1 From 33c5e1a802583cd84b55a4c5270e9d7753ac29bf Mon Sep 17 00:00:00 2001
2 From: Eli Cohen <eli@mellanox.co.il>
3 Date: Tue, 15 Jan 2008 18:57:09 +0200
4 Subject: [PATCH] Add LSO support to mlx4
9 Adjusted setting "reserve" value in set_kernel_sq_size to fit changes
12 Signed-off-by: Eli Cohen <eli@mellnaox.co.il>
13 Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
16 drivers/infiniband/hw/mlx4/cq.c | 3 ++
17 drivers/infiniband/hw/mlx4/main.c | 4 +++
18 drivers/infiniband/hw/mlx4/qp.c | 52 +++++++++++++++++++++++++++++++++---
19 drivers/net/mlx4/fw.c | 9 ++++++
20 drivers/net/mlx4/fw.h | 1 +
21 drivers/net/mlx4/main.c | 1 +
22 include/linux/mlx4/device.h | 1 +
23 include/linux/mlx4/qp.h | 5 +++
24 8 files changed, 71 insertions(+), 5 deletions(-)
26 Index: ofed_kernel/drivers/infiniband/hw/mlx4/cq.c
27 ===================================================================
28 --- ofed_kernel.orig/drivers/infiniband/hw/mlx4/cq.c 2008-01-23 16:01:48.392614000 +0200
29 +++ ofed_kernel/drivers/infiniband/hw/mlx4/cq.c 2008-01-23 16:05:20.076983000 +0200
30 @@ -408,6 +408,9 @@ static int mlx4_ib_poll_one(struct mlx4_
31 case MLX4_OPCODE_BIND_MW:
32 wc->opcode = IB_WC_BIND_MW;
34 + case MLX4_OPCODE_LSO:
35 + wc->opcode = IB_WC_LSO;
39 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
40 Index: ofed_kernel/drivers/infiniband/hw/mlx4/main.c
41 ===================================================================
42 --- ofed_kernel.orig/drivers/infiniband/hw/mlx4/main.c 2008-01-23 16:01:48.398613000 +0200
43 +++ ofed_kernel/drivers/infiniband/hw/mlx4/main.c 2008-01-23 16:05:20.081982000 +0200
44 @@ -102,6 +102,8 @@ static int mlx4_ib_query_device(struct i
45 props->device_cap_flags |= IB_DEVICE_UD_AV_PORT_ENFORCE;
46 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_IPOIB_CSUM)
47 props->device_cap_flags |= IB_DEVICE_IP_CSUM;
48 + if (dev->dev->caps.max_gso_sz)
49 + props->device_cap_flags |= IB_DEVICE_TCP_TSO;
51 props->vendor_id = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
53 @@ -617,6 +619,8 @@ static void *mlx4_ib_add(struct mlx4_dev
55 if (ibdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_IPOIB_CSUM)
56 ibdev->ib_dev.flags |= IB_DEVICE_IP_CSUM;
57 + if (ibdev->dev->caps.max_gso_sz)
58 + ibdev->ib_dev.flags |= IB_DEVICE_TCP_TSO;
60 if (init_node_data(ibdev))
62 Index: ofed_kernel/drivers/infiniband/hw/mlx4/qp.c
63 ===================================================================
64 --- ofed_kernel.orig/drivers/infiniband/hw/mlx4/qp.c 2008-01-23 16:01:51.101506000 +0200
65 +++ ofed_kernel/drivers/infiniband/hw/mlx4/qp.c 2008-01-23 16:08:04.078114000 +0200
66 @@ -69,6 +69,7 @@ enum {
68 static const __be32 mlx4_ib_opcode[] = {
69 [IB_WR_SEND] = __constant_cpu_to_be32(MLX4_OPCODE_SEND),
70 + [IB_WR_LSO] = __constant_cpu_to_be32(MLX4_OPCODE_LSO),
71 [IB_WR_SEND_WITH_IMM] = __constant_cpu_to_be32(MLX4_OPCODE_SEND_IMM),
72 [IB_WR_RDMA_WRITE] = __constant_cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
73 [IB_WR_RDMA_WRITE_WITH_IMM] = __constant_cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
74 @@ -244,6 +245,7 @@ static int set_kernel_sq_size(struct mlx
76 struct ib_qp_cap *cap = &init_attr->cap;
77 enum ib_qp_type type = init_attr->qp_type;
80 /* Sanity check SQ size before proceeding */
81 if (cap->max_send_wr > dev->dev->caps.max_wqes ||
82 @@ -260,12 +262,16 @@ static int set_kernel_sq_size(struct mlx
83 cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
86 + if (qp->flags & MLX4_QP_LSO)
89 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(max(cap->max_send_sge *
90 - sizeof (struct mlx4_wqe_data_seg),
91 + sizeof (struct mlx4_wqe_data_seg) +
93 cap->max_inline_data +
94 sizeof (struct mlx4_wqe_inline_seg)) +
95 send_wqe_overhead(type)));
96 - qp->sq.max_gs = ((1 << qp->sq.wqe_shift) - send_wqe_overhead(type)) /
97 + qp->sq.max_gs = ((1 << qp->sq.wqe_shift) -reserve - send_wqe_overhead(type)) /
98 sizeof (struct mlx4_wqe_data_seg);
101 @@ -756,9 +764,11 @@ static int __mlx4_ib_modify_qp(struct ib
105 - if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
106 - ibqp->qp_type == IB_QPT_UD)
107 + if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
108 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
109 + else if (ibqp->qp_type == IB_QPT_UD)
110 + context->mtu_msgmax = (IB_MTU_4096 << 5) |
111 + ilog2(dev->dev->caps.max_gso_sz);
112 else if (attr_mask & IB_QP_PATH_MTU) {
113 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
114 printk(KERN_ERR "path MTU (%u) is invalid\n",
115 @@ -1276,6 +1286,28 @@ static void __set_data_seg(struct mlx4_w
116 dseg->addr = cpu_to_be64(sg->addr);
119 +static int build_lso_seg(struct mlx4_lso_seg *wqe, struct ib_send_wr *wr,
120 + struct mlx4_ib_qp *qp, int *lso_seg_len)
124 + halign = ALIGN(wr->wr.ud.hlen, 16);
125 + if (unlikely(!(qp->flags & MLX4_QP_LSO) && wr->num_sge > qp->sq.max_gs - (halign >> 4)))
128 + memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
130 + /* make sure LSO header is written before
131 + overwriting stamping */
134 + wqe->mss_hdr_size = cpu_to_be32(((wr->wr.ud.mss - wr->wr.ud.hlen)
135 + << 16) | wr->wr.ud.hlen);
137 + *lso_seg_len = halign;
141 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
142 struct ib_send_wr **bad_wr)
144 @@ -1366,6 +1398,19 @@ int mlx4_ib_post_send(struct ib_qp *ibqp
145 set_datagram_seg(wqe, wr);
146 wqe += sizeof (struct mlx4_wqe_datagram_seg);
147 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
149 + if (wr->opcode == IB_WR_LSO) {
152 + err = build_lso_seg(wqe, wr, qp, &hlen);
164 Index: ofed_kernel/drivers/net/mlx4/fw.c
165 ===================================================================
166 --- ofed_kernel.orig/drivers/net/mlx4/fw.c 2008-01-23 16:01:48.430615000 +0200
167 +++ ofed_kernel/drivers/net/mlx4/fw.c 2008-01-23 16:05:20.106981000 +0200
168 @@ -133,6 +133,7 @@ int mlx4_QUERY_DEV_CAP(struct mlx4_dev *
169 #define QUERY_DEV_CAP_MAX_AV_OFFSET 0x27
170 #define QUERY_DEV_CAP_MAX_REQ_QP_OFFSET 0x29
171 #define QUERY_DEV_CAP_MAX_RES_QP_OFFSET 0x2b
172 +#define QUERY_DEV_CAP_MAX_GSO_OFFSET 0x2d
173 #define QUERY_DEV_CAP_MAX_RDMA_OFFSET 0x2f
174 #define QUERY_DEV_CAP_RSZ_SRQ_OFFSET 0x33
175 #define QUERY_DEV_CAP_ACK_DELAY_OFFSET 0x35
176 @@ -215,6 +216,13 @@ int mlx4_QUERY_DEV_CAP(struct mlx4_dev *
177 dev_cap->max_requester_per_qp = 1 << (field & 0x3f);
178 MLX4_GET(field, outbox, QUERY_DEV_CAP_MAX_RES_QP_OFFSET);
179 dev_cap->max_responder_per_qp = 1 << (field & 0x3f);
180 + MLX4_GET(field, outbox, QUERY_DEV_CAP_MAX_GSO_OFFSET);
183 + dev_cap->max_gso_sz = 0;
185 + dev_cap->max_gso_sz = 1 << field;
187 MLX4_GET(field, outbox, QUERY_DEV_CAP_MAX_RDMA_OFFSET);
188 dev_cap->max_rdma_global = 1 << (field & 0x3f);
189 MLX4_GET(field, outbox, QUERY_DEV_CAP_ACK_DELAY_OFFSET);
190 @@ -377,6 +385,7 @@ int mlx4_QUERY_DEV_CAP(struct mlx4_dev *
191 dev_cap->max_sq_desc_sz, dev_cap->max_sq_sg);
192 mlx4_dbg(dev, "Max RQ desc size: %d, max RQ S/G: %d\n",
193 dev_cap->max_rq_desc_sz, dev_cap->max_rq_sg);
194 + mlx4_dbg(dev, "Max GSO size: %d\n", dev_cap->max_gso_sz);
196 dump_dev_cap_flags(dev, dev_cap->flags);
198 Index: ofed_kernel/drivers/net/mlx4/fw.h
199 ===================================================================
200 --- ofed_kernel.orig/drivers/net/mlx4/fw.h 2008-01-23 15:58:48.837059000 +0200
201 +++ ofed_kernel/drivers/net/mlx4/fw.h 2008-01-23 16:05:20.109984000 +0200
202 @@ -96,6 +96,7 @@ struct mlx4_dev_cap {
209 struct mlx4_adapter {
210 Index: ofed_kernel/drivers/net/mlx4/main.c
211 ===================================================================
212 --- ofed_kernel.orig/drivers/net/mlx4/main.c 2008-01-23 15:58:48.841058000 +0200
213 +++ ofed_kernel/drivers/net/mlx4/main.c 2008-01-23 16:05:20.115981000 +0200
214 @@ -159,6 +159,7 @@ static int mlx4_dev_cap(struct mlx4_dev
215 dev->caps.page_size_cap = ~(u32) (dev_cap->min_page_sz - 1);
216 dev->caps.flags = dev_cap->flags;
217 dev->caps.stat_rate_support = dev_cap->stat_rate_support;
218 + dev->caps.max_gso_sz = dev_cap->max_gso_sz;
222 Index: ofed_kernel/include/linux/mlx4/device.h
223 ===================================================================
224 --- ofed_kernel.orig/include/linux/mlx4/device.h 2008-01-23 15:58:48.844060000 +0200
225 +++ ofed_kernel/include/linux/mlx4/device.h 2008-01-23 16:05:20.138984000 +0200
226 @@ -181,6 +181,7 @@ struct mlx4_caps {
228 u16 stat_rate_support;
229 u8 port_width_cap[MLX4_MAX_PORTS + 1];
233 struct mlx4_buf_list {
234 Index: ofed_kernel/include/linux/mlx4/qp.h
235 ===================================================================
236 --- ofed_kernel.orig/include/linux/mlx4/qp.h 2008-01-23 16:01:48.448613000 +0200
237 +++ ofed_kernel/include/linux/mlx4/qp.h 2008-01-23 16:05:20.142981000 +0200
238 @@ -215,6 +215,11 @@ struct mlx4_wqe_datagram_seg {
242 +struct mlx4_lso_seg {
243 + __be32 mss_hdr_size;
247 struct mlx4_wqe_bind_seg {