NFS: Fix asynchronous read error handling
We must always call ->read_done() before we truncate the page data, or
decide to flag an error. The reasons are that
in NFSv2, ->read_done() is where the eof flag gets set.
in NFSv3/v4 ->read_done() handles EJUKEBOX-type errors, and
v4 state recovery.
However, we need to mark the pages as uptodate before we deal with short
read errors, since we may need to modify the nfs_read_data arguments.
We therefore split the current nfs_readpage_result() into two parts:
nfs_readpage_result(), which calls ->read_done() etc, and
nfs_readpage_retry(), which subsequently handles short reads.
Note: Removing the code that retries in case of a short read also fixes a
bug in nfs_direct_read_result(), which used to return a corrupted number of
bytes.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
diff --git a/fs/nfs/read.c b/fs/nfs/read.c
index 8b58bbf..e879ee6e 100644
--- a/fs/nfs/read.c
+++ b/fs/nfs/read.c
@@ -461,6 +461,55 @@
}
/*
+ * This is the callback from RPC telling us whether a reply was
+ * received or some error occurred (timeout or socket shutdown).
+ */
+int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
+{
+ int status;
+
+ dprintk("%s: %4d, (status %d)\n", __FUNCTION__, task->tk_pid,
+ task->tk_status);
+
+ status = NFS_PROTO(data->inode)->read_done(task, data);
+ if (status != 0)
+ return status;
+
+ nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
+
+ if (task->tk_status == -ESTALE) {
+ set_bit(NFS_INO_STALE, &NFS_FLAGS(data->inode));
+ nfs_mark_for_revalidate(data->inode);
+ }
+ spin_lock(&data->inode->i_lock);
+ NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
+ spin_unlock(&data->inode->i_lock);
+ return 0;
+}
+
+static int nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
+{
+ struct nfs_readargs *argp = &data->args;
+ struct nfs_readres *resp = &data->res;
+
+ if (resp->eof || resp->count == argp->count)
+ return 0;
+
+ /* This is a short read! */
+ nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
+ /* Has the server at least made some progress? */
+ if (resp->count == 0)
+ return 0;
+
+ /* Yes, so retry the read at the end of the data */
+ argp->offset += resp->count;
+ argp->pgbase += resp->count;
+ argp->count -= resp->count;
+ rpc_restart_call(task);
+ return -EAGAIN;
+}
+
+/*
* Handle a read reply that fills part of a page.
*/
static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
@@ -469,12 +518,16 @@
struct nfs_page *req = data->req;
struct page *page = req->wb_page;
- if (likely(task->tk_status >= 0))
- nfs_readpage_truncate_uninitialised_page(data);
- else
- SetPageError(page);
if (nfs_readpage_result(task, data) != 0)
return;
+
+ if (likely(task->tk_status >= 0)) {
+ nfs_readpage_truncate_uninitialised_page(data);
+ if (nfs_readpage_retry(task, data) != 0)
+ return;
+ }
+ if (unlikely(task->tk_status < 0))
+ SetPageError(page);
if (atomic_dec_and_test(&req->wb_complete)) {
if (!PageError(page))
SetPageUptodate(page);
@@ -502,25 +555,13 @@
count += base;
for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
SetPageUptodate(*pages);
- if (count != 0)
+ if (count == 0)
+ return;
+ /* Was this a short read? */
+ if (data->res.eof || data->res.count == data->args.count)
SetPageUptodate(*pages);
}
-static void nfs_readpage_set_pages_error(struct nfs_read_data *data)
-{
- unsigned int count = data->args.count;
- unsigned int base = data->args.pgbase;
- struct page **pages;
-
- pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
- base &= ~PAGE_CACHE_MASK;
- count += base;
- for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
- SetPageError(*pages);
- if (count != 0)
- SetPageError(*pages);
-}
-
/*
* This is the callback from RPC telling us whether a reply was
* received or some error occurred (timeout or socket shutdown).
@@ -529,19 +570,20 @@
{
struct nfs_read_data *data = calldata;
+ if (nfs_readpage_result(task, data) != 0)
+ return;
/*
- * Note: nfs_readpage_result may change the values of
+ * Note: nfs_readpage_retry may change the values of
* data->args. In the multi-page case, we therefore need
- * to ensure that we call the next nfs_readpage_set_page_uptodate()
- * first in the multi-page case.
+ * to ensure that we call nfs_readpage_set_pages_uptodate()
+ * first.
*/
if (likely(task->tk_status >= 0)) {
nfs_readpage_truncate_uninitialised_page(data);
nfs_readpage_set_pages_uptodate(data);
- } else
- nfs_readpage_set_pages_error(data);
- if (nfs_readpage_result(task, data) != 0)
- return;
+ if (nfs_readpage_retry(task, data) != 0)
+ return;
+ }
while (!list_empty(&data->pages)) {
struct nfs_page *req = nfs_list_entry(data->pages.next);
@@ -556,50 +598,6 @@
};
/*
- * This is the callback from RPC telling us whether a reply was
- * received or some error occurred (timeout or socket shutdown).
- */
-int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
-{
- struct nfs_readargs *argp = &data->args;
- struct nfs_readres *resp = &data->res;
- int status;
-
- dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
- task->tk_pid, task->tk_status);
-
- status = NFS_PROTO(data->inode)->read_done(task, data);
- if (status != 0)
- return status;
-
- nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, resp->count);
-
- if (task->tk_status < 0) {
- if (task->tk_status == -ESTALE) {
- set_bit(NFS_INO_STALE, &NFS_FLAGS(data->inode));
- nfs_mark_for_revalidate(data->inode);
- }
- } else if (resp->count < argp->count && !resp->eof) {
- /* This is a short read! */
- nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
- /* Has the server at least made some progress? */
- if (resp->count != 0) {
- /* Yes, so retry the read at the end of the data */
- argp->offset += resp->count;
- argp->pgbase += resp->count;
- argp->count -= resp->count;
- rpc_restart_call(task);
- return -EAGAIN;
- }
- task->tk_status = -EIO;
- }
- spin_lock(&data->inode->i_lock);
- NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
- spin_unlock(&data->inode->i_lock);
- return 0;
-}
-
-/*
* Read a page over NFS.
* We read the page synchronously in the following case:
* - The error flag is set for this page. This happens only when a
