[PATCH 1/2] pNFS/flexfiles: Don't tie up all the rpciod threads in resends

[PATCH 1/2] pNFS/flexfiles: Don't tie up all the rpciod threads in resends

[Trond Myklebust]

We do not want to have rpciod threads perform recursive calls into the
RPC layer since that can deadlock. In particular, having to wait for
a layoutget can be nasty... We want rather to defer scheduling those
retries until we're in the rpc_release() callback, since that is
called from the nfsiod workqueue.

Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
---
 fs/nfs/flexfilelayout/flexfilelayout.c | 11 ++++++++---
 include/linux/nfs_xdr.h                |  2 ++
 2 files changed, 10 insertions(+), 3 deletions(-)

diff --git a/fs/nfs/flexfilelayout/flexfilelayout.c b/fs/nfs/flexfilelayout/flexfilelayout.c
index 3ae038d9c292..336b4d560e2c 100644
--- a/fs/nfs/flexfilelayout/flexfilelayout.c
+++ b/fs/nfs/flexfilelayout/flexfilelayout.c
@@ -1243,17 +1243,18 @@ static int ff_layout_read_done_cb(struct rpc_task *task,
 					   hdr->ds_clp, hdr->lseg,
 					   hdr->pgio_mirror_idx);
 
+	clear_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags);
+	clear_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags);
 	switch (err) {
 	case -NFS4ERR_RESET_TO_PNFS:
 		if (ff_layout_choose_best_ds_for_read(hdr->lseg,
 					hdr->pgio_mirror_idx + 1,
 					&hdr->pgio_mirror_idx))
 			goto out_eagain;
-		ff_layout_read_record_layoutstats_done(task, hdr);
-		pnfs_read_resend_pnfs(hdr);
+		set_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags);
 		return task->tk_status;
 	case -NFS4ERR_RESET_TO_MDS:
-		ff_layout_reset_read(hdr);
+		set_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags);
 		return task->tk_status;
 	case -EAGAIN:
 		goto out_eagain;
@@ -1403,6 +1404,10 @@ static void ff_layout_read_release(void *data)
 	struct nfs_pgio_header *hdr = data;
 
 	ff_layout_read_record_layoutstats_done(&hdr->task, hdr);
+	if (test_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags))
+		pnfs_read_resend_pnfs(hdr);
+	else if (test_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags))
+		ff_layout_reset_read(hdr);
 	pnfs_generic_rw_release(data);
 }
 
diff --git a/include/linux/nfs_xdr.h b/include/linux/nfs_xdr.h
index 9dee3c23895d..712eed156d09 100644
--- a/include/linux/nfs_xdr.h
+++ b/include/linux/nfs_xdr.h
@@ -1438,6 +1438,8 @@ enum {
 	NFS_IOHDR_EOF,
 	NFS_IOHDR_REDO,
 	NFS_IOHDR_STAT,
+	NFS_IOHDR_RESEND_PNFS,
+	NFS_IOHDR_RESEND_MDS,
 };
 
 struct nfs_io_completion;
-- 
2.17.1

[PATCH 2/2] pNFS/flexfiles: Process writeback resends from nfsiod context as well

[Trond Myklebust]

Although the writeback resends are more robust than the reads, since they
are not immediately rescheduled by the same thread, we are better off
processing them in the same place as the reads.

Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
---
 fs/nfs/flexfilelayout/flexfilelayout.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/fs/nfs/flexfilelayout/flexfilelayout.c b/fs/nfs/flexfilelayout/flexfilelayout.c
index 336b4d560e2c..1386b774ec95 100644
--- a/fs/nfs/flexfilelayout/flexfilelayout.c
+++ b/fs/nfs/flexfilelayout/flexfilelayout.c
@@ -1428,12 +1428,14 @@ static int ff_layout_write_done_cb(struct rpc_task *task,
 					   hdr->ds_clp, hdr->lseg,
 					   hdr->pgio_mirror_idx);
 
+	clear_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags);
+	clear_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags);
 	switch (err) {
 	case -NFS4ERR_RESET_TO_PNFS:
-		ff_layout_reset_write(hdr, true);
+		set_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags);
 		return task->tk_status;
 	case -NFS4ERR_RESET_TO_MDS:
-		ff_layout_reset_write(hdr, false);
+		set_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags);
 		return task->tk_status;
 	case -EAGAIN:
 		return -EAGAIN;
@@ -1580,6 +1582,10 @@ static void ff_layout_write_release(void *data)
 	struct nfs_pgio_header *hdr = data;
 
 	ff_layout_write_record_layoutstats_done(&hdr->task, hdr);
+	if (test_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags))
+		ff_layout_reset_write(hdr, true);
+	else if (test_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags))
+		ff_layout_reset_write(hdr, false);
 	pnfs_generic_rw_release(data);
 }
 
-- 
2.17.1