate(mm, pgd, new);
3359         spin_unlock(&mm->page_table_lock);
3360         return 0;
3361 }
3362 #endif /* __PAGETABLE_PUD_FOLDED */
3363 
3364 #ifndef __PAGETABLE_PMD_FOLDED
3365 /*
3366  * Allocate page middle directory.
3367  * We've already handled the fast-path in-line.
3368  */
3369 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3370 {
3371         pmd_t *new = pmd_alloc_one(mm, address);
3372         if (!new)
3373                 return -ENOMEM;
3374 
3375         smp_wmb(); /* See comment in __pte_alloc */
3376 
3377         spin_lock(&mm->page_table_lock);
3378 #ifndef __ARCH_HAS_4LEVEL_HACK
3379         if (pud_present(*pud))          /* Another has populated it */
3380                 pmd_free(mm, new);
3381         else
3382                 pud_populate(mm, pud, new);
3383 #else
3384         if (pgd_present(*pud))          /* Another has populated it */
3385                 pmd_free(mm, new);
3386         else
3387                 pgd_populate(mm, pud, new);
3388 #endif /* __ARCH_HAS_4LEVEL_HACK */
3389         spin_unlock(&mm->page_table_lock);
3390         return 0;
3391 }
3392 #endif /* __PAGETABLE_PMD_FOLDED */
3393 
3394 int make_pages_present(unsigned long addr, unsigned long end)
3395 {
3396         int ret, len, write;
3397         struct vm_area_struct * vma;
3398 
3399         vma = find_vma(current->mm, addr);
3400         if (!vma)
3401                 return -ENOMEM;
3402         /*
3403          * We want to touch writable mappings with a write fault in order
3404          * to break COW, except for shared mappings because these don't COW
3405          * and we would not want to dirty them for nothing.
3406          */
3407         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3408         BUG_ON(addr >= end);
3409         BUG_ON(end > vma->vm_end);
3410         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3411         ret = get_user_pages(current, current->mm, addr,
3412                         len, write, 0, NULL, NULL);
3413         if (ret < 0)
3414                 return ret;
3415         return ret == len ? 0 : -EFAULT;
3416 }
3417 
3418 #if !defined(__HAVE_ARCH_GATE_AREA)
3419 
3420 #if defined(AT_SYSINFO_EHDR)
3421 static struct vm_area_struct gate_vma;
3422 
3423 static int __init gate_vma_init(void)
3424 {
3425         gate_vma.vm_mm = NULL;
3426         gate_vma.vm_start = FIXADDR_USER_START;
3427         gate_vma.vm_end = FIXADDR_USER_END;
3428         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3429         gate_vma.vm_page_prot = __P101;
3430         /*
3431          * Make sure the vDSO gets into every core dump.
3432          * Dumping its contents makes post-mortem fully interpretable later
3433          * without matching up the same kernel and hardware config to see
3434          * what PC values meant.
3435          */
3436         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3437         return 0;
3438 }
3439 __initcall(gate_vma_init);
3440 #endif
3441 
3442 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3443 {
3444 #ifdef AT_SYSINFO_EHDR
3445         return &gate_vma;
3446 #else
3447         return NULL;
3448 #endif
3449 }
3450 
3451 int in_gate_area_no_task(unsigned long addr)
3452 {
3453 #ifdef AT_SYSINFO_EHDR
3454         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3455                 return 1;
3456 #endif
3457         return 0;
3458 }
3459 
3460 #endif  /* __HAVE_ARCH_GATE_AREA */
3461 
3462 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3463                 pte_t **ptepp, spinlock_t **ptlp)
3464 {
3465         pgd_t *pgd;
3466         pud_t *pud;
3467         pmd_t *pmd;
3468         pte_t *ptep;
3469 
3470         pgd = pgd_offset(mm, address);
3471         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3472                 goto out;
3473 
3474         pud = pud_offset(pgd, address);
3475         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3476                 goto out;
3477 
3478         pmd = pmd_offset(pud, address);
3479         VM_BUG_ON(pmd_trans_huge(*pmd));
3480         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3481                 goto out;
3482 
3483         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3484         if (pmd_huge(*pmd))
3485                 goto out;
3486 
3487         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3488         if (!ptep)
3489                 goto out;
3490         if (!pte_present(*ptep))
3491                 goto unlock;
3492         *ptepp = ptep;
3493         return 0;
3494 unlock:
3495         pte_unmap_unlock(ptep, *ptlp);
3496 out:
3497         return -EINVAL;
3498 }
3499 
3500 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3501                              pte_t **ptepp, spinlock_t **ptlp)
3502 {
3503         int res;
3504 
3505         /* (void) is needed to make gcc happy */
3506         (void) __cond_lock(*ptlp,
3507                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3508         return res;
3509 }
3510 
3511 /**
3512  * follow_pfn - look up PFN at a user virtual address
3513  * @vma: memory mapping
3514  * @address: user virtual address
3515  * @pfn: location to store found PFN
3516  *
3517  * Only IO mappings and raw PFN mappings are allowed.
3518  *
3519  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3520  */
3521 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3522         unsigned long *pfn)
3523 {
3524         int ret = -EINVAL;
3525         spinlock_t *ptl;
3526         pte_t *ptep;
3527 
3528         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3529                 return ret;
3530 
3531         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3532         if (ret)
3533                 return ret;
3534         *pfn = pte_pfn(*ptep);
3535         pte_unmap_unlock(ptep, ptl);
3536         return 0;
3537 }
3538 EXPORT_SYMBOL(follow_pfn);
3539 
3540 #ifdef CONFIG_HAVE_IOREMAP_PROT
3541 int follow_phys(struct vm_area_struct *vma,
3542                 unsigned long address, unsigned int flags,
3543                 unsigned long *prot, resource_size_t *phys)
3544 {
3545         int ret = -EINVAL;
3546         pte_t *ptep, pte;
3547         spinlock_t *ptl;
3548 
3549         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3550                 goto out;
3551 
3552         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3553                 goto out;
3554         pte = *ptep;
3555 
3556         if ((flags & FOLL_WRITE) && !pte_write(pte))
3557                 goto unlock;
3558 
3559         *prot = pgprot_val(pte_pgprot(pte));
3560         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3561 
3562         ret = 0;
3563 unlock:
3564         pte_unmap_unlock(ptep, ptl);
3565 out:
3566         return ret;
3567 }
3568 
3569 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3570                         void *buf, int len, int write)
3571 {
3572         resource_size_t phys_addr;
3573         unsigned long prot = 0;
3574         void __iomem *maddr;
3575         int offset = addr & (PAGE_SIZE-1);
3576 
3577         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3578                 return -EINVAL;
3579 
3580         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3581         if (write)
3582                 memcpy_toio(maddr + offset, buf, len);
3583         else
3584                 memcpy_fromio(buf, maddr + offset, len);
3585         iounmap(maddr);
3586 
3587         return len;
3588 }
3589 #endif
3590 
3591 /*
3592  * Access another process' address space.
3593  * Source/target buffer must be kernel space,
3594  * Do not walk the page table directly, use get_user_pages
3595  */
3596 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3597 {
3598         struct mm_struct *mm;
3599         struct vm_area_struct *vma;
3600         void *old_buf = buf;
3601 
3602         mm = get_task_mm(tsk);
3603         if (!mm)
3604                 return 0;
3605 
3606         down_read(&mm->mmap_sem);
3607         /* ignore errors, just check how much was successfully transferred */
3608         while (len) {
3609                 int bytes, ret, offset;
3610                 void *maddr;
3611                 struct page *page = NULL;
3612 
3613                 ret = get_user_pages(tsk, mm, addr, 1,
3614                                 write, 1, &page, &vma);
3615                 if (ret <= 0) {
3616                         /*
3617                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3618                          * we can access using slightly different code.
3619                          */
3620 #ifdef CONFIG_HAVE_IOREMAP_PROT
3621                         vma = find_vma(mm, addr);
3622                         if (!vma)
3623                                 break;
3624                         if (vma->vm_ops && vma->vm_ops->access)
3625                                 ret = vma->vm_ops->access(vma, addr, buf,
3626                                                           len, write);
3627                         if (ret <= 0)
3628 #endif
3629                                 break;
3630                         bytes = ret;
3631                 } else {
3632                         bytes = len;
3633                         offset = addr & (PAGE_SIZE-1);
3634                         if (bytes > PAGE_SIZE-offset)
3635                                 bytes = PAGE_SIZE-offset;
3636 
3637                         maddr = kmap(page);
3638                         if (write) {
3639                                 copy_to_user_page(vma, page, addr,
3640                                                   maddr + offset, buf, bytes);
3641                                 set_page_dirty_lock(page);
3642                         } else {
3643                                 copy_from_user_page(vma, page, addr,
3644                                                     buf, maddr + offset, bytes);
3645                         }
3646                         kunmap(page);
3647                         page_cache_release(page);
3648                 }
3649                 len -= bytes;
3650                 buf += bytes;
3651                 addr += bytes;
3652         }
3653         up_read(&mm->mmap_sem);
3654         mmput(mm);
3655 
3656         return buf - old_buf;
3657 }
3658 
3659 /*
3660  * Print the name of a VMA.
3661  */
3662 void print_vma_addr(char *prefix, unsigned long ip)
3663 {
3664         struct mm_struct *mm = current->mm;
3665         struct vm_area_struct *vma;
3666 
3667         /*
3668          * Do not print if we are in atomic
3669          * contexts (in exception stacks, etc.):
3670          */
3671         if (preempt_count())
3672                 return;
3673 
3674         down_read(&mm->mmap_sem);
3675         vma = find_vma(mm, ip);
3676         if (vma && vma->vm_file) {
3677                 struct file *f = vma->vm_file;
3678                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3679                 if (buf) {
3680                         char *p, *s;
3681 
3682                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3683                         if (IS_ERR(p))
3684                                 p = "?";
3685                         s = strrchr(p, '/');
3686                         if (s)
3687                                 p = s+1;
3688                         printk("%s%s[%lx+%lx]", prefix, p,
3689                                         vma->vm_start,
3690                                         vma->vm_end - vma->vm_start);
3691                         free_page((unsigned long)buf);
3692                 }
3693         }
3694         up_read(&current->mm->mmap_sem);
3695 }
3696 
3697 #ifdef CONFIG_PROVE_LOCKING
3698 void might_fault(void)
3699 {
3700         /*
3701          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3702          * holding the mmap_sem, this is safe because kernel memory doesn't
3703          * get paged out, therefore we'll never actually fault, and the
3704          * below annotations will generate false positives.
3705          */
3706         if (segment_eq(get_fs(), KERNEL_DS))
3707                 return;
3708 
3709         might_sleep();
3710         /*
3711          * it would be nicer only to annotate paths which are not under
3712          * pagefault_disable, however that requires a larger audit and
3713          * providing helpers like get_user_atomic.
3714          */
3715         if (!in_atomic() && current->mm)
3716                 might_lock_read(&current->mm->mmap_sem);
3717 }
3718 EXPORT_SYMBOL(might_fault);
3719 #endif
3720 
3721 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3722 static void clear_gigantic_page(struct page *page,
3723                                 unsigned long addr,
3724                                 unsigned int pages_per_huge_page)
3725 {
3726         int i;
3727         struct page *p = page;
3728 
3729         might_sleep();
3730         for (i = 0; i < pages_per_huge_page;
3731              i++, p = mem_map_next(p, page, i)) {
3732                 cond_resched();
3733                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3734         }
3735 }
3736 void clear_huge_page(struct page *page,
3737                      unsigned long addr, unsigned int pages_per_huge_page)
3738 {
3739         int i;
3740 
3741         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3742                 clear_gigantic_page(page, addr, pages_per_huge_page);
3743                 return;
3744         }
3745 
3746         might_sleep();
3747         for (i = 0; i < pages_per_huge_page; i++) {
3748                 cond_resched();
3749                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3750         }
3751 }
3752 
3753 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3754                                     unsigned long addr,
3755                                     struct vm_area_struct *vma,
3756                                     unsigned int pages_per_huge_page)
3757 {
3758         int i;
3759         struct page *dst_base = dst;
3760         struct page *src_base = src;
3761 
3762         for (i = 0; i < pages_per_huge_page; ) {
3763                 cond_resched();
3764                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3765 
3766                 i++;
3767                 dst = mem_map_next(dst, dst_base, i);
3768                 src = mem_map_next(src, src_base, i);
3769         }
3770 }
3771 
3772 void copy_user_huge_page(struct page *dst, struct page *src,
3773                          unsigned long addr, struct vm_area_struct *vma,
3774                          unsigned int pages_per_huge_page)
3775 {
3776         int i;
3777 
3778         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3779                 copy_user_gigantic_page(dst, src, addr, vma,
3780                                         pages_per_huge_page);
3781                 return;
3782         }
3783 
3784         might_sleep();
3785         for (i = 0; i < pages_per_huge_page; i++) {
3786                 cond_resched();
3787                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3788         }
3789 }
3790 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3791 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us