LXR RHEL4/mm/msync.c

   1/*
   2 *      linux/mm/msync.c
   3 *
   4 * Copyright (C) 1994-1999  Linus Torvalds
   5 */
   6
   7/*
   8 * The msync() system call.
   9 */
  10#include <linux/slab.h>
  11#include <linux/pagemap.h>
  12#include <linux/mm.h>
  13#include <linux/mman.h>
  14#include <linux/hugetlb.h>
  15
  16#include <asm/pgtable.h>
  17#include <asm/tlbflush.h>
  18
  19/*
  20 * Called with mm->page_table_lock held to protect against other
  21 * threads/the swapper from ripping pte's out from under us.
  22 */
  23static int filemap_sync_pte(pte_t *ptep, struct vm_area_struct *vma,
  24        unsigned long address, unsigned int flags)
  25{
  26        pte_t pte = *ptep;
  27        unsigned long pfn = pte_pfn(pte);
  28        struct page *page;
  29
  30        if (pte_present(pte) && pfn_valid(pfn)) {
  31                page = pfn_to_page(pfn);
  32                if (!PageReserved(page) &&
  33                    (ptep_clear_flush_dirty(vma, address, ptep) ||
  34                     page_test_and_clear_dirty(page)))
  35                        set_page_dirty(page);
  36        }
  37        return 0;
  38}
  39
  40static int filemap_sync_pte_range(pmd_t * pmd,
  41        unsigned long address, unsigned long end, 
  42        struct vm_area_struct *vma, unsigned int flags)
  43{
  44        pte_t *pte;
  45        int error;
  46
  47        if (pmd_none(*pmd))
  48                return 0;
  49        if (pmd_bad(*pmd)) {
  50                pmd_ERROR(*pmd);
  51                pmd_clear(pmd);
  52                return 0;
  53        }
  54        pte = pte_offset_map(pmd, address);
  55        if ((address & PMD_MASK) != (end & PMD_MASK))
  56                end = (address & PMD_MASK) + PMD_SIZE;
  57        error = 0;
  58        do {
  59                error |= filemap_sync_pte(pte, vma, address, flags);
  60                address += PAGE_SIZE;
  61                pte++;
  62        } while (address && (address < end));
  63
  64        pte_unmap(pte - 1);
  65
  66        return error;
  67}
  68
  69static inline int filemap_sync_pmd_range(pgd_t * pgd,
  70        unsigned long address, unsigned long end, 
  71        struct vm_area_struct *vma, unsigned int flags)
  72{
  73        pmd_t * pmd;
  74        int error;
  75
  76        if (pgd_none(*pgd))
  77                return 0;
  78        if (pgd_bad(*pgd)) {
  79                pgd_ERROR(*pgd);
  80                pgd_clear(pgd);
  81                return 0;
  82        }
  83        pmd = pmd_offset(pgd, address);
  84        if ((address & PGDIR_MASK) != (end & PGDIR_MASK))
  85                end = (address & PGDIR_MASK) + PGDIR_SIZE;
  86        error = 0;
  87        do {
  88                error |= filemap_sync_pte_range(pmd, address, end, vma, flags);
  89                address = (address + PMD_SIZE) & PMD_MASK;
  90                pmd++;
  91        } while (address && (address < end));
  92        return error;
  93}
  94
  95static int filemap_sync(struct vm_area_struct * vma, unsigned long address,
  96        size_t size, unsigned int flags)
  97{
  98        pgd_t * dir;
  99        unsigned long end = address + size;
 100        int error = 0;
 101
 102        /* Aquire the lock early; it may be possible to avoid dropping
 103         * and reaquiring it repeatedly.
 104         */
 105        spin_lock(&vma->vm_mm->page_table_lock);
 106
 107        dir = pgd_offset(vma->vm_mm, address);
 108        flush_cache_range(vma, address, end);
 109
 110        /* For hugepages we can't go walking the page table normally,
 111         * but that's ok, hugetlbfs is memory based, so we don't need
 112         * to do anything more on an msync() */
 113        if (is_vm_hugetlb_page(vma))
 114                goto out;
 115
 116        if (address >= end)
 117                BUG();
 118        do {
 119                error |= filemap_sync_pmd_range(dir, address, end, vma, flags);
 120                address = (address + PGDIR_SIZE) & PGDIR_MASK;
 121                dir++;
 122        } while (address && (address < end));
 123        /*
 124         * Why flush ? filemap_sync_pte already flushed the tlbs with the
 125         * dirty bits.
 126         */
 127        flush_tlb_range(vma, end - size, end);
 128 out:
 129        spin_unlock(&vma->vm_mm->page_table_lock);
 130
 131        return error;
 132}
 133
 134/*
 135 * MS_SYNC syncs the entire file - including mappings.
 136 *
 137 * MS_ASYNC does not start I/O (it used to, up to 2.5.67).  Instead, it just
 138 * marks the relevant pages dirty.  The application may now run fsync() to
 139 * write out the dirty pages and wait on the writeout and check the result.
 140 * Or the application may run fadvise(FADV_DONTNEED) against the fd to start
 141 * async writeout immediately.
 142 * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to
 143 * applications.
 144 */
 145static int msync_interval(struct vm_area_struct * vma,
 146        unsigned long start, unsigned long end, int flags)
 147{
 148        int ret = 0;
 149        struct file * file = vma->vm_file;
 150
 151        if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED))
 152                return -EBUSY;
 153
 154        if (file && (vma->vm_flags & VM_SHARED)) {
 155                ret = filemap_sync(vma, start, end-start, flags);
 156
 157                if (!ret && (flags & MS_SYNC)) {
 158                        struct address_space *mapping = file->f_mapping;
 159                        int err;
 160
 161                        down(&mapping->host->i_sem);
 162                        ret = filemap_fdatawrite(mapping);
 163                        if (file->f_op && file->f_op->fsync) {
 164                                err = file->f_op->fsync(file,file->f_dentry,1);
 165                                if (err && !ret)
 166                                        ret = err;
 167                        }
 168                        err = filemap_fdatawait(mapping);
 169                        if (!ret)
 170                                ret = err;
 171                        up(&mapping->host->i_sem);
 172                }
 173        }
 174        return ret;
 175}
 176
 177asmlinkage long sys_msync(unsigned long start, size_t len, int flags)
 178{
 179        unsigned long end;
 180        struct vm_area_struct * vma;
 181        int unmapped_error, error = -EINVAL;
 182
 183        down_read(&current->mm->mmap_sem);
 184        if (flags & ~(MS_ASYNC | MS_INVALIDATE | MS_SYNC))
 185                goto out;
 186        if (start & ~PAGE_MASK)
 187                goto out;
 188        if ((flags & MS_ASYNC) && (flags & MS_SYNC))
 189                goto out;
 190        error = -ENOMEM;
 191        len = (len + ~PAGE_MASK) & PAGE_MASK;
 192        end = start + len;
 193        if (end < start)
 194                goto out;
 195        error = 0;
 196        if (end == start)
 197                goto out;
 198        /*
 199         * If the interval [start,end) covers some unmapped address ranges,
 200         * just ignore them, but return -ENOMEM at the end.
 201         */
 202        vma = find_vma(current->mm, start);
 203        unmapped_error = 0;
 204        for (;;) {
 205                /* Still start < end. */
 206                error = -ENOMEM;
 207                if (!vma)
 208                        goto out;
 209                /* Here start < vma->vm_end. */
 210                if (start < vma->vm_start) {
 211                        unmapped_error = -ENOMEM;
 212                        start = vma->vm_start;
 213                }
 214                /* Here vma->vm_start <= start < vma->vm_end. */
 215                if (end <= vma->vm_end) {
 216                        if (start < end) {
 217                                error = msync_interval(vma, start, end, flags);
 218                                if (error)
 219                                        goto out;
 220                        }
 221                        error = unmapped_error;
 222                        goto out;
 223                }
 224                /* Here vma->vm_start <= start < vma->vm_end < end. */
 225                error = msync_interval(vma, start, vma->vm_end, flags);
 226                if (error)
 227                        goto out;
 228                start = vma->vm_end;
 229                vma = vma->vm_next;
 230        }
 231out:
 232        up_read(&current->mm->mmap_sem);
 233        return error;
 234}
 235