Root/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 */
72unsigned long max_mapnr;
73struct page *mem_map;
74
75EXPORT_SYMBOL(max_mapnr);
76EXPORT_SYMBOL(mem_map);
77#endif
78
79unsigned 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 */
87void * high_memory;
88
89EXPORT_SYMBOL(num_physpages);
90EXPORT_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 */
98int randomize_va_space __read_mostly =
99#ifdef CONFIG_COMPAT_BRK
100                    1;
101#else
102                    2;
103#endif
104
105static int __init disable_randmaps(char *s)
106{
107    randomize_va_space = 0;
108    return 1;
109}
110__setup("norandmaps", disable_randmaps);
111
112unsigned long zero_pfn __read_mostly;
113unsigned long highest_memmap_pfn __read_mostly;
114
115/*
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117 */
118static int __init init_zero_pfn(void)
119{
120    zero_pfn = page_to_pfn(ZERO_PAGE(0));
121    return 0;
122}
123core_initcall(init_zero_pfn);
124
125
126#if defined(SPLIT_RSS_COUNTING)
127
128static 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
141static 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)
155static 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
163unsigned 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
181void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182{
183    __sync_task_rss_stat(task, mm);
184}
185#else /* SPLIT_RSS_COUNTING */
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
190static void check_sync_rss_stat(struct task_struct *task)
191{
192}
193
194#endif /* SPLIT_RSS_COUNTING */
195
196#ifdef HAVE_GENERIC_MMU_GATHER
197
198static int tlb_next_batch(struct mmu_gather *tlb)
199{
200    struct mmu_gather_batch *batch;
201
202    batch = tlb->active;
203    if (batch->next) {
204        tlb->active = batch->next;
205        return 1;
206    }
207
208    batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
209    if (!batch)
210        return 0;
211
212    batch->next = NULL;
213    batch->nr = 0;
214    batch->max = MAX_GATHER_BATCH;
215
216    tlb->active->next = batch;
217    tlb->active = batch;
218
219    return 1;
220}
221
222/* tlb_gather_mmu
223 * Called to initialize an (on-stack) mmu_gather structure for page-table
224 * tear-down from @mm. The @fullmm argument is used when @mm is without
225 * users and we're going to destroy the full address space (exit/execve).
226 */
227void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
228{
229    tlb->mm = mm;
230
231    tlb->fullmm = fullmm;
232    tlb->need_flush = 0;
233    tlb->fast_mode = (num_possible_cpus() == 1);
234    tlb->local.next = NULL;
235    tlb->local.nr = 0;
236    tlb->local.max = ARRAY_SIZE(tlb->__pages);
237    tlb->active = &tlb->local;
238
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240    tlb->batch = NULL;
241#endif
242}
243
244void tlb_flush_mmu(struct mmu_gather *tlb)
245{
246    struct mmu_gather_batch *batch;
247
248    if (!tlb->need_flush)
249        return;
250    tlb->need_flush = 0;
251    tlb_flush(tlb);
252#ifdef CONFIG_HAVE_RCU_TABLE_FREE
253    tlb_table_flush(tlb);
254#endif
255
256    if (tlb_fast_mode(tlb))
257        return;
258
259    for (batch = &tlb->local; batch; batch = batch->next) {
260        free_pages_and_swap_cache(batch->pages, batch->nr);
261        batch->nr = 0;
262    }
263    tlb->active = &tlb->local;
264}
265
266/* tlb_finish_mmu
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
269 */
270void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
271{
272    struct mmu_gather_batch *batch, *next;
273
274    tlb_flush_mmu(tlb);
275
276    /* keep the page table cache within bounds */
277    check_pgt_cache();
278
279    for (batch = tlb->local.next; batch; batch = next) {
280        next = batch->next;
281        free_pages((unsigned long)batch, 0);
282    }
283    tlb->local.next = NULL;
284}
285
286/* __tlb_remove_page
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
291 */
292int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
293{
294    struct mmu_gather_batch *batch;
295
296    tlb->need_flush = 1;
297
298    if (tlb_fast_mode(tlb)) {
299        free_page_and_swap_cache(page);
300        return 1; /* avoid calling tlb_flush_mmu() */
301    }
302
303    batch = tlb->active;
304    batch->pages[batch->nr++] = page;
305    if (batch->nr == batch->max) {
306        if (!tlb_next_batch(tlb))
307            return 0;
308        batch = tlb->active;
309    }
310    VM_BUG_ON(batch->nr > batch->max);
311
312    return batch->max - batch->nr;
313}
314
315#endif /* HAVE_GENERIC_MMU_GATHER */
316
317#ifdef CONFIG_HAVE_RCU_TABLE_FREE
318
319/*
320 * See the comment near struct mmu_table_batch.
321 */
322
323static void tlb_remove_table_smp_sync(void *arg)
324{
325    /* Simply deliver the interrupt */
326}
327
328static void tlb_remove_table_one(void *table)
329{
330    /*
331     * This isn't an RCU grace period and hence the page-tables cannot be
332     * assumed to be actually RCU-freed.
333     *
334     * It is however sufficient for software page-table walkers that rely on
335     * IRQ disabling. See the comment near struct mmu_table_batch.
336     */
337    smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
338    __tlb_remove_table(table);
339}
340
341static void tlb_remove_table_rcu(struct rcu_head *head)
342{
343    struct mmu_table_batch *batch;
344    int i;
345
346    batch = container_of(head, struct mmu_table_batch, rcu);
347
348    for (i = 0; i < batch->nr; i++)
349        __tlb_remove_table(batch->tables[i]);
350
351    free_page((unsigned long)batch);
352}
353
354void tlb_table_flush(struct mmu_gather *tlb)
355{
356    struct mmu_table_batch **batch = &tlb->batch;
357
358    if (*batch) {
359        call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
360        *batch = NULL;
361    }
362}
363
364void tlb_remove_table(struct mmu_gather *tlb, void *table)
365{
366    struct mmu_table_batch **batch = &tlb->batch;
367
368    tlb->need_flush = 1;
369
370    /*
371     * When there's less then two users of this mm there cannot be a
372     * concurrent page-table walk.
373     */
374    if (atomic_read(&tlb->mm->mm_users) < 2) {
375        __tlb_remove_table(table);
376        return;
377    }
378
379    if (*batch == NULL) {
380        *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
381        if (*batch == NULL) {
382            tlb_remove_table_one(table);
383            return;
384        }
385        (*batch)->nr = 0;
386    }
387    (*batch)->tables[(*batch)->nr++] = table;
388    if ((*batch)->nr == MAX_TABLE_BATCH)
389        tlb_table_flush(tlb);
390}
391
392#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
393
394/*
395 * If a p?d_bad entry is found while walking page tables, report
396 * the error, before resetting entry to p?d_none. Usually (but
397 * very seldom) called out from the p?d_none_or_clear_bad macros.
398 */
399
400void pgd_clear_bad(pgd_t *pgd)
401{
402    pgd_ERROR(*pgd);
403    pgd_clear(pgd);
404}
405
406void pud_clear_bad(pud_t *pud)
407{
408    pud_ERROR(*pud);
409    pud_clear(pud);
410}
411
412void pmd_clear_bad(pmd_t *pmd)
413{
414    pmd_ERROR(*pmd);
415    pmd_clear(pmd);
416}
417
418/*
419 * Note: this doesn't free the actual pages themselves. That
420 * has been handled earlier when unmapping all the memory regions.
421 */
422static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
423               unsigned long addr)
424{
425    pgtable_t token = pmd_pgtable(*pmd);
426    pmd_clear(pmd);
427    pte_free_tlb(tlb, token, addr);
428    tlb->mm->nr_ptes--;
429}
430
431static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
432                unsigned long addr, unsigned long end,
433                unsigned long floor, unsigned long ceiling)
434{
435    pmd_t *pmd;
436    unsigned long next;
437    unsigned long start;
438
439    start = addr;
440    pmd = pmd_offset(pud, addr);
441    do {
442        next = pmd_addr_end(addr, end);
443        if (pmd_none_or_clear_bad(pmd))
444            continue;
445        free_pte_range(tlb, pmd, addr);
446    } while (pmd++, addr = next, addr != end);
447
448    start &= PUD_MASK;
449    if (start < floor)
450        return;
451    if (ceiling) {
452        ceiling &= PUD_MASK;
453        if (!ceiling)
454            return;
455    }
456    if (end - 1 > ceiling - 1)
457        return;
458
459    pmd = pmd_offset(pud, start);
460    pud_clear(pud);
461    pmd_free_tlb(tlb, pmd, start);
462}
463
464static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
465                unsigned long addr, unsigned long end,
466                unsigned long floor, unsigned long ceiling)
467{
468    pud_t *pud;
469    unsigned long next;
470    unsigned long start;
471
472    start = addr;
473    pud = pud_offset(pgd, addr);
474    do {
475        next = pud_addr_end(addr, end);
476        if (pud_none_or_clear_bad(pud))
477            continue;
478        free_pmd_range(tlb, pud, addr, next, floor, ceiling);
479    } while (pud++, addr = next, addr != end);
480
481    start &= PGDIR_MASK;
482    if (start < floor)
483        return;
484    if (ceiling) {
485        ceiling &= PGDIR_MASK;
486        if (!ceiling)
487            return;
488    }
489    if (end - 1 > ceiling - 1)
490        return;
491
492    pud = pud_offset(pgd, start);
493    pgd_clear(pgd);
494    pud_free_tlb(tlb, pud, start);
495}
496
497/*
498 * This function frees user-level page tables of a process.
499 *
500 * Must be called with pagetable lock held.
501 */
502void free_pgd_range(struct mmu_gather *tlb,
503            unsigned long addr, unsigned long end,
504            unsigned long floor, unsigned long ceiling)
505{
506    pgd_t *pgd;
507    unsigned long next;
508
509    /*
510     * The next few lines have given us lots of grief...
511     *
512     * Why are we testing PMD* at this top level? Because often
513     * there will be no work to do at all, and we'd prefer not to
514     * go all the way down to the bottom just to discover that.
515     *
516     * Why all these "- 1"s? Because 0 represents both the bottom
517     * of the address space and the top of it (using -1 for the
518     * top wouldn't help much: the masks would do the wrong thing).
519     * The rule is that addr 0 and floor 0 refer to the bottom of
520     * the address space, but end 0 and ceiling 0 refer to the top
521     * Comparisons need to use "end - 1" and "ceiling - 1" (though
522     * that end 0 case should be mythical).
523     *
524     * Wherever addr is brought up or ceiling brought down, we must
525     * be careful to reject "the opposite 0" before it confuses the
526     * subsequent tests. But what about where end is brought down
527     * by PMD_SIZE below? no, end can't go down to 0 there.
528     *
529     * Whereas we round start (addr) and ceiling down, by different
530     * masks at different levels, in order to test whether a table
531     * now has no other vmas using it, so can be freed, we don't
532     * bother to round floor or end up - the tests don't need that.
533     */
534
535    addr &= PMD_MASK;
536    if (addr < floor) {
537        addr += PMD_SIZE;
538        if (!addr)
539            return;
540    }
541    if (ceiling) {
542        ceiling &= PMD_MASK;
543        if (!ceiling)
544            return;
545    }
546    if (end - 1 > ceiling - 1)
547        end -= PMD_SIZE;
548    if (addr > end - 1)
549        return;
550
551    pgd = pgd_offset(tlb->mm, addr);
552    do {
553        next = pgd_addr_end(addr, end);
554        if (pgd_none_or_clear_bad(pgd))
555            continue;
556        free_pud_range(tlb, pgd, addr, next, floor, ceiling);
557    } while (pgd++, addr = next, addr != end);
558}
559
560void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
561        unsigned long floor, unsigned long ceiling)
562{
563    while (vma) {
564        struct vm_area_struct *next = vma->vm_next;
565        unsigned long addr = vma->vm_start;
566
567        /*
568         * Hide vma from rmap and truncate_pagecache before freeing
569         * pgtables
570         */
571        unlink_anon_vmas(vma);
572        unlink_file_vma(vma);
573
574        if (is_vm_hugetlb_page(vma)) {
575            hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
576                floor, next? next->vm_start: ceiling);
577        } else {
578            /*
579             * Optimization: gather nearby vmas into one call down
580             */
581            while (next && next->vm_start <= vma->vm_end + PMD_SIZE
582                   && !is_vm_hugetlb_page(next)) {
583                vma = next;
584                next = vma->vm_next;
585                unlink_anon_vmas(vma);
586                unlink_file_vma(vma);
587            }
588            free_pgd_range(tlb, addr, vma->vm_end,
589                floor, next? next->vm_start: ceiling);
590        }
591        vma = next;
592    }
593}
594
595int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
596        pmd_t *pmd, unsigned long address)
597{
598    pgtable_t new = pte_alloc_one(mm, address);
599    int wait_split_huge_page;
600    if (!new)
601        return -ENOMEM;
602
603    /*
604     * Ensure all pte setup (eg. pte page lock and page clearing) are
605     * visible before the pte is made visible to other CPUs by being
606     * put into page tables.
607     *
608     * The other side of the story is the pointer chasing in the page
609     * table walking code (when walking the page table without locking;
610     * ie. most of the time). Fortunately, these data accesses consist
611     * of a chain of data-dependent loads, meaning most CPUs (alpha
612     * being the notable exception) will already guarantee loads are
613     * seen in-order. See the alpha page table accessors for the
614     * smp_read_barrier_depends() barriers in page table walking code.
615     */
616    smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
617
618    spin_lock(&mm->page_table_lock);
619    wait_split_huge_page = 0;
620    if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
621        mm->nr_ptes++;
622        pmd_populate(mm, pmd, new);
623        new = NULL;
624    } else if (unlikely(pmd_trans_splitting(*pmd)))
625        wait_split_huge_page = 1;
626    spin_unlock(&mm->page_table_lock);
627    if (new)
628        pte_free(mm, new);
629    if (wait_split_huge_page)
630        wait_split_huge_page(vma->anon_vma, pmd);
631    return 0;
632}
633
634int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
635{
636    pte_t *new = pte_alloc_one_kernel(&init_mm, address);
637    if (!new)
638        return -ENOMEM;
639
640    smp_wmb(); /* See comment in __pte_alloc */
641
642    spin_lock(&init_mm.page_table_lock);
643    if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
644        pmd_populate_kernel(&init_mm, pmd, new);
645        new = NULL;
646    } else
647        VM_BUG_ON(pmd_trans_splitting(*pmd));
648    spin_unlock(&init_mm.page_table_lock);
649    if (new)
650        pte_free_kernel(&init_mm, new);
651    return 0;
652}
653
654static inline void init_rss_vec(int *rss)
655{
656    memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
657}
658
659static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
660{
661    int i;
662
663    if (current->mm == mm)
664        sync_mm_rss(current, mm);
665    for (i = 0; i < NR_MM_COUNTERS; i++)
666        if (rss[i])
667            add_mm_counter(mm, i, rss[i]);
668}
669
670/*
671 * This function is called to print an error when a bad pte
672 * is found. For example, we might have a PFN-mapped pte in
673 * a region that doesn't allow it.
674 *
675 * The calling function must still handle the error.
676 */
677static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
678              pte_t pte, struct page *page)
679{
680    pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
681    pud_t *pud = pud_offset(pgd, addr);
682    pmd_t *pmd = pmd_offset(pud, addr);
683    struct address_space *mapping;
684    pgoff_t index;
685    static unsigned long resume;
686    static unsigned long nr_shown;
687    static unsigned long nr_unshown;
688
689    /*
690     * Allow a burst of 60 reports, then keep quiet for that minute;
691     * or allow a steady drip of one report per second.
692     */
693    if (nr_shown == 60) {
694        if (time_before(jiffies, resume)) {
695            nr_unshown++;
696            return;
697        }
698        if (nr_unshown) {
699            printk(KERN_ALERT
700                "BUG: Bad page map: %lu messages suppressed\n",
701                nr_unshown);
702            nr_unshown = 0;
703        }
704        nr_shown = 0;
705    }
706    if (nr_shown++ == 0)
707        resume = jiffies + 60 * HZ;
708
709    mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
710    index = linear_page_index(vma, addr);
711
712    printk(KERN_ALERT
713        "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
714        current->comm,
715        (long long)pte_val(pte), (long long)pmd_val(*pmd));
716    if (page)
717        dump_page(page);
718    printk(KERN_ALERT
719        "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
720        (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
721    /*
722     * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
723     */
724    if (vma->vm_ops)
725        print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
726                (unsigned long)vma->vm_ops->fault);
727    if (vma->vm_file && vma->vm_file->f_op)
728        print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
729                (unsigned long)vma->vm_file->f_op->mmap);
730    dump_stack();
731    add_taint(TAINT_BAD_PAGE);
732}
733
734static inline int is_cow_mapping(vm_flags_t flags)
735{
736    return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
737}
738
739#ifndef is_zero_pfn
740static inline int is_zero_pfn(unsigned long pfn)
741{
742    return pfn == zero_pfn;
743}
744#endif
745
746#ifndef my_zero_pfn
747static inline unsigned long my_zero_pfn(unsigned long addr)
748{
749    return zero_pfn;
750}
751#endif
752
753/*
754 * vm_normal_page -- This function gets the "struct page" associated with a pte.
755 *
756 * "Special" mappings do not wish to be associated with a "struct page" (either
757 * it doesn't exist, or it exists but they don't want to touch it). In this
758 * case, NULL is returned here. "Normal" mappings do have a struct page.
759 *
760 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
761 * pte bit, in which case this function is trivial. Secondly, an architecture
762 * may not have a spare pte bit, which requires a more complicated scheme,
763 * described below.
764 *
765 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
766 * special mapping (even if there are underlying and valid "struct pages").
767 * COWed pages of a VM_PFNMAP are always normal.
768 *
769 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
770 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
771 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
772 * mapping will always honor the rule
773 *
774 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
775 *
776 * And for normal mappings this is false.
777 *
778 * This restricts such mappings to be a linear translation from virtual address
779 * to pfn. To get around this restriction, we allow arbitrary mappings so long
780 * as the vma is not a COW mapping; in that case, we know that all ptes are
781 * special (because none can have been COWed).
782 *
783 *
784 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
785 *
786 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
787 * page" backing, however the difference is that _all_ pages with a struct
788 * page (that is, those where pfn_valid is true) are refcounted and considered
789 * normal pages by the VM. The disadvantage is that pages are refcounted
790 * (which can be slower and simply not an option for some PFNMAP users). The
791 * advantage is that we don't have to follow the strict linearity rule of
792 * PFNMAP mappings in order to support COWable mappings.
793 *
794 */
795#ifdef __HAVE_ARCH_PTE_SPECIAL
796# define HAVE_PTE_SPECIAL 1
797#else
798# define HAVE_PTE_SPECIAL 0
799#endif
800struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
801                pte_t pte)
802{
803    unsigned long pfn = pte_pfn(pte);
804
805    if (HAVE_PTE_SPECIAL) {
806        if (likely(!pte_special(pte)))
807            goto check_pfn;
808        if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
809            return NULL;
810        if (!is_zero_pfn(pfn))
811            print_bad_pte(vma, addr, pte, NULL);
812        return NULL;
813    }
814
815    /* !HAVE_PTE_SPECIAL case follows: */
816
817    if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
818        if (vma->vm_flags & VM_MIXEDMAP) {
819            if (!pfn_valid(pfn))
820                return NULL;
821            goto out;
822        } else {
823            unsigned long off;
824            off = (addr - vma->vm_start) >> PAGE_SHIFT;
825            if (pfn == vma->vm_pgoff + off)
826                return NULL;
827            if (!is_cow_mapping(vma->vm_flags))
828                return NULL;
829        }
830    }
831
832    if (is_zero_pfn(pfn))
833        return NULL;
834check_pfn:
835    if (unlikely(pfn > highest_memmap_pfn)) {
836        print_bad_pte(vma, addr, pte, NULL);
837        return NULL;
838    }
839
840    /*
841     * NOTE! We still have PageReserved() pages in the page tables.
842     * eg. VDSO mappings can cause them to exist.
843     */
844out:
845    return pfn_to_page(pfn);
846}
847
848/*
849 * copy one vm_area from one task to the other. Assumes the page tables
850 * already present in the new task to be cleared in the whole range
851 * covered by this vma.
852 */
853
854static inline unsigned long
855copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
856        pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
857        unsigned long addr, int *rss)
858{
859    unsigned long vm_flags = vma->vm_flags;
860    pte_t pte = *src_pte;
861    struct page *page;
862
863    /* pte contains position in swap or file, so copy. */
864    if (unlikely(!pte_present(pte))) {
865        if (!pte_file(pte)) {
866            swp_entry_t entry = pte_to_swp_entry(pte);
867
868            if (swap_duplicate(entry) < 0)
869                return entry.val;
870
871            /* make sure dst_mm is on swapoff's mmlist. */
872            if (unlikely(list_empty(&dst_mm->mmlist))) {
873                spin_lock(&mmlist_lock);
874                if (list_empty(&dst_mm->mmlist))
875                    list_add(&dst_mm->mmlist,
876                         &src_mm->mmlist);
877                spin_unlock(&mmlist_lock);
878            }
879            if (likely(!non_swap_entry(entry)))
880                rss[MM_SWAPENTS]++;
881            else if (is_write_migration_entry(entry) &&
882                    is_cow_mapping(vm_flags)) {
883                /*
884                 * COW mappings require pages in both parent
885                 * and child to be set to read.
886                 */
887                make_migration_entry_read(&entry);
888                pte = swp_entry_to_pte(entry);
889                set_pte_at(src_mm, addr, src_pte, pte);
890            }
891        }
892        goto out_set_pte;
893    }
894
895    /*
896     * If it's a COW mapping, write protect it both
897     * in the parent and the child
898     */
899    if (is_cow_mapping(vm_flags)) {
900        ptep_set_wrprotect(src_mm, addr, src_pte);
901        pte = pte_wrprotect(pte);
902    }
903
904    /*
905     * If it's a shared mapping, mark it clean in
906     * the child
907     */
908    if (vm_flags & VM_SHARED)
909        pte = pte_mkclean(pte);
910    pte = pte_mkold(pte);
911
912    page = vm_normal_page(vma, addr, pte);
913    if (page) {
914        get_page(page);
915        page_dup_rmap(page);
916        if (PageAnon(page))
917            rss[MM_ANONPAGES]++;
918        else
919            rss[MM_FILEPAGES]++;
920    }
921
922out_set_pte:
923    set_pte_at(dst_mm, addr, dst_pte, pte);
924    return 0;
925}
926
927int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
928           pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
929           unsigned long addr, unsigned long end)
930{
931    pte_t *orig_src_pte, *orig_dst_pte;
932    pte_t *src_pte, *dst_pte;
933    spinlock_t *src_ptl, *dst_ptl;
934    int progress = 0;
935    int rss[NR_MM_COUNTERS];
936    swp_entry_t entry = (swp_entry_t){0};
937
938again:
939    init_rss_vec(rss);
940
941    dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
942    if (!dst_pte)
943        return -ENOMEM;
944    src_pte = pte_offset_map(src_pmd, addr);
945    src_ptl = pte_lockptr(src_mm, src_pmd);
946    spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
947    orig_src_pte = src_pte;
948    orig_dst_pte = dst_pte;
949    arch_enter_lazy_mmu_mode();
950
951    do {
952        /*
953         * We are holding two locks at this point - either of them
954         * could generate latencies in another task on another CPU.
955         */
956        if (progress >= 32) {
957            progress = 0;
958            if (need_resched() ||
959                spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
960                break;
961        }
962        if (pte_none(*src_pte)) {
963            progress++;
964            continue;
965        }
966        entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
967                            vma, addr, rss);
968        if (entry.val)
969            break;
970        progress += 8;
971    } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
972
973    arch_leave_lazy_mmu_mode();
974    spin_unlock(src_ptl);
975    pte_unmap(orig_src_pte);
976    add_mm_rss_vec(dst_mm, rss);
977    pte_unmap_unlock(orig_dst_pte, dst_ptl);
978    cond_resched();
979
980    if (entry.val) {
981        if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
982            return -ENOMEM;
983        progress = 0;
984    }
985    if (addr != end)
986        goto again;
987    return 0;
988}
989
990static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
991        pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
992        unsigned long addr, unsigned long end)
993{
994    pmd_t *src_pmd, *dst_pmd;
995    unsigned long next;
996
997    dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
998    if (!dst_pmd)
999        return -ENOMEM;
1000    src_pmd = pmd_offset(src_pud, addr);
1001    do {
1002        next = pmd_addr_end(addr, end);
1003        if (pmd_trans_huge(*src_pmd)) {
1004            int err;
1005            VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1006            err = copy_huge_pmd(dst_mm, src_mm,
1007                        dst_pmd, src_pmd, addr, vma);
1008            if (err == -ENOMEM)
1009                return -ENOMEM;
1010            if (!err)
1011                continue;
1012            /* fall through */
1013        }
1014        if (pmd_none_or_clear_bad(src_pmd))
1015            continue;
1016        if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1017                        vma, addr, next))
1018            return -ENOMEM;
1019    } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1020    return 0;
1021}
1022
1023static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1024        pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1025        unsigned long addr, unsigned long end)
1026{
1027    pud_t *src_pud, *dst_pud;
1028    unsigned long next;
1029
1030    dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1031    if (!dst_pud)
1032        return -ENOMEM;
1033    src_pud = pud_offset(src_pgd, addr);
1034    do {
1035        next = pud_addr_end(addr, end);
1036        if (pud_none_or_clear_bad(src_pud))
1037            continue;
1038        if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1039                        vma, addr, next))
1040            return -ENOMEM;
1041    } while (dst_pud++, src_pud++, addr = next, addr != end);
1042    return 0;
1043}
1044
1045int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1046        struct vm_area_struct *vma)
1047{
1048    pgd_t *src_pgd, *dst_pgd;
1049    unsigned long next;
1050    unsigned long addr = vma->vm_start;
1051    unsigned long end = vma->vm_end;
1052    int ret;
1053
1054    /*
1055     * Don't copy ptes where a page fault will fill them correctly.
1056     * Fork becomes much lighter when there are big shared or private
1057     * readonly mappings. The tradeoff is that copy_page_range is more
1058     * efficient than faulting.
1059     */
1060    if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1061        if (!vma->anon_vma)
1062            return 0;
1063    }
1064
1065    if (is_vm_hugetlb_page(vma))
1066        return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1067
1068    if (unlikely(is_pfn_mapping(vma))) {
1069        /*
1070         * We do not free on error cases below as remove_vma
1071         * gets called on error from higher level routine
1072         */
1073        ret = track_pfn_vma_copy(vma);
1074        if (ret)
1075            return ret;
1076    }
1077
1078    /*
1079     * We need to invalidate the secondary MMU mappings only when
1080     * there could be a permission downgrade on the ptes of the
1081     * parent mm. And a permission downgrade will only happen if
1082     * is_cow_mapping() returns true.
1083     */
1084    if (is_cow_mapping(vma->vm_flags))
1085        mmu_notifier_invalidate_range_start(src_mm, addr, end);
1086
1087    ret = 0;
1088    dst_pgd = pgd_offset(dst_mm, addr);
1089    src_pgd = pgd_offset(src_mm, addr);
1090    do {
1091        next = pgd_addr_end(addr, end);
1092        if (pgd_none_or_clear_bad(src_pgd))
1093            continue;
1094        if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1095                        vma, addr, next))) {
1096            ret = -ENOMEM;
1097            break;
1098        }
1099    } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1100
1101    if (is_cow_mapping(vma->vm_flags))
1102        mmu_notifier_invalidate_range_end(src_mm,
1103                          vma->vm_start, end);
1104    return ret;
1105}
1106
1107static unsigned long zap_pte_range(struct mmu_gather *tlb,
1108                struct vm_area_struct *vma, pmd_t *pmd,
1109                unsigned long addr, unsigned long end,
1110                struct zap_details *details)
1111{
1112    struct mm_struct *mm = tlb->mm;
1113    int force_flush = 0;
1114    int rss[NR_MM_COUNTERS];
1115    spinlock_t *ptl;
1116    pte_t *start_pte;
1117    pte_t *pte;
1118
1119again:
1120    init_rss_vec(rss);
1121    start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1122    pte = start_pte;
1123    arch_enter_lazy_mmu_mode();
1124    do {
1125        pte_t ptent = *pte;
1126        if (pte_none(ptent)) {
1127            continue;
1128        }
1129
1130        if (pte_present(ptent)) {
1131            struct page *page;
1132
1133            page = vm_normal_page(vma, addr, ptent);
1134            if (unlikely(details) && page) {
1135                /*
1136                 * unmap_shared_mapping_pages() wants to
1137                 * invalidate cache without truncating:
1138                 * unmap shared but keep private pages.
1139                 */
1140                if (details->check_mapping &&
1141                    details->check_mapping != page->mapping)
1142                    continue;
1143                /*
1144                 * Each page->index must be checked when
1145                 * invalidating or truncating nonlinear.
1146                 */
1147                if (details->nonlinear_vma &&
1148                    (page->index < details->first_index ||
1149                     page->index > details->last_index))
1150                    continue;
1151            }
1152            ptent = ptep_get_and_clear_full(mm, addr, pte,
1153                            tlb->fullmm);
1154            tlb_remove_tlb_entry(tlb, pte, addr);
1155            if (unlikely(!page))
1156                continue;
1157            if (unlikely(details) && details->nonlinear_vma
1158                && linear_page_index(details->nonlinear_vma,
1159                        addr) != page->index)
1160                set_pte_at(mm, addr, pte,
1161                       pgoff_to_pte(page->index));
1162            if (PageAnon(page))
1163                rss[MM_ANONPAGES]--;
1164            else {
1165                if (pte_dirty(ptent))
1166                    set_page_dirty(page);
1167                if (pte_young(ptent) &&
1168                    likely(!VM_SequentialReadHint(vma)))
1169                    mark_page_accessed(page);
1170                rss[MM_FILEPAGES]--;
1171            }
1172            page_remove_rmap(page);
1173            if (unlikely(page_mapcount(page) < 0))
1174                print_bad_pte(vma, addr, ptent, page);
1175            force_flush = !__tlb_remove_page(tlb, page);
1176            if (force_flush)
1177                break;
1178            continue;
1179        }
1180        /*
1181         * If details->check_mapping, we leave swap entries;
1182         * if details->nonlinear_vma, we leave file entries.
1183         */
1184        if (unlikely(details))
1185            continue;
1186        if (pte_file(ptent)) {
1187            if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1188                print_bad_pte(vma, addr, ptent, NULL);
1189        } else {
1190            swp_entry_t entry = pte_to_swp_entry(ptent);
1191
1192            if (!non_swap_entry(entry))
1193                rss[MM_SWAPENTS]--;
1194            if (unlikely(!free_swap_and_cache(entry)))
1195                print_bad_pte(vma, addr, ptent, NULL);
1196        }
1197        pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198    } while (pte++, addr += PAGE_SIZE, addr != end);
1199
1200    add_mm_rss_vec(mm, rss);
1201    arch_leave_lazy_mmu_mode();
1202    pte_unmap_unlock(start_pte, ptl);
1203
1204    /*
1205     * mmu_gather ran out of room to batch pages, we break out of
1206     * the PTE lock to avoid doing the potential expensive TLB invalidate
1207     * and page-free while holding it.
1208     */
1209    if (force_flush) {
1210        force_flush = 0;
1211        tlb_flush_mmu(tlb);
1212        if (addr != end)
1213            goto again;
1214    }
1215
1216    return addr;
1217}
1218
1219static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1220                struct vm_area_struct *vma, pud_t *pud,
1221                unsigned long addr, unsigned long end,
1222                struct zap_details *details)
1223{
1224    pmd_t *pmd;
1225    unsigned long next;
1226
1227    pmd = pmd_offset(pud, addr);
1228    do {
1229        next = pmd_addr_end(addr, end);
1230        if (pmd_trans_huge(*pmd)) {
1231            if (next-addr != HPAGE_PMD_SIZE) {
1232                VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1233                split_huge_page_pmd(vma->vm_mm, pmd);
1234            } else if (zap_huge_pmd(tlb, vma, pmd))
1235                continue;
1236            /* fall through */
1237        }
1238        if (pmd_none_or_clear_bad(pmd))
1239            continue;
1240        next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1241        cond_resched();
1242    } while (pmd++, addr = next, addr != end);
1243
1244    return addr;
1245}
1246
1247static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1248                struct vm_area_struct *vma, pgd_t *pgd,
1249                unsigned long addr, unsigned long end,
1250                struct zap_details *details)
1251{
1252    pud_t *pud;
1253    unsigned long next;
1254
1255    pud = pud_offset(pgd, addr);
1256    do {
1257        next = pud_addr_end(addr, end);
1258        if (pud_none_or_clear_bad(pud))
1259            continue;
1260        next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1261    } while (pud++, addr = next, addr != end);
1262
1263    return addr;
1264}
1265
1266static unsigned long unmap_page_range(struct mmu_gather *tlb,
1267                struct vm_area_struct *vma,
1268                unsigned long addr, unsigned long end,
1269                struct zap_details *details)
1270{
1271    pgd_t *pgd;
1272    unsigned long next;
1273
1274    if (details && !details->check_mapping && !details->nonlinear_vma)
1275        details = NULL;
1276
1277    BUG_ON(addr >= end);
1278    mem_cgroup_uncharge_start();
1279    tlb_start_vma(tlb, vma);
1280    pgd = pgd_offset(vma->vm_mm, addr);
1281    do {
1282        next = pgd_addr_end(addr, end);
1283        if (pgd_none_or_clear_bad(pgd))
1284            continue;
1285        next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1286    } while (pgd++, addr = next, addr != end);
1287    tlb_end_vma(tlb, vma);
1288    mem_cgroup_uncharge_end();
1289
1290    return addr;
1291}
1292
1293#ifdef CONFIG_PREEMPT
1294# define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1295#else
1296/* No preempt: go for improved straight-line efficiency */
1297# define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1298#endif
1299
1300/**
1301 * unmap_vmas - unmap a range of memory covered by a list of vma's
1302 * @tlb: address of the caller's struct mmu_gather
1303 * @vma: the starting vma
1304 * @start_addr: virtual address at which to start unmapping
1305 * @end_addr: virtual address at which to end unmapping
1306 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1307 * @details: details of nonlinear truncation or shared cache invalidation
1308 *
1309 * Returns the end address of the unmapping (restart addr if interrupted).
1310 *
1311 * Unmap all pages in the vma list.
1312 *
1313 * We aim to not hold locks for too long (for scheduling latency reasons).
1314 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
1315 * return the ending mmu_gather to the caller.
1316 *
1317 * Only addresses between `start' and `end' will be unmapped.
1318 *
1319 * The VMA list must be sorted in ascending virtual address order.
1320 *
1321 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1322 * range after unmap_vmas() returns. So the only responsibility here is to
1323 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1324 * drops the lock and schedules.
1325 */
1326unsigned long unmap_vmas(struct mmu_gather *tlb,
1327        struct vm_area_struct *vma, unsigned long start_addr,
1328        unsigned long end_addr, unsigned long *nr_accounted,
1329        struct zap_details *details)
1330{
1331    unsigned long start = start_addr;
1332    struct mm_struct *mm = vma->vm_mm;
1333
1334    mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1335    for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1336        unsigned long end;
1337
1338        start = max(vma->vm_start, start_addr);
1339        if (start >= vma->vm_end)
1340            continue;
1341        end = min(vma->vm_end, end_addr);
1342        if (end <= vma->vm_start)
1343            continue;
1344
1345        if (vma->vm_flags & VM_ACCOUNT)
1346            *nr_accounted += (end - start) >> PAGE_SHIFT;
1347
1348        if (unlikely(is_pfn_mapping(vma)))
1349            untrack_pfn_vma(vma, 0, 0);
1350
1351        while (start != end) {
1352            if (unlikely(is_vm_hugetlb_page(vma))) {
1353                /*
1354                 * It is undesirable to test vma->vm_file as it
1355                 * should be non-null for valid hugetlb area.
1356                 * However, vm_file will be NULL in the error
1357                 * cleanup path of do_mmap_pgoff. When
1358                 * hugetlbfs ->mmap method fails,
1359                 * do_mmap_pgoff() nullifies vma->vm_file
1360                 * before calling this function to clean up.
1361                 * Since no pte has actually been setup, it is
1362                 * safe to do nothing in this case.
1363                 */
1364                if (vma->vm_file)
1365                    unmap_hugepage_range(vma, start, end, NULL);
1366
1367                start = end;
1368            } else
1369                start = unmap_page_range(tlb, vma, start, end, details);
1370        }
1371    }
1372
1373    mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1374    return start; /* which is now the end (or restart) address */
1375}
1376
1377/**
1378 * zap_page_range - remove user pages in a given range
1379 * @vma: vm_area_struct holding the applicable pages
1380 * @address: starting address of pages to zap
1381 * @size: number of bytes to zap
1382 * @details: details of nonlinear truncation or shared cache invalidation
1383 */
1384unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1385        unsigned long size, struct zap_details *details)
1386{
1387    struct mm_struct *mm = vma->vm_mm;
1388    struct mmu_gather tlb;
1389    unsigned long end = address + size;
1390    unsigned long nr_accounted = 0;
1391
1392    lru_add_drain();
1393    tlb_gather_mmu(&tlb, mm, 0);
1394    update_hiwater_rss(mm);
1395    end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1396    tlb_finish_mmu(&tlb, address, end);
1397    return end;
1398}
1399
1400/**
1401 * zap_vma_ptes - remove ptes mapping the vma
1402 * @vma: vm_area_struct holding ptes to be zapped
1403 * @address: starting address of pages to zap
1404 * @size: number of bytes to zap
1405 *
1406 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1407 *
1408 * The entire address range must be fully contained within the vma.
1409 *
1410 * Returns 0 if successful.
1411 */
1412int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1413        unsigned long size)
1414{
1415    if (address < vma->vm_start || address + size > vma->vm_end ||
1416                !(vma->vm_flags & VM_PFNMAP))
1417        return -1;
1418    zap_page_range(vma, address, size, NULL);
1419    return 0;
1420}
1421EXPORT_SYMBOL_GPL(zap_vma_ptes);
1422
1423/**
1424 * follow_page - look up a page descriptor from a user-virtual address
1425 * @vma: vm_area_struct mapping @address
1426 * @address: virtual address to look up
1427 * @flags: flags modifying lookup behaviour
1428 *
1429 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1430 *
1431 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1432 * an error pointer if there is a mapping to something not represented
1433 * by a page descriptor (see also vm_normal_page()).
1434 */
1435struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1436            unsigned int flags)
1437{
1438    pgd_t *pgd;
1439    pud_t *pud;
1440    pmd_t *pmd;
1441    pte_t *ptep, pte;
1442    spinlock_t *ptl;
1443    struct page *page;
1444    struct mm_struct *mm = vma->vm_mm;
1445
1446    page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1447    if (!IS_ERR(page)) {
1448        BUG_ON(flags & FOLL_GET);
1449        goto out;
1450    }
1451
1452    page = NULL;
1453    pgd = pgd_offset(mm, address);
1454    if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1455        goto no_page_table;
1456
1457    pud = pud_offset(pgd, address);
1458    if (pud_none(*pud))
1459        goto no_page_table;
1460    if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1461        BUG_ON(flags & FOLL_GET);
1462        page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1463        goto out;
1464    }
1465    if (unlikely(pud_bad(*pud)))
1466        goto no_page_table;
1467
1468    pmd = pmd_offset(pud, address);
1469    if (pmd_none(*pmd))
1470        goto no_page_table;
1471    if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1472        BUG_ON(flags & FOLL_GET);
1473        page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1474        goto out;
1475    }
1476    if (pmd_trans_huge(*pmd)) {
1477        if (flags & FOLL_SPLIT) {
1478            split_huge_page_pmd(mm, pmd);
1479            goto split_fallthrough;
1480        }
1481        spin_lock(&mm->page_table_lock);
1482        if (likely(pmd_trans_huge(*pmd))) {
1483            if (unlikely(pmd_trans_splitting(*pmd))) {
1484                spin_unlock(&mm->page_table_lock);
1485                wait_split_huge_page(vma->anon_vma, pmd);
1486            } else {
1487                page = follow_trans_huge_pmd(mm, address,
1488                                 pmd, flags);
1489                spin_unlock(&mm->page_table_lock);
1490                goto out;
1491            }
1492        } else
1493            spin_unlock(&mm->page_table_lock);
1494        /* fall through */
1495    }
1496split_fallthrough:
1497    if (unlikely(pmd_bad(*pmd)))
1498        goto no_page_table;
1499
1500    ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1501
1502    pte = *ptep;
1503    if (!pte_present(pte))
1504        goto no_page;
1505    if ((flags & FOLL_WRITE) && !pte_write(pte))
1506        goto unlock;
1507
1508    page = vm_normal_page(vma, address, pte);
1509    if (unlikely(!page)) {
1510        if ((flags & FOLL_DUMP) ||
1511            !is_zero_pfn(pte_pfn(pte)))
1512            goto bad_page;
1513        page = pte_page(pte);
1514    }
1515
1516    if (flags & FOLL_GET)
1517        get_page(page);
1518    if (flags & FOLL_TOUCH) {
1519        if ((flags & FOLL_WRITE) &&
1520            !pte_dirty(pte) && !PageDirty(page))
1521            set_page_dirty(page);
1522        /*
1523         * pte_mkyoung() would be more correct here, but atomic care
1524         * is needed to avoid losing the dirty bit: it is easier to use
1525         * mark_page_accessed().
1526         */
1527        mark_page_accessed(page);
1528    }
1529    if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1530        /*
1531         * The preliminary mapping check is mainly to avoid the
1532         * pointless overhead of lock_page on the ZERO_PAGE
1533         * which might bounce very badly if there is contention.
1534         *
1535         * If the page is already locked, we don't need to
1536         * handle it now - vmscan will handle it later if and
1537         * when it attempts to reclaim the page.
1538         */
1539        if (page->mapping && trylock_page(page)) {
1540            lru_add_drain(); /* push cached pages to LRU */
1541            /*
1542             * Because we lock page here and migration is
1543             * blocked by the pte's page reference, we need
1544             * only check for file-cache page truncation.
1545             */
1546            if (page->mapping)
1547                mlock_vma_page(page);
1548            unlock_page(page);
1549        }
1550    }
1551unlock:
1552    pte_unmap_unlock(ptep, ptl);
1553out:
1554    return page;
1555
1556bad_page:
1557    pte_unmap_unlock(ptep, ptl);
1558    return ERR_PTR(-EFAULT);
1559
1560no_page:
1561    pte_unmap_unlock(ptep, ptl);
1562    if (!pte_none(pte))
1563        return page;
1564
1565no_page_table:
1566    /*
1567     * When core dumping an enormous anonymous area that nobody
1568     * has touched so far, we don't want to allocate unnecessary pages or
1569     * page tables. Return error instead of NULL to skip handle_mm_fault,
1570     * then get_dump_page() will return NULL to leave a hole in the dump.
1571     * But we can only make this optimization where a hole would surely
1572     * be zero-filled if handle_mm_fault() actually did handle it.
1573     */
1574    if ((flags & FOLL_DUMP) &&
1575        (!vma->vm_ops || !vma->vm_ops->fault))
1576        return ERR_PTR(-EFAULT);
1577    return page;
1578}
1579
1580static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1581{
1582    return stack_guard_page_start(vma, addr) ||
1583           stack_guard_page_end(vma, addr+PAGE_SIZE);
1584}
1585
1586/**
1587 * __get_user_pages() - pin user pages in memory
1588 * @tsk: task_struct of target task
1589 * @mm: mm_struct of target mm
1590 * @start: starting user address
1591 * @nr_pages: number of pages from start to pin
1592 * @gup_flags: flags modifying pin behaviour
1593 * @pages: array that receives pointers to the pages pinned.
1594 * Should be at least nr_pages long. Or NULL, if caller
1595 * only intends to ensure the pages are faulted in.
1596 * @vmas: array of pointers to vmas corresponding to each page.
1597 * Or NULL if the caller does not require them.
1598 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1599 *
1600 * Returns number of pages pinned. This may be fewer than the number
1601 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1602 * were pinned, returns -errno. Each page returned must be released
1603 * with a put_page() call when it is finished with. vmas will only
1604 * remain valid while mmap_sem is held.
1605 *
1606 * Must be called with mmap_sem held for read or write.
1607 *
1608 * __get_user_pages walks a process's page tables and takes a reference to
1609 * each struct page that each user address corresponds to at a given
1610 * instant. That is, it takes the page that would be accessed if a user
1611 * thread accesses the given user virtual address at that instant.
1612 *
1613 * This does not guarantee that the page exists in the user mappings when
1614 * __get_user_pages returns, and there may even be a completely different
1615 * page there in some cases (eg. if mmapped pagecache has been invalidated
1616 * and subsequently re faulted). However it does guarantee that the page
1617 * won't be freed completely. And mostly callers simply care that the page
1618 * contains data that was valid *at some point in time*. Typically, an IO
1619 * or similar operation cannot guarantee anything stronger anyway because
1620 * locks can't be held over the syscall boundary.
1621 *
1622 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1623 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1624 * appropriate) must be called after the page is finished with, and
1625 * before put_page is called.
1626 *
1627 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1628 * or mmap_sem contention, and if waiting is needed to pin all pages,
1629 * *@nonblocking will be set to 0.
1630 *
1631 * In most cases, get_user_pages or get_user_pages_fast should be used
1632 * instead of __get_user_pages. __get_user_pages should be used only if
1633 * you need some special @gup_flags.
1634 */
1635int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1636             unsigned long start, int nr_pages, unsigned int gup_flags,
1637             struct page **pages, struct vm_area_struct **vmas,
1638             int *nonblocking)
1639{
1640    int i;
1641    unsigned long vm_flags;
1642
1643    if (nr_pages <= 0)
1644        return 0;
1645
1646    VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1647
1648    /*
1649     * Require read or write permissions.
1650     * If FOLL_FORCE is set, we only require the "MAY" flags.
1651     */
1652    vm_flags = (gup_flags & FOLL_WRITE) ?
1653            (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1654    vm_flags &= (gup_flags & FOLL_FORCE) ?
1655            (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1656    i = 0;
1657
1658    do {
1659        struct vm_area_struct *vma;
1660
1661        vma = find_extend_vma(mm, start);
1662        if (!vma && in_gate_area(mm, start)) {
1663            unsigned long pg = start & PAGE_MASK;
1664            pgd_t *pgd;
1665            pud_t *pud;
1666            pmd_t *pmd;
1667            pte_t *pte;
1668
1669            /* user gate pages are read-only */
1670            if (gup_flags & FOLL_WRITE)
1671                return i ? : -EFAULT;
1672            if (pg > TASK_SIZE)
1673                pgd = pgd_offset_k(pg);
1674            else
1675                pgd = pgd_offset_gate(mm, pg);
1676            BUG_ON(pgd_none(*pgd));
1677            pud = pud_offset(pgd, pg);
1678            BUG_ON(pud_none(*pud));
1679            pmd = pmd_offset(pud, pg);
1680            if (pmd_none(*pmd))
1681                return i ? : -EFAULT;
1682            VM_BUG_ON(pmd_trans_huge(*pmd));
1683            pte = pte_offset_map(pmd, pg);
1684            if (pte_none(*pte)) {
1685                pte_unmap(pte);
1686                return i ? : -EFAULT;
1687            }
1688            vma = get_gate_vma(mm);
1689            if (pages) {
1690                struct page *page;
1691
1692                page = vm_normal_page(vma, start, *pte);
1693                if (!page) {
1694                    if (!(gup_flags & FOLL_DUMP) &&
1695                         is_zero_pfn(pte_pfn(*pte)))
1696                        page = pte_page(*pte);
1697                    else {
1698                        pte_unmap(pte);
1699                        return i ? : -EFAULT;
1700                    }
1701                }
1702                pages[i] = page;
1703                get_page(page);
1704            }
1705            pte_unmap(pte);
1706            goto next_page;
1707        }
1708
1709        if (!vma ||
1710            (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1711            !(vm_flags & vma->vm_flags))
1712            return i ? : -EFAULT;
1713
1714        if (is_vm_hugetlb_page(vma)) {
1715            i = follow_hugetlb_page(mm, vma, pages, vmas,
1716                    &start, &nr_pages, i, gup_flags);
1717            continue;
1718        }
1719
1720        do {
1721            struct page *page;
1722            unsigned int foll_flags = gup_flags;
1723
1724            /*
1725             * If we have a pending SIGKILL, don't keep faulting
1726             * pages and potentially allocating memory.
1727             */
1728            if (unlikely(fatal_signal_pending(current)))
1729                return i ? i : -ERESTARTSYS;
1730
1731            cond_resched();
1732            while (!(page = follow_page(vma, start, foll_flags))) {
1733                int ret;
1734                unsigned int fault_flags = 0;
1735
1736                /* For mlock, just skip the stack guard page. */
1737                if (foll_flags & FOLL_MLOCK) {
1738                    if (stack_guard_page(vma, start))
1739                        goto next_page;
1740                }
1741                if (foll_flags & FOLL_WRITE)
1742                    fault_flags |= FAULT_FLAG_WRITE;
1743                if (nonblocking)
1744                    fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1745                if (foll_flags & FOLL_NOWAIT)
1746                    fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1747
1748                ret = handle_mm_fault(mm, vma, start,
1749                            fault_flags);
1750
1751                if (ret & VM_FAULT_ERROR) {
1752                    if (ret & VM_FAULT_OOM)
1753                        return i ? i : -ENOMEM;
1754                    if (ret & (VM_FAULT_HWPOISON |
1755                           VM_FAULT_HWPOISON_LARGE)) {
1756                        if (i)
1757                            return i;
1758                        else if (gup_flags & FOLL_HWPOISON)
1759                            return -EHWPOISON;
1760                        else
1761                            return -EFAULT;
1762                    }
1763                    if (ret & VM_FAULT_SIGBUS)
1764                        return i ? i : -EFAULT;
1765                    BUG();
1766                }
1767
1768                if (tsk) {
1769                    if (ret & VM_FAULT_MAJOR)
1770                        tsk->maj_flt++;
1771                    else
1772                        tsk->min_flt++;
1773                }
1774
1775                if (ret & VM_FAULT_RETRY) {
1776                    if (nonblocking)
1777                        *nonblocking = 0;
1778                    return i;
1779                }
1780
1781                /*
1782                 * The VM_FAULT_WRITE bit tells us that
1783                 * do_wp_page has broken COW when necessary,
1784                 * even if maybe_mkwrite decided not to set
1785                 * pte_write. We can thus safely do subsequent
1786                 * page lookups as if they were reads. But only
1787                 * do so when looping for pte_write is futile:
1788                 * in some cases userspace may also be wanting
1789                 * to write to the gotten user page, which a
1790                 * read fault here might prevent (a readonly
1791                 * page might get reCOWed by userspace write).
1792                 */
1793                if ((ret & VM_FAULT_WRITE) &&
1794                    !(vma->vm_flags & VM_WRITE))
1795                    foll_flags &= ~FOLL_WRITE;
1796
1797                cond_resched();
1798            }
1799            if (IS_ERR(page))
1800                return i ? i : PTR_ERR(page);
1801            if (pages) {
1802                pages[i] = page;
1803
1804                flush_anon_page(vma, page, start);
1805                flush_dcache_page(page);
1806            }
1807next_page:
1808            if (vmas)
1809                vmas[i] = vma;
1810            i++;
1811            start += PAGE_SIZE;
1812            nr_pages--;
1813        } while (nr_pages && start < vma->vm_end);
1814    } while (nr_pages);
1815    return i;
1816}
1817EXPORT_SYMBOL(__get_user_pages);
1818
1819/**
1820 * get_user_pages() - pin user pages in memory
1821 * @tsk: the task_struct to use for page fault accounting, or
1822 * NULL if faults are not to be recorded.
1823 * @mm: mm_struct of target mm
1824 * @start: starting user address
1825 * @nr_pages: number of pages from start to pin
1826 * @write: whether pages will be written to by the caller
1827 * @force: whether to force write access even if user mapping is
1828 * readonly. This will result in the page being COWed even
1829 * in MAP_SHARED mappings. You do not want this.
1830 * @pages: array that receives pointers to the pages pinned.
1831 * Should be at least nr_pages long. Or NULL, if caller
1832 * only intends to ensure the pages are faulted in.
1833 * @vmas: array of pointers to vmas corresponding to each page.
1834 * Or NULL if the caller does not require them.
1835 *
1836 * Returns number of pages pinned. This may be fewer than the number
1837 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1838 * were pinned, returns -errno. Each page returned must be released
1839 * with a put_page() call when it is finished with. vmas will only
1840 * remain valid while mmap_sem is held.
1841 *
1842 * Must be called with mmap_sem held for read or write.
1843 *
1844 * get_user_pages walks a process's page tables and takes a reference to
1845 * each struct page that each user address corresponds to at a given
1846 * instant. That is, it takes the page that would be accessed if a user
1847 * thread accesses the given user virtual address at that instant.
1848 *
1849 * This does not guarantee that the page exists in the user mappings when
1850 * get_user_pages returns, and there may even be a completely different
1851 * page there in some cases (eg. if mmapped pagecache has been invalidated
1852 * and subsequently re faulted). However it does guarantee that the page
1853 * won't be freed completely. And mostly callers simply care that the page
1854 * contains data that was valid *at some point in time*. Typically, an IO
1855 * or similar operation cannot guarantee anything stronger anyway because
1856 * locks can't be held over the syscall boundary.
1857 *
1858 * If write=0, the page must not be written to. If the page is written to,
1859 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1860 * after the page is finished with, and before put_page is called.
1861 *
1862 * get_user_pages is typically used for fewer-copy IO operations, to get a
1863 * handle on the memory by some means other than accesses via the user virtual
1864 * addresses. The pages may be submitted for DMA to devices or accessed via
1865 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1866 * use the correct cache flushing APIs.
1867 *
1868 * See also get_user_pages_fast, for performance critical applications.
1869 */
1870int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1871        unsigned long start, int nr_pages, int write, int force,
1872        struct page **pages, struct vm_area_struct **vmas)
1873{
1874    int flags = FOLL_TOUCH;
1875
1876    if (pages)
1877        flags |= FOLL_GET;
1878    if (write)
1879        flags |= FOLL_WRITE;
1880    if (force)
1881        flags |= FOLL_FORCE;
1882
1883    return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1884                NULL);
1885}
1886EXPORT_SYMBOL(get_user_pages);
1887
1888/**
1889 * get_dump_page() - pin user page in memory while writing it to core dump
1890 * @addr: user address
1891 *
1892 * Returns struct page pointer of user page pinned for dump,
1893 * to be freed afterwards by page_cache_release() or put_page().
1894 *
1895 * Returns NULL on any kind of failure - a hole must then be inserted into
1896 * the corefile, to preserve alignment with its headers; and also returns
1897 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1898 * allowing a hole to be left in the corefile to save diskspace.
1899 *
1900 * Called without mmap_sem, but after all other threads have been killed.
1901 */
1902#ifdef CONFIG_ELF_CORE
1903struct page *get_dump_page(unsigned long addr)
1904{
1905    struct vm_area_struct *vma;
1906    struct page *page;
1907
1908    if (__get_user_pages(current, current->mm, addr, 1,
1909                 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1910                 NULL) < 1)
1911        return NULL;
1912    flush_cache_page(vma, addr, page_to_pfn(page));
1913    return page;
1914}
1915#endif /* CONFIG_ELF_CORE */
1916
1917pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1918            spinlock_t **ptl)
1919{
1920    pgd_t * pgd = pgd_offset(mm, addr);
1921    pud_t * pud = pud_alloc(mm, pgd, addr);
1922    if (pud) {
1923        pmd_t * pmd = pmd_alloc(mm, pud, addr);
1924        if (pmd) {
1925            VM_BUG_ON(pmd_trans_huge(*pmd));
1926            return pte_alloc_map_lock(mm, pmd, addr, ptl);
1927        }
1928    }
1929    return NULL;
1930}
1931
1932/*
1933 * This is the old fallback for page remapping.
1934 *
1935 * For historical reasons, it only allows reserved pages. Only
1936 * old drivers should use this, and they needed to mark their
1937 * pages reserved for the old functions anyway.
1938 */
1939static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1940            struct page *page, pgprot_t prot)
1941{
1942    struct mm_struct *mm = vma->vm_mm;
1943    int retval;
1944    pte_t *pte;
1945    spinlock_t *ptl;
1946
1947    retval = -EINVAL;
1948    if (PageAnon(page))
1949        goto out;
1950    retval = -ENOMEM;
1951    flush_dcache_page(page);
1952    pte = get_locked_pte(mm, addr, &ptl);
1953    if (!pte)
1954        goto out;
1955    retval = -EBUSY;
1956    if (!pte_none(*pte))
1957        goto out_unlock;
1958
1959    /* Ok, finally just insert the thing.. */
1960    get_page(page);
1961    inc_mm_counter_fast(mm, MM_FILEPAGES);
1962    page_add_file_rmap(page);
1963    set_pte_at(mm, addr, pte, mk_pte(page, prot));
1964
1965    retval = 0;
1966    pte_unmap_unlock(pte, ptl);
1967    return retval;
1968out_unlock:
1969    pte_unmap_unlock(pte, ptl);
1970out:
1971    return retval;
1972}
1973
1974/**
1975 * vm_insert_page - insert single page into user vma
1976 * @vma: user vma to map to
1977 * @addr: target user address of this page
1978 * @page: source kernel page
1979 *
1980 * This allows drivers to insert individual pages they've allocated
1981 * into a user vma.
1982 *
1983 * The page has to be a nice clean _individual_ kernel allocation.
1984 * If you allocate a compound page, you need to have marked it as
1985 * such (__GFP_COMP), or manually just split the page up yourself
1986 * (see split_page()).
1987 *
1988 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1989 * took an arbitrary page protection parameter. This doesn't allow
1990 * that. Your vma protection will have to be set up correctly, which
1991 * means that if you want a shared writable mapping, you'd better
1992 * ask for a shared writable mapping!
1993 *
1994 * The page does not need to be reserved.
1995 */
1996int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1997            struct page *page)
1998{
1999    if (addr < vma->vm_start || addr >= vma->vm_end)
2000        return -EFAULT;
2001    if (!page_count(page))
2002        return -EINVAL;
2003    vma->vm_flags |= VM_INSERTPAGE;
2004    return insert_page(vma, addr, page, vma->vm_page_prot);
2005}
2006EXPORT_SYMBOL(vm_insert_page);
2007
2008static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2009            unsigned long pfn, pgprot_t prot)
2010{
2011    struct mm_struct *mm = vma->vm_mm;
2012    int retval;
2013    pte_t *pte, entry;
2014    spinlock_t *ptl;
2015
2016    retval = -ENOMEM;
2017    pte = get_locked_pte(mm, addr, &ptl);
2018    if (!pte)
2019        goto out;
2020    retval = -EBUSY;
2021    if (!pte_none(*pte))
2022        goto out_unlock;
2023
2024    /* Ok, finally just insert the thing.. */
2025    entry = pte_mkspecial(pfn_pte(pfn, prot));
2026    set_pte_at(mm, addr, pte, entry);
2027    update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2028
2029    retval = 0;
2030out_unlock:
2031    pte_unmap_unlock(pte, ptl);
2032out:
2033    return retval;
2034}
2035
2036/**
2037 * vm_insert_pfn - insert single pfn into user vma
2038 * @vma: user vma to map to
2039 * @addr: target user address of this page
2040 * @pfn: source kernel pfn
2041 *
2042 * Similar to vm_inert_page, this allows drivers to insert individual pages
2043 * they've allocated into a user vma. Same comments apply.
2044 *
2045 * This function should only be called from a vm_ops->fault handler, and
2046 * in that case the handler should return NULL.
2047 *
2048 * vma cannot be a COW mapping.
2049 *
2050 * As this is called only for pages that do not currently exist, we
2051 * do not need to flush old virtual caches or the TLB.
2052 */
2053int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2054            unsigned long pfn)
2055{
2056    int ret;
2057    pgprot_t pgprot = vma->vm_page_prot;
2058    /*
2059     * Technically, architectures with pte_special can avoid all these
2060     * restrictions (same for remap_pfn_range). However we would like
2061     * consistency in testing and feature parity among all, so we should
2062     * try to keep these invariants in place for everybody.
2063     */
2064    BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2065    BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2066                        (VM_PFNMAP|VM_MIXEDMAP));
2067    BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2068    BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2069
2070    if (addr < vma->vm_start || addr >= vma->vm_end)
2071        return -EFAULT;
2072    if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2073        return -EINVAL;
2074
2075    ret = insert_pfn(vma, addr, pfn, pgprot);
2076
2077    if (ret)
2078        untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2079
2080    return ret;
2081}
2082EXPORT_SYMBOL(vm_insert_pfn);
2083
2084int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2085            unsigned long pfn)
2086{
2087    BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2088
2089    if (addr < vma->vm_start || addr >= vma->vm_end)
2090        return -EFAULT;
2091
2092    /*
2093     * If we don't have pte special, then we have to use the pfn_valid()
2094     * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2095     * refcount the page if pfn_valid is true (hence insert_page rather
2096     * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2097     * without pte special, it would there be refcounted as a normal page.
2098     */
2099    if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2100        struct page *page;
2101
2102        page = pfn_to_page(pfn);
2103        return insert_page(vma, addr, page, vma->vm_page_prot);
2104    }
2105    return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2106}
2107EXPORT_SYMBOL(vm_insert_mixed);
2108
2109/*
2110 * maps a range of physical memory into the requested pages. the old
2111 * mappings are removed. any references to nonexistent pages results
2112 * in null mappings (currently treated as "copy-on-access")
2113 */
2114static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2115            unsigned long addr, unsigned long end,
2116            unsigned long pfn, pgprot_t prot)
2117{
2118    pte_t *pte;
2119    spinlock_t *ptl;
2120
2121    pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2122    if (!pte)
2123        return -ENOMEM;
2124    arch_enter_lazy_mmu_mode();
2125    do {
2126        BUG_ON(!pte_none(*pte));
2127        set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2128        pfn++;
2129    } while (pte++, addr += PAGE_SIZE, addr != end);
2130    arch_leave_lazy_mmu_mode();
2131    pte_unmap_unlock(pte - 1, ptl);
2132    return 0;
2133}
2134
2135static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2136            unsigned long addr, unsigned long end,
2137            unsigned long pfn, pgprot_t prot)
2138{
2139    pmd_t *pmd;
2140    unsigned long next;
2141
2142    pfn -= addr >> PAGE_SHIFT;
2143    pmd = pmd_alloc(mm, pud, addr);
2144    if (!pmd)
2145        return -ENOMEM;
2146    VM_BUG_ON(pmd_trans_huge(*pmd));
2147    do {
2148        next = pmd_addr_end(addr, end);
2149        if (remap_pte_range(mm, pmd, addr, next,
2150                pfn + (addr >> PAGE_SHIFT), prot))
2151            return -ENOMEM;
2152    } while (pmd++, addr = next, addr != end);
2153    return 0;
2154}
2155
2156static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2157            unsigned long addr, unsigned long end,
2158            unsigned long pfn, pgprot_t prot)
2159{
2160    pud_t *pud;
2161    unsigned long next;
2162
2163    pfn -= addr >> PAGE_SHIFT;
2164    pud = pud_alloc(mm, pgd, addr);
2165    if (!pud)
2166        return -ENOMEM;
2167    do {
2168        next = pud_addr_end(addr, end);
2169        if (remap_pmd_range(mm, pud, addr, next,
2170                pfn + (addr >> PAGE_SHIFT), prot))
2171            return -ENOMEM;
2172    } while (pud++, addr = next, addr != end);
2173    return 0;
2174}
2175
2176/**
2177 * remap_pfn_range - remap kernel memory to userspace
2178 * @vma: user vma to map to
2179 * @addr: target user address to start at
2180 * @pfn: physical address of kernel memory
2181 * @size: size of map area
2182 * @prot: page protection flags for this mapping
2183 *
2184 * Note: this is only safe if the mm semaphore is held when called.
2185 */
2186int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2187            unsigned long pfn, unsigned long size, pgprot_t prot)
2188{
2189    pgd_t *pgd;
2190    unsigned long next;
2191    unsigned long end = addr + PAGE_ALIGN(size);
2192    struct mm_struct *mm = vma->vm_mm;
2193    int err;
2194
2195    /*
2196     * Physically remapped pages are special. Tell the
2197     * rest of the world about it:
2198     * VM_IO tells people not to look at these pages
2199     * (accesses can have side effects).
2200     * VM_RESERVED is specified all over the place, because
2201     * in 2.4 it kept swapout's vma scan off this vma; but
2202     * in 2.6 the LRU scan won't even find its pages, so this
2203     * flag means no more than count its pages in reserved_vm,
2204     * and omit it from core dump, even when VM_IO turned off.
2205     * VM_PFNMAP tells the core MM that the base pages are just
2206     * raw PFN mappings, and do not have a "struct page" associated
2207     * with them.
2208     *
2209     * There's a horrible special case to handle copy-on-write
2210     * behaviour that some programs depend on. We mark the "original"
2211     * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2212     */
2213    if (addr == vma->vm_start && end == vma->vm_end) {
2214        vma->vm_pgoff = pfn;
2215        vma->vm_flags |= VM_PFN_AT_MMAP;
2216    } else if (is_cow_mapping(vma->vm_flags))
2217        return -EINVAL;
2218
2219    vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2220
2221    err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2222    if (err) {
2223        /*
2224         * To indicate that track_pfn related cleanup is not
2225         * needed from higher level routine calling unmap_vmas
2226         */
2227        vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2228        vma->vm_flags &= ~VM_PFN_AT_MMAP;
2229        return -EINVAL;
2230    }
2231
2232    BUG_ON(addr >= end);
2233    pfn -= addr >> PAGE_SHIFT;
2234    pgd = pgd_offset(mm, addr);
2235    flush_cache_range(vma, addr, end);
2236    do {
2237        next = pgd_addr_end(addr, end);
2238        err = remap_pud_range(mm, pgd, addr, next,
2239                pfn + (addr >> PAGE_SHIFT), prot);
2240        if (err)
2241            break;
2242    } while (pgd++, addr = next, addr != end);
2243
2244    if (err)
2245        untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2246
2247    return err;
2248}
2249EXPORT_SYMBOL(remap_pfn_range);
2250
2251static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2252                     unsigned long addr, unsigned long end,
2253                     pte_fn_t fn, void *data)
2254{
2255    pte_t *pte;
2256    int err;
2257    pgtable_t token;
2258    spinlock_t *uninitialized_var(ptl);
2259
2260    pte = (mm == &init_mm) ?
2261        pte_alloc_kernel(pmd, addr) :
2262        pte_alloc_map_lock(mm, pmd, addr, &ptl);
2263    if (!pte)
2264        return -ENOMEM;
2265
2266    BUG_ON(pmd_huge(*pmd));
2267
2268    arch_enter_lazy_mmu_mode();
2269
2270    token = pmd_pgtable(*pmd);
2271
2272    do {
2273        err = fn(pte++, token, addr, data);
2274        if (err)
2275            break;
2276    } while (addr += PAGE_SIZE, addr != end);
2277
2278    arch_leave_lazy_mmu_mode();
2279
2280    if (mm != &init_mm)
2281        pte_unmap_unlock(pte-1, ptl);
2282    return err;
2283}
2284
2285static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2286                     unsigned long addr, unsigned long end,
2287                     pte_fn_t fn, void *data)
2288{
2289    pmd_t *pmd;
2290    unsigned long next;
2291    int err;
2292
2293    BUG_ON(pud_huge(*pud));
2294
2295    pmd = pmd_alloc(mm, pud, addr);
2296    if (!pmd)
2297        return -ENOMEM;
2298    do {
2299        next = pmd_addr_end(addr, end);
2300        err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2301        if (err)
2302            break;
2303    } while (pmd++, addr = next, addr != end);
2304    return err;
2305}
2306
2307static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2308                     unsigned long addr, unsigned long end,
2309                     pte_fn_t fn, void *data)
2310{
2311    pud_t *pud;
2312    unsigned long next;
2313    int err;
2314
2315    pud = pud_alloc(mm, pgd, addr);
2316    if (!pud)
2317        return -ENOMEM;
2318    do {
2319        next = pud_addr_end(addr, end);
2320        err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2321        if (err)
2322            break;
2323    } while (pud++, addr = next, addr != end);
2324    return err;
2325}
2326
2327/*
2328 * Scan a region of virtual memory, filling in page tables as necessary
2329 * and calling a provided function on each leaf page table.
2330 */
2331int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2332            unsigned long size, pte_fn_t fn, void *data)
2333{
2334    pgd_t *pgd;
2335    unsigned long next;
2336    unsigned long end = addr + size;
2337    int err;
2338
2339    BUG_ON(addr >= end);
2340    pgd = pgd_offset(mm, addr);
2341    do {
2342        next = pgd_addr_end(addr, end);
2343        err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2344        if (err)
2345            break;
2346    } while (pgd++, addr = next, addr != end);
2347
2348    return err;
2349}
2350EXPORT_SYMBOL_GPL(apply_to_page_range);
2351
2352/*
2353 * handle_pte_fault chooses page fault handler according to an entry
2354 * which was read non-atomically. Before making any commitment, on
2355 * those architectures or configurations (e.g. i386 with PAE) which
2356 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2357 * must check under lock before unmapping the pte and proceeding
2358 * (but do_wp_page is only called after already making such a check;
2359 * and do_anonymous_page can safely check later on).
2360 */
2361static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2362                pte_t *page_table, pte_t orig_pte)
2363{
2364    int same = 1;
2365#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2366    if (sizeof(pte_t) > sizeof(unsigned long)) {
2367        spinlock_t *ptl = pte_lockptr(mm, pmd);
2368        spin_lock(ptl);
2369        same = pte_same(*page_table, orig_pte);
2370        spin_unlock(ptl);
2371    }
2372#endif
2373    pte_unmap(page_table);
2374    return same;
2375}
2376
2377static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2378{
2379    /*
2380     * If the source page was a PFN mapping, we don't have
2381     * a "struct page" for it. We do a best-effort copy by
2382     * just copying from the original user address. If that
2383     * fails, we just zero-fill it. Live with it.
2384     */
2385    if (unlikely(!src)) {
2386        void *kaddr = kmap_atomic(dst, KM_USER0);
2387        void __user *uaddr = (void __user *)(va & PAGE_MASK);
2388
2389        /*
2390         * This really shouldn't fail, because the page is there
2391         * in the page tables. But it might just be unreadable,
2392         * in which case we just give up and fill the result with
2393         * zeroes.
2394         */
2395        if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2396            clear_page(kaddr);
2397        kunmap_atomic(kaddr, KM_USER0);
2398        flush_dcache_page(dst);
2399    } else
2400        copy_user_highpage(dst, src, va, vma);
2401}
2402
2403/*
2404 * This routine handles present pages, when users try to write
2405 * to a shared page. It is done by copying the page to a new address
2406 * and decrementing the shared-page counter for the old page.
2407 *
2408 * Note that this routine assumes that the protection checks have been
2409 * done by the caller (the low-level page fault routine in most cases).
2410 * Thus we can safely just mark it writable once we've done any necessary
2411 * COW.
2412 *
2413 * We also mark the page dirty at this point even though the page will
2414 * change only once the write actually happens. This avoids a few races,
2415 * and potentially makes it more efficient.
2416 *
2417 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2418 * but allow concurrent faults), with pte both mapped and locked.
2419 * We return with mmap_sem still held, but pte unmapped and unlocked.
2420 */
2421static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2422        unsigned long address, pte_t *page_table, pmd_t *pmd,
2423        spinlock_t *ptl, pte_t orig_pte)
2424    __releases(ptl)
2425{
2426    struct page *old_page, *new_page;
2427    pte_t entry;
2428    int ret = 0;
2429    int page_mkwrite = 0;
2430    struct page *dirty_page = NULL;
2431
2432    old_page = vm_normal_page(vma, address, orig_pte);
2433    if (!old_page) {
2434        /*
2435         * VM_MIXEDMAP !pfn_valid() case
2436         *
2437         * We should not cow pages in a shared writeable mapping.
2438         * Just mark the pages writable as we can't do any dirty
2439         * accounting on raw pfn maps.
2440         */
2441        if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2442                     (VM_WRITE|VM_SHARED))
2443            goto reuse;
2444        goto gotten;
2445    }
2446
2447    /*
2448     * Take out anonymous pages first, anonymous shared vmas are
2449     * not dirty accountable.
2450     */
2451    if (PageAnon(old_page) && !PageKsm(old_page)) {
2452        if (!trylock_page(old_page)) {
2453            page_cache_get(old_page);
2454            pte_unmap_unlock(page_table, ptl);
2455            lock_page(old_page);
2456            page_table = pte_offset_map_lock(mm, pmd, address,
2457                             &ptl);
2458            if (!pte_same(*page_table, orig_pte)) {
2459                unlock_page(old_page);
2460                goto unlock;
2461            }
2462            page_cache_release(old_page);
2463        }
2464        if (reuse_swap_page(old_page)) {
2465            /*
2466             * The page is all ours. Move it to our anon_vma so
2467             * the rmap code will not search our parent or siblings.
2468             * Protected against the rmap code by the page lock.
2469             */
2470            page_move_anon_rmap(old_page, vma, address);
2471            unlock_page(old_page);
2472            goto reuse;
2473        }
2474        unlock_page(old_page);
2475    } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2476                    (VM_WRITE|VM_SHARED))) {
2477        /*
2478         * Only catch write-faults on shared writable pages,
2479         * read-only shared pages can get COWed by
2480         * get_user_pages(.write=1, .force=1).
2481         */
2482        if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2483            struct vm_fault vmf;
2484            int tmp;
2485
2486            vmf.virtual_address = (void __user *)(address &
2487                                PAGE_MASK);
2488            vmf.pgoff = old_page->index;
2489            vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2490            vmf.page = old_page;
2491
2492            /*
2493             * Notify the address space that the page is about to
2494             * become writable so that it can prohibit this or wait
2495             * for the page to get into an appropriate state.
2496             *
2497             * We do this without the lock held, so that it can
2498             * sleep if it needs to.
2499             */
2500            page_cache_get(old_page);
2501            pte_unmap_unlock(page_table, ptl);
2502
2503            tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2504            if (unlikely(tmp &
2505                    (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2506                ret = tmp;
2507                goto unwritable_page;
2508            }
2509            if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2510                lock_page(old_page);
2511                if (!old_page->mapping) {
2512                    ret = 0; /* retry the fault */
2513                    unlock_page(old_page);
2514                    goto unwritable_page;
2515                }
2516            } else
2517                VM_BUG_ON(!PageLocked(old_page));
2518
2519            /*
2520             * Since we dropped the lock we need to revalidate
2521             * the PTE as someone else may have changed it. If
2522             * they did, we just return, as we can count on the
2523             * MMU to tell us if they didn't also make it writable.
2524             */
2525            page_table = pte_offset_map_lock(mm, pmd, address,
2526                             &ptl);
2527            if (!pte_same(*page_table, orig_pte)) {
2528                unlock_page(old_page);
2529                goto unlock;
2530            }
2531
2532            page_mkwrite = 1;
2533        }
2534        dirty_page = old_page;
2535        get_page(dirty_page);
2536
2537reuse:
2538        flush_cache_page(vma, address, pte_pfn(orig_pte));
2539        entry = pte_mkyoung(orig_pte);
2540        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2541        if (ptep_set_access_flags(vma, address, page_table, entry,1))
2542            update_mmu_cache(vma, address, page_table);
2543        pte_unmap_unlock(page_table, ptl);
2544        ret |= VM_FAULT_WRITE;
2545
2546        if (!dirty_page)
2547            return ret;
2548
2549        /*
2550         * Yes, Virginia, this is actually required to prevent a race
2551         * with clear_page_dirty_for_io() from clearing the page dirty
2552         * bit after it clear all dirty ptes, but before a racing
2553         * do_wp_page installs a dirty pte.
2554         *
2555         * __do_fault is protected similarly.
2556         */
2557        if (!page_mkwrite) {
2558            wait_on_page_locked(dirty_page);
2559            set_page_dirty_balance(dirty_page, page_mkwrite);
2560        }
2561        put_page(dirty_page);
2562        if (page_mkwrite) {
2563            struct address_space *mapping = dirty_page->mapping;
2564
2565            set_page_dirty(dirty_page);
2566            unlock_page(dirty_page);
2567            page_cache_release(dirty_page);
2568            if (mapping) {
2569                /*
2570                 * Some device drivers do not set page.mapping
2571                 * but still dirty their pages
2572                 */
2573                balance_dirty_pages_ratelimited(mapping);
2574            }
2575        }
2576
2577        /* file_update_time outside page_lock */
2578        if (vma->vm_file)
2579            file_update_time(vma->vm_file);
2580
2581        return ret;
2582    }
2583
2584    /*
2585     * Ok, we need to copy. Oh, well..
2586     */
2587    page_cache_get(old_page);
2588gotten:
2589    pte_unmap_unlock(page_table, ptl);
2590
2591    if (unlikely(anon_vma_prepare(vma)))
2592        goto oom;
2593
2594    if (is_zero_pfn(pte_pfn(orig_pte))) {
2595        new_page = alloc_zeroed_user_highpage_movable(vma, address);
2596        if (!new_page)
2597            goto oom;
2598    } else {
2599        new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2600        if (!new_page)
2601            goto oom;
2602        cow_user_page(new_page, old_page, address, vma);
2603    }
2604    __SetPageUptodate(new_page);
2605
2606    if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2607        goto oom_free_new;
2608
2609    /*
2610     * Re-check the pte - we dropped the lock
2611     */
2612    page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2613    if (likely(pte_same(*page_table, orig_pte))) {
2614        if (old_page) {
2615            if (!PageAnon(old_page)) {
2616                dec_mm_counter_fast(mm, MM_FILEPAGES);
2617                inc_mm_counter_fast(mm, MM_ANONPAGES);
2618            }
2619        } else
2620            inc_mm_counter_fast(mm, MM_ANONPAGES);
2621        flush_cache_page(vma, address, pte_pfn(orig_pte));
2622        entry = mk_pte(new_page, vma->vm_page_prot);
2623        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2624        /*
2625         * Clear the pte entry and flush it first, before updating the
2626         * pte with the new entry. This will avoid a race condition
2627         * seen in the presence of one thread doing SMC and another
2628         * thread doing COW.
2629         */
2630        ptep_clear_flush(vma, address, page_table);
2631        page_add_new_anon_rmap(new_page, vma, address);
2632        /*
2633         * We call the notify macro here because, when using secondary
2634         * mmu page tables (such as kvm shadow page tables), we want the
2635         * new page to be mapped directly into the secondary page table.
2636         */
2637        set_pte_at_notify(mm, address, page_table, entry);
2638        update_mmu_cache(vma, address, page_table);
2639        if (old_page) {
2640            /*
2641             * Only after switching the pte to the new page may
2642             * we remove the mapcount here. Otherwise another
2643             * process may come and find the rmap count decremented
2644             * before the pte is switched to the new page, and
2645             * "reuse" the old page writing into it while our pte
2646             * here still points into it and can be read by other
2647             * threads.
2648             *
2649             * The critical issue is to order this
2650             * page_remove_rmap with the ptp_clear_flush above.
2651             * Those stores are ordered by (if nothing else,)
2652             * the barrier present in the atomic_add_negative
2653             * in page_remove_rmap.
2654             *
2655             * Then the TLB flush in ptep_clear_flush ensures that
2656             * no process can access the old page before the
2657             * decremented mapcount is visible. And the old page
2658             * cannot be reused until after the decremented
2659             * mapcount is visible. So transitively, TLBs to
2660             * old page will be flushed before it can be reused.
2661             */
2662            page_remove_rmap(old_page);
2663        }
2664
2665        /* Free the old page.. */
2666        new_page = old_page;
2667        ret |= VM_FAULT_WRITE;
2668    } else
2669        mem_cgroup_uncharge_page(new_page);
2670
2671    if (new_page)
2672        page_cache_release(new_page);
2673unlock:
2674    pte_unmap_unlock(page_table, ptl);
2675    if (old_page) {
2676        /*
2677         * Don't let another task, with possibly unlocked vma,
2678         * keep the mlocked page.
2679         */
2680        if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2681            lock_page(old_page); /* LRU manipulation */
2682            munlock_vma_page(old_page);
2683            unlock_page(old_page);
2684        }
2685        page_cache_release(old_page);
2686    }
2687    return ret;
2688oom_free_new:
2689    page_cache_release(new_page);
2690oom:
2691    if (old_page) {
2692        if (page_mkwrite) {
2693            unlock_page(old_page);
2694            page_cache_release(old_page);
2695        }
2696        page_cache_release(old_page);
2697    }
2698    return VM_FAULT_OOM;
2699
2700unwritable_page:
2701    page_cache_release(old_page);
2702    return ret;
2703}
2704
2705static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2706        unsigned long start_addr, unsigned long end_addr,
2707        struct zap_details *details)
2708{
2709    zap_page_range(vma, start_addr, end_addr - start_addr, details);
2710}
2711
2712static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2713                        struct zap_details *details)
2714{
2715    struct vm_area_struct *vma;
2716    struct prio_tree_iter iter;
2717    pgoff_t vba, vea, zba, zea;
2718
2719    vma_prio_tree_foreach(vma, &iter, root,
2720            details->first_index, details->last_index) {
2721
2722        vba = vma->vm_pgoff;
2723        vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2724        /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2725        zba = details->first_index;
2726        if (zba < vba)
2727            zba = vba;
2728        zea = details->last_index;
2729        if (zea > vea)
2730            zea = vea;
2731
2732        unmap_mapping_range_vma(vma,
2733            ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2734            ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2735                details);
2736    }
2737}
2738
2739static inline void unmap_mapping_range_list(struct list_head *head,
2740                        struct zap_details *details)
2741{
2742    struct vm_area_struct *vma;
2743
2744    /*
2745     * In nonlinear VMAs there is no correspondence between virtual address
2746     * offset and file offset. So we must perform an exhaustive search
2747     * across *all* the pages in each nonlinear VMA, not just the pages
2748     * whose virtual address lies outside the file truncation point.
2749     */
2750    list_for_each_entry(vma, head, shared.vm_set.list) {
2751        details->nonlinear_vma = vma;
2752        unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2753    }
2754}
2755
2756/**
2757 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2758 * @mapping: the address space containing mmaps to be unmapped.
2759 * @holebegin: byte in first page to unmap, relative to the start of
2760 * the underlying file. This will be rounded down to a PAGE_SIZE
2761 * boundary. Note that this is different from truncate_pagecache(), which
2762 * must keep the partial page. In contrast, we must get rid of
2763 * partial pages.
2764 * @holelen: size of prospective hole in bytes. This will be rounded
2765 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2766 * end of the file.
2767 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2768 * but 0 when invalidating pagecache, don't throw away private data.
2769 */
2770void unmap_mapping_range(struct address_space *mapping,
2771        loff_t const holebegin, loff_t const holelen, int even_cows)
2772{
2773    struct zap_details details;
2774    pgoff_t hba = holebegin >> PAGE_SHIFT;
2775    pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2776
2777    /* Check for overflow. */
2778    if (sizeof(holelen) > sizeof(hlen)) {
2779        long long holeend =
2780            (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2781        if (holeend & ~(long long)ULONG_MAX)
2782            hlen = ULONG_MAX - hba + 1;
2783    }
2784
2785    details.check_mapping = even_cows? NULL: mapping;
2786    details.nonlinear_vma = NULL;
2787    details.first_index = hba;
2788    details.last_index = hba + hlen - 1;
2789    if (details.last_index < details.first_index)
2790        details.last_index = ULONG_MAX;
2791
2792
2793    mutex_lock(&mapping->i_mmap_mutex);
2794    if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2795        unmap_mapping_range_tree(&mapping->i_mmap, &details);
2796    if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2797        unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2798    mutex_unlock(&mapping->i_mmap_mutex);
2799}
2800EXPORT_SYMBOL(unmap_mapping_range);
2801
2802/*
2803 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2804 * but allow concurrent faults), and pte mapped but not yet locked.
2805 * We return with mmap_sem still held, but pte unmapped and unlocked.
2806 */
2807static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2808        unsigned long address, pte_t *page_table, pmd_t *pmd,
2809        unsigned int flags, pte_t orig_pte)
2810{
2811    spinlock_t *ptl;
2812    struct page *page, *swapcache = NULL;
2813    swp_entry_t entry;
2814    pte_t pte;
2815    int locked;
2816    struct mem_cgroup *ptr;
2817    int exclusive = 0;
2818    int ret = 0;
2819
2820    if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2821        goto out;
2822
2823    entry = pte_to_swp_entry(orig_pte);
2824    if (unlikely(non_swap_entry(entry))) {
2825        if (is_migration_entry(entry)) {
2826            migration_entry_wait(mm, pmd, address);
2827        } else if (is_hwpoison_entry(entry)) {
2828            ret = VM_FAULT_HWPOISON;
2829        } else {
2830            print_bad_pte(vma, address, orig_pte, NULL);
2831            ret = VM_FAULT_SIGBUS;
2832        }
2833        goto out;
2834    }
2835    delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2836    page = lookup_swap_cache(entry);
2837    if (!page) {
2838        grab_swap_token(mm); /* Contend for token _before_ read-in */
2839        page = swapin_readahead(entry,
2840                    GFP_HIGHUSER_MOVABLE, vma, address);
2841        if (!page) {
2842            /*
2843             * Back out if somebody else faulted in this pte
2844             * while we released the pte lock.
2845             */
2846            page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2847            if (likely(pte_same(*page_table, orig_pte)))
2848                ret = VM_FAULT_OOM;
2849            delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2850            goto unlock;
2851        }
2852
2853        /* Had to read the page from swap area: Major fault */
2854        ret = VM_FAULT_MAJOR;
2855        count_vm_event(PGMAJFAULT);
2856        mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2857    } else if (PageHWPoison(page)) {
2858        /*
2859         * hwpoisoned dirty swapcache pages are kept for killing
2860         * owner processes (which may be unknown at hwpoison time)
2861         */
2862        ret = VM_FAULT_HWPOISON;
2863        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2864        goto out_release;
2865    }
2866
2867    locked = lock_page_or_retry(page, mm, flags);
2868    delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2869    if (!locked) {
2870        ret |= VM_FAULT_RETRY;
2871        goto out_release;
2872    }
2873
2874    /*
2875     * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2876     * release the swapcache from under us. The page pin, and pte_same
2877     * test below, are not enough to exclude that. Even if it is still
2878     * swapcache, we need to check that the page's swap has not changed.
2879     */
2880    if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2881        goto out_page;
2882
2883    if (ksm_might_need_to_copy(page, vma, address)) {
2884        swapcache = page;
2885        page = ksm_does_need_to_copy(page, vma, address);
2886
2887        if (unlikely(!page)) {
2888            ret = VM_FAULT_OOM;
2889            page = swapcache;
2890            swapcache = NULL;
2891            goto out_page;
2892        }
2893    }
2894
2895    if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2896        ret = VM_FAULT_OOM;
2897        goto out_page;
2898    }
2899
2900    /*
2901     * Back out if somebody else already faulted in this pte.
2902     */
2903    page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2904    if (unlikely(!pte_same(*page_table, orig_pte)))
2905        goto out_nomap;
2906
2907    if (unlikely(!PageUptodate(page))) {
2908        ret = VM_FAULT_SIGBUS;
2909        goto out_nomap;
2910    }
2911
2912    /*
2913     * The page isn't present yet, go ahead with the fault.
2914     *
2915     * Be careful about the sequence of operations here.
2916     * To get its accounting right, reuse_swap_page() must be called
2917     * while the page is counted on swap but not yet in mapcount i.e.
2918     * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2919     * must be called after the swap_free(), or it will never succeed.
2920     * Because delete_from_swap_page() may be called by reuse_swap_page(),
2921     * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2922     * in page->private. In this case, a record in swap_cgroup is silently
2923     * discarded at swap_free().
2924     */
2925
2926    inc_mm_counter_fast(mm, MM_ANONPAGES);
2927    dec_mm_counter_fast(mm, MM_SWAPENTS);
2928    pte = mk_pte(page, vma->vm_page_prot);
2929    if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2930        pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2931        flags &= ~FAULT_FLAG_WRITE;
2932        ret |= VM_FAULT_WRITE;
2933        exclusive = 1;
2934    }
2935    flush_icache_page(vma, page);
2936    set_pte_at(mm, address, page_table, pte);
2937    do_page_add_anon_rmap(page, vma, address, exclusive);
2938    /* It's better to call commit-charge after rmap is established */
2939    mem_cgroup_commit_charge_swapin(page, ptr);
2940
2941    swap_free(entry);
2942    if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2943        try_to_free_swap(page);
2944    unlock_page(page);
2945    if (swapcache) {
2946        /*
2947         * Hold the lock to avoid the swap entry to be reused
2948         * until we take the PT lock for the pte_same() check
2949         * (to avoid false positives from pte_same). For
2950         * further safety release the lock after the swap_free
2951         * so that the swap count won't change under a
2952         * parallel locked swapcache.
2953         */
2954        unlock_page(swapcache);
2955        page_cache_release(swapcache);
2956    }
2957
2958    if (flags & FAULT_FLAG_WRITE) {
2959        ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2960        if (ret & VM_FAULT_ERROR)
2961            ret &= VM_FAULT_ERROR;
2962        goto out;
2963    }
2964
2965    /* No need to invalidate - it was non-present before */
2966    update_mmu_cache(vma, address, page_table);
2967unlock:
2968    pte_unmap_unlock(page_table, ptl);
2969out:
2970    return ret;
2971out_nomap:
2972    mem_cgroup_cancel_charge_swapin(ptr);
2973    pte_unmap_unlock(page_table, ptl);
2974out_page:
2975    unlock_page(page);
2976out_release:
2977    page_cache_release(page);
2978    if (swapcache) {
2979        unlock_page(swapcache);
2980        page_cache_release(swapcache);
2981    }
2982    return ret;
2983}
2984
2985/*
2986 * This is like a special single-page "expand_{down|up}wards()",
2987 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2988 * doesn't hit another vma.
2989 */
2990static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2991{
2992    address &= PAGE_MASK;
2993    if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2994        struct vm_area_struct *prev = vma->vm_prev;
2995
2996        /*
2997         * Is there a mapping abutting this one below?
2998         *
2999         * That's only ok if it's the same stack mapping
3000         * that has gotten split..
3001         */
3002        if (prev && prev->vm_end == address)
3003            return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3004
3005        expand_downwards(vma, address - PAGE_SIZE);
3006    }
3007    if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3008        struct vm_area_struct *next = vma->vm_next;
3009
3010        /* As VM_GROWSDOWN but s/below/above/ */
3011        if (next && next->vm_start == address + PAGE_SIZE)
3012            return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3013
3014        expand_upwards(vma, address + PAGE_SIZE);
3015    }
3016    return 0;
3017}
3018
3019/*
3020 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3021 * but allow concurrent faults), and pte mapped but not yet locked.
3022 * We return with mmap_sem still held, but pte unmapped and unlocked.
3023 */
3024static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3025        unsigned long address, pte_t *page_table, pmd_t *pmd,
3026        unsigned int flags)
3027{
3028    struct page *page;
3029    spinlock_t *ptl;
3030    pte_t entry;
3031
3032    pte_unmap(page_table);
3033
3034    /* Check if we need to add a guard page to the stack */
3035    if (check_stack_guard_page(vma, address) < 0)
3036        return VM_FAULT_SIGBUS;
3037
3038    /* Use the zero-page for reads */
3039    if (!(flags & FAULT_FLAG_WRITE)) {
3040        entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3041                        vma->vm_page_prot));
3042        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3043        if (!pte_none(*page_table))
3044            goto unlock;
3045        goto setpte;
3046    }
3047
3048    /* Allocate our own private page. */
3049    if (unlikely(anon_vma_prepare(vma)))
3050        goto oom;
3051    page = alloc_zeroed_user_highpage_movable(vma, address);
3052    if (!page)
3053        goto oom;
3054    __SetPageUptodate(page);
3055
3056    if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3057        goto oom_free_page;
3058
3059    entry = mk_pte(page, vma->vm_page_prot);
3060    if (vma->vm_flags & VM_WRITE)
3061        entry = pte_mkwrite(pte_mkdirty(entry));
3062
3063    page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3064    if (!pte_none(*page_table))
3065        goto release;
3066
3067    inc_mm_counter_fast(mm, MM_ANONPAGES);
3068    page_add_new_anon_rmap(page, vma, address);
3069setpte:
3070    set_pte_at(mm, address, page_table, entry);
3071
3072    /* No need to invalidate - it was non-present before */
3073    update_mmu_cache(vma, address, page_table);
3074unlock:
3075    pte_unmap_unlock(page_table, ptl);
3076    return 0;
3077release:
3078    mem_cgroup_uncharge_page(page);
3079    page_cache_release(page);
3080    goto unlock;
3081oom_free_page:
3082    page_cache_release(page);
3083oom:
3084    return VM_FAULT_OOM;
3085}
3086
3087/*
3088 * __do_fault() tries to create a new page mapping. It aggressively
3089 * tries to share with existing pages, but makes a separate copy if
3090 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3091 * the next page fault.
3092 *
3093 * As this is called only for pages that do not currently exist, we
3094 * do not need to flush old virtual caches or the TLB.
3095 *
3096 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3097 * but allow concurrent faults), and pte neither mapped nor locked.
3098 * We return with mmap_sem still held, but pte unmapped and unlocked.
3099 */
3100static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3101        unsigned long address, pmd_t *pmd,
3102        pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3103{
3104    pte_t *page_table;
3105    spinlock_t *ptl;
3106    struct page *page;
3107    pte_t entry;
3108    int anon = 0;
3109    int charged = 0;
3110    struct page *dirty_page = NULL;
3111    struct vm_fault vmf;
3112    int ret;
3113    int page_mkwrite = 0;
3114
3115    vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3116    vmf.pgoff = pgoff;
3117    vmf.flags = flags;
3118    vmf.page = NULL;
3119
3120    ret = vma->vm_ops->fault(vma, &vmf);
3121    if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3122                VM_FAULT_RETRY)))
3123        return ret;
3124
3125    if (unlikely(PageHWPoison(vmf.page))) {
3126        if (ret & VM_FAULT_LOCKED)
3127            unlock_page(vmf.page);
3128        return VM_FAULT_HWPOISON;
3129    }
3130
3131    /*
3132     * For consistency in subsequent calls, make the faulted page always
3133     * locked.
3134     */
3135    if (unlikely(!(ret & VM_FAULT_LOCKED)))
3136        lock_page(vmf.page);
3137    else
3138        VM_BUG_ON(!PageLocked(vmf.page));
3139
3140    /*
3141     * Should we do an early C-O-W break?
3142     */
3143    page = vmf.page;
3144    if (flags & FAULT_FLAG_WRITE) {
3145        if (!(vma->vm_flags & VM_SHARED)) {
3146            anon = 1;
3147            if (unlikely(anon_vma_prepare(vma))) {
3148                ret = VM_FAULT_OOM;
3149                goto out;
3150            }
3151            page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3152                        vma, address);
3153            if (!page) {
3154                ret = VM_FAULT_OOM;
3155                goto out;
3156            }
3157            if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3158                ret = VM_FAULT_OOM;
3159                page_cache_release(page);
3160                goto out;
3161            }
3162            charged = 1;
3163            copy_user_highpage(page, vmf.page, address, vma);
3164            __SetPageUptodate(page);
3165        } else {
3166            /*
3167             * If the page will be shareable, see if the backing
3168             * address space wants to know that the page is about
3169             * to become writable
3170             */
3171            if (vma->vm_ops->page_mkwrite) {
3172                int tmp;
3173
3174                unlock_page(page);
3175                vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3176                tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3177                if (unlikely(tmp &
3178                      (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3179                    ret = tmp;
3180                    goto unwritable_page;
3181                }
3182                if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3183                    lock_page(page);
3184                    if (!page->mapping) {
3185                        ret = 0; /* retry the fault */
3186                        unlock_page(page);
3187                        goto unwritable_page;
3188                    }
3189                } else
3190                    VM_BUG_ON(!PageLocked(page));
3191                page_mkwrite = 1;
3192            }
3193        }
3194
3195    }
3196
3197    page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3198
3199    /*
3200     * This silly early PAGE_DIRTY setting removes a race
3201     * due to the bad i386 page protection. But it's valid
3202     * for other architectures too.
3203     *
3204     * Note that if FAULT_FLAG_WRITE is set, we either now have
3205     * an exclusive copy of the page, or this is a shared mapping,
3206     * so we can make it writable and dirty to avoid having to
3207     * handle that later.
3208     */
3209    /* Only go through if we didn't race with anybody else... */
3210    if (likely(pte_same(*page_table, orig_pte))) {
3211        flush_icache_page(vma, page);
3212        entry = mk_pte(page, vma->vm_page_prot);
3213        if (flags & FAULT_FLAG_WRITE)
3214            entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3215        if (anon) {
3216            inc_mm_counter_fast(mm, MM_ANONPAGES);
3217            page_add_new_anon_rmap(page, vma, address);
3218        } else {
3219            inc_mm_counter_fast(mm, MM_FILEPAGES);
3220            page_add_file_rmap(page);
3221            if (flags & FAULT_FLAG_WRITE) {
3222                dirty_page = page;
3223                get_page(dirty_page);
3224            }
3225        }
3226        set_pte_at(mm, address, page_table, entry);
3227
3228        /* no need to invalidate: a not-present page won't be cached */
3229        update_mmu_cache(vma, address, page_table);
3230    } else {
3231        if (charged)
3232            mem_cgroup_uncharge_page(page);
3233        if (anon)
3234            page_cache_release(page);
3235        else
3236            anon = 1; /* no anon but release faulted_page */
3237    }
3238
3239    pte_unmap_unlock(page_table, ptl);
3240
3241out:
3242    if (dirty_page) {
3243        struct address_space *mapping = page->mapping;
3244
3245        if (set_page_dirty(dirty_page))
3246            page_mkwrite = 1;
3247        unlock_page(dirty_page);
3248        put_page(dirty_page);
3249        if (page_mkwrite && mapping) {
3250            /*
3251             * Some device drivers do not set page.mapping but still
3252             * dirty their pages
3253             */
3254            balance_dirty_pages_ratelimited(mapping);
3255        }
3256
3257        /* file_update_time outside page_lock */
3258        if (vma->vm_file)
3259            file_update_time(vma->vm_file);
3260    } else {
3261        unlock_page(vmf.page);
3262        if (anon)
3263            page_cache_release(vmf.page);
3264    }
3265
3266    return ret;
3267
3268unwritable_page:
3269    page_cache_release(page);
3270    return ret;
3271}
3272
3273static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3274        unsigned long address, pte_t *page_table, pmd_t *pmd,
3275        unsigned int flags, pte_t orig_pte)
3276{
3277    pgoff_t pgoff = (((address & PAGE_MASK)
3278            - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3279
3280    pte_unmap(page_table);
3281    return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3282}
3283
3284/*
3285 * Fault of a previously existing named mapping. Repopulate the pte
3286 * from the encoded file_pte if possible. This enables swappable
3287 * nonlinear vmas.
3288 *
3289 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3290 * but allow concurrent faults), and pte mapped but not yet locked.
3291 * We return with mmap_sem still held, but pte unmapped and unlocked.
3292 */
3293static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3294        unsigned long address, pte_t *page_table, pmd_t *pmd,
3295        unsigned int flags, pte_t orig_pte)
3296{
3297    pgoff_t pgoff;
3298
3299    flags |= FAULT_FLAG_NONLINEAR;
3300
3301    if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3302        return 0;
3303
3304    if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3305        /*
3306         * Page table corrupted: show pte and kill process.
3307         */
3308        print_bad_pte(vma, address, orig_pte, NULL);
3309        return VM_FAULT_SIGBUS;
3310    }
3311
3312    pgoff = pte_to_pgoff(orig_pte);
3313    return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3314}
3315
3316/*
3317 * These routines also need to handle stuff like marking pages dirty
3318 * and/or accessed for architectures that don't do it in hardware (most
3319 * RISC architectures). The early dirtying is also good on the i386.
3320 *
3321 * There is also a hook called "update_mmu_cache()" that architectures
3322 * with external mmu caches can use to update those (ie the Sparc or
3323 * PowerPC hashed page tables that act as extended TLBs).
3324 *
3325 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3326 * but allow concurrent faults), and pte mapped but not yet locked.
3327 * We return with mmap_sem still held, but pte unmapped and unlocked.
3328 */
3329int handle_pte_fault(struct mm_struct *mm,
3330             struct vm_area_struct *vma, unsigned long address,
3331             pte_t *pte, pmd_t *pmd, unsigned int flags)
3332{
3333    pte_t entry;
3334    spinlock_t *ptl;
3335
3336    entry = *pte;
3337    if (!pte_present(entry)) {
3338        if (pte_none(entry)) {
3339            if (vma->vm_ops) {
3340                if (likely(vma->vm_ops->fault))
3341                    return do_linear_fault(mm, vma, address,
3342                        pte, pmd, flags, entry);
3343            }
3344            return do_anonymous_page(mm, vma, address,
3345                         pte, pmd, flags);
3346        }
3347        if (pte_file(entry))
3348            return do_nonlinear_fault(mm, vma, address,
3349                    pte, pmd, flags, entry);
3350        return do_swap_page(mm, vma, address,
3351                    pte, pmd, flags, entry);
3352    }
3353
3354    ptl = pte_lockptr(mm, pmd);
3355    spin_lock(ptl);
3356    if (unlikely(!pte_same(*pte, entry)))
3357        goto unlock;
3358    if (flags & FAULT_FLAG_WRITE) {
3359        if (!pte_write(entry))
3360            return do_wp_page(mm, vma, address,
3361                    pte, pmd, ptl, entry);
3362        entry = pte_mkdirty(entry);
3363    }
3364    entry = pte_mkyoung(entry);
3365    if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3366        update_mmu_cache(vma, address, pte);
3367    } else {
3368        /*
3369         * This is needed only for protection faults but the arch code
3370         * is not yet telling us if this is a protection fault or not.
3371         * This still avoids useless tlb flushes for .text page faults
3372         * with threads.
3373         */
3374        if (flags & FAULT_FLAG_WRITE)
3375            flush_tlb_fix_spurious_fault(vma, address);
3376    }
3377unlock:
3378    pte_unmap_unlock(pte, ptl);
3379    return 0;
3380}
3381
3382/*
3383 * By the time we get here, we already hold the mm semaphore
3384 */
3385int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3386        unsigned long address, unsigned int flags)
3387{
3388    pgd_t *pgd;
3389    pud_t *pud;
3390    pmd_t *pmd;
3391    pte_t *pte;
3392
3393    __set_current_state(TASK_RUNNING);
3394
3395    count_vm_event(PGFAULT);
3396    mem_cgroup_count_vm_event(mm, PGFAULT);
3397
3398    /* do counter updates before entering really critical section. */
3399    check_sync_rss_stat(current);
3400
3401    if (unlikely(is_vm_hugetlb_page(vma)))
3402        return hugetlb_fault(mm, vma, address, flags);
3403
3404    pgd = pgd_offset(mm, address);
3405    pud = pud_alloc(mm, pgd, address);
3406    if (!pud)
3407        return VM_FAULT_OOM;
3408    pmd = pmd_alloc(mm, pud, address);
3409    if (!pmd)
3410        return VM_FAULT_OOM;
3411    if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3412        if (!vma->vm_ops)
3413            return do_huge_pmd_anonymous_page(mm, vma, address,
3414                              pmd, flags);
3415    } else {
3416        pmd_t orig_pmd = *pmd;
3417        barrier();
3418        if (pmd_trans_huge(orig_pmd)) {
3419            if (flags & FAULT_FLAG_WRITE &&
3420                !pmd_write(orig_pmd) &&
3421                !pmd_trans_splitting(orig_pmd))
3422                return do_huge_pmd_wp_page(mm, vma, address,
3423                               pmd, orig_pmd);
3424            return 0;
3425        }
3426    }
3427
3428    /*
3429     * Use __pte_alloc instead of pte_alloc_map, because we can't
3430     * run pte_offset_map on the pmd, if an huge pmd could
3431     * materialize from under us from a different thread.
3432     */
3433    if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3434        return VM_FAULT_OOM;
3435    /* if an huge pmd materialized from under us just retry later */
3436    if (unlikely(pmd_trans_huge(*pmd)))
3437        return 0;
3438    /*
3439     * A regular pmd is established and it can't morph into a huge pmd
3440     * from under us anymore at this point because we hold the mmap_sem
3441     * read mode and khugepaged takes it in write mode. So now it's
3442     * safe to run pte_offset_map().
3443     */
3444    pte = pte_offset_map(pmd, address);
3445
3446    return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3447}
3448
3449#ifndef __PAGETABLE_PUD_FOLDED
3450/*
3451 * Allocate page upper directory.
3452 * We've already handled the fast-path in-line.
3453 */
3454int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3455{
3456    pud_t *new = pud_alloc_one(mm, address);
3457    if (!new)
3458        return -ENOMEM;
3459
3460    smp_wmb(); /* See comment in __pte_alloc */
3461
3462    spin_lock(&mm->page_table_lock);
3463    if (pgd_present(*pgd)) /* Another has populated it */
3464        pud_free(mm, new);
3465    else
3466        pgd_populate(mm, pgd, new);
3467    spin_unlock(&mm->page_table_lock);
3468    return 0;
3469}
3470#endif /* __PAGETABLE_PUD_FOLDED */
3471
3472#ifndef __PAGETABLE_PMD_FOLDED
3473/*
3474 * Allocate page middle directory.
3475 * We've already handled the fast-path in-line.
3476 */
3477int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3478{
3479    pmd_t *new = pmd_alloc_one(mm, address);
3480    if (!new)
3481        return -ENOMEM;
3482
3483    smp_wmb(); /* See comment in __pte_alloc */
3484
3485    spin_lock(&mm->page_table_lock);
3486#ifndef __ARCH_HAS_4LEVEL_HACK
3487    if (pud_present(*pud)) /* Another has populated it */
3488        pmd_free(mm, new);
3489    else
3490        pud_populate(mm, pud, new);
3491#else
3492    if (pgd_present(*pud)) /* Another has populated it */
3493        pmd_free(mm, new);
3494    else
3495        pgd_populate(mm, pud, new);
3496#endif /* __ARCH_HAS_4LEVEL_HACK */
3497    spin_unlock(&mm->page_table_lock);
3498    return 0;
3499}
3500#endif /* __PAGETABLE_PMD_FOLDED */
3501
3502int make_pages_present(unsigned long addr, unsigned long end)
3503{
3504    int ret, len, write;
3505    struct vm_area_struct * vma;
3506
3507    vma = find_vma(current->mm, addr);
3508    if (!vma)
3509        return -ENOMEM;
3510    /*
3511     * We want to touch writable mappings with a write fault in order
3512     * to break COW, except for shared mappings because these don't COW
3513     * and we would not want to dirty them for nothing.
3514     */
3515    write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3516    BUG_ON(addr >= end);
3517    BUG_ON(end > vma->vm_end);
3518    len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3519    ret = get_user_pages(current, current->mm, addr,
3520            len, write, 0, NULL, NULL);
3521    if (ret < 0)
3522        return ret;
3523    return ret == len ? 0 : -EFAULT;
3524}
3525
3526#if !defined(__HAVE_ARCH_GATE_AREA)
3527
3528#if defined(AT_SYSINFO_EHDR)
3529static struct vm_area_struct gate_vma;
3530
3531static int __init gate_vma_init(void)
3532{
3533    gate_vma.vm_mm = NULL;
3534    gate_vma.vm_start = FIXADDR_USER_START;
3535    gate_vma.vm_end = FIXADDR_USER_END;
3536    gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3537    gate_vma.vm_page_prot = __P101;
3538    /*
3539     * Make sure the vDSO gets into every core dump.
3540     * Dumping its contents makes post-mortem fully interpretable later
3541     * without matching up the same kernel and hardware config to see
3542     * what PC values meant.
3543     */
3544    gate_vma.vm_flags |= VM_ALWAYSDUMP;
3545    return 0;
3546}
3547__initcall(gate_vma_init);
3548#endif
3549
3550struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3551{
3552#ifdef AT_SYSINFO_EHDR
3553    return &gate_vma;
3554#else
3555    return NULL;
3556#endif
3557}
3558
3559int in_gate_area_no_mm(unsigned long addr)
3560{
3561#ifdef AT_SYSINFO_EHDR
3562    if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3563        return 1;
3564#endif
3565    return 0;
3566}
3567
3568#endif /* __HAVE_ARCH_GATE_AREA */
3569
3570static int __follow_pte(struct mm_struct *mm, unsigned long address,
3571        pte_t **ptepp, spinlock_t **ptlp)
3572{
3573    pgd_t *pgd;
3574    pud_t *pud;
3575    pmd_t *pmd;
3576    pte_t *ptep;
3577
3578    pgd = pgd_offset(mm, address);
3579    if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3580        goto out;
3581
3582    pud = pud_offset(pgd, address);
3583    if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3584        goto out;
3585
3586    pmd = pmd_offset(pud, address);
3587    VM_BUG_ON(pmd_trans_huge(*pmd));
3588    if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3589        goto out;
3590
3591    /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3592    if (pmd_huge(*pmd))
3593        goto out;
3594
3595    ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3596    if (!ptep)
3597        goto out;
3598    if (!pte_present(*ptep))
3599        goto unlock;
3600    *ptepp = ptep;
3601    return 0;
3602unlock:
3603    pte_unmap_unlock(ptep, *ptlp);
3604out:
3605    return -EINVAL;
3606}
3607
3608static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3609                 pte_t **ptepp, spinlock_t **ptlp)
3610{
3611    int res;
3612
3613    /* (void) is needed to make gcc happy */
3614    (void) __cond_lock(*ptlp,
3615               !(res = __follow_pte(mm, address, ptepp, ptlp)));
3616    return res;
3617}
3618
3619/**
3620 * follow_pfn - look up PFN at a user virtual address
3621 * @vma: memory mapping
3622 * @address: user virtual address
3623 * @pfn: location to store found PFN
3624 *
3625 * Only IO mappings and raw PFN mappings are allowed.
3626 *
3627 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3628 */
3629int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3630    unsigned long *pfn)
3631{
3632    int ret = -EINVAL;
3633    spinlock_t *ptl;
3634    pte_t *ptep;
3635
3636    if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3637        return ret;
3638
3639    ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3640    if (ret)
3641        return ret;
3642    *pfn = pte_pfn(*ptep);
3643    pte_unmap_unlock(ptep, ptl);
3644    return 0;
3645}
3646EXPORT_SYMBOL(follow_pfn);
3647
3648#ifdef CONFIG_HAVE_IOREMAP_PROT
3649int follow_phys(struct vm_area_struct *vma,
3650        unsigned long address, unsigned int flags,
3651        unsigned long *prot, resource_size_t *phys)
3652{
3653    int ret = -EINVAL;
3654    pte_t *ptep, pte;
3655    spinlock_t *ptl;
3656
3657    if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3658        goto out;
3659
3660    if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3661        goto out;
3662    pte = *ptep;
3663
3664    if ((flags & FOLL_WRITE) && !pte_write(pte))
3665        goto unlock;
3666
3667    *prot = pgprot_val(pte_pgprot(pte));
3668    *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3669
3670    ret = 0;
3671unlock:
3672    pte_unmap_unlock(ptep, ptl);
3673out:
3674    return ret;
3675}
3676
3677int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3678            void *buf, int len, int write)
3679{
3680    resource_size_t phys_addr;
3681    unsigned long prot = 0;
3682    void __iomem *maddr;
3683    int offset = addr & (PAGE_SIZE-1);
3684
3685    if (follow_phys(vma, addr, write, &prot, &phys_addr))
3686        return -EINVAL;
3687
3688    maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3689    if (write)
3690        memcpy_toio(maddr + offset, buf, len);
3691    else
3692        memcpy_fromio(buf, maddr + offset, len);
3693    iounmap(maddr);
3694
3695    return len;
3696}
3697#endif
3698
3699/*
3700 * Access another process' address space as given in mm. If non-NULL, use the
3701 * given task for page fault accounting.
3702 */
3703static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3704        unsigned long addr, void *buf, int len, int write)
3705{
3706    struct vm_area_struct *vma;
3707    void *old_buf = buf;
3708
3709    down_read(&mm->mmap_sem);
3710    /* ignore errors, just check how much was successfully transferred */
3711    while (len) {
3712        int bytes, ret, offset;
3713        void *maddr;
3714        struct page *page = NULL;
3715
3716        ret = get_user_pages(tsk, mm, addr, 1,
3717                write, 1, &page, &vma);
3718        if (ret <= 0) {
3719            /*
3720             * Check if this is a VM_IO | VM_PFNMAP VMA, which
3721             * we can access using slightly different code.
3722             */
3723#ifdef CONFIG_HAVE_IOREMAP_PROT
3724            vma = find_vma(mm, addr);
3725            if (!vma || vma->vm_start > addr)
3726                break;
3727            if (vma->vm_ops && vma->vm_ops->access)
3728                ret = vma->vm_ops->access(vma, addr, buf,
3729                              len, write);
3730            if (ret <= 0)
3731#endif
3732                break;
3733            bytes = ret;
3734        } else {
3735            bytes = len;
3736            offset = addr & (PAGE_SIZE-1);
3737            if (bytes > PAGE_SIZE-offset)
3738                bytes = PAGE_SIZE-offset;
3739
3740            maddr = kmap(page);
3741            if (write) {
3742                copy_to_user_page(vma, page, addr,
3743                          maddr + offset, buf, bytes);
3744                set_page_dirty_lock(page);
3745            } else {
3746                copy_from_user_page(vma, page, addr,
3747                            buf, maddr + offset, bytes);
3748            }
3749            kunmap(page);
3750            page_cache_release(page);
3751        }
3752        len -= bytes;
3753        buf += bytes;
3754        addr += bytes;
3755    }
3756    up_read(&mm->mmap_sem);
3757
3758    return buf - old_buf;
3759}
3760
3761/**
3762 * access_remote_vm - access another process' address space
3763 * @mm: the mm_struct of the target address space
3764 * @addr: start address to access
3765 * @buf: source or destination buffer
3766 * @len: number of bytes to transfer
3767 * @write: whether the access is a write
3768 *
3769 * The caller must hold a reference on @mm.
3770 */
3771int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3772        void *buf, int len, int write)
3773{
3774    return __access_remote_vm(NULL, mm, addr, buf, len, write);
3775}
3776
3777/*
3778 * Access another process' address space.
3779 * Source/target buffer must be kernel space,
3780 * Do not walk the page table directly, use get_user_pages
3781 */
3782int access_process_vm(struct task_struct *tsk, unsigned long addr,
3783        void *buf, int len, int write)
3784{
3785    struct mm_struct *mm;
3786    int ret;
3787
3788    mm = get_task_mm(tsk);
3789    if (!mm)
3790        return 0;
3791
3792    ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3793    mmput(mm);
3794
3795    return ret;
3796}
3797
3798/*
3799 * Print the name of a VMA.
3800 */
3801void print_vma_addr(char *prefix, unsigned long ip)
3802{
3803    struct mm_struct *mm = current->mm;
3804    struct vm_area_struct *vma;
3805
3806    /*
3807     * Do not print if we are in atomic
3808     * contexts (in exception stacks, etc.):
3809     */
3810    if (preempt_count())
3811        return;
3812
3813    down_read(&mm->mmap_sem);
3814    vma = find_vma(mm, ip);
3815    if (vma && vma->vm_file) {
3816        struct file *f = vma->vm_file;
3817        char *buf = (char *)__get_free_page(GFP_KERNEL);
3818        if (buf) {
3819            char *p, *s;
3820
3821            p = d_path(&f->f_path, buf, PAGE_SIZE);
3822            if (IS_ERR(p))
3823                p = "?";
3824            s = strrchr(p, '/');
3825            if (s)
3826                p = s+1;
3827            printk("%s%s[%lx+%lx]", prefix, p,
3828                    vma->vm_start,
3829                    vma->vm_end - vma->vm_start);
3830            free_page((unsigned long)buf);
3831        }
3832    }
3833    up_read(&current->mm->mmap_sem);
3834}
3835
3836#ifdef CONFIG_PROVE_LOCKING
3837void might_fault(void)
3838{
3839    /*
3840     * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3841     * holding the mmap_sem, this is safe because kernel memory doesn't
3842     * get paged out, therefore we'll never actually fault, and the
3843     * below annotations will generate false positives.
3844     */
3845    if (segment_eq(get_fs(), KERNEL_DS))
3846        return;
3847
3848    might_sleep();
3849    /*
3850     * it would be nicer only to annotate paths which are not under
3851     * pagefault_disable, however that requires a larger audit and
3852     * providing helpers like get_user_atomic.
3853     */
3854    if (!in_atomic() && current->mm)
3855        might_lock_read(&current->mm->mmap_sem);
3856}
3857EXPORT_SYMBOL(might_fault);
3858#endif
3859
3860#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3861static void clear_gigantic_page(struct page *page,
3862                unsigned long addr,
3863                unsigned int pages_per_huge_page)
3864{
3865    int i;
3866    struct page *p = page;
3867
3868    might_sleep();
3869    for (i = 0; i < pages_per_huge_page;
3870         i++, p = mem_map_next(p, page, i)) {
3871        cond_resched();
3872        clear_user_highpage(p, addr + i * PAGE_SIZE);
3873    }
3874}
3875void clear_huge_page(struct page *page,
3876             unsigned long addr, unsigned int pages_per_huge_page)
3877{
3878    int i;
3879
3880    if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3881        clear_gigantic_page(page, addr, pages_per_huge_page);
3882        return;
3883    }
3884
3885    might_sleep();
3886    for (i = 0; i < pages_per_huge_page; i++) {
3887        cond_resched();
3888        clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3889    }
3890}
3891
3892static void copy_user_gigantic_page(struct page *dst, struct page *src,
3893                    unsigned long addr,
3894                    struct vm_area_struct *vma,
3895                    unsigned int pages_per_huge_page)
3896{
3897    int i;
3898    struct page *dst_base = dst;
3899    struct page *src_base = src;
3900
3901    for (i = 0; i < pages_per_huge_page; ) {
3902        cond_resched();
3903        copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3904
3905        i++;
3906        dst = mem_map_next(dst, dst_base, i);
3907        src = mem_map_next(src, src_base, i);
3908    }
3909}
3910
3911void copy_user_huge_page(struct page *dst, struct page *src,
3912             unsigned long addr, struct vm_area_struct *vma,
3913             unsigned int pages_per_huge_page)
3914{
3915    int i;
3916
3917    if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3918        copy_user_gigantic_page(dst, src, addr, vma,
3919                    pages_per_huge_page);
3920        return;
3921    }
3922
3923    might_sleep();
3924    for (i = 0; i < pages_per_huge_page; i++) {
3925        cond_resched();
3926        copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3927    }
3928}
3929#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3930

Archive Download this file



interactive