[PATCH] scsi_error: send device reset if asynchronous abort failed

[Hannes Reinecke <hare@kernel.org>]

The current SCSI EH implementation requires us to enter the error
handler before a device reset can be send, requiring us to wait for
all outstanding commands on the _host_ before a device reset can be
issued. This is causing a noticeable I/O stall on SAS JBODs if an
individual device fails.

But a device reset has a very well defined scope, so there really should
not be any implications to _other_ devices on the same host when the
device reset is issued.
So it should be possible to issue a device reset directly from
scsi_eh_abort_handler() once a command abort fails without having to
wait for outstanding commands on any other devices on the same host.

Signed-off-by: Hannes Reinecke <hare@kernel.org>
---
 drivers/scsi/scsi_error.c | 17 ++++++++++++++++-
 drivers/scsi/scsi_priv.h  |  1 +
 2 files changed, 17 insertions(+), 1 deletion(-)

diff --git a/drivers/scsi/scsi_error.c b/drivers/scsi/scsi_error.c
index 147127fb4db9..23dc50332f58 100644
--- a/drivers/scsi/scsi_error.c
+++ b/drivers/scsi/scsi_error.c
@@ -58,6 +58,7 @@
 #define HOST_RESET_SETTLE_TIME  (10)
 
 static int scsi_eh_try_stu(struct scsi_cmnd *scmd);
+static enum scsi_disposition scsi_try_bus_device_reset(struct scsi_cmnd *scmd);
 static enum scsi_disposition scsi_try_to_abort_cmd(const struct scsi_host_template *,
 						   struct scsi_cmnd *);
 
@@ -170,7 +171,16 @@ scmd_eh_abort_handler(struct work_struct *work)
 				    "cmd abort %s\n",
 				    (rtn == FAST_IO_FAIL) ?
 				    "not send" : "failed"));
-		goto out;
+		if (scsi_host_eh_past_deadline(shost))
+			goto out;
+		rtn = scsi_try_bus_device_reset(scmd);
+		if (rtn != SUCCESS) {
+			scmd_printk(KERN_INFO, scmd,
+				    "device reset %s\n",
+				    (rtn == FAST_IO_FAIL) ?
+				    "not send" : "failed");
+			goto out;
+		}
 	}
 	set_host_byte(scmd, DID_TIME_OUT);
 	if (scsi_host_eh_past_deadline(shost)) {
@@ -1013,6 +1023,11 @@ static enum scsi_disposition scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
 	if (!hostt->eh_device_reset_handler)
 		return FAILED;
 
+	/* Device reset had already been issued, escalate to next level */
+	if (scmd->eh_eflags & SCSI_EH_RESET_SCHEDULED)
+		return FAILED;
+
+	scmd->eh_eflags |= SCSI_EH_RESET_SCHEDULED;
 	rtn = hostt->eh_device_reset_handler(scmd);
 	if (rtn == SUCCESS)
 		__scsi_report_device_reset(scmd->device, NULL);
diff --git a/drivers/scsi/scsi_priv.h b/drivers/scsi/scsi_priv.h
index 7a193cc04e5b..be0570cea445 100644
--- a/drivers/scsi/scsi_priv.h
+++ b/drivers/scsi/scsi_priv.h
@@ -39,6 +39,7 @@ static inline u8 scsi_ml_byte(int result)
  * Scsi Error Handler Flags
  */
 #define SCSI_EH_ABORT_SCHEDULED	0x0002	/* Abort has been scheduled */
+#define SCSI_EH_RESET_SCHEDULED	0x0004  /* Reset has been scheduled */
 
 #define SCSI_SENSE_VALID(scmd) \
 	(((scmd)->sense_buffer[0] & 0x70) == 0x70)
-- 
2.51.0