[Infiniband] Add preliminary multiple port support for Hermon cards

Infiniband devices no longer block waiting for link-up in
register_ibdev().

Hermon driver needs to create an event queue and poll for link-up events.

Infiniband core needs to reread MAD parameters when link state changes.

IPoIB needs to cope with Infiniband link parameters being only partially
available at probe and open time.

src/drivers/infiniband/hermon.c

@@ -30,6 +30,7 @@
 #include <gpxe/umalloc.h>
 #include <gpxe/iobuf.h>
 #include <gpxe/netdevice.h>
+#include <gpxe/process.h>
 #include <gpxe/infiniband.h>
 #include "hermon.h"
 
@@ -316,20 +317,31 @@ hermon_cmd_write_mtt ( struct hermon *hermon,
 			    0, write_mtt, 1, NULL );
 }
 
+static inline int
+hermon_cmd_map_eq ( struct hermon *hermon, unsigned long index_map,
+		    const struct hermonprm_event_mask *mask ) {
+	return hermon_cmd ( hermon,
+			    HERMON_HCR_IN_CMD ( HERMON_HCR_MAP_EQ,
+						0, sizeof ( *mask ) ),
+			    0, mask, index_map, NULL );
+}
+
 static inline int
 hermon_cmd_sw2hw_eq ( struct hermon *hermon, unsigned int index,
-		      const struct hermonprm_eqc *eqc ) {
+		      const struct hermonprm_eqc *eqctx ) {
 	return hermon_cmd ( hermon,
 			    HERMON_HCR_IN_CMD ( HERMON_HCR_SW2HW_EQ,
-						1, sizeof ( *eqc ) ),
-			    0, eqc, index, NULL );
+						1, sizeof ( *eqctx ) ),
+			    0, eqctx, index, NULL );
 }
 
 static inline int
-hermon_cmd_hw2sw_eq ( struct hermon *hermon, unsigned int index ) {
+hermon_cmd_hw2sw_eq ( struct hermon *hermon, unsigned int index,
+		      struct hermonprm_eqc *eqctx ) {
 	return hermon_cmd ( hermon,
-			    HERMON_HCR_VOID_CMD ( HERMON_HCR_HW2SW_EQ ),
-			    1, NULL, index, NULL );
+			    HERMON_HCR_OUT_CMD ( HERMON_HCR_HW2SW_EQ,
+						 1, sizeof ( *eqctx ) ),
+			    1, NULL, index, eqctx );
 }
 
 static inline int
@@ -377,6 +389,15 @@ hermon_cmd_rtr2rts_qp ( struct hermon *hermon, unsigned long qpn,
 			    0, ctx, qpn, NULL );
 }
 
+static inline int
+hermon_cmd_rts2rts_qp ( struct hermon *hermon, unsigned long qpn,
+			const struct hermonprm_qp_ee_state_transitions *ctx ) {
+	return hermon_cmd ( hermon,
+			    HERMON_HCR_IN_CMD ( HERMON_HCR_RTS2RTS_QP,
+						1, sizeof ( *ctx ) ),
+			    0, ctx, qpn, NULL );
+}
+
 static inline int
 hermon_cmd_2rst_qp ( struct hermon *hermon, unsigned long qpn ) {
 	return hermon_cmd ( hermon,
@@ -859,6 +880,39 @@ static int hermon_create_qp ( struct ib_device *ibdev,
 	return rc;
 }
 
+/**
+ * Modify queue pair
+ *
+ * @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 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, 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",
+		       hermon, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+}
+
 /**
  * Destroy queue pair
  *
@@ -1356,6 +1410,7 @@ static struct ib_device_operations hermon_ib_operations = {
 	.create_cq	= hermon_create_cq,
 	.destroy_cq	= hermon_destroy_cq,
 	.create_qp	= hermon_create_qp,
+	.modify_qp	= hermon_modify_qp,
 	.destroy_qp	= hermon_destroy_qp,
 	.post_send	= hermon_post_send,
 	.post_recv	= hermon_post_recv,
@@ -1367,6 +1422,211 @@ static struct ib_device_operations hermon_ib_operations = {
 	.mad		= hermon_mad,
 };
 
+/***************************************************************************
+ *
+ * Event queues
+ *
+ ***************************************************************************
+ */
+
+/**
+ * Create event queue
+ *
+ * @v hermon		Hermon device
+ * @ret rc		Return status code
+ */
+static int hermon_create_eq ( struct hermon *hermon ) {
+	struct hermon_event_queue *hermon_eq = &hermon->eq;
+	struct hermonprm_eqc eqctx;
+	struct hermonprm_event_mask mask;
+	unsigned int i;
+	int rc;
+
+	/* Allocate event queue itself */
+	hermon_eq->eqe_size =
+		( HERMON_NUM_EQES * sizeof ( hermon_eq->eqe[0] ) );
+	hermon_eq->eqe = malloc_dma ( hermon_eq->eqe_size,
+				      sizeof ( hermon_eq->eqe[0] ) );
+	if ( ! hermon_eq->eqe ) {
+		rc = -ENOMEM;
+		goto err_eqe;
+	}
+	memset ( hermon_eq->eqe, 0, hermon_eq->eqe_size );
+	for ( i = 0 ; i < HERMON_NUM_EQES ; i++ ) {
+		MLX_FILL_1 ( &hermon_eq->eqe[i].generic, 7, owner, 1 );
+	}
+	barrier();
+
+	/* Allocate MTT entries */
+	if ( ( rc = hermon_alloc_mtt ( hermon, hermon_eq->eqe,
+				       hermon_eq->eqe_size,
+				       &hermon_eq->mtt ) ) != 0 )
+		goto err_alloc_mtt;
+
+	/* Hand queue over to hardware */
+	memset ( &eqctx, 0, sizeof ( eqctx ) );
+	MLX_FILL_1 ( &eqctx, 0, st, 0xa /* "Fired" */ );
+	MLX_FILL_1 ( &eqctx, 2,
+		     page_offset, ( hermon_eq->mtt.page_offset >> 5 ) );
+	MLX_FILL_1 ( &eqctx, 3, log_eq_size, fls ( HERMON_NUM_EQES - 1 ) );
+	MLX_FILL_1 ( &eqctx, 7, mtt_base_addr_l,
+		     ( hermon_eq->mtt.mtt_base_addr >> 3 ) );
+	if ( ( rc = hermon_cmd_sw2hw_eq ( hermon, 0, &eqctx ) ) != 0 ) {
+		DBGC ( hermon, "Hermon %p SW2HW_EQ failed: %s\n",
+		       hermon, strerror ( rc ) );
+		goto err_sw2hw_eq;
+	}
+
+	/* Map events to this event queue */
+	memset ( &mask, 0, sizeof ( mask ) );
+	MLX_FILL_1 ( &mask, 1, port_state_change, 1 );
+	if ( ( rc = hermon_cmd_map_eq ( hermon, ( HERMON_MAP_EQ_MAP | 0 ),
+					&mask ) ) != 0 ) {
+		DBGC ( hermon, "Hermon %p MAP_EQ failed: %s\n",
+		       hermon, strerror ( rc )  );
+		goto err_map_eq;
+	}
+
+	return 0;
+
+ err_map_eq:
+	hermon_cmd_hw2sw_eq ( hermon, 0, &eqctx );
+ err_sw2hw_eq:
+	hermon_free_mtt ( hermon, &hermon_eq->mtt );
+ err_alloc_mtt:
+	free_dma ( hermon_eq->eqe, hermon_eq->eqe_size );
+ err_eqe:
+	memset ( hermon_eq, 0, sizeof ( *hermon_eq ) );
+	return rc;
+}
+
+/**
+ * Destroy event queue
+ *
+ * @v hermon		Hermon device
+ */
+static void hermon_destroy_eq ( struct hermon *hermon ) {
+	struct hermon_event_queue *hermon_eq = &hermon->eq;
+	struct hermonprm_eqc eqctx;
+	struct hermonprm_event_mask mask;
+	int rc;
+
+	/* Unmap events from event queue */
+	memset ( &mask, 0, sizeof ( mask ) );
+	MLX_FILL_1 ( &mask, 1, port_state_change, 1 );
+	if ( ( rc = hermon_cmd_map_eq ( hermon, ( HERMON_MAP_EQ_UNMAP | 0 ),
+					&mask ) ) != 0 ) {
+		DBGC ( hermon, "Hermon %p FATAL MAP_EQ failed to unmap: %s\n",
+		       hermon, strerror ( rc ) );
+		/* Continue; HCA may die but system should survive */
+	}
+
+	/* Take ownership back from hardware */
+	if ( ( rc = hermon_cmd_hw2sw_eq ( hermon, 0, &eqctx ) ) != 0 ) {
+		DBGC ( hermon, "Hermon %p FATAL HW2SW_EQ failed: %s\n",
+		       hermon, strerror ( rc ) );
+		/* Leak memory and return; at least we avoid corruption */
+		return;
+	}
+
+	/* Free MTT entries */
+	hermon_free_mtt ( hermon, &hermon_eq->mtt );
+
+	/* Free memory */
+	free_dma ( hermon_eq->eqe, hermon_eq->eqe_size );
+	memset ( hermon_eq, 0, sizeof ( *hermon_eq ) );
+}
+
+/**
+ * Handle port state event
+ *
+ * @v hermon		Hermon device
+ * @v eqe		Port state change event queue entry
+ */
+static void hermon_event_port_state_change ( struct hermon *hermon,
+					     union hermonprm_event_entry *eqe){
+	unsigned int port;
+	int link_up;
+
+	/* Get port and link status */
+	port = ( MLX_GET ( &eqe->port_state_change, data.p ) - 1 );
+	link_up = ( MLX_GET ( &eqe->generic, event_sub_type ) & 0x04 );
+	DBGC ( hermon, "Hermon %p port %d link %s\n", hermon, ( port + 1 ),
+	       ( link_up ? "up" : "down" ) );
+
+	/* Sanity check */
+	if ( port >= HERMON_NUM_PORTS ) {
+		DBGC ( hermon, "Hermon %p port %d does not exist!\n",
+		       hermon, ( port + 1 ) );
+		return;
+	}
+
+	/* Notify Infiniband core of link state change */
+	ib_link_state_changed ( hermon->ibdev[port] );
+}
+
+/**
+ * Poll event queue
+ *
+ * @v hermon		Hermon device
+ */
+static void hermon_poll_eq ( struct hermon *hermon ) {
+	struct hermon_event_queue *hermon_eq = &hermon->eq;
+	union hermonprm_event_entry *eqe;
+	union hermonprm_doorbell_register db_reg;
+	unsigned int eqe_idx_mask;
+	unsigned int event_type;
+
+	while ( 1 ) {
+		eqe_idx_mask = ( HERMON_NUM_EQES - 1 );
+		eqe = &hermon_eq->eqe[hermon_eq->next_idx & eqe_idx_mask];
+		if ( MLX_GET ( &eqe->generic, owner ) ^
+		     ( ( hermon_eq->next_idx & HERMON_NUM_EQES ) ? 1 : 0 ) ) {
+			/* Entry still owned by hardware; end of poll */
+			break;
+		}
+		DBGCP ( hermon, "Hermon %p event:\n", hermon );
+		DBGCP_HD ( hermon, eqe, sizeof ( *eqe ) );
+
+		/* Handle event */
+		event_type = MLX_GET ( &eqe->generic, event_type );
+		switch ( event_type ) {
+		case HERMON_EV_PORT_STATE_CHANGE:
+			hermon_event_port_state_change ( hermon, eqe );
+			break;
+		default:
+			DBGC ( hermon, "Hermon %p unrecognised event type "
+			       "%#x:\n", hermon, event_type );
+			DBGC_HD ( hermon, eqe, sizeof ( *eqe ) );
+			break;
+		}
+
+		/* Update event queue's index */
+		hermon_eq->next_idx++;
+
+		/* Ring doorbell */
+		memset ( &db_reg, 0, sizeof ( db_reg ) );
+		MLX_FILL_1 ( &db_reg.event, 0, ci, hermon_eq->next_idx );
+		DBGCP ( hermon, "Ringing doorbell %08lx with %08lx\n",
+			virt_to_phys ( hermon->uar + HERMON_DB_EQ0_OFFSET ),
+			db_reg.dword[0] );
+		writel ( db_reg.dword[0],
+			 ( hermon->uar + HERMON_DB_EQ0_OFFSET ) );
+	}
+}
+
+/**
+ * Event queue poll processor
+ *
+ * @v process		Hermon event queue process
+ */
+static void hermon_step ( struct process *process ) {
+	struct hermon *hermon =
+		container_of ( process, struct hermon, event_process );
+
+	hermon_poll_eq ( hermon );
+}
+
 /***************************************************************************
  *
  * Firmware control
@@ -1879,6 +2139,7 @@ static int hermon_probe ( struct pci_device *pci,
 		goto err_alloc_hermon;
 	}
 	pci_set_drvdata ( pci, hermon );
+	process_init ( &hermon->event_process, hermon_step, NULL );
 
 	/* Allocate Infiniband devices */
 	for ( i = 0 ; i < HERMON_NUM_PORTS ; i++ ) {
@@ -1945,6 +2206,10 @@ static int hermon_probe ( struct pci_device *pci,
 	if ( ( rc = hermon_setup_mpt ( hermon ) ) != 0 )
 		goto err_setup_mpt;
 
+	/* Set up event queue */
+	if ( ( rc = hermon_create_eq ( hermon ) ) != 0 )
+		goto err_create_eq;
+
 	/* Register Infiniband devices */
 	for ( i = 0 ; i < HERMON_NUM_PORTS ; i++ ) {
 		if ( ( rc = register_ibdev ( hermon->ibdev[i] ) ) != 0 ) {
@@ -1960,6 +2225,8 @@ static int hermon_probe ( struct pci_device *pci,
  err_register_ibdev:
 	for ( ; i >= 0 ; i-- )
 		unregister_ibdev ( hermon->ibdev[i] );
+	hermon_destroy_eq ( hermon );
+ err_create_eq:
  err_setup_mpt:
 	hermon_cmd_close_hca ( hermon );
  err_init_hca:
@@ -1976,6 +2243,7 @@ static int hermon_probe ( struct pci_device *pci,
  err_alloc_ibdev:
 	for ( ; i >= 0 ; i-- )
 		free_ibdev ( hermon->ibdev[i] );
+	process_del ( &hermon->event_process );
 	free ( hermon );
  err_alloc_hermon:
 	return rc;
@@ -1992,6 +2260,7 @@ static void hermon_remove ( struct pci_device *pci ) {
 
 	for ( i = ( HERMON_NUM_PORTS - 1 ) ; i >= 0 ; i-- )
 		unregister_ibdev ( hermon->ibdev[i] );
+	hermon_destroy_eq ( hermon );
 	hermon_cmd_close_hca ( hermon );
 	hermon_free_icm ( hermon );
 	hermon_stop_firmware ( hermon );
@@ -2000,6 +2269,7 @@ static void hermon_remove ( struct pci_device *pci ) {
 	free_dma ( hermon->mailbox_in, HERMON_MBOX_SIZE );
 	for ( i = ( HERMON_NUM_PORTS - 1 ) ; i >= 0 ; i-- )
 		free_ibdev ( hermon->ibdev[i] );
+	process_del ( &hermon->event_process );
 	free ( hermon );
 }
 

src/drivers/infiniband/hermon.h

@@ -9,6 +9,7 @@
 
 #include <stdint.h>
 #include <gpxe/uaccess.h>
+#include <gpxe/process.h>
 #include "mlx_bitops.h"
 #include "MT25408_PRM.h"
 
@@ -18,7 +19,7 @@
  */
 
 /* Ports in existence */
-#define HERMON_NUM_PORTS		1
+#define HERMON_NUM_PORTS		2
 #define HERMON_PORT_BASE		1
 
 /* PCI BARs */
@@ -48,6 +49,7 @@
 #define HERMON_HCR_RST2INIT_QP		0x0019
 #define HERMON_HCR_INIT2RTR_QP		0x001a
 #define HERMON_HCR_RTR2RTS_QP		0x001b
+#define HERMON_HCR_RTS2RTS_QP		0x001c
 #define HERMON_HCR_2RST_QP		0x0021
 #define HERMON_HCR_MAD_IFC		0x0024
 #define HERMON_HCR_READ_MCG		0x0025
@@ -75,6 +77,14 @@
 #define HERMON_PAGE_SIZE		4096
 
 #define HERMON_DB_POST_SND_OFFSET	0x14
+#define HERMON_DB_EQ0_OFFSET		0x800
+
+#define HERMON_QP_OPT_PARAM_QKEY	0x00000020UL
+
+#define HERMON_MAP_EQ_MAP		( 0UL << 31 )
+#define HERMON_MAP_EQ_UNMAP		( 1UL << 31 )
+
+#define HERMON_EV_PORT_STATE_CHANGE	0x09
 
 /*
  * Datatypes that seem to be missing from the autogenerated documentation
@@ -108,12 +118,32 @@ struct hermonprm_send_db_register_st {
 	pseudo_bit_t qn[0x00018];
 } __attribute__ (( packed ));
 
+struct hermonprm_event_db_register_st {
+	pseudo_bit_t ci[0x00018];
+	pseudo_bit_t reserver[0x00007];
+	pseudo_bit_t a[0x00001];
+} __attribute__ (( packed ));
+
 struct hermonprm_scalar_parameter_st {
 	pseudo_bit_t value_hi[0x00020];
 /* -------------- */
 	pseudo_bit_t value[0x00020];
 } __attribute__ (( packed ));
 
+struct hermonprm_event_mask_st {
+	pseudo_bit_t reserved0[0x00020];
+/* -------------- */
+	pseudo_bit_t completion[0x00001];
+	pseudo_bit_t reserved1[0x0008];
+	pseudo_bit_t port_state_change[0x00001];
+	pseudo_bit_t reserved2[0x00016];
+} __attribute__ (( packed ));
+
+struct hermonprm_port_state_change_event_st {
+	pseudo_bit_t reserved[0x00020];
+	struct hermonprm_port_state_change_st data;
+} __attribute__ (( packed ));
+
 /*
  * Wrapper structures for hardware datatypes
  *
@@ -124,6 +154,9 @@ struct MLX_DECLARE_STRUCT ( hermonprm_completion_queue_entry );
 struct MLX_DECLARE_STRUCT ( hermonprm_completion_with_error );
 struct MLX_DECLARE_STRUCT ( hermonprm_cq_db_record );
 struct MLX_DECLARE_STRUCT ( hermonprm_eqc );
+struct MLX_DECLARE_STRUCT ( hermonprm_event_db_register );
+struct MLX_DECLARE_STRUCT ( hermonprm_event_mask );
+struct MLX_DECLARE_STRUCT ( hermonprm_event_queue_entry );
 struct MLX_DECLARE_STRUCT ( hermonprm_hca_command_register );
 struct MLX_DECLARE_STRUCT ( hermonprm_init_hca );
 struct MLX_DECLARE_STRUCT ( hermonprm_init_port );
@@ -132,6 +165,7 @@ struct MLX_DECLARE_STRUCT ( hermonprm_mcg_entry );
 struct MLX_DECLARE_STRUCT ( hermonprm_mgm_hash );
 struct MLX_DECLARE_STRUCT ( hermonprm_mpt );
 struct MLX_DECLARE_STRUCT ( hermonprm_mtt );
+struct MLX_DECLARE_STRUCT ( hermonprm_port_state_change_event );
 struct MLX_DECLARE_STRUCT ( hermonprm_qp_db_record );
 struct MLX_DECLARE_STRUCT ( hermonprm_qp_ee_state_transitions );
 struct MLX_DECLARE_STRUCT ( hermonprm_query_dev_cap );
@@ -175,8 +209,14 @@ union hermonprm_completion_entry {
 	struct hermonprm_completion_with_error error;
 } __attribute__ (( packed ));
 
+union hermonprm_event_entry {
+	struct hermonprm_event_queue_entry generic;
+	struct hermonprm_port_state_change_event port_state_change;
+} __attribute__ (( packed ));
+
 union hermonprm_doorbell_register {
 	struct hermonprm_send_db_register send;
+	struct hermonprm_event_db_register event;
 	uint32_t dword[1];
 } __attribute__ (( packed ));
 
@@ -362,6 +402,24 @@ struct hermon_completion_queue {
  */
 #define HERMON_MAX_EQS		4
 
+/** A Hermon event queue */
+struct hermon_event_queue {
+	/** Event queue entries */
+	union hermonprm_event_entry *eqe;
+	/** Size of event queue */
+	size_t eqe_size;
+	/** MTT descriptor */
+	struct hermon_mtt mtt;
+	/** Next event queue entry index */
+	unsigned long next_idx;
+};
+
+/** Number of event queue entries
+ *
+ * This is a policy decision.
+ */
+#define HERMON_NUM_EQES		4
+
 /** A Hermon resource bitmask */
 typedef uint32_t hermon_bitmask_t;
 
@@ -397,6 +455,11 @@ struct hermon {
 	 */
 	unsigned long reserved_lkey;
 
+	/** Event queue */
+	struct hermon_event_queue eq;
+	/** Event queue process */
+	struct process event_process;
+
 	/** Completion queue in-use bitmask */
 	hermon_bitmask_t cq_inuse[ HERMON_BITMASK_SIZE ( HERMON_MAX_CQS ) ];
 	/** Queue pair in-use bitmask */

src/drivers/net/ipoib.c

@@ -80,10 +80,14 @@ struct ipoib_device {
 	struct ib_gid broadcast_gid;
 	/** Broadcast LID */
 	unsigned int broadcast_lid;
-	/** Joined to broadcast group */
-	int broadcast_joined;
 	/** Data queue key */
 	unsigned long data_qkey;
+	/** Attached to multicast group
+	 *
+	 * This flag indicates whether or not we have attached our
+	 * data queue pair to the broadcast multicast GID.
+	 */
+	int broadcast_attached;
 };
 
 /**
@@ -272,6 +276,10 @@ static int ipoib_create_qset ( struct ipoib_device *ipoib,
 	struct ib_device *ibdev = ipoib->ibdev;
 	int rc;
 
+	/* Sanity check */
+	assert ( qset->cq == NULL );
+	assert ( qset->qp == NULL );
+
 	/* Store queue parameters */
 	qset->recv_max_fill = num_recv_wqes;
 
@@ -617,14 +625,24 @@ static void ipoib_recv_path_record ( struct ipoib_device *ipoib __unused,
  */
 static void ipoib_recv_mc_member_record ( struct ipoib_device *ipoib,
 			  struct ib_mad_mc_member_record *mc_member_record ) {
+	int joined;
+	int rc;
+
 	/* Record parameters */
-	ipoib->broadcast_joined =
-		( mc_member_record->scope__join_state & 0x0f );
+	joined = ( mc_member_record->scope__join_state & 0x0f );
 	ipoib->data_qkey = ntohl ( mc_member_record->qkey );
 	ipoib->broadcast_lid = ntohs ( mc_member_record->mlid );
 	DBGC ( ipoib, "IPoIB %p %s broadcast group: qkey %lx mlid %x\n",
-	       ipoib, ( ipoib->broadcast_joined ? "joined" : "left" ),
-	       ipoib->data_qkey, ipoib->broadcast_lid );
+	       ipoib, ( joined ? "joined" : "left" ), ipoib->data_qkey,
+	       ipoib->broadcast_lid );
+
+	/* Update data queue pair qkey */
+	if ( ( rc = ib_modify_qp ( ipoib->ibdev, ipoib->data.qp,
+				   IB_MODIFY_QKEY, ipoib->data_qkey ) ) != 0 ){
+		DBGC ( ipoib, "IPoIB %p could not update data qkey: %s\n",
+		       ipoib, strerror ( rc ) );
+		return;
+	}
 }
 
 /**
@@ -741,6 +759,56 @@ static void ipoib_irq ( struct net_device *netdev __unused,
 	/* No implementation */
 }
 
+/**
+ * Join IPv4 broadcast multicast group
+ *
+ * @v ipoib		IPoIB device
+ * @ret rc		Return status code
+ */
+static int ipoib_join_broadcast_group ( struct ipoib_device *ipoib ) {
+	int rc;
+
+	/* Sanity check */
+	if ( ! ipoib->data.qp )
+		return 0;
+
+	/* Attach data queue to broadcast multicast GID */
+	assert ( ipoib->broadcast_attached == 0 );
+	if ( ( rc = ib_mcast_attach ( ipoib->ibdev, ipoib->data.qp,
+				      &ipoib->broadcast_gid ) ) != 0 ){
+		DBGC ( ipoib, "IPoIB %p could not attach to broadcast GID: "
+		       "%s\n", ipoib, strerror ( rc ) );
+		return rc;
+	}
+	ipoib->broadcast_attached = 1;
+
+	/* Initiate broadcast group join */
+	if ( ( rc = ipoib_mc_member_record ( ipoib, &ipoib->broadcast_gid,
+					     1 ) ) != 0 ) {
+		DBGC ( ipoib, "IPoIB %p could not send broadcast join: %s\n",
+		       ipoib, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+}
+
+/**
+ * Leave IPv4 broadcast multicast group
+ *
+ * @v ipoib		IPoIB device
+ */
+static void ipoib_leave_broadcast_group ( struct ipoib_device *ipoib ) {
+
+	/* Detach data queue from broadcast multicast GID */
+	if ( ipoib->broadcast_attached ) {
+		assert ( ipoib->data.qp != NULL );
+		ib_mcast_detach ( ipoib->ibdev, ipoib->data.qp,
+				  &ipoib->broadcast_gid );
+		ipoib->broadcast_attached = 0;
+	}
+}
+
 /**
  * Open IPoIB network device
  *
@@ -749,22 +817,53 @@ static void ipoib_irq ( struct net_device *netdev __unused,
  */
 static int ipoib_open ( struct net_device *netdev ) {
 	struct ipoib_device *ipoib = netdev->priv;
-	struct ib_device *ibdev = ipoib->ibdev;
+	struct ipoib_mac *mac = ( ( struct ipoib_mac * ) netdev->ll_addr );
 	int rc;
 
-	/* Attach to broadcast multicast GID */
-	if ( ( rc = ib_mcast_attach ( ibdev, ipoib->data.qp,
-				      &ipoib->broadcast_gid ) ) != 0 ) {
-		DBG ( "Could not attach to broadcast GID: %s\n",
-		      strerror ( rc ) );
-		return rc;
+	/* Allocate metadata queue set */
+	if ( ( rc = ipoib_create_qset ( ipoib, &ipoib->meta,
+					IPOIB_META_NUM_CQES,
+					IPOIB_META_NUM_SEND_WQES,
+					IPOIB_META_NUM_RECV_WQES,
+					IB_GLOBAL_QKEY ) ) != 0 ) {
+		DBGC ( ipoib, "IPoIB %p could not allocate metadata QP: %s\n",
+		       ipoib, strerror ( rc ) );
+		goto err_create_meta_qset;
 	}
 
+	/* Allocate data queue set */
+	if ( ( rc = ipoib_create_qset ( ipoib, &ipoib->data,
+					IPOIB_DATA_NUM_CQES,
+					IPOIB_DATA_NUM_SEND_WQES,
+					IPOIB_DATA_NUM_RECV_WQES,
+					IB_GLOBAL_QKEY ) ) != 0 ) {
+		DBGC ( ipoib, "IPoIB %p could not allocate data QP: %s\n",
+		       ipoib, strerror ( rc ) );
+		goto err_create_data_qset;
+	}
+
+	/* Update MAC address with data QPN */
+	mac->qpn = htonl ( ipoib->data.qp->qpn );
+
 	/* Fill receive rings */
 	ipoib_refill_recv ( ipoib, &ipoib->meta );
 	ipoib_refill_recv ( ipoib, &ipoib->data );
 
+	/* Join broadcast group */
+	if ( ( rc = ipoib_join_broadcast_group ( ipoib ) ) != 0 ) {
+		DBGC ( ipoib, "IPoIB %p could not join broadcast group: %s\n",
+		       ipoib, strerror ( rc ) );
+		goto err_join_broadcast;
+	}
+
 	return 0;
+
+ err_join_broadcast:
+	ipoib_destroy_qset ( ipoib, &ipoib->data );
+ err_create_data_qset:
+	ipoib_destroy_qset ( ipoib, &ipoib->meta );
+ err_create_meta_qset:
+	return rc;
 }
 
 /**
@@ -774,12 +873,17 @@ static int ipoib_open ( struct net_device *netdev ) {
  */
 static void ipoib_close ( struct net_device *netdev ) {
 	struct ipoib_device *ipoib = netdev->priv;
-	struct ib_device *ibdev = ipoib->ibdev;
+	struct ipoib_mac *mac = ( ( struct ipoib_mac * ) netdev->ll_addr );
 
-	/* Detach from broadcast multicast GID */
-	ib_mcast_detach ( ibdev, ipoib->data.qp, &ipoib->broadcast_gid );
+	/* Leave broadcast group */
+	ipoib_leave_broadcast_group ( ipoib );
 
-	/* FIXME: should probably flush the receive ring */
+	/* Remove data QPN from MAC address */
+	mac->qpn = 0;
+
+	/* Tear down the queues */
+	ipoib_destroy_qset ( ipoib, &ipoib->data );
+	ipoib_destroy_qset ( ipoib, &ipoib->meta );
 }
 
 /** IPoIB network device operations */
@@ -792,44 +896,53 @@ static struct net_device_operations ipoib_operations = {
 };
 
 /**
- * Join IPoIB broadcast group
+ * Update IPoIB dynamic Infiniband parameters
  *
  * @v ipoib		IPoIB device
- * @ret rc		Return status code
+ *
+ * The Infiniband port GID and partition key will change at runtime,
+ * when the link is established (or lost).  The MAC address is based
+ * on the port GID, and the broadcast GID is based on the partition
+ * key.  This function recalculates these IPoIB device parameters.
  */
-static int ipoib_join_broadcast_group ( struct ipoib_device *ipoib ) {
+static void ipoib_set_ib_params ( struct ipoib_device *ipoib ) {
 	struct ib_device *ibdev = ipoib->ibdev;
-	unsigned int delay_ms;
-	int rc;
+	struct ipoib_mac *mac;
 
-	/* Make sure we have some receive descriptors */
-	ipoib_refill_recv ( ipoib, &ipoib->meta );
+	/* Calculate GID portion of MAC address based on port GID */
+	mac = ( ( struct ipoib_mac * ) ipoib->netdev->ll_addr );
+	memcpy ( &mac->gid, &ibdev->port_gid, sizeof ( mac->gid ) );
 
-	/* Send join request */
-	if ( ( rc = ipoib_mc_member_record ( ipoib, &ipoib->broadcast_gid,
-					     1 ) ) != 0 ) {
-		DBGC ( ipoib, "IPoIB %p could not send broadcast join: %s\n",
-		       ipoib, strerror ( rc ) );
-		return rc;
-	}
+	/* Calculate broadcast GID based on partition key */
+	memcpy ( &ipoib->broadcast_gid, &ipv4_broadcast_gid,
+		 sizeof ( ipoib->broadcast_gid ) );
+	ipoib->broadcast_gid.u.words[2] = htons ( ibdev->pkey );
+}
+
+/**
+ * Handle link status change
+ *
+ * @v ibdev		Infiniband device
+ */
+void ipoib_link_state_changed ( struct ib_device *ibdev ) {
+	struct net_device *netdev = ib_get_ownerdata ( ibdev );
+	struct ipoib_device *ipoib = netdev->priv;
+	int rc;
 
-	/* Wait for join to complete.  Ideally we wouldn't delay for
-	 * this long, but we need the queue key before we can set up
-	 * the data queue pair, which we need before we can know the
-	 * MAC address.
+	/* Leave existing broadcast group */
+	ipoib_leave_broadcast_group ( ipoib );
+
+	/* Update MAC address and broadcast GID based on new port GID
+	 * and partition key.
 	 */
-	for ( delay_ms = IPOIB_JOIN_MAX_DELAY_MS ; delay_ms ; delay_ms-- ) {
-		mdelay ( 1 );
-		ib_poll_cq ( ibdev, ipoib->meta.cq, ipoib_meta_complete_send,
-			     ipoib_meta_complete_recv );
-		ipoib_refill_recv ( ipoib, &ipoib->meta );
-		if ( ipoib->broadcast_joined )
-			return 0;
-	}
-	DBGC ( ipoib, "IPoIB %p timed out waiting for broadcast join\n",
-	       ipoib );
+	ipoib_set_ib_params ( ipoib );
 
-	return -ETIMEDOUT;
+	/* Join new broadcast group */
+	if ( ( rc = ipoib_join_broadcast_group ( ipoib ) ) != 0 ) {
+		DBGC ( ipoib, "IPoIB %p could not rejoin broadcast group: "
+		       "%s\n", ipoib, strerror ( rc ) );
+		return;
+	}
 }
 
 /**
@@ -841,7 +954,6 @@ static int ipoib_join_broadcast_group ( struct ipoib_device *ipoib ) {
 int ipoib_probe ( struct ib_device *ibdev ) {
 	struct net_device *netdev;
 	struct ipoib_device *ipoib;
-	struct ipoib_mac *mac;
 	int rc;
 
 	/* Allocate network device */
@@ -856,44 +968,11 @@ int ipoib_probe ( struct ib_device *ibdev ) {
 	ipoib->netdev = netdev;
 	ipoib->ibdev = ibdev;
 
-	/* Calculate broadcast GID */
-	memcpy ( &ipoib->broadcast_gid, &ipv4_broadcast_gid,
-		 sizeof ( ipoib->broadcast_gid ) );
-	ipoib->broadcast_gid.u.words[2] = htons ( ibdev->pkey );
-
-	/* Allocate metadata queue set */
-	if ( ( rc = ipoib_create_qset ( ipoib, &ipoib->meta,
-					IPOIB_META_NUM_CQES,
-					IPOIB_META_NUM_SEND_WQES,
-					IPOIB_META_NUM_RECV_WQES,
-					IB_GLOBAL_QKEY ) ) != 0 ) {
-		DBGC ( ipoib, "IPoIB %p could not allocate metadata QP: %s\n",
-		       ipoib, strerror ( rc ) );
-		goto err_create_meta_qset;
-	}
-
-	/* Join broadcast group */
-	if ( ( rc = ipoib_join_broadcast_group ( ipoib ) ) != 0 ) {
-		DBGC ( ipoib, "IPoIB %p could not join broadcast group: %s\n",
-		       ipoib, strerror ( rc ) );
-		goto err_join_broadcast_group;
-	}
-
-	/* Allocate data queue set */
-	if ( ( rc = ipoib_create_qset ( ipoib, &ipoib->data,
-					IPOIB_DATA_NUM_CQES,
-					IPOIB_DATA_NUM_SEND_WQES,
-					IPOIB_DATA_NUM_RECV_WQES,
-					ipoib->data_qkey ) ) != 0 ) {
-		DBGC ( ipoib, "IPoIB %p could not allocate data QP: %s\n",
-		       ipoib, strerror ( rc ) );
-		goto err_create_data_qset;
-	}
-
-	/* Construct MAC address */
-	mac = ( ( struct ipoib_mac * ) netdev->ll_addr );
-	mac->qpn = htonl ( ipoib->data.qp->qpn );
-	memcpy ( &mac->gid, &ibdev->port_gid, sizeof ( mac->gid ) );
+	/* Calculate as much of the broadcast GID and the MAC address
+	 * as we can.  We won't know either of these in full until we
+	 * have link-up.
+	 */
+	ipoib_set_ib_params ( ipoib );
 
 	/* Register network device */
 	if ( ( rc = register_netdev ( netdev ) ) != 0 )
@@ -902,11 +981,6 @@ int ipoib_probe ( struct ib_device *ibdev ) {
 	return 0;
 
  err_register_netdev:
-	ipoib_destroy_qset ( ipoib, &ipoib->data );
- err_join_broadcast_group:
- err_create_data_qset:
-	ipoib_destroy_qset ( ipoib, &ipoib->meta );
- err_create_meta_qset:
 	netdev_nullify ( netdev );
 	netdev_put ( netdev );
 	return rc;
@@ -919,11 +993,8 @@ int ipoib_probe ( struct ib_device *ibdev ) {
  */
 void ipoib_remove ( struct ib_device *ibdev ) {
 	struct net_device *netdev = ib_get_ownerdata ( ibdev );
-	struct ipoib_device *ipoib = netdev->priv;
 
 	unregister_netdev ( netdev );
-	ipoib_destroy_qset ( ipoib, &ipoib->data );
-	ipoib_destroy_qset ( ipoib, &ipoib->meta );
 	netdev_nullify ( netdev );
 	netdev_put ( netdev );
 }

src/include/gpxe/infiniband.h

@@ -95,6 +95,11 @@ struct ib_queue_pair {
 	void *owner_priv;
 };
 
+/** Infiniband queue pair modification flags */
+enum ib_queue_pair_mods {
+	IB_MODIFY_QKEY = 0x0001,
+};
+
 /** An Infiniband Completion Queue */
 struct ib_completion_queue {
 	/** Completion queue number */
@@ -187,6 +192,16 @@ struct ib_device_operations {
 	 */
 	int ( * create_qp ) ( struct ib_device *ibdev,
 			      struct ib_queue_pair *qp );
+	/** Modify queue pair
+	 *
+	 * @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 );
 	/** Destroy queue pair
 	 *
 	 * @v ibdev		Infiniband device
@@ -291,6 +306,8 @@ struct ib_device {
 	struct ib_device_operations *op;
 	/** Port number */
 	unsigned int port;
+	/** Link state */
+	int link_up;
 	/** Port GID */
 	struct ib_gid port_gid;
 	/** Subnet manager LID */
@@ -311,6 +328,8 @@ extern struct ib_queue_pair *
 ib_create_qp ( struct ib_device *ibdev, 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 );
 extern void ib_destroy_qp ( struct ib_device *ibdev,
 			    struct ib_queue_pair *qp );
 extern struct ib_work_queue * ib_find_wq ( struct ib_completion_queue *cq,
@@ -319,6 +338,7 @@ 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 );
 extern void free_ibdev ( struct ib_device *ibdev );
+extern void ib_link_state_changed ( struct ib_device *ibdev );
 
 /**
  * Post send work queue entry

src/include/gpxe/ipoib.h

@@ -72,6 +72,7 @@ static inline struct net_device * alloc_ipoibdev ( size_t priv_size ) {
 	return netdev;
 }
 
+extern void ipoib_link_state_changed ( struct ib_device *ibdev );
 extern int ipoib_probe ( struct ib_device *ibdev );
 extern void ipoib_remove ( struct ib_device *ibdev );
 

src/net/infiniband.c

@@ -152,15 +152,41 @@ struct ib_queue_pair * ib_create_qp ( struct ib_device *ibdev,
 	return qp;
 }
 
+/**
+ * Modify queue pair
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v mod_list		Modification list
+ * @v qkey		New queue key, 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 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 ) {
+		DBGC ( ibdev, "IBDEV %p could not modify QPN %#lx: %s\n",
+		       ibdev, qp->qpn, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+}
+
 /**
  * Destroy queue pair
  *
  * @v ibdev		Infiniband device
  * @v qp		Queue pair
  */
-void ib_destroy_qp ( struct ib_device *ibdev,
-		     struct ib_queue_pair *qp ) {
-	DBGC ( ibdev, "IBDEV %p destroying queue pair %#lx\n",
+void ib_destroy_qp ( struct ib_device *ibdev, struct ib_queue_pair *qp ) {
+	DBGC ( ibdev, "IBDEV %p destroying QPN %#lx\n",
 	       ibdev, qp->qpn );
 	ibdev->op->destroy_qp ( ibdev, qp );
 	list_del ( &qp->send.list );
@@ -279,38 +305,6 @@ static int ib_get_pkey_table ( struct ib_device *ibdev,
 	return 0;
 }
 
-/**
- * Wait for link up
- *
- * @v ibdev		Infiniband device
- * @ret rc		Return status code
- *
- * This function shouldn't really exist.  Unfortunately, IB links take
- * a long time to come up, and we can't get various key parameters
- * e.g. our own IPoIB MAC address without information from the subnet
- * manager).  We should eventually make link-up an asynchronous event.
- */
-static int ib_wait_for_link ( struct ib_device *ibdev ) {
-	struct ib_mad_port_info port_info;
-	unsigned int retries;
-	int rc;
-
-	printf ( "Waiting for Infiniband link-up..." );
-	for ( retries = 20 ; retries ; retries-- ) {
-		if ( ( rc = ib_get_port_info ( ibdev, &port_info ) ) != 0 )
-			continue;
-		if ( ( ( port_info.port_state__link_speed_supported ) & 0xf )
-		     == 4 ) {
-			printf ( "ok\n" );
-			return 0;
-		}
-		printf ( "." );
-		sleep ( 1 );
-	}
-	printf ( "failed\n" );
-	return -ENODEV;
-};
-
 /**
  * Get MAD parameters
  *
@@ -326,9 +320,13 @@ static int ib_get_mad_params ( struct ib_device *ibdev ) {
 	} u;
 	int rc;
 
-	/* Port info gives us the first half of the port GID and the SM LID */
+	/* Port info gives us the link state, the first half of the
+	 * port GID and the SM LID.
+	 */
 	if ( ( rc = ib_get_port_info ( ibdev, &u.port_info ) ) != 0 )
 		return rc;
+	ibdev->link_up = ( ( u.port_info.port_state__link_speed_supported
+			     & 0xf ) == 4 );
 	memcpy ( &ibdev->port_gid.u.bytes[0], u.port_info.gid_prefix, 8 );
 	ibdev->sm_lid = ntohs ( u.port_info.mastersm_lid );
 
@@ -391,10 +389,6 @@ int register_ibdev ( struct ib_device *ibdev ) {
 	if ( ( rc = ib_open ( ibdev ) ) != 0 )
 		goto err_open;
 
-	/* Wait for link */
-	if ( ( rc = ib_wait_for_link ( ibdev ) ) != 0 )
-		goto err_wait_for_link;
-
 	/* Get MAD parameters */
 	if ( ( rc = ib_get_mad_params ( ibdev ) ) != 0 )
 		goto err_get_mad_params;
@@ -410,7 +404,6 @@ int register_ibdev ( struct ib_device *ibdev ) {
 
  err_ipoib_probe:
  err_get_mad_params:
- err_wait_for_link:
 	ib_close ( ibdev );
  err_open:
 	return rc;
@@ -435,3 +428,21 @@ void free_ibdev ( struct ib_device *ibdev ) {
 	free ( ibdev );
 }
 
+/**
+ * Handle Infiniband link state change
+ *
+ * @v ibdev		Infiniband device
+ */
+void ib_link_state_changed ( struct ib_device *ibdev ) {
+	int rc;
+
+	/* Update MAD parameters */
+	if ( ( rc = ib_get_mad_params ( ibdev ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not update MAD parameters: %s\n",
+		       ibdev, strerror ( rc ) );
+		return;
+	}
+
+	/* Notify IPoIB of link state change */
+	ipoib_link_state_changed ( ibdev );
+}