NFS: Remove pNFS bloat from the generic write path

We have no business doing any this in the standard write release path.
Get rid of it, and put it in the pNFS layer.

Also, while we're at it, get rid of the completely bogus unlock/relock
semantics that were present in nfs_writeback_release_full(). It is
not only unnecessary, but actually dangerous to release the write lock
just in order to take it again in nfs_page_async_flush(). Better just
to open code the pgio operations in a pnfs helper.

Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

1 files changed, 28 insertions, 2 deletions

diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c
index f881a6387942..17149a490065 100644
--- a/fs/nfs/pnfs.c
+++ b/fs/nfs/pnfs.c
@@ -1166,6 +1166,33 @@ pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 }
 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
 
+static int pnfs_write_done_resend_to_mds(struct inode *inode, struct list_head *head)
+{
+        struct nfs_pageio_descriptor pgio;
+        LIST_HEAD(failed);
+
+        /* Resend all requests through the MDS */
+        nfs_pageio_init_write_mds(&pgio, inode, FLUSH_STABLE);
+        while (!list_empty(head)) {
+                struct nfs_page *req = nfs_list_entry(head->next);
+
+                nfs_list_remove_request(req);
+                if (!nfs_pageio_add_request(&pgio, req))
+                        nfs_list_add_request(req, &failed);
+        }
+        nfs_pageio_complete(&pgio);
+
+        if (!list_empty(&failed)) {
+                /* For some reason our attempt to resend pages. Mark the
+                 * overall send request as having failed, and let
+                 * nfs_writeback_release_full deal with the error.
+                 */
+                list_move(&failed, head);
+                return -EIO;
+        }
+        return 0;
+}
+
 /*
  * Called by non rpc-based layout drivers
  */
@@ -1175,8 +1202,6 @@ void pnfs_ld_write_done(struct nfs_write_data *data)
                 pnfs_set_layoutcommit(data);
                 data->mds_ops->rpc_call_done(&data->task, data);
         } else {
-                put_lseg(data->lseg);
-                data->lseg = NULL;
                 dprintk("pnfs write error = %d\n", data->pnfs_error);
                 if (NFS_SERVER(data->inode)->pnfs_curr_ld->flags &
                                                 PNFS_LAYOUTRET_ON_ERROR) {
@@ -1187,6 +1212,7 @@ void pnfs_ld_write_done(struct nfs_write_data *data)
                                   &NFS_I(data->inode)->flags);
                         pnfs_return_layout(data->inode);
                 }
+                data->task.tk_status = pnfs_write_done_resend_to_mds(data->inode, &data->pages);
         }
         data->mds_ops->rpc_release(data);
 }