fs/ocfs2/mmap.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * mmap.c
   5  *
   6  * Code to deal with the mess that is clustered mmap.
   7  *
   8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  */
  25
  26 #include <linux/fs.h>
  27 #include <linux/types.h>
  28 #include <linux/slab.h>
  29 #include <linux/highmem.h>
  30 #include <linux/pagemap.h>
  31 #include <linux/uio.h>
  32 #include <linux/signal.h>
  33 #include <linux/rbtree.h>
  34
  35 #define MLOG_MASK_PREFIX ML_FILE_IO
  36 #include <cluster/masklog.h>
  37
  38 #include "ocfs2.h"
  39
  40 #include "aops.h"
  41 #include "dlmglue.h"
  42 #include "file.h"
  43 #include "inode.h"
  44 #include "mmap.h"
  45
  46 static inline int ocfs2_vm_op_block_sigs(sigset_t *blocked, sigset_t *oldset)
  47 {
  48         /* The best way to deal with signals in the vm path is
  49          * to block them upfront, rather than allowing the
  50          * locking paths to return -ERESTARTSYS. */
  51         sigfillset(blocked);
  52
  53         /* We should technically never get a bad return value
  54          * from sigprocmask */
  55         return sigprocmask(SIG_BLOCK, blocked, oldset);
  56 }
  57
  58 static inline int ocfs2_vm_op_unblock_sigs(sigset_t *oldset)
  59 {
  60         return sigprocmask(SIG_SETMASK, oldset, NULL);
  61 }
  62
  63 static int ocfs2_fault(struct vm_area_struct *area, struct vm_fault *vmf)
  64 {
  65         sigset_t blocked, oldset;
  66         int error, ret;
  67
  68         mlog_entry("(area=%p, page offset=%lu)\n", area, vmf->pgoff);
  69
  70         error = ocfs2_vm_op_block_sigs(&blocked, &oldset);
  71         if (error < 0) {
  72                 mlog_errno(error);
  73                 ret = VM_FAULT_SIGBUS;
  74                 goto out;
  75         }
  76
  77         ret = filemap_fault(area, vmf);
  78
  79         error = ocfs2_vm_op_unblock_sigs(&oldset);
  80         if (error < 0)
  81                 mlog_errno(error);
  82 out:
  83         mlog_exit_ptr(vmf->page);
  84         return ret;
  85 }
  86
  87 static int __ocfs2_page_mkwrite(struct inode *inode, struct buffer_head *di_bh,
  88                                 struct page *page)
  89 {
  90         int ret;
  91         struct address_space *mapping = inode->i_mapping;
  92         loff_t pos = page_offset(page);
  93         unsigned int len = PAGE_CACHE_SIZE;
  94         pgoff_t last_index;
  95         struct page *locked_page = NULL;
  96         void *fsdata;
  97         loff_t size = i_size_read(inode);
  98
  99         /*
 100          * Another node might have truncated while we were waiting on
 101          * cluster locks.
 102          */
 103         last_index = size >> PAGE_CACHE_SHIFT;
 104         if (page->index > last_index) {
 105                 ret = -EINVAL;
 106                 goto out;
 107         }
 108
 109         /*
 110          * The i_size check above doesn't catch the case where nodes
 111          * truncated and then re-extended the file. We'll re-check the
 112          * page mapping after taking the page lock inside of
 113          * ocfs2_write_begin_nolock().
 114          */
 115         if (!PageUptodate(page) || page->mapping != inode->i_mapping) {
 116                 ret = -EINVAL;
 117                 goto out;
 118         }
 119
 120         /*
 121          * Call ocfs2_write_begin() and ocfs2_write_end() to take
 122          * advantage of the allocation code there. We pass a write
 123          * length of the whole page (chopped to i_size) to make sure
 124          * the whole thing is allocated.
 125          *
 126          * Since we know the page is up to date, we don't have to
 127          * worry about ocfs2_write_begin() skipping some buffer reads
 128          * because the "write" would invalidate their data.
 129          */
 130         if (page->index == last_index)
 131                 len = size & ~PAGE_CACHE_MASK;
 132
 133         ret = ocfs2_write_begin_nolock(mapping, pos, len, 0, &locked_page,
 134                                        &fsdata, di_bh, page);
 135         if (ret) {
 136                 if (ret != -ENOSPC)
 137                         mlog_errno(ret);
 138                 goto out;
 139         }
 140
 141         ret = ocfs2_write_end_nolock(mapping, pos, len, len, locked_page,
 142                                      fsdata);
 143         if (ret < 0) {
 144                 mlog_errno(ret);
 145                 goto out;
 146         }
 147         BUG_ON(ret != len);
 148         ret = 0;
 149 out:
 150         return ret;
 151 }
 152
 153 static int ocfs2_page_mkwrite(struct vm_area_struct *vma, struct page *page)
 154 {
 155         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
 156         struct buffer_head *di_bh = NULL;
 157         sigset_t blocked, oldset;
 158         int ret, ret2;
 159
 160         ret = ocfs2_vm_op_block_sigs(&blocked, &oldset);
 161         if (ret < 0) {
 162                 mlog_errno(ret);
 163                 return ret;
 164         }
 165
 166         /*
 167          * The cluster locks taken will block a truncate from another
 168          * node. Taking the data lock will also ensure that we don't
 169          * attempt page truncation as part of a downconvert.
 170          */
 171         ret = ocfs2_meta_lock(inode, &di_bh, 1);
 172         if (ret < 0) {
 173                 mlog_errno(ret);
 174                 goto out;
 175         }
 176
 177         /*
 178          * The alloc sem should be enough to serialize with
 179          * ocfs2_truncate_file() changing i_size as well as any thread
 180          * modifying the inode btree.
 181          */
 182         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 183
 184         ret = ocfs2_data_lock(inode, 1);
 185         if (ret < 0) {
 186                 mlog_errno(ret);
 187                 goto out_meta_unlock;
 188         }
 189
 190         ret = __ocfs2_page_mkwrite(inode, di_bh, page);
 191
 192         ocfs2_data_unlock(inode, 1);
 193
 194 out_meta_unlock:
 195         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 196
 197         brelse(di_bh);
 198         ocfs2_meta_unlock(inode, 1);
 199
 200 out:
 201         ret2 = ocfs2_vm_op_unblock_sigs(&oldset);
 202         if (ret2 < 0)
 203                 mlog_errno(ret2);
 204
 205         return ret;
 206 }
 207
 208 static struct vm_operations_struct ocfs2_file_vm_ops = {
 209         .fault          = ocfs2_fault,
 210         .page_mkwrite   = ocfs2_page_mkwrite,
 211 };
 212
 213 int ocfs2_mmap(struct file *file, struct vm_area_struct *vma)
 214 {
 215         int ret = 0, lock_level = 0;
 216
 217         ret = ocfs2_meta_lock_atime(file->f_dentry->d_inode,
 218                                     file->f_vfsmnt, &lock_level);
 219         if (ret < 0) {
 220                 mlog_errno(ret);
 221                 goto out;
 222         }
 223         ocfs2_meta_unlock(file->f_dentry->d_inode, lock_level);
 224 out:
 225         vma->vm_ops = &ocfs2_file_vm_ops;
 226         vma->vm_flags |= VM_CAN_NONLINEAR;
 227         return 0;
 228 }
 229