path: root/drivers/infiniband/hw/ipath/ipath_verbs.h

/*
 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * OpenIB.org BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */

#ifndef IPATH_VERBS_H
#define IPATH_VERBS_H

#include <linux/types.h>
#include <linux/spinlock.h>
#include <linux/kernel.h>
#include <linux/interrupt.h>
#include <rdma/ib_pack.h>

#include "ipath_layer.h"
#include "verbs_debug.h"

#define QPN_MAX                 (1 << 24)
#define QPNMAP_ENTRIES          (QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)

/*
 * Increment this value if any changes that break userspace ABI
 * compatibility are made.
 */
#define IPATH_UVERBS_ABI_VERSION       1

/*
 * Define an ib_cq_notify value that is not valid so we know when CQ
 * notifications are armed.
 */
#define IB_CQ_NONE	(IB_CQ_NEXT_COMP + 1)

#define IB_RNR_NAK			0x20
#define IB_NAK_PSN_ERROR		0x60
#define IB_NAK_INVALID_REQUEST		0x61
#define IB_NAK_REMOTE_ACCESS_ERROR	0x62
#define IB_NAK_REMOTE_OPERATIONAL_ERROR 0x63
#define IB_NAK_INVALID_RD_REQUEST	0x64

#define IPATH_POST_SEND_OK		0x01
#define IPATH_POST_RECV_OK		0x02
#define IPATH_PROCESS_RECV_OK		0x04
#define IPATH_PROCESS_SEND_OK		0x08

/* IB Performance Manager status values */
#define IB_PMA_SAMPLE_STATUS_DONE	0x00
#define IB_PMA_SAMPLE_STATUS_STARTED	0x01
#define IB_PMA_SAMPLE_STATUS_RUNNING	0x02

/* Mandatory IB performance counter select values. */
#define IB_PMA_PORT_XMIT_DATA	__constant_htons(0x0001)
#define IB_PMA_PORT_RCV_DATA	__constant_htons(0x0002)
#define IB_PMA_PORT_XMIT_PKTS	__constant_htons(0x0003)
#define IB_PMA_PORT_RCV_PKTS	__constant_htons(0x0004)
#define IB_PMA_PORT_XMIT_WAIT	__constant_htons(0x0005)

struct ib_reth {
	__be64 vaddr;
	__be32 rkey;
	__be32 length;
} __attribute__ ((packed));

struct ib_atomic_eth {
	__be64 vaddr;
	__be32 rkey;
	__be64 swap_data;
	__be64 compare_data;
} __attribute__ ((packed));

struct ipath_other_headers {
	__be32 bth[3];
	union {
		struct {
			__be32 deth[2];
			__be32 imm_data;
		} ud;
		struct {
			struct ib_reth reth;
			__be32 imm_data;
		} rc;
		struct {
			__be32 aeth;
			__be64 atomic_ack_eth;
		} at;
		__be32 imm_data;
		__be32 aeth;
		struct ib_atomic_eth atomic_eth;
	} u;
} __attribute__ ((packed));

/*
 * Note that UD packets with a GRH header are 8+40+12+8 = 68 bytes
 * long (72 w/ imm_data).  Only the first 56 bytes of the IB header
 * will be in the eager header buffer.  The remaining 12 or 16 bytes
 * are in the data buffer.
 */
struct ipath_ib_header {
	__be16 lrh[4];
	union {
		struct {
			struct ib_grh grh;
			struct ipath_other_headers oth;
		} l;
		struct ipath_other_headers oth;
	} u;
} __attribute__ ((packed));

/*
 * There is one struct ipath_mcast for each multicast GID.
 * All attached QPs are then stored as a list of
 * struct ipath_mcast_qp.
 */
struct ipath_mcast_qp {
	struct list_head list;
	struct ipath_qp *qp;
};

struct ipath_mcast {
	struct rb_node rb_node;
	union ib_gid mgid;
	struct list_head qp_list;
	wait_queue_head_t wait;
	atomic_t refcount;
};

/* Memory region */
struct ipath_mr {
	struct ib_mr ibmr;
	struct ipath_mregion mr;	/* must be last */
};

/* Fast memory region */
struct ipath_fmr {
	struct ib_fmr ibfmr;
	u8 page_shift;
	struct ipath_mregion mr;	/* must be last */
};

/* Protection domain */
struct ipath_pd {
	struct ib_pd ibpd;
	int user;		/* non-zero if created from user space */
};

/* Address Handle */
struct ipath_ah {
	struct ib_ah ibah;
	struct ib_ah_attr attr;
};

/*
 * Quick description of our CQ/QP locking scheme:
 *
 * We have one global lock that protects dev->cq/qp_table.  Each
 * struct ipath_cq/qp also has its own lock.  An individual qp lock
 * may be taken inside of an individual cq lock.  Both cqs attached to
 * a qp may be locked, with the send cq locked first.  No other
 * nesting should be done.
 *
 * Each struct ipath_cq/qp also has an atomic_t ref count.  The
 * pointer from the cq/qp_table to the struct counts as one reference.
 * This reference also is good for access through the consumer API, so
 * modifying the CQ/QP etc doesn't need to take another reference.
 * Access because of a completion being polled does need a reference.
 *
 * Finally, each struct ipath_cq/qp has a wait_queue_head_t for the
 * destroy function to sleep on.
 *
 * This means that access from the consumer API requires nothing but
 * taking the struct's lock.
 *
 * Access because of a completion event should go as follows:
 * - lock cq/qp_table and look up struct
 * - increment ref count in struct
 * - drop cq/qp_table lock
 * - lock struct, do your thing, and unlock struct
 * - decrement ref count; if zero, wake up waiters
 *
 * To destroy a CQ/QP, we can do the following:
 * - lock cq/qp_table, remove pointer, unlock cq/qp_table lock
 * - decrement ref count
 * - wait_event until ref count is zero
 *
 * It is the consumer's responsibilty to make sure that no QP
 * operations (WQE posting or state modification) are pending when the
 * QP is destroyed.  Also, the consumer must make sure that calls to
 * qp_modify are serialized.
 *
 * Possible optimizations (wait for profile data to see if/where we
 * have locks bouncing between CPUs):
 * - split cq/qp table lock into n separate (cache-aligned) locks,
 *   indexed (say) by the page in the table
 */

struct ipath_cq {
	struct ib_cq ibcq;
	struct tasklet_struct comptask;
	spinlock_t lock;
	u8 notify;
	u8 triggered;
	u32 head;		/* new records added to the head */
	u32 tail;		/* poll_cq() reads from here. */
	struct ib_wc *queue;	/* this is actually ibcq.cqe + 1 */
};

/*
 * Send work request queue entry.
 * The size of the sg_list is determined when the QP is created and stored
 * in qp->s_max_sge.
 */
struct ipath_swqe {
	struct ib_send_wr wr;	/* don't use wr.sg_list */
	u32 psn;		/* first packet sequence number */
	u32 lpsn;		/* last packet sequence number */
	u32 ssn;		/* send sequence number */
	u32 length;		/* total length of data in sg_list */
	struct ipath_sge sg_list[0];
};

/*
 * Receive work request queue entry.
 * The size of the sg_list is determined when the QP is created and stored
 * in qp->r_max_sge.
 */
struct ipath_rwqe {
	u64 wr_id;
	u32 length;		/* total length of data in sg_list */
	u8 num_sge;
	struct ipath_sge sg_list[0];
};

struct ipath_rq {
	spinlock_t lock;
	u32 head;		/* new work requests posted to the head */
	u32 tail;		/* receives pull requests from here. */
	u32 size;		/* size of RWQE array */
	u8 max_sge;
	struct ipath_rwqe *wq;	/* RWQE array */
};

struct ipath_srq {
	struct ib_srq ibsrq;
	struct ipath_rq rq;
	/* send signal when number of RWQEs < limit */
	u32 limit;
};

/*
 * Variables prefixed with s_ are for the requester (sender).
 * Variables prefixed with r_ are for the responder (receiver).
 * Variables prefixed with ack_ are for responder replies.
 *
 * Common variables are protected by both r_rq.lock and s_lock in that order
 * which only happens in modify_qp() or changing the QP 'state'.
 */
struct ipath_qp {
	struct ib_qp ibqp;
	struct ipath_qp *next;		/* link list for QPN hash table */
	struct ipath_qp *timer_next;	/* link list for ipath_ib_timer() */
	struct list_head piowait;	/* link for wait PIO buf */
	struct list_head timerwait;	/* link for waiting for timeouts */
	struct ib_ah_attr remote_ah_attr;
	struct ipath_ib_header s_hdr;	/* next packet header to send */
	atomic_t refcount;
	wait_queue_head_t wait;
	struct tasklet_struct s_task;
	struct ipath_sge_state *s_cur_sge;
	struct ipath_sge_state s_sge;	/* current send request data */
	/* current RDMA read send data */
	struct ipath_sge_state s_rdma_sge;
	struct ipath_sge_state r_sge;	/* current receive data */
	spinlock_t s_lock;
	unsigned long s_flags;
	u32 s_hdrwords;		/* size of s_hdr in 32 bit words */
	u32 s_cur_size;		/* size of send packet in bytes */
	u32 s_len;		/* total length of s_sge */
	u32 s_rdma_len;		/* total length of s_rdma_sge */
	u32 s_next_psn;		/* PSN for next request */
	u32 s_last_psn;		/* last response PSN processed */
	u32 s_psn;		/* current packet sequence number */
	u32 s_rnr_timeout;	/* number of milliseconds for RNR timeout */
	u32 s_ack_psn;		/* PSN for next ACK or RDMA_READ */
	u64 s_ack_atomic;	/* data for atomic ACK */
	u64 r_wr_id;		/* ID for current receive WQE */
	u64 r_atomic_data;	/* data for last atomic op */
	u32 r_atomic_psn;	/* PSN of last atomic op */
	u32 r_len;		/* total length of r_sge */
	u32 r_rcv_len;		/* receive data len processed */
	u32 r_psn;		/* expected rcv packet sequence number */
	u8 state;		/* QP state */
	u8 s_state;		/* opcode of last packet sent */