]> bbs.cooldavid.org Git - net-next-2.6.git/commitdiff
ocfs2: When zero extending, do it by page.
authorJoel Becker <joel.becker@oracle.com>
Tue, 6 Jul 2010 21:36:06 +0000 (14:36 -0700)
committerJoel Becker <joel.becker@oracle.com>
Thu, 8 Jul 2010 20:24:49 +0000 (13:24 -0700)
ocfs2_zero_extend() does its zeroing block by block, but it calls a
function named ocfs2_write_zero_page().  Let's have
ocfs2_write_zero_page() handle the page level.  From
ocfs2_zero_extend()'s perspective, it is now page-at-a-time.

Signed-off-by: Joel Becker <joel.becker@oracle.com>
Cc: stable@kernel.org
fs/ocfs2/aops.c
fs/ocfs2/file.c

index 3623ca20cc186046cfbf840f03ce9705146b17ba..9a5c931439bd2f118bbf3d87a838c17cc89ddacf 100644 (file)
@@ -459,36 +459,6 @@ int walk_page_buffers(     handle_t *handle,
        return ret;
 }
 
-handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
-                                                        struct page *page,
-                                                        unsigned from,
-                                                        unsigned to)
-{
-       struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
-       handle_t *handle;
-       int ret = 0;
-
-       handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
-       if (IS_ERR(handle)) {
-               ret = -ENOMEM;
-               mlog_errno(ret);
-               goto out;
-       }
-
-       if (ocfs2_should_order_data(inode)) {
-               ret = ocfs2_jbd2_file_inode(handle, inode);
-               if (ret < 0)
-                       mlog_errno(ret);
-       }
-out:
-       if (ret) {
-               if (!IS_ERR(handle))
-                       ocfs2_commit_trans(osb, handle);
-               handle = ERR_PTR(ret);
-       }
-       return handle;
-}
-
 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 {
        sector_t status;
index 6a13ea64c44773fc239ad55a5237422ddee18e94..4cfc976a90675195a33df6b12df13732c51ad005 100644 (file)
@@ -724,28 +724,55 @@ leave:
        return status;
 }
 
+/*
+ * While a write will already be ordering the data, a truncate will not.
+ * Thus, we need to explicitly order the zeroed pages.
+ */
+static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode)
+{
+       struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
+       handle_t *handle = NULL;
+       int ret = 0;
+
+       if (!ocfs2_should_order_data(inode))
+               goto out;
+
+       handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
+       if (IS_ERR(handle)) {
+               ret = -ENOMEM;
+               mlog_errno(ret);
+               goto out;
+       }
+
+       ret = ocfs2_jbd2_file_inode(handle, inode);
+       if (ret < 0)
+               mlog_errno(ret);
+
+out:
+       if (ret) {
+               if (!IS_ERR(handle))
+                       ocfs2_commit_trans(osb, handle);
+               handle = ERR_PTR(ret);
+       }
+       return handle;
+}
+
 /* 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 ->write_begin() and ->write_end(). */
-static int ocfs2_write_zero_page(struct inode *inode,
-                                u64 size)
+static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
+                                u64 abs_to)
 {
        struct address_space *mapping = inode->i_mapping;
        struct page *page;
-       unsigned long index;
-       unsigned int offset;
+       unsigned long index = abs_from >> PAGE_CACHE_SHIFT;
        handle_t *handle = NULL;
        int ret;
+       unsigned zero_from, zero_to, block_start, block_end;
 
-       offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
-       /* ugh.  in prepare/commit_write, if from==to==start of block, we
-       ** skip the prepare.  make sure we never send an offset for the start
-       ** of a block
-       */
-       if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
-               offset++;
-       }
-       index = size >> PAGE_CACHE_SHIFT;
+       BUG_ON(abs_from >= abs_to);
+       BUG_ON(abs_to > (((u64)index + 1) << PAGE_CACHE_SHIFT));
+       BUG_ON(abs_from & (inode->i_blkbits - 1));
 
        page = grab_cache_page(mapping, index);
        if (!page) {
@@ -754,31 +781,51 @@ static int ocfs2_write_zero_page(struct inode *inode,
                goto out;
        }
 
-       ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
-       if (ret < 0) {
-               mlog_errno(ret);
-               goto out_unlock;
-       }
+       /* Get the offsets within the page that we want to zero */
+       zero_from = abs_from & (PAGE_CACHE_SIZE - 1);
+       zero_to = abs_to & (PAGE_CACHE_SIZE - 1);
+       if (!zero_to)
+               zero_to = PAGE_CACHE_SIZE;
 
-       if (ocfs2_should_order_data(inode)) {
-               handle = ocfs2_start_walk_page_trans(inode, page, offset,
-                                                    offset);
-               if (IS_ERR(handle)) {
-                       ret = PTR_ERR(handle);
-                       handle = NULL;
+       /* We know that zero_from is block aligned */
+       for (block_start = zero_from; block_start < zero_to;
+            block_start = block_end) {
+               block_end = block_start + (1 << inode->i_blkbits);
+
+               /*
+                * block_start is block-aligned.  Bump it by one to
+                * force ocfs2_{prepare,commit}_write() to zero the
+                * whole block.
+                */
+               ret = ocfs2_prepare_write_nolock(inode, page,
+                                                block_start + 1,
+                                                block_start + 1);
+               if (ret < 0) {
+                       mlog_errno(ret);
                        goto out_unlock;
                }
-       }
 
-       /* must not update i_size! */
-       ret = block_commit_write(page, offset, offset);
-       if (ret < 0)
-               mlog_errno(ret);
-       else
-               ret = 0;
+               if (!handle) {
+                       handle = ocfs2_zero_start_ordered_transaction(inode);
+                       if (IS_ERR(handle)) {
+                               ret = PTR_ERR(handle);
+                               handle = NULL;
+                               break;
+                       }
+               }
+
+               /* must not update i_size! */
+               ret = block_commit_write(page, block_start + 1,
+                                        block_start + 1);
+               if (ret < 0)
+                       mlog_errno(ret);
+               else
+                       ret = 0;
+       }
 
        if (handle)
                ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
+
 out_unlock:
        unlock_page(page);
        page_cache_release(page);
@@ -790,18 +837,21 @@ static int ocfs2_zero_extend(struct inode *inode,
                             u64 zero_to_size)
 {
        int ret = 0;
-       u64 start_off;
+       u64 start_off, next_off;
        struct super_block *sb = inode->i_sb;
 
        start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
        while (start_off < zero_to_size) {
-               ret = ocfs2_write_zero_page(inode, start_off);
+               next_off = (start_off & PAGE_CACHE_MASK) + PAGE_CACHE_SIZE;
+               if (next_off > zero_to_size)
+                       next_off = zero_to_size;
+               ret = ocfs2_write_zero_page(inode, start_off, next_off);
                if (ret < 0) {
                        mlog_errno(ret);
                        goto out;
                }
 
-               start_off += sb->s_blocksize;
+               start_off = next_off;
 
                /*
                 * Very large extends have the potential to lock up