struct Scsi_Host *shost = target->scsi_host;
        int i, ret;
 
-       if (target->state != SRP_TARGET_LIVE)
-               return -EAGAIN;
-
        scsi_target_block(&shost->shost_gendev);
 
        srp_disconnect_target(target);
        /*
-        * Now get a new local CM ID so that we avoid confusing the
-        * target in case things are really fouled up.
+        * Now get a new local CM ID so that we avoid confusing the target in
+        * case things are really fouled up. Doing so also ensures that all CM
+        * callbacks will have finished before a new QP is allocated.
         */
        ret = srp_new_cm_id(target);
-       if (ret)
-               goto unblock;
-
-       ret = srp_create_target_ib(target);
-       if (ret)
-               goto unblock;
+       /*
+        * Whether or not creating a new CM ID succeeded, create a new
+        * QP. This guarantees that all completion callback function
+        * invocations have finished before request resetting starts.
+        */
+       if (ret == 0)
+               ret = srp_create_target_ib(target);
+       else
+               srp_create_target_ib(target);
 
        for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
                struct srp_request *req = &target->req_ring[i];
        for (i = 0; i < SRP_SQ_SIZE; ++i)
                list_add(&target->tx_ring[i]->list, &target->free_tx);
 
-       ret = srp_connect_target(target);
+       if (ret == 0)
+               ret = srp_connect_target(target);
 
-unblock:
        scsi_target_unblock(&shost->shost_gendev, ret == 0 ? SDEV_RUNNING :
                            SDEV_TRANSPORT_OFFLINE);
 
 
        shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
 
-       if (!req || target->qp_in_error || !srp_claim_req(target, req, scmnd))
+       if (!req || !srp_claim_req(target, req, scmnd))
                return FAILED;
        srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
                          SRP_TSK_ABORT_TASK);