2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/highmem.h>
34 #include <linux/module.h>
35 #include <linux/init.h>
36 #include <linux/errno.h>
37 #include <linux/pci.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/slab.h>
40 #include <linux/io-mapping.h>
41 #include <linux/sched.h>
42 #include <rdma/ib_user_verbs.h>
43 #include <rdma/ib_addr.h>
44 #include <rdma/ib_cache.h>
45 #include <linux/mlx5/vport.h>
46 #include <rdma/ib_smi.h>
47 #include <rdma/ib_umem.h>
49 #include <linux/etherdevice.h>
50 #include <linux/mlx5/fs.h>
54 #define DRIVER_NAME "mlx5_ib"
55 #define DRIVER_VERSION "2.2-1"
56 #define DRIVER_RELDATE "Feb 2014"
58 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
59 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
60 MODULE_LICENSE("Dual BSD/GPL");
61 MODULE_VERSION(DRIVER_VERSION);
63 static int deprecated_prof_sel = 2;
64 module_param_named(prof_sel, deprecated_prof_sel, int, 0444);
65 MODULE_PARM_DESC(prof_sel, "profile selector. Deprecated here. Moved to module mlx5_core");
67 static char mlx5_version[] =
68 DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
69 DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
72 MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
75 static enum rdma_link_layer
76 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
78 switch (port_type_cap) {
79 case MLX5_CAP_PORT_TYPE_IB:
80 return IB_LINK_LAYER_INFINIBAND;
81 case MLX5_CAP_PORT_TYPE_ETH:
82 return IB_LINK_LAYER_ETHERNET;
84 return IB_LINK_LAYER_UNSPECIFIED;
88 static enum rdma_link_layer
89 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
91 struct mlx5_ib_dev *dev = to_mdev(device);
92 int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
94 return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
97 static int mlx5_netdev_event(struct notifier_block *this,
98 unsigned long event, void *ptr)
100 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
101 struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
104 if ((event != NETDEV_UNREGISTER) && (event != NETDEV_REGISTER))
107 write_lock(&ibdev->roce.netdev_lock);
108 if (ndev->dev.parent == &ibdev->mdev->pdev->dev)
109 ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ? NULL : ndev;
110 write_unlock(&ibdev->roce.netdev_lock);
115 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
118 struct mlx5_ib_dev *ibdev = to_mdev(device);
119 struct net_device *ndev;
121 /* Ensure ndev does not disappear before we invoke dev_hold()
123 read_lock(&ibdev->roce.netdev_lock);
124 ndev = ibdev->roce.netdev;
127 read_unlock(&ibdev->roce.netdev_lock);
132 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
133 struct ib_port_attr *props)
135 struct mlx5_ib_dev *dev = to_mdev(device);
136 struct net_device *ndev;
137 enum ib_mtu ndev_ib_mtu;
140 memset(props, 0, sizeof(*props));
142 props->port_cap_flags |= IB_PORT_CM_SUP;
143 props->port_cap_flags |= IB_PORT_IP_BASED_GIDS;
145 props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
146 roce_address_table_size);
147 props->max_mtu = IB_MTU_4096;
148 props->max_msg_sz = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
149 props->pkey_tbl_len = 1;
150 props->state = IB_PORT_DOWN;
151 props->phys_state = 3;
153 mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
154 props->qkey_viol_cntr = qkey_viol_cntr;
156 ndev = mlx5_ib_get_netdev(device, port_num);
160 if (netif_running(ndev) && netif_carrier_ok(ndev)) {
161 props->state = IB_PORT_ACTIVE;
162 props->phys_state = 5;
165 ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
169 props->active_mtu = min(props->max_mtu, ndev_ib_mtu);
171 props->active_width = IB_WIDTH_4X; /* TODO */
172 props->active_speed = IB_SPEED_QDR; /* TODO */
177 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
178 const struct ib_gid_attr *attr,
181 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
182 char *mlx5_addr_l3_addr = MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
184 void *mlx5_addr_mac = MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
190 ether_addr_copy(mlx5_addr_mac, attr->ndev->dev_addr);
192 if (is_vlan_dev(attr->ndev)) {
193 MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
194 MLX5_SET_RA(mlx5_addr, vlan_id, vlan_dev_vlan_id(attr->ndev));
197 switch (attr->gid_type) {
199 MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
201 case IB_GID_TYPE_ROCE_UDP_ENCAP:
202 MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
209 if (attr->gid_type != IB_GID_TYPE_IB) {
210 if (ipv6_addr_v4mapped((void *)gid))
211 MLX5_SET_RA(mlx5_addr, roce_l3_type,
212 MLX5_ROCE_L3_TYPE_IPV4);
214 MLX5_SET_RA(mlx5_addr, roce_l3_type,
215 MLX5_ROCE_L3_TYPE_IPV6);
218 if ((attr->gid_type == IB_GID_TYPE_IB) ||
219 !ipv6_addr_v4mapped((void *)gid))
220 memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
222 memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
225 static int set_roce_addr(struct ib_device *device, u8 port_num,
227 const union ib_gid *gid,
228 const struct ib_gid_attr *attr)
230 struct mlx5_ib_dev *dev = to_mdev(device);
231 u32 in[MLX5_ST_SZ_DW(set_roce_address_in)];
232 u32 out[MLX5_ST_SZ_DW(set_roce_address_out)];
233 void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
234 enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
236 if (ll != IB_LINK_LAYER_ETHERNET)
239 memset(in, 0, sizeof(in));
241 ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
243 MLX5_SET(set_roce_address_in, in, roce_address_index, index);
244 MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
246 memset(out, 0, sizeof(out));
247 return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
250 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
251 unsigned int index, const union ib_gid *gid,
252 const struct ib_gid_attr *attr,
253 __always_unused void **context)
255 return set_roce_addr(device, port_num, index, gid, attr);
258 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
259 unsigned int index, __always_unused void **context)
261 return set_roce_addr(device, port_num, index, NULL, NULL);
264 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
267 struct ib_gid_attr attr;
270 if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
278 if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
281 return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
284 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
286 return !dev->mdev->issi;
290 MLX5_VPORT_ACCESS_METHOD_MAD,
291 MLX5_VPORT_ACCESS_METHOD_HCA,
292 MLX5_VPORT_ACCESS_METHOD_NIC,
295 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
297 if (mlx5_use_mad_ifc(to_mdev(ibdev)))
298 return MLX5_VPORT_ACCESS_METHOD_MAD;
300 if (mlx5_ib_port_link_layer(ibdev, 1) ==
301 IB_LINK_LAYER_ETHERNET)
302 return MLX5_VPORT_ACCESS_METHOD_NIC;
304 return MLX5_VPORT_ACCESS_METHOD_HCA;
307 static void get_atomic_caps(struct mlx5_ib_dev *dev,
308 struct ib_device_attr *props)
311 u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
312 u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
313 u8 atomic_req_8B_endianness_mode =
314 MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
316 /* Check if HW supports 8 bytes standard atomic operations and capable
317 * of host endianness respond
319 tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
320 if (((atomic_operations & tmp) == tmp) &&
321 (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
322 (atomic_req_8B_endianness_mode)) {
323 props->atomic_cap = IB_ATOMIC_HCA;
325 props->atomic_cap = IB_ATOMIC_NONE;
329 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
330 __be64 *sys_image_guid)
332 struct mlx5_ib_dev *dev = to_mdev(ibdev);
333 struct mlx5_core_dev *mdev = dev->mdev;
337 switch (mlx5_get_vport_access_method(ibdev)) {
338 case MLX5_VPORT_ACCESS_METHOD_MAD:
339 return mlx5_query_mad_ifc_system_image_guid(ibdev,
342 case MLX5_VPORT_ACCESS_METHOD_HCA:
343 err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
346 case MLX5_VPORT_ACCESS_METHOD_NIC:
347 err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
355 *sys_image_guid = cpu_to_be64(tmp);
361 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
364 struct mlx5_ib_dev *dev = to_mdev(ibdev);
365 struct mlx5_core_dev *mdev = dev->mdev;
367 switch (mlx5_get_vport_access_method(ibdev)) {
368 case MLX5_VPORT_ACCESS_METHOD_MAD:
369 return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
371 case MLX5_VPORT_ACCESS_METHOD_HCA:
372 case MLX5_VPORT_ACCESS_METHOD_NIC:
373 *max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
382 static int mlx5_query_vendor_id(struct ib_device *ibdev,
385 struct mlx5_ib_dev *dev = to_mdev(ibdev);
387 switch (mlx5_get_vport_access_method(ibdev)) {
388 case MLX5_VPORT_ACCESS_METHOD_MAD:
389 return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
391 case MLX5_VPORT_ACCESS_METHOD_HCA:
392 case MLX5_VPORT_ACCESS_METHOD_NIC:
393 return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
400 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
406 switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
407 case MLX5_VPORT_ACCESS_METHOD_MAD:
408 return mlx5_query_mad_ifc_node_guid(dev, node_guid);
410 case MLX5_VPORT_ACCESS_METHOD_HCA:
411 err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
414 case MLX5_VPORT_ACCESS_METHOD_NIC:
415 err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
423 *node_guid = cpu_to_be64(tmp);
428 struct mlx5_reg_node_desc {
432 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
434 struct mlx5_reg_node_desc in;
436 if (mlx5_use_mad_ifc(dev))
437 return mlx5_query_mad_ifc_node_desc(dev, node_desc);
439 memset(&in, 0, sizeof(in));
441 return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
442 sizeof(struct mlx5_reg_node_desc),
443 MLX5_REG_NODE_DESC, 0, 0);
446 static int mlx5_ib_query_device(struct ib_device *ibdev,
447 struct ib_device_attr *props,
448 struct ib_udata *uhw)
450 struct mlx5_ib_dev *dev = to_mdev(ibdev);
451 struct mlx5_core_dev *mdev = dev->mdev;
455 u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
457 if (uhw->inlen || uhw->outlen)
460 memset(props, 0, sizeof(*props));
461 err = mlx5_query_system_image_guid(ibdev,
462 &props->sys_image_guid);
466 err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
470 err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
474 props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
475 (fw_rev_min(dev->mdev) << 16) |
476 fw_rev_sub(dev->mdev);
477 props->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT |
478 IB_DEVICE_PORT_ACTIVE_EVENT |
479 IB_DEVICE_SYS_IMAGE_GUID |
480 IB_DEVICE_RC_RNR_NAK_GEN;
482 if (MLX5_CAP_GEN(mdev, pkv))
483 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
484 if (MLX5_CAP_GEN(mdev, qkv))
485 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
486 if (MLX5_CAP_GEN(mdev, apm))
487 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
488 if (MLX5_CAP_GEN(mdev, xrc))
489 props->device_cap_flags |= IB_DEVICE_XRC;
490 props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
491 if (MLX5_CAP_GEN(mdev, sho)) {
492 props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
493 /* At this stage no support for signature handover */
494 props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
495 IB_PROT_T10DIF_TYPE_2 |
496 IB_PROT_T10DIF_TYPE_3;
497 props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
498 IB_GUARD_T10DIF_CSUM;
500 if (MLX5_CAP_GEN(mdev, block_lb_mc))
501 props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
503 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
504 (MLX5_CAP_ETH(dev->mdev, csum_cap)))
505 props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
507 props->vendor_part_id = mdev->pdev->device;
508 props->hw_ver = mdev->pdev->revision;
510 props->max_mr_size = ~0ull;
511 props->page_size_cap = ~(min_page_size - 1);
512 props->max_qp = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
513 props->max_qp_wr = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
514 max_rq_sg = MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
515 sizeof(struct mlx5_wqe_data_seg);
516 max_sq_sg = (MLX5_CAP_GEN(mdev, max_wqe_sz_sq) -
517 sizeof(struct mlx5_wqe_ctrl_seg)) /
518 sizeof(struct mlx5_wqe_data_seg);
519 props->max_sge = min(max_rq_sg, max_sq_sg);
520 props->max_sge_rd = props->max_sge;
521 props->max_cq = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
522 props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
523 props->max_mr = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
524 props->max_pd = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
525 props->max_qp_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
526 props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
527 props->max_srq = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
528 props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
529 props->local_ca_ack_delay = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
530 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp;
531 props->max_srq_sge = max_rq_sg - 1;
532 props->max_fast_reg_page_list_len = (unsigned int)-1;
533 get_atomic_caps(dev, props);
534 props->masked_atomic_cap = IB_ATOMIC_NONE;
535 props->max_mcast_grp = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
536 props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
537 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
538 props->max_mcast_grp;
539 props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
540 props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
541 props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
543 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
544 if (MLX5_CAP_GEN(mdev, pg))
545 props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
546 props->odp_caps = dev->odp_caps;
549 if (MLX5_CAP_GEN(mdev, cd))
550 props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
556 MLX5_IB_WIDTH_1X = 1 << 0,
557 MLX5_IB_WIDTH_2X = 1 << 1,
558 MLX5_IB_WIDTH_4X = 1 << 2,
559 MLX5_IB_WIDTH_8X = 1 << 3,
560 MLX5_IB_WIDTH_12X = 1 << 4
563 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
566 struct mlx5_ib_dev *dev = to_mdev(ibdev);
569 if (active_width & MLX5_IB_WIDTH_1X) {
570 *ib_width = IB_WIDTH_1X;
571 } else if (active_width & MLX5_IB_WIDTH_2X) {
572 mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
575 } else if (active_width & MLX5_IB_WIDTH_4X) {
576 *ib_width = IB_WIDTH_4X;
577 } else if (active_width & MLX5_IB_WIDTH_8X) {
578 *ib_width = IB_WIDTH_8X;
579 } else if (active_width & MLX5_IB_WIDTH_12X) {
580 *ib_width = IB_WIDTH_12X;
582 mlx5_ib_dbg(dev, "Invalid active_width %d\n",
590 static int mlx5_mtu_to_ib_mtu(int mtu)
599 pr_warn("invalid mtu\n");
609 __IB_MAX_VL_0_14 = 5,
612 enum mlx5_vl_hw_cap {
624 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
629 *max_vl_num = __IB_MAX_VL_0;
632 *max_vl_num = __IB_MAX_VL_0_1;
635 *max_vl_num = __IB_MAX_VL_0_3;
638 *max_vl_num = __IB_MAX_VL_0_7;
640 case MLX5_VL_HW_0_14:
641 *max_vl_num = __IB_MAX_VL_0_14;
651 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
652 struct ib_port_attr *props)
654 struct mlx5_ib_dev *dev = to_mdev(ibdev);
655 struct mlx5_core_dev *mdev = dev->mdev;
656 struct mlx5_hca_vport_context *rep;
660 u8 ib_link_width_oper;
663 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
669 memset(props, 0, sizeof(*props));
671 err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
675 props->lid = rep->lid;
676 props->lmc = rep->lmc;
677 props->sm_lid = rep->sm_lid;
678 props->sm_sl = rep->sm_sl;
679 props->state = rep->vport_state;
680 props->phys_state = rep->port_physical_state;
681 props->port_cap_flags = rep->cap_mask1;
682 props->gid_tbl_len = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
683 props->max_msg_sz = 1 << MLX5_CAP_GEN(mdev, log_max_msg);
684 props->pkey_tbl_len = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
685 props->bad_pkey_cntr = rep->pkey_violation_counter;
686 props->qkey_viol_cntr = rep->qkey_violation_counter;
687 props->subnet_timeout = rep->subnet_timeout;
688 props->init_type_reply = rep->init_type_reply;
690 err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
694 err = translate_active_width(ibdev, ib_link_width_oper,
695 &props->active_width);
698 err = mlx5_query_port_proto_oper(mdev, &props->active_speed, MLX5_PTYS_IB,
703 mlx5_query_port_max_mtu(mdev, &max_mtu, port);
705 props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
707 mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
709 props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
711 err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
715 err = translate_max_vl_num(ibdev, vl_hw_cap,
722 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
723 struct ib_port_attr *props)
725 switch (mlx5_get_vport_access_method(ibdev)) {
726 case MLX5_VPORT_ACCESS_METHOD_MAD:
727 return mlx5_query_mad_ifc_port(ibdev, port, props);
729 case MLX5_VPORT_ACCESS_METHOD_HCA:
730 return mlx5_query_hca_port(ibdev, port, props);
732 case MLX5_VPORT_ACCESS_METHOD_NIC:
733 return mlx5_query_port_roce(ibdev, port, props);
740 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
743 struct mlx5_ib_dev *dev = to_mdev(ibdev);
744 struct mlx5_core_dev *mdev = dev->mdev;
746 switch (mlx5_get_vport_access_method(ibdev)) {
747 case MLX5_VPORT_ACCESS_METHOD_MAD:
748 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
750 case MLX5_VPORT_ACCESS_METHOD_HCA:
751 return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
759 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
762 struct mlx5_ib_dev *dev = to_mdev(ibdev);
763 struct mlx5_core_dev *mdev = dev->mdev;
765 switch (mlx5_get_vport_access_method(ibdev)) {
766 case MLX5_VPORT_ACCESS_METHOD_MAD:
767 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
769 case MLX5_VPORT_ACCESS_METHOD_HCA:
770 case MLX5_VPORT_ACCESS_METHOD_NIC:
771 return mlx5_query_hca_vport_pkey(mdev, 0, port, 0, index,
778 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
779 struct ib_device_modify *props)
781 struct mlx5_ib_dev *dev = to_mdev(ibdev);
782 struct mlx5_reg_node_desc in;
783 struct mlx5_reg_node_desc out;
786 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
789 if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
793 * If possible, pass node desc to FW, so it can generate
794 * a 144 trap. If cmd fails, just ignore.
796 memcpy(&in, props->node_desc, 64);
797 err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
798 sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
802 memcpy(ibdev->node_desc, props->node_desc, 64);
807 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
808 struct ib_port_modify *props)
810 struct mlx5_ib_dev *dev = to_mdev(ibdev);
811 struct ib_port_attr attr;
815 mutex_lock(&dev->cap_mask_mutex);
817 err = mlx5_ib_query_port(ibdev, port, &attr);
821 tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
822 ~props->clr_port_cap_mask;
824 err = mlx5_set_port_caps(dev->mdev, port, tmp);
827 mutex_unlock(&dev->cap_mask_mutex);
831 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
832 struct ib_udata *udata)
834 struct mlx5_ib_dev *dev = to_mdev(ibdev);
835 struct mlx5_ib_alloc_ucontext_req_v2 req = {};
836 struct mlx5_ib_alloc_ucontext_resp resp = {};
837 struct mlx5_ib_ucontext *context;
838 struct mlx5_uuar_info *uuari;
839 struct mlx5_uar *uars;
847 size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
851 return ERR_PTR(-EAGAIN);
853 if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
854 return ERR_PTR(-EINVAL);
856 reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
857 if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
859 else if (reqlen >= min_req_v2)
862 return ERR_PTR(-EINVAL);
864 err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
869 return ERR_PTR(-EINVAL);
871 if (req.total_num_uuars > MLX5_MAX_UUARS)
872 return ERR_PTR(-ENOMEM);
874 if (req.total_num_uuars == 0)
875 return ERR_PTR(-EINVAL);
877 if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
878 return ERR_PTR(-EOPNOTSUPP);
880 if (reqlen > sizeof(req) &&
881 !ib_is_udata_cleared(udata, sizeof(req),
882 reqlen - sizeof(req)))
883 return ERR_PTR(-EOPNOTSUPP);
885 req.total_num_uuars = ALIGN(req.total_num_uuars,
886 MLX5_NON_FP_BF_REGS_PER_PAGE);
887 if (req.num_low_latency_uuars > req.total_num_uuars - 1)
888 return ERR_PTR(-EINVAL);
890 num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
891 gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
892 resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
893 resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
894 resp.cache_line_size = L1_CACHE_BYTES;
895 resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
896 resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
897 resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
898 resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
899 resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
900 resp.cqe_version = min_t(__u8,
901 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
902 req.max_cqe_version);
903 resp.response_length = min(offsetof(typeof(resp), response_length) +
904 sizeof(resp.response_length), udata->outlen);
906 context = kzalloc(sizeof(*context), GFP_KERNEL);
908 return ERR_PTR(-ENOMEM);
910 uuari = &context->uuari;
911 mutex_init(&uuari->lock);
912 uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
918 uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
919 sizeof(*uuari->bitmap),
921 if (!uuari->bitmap) {
926 * clear all fast path uuars
928 for (i = 0; i < gross_uuars; i++) {
930 if (uuarn == 2 || uuarn == 3)
931 set_bit(i, uuari->bitmap);
934 uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
940 for (i = 0; i < num_uars; i++) {
941 err = mlx5_cmd_alloc_uar(dev->mdev, &uars[i].index);
946 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
947 context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
950 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
951 err = mlx5_core_alloc_transport_domain(dev->mdev,
957 INIT_LIST_HEAD(&context->db_page_list);
958 mutex_init(&context->db_page_mutex);
960 resp.tot_uuars = req.total_num_uuars;
961 resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
963 if (field_avail(typeof(resp), cqe_version, udata->outlen))
964 resp.response_length += sizeof(resp.cqe_version);
966 if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
968 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
969 resp.hca_core_clock_offset =
970 offsetof(struct mlx5_init_seg, internal_timer_h) %
972 resp.response_length += sizeof(resp.hca_core_clock_offset) +
973 sizeof(resp.reserved2) +
974 sizeof(resp.reserved3);
977 err = ib_copy_to_udata(udata, &resp, resp.response_length);
982 uuari->num_low_latency_uuars = req.num_low_latency_uuars;
984 uuari->num_uars = num_uars;
985 context->cqe_version = resp.cqe_version;
987 return &context->ibucontext;
990 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
991 mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
994 for (i--; i >= 0; i--)
995 mlx5_cmd_free_uar(dev->mdev, uars[i].index);
1000 kfree(uuari->bitmap);
1007 return ERR_PTR(err);
1010 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1012 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1013 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1014 struct mlx5_uuar_info *uuari = &context->uuari;
1017 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1018 mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1020 for (i = 0; i < uuari->num_uars; i++) {
1021 if (mlx5_cmd_free_uar(dev->mdev, uuari->uars[i].index))
1022 mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
1025 kfree(uuari->count);
1026 kfree(uuari->bitmap);
1033 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
1035 return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + index;
1038 static int get_command(unsigned long offset)
1040 return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1043 static int get_arg(unsigned long offset)
1045 return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1048 static int get_index(unsigned long offset)
1050 return get_arg(offset);
1053 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1055 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1056 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1057 struct mlx5_uuar_info *uuari = &context->uuari;
1058 unsigned long command;
1062 command = get_command(vma->vm_pgoff);
1064 case MLX5_IB_MMAP_REGULAR_PAGE:
1065 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1068 idx = get_index(vma->vm_pgoff);
1069 if (idx >= uuari->num_uars)
1072 pfn = uar_index2pfn(dev, uuari->uars[idx].index);
1073 mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn 0x%llx\n", idx,
1074 (unsigned long long)pfn);
1076 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
1077 if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1078 PAGE_SIZE, vma->vm_page_prot))
1081 mlx5_ib_dbg(dev, "mapped WC at 0x%lx, PA 0x%llx\n",
1083 (unsigned long long)pfn << PAGE_SHIFT);
1086 case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1089 case MLX5_IB_MMAP_CORE_CLOCK:
1090 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1093 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1096 /* Don't expose to user-space information it shouldn't have */
1097 if (PAGE_SIZE > 4096)
1100 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1101 pfn = (dev->mdev->iseg_base +
1102 offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1104 if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1105 PAGE_SIZE, vma->vm_page_prot))
1108 mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
1110 (unsigned long long)pfn << PAGE_SHIFT);
1120 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1121 struct ib_ucontext *context,
1122 struct ib_udata *udata)
1124 struct mlx5_ib_alloc_pd_resp resp;
1125 struct mlx5_ib_pd *pd;
1128 pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1130 return ERR_PTR(-ENOMEM);
1132 err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1135 return ERR_PTR(err);
1140 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1141 mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1143 return ERR_PTR(-EFAULT);
1150 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1152 struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1153 struct mlx5_ib_pd *mpd = to_mpd(pd);
1155 mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1161 static bool outer_header_zero(u32 *match_criteria)
1163 int size = MLX5_ST_SZ_BYTES(fte_match_param);
1164 char *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_criteria,
1167 return outer_headers_c[0] == 0 && !memcmp(outer_headers_c,
1168 outer_headers_c + 1,
1172 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1173 union ib_flow_spec *ib_spec)
1175 void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1177 void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1179 switch (ib_spec->type) {
1180 case IB_FLOW_SPEC_ETH:
1181 if (ib_spec->size != sizeof(ib_spec->eth))
1184 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1186 ib_spec->eth.mask.dst_mac);
1187 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1189 ib_spec->eth.val.dst_mac);
1191 if (ib_spec->eth.mask.vlan_tag) {
1192 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1194 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1197 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1198 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1199 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1200 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1202 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1204 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1205 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1207 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1209 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1211 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1212 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1214 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1216 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1217 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1218 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1219 ethertype, ntohs(ib_spec->eth.val.ether_type));
1221 case IB_FLOW_SPEC_IPV4:
1222 if (ib_spec->size != sizeof(ib_spec->ipv4))
1225 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1227 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1228 ethertype, ETH_P_IP);
1230 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1231 src_ipv4_src_ipv6.ipv4_layout.ipv4),
1232 &ib_spec->ipv4.mask.src_ip,
1233 sizeof(ib_spec->ipv4.mask.src_ip));
1234 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1235 src_ipv4_src_ipv6.ipv4_layout.ipv4),
1236 &ib_spec->ipv4.val.src_ip,
1237 sizeof(ib_spec->ipv4.val.src_ip));
1238 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1239 dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1240 &ib_spec->ipv4.mask.dst_ip,
1241 sizeof(ib_spec->ipv4.mask.dst_ip));
1242 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1243 dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1244 &ib_spec->ipv4.val.dst_ip,
1245 sizeof(ib_spec->ipv4.val.dst_ip));
1247 case IB_FLOW_SPEC_TCP:
1248 if (ib_spec->size != sizeof(ib_spec->tcp_udp))
1251 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1253 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1256 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport,
1257 ntohs(ib_spec->tcp_udp.mask.src_port));
1258 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport,
1259 ntohs(ib_spec->tcp_udp.val.src_port));
1261 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport,
1262 ntohs(ib_spec->tcp_udp.mask.dst_port));
1263 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport,
1264 ntohs(ib_spec->tcp_udp.val.dst_port));
1266 case IB_FLOW_SPEC_UDP:
1267 if (ib_spec->size != sizeof(ib_spec->tcp_udp))
1270 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1272 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1275 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport,
1276 ntohs(ib_spec->tcp_udp.mask.src_port));
1277 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport,
1278 ntohs(ib_spec->tcp_udp.val.src_port));
1280 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport,
1281 ntohs(ib_spec->tcp_udp.mask.dst_port));
1282 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport,
1283 ntohs(ib_spec->tcp_udp.val.dst_port));
1292 /* If a flow could catch both multicast and unicast packets,
1293 * it won't fall into the multicast flow steering table and this rule
1294 * could steal other multicast packets.
1296 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1298 struct ib_flow_spec_eth *eth_spec;
1300 if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1301 ib_attr->size < sizeof(struct ib_flow_attr) +
1302 sizeof(struct ib_flow_spec_eth) ||
1303 ib_attr->num_of_specs < 1)
1306 eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1307 if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1308 eth_spec->size != sizeof(*eth_spec))
1311 return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1312 is_multicast_ether_addr(eth_spec->val.dst_mac);
1315 static bool is_valid_attr(struct ib_flow_attr *flow_attr)
1317 union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1318 bool has_ipv4_spec = false;
1319 bool eth_type_ipv4 = true;
1320 unsigned int spec_index;
1322 /* Validate that ethertype is correct */
1323 for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1324 if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1325 ib_spec->eth.mask.ether_type) {
1326 if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1327 ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1328 eth_type_ipv4 = false;
1329 } else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1330 has_ipv4_spec = true;
1332 ib_spec = (void *)ib_spec + ib_spec->size;
1334 return !has_ipv4_spec || eth_type_ipv4;
1337 static void put_flow_table(struct mlx5_ib_dev *dev,
1338 struct mlx5_ib_flow_prio *prio, bool ft_added)
1340 prio->refcount -= !!ft_added;
1341 if (!prio->refcount) {
1342 mlx5_destroy_flow_table(prio->flow_table);
1343 prio->flow_table = NULL;
1347 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
1349 struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
1350 struct mlx5_ib_flow_handler *handler = container_of(flow_id,
1351 struct mlx5_ib_flow_handler,
1353 struct mlx5_ib_flow_handler *iter, *tmp;
1355 mutex_lock(&dev->flow_db.lock);
1357 list_for_each_entry_safe(iter, tmp, &handler->list, list) {
1358 mlx5_del_flow_rule(iter->rule);
1359 list_del(&iter->list);
1363 mlx5_del_flow_rule(handler->rule);
1364 put_flow_table(dev, &dev->flow_db.prios[handler->prio], true);
1365 mutex_unlock(&dev->flow_db.lock);
1372 #define MLX5_FS_MAX_TYPES 10
1373 #define MLX5_FS_MAX_ENTRIES 32000UL
1374 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
1375 struct ib_flow_attr *flow_attr)
1377 struct mlx5_flow_namespace *ns = NULL;
1378 struct mlx5_ib_flow_prio *prio;
1379 struct mlx5_flow_table *ft;
1385 if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1386 if (flow_is_multicast_only(flow_attr))
1387 priority = MLX5_IB_FLOW_MCAST_PRIO;
1389 priority = flow_attr->priority;
1390 ns = mlx5_get_flow_namespace(dev->mdev,
1391 MLX5_FLOW_NAMESPACE_BYPASS);
1392 num_entries = MLX5_FS_MAX_ENTRIES;
1393 num_groups = MLX5_FS_MAX_TYPES;
1394 prio = &dev->flow_db.prios[priority];
1395 } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1396 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1397 ns = mlx5_get_flow_namespace(dev->mdev,
1398 MLX5_FLOW_NAMESPACE_LEFTOVERS);
1399 build_leftovers_ft_param(&priority,
1402 prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
1406 return ERR_PTR(-ENOTSUPP);
1408 ft = prio->flow_table;
1410 ft = mlx5_create_auto_grouped_flow_table(ns, priority,
1416 prio->flow_table = ft;
1422 return err ? ERR_PTR(err) : prio;
1425 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
1426 struct mlx5_ib_flow_prio *ft_prio,
1427 struct ib_flow_attr *flow_attr,
1428 struct mlx5_flow_destination *dst)
1430 struct mlx5_flow_table *ft = ft_prio->flow_table;
1431 struct mlx5_ib_flow_handler *handler;
1432 void *ib_flow = flow_attr + 1;
1433 u8 match_criteria_enable = 0;
1434 unsigned int spec_index;
1439 if (!is_valid_attr(flow_attr))
1440 return ERR_PTR(-EINVAL);
1442 match_c = kzalloc(MLX5_ST_SZ_BYTES(fte_match_param), GFP_KERNEL);
1443 match_v = kzalloc(MLX5_ST_SZ_BYTES(fte_match_param), GFP_KERNEL);
1444 handler = kzalloc(sizeof(*handler), GFP_KERNEL);
1445 if (!handler || !match_c || !match_v) {
1450 INIT_LIST_HEAD(&handler->list);
1452 for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1453 err = parse_flow_attr(match_c, match_v, ib_flow);
1457 ib_flow += ((union ib_flow_spec *)ib_flow)->size;
1460 /* Outer header support only */
1461 match_criteria_enable = (!outer_header_zero(match_c)) << 0;
1462 handler->rule = mlx5_add_flow_rule(ft, match_criteria_enable,
1464 MLX5_FLOW_CONTEXT_ACTION_FWD_DEST,
1465 MLX5_FS_DEFAULT_FLOW_TAG,
1468 if (IS_ERR(handler->rule)) {
1469 err = PTR_ERR(handler->rule);
1473 handler->prio = ft_prio - dev->flow_db.prios;
1475 ft_prio->flow_table = ft;
1481 return err ? ERR_PTR(err) : handler;
1489 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
1490 struct mlx5_ib_flow_prio *ft_prio,
1491 struct ib_flow_attr *flow_attr,
1492 struct mlx5_flow_destination *dst)
1494 struct mlx5_ib_flow_handler *handler_ucast = NULL;
1495 struct mlx5_ib_flow_handler *handler = NULL;
1498 struct ib_flow_attr flow_attr;
1499 struct ib_flow_spec_eth eth_flow;
1500 } leftovers_specs[] = {
1504 .size = sizeof(leftovers_specs[0])
1507 .type = IB_FLOW_SPEC_ETH,
1508 .size = sizeof(struct ib_flow_spec_eth),
1509 .mask = {.dst_mac = {0x1} },
1510 .val = {.dst_mac = {0x1} }
1516 .size = sizeof(leftovers_specs[0])
1519 .type = IB_FLOW_SPEC_ETH,
1520 .size = sizeof(struct ib_flow_spec_eth),
1521 .mask = {.dst_mac = {0x1} },
1522 .val = {.dst_mac = {} }
1527 handler = create_flow_rule(dev, ft_prio,
1528 &leftovers_specs[LEFTOVERS_MC].flow_attr,
1530 if (!IS_ERR(handler) &&
1531 flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
1532 handler_ucast = create_flow_rule(dev, ft_prio,
1533 &leftovers_specs[LEFTOVERS_UC].flow_attr,
1535 if (IS_ERR(handler_ucast)) {
1537 handler = handler_ucast;
1539 list_add(&handler_ucast->list, &handler->list);
1546 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
1547 struct ib_flow_attr *flow_attr,
1550 struct mlx5_ib_dev *dev = to_mdev(qp->device);
1551 struct mlx5_ib_flow_handler *handler = NULL;
1552 struct mlx5_flow_destination *dst = NULL;
1553 struct mlx5_ib_flow_prio *ft_prio;
1556 if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
1557 return ERR_PTR(-ENOSPC);
1559 if (domain != IB_FLOW_DOMAIN_USER ||
1560 flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
1562 return ERR_PTR(-EINVAL);
1564 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1566 return ERR_PTR(-ENOMEM);
1568 mutex_lock(&dev->flow_db.lock);
1570 ft_prio = get_flow_table(dev, flow_attr);
1571 if (IS_ERR(ft_prio)) {
1572 err = PTR_ERR(ft_prio);
1576 dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
1577 dst->tir_num = to_mqp(qp)->raw_packet_qp.rq.tirn;
1579 if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1580 handler = create_flow_rule(dev, ft_prio, flow_attr,
1582 } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1583 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1584 handler = create_leftovers_rule(dev, ft_prio, flow_attr,
1591 if (IS_ERR(handler)) {
1592 err = PTR_ERR(handler);
1597 ft_prio->refcount++;
1598 mutex_unlock(&dev->flow_db.lock);
1601 return &handler->ibflow;
1604 put_flow_table(dev, ft_prio, false);
1606 mutex_unlock(&dev->flow_db.lock);
1609 return ERR_PTR(err);
1612 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
1614 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1617 err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
1619 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
1620 ibqp->qp_num, gid->raw);
1625 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
1627 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1630 err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
1632 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
1633 ibqp->qp_num, gid->raw);
1638 static int init_node_data(struct mlx5_ib_dev *dev)
1642 err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
1646 dev->mdev->rev_id = dev->mdev->pdev->revision;
1648 return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
1651 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
1654 struct mlx5_ib_dev *dev =
1655 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1657 return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
1660 static ssize_t show_reg_pages(struct device *device,
1661 struct device_attribute *attr, char *buf)
1663 struct mlx5_ib_dev *dev =
1664 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1666 return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
1669 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
1672 struct mlx5_ib_dev *dev =
1673 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1674 return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
1677 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
1680 struct mlx5_ib_dev *dev =
1681 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1682 return sprintf(buf, "%d.%d.%d\n", fw_rev_maj(dev->mdev),
1683 fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
1686 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
1689 struct mlx5_ib_dev *dev =
1690 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1691 return sprintf(buf, "%x\n", dev->mdev->rev_id);
1694 static ssize_t show_board(struct device *device, struct device_attribute *attr,
1697 struct mlx5_ib_dev *dev =
1698 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1699 return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
1700 dev->mdev->board_id);
1703 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
1704 static DEVICE_ATTR(fw_ver, S_IRUGO, show_fw_ver, NULL);
1705 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
1706 static DEVICE_ATTR(board_id, S_IRUGO, show_board, NULL);
1707 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
1708 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
1710 static struct device_attribute *mlx5_class_attributes[] = {
1716 &dev_attr_reg_pages,
1719 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
1720 enum mlx5_dev_event event, unsigned long param)
1722 struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
1723 struct ib_event ibev;
1728 case MLX5_DEV_EVENT_SYS_ERROR:
1729 ibdev->ib_active = false;
1730 ibev.event = IB_EVENT_DEVICE_FATAL;
1733 case MLX5_DEV_EVENT_PORT_UP:
1734 ibev.event = IB_EVENT_PORT_ACTIVE;
1738 case MLX5_DEV_EVENT_PORT_DOWN:
1739 ibev.event = IB_EVENT_PORT_ERR;
1743 case MLX5_DEV_EVENT_PORT_INITIALIZED:
1744 /* not used by ULPs */
1747 case MLX5_DEV_EVENT_LID_CHANGE:
1748 ibev.event = IB_EVENT_LID_CHANGE;
1752 case MLX5_DEV_EVENT_PKEY_CHANGE:
1753 ibev.event = IB_EVENT_PKEY_CHANGE;
1757 case MLX5_DEV_EVENT_GUID_CHANGE:
1758 ibev.event = IB_EVENT_GID_CHANGE;
1762 case MLX5_DEV_EVENT_CLIENT_REREG:
1763 ibev.event = IB_EVENT_CLIENT_REREGISTER;
1768 ibev.device = &ibdev->ib_dev;
1769 ibev.element.port_num = port;
1771 if (port < 1 || port > ibdev->num_ports) {
1772 mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
1776 if (ibdev->ib_active)
1777 ib_dispatch_event(&ibev);
1780 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
1784 for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
1785 mlx5_query_ext_port_caps(dev, port);
1788 static int get_port_caps(struct mlx5_ib_dev *dev)
1790 struct ib_device_attr *dprops = NULL;
1791 struct ib_port_attr *pprops = NULL;
1794 struct ib_udata uhw = {.inlen = 0, .outlen = 0};
1796 pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
1800 dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
1804 err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
1806 mlx5_ib_warn(dev, "query_device failed %d\n", err);
1810 for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
1811 err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
1813 mlx5_ib_warn(dev, "query_port %d failed %d\n",
1817 dev->mdev->port_caps[port - 1].pkey_table_len =
1819 dev->mdev->port_caps[port - 1].gid_table_len =
1820 pprops->gid_tbl_len;
1821 mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
1822 dprops->max_pkeys, pprops->gid_tbl_len);
1832 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
1836 err = mlx5_mr_cache_cleanup(dev);
1838 mlx5_ib_warn(dev, "mr cache cleanup failed\n");
1840 mlx5_ib_destroy_qp(dev->umrc.qp);
1841 ib_destroy_cq(dev->umrc.cq);
1842 ib_dealloc_pd(dev->umrc.pd);
1849 static int create_umr_res(struct mlx5_ib_dev *dev)
1851 struct ib_qp_init_attr *init_attr = NULL;
1852 struct ib_qp_attr *attr = NULL;
1856 struct ib_cq_init_attr cq_attr = {};
1859 attr = kzalloc(sizeof(*attr), GFP_KERNEL);
1860 init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
1861 if (!attr || !init_attr) {
1866 pd = ib_alloc_pd(&dev->ib_dev);
1868 mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
1874 cq = ib_create_cq(&dev->ib_dev, mlx5_umr_cq_handler, NULL, NULL,
1877 mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
1881 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1883 init_attr->send_cq = cq;
1884 init_attr->recv_cq = cq;
1885 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
1886 init_attr->cap.max_send_wr = MAX_UMR_WR;
1887 init_attr->cap.max_send_sge = 1;
1888 init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
1889 init_attr->port_num = 1;
1890 qp = mlx5_ib_create_qp(pd, init_attr, NULL);
1892 mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
1896 qp->device = &dev->ib_dev;
1899 qp->qp_type = MLX5_IB_QPT_REG_UMR;
1901 attr->qp_state = IB_QPS_INIT;
1903 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
1906 mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
1910 memset(attr, 0, sizeof(*attr));
1911 attr->qp_state = IB_QPS_RTR;
1912 attr->path_mtu = IB_MTU_256;
1914 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
1916 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
1920 memset(attr, 0, sizeof(*attr));
1921 attr->qp_state = IB_QPS_RTS;
1922 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
1924 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
1932 sema_init(&dev->umrc.sem, MAX_UMR_WR);
1933 ret = mlx5_mr_cache_init(dev);
1935 mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
1945 mlx5_ib_destroy_qp(qp);
1959 static int create_dev_resources(struct mlx5_ib_resources *devr)
1961 struct ib_srq_init_attr attr;
1962 struct mlx5_ib_dev *dev;
1963 struct ib_cq_init_attr cq_attr = {.cqe = 1};
1966 dev = container_of(devr, struct mlx5_ib_dev, devr);
1968 devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
1969 if (IS_ERR(devr->p0)) {
1970 ret = PTR_ERR(devr->p0);
1973 devr->p0->device = &dev->ib_dev;
1974 devr->p0->uobject = NULL;
1975 atomic_set(&devr->p0->usecnt, 0);
1977 devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
1978 if (IS_ERR(devr->c0)) {
1979 ret = PTR_ERR(devr->c0);
1982 devr->c0->device = &dev->ib_dev;
1983 devr->c0->uobject = NULL;
1984 devr->c0->comp_handler = NULL;
1985 devr->c0->event_handler = NULL;
1986 devr->c0->cq_context = NULL;
1987 atomic_set(&devr->c0->usecnt, 0);
1989 devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
1990 if (IS_ERR(devr->x0)) {
1991 ret = PTR_ERR(devr->x0);
1994 devr->x0->device = &dev->ib_dev;
1995 devr->x0->inode = NULL;
1996 atomic_set(&devr->x0->usecnt, 0);
1997 mutex_init(&devr->x0->tgt_qp_mutex);
1998 INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2000 devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2001 if (IS_ERR(devr->x1)) {
2002 ret = PTR_ERR(devr->x1);
2005 devr->x1->device = &dev->ib_dev;
2006 devr->x1->inode = NULL;
2007 atomic_set(&devr->x1->usecnt, 0);
2008 mutex_init(&devr->x1->tgt_qp_mutex);
2009 INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2011 memset(&attr, 0, sizeof(attr));
2012 attr.attr.max_sge = 1;
2013 attr.attr.max_wr = 1;
2014 attr.srq_type = IB_SRQT_XRC;
2015 attr.ext.xrc.cq = devr->c0;
2016 attr.ext.xrc.xrcd = devr->x0;
2018 devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2019 if (IS_ERR(devr->s0)) {
2020 ret = PTR_ERR(devr->s0);
2023 devr->s0->device = &dev->ib_dev;
2024 devr->s0->pd = devr->p0;
2025 devr->s0->uobject = NULL;
2026 devr->s0->event_handler = NULL;
2027 devr->s0->srq_context = NULL;
2028 devr->s0->srq_type = IB_SRQT_XRC;
2029 devr->s0->ext.xrc.xrcd = devr->x0;
2030 devr->s0->ext.xrc.cq = devr->c0;
2031 atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2032 atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2033 atomic_inc(&devr->p0->usecnt);
2034 atomic_set(&devr->s0->usecnt, 0);
2036 memset(&attr, 0, sizeof(attr));
2037 attr.attr.max_sge = 1;
2038 attr.attr.max_wr = 1;
2039 attr.srq_type = IB_SRQT_BASIC;
2040 devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2041 if (IS_ERR(devr->s1)) {
2042 ret = PTR_ERR(devr->s1);
2045 devr->s1->device = &dev->ib_dev;
2046 devr->s1->pd = devr->p0;
2047 devr->s1->uobject = NULL;
2048 devr->s1->event_handler = NULL;
2049 devr->s1->srq_context = NULL;
2050 devr->s1->srq_type = IB_SRQT_BASIC;
2051 devr->s1->ext.xrc.cq = devr->c0;
2052 atomic_inc(&devr->p0->usecnt);
2053 atomic_set(&devr->s0->usecnt, 0);
2058 mlx5_ib_destroy_srq(devr->s0);
2060 mlx5_ib_dealloc_xrcd(devr->x1);
2062 mlx5_ib_dealloc_xrcd(devr->x0);
2064 mlx5_ib_destroy_cq(devr->c0);
2066 mlx5_ib_dealloc_pd(devr->p0);
2071 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2073 mlx5_ib_destroy_srq(devr->s1);
2074 mlx5_ib_destroy_srq(devr->s0);
2075 mlx5_ib_dealloc_xrcd(devr->x0);
2076 mlx5_ib_dealloc_xrcd(devr->x1);
2077 mlx5_ib_destroy_cq(devr->c0);
2078 mlx5_ib_dealloc_pd(devr->p0);
2081 static u32 get_core_cap_flags(struct ib_device *ibdev)
2083 struct mlx5_ib_dev *dev = to_mdev(ibdev);
2084 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2085 u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2086 u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2089 if (ll == IB_LINK_LAYER_INFINIBAND)
2090 return RDMA_CORE_PORT_IBA_IB;
2092 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2095 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2098 if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2099 ret |= RDMA_CORE_PORT_IBA_ROCE;
2101 if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2102 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2107 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
2108 struct ib_port_immutable *immutable)
2110 struct ib_port_attr attr;
2113 err = mlx5_ib_query_port(ibdev, port_num, &attr);
2117 immutable->pkey_tbl_len = attr.pkey_tbl_len;
2118 immutable->gid_tbl_len = attr.gid_tbl_len;
2119 immutable->core_cap_flags = get_core_cap_flags(ibdev);
2120 immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2125 static int mlx5_enable_roce(struct mlx5_ib_dev *dev)
2129 dev->roce.nb.notifier_call = mlx5_netdev_event;
2130 err = register_netdevice_notifier(&dev->roce.nb);
2134 err = mlx5_nic_vport_enable_roce(dev->mdev);
2136 goto err_unregister_netdevice_notifier;
2140 err_unregister_netdevice_notifier:
2141 unregister_netdevice_notifier(&dev->roce.nb);
2145 static void mlx5_disable_roce(struct mlx5_ib_dev *dev)
2147 mlx5_nic_vport_disable_roce(dev->mdev);
2148 unregister_netdevice_notifier(&dev->roce.nb);
2151 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
2153 struct mlx5_ib_dev *dev;
2154 enum rdma_link_layer ll;
2159 port_type_cap = MLX5_CAP_GEN(mdev, port_type);
2160 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
2162 if ((ll == IB_LINK_LAYER_ETHERNET) && !MLX5_CAP_GEN(mdev, roce))
2165 printk_once(KERN_INFO "%s", mlx5_version);
2167 dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
2173 rwlock_init(&dev->roce.netdev_lock);
2174 err = get_port_caps(dev);
2178 if (mlx5_use_mad_ifc(dev))
2179 get_ext_port_caps(dev);
2181 MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
2183 strlcpy(dev->ib_dev.name, "mlx5_%d", IB_DEVICE_NAME_MAX);
2184 dev->ib_dev.owner = THIS_MODULE;
2185 dev->ib_dev.node_type = RDMA_NODE_IB_CA;
2186 dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
2187 dev->num_ports = MLX5_CAP_GEN(mdev, num_ports);
2188 dev->ib_dev.phys_port_cnt = dev->num_ports;
2189 dev->ib_dev.num_comp_vectors =
2190 dev->mdev->priv.eq_table.num_comp_vectors;
2191 dev->ib_dev.dma_device = &mdev->pdev->dev;
2193 dev->ib_dev.uverbs_abi_ver = MLX5_IB_UVERBS_ABI_VERSION;
2194 dev->ib_dev.uverbs_cmd_mask =
2195 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
2196 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
2197 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
2198 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
2199 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
2200 (1ull << IB_USER_VERBS_CMD_REG_MR) |
2201 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
2202 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2203 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
2204 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
2205 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
2206 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
2207 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
2208 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
2209 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
2210 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
2211 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) |
2212 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
2213 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
2214 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
2215 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) |
2216 (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ) |
2217 (1ull << IB_USER_VERBS_CMD_OPEN_QP);
2218 dev->ib_dev.uverbs_ex_cmd_mask =
2219 (1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE) |
2220 (1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ) |
2221 (1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
2223 dev->ib_dev.query_device = mlx5_ib_query_device;
2224 dev->ib_dev.query_port = mlx5_ib_query_port;
2225 dev->ib_dev.get_link_layer = mlx5_ib_port_link_layer;
2226 if (ll == IB_LINK_LAYER_ETHERNET)
2227 dev->ib_dev.get_netdev = mlx5_ib_get_netdev;
2228 dev->ib_dev.query_gid = mlx5_ib_query_gid;
2229 dev->ib_dev.add_gid = mlx5_ib_add_gid;
2230 dev->ib_dev.del_gid = mlx5_ib_del_gid;
2231 dev->ib_dev.query_pkey = mlx5_ib_query_pkey;
2232 dev->ib_dev.modify_device = mlx5_ib_modify_device;
2233 dev->ib_dev.modify_port = mlx5_ib_modify_port;
2234 dev->ib_dev.alloc_ucontext = mlx5_ib_alloc_ucontext;
2235 dev->ib_dev.dealloc_ucontext = mlx5_ib_dealloc_ucontext;
2236 dev->ib_dev.mmap = mlx5_ib_mmap;
2237 dev->ib_dev.alloc_pd = mlx5_ib_alloc_pd;
2238 dev->ib_dev.dealloc_pd = mlx5_ib_dealloc_pd;
2239 dev->ib_dev.create_ah = mlx5_ib_create_ah;
2240 dev->ib_dev.query_ah = mlx5_ib_query_ah;
2241 dev->ib_dev.destroy_ah = mlx5_ib_destroy_ah;
2242 dev->ib_dev.create_srq = mlx5_ib_create_srq;
2243 dev->ib_dev.modify_srq = mlx5_ib_modify_srq;
2244 dev->ib_dev.query_srq = mlx5_ib_query_srq;
2245 dev->ib_dev.destroy_srq = mlx5_ib_destroy_srq;
2246 dev->ib_dev.post_srq_recv = mlx5_ib_post_srq_recv;
2247 dev->ib_dev.create_qp = mlx5_ib_create_qp;
2248 dev->ib_dev.modify_qp = mlx5_ib_modify_qp;
2249 dev->ib_dev.query_qp = mlx5_ib_query_qp;
2250 dev->ib_dev.destroy_qp = mlx5_ib_destroy_qp;
2251 dev->ib_dev.post_send = mlx5_ib_post_send;
2252 dev->ib_dev.post_recv = mlx5_ib_post_recv;
2253 dev->ib_dev.create_cq = mlx5_ib_create_cq;
2254 dev->ib_dev.modify_cq = mlx5_ib_modify_cq;
2255 dev->ib_dev.resize_cq = mlx5_ib_resize_cq;
2256 dev->ib_dev.destroy_cq = mlx5_ib_destroy_cq;
2257 dev->ib_dev.poll_cq = mlx5_ib_poll_cq;
2258 dev->ib_dev.req_notify_cq = mlx5_ib_arm_cq;
2259 dev->ib_dev.get_dma_mr = mlx5_ib_get_dma_mr;
2260 dev->ib_dev.reg_user_mr = mlx5_ib_reg_user_mr;
2261 dev->ib_dev.dereg_mr = mlx5_ib_dereg_mr;
2262 dev->ib_dev.attach_mcast = mlx5_ib_mcg_attach;
2263 dev->ib_dev.detach_mcast = mlx5_ib_mcg_detach;
2264 dev->ib_dev.process_mad = mlx5_ib_process_mad;
2265 dev->ib_dev.alloc_mr = mlx5_ib_alloc_mr;
2266 dev->ib_dev.map_mr_sg = mlx5_ib_map_mr_sg;
2267 dev->ib_dev.check_mr_status = mlx5_ib_check_mr_status;
2268 dev->ib_dev.get_port_immutable = mlx5_port_immutable;
2270 mlx5_ib_internal_fill_odp_caps(dev);
2272 if (MLX5_CAP_GEN(mdev, xrc)) {
2273 dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
2274 dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
2275 dev->ib_dev.uverbs_cmd_mask |=
2276 (1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
2277 (1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
2280 if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
2281 IB_LINK_LAYER_ETHERNET) {
2282 dev->ib_dev.create_flow = mlx5_ib_create_flow;
2283 dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
2284 dev->ib_dev.uverbs_ex_cmd_mask |=
2285 (1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
2286 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
2288 err = init_node_data(dev);
2292 mutex_init(&dev->flow_db.lock);
2293 mutex_init(&dev->cap_mask_mutex);
2295 if (ll == IB_LINK_LAYER_ETHERNET) {
2296 err = mlx5_enable_roce(dev);
2301 err = create_dev_resources(&dev->devr);
2303 goto err_disable_roce;
2305 err = mlx5_ib_odp_init_one(dev);
2309 err = ib_register_device(&dev->ib_dev, NULL);
2313 err = create_umr_res(dev);
2317 for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
2318 err = device_create_file(&dev->ib_dev.dev,
2319 mlx5_class_attributes[i]);
2324 dev->ib_active = true;
2329 destroy_umrc_res(dev);
2332 ib_unregister_device(&dev->ib_dev);
2335 mlx5_ib_odp_remove_one(dev);
2338 destroy_dev_resources(&dev->devr);
2341 if (ll == IB_LINK_LAYER_ETHERNET)
2342 mlx5_disable_roce(dev);
2345 ib_dealloc_device((struct ib_device *)dev);
2350 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
2352 struct mlx5_ib_dev *dev = context;
2353 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
2355 ib_unregister_device(&dev->ib_dev);
2356 destroy_umrc_res(dev);
2357 mlx5_ib_odp_remove_one(dev);
2358 destroy_dev_resources(&dev->devr);
2359 if (ll == IB_LINK_LAYER_ETHERNET)
2360 mlx5_disable_roce(dev);
2361 ib_dealloc_device(&dev->ib_dev);
2364 static struct mlx5_interface mlx5_ib_interface = {
2366 .remove = mlx5_ib_remove,
2367 .event = mlx5_ib_event,
2368 .protocol = MLX5_INTERFACE_PROTOCOL_IB,
2371 static int __init mlx5_ib_init(void)
2375 if (deprecated_prof_sel != 2)
2376 pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
2378 err = mlx5_ib_odp_init();
2382 err = mlx5_register_interface(&mlx5_ib_interface);
2389 mlx5_ib_odp_cleanup();
2393 static void __exit mlx5_ib_cleanup(void)
2395 mlx5_unregister_interface(&mlx5_ib_interface);
2396 mlx5_ib_odp_cleanup();
2399 module_init(mlx5_ib_init);
2400 module_exit(mlx5_ib_cleanup);