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/export.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#include <linux/migrate.h>
61#include <linux/string.h>
62#include <linux/dma-debug.h>
63#include <linux/debugfs.h>
64
65#include <asm/io.h>
66#include <asm/pgalloc.h>
67#include <asm/uaccess.h>
68#include <asm/tlb.h>
69#include <asm/tlbflush.h>
70#include <asm/pgtable.h>
71
72#include "internal.h"
73
74#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
76#endif
77
78#ifndef CONFIG_NEED_MULTIPLE_NODES
79/* use the per-pgdat data instead for discontigmem - mbligh */
80unsigned long max_mapnr;
81struct page *mem_map;
82
83EXPORT_SYMBOL(max_mapnr);
84EXPORT_SYMBOL(mem_map);
85#endif
86
87/*
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92 * and ZONE_HIGHMEM.
93 */
94void * high_memory;
95
96EXPORT_SYMBOL(high_memory);
97
98/*
99 * Randomize the address space (stacks, mmaps, brk, etc.).
100 *
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
103 */
104int randomize_va_space __read_mostly =
105#ifdef CONFIG_COMPAT_BRK
106                    1;
107#else
108                    2;
109#endif
110
111static int __init disable_randmaps(char *s)
112{
113    randomize_va_space = 0;
114    return 1;
115}
116__setup("norandmaps", disable_randmaps);
117
118unsigned long zero_pfn __read_mostly;
119unsigned long highest_memmap_pfn __read_mostly;
120
121/*
122 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123 */
124static int __init init_zero_pfn(void)
125{
126    zero_pfn = page_to_pfn(ZERO_PAGE(0));
127    return 0;
128}
129core_initcall(init_zero_pfn);
130
131
132#if defined(SPLIT_RSS_COUNTING)
133
134void sync_mm_rss(struct mm_struct *mm)
135{
136    int i;
137
138    for (i = 0; i < NR_MM_COUNTERS; i++) {
139        if (current->rss_stat.count[i]) {
140            add_mm_counter(mm, i, current->rss_stat.count[i]);
141            current->rss_stat.count[i] = 0;
142        }
143    }
144    current->rss_stat.events = 0;
145}
146
147static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148{
149    struct task_struct *task = current;
150
151    if (likely(task->mm == mm))
152        task->rss_stat.count[member] += val;
153    else
154        add_mm_counter(mm, member, val);
155}
156#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159/* sync counter once per 64 page faults */
160#define TASK_RSS_EVENTS_THRESH (64)
161static void check_sync_rss_stat(struct task_struct *task)
162{
163    if (unlikely(task != current))
164        return;
165    if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
166        sync_mm_rss(task->mm);
167}
168#else /* SPLIT_RSS_COUNTING */
169
170#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173static void check_sync_rss_stat(struct task_struct *task)
174{
175}
176
177#endif /* SPLIT_RSS_COUNTING */
178
179#ifdef HAVE_GENERIC_MMU_GATHER
180
181static int tlb_next_batch(struct mmu_gather *tlb)
182{
183    struct mmu_gather_batch *batch;
184
185    batch = tlb->active;
186    if (batch->next) {
187        tlb->active = batch->next;
188        return 1;
189    }
190
191    if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192        return 0;
193
194    batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195    if (!batch)
196        return 0;
197
198    tlb->batch_count++;
199    batch->next = NULL;
200    batch->nr = 0;
201    batch->max = MAX_GATHER_BATCH;
202
203    tlb->active->next = batch;
204    tlb->active = batch;
205
206    return 1;
207}
208
209/* tlb_gather_mmu
210 * Called to initialize an (on-stack) mmu_gather structure for page-table
211 * tear-down from @mm. The @fullmm argument is used when @mm is without
212 * users and we're going to destroy the full address space (exit/execve).
213 */
214void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
215{
216    tlb->mm = mm;
217
218    /* Is it from 0 to ~0? */
219    tlb->fullmm = !(start | (end+1));
220    tlb->need_flush_all = 0;
221    tlb->start = start;
222    tlb->end = end;
223    tlb->need_flush = 0;
224    tlb->local.next = NULL;
225    tlb->local.nr = 0;
226    tlb->local.max = ARRAY_SIZE(tlb->__pages);
227    tlb->active = &tlb->local;
228    tlb->batch_count = 0;
229
230#ifdef CONFIG_HAVE_RCU_TABLE_FREE
231    tlb->batch = NULL;
232#endif
233}
234
235static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
236{
237    tlb->need_flush = 0;
238    tlb_flush(tlb);
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240    tlb_table_flush(tlb);
241#endif
242}
243
244static void tlb_flush_mmu_free(struct mmu_gather *tlb)
245{
246    struct mmu_gather_batch *batch;
247
248    for (batch = &tlb->local; batch; batch = batch->next) {
249        free_pages_and_swap_cache(batch->pages, batch->nr);
250        batch->nr = 0;
251    }
252    tlb->active = &tlb->local;
253}
254
255void tlb_flush_mmu(struct mmu_gather *tlb)
256{
257    if (!tlb->need_flush)
258        return;
259    tlb_flush_mmu_tlbonly(tlb);
260    tlb_flush_mmu_free(tlb);
261}
262
263/* tlb_finish_mmu
264 * Called at the end of the shootdown operation to free up any resources
265 * that were required.
266 */
267void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
268{
269    struct mmu_gather_batch *batch, *next;
270
271    tlb_flush_mmu(tlb);
272
273    /* keep the page table cache within bounds */
274    check_pgt_cache();
275
276    for (batch = tlb->local.next; batch; batch = next) {
277        next = batch->next;
278        free_pages((unsigned long)batch, 0);
279    }
280    tlb->local.next = NULL;
281}
282
283/* __tlb_remove_page
284 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
285 * handling the additional races in SMP caused by other CPUs caching valid
286 * mappings in their TLBs. Returns the number of free page slots left.
287 * When out of page slots we must call tlb_flush_mmu().
288 */
289int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
290{
291    struct mmu_gather_batch *batch;
292
293    VM_BUG_ON(!tlb->need_flush);
294
295    batch = tlb->active;
296    batch->pages[batch->nr++] = page;
297    if (batch->nr == batch->max) {
298        if (!tlb_next_batch(tlb))
299            return 0;
300        batch = tlb->active;
301    }
302    VM_BUG_ON_PAGE(batch->nr > batch->max, page);
303
304    return batch->max - batch->nr;
305}
306
307#endif /* HAVE_GENERIC_MMU_GATHER */
308
309#ifdef CONFIG_HAVE_RCU_TABLE_FREE
310
311/*
312 * See the comment near struct mmu_table_batch.
313 */
314
315static void tlb_remove_table_smp_sync(void *arg)
316{
317    /* Simply deliver the interrupt */
318}
319
320static void tlb_remove_table_one(void *table)
321{
322    /*
323     * This isn't an RCU grace period and hence the page-tables cannot be
324     * assumed to be actually RCU-freed.
325     *
326     * It is however sufficient for software page-table walkers that rely on
327     * IRQ disabling. See the comment near struct mmu_table_batch.
328     */
329    smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
330    __tlb_remove_table(table);
331}
332
333static void tlb_remove_table_rcu(struct rcu_head *head)
334{
335    struct mmu_table_batch *batch;
336    int i;
337
338    batch = container_of(head, struct mmu_table_batch, rcu);
339
340    for (i = 0; i < batch->nr; i++)
341        __tlb_remove_table(batch->tables[i]);
342
343    free_page((unsigned long)batch);
344}
345
346void tlb_table_flush(struct mmu_gather *tlb)
347{
348    struct mmu_table_batch **batch = &tlb->batch;
349
350    if (*batch) {
351        call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
352        *batch = NULL;
353    }
354}
355
356void tlb_remove_table(struct mmu_gather *tlb, void *table)
357{
358    struct mmu_table_batch **batch = &tlb->batch;
359
360    tlb->need_flush = 1;
361
362    /*
363     * When there's less then two users of this mm there cannot be a
364     * concurrent page-table walk.
365     */
366    if (atomic_read(&tlb->mm->mm_users) < 2) {
367        __tlb_remove_table(table);
368        return;
369    }
370
371    if (*batch == NULL) {
372        *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
373        if (*batch == NULL) {
374            tlb_remove_table_one(table);
375            return;
376        }
377        (*batch)->nr = 0;
378    }
379    (*batch)->tables[(*batch)->nr++] = table;
380    if ((*batch)->nr == MAX_TABLE_BATCH)
381        tlb_table_flush(tlb);
382}
383
384#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
385
386/*
387 * Note: this doesn't free the actual pages themselves. That
388 * has been handled earlier when unmapping all the memory regions.
389 */
390static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
391               unsigned long addr)
392{
393    pgtable_t token = pmd_pgtable(*pmd);
394    pmd_clear(pmd);
395    pte_free_tlb(tlb, token, addr);
396    atomic_long_dec(&tlb->mm->nr_ptes);
397}
398
399static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
400                unsigned long addr, unsigned long end,
401                unsigned long floor, unsigned long ceiling)
402{
403    pmd_t *pmd;
404    unsigned long next;
405    unsigned long start;
406
407    start = addr;
408    pmd = pmd_offset(pud, addr);
409    do {
410        next = pmd_addr_end(addr, end);
411        if (pmd_none_or_clear_bad(pmd))
412            continue;
413        free_pte_range(tlb, pmd, addr);
414    } while (pmd++, addr = next, addr != end);
415
416    start &= PUD_MASK;
417    if (start < floor)
418        return;
419    if (ceiling) {
420        ceiling &= PUD_MASK;
421        if (!ceiling)
422            return;
423    }
424    if (end - 1 > ceiling - 1)
425        return;
426
427    pmd = pmd_offset(pud, start);
428    pud_clear(pud);
429    pmd_free_tlb(tlb, pmd, start);
430}
431
432static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
433                unsigned long addr, unsigned long end,
434                unsigned long floor, unsigned long ceiling)
435{
436    pud_t *pud;
437    unsigned long next;
438    unsigned long start;
439
440    start = addr;
441    pud = pud_offset(pgd, addr);
442    do {
443        next = pud_addr_end(addr, end);
444        if (pud_none_or_clear_bad(pud))
445            continue;
446        free_pmd_range(tlb, pud, addr, next, floor, ceiling);
447    } while (pud++, addr = next, addr != end);
448
449    start &= PGDIR_MASK;
450    if (start < floor)
451        return;
452    if (ceiling) {
453        ceiling &= PGDIR_MASK;
454        if (!ceiling)
455            return;
456    }
457    if (end - 1 > ceiling - 1)
458        return;
459
460    pud = pud_offset(pgd, start);
461    pgd_clear(pgd);
462    pud_free_tlb(tlb, pud, start);
463}
464
465/*
466 * This function frees user-level page tables of a process.
467 */
468void free_pgd_range(struct mmu_gather *tlb,
469            unsigned long addr, unsigned long end,
470            unsigned long floor, unsigned long ceiling)
471{
472    pgd_t *pgd;
473    unsigned long next;
474
475    /*
476     * The next few lines have given us lots of grief...
477     *
478     * Why are we testing PMD* at this top level? Because often
479     * there will be no work to do at all, and we'd prefer not to
480     * go all the way down to the bottom just to discover that.
481     *
482     * Why all these "- 1"s? Because 0 represents both the bottom
483     * of the address space and the top of it (using -1 for the
484     * top wouldn't help much: the masks would do the wrong thing).
485     * The rule is that addr 0 and floor 0 refer to the bottom of
486     * the address space, but end 0 and ceiling 0 refer to the top
487     * Comparisons need to use "end - 1" and "ceiling - 1" (though
488     * that end 0 case should be mythical).
489     *
490     * Wherever addr is brought up or ceiling brought down, we must
491     * be careful to reject "the opposite 0" before it confuses the
492     * subsequent tests. But what about where end is brought down
493     * by PMD_SIZE below? no, end can't go down to 0 there.
494     *
495     * Whereas we round start (addr) and ceiling down, by different
496     * masks at different levels, in order to test whether a table
497     * now has no other vmas using it, so can be freed, we don't
498     * bother to round floor or end up - the tests don't need that.
499     */
500
501    addr &= PMD_MASK;
502    if (addr < floor) {
503        addr += PMD_SIZE;
504        if (!addr)
505            return;
506    }
507    if (ceiling) {
508        ceiling &= PMD_MASK;
509        if (!ceiling)
510            return;
511    }
512    if (end - 1 > ceiling - 1)
513        end -= PMD_SIZE;
514    if (addr > end - 1)
515        return;
516
517    pgd = pgd_offset(tlb->mm, addr);
518    do {
519        next = pgd_addr_end(addr, end);
520        if (pgd_none_or_clear_bad(pgd))
521            continue;
522        free_pud_range(tlb, pgd, addr, next, floor, ceiling);
523    } while (pgd++, addr = next, addr != end);
524}
525
526void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
527        unsigned long floor, unsigned long ceiling)
528{
529    while (vma) {
530        struct vm_area_struct *next = vma->vm_next;
531        unsigned long addr = vma->vm_start;
532
533        /*
534         * Hide vma from rmap and truncate_pagecache before freeing
535         * pgtables
536         */
537        unlink_anon_vmas(vma);
538        unlink_file_vma(vma);
539
540        if (is_vm_hugetlb_page(vma)) {
541            hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
542                floor, next? next->vm_start: ceiling);
543        } else {
544            /*
545             * Optimization: gather nearby vmas into one call down
546             */
547            while (next && next->vm_start <= vma->vm_end + PMD_SIZE
548                   && !is_vm_hugetlb_page(next)) {
549                vma = next;
550                next = vma->vm_next;
551                unlink_anon_vmas(vma);
552                unlink_file_vma(vma);
553            }
554            free_pgd_range(tlb, addr, vma->vm_end,
555                floor, next? next->vm_start: ceiling);
556        }
557        vma = next;
558    }
559}
560
561int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
562        pmd_t *pmd, unsigned long address)
563{
564    spinlock_t *ptl;
565    pgtable_t new = pte_alloc_one(mm, address);
566    int wait_split_huge_page;
567    if (!new)
568        return -ENOMEM;
569
570    /*
571     * Ensure all pte setup (eg. pte page lock and page clearing) are
572     * visible before the pte is made visible to other CPUs by being
573     * put into page tables.
574     *
575     * The other side of the story is the pointer chasing in the page
576     * table walking code (when walking the page table without locking;
577     * ie. most of the time). Fortunately, these data accesses consist
578     * of a chain of data-dependent loads, meaning most CPUs (alpha
579     * being the notable exception) will already guarantee loads are
580     * seen in-order. See the alpha page table accessors for the
581     * smp_read_barrier_depends() barriers in page table walking code.
582     */
583    smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
584
585    ptl = pmd_lock(mm, pmd);
586    wait_split_huge_page = 0;
587    if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
588        atomic_long_inc(&mm->nr_ptes);
589        pmd_populate(mm, pmd, new);
590        new = NULL;
591    } else if (unlikely(pmd_trans_splitting(*pmd)))
592        wait_split_huge_page = 1;
593    spin_unlock(ptl);
594    if (new)
595        pte_free(mm, new);
596    if (wait_split_huge_page)
597        wait_split_huge_page(vma->anon_vma, pmd);
598    return 0;
599}
600
601int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
602{
603    pte_t *new = pte_alloc_one_kernel(&init_mm, address);
604    if (!new)
605        return -ENOMEM;
606
607    smp_wmb(); /* See comment in __pte_alloc */
608
609    spin_lock(&init_mm.page_table_lock);
610    if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
611        pmd_populate_kernel(&init_mm, pmd, new);
612        new = NULL;
613    } else
614        VM_BUG_ON(pmd_trans_splitting(*pmd));
615    spin_unlock(&init_mm.page_table_lock);
616    if (new)
617        pte_free_kernel(&init_mm, new);
618    return 0;
619}
620
621static inline void init_rss_vec(int *rss)
622{
623    memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
624}
625
626static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
627{
628    int i;
629
630    if (current->mm == mm)
631        sync_mm_rss(mm);
632    for (i = 0; i < NR_MM_COUNTERS; i++)
633        if (rss[i])
634            add_mm_counter(mm, i, rss[i]);
635}
636
637/*
638 * This function is called to print an error when a bad pte
639 * is found. For example, we might have a PFN-mapped pte in
640 * a region that doesn't allow it.
641 *
642 * The calling function must still handle the error.
643 */
644static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
645              pte_t pte, struct page *page)
646{
647    pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
648    pud_t *pud = pud_offset(pgd, addr);
649    pmd_t *pmd = pmd_offset(pud, addr);
650    struct address_space *mapping;
651    pgoff_t index;
652    static unsigned long resume;
653    static unsigned long nr_shown;
654    static unsigned long nr_unshown;
655
656    /*
657     * Allow a burst of 60 reports, then keep quiet for that minute;
658     * or allow a steady drip of one report per second.
659     */
660    if (nr_shown == 60) {
661        if (time_before(jiffies, resume)) {
662            nr_unshown++;
663            return;
664        }
665        if (nr_unshown) {
666            printk(KERN_ALERT
667                "BUG: Bad page map: %lu messages suppressed\n",
668                nr_unshown);
669            nr_unshown = 0;
670        }
671        nr_shown = 0;
672    }
673    if (nr_shown++ == 0)
674        resume = jiffies + 60 * HZ;
675
676    mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
677    index = linear_page_index(vma, addr);
678
679    printk(KERN_ALERT
680        "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
681        current->comm,
682        (long long)pte_val(pte), (long long)pmd_val(*pmd));
683    if (page)
684        dump_page(page, "bad pte");
685    printk(KERN_ALERT
686        "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
687        (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
688    /*
689     * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
690     */
691    if (vma->vm_ops)
692        printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
693               vma->vm_ops->fault);
694    if (vma->vm_file)
695        printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
696               vma->vm_file->f_op->mmap);
697    dump_stack();
698    add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
699}
700
701/*
702 * vm_normal_page -- This function gets the "struct page" associated with a pte.
703 *
704 * "Special" mappings do not wish to be associated with a "struct page" (either
705 * it doesn't exist, or it exists but they don't want to touch it). In this
706 * case, NULL is returned here. "Normal" mappings do have a struct page.
707 *
708 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
709 * pte bit, in which case this function is trivial. Secondly, an architecture
710 * may not have a spare pte bit, which requires a more complicated scheme,
711 * described below.
712 *
713 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
714 * special mapping (even if there are underlying and valid "struct pages").
715 * COWed pages of a VM_PFNMAP are always normal.
716 *
717 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
718 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
719 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
720 * mapping will always honor the rule
721 *
722 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
723 *
724 * And for normal mappings this is false.
725 *
726 * This restricts such mappings to be a linear translation from virtual address
727 * to pfn. To get around this restriction, we allow arbitrary mappings so long
728 * as the vma is not a COW mapping; in that case, we know that all ptes are
729 * special (because none can have been COWed).
730 *
731 *
732 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
733 *
734 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
735 * page" backing, however the difference is that _all_ pages with a struct
736 * page (that is, those where pfn_valid is true) are refcounted and considered
737 * normal pages by the VM. The disadvantage is that pages are refcounted
738 * (which can be slower and simply not an option for some PFNMAP users). The
739 * advantage is that we don't have to follow the strict linearity rule of
740 * PFNMAP mappings in order to support COWable mappings.
741 *
742 */
743#ifdef __HAVE_ARCH_PTE_SPECIAL
744# define HAVE_PTE_SPECIAL 1
745#else
746# define HAVE_PTE_SPECIAL 0
747#endif
748struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
749                pte_t pte)
750{
751    unsigned long pfn = pte_pfn(pte);
752
753    if (HAVE_PTE_SPECIAL) {
754        if (likely(!pte_special(pte) || pte_numa(pte)))
755            goto check_pfn;
756        if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
757            return NULL;
758        if (!is_zero_pfn(pfn))
759            print_bad_pte(vma, addr, pte, NULL);
760        return NULL;
761    }
762
763    /* !HAVE_PTE_SPECIAL case follows: */
764
765    if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
766        if (vma->vm_flags & VM_MIXEDMAP) {
767            if (!pfn_valid(pfn))
768                return NULL;
769            goto out;
770        } else {
771            unsigned long off;
772            off = (addr - vma->vm_start) >> PAGE_SHIFT;
773            if (pfn == vma->vm_pgoff + off)
774                return NULL;
775            if (!is_cow_mapping(vma->vm_flags))
776                return NULL;
777        }
778    }
779
780check_pfn:
781    if (unlikely(pfn > highest_memmap_pfn)) {
782        print_bad_pte(vma, addr, pte, NULL);
783        return NULL;
784    }
785
786    if (is_zero_pfn(pfn))
787        return NULL;
788
789    /*
790     * NOTE! We still have PageReserved() pages in the page tables.
791     * eg. VDSO mappings can cause them to exist.
792     */
793out:
794    return pfn_to_page(pfn);
795}
796
797/*
798 * copy one vm_area from one task to the other. Assumes the page tables
799 * already present in the new task to be cleared in the whole range
800 * covered by this vma.
801 */
802
803static inline unsigned long
804copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
805        pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
806        unsigned long addr, int *rss)
807{
808    unsigned long vm_flags = vma->vm_flags;
809    pte_t pte = *src_pte;
810    struct page *page;
811
812    /* pte contains position in swap or file, so copy. */
813    if (unlikely(!pte_present(pte))) {
814        if (!pte_file(pte)) {
815            swp_entry_t entry = pte_to_swp_entry(pte);
816
817            if (swap_duplicate(entry) < 0)
818                return entry.val;
819
820            /* make sure dst_mm is on swapoff's mmlist. */
821            if (unlikely(list_empty(&dst_mm->mmlist))) {
822                spin_lock(&mmlist_lock);
823                if (list_empty(&dst_mm->mmlist))
824                    list_add(&dst_mm->mmlist,
825                         &src_mm->mmlist);
826                spin_unlock(&mmlist_lock);
827            }
828            if (likely(!non_swap_entry(entry)))
829                rss[MM_SWAPENTS]++;
830            else if (is_migration_entry(entry)) {
831                page = migration_entry_to_page(entry);
832
833                if (PageAnon(page))
834                    rss[MM_ANONPAGES]++;
835                else
836                    rss[MM_FILEPAGES]++;
837
838                if (is_write_migration_entry(entry) &&
839                    is_cow_mapping(vm_flags)) {
840                    /*
841                     * COW mappings require pages in both
842                     * parent and child to be set to read.
843                     */
844                    make_migration_entry_read(&entry);
845                    pte = swp_entry_to_pte(entry);
846                    if (pte_swp_soft_dirty(*src_pte))
847                        pte = pte_swp_mksoft_dirty(pte);
848                    set_pte_at(src_mm, addr, src_pte, pte);
849                }
850            }
851        }
852        goto out_set_pte;
853    }
854
855    /*
856     * If it's a COW mapping, write protect it both
857     * in the parent and the child
858     */
859    if (is_cow_mapping(vm_flags)) {
860        ptep_set_wrprotect(src_mm, addr, src_pte);
861        pte = pte_wrprotect(pte);
862    }
863
864    /*
865     * If it's a shared mapping, mark it clean in
866     * the child
867     */
868    if (vm_flags & VM_SHARED)
869        pte = pte_mkclean(pte);
870    pte = pte_mkold(pte);
871
872    page = vm_normal_page(vma, addr, pte);
873    if (page) {
874        get_page(page);
875        page_dup_rmap(page);
876        if (PageAnon(page))
877            rss[MM_ANONPAGES]++;
878        else
879            rss[MM_FILEPAGES]++;
880    }
881
882out_set_pte:
883    set_pte_at(dst_mm, addr, dst_pte, pte);
884    return 0;
885}
886
887int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
888           pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
889           unsigned long addr, unsigned long end)
890{
891    pte_t *orig_src_pte, *orig_dst_pte;
892    pte_t *src_pte, *dst_pte;
893    spinlock_t *src_ptl, *dst_ptl;
894    int progress = 0;
895    int rss[NR_MM_COUNTERS];
896    swp_entry_t entry = (swp_entry_t){0};
897
898again:
899    init_rss_vec(rss);
900
901    dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
902    if (!dst_pte)
903        return -ENOMEM;
904    src_pte = pte_offset_map(src_pmd, addr);
905    src_ptl = pte_lockptr(src_mm, src_pmd);
906    spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
907    orig_src_pte = src_pte;
908    orig_dst_pte = dst_pte;
909    arch_enter_lazy_mmu_mode();
910
911    do {
912        /*
913         * We are holding two locks at this point - either of them
914         * could generate latencies in another task on another CPU.
915         */
916        if (progress >= 32) {
917            progress = 0;
918            if (need_resched() ||
919                spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
920                break;
921        }
922        if (pte_none(*src_pte)) {
923            progress++;
924            continue;
925        }
926        entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
927                            vma, addr, rss);
928        if (entry.val)
929            break;
930        progress += 8;
931    } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
932
933    arch_leave_lazy_mmu_mode();
934    spin_unlock(src_ptl);
935    pte_unmap(orig_src_pte);
936    add_mm_rss_vec(dst_mm, rss);
937    pte_unmap_unlock(orig_dst_pte, dst_ptl);
938    cond_resched();
939
940    if (entry.val) {
941        if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
942            return -ENOMEM;
943        progress = 0;
944    }
945    if (addr != end)
946        goto again;
947    return 0;
948}
949
950static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
951        pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
952        unsigned long addr, unsigned long end)
953{
954    pmd_t *src_pmd, *dst_pmd;
955    unsigned long next;
956
957    dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
958    if (!dst_pmd)
959        return -ENOMEM;
960    src_pmd = pmd_offset(src_pud, addr);
961    do {
962        next = pmd_addr_end(addr, end);
963        if (pmd_trans_huge(*src_pmd)) {
964            int err;
965            VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
966            err = copy_huge_pmd(dst_mm, src_mm,
967                        dst_pmd, src_pmd, addr, vma);
968            if (err == -ENOMEM)
969                return -ENOMEM;
970            if (!err)
971                continue;
972            /* fall through */
973        }
974        if (pmd_none_or_clear_bad(src_pmd))
975            continue;
976        if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
977                        vma, addr, next))
978            return -ENOMEM;
979    } while (dst_pmd++, src_pmd++, addr = next, addr != end);
980    return 0;
981}
982
983static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
984        pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
985        unsigned long addr, unsigned long end)
986{
987    pud_t *src_pud, *dst_pud;
988    unsigned long next;
989
990    dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
991    if (!dst_pud)
992        return -ENOMEM;
993    src_pud = pud_offset(src_pgd, addr);
994    do {
995        next = pud_addr_end(addr, end);
996        if (pud_none_or_clear_bad(src_pud))
997            continue;
998        if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
999                        vma, addr, next))
1000            return -ENOMEM;
1001    } while (dst_pud++, src_pud++, addr = next, addr != end);
1002    return 0;
1003}
1004
1005int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1006        struct vm_area_struct *vma)
1007{
1008    pgd_t *src_pgd, *dst_pgd;
1009    unsigned long next;
1010    unsigned long addr = vma->vm_start;
1011    unsigned long end = vma->vm_end;
1012    unsigned long mmun_start; /* For mmu_notifiers */
1013    unsigned long mmun_end; /* For mmu_notifiers */
1014    bool is_cow;
1015    int ret;
1016
1017    /*
1018     * Don't copy ptes where a page fault will fill them correctly.
1019     * Fork becomes much lighter when there are big shared or private
1020     * readonly mappings. The tradeoff is that copy_page_range is more
1021     * efficient than faulting.
1022     */
1023    if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1024                   VM_PFNMAP | VM_MIXEDMAP))) {
1025        if (!vma->anon_vma)
1026            return 0;
1027    }
1028
1029    if (is_vm_hugetlb_page(vma))
1030        return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1031
1032    if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1033        /*
1034         * We do not free on error cases below as remove_vma
1035         * gets called on error from higher level routine
1036         */
1037        ret = track_pfn_copy(vma);
1038        if (ret)
1039            return ret;
1040    }
1041
1042    /*
1043     * We need to invalidate the secondary MMU mappings only when
1044     * there could be a permission downgrade on the ptes of the
1045     * parent mm. And a permission downgrade will only happen if
1046     * is_cow_mapping() returns true.
1047     */
1048    is_cow = is_cow_mapping(vma->vm_flags);
1049    mmun_start = addr;
1050    mmun_end = end;
1051    if (is_cow)
1052        mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1053                            mmun_end);
1054
1055    ret = 0;
1056    dst_pgd = pgd_offset(dst_mm, addr);
1057    src_pgd = pgd_offset(src_mm, addr);
1058    do {
1059        next = pgd_addr_end(addr, end);
1060        if (pgd_none_or_clear_bad(src_pgd))
1061            continue;
1062        if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1063                        vma, addr, next))) {
1064            ret = -ENOMEM;
1065            break;
1066        }
1067    } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1068
1069    if (is_cow)
1070        mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1071    return ret;
1072}
1073
1074static unsigned long zap_pte_range(struct mmu_gather *tlb,
1075                struct vm_area_struct *vma, pmd_t *pmd,
1076                unsigned long addr, unsigned long end,
1077                struct zap_details *details)
1078{
1079    struct mm_struct *mm = tlb->mm;
1080    int force_flush = 0;
1081    int rss[NR_MM_COUNTERS];
1082    spinlock_t *ptl;
1083    pte_t *start_pte;
1084    pte_t *pte;
1085
1086again:
1087    init_rss_vec(rss);
1088    start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1089    pte = start_pte;
1090    arch_enter_lazy_mmu_mode();
1091    do {
1092        pte_t ptent = *pte;
1093        if (pte_none(ptent)) {
1094            continue;
1095        }
1096
1097        if (pte_present(ptent)) {
1098            struct page *page;
1099
1100            page = vm_normal_page(vma, addr, ptent);
1101            if (unlikely(details) && page) {
1102                /*
1103                 * unmap_shared_mapping_pages() wants to
1104                 * invalidate cache without truncating:
1105                 * unmap shared but keep private pages.
1106                 */
1107                if (details->check_mapping &&
1108                    details->check_mapping != page->mapping)
1109                    continue;
1110                /*
1111                 * Each page->index must be checked when
1112                 * invalidating or truncating nonlinear.
1113                 */
1114                if (details->nonlinear_vma &&
1115                    (page->index < details->first_index ||
1116                     page->index > details->last_index))
1117                    continue;
1118            }
1119            ptent = ptep_get_and_clear_full(mm, addr, pte,
1120                            tlb->fullmm);
1121            tlb_remove_tlb_entry(tlb, pte, addr);
1122            if (unlikely(!page))
1123                continue;
1124            if (unlikely(details) && details->nonlinear_vma
1125                && linear_page_index(details->nonlinear_vma,
1126                        addr) != page->index) {
1127                pte_t ptfile = pgoff_to_pte(page->index);
1128                if (pte_soft_dirty(ptent))
1129                    pte_file_mksoft_dirty(ptfile);
1130                set_pte_at(mm, addr, pte, ptfile);
1131            }
1132            if (PageAnon(page))
1133                rss[MM_ANONPAGES]--;
1134            else {
1135                if (pte_dirty(ptent)) {
1136                    force_flush = 1;
1137                    set_page_dirty(page);
1138                }
1139                if (pte_young(ptent) &&
1140                    likely(!(vma->vm_flags & VM_SEQ_READ)))
1141                    mark_page_accessed(page);
1142                rss[MM_FILEPAGES]--;
1143            }
1144            page_remove_rmap(page);
1145            if (unlikely(page_mapcount(page) < 0))
1146                print_bad_pte(vma, addr, ptent, page);
1147            if (unlikely(!__tlb_remove_page(tlb, page))) {
1148                force_flush = 1;
1149                break;
1150            }
1151            continue;
1152        }
1153        /*
1154         * If details->check_mapping, we leave swap entries;
1155         * if details->nonlinear_vma, we leave file entries.
1156         */
1157        if (unlikely(details))
1158            continue;
1159        if (pte_file(ptent)) {
1160            if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1161                print_bad_pte(vma, addr, ptent, NULL);
1162        } else {
1163            swp_entry_t entry = pte_to_swp_entry(ptent);
1164
1165            if (!non_swap_entry(entry))
1166                rss[MM_SWAPENTS]--;
1167            else if (is_migration_entry(entry)) {
1168                struct page *page;
1169
1170                page = migration_entry_to_page(entry);
1171
1172                if (PageAnon(page))
1173                    rss[MM_ANONPAGES]--;
1174                else
1175                    rss[MM_FILEPAGES]--;
1176            }
1177            if (unlikely(!free_swap_and_cache(entry)))
1178                print_bad_pte(vma, addr, ptent, NULL);
1179        }
1180        pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1181    } while (pte++, addr += PAGE_SIZE, addr != end);
1182
1183    add_mm_rss_vec(mm, rss);
1184    arch_leave_lazy_mmu_mode();
1185
1186    /* Do the actual TLB flush before dropping ptl */
1187    if (force_flush) {
1188        unsigned long old_end;
1189
1190        /*
1191         * Flush the TLB just for the previous segment,
1192         * then update the range to be the remaining
1193         * TLB range.
1194         */
1195        old_end = tlb->end;
1196        tlb->end = addr;
1197        tlb_flush_mmu_tlbonly(tlb);
1198        tlb->start = addr;
1199        tlb->end = old_end;
1200    }
1201    pte_unmap_unlock(start_pte, ptl);
1202
1203    /*
1204     * If we forced a TLB flush (either due to running out of
1205     * batch buffers or because we needed to flush dirty TLB
1206     * entries before releasing the ptl), free the batched
1207     * memory too. Restart if we didn't do everything.
1208     */
1209    if (force_flush) {
1210        force_flush = 0;
1211        tlb_flush_mmu_free(tlb);
1212
1213        if (addr != end)
1214            goto again;
1215    }
1216
1217    return addr;
1218}
1219
1220static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1221                struct vm_area_struct *vma, pud_t *pud,
1222                unsigned long addr, unsigned long end,
1223                struct zap_details *details)
1224{
1225    pmd_t *pmd;
1226    unsigned long next;
1227
1228    pmd = pmd_offset(pud, addr);
1229    do {
1230        next = pmd_addr_end(addr, end);
1231        if (pmd_trans_huge(*pmd)) {
1232            if (next - addr != HPAGE_PMD_SIZE) {
1233#ifdef CONFIG_DEBUG_VM
1234                if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1235                    pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1236                        __func__, addr, end,
1237                        vma->vm_start,
1238                        vma->vm_end);
1239                    BUG();
1240                }
1241#endif
1242                split_huge_page_pmd(vma, addr, pmd);
1243            } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1244                goto next;
1245            /* fall through */
1246        }
1247        /*
1248         * Here there can be other concurrent MADV_DONTNEED or
1249         * trans huge page faults running, and if the pmd is
1250         * none or trans huge it can change under us. This is
1251         * because MADV_DONTNEED holds the mmap_sem in read
1252         * mode.
1253         */
1254        if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1255            goto next;
1256        next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1257next:
1258        cond_resched();
1259    } while (pmd++, addr = next, addr != end);
1260
1261    return addr;
1262}
1263
1264static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1265                struct vm_area_struct *vma, pgd_t *pgd,
1266                unsigned long addr, unsigned long end,
1267                struct zap_details *details)
1268{
1269    pud_t *pud;
1270    unsigned long next;
1271
1272    pud = pud_offset(pgd, addr);
1273    do {
1274        next = pud_addr_end(addr, end);
1275        if (pud_none_or_clear_bad(pud))
1276            continue;
1277        next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1278    } while (pud++, addr = next, addr != end);
1279
1280    return addr;
1281}
1282
1283static void unmap_page_range(struct mmu_gather *tlb,
1284                 struct vm_area_struct *vma,
1285                 unsigned long addr, unsigned long end,
1286                 struct zap_details *details)
1287{
1288    pgd_t *pgd;
1289    unsigned long next;
1290
1291    if (details && !details->check_mapping && !details->nonlinear_vma)
1292        details = NULL;
1293
1294    BUG_ON(addr >= end);
1295    mem_cgroup_uncharge_start();
1296    tlb_start_vma(tlb, vma);
1297    pgd = pgd_offset(vma->vm_mm, addr);
1298    do {
1299        next = pgd_addr_end(addr, end);
1300        if (pgd_none_or_clear_bad(pgd))
1301            continue;
1302        next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1303    } while (pgd++, addr = next, addr != end);
1304    tlb_end_vma(tlb, vma);
1305    mem_cgroup_uncharge_end();
1306}
1307
1308
1309static void unmap_single_vma(struct mmu_gather *tlb,
1310        struct vm_area_struct *vma, unsigned long start_addr,
1311        unsigned long end_addr,
1312        struct zap_details *details)
1313{
1314    unsigned long start = max(vma->vm_start, start_addr);
1315    unsigned long end;
1316
1317    if (start >= vma->vm_end)
1318        return;
1319    end = min(vma->vm_end, end_addr);
1320    if (end <= vma->vm_start)
1321        return;
1322
1323    if (vma->vm_file)
1324        uprobe_munmap(vma, start, end);
1325
1326    if (unlikely(vma->vm_flags & VM_PFNMAP))
1327        untrack_pfn(vma, 0, 0);
1328
1329    if (start != end) {
1330        if (unlikely(is_vm_hugetlb_page(vma))) {
1331            /*
1332             * It is undesirable to test vma->vm_file as it
1333             * should be non-null for valid hugetlb area.
1334             * However, vm_file will be NULL in the error
1335             * cleanup path of mmap_region. When
1336             * hugetlbfs ->mmap method fails,
1337             * mmap_region() nullifies vma->vm_file
1338             * before calling this function to clean up.
1339             * Since no pte has actually been setup, it is
1340             * safe to do nothing in this case.
1341             */
1342            if (vma->vm_file) {
1343                mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1344                __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1345                mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1346            }
1347        } else
1348            unmap_page_range(tlb, vma, start, end, details);
1349    }
1350}
1351
1352/**
1353 * unmap_vmas - unmap a range of memory covered by a list of vma's
1354 * @tlb: address of the caller's struct mmu_gather
1355 * @vma: the starting vma
1356 * @start_addr: virtual address at which to start unmapping
1357 * @end_addr: virtual address at which to end unmapping
1358 *
1359 * Unmap all pages in the vma list.
1360 *
1361 * Only addresses between `start' and `end' will be unmapped.
1362 *
1363 * The VMA list must be sorted in ascending virtual address order.
1364 *
1365 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1366 * range after unmap_vmas() returns. So the only responsibility here is to
1367 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1368 * drops the lock and schedules.
1369 */
1370void unmap_vmas(struct mmu_gather *tlb,
1371        struct vm_area_struct *vma, unsigned long start_addr,
1372        unsigned long end_addr)
1373{
1374    struct mm_struct *mm = vma->vm_mm;
1375
1376    mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1377    for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1378        unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1379    mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1380}
1381
1382/**
1383 * zap_page_range - remove user pages in a given range
1384 * @vma: vm_area_struct holding the applicable pages
1385 * @start: starting address of pages to zap
1386 * @size: number of bytes to zap
1387 * @details: details of nonlinear truncation or shared cache invalidation
1388 *
1389 * Caller must protect the VMA list
1390 */
1391void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1392        unsigned long size, struct zap_details *details)
1393{
1394    struct mm_struct *mm = vma->vm_mm;
1395    struct mmu_gather tlb;
1396    unsigned long end = start + size;
1397
1398    lru_add_drain();
1399    tlb_gather_mmu(&tlb, mm, start, end);
1400    update_hiwater_rss(mm);
1401    mmu_notifier_invalidate_range_start(mm, start, end);
1402    for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1403        unmap_single_vma(&tlb, vma, start, end, details);
1404    mmu_notifier_invalidate_range_end(mm, start, end);
1405    tlb_finish_mmu(&tlb, start, end);
1406}
1407
1408/**
1409 * zap_page_range_single - remove user pages in a given range
1410 * @vma: vm_area_struct holding the applicable pages
1411 * @address: starting address of pages to zap
1412 * @size: number of bytes to zap
1413 * @details: details of nonlinear truncation or shared cache invalidation
1414 *
1415 * The range must fit into one VMA.
1416 */
1417static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1418        unsigned long size, struct zap_details *details)
1419{
1420    struct mm_struct *mm = vma->vm_mm;
1421    struct mmu_gather tlb;
1422    unsigned long end = address + size;
1423
1424    lru_add_drain();
1425    tlb_gather_mmu(&tlb, mm, address, end);
1426    update_hiwater_rss(mm);
1427    mmu_notifier_invalidate_range_start(mm, address, end);
1428    unmap_single_vma(&tlb, vma, address, end, details);
1429    mmu_notifier_invalidate_range_end(mm, address, end);
1430    tlb_finish_mmu(&tlb, address, end);
1431}
1432
1433/**
1434 * zap_vma_ptes - remove ptes mapping the vma
1435 * @vma: vm_area_struct holding ptes to be zapped
1436 * @address: starting address of pages to zap
1437 * @size: number of bytes to zap
1438 *
1439 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1440 *
1441 * The entire address range must be fully contained within the vma.
1442 *
1443 * Returns 0 if successful.
1444 */
1445int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1446        unsigned long size)
1447{
1448    if (address < vma->vm_start || address + size > vma->vm_end ||
1449                !(vma->vm_flags & VM_PFNMAP))
1450        return -1;
1451    zap_page_range_single(vma, address, size, NULL);
1452    return 0;
1453}
1454EXPORT_SYMBOL_GPL(zap_vma_ptes);
1455
1456pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1457            spinlock_t **ptl)
1458{
1459    pgd_t * pgd = pgd_offset(mm, addr);
1460    pud_t * pud = pud_alloc(mm, pgd, addr);
1461    if (pud) {
1462        pmd_t * pmd = pmd_alloc(mm, pud, addr);
1463        if (pmd) {
1464            VM_BUG_ON(pmd_trans_huge(*pmd));
1465            return pte_alloc_map_lock(mm, pmd, addr, ptl);
1466        }
1467    }
1468    return NULL;
1469}
1470
1471/*
1472 * This is the old fallback for page remapping.
1473 *
1474 * For historical reasons, it only allows reserved pages. Only
1475 * old drivers should use this, and they needed to mark their
1476 * pages reserved for the old functions anyway.
1477 */
1478static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1479            struct page *page, pgprot_t prot)
1480{
1481    struct mm_struct *mm = vma->vm_mm;
1482    int retval;
1483    pte_t *pte;
1484    spinlock_t *ptl;
1485
1486    retval = -EINVAL;
1487    if (PageAnon(page))
1488        goto out;
1489    retval = -ENOMEM;
1490    flush_dcache_page(page);
1491    pte = get_locked_pte(mm, addr, &ptl);
1492    if (!pte)
1493        goto out;
1494    retval = -EBUSY;
1495    if (!pte_none(*pte))
1496        goto out_unlock;
1497
1498    /* Ok, finally just insert the thing.. */
1499    get_page(page);
1500    inc_mm_counter_fast(mm, MM_FILEPAGES);
1501    page_add_file_rmap(page);
1502    set_pte_at(mm, addr, pte, mk_pte(page, prot));
1503
1504    retval = 0;
1505    pte_unmap_unlock(pte, ptl);
1506    return retval;
1507out_unlock:
1508    pte_unmap_unlock(pte, ptl);
1509out:
1510    return retval;
1511}
1512
1513/**
1514 * vm_insert_page - insert single page into user vma
1515 * @vma: user vma to map to
1516 * @addr: target user address of this page
1517 * @page: source kernel page
1518 *
1519 * This allows drivers to insert individual pages they've allocated
1520 * into a user vma.
1521 *
1522 * The page has to be a nice clean _individual_ kernel allocation.
1523 * If you allocate a compound page, you need to have marked it as
1524 * such (__GFP_COMP), or manually just split the page up yourself
1525 * (see split_page()).
1526 *
1527 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1528 * took an arbitrary page protection parameter. This doesn't allow
1529 * that. Your vma protection will have to be set up correctly, which
1530 * means that if you want a shared writable mapping, you'd better
1531 * ask for a shared writable mapping!
1532 *
1533 * The page does not need to be reserved.
1534 *
1535 * Usually this function is called from f_op->mmap() handler
1536 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1537 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1538 * function from other places, for example from page-fault handler.
1539 */
1540int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1541            struct page *page)
1542{
1543    if (addr < vma->vm_start || addr >= vma->vm_end)
1544        return -EFAULT;
1545    if (!page_count(page))
1546        return -EINVAL;
1547    if (!(vma->vm_flags & VM_MIXEDMAP)) {
1548        BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1549        BUG_ON(vma->vm_flags & VM_PFNMAP);
1550        vma->vm_flags |= VM_MIXEDMAP;
1551    }
1552    return insert_page(vma, addr, page, vma->vm_page_prot);
1553}
1554EXPORT_SYMBOL(vm_insert_page);
1555
1556static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1557            unsigned long pfn, pgprot_t prot)
1558{
1559    struct mm_struct *mm = vma->vm_mm;
1560    int retval;
1561    pte_t *pte, entry;
1562    spinlock_t *ptl;
1563
1564    retval = -ENOMEM;
1565    pte = get_locked_pte(mm, addr, &ptl);
1566    if (!pte)
1567        goto out;
1568    retval = -EBUSY;
1569    if (!pte_none(*pte))
1570        goto out_unlock;
1571
1572    /* Ok, finally just insert the thing.. */
1573    entry = pte_mkspecial(pfn_pte(pfn, prot));
1574    set_pte_at(mm, addr, pte, entry);
1575    update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1576
1577    retval = 0;
1578out_unlock:
1579    pte_unmap_unlock(pte, ptl);
1580out:
1581    return retval;
1582}
1583
1584/**
1585 * vm_insert_pfn - insert single pfn into user vma
1586 * @vma: user vma to map to
1587 * @addr: target user address of this page
1588 * @pfn: source kernel pfn
1589 *
1590 * Similar to vm_insert_page, this allows drivers to insert individual pages
1591 * they've allocated into a user vma. Same comments apply.
1592 *
1593 * This function should only be called from a vm_ops->fault handler, and
1594 * in that case the handler should return NULL.
1595 *
1596 * vma cannot be a COW mapping.
1597 *
1598 * As this is called only for pages that do not currently exist, we
1599 * do not need to flush old virtual caches or the TLB.
1600 */
1601int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1602            unsigned long pfn)
1603{
1604    int ret;
1605    pgprot_t pgprot = vma->vm_page_prot;
1606    /*
1607     * Technically, architectures with pte_special can avoid all these
1608     * restrictions (same for remap_pfn_range). However we would like
1609     * consistency in testing and feature parity among all, so we should
1610     * try to keep these invariants in place for everybody.
1611     */
1612    BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1613    BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1614                        (VM_PFNMAP|VM_MIXEDMAP));
1615    BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1616    BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1617
1618    if (addr < vma->vm_start || addr >= vma->vm_end)
1619        return -EFAULT;
1620    if (track_pfn_insert(vma, &pgprot, pfn))
1621        return -EINVAL;
1622
1623    ret = insert_pfn(vma, addr, pfn, pgprot);
1624
1625    return ret;
1626}
1627EXPORT_SYMBOL(vm_insert_pfn);
1628
1629int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1630            unsigned long pfn)
1631{
1632    BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1633
1634    if (addr < vma->vm_start || addr >= vma->vm_end)
1635        return -EFAULT;
1636
1637    /*
1638     * If we don't have pte special, then we have to use the pfn_valid()
1639     * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1640     * refcount the page if pfn_valid is true (hence insert_page rather
1641     * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1642     * without pte special, it would there be refcounted as a normal page.
1643     */
1644    if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1645        struct page *page;
1646
1647        page = pfn_to_page(pfn);
1648        return insert_page(vma, addr, page, vma->vm_page_prot);
1649    }
1650    return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1651}
1652EXPORT_SYMBOL(vm_insert_mixed);
1653
1654/*
1655 * maps a range of physical memory into the requested pages. the old
1656 * mappings are removed. any references to nonexistent pages results
1657 * in null mappings (currently treated as "copy-on-access")
1658 */
1659static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1660            unsigned long addr, unsigned long end,
1661            unsigned long pfn, pgprot_t prot)
1662{
1663    pte_t *pte;
1664    spinlock_t *ptl;
1665
1666    pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1667    if (!pte)
1668        return -ENOMEM;
1669    arch_enter_lazy_mmu_mode();
1670    do {
1671        BUG_ON(!pte_none(*pte));
1672        set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1673        pfn++;
1674    } while (pte++, addr += PAGE_SIZE, addr != end);
1675    arch_leave_lazy_mmu_mode();
1676    pte_unmap_unlock(pte - 1, ptl);
1677    return 0;
1678}
1679
1680static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1681            unsigned long addr, unsigned long end,
1682            unsigned long pfn, pgprot_t prot)
1683{
1684    pmd_t *pmd;
1685    unsigned long next;
1686
1687    pfn -= addr >> PAGE_SHIFT;
1688    pmd = pmd_alloc(mm, pud, addr);
1689    if (!pmd)
1690        return -ENOMEM;
1691    VM_BUG_ON(pmd_trans_huge(*pmd));
1692    do {
1693        next = pmd_addr_end(addr, end);
1694        if (remap_pte_range(mm, pmd, addr, next,
1695                pfn + (addr >> PAGE_SHIFT), prot))
1696            return -ENOMEM;
1697    } while (pmd++, addr = next, addr != end);
1698    return 0;
1699}
1700
1701static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1702            unsigned long addr, unsigned long end,
1703            unsigned long pfn, pgprot_t prot)
1704{
1705    pud_t *pud;
1706    unsigned long next;
1707
1708    pfn -= addr >> PAGE_SHIFT;
1709    pud = pud_alloc(mm, pgd, addr);
1710    if (!pud)
1711        return -ENOMEM;
1712    do {
1713        next = pud_addr_end(addr, end);
1714        if (remap_pmd_range(mm, pud, addr, next,
1715                pfn + (addr >> PAGE_SHIFT), prot))
1716            return -ENOMEM;
1717    } while (pud++, addr = next, addr != end);
1718    return 0;
1719}
1720
1721/**
1722 * remap_pfn_range - remap kernel memory to userspace
1723 * @vma: user vma to map to
1724 * @addr: target user address to start at
1725 * @pfn: physical address of kernel memory
1726 * @size: size of map area
1727 * @prot: page protection flags for this mapping
1728 *
1729 * Note: this is only safe if the mm semaphore is held when called.
1730 */
1731int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1732            unsigned long pfn, unsigned long size, pgprot_t prot)
1733{
1734    pgd_t *pgd;
1735    unsigned long next;
1736    unsigned long end = addr + PAGE_ALIGN(size);
1737    struct mm_struct *mm = vma->vm_mm;
1738    int err;
1739
1740    /*
1741     * Physically remapped pages are special. Tell the
1742     * rest of the world about it:
1743     * VM_IO tells people not to look at these pages
1744     * (accesses can have side effects).
1745     * VM_PFNMAP tells the core MM that the base pages are just
1746     * raw PFN mappings, and do not have a "struct page" associated
1747     * with them.
1748     * VM_DONTEXPAND
1749     * Disable vma merging and expanding with mremap().
1750     * VM_DONTDUMP
1751     * Omit vma from core dump, even when VM_IO turned off.
1752     *
1753     * There's a horrible special case to handle copy-on-write
1754     * behaviour that some programs depend on. We mark the "original"
1755     * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1756     * See vm_normal_page() for details.
1757     */
1758    if (is_cow_mapping(vma->vm_flags)) {
1759        if (addr != vma->vm_start || end != vma->vm_end)
1760            return -EINVAL;
1761        vma->vm_pgoff = pfn;
1762    }
1763
1764    err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1765    if (err)
1766        return -EINVAL;
1767
1768    vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1769
1770    BUG_ON(addr >= end);
1771    pfn -= addr >> PAGE_SHIFT;
1772    pgd = pgd_offset(mm, addr);
1773    flush_cache_range(vma, addr, end);
1774    do {
1775        next = pgd_addr_end(addr, end);
1776        err = remap_pud_range(mm, pgd, addr, next,
1777                pfn + (addr >> PAGE_SHIFT), prot);
1778        if (err)
1779            break;
1780    } while (pgd++, addr = next, addr != end);
1781
1782    if (err)
1783        untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1784
1785    return err;
1786}
1787EXPORT_SYMBOL(remap_pfn_range);
1788
1789/**
1790 * vm_iomap_memory - remap memory to userspace
1791 * @vma: user vma to map to
1792 * @start: start of area
1793 * @len: size of area
1794 *
1795 * This is a simplified io_remap_pfn_range() for common driver use. The
1796 * driver just needs to give us the physical memory range to be mapped,
1797 * we'll figure out the rest from the vma information.
1798 *
1799 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1800 * whatever write-combining details or similar.
1801 */
1802int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1803{
1804    unsigned long vm_len, pfn, pages;
1805
1806    /* Check that the physical memory area passed in looks valid */
1807    if (start + len < start)
1808        return -EINVAL;
1809    /*
1810     * You *really* shouldn't map things that aren't page-aligned,
1811     * but we've historically allowed it because IO memory might
1812     * just have smaller alignment.
1813     */
1814    len += start & ~PAGE_MASK;
1815    pfn = start >> PAGE_SHIFT;
1816    pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1817    if (pfn + pages < pfn)
1818        return -EINVAL;
1819
1820    /* We start the mapping 'vm_pgoff' pages into the area */
1821    if (vma->vm_pgoff > pages)
1822        return -EINVAL;
1823    pfn += vma->vm_pgoff;
1824    pages -= vma->vm_pgoff;
1825
1826    /* Can we fit all of the mapping? */
1827    vm_len = vma->vm_end - vma->vm_start;
1828    if (vm_len >> PAGE_SHIFT > pages)
1829        return -EINVAL;
1830
1831    /* Ok, let it rip */
1832    return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1833}
1834EXPORT_SYMBOL(vm_iomap_memory);
1835
1836static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1837                     unsigned long addr, unsigned long end,
1838                     pte_fn_t fn, void *data)
1839{
1840    pte_t *pte;
1841    int err;
1842    pgtable_t token;
1843    spinlock_t *uninitialized_var(ptl);
1844
1845    pte = (mm == &init_mm) ?
1846        pte_alloc_kernel(pmd, addr) :
1847        pte_alloc_map_lock(mm, pmd, addr, &ptl);
1848    if (!pte)
1849        return -ENOMEM;
1850
1851    BUG_ON(pmd_huge(*pmd));
1852
1853    arch_enter_lazy_mmu_mode();
1854
1855    token = pmd_pgtable(*pmd);
1856
1857    do {
1858        err = fn(pte++, token, addr, data);
1859        if (err)
1860            break;
1861    } while (addr += PAGE_SIZE, addr != end);
1862
1863    arch_leave_lazy_mmu_mode();
1864
1865    if (mm != &init_mm)
1866        pte_unmap_unlock(pte-1, ptl);
1867    return err;
1868}
1869
1870static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1871                     unsigned long addr, unsigned long end,
1872                     pte_fn_t fn, void *data)
1873{
1874    pmd_t *pmd;
1875    unsigned long next;
1876    int err;
1877
1878    BUG_ON(pud_huge(*pud));
1879
1880    pmd = pmd_alloc(mm, pud, addr);
1881    if (!pmd)
1882        return -ENOMEM;
1883    do {
1884        next = pmd_addr_end(addr, end);
1885        err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1886        if (err)
1887            break;
1888    } while (pmd++, addr = next, addr != end);
1889    return err;
1890}
1891
1892static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1893                     unsigned long addr, unsigned long end,
1894                     pte_fn_t fn, void *data)
1895{
1896    pud_t *pud;
1897    unsigned long next;
1898    int err;
1899
1900    pud = pud_alloc(mm, pgd, addr);
1901    if (!pud)
1902        return -ENOMEM;
1903    do {
1904        next = pud_addr_end(addr, end);
1905        err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1906        if (err)
1907            break;
1908    } while (pud++, addr = next, addr != end);
1909    return err;
1910}
1911
1912/*
1913 * Scan a region of virtual memory, filling in page tables as necessary
1914 * and calling a provided function on each leaf page table.
1915 */
1916int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1917            unsigned long size, pte_fn_t fn, void *data)
1918{
1919    pgd_t *pgd;
1920    unsigned long next;
1921    unsigned long end = addr + size;
1922    int err;
1923
1924    BUG_ON(addr >= end);
1925    pgd = pgd_offset(mm, addr);
1926    do {
1927        next = pgd_addr_end(addr, end);
1928        err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1929        if (err)
1930            break;
1931    } while (pgd++, addr = next, addr != end);
1932
1933    return err;
1934}
1935EXPORT_SYMBOL_GPL(apply_to_page_range);
1936
1937/*
1938 * handle_pte_fault chooses page fault handler according to an entry
1939 * which was read non-atomically. Before making any commitment, on
1940 * those architectures or configurations (e.g. i386 with PAE) which
1941 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
1942 * must check under lock before unmapping the pte and proceeding
1943 * (but do_wp_page is only called after already making such a check;
1944 * and do_anonymous_page can safely check later on).
1945 */
1946static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1947                pte_t *page_table, pte_t orig_pte)
1948{
1949    int same = 1;
1950#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1951    if (sizeof(pte_t) > sizeof(unsigned long)) {
1952        spinlock_t *ptl = pte_lockptr(mm, pmd);
1953        spin_lock(ptl);
1954        same = pte_same(*page_table, orig_pte);
1955        spin_unlock(ptl);
1956    }
1957#endif
1958    pte_unmap(page_table);
1959    return same;
1960}
1961
1962static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1963{
1964    debug_dma_assert_idle(src);
1965
1966    /*
1967     * If the source page was a PFN mapping, we don't have
1968     * a "struct page" for it. We do a best-effort copy by
1969     * just copying from the original user address. If that
1970     * fails, we just zero-fill it. Live with it.
1971     */
1972    if (unlikely(!src)) {
1973        void *kaddr = kmap_atomic(dst);
1974        void __user *uaddr = (void __user *)(va & PAGE_MASK);
1975
1976        /*
1977         * This really shouldn't fail, because the page is there
1978         * in the page tables. But it might just be unreadable,
1979         * in which case we just give up and fill the result with
1980         * zeroes.
1981         */
1982        if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1983            clear_page(kaddr);
1984        kunmap_atomic(kaddr);
1985        flush_dcache_page(dst);
1986    } else
1987        copy_user_highpage(dst, src, va, vma);
1988}
1989
1990/*
1991 * Notify the address space that the page is about to become writable so that
1992 * it can prohibit this or wait for the page to get into an appropriate state.
1993 *
1994 * We do this without the lock held, so that it can sleep if it needs to.
1995 */
1996static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1997           unsigned long address)
1998{
1999    struct vm_fault vmf;
2000    int ret;
2001
2002    vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2003    vmf.pgoff = page->index;
2004    vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2005    vmf.page = page;
2006
2007    ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2008    if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2009        return ret;
2010    if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2011        lock_page(page);
2012        if (!page->mapping) {
2013            unlock_page(page);
2014            return 0; /* retry */
2015        }
2016        ret |= VM_FAULT_LOCKED;
2017    } else
2018        VM_BUG_ON_PAGE(!PageLocked(page), page);
2019    return ret;
2020}
2021
2022/*
2023 * This routine handles present pages, when users try to write
2024 * to a shared page. It is done by copying the page to a new address
2025 * and decrementing the shared-page counter for the old page.
2026 *
2027 * Note that this routine assumes that the protection checks have been
2028 * done by the caller (the low-level page fault routine in most cases).
2029 * Thus we can safely just mark it writable once we've done any necessary
2030 * COW.
2031 *
2032 * We also mark the page dirty at this point even though the page will
2033 * change only once the write actually happens. This avoids a few races,
2034 * and potentially makes it more efficient.
2035 *
2036 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2037 * but allow concurrent faults), with pte both mapped and locked.
2038 * We return with mmap_sem still held, but pte unmapped and unlocked.
2039 */
2040static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2041        unsigned long address, pte_t *page_table, pmd_t *pmd,
2042        spinlock_t *ptl, pte_t orig_pte)
2043    __releases(ptl)
2044{
2045    struct page *old_page, *new_page = NULL;
2046    pte_t entry;
2047    int ret = 0;
2048    int page_mkwrite = 0;
2049    struct page *dirty_page = NULL;
2050    unsigned long mmun_start = 0; /* For mmu_notifiers */
2051    unsigned long mmun_end = 0; /* For mmu_notifiers */
2052
2053    old_page = vm_normal_page(vma, address, orig_pte);
2054    if (!old_page) {
2055        /*
2056         * VM_MIXEDMAP !pfn_valid() case
2057         *
2058         * We should not cow pages in a shared writeable mapping.
2059         * Just mark the pages writable as we can't do any dirty
2060         * accounting on raw pfn maps.
2061         */
2062        if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2063                     (VM_WRITE|VM_SHARED))
2064            goto reuse;
2065        goto gotten;
2066    }
2067
2068    /*
2069     * Take out anonymous pages first, anonymous shared vmas are
2070     * not dirty accountable.
2071     */
2072    if (PageAnon(old_page) && !PageKsm(old_page)) {
2073        if (!trylock_page(old_page)) {
2074            page_cache_get(old_page);
2075            pte_unmap_unlock(page_table, ptl);
2076            lock_page(old_page);
2077            page_table = pte_offset_map_lock(mm, pmd, address,
2078                             &ptl);
2079            if (!pte_same(*page_table, orig_pte)) {
2080                unlock_page(old_page);
2081                goto unlock;
2082            }
2083            page_cache_release(old_page);
2084        }
2085        if (reuse_swap_page(old_page)) {
2086            /*
2087             * The page is all ours. Move it to our anon_vma so
2088             * the rmap code will not search our parent or siblings.
2089             * Protected against the rmap code by the page lock.
2090             */
2091            page_move_anon_rmap(old_page, vma, address);
2092            unlock_page(old_page);
2093            goto reuse;
2094        }
2095        unlock_page(old_page);
2096    } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2097                    (VM_WRITE|VM_SHARED))) {
2098        /*
2099         * Only catch write-faults on shared writable pages,
2100         * read-only shared pages can get COWed by
2101         * get_user_pages(.write=1, .force=1).
2102         */
2103        if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2104            int tmp;
2105            page_cache_get(old_page);
2106            pte_unmap_unlock(page_table, ptl);
2107            tmp = do_page_mkwrite(vma, old_page, address);
2108            if (unlikely(!tmp || (tmp &
2109                    (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2110                page_cache_release(old_page);
2111                return tmp;
2112            }
2113            /*
2114             * Since we dropped the lock we need to revalidate
2115             * the PTE as someone else may have changed it. If
2116             * they did, we just return, as we can count on the
2117             * MMU to tell us if they didn't also make it writable.
2118             */
2119            page_table = pte_offset_map_lock(mm, pmd, address,
2120                             &ptl);
2121            if (!pte_same(*page_table, orig_pte)) {
2122                unlock_page(old_page);
2123                goto unlock;
2124            }
2125
2126            page_mkwrite = 1;
2127        }
2128        dirty_page = old_page;
2129        get_page(dirty_page);
2130
2131reuse:
2132        /*
2133         * Clear the pages cpupid information as the existing
2134         * information potentially belongs to a now completely
2135         * unrelated process.
2136         */
2137        if (old_page)
2138            page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2139
2140        flush_cache_page(vma, address, pte_pfn(orig_pte));
2141        entry = pte_mkyoung(orig_pte);
2142        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2143        if (ptep_set_access_flags(vma, address, page_table, entry,1))
2144            update_mmu_cache(vma, address, page_table);
2145        pte_unmap_unlock(page_table, ptl);
2146        ret |= VM_FAULT_WRITE;
2147
2148        if (!dirty_page)
2149            return ret;
2150
2151        /*
2152         * Yes, Virginia, this is actually required to prevent a race
2153         * with clear_page_dirty_for_io() from clearing the page dirty
2154         * bit after it clear all dirty ptes, but before a racing
2155         * do_wp_page installs a dirty pte.
2156         *
2157         * do_shared_fault is protected similarly.
2158         */
2159        if (!page_mkwrite) {
2160            wait_on_page_locked(dirty_page);
2161            set_page_dirty_balance(dirty_page);
2162            /* file_update_time outside page_lock */
2163            if (vma->vm_file)
2164                file_update_time(vma->vm_file);
2165        }
2166        put_page(dirty_page);
2167        if (page_mkwrite) {
2168            struct address_space *mapping = dirty_page->mapping;
2169
2170            set_page_dirty(dirty_page);
2171            unlock_page(dirty_page);
2172            page_cache_release(dirty_page);
2173            if (mapping) {
2174                /*
2175                 * Some device drivers do not set page.mapping
2176                 * but still dirty their pages
2177                 */
2178                balance_dirty_pages_ratelimited(mapping);
2179            }
2180        }
2181
2182        return ret;
2183    }
2184
2185    /*
2186     * Ok, we need to copy. Oh, well..
2187     */
2188    page_cache_get(old_page);
2189gotten:
2190    pte_unmap_unlock(page_table, ptl);
2191
2192    if (unlikely(anon_vma_prepare(vma)))
2193        goto oom;
2194
2195    if (is_zero_pfn(pte_pfn(orig_pte))) {
2196        new_page = alloc_zeroed_user_highpage_movable(vma, address);
2197        if (!new_page)
2198            goto oom;
2199    } else {
2200        new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2201        if (!new_page)
2202            goto oom;
2203        cow_user_page(new_page, old_page, address, vma);
2204    }
2205    __SetPageUptodate(new_page);
2206
2207    if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))
2208        goto oom_free_new;
2209
2210    mmun_start = address & PAGE_MASK;
2211    mmun_end = mmun_start + PAGE_SIZE;
2212    mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2213
2214    /*
2215     * Re-check the pte - we dropped the lock
2216     */
2217    page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2218    if (likely(pte_same(*page_table, orig_pte))) {
2219        if (old_page) {
2220            if (!PageAnon(old_page)) {
2221                dec_mm_counter_fast(mm, MM_FILEPAGES);
2222                inc_mm_counter_fast(mm, MM_ANONPAGES);
2223            }
2224        } else
2225            inc_mm_counter_fast(mm, MM_ANONPAGES);
2226        flush_cache_page(vma, address, pte_pfn(orig_pte));
2227        entry = mk_pte(new_page, vma->vm_page_prot);
2228        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2229        /*
2230         * Clear the pte entry and flush it first, before updating the
2231         * pte with the new entry. This will avoid a race condition
2232         * seen in the presence of one thread doing SMC and another
2233         * thread doing COW.
2234         */
2235        ptep_clear_flush(vma, address, page_table);
2236        page_add_new_anon_rmap(new_page, vma, address);
2237        /*
2238         * We call the notify macro here because, when using secondary
2239         * mmu page tables (such as kvm shadow page tables), we want the
2240         * new page to be mapped directly into the secondary page table.
2241         */
2242        set_pte_at_notify(mm, address, page_table, entry);
2243        update_mmu_cache(vma, address, page_table);
2244        if (old_page) {
2245            /*
2246             * Only after switching the pte to the new page may
2247             * we remove the mapcount here. Otherwise another
2248             * process may come and find the rmap count decremented
2249             * before the pte is switched to the new page, and
2250             * "reuse" the old page writing into it while our pte
2251             * here still points into it and can be read by other
2252             * threads.
2253             *
2254             * The critical issue is to order this
2255             * page_remove_rmap with the ptp_clear_flush above.
2256             * Those stores are ordered by (if nothing else,)
2257             * the barrier present in the atomic_add_negative
2258             * in page_remove_rmap.
2259             *
2260             * Then the TLB flush in ptep_clear_flush ensures that
2261             * no process can access the old page before the
2262             * decremented mapcount is visible. And the old page
2263             * cannot be reused until after the decremented
2264             * mapcount is visible. So transitively, TLBs to
2265             * old page will be flushed before it can be reused.
2266             */
2267            page_remove_rmap(old_page);
2268        }
2269
2270        /* Free the old page.. */
2271        new_page = old_page;
2272        ret |= VM_FAULT_WRITE;
2273    } else
2274        mem_cgroup_uncharge_page(new_page);
2275
2276    if (new_page)
2277        page_cache_release(new_page);
2278unlock:
2279    pte_unmap_unlock(page_table, ptl);
2280    if (mmun_end > mmun_start)
2281        mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2282    if (old_page) {
2283        /*
2284         * Don't let another task, with possibly unlocked vma,
2285         * keep the mlocked page.
2286         */
2287        if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2288            lock_page(old_page); /* LRU manipulation */
2289            munlock_vma_page(old_page);
2290            unlock_page(old_page);
2291        }
2292        page_cache_release(old_page);
2293    }
2294    return ret;
2295oom_free_new:
2296    page_cache_release(new_page);
2297oom:
2298    if (old_page)
2299        page_cache_release(old_page);
2300    return VM_FAULT_OOM;
2301}
2302
2303static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2304        unsigned long start_addr, unsigned long end_addr,
2305        struct zap_details *details)
2306{
2307    zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2308}
2309
2310static inline void unmap_mapping_range_tree(struct rb_root *root,
2311                        struct zap_details *details)
2312{
2313    struct vm_area_struct *vma;
2314    pgoff_t vba, vea, zba, zea;
2315
2316    vma_interval_tree_foreach(vma, root,
2317            details->first_index, details->last_index) {
2318
2319        vba = vma->vm_pgoff;
2320        vea = vba + vma_pages(vma) - 1;
2321        /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2322        zba = details->first_index;
2323        if (zba < vba)
2324            zba = vba;
2325        zea = details->last_index;
2326        if (zea > vea)
2327            zea = vea;
2328
2329        unmap_mapping_range_vma(vma,
2330            ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2331            ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2332                details);
2333    }
2334}
2335
2336static inline void unmap_mapping_range_list(struct list_head *head,
2337                        struct zap_details *details)
2338{
2339    struct vm_area_struct *vma;
2340
2341    /*
2342     * In nonlinear VMAs there is no correspondence between virtual address
2343     * offset and file offset. So we must perform an exhaustive search
2344     * across *all* the pages in each nonlinear VMA, not just the pages
2345     * whose virtual address lies outside the file truncation point.
2346     */
2347    list_for_each_entry(vma, head, shared.nonlinear) {
2348        details->nonlinear_vma = vma;
2349        unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2350    }
2351}
2352
2353/**
2354 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2355 * @mapping: the address space containing mmaps to be unmapped.
2356 * @holebegin: byte in first page to unmap, relative to the start of
2357 * the underlying file. This will be rounded down to a PAGE_SIZE
2358 * boundary. Note that this is different from truncate_pagecache(), which
2359 * must keep the partial page. In contrast, we must get rid of
2360 * partial pages.
2361 * @holelen: size of prospective hole in bytes. This will be rounded
2362 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2363 * end of the file.
2364 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2365 * but 0 when invalidating pagecache, don't throw away private data.
2366 */
2367void unmap_mapping_range(struct address_space *mapping,
2368        loff_t const holebegin, loff_t const holelen, int even_cows)
2369{
2370    struct zap_details details;
2371    pgoff_t hba = holebegin >> PAGE_SHIFT;
2372    pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2373
2374    /* Check for overflow. */
2375    if (sizeof(holelen) > sizeof(hlen)) {
2376        long long holeend =
2377            (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2378        if (holeend & ~(long long)ULONG_MAX)
2379            hlen = ULONG_MAX - hba + 1;
2380    }
2381
2382    details.check_mapping = even_cows? NULL: mapping;
2383    details.nonlinear_vma = NULL;
2384    details.first_index = hba;
2385    details.last_index = hba + hlen - 1;
2386    if (details.last_index < details.first_index)
2387        details.last_index = ULONG_MAX;
2388
2389
2390    mutex_lock(&mapping->i_mmap_mutex);
2391    if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2392        unmap_mapping_range_tree(&mapping->i_mmap, &details);
2393    if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2394        unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2395    mutex_unlock(&mapping->i_mmap_mutex);
2396}
2397EXPORT_SYMBOL(unmap_mapping_range);
2398
2399/*
2400 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2401 * but allow concurrent faults), and pte mapped but not yet locked.
2402 * We return with mmap_sem still held, but pte unmapped and unlocked.
2403 */
2404static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2405        unsigned long address, pte_t *page_table, pmd_t *pmd,
2406        unsigned int flags, pte_t orig_pte)
2407{
2408    spinlock_t *ptl;
2409    struct page *page, *swapcache;
2410    swp_entry_t entry;
2411    pte_t pte;
2412    int locked;
2413    struct mem_cgroup *ptr;
2414    int exclusive = 0;
2415    int ret = 0;
2416
2417    if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2418        goto out;
2419
2420    entry = pte_to_swp_entry(orig_pte);
2421    if (unlikely(non_swap_entry(entry))) {
2422        if (is_migration_entry(entry)) {
2423            migration_entry_wait(mm, pmd, address);
2424        } else if (is_hwpoison_entry(entry)) {
2425            ret = VM_FAULT_HWPOISON;
2426        } else {
2427            print_bad_pte(vma, address, orig_pte, NULL);
2428            ret = VM_FAULT_SIGBUS;
2429        }
2430        goto out;
2431    }
2432    delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2433    page = lookup_swap_cache(entry);
2434    if (!page) {
2435        page = swapin_readahead(entry,
2436                    GFP_HIGHUSER_MOVABLE, vma, address);
2437        if (!page) {
2438            /*
2439             * Back out if somebody else faulted in this pte
2440             * while we released the pte lock.
2441             */
2442            page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2443            if (likely(pte_same(*page_table, orig_pte)))
2444                ret = VM_FAULT_OOM;
2445            delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2446            goto unlock;
2447        }
2448
2449        /* Had to read the page from swap area: Major fault */
2450        ret = VM_FAULT_MAJOR;
2451        count_vm_event(PGMAJFAULT);
2452        mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2453    } else if (PageHWPoison(page)) {
2454        /*
2455         * hwpoisoned dirty swapcache pages are kept for killing
2456         * owner processes (which may be unknown at hwpoison time)
2457         */
2458        ret = VM_FAULT_HWPOISON;
2459        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2460        swapcache = page;
2461        goto out_release;
2462    }
2463
2464    swapcache = page;
2465    locked = lock_page_or_retry(page, mm, flags);
2466
2467    delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2468    if (!locked) {
2469        ret |= VM_FAULT_RETRY;
2470        goto out_release;
2471    }
2472
2473    /*
2474     * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2475     * release the swapcache from under us. The page pin, and pte_same
2476     * test below, are not enough to exclude that. Even if it is still
2477     * swapcache, we need to check that the page's swap has not changed.
2478     */
2479    if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2480        goto out_page;
2481
2482    page = ksm_might_need_to_copy(page, vma, address);
2483    if (unlikely(!page)) {
2484        ret = VM_FAULT_OOM;
2485        page = swapcache;
2486        goto out_page;
2487    }
2488
2489    if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2490        ret = VM_FAULT_OOM;
2491        goto out_page;
2492    }
2493
2494    /*
2495     * Back out if somebody else already faulted in this pte.
2496     */
2497    page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2498    if (unlikely(!pte_same(*page_table, orig_pte)))
2499        goto out_nomap;
2500
2501    if (unlikely(!PageUptodate(page))) {
2502        ret = VM_FAULT_SIGBUS;
2503        goto out_nomap;
2504    }
2505
2506    /*
2507     * The page isn't present yet, go ahead with the fault.
2508     *
2509     * Be careful about the sequence of operations here.
2510     * To get its accounting right, reuse_swap_page() must be called
2511     * while the page is counted on swap but not yet in mapcount i.e.
2512     * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2513     * must be called after the swap_free(), or it will never succeed.
2514     * Because delete_from_swap_page() may be called by reuse_swap_page(),
2515     * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2516     * in page->private. In this case, a record in swap_cgroup is silently
2517     * discarded at swap_free().
2518     */
2519
2520    inc_mm_counter_fast(mm, MM_ANONPAGES);
2521    dec_mm_counter_fast(mm, MM_SWAPENTS);
2522    pte = mk_pte(page, vma->vm_page_prot);
2523    if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2524        pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2525        flags &= ~FAULT_FLAG_WRITE;
2526        ret |= VM_FAULT_WRITE;
2527        exclusive = 1;
2528    }
2529    flush_icache_page(vma, page);
2530    if (pte_swp_soft_dirty(orig_pte))
2531        pte = pte_mksoft_dirty(pte);
2532    set_pte_at(mm, address, page_table, pte);
2533    if (page == swapcache)
2534        do_page_add_anon_rmap(page, vma, address, exclusive);
2535    else /* ksm created a completely new copy */
2536        page_add_new_anon_rmap(page, vma, address);
2537    /* It's better to call commit-charge after rmap is established */
2538    mem_cgroup_commit_charge_swapin(page, ptr);
2539
2540    swap_free(entry);
2541    if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2542        try_to_free_swap(page);
2543    unlock_page(page);
2544    if (page != swapcache) {
2545        /*
2546         * Hold the lock to avoid the swap entry to be reused
2547         * until we take the PT lock for the pte_same() check
2548         * (to avoid false positives from pte_same). For
2549         * further safety release the lock after the swap_free
2550         * so that the swap count won't change under a
2551         * parallel locked swapcache.
2552         */
2553        unlock_page(swapcache);
2554        page_cache_release(swapcache);
2555    }
2556
2557    if (flags & FAULT_FLAG_WRITE) {
2558        ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2559        if (ret & VM_FAULT_ERROR)
2560            ret &= VM_FAULT_ERROR;
2561        goto out;
2562    }
2563
2564    /* No need to invalidate - it was non-present before */
2565    update_mmu_cache(vma, address, page_table);
2566unlock:
2567    pte_unmap_unlock(page_table, ptl);
2568out:
2569    return ret;
2570out_nomap:
2571    mem_cgroup_cancel_charge_swapin(ptr);
2572    pte_unmap_unlock(page_table, ptl);
2573out_page:
2574    unlock_page(page);
2575out_release:
2576    page_cache_release(page);
2577    if (page != swapcache) {
2578        unlock_page(swapcache);
2579        page_cache_release(swapcache);
2580    }
2581    return ret;
2582}
2583
2584/*
2585 * This is like a special single-page "expand_{down|up}wards()",
2586 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2587 * doesn't hit another vma.
2588 */
2589static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2590{
2591    address &= PAGE_MASK;
2592    if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2593        struct vm_area_struct *prev = vma->vm_prev;
2594
2595        /*
2596         * Is there a mapping abutting this one below?
2597         *
2598         * That's only ok if it's the same stack mapping
2599         * that has gotten split..
2600         */
2601        if (prev && prev->vm_end == address)
2602            return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2603
2604        expand_downwards(vma, address - PAGE_SIZE);
2605    }
2606    if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2607        struct vm_area_struct *next = vma->vm_next;
2608
2609        /* As VM_GROWSDOWN but s/below/above/ */
2610        if (next && next->vm_start == address + PAGE_SIZE)
2611            return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2612
2613        expand_upwards(vma, address + PAGE_SIZE);
2614    }
2615    return 0;
2616}
2617
2618/*
2619 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2620 * but allow concurrent faults), and pte mapped but not yet locked.
2621 * We return with mmap_sem still held, but pte unmapped and unlocked.
2622 */
2623static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2624        unsigned long address, pte_t *page_table, pmd_t *pmd,
2625        unsigned int flags)
2626{
2627    struct page *page;
2628    spinlock_t *ptl;
2629    pte_t entry;
2630
2631    pte_unmap(page_table);
2632
2633    /* Check if we need to add a guard page to the stack */
2634    if (check_stack_guard_page(vma, address) < 0)
2635        return VM_FAULT_SIGBUS;
2636
2637    /* Use the zero-page for reads */
2638    if (!(flags & FAULT_FLAG_WRITE)) {
2639        entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2640                        vma->vm_page_prot));
2641        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2642        if (!pte_none(*page_table))
2643            goto unlock;
2644        goto setpte;
2645    }
2646
2647    /* Allocate our own private page. */
2648    if (unlikely(anon_vma_prepare(vma)))
2649        goto oom;
2650    page = alloc_zeroed_user_highpage_movable(vma, address);
2651    if (!page)
2652        goto oom;
2653    /*
2654     * The memory barrier inside __SetPageUptodate makes sure that
2655     * preceeding stores to the page contents become visible before
2656     * the set_pte_at() write.
2657     */
2658    __SetPageUptodate(page);
2659
2660    if (mem_cgroup_charge_anon(page, mm, GFP_KERNEL))
2661        goto oom_free_page;
2662
2663    entry = mk_pte(page, vma->vm_page_prot);
2664    if (vma->vm_flags & VM_WRITE)
2665        entry = pte_mkwrite(pte_mkdirty(entry));
2666
2667    page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2668    if (!pte_none(*page_table))
2669        goto release;
2670
2671    inc_mm_counter_fast(mm, MM_ANONPAGES);
2672    page_add_new_anon_rmap(page, vma, address);
2673setpte:
2674    set_pte_at(mm, address, page_table, entry);
2675
2676    /* No need to invalidate - it was non-present before */
2677    update_mmu_cache(vma, address, page_table);
2678unlock:
2679    pte_unmap_unlock(page_table, ptl);
2680    return 0;
2681release:
2682    mem_cgroup_uncharge_page(page);
2683    page_cache_release(page);
2684    goto unlock;
2685oom_free_page:
2686    page_cache_release(page);
2687oom:
2688    return VM_FAULT_OOM;
2689}
2690
2691static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2692        pgoff_t pgoff, unsigned int flags, struct page **page)
2693{
2694    struct vm_fault vmf;
2695    int ret;
2696
2697    vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2698    vmf.pgoff = pgoff;
2699    vmf.flags = flags;
2700    vmf.page = NULL;
2701
2702    ret = vma->vm_ops->fault(vma, &vmf);
2703    if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2704        return ret;
2705
2706    if (unlikely(PageHWPoison(vmf.page))) {
2707        if (ret & VM_FAULT_LOCKED)
2708            unlock_page(vmf.page);
2709        page_cache_release(vmf.page);
2710        return VM_FAULT_HWPOISON;
2711    }
2712
2713    if (unlikely(!(ret & VM_FAULT_LOCKED)))
2714        lock_page(vmf.page);
2715    else
2716        VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2717
2718    *page = vmf.page;
2719    return ret;
2720}
2721
2722/**
2723 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2724 *
2725 * @vma: virtual memory area
2726 * @address: user virtual address
2727 * @page: page to map
2728 * @pte: pointer to target page table entry
2729 * @write: true, if new entry is writable
2730 * @anon: true, if it's anonymous page
2731 *
2732 * Caller must hold page table lock relevant for @pte.
2733 *
2734 * Target users are page handler itself and implementations of
2735 * vm_ops->map_pages.
2736 */
2737void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2738        struct page *page, pte_t *pte, bool write, bool anon)
2739{
2740    pte_t entry;
2741
2742    flush_icache_page(vma, page);
2743    entry = mk_pte(page, vma->vm_page_prot);
2744    if (write)
2745        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2746    else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
2747        pte_mksoft_dirty(entry);
2748    if (anon) {
2749        inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2750        page_add_new_anon_rmap(page, vma, address);
2751    } else {
2752        inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2753        page_add_file_rmap(page);
2754    }
2755    set_pte_at(vma->vm_mm, address, pte, entry);
2756
2757    /* no need to invalidate: a not-present page won't be cached */
2758    update_mmu_cache(vma, address, pte);
2759}
2760
2761static unsigned long fault_around_bytes = rounddown_pow_of_two(65536);
2762
2763static inline unsigned long fault_around_pages(void)
2764{
2765    return fault_around_bytes >> PAGE_SHIFT;
2766}
2767
2768static inline unsigned long fault_around_mask(void)
2769{
2770    return ~(fault_around_bytes - 1) & PAGE_MASK;
2771}
2772
2773#ifdef CONFIG_DEBUG_FS
2774static int fault_around_bytes_get(void *data, u64 *val)
2775{
2776    *val = fault_around_bytes;
2777    return 0;
2778}
2779
2780/*
2781 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2782 * rounded down to nearest page order. It's what do_fault_around() expects to
2783 * see.
2784 */
2785static int fault_around_bytes_set(void *data, u64 val)
2786{
2787    if (val / PAGE_SIZE > PTRS_PER_PTE)
2788        return -EINVAL;
2789    if (val > PAGE_SIZE)
2790        fault_around_bytes = rounddown_pow_of_two(val);
2791    else
2792        fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2793    return 0;
2794}
2795DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2796        fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2797
2798static int __init fault_around_debugfs(void)
2799{
2800    void *ret;
2801
2802    ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2803            &fault_around_bytes_fops);
2804    if (!ret)
2805        pr_warn("Failed to create fault_around_bytes in debugfs");
2806    return 0;
2807}
2808late_initcall(fault_around_debugfs);
2809#endif
2810
2811/*
2812 * do_fault_around() tries to map few pages around the fault address. The hope
2813 * is that the pages will be needed soon and this will lower the number of
2814 * faults to handle.
2815 *
2816 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2817 * not ready to be mapped: not up-to-date, locked, etc.
2818 *
2819 * This function is called with the page table lock taken. In the split ptlock
2820 * case the page table lock only protects only those entries which belong to
2821 * the page table corresponding to the fault address.
2822 *
2823 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2824 * only once.
2825 *
2826 * fault_around_pages() defines how many pages we'll try to map.
2827 * do_fault_around() expects it to return a power of two less than or equal to
2828 * PTRS_PER_PTE.
2829 *
2830 * The virtual address of the area that we map is naturally aligned to the
2831 * fault_around_pages() value (and therefore to page order). This way it's
2832 * easier to guarantee that we don't cross page table boundaries.
2833 */
2834static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2835        pte_t *pte, pgoff_t pgoff, unsigned int flags)
2836{
2837    unsigned long start_addr;
2838    pgoff_t max_pgoff;
2839    struct vm_fault vmf;
2840    int off;
2841
2842    start_addr = max(address & fault_around_mask(), vma->vm_start);
2843    off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2844    pte -= off;
2845    pgoff -= off;
2846
2847    /*
2848     * max_pgoff is either end of page table or end of vma
2849     * or fault_around_pages() from pgoff, depending what is nearest.
2850     */
2851    max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2852        PTRS_PER_PTE - 1;
2853    max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2854            pgoff + fault_around_pages() - 1);
2855
2856    /* Check if it makes any sense to call ->map_pages */
2857    while (!pte_none(*pte)) {
2858        if (++pgoff > max_pgoff)
2859            return;
2860        start_addr += PAGE_SIZE;
2861        if (start_addr >= vma->vm_end)
2862            return;
2863        pte++;
2864    }
2865
2866    vmf.virtual_address = (void __user *) start_addr;
2867    vmf.pte = pte;
2868    vmf.pgoff = pgoff;
2869    vmf.max_pgoff = max_pgoff;
2870    vmf.flags = flags;
2871    vma->vm_ops->map_pages(vma, &vmf);
2872}
2873
2874static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2875        unsigned long address, pmd_t *pmd,
2876        pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2877{
2878    struct page *fault_page;
2879    spinlock_t *ptl;
2880    pte_t *pte;
2881    int ret = 0;
2882
2883    /*
2884     * Let's call ->map_pages() first and use ->fault() as fallback
2885     * if page by the offset is not ready to be mapped (cold cache or
2886     * something).
2887     */
2888    if (vma->vm_ops->map_pages && !(flags & FAULT_FLAG_NONLINEAR) &&
2889        fault_around_pages() > 1) {
2890        pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2891        do_fault_around(vma, address, pte, pgoff, flags);
2892        if (!pte_same(*pte, orig_pte))
2893            goto unlock_out;
2894        pte_unmap_unlock(pte, ptl);
2895    }
2896
2897    ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2898    if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2899        return ret;
2900
2901    pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2902    if (unlikely(!pte_same(*pte, orig_pte))) {
2903        pte_unmap_unlock(pte, ptl);
2904        unlock_page(fault_page);
2905        page_cache_release(fault_page);
2906        return ret;
2907    }
2908    do_set_pte(vma, address, fault_page, pte, false, false);
2909    unlock_page(fault_page);
2910unlock_out:
2911    pte_unmap_unlock(pte, ptl);
2912    return ret;
2913}
2914
2915static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2916        unsigned long address, pmd_t *pmd,
2917        pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2918{
2919    struct page *fault_page, *new_page;
2920    spinlock_t *ptl;
2921    pte_t *pte;
2922    int ret;
2923
2924    if (unlikely(anon_vma_prepare(vma)))
2925        return VM_FAULT_OOM;
2926
2927    new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2928    if (!new_page)
2929        return VM_FAULT_OOM;
2930
2931    if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)) {
2932        page_cache_release(new_page);
2933        return VM_FAULT_OOM;
2934    }
2935
2936    ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2937    if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2938        goto uncharge_out;
2939
2940    copy_user_highpage(new_page, fault_page, address, vma);
2941    __SetPageUptodate(new_page);
2942
2943    pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2944    if (unlikely(!pte_same(*pte, orig_pte))) {
2945        pte_unmap_unlock(pte, ptl);
2946        unlock_page(fault_page);
2947        page_cache_release(fault_page);
2948        goto uncharge_out;
2949    }
2950    do_set_pte(vma, address, new_page, pte, true, true);
2951    pte_unmap_unlock(pte, ptl);
2952    unlock_page(fault_page);
2953    page_cache_release(fault_page);
2954    return ret;
2955uncharge_out:
2956    mem_cgroup_uncharge_page(new_page);
2957    page_cache_release(new_page);
2958    return ret;
2959}
2960
2961static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2962        unsigned long address, pmd_t *pmd,
2963        pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2964{
2965    struct page *fault_page;
2966    struct address_space *mapping;
2967    spinlock_t *ptl;
2968    pte_t *pte;
2969    int dirtied = 0;
2970    int ret, tmp;
2971
2972    ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2973    if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2974        return ret;
2975
2976    /*
2977     * Check if the backing address space wants to know that the page is
2978     * about to become writable
2979     */
2980    if (vma->vm_ops->page_mkwrite) {
2981        unlock_page(fault_page);
2982        tmp = do_page_mkwrite(vma, fault_page, address);
2983        if (unlikely(!tmp ||
2984                (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2985            page_cache_release(fault_page);
2986            return tmp;
2987        }
2988    }
2989
2990    pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2991    if (unlikely(!pte_same(*pte, orig_pte))) {
2992        pte_unmap_unlock(pte, ptl);
2993        unlock_page(fault_page);
2994        page_cache_release(fault_page);
2995        return ret;
2996    }
2997    do_set_pte(vma, address, fault_page, pte, true, false);
2998    pte_unmap_unlock(pte, ptl);
2999
3000    if (set_page_dirty(fault_page))
3001        dirtied = 1;
3002    mapping = fault_page->mapping;
3003    unlock_page(fault_page);
3004    if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3005        /*
3006         * Some device drivers do not set page.mapping but still
3007         * dirty their pages
3008         */
3009        balance_dirty_pages_ratelimited(mapping);
3010    }
3011
3012    /* file_update_time outside page_lock */
3013    if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3014        file_update_time(vma->vm_file);
3015
3016    return ret;
3017}
3018
3019static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3020        unsigned long address, pte_t *page_table, pmd_t *pmd,
3021        unsigned int flags, pte_t orig_pte)
3022{
3023    pgoff_t pgoff = (((address & PAGE_MASK)
3024            - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3025
3026    pte_unmap(page_table);
3027    if (!(flags & FAULT_FLAG_WRITE))
3028        return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3029                orig_pte);
3030    if (!(vma->vm_flags & VM_SHARED))
3031        return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3032                orig_pte);
3033    return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3034}
3035
3036/*
3037 * Fault of a previously existing named mapping. Repopulate the pte
3038 * from the encoded file_pte if possible. This enables swappable
3039 * nonlinear vmas.
3040 *
3041 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3042 * but allow concurrent faults), and pte mapped but not yet locked.
3043 * We return with mmap_sem still held, but pte unmapped and unlocked.
3044 */
3045static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3046        unsigned long address, pte_t *page_table, pmd_t *pmd,
3047        unsigned int flags, pte_t orig_pte)
3048{
3049    pgoff_t pgoff;
3050
3051    flags |= FAULT_FLAG_NONLINEAR;
3052
3053    if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3054        return 0;
3055
3056    if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3057        /*
3058         * Page table corrupted: show pte and kill process.
3059         */
3060        print_bad_pte(vma, address, orig_pte, NULL);
3061        return VM_FAULT_SIGBUS;
3062    }
3063
3064    pgoff = pte_to_pgoff(orig_pte);
3065    if (!(flags & FAULT_FLAG_WRITE))
3066        return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3067                orig_pte);
3068    if (!(vma->vm_flags & VM_SHARED))
3069        return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3070                orig_pte);
3071    return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3072}
3073
3074static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3075                unsigned long addr, int page_nid,
3076                int *flags)
3077{
3078    get_page(page);
3079
3080    count_vm_numa_event(NUMA_HINT_FAULTS);
3081    if (page_nid == numa_node_id()) {
3082        count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3083        *flags |= TNF_FAULT_LOCAL;
3084    }
3085
3086    return mpol_misplaced(page, vma, addr);
3087}
3088
3089static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3090           unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3091{
3092    struct page *page = NULL;
3093    spinlock_t *ptl;
3094    int page_nid = -1;
3095    int last_cpupid;
3096    int target_nid;
3097    bool migrated = false;
3098    int flags = 0;
3099
3100    /*
3101    * The "pte" at this point cannot be used safely without
3102    * validation through pte_unmap_same(). It's of NUMA type but
3103    * the pfn may be screwed if the read is non atomic.
3104    *
3105    * ptep_modify_prot_start is not called as this is clearing
3106    * the _PAGE_NUMA bit and it is not really expected that there
3107    * would be concurrent hardware modifications to the PTE.
3108    */
3109    ptl = pte_lockptr(mm, pmd);
3110    spin_lock(ptl);
3111    if (unlikely(!pte_same(*ptep, pte))) {
3112        pte_unmap_unlock(ptep, ptl);
3113        goto out;
3114    }
3115
3116    pte = pte_mknonnuma(pte);
3117    set_pte_at(mm, addr, ptep, pte);
3118    update_mmu_cache(vma, addr, ptep);
3119
3120    page = vm_normal_page(vma, addr, pte);
3121    if (!page) {
3122        pte_unmap_unlock(ptep, ptl);
3123        return 0;
3124    }
3125    BUG_ON(is_zero_pfn(page_to_pfn(page)));
3126
3127    /*
3128     * Avoid grouping on DSO/COW pages in specific and RO pages
3129     * in general, RO pages shouldn't hurt as much anyway since
3130     * they can be in shared cache state.
3131     */
3132    if (!pte_write(pte))
3133        flags |= TNF_NO_GROUP;
3134
3135    /*
3136     * Flag if the page is shared between multiple address spaces. This
3137     * is later used when determining whether to group tasks together
3138     */
3139    if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3140        flags |= TNF_SHARED;
3141
3142    last_cpupid = page_cpupid_last(page);
3143    page_nid = page_to_nid(page);
3144    target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3145    pte_unmap_unlock(ptep, ptl);
3146    if (target_nid == -1) {
3147        put_page(page);
3148        goto out;
3149    }
3150
3151    /* Migrate to the requested node */
3152    migrated = migrate_misplaced_page(page, vma, target_nid);
3153    if (migrated) {
3154        page_nid = target_nid;
3155        flags |= TNF_MIGRATED;
3156    }
3157
3158out:
3159    if (page_nid != -1)
3160        task_numa_fault(last_cpupid, page_nid, 1, flags);
3161    return 0;
3162}
3163
3164/*
3165 * These routines also need to handle stuff like marking pages dirty
3166 * and/or accessed for architectures that don't do it in hardware (most
3167 * RISC architectures). The early dirtying is also good on the i386.
3168 *
3169 * There is also a hook called "update_mmu_cache()" that architectures
3170 * with external mmu caches can use to update those (ie the Sparc or
3171 * PowerPC hashed page tables that act as extended TLBs).
3172 *
3173 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3174 * but allow concurrent faults), and pte mapped but not yet locked.
3175 * We return with mmap_sem still held, but pte unmapped and unlocked.
3176 */
3177static int handle_pte_fault(struct mm_struct *mm,
3178             struct vm_area_struct *vma, unsigned long address,
3179             pte_t *pte, pmd_t *pmd, unsigned int flags)
3180{
3181    pte_t entry;
3182    spinlock_t *ptl;
3183
3184    entry = *pte;
3185    if (!pte_present(entry)) {
3186        if (pte_none(entry)) {
3187            if (vma->vm_ops) {
3188                if (likely(vma->vm_ops->fault))
3189                    return do_linear_fault(mm, vma, address,
3190                        pte, pmd, flags, entry);
3191            }
3192            return do_anonymous_page(mm, vma, address,
3193                         pte, pmd, flags);
3194        }
3195        if (pte_file(entry))
3196            return do_nonlinear_fault(mm, vma, address,
3197                    pte, pmd, flags, entry);
3198        return do_swap_page(mm, vma, address,
3199                    pte, pmd, flags, entry);
3200    }
3201
3202    if (pte_numa(entry))
3203        return do_numa_page(mm, vma, address, entry, pte, pmd);
3204
3205    ptl = pte_lockptr(mm, pmd);
3206    spin_lock(ptl);
3207    if (unlikely(!pte_same(*pte, entry)))
3208        goto unlock;
3209    if (flags & FAULT_FLAG_WRITE) {
3210        if (!pte_write(entry))
3211            return do_wp_page(mm, vma, address,
3212                    pte, pmd, ptl, entry);
3213        entry = pte_mkdirty(entry);
3214    }
3215    entry = pte_mkyoung(entry);
3216    if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3217        update_mmu_cache(vma, address, pte);
3218    } else {
3219        /*
3220         * This is needed only for protection faults but the arch code
3221         * is not yet telling us if this is a protection fault or not.
3222         * This still avoids useless tlb flushes for .text page faults
3223         * with threads.
3224         */
3225        if (flags & FAULT_FLAG_WRITE)
3226            flush_tlb_fix_spurious_fault(vma, address);
3227    }
3228unlock:
3229    pte_unmap_unlock(pte, ptl);
3230    return 0;
3231}
3232
3233/*
3234 * By the time we get here, we already hold the mm semaphore
3235 */
3236static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3237                 unsigned long address, unsigned int flags)
3238{
3239    pgd_t *pgd;
3240    pud_t *pud;
3241    pmd_t *pmd;
3242    pte_t *pte;
3243
3244    if (unlikely(is_vm_hugetlb_page(vma)))
3245        return hugetlb_fault(mm, vma, address, flags);
3246
3247    pgd = pgd_offset(mm, address);
3248    pud = pud_alloc(mm, pgd, address);
3249    if (!pud)
3250        return VM_FAULT_OOM;
3251    pmd = pmd_alloc(mm, pud, address);
3252    if (!pmd)
3253        return VM_FAULT_OOM;
3254    if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3255        int ret = VM_FAULT_FALLBACK;
3256        if (!vma->vm_ops)
3257            ret = do_huge_pmd_anonymous_page(mm, vma, address,
3258                    pmd, flags);
3259        if (!(ret & VM_FAULT_FALLBACK))
3260            return ret;
3261    } else {
3262        pmd_t orig_pmd = *pmd;
3263        int ret;
3264
3265        barrier();
3266        if (pmd_trans_huge(orig_pmd)) {
3267            unsigned int dirty = flags & FAULT_FLAG_WRITE;
3268
3269            /*
3270             * If the pmd is splitting, return and retry the
3271             * the fault. Alternative: wait until the split
3272             * is done, and goto retry.
3273             */
3274            if (pmd_trans_splitting(orig_pmd))
3275                return 0;
3276
3277            if (pmd_numa(orig_pmd))
3278                return do_huge_pmd_numa_page(mm, vma, address,
3279                                 orig_pmd, pmd);
3280
3281            if (dirty && !pmd_write(orig_pmd)) {
3282                ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3283                              orig_pmd);
3284                if (!(ret & VM_FAULT_FALLBACK))
3285                    return ret;
3286            } else {
3287                huge_pmd_set_accessed(mm, vma, address, pmd,
3288                              orig_pmd, dirty);
3289                return 0;
3290            }
3291        }
3292    }
3293
3294    /*
3295     * Use __pte_alloc instead of pte_alloc_map, because we can't
3296     * run pte_offset_map on the pmd, if an huge pmd could
3297     * materialize from under us from a different thread.
3298     */
3299    if (unlikely(pmd_none(*pmd)) &&
3300        unlikely(__pte_alloc(mm, vma, pmd, address)))
3301        return VM_FAULT_OOM;
3302    /* if an huge pmd materialized from under us just retry later */
3303    if (unlikely(pmd_trans_huge(*pmd)))
3304        return 0;
3305    /*
3306     * A regular pmd is established and it can't morph into a huge pmd
3307     * from under us anymore at this point because we hold the mmap_sem
3308     * read mode and khugepaged takes it in write mode. So now it's
3309     * safe to run pte_offset_map().
3310     */
3311    pte = pte_offset_map(pmd, address);
3312
3313    return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3314}
3315
3316int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3317            unsigned long address, unsigned int flags)
3318{
3319    int ret;
3320
3321    __set_current_state(TASK_RUNNING);
3322
3323    count_vm_event(PGFAULT);
3324    mem_cgroup_count_vm_event(mm, PGFAULT);
3325
3326    /* do counter updates before entering really critical section. */
3327    check_sync_rss_stat(current);
3328
3329    /*
3330     * Enable the memcg OOM handling for faults triggered in user
3331     * space. Kernel faults are handled more gracefully.
3332     */
3333    if (flags & FAULT_FLAG_USER)
3334        mem_cgroup_oom_enable();
3335
3336    ret = __handle_mm_fault(mm, vma, address, flags);
3337
3338    if (flags & FAULT_FLAG_USER) {
3339        mem_cgroup_oom_disable();
3340                /*
3341                 * The task may have entered a memcg OOM situation but
3342                 * if the allocation error was handled gracefully (no
3343                 * VM_FAULT_OOM), there is no need to kill anything.
3344                 * Just clean up the OOM state peacefully.
3345                 */
3346                if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3347                        mem_cgroup_oom_synchronize(false);
3348    }
3349
3350    return ret;
3351}
3352
3353#ifndef __PAGETABLE_PUD_FOLDED
3354/*
3355 * Allocate page upper directory.
3356 * We've already handled the fast-path in-line.
3357 */
3358int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3359{
3360    pud_t *new = pud_alloc_one(mm, address);
3361    if (!new)
3362        return -ENOMEM;
3363
3364    smp_wmb(); /* See comment in __pte_alloc */
3365
3366    spin_lock(&mm->page_table_lock);
3367    if (pgd_present(*pgd)) /* Another has populated it */
3368        pud_free(mm, new);
3369    else
3370        pgd_populate(mm, pgd, new);
3371    spin_unlock(&mm->page_table_lock);
3372    return 0;
3373}
3374#endif /* __PAGETABLE_PUD_FOLDED */
3375
3376#ifndef __PAGETABLE_PMD_FOLDED
3377/*
3378 * Allocate page middle directory.
3379 * We've already handled the fast-path in-line.
3380 */
3381int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3382{
3383    pmd_t *new = pmd_alloc_one(mm, address);
3384    if (!new)
3385        return -ENOMEM;
3386
3387    smp_wmb(); /* See comment in __pte_alloc */
3388
3389    spin_lock(&mm->page_table_lock);
3390#ifndef __ARCH_HAS_4LEVEL_HACK
3391    if (pud_present(*pud)) /* Another has populated it */
3392        pmd_free(mm, new);
3393    else
3394        pud_populate(mm, pud, new);
3395#else
3396    if (pgd_present(*pud)) /* Another has populated it */
3397        pmd_free(mm, new);
3398    else
3399        pgd_populate(mm, pud, new);
3400#endif /* __ARCH_HAS_4LEVEL_HACK */
3401    spin_unlock(&mm->page_table_lock);
3402    return 0;
3403}
3404#endif /* __PAGETABLE_PMD_FOLDED */
3405
3406#if !defined(__HAVE_ARCH_GATE_AREA)
3407
3408#if defined(AT_SYSINFO_EHDR)
3409static struct vm_area_struct gate_vma;
3410
3411static int __init gate_vma_init(void)
3412{
3413    gate_vma.vm_mm = NULL;
3414    gate_vma.vm_start = FIXADDR_USER_START;
3415    gate_vma.vm_end = FIXADDR_USER_END;
3416    gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3417    gate_vma.vm_page_prot = __P101;
3418
3419    return 0;
3420}
3421__initcall(gate_vma_init);
3422#endif
3423
3424struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3425{
3426#ifdef AT_SYSINFO_EHDR
3427    return &gate_vma;
3428#else
3429    return NULL;
3430#endif
3431}
3432
3433int in_gate_area_no_mm(unsigned long addr)
3434{
3435#ifdef AT_SYSINFO_EHDR
3436    if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3437        return 1;
3438#endif
3439    return 0;
3440}
3441
3442#endif /* __HAVE_ARCH_GATE_AREA */
3443
3444static int __follow_pte(struct mm_struct *mm, unsigned long address,
3445        pte_t **ptepp, spinlock_t **ptlp)
3446{
3447    pgd_t *pgd;
3448    pud_t *pud;
3449    pmd_t *pmd;
3450    pte_t *ptep;
3451
3452    pgd = pgd_offset(mm, address);
3453    if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3454        goto out;
3455
3456    pud = pud_offset(pgd, address);
3457    if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3458        goto out;
3459
3460    pmd = pmd_offset(pud, address);
3461    VM_BUG_ON(pmd_trans_huge(*pmd));
3462    if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3463        goto out;
3464
3465    /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3466    if (pmd_huge(*pmd))
3467        goto out;
3468
3469    ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3470    if (!ptep)
3471        goto out;
3472    if (!pte_present(*ptep))
3473        goto unlock;
3474    *ptepp = ptep;
3475    return 0;
3476unlock:
3477    pte_unmap_unlock(ptep, *ptlp);
3478out:
3479    return -EINVAL;
3480}
3481
3482static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3483                 pte_t **ptepp, spinlock_t **ptlp)
3484{
3485    int res;
3486
3487    /* (void) is needed to make gcc happy */
3488    (void) __cond_lock(*ptlp,
3489               !(res = __follow_pte(mm, address, ptepp, ptlp)));
3490    return res;
3491}
3492
3493/**
3494 * follow_pfn - look up PFN at a user virtual address
3495 * @vma: memory mapping
3496 * @address: user virtual address
3497 * @pfn: location to store found PFN
3498 *
3499 * Only IO mappings and raw PFN mappings are allowed.
3500 *
3501 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3502 */
3503int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3504    unsigned long *pfn)
3505{
3506    int ret = -EINVAL;
3507    spinlock_t *ptl;
3508    pte_t *ptep;
3509
3510    if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3511        return ret;
3512
3513    ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3514    if (ret)
3515        return ret;
3516    *pfn = pte_pfn(*ptep);
3517    pte_unmap_unlock(ptep, ptl);
3518    return 0;
3519}
3520EXPORT_SYMBOL(follow_pfn);
3521
3522#ifdef CONFIG_HAVE_IOREMAP_PROT
3523int follow_phys(struct vm_area_struct *vma,
3524        unsigned long address, unsigned int flags,
3525        unsigned long *prot, resource_size_t *phys)
3526{
3527    int ret = -EINVAL;
3528    pte_t *ptep, pte;
3529    spinlock_t *ptl;
3530
3531    if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3532        goto out;
3533
3534    if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3535        goto out;
3536    pte = *ptep;
3537
3538    if ((flags & FOLL_WRITE) && !pte_write(pte))
3539        goto unlock;
3540
3541    *prot = pgprot_val(pte_pgprot(pte));
3542    *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3543
3544    ret = 0;
3545unlock:
3546    pte_unmap_unlock(ptep, ptl);
3547out:
3548    return ret;
3549}
3550
3551int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3552            void *buf, int len, int write)
3553{
3554    resource_size_t phys_addr;
3555    unsigned long prot = 0;
3556    void __iomem *maddr;
3557    int offset = addr & (PAGE_SIZE-1);
3558
3559    if (follow_phys(vma, addr, write, &prot, &phys_addr))
3560        return -EINVAL;
3561
3562    maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3563    if (write)
3564        memcpy_toio(maddr + offset, buf, len);
3565    else
3566        memcpy_fromio(buf, maddr + offset, len);
3567    iounmap(maddr);
3568
3569    return len;
3570}
3571EXPORT_SYMBOL_GPL(generic_access_phys);
3572#endif
3573
3574/*
3575 * Access another process' address space as given in mm. If non-NULL, use the
3576 * given task for page fault accounting.
3577 */
3578static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3579        unsigned long addr, void *buf, int len, int write)
3580{
3581    struct vm_area_struct *vma;
3582    void *old_buf = buf;
3583
3584    down_read(&mm->mmap_sem);
3585    /* ignore errors, just check how much was successfully transferred */
3586    while (len) {
3587        int bytes, ret, offset;
3588        void *maddr;
3589        struct page *page = NULL;
3590
3591        ret = get_user_pages(tsk, mm, addr, 1,
3592                write, 1, &page, &vma);
3593        if (ret <= 0) {
3594            /*
3595             * Check if this is a VM_IO | VM_PFNMAP VMA, which
3596             * we can access using slightly different code.
3597             */
3598#ifdef CONFIG_HAVE_IOREMAP_PROT
3599            vma = find_vma(mm, addr);
3600            if (!vma || vma->vm_start > addr)
3601                break;
3602            if (vma->vm_ops && vma->vm_ops->access)
3603                ret = vma->vm_ops->access(vma, addr, buf,
3604                              len, write);
3605            if (ret <= 0)
3606#endif
3607                break;
3608            bytes = ret;
3609        } else {
3610            bytes = len;
3611            offset = addr & (PAGE_SIZE-1);
3612            if (bytes > PAGE_SIZE-offset)
3613                bytes = PAGE_SIZE-offset;
3614
3615            maddr = kmap(page);
3616            if (write) {
3617                copy_to_user_page(vma, page, addr,
3618                          maddr + offset, buf, bytes);
3619                set_page_dirty_lock(page);
3620            } else {
3621                copy_from_user_page(vma, page, addr,
3622                            buf, maddr + offset, bytes);
3623            }
3624            kunmap(page);
3625            page_cache_release(page);
3626        }
3627        len -= bytes;
3628        buf += bytes;
3629        addr += bytes;
3630    }
3631    up_read(&mm->mmap_sem);
3632
3633    return buf - old_buf;
3634}
3635
3636/**
3637 * access_remote_vm - access another process' address space
3638 * @mm: the mm_struct of the target address space
3639 * @addr: start address to access
3640 * @buf: source or destination buffer
3641 * @len: number of bytes to transfer
3642 * @write: whether the access is a write
3643 *
3644 * The caller must hold a reference on @mm.
3645 */
3646int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3647        void *buf, int len, int write)
3648{
3649    return __access_remote_vm(NULL, mm, addr, buf, len, write);
3650}
3651
3652/*
3653 * Access another process' address space.
3654 * Source/target buffer must be kernel space,
3655 * Do not walk the page table directly, use get_user_pages
3656 */
3657int access_process_vm(struct task_struct *tsk, unsigned long addr,
3658        void *buf, int len, int write)
3659{
3660    struct mm_struct *mm;
3661    int ret;
3662
3663    mm = get_task_mm(tsk);
3664    if (!mm)
3665        return 0;
3666
3667    ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3668    mmput(mm);
3669
3670    return ret;
3671}
3672
3673/*
3674 * Print the name of a VMA.
3675 */
3676void print_vma_addr(char *prefix, unsigned long ip)
3677{
3678    struct mm_struct *mm = current->mm;
3679    struct vm_area_struct *vma;
3680
3681    /*
3682     * Do not print if we are in atomic
3683     * contexts (in exception stacks, etc.):
3684     */
3685    if (preempt_count())
3686        return;
3687
3688    down_read(&mm->mmap_sem);
3689    vma = find_vma(mm, ip);
3690    if (vma && vma->vm_file) {
3691        struct file *f = vma->vm_file;
3692        char *buf = (char *)__get_free_page(GFP_KERNEL);
3693        if (buf) {
3694            char *p;
3695
3696            p = d_path(&f->f_path, buf, PAGE_SIZE);
3697            if (IS_ERR(p))
3698                p = "?";
3699            printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3700                    vma->vm_start,
3701                    vma->vm_end - vma->vm_start);
3702            free_page((unsigned long)buf);
3703        }
3704    }
3705    up_read(&mm->mmap_sem);
3706}
3707
3708#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3709void might_fault(void)
3710{
3711    /*
3712     * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3713     * holding the mmap_sem, this is safe because kernel memory doesn't
3714     * get paged out, therefore we'll never actually fault, and the
3715     * below annotations will generate false positives.
3716     */
3717    if (segment_eq(get_fs(), KERNEL_DS))
3718        return;
3719
3720    /*
3721     * it would be nicer only to annotate paths which are not under
3722     * pagefault_disable, however that requires a larger audit and
3723     * providing helpers like get_user_atomic.
3724     */
3725    if (in_atomic())
3726        return;
3727
3728    __might_sleep(__FILE__, __LINE__, 0);
3729
3730    if (current->mm)
3731        might_lock_read(&current->mm->mmap_sem);
3732}
3733EXPORT_SYMBOL(might_fault);
3734#endif
3735
3736#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3737static void clear_gigantic_page(struct page *page,
3738                unsigned long addr,
3739                unsigned int pages_per_huge_page)
3740{
3741    int i;
3742    struct page *p = page;
3743
3744    might_sleep();
3745    for (i = 0; i < pages_per_huge_page;
3746         i++, p = mem_map_next(p, page, i)) {
3747        cond_resched();
3748        clear_user_highpage(p, addr + i * PAGE_SIZE);
3749    }
3750}
3751void clear_huge_page(struct page *page,
3752             unsigned long addr, unsigned int pages_per_huge_page)
3753{
3754    int i;
3755
3756    if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3757        clear_gigantic_page(page, addr, pages_per_huge_page);
3758        return;
3759    }
3760
3761    might_sleep();
3762    for (i = 0; i < pages_per_huge_page; i++) {
3763        cond_resched();
3764        clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3765    }
3766}
3767
3768static void copy_user_gigantic_page(struct page *dst, struct page *src,
3769                    unsigned long addr,
3770                    struct vm_area_struct *vma,
3771                    unsigned int pages_per_huge_page)
3772{
3773    int i;
3774    struct page *dst_base = dst;
3775    struct page *src_base = src;
3776
3777    for (i = 0; i < pages_per_huge_page; ) {
3778        cond_resched();
3779        copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3780
3781        i++;
3782        dst = mem_map_next(dst, dst_base, i);
3783        src = mem_map_next(src, src_base, i);
3784    }
3785}
3786
3787void copy_user_huge_page(struct page *dst, struct page *src,
3788             unsigned long addr, struct vm_area_struct *vma,
3789             unsigned int pages_per_huge_page)
3790{
3791    int i;
3792
3793    if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3794        copy_user_gigantic_page(dst, src, addr, vma,
3795                    pages_per_huge_page);
3796        return;
3797    }
3798
3799    might_sleep();
3800    for (i = 0; i < pages_per_huge_page; i++) {
3801        cond_resched();
3802        copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3803    }
3804}
3805#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3806
3807#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3808
3809static struct kmem_cache *page_ptl_cachep;
3810
3811void __init ptlock_cache_init(void)
3812{
3813    page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3814            SLAB_PANIC, NULL);
3815}
3816
3817bool ptlock_alloc(struct page *page)
3818{
3819    spinlock_t *ptl;
3820
3821    ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3822    if (!ptl)
3823        return false;
3824    page->ptl = ptl;
3825    return true;
3826}
3827
3828void ptlock_free(struct page *page)
3829{
3830    kmem_cache_free(page_ptl_cachep, page->ptl);
3831}
3832#endif
3833

Archive Download this file



interactive