From: Linda Dunaphant <linda.dunaphant@ccur.com>
One of our applications running on a Linux client experiences partial file
data corruption (nulls) if the NFS filesystem on a non-Linux server is
mounted as NFS3. It works properly if mounted NFS2. The Linux NFS client
creates a file on the server and issues the following sequence of lseeks,
writes, and read calls: init write buffer to non-null values
lseek: SEEK_SET to 52
write: 224 bytes
lseek: SEEK_SET to 4096
write: 224 bytes
lseek: SEEK_SET to 0
read: 276 bytes
Using ethereal, I found that the data read back over the wire from the
server for bytes 0 to 276 was correct (bytes 52-275 were not null).
However, the data returned to the application for bytes 52 to 275 was null.
We have several non-Linux NFS servers that do not return the EOF status
flag on a read when data is being returned. A second read at the updated
offset returns the EOF flag. This causes the nfs_readpage_result() short
read logic to be used in this case. If some progress is made on the read
(count != 0), nfs_readpage_result() adjusts the starting position in pgbase
and offset, and decrements the count by the amount that was completed. It
then restarts the read call. Once the original count is satified or an EOF
occurs, nfs_readpage_result_full() is called to null fill the end of any
page(s) that were not satisfied by the read(s).
I found nfs_readpage_result_full() is not taking into account a short read
situation may have occurred. In my example, data->res.count was 0 after
the second read (for bytes 276 to 4095) due to the EOF. Because
req->wb_pgbase was 0 and req->wb_bytes was 4096, memclear_highpage_flush()
cleared all 4096 bytes of the page even though 276 bytes had been read by
the first short read. This caused bytes 52 to 275 to be clobbered by
nulls.
I changed count in nfs_readpage_result_full() to be a total of all the
reads by adding data->args.pgbase to data->res.count (count of the last
read). I found that data->args.pgbase contained the sum of any previous
short reads that may have occurred, and 0 otherwise. In my example, bytes
276 to 4095 are now cleared by memclear_highpage_flush() instead of bytes 0
to 4095 which fixes the problem for this application.
Signed-off-by: Andrew Morton <akpm@osdl.org>
---
fs/nfs/read.c | 2 +-
1 files changed, 1 insertion(+), 1 deletion(-)
diff -puN fs/nfs/read.c~nfs-nfs3-page-null-fill-in-a-short-read-situation fs/nfs/read.c
--- devel/fs/nfs/read.c~nfs-nfs3-page-null-fill-in-a-short-read-situation 2005-08-21 23:48:51.000000000 -0700
+++ devel-akpm/fs/nfs/read.c 2005-08-21 23:48:51.000000000 -0700
@@ -424,7 +424,7 @@ static void nfs_readpage_result_partial(
*/
static void nfs_readpage_result_full(struct nfs_read_data *data, int status)
{
- unsigned int count = data->res.count;
+ unsigned int count = data->res.count + data->args.pgbase;
while (!list_empty(&data->pages)) {
struct nfs_page *req = nfs_list_entry(data->pages.next);
_