Merge branch '3leaf'

src/Makefile

@@ -152,6 +152,7 @@ SRCDIRS		+= drivers/scsi
 SRCDIRS		+= drivers/ata
 SRCDIRS		+= drivers/nvs
 SRCDIRS		+= drivers/bitbash
+SRCDIRS		+= drivers/infiniband
 SRCDIRS		+= interface/pxe
 SRCDIRS		+= tests
 SRCDIRS		+= crypto crypto/axtls crypto/matrixssl

src/drivers/infiniband/arbel.h

@@ -0,0 +1,461 @@
+#ifndef _ARBEL_H
+#define _ARBEL_H
+
+/** @file
+ *
+ * Mellanox Arbel Infiniband HCA driver
+ *
+ */
+
+#include <stdint.h>
+#include <gpxe/uaccess.h>
+#include "mlx_bitops.h"
+#include "MT25218_PRM.h"
+
+/*
+ * Hardware constants
+ *
+ */
+
+/* PCI BARs */
+#define ARBEL_PCI_CONFIG_BAR		PCI_BASE_ADDRESS_0
+#define ARBEL_PCI_CONFIG_BAR_SIZE	0x100000
+#define ARBEL_PCI_UAR_BAR		PCI_BASE_ADDRESS_2
+#define ARBEL_PCI_UAR_IDX		1
+#define ARBEL_PCI_UAR_SIZE		0x1000
+
+/* UAR context table (UCE) resource types */
+#define ARBEL_UAR_RES_NONE		0x00
+#define ARBEL_UAR_RES_CQ_CI		0x01
+#define ARBEL_UAR_RES_CQ_ARM		0x02
+#define ARBEL_UAR_RES_SQ		0x03
+#define ARBEL_UAR_RES_RQ		0x04
+#define ARBEL_UAR_RES_GROUP_SEP		0x07
+
+/* Work queue entry and completion queue entry opcodes */
+#define ARBEL_OPCODE_SEND		0x0a
+#define ARBEL_OPCODE_RECV_ERROR		0xfe
+#define ARBEL_OPCODE_SEND_ERROR		0xff
+
+/* HCA command register opcodes */
+#define ARBEL_HCR_QUERY_DEV_LIM		0x0003
+#define ARBEL_HCR_QUERY_FW		0x0004
+#define ARBEL_HCR_INIT_HCA		0x0007
+#define ARBEL_HCR_CLOSE_HCA		0x0008
+#define ARBEL_HCR_INIT_IB		0x0009
+#define ARBEL_HCR_CLOSE_IB		0x000a
+#define ARBEL_HCR_SW2HW_MPT		0x000d
+#define ARBEL_HCR_MAP_EQ		0x0012
+#define ARBEL_HCR_SW2HW_EQ		0x0013
+#define ARBEL_HCR_HW2SW_EQ		0x0014
+#define ARBEL_HCR_SW2HW_CQ		0x0016
+#define ARBEL_HCR_HW2SW_CQ		0x0017
+#define ARBEL_HCR_RST2INIT_QPEE		0x0019
+#define ARBEL_HCR_INIT2RTR_QPEE		0x001a
+#define ARBEL_HCR_RTR2RTS_QPEE		0x001b
+#define ARBEL_HCR_2RST_QPEE		0x0021
+#define ARBEL_HCR_MAD_IFC		0x0024
+#define ARBEL_HCR_READ_MGM		0x0025
+#define ARBEL_HCR_WRITE_MGM		0x0026
+#define ARBEL_HCR_MGID_HASH		0x0027
+#define ARBEL_HCR_RUN_FW		0x0ff6
+#define ARBEL_HCR_DISABLE_LAM		0x0ff7
+#define ARBEL_HCR_ENABLE_LAM		0x0ff8
+#define ARBEL_HCR_UNMAP_ICM		0x0ff9
+#define ARBEL_HCR_MAP_ICM		0x0ffa
+#define ARBEL_HCR_UNMAP_ICM_AUX		0x0ffb
+#define ARBEL_HCR_MAP_ICM_AUX		0x0ffc
+#define ARBEL_HCR_SET_ICM_SIZE		0x0ffd
+#define ARBEL_HCR_UNMAP_FA		0x0ffe
+#define ARBEL_HCR_MAP_FA		0x0fff
+
+/* Service types */
+#define ARBEL_ST_UD			0x03
+
+/* MTUs */
+#define ARBEL_MTU_2048			0x04
+
+#define ARBEL_NO_EQ			64
+
+#define ARBEL_INVALID_LKEY		0x00000100UL
+
+#define ARBEL_PAGE_SIZE			4096
+
+#define ARBEL_DB_POST_SND_OFFSET	0x10
+
+/*
+ * Datatypes that seem to be missing from the autogenerated documentation
+ *
+ */
+struct arbelprm_mgm_hash_st {
+	pseudo_bit_t reserved0[0x00020];
+/* -------------- */
+	pseudo_bit_t hash[0x00010];
+	pseudo_bit_t reserved1[0x00010];
+} __attribute__ (( packed ));
+
+struct arbelprm_scalar_parameter_st {
+	pseudo_bit_t reserved0[0x00020];
+/* -------------- */
+	pseudo_bit_t value[0x00020];
+} __attribute__ (( packed ));
+
+/*
+ * Wrapper structures for hardware datatypes
+ *
+ */
+
+struct MLX_DECLARE_STRUCT ( arbelprm_access_lam );
+struct MLX_DECLARE_STRUCT ( arbelprm_completion_queue_context );
+struct MLX_DECLARE_STRUCT ( arbelprm_completion_queue_entry );
+struct MLX_DECLARE_STRUCT ( arbelprm_completion_with_error );
+struct MLX_DECLARE_STRUCT ( arbelprm_cq_arm_db_record );
+struct MLX_DECLARE_STRUCT ( arbelprm_cq_ci_db_record );
+struct MLX_DECLARE_STRUCT ( arbelprm_eqc );
+struct MLX_DECLARE_STRUCT ( arbelprm_hca_command_register );
+struct MLX_DECLARE_STRUCT ( arbelprm_init_hca );
+struct MLX_DECLARE_STRUCT ( arbelprm_init_ib );
+struct MLX_DECLARE_STRUCT ( arbelprm_mad_ifc );
+struct MLX_DECLARE_STRUCT ( arbelprm_mgm_entry );
+struct MLX_DECLARE_STRUCT ( arbelprm_mgm_hash );
+struct MLX_DECLARE_STRUCT ( arbelprm_mpt );
+struct MLX_DECLARE_STRUCT ( arbelprm_qp_db_record );
+struct MLX_DECLARE_STRUCT ( arbelprm_qp_ee_state_transitions );
+struct MLX_DECLARE_STRUCT ( arbelprm_query_dev_lim );
+struct MLX_DECLARE_STRUCT ( arbelprm_query_fw );
+struct MLX_DECLARE_STRUCT ( arbelprm_queue_pair_ee_context_entry );
+struct MLX_DECLARE_STRUCT ( arbelprm_recv_wqe_segment_next );
+struct MLX_DECLARE_STRUCT ( arbelprm_scalar_parameter );
+struct MLX_DECLARE_STRUCT ( arbelprm_send_doorbell );
+struct MLX_DECLARE_STRUCT ( arbelprm_ud_address_vector );
+struct MLX_DECLARE_STRUCT ( arbelprm_virtual_physical_mapping );
+struct MLX_DECLARE_STRUCT ( arbelprm_wqe_segment_ctrl_send );
+struct MLX_DECLARE_STRUCT ( arbelprm_wqe_segment_data_ptr );
+struct MLX_DECLARE_STRUCT ( arbelprm_wqe_segment_next );
+struct MLX_DECLARE_STRUCT ( arbelprm_wqe_segment_ud );
+
+/*
+ * Composite hardware datatypes
+ *
+ */
+
+#define ARBEL_MAX_GATHER 1
+
+struct arbelprm_ud_send_wqe {
+	struct arbelprm_wqe_segment_next next;
+	struct arbelprm_wqe_segment_ctrl_send ctrl;
+	struct arbelprm_wqe_segment_ud ud;
+	struct arbelprm_wqe_segment_data_ptr data[ARBEL_MAX_GATHER];
+} __attribute__ (( packed ));
+
+#define ARBEL_MAX_SCATTER 1
+
+struct arbelprm_recv_wqe {
+	/* The autogenerated header is inconsistent between send and
+	 * receive WQEs.  The "ctrl" structure for receive WQEs is
+	 * defined to include the "next" structure.  Since the "ctrl"
+	 * part of the "ctrl" structure contains only "reserved, must
+	 * be zero" bits, we ignore its definition and provide
+	 * something more usable.
+	 */
+	struct arbelprm_recv_wqe_segment_next next;
+	uint32_t ctrl[2]; /* All "reserved, must be zero" */
+	struct arbelprm_wqe_segment_data_ptr data[ARBEL_MAX_SCATTER];
+} __attribute__ (( packed ));
+
+union arbelprm_completion_entry {
+	struct arbelprm_completion_queue_entry normal;
+	struct arbelprm_completion_with_error error;
+} __attribute__ (( packed ));
+
+union arbelprm_doorbell_record {
+	struct arbelprm_cq_arm_db_record cq_arm;
+	struct arbelprm_cq_ci_db_record cq_ci;
+	struct arbelprm_qp_db_record qp;
+} __attribute__ (( packed ));
+
+union arbelprm_doorbell_register {
+	struct arbelprm_send_doorbell send;
+	uint32_t dword[2];
+} __attribute__ (( packed ));
+
+union arbelprm_mad {
+	struct arbelprm_mad_ifc ifc;
+	union ib_mad mad;
+} __attribute__ (( packed ));
+
+/*
+ * gPXE-specific definitions
+ *
+ */
+
+/** Arbel device limits */
+struct arbel_dev_limits {
+	/** Number of reserved QPs */
+	unsigned int reserved_qps;
+	/** QP context entry size */
+	size_t qpc_entry_size;
+	/** Extended QP context entry size */
+	size_t eqpc_entry_size;
+	/** Number of reserved SRQs */
+	unsigned int reserved_srqs;
+	/** SRQ context entry size */
+	size_t srqc_entry_size;
+	/** Number of reserved EEs */
+	unsigned int reserved_ees;
+	/** EE context entry size */
+	size_t eec_entry_size;
+	/** Extended EE context entry size */
+	size_t eeec_entry_size;
+	/** Number of reserved CQs */
+	unsigned int reserved_cqs;
+	/** CQ context entry size */
+	size_t cqc_entry_size;
+	/** Number of reserved MTTs */
+	unsigned int reserved_mtts;
+	/** MTT entry size */
+	size_t mtt_entry_size;
+	/** Number of reserved MRWs */
+	unsigned int reserved_mrws;
+	/** MPT entry size */
+	size_t mpt_entry_size;
+	/** Number of reserved RDBs */
+	unsigned int reserved_rdbs;
+	/** EQ context entry size */
+	size_t eqc_entry_size;
+	/** Number of reserved UARs */
+	unsigned int reserved_uars;
+};
+
+/** Alignment of Arbel send work queue entries */
+#define ARBEL_SEND_WQE_ALIGN 128
+
+/** An Arbel send work queue entry */
+union arbel_send_wqe {
+	struct arbelprm_ud_send_wqe ud;
+	uint8_t force_align[ARBEL_SEND_WQE_ALIGN];
+} __attribute__ (( packed ));
+
+/** An Arbel send work queue */
+struct arbel_send_work_queue {
+	/** Doorbell record number */
+	unsigned int doorbell_idx;
+	/** Work queue entries */
+	union arbel_send_wqe *wqe;
+	/** Size of work queue */
+	size_t wqe_size;
+};
+
+/** Alignment of Arbel receive work queue entries */
+#define ARBEL_RECV_WQE_ALIGN 64
+
+/** An Arbel receive work queue entry */
+union arbel_recv_wqe {
+	struct arbelprm_recv_wqe recv;
+	uint8_t force_align[ARBEL_RECV_WQE_ALIGN];
+} __attribute__ (( packed ));
+
+/** An Arbel receive work queue */
+struct arbel_recv_work_queue {
+	/** Doorbell record number */
+	unsigned int doorbell_idx;
+	/** Work queue entries */
+	union arbel_recv_wqe *wqe;
+	/** Size of work queue */
+	size_t wqe_size;
+};
+
+/** Maximum number of allocatable queue pairs
+ *
+ * This is a policy decision, not a device limit.
+ */
+#define ARBEL_MAX_QPS		8
+
+/** Base queue pair number */
+#define ARBEL_QPN_BASE 0x550000
+
+/** An Arbel queue pair */
+struct arbel_queue_pair {
+	/** Send work queue */
+	struct arbel_send_work_queue send;
+	/** Receive work queue */
+	struct arbel_recv_work_queue recv;
+};
+
+/** Maximum number of allocatable completion queues
+ *
+ * This is a policy decision, not a device limit.
+ */
+#define ARBEL_MAX_CQS		8
+
+/** An Arbel completion queue */
+struct arbel_completion_queue {
+	/** Consumer counter doorbell record number */
+	unsigned int ci_doorbell_idx;
+	/** Arm queue doorbell record number */
+	unsigned int arm_doorbell_idx;
+	/** Completion queue entries */
+	union arbelprm_completion_entry *cqe;
+	/** Size of completion queue */
+	size_t cqe_size;
+};
+
+/** An Arbel resource bitmask */
+typedef uint32_t arbel_bitmask_t;
+
+/** Size of an Arbel resource bitmask */
+#define ARBEL_BITMASK_SIZE(max_entries)					     \
+	( ( (max_entries) + ( 8 * sizeof ( arbel_bitmask_t ) ) - 1 ) /	     \
+	  ( 8 * sizeof ( arbel_bitmask_t ) ) )
+
+/** An Arbel device */
+struct arbel {
+	/** PCI configuration registers */
+	void *config;
+	/** PCI user Access Region */
+	void *uar;
+
+	/** Command input mailbox */
+	void *mailbox_in;
+	/** Command output mailbox */
+	void *mailbox_out;
+
+	/** Firmware area in external memory */
+	userptr_t firmware_area;
+	/** ICM size */
+	size_t icm_len;
+	/** ICM AUX size */
+	size_t icm_aux_len;
+	/** ICM area */
+	userptr_t icm;
+
+	/** Doorbell records */
+	union arbelprm_doorbell_record *db_rec;
+	/** Reserved LKey
+	 *
+	 * Used to get unrestricted memory access.
+	 */
+	unsigned long reserved_lkey;
+
+	/** Completion queue in-use bitmask */
+	arbel_bitmask_t cq_inuse[ ARBEL_BITMASK_SIZE ( ARBEL_MAX_CQS ) ];
+	/** Queue pair in-use bitmask */
+	arbel_bitmask_t qp_inuse[ ARBEL_BITMASK_SIZE ( ARBEL_MAX_QPS ) ];
+	
+	/** Device limits */
+	struct arbel_dev_limits limits;
+};
+
+/** Global protection domain */
+#define ARBEL_GLOBAL_PD			0x123456
+
+/** Memory key prefix */
+#define ARBEL_MKEY_PREFIX		0x77000000UL
+
+/*
+ * HCA commands
+ *
+ */
+
+#define ARBEL_HCR_BASE			0x80680
+#define ARBEL_HCR_REG(x)		( ARBEL_HCR_BASE + 4 * (x) )
+#define ARBEL_HCR_MAX_WAIT_MS		2000
+#define ARBEL_MBOX_ALIGN		4096
+#define ARBEL_MBOX_SIZE			512
+
+/* HCA command is split into
+ *
+ * bits  11:0	Opcode
+ * bit     12	Input uses mailbox
+ * bit     13	Output uses mailbox
+ * bits 22:14	Input parameter length (in dwords)
+ * bits 31:23	Output parameter length (in dwords)
+ *
+ * Encoding the information in this way allows us to cut out several
+ * parameters to the arbel_command() call.
+ */
+#define ARBEL_HCR_IN_MBOX		0x00001000UL
+#define ARBEL_HCR_OUT_MBOX		0x00002000UL
+#define ARBEL_HCR_OPCODE( _command )	( (_command) & 0xfff )
+#define ARBEL_HCR_IN_LEN( _command )	( ( (_command) >> 12 ) & 0x7fc )
+#define ARBEL_HCR_OUT_LEN( _command )	( ( (_command) >> 21 ) & 0x7fc )
+
+/** Build HCR command from component parts */
+#define ARBEL_HCR_INOUT_CMD( _opcode, _in_mbox, _in_len,		     \
+			     _out_mbox, _out_len )			     \
+	( (_opcode) |							     \
+	  ( (_in_mbox) ? ARBEL_HCR_IN_MBOX : 0 ) |			     \
+	  ( ( (_in_len) / 4 ) << 14 ) |					     \
+	  ( (_out_mbox) ? ARBEL_HCR_OUT_MBOX : 0 ) |			     \
+	  ( ( (_out_len) / 4 ) << 23 ) )
+
+#define ARBEL_HCR_IN_CMD( _opcode, _in_mbox, _in_len )			     \
+	ARBEL_HCR_INOUT_CMD ( _opcode, _in_mbox, _in_len, 0, 0 )
+
+#define ARBEL_HCR_OUT_CMD( _opcode, _out_mbox, _out_len )		     \
+	ARBEL_HCR_INOUT_CMD ( _opcode, 0, 0, _out_mbox, _out_len )
+
+#define ARBEL_HCR_VOID_CMD( _opcode )					     \
+	ARBEL_HCR_INOUT_CMD ( _opcode, 0, 0, 0, 0 )
+
+/*
+ * Doorbell record allocation
+ *
+ * The doorbell record map looks like:
+ *
+ *    ARBEL_MAX_CQS * Arm completion queue doorbell
+ *    ARBEL_MAX_QPS * Send work request doorbell
+ *    Group separator
+ *    ...(empty space)...
+ *    ARBEL_MAX_QPS * Receive work request doorbell
+ *    ARBEL_MAX_CQS * Completion queue consumer counter update doorbell
+ */
+
+#define ARBEL_MAX_DOORBELL_RECORDS 512
+#define ARBEL_GROUP_SEPARATOR_DOORBELL ( ARBEL_MAX_CQS + ARBEL_MAX_QPS )
+
+/**
+ * Get arm completion queue doorbell index
+ *
+ * @v cqn_offset	Completion queue number offset
+ * @ret doorbell_idx	Doorbell index
+ */
+static inline unsigned int
+arbel_cq_arm_doorbell_idx ( unsigned int cqn_offset ) {
+	return cqn_offset;
+}
+
+/**
+ * Get send work request doorbell index
+ *
+ * @v qpn_offset	Queue pair number offset
+ * @ret doorbell_idx	Doorbell index
+ */
+static inline unsigned int
+arbel_send_doorbell_idx ( unsigned int qpn_offset ) {
+	return ( ARBEL_MAX_CQS + qpn_offset );
+}
+
+/**
+ * Get receive work request doorbell index
+ *
+ * @v qpn_offset	Queue pair number offset
+ * @ret doorbell_idx	Doorbell index
+ */
+static inline unsigned int
+arbel_recv_doorbell_idx ( unsigned int qpn_offset ) {
+	return ( ARBEL_MAX_DOORBELL_RECORDS - ARBEL_MAX_CQS - qpn_offset - 1 );
+}
+
+/**
+ * Get completion queue consumer counter doorbell index
+ *
+ * @v cqn_offset	Completion queue number offset
+ * @ret doorbell_idx	Doorbell index
+ */
+static inline unsigned int
+arbel_cq_ci_doorbell_idx ( unsigned int cqn_offset ) {
+	return ( ARBEL_MAX_DOORBELL_RECORDS - cqn_offset - 1 );
+}
+
+#endif /* _ARBEL_H */

src/drivers/infiniband/mlx_bitops.h

@@ -0,0 +1,209 @@
+#ifndef _MLX_BITOPS_H
+#define _MLX_BITOPS_H
+
+/*
+ * Copyright (C) 2007 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+/**
+ * @file
+ *
+ * Mellanox bit operations
+ *
+ */
+
+/* Datatype used to represent a bit in the Mellanox autogenerated headers */
+typedef unsigned char pseudo_bit_t;
+
+/**
+ * Wrapper structure for pseudo_bit_t structures
+ *
+ * This structure provides a wrapper around the autogenerated
+ * pseudo_bit_t structures.  It has the correct size, and also
+ * encapsulates type information about the underlying pseudo_bit_t
+ * structure, which allows the MLX_FILL etc. macros to work without
+ * requiring explicit type information.
+ */
+#define MLX_DECLARE_STRUCT( _structure )				     \
+	_structure {							     \
+	    union {							     \
+		uint8_t bytes[ sizeof ( struct _structure ## _st ) / 8 ];    \
+		uint32_t dwords[ sizeof ( struct _structure ## _st ) / 32 ]; \
+		struct _structure ## _st *dummy[0];			     \
+	    } u;							     \
+	}
+
+/** Get pseudo_bit_t structure type from wrapper structure pointer */
+#define MLX_PSEUDO_STRUCT( _ptr )					     \
+	typeof ( *((_ptr)->u.dummy[0]) )
+
+/** Bit offset of a field within a pseudo_bit_t structure */
+#define MLX_BIT_OFFSET( _structure_st, _field )				     \
+	offsetof ( _structure_st, _field )
+
+/** Dword offset of a field within a pseudo_bit_t structure */
+#define MLX_DWORD_OFFSET( _structure_st, _field )			     \
+	( MLX_BIT_OFFSET ( _structure_st, _field ) / 32 )
+
+/** Dword bit offset of a field within a pseudo_bit_t structure
+ *
+ * Yes, using mod-32 would work, but would lose the check for the
+ * error of specifying a mismatched field name and dword index.
+ */
+#define MLX_DWORD_BIT_OFFSET( _structure_st, _index, _field )		     \
+	( MLX_BIT_OFFSET ( _structure_st, _field ) - ( 32 * (_index) ) )
+
+/** Bit width of a field within a pseudo_bit_t structure */
+#define MLX_BIT_WIDTH( _structure_st, _field )				     \
+	sizeof ( ( ( _structure_st * ) NULL )->_field )
+
+/** Bit mask for a field within a pseudo_bit_t structure */
+#define MLX_BIT_MASK( _structure_st, _field )				     \
+	( ( ~( ( uint32_t ) 0 ) ) >>					     \
+	  ( 32 - MLX_BIT_WIDTH ( _structure_st, _field ) ) )
+
+/*
+ * Assemble native-endian dword from named fields and values
+ *
+ */
+
+#define MLX_ASSEMBLE_1( _structure_st, _index, _field, _value )		     \
+	( (_value) << MLX_DWORD_BIT_OFFSET ( _structure_st, _index, _field ) )
+
+#define MLX_ASSEMBLE_2( _structure_st, _index, _field, _value, ... )	     \
+	( MLX_ASSEMBLE_1 ( _structure_st, _index, _field, _value ) |	     \
+	  MLX_ASSEMBLE_1 ( _structure_st, _index, __VA_ARGS__ ) )
+
+#define MLX_ASSEMBLE_3( _structure_st, _index, _field, _value, ... )	     \
+	( MLX_ASSEMBLE_1 ( _structure_st, _index, _field, _value ) |	     \
+	  MLX_ASSEMBLE_2 ( _structure_st, _index, __VA_ARGS__ ) )
+
+#define MLX_ASSEMBLE_4( _structure_st, _index, _field, _value, ... )	     \
+	( MLX_ASSEMBLE_1 ( _structure_st, _index, _field, _value ) |	     \
+	  MLX_ASSEMBLE_3 ( _structure_st, _index, __VA_ARGS__ ) )
+
+#define MLX_ASSEMBLE_5( _structure_st, _index, _field, _value, ... )	     \
+	( MLX_ASSEMBLE_1 ( _structure_st, _index, _field, _value ) |	     \
+	  MLX_ASSEMBLE_4 ( _structure_st, _index, __VA_ARGS__ ) )
+
+#define MLX_ASSEMBLE_6( _structure_st, _index, _field, _value, ... )	     \
+	( MLX_ASSEMBLE_1 ( _structure_st, _index, _field, _value ) |	     \
+	  MLX_ASSEMBLE_5 ( _structure_st, _index, __VA_ARGS__ ) )
+
+/*
+ * Build native-endian (positive) dword bitmasks from named fields
+ *
+ */
+
+#define MLX_MASK_1( _structure_st, _index, _field )			     \
+	( MLX_BIT_MASK ( _structure_st, _field ) <<			     \
+	  MLX_DWORD_BIT_OFFSET ( _structure_st, _index, _field ) )
+
+#define MLX_MASK_2( _structure_st, _index, _field, ... )		     \
+	( MLX_MASK_1 ( _structure_st, _index, _field ) |		     \
+	  MLX_MASK_1 ( _structure_st, _index, __VA_ARGS__ ) )
+
+#define MLX_MASK_3( _structure_st, _index, _field, ... )		     \
+	( MLX_MASK_1 ( _structure_st, _index, _field ) |		     \
+	  MLX_MASK_2 ( _structure_st, _index, __VA_ARGS__ ) )
+
+#define MLX_MASK_4( _structure_st, _index, _field, ... )		     \
+	( MLX_MASK_1 ( _structure_st, _index, _field ) |		     \
+	  MLX_MASK_3 ( _structure_st, _index, __VA_ARGS__ ) )
+
+#define MLX_MASK_5( _structure_st, _index, _field, ... )		     \
+	( MLX_MASK_1 ( _structure_st, _index, _field ) |		     \
+	  MLX_MASK_4 ( _structure_st, _index, __VA_ARGS__ ) )
+
+#define MLX_MASK_6( _structure_st, _index, _field, ... )		     \
+	( MLX_MASK_1 ( _structure_st, _index, _field ) |		     \
+	  MLX_MASK_5 ( _structure_st, _index, __VA_ARGS__ ) )
+
+/*
+ * Populate big-endian dwords from named fields and values
+ *
+ */
+
+#define MLX_FILL( _ptr, _index, _assembled )				     \
+	do {								     \
+		uint32_t *__ptr = &(_ptr)->u.dwords[(_index)];		     \
+		uint32_t __assembled = (_assembled);			     \
+		*__ptr = cpu_to_be32 ( __assembled );			     \
+	} while ( 0 )
+
+#define MLX_FILL_1( _ptr, _index, ... )					     \
+	MLX_FILL ( _ptr, _index, MLX_ASSEMBLE_1 ( MLX_PSEUDO_STRUCT ( _ptr ),\
+						  _index, __VA_ARGS__ ) )
+
+#define MLX_FILL_2( _ptr, _index, ... )					     \
+	MLX_FILL ( _ptr, _index, MLX_ASSEMBLE_2 ( MLX_PSEUDO_STRUCT ( _ptr ),\
+						  _index, __VA_ARGS__ ) )
+
+#define MLX_FILL_3( _ptr, _index, ... )					     \
+	MLX_FILL ( _ptr, _index, MLX_ASSEMBLE_3 ( MLX_PSEUDO_STRUCT ( _ptr ),\
+						  _index, __VA_ARGS__ ) )
+
+#define MLX_FILL_4( _ptr, _index, ... )					     \
+	MLX_FILL ( _ptr, _index, MLX_ASSEMBLE_4 ( MLX_PSEUDO_STRUCT ( _ptr ),\
+						  _index, __VA_ARGS__ ) )
+
+#define MLX_FILL_5( _ptr, _index, ... )					     \
+	MLX_FILL ( _ptr, _index, MLX_ASSEMBLE_5 ( MLX_PSEUDO_STRUCT ( _ptr ),\
+						  _index, __VA_ARGS__ ) )
+
+#define MLX_FILL_6( _ptr, _index, ... )					     \
+	MLX_FILL ( _ptr, _index, MLX_ASSEMBLE_6 ( MLX_PSEUDO_STRUCT ( _ptr ),\
+						  _index, __VA_ARGS__ ) )
+
+/*
+ * Modify big-endian dword using named field and value
+ *
+ */
+
+#define MLX_SET( _ptr, _field, _value )					     \
+	do {								     \
+		unsigned int __index = 					     \
+		    MLX_DWORD_OFFSET ( MLX_PSEUDO_STRUCT ( _ptr ), _field ); \
+		uint32_t *__ptr = &(_ptr)->u.dwords[__index];		     \
+		uint32_t __value = be32_to_cpu ( *__ptr );		     \
+		__value &= ~( MLX_MASK_1 ( MLX_PSEUDO_STRUCT ( _ptr ),	     \
+					   __index, _field ) );		     \
+		__value |= MLX_ASSEMBLE_1 ( MLX_PSEUDO_STRUCT ( _ptr ),	     \
+					    __index, _field, _value );	     \
+		*__ptr = cpu_to_be32 ( __value );			     \
+	} while ( 0 )
+
+/*
+ * Extract value of named field
+ *
+ */
+
+#define MLX_GET( _ptr, _field )						     \
+	( {								     \
+		unsigned int __index = 					     \
+		    MLX_DWORD_OFFSET ( MLX_PSEUDO_STRUCT ( _ptr ), _field ); \
+		uint32_t *__ptr = &(_ptr)->u.dwords[__index];		     \
+		uint32_t __value = be32_to_cpu ( *__ptr );		     \
+		__value >>=						     \
+		    MLX_DWORD_BIT_OFFSET ( MLX_PSEUDO_STRUCT ( _ptr ),	     \
+					    __index, _field );		     \
+		__value &=						     \
+		    MLX_BIT_MASK ( MLX_PSEUDO_STRUCT ( _ptr ), _field );     \
+		__value;						     \
+	} )
+
+#endif /* _MLX_BITOPS_H */

src/drivers/net/ipoib.c

@@ -0,0 +1,930 @@
+/*
+ * Copyright (C) 2007 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+#include <stdint.h>
+#include <stdio.h>
+#include <unistd.h>
+#include <string.h>
+#include <byteswap.h>
+#include <errno.h>
+#include "timer.h"
+#include <gpxe/if_arp.h>
+#include <gpxe/iobuf.h>
+#include <gpxe/netdevice.h>
+#include <gpxe/infiniband.h>
+#include <gpxe/ipoib.h>
+
+/** @file
+ *
+ * IP over Infiniband
+ */
+
+/** IPoIB MTU */
+#define IPOIB_MTU 2048
+
+/** Number of IPoIB data send work queue entries */
+#define IPOIB_DATA_NUM_SEND_WQES 2
+
+/** Number of IPoIB data receive work queue entries */
+#define IPOIB_DATA_NUM_RECV_WQES 4
+
+/** Number of IPoIB data completion entries */
+#define IPOIB_DATA_NUM_CQES 8
+
+/** Number of IPoIB metadata send work queue entries */
+#define IPOIB_META_NUM_SEND_WQES 2
+
+/** Number of IPoIB metadata receive work queue entries */
+#define IPOIB_META_NUM_RECV_WQES 2
+
+/** Number of IPoIB metadata completion entries */
+#define IPOIB_META_NUM_CQES 8
+
+/** An IPoIB queue set */
+struct ipoib_queue_set {
+	/** Completion queue */
+	struct ib_completion_queue *cq;
+	/** Queue pair */
+	struct ib_queue_pair *qp;
+	/** Receive work queue fill level */
+	unsigned int recv_fill;
+	/** Receive work queue maximum fill level */
+	unsigned int recv_max_fill;
+};
+
+/** An IPoIB device */
+struct ipoib_device {
+	/** Network device */
+	struct net_device *netdev;
+	/** Underlying Infiniband device */
+	struct ib_device *ibdev;
+	/** Data queue set */
+	struct ipoib_queue_set data;
+	/** Data queue set */
+	struct ipoib_queue_set meta;
+	/** Broadcast GID */
+	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;
+};
+
+/**
+ * IPoIB path cache entry
+ *
+ * This serves a similar role to the ARP cache for Ethernet.  (ARP
+ * *is* used on IPoIB; we have two caches to maintain.)
+ */
+struct ipoib_cached_path {
+	/** Destination GID */
+	struct ib_gid gid;
+	/** Destination LID */
+	unsigned int dlid;
+	/** Service level */
+	unsigned int sl;
+	/** Rate */
+	unsigned int rate;
+};
+
+/** Number of IPoIB path cache entries */
+#define IPOIB_NUM_CACHED_PATHS 2
+
+/** IPoIB path cache */
+static struct ipoib_cached_path ipoib_path_cache[IPOIB_NUM_CACHED_PATHS];
+
+/** Oldest IPoIB path cache entry index */
+static unsigned int ipoib_path_cache_idx = 0;
+
+/** TID half used to identify get path record replies */
+#define IPOIB_TID_GET_PATH_REC 0x11111111UL
+
+/** TID half used to identify multicast member record replies */
+#define IPOIB_TID_MC_MEMBER_REC 0x22222222UL
+
+/** IPoIB metadata TID */
+static uint32_t ipoib_meta_tid = 0;
+
+/** IPv4 broadcast GID */
+static const struct ib_gid ipv4_broadcast_gid = {
+	{ { 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
+	    0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff } }
+};
+
+/** Maximum time we will wait for the broadcast join to succeed */
+#define IPOIB_JOIN_MAX_DELAY_MS 1000
+
+/****************************************************************************
+ *
+ * IPoIB link layer
+ *
+ ****************************************************************************
+ */
+
+/** Broadcast QPN used in IPoIB MAC addresses
+ *
+ * This is a guaranteed invalid real QPN
+ */
+#define IPOIB_BROADCAST_QPN 0xffffffffUL
+
+/** Broadcast IPoIB address */
+static struct ipoib_mac ipoib_broadcast = {
+	.qpn = ntohl ( IPOIB_BROADCAST_QPN ),
+};
+
+/**
+ * Transmit IPoIB packet
+ *
+ * @v iobuf		I/O buffer
+ * @v netdev		Network device
+ * @v net_protocol	Network-layer protocol
+ * @v ll_dest		Link-layer destination address
+ *
+ * Prepends the IPoIB link-layer header and transmits the packet.
+ */
+static int ipoib_tx ( struct io_buffer *iobuf, struct net_device *netdev,
+		      struct net_protocol *net_protocol,
+		      const void *ll_dest ) {
+	struct ipoib_hdr *ipoib_hdr =
+		iob_push ( iobuf, sizeof ( *ipoib_hdr ) );
+
+	/* Build IPoIB header */
+	memcpy ( &ipoib_hdr->pseudo.peer, ll_dest,
+		 sizeof ( ipoib_hdr->pseudo.peer ) );
+	ipoib_hdr->real.proto = net_protocol->net_proto;
+	ipoib_hdr->real.reserved = 0;
+
+	/* Hand off to network device */
+	return netdev_tx ( netdev, iobuf );
+}
+
+/**
+ * Process received IPoIB packet
+ *
+ * @v iobuf	I/O buffer
+ * @v netdev	Network device
+ *
+ * Strips off the IPoIB link-layer header and passes up to the
+ * network-layer protocol.
+ */
+static int ipoib_rx ( struct io_buffer *iobuf, struct net_device *netdev ) {
+	struct ipoib_hdr *ipoib_hdr = iobuf->data;
+
+	/* Sanity check */
+	if ( iob_len ( iobuf ) < sizeof ( *ipoib_hdr ) ) {
+		DBG ( "IPoIB packet too short for link-layer header\n" );
+		DBG_HD ( iobuf->data, iob_len ( iobuf ) );
+		free_iob ( iobuf );
+		return -EINVAL;
+	}
+
+	/* Strip off IPoIB header */
+	iob_pull ( iobuf, sizeof ( *ipoib_hdr ) );
+
+	/* Hand off to network-layer protocol */
+	return net_rx ( iobuf, netdev, ipoib_hdr->real.proto,
+			&ipoib_hdr->pseudo.peer );
+}
+
+/**
+ * Transcribe IPoIB address
+ *
+ * @v ll_addr	Link-layer address
+ * @ret string	Link-layer address in human-readable format
+ */
+const char * ipoib_ntoa ( const void *ll_addr ) {
+	static char buf[45];
+	const struct ipoib_mac *mac = ll_addr;
+
+	snprintf ( buf, sizeof ( buf ), "%08lx:%08lx:%08lx:%08lx:%08lx",
+		   htonl ( mac->qpn ), htonl ( mac->gid.u.dwords[0] ),
+		   htonl ( mac->gid.u.dwords[1] ),
+		   htonl ( mac->gid.u.dwords[2] ),
+		   htonl ( mac->gid.u.dwords[3] ) );
+	return buf;
+}
+
+/** IPoIB protocol */
+struct ll_protocol ipoib_protocol __ll_protocol = {
+	.name		= "IPoIB",
+	.ll_proto	= htons ( ARPHRD_INFINIBAND ),
+	.ll_addr_len	= IPOIB_ALEN,
+	.ll_header_len	= IPOIB_HLEN,
+	.ll_broadcast	= ( uint8_t * ) &ipoib_broadcast,
+	.tx		= ipoib_tx,
+	.rx		= ipoib_rx,
+	.ntoa		= ipoib_ntoa,
+};
+
+/****************************************************************************
+ *
+ * IPoIB network device
+ *
+ ****************************************************************************
+ */
+
+/**
+ * Destroy queue set
+ *
+ * @v ipoib		IPoIB device
+ * @v qset		Queue set
+ */
+static void ipoib_destroy_qset ( struct ipoib_device *ipoib,
+				 struct ipoib_queue_set *qset ) {
+	struct ib_device *ibdev = ipoib->ibdev;
+
+	if ( qset->qp )
+		ib_destroy_qp ( ibdev, qset->qp );
+	if ( qset->cq )
+		ib_destroy_cq ( ibdev, qset->cq );
+	memset ( qset, 0, sizeof ( *qset ) );
+}
+
+/**
+ * Create queue set
+ *
+ * @v ipoib		IPoIB device
+ * @v qset		Queue set
+ * @ret rc		Return status code
+ */
+static int ipoib_create_qset ( struct ipoib_device *ipoib,
+			       struct ipoib_queue_set *qset,
+			       unsigned int num_cqes,
+			       unsigned int num_send_wqes,
+			       unsigned int num_recv_wqes,
+			       unsigned long qkey ) {
+	struct ib_device *ibdev = ipoib->ibdev;
+	int rc;
+
+	/* Store queue parameters */
+	qset->recv_max_fill = num_recv_wqes;
+
+	/* Allocate completion queue */
+	qset->cq = ib_create_cq ( ibdev, num_cqes );
+	if ( ! qset->cq ) {
+		DBGC ( ipoib, "IPoIB %p could not allocate completion queue\n",
+		       ipoib );
+		rc = -ENOMEM;
+		goto err;
+	}
+
+	/* Allocate queue pair */
+	qset->qp = ib_create_qp ( ibdev, num_send_wqes, qset->cq,
+				  num_recv_wqes, qset->cq, qkey );
+	if ( ! qset->qp ) {
+		DBGC ( ipoib, "IPoIB %p could not allocate queue pair\n",
+		       ipoib );
+		rc = -ENOMEM;
+		goto err;
+	}
+	qset->qp->owner_priv = ipoib->netdev;
+
+	return 0;
+
+ err:
+	ipoib_destroy_qset ( ipoib, qset );
+	return rc;
+}
+
+/**
+ * Find path cache entry by GID
+ *
+ * @v gid		GID
+ * @ret entry		Path cache entry, or NULL
+ */
+static struct ipoib_cached_path *
+ipoib_find_cached_path ( struct ib_gid *gid ) {
+	struct ipoib_cached_path *path;
+	unsigned int i;
+
+	for ( i = 0 ; i < IPOIB_NUM_CACHED_PATHS ; i++ ) {
+		path = &ipoib_path_cache[i];
+		if ( memcmp ( &path->gid, gid, sizeof ( *gid ) ) == 0 )
+			return path;
+	}
+	DBG ( "IPoIB %08lx:%08lx:%08lx:%08lx cache miss\n",
+	      htonl ( gid->u.dwords[0] ), htonl ( gid->u.dwords[1] ),
+	      htonl ( gid->u.dwords[2] ), htonl ( gid->u.dwords[3] ) );
+	return NULL;
+}
+
+/**
+ * Transmit path record request
+ *
+ * @v ipoib		IPoIB device
+ * @v gid		Destination GID
+ * @ret rc		Return status code
+ */
+static int ipoib_get_path_record ( struct ipoib_device *ipoib,
+				   struct ib_gid *gid ) {
+	struct ib_device *ibdev = ipoib->ibdev;
+	struct io_buffer *iobuf;
+	struct ib_mad_path_record *path_record;
+	struct ib_address_vector av;
+	int rc;
+
+	/* Allocate I/O buffer */
+	iobuf = alloc_iob ( sizeof ( *path_record ) );
+	if ( ! iobuf )
+		return -ENOMEM;
+	iob_put ( iobuf, sizeof ( *path_record ) );
+	path_record = iobuf->data;
+	memset ( path_record, 0, sizeof ( *path_record ) );
+
+	/* Construct path record request */
+	path_record->mad_hdr.base_version = IB_MGMT_BASE_VERSION;
+	path_record->mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
+	path_record->mad_hdr.class_version = 2;
+	path_record->mad_hdr.method = IB_MGMT_METHOD_GET;
+	path_record->mad_hdr.attr_id = htons ( IB_SA_ATTR_PATH_REC );
+	path_record->mad_hdr.tid[0] = IPOIB_TID_GET_PATH_REC;
+	path_record->mad_hdr.tid[1] = ipoib_meta_tid++;
+	path_record->sa_hdr.comp_mask[1] =
+		htonl ( IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID );
+	memcpy ( &path_record->dgid, gid, sizeof ( path_record->dgid ) );
+	memcpy ( &path_record->sgid, &ibdev->port_gid,
+		 sizeof ( path_record->sgid ) );
+
+	/* Construct address vector */
+	memset ( &av, 0, sizeof ( av ) );
+	av.dlid = ibdev->sm_lid;
+	av.dest_qp = IB_SA_QPN;
+	av.qkey = IB_GLOBAL_QKEY;
+
+	/* Post send request */
+	if ( ( rc = ib_post_send ( ibdev, ipoib->meta.qp, &av,
+				   iobuf ) ) != 0 ) {
+		DBGC ( ipoib, "IPoIB %p could not send get path record: %s\n",
+		       ipoib, strerror ( rc ) );
+		free_iob ( iobuf );
+		return rc;
+	}
+
+	return 0;
+}
+
+/**
+ * Transmit multicast group membership request
+ *
+ * @v ipoib		IPoIB device
+ * @v gid		Multicast GID
+ * @v join		Join (rather than leave) group
+ * @ret rc		Return status code
+ */
+static int ipoib_mc_member_record ( struct ipoib_device *ipoib,
+				    struct ib_gid *gid, int join ) {
+	struct ib_device *ibdev = ipoib->ibdev;
+	struct io_buffer *iobuf;
+	struct ib_mad_mc_member_record *mc_member_record;
+	struct ib_address_vector av;
+	int rc;
+
+	/* Allocate I/O buffer */
+	iobuf = alloc_iob ( sizeof ( *mc_member_record ) );
+	if ( ! iobuf )
+		return -ENOMEM;
+	iob_put ( iobuf, sizeof ( *mc_member_record ) );
+	mc_member_record = iobuf->data;
+	memset ( mc_member_record, 0, sizeof ( *mc_member_record ) );
+
+	/* Construct path record request */
+	mc_member_record->mad_hdr.base_version = IB_MGMT_BASE_VERSION;
+	mc_member_record->mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
+	mc_member_record->mad_hdr.class_version = 2;
+	mc_member_record->mad_hdr.method = 
+		( join ? IB_MGMT_METHOD_SET : IB_MGMT_METHOD_DELETE );
+	mc_member_record->mad_hdr.attr_id = htons ( IB_SA_ATTR_MC_MEMBER_REC );
+	mc_member_record->mad_hdr.tid[0] = IPOIB_TID_MC_MEMBER_REC;
+	mc_member_record->mad_hdr.tid[1] = ipoib_meta_tid++;
+	mc_member_record->sa_hdr.comp_mask[1] =
+		htonl ( IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
+			IB_SA_MCMEMBER_REC_JOIN_STATE );
+	mc_member_record->scope__join_state = 1;
+	memcpy ( &mc_member_record->mgid, gid,
+		 sizeof ( mc_member_record->mgid ) );
+	memcpy ( &mc_member_record->port_gid, &ibdev->port_gid,
+		 sizeof ( mc_member_record->port_gid ) );
+
+	/* Construct address vector */
+	memset ( &av, 0, sizeof ( av ) );
+	av.dlid = ibdev->sm_lid;
+	av.dest_qp = IB_SA_QPN;
+	av.qkey = IB_GLOBAL_QKEY;
+
+	/* Post send request */
+	if ( ( rc = ib_post_send ( ibdev, ipoib->meta.qp, &av,
+				   iobuf ) ) != 0 ) {
+		DBGC ( ipoib, "IPoIB %p could not send get path record: %s\n",
+		       ipoib, strerror ( rc ) );
+		free_iob ( iobuf );
+		return rc;
+	}
+
+	return 0;
+}
+
+/**
+ * Transmit packet via IPoIB network device
+ *
+ * @v netdev		Network device
+ * @v iobuf		I/O buffer
+ * @ret rc		Return status code
+ */
+static int ipoib_transmit ( struct net_device *netdev,
+			    struct io_buffer *iobuf ) {
+	struct ipoib_device *ipoib = netdev->priv;
+	struct ib_device *ibdev = ipoib->ibdev;
+	struct ipoib_pseudo_hdr *ipoib_pshdr = iobuf->data;
+	struct ib_address_vector av;
+	struct ib_gid *gid;
+	struct ipoib_cached_path *path;
+	int rc;
+
+	/* Sanity check */
+	if ( iob_len ( iobuf ) < sizeof ( *ipoib_pshdr ) ) {
+		DBGC ( ipoib, "IPoIB %p buffer too short\n", ipoib );
+		return -EINVAL;
+	}
+	iob_pull ( iobuf, ( sizeof ( *ipoib_pshdr ) ) );
+
+	/* Construct address vector */
+	memset ( &av, 0, sizeof ( av ) );
+	av.qkey = IB_GLOBAL_QKEY;
+	av.gid_present = 1;
+	if ( ipoib_pshdr->peer.qpn == htonl ( IPOIB_BROADCAST_QPN ) ) {
+		/* Broadcast address */
+		av.dest_qp = IB_BROADCAST_QPN;
+		av.dlid = ipoib->broadcast_lid;
+		gid = &ipoib->broadcast_gid;
+	} else {
+		/* Unicast - look in path cache */
+		path = ipoib_find_cached_path ( &ipoib_pshdr->peer.gid );
+		if ( ! path ) {
+			/* No path entry - get path record */
+			rc = ipoib_get_path_record ( ipoib,
+						     &ipoib_pshdr->peer.gid );
+			netdev_tx_complete ( netdev, iobuf );
+			return rc;
+		}
+		av.dest_qp = ntohl ( ipoib_pshdr->peer.qpn );
+		av.dlid = path->dlid;
+		av.rate = path->rate;
+		av.sl = path->sl;
+		gid = &ipoib_pshdr->peer.gid;
+	}
+	memcpy ( &av.gid, gid, sizeof ( av.gid ) );
+
+	return ib_post_send ( ibdev, ipoib->data.qp, &av, iobuf );
+}
+
+/**
+ * Handle IPoIB data send completion
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v completion	Completion
+ * @v iobuf		I/O buffer
+ */
+static void ipoib_data_complete_send ( struct ib_device *ibdev __unused,
+				       struct ib_queue_pair *qp,
+				       struct ib_completion *completion,
+				       struct io_buffer *iobuf ) {
+	struct net_device *netdev = qp->owner_priv;
+
+	netdev_tx_complete_err ( netdev, iobuf,
+				 ( completion->syndrome ? -EIO : 0 ) );
+}
+
+/**
+ * Handle IPoIB data receive completion
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v completion	Completion
+ * @v iobuf		I/O buffer
+ */
+static void ipoib_data_complete_recv ( struct ib_device *ibdev __unused,
+				       struct ib_queue_pair *qp,
+				       struct ib_completion *completion,
+				       struct io_buffer *iobuf ) {
+	struct net_device *netdev = qp->owner_priv;
+	struct ipoib_device *ipoib = netdev->priv;
+	struct ipoib_pseudo_hdr *ipoib_pshdr;
+
+	if ( completion->syndrome ) {
+		netdev_rx_err ( netdev, iobuf, -EIO );
+		goto done;
+	}
+
+	iob_put ( iobuf, completion->len );
+	if ( iob_len ( iobuf ) < sizeof ( struct ib_global_route_header ) ) {
+		DBGC ( ipoib, "IPoIB %p received data packet too short to "
+		       "contain GRH\n", ipoib );
+		DBGC_HD ( ipoib, iobuf->data, iob_len ( iobuf ) );
+		netdev_rx_err ( netdev, iobuf, -EIO );
+		goto done;
+	}
+	iob_pull ( iobuf, sizeof ( struct ib_global_route_header ) );
+
+	if ( iob_len ( iobuf ) < sizeof ( struct ipoib_real_hdr ) ) {
+		DBGC ( ipoib, "IPoIB %p received data packet too short to "
+		       "contain IPoIB header\n", ipoib );
+		DBGC_HD ( ipoib, iobuf->data, iob_len ( iobuf ) );
+		netdev_rx_err ( netdev, iobuf, -EIO );
+		goto done;
+	}
+
+	ipoib_pshdr = iob_push ( iobuf, sizeof ( *ipoib_pshdr ) );
+	/* FIXME: fill in a MAC address for the sake of AoE! */
+
+	netdev_rx ( netdev, iobuf );
+
+ done:
+	ipoib->data.recv_fill--;
+}
+
+/**
+ * Handle IPoIB metadata send completion
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v completion	Completion
+ * @v iobuf		I/O buffer
+ */
+static void ipoib_meta_complete_send ( struct ib_device *ibdev __unused,
+				       struct ib_queue_pair *qp,
+				       struct ib_completion *completion,
+				       struct io_buffer *iobuf ) {
+	struct net_device *netdev = qp->owner_priv;
+	struct ipoib_device *ipoib = netdev->priv;
+
+	if ( completion->syndrome ) {
+		DBGC ( ipoib, "IPoIB %p metadata TX completion error %x\n",
+		       ipoib, completion->syndrome );
+	}
+	free_iob ( iobuf );
+}
+
+/**
+ * Handle received IPoIB path record
+ *
+ * @v ipoib		IPoIB device
+ * @v path_record	Path record
+ */
+static void ipoib_recv_path_record ( struct ipoib_device *ipoib __unused,
+				     struct ib_mad_path_record *path_record ) {
+	struct ipoib_cached_path *path;
+
+	/* Update path cache entry */
+	path = &ipoib_path_cache[ipoib_path_cache_idx];
+	memcpy ( &path->gid, &path_record->dgid, sizeof ( path->gid ) );
+	path->dlid = ntohs ( path_record->dlid );
+	path->sl = ( path_record->reserved__sl & 0x0f );
+	path->rate = ( path_record->rate_selector__rate & 0x3f );
+
+	DBG ( "IPoIB %08lx:%08lx:%08lx:%08lx dlid %x sl %x rate %x\n",
+	      htonl ( path->gid.u.dwords[0] ), htonl ( path->gid.u.dwords[1] ),
+	      htonl ( path->gid.u.dwords[2] ), htonl ( path->gid.u.dwords[3] ),
+	      path->dlid, path->sl, path->rate );
+	
+	/* Update path cache index */
+	ipoib_path_cache_idx++;
+	if ( ipoib_path_cache_idx == IPOIB_NUM_CACHED_PATHS )
+		ipoib_path_cache_idx = 0;
+}
+
+/**
+ * Handle received IPoIB multicast membership record
+ *
+ * @v ipoib		IPoIB device
+ * @v mc_member_record	Multicast membership record
+ */
+static void ipoib_recv_mc_member_record ( struct ipoib_device *ipoib,
+			  struct ib_mad_mc_member_record *mc_member_record ) {
+	/* Record parameters */
+	ipoib->broadcast_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 );
+}
+
+/**
+ * Handle IPoIB metadata receive completion
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v completion	Completion
+ * @v iobuf		I/O buffer
+ */
+static void ipoib_meta_complete_recv ( struct ib_device *ibdev __unused,
+				       struct ib_queue_pair *qp,
+				       struct ib_completion *completion,
+				       struct io_buffer *iobuf ) {
+	struct net_device *netdev = qp->owner_priv;
+	struct ipoib_device *ipoib = netdev->priv;
+	union ib_mad *mad;
+
+	if ( completion->syndrome ) {
+		DBGC ( ipoib, "IPoIB %p metadata RX completion error %x\n",
+		       ipoib, completion->syndrome );
+		goto done;
+	}
+
+	iob_put ( iobuf, completion->len );
+	if ( iob_len ( iobuf ) < sizeof ( struct ib_global_route_header ) ) {
+		DBGC ( ipoib, "IPoIB %p received metadata packet too short "
+		       "to contain GRH\n", ipoib );
+		DBGC_HD ( ipoib, iobuf->data, iob_len ( iobuf ) );
+		goto done;
+	}
+	iob_pull ( iobuf, sizeof ( struct ib_global_route_header ) );
+	if ( iob_len ( iobuf ) < sizeof ( *mad ) ) {
+		DBGC ( ipoib, "IPoIB %p received metadata packet too short "
+		       "to contain reply\n", ipoib );
+		DBGC_HD ( ipoib, iobuf->data, iob_len ( iobuf ) );
+		goto done;
+	}
+	mad = iobuf->data;
+
+	if ( mad->mad_hdr.status != 0 ) {
+		DBGC ( ipoib, "IPoIB %p metadata RX err status %04x\n",
+		       ipoib, ntohs ( mad->mad_hdr.status ) );
+		goto done;
+	}
+
+	switch ( mad->mad_hdr.tid[0] ) {
+	case IPOIB_TID_GET_PATH_REC:
+		ipoib_recv_path_record ( ipoib, &mad->path_record );
+		break;
+	case IPOIB_TID_MC_MEMBER_REC:
+		ipoib_recv_mc_member_record ( ipoib, &mad->mc_member_record );
+		break;
+	default:
+		DBGC ( ipoib, "IPoIB %p unwanted response:\n",
+		       ipoib );
+		DBGC_HD ( ipoib, mad, sizeof ( *mad ) );
+		break;
+	}
+
+ done:
+	ipoib->meta.recv_fill--;
+	free_iob ( iobuf );
+}
+
+/**
+ * Refill IPoIB receive ring
+ *
+ * @v ipoib		IPoIB device
+ */
+static void ipoib_refill_recv ( struct ipoib_device *ipoib,
+				struct ipoib_queue_set *qset ) {
+	struct ib_device *ibdev = ipoib->ibdev;
+	struct io_buffer *iobuf;
+	int rc;
+
+	while ( qset->recv_fill < qset->recv_max_fill ) {
+		iobuf = alloc_iob ( IPOIB_MTU );
+		if ( ! iobuf )
+			break;
+		if ( ( rc = ib_post_recv ( ibdev, qset->qp, iobuf ) ) != 0 ) {
+			free_iob ( iobuf );
+			break;
+		}
+		qset->recv_fill++;
+	}
+}
+
+/**
+ * Poll IPoIB network device
+ *
+ * @v netdev		Network device
+ */
+static void ipoib_poll ( struct net_device *netdev ) {
+	struct ipoib_device *ipoib = netdev->priv;
+	struct ib_device *ibdev = ipoib->ibdev;
+
+	ib_poll_cq ( ibdev, ipoib->meta.cq, ipoib_meta_complete_send,
+		     ipoib_meta_complete_recv );
+	ib_poll_cq ( ibdev, ipoib->data.cq, ipoib_data_complete_send,
+		     ipoib_data_complete_recv );
+	ipoib_refill_recv ( ipoib, &ipoib->meta );
+	ipoib_refill_recv ( ipoib, &ipoib->data );
+}
+
+/**
+ * Enable/disable interrupts on IPoIB network device
+ *
+ * @v netdev		Network device
+ * @v enable		Interrupts should be enabled
+ */
+static void ipoib_irq ( struct net_device *netdev __unused,
+			int enable __unused ) {
+	/* No implementation */
+}
+
+/**
+ * Open IPoIB network device
+ *
+ * @v netdev		Network device
+ * @ret rc		Return status code
+ */
+static int ipoib_open ( struct net_device *netdev ) {
+	struct ipoib_device *ipoib = netdev->priv;
+	struct ib_device *ibdev = ipoib->ibdev;
+	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;
+	}
+
+	/* Fill receive rings */
+	ipoib_refill_recv ( ipoib, &ipoib->meta );
+	ipoib_refill_recv ( ipoib, &ipoib->data );
+
+	return 0;
+}
+
+/**
+ * Close IPoIB network device
+ *
+ * @v netdev		Network device
+ */
+static void ipoib_close ( struct net_device *netdev ) {
+	struct ipoib_device *ipoib = netdev->priv;
+	struct ib_device *ibdev = ipoib->ibdev;
+
+	/* Detach from broadcast multicast GID */
+	ib_mcast_detach ( ibdev, ipoib->data.qp, &ipoib->broadcast_gid );
+
+	/* FIXME: should probably flush the receive ring */
+}
+
+/** IPoIB network device operations */
+static struct net_device_operations ipoib_operations = {
+	.open		= ipoib_open,
+	.close		= ipoib_close,
+	.transmit	= ipoib_transmit,
+	.poll		= ipoib_poll,
+	.irq		= ipoib_irq,
+};
+
+/**
+ * Join IPoIB broadcast group
+ *
+ * @v ipoib		IPoIB device
+ * @ret rc		Return status code
+ */
+static int ipoib_join_broadcast_group ( struct ipoib_device *ipoib ) {
+	struct ib_device *ibdev = ipoib->ibdev;
+	unsigned int delay_ms;
+	int rc;
+
+	/* Make sure we have some receive descriptors */
+	ipoib_refill_recv ( ipoib, &ipoib->meta );
+
+	/* 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;
+	}
+
+	/* 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.
+	 */
+	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 );
+
+	return -ETIMEDOUT;
+}
+
+/**
+ * Probe IPoIB device
+ *
+ * @v ibdev		Infiniband device
+ * @ret rc		Return status code
+ */
+int ipoib_probe ( struct ib_device *ibdev ) {
+	struct net_device *netdev;
+	struct ipoib_device *ipoib;
+	struct ipoib_mac *mac;
+	int rc;
+
+	/* Allocate network device */
+	netdev = alloc_ipoibdev ( sizeof ( *ipoib ) );
+	if ( ! netdev )
+		return -ENOMEM;
+	netdev_init ( netdev, &ipoib_operations );
+	ipoib = netdev->priv;
+	ib_set_ownerdata ( ibdev, netdev );
+	netdev->dev = ibdev->dev;
+	memset ( ipoib, 0, sizeof ( *ipoib ) );
+	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 ) );
+
+	/* Register network device */
+	if ( ( rc = register_netdev ( netdev ) ) != 0 )
+		goto err_register_netdev;
+
+	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;
+}
+
+/**
+ * Remove IPoIB device
+ *
+ * @v ibdev		Infiniband device
+ */
+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

@@ -0,0 +1,578 @@
+#ifndef _GPXE_INFINIBAND_H
+#define _GPXE_INFINIBAND_H
+
+/** @file
+ *
+ * Infiniband protocol
+ *
+ */
+
+#include <stdint.h>
+#include <gpxe/device.h>
+
+/** Subnet administrator QPN */
+#define IB_SA_QPN 1
+
+/** Broadcast QPN */
+#define IB_BROADCAST_QPN 0xffffffUL
+
+/** Subnet administrator queue key */
+#define IB_GLOBAL_QKEY 0x80010000UL
+
+/** An Infiniband Global Identifier */
+struct ib_gid {
+	union {
+		uint8_t bytes[16];
+		uint16_t words[8];
+		uint32_t dwords[4];
+	} u;
+};
+
+/** An Infiniband Global Route Header */
+struct ib_global_route_header {
+	/** IP version, traffic class, and flow label
+	 *
+	 *  4 bits : Version of the GRH
+	 *  8 bits : Traffic class
+	 * 20 bits : Flow label
+	 */
+	uint32_t ipver_tclass_flowlabel;
+	/** Payload length */
+	uint16_t paylen;
+	/** Next header */
+	uint8_t nxthdr;
+	/** Hop limit */
+	uint8_t hoplmt;
+	/** Source GID */
+	struct ib_gid sgid;
+	/** Destiniation GID */
+	struct ib_gid dgid;
+} __attribute__ (( packed ));
+
+struct ib_device;
+struct ib_queue_pair;
+struct ib_completion_queue;
+
+/** An Infiniband Work Queue */
+struct ib_work_queue {
+	/** Containing queue pair */
+	struct ib_queue_pair *qp;
+	/** "Is a send queue" flag */
+	int is_send;
+	/** Associated completion queue */
+	struct ib_completion_queue *cq;
+	/** List of work queues on this completion queue */
+	struct list_head list;
+	/** Number of work queue entries */
+	unsigned int num_wqes;
+	/** Next work queue entry index
+	 *
+	 * This is the index of the next entry to be filled (i.e. the
+	 * first empty entry).  This value is not bounded by num_wqes;
+	 * users must logical-AND with (num_wqes-1) to generate an
+	 * array index.
+	 */
+	unsigned long next_idx;
+	/** I/O buffers assigned to work queue */
+	struct io_buffer **iobufs;
+	/** Device private data */
+	void *dev_priv;
+};
+
+/** An Infiniband Queue Pair */
+struct ib_queue_pair {
+	/** Queue Pair Number */
+	unsigned long qpn;
+	/** Queue key */
+	unsigned long qkey;
+	/** Send queue */
+	struct ib_work_queue send;
+	/** Receive queue */
+	struct ib_work_queue recv;
+	/** Device private data */
+	void *dev_priv;
+	/** Queue owner private data */
+	void *owner_priv;
+};
+
+/** An Infiniband Completion Queue */
+struct ib_completion_queue {
+	/** Completion queue number */
+	unsigned long cqn;
+	/** Number of completion queue entries */
+	unsigned int num_cqes;
+	/** Next completion queue entry index
+	 *
+	 * This is the index of the next entry to be filled (i.e. the
+	 * first empty entry).  This value is not bounded by num_wqes;
+	 * users must logical-AND with (num_wqes-1) to generate an
+	 * array index.
+	 */
+	unsigned long next_idx;
+	/** List of work queues completing to this queue */
+	struct list_head work_queues;
+	/** Device private data */
+	void *dev_priv;
+};
+
+/** An Infiniband completion */
+struct ib_completion {
+	/** Syndrome
+	 *
+	 * If non-zero, then the completion is in error.
+	 */
+	unsigned int syndrome;
+	/** Length */
+	size_t len;
+};
+
+/** An Infiniband completion handler
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v completion	Completion
+ * @v iobuf		I/O buffer
+ */
+typedef void ( * ib_completer_t ) ( struct ib_device *ibdev,
+				    struct ib_queue_pair *qp,
+				    struct ib_completion *completion,
+				    struct io_buffer *iobuf );
+
+/** An Infiniband Address Vector */
+struct ib_address_vector {
+	/** Destination Queue Pair */
+	unsigned int dest_qp;
+	/** Queue key */
+	unsigned long qkey;
+	/** Destination Local ID */
+	unsigned int dlid;
+	/** Rate */
+	unsigned int rate;
+	/** Service level */
+	unsigned int sl;
+	/** GID is present */
+	unsigned int gid_present;
+	/** GID */
+	struct ib_gid gid;
+};
+
+/**
+ * Infiniband device operations
+ *
+ * These represent a subset of the Infiniband Verbs.
+ */
+struct ib_device_operations {
+	/** Create completion queue
+	 *
+	 * @v ibdev		Infiniband device
+	 * @v cq		Completion queue
+	 * @ret rc		Return status code
+	 */
+	int ( * create_cq ) ( struct ib_device *ibdev,
+			      struct ib_completion_queue *cq );
+	/** Destroy completion queue
+	 *
+	 * @v ibdev		Infiniband device
+	 * @v cq		Completion queue
+	 */
+	void ( * destroy_cq ) ( struct ib_device *ibdev,
+				struct ib_completion_queue *cq );
+	/** Create queue pair
+	 *
+	 * @v ibdev		Infiniband device
+	 * @v qp		Queue pair
+	 * @ret rc		Return status code
+	 */
+	int ( * create_qp ) ( struct ib_device *ibdev,
+			      struct ib_queue_pair *qp );
+	/** Destroy queue pair
+	 *
+	 * @v ibdev		Infiniband device
+	 * @v qp		Queue pair
+	 */
+	void ( * destroy_qp ) ( struct ib_device *ibdev,
+				struct ib_queue_pair *qp );
+	/** Post send work queue entry
+	 *
+	 * @v ibdev		Infiniband device
+	 * @v qp		Queue pair
+	 * @v av		Address vector
+	 * @v iobuf		I/O buffer
+	 * @ret rc		Return status code
+	 *
+	 * If this method returns success, the I/O buffer remains
+	 * owned by the queue pair.  If this method returns failure,
+	 * the I/O buffer is immediately released; the failure is
+	 * interpreted as "failure to enqueue buffer".
+	 */
+	int ( * post_send ) ( struct ib_device *ibdev,
+			      struct ib_queue_pair *qp,
+			      struct ib_address_vector *av,
+			      struct io_buffer *iobuf );
+	/** Post receive work queue entry
+	 *
+	 * @v ibdev		Infiniband device
+	 * @v qp		Queue pair
+	 * @v iobuf		I/O buffer
+	 * @ret rc		Return status code
+	 *
+	 * If this method returns success, the I/O buffer remains
+	 * owned by the queue pair.  If this method returns failure,
+	 * the I/O buffer is immediately released; the failure is
+	 * interpreted as "failure to enqueue buffer".
+	 */
+	int ( * post_recv ) ( struct ib_device *ibdev,
+			      struct ib_queue_pair *qp,
+			      struct io_buffer *iobuf );
+	/** Poll completion queue
+	 *
+	 * @v ibdev		Infiniband device
+	 * @v cq		Completion queue
+	 * @v complete_send	Send completion handler
+	 * @v complete_recv	Receive completion handler
+	 *
+	 * The completion handler takes ownership of the I/O buffer.
+	 */
+	void ( * poll_cq ) ( struct ib_device *ibdev,
+			     struct ib_completion_queue *cq,
+			     ib_completer_t complete_send,
+			     ib_completer_t complete_recv );
+	/** Attach to multicast group
+	 *
+	 * @v ibdev		Infiniband device
+	 * @v qp		Queue pair
+	 * @v gid		Multicast GID
+	 * @ret rc		Return status code
+	 */
+	int ( * mcast_attach ) ( struct ib_device *ibdev,
+				 struct ib_queue_pair *qp,
+				 struct ib_gid *gid );
+	/** Detach from multicast group
+	 *
+	 * @v ibdev		Infiniband device
+	 * @v qp		Queue pair
+	 * @v gid		Multicast GID
+	 */
+	void ( * mcast_detach ) ( struct ib_device *ibdev,
+				  struct ib_queue_pair *qp,
+				  struct ib_gid *gid );
+};
+
+/** An Infiniband device */
+struct ib_device {
+	/** Port GID */
+	struct ib_gid port_gid;
+	/** Subnet manager LID */
+	unsigned long sm_lid;
+	/** Partition key */
+	unsigned int pkey;
+	/** Underlying device */
+	struct device *dev;
+	/** Infiniband operations */
+	struct ib_device_operations *op;
+	/** Device private data */
+	void *dev_priv;
+	/** Owner private data */
+	void *owner_priv;
+};
+
+extern struct ib_completion_queue * ib_create_cq ( struct ib_device *ibdev,
+						   unsigned int num_cqes );
+extern void ib_destroy_cq ( struct ib_device *ibdev,
+			    struct ib_completion_queue *cq );
+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 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,
+					   unsigned long qpn, int is_send );
+extern struct ib_device * alloc_ibdev ( size_t priv_size );
+extern void free_ibdev ( struct ib_device *ibdev );
+
+/**
+ * Post send work queue entry
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v av		Address vector
+ * @v iobuf		I/O buffer
+ * @ret rc		Return status code
+ */
+static inline __attribute__ (( always_inline )) int
+ib_post_send ( struct ib_device *ibdev, struct ib_queue_pair *qp,
+	       struct ib_address_vector *av, struct io_buffer *iobuf ) {
+	return ibdev->op->post_send ( ibdev, qp, av, iobuf );
+}
+
+/**
+ * Post receive work queue entry
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v iobuf		I/O buffer
+ * @ret rc		Return status code
+ */
+static inline __attribute__ (( always_inline )) int
+ib_post_recv ( struct ib_device *ibdev, struct ib_queue_pair *qp,
+	       struct io_buffer *iobuf ) {
+	return ibdev->op->post_recv ( ibdev, qp, iobuf );
+}
+
+/**
+ * Poll completion queue
+ *
+ * @v ibdev		Infiniband device
+ * @v cq		Completion queue
+ * @v complete_send	Send completion handler
+ * @v complete_recv	Receive completion handler
+ */
+static inline __attribute__ (( always_inline )) void
+ib_poll_cq ( struct ib_device *ibdev, struct ib_completion_queue *cq,
+	     ib_completer_t complete_send, ib_completer_t complete_recv ) {
+	ibdev->op->poll_cq ( ibdev, cq, complete_send, complete_recv );
+}
+
+
+/**
+ * Attach to multicast group
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v gid		Multicast GID
+ * @ret rc		Return status code
+ */
+static inline __attribute__ (( always_inline )) int
+ib_mcast_attach ( struct ib_device *ibdev, struct ib_queue_pair *qp,
+		  struct ib_gid *gid ) {
+	return ibdev->op->mcast_attach ( ibdev, qp, gid );
+}
+
+/**
+ * Detach from multicast group
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v gid		Multicast GID
+ */
+static inline __attribute__ (( always_inline )) void
+ib_mcast_detach ( struct ib_device *ibdev, struct ib_queue_pair *qp,
+		  struct ib_gid *gid ) {
+	ibdev->op->mcast_detach ( ibdev, qp, gid );
+}
+
+/**
+ * Set Infiniband owner-private data
+ *
+ * @v pci		Infiniband device
+ * @v priv		Private data
+ */
+static inline void ib_set_ownerdata ( struct ib_device *ibdev,
+				      void *owner_priv ) {
+	ibdev->owner_priv = owner_priv;
+}
+
+/**
+ * Get Infiniband owner-private data
+ *
+ * @v pci		Infiniband device
+ * @ret priv		Private data
+ */
+static inline void * ib_get_ownerdata ( struct ib_device *ibdev ) {
+	return ibdev->owner_priv;
+}
+
+/*****************************************************************************
+ *
+ * Management datagrams
+ *
+ * Portions Copyright (c) 2004 Mellanox Technologies Ltd.  All rights
+ * reserved.
+ *
+ */
+
+/* Management base version */
+#define IB_MGMT_BASE_VERSION			1
+
+/* Management classes */
+#define IB_MGMT_CLASS_SUBN_LID_ROUTED		0x01
+#define IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE	0x81
+#define IB_MGMT_CLASS_SUBN_ADM			0x03
+#define IB_MGMT_CLASS_PERF_MGMT			0x04
+#define IB_MGMT_CLASS_BM			0x05
+#define IB_MGMT_CLASS_DEVICE_MGMT		0x06
+#define IB_MGMT_CLASS_CM			0x07
+#define IB_MGMT_CLASS_SNMP			0x08
+#define IB_MGMT_CLASS_VENDOR_RANGE2_START	0x30
+#define IB_MGMT_CLASS_VENDOR_RANGE2_END		0x4F
+
+/* Management methods */
+#define IB_MGMT_METHOD_GET			0x01
+#define IB_MGMT_METHOD_SET			0x02
+#define IB_MGMT_METHOD_GET_RESP			0x81
+#define IB_MGMT_METHOD_SEND			0x03
+#define IB_MGMT_METHOD_TRAP			0x05
+#define IB_MGMT_METHOD_REPORT			0x06
+#define IB_MGMT_METHOD_REPORT_RESP		0x86
+#define IB_MGMT_METHOD_TRAP_REPRESS		0x07
+#define IB_MGMT_METHOD_DELETE			0x15
+#define IB_MGMT_METHOD_RESP			0x80
+
+/* Subnet management attributes */
+#define IB_SMP_ATTR_NOTICE			0x0002
+#define IB_SMP_ATTR_NODE_DESC			0x0010
+#define IB_SMP_ATTR_NODE_INFO			0x0011
+#define IB_SMP_ATTR_SWITCH_INFO			0x0012
+#define IB_SMP_ATTR_GUID_INFO			0x0014
+#define IB_SMP_ATTR_PORT_INFO			0x0015
+#define IB_SMP_ATTR_PKEY_TABLE			0x0016
+#define IB_SMP_ATTR_SL_TO_VL_TABLE		0x0017
+#define IB_SMP_ATTR_VL_ARB_TABLE		0x0018
+#define IB_SMP_ATTR_LINEAR_FORWARD_TABLE	0x0019
+#define IB_SMP_ATTR_RANDOM_FORWARD_TABLE	0x001A
+#define IB_SMP_ATTR_MCAST_FORWARD_TABLE		0x001B
+#define IB_SMP_ATTR_SM_INFO			0x0020
+#define IB_SMP_ATTR_VENDOR_DIAG			0x0030
+#define IB_SMP_ATTR_LED_INFO			0x0031
+#define IB_SMP_ATTR_VENDOR_MASK			0xFF00
+
+#define IB_SA_ATTR_MC_MEMBER_REC		0x38
+#define IB_SA_ATTR_PATH_REC			0x35
+
+#define IB_SA_MCMEMBER_REC_MGID			(1<<0)
+#define IB_SA_MCMEMBER_REC_PORT_GID		(1<<1)
+#define IB_SA_MCMEMBER_REC_QKEY			(1<<2)
+#define IB_SA_MCMEMBER_REC_MLID			(1<<3)
+#define IB_SA_MCMEMBER_REC_MTU_SELECTOR		(1<<4)
+#define IB_SA_MCMEMBER_REC_MTU			(1<<5)
+#define IB_SA_MCMEMBER_REC_TRAFFIC_CLASS	(1<<6)
+#define IB_SA_MCMEMBER_REC_PKEY			(1<<7)
+#define IB_SA_MCMEMBER_REC_RATE_SELECTOR	(1<<8)
+#define IB_SA_MCMEMBER_REC_RATE			(1<<9)
+#define IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR	(1<<10)
+#define IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME	(1<<11)
+#define IB_SA_MCMEMBER_REC_SL			(1<<12)
+#define IB_SA_MCMEMBER_REC_FLOW_LABEL		(1<<13)
+#define IB_SA_MCMEMBER_REC_HOP_LIMIT		(1<<14)
+#define IB_SA_MCMEMBER_REC_SCOPE		(1<<15)
+#define IB_SA_MCMEMBER_REC_JOIN_STATE		(1<<16)
+#define IB_SA_MCMEMBER_REC_PROXY_JOIN		(1<<17)
+
+#define IB_SA_PATH_REC_DGID			(1<<2)
+#define IB_SA_PATH_REC_SGID			(1<<3)
+
+struct ib_mad_hdr {
+	uint8_t base_version;
+	uint8_t mgmt_class;
+	uint8_t class_version;
+	uint8_t method;
+	uint16_t status;
+	uint16_t class_specific;
+	uint32_t tid[2];
+	uint16_t attr_id;
+	uint16_t resv;
+	uint32_t attr_mod;
+} __attribute__ (( packed ));
+
+struct ib_sa_hdr {
+	uint32_t sm_key[2];
+	uint16_t reserved;
+	uint16_t attrib_offset;
+	uint32_t comp_mask[2];
+} __attribute__ (( packed ));
+
+struct ib_rmpp_hdr {
+	uint32_t raw[3];
+} __attribute__ (( packed ));
+
+struct ib_mad_data {
+	struct ib_mad_hdr mad_hdr;
+	uint8_t data[232];
+} __attribute__ (( packed ));
+
+struct ib_mad_guid_info {
+	struct ib_mad_hdr mad_hdr;
+	uint32_t mkey[2];
+	uint32_t reserved[8];
+	uint8_t gid_local[8];
+} __attribute__ (( packed ));
+
+struct ib_mad_port_info {
+	struct ib_mad_hdr mad_hdr;
+	uint32_t mkey[2];
+	uint32_t reserved[8];
+	uint32_t mkey2[2];
+	uint8_t gid_prefix[8];
+	uint16_t lid;
+	uint16_t mastersm_lid;
+	uint32_t cap_mask;
+	uint16_t diag_code;
+	uint16_t mkey_lease_period;
+	uint8_t local_port_num;
+	uint8_t link_width_enabled;
+	uint8_t link_width_supported;
+	uint8_t link_width_active;
+	uint8_t port_state__link_speed_supported;
+	uint8_t link_down_def_state__port_phys_state;
+	uint8_t lmc__r1__mkey_prot_bits;
+	uint8_t link_speed_enabled__link_speed_active;
+} __attribute__ (( packed ));
+
+struct ib_mad_pkey_table {
+	struct ib_mad_hdr mad_hdr;
+	uint32_t mkey[2];
+	uint32_t reserved[8];
+	uint16_t pkey[16][2];
+} __attribute__ (( packed ));
+
+struct ib_mad_path_record {
+	struct ib_mad_hdr mad_hdr;
+	struct ib_rmpp_hdr rmpp_hdr;
+	struct ib_sa_hdr sa_hdr;
+	uint32_t reserved0[2];
+	struct ib_gid dgid;
+	struct ib_gid sgid;
+	uint16_t dlid;
+	uint16_t slid;
+	uint32_t hop_limit__flow_label__raw_traffic;
+	uint32_t pkey__numb_path__reversible__tclass;
+	uint8_t reserved1;
+	uint8_t reserved__sl;
+	uint8_t mtu_selector__mtu;
+	uint8_t rate_selector__rate;
+	uint32_t preference__packet_lifetime__packet_lifetime_selector;
+	uint32_t reserved2[35];
+} __attribute__ (( packed ));
+
+struct ib_mad_mc_member_record {
+	struct ib_mad_hdr mad_hdr;
+	struct ib_rmpp_hdr rmpp_hdr;
+	struct ib_sa_hdr sa_hdr;
+	struct ib_gid mgid;
+	struct ib_gid port_gid;
+	uint32_t qkey;
+	uint16_t mlid;
+	uint8_t mtu_selector__mtu;
+	uint8_t tclass;
+	uint16_t pkey;
+	uint8_t rate_selector__rate;
+	uint8_t packet_lifetime_selector__packet_lifetime;
+	uint32_t sl__flow_label__hop_limit;
+	uint8_t scope__join_state;
+	uint8_t proxy_join__reserved;
+	uint16_t reserved0;
+	uint32_t reserved1[37];
+} __attribute__ (( packed ));
+
+union ib_mad {
+	struct ib_mad_hdr mad_hdr;
+	struct ib_mad_data data;
+	struct ib_mad_guid_info guid_info;
+	struct ib_mad_port_info port_info;
+	struct ib_mad_pkey_table pkey_table;
+	struct ib_mad_path_record path_record;
+	struct ib_mad_mc_member_record mc_member_record;
+} __attribute__ (( packed ));
+
+#endif /* _GPXE_INFINIBAND_H */

src/include/gpxe/ipoib.h

@@ -0,0 +1,78 @@
+#ifndef _GPXE_IPOIB_H
+#define _GPXE_IPOIB_H
+
+/** @file
+ *
+ * IP over Infiniband
+ */
+
+#include <gpxe/infiniband.h>
+
+/** IPoIB MAC address length */
+#define IPOIB_ALEN 20
+
+/** An IPoIB MAC address */
+struct ipoib_mac {
+	/** Queue pair number
+	 *
+	 * MSB must be zero; QPNs are only 24-bit.
+	 */
+	uint32_t qpn;
+	/** Port GID */
+	struct ib_gid gid;
+} __attribute__ (( packed ));
+
+/** IPoIB link-layer header length */
+#define IPOIB_HLEN 24
+
+/**
+ * IPoIB link-layer header pseudo portion
+ *
+ * This part doesn't actually exist on the wire, but it provides a
+ * convenient way to fit into the typical network device model.
+ */
+struct ipoib_pseudo_hdr {
+	/** Peer address */
+	struct ipoib_mac peer;
+} __attribute__ (( packed ));
+
+/** IPoIB link-layer header real portion */
+struct ipoib_real_hdr {
+	/** Network-layer protocol */
+	uint16_t proto;
+	/** Reserved, must be zero */
+	uint16_t reserved;
+} __attribute__ (( packed ));
+
+/** An IPoIB link-layer header */
+struct ipoib_hdr {
+	/** Pseudo portion */
+	struct ipoib_pseudo_hdr pseudo;
+	/** Real portion */
+	struct ipoib_real_hdr real;
+} __attribute__ (( packed ));
+
+extern struct ll_protocol ipoib_protocol;
+
+extern const char * ipoib_ntoa ( const void *ll_addr );
+
+/**
+ * Allocate IPoIB device
+ *
+ * @v priv_size		Size of driver private data
+ * @ret netdev		Network device, or NULL
+ */
+static inline struct net_device * alloc_ipoibdev ( size_t priv_size ) {
+	struct net_device *netdev;
+
+	netdev = alloc_netdev ( priv_size );
+	if ( netdev ) {
+		netdev->ll_protocol = &ipoib_protocol;
+	}
+	return netdev;
+}
+
+extern int ipoib_probe ( struct ib_device *ibdev );
+extern void ipoib_remove ( struct ib_device *ibdev );
+
+#endif /* _GPXE_IPOIB_H */

src/include/gpxe/tcp.h

@@ -275,7 +275,8 @@ struct tcp_options {
  * actually use 65536, we use a window size of (65536-4) to ensure
  * that payloads remain dword-aligned.
  */
-#define TCP_MAX_WINDOW_SIZE	( 65536 - 4 )
+//#define TCP_MAX_WINDOW_SIZE	( 65536 - 4 )
+#define TCP_MAX_WINDOW_SIZE	4096
 
 /**
  * Path MTU

src/net/infiniband.c

@@ -0,0 +1,210 @@
+/*
+ * Copyright (C) 2007 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <byteswap.h>
+#include <errno.h>
+#include <assert.h>
+#include <gpxe/list.h>
+#include <gpxe/if_arp.h>
+#include <gpxe/netdevice.h>
+#include <gpxe/iobuf.h>
+#include <gpxe/infiniband.h>
+
+/** @file
+ *
+ * Infiniband protocol
+ *
+ */
+
+/**
+ * Create completion queue
+ *
+ * @v ibdev		Infiniband device
+ * @v num_cqes		Number of completion queue entries
+ * @ret cq		New completion queue
+ */
+struct ib_completion_queue * ib_create_cq ( struct ib_device *ibdev,
+					    unsigned int num_cqes ) {
+	struct ib_completion_queue *cq;
+	int rc;
+
+	DBGC ( ibdev, "IBDEV %p creating completion queue\n", ibdev );
+
+	/* Allocate and initialise data structure */
+	cq = zalloc ( sizeof ( *cq ) );
+	if ( ! cq )
+		return NULL;
+	cq->num_cqes = num_cqes;
+	INIT_LIST_HEAD ( &cq->work_queues );
+
+	/* Perform device-specific initialisation and get CQN */
+	if ( ( rc = ibdev->op->create_cq ( ibdev, cq ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not initialise completion "
+		       "queue: %s\n", ibdev, strerror ( rc ) );
+		free ( cq );
+		return NULL;
+	}
+
+	DBGC ( ibdev, "IBDEV %p created %d-entry completion queue %p (%p) "
+	       "with CQN %#lx\n", ibdev, num_cqes, cq, cq->dev_priv, cq->cqn );
+	return cq;
+}
+
+/**
+ * Destroy completion queue
+ *
+ * @v ibdev		Infiniband device
+ * @v cq		Completion queue
+ */
+void ib_destroy_cq ( struct ib_device *ibdev,
+		     struct ib_completion_queue *cq ) {
+	DBGC ( ibdev, "IBDEV %p destroying completion queue %#lx\n",
+	       ibdev, cq->cqn );
+	assert ( list_empty ( &cq->work_queues ) );
+	ibdev->op->destroy_cq ( ibdev, cq );
+	free ( cq );
+}
+
+/**
+ * Create queue pair
+ *
+ * @v ibdev		Infiniband device
+ * @v num_send_wqes	Number of send work queue entries
+ * @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
+ */
+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 ) {
+	struct ib_queue_pair *qp;
+	size_t total_size;
+	int rc;
+
+	DBGC ( ibdev, "IBDEV %p creating queue pair\n", ibdev );
+
+	/* Allocate and initialise data structure */
+	total_size = ( sizeof ( *qp ) +
+		       ( num_send_wqes * sizeof ( qp->send.iobufs[0] ) ) +
+		       ( num_recv_wqes * sizeof ( qp->recv.iobufs[0] ) ) );
+	qp = zalloc ( total_size );
+	if ( ! qp )
+		return NULL;
+	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.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.num_wqes = num_recv_wqes;
+	qp->recv.iobufs = ( ( ( void * ) qp ) + sizeof ( *qp ) +
+			    ( num_send_wqes * sizeof ( qp->send.iobufs[0] ) ));
+
+	/* Perform device-specific initialisation and get QPN */
+	if ( ( rc = ibdev->op->create_qp ( ibdev, qp ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not initialise queue pair: "
+		       "%s\n", ibdev, strerror ( rc ) );
+		free ( qp );
+		return NULL;
+	}
+
+	DBGC ( ibdev, "IBDEV %p created queue pair %p (%p) with QPN %#lx\n",
+	       ibdev, qp, qp->dev_priv, qp->qpn );
+	DBGC ( ibdev, "IBDEV %p QPN %#lx has %d send entries at [%p,%p)\n",
+	       ibdev, qp->qpn, num_send_wqes, qp->send.iobufs,
+	       qp->recv.iobufs );
+	DBGC ( ibdev, "IBDEV %p QPN %#lx has %d receive entries at [%p,%p)\n",
+	       ibdev, qp->qpn, num_send_wqes, qp->recv.iobufs,
+	       ( ( ( void * ) qp ) + total_size ) );
+	return qp;
+}
+
+/**
+ * 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",
+	       ibdev, qp->qpn );
+	ibdev->op->destroy_qp ( ibdev, qp );
+	list_del ( &qp->send.list );
+	list_del ( &qp->recv.list );
+	free ( qp );
+}
+
+/**
+ * Find work queue belonging to completion queue
+ *
+ * @v cq		Completion queue
+ * @v qpn		Queue pair number
+ * @v is_send		Find send work queue (rather than receive)
+ * @ret wq		Work queue, or NULL if not found
+ */
+struct ib_work_queue * ib_find_wq ( struct ib_completion_queue *cq,
+				    unsigned long qpn, int is_send ) {
+	struct ib_work_queue *wq;
+
+	list_for_each_entry ( wq, &cq->work_queues, list ) {
+		if ( ( wq->qp->qpn == qpn ) && ( wq->is_send == is_send ) )
+			return wq;
+	}
+	return NULL;
+}
+
+/**
+ * Allocate Infiniband device
+ *
+ * @v priv_size		Size of private data area
+ * @ret ibdev		Infiniband device, or NULL
+ */
+struct ib_device * alloc_ibdev ( size_t priv_size ) {
+	struct ib_device *ibdev;
+	size_t total_len;
+
+	total_len = ( sizeof ( *ibdev ) + priv_size );
+	ibdev = zalloc ( total_len );
+	if ( ibdev ) {
+		ibdev->dev_priv = ( ( ( void * ) ibdev ) + sizeof ( *ibdev ) );
+	}
+	return ibdev;
+}
+
+/**
+ * Free Infiniband device
+ *
+ * @v ibdev		Infiniband device
+ */
+void free_ibdev ( struct ib_device *ibdev ) {
+	free ( ibdev );
+}