@@ -1269,6 +1269,7 @@ static int
qla2xxx_eh_abort(struct scsi_cmnd *cmd)
{
scsi_qla_host_t *vha = shost_priv(cmd->device->host);
+ DECLARE_COMPLETION_ONSTACK(comp);
srb_t *sp;
int ret;
unsigned int id;
@@ -1307,6 +1308,7 @@ qla2xxx_eh_abort(struct scsi_cmnd *cmd)
return SUCCESS;
}
+ /* Get a reference to the sp and drop the lock. */
if (sp_get(sp)){
/* ref_count is already 0 */
spin_unlock_irqrestore(qpair->qp_lock_ptr, flags);
@@ -1334,6 +1336,23 @@ qla2xxx_eh_abort(struct scsi_cmnd *cmd)
sp->done(sp, DID_ABORT << 16);
ret = SUCCESS;
break;
+ case QLA_FUNCTION_PARAMETER_ERROR: {
+ /* Wait for the command completion. */
+ uint32_t ratov = ha->r_a_tov/10;
+ uint32_t ratov_j = msecs_to_jiffies(4 * ratov * 1000);
+
+ WARN_ON_ONCE(sp->comp);
+ sp->comp = ∁
+ if (!wait_for_completion_timeout(&comp, ratov_j)) {
+ ql_dbg(ql_dbg_taskm, vha, 0xffff,
+ "%s: Abort wait timer (4 * R_A_TOV[%d]) expired\n",
+ __func__, ha->r_a_tov);
+ ret = FAILED;
+ } else {
+ ret = SUCCESS;
+ }
+ break;
+ }
default:
/*
* Either abort failed or abort and completion raced. Let
@@ -1343,6 +1362,8 @@ qla2xxx_eh_abort(struct scsi_cmnd *cmd)
break;
}
+ sp->comp = NULL;
+ atomic_dec(&sp->ref_count);
ql_log(ql_log_info, vha, 0x801c,
"Abort command issued nexus=%ld:%d:%llu -- %x.\n",
vha->host_no, id, lun, ret);
I'm not sure how this happened but the patch that was intended to fix abort handling was incomplete. This patch fixes that patch as follows: - If aborting the SCSI command failed, wait until the SCSI command completes. - Return SUCCESS instead of FAILED if an abort attempt races with SCSI command completion. - Since qla2xxx_eh_abort() increments the sp reference count by calling sp_get(), decrement the sp reference count before returning. Cc: Himanshu Madhani <hmadhani@marvell.com> Fixes: 219d27d7147e ("scsi: qla2xxx: Fix race conditions in the code for aborting SCSI commands") Signed-off-by: Bart Van Assche <bvanassche@acm.org> --- drivers/scsi/qla2xxx/qla_os.c | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+)