]> err.no Git - linux-2.6/commitdiff
NFS: Fix asynchronous read error handling
authorTrond Myklebust <Trond.Myklebust@netapp.com>
Tue, 14 Nov 2006 21:12:23 +0000 (16:12 -0500)
committerTrond Myklebust <Trond.Myklebust@netapp.com>
Wed, 6 Dec 2006 15:46:27 +0000 (10:46 -0500)
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>
fs/nfs/read.c

index 8b58bbf6e39e6fbdddcef66e31ed7ffe77745c6f..e879ee6e1385608fee18194d27915edd5fb42531 100644 (file)
@@ -460,6 +460,55 @@ nfs_pagein_list(struct list_head *head, int rpages)
        return error;
 }
 
+/*
+ * 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.
  */
@@ -469,12 +518,16 @@ static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
        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 @@ static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
        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 @@ static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
 {
        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);
 
@@ -555,50 +597,6 @@ static const struct rpc_call_ops nfs_read_full_ops = {
        .rpc_release = nfs_readdata_release,
 };
 
-/*
- * 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: