ocfs2: take data locks around extend
authorMark Fasheh <mark.fasheh@oracle.com>
Sat, 6 May 2006 02:04:03 +0000 (19:04 -0700)
committerMark Fasheh <mark.fasheh@oracle.com>
Wed, 17 May 2006 21:38:47 +0000 (14:38 -0700)
We need to take a data lock around extends to protect the pages that
ocfs2_zero_extend is going to be pulling into the page cache. Otherwise an
extend on one node might populate the page cache with data pages that have
no lock coverage.

Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
fs/ocfs2/aops.c
fs/ocfs2/aops.h
fs/ocfs2/file.c

index 0d858d0b25be70f496e1fb0ec863ec4778ae3c57..47152bf9a7f26cdfb82cd02240e5546ba3e47485 100644 (file)
@@ -276,13 +276,29 @@ static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
        return ret;
 }
 
+/* This can also be called from ocfs2_write_zero_page() which has done
+ * it's own cluster locking. */
+int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
+                              unsigned from, unsigned to)
+{
+       int ret;
+
+       down_read(&OCFS2_I(inode)->ip_alloc_sem);
+
+       ret = block_prepare_write(page, from, to, ocfs2_get_block);
+
+       up_read(&OCFS2_I(inode)->ip_alloc_sem);
+
+       return ret;
+}
+
 /*
  * ocfs2_prepare_write() can be an outer-most ocfs2 call when it is called
  * from loopback.  It must be able to perform its own locking around
  * ocfs2_get_block().
  */
-int ocfs2_prepare_write(struct file *file, struct page *page,
-                       unsigned from, unsigned to)
+static int ocfs2_prepare_write(struct file *file, struct page *page,
+                              unsigned from, unsigned to)
 {
        struct inode *inode = page->mapping->host;
        int ret;
@@ -295,11 +311,7 @@ int ocfs2_prepare_write(struct file *file, struct page *page,
                goto out;
        }
 
-       down_read(&OCFS2_I(inode)->ip_alloc_sem);
-
-       ret = block_prepare_write(page, from, to, ocfs2_get_block);
-
-       up_read(&OCFS2_I(inode)->ip_alloc_sem);
+       ret = ocfs2_prepare_write_nolock(inode, page, from, to);
 
        ocfs2_meta_unlock(inode, 0);
 out:
@@ -625,11 +637,31 @@ static ssize_t ocfs2_direct_IO(int rw,
        int ret;
 
        mlog_entry_void();
+
+       /*
+        * We get PR data locks even for O_DIRECT.  This allows
+        * concurrent O_DIRECT I/O but doesn't let O_DIRECT with
+        * extending and buffered zeroing writes race.  If they did
+        * race then the buffered zeroing could be written back after
+        * the O_DIRECT I/O.  It's one thing to tell people not to mix
+        * buffered and O_DIRECT writes, but expecting them to
+        * understand that file extension is also an implicit buffered
+        * write is too much.  By getting the PR we force writeback of
+        * the buffered zeroing before proceeding.
+        */
+       ret = ocfs2_data_lock(inode, 0);
+       if (ret < 0) {
+               mlog_errno(ret);
+               goto out;
+       }
+       ocfs2_data_unlock(inode, 0);
+
        ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
                                            inode->i_sb->s_bdev, iov, offset,
                                            nr_segs, 
                                            ocfs2_direct_IO_get_blocks,
                                            ocfs2_dio_end_io);
+out:
        mlog_exit(ret);
        return ret;
 }
index d40456d509a00e8ddecac4e28a26d1482140f23f..e88c3f0b8fa9f3bfc5078f992360cd3cbea8f69c 100644 (file)
@@ -22,8 +22,8 @@
 #ifndef OCFS2_AOPS_H
 #define OCFS2_AOPS_H
 
-int ocfs2_prepare_write(struct file *file, struct page *page,
-                       unsigned from, unsigned to);
+int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
+                              unsigned from, unsigned to);
 
 struct ocfs2_journal_handle *ocfs2_start_walk_page_trans(struct inode *inode,
                                                         struct page *page,
index 581eb451a41a36796a2a2d55b7463028d4dae10e..20fffeed630b4c1cd437742ad900205718a80b13 100644 (file)
@@ -613,7 +613,8 @@ leave:
 
 /* Some parts of this taken from generic_cont_expand, which turned out
  * to be too fragile to do exactly what we need without us having to
- * worry about recursive locking in ->commit_write(). */
+ * worry about recursive locking in ->prepare_write() and
+ * ->commit_write(). */
 static int ocfs2_write_zero_page(struct inode *inode,
                                 u64 size)
 {
@@ -641,7 +642,7 @@ static int ocfs2_write_zero_page(struct inode *inode,
                goto out;
        }
 
-       ret = ocfs2_prepare_write(NULL, page, offset, offset);
+       ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
        if (ret < 0) {
                mlog_errno(ret);
                goto out_unlock;
@@ -695,13 +696,26 @@ out:
        return ret;
 }
 
+/* 
+ * A tail_to_skip value > 0 indicates that we're being called from
+ * ocfs2_file_aio_write(). This has the following implications:
+ *
+ * - we don't want to update i_size
+ * - di_bh will be NULL, which is fine because it's only used in the
+ *   case where we want to update i_size.
+ * - ocfs2_zero_extend() will then only be filling the hole created
+ *   between i_size and the start of the write.
+ */
 static int ocfs2_extend_file(struct inode *inode,
                             struct buffer_head *di_bh,
-                            u64 new_i_size)
+                            u64 new_i_size,
+                            size_t tail_to_skip)
 {
        int ret = 0;
        u32 clusters_to_add;
 
+       BUG_ON(!tail_to_skip && !di_bh);
+
        /* setattr sometimes calls us like this. */
        if (new_i_size == 0)
                goto out;
@@ -714,27 +728,44 @@ static int ocfs2_extend_file(struct inode *inode,
                OCFS2_I(inode)->ip_clusters;
 
        if (clusters_to_add) {
-               ret = ocfs2_extend_allocation(inode, clusters_to_add);
+               /* 
+                * protect the pages that ocfs2_zero_extend is going to
+                * be pulling into the page cache.. we do this before the
+                * metadata extend so that we don't get into the situation
+                * where we've extended the metadata but can't get the data
+                * lock to zero.
+                */
+               ret = ocfs2_data_lock(inode, 1);
                if (ret < 0) {
                        mlog_errno(ret);
                        goto out;
                }
 
-               ret = ocfs2_zero_extend(inode, new_i_size);
+               ret = ocfs2_extend_allocation(inode, clusters_to_add);
                if (ret < 0) {
                        mlog_errno(ret);
-                       goto out;
+                       goto out_unlock;
                }
-       } 
 
-       /* No allocation required, we just use this helper to
-        * do a trivial update of i_size. */
-       ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
-       if (ret < 0) {
-               mlog_errno(ret);
-               goto out;
+               ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
+               if (ret < 0) {
+                       mlog_errno(ret);
+                       goto out_unlock;
+               }
+       }
+
+       if (!tail_to_skip) {
+               /* We're being called from ocfs2_setattr() which wants
+                * us to update i_size */
+               ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
+               if (ret < 0)
+                       mlog_errno(ret);
        }
 
+out_unlock:
+       if (clusters_to_add) /* this is the only case in which we lock */
+               ocfs2_data_unlock(inode, 1);
+
 out:
        return ret;
 }
@@ -793,7 +824,7 @@ int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
                if (i_size_read(inode) > attr->ia_size)
                        status = ocfs2_truncate_file(inode, bh, attr->ia_size);
                else
-                       status = ocfs2_extend_file(inode, bh, attr->ia_size);
+                       status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
                if (status < 0) {
                        if (status != -ENOSPC)
                                mlog_errno(status);
@@ -1049,21 +1080,12 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
                if (!clusters)
                        break;
 
-               ret = ocfs2_extend_allocation(inode, clusters);
+               ret = ocfs2_extend_file(inode, NULL, newsize, count);
                if (ret < 0) {
                        if (ret != -ENOSPC)
                                mlog_errno(ret);
                        goto out;
                }
-
-               /* Fill any holes which would've been created by this
-                * write. If we're O_APPEND, this will wind up
-                * (correctly) being a noop. */
-               ret = ocfs2_zero_extend(inode, (u64) newsize - count);
-               if (ret < 0) {
-                       mlog_errno(ret);
-                       goto out;
-               }
                break;
        }