Skip to content

Commit 28d6137

Browse files
yishaihdledford
yishaih
authored and
dledford
committed
IB/mlx5: Add RSS QP support
Add support for Raw Ethernet RX HASH QP. Currently, creation and destruction of such a QP are supported. This QP is implemented as a simple TIR object which points to the receive RQ indirection table. The given hashing configuration is used to configure the TIR and by that it chooses the right RQ from the RQ indirection table. Signed-off-by: Yishai Hadas <[email protected]> Signed-off-by: Matan Barak <[email protected]> Reviewed-by: Sagi Grimberg <[email protected]> Signed-off-by: Doug Ledford <[email protected]>
1 parent c70285f commit 28d6137

File tree

3 files changed

+239
-0
lines changed

3 files changed

+239
-0
lines changed

‎drivers/infiniband/hw/mlx5/mlx5_ib.h

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -295,6 +295,10 @@ struct mlx5_ib_qp_trans {
295295
u8 resp_depth;
296296
};
297297

298+
struct mlx5_ib_rss_qp {
299+
u32 tirn;
300+
};
301+
298302
struct mlx5_ib_rq {
299303
struct mlx5_ib_qp_base base;
300304
struct mlx5_ib_wq *rq;
@@ -323,6 +327,7 @@ struct mlx5_ib_qp {
323327
union {
324328
struct mlx5_ib_qp_trans trans_qp;
325329
struct mlx5_ib_raw_packet_qp raw_packet_qp;
330+
struct mlx5_ib_rss_qp rss_qp;
326331
};
327332
struct mlx5_buf buf;
328333

‎drivers/infiniband/hw/mlx5/qp.c

Lines changed: 200 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1266,6 +1266,187 @@ static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
12661266
rq->doorbell = &qp->db;
12671267
}
12681268

1269+
static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1270+
{
1271+
mlx5_core_destroy_tir(dev->mdev, qp->rss_qp.tirn);
1272+
}
1273+
1274+
static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1275+
struct ib_pd *pd,
1276+
struct ib_qp_init_attr *init_attr,
1277+
struct ib_udata *udata)
1278+
{
1279+
struct ib_uobject *uobj = pd->uobject;
1280+
struct ib_ucontext *ucontext = uobj->context;
1281+
struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1282+
struct mlx5_ib_create_qp_resp resp = {};
1283+
int inlen;
1284+
int err;
1285+
u32 *in;
1286+
void *tirc;
1287+
void *hfso;
1288+
u32 selected_fields = 0;
1289+
size_t min_resp_len;
1290+
u32 tdn = mucontext->tdn;
1291+
struct mlx5_ib_create_qp_rss ucmd = {};
1292+
size_t required_cmd_sz;
1293+
1294+
if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1295+
return -EOPNOTSUPP;
1296+
1297+
if (init_attr->create_flags || init_attr->send_cq)
1298+
return -EINVAL;
1299+
1300+
min_resp_len = offsetof(typeof(resp), uuar_index) + sizeof(resp.uuar_index);
1301+
if (udata->outlen < min_resp_len)
1302+
return -EINVAL;
1303+
1304+
required_cmd_sz = offsetof(typeof(ucmd), reserved1) + sizeof(ucmd.reserved1);
1305+
if (udata->inlen < required_cmd_sz) {
1306+
mlx5_ib_dbg(dev, "invalid inlen\n");
1307+
return -EINVAL;
1308+
}
1309+
1310+
if (udata->inlen > sizeof(ucmd) &&
1311+
!ib_is_udata_cleared(udata, sizeof(ucmd),
1312+
udata->inlen - sizeof(ucmd))) {
1313+
mlx5_ib_dbg(dev, "inlen is not supported\n");
1314+
return -EOPNOTSUPP;
1315+
}
1316+
1317+
if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1318+
mlx5_ib_dbg(dev, "copy failed\n");
1319+
return -EFAULT;
1320+
}
1321+
1322+
if (ucmd.comp_mask) {
1323+
mlx5_ib_dbg(dev, "invalid comp mask\n");
1324+
return -EOPNOTSUPP;
1325+
}
1326+
1327+
if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)) || ucmd.reserved1) {
1328+
mlx5_ib_dbg(dev, "invalid reserved\n");
1329+
return -EOPNOTSUPP;
1330+
}
1331+
1332+
err = ib_copy_to_udata(udata, &resp, min_resp_len);
1333+
if (err) {
1334+
mlx5_ib_dbg(dev, "copy failed\n");
1335+
return -EINVAL;
1336+
}
1337+
1338+
inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1339+
in = mlx5_vzalloc(inlen);
1340+
if (!in)
1341+
return -ENOMEM;
1342+
1343+
tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1344+
MLX5_SET(tirc, tirc, disp_type,
1345+
MLX5_TIRC_DISP_TYPE_INDIRECT);
1346+
MLX5_SET(tirc, tirc, indirect_table,
1347+
init_attr->rwq_ind_tbl->ind_tbl_num);
1348+
MLX5_SET(tirc, tirc, transport_domain, tdn);
1349+
1350+
hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1351+
switch (ucmd.rx_hash_function) {
1352+
case MLX5_RX_HASH_FUNC_TOEPLITZ:
1353+
{
1354+
void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1355+
size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1356+
1357+
if (len != ucmd.rx_key_len) {
1358+
err = -EINVAL;
1359+
goto err;
1360+
}
1361+
1362+
MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1363+
MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1364+
memcpy(rss_key, ucmd.rx_hash_key, len);
1365+
break;
1366+
}
1367+
default:
1368+
err = -EOPNOTSUPP;
1369+
goto err;
1370+
}
1371+
1372+
if (!ucmd.rx_hash_fields_mask) {
1373+
/* special case when this TIR serves as steering entry without hashing */
1374+
if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1375+
goto create_tir;
1376+
err = -EINVAL;
1377+
goto err;
1378+
}
1379+
1380+
if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1381+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1382+
((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1383+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1384+
err = -EINVAL;
1385+
goto err;
1386+
}
1387+
1388+
/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1389+
if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1390+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1391+
MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1392+
MLX5_L3_PROT_TYPE_IPV4);
1393+
else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1394+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1395+
MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1396+
MLX5_L3_PROT_TYPE_IPV6);
1397+
1398+
if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1399+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) &&
1400+
((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1401+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))) {
1402+
err = -EINVAL;
1403+
goto err;
1404+
}
1405+
1406+
/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1407+
if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1408+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1409+
MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1410+
MLX5_L4_PROT_TYPE_TCP);
1411+
else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1412+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1413+
MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1414+
MLX5_L4_PROT_TYPE_UDP);
1415+
1416+
if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1417+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1418+
selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1419+
1420+
if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1421+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1422+
selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1423+
1424+
if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1425+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1426+
selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1427+
1428+
if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1429+
(ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1430+
selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1431+
1432+
MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1433+
1434+
create_tir:
1435+
err = mlx5_core_create_tir(dev->mdev, in, inlen, &qp->rss_qp.tirn);
1436+
1437+
if (err)
1438+
goto err;
1439+
1440+
kvfree(in);
1441+
/* qpn is reserved for that QP */
1442+
qp->trans_qp.base.mqp.qpn = 0;
1443+
return 0;
1444+
1445+
err:
1446+
kvfree(in);
1447+
return err;
1448+
}
1449+
12691450
static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
12701451
struct ib_qp_init_attr *init_attr,
12711452
struct ib_udata *udata, struct mlx5_ib_qp *qp)
@@ -1292,6 +1473,14 @@ static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
12921473
spin_lock_init(&qp->sq.lock);
12931474
spin_lock_init(&qp->rq.lock);
12941475

1476+
if (init_attr->rwq_ind_tbl) {
1477+
if (!udata)
1478+
return -ENOSYS;
1479+
1480+
err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
1481+
return err;
1482+
}
1483+
12951484
if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
12961485
if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
12971486
mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
@@ -1644,6 +1833,11 @@ static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
16441833
struct mlx5_modify_qp_mbox_in *in;
16451834
int err;
16461835

1836+
if (qp->ibqp.rwq_ind_tbl) {
1837+
destroy_rss_raw_qp_tir(dev, qp);
1838+
return;
1839+
}
1840+
16471841
base = qp->ibqp.qp_type == IB_QPT_RAW_PACKET ?
16481842
&qp->raw_packet_qp.rq.base :
16491843
&qp->trans_qp.base;
@@ -2504,6 +2698,9 @@ int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
25042698
int port;
25052699
enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
25062700

2701+
if (ibqp->rwq_ind_tbl)
2702+
return -ENOSYS;
2703+
25072704
if (unlikely(ibqp->qp_type == IB_QPT_GSI))
25082705
return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
25092706

@@ -4119,6 +4316,9 @@ int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
41194316
int err = 0;
41204317
u8 raw_packet_qp_state;
41214318

4319+
if (ibqp->rwq_ind_tbl)
4320+
return -ENOSYS;
4321+
41224322
if (unlikely(ibqp->qp_type == IB_QPT_GSI))
41234323
return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
41244324
qp_init_attr);

‎drivers/infiniband/hw/mlx5/user.h

Lines changed: 34 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -152,6 +152,40 @@ struct mlx5_ib_create_qp {
152152
__u64 sq_buf_addr;
153153
};
154154

155+
/* RX Hash function flags */
156+
enum mlx5_rx_hash_function_flags {
157+
MLX5_RX_HASH_FUNC_TOEPLITZ = 1 << 0,
158+
};
159+
160+
/*
161+
* RX Hash flags, these flags allows to set which incoming packet's field should
162+
* participates in RX Hash. Each flag represent certain packet's field,
163+
* when the flag is set the field that is represented by the flag will
164+
* participate in RX Hash calculation.
165+
* Note: *IPV4 and *IPV6 flags can't be enabled together on the same QP
166+
* and *TCP and *UDP flags can't be enabled together on the same QP.
167+
*/
168+
enum mlx5_rx_hash_fields {
169+
MLX5_RX_HASH_SRC_IPV4 = 1 << 0,
170+
MLX5_RX_HASH_DST_IPV4 = 1 << 1,
171+
MLX5_RX_HASH_SRC_IPV6 = 1 << 2,
172+
MLX5_RX_HASH_DST_IPV6 = 1 << 3,
173+
MLX5_RX_HASH_SRC_PORT_TCP = 1 << 4,
174+
MLX5_RX_HASH_DST_PORT_TCP = 1 << 5,
175+
MLX5_RX_HASH_SRC_PORT_UDP = 1 << 6,
176+
MLX5_RX_HASH_DST_PORT_UDP = 1 << 7
177+
};
178+
179+
struct mlx5_ib_create_qp_rss {
180+
__u64 rx_hash_fields_mask; /* enum mlx5_rx_hash_fields */
181+
__u8 rx_hash_function; /* enum mlx5_rx_hash_function_flags */
182+
__u8 rx_key_len; /* valid only for Toeplitz */
183+
__u8 reserved[6];
184+
__u8 rx_hash_key[128]; /* valid only for Toeplitz */
185+
__u32 comp_mask;
186+
__u32 reserved1;
187+
};
188+
155189
struct mlx5_ib_create_qp_resp {
156190
__u32 uuar_index;
157191
};

0 commit comments

Comments
 (0)