[SCSI] zfcp: Fix deadlock between zfcp ERP and SCSI
The SCSI stack requires low level drivers to register and
unregister devices. For zfcp this leads to the situation where
zfcp calls the SCSI stack, the SCSI tries to scan the new device
and the scan SCSI command fails. This would require the zfcp erp,
but the erp thread is already blocked in the register call.
The fix is to make sure that the calls from the ERP thread to
the SCSI stack do not block the ERP thread. In detail:
1) Use a workqueue to avoid blocking of the scsi_scan_target calls.
2) When removing a unit make sure that no scsi_scan_target call is
pending.
3) Replace scsi_flush_work with scsi_target_unblock. This avoids
blocking and has the same result.
Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Signed-off-by: Swen Schillig <swen@vnet.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
diff --git a/drivers/s390/scsi/zfcp_scsi.c b/drivers/s390/scsi/zfcp_scsi.c
index 99db020..e742b3d 100644
--- a/drivers/s390/scsi/zfcp_scsi.c
+++ b/drivers/s390/scsi/zfcp_scsi.c
@@ -22,6 +22,7 @@
#define ZFCP_LOG_AREA ZFCP_LOG_AREA_SCSI
#include "zfcp_ext.h"
+#include <asm/atomic.h>
static void zfcp_scsi_slave_destroy(struct scsi_device *sdp);
static int zfcp_scsi_slave_alloc(struct scsi_device *sdp);
@@ -179,6 +180,10 @@
struct zfcp_unit *unit = (struct zfcp_unit *) sdpnt->hostdata;
if (unit) {
+ zfcp_erp_wait(unit->port->adapter);
+ wait_event(unit->scsi_scan_wq,
+ atomic_test_mask(ZFCP_STATUS_UNIT_SCSI_WORK_PENDING,
+ &unit->status) == 0);
atomic_clear_mask(ZFCP_STATUS_UNIT_REGISTERED, &unit->status);
sdpnt->hostdata = NULL;
unit->device = NULL;