[infiniband] Add infrastructure for RC queue pairs

Queue pairs are now assumed to be created in the INIT state, with a
call to ib_modify_qp() required to bring the queue pair to the RTS
state.

ib_modify_qp() no longer takes a modification list; callers should
modify the relevant queue pair parameters (e.g. qkey) directly and
then call ib_modify_qp() to synchronise the changes to the hardware.

The packet sequence number is now a property of the queue pair, rather
than of the device.

Each queue pair may have an associated address vector.  For RC queue
pairs, this is the address vector that will be programmed in to the
hardware as the remote address.  For UD queue pairs, it will be used
as the default address vector if none is supplied to ib_post_send().

src/drivers/infiniband/arbel.c

 		     ( virt_to_bus ( arbel_qp->recv.wqe ) >> 6 ) );
 	MLX_FILL_1 ( &qpctx, 43, qpc_eec_data.rcv_db_record_index,
 		     arbel_qp->recv.doorbell_idx );
-	MLX_FILL_1 ( &qpctx, 44, qpc_eec_data.q_key, qp->qkey );
 	if ( ( rc = arbel_cmd_rst2init_qpee ( arbel, qp->qpn, &qpctx )) != 0 ){
 		DBGC ( arbel, "Arbel %p RST2INIT_QPEE failed: %s\n",
 		       arbel, strerror ( rc ) );
  *
  * @v ibdev		Infiniband device
  * @v qp		Queue pair
- * @v mod_list		Modification list
  * @ret rc		Return status code
  */
 static int arbel_modify_qp ( struct ib_device *ibdev,
-			     struct ib_queue_pair *qp,
-			     unsigned long mod_list ) {
+			     struct ib_queue_pair *qp ) {
 	struct arbel *arbel = ib_get_drvdata ( ibdev );
 	struct arbelprm_qp_ee_state_transitions qpctx;
-	unsigned long optparammask = 0;
 	int rc;
 
-	/* Construct optparammask */
-	if ( mod_list & IB_MODIFY_QKEY )
-		optparammask |= ARBEL_QPEE_OPT_PARAM_QKEY;
-
 	/* Issue RTS2RTS_QP */
 	memset ( &qpctx, 0, sizeof ( qpctx ) );
-	MLX_FILL_1 ( &qpctx, 0, opt_param_mask, optparammask );
+	MLX_FILL_1 ( &qpctx, 0, opt_param_mask, ARBEL_QPEE_OPT_PARAM_QKEY );
 	MLX_FILL_1 ( &qpctx, 44, qpc_eec_data.q_key, qp->qkey );
 	if ( ( rc = arbel_cmd_rts2rts_qp ( arbel, qp->qpn, &qpctx ) ) != 0 ){
 		DBGC ( arbel, "Arbel %p RTS2RTS_QP failed: %s\n",

src/drivers/infiniband/hermon.c

 	MLX_FILL_1 ( &qpctx, 41, qpc_eec_data.cqn_rcv, qp->recv.cq->cqn );
 	MLX_FILL_1 ( &qpctx, 43, qpc_eec_data.db_record_addr_l,
 		     ( virt_to_phys ( &hermon_qp->recv.doorbell ) >> 2 ) );
-	MLX_FILL_1 ( &qpctx, 44, qpc_eec_data.q_key, qp->qkey );
 	MLX_FILL_1 ( &qpctx, 53, qpc_eec_data.mtt_base_addr_l,
 		     ( hermon_qp->mtt.mtt_base_addr >> 3 ) );
 	if ( ( rc = hermon_cmd_rst2init_qp ( hermon, qp->qpn,
  *
  * @v ibdev		Infiniband device
  * @v qp		Queue pair
- * @v mod_list		Modification list
  * @ret rc		Return status code
  */
 static int hermon_modify_qp ( struct ib_device *ibdev,
-			      struct ib_queue_pair *qp,
-			      unsigned long mod_list ) {
+			      struct ib_queue_pair *qp ) {
 	struct hermon *hermon = ib_get_drvdata ( ibdev );
 	struct hermonprm_qp_ee_state_transitions qpctx;
-	unsigned long optparammask = 0;
 	int rc;
 
-	/* Construct optparammask */
-	if ( mod_list & IB_MODIFY_QKEY )
-		optparammask |= HERMON_QP_OPT_PARAM_QKEY;
-
 	/* Issue RTS2RTS_QP */
 	memset ( &qpctx, 0, sizeof ( qpctx ) );
-	MLX_FILL_1 ( &qpctx, 0, opt_param_mask, optparammask );
+	MLX_FILL_1 ( &qpctx, 0, opt_param_mask, HERMON_QP_OPT_PARAM_QKEY );
 	MLX_FILL_1 ( &qpctx, 44, qpc_eec_data.q_key, qp->qkey );
 	if ( ( rc = hermon_cmd_rts2rts_qp ( hermon, qp->qpn, &qpctx ) ) != 0 ){
 		DBGC ( hermon, "Hermon %p RTS2RTS_QP failed: %s\n",

src/drivers/infiniband/linda.c

 	return 0;
 }
 
+/**
+ * Set partition key table
+ *
+ * @v ibdev		Infiniband device
+ * @v mad		Set partition key table MAD
+ */
+static int linda_set_pkey_table ( struct ib_device *ibdev __unused,
+				  union ib_mad *mad __unused ) {
+	/* Nothing to do */
+	return 0;
+}
+
 /***************************************************************************
  *
  * Context allocation
  *
  * @v ibdev		Infiniband device
  * @v qp		Queue pair
- * @v mod_list		Modification list
  * @ret rc		Return status code
  */
 static int linda_modify_qp ( struct ib_device *ibdev,
-			     struct ib_queue_pair *qp,
-			     unsigned long mod_list __unused ) {
+			     struct ib_queue_pair *qp ) {
 	struct linda *linda = ib_get_drvdata ( ibdev );
 
 	/* Nothing to do; the hardware doesn't have a notion of queue
 	.mcast_attach	= linda_mcast_attach,
 	.mcast_detach	= linda_mcast_detach,
 	.set_port_info	= linda_set_port_info,
+	.set_pkey_table	= linda_set_pkey_table,
 };
 
 /***************************************************************************

src/drivers/net/ipoib.c

 	/* Allocate queue pair */
 	ipoib->qp = ib_create_qp ( ibdev, IB_QPT_UD,
 				   IPOIB_NUM_SEND_WQES, ipoib->cq,
-				   IPOIB_NUM_RECV_WQES, ipoib->cq, 0 );
+				   IPOIB_NUM_RECV_WQES, ipoib->cq );
 	if ( ! ipoib->qp ) {
 		DBGC ( ipoib, "IPoIB %p could not allocate queue pair\n",
 		       ipoib );

src/include/gpxe/infiniband.h

 struct ib_completion_queue;
 struct ib_gma;
 
+/** Infiniband transmission rates */
+enum ib_rate {
+	IB_RATE_2_5 = 2,
+	IB_RATE_10 = 3,
+	IB_RATE_30 = 4,
+	IB_RATE_5 = 5,
+	IB_RATE_20 = 6,
+	IB_RATE_40 = 7,
+	IB_RATE_60 = 8,
+	IB_RATE_80 = 9,
+	IB_RATE_120 = 10,
+};
+
+/** An Infiniband Address Vector */
+struct ib_address_vector {
+	/** Queue Pair Number */
+	unsigned long qpn;
+	/** Queue key
+	 *
+	 * Not specified for received packets.
+	 */
+	unsigned long qkey;
+	/** Local ID */
+	unsigned int lid;
+	/** Rate
+	 *
+	 * Not specified for received packets.
+	 */
+	enum ib_rate rate;
+	/** Service level */
+	unsigned int sl;
+	/** GID is present */
+	unsigned int gid_present;
+	/** GID, if present */
+	struct ib_gid gid;
+};
+
 /** An Infiniband Work Queue */
 struct ib_work_queue {
 	/** Containing queue pair */
 	struct ib_completion_queue *cq;
 	/** List of work queues on this completion queue */
 	struct list_head list;
+	/** Packet sequence number */
+	uint32_t psn;
 	/** Number of work queue entries */
 	unsigned int num_wqes;
 	/** Number of occupied work queue entries */
 	IB_QPT_SMA,
 	IB_QPT_GMA,
 	IB_QPT_UD,
+	IB_QPT_RC,
 };
 
 /** An Infiniband Queue Pair */
 	struct ib_work_queue recv;
 	/** List of multicast GIDs */
 	struct list_head mgids;
+	/** Address vector */
+	struct ib_address_vector av;
 	/** Driver private data */
 	void *drv_priv;
 	/** Queue owner private data */
 	void *owner_priv;
 };
 
-/** Infiniband queue pair modification flags */
-enum ib_queue_pair_mods {
-	IB_MODIFY_QKEY = 0x0001,
-};
-
-/** An Infiniband Address Vector */
-struct ib_address_vector {
-	/** Queue Pair Number */
-	unsigned long qpn;
-	/** Queue key
-	 *
-	 * Not specified for received packets.
-	 */
-	unsigned long qkey;
-	/** Local ID */
-	unsigned int lid;
-	/** Rate
-	 *
-	 * Not specified for received packets.
-	 */
-	unsigned int rate;
-	/** Service level */
-	unsigned int sl;
-	/** GID is present */
-	unsigned int gid_present;
-	/** GID, if present */
-	struct ib_gid gid;
-};
-
-/** Infiniband transmission rates */
-enum ib_rate {
-	IB_RATE_2_5 = 2,
-	IB_RATE_10 = 3,
-	IB_RATE_30 = 4,
-	IB_RATE_5 = 5,
-	IB_RATE_20 = 6,
-	IB_RATE_40 = 7,
-	IB_RATE_60 = 8,
-	IB_RATE_80 = 9,
-	IB_RATE_120 = 10,
-};
-
 /** Infiniband completion queue operations */
 struct ib_completion_queue_operations {
 	/**
 	 *
 	 * @v ibdev		Infiniband device
 	 * @v qp		Queue pair
-	 * @v mod_list		Modification list
 	 * @ret rc		Return status code
 	 */
 	int ( * modify_qp ) ( struct ib_device *ibdev,
-			      struct ib_queue_pair *qp,
-			      unsigned long mod_list );
+			      struct ib_queue_pair *qp );
 	/** Destroy queue pair
 	 *
 	 * @v ibdev		Infiniband device
 	 * an embedded SMA.
 	 */
 	int ( * set_port_info ) ( struct ib_device *ibdev, union ib_mad *mad );
+	/** Set partition key table
+	 *
+	 * @v ibdev		Infiniband device
+	 * @v mad		Set partition key table MAD
+	 *
+	 * This method is required only by adapters that do not have
+	 * an embedded SMA.
+	 */
+	int ( * set_pkey_table ) ( struct ib_device *ibdev,
+				   union ib_mad *mad );
 };
 
 /** An Infiniband device */
 	/** Partition key */
 	uint16_t pkey;
 
-	/** Outbound packet sequence number */
-	uint32_t psn;
+	/** RDMA key
+	 *
+	 * This is a single key allowing unrestricted access to
+	 * memory.
+	 */
+	uint32_t rdma_key;
 
 	/** Subnet management agent */
 	struct ib_gma *sma;
 extern struct ib_queue_pair *
 ib_create_qp ( struct ib_device *ibdev, enum ib_queue_pair_type type,
 	       unsigned int num_send_wqes, struct ib_completion_queue *send_cq,
-	       unsigned int num_recv_wqes, struct ib_completion_queue *recv_cq,
-	       unsigned long qkey );
-extern int ib_modify_qp ( struct ib_device *ibdev, struct ib_queue_pair *qp,
-			  unsigned long mod_list, unsigned long qkey );
+	       unsigned int num_recv_wqes,
+	       struct ib_completion_queue *recv_cq );
+extern int ib_modify_qp ( struct ib_device *ibdev, struct ib_queue_pair *qp );
 extern void ib_destroy_qp ( struct ib_device *ibdev,
 			    struct ib_queue_pair *qp );
 extern struct ib_queue_pair * ib_find_qp_qpn ( struct ib_device *ibdev,
 extern int ib_get_hca_info ( struct ib_device *ibdev,
 			     struct ib_gid_half *hca_guid );
 extern int ib_set_port_info ( struct ib_device *ibdev, union ib_mad *mad );
+extern int ib_set_pkey_table ( struct ib_device *ibdev, union ib_mad *mad );
 extern struct ib_device * alloc_ibdev ( size_t priv_size );
 extern int register_ibdev ( struct ib_device *ibdev );
 extern void unregister_ibdev ( struct ib_device *ibdev );

src/net/infiniband.c

  * @v send_cq		Send completion queue
  * @v num_recv_wqes	Number of receive work queue entries
  * @v recv_cq		Receive completion queue
- * @v qkey		Queue key
  * @ret qp		Queue pair
+ *
+ * The queue pair will be left in the INIT state; you must call
+ * ib_modify_qp() before it is ready to use for sending and receiving.
  */
 struct ib_queue_pair * ib_create_qp ( struct ib_device *ibdev,
 				      enum ib_queue_pair_type type,
 				      unsigned int num_send_wqes,
 				      struct ib_completion_queue *send_cq,
 				      unsigned int num_recv_wqes,
-				      struct ib_completion_queue *recv_cq,
-				      unsigned long qkey ) {
+				      struct ib_completion_queue *recv_cq ) {
 	struct ib_queue_pair *qp;
 	size_t total_size;
 	int rc;
 	qp->ibdev = ibdev;
 	list_add ( &qp->list, &ibdev->qps );
 	qp->type = type;
-	qp->qkey = qkey;
 	qp->send.qp = qp;
 	qp->send.is_send = 1;
 	qp->send.cq = send_cq;
 	list_add ( &qp->send.list, &send_cq->work_queues );
+	qp->send.psn = ( random() & 0xffffffUL );
 	qp->send.num_wqes = num_send_wqes;
 	qp->send.iobufs = ( ( ( void * ) qp ) + sizeof ( *qp ) );
 	qp->recv.qp = qp;
 	qp->recv.cq = recv_cq;
 	list_add ( &qp->recv.list, &recv_cq->work_queues );
+	qp->recv.psn = ( random() & 0xffffffUL );
 	qp->recv.num_wqes = num_recv_wqes;
 	qp->recv.iobufs = ( ( ( void * ) qp ) + sizeof ( *qp ) +
 			    ( num_send_wqes * sizeof ( qp->send.iobufs[0] ) ));
  *
  * @v ibdev		Infiniband device
  * @v qp		Queue pair
- * @v mod_list		Modification list
- * @v qkey		New queue key, if applicable
+ * @v av		New address vector, if applicable
  * @ret rc		Return status code
  */
-int ib_modify_qp ( struct ib_device *ibdev, struct ib_queue_pair *qp,
-		   unsigned long mod_list, unsigned long qkey ) {
+int ib_modify_qp ( struct ib_device *ibdev, struct ib_queue_pair *qp ) {
 	int rc;
 
 	DBGC ( ibdev, "IBDEV %p modifying QPN %#lx\n", ibdev, qp->qpn );
 
-	if ( mod_list & IB_MODIFY_QKEY )
-		qp->qkey = qkey;
-
-	if ( ( rc = ibdev->op->modify_qp ( ibdev, qp, mod_list ) ) != 0 ) {
+	if ( ( rc = ibdev->op->modify_qp ( ibdev, qp ) ) != 0 ) {
 		DBGC ( ibdev, "IBDEV %p could not modify QPN %#lx: %s\n",
 		       ibdev, qp->qpn, strerror ( rc ) );
 		return rc;
 int ib_post_send ( struct ib_device *ibdev, struct ib_queue_pair *qp,
 		   struct ib_address_vector *av,
 		   struct io_buffer *iobuf ) {
+	struct ib_address_vector av_copy;
 	int rc;
 
 	/* Check queue fill level */
 		return -ENOBUFS;
 	}
 
+	/* Use default address vector if none specified */
+	if ( ! av )
+		av = &qp->av;
+
+	/* Make modifiable copy of address vector */
+	memcpy ( &av_copy, av, sizeof ( av_copy ) );
+	av = &av_copy;
+
 	/* Fill in optional parameters in address vector */
 	if ( ! av->qkey )
 		av->qkey = qp->qkey;
 	return 0;
 };
 
+/**
+ * Set partition key table
+ *
+ * @v ibdev		Infiniband device
+ * @v mad		Set partition key table MAD
+ */
+int ib_set_pkey_table ( struct ib_device *ibdev, union ib_mad *mad ) {
+	int rc;
+
+	/* Adapters with embedded SMAs do not need to support this method */
+	if ( ! ibdev->op->set_pkey_table ) {
+		DBGC ( ibdev, "IBDEV %p does not support setting partition "
+		       "key table\n", ibdev );
+		return -ENOTSUP;
+	}
+
+	if ( ( rc = ibdev->op->set_pkey_table ( ibdev, mad ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not set partition key table: "
+		       "%s\n", ibdev, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+};
+
 /***************************************************************************
  *
  * Event queues

src/net/infiniband/ib_gma.c

 					      union ib_mad *mad ) {
 	struct ib_device *ibdev = gma->ibdev;
 	struct ib_pkey_table *pkey_table = &mad->smp.smp_data.pkey_table;
+	int rc;
 
 	ibdev->pkey = ntohs ( pkey_table->pkey[0] );
+	DBGC ( gma, "GMA %p set pkey %04x\n", gma, ibdev->pkey );
+
+	if ( ( rc = ib_set_pkey_table ( ibdev, mad ) ) != 0 ) {
+		DBGC ( gma, "GMA %p could not set pkey table: %s\n",
+		       gma, strerror ( rc ) );
+		mad->hdr.status =
+			htons ( IB_MGMT_STATUS_UNSUPPORTED_METHOD_ATTR );
+	}
 
 	return ib_sma_get_pkey_table ( gma, mad );
 }
 struct ib_gma * ib_create_gma ( struct ib_device *ibdev,
 				enum ib_queue_pair_type type ) {
 	struct ib_gma *gma;
-	unsigned long qkey;
+	int rc;
 
 	/* Allocate and initialise fields */
 	gma = zalloc ( sizeof ( *gma ) );
 	}
 
 	/* Create queue pair */
-	qkey = ( ( type == IB_QPT_SMA ) ? IB_QKEY_SMA : IB_QKEY_GMA );
 	gma->qp = ib_create_qp ( ibdev, type, IB_GMA_NUM_SEND_WQES, gma->cq,
-				 IB_GMA_NUM_RECV_WQES, gma->cq, qkey );
+				 IB_GMA_NUM_RECV_WQES, gma->cq );
 	if ( ! gma->qp ) {
 		DBGC ( gma, "GMA %p could not allocate queue pair\n", gma );
 		goto err_create_qp;
 	}
 	ib_qp_set_ownerdata ( gma->qp, gma );
-
 	DBGC ( gma, "GMA %p running on QPN %#lx\n", gma, gma->qp->qpn );
 
+	/* Set queue key */
+	gma->qp->qkey = ( ( type == IB_QPT_SMA ) ? IB_QKEY_SMA : IB_QKEY_GMA );
+	if ( ( rc = ib_modify_qp ( ibdev, gma->qp ) ) != 0 ) {
+		DBGC ( gma, "GMA %p could not set queue key: %s\n",
+		       gma, strerror ( rc ) );
+		goto err_modify_qp;
+	}
+
 	/* Fill receive ring */
 	ib_refill_recv ( ibdev, gma->qp );
 	return gma;
 
+ err_modify_qp:
 	ib_destroy_qp ( ibdev, gma->qp );
  err_create_qp:
 	ib_destroy_cq ( ibdev, gma->cq );

src/net/infiniband/ib_mcast.c

 	       ntohl ( gid->u.dwords[3] ), qkey );
 
 	/* Set queue key */
-	if ( ( rc = ib_modify_qp ( ibdev, qp, IB_MODIFY_QKEY, qkey ) ) != 0 ) {
+	qp->qkey = qkey;
+	if ( ( rc = ib_modify_qp ( ibdev, qp ) ) != 0 ) {
 		DBGC ( gma, "GMA %p QPN %lx could not modify qkey: %s\n",
 		       gma, qp->qpn, strerror ( rc ) );
 		return NULL;

src/net/infiniband/ib_packet.c

 	bth->se__m__padcnt__tver = ( pad_len << 4 );
 	bth->pkey = htons ( ibdev->pkey );
 	bth->dest_qp = htonl ( av->qpn );
-	bth->ack__psn = htonl ( ( ibdev->psn++ ) & 0xffffffUL );
+	bth->ack__psn = htonl ( ( qp->send.psn++ ) & 0xffffffUL );
 
 	/* Construct DETH */
 	deth->qkey = htonl ( av->qkey );