nfs/blocklayout: Limit repeat device registration on failure

[ Upstream commit 614733f ]

Every pNFS SCSI IO wants to do LAYOUTGET, then within the layout find the
device which can drive GETDEVINFO, then finally may need to prep the device
with a reservation.  This slow work makes a mess of IO latencies if one of
the later steps is going to fail for awhile.

If we're unable to register a SCSI device, ensure we mark the device as
unavailable so that it will timeout and be re-added via GETDEVINFO.  This
avoids repeated doomed attempts to register a device in the IO path.

Add some clarifying comments as well.

Fixes: d869da9 ("nfs/blocklayout: Fix premature PR key unregistration")
Signed-off-by: Benjamin Coddington <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Reviewed-by: Chuck Lever <[email protected]>
Signed-off-by: Trond Myklebust <[email protected]>
Signed-off-by: Sasha Levin <[email protected]>

Author: bcodding-rh

fs/nfs/blocklayout/blocklayout.c

@@ -571,19 +571,32 @@ bl_find_get_deviceid(struct nfs_server *server,
 	if (!node)
 		return ERR_PTR(-ENODEV);
 
+	/*
+	 * Devices that are marked unavailable are left in the cache with a
+	 * timeout to avoid sending GETDEVINFO after every LAYOUTGET, or
+	 * constantly attempting to register the device.  Once marked as
+	 * unavailable they must be deleted and never reused.
+	 */
 	if (test_bit(NFS_DEVICEID_UNAVAILABLE, &node->flags)) {
 		unsigned long end = jiffies;
 		unsigned long start = end - PNFS_DEVICE_RETRY_TIMEOUT;
 
 		if (!time_in_range(node->timestamp_unavailable, start, end)) {
+			/* Uncork subsequent GETDEVINFO operations for this device */
 			nfs4_delete_deviceid(node->ld, node->nfs_client, id);
 			goto retry;
 		}
 		goto out_put;
 	}
 
-	if (!bl_register_dev(container_of(node, struct pnfs_block_dev, node)))
+	if (!bl_register_dev(container_of(node, struct pnfs_block_dev, node))) {
+		/*
+		 * If we cannot register, treat this device as transient:
+		 * Make a negative cache entry for the device
+		 */
+		nfs4_mark_deviceid_unavailable(node);
 		goto out_put;
+	}
 
 	return node;