IB/srpt: Fix how aborted commands are processed
authorBart Van Assche <bart.vanassche@sandisk.com>
Thu, 11 Feb 2016 19:07:11 +0000 (11:07 -0800)
committerDoug Ledford <dledford@redhat.com>
Mon, 29 Feb 2016 22:12:35 +0000 (17:12 -0500)
srpt_abort_cmd() must not be called in state SRPT_STATE_DATA_IN. Issue
a warning if this occurs.

srpt_abort_cmd() must not invoke target_put_sess_cmd() for commands
in state SRPT_STATE_DONE because the srpt_abort_cmd() callers already
do this when necessary. Hence remove this call.

If an RDMA read fails the corresponding SCSI command must fail. Hence
add a transport_generic_request_failure() call.

Remove an incorrect srpt_abort_cmd() call from srpt_rdma_write_done().

Avoid that srpt_send_done() calls srpt_abort_cmd() for finished SCSI
commands.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Estrin <alex.estrin@intel.com>
Cc: Sagi Grimberg <sagig@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/ulp/srpt/ib_srpt.c

index 5b0fbc1bfce26023638eea6fdd1c3ce5a872f60d..64e9107f6f4f43a2f51446b50ae176b608e63b94 100644 (file)
@@ -1267,10 +1267,7 @@ static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
 
        /*
         * If the command is in a state where the target core is waiting for
-        * the ib_srpt driver, change the state to the next state. Changing
-        * the state of the command from SRPT_STATE_NEED_DATA to
-        * SRPT_STATE_DATA_IN ensures that srpt_xmit_response() will call this
-        * function a second time.
+        * the ib_srpt driver, change the state to the next state.
         */
 
        spin_lock_irqsave(&ioctx->spinlock, flags);
@@ -1279,25 +1276,17 @@ static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
        case SRPT_STATE_NEED_DATA:
                ioctx->state = SRPT_STATE_DATA_IN;
                break;
-       case SRPT_STATE_DATA_IN:
        case SRPT_STATE_CMD_RSP_SENT:
        case SRPT_STATE_MGMT_RSP_SENT:
                ioctx->state = SRPT_STATE_DONE;
                break;
        default:
+               WARN_ONCE(true, "%s: unexpected I/O context state %d\n",
+                         __func__, state);
                break;
        }
        spin_unlock_irqrestore(&ioctx->spinlock, flags);
 
-       if (state == SRPT_STATE_DONE) {
-               struct srpt_rdma_ch *ch = ioctx->ch;
-
-               BUG_ON(ch->sess == NULL);
-
-               target_put_sess_cmd(&ioctx->cmd);
-               goto out;
-       }
-
        pr_debug("Aborting cmd with state %d and tag %lld\n", state,
                 ioctx->cmd.tag);
 
@@ -1305,19 +1294,16 @@ static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
        case SRPT_STATE_NEW:
        case SRPT_STATE_DATA_IN:
        case SRPT_STATE_MGMT:
+       case SRPT_STATE_DONE:
                /*
                 * Do nothing - defer abort processing until
                 * srpt_queue_response() is invoked.
                 */
-               WARN_ON(!transport_check_aborted_status(&ioctx->cmd, false));
                break;
        case SRPT_STATE_NEED_DATA:
-               /* DMA_TO_DEVICE (write) - RDMA read error. */
-
-               /* XXX(hch): this is a horrible layering violation.. */
-               spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
-               ioctx->cmd.transport_state &= ~CMD_T_ACTIVE;
-               spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
+               pr_debug("tag %#llx: RDMA read error\n", ioctx->cmd.tag);
+               transport_generic_request_failure(&ioctx->cmd,
+                                       TCM_CHECK_CONDITION_ABORT_CMD);
                break;
        case SRPT_STATE_CMD_RSP_SENT:
                /*
@@ -1325,18 +1311,16 @@ static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
                 * not been received in time.
                 */
                srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
-               target_put_sess_cmd(&ioctx->cmd);
+               transport_generic_free_cmd(&ioctx->cmd, 0);
                break;
        case SRPT_STATE_MGMT_RSP_SENT:
-               srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
-               target_put_sess_cmd(&ioctx->cmd);
+               transport_generic_free_cmd(&ioctx->cmd, 0);
                break;
        default:
                WARN(1, "Unexpected command state (%d)", state);
                break;
        }
 
-out:
        return state;
 }
 
@@ -1376,9 +1360,14 @@ static void srpt_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc)
                container_of(wc->wr_cqe, struct srpt_send_ioctx, rdma_cqe);
 
        if (unlikely(wc->status != IB_WC_SUCCESS)) {
+               /*
+                * Note: if an RDMA write error completion is received that
+                * means that a SEND also has been posted. Defer further
+                * processing of the associated command until the send error
+                * completion has been received.
+                */
                pr_info("RDMA_WRITE for ioctx 0x%p failed with status %d\n",
                        ioctx, wc->status);
-               srpt_abort_cmd(ioctx);
        }
 }
 
@@ -1721,15 +1710,10 @@ static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc)
 
        atomic_inc(&ch->sq_wr_avail);
 
-       if (wc->status != IB_WC_SUCCESS) {
+       if (wc->status != IB_WC_SUCCESS)
                pr_info("sending response for ioctx 0x%p failed"
                        " with status %d\n", ioctx, wc->status);
 
-               atomic_dec(&ch->req_lim);
-               srpt_abort_cmd(ioctx);
-               goto out;
-       }
-
        if (state != SRPT_STATE_DONE) {
                srpt_unmap_sg_to_ib_sge(ch, ioctx);
                transport_generic_free_cmd(&ioctx->cmd, 0);
@@ -1738,7 +1722,6 @@ static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc)
                       " wr_id = %u.\n", ioctx->ioctx.index);
        }
 
-out:
        while (!list_empty(&ch->cmd_wait_list) &&
               ch->state == CH_LIVE &&
               (ioctx = srpt_get_send_ioctx(ch)) != NULL) {