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 struct page *ocfs2_fault(struct vm_area_struct *area,
  64                                                 struct fault_data *fdata)
  65 {
  66         struct page *page = NULL;
  67         sigset_t blocked, oldset;
  68         int ret;
  69
  70         mlog_entry("(area=%p, page offset=%lu)\n", area, fdata->pgoff);
  71
  72         ret = ocfs2_vm_op_block_sigs(&blocked, &oldset);
  73         if (ret < 0) {
  74                 fdata->type = VM_FAULT_SIGBUS;
  75                 mlog_errno(ret);
  76                 goto out;
  77         }
  78
  79         page = filemap_fault(area, fdata);
  80
  81         ret = ocfs2_vm_op_unblock_sigs(&oldset);
  82         if (ret < 0)
  83                 mlog_errno(ret);
  84 out:
  85         mlog_exit_ptr(page);
  86         return page;
  87 }
  88
  89 static int __ocfs2_page_mkwrite(struct inode *inode, struct buffer_head *di_bh,
  90                                 struct page *page)
  91 {
  92         int ret;
  93         struct address_space *mapping = inode->i_mapping;
  94         loff_t pos = page->index << PAGE_CACHE_SHIFT;
  95         unsigned int len = PAGE_CACHE_SIZE;
  96         pgoff_t last_index;
  97         struct page *locked_page = NULL;
  98         void *fsdata;
  99         loff_t size = i_size_read(inode);
 100
 101         /*
 102          * Another node might have truncated while we were waiting on
 103          * cluster locks.
 104          */
 105         last_index = size >> PAGE_CACHE_SHIFT;
 106         if (page->index > last_index) {
 107                 ret = -EINVAL;
 108                 goto out;
 109         }
 110
 111         /*
 112          * The i_size check above doesn't catch the case where nodes
 113          * truncated and then re-extended the file. We'll re-check the
 114          * page mapping after taking the page lock inside of
 115          * ocfs2_write_begin_nolock().
 116          */
 117         if (!PageUptodate(page) || page->mapping != inode->i_mapping) {
 118                 ret = -EINVAL;
 119                 goto out;
 120         }
 121
 122         /*
 123          * Call ocfs2_write_begin() and ocfs2_write_end() to take
 124          * advantage of the allocation code there. We pass a write
 125          * length of the whole page (chopped to i_size) to make sure
 126          * the whole thing is allocated.
 127          *
 128          * Since we know the page is up to date, we don't have to
 129          * worry about ocfs2_write_begin() skipping some buffer reads
 130          * because the "write" would invalidate their data.
 131          */
 132         if (page->index == last_index)
 133                 len = size & ~PAGE_CACHE_MASK;
 134
 135         ret = ocfs2_write_begin_nolock(mapping, pos, len, 0, &locked_page,
 136                                        &fsdata, di_bh, page);
 137         if (ret) {
 138                 if (ret != -ENOSPC)
 139                         mlog_errno(ret);
 140                 goto out;
 141         }
 142
 143         ret = ocfs2_write_end_nolock(mapping, pos, len, len, locked_page,
 144                                      fsdata);
 145         if (ret < 0) {
 146                 mlog_errno(ret);
 147                 goto out;
 148         }
 149         BUG_ON(ret != len);
 150         ret = 0;
 151 out:
 152         return ret;
 153 }
 154
 155 static int ocfs2_page_mkwrite(struct vm_area_struct *vma, struct page *page)
 156 {
 157         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
 158         struct buffer_head *di_bh = NULL;
 159         sigset_t blocked, oldset;
 160         int ret, ret2;
 161
 162         ret = ocfs2_vm_op_block_sigs(&blocked, &oldset);
 163         if (ret < 0) {
 164                 mlog_errno(ret);
 165                 return ret;
 166         }
 167
 168         /*
 169          * The cluster locks taken will block a truncate from another
 170          * node. Taking the data lock will also ensure that we don't
 171          * attempt page truncation as part of a downconvert.
 172          */
 173         ret = ocfs2_meta_lock(inode, &di_bh, 1);
 174         if (ret < 0) {
 175                 mlog_errno(ret);
 176                 goto out;
 177         }
 178
 179         /*
 180          * The alloc sem should be enough to serialize with
 181          * ocfs2_truncate_file() changing i_size as well as any thread
 182          * modifying the inode btree.
 183          */
 184         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 185
 186         ret = ocfs2_data_lock(inode, 1);
 187         if (ret < 0) {
 188                 mlog_errno(ret);
 189                 goto out_meta_unlock;
 190         }
 191
 192         ret = __ocfs2_page_mkwrite(inode, di_bh, page);
 193
 194         ocfs2_data_unlock(inode, 1);
 195
 196 out_meta_unlock:
 197         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 198
 199         brelse(di_bh);
 200         ocfs2_meta_unlock(inode, 1);
 201
 202 out:
 203         ret2 = ocfs2_vm_op_unblock_sigs(&oldset);
 204         if (ret2 < 0)
 205                 mlog_errno(ret2);
 206
 207         return ret;
 208 }
 209
 210 static struct vm_operations_struct ocfs2_file_vm_ops = {
 211         .fault          = ocfs2_fault,
 212         .page_mkwrite   = ocfs2_page_mkwrite,
 213 };
 214
 215 int ocfs2_mmap(struct file *file, struct vm_area_struct *vma)
 216 {
 217         int ret = 0, lock_level = 0;
 218
 219         ret = ocfs2_meta_lock_atime(file->f_dentry->d_inode,
 220                                     file->f_vfsmnt, &lock_level);
 221         if (ret < 0) {
 222                 mlog_errno(ret);
 223                 goto out;
 224         }
 225         ocfs2_meta_unlock(file->f_dentry->d_inode, lock_level);
 226 out:
 227         vma->vm_ops = &ocfs2_file_vm_ops;
 228         vma->vm_flags |= VM_CAN_INVALIDATE | VM_CAN_NONLINEAR;
 229         return 0;
 230 }
 231