pnfs: recoalesce when ld write pagelist fails
Peng Tao [Fri, 23 Sep 2011 01:50:14 +0000 (21:50 -0400)]
commit 8ce160c5ef06cc89c2b6b26bfa5ef7a5ce2c93e0 upstream.

For pnfs pagelist write failure, we need to pg_recoalesce and resend IO to
mds.

Signed-off-by: Peng Tao <peng_tao@emc.com>
Signed-off-by: Jim Rees <rees@umich.edu>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>

fs/nfs/pnfs.c
fs/nfs/pnfs.h
fs/nfs/write.c

index 6b19fff..a205c8e 100644 (file)
@@ -1168,23 +1168,17 @@ EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
 /*
  * Called by non rpc-based layout drivers
  */
-int
-pnfs_ld_write_done(struct nfs_write_data *data)
+void pnfs_ld_write_done(struct nfs_write_data *data)
 {
-       int status;
-
-       if (!data->pnfs_error) {
+       if (likely(!data->pnfs_error)) {
                pnfs_set_layoutcommit(data);
                data->mds_ops->rpc_call_done(&data->task, data);
-               data->mds_ops->rpc_release(data);
-               return 0;
+       } else {
+               put_lseg(data->lseg);
+               data->lseg = NULL;
+               dprintk("pnfs write error = %d\n", data->pnfs_error);
        }
-
-       dprintk("%s: pnfs_error=%d, retry via MDS\n", __func__,
-               data->pnfs_error);
-       status = nfs_initiate_write(data, NFS_CLIENT(data->inode),
-                                   data->mds_ops, NFS_FILE_SYNC);
-       return status ? : -EAGAIN;
+       data->mds_ops->rpc_release(data);
 }
 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
 
index 94e760e..71c23d4 100644 (file)
@@ -201,7 +201,7 @@ void pnfs_set_layoutcommit(struct nfs_write_data *wdata);
 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data);
 int pnfs_layoutcommit_inode(struct inode *inode, bool sync);
 int _pnfs_return_layout(struct inode *);
-int pnfs_ld_write_done(struct nfs_write_data *);
+void pnfs_ld_write_done(struct nfs_write_data *);
 int pnfs_ld_read_done(struct nfs_read_data *);
 struct pnfs_layout_segment *pnfs_update_layout(struct inode *ino,
                                               struct nfs_open_context *ctx,
index c9bd2a6..62eb5c5 100644 (file)
@@ -1165,7 +1165,13 @@ static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
 static void nfs_writeback_release_full(void *calldata)
 {
        struct nfs_write_data   *data = calldata;
-       int status = data->task.tk_status;
+       int ret, status = data->task.tk_status;
+       struct nfs_pageio_descriptor pgio;
+
+       if (data->pnfs_error) {
+               nfs_pageio_init_write_mds(&pgio, data->inode, FLUSH_STABLE);
+               pgio.pg_recoalesce = 1;
+       }
 
        /* Update attributes as result of writeback. */
        while (!list_empty(&data->pages)) {
@@ -1181,6 +1187,11 @@ static void nfs_writeback_release_full(void *calldata)
                        req->wb_bytes,
                        (long long)req_offset(req));
 
+               if (data->pnfs_error) {
+                       dprintk(", pnfs error = %d\n", data->pnfs_error);
+                       goto next;
+               }
+
                if (status < 0) {
                        nfs_set_pageerror(page);
                        nfs_context_set_write_error(req->wb_context, status);
@@ -1200,7 +1211,19 @@ remove_request:
        next:
                nfs_clear_page_tag_locked(req);
                nfs_end_page_writeback(page);
+               if (data->pnfs_error) {
+                       lock_page(page);
+                       nfs_pageio_cond_complete(&pgio, page->index);
+                       ret = nfs_page_async_flush(&pgio, page, 0);
+                       if (ret) {
+                               nfs_set_pageerror(page);
+                               dprintk("rewrite to MDS error = %d\n", ret);
+                       }
+                       unlock_page(page);
+               }
        }
+       if (data->pnfs_error)
+               nfs_pageio_complete(&pgio);
        nfs_writedata_release(calldata);
 }