upstream/oracle/userland-gate: comparison components/open-fabrics/libmlx4/patches/base.patch

equal deleted inserted replaced

-:b205ca9f0d84
+:eed3ed08f692
 #define MLX4_UVERBS_MAX_ABI_VERSION	3
 diff -r -u /tmp/839450/libmlx4-1.0.1/src/verbs.c libmlx4-1.0.1/src/verbs.c
 --- /tmp/839450/libmlx4-1.0.1/src/verbs.c	Thu Mar 10 04:48:34 2011
 +++ libmlx4-1.0.1/src/verbs.c	Fri Mar 11 14:40:18 2011
-@@ -56,6 +56,15 @@
+@@ -56,6 +56,14 @@
 	if (ret)
 		return ret;
 +#if defined(__SVR4) && defined(__sun)
 +	/*
 +	 * we need to report the max as actual max  less 1. In OFED this is
 +	 * done in the HCA driver.
 +	 */
 +	attr->max_srq_wr -=1;
 +#endif
-+
 	major     = (raw_fw_ver >> 32) & 0xffff;
 	minor     = (raw_fw_ver >> 16) & 0xffff;
 	sub_minor = raw_fw_ver & 0xffff;
-@@ -79,6 +88,9 @@
+@@ -79,6 +87,9 @@
 	struct ibv_alloc_pd       cmd;
 	struct mlx4_alloc_pd_resp resp;
 	struct mlx4_pd		 *pd;
 +#if defined(__SVR4) && defined(__sun)
 +	mlnx_umap_pd_data_out_t   *mdd;
 +#endif
 	pd = malloc(sizeof *pd);
 	if (!pd)
-@@ -90,7 +102,16 @@
+@@ -90,7 +101,16 @@
 		return NULL;
 	}
 +#if defined(__SVR4) && defined(__sun)
 +	/*
-+	 * The kernel driver passes back the PD table index as opaque data.  This
++	 * kernel driver passes back the PD table index as opaque data.  This
-+	 * index is required for specifying the PD in user space address vectors.
++	 * is required for specifying the PD in user space address vectors.
 +	 */
 +	mdd     = (mlnx_umap_pd_data_out_t *) &resp.ibv_resp.drv_out;
 +	pd->pdn = mdd->mpd_pdnum;
 +#else
 	pd->pdn = resp.pdn;
 +#endif
 	return &pd->ibv_pd;
 }
-@@ -168,6 +189,10 @@
+@@ -168,6 +188,10 @@
 	struct mlx4_create_cq_resp resp;
 	struct mlx4_cq		  *cq;
 	int			   ret;
 +#if defined(__SVR4) && defined(__sun)
 +	void                      *cqbuf;
 +	mlnx_umap_cq_data_out_t   *mdd;
 +#endif
 	/* Sanity check CQ size before proceeding */
 	if (cqe > 0x3fffff)
-@@ -184,7 +209,8 @@
+@@ -184,7 +208,8 @@
 	cqe = align_queue_size(cqe + 1);
 -	if (mlx4_alloc_cq_buf(to_mdev(context->device), &cq->buf, cqe))
 +#if !(defined(__SVR4) && defined(__sun))
 +	if (mlx4_alloc_cq_buf((to_mdev(context->device), &cq->buf, cqe))
 		goto err;
 	cq->set_ci_db  = mlx4_alloc_db(to_mctx(context), MLX4_DB_TYPE_CQ);
-@@ -198,15 +224,84 @@
+@@ -198,15 +223,78 @@
 	cmd.buf_addr = (uintptr_t) cq->buf.buf;
 	cmd.db_addr  = (uintptr_t) cq->set_ci_db;
 +#else
 +	cq->buf.buf    = NULL;
 +#endif
 	ret = ibv_cmd_create_cq(context, cqe - 1, channel, comp_vector,
 				&cq->ibv_cq, &cmd.ibv_cmd, sizeof cmd,
 				&resp.ibv_resp, sizeof resp);
-+
++#if defined(__SVR4) && defined(__sun)
-+#if defined(__SVR4) && defined(__sun)
+	if (ret)
-+	if (ret) {
 +		goto err;
-+	}
++#else
-+#else
++	if (ret)
-	if (ret)
 		goto err_db;
 +#endif
 	cq->cqn = resp.cqn;
 +	}
 +
 +        cqbuf = mmap64((void *)0, mdd->mcq_maplen, (PROT_READ | PROT_WRITE),
 +                    MAP_SHARED, context->mmap_fd, mdd->mcq_mapoffset);
 +
-+        if (cqbuf == MAP_FAILED) {
++        if (cqbuf == MAP_FAILED)
 +                goto err_destroy;
-+        }
 +
 +        /*
 +         * Extract hardware driver values for the number of CQEs and the
 +	 * hardware CQ number to use (needed for user space doorbells).
 +         */
-+        cqe            = mdd->mcq_numcqe;
++	cqe            = mdd->mcq_numcqe;
-+        cq->cqn        = mdd->mcq_cqnum;
++	cq->cqn        = mdd->mcq_cqnum;
-+        cq->buf.buf    = cqbuf;
++	cq->buf.buf    = cqbuf;
-+        cq->buf.length = mdd->mcq_maplen;
++	cq->buf.length = mdd->mcq_maplen;
 +	cq->ibv_cq.cqe = cqe-1;
 +
 +	/*
 +	 * We map both poll and arm as seperate doorbells (OFED assumes 1 word
 +	 * offset and just bumpts the address) since Solaris provides a
 +	 */
 +	cq->set_ci_db = mlx4_alloc_db(to_mctx(context),
 +	                              mdd->mcq_polldbr_mapoffset,
 +	                              mdd->mcq_polldbr_maplen,
 +	                              mdd->mcq_polldbr_offset);
-+        if (cq->set_ci_db == NULL) {
++        if (cq->set_ci_db == NULL)
 +                goto err_buf;
-+        }
 +
 +	cq->arm_db = mlx4_alloc_db(to_mctx(context),
 +	                           mdd->mcq_armdbr_mapoffset,
 +	                           mdd->mcq_armdbr_maplen,
 +	                           mdd->mcq_armdbr_offset);
-+        if (cq->arm_db == NULL) {
++        if (cq->arm_db == NULL)
 +                goto err_db;
-+        }
 +
 +	*cq->arm_db    = 0;
 +	cq->arm_sn     = 1;
 +	*cq->set_ci_db = 0;
 +#endif
-+
 	return &cq->ibv_cq;
 err_db:
-@@ -215,6 +310,22 @@
+@@ -215,6 +303,21 @@
 err_buf:
 	mlx4_free_buf(&cq->buf);
 +#if defined(__SVR4) && defined(__sun)
 +err_destroy:
 +	cq->ibv_cq.comp_events_completed = 0;
 +	cq->ibv_cq.async_events_completed = 0;
 +
 +	ibv_cmd_destroy_cq(&cq->ibv_cq);
 +#endif
-+
 err:
 	free(cq);
-@@ -225,12 +336,17 @@
+@@ -225,12 +328,16 @@
 {
 	struct mlx4_cq *cq = to_mcq(ibcq);
 	struct mlx4_resize_cq cmd;
 +	struct ibv_resize_cq_resp resp;
 	struct mlx4_buf buf;
 -
 +#if defined(__SVR4) && defined(__sun)
 +	void			*cqbuf;
 +	mlnx_umap_cq_data_out_t	*mdd;
 +#endif
-+
 	/* Sanity check CQ size before proceeding */
 	if (cqe > 0x3fffff)
 -		return EINVAL;
 + 		return EINVAL;
 	pthread_spin_lock(&cq->lock);
-@@ -247,32 +363,65 @@
+@@ -247,32 +354,76 @@
 		goto out;
 	}
 +#if !(defined(__SVR4) && defined(__sun))
 	ret = mlx4_alloc_cq_buf(to_mdev(ibcq->context->device), &buf, cqe);
 	if (ret)
 		goto out;
+-
 -	old_cqe = ibcq->cqe;
--	cmd.buf_addr = (uintptr_t) buf.buf;
+	cmd.buf_addr = (uintptr_t) buf.buf;
-+        cmd.buf_addr = (uintptr_t) buf.buf;
 +#endif
 +        old_cqe = ibcq->cqe;
 #ifdef IBV_CMD_RESIZE_CQ_HAS_RESP_PARAMS
 -	{
 #endif
 -	if (ret) {
 -		mlx4_free_buf(&buf);
 +
 +        if (ret) {
-+#if ! (defined(__SVR4) && defined(__sun))
++#if !(defined(__SVR4) && defined(__sun))
 +                mlx4_free_buf(&buf);
-+#endif
 		goto out;
 	}
 -	mlx4_cq_resize_copy_cqes(cq, buf.buf, old_cqe);
-+#if defined(__SVR4) && defined(__sun)
++        mlx4_cq_resize_copy_cqes(cq, buf.buf, old_cqe);
++        mlx4_free_buf(&cq->buf);
++        cq->buf = buf;
++#else
++		goto out;
++	}
++        if (cq->buf.buf != NULL) {
++        	buf.buf = malloc(cq->buf.length);
++        	if (!buf.buf) {
++                	ret = ENOMEM;
++                	goto out;
++        	}
+-	mlx4_free_buf(&cq->buf);
+-	cq->buf = buf;
++        	memcpy(buf.buf, cq->buf.buf, cq->buf.length);
++        	buf.length =  cq->buf.length;
++                ret = munmap((char *)cq->buf.buf, cq->buf.length);
++                if (ret) {
++                        free(buf.buf);
++                        goto out;
++                }
++        }
 +	/*
 +	 * For Solaris the kernel driver passes back mmap information for
 +	 * mapping the CQ memory it allocated.
 +	 */
 +	mdd = (mlnx_umap_cq_data_out_t *) &resp.drv_out;
 +		    "rev mismatch (kernel rev=%d)\n", mdd->mcq_rev);
 +		ret = EINVAL;
 +		goto out;
 +	}
--	mlx4_free_buf(&cq->buf);
--	cq->buf = buf;
 +	cqbuf = mmap64((void *)0, mdd->mcq_maplen, (PROT_READ | PROT_WRITE),
 +	     MAP_SHARED, ibcq->context->mmap_fd, mdd->mcq_mapoffset);
++
 +	if (cqbuf == MAP_FAILED) {
 +		ret = EINVAL;
 +		goto out;
 +	}
-+#endif
++	cq->buf.buf    = buf.buf;
-+
++	cq->buf.length = buf.length;
-+        mlx4_cq_resize_copy_cqes(cq, buf.buf, old_cqe);
++	mlx4_cq_resize_copy_cqes(cq, cqbuf, old_cqe);
-+
-+#if !(defined(__SVR4) && defined(__sun))
-+        mlx4_free_buf(&cq->buf);
-+#endif
-+        cq->buf = buf;
-+
-+#if defined(__SVR4) && defined(__sun)
-+	cqe            = mdd->mcq_numcqe;
-+	cq->cqn        = mdd->mcq_cqnum;
 +	cq->buf.buf    = cqbuf;
 +	cq->buf.length = mdd->mcq_maplen;
-+	cq->ibv_cq.cqe = cqe-1;
++	free(buf.buf);
++	cq->ibv_cq.cqe =  mdd->mcq_numcqe - 1;
++	cq->cqn        = mdd->mcq_cqnum;
 +#endif
 out:
 	pthread_spin_unlock(&cq->lock);
 	return ret;
-@@ -287,6 +436,9 @@
+@@ -287,6 +438,9 @@
 		return ret;
 	mlx4_free_db(to_mctx(cq->context), MLX4_DB_TYPE_CQ, to_mcq(cq)->set_ci_db);
 +#if defined(__SVR4) && defined(__sun)
 +	mlx4_free_db(to_mctx(cq->context), MLX4_DB_TYPE_CQ, to_mcq(cq)->arm_db);
 +#endif
 	mlx4_free_buf(&to_mcq(cq)->buf);
 	free(to_mcq(cq));
-@@ -300,6 +452,10 @@
+@@ -300,6 +454,10 @@
 	struct mlx4_create_srq_resp resp;
 	struct mlx4_srq		   *srq;
 	int			    ret;
 +#if defined(__SVR4) && defined(__sun)
 +	mlnx_umap_srq_data_out_t   *mdd;
 +	void                       *srqbuf;
 +#endif
 	/* Sanity check SRQ size before proceeding */
 	if (attr->attr.max_wr > 1 << 16 || attr->attr.max_sge > 64)
-@@ -312,6 +468,7 @@
+@@ -312,6 +470,7 @@
 	if (pthread_spin_init(&srq->lock, PTHREAD_PROCESS_PRIVATE))
 		goto err;
 +#if !(defined(__SVR4) && defined(__sun))
 	srq->max     = align_queue_size(attr->attr.max_wr + 1);
 	srq->max_gs  = attr->attr.max_sge;
 	srq->counter = 0;
-@@ -324,7 +481,23 @@
+@@ -324,7 +483,23 @@
 		goto err_free;
 	*srq->db = 0;
 +#else
 +	/*
 +	attr->attr.max_wr += 1;
 +#endif
 	cmd.buf_addr = (uintptr_t) srq->buf.buf;
 	cmd.db_addr  = (uintptr_t) srq->db;
-@@ -331,19 +504,97 @@
+@@ -331,19 +506,97 @@
 	ret = ibv_cmd_create_srq(pd, &srq->ibv_srq, attr,
 				 &cmd.ibv_cmd, sizeof cmd,
 				 &resp.ibv_resp, sizeof resp);
 +#if defined(__SVR4) && defined(__sun)
 +	if (ret) {
 	mlx4_free_buf(&srq->buf);
 +#endif
 err:
 	free(srq);
-@@ -357,7 +608,16 @@
+@@ -357,7 +610,16 @@
 {
 	struct ibv_modify_srq cmd;
 +#if !(defined(__SVR4) && defined(__sun))
 	return ibv_cmd_modify_srq(srq, attr, attr_mask, &cmd, sizeof cmd);
 +	return (ret);
 +#endif
 }
 int mlx4_query_srq(struct ibv_srq *srq,
-@@ -365,7 +625,17 @@
+@@ -365,7 +627,17 @@
 {
 	struct ibv_query_srq cmd;
 +#if !(defined(__SVR4) && defined(__sun))
 	return ibv_cmd_query_srq(srq, attr, &cmd, sizeof cmd);
 +	return (ret);
 +#endif
 }
 int mlx4_destroy_srq(struct ibv_srq *ibsrq)
-@@ -414,7 +684,11 @@
+@@ -447,6 +719,10 @@
-		return -1;
-	if (attr->cap.max_inline_data) {
-+#if !(defined(__SVR4) && defined(__sun))
-		nsegs = num_inline_segs(attr->cap.max_inline_data, attr->qp_type);
-+#else
-+		nsegs = mlx4_num_inline_segs(attr->cap.max_inline_data, attr->qp_type);
-+#endif
-		size = MLX4_MAX_WQE_SIZE - nsegs * sizeof (struct mlx4_wqe_inline_seg);
-		switch (attr->qp_type) {
-		case IBV_QPT_UD:
-@@ -447,6 +721,12 @@
 	struct mlx4_qp		 *qp;
 	int			  ret;
 	struct mlx4_context	 *context = to_mctx(pd->context);
 +#if defined(__SVR4) && defined(__sun)
 +	mlnx_umap_qp_data_out_t	*mdd;
 +	void			*qpbuf;
-+	int			max_send_sge;
-+	int			max_inline_data;
 +#endif
 	/* Sanity check QP size before proceeding */
-@@ -457,6 +737,7 @@
+@@ -457,6 +733,7 @@
 	if (!qp)
 		return NULL;
 +#if !(defined(__SVR4) && defined(__sun))
 	mlx4_calc_sq_wqe_size(&attr->cap, attr->qp_type, qp);
 	/*
-@@ -466,6 +747,7 @@
+@@ -466,6 +743,7 @@
 	qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + 1;
 	qp->sq.wqe_cnt = align_queue_size(attr->cap.max_send_wr + qp->sq_spare_wqes);
 	qp->rq.wqe_cnt = align_queue_size(attr->cap.max_recv_wr);
 +#endif
 	if (attr->srq || attr->qp_type == IBV_QPT_XRC)
 		attr->cap.max_recv_wr = qp->rq.wqe_cnt = 0;
-@@ -476,6 +758,46 @@
+@@ -476,6 +754,22 @@
 			attr->cap.max_recv_wr = 1;
 	}
 +#if defined(__SVR4) && defined(__sun)
 +	if (pthread_spin_init(&qp->sq.lock, PTHREAD_PROCESS_PRIVATE) ||
-+	    pthread_spin_init(&qp->rq.lock, PTHREAD_PROCESS_PRIVATE)) {
++	    pthread_spin_init(&qp->rq.lock, PTHREAD_PROCESS_PRIVATE))
 +		goto err;
-+	}
-+
-+	/*
-+	 * We adjust the number of send SGL entries to force the kernel to
-+	 * allocate a larger WQE that will fit the inline data requested.
-+	 * The Solaris Hermon driver does not look at inline data size when
-+	 * calculating the send WQE size, so this allows us to get closer
-+	 * to what the user has requested.
-+	 */
-+	max_send_sge = align(attr->cap.max_inline_data +
-+			mlx4_num_inline_segs(attr->cap.max_inline_data,
-+			attr->qp_type) * sizeof (struct mlx4_wqe_inline_seg),
-+			sizeof( struct mlx4_wqe_data_seg)) /
-+	                          sizeof(struct mlx4_wqe_data_seg);
-+
-+
-+	if (max_send_sge > attr->cap.max_send_sge)
-+		attr->cap.max_send_sge = max_send_sge;
-+
-+	if (attr->cap.max_send_sge > context->max_sge) {
-+		free(qp);
-+		return (NULL);
-+	};
-+
 +
 +	/*
 +	 * Solaris QP work queue memory is supplied by the kernel, so
 +	 * we will update this after creation.
 +	 */
 +	memset(&cmd, 0, sizeof(cmd));
 +#else
 	if (mlx4_alloc_qp_buf(pd, &attr->cap, attr->qp_type, qp))
 		goto err;
-@@ -505,11 +827,120 @@
+@@ -505,17 +799,84 @@
 		; /* nothing */
 	cmd.sq_no_prefetch = 0;	/* OK for ABI 2: just a reserved field */
 	memset(cmd.reserved, 0, sizeof cmd.reserved);
 +#endif
 	pthread_mutex_lock(&to_mctx(pd->context)->qp_table_mutex);
 	ret = ibv_cmd_create_qp(pd, &qp->ibv_qp, attr, &cmd.ibv_cmd, sizeof cmd,
 				&resp, sizeof resp);
-+
++#if defined(__SVR4) && defined(__sun)
-+#if defined(__SVR4) && defined(__sun)
+	if (ret)
-+	if (ret) {
 +		goto err_free;
-+	}
 +
 +        /*
 +         * The kernel driver passes back mmap information for mapping the
 +         * QP work queue memory it allocated back into user space.
 +         */
 +		goto err_destroy;
 +	}
 +	qpbuf = mmap64((void *)0, mdd->mqp_maplen, (PROT_READ | PROT_WRITE),
 +	                MAP_SHARED, pd->context->mmap_fd, mdd->mqp_mapoffset);
 +
-+	if (qpbuf == MAP_FAILED) {
++	if (qpbuf == MAP_FAILED)
 +		goto err_destroy;
-+	}
 +
 +	/*
 +	 * Need to set qp->buf here in case alloc_db fails then
 +	 * we'll call mlx4_free_buf() to umap.
 +	 */
 +	if (!attr->srq && attr->qp_type != IBV_QPT_XRC) {
 +		qp->db = mlx4_alloc_db(to_mctx(pd->context),
 +		                       mdd->mqp_rdbr_mapoffset,
 +		                       mdd->mqp_rdbr_maplen,
 +		                       mdd->mqp_rdbr_offset);
-+		if (qp->db == NULL) {
++		if (qp->db == NULL)
 +			goto err_buf;
-+		}
++
 +		*qp->db = 0;
 +	}
-+
-+	/*
-+	 * Calculate the official maximum inline data size, this is not done
-+	 * by the kernel driver, so we do it here and update the qp struct.
-+	 */
-+	max_inline_data =
-+			mdd->mqp_sq_wqesz - sizeof(struct mlx4_wqe_inline_seg);
-+	max_inline_data -= sizeof(struct mlx4_wqe_ctrl_seg);
-+
-+	switch (attr->qp_type) {
-+	case IBV_QPT_UD:
-+		max_inline_data -= sizeof(struct mlx4_wqe_datagram_seg);
-+		break;
-+
-+	case IBV_QPT_UC:
-+		max_inline_data -= sizeof(struct mlx4_wqe_raddr_seg);
-+		break;
-+
-+	case IBV_QPT_RC:
-+		max_inline_data -= sizeof(struct mlx4_wqe_raddr_seg);
-+		if (max_inline_data > (sizeof(struct mlx4_wqe_atomic_seg) +
-+		                       sizeof(struct mlx4_wqe_raddr_seg) +
-+		                       sizeof(struct mlx4_wqe_data_seg))) {
-+			max_inline_data -= sizeof(struct mlx4_wqe_atomic_seg) +
-+		                           sizeof(struct mlx4_wqe_raddr_seg) +
-+		                           sizeof(struct mlx4_wqe_data_seg);
-+		} else {
-+			max_inline_data = 0;
-+		}
-+		break;
-+
-+	default:
-+		break;
-+	}
-+
-+	attr->cap.max_inline_data = max_inline_data;
 +
 +	/*
 +	 * Retrieve sendqueue actual size, and the number of headroom WQEs
 +	 * that were required based on kernel setup of prefetch or not for
 +	 * send queue.
-+	 * 	Note: mqp_sq_numwqe includes the head room wqes.
++	 * 	Note: mqp_sq_numwqe includes the head room wqes. The private
-+	 *	      The private wqe.cnt also includes headroom wqes,
++	 *	      wqe.cnt also includes headroom wqes, the verbs count
-+	 *	      the verbs count should reflect the wqe count that
++	 *	      should reflect the wqe count that is usable.
-+	 *	      is usable.
 +	 */
 +	qp->sq_spare_wqes = mdd->mqp_sq_headroomwqes;
 +	qp->sq.wqe_cnt    = mdd->mqp_sq_numwqe;
 +
-+	if (attr->srq) {
++	if (attr->srq)
 +		qp->rq.wqe_cnt  = 0;
-+	} else {
++	else
 +		qp->rq.wqe_cnt  = mdd->mqp_rq_numwqe;
-+	}
 +
 +	if (mlx4_set_qp_buf(pd, qp, qpbuf, mdd->mqp_maplen,
 +	                    mdd->mqp_rq_wqesz, mdd->mqp_rq_off,
-+	                    mdd->mqp_sq_wqesz, mdd->mqp_sq_off)) {
++	                    mdd->mqp_sq_wqesz, mdd->mqp_sq_off))
-+		goto err_rq_db;
+		goto err_rq_db;
-+	}
-+
 +	mlx4_init_qp_indices(qp);
 +
-+	ret = mlx4_store_qp(to_mctx(pd->context), qp->ibv_qp.qp_num, qp);
-+	if (ret) {
-+		goto err_rq_db;
-+	}
-+#else
-	if (ret)
-		goto err_rq_db;
-@@ -516,6 +947,7 @@
 	ret = mlx4_store_qp(to_mctx(pd->context), qp->ibv_qp.qp_num, qp);
 	if (ret)
++		goto err_rq_db;
++#else
++	if (ret)
++		goto err_rq_db;
++
++	ret = mlx4_store_qp(to_mctx(pd->context), qp->ibv_qp.qp_num, qp);
++	if (ret)
 		goto err_destroy;
 +#endif
 	pthread_mutex_unlock(&to_mctx(pd->context)->qp_table_mutex);
 	qp->rq.wqe_cnt = attr->cap.max_recv_wr;
-@@ -536,9 +968,42 @@
+@@ -536,9 +897,38 @@
 	return &qp->ibv_qp;
 +#if defined(__SVR4) && defined(__sun)
 +err_rq_db:
-+	if (!attr->srq && attr->qp_type != IBV_QPT_XRC) {
++	if (!attr->srq && attr->qp_type != IBV_QPT_XRC)
 +		mlx4_free_db(to_mctx(pd->context), MLX4_DB_TYPE_RQ, qp->db);
-+	}
-+
 +err_buf:
 +	mlx4_free_buf(&qp->buf);
 +
 err_destroy:
 +	/*
 +	 */
 +	pthread_mutex_init(&(qp->ibv_qp.mutex), NULL);
 +	pthread_cond_init(&(qp->ibv_qp.cond), NULL);
 +	qp->ibv_qp.events_completed = 0;
 	ibv_cmd_destroy_qp(&qp->ibv_qp);
 +err_free:
 +	pthread_mutex_unlock(&to_mctx(pd->context)->qp_table_mutex);
-+
 +	if (qp->sq.wrid)
 +		free(qp->sq.wrid);
 +
 +	if (qp->rq.wrid)
 +		free(qp->rq.wrid);
-+
 +err:
 +	free(qp);
 +#else
 +err_destroy:
 +	ibv_cmd_destroy_qp(&qp->ibv_qp);
 +
 err_rq_db:
 	pthread_mutex_unlock(&to_mctx(pd->context)->qp_table_mutex);
 	if (!attr->srq && attr->qp_type != IBV_QPT_XRC)
-@@ -552,6 +1017,7 @@
+@@ -552,6 +942,7 @@
 err:
 	free(qp);
 +#endif
 	return NULL;
 }
-@@ -745,6 +1211,13 @@
+@@ -745,6 +1136,13 @@
 				    struct ibv_cq *xrc_cq,
 				    struct ibv_srq_init_attr *attr)
 {
 +#if defined(__SVR4) && defined(__sun)
 +	/*
 +	return NULL;
 +#else
 	struct mlx4_create_xrc_srq  cmd;
 	struct mlx4_create_srq_resp resp;
 	struct mlx4_srq		   *srq;
-@@ -807,6 +1280,7 @@
+@@ -807,6 +1205,7 @@
 	free(srq);
 	return NULL;
 +#endif
 }
 struct ibv_xrc_domain *mlx4_open_xrc_domain(struct ibv_context *context,
-@@ -893,5 +1367,4 @@
-{
-	return ibv_cmd_unreg_xrc_rcv_qp(xrc_domain, xrc_qp_num);
-}
--
-#endif
 diff -r -u /tmp/839450/libmlx4-1.0.1/src/qp.c libmlx4-1.0.1/src/qp.c
 --- /tmp/839450/libmlx4-1.0.1/src/qp.c	Thu Mar 10 04:48:34 2011
 +++ libmlx4-1.0.1/src/qp.c	Tue Mar 15 07:09:43 2011
-@@ -511,7 +511,17 @@
+@@ -589,6 +589,58 @@
-	return ret;
-}
-+#if defined(__SVR4) && defined(__sun)
-+/*
-+ * Create a non-static version that can be called externally;
-+ * default file local calls to now use the name of the non-static
-+ * version.
-+ */
-+#define	num_inline_segs mlx4_num_inline_segs
-+int mlx4_num_inline_segs(int data, enum ibv_qp_type type)
-+#else
-int num_inline_segs(int data, enum ibv_qp_type type)
-+#endif
-{
-	/*
-	 * Inline data segments are not allowed to cross 64 byte
-@@ -589,6 +599,58 @@
 		; /* nothing */
 }
 +#if defined(__SVR4) && defined(__sun)
 +int mlx4_set_qp_buf(struct ibv_pd *pd, struct mlx4_qp *qp, void *qpbuf,
 			struct mlx4_srq *srq);
 +#endif
 void mlx4_free_srq_wqe(struct mlx4_srq *srq, int ind);
 int mlx4_post_srq_recv(struct ibv_srq *ibsrq,
 		       struct ibv_recv_wr *wr,
-@@ -399,6 +413,9 @@
+@@ -415,8 +429,14 @@
-		       struct mlx4_srq *srq);
-void mlx4_clear_xrc_srq(struct mlx4_context *ctx, uint32_t xrc_srqn);
-+#if defined(__SVR4) && defined(__sun)
-+int mlx4_num_inline_segs(int data, enum ibv_qp_type type);
-+#endif
-struct ibv_qp *mlx4_create_qp(struct ibv_pd *pd, struct ibv_qp_init_attr *attr);
-int mlx4_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
-		   int attr_mask,
-@@ -415,8 +432,14 @@
 void mlx4_calc_sq_wqe_size(struct ibv_qp_cap *cap, enum ibv_qp_type type,
 			   struct mlx4_qp *qp);
 int num_inline_segs(int data, enum ibv_qp_type type);
 +#if defined(__SVR4) && defined(__sun)
 +int mlx4_set_qp_buf(struct ibv_pd *pd, struct mlx4_qp *qp,
 +	uint32_t                        temp_qp_num;
 +#endif
 	context = calloc(1, sizeof *context);
 	if (!context)
-@@ -150,11 +155,32 @@
+@@ -150,11 +155,30 @@
 		return NULL;
 	context->ibv_ctx.cmd_fd = cmd_fd;
 +#if defined(__SVR4) && defined(__sun)
 +	context->ibv_ctx.device = ibdev;
 +#if defined(__SVR4) && defined(__sun)
 +	/*
 +	 * OFED expects power of two, round up here to make user table
 +	 * large enough.
 +	 */
-+	for (temp_qp_num = 1; temp_qp_num < resp.qp_tab_size; temp_qp_num <<= 1) {
++	for (temp_qp_num = 1; temp_qp_num < resp.qp_tab_size; temp_qp_num <<= 1)
 +		;
-+	}
-+
 +	resp.qp_tab_size = temp_qp_num;
 +
 +	/*
 +	 * NOTE: In Solaris this value is not returned in the channel interface
 +	 * opaque data and is assumed to be 2*256 by the dapl code.  We have
 +	resp.bf_reg_size = 512;
 +#endif
 	context->num_qps	= resp.qp_tab_size;
 	context->qp_table_shift = ffs(context->num_qps) - 1 - MLX4_QP_TABLE_BITS;
 	context->qp_table_mask	= (1 << context->qp_table_shift) - 1;
-@@ -172,20 +198,44 @@
+@@ -172,20 +196,44 @@
 	for (i = 0; i < MLX4_XRC_SRQ_TABLE_SIZE; ++i)
 		context->xrc_srq_table[i].refcnt = 0;
 +#if defined(__SVR4) && defined(__sun)
 +	context->db_page_list = NULL;
 					to_mdev(ibdev)->page_size);
 +#endif
 		if (context->bf_page == MAP_FAILED) {
 			fprintf(stderr, PFX "Warning: BlueFlame available, "
 				"but failed to mmap() BlueFlame page.\n");
-@@ -214,6 +264,7 @@
+@@ -214,6 +262,7 @@
 	context->max_qp_wr = dev_attrs.max_qp_wr;
 	context->max_sge = dev_attrs.max_sge;
 	context->max_cqe = dev_attrs.max_cqe;
 +#ifdef HAVE_IBV_XRC_OPS
 	if (!(dev_attrs.device_cap_flags & IBV_DEVICE_XRC)) {
 		fprintf(stderr, PFX "There is a mismatch between "
 		        "the kernel and the userspace libraries: "
-@@ -220,6 +271,7 @@
+@@ -220,6 +269,7 @@
 			"Kernel does not support XRC. Exiting.\n");
 		goto query_free;
 	}
 +#endif
 	return &context->ibv_ctx;
-@@ -227,7 +279,6 @@
+@@ -240,6 +290,7 @@
-	munmap(context->uar, to_mdev(ibdev)->page_size);
-	if (context->bf_page)
-		munmap(context->bf_page, to_mdev(ibdev)->page_size);
--
-err_free:
-	free(context);
-	return NULL;
-@@ -240,6 +291,7 @@
 	munmap(context->uar, to_mdev(ibctx->device)->page_size);
 	if (context->bf_page)
 		munmap(context->bf_page, to_mdev(ibctx->device)->page_size);
 +
 	free(context);
 @@ -6,7 +6,7 @@
 Group: System Environment/Libraries
 License: GPLv2 or BSD
 Url: http://openfabrics.org/
 -Source: http://openfabrics.org/downloads/libmlx4/libmlx4-1.0.1.tar.gz
-+Source: http://openfabrics.org/downloads/mlx4/libmlx4-1.0.1.tar.gz
++Source: http://openfabrics.org/downloads/libmlx4-1.0.1.tar.gz
 BuildRoot: %(mktemp -ud %{_tmppath}/%{name}-%{version}-%{release}-XXXXXX)
 BuildRequires: libibverbs-devel >= 1.1-0.1.rc2
 diff -r -u /tmp/839450/libmlx4-1.0.1/configure libmlx4-1.0.1/configure
 --- /tmp/839450/libmlx4-1.0.1/configure	Thu Mar 10 04:48:41 2011
 +++ libmlx4-1.0.1/configure	Tue Mar 15 07:35:49 2011
+@@ -3899,13 +3899,13 @@
+CFLAGS=$ac_save_CFLAGS
+elif test $ac_cv_prog_cc_g = yes; then
+if test "$GCC" = yes; then
+-    CFLAGS="-g -O2"
++    CFLAGS="-g -O3"
+else
+CFLAGS="-g"
+fi
+else
+if test "$GCC" = yes; then
+-    CFLAGS="-O2"
++    CFLAGS="-O3"
+else
+CFLAGS=
+fi
 @@ -8890,6 +8890,7 @@
 	;;
 esac
 link_all_deplibs=yes
 +	hardcode_libdir_flag_spec=
 ;;
 sunos4*)
-@@ -11616,6 +11617,14 @@
+@@ -11113,13 +11114,13 @@
-# This bug is HP SR number 8606223364.
+CFLAGS=$ac_save_CFLAGS
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking size of long" >&5
+elif test $ac_cv_prog_cc_g = yes; then
-$as_echo_n "checking size of long... " >&6; }
+if test "$GCC" = yes; then
-+
+-    CFLAGS="-g -O2"
-+echo $CFLAGS | grep 64 > /dev/null
++    CFLAGS="-g -O3"
-+if [ $? -eq 0 ]; then
+else
-+	ac_cv_sizeof_long=8
+CFLAGS="-g"
-+else
+fi
-+	ac_cv_sizeof_long=4
-+fi
-+
-if test "${ac_cv_sizeof_long+set}" = set; then :
-$as_echo_n "(cached) " >&6
 else
-@@ -11654,11 +11663,11 @@
+if test "$GCC" = yes; then
+-    CFLAGS="-O2"
++    CFLAGS="-O3"
+else
+CFLAGS=
+fi
+@@ -11654,11 +11655,11 @@
 ac_fn_c_check_member "$LINENO" "struct ibv_more_ops" "create_xrc_srq" "ac_cv_member_struct_ibv_more_ops_create_xrc_srq" "#include <infiniband/verbs.h>
 "
 -if test "x$ac_cv_member_struct_ibv_more_ops_create_xrc_srq" = x""yes; then :
 +#if test "x$ac_cv_member_struct_ibv_more_ops_create_xrc_srq" = x""yes; then :

changeset 715	eed3ed08f692
parent 673	bb9df3c906be
child 817	f45ca7242301