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/debugfs.h>
34 #include <linux/highmem.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/errno.h>
38 #include <linux/pci.h>
39 #include <linux/dma-mapping.h>
40 #include <linux/slab.h>
41 #if defined(CONFIG_X86)
44 #include <linux/sched.h>
45 #include <linux/sched/mm.h>
46 #include <linux/sched/task.h>
47 #include <linux/delay.h>
48 #include <rdma/ib_user_verbs.h>
49 #include <rdma/ib_addr.h>
50 #include <rdma/ib_cache.h>
51 #include <linux/mlx5/port.h>
52 #include <linux/mlx5/vport.h>
53 #include <linux/list.h>
54 #include <rdma/ib_smi.h>
55 #include <rdma/ib_umem.h>
57 #include <linux/etherdevice.h>
58 #include <linux/mlx5/fs.h>
59 #include <linux/mlx5/vport.h>
62 #include <linux/mlx5/vport.h>
64 #define DRIVER_NAME "mlx5_ib"
65 #define DRIVER_VERSION "5.0-0"
67 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
68 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
69 MODULE_LICENSE("Dual BSD/GPL");
71 static char mlx5_version[] =
72 DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
76 MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
79 static enum rdma_link_layer
80 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
82 switch (port_type_cap) {
83 case MLX5_CAP_PORT_TYPE_IB:
84 return IB_LINK_LAYER_INFINIBAND;
85 case MLX5_CAP_PORT_TYPE_ETH:
86 return IB_LINK_LAYER_ETHERNET;
88 return IB_LINK_LAYER_UNSPECIFIED;
92 static enum rdma_link_layer
93 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
95 struct mlx5_ib_dev *dev = to_mdev(device);
96 int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
98 return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
101 static int get_port_state(struct ib_device *ibdev,
103 enum ib_port_state *state)
105 struct ib_port_attr attr;
108 memset(&attr, 0, sizeof(attr));
109 ret = mlx5_ib_query_port(ibdev, port_num, &attr);
115 static int mlx5_netdev_event(struct notifier_block *this,
116 unsigned long event, void *ptr)
118 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
119 struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
123 case NETDEV_REGISTER:
124 case NETDEV_UNREGISTER:
125 write_lock(&ibdev->roce.netdev_lock);
126 if (ndev->dev.parent == &ibdev->mdev->pdev->dev)
127 ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
129 write_unlock(&ibdev->roce.netdev_lock);
135 struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
136 struct net_device *upper = NULL;
139 upper = netdev_master_upper_dev_get(lag_ndev);
143 if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
144 && ibdev->ib_active) {
145 struct ib_event ibev = { };
146 enum ib_port_state port_state;
148 if (get_port_state(&ibdev->ib_dev, 1, &port_state))
151 if (ibdev->roce.last_port_state == port_state)
154 ibdev->roce.last_port_state = port_state;
155 ibev.device = &ibdev->ib_dev;
156 if (port_state == IB_PORT_DOWN)
157 ibev.event = IB_EVENT_PORT_ERR;
158 else if (port_state == IB_PORT_ACTIVE)
159 ibev.event = IB_EVENT_PORT_ACTIVE;
163 ibev.element.port_num = 1;
164 ib_dispatch_event(&ibev);
176 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
179 struct mlx5_ib_dev *ibdev = to_mdev(device);
180 struct net_device *ndev;
182 ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
186 /* Ensure ndev does not disappear before we invoke dev_hold()
188 read_lock(&ibdev->roce.netdev_lock);
189 ndev = ibdev->roce.netdev;
192 read_unlock(&ibdev->roce.netdev_lock);
197 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
200 switch (eth_proto_oper) {
201 case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
202 case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
203 case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
204 case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
205 *active_width = IB_WIDTH_1X;
206 *active_speed = IB_SPEED_SDR;
208 case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
209 case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
210 case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
211 case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
212 case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
213 case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
214 case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
215 *active_width = IB_WIDTH_1X;
216 *active_speed = IB_SPEED_QDR;
218 case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
219 case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
220 case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
221 *active_width = IB_WIDTH_1X;
222 *active_speed = IB_SPEED_EDR;
224 case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
225 case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
226 case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
227 case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
228 *active_width = IB_WIDTH_4X;
229 *active_speed = IB_SPEED_QDR;
231 case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
232 case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
233 case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
234 *active_width = IB_WIDTH_1X;
235 *active_speed = IB_SPEED_HDR;
237 case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
238 *active_width = IB_WIDTH_4X;
239 *active_speed = IB_SPEED_FDR;
241 case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
242 case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
243 case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
244 case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
245 *active_width = IB_WIDTH_4X;
246 *active_speed = IB_SPEED_EDR;
255 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
256 struct ib_port_attr *props)
258 struct mlx5_ib_dev *dev = to_mdev(device);
259 struct mlx5_core_dev *mdev = dev->mdev;
260 struct net_device *ndev, *upper;
261 enum ib_mtu ndev_ib_mtu;
266 /* Possible bad flows are checked before filling out props so in case
267 * of an error it will still be zeroed out.
269 err = mlx5_query_port_eth_proto_oper(mdev, ð_prot_oper, port_num);
273 translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
274 &props->active_width);
276 props->port_cap_flags |= IB_PORT_CM_SUP;
277 props->port_cap_flags |= IB_PORT_IP_BASED_GIDS;
279 props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
280 roce_address_table_size);
281 props->max_mtu = IB_MTU_4096;
282 props->max_msg_sz = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
283 props->pkey_tbl_len = 1;
284 props->state = IB_PORT_DOWN;
285 props->phys_state = 3;
287 mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
288 props->qkey_viol_cntr = qkey_viol_cntr;
290 ndev = mlx5_ib_get_netdev(device, port_num);
294 if (mlx5_lag_is_active(dev->mdev)) {
296 upper = netdev_master_upper_dev_get_rcu(ndev);
305 if (netif_running(ndev) && netif_carrier_ok(ndev)) {
306 props->state = IB_PORT_ACTIVE;
307 props->phys_state = 5;
310 ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
314 props->active_mtu = min(props->max_mtu, ndev_ib_mtu);
318 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
319 unsigned int index, const union ib_gid *gid,
320 const struct ib_gid_attr *attr)
322 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
330 gid_type = attr->gid_type;
331 ether_addr_copy(mac, attr->ndev->dev_addr);
333 if (is_vlan_dev(attr->ndev)) {
335 vlan_id = vlan_dev_vlan_id(attr->ndev);
341 roce_version = MLX5_ROCE_VERSION_1;
343 case IB_GID_TYPE_ROCE_UDP_ENCAP:
344 roce_version = MLX5_ROCE_VERSION_2;
345 if (ipv6_addr_v4mapped((void *)gid))
346 roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
348 roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
352 mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
355 return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
356 roce_l3_type, gid->raw, mac, vlan,
360 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
361 unsigned int index, const union ib_gid *gid,
362 const struct ib_gid_attr *attr,
363 __always_unused void **context)
365 return set_roce_addr(to_mdev(device), port_num, index, gid, attr);
368 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
369 unsigned int index, __always_unused void **context)
371 return set_roce_addr(to_mdev(device), port_num, index, NULL, NULL);
374 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
377 struct ib_gid_attr attr;
380 if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
388 if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
391 return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
394 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
395 int index, enum ib_gid_type *gid_type)
397 struct ib_gid_attr attr;
401 ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
410 *gid_type = attr.gid_type;
415 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
417 if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
418 return !MLX5_CAP_GEN(dev->mdev, ib_virt);
423 MLX5_VPORT_ACCESS_METHOD_MAD,
424 MLX5_VPORT_ACCESS_METHOD_HCA,
425 MLX5_VPORT_ACCESS_METHOD_NIC,
428 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
430 if (mlx5_use_mad_ifc(to_mdev(ibdev)))
431 return MLX5_VPORT_ACCESS_METHOD_MAD;
433 if (mlx5_ib_port_link_layer(ibdev, 1) ==
434 IB_LINK_LAYER_ETHERNET)
435 return MLX5_VPORT_ACCESS_METHOD_NIC;
437 return MLX5_VPORT_ACCESS_METHOD_HCA;
440 static void get_atomic_caps(struct mlx5_ib_dev *dev,
441 struct ib_device_attr *props)
444 u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
445 u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
446 u8 atomic_req_8B_endianness_mode =
447 MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
449 /* Check if HW supports 8 bytes standard atomic operations and capable
450 * of host endianness respond
452 tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
453 if (((atomic_operations & tmp) == tmp) &&
454 (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
455 (atomic_req_8B_endianness_mode)) {
456 props->atomic_cap = IB_ATOMIC_HCA;
458 props->atomic_cap = IB_ATOMIC_NONE;
462 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
463 __be64 *sys_image_guid)
465 struct mlx5_ib_dev *dev = to_mdev(ibdev);
466 struct mlx5_core_dev *mdev = dev->mdev;
470 switch (mlx5_get_vport_access_method(ibdev)) {
471 case MLX5_VPORT_ACCESS_METHOD_MAD:
472 return mlx5_query_mad_ifc_system_image_guid(ibdev,
475 case MLX5_VPORT_ACCESS_METHOD_HCA:
476 err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
479 case MLX5_VPORT_ACCESS_METHOD_NIC:
480 err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
488 *sys_image_guid = cpu_to_be64(tmp);
494 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
497 struct mlx5_ib_dev *dev = to_mdev(ibdev);
498 struct mlx5_core_dev *mdev = dev->mdev;
500 switch (mlx5_get_vport_access_method(ibdev)) {
501 case MLX5_VPORT_ACCESS_METHOD_MAD:
502 return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
504 case MLX5_VPORT_ACCESS_METHOD_HCA:
505 case MLX5_VPORT_ACCESS_METHOD_NIC:
506 *max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
515 static int mlx5_query_vendor_id(struct ib_device *ibdev,
518 struct mlx5_ib_dev *dev = to_mdev(ibdev);
520 switch (mlx5_get_vport_access_method(ibdev)) {
521 case MLX5_VPORT_ACCESS_METHOD_MAD:
522 return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
524 case MLX5_VPORT_ACCESS_METHOD_HCA:
525 case MLX5_VPORT_ACCESS_METHOD_NIC:
526 return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
533 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
539 switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
540 case MLX5_VPORT_ACCESS_METHOD_MAD:
541 return mlx5_query_mad_ifc_node_guid(dev, node_guid);
543 case MLX5_VPORT_ACCESS_METHOD_HCA:
544 err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
547 case MLX5_VPORT_ACCESS_METHOD_NIC:
548 err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
556 *node_guid = cpu_to_be64(tmp);
561 struct mlx5_reg_node_desc {
562 u8 desc[IB_DEVICE_NODE_DESC_MAX];
565 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
567 struct mlx5_reg_node_desc in;
569 if (mlx5_use_mad_ifc(dev))
570 return mlx5_query_mad_ifc_node_desc(dev, node_desc);
572 memset(&in, 0, sizeof(in));
574 return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
575 sizeof(struct mlx5_reg_node_desc),
576 MLX5_REG_NODE_DESC, 0, 0);
579 static int mlx5_ib_query_device(struct ib_device *ibdev,
580 struct ib_device_attr *props,
581 struct ib_udata *uhw)
583 struct mlx5_ib_dev *dev = to_mdev(ibdev);
584 struct mlx5_core_dev *mdev = dev->mdev;
589 u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
590 struct mlx5_ib_query_device_resp resp = {};
594 resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
595 if (uhw->outlen && uhw->outlen < resp_len)
598 resp.response_length = resp_len;
600 if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
603 memset(props, 0, sizeof(*props));
604 err = mlx5_query_system_image_guid(ibdev,
605 &props->sys_image_guid);
609 err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
613 err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
617 props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
618 (fw_rev_min(dev->mdev) << 16) |
619 fw_rev_sub(dev->mdev);
620 props->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT |
621 IB_DEVICE_PORT_ACTIVE_EVENT |
622 IB_DEVICE_SYS_IMAGE_GUID |
623 IB_DEVICE_RC_RNR_NAK_GEN;
625 if (MLX5_CAP_GEN(mdev, pkv))
626 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
627 if (MLX5_CAP_GEN(mdev, qkv))
628 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
629 if (MLX5_CAP_GEN(mdev, apm))
630 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
631 if (MLX5_CAP_GEN(mdev, xrc))
632 props->device_cap_flags |= IB_DEVICE_XRC;
633 if (MLX5_CAP_GEN(mdev, imaicl)) {
634 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
635 IB_DEVICE_MEM_WINDOW_TYPE_2B;
636 props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
637 /* We support 'Gappy' memory registration too */
638 props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
640 props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
641 if (MLX5_CAP_GEN(mdev, sho)) {
642 props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
643 /* At this stage no support for signature handover */
644 props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
645 IB_PROT_T10DIF_TYPE_2 |
646 IB_PROT_T10DIF_TYPE_3;
647 props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
648 IB_GUARD_T10DIF_CSUM;
650 if (MLX5_CAP_GEN(mdev, block_lb_mc))
651 props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
653 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
654 if (MLX5_CAP_ETH(mdev, csum_cap)) {
655 /* Legacy bit to support old userspace libraries */
656 props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
657 props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
660 if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
661 props->raw_packet_caps |=
662 IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
664 if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
665 max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
667 resp.tso_caps.max_tso = 1 << max_tso;
668 resp.tso_caps.supported_qpts |=
669 1 << IB_QPT_RAW_PACKET;
670 resp.response_length += sizeof(resp.tso_caps);
674 if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
675 resp.rss_caps.rx_hash_function =
676 MLX5_RX_HASH_FUNC_TOEPLITZ;
677 resp.rss_caps.rx_hash_fields_mask =
678 MLX5_RX_HASH_SRC_IPV4 |
679 MLX5_RX_HASH_DST_IPV4 |
680 MLX5_RX_HASH_SRC_IPV6 |
681 MLX5_RX_HASH_DST_IPV6 |
682 MLX5_RX_HASH_SRC_PORT_TCP |
683 MLX5_RX_HASH_DST_PORT_TCP |
684 MLX5_RX_HASH_SRC_PORT_UDP |
685 MLX5_RX_HASH_DST_PORT_UDP;
686 resp.response_length += sizeof(resp.rss_caps);
689 if (field_avail(typeof(resp), tso_caps, uhw->outlen))
690 resp.response_length += sizeof(resp.tso_caps);
691 if (field_avail(typeof(resp), rss_caps, uhw->outlen))
692 resp.response_length += sizeof(resp.rss_caps);
695 if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
696 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
697 props->device_cap_flags |= IB_DEVICE_UD_TSO;
700 if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
701 MLX5_CAP_GEN(dev->mdev, general_notification_event))
702 props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
704 if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
705 MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
706 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
708 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
709 MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
710 /* Legacy bit to support old userspace libraries */
711 props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
712 props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
715 if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
716 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
718 if (MLX5_CAP_GEN(mdev, end_pad))
719 props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
721 props->vendor_part_id = mdev->pdev->device;
722 props->hw_ver = mdev->pdev->revision;
724 props->max_mr_size = ~0ull;
725 props->page_size_cap = ~(min_page_size - 1);
726 props->max_qp = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
727 props->max_qp_wr = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
728 max_rq_sg = MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
729 sizeof(struct mlx5_wqe_data_seg);
730 max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
731 max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
732 sizeof(struct mlx5_wqe_raddr_seg)) /
733 sizeof(struct mlx5_wqe_data_seg);
734 props->max_sge = min(max_rq_sg, max_sq_sg);
735 props->max_sge_rd = MLX5_MAX_SGE_RD;
736 props->max_cq = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
737 props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
738 props->max_mr = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
739 props->max_pd = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
740 props->max_qp_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
741 props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
742 props->max_srq = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
743 props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
744 props->local_ca_ack_delay = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
745 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp;
746 props->max_srq_sge = max_rq_sg - 1;
747 props->max_fast_reg_page_list_len =
748 1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
749 get_atomic_caps(dev, props);
750 props->masked_atomic_cap = IB_ATOMIC_NONE;
751 props->max_mcast_grp = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
752 props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
753 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
754 props->max_mcast_grp;
755 props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
756 props->max_ah = INT_MAX;
757 props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
758 props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
760 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
761 if (MLX5_CAP_GEN(mdev, pg))
762 props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
763 props->odp_caps = dev->odp_caps;
766 if (MLX5_CAP_GEN(mdev, cd))
767 props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
769 if (!mlx5_core_is_pf(mdev))
770 props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
772 if (mlx5_ib_port_link_layer(ibdev, 1) ==
773 IB_LINK_LAYER_ETHERNET) {
774 props->rss_caps.max_rwq_indirection_tables =
775 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
776 props->rss_caps.max_rwq_indirection_table_size =
777 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
778 props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
779 props->max_wq_type_rq =
780 1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
783 if (MLX5_CAP_GEN(mdev, tag_matching)) {
784 props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
785 props->tm_caps.max_num_tags =
786 (1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
787 props->tm_caps.flags = IB_TM_CAP_RC;
788 props->tm_caps.max_ops =
789 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
790 props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
793 if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
794 props->cq_caps.max_cq_moderation_count =
796 props->cq_caps.max_cq_moderation_period =
800 if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
801 resp.cqe_comp_caps.max_num =
802 MLX5_CAP_GEN(dev->mdev, cqe_compression) ?
803 MLX5_CAP_GEN(dev->mdev, cqe_compression_max_num) : 0;
804 resp.cqe_comp_caps.supported_format =
805 MLX5_IB_CQE_RES_FORMAT_HASH |
806 MLX5_IB_CQE_RES_FORMAT_CSUM;
807 resp.response_length += sizeof(resp.cqe_comp_caps);
810 if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen)) {
811 if (MLX5_CAP_QOS(mdev, packet_pacing) &&
812 MLX5_CAP_GEN(mdev, qos)) {
813 resp.packet_pacing_caps.qp_rate_limit_max =
814 MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
815 resp.packet_pacing_caps.qp_rate_limit_min =
816 MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
817 resp.packet_pacing_caps.supported_qpts |=
818 1 << IB_QPT_RAW_PACKET;
820 resp.response_length += sizeof(resp.packet_pacing_caps);
823 if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
825 if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
826 resp.mlx5_ib_support_multi_pkt_send_wqes =
829 if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
830 resp.mlx5_ib_support_multi_pkt_send_wqes |=
831 MLX5_IB_SUPPORT_EMPW;
833 resp.response_length +=
834 sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
837 if (field_avail(typeof(resp), flags, uhw->outlen)) {
838 resp.response_length += sizeof(resp.flags);
840 if (MLX5_CAP_GEN(mdev, cqe_compression_128))
842 MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
844 if (MLX5_CAP_GEN(mdev, cqe_128_always))
845 resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
848 if (field_avail(typeof(resp), sw_parsing_caps,
850 resp.response_length += sizeof(resp.sw_parsing_caps);
851 if (MLX5_CAP_ETH(mdev, swp)) {
852 resp.sw_parsing_caps.sw_parsing_offloads |=
855 if (MLX5_CAP_ETH(mdev, swp_csum))
856 resp.sw_parsing_caps.sw_parsing_offloads |=
857 MLX5_IB_SW_PARSING_CSUM;
859 if (MLX5_CAP_ETH(mdev, swp_lso))
860 resp.sw_parsing_caps.sw_parsing_offloads |=
861 MLX5_IB_SW_PARSING_LSO;
863 if (resp.sw_parsing_caps.sw_parsing_offloads)
864 resp.sw_parsing_caps.supported_qpts =
865 BIT(IB_QPT_RAW_PACKET);
869 if (field_avail(typeof(resp), striding_rq_caps, uhw->outlen)) {
870 resp.response_length += sizeof(resp.striding_rq_caps);
871 if (MLX5_CAP_GEN(mdev, striding_rq)) {
872 resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
873 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
874 resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
875 MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
876 resp.striding_rq_caps.min_single_wqe_log_num_of_strides =
877 MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
878 resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
879 MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
880 resp.striding_rq_caps.supported_qpts =
881 BIT(IB_QPT_RAW_PACKET);
885 if (field_avail(typeof(resp), tunnel_offloads_caps,
887 resp.response_length += sizeof(resp.tunnel_offloads_caps);
888 if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
889 resp.tunnel_offloads_caps |=
890 MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
891 if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
892 resp.tunnel_offloads_caps |=
893 MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
894 if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
895 resp.tunnel_offloads_caps |=
896 MLX5_IB_TUNNELED_OFFLOADS_GRE;
900 err = ib_copy_to_udata(uhw, &resp, resp.response_length);
910 MLX5_IB_WIDTH_1X = 1 << 0,
911 MLX5_IB_WIDTH_2X = 1 << 1,
912 MLX5_IB_WIDTH_4X = 1 << 2,
913 MLX5_IB_WIDTH_8X = 1 << 3,
914 MLX5_IB_WIDTH_12X = 1 << 4
917 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
920 struct mlx5_ib_dev *dev = to_mdev(ibdev);
923 if (active_width & MLX5_IB_WIDTH_1X) {
924 *ib_width = IB_WIDTH_1X;
925 } else if (active_width & MLX5_IB_WIDTH_2X) {
926 mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
929 } else if (active_width & MLX5_IB_WIDTH_4X) {
930 *ib_width = IB_WIDTH_4X;
931 } else if (active_width & MLX5_IB_WIDTH_8X) {
932 *ib_width = IB_WIDTH_8X;
933 } else if (active_width & MLX5_IB_WIDTH_12X) {
934 *ib_width = IB_WIDTH_12X;
936 mlx5_ib_dbg(dev, "Invalid active_width %d\n",
944 static int mlx5_mtu_to_ib_mtu(int mtu)
953 pr_warn("invalid mtu\n");
963 __IB_MAX_VL_0_14 = 5,
966 enum mlx5_vl_hw_cap {
978 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
983 *max_vl_num = __IB_MAX_VL_0;
986 *max_vl_num = __IB_MAX_VL_0_1;
989 *max_vl_num = __IB_MAX_VL_0_3;
992 *max_vl_num = __IB_MAX_VL_0_7;
994 case MLX5_VL_HW_0_14:
995 *max_vl_num = __IB_MAX_VL_0_14;
1005 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1006 struct ib_port_attr *props)
1008 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1009 struct mlx5_core_dev *mdev = dev->mdev;
1010 struct mlx5_hca_vport_context *rep;
1014 u8 ib_link_width_oper;
1017 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1023 /* props being zeroed by the caller, avoid zeroing it here */
1025 err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1029 props->lid = rep->lid;
1030 props->lmc = rep->lmc;
1031 props->sm_lid = rep->sm_lid;
1032 props->sm_sl = rep->sm_sl;
1033 props->state = rep->vport_state;
1034 props->phys_state = rep->port_physical_state;
1035 props->port_cap_flags = rep->cap_mask1;
1036 props->gid_tbl_len = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1037 props->max_msg_sz = 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1038 props->pkey_tbl_len = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1039 props->bad_pkey_cntr = rep->pkey_violation_counter;
1040 props->qkey_viol_cntr = rep->qkey_violation_counter;
1041 props->subnet_timeout = rep->subnet_timeout;
1042 props->init_type_reply = rep->init_type_reply;
1043 props->grh_required = rep->grh_required;
1045 err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
1049 err = translate_active_width(ibdev, ib_link_width_oper,
1050 &props->active_width);
1053 err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
1057 mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1059 props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1061 mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1063 props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1065 err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1069 err = translate_max_vl_num(ibdev, vl_hw_cap,
1070 &props->max_vl_num);
1076 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1077 struct ib_port_attr *props)
1082 switch (mlx5_get_vport_access_method(ibdev)) {
1083 case MLX5_VPORT_ACCESS_METHOD_MAD:
1084 ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1087 case MLX5_VPORT_ACCESS_METHOD_HCA:
1088 ret = mlx5_query_hca_port(ibdev, port, props);
1091 case MLX5_VPORT_ACCESS_METHOD_NIC:
1092 ret = mlx5_query_port_roce(ibdev, port, props);
1099 if (!ret && props) {
1100 count = mlx5_core_reserved_gids_count(to_mdev(ibdev)->mdev);
1101 props->gid_tbl_len -= count;
1106 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1109 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1110 struct mlx5_core_dev *mdev = dev->mdev;
1112 switch (mlx5_get_vport_access_method(ibdev)) {
1113 case MLX5_VPORT_ACCESS_METHOD_MAD:
1114 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1116 case MLX5_VPORT_ACCESS_METHOD_HCA:
1117 return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1125 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1128 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1129 struct mlx5_core_dev *mdev = dev->mdev;
1131 switch (mlx5_get_vport_access_method(ibdev)) {
1132 case MLX5_VPORT_ACCESS_METHOD_MAD:
1133 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1135 case MLX5_VPORT_ACCESS_METHOD_HCA:
1136 case MLX5_VPORT_ACCESS_METHOD_NIC:
1137 return mlx5_query_hca_vport_pkey(mdev, 0, port, 0, index,
1144 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1145 struct ib_device_modify *props)
1147 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1148 struct mlx5_reg_node_desc in;
1149 struct mlx5_reg_node_desc out;
1152 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1155 if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1159 * If possible, pass node desc to FW, so it can generate
1160 * a 144 trap. If cmd fails, just ignore.
1162 memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1163 err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1164 sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1168 memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1173 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1176 struct mlx5_hca_vport_context ctx = {};
1179 err = mlx5_query_hca_vport_context(dev->mdev, 0,
1184 if (~ctx.cap_mask1_perm & mask) {
1185 mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1186 mask, ctx.cap_mask1_perm);
1190 ctx.cap_mask1 = value;
1191 ctx.cap_mask1_perm = mask;
1192 err = mlx5_core_modify_hca_vport_context(dev->mdev, 0,
1198 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1199 struct ib_port_modify *props)
1201 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1202 struct ib_port_attr attr;
1207 bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1208 IB_LINK_LAYER_INFINIBAND);
1210 /* CM layer calls ib_modify_port() regardless of the link layer. For
1211 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1216 if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1217 change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1218 value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1219 return set_port_caps_atomic(dev, port, change_mask, value);
1222 mutex_lock(&dev->cap_mask_mutex);
1224 err = ib_query_port(ibdev, port, &attr);
1228 tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1229 ~props->clr_port_cap_mask;
1231 err = mlx5_set_port_caps(dev->mdev, port, tmp);
1234 mutex_unlock(&dev->cap_mask_mutex);
1238 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1240 mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1241 caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1244 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1245 struct mlx5_ib_alloc_ucontext_req_v2 *req,
1248 int uars_per_sys_page;
1249 int bfregs_per_sys_page;
1250 int ref_bfregs = req->total_num_bfregs;
1252 if (req->total_num_bfregs == 0)
1255 BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1256 BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1258 if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1261 uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1262 bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1263 req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1264 *num_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1266 if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1269 mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, allocated %d, using %d sys pages\n",
1270 MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1271 lib_uar_4k ? "yes" : "no", ref_bfregs,
1272 req->total_num_bfregs, *num_sys_pages);
1277 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1279 struct mlx5_bfreg_info *bfregi;
1283 bfregi = &context->bfregi;
1284 for (i = 0; i < bfregi->num_sys_pages; i++) {
1285 err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1289 mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1294 for (--i; i >= 0; i--)
1295 if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1296 mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1301 static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1303 struct mlx5_bfreg_info *bfregi;
1307 bfregi = &context->bfregi;
1308 for (i = 0; i < bfregi->num_sys_pages; i++) {
1309 err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1311 mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1318 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn)
1322 err = mlx5_core_alloc_transport_domain(dev->mdev, tdn);
1326 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1327 !MLX5_CAP_GEN(dev->mdev, disable_local_lb))
1330 mutex_lock(&dev->lb_mutex);
1333 if (dev->user_td == 2)
1334 err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1336 mutex_unlock(&dev->lb_mutex);
1340 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn)
1342 mlx5_core_dealloc_transport_domain(dev->mdev, tdn);
1344 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1345 !MLX5_CAP_GEN(dev->mdev, disable_local_lb))
1348 mutex_lock(&dev->lb_mutex);
1351 if (dev->user_td < 2)
1352 mlx5_nic_vport_update_local_lb(dev->mdev, false);
1354 mutex_unlock(&dev->lb_mutex);
1357 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1358 struct ib_udata *udata)
1360 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1361 struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1362 struct mlx5_ib_alloc_ucontext_resp resp = {};
1363 struct mlx5_ib_ucontext *context;
1364 struct mlx5_bfreg_info *bfregi;
1367 size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1371 if (!dev->ib_active)
1372 return ERR_PTR(-EAGAIN);
1374 if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1376 else if (udata->inlen >= min_req_v2)
1379 return ERR_PTR(-EINVAL);
1381 err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1383 return ERR_PTR(err);
1386 return ERR_PTR(-EINVAL);
1388 if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1389 return ERR_PTR(-EOPNOTSUPP);
1391 req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1392 MLX5_NON_FP_BFREGS_PER_UAR);
1393 if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1394 return ERR_PTR(-EINVAL);
1396 resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1397 if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1398 resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1399 resp.cache_line_size = cache_line_size();
1400 resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1401 resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1402 resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1403 resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1404 resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1405 resp.cqe_version = min_t(__u8,
1406 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1407 req.max_cqe_version);
1408 resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1409 MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1410 resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1411 MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1412 resp.response_length = min(offsetof(typeof(resp), response_length) +
1413 sizeof(resp.response_length), udata->outlen);
1415 context = kzalloc(sizeof(*context), GFP_KERNEL);
1417 return ERR_PTR(-ENOMEM);
1419 lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1420 bfregi = &context->bfregi;
1422 /* updates req->total_num_bfregs */
1423 err = calc_total_bfregs(dev, lib_uar_4k, &req, &bfregi->num_sys_pages);
1427 mutex_init(&bfregi->lock);
1428 bfregi->lib_uar_4k = lib_uar_4k;
1429 bfregi->count = kcalloc(req.total_num_bfregs, sizeof(*bfregi->count),
1431 if (!bfregi->count) {
1436 bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1437 sizeof(*bfregi->sys_pages),
1439 if (!bfregi->sys_pages) {
1444 err = allocate_uars(dev, context);
1448 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1449 context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1452 context->upd_xlt_page = __get_free_page(GFP_KERNEL);
1453 if (!context->upd_xlt_page) {
1457 mutex_init(&context->upd_xlt_page_mutex);
1459 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1460 err = mlx5_ib_alloc_transport_domain(dev, &context->tdn);
1465 INIT_LIST_HEAD(&context->vma_private_list);
1466 INIT_LIST_HEAD(&context->db_page_list);
1467 mutex_init(&context->db_page_mutex);
1469 resp.tot_bfregs = req.total_num_bfregs;
1470 resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1472 if (field_avail(typeof(resp), cqe_version, udata->outlen))
1473 resp.response_length += sizeof(resp.cqe_version);
1475 if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1476 resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1477 MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1478 resp.response_length += sizeof(resp.cmds_supp_uhw);
1481 if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1482 if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1483 mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1484 resp.eth_min_inline++;
1486 resp.response_length += sizeof(resp.eth_min_inline);
1490 * We don't want to expose information from the PCI bar that is located
1491 * after 4096 bytes, so if the arch only supports larger pages, let's
1492 * pretend we don't support reading the HCA's core clock. This is also
1493 * forced by mmap function.
1495 if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1496 if (PAGE_SIZE <= 4096) {
1498 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1499 resp.hca_core_clock_offset =
1500 offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1502 resp.response_length += sizeof(resp.hca_core_clock_offset) +
1503 sizeof(resp.reserved2);
1506 if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1507 resp.response_length += sizeof(resp.log_uar_size);
1509 if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1510 resp.response_length += sizeof(resp.num_uars_per_page);
1512 err = ib_copy_to_udata(udata, &resp, resp.response_length);
1517 bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1518 context->cqe_version = resp.cqe_version;
1519 context->lib_caps = req.lib_caps;
1520 print_lib_caps(dev, context->lib_caps);
1522 return &context->ibucontext;
1525 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1526 mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1529 free_page(context->upd_xlt_page);
1532 deallocate_uars(dev, context);
1535 kfree(bfregi->sys_pages);
1538 kfree(bfregi->count);
1543 return ERR_PTR(err);
1546 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1548 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1549 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1550 struct mlx5_bfreg_info *bfregi;
1552 bfregi = &context->bfregi;
1553 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1554 mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1556 free_page(context->upd_xlt_page);
1557 deallocate_uars(dev, context);
1558 kfree(bfregi->sys_pages);
1559 kfree(bfregi->count);
1565 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1566 struct mlx5_bfreg_info *bfregi,
1569 int fw_uars_per_page;
1571 fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1573 return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) +
1574 bfregi->sys_pages[idx] / fw_uars_per_page;
1577 static int get_command(unsigned long offset)
1579 return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1582 static int get_arg(unsigned long offset)
1584 return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1587 static int get_index(unsigned long offset)
1589 return get_arg(offset);
1592 static void mlx5_ib_vma_open(struct vm_area_struct *area)
1594 /* vma_open is called when a new VMA is created on top of our VMA. This
1595 * is done through either mremap flow or split_vma (usually due to
1596 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1597 * as this VMA is strongly hardware related. Therefore we set the
1598 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1599 * calling us again and trying to do incorrect actions. We assume that
1600 * the original VMA size is exactly a single page, and therefore all
1601 * "splitting" operation will not happen to it.
1603 area->vm_ops = NULL;
1606 static void mlx5_ib_vma_close(struct vm_area_struct *area)
1608 struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1610 /* It's guaranteed that all VMAs opened on a FD are closed before the
1611 * file itself is closed, therefore no sync is needed with the regular
1612 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1613 * However need a sync with accessing the vma as part of
1614 * mlx5_ib_disassociate_ucontext.
1615 * The close operation is usually called under mm->mmap_sem except when
1616 * process is exiting.
1617 * The exiting case is handled explicitly as part of
1618 * mlx5_ib_disassociate_ucontext.
1620 mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1622 /* setting the vma context pointer to null in the mlx5_ib driver's
1623 * private data, to protect a race condition in
1624 * mlx5_ib_disassociate_ucontext().
1626 mlx5_ib_vma_priv_data->vma = NULL;
1627 list_del(&mlx5_ib_vma_priv_data->list);
1628 kfree(mlx5_ib_vma_priv_data);
1631 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1632 .open = mlx5_ib_vma_open,
1633 .close = mlx5_ib_vma_close
1636 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1637 struct mlx5_ib_ucontext *ctx)
1639 struct mlx5_ib_vma_private_data *vma_prv;
1640 struct list_head *vma_head = &ctx->vma_private_list;
1642 vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1647 vma->vm_private_data = vma_prv;
1648 vma->vm_ops = &mlx5_ib_vm_ops;
1650 list_add(&vma_prv->list, vma_head);
1655 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1658 struct vm_area_struct *vma;
1659 struct mlx5_ib_vma_private_data *vma_private, *n;
1660 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1661 struct task_struct *owning_process = NULL;
1662 struct mm_struct *owning_mm = NULL;
1664 owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1665 if (!owning_process)
1668 owning_mm = get_task_mm(owning_process);
1670 pr_info("no mm, disassociate ucontext is pending task termination\n");
1672 put_task_struct(owning_process);
1673 usleep_range(1000, 2000);
1674 owning_process = get_pid_task(ibcontext->tgid,
1676 if (!owning_process ||
1677 owning_process->state == TASK_DEAD) {
1678 pr_info("disassociate ucontext done, task was terminated\n");
1679 /* in case task was dead need to release the
1683 put_task_struct(owning_process);
1689 /* need to protect from a race on closing the vma as part of
1690 * mlx5_ib_vma_close.
1692 down_write(&owning_mm->mmap_sem);
1693 list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1695 vma = vma_private->vma;
1696 ret = zap_vma_ptes(vma, vma->vm_start,
1698 WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1699 /* context going to be destroyed, should
1700 * not access ops any more.
1702 vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
1704 list_del(&vma_private->list);
1707 up_write(&owning_mm->mmap_sem);
1709 put_task_struct(owning_process);
1712 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1715 case MLX5_IB_MMAP_WC_PAGE:
1717 case MLX5_IB_MMAP_REGULAR_PAGE:
1718 return "best effort WC";
1719 case MLX5_IB_MMAP_NC_PAGE:
1726 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1727 struct vm_area_struct *vma,
1728 struct mlx5_ib_ucontext *context)
1730 struct mlx5_bfreg_info *bfregi = &context->bfregi;
1733 phys_addr_t pfn, pa;
1737 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1740 uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
1741 idx = get_index(vma->vm_pgoff);
1742 if (idx % uars_per_page ||
1743 idx * uars_per_page >= bfregi->num_sys_pages) {
1744 mlx5_ib_warn(dev, "invalid uar index %lu\n", idx);
1749 case MLX5_IB_MMAP_WC_PAGE:
1750 /* Some architectures don't support WC memory */
1751 #if defined(CONFIG_X86)
1754 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1758 case MLX5_IB_MMAP_REGULAR_PAGE:
1759 /* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1760 prot = pgprot_writecombine(vma->vm_page_prot);
1762 case MLX5_IB_MMAP_NC_PAGE:
1763 prot = pgprot_noncached(vma->vm_page_prot);
1769 pfn = uar_index2pfn(dev, bfregi, idx);
1770 mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1772 vma->vm_page_prot = prot;
1773 err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1774 PAGE_SIZE, vma->vm_page_prot);
1776 mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
1777 err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
1781 pa = pfn << PAGE_SHIFT;
1782 mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
1783 vma->vm_start, &pa);
1785 return mlx5_ib_set_vma_data(vma, context);
1788 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1790 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1791 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1792 unsigned long command;
1795 command = get_command(vma->vm_pgoff);
1797 case MLX5_IB_MMAP_WC_PAGE:
1798 case MLX5_IB_MMAP_NC_PAGE:
1799 case MLX5_IB_MMAP_REGULAR_PAGE:
1800 return uar_mmap(dev, command, vma, context);
1802 case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1805 case MLX5_IB_MMAP_CORE_CLOCK:
1806 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1809 if (vma->vm_flags & VM_WRITE)
1812 /* Don't expose to user-space information it shouldn't have */
1813 if (PAGE_SIZE > 4096)
1816 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1817 pfn = (dev->mdev->iseg_base +
1818 offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1820 if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1821 PAGE_SIZE, vma->vm_page_prot))
1824 mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
1826 (unsigned long long)pfn << PAGE_SHIFT);
1836 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1837 struct ib_ucontext *context,
1838 struct ib_udata *udata)
1840 struct mlx5_ib_alloc_pd_resp resp;
1841 struct mlx5_ib_pd *pd;
1844 pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1846 return ERR_PTR(-ENOMEM);
1848 err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1851 return ERR_PTR(err);
1856 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1857 mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1859 return ERR_PTR(-EFAULT);
1866 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1868 struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1869 struct mlx5_ib_pd *mpd = to_mpd(pd);
1871 mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1878 MATCH_CRITERIA_ENABLE_OUTER_BIT,
1879 MATCH_CRITERIA_ENABLE_MISC_BIT,
1880 MATCH_CRITERIA_ENABLE_INNER_BIT
1883 #define HEADER_IS_ZERO(match_criteria, headers) \
1884 !(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1885 0, MLX5_FLD_SZ_BYTES(fte_match_param, headers))) \
1887 static u8 get_match_criteria_enable(u32 *match_criteria)
1889 u8 match_criteria_enable;
1891 match_criteria_enable =
1892 (!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1893 MATCH_CRITERIA_ENABLE_OUTER_BIT;
1894 match_criteria_enable |=
1895 (!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1896 MATCH_CRITERIA_ENABLE_MISC_BIT;
1897 match_criteria_enable |=
1898 (!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1899 MATCH_CRITERIA_ENABLE_INNER_BIT;
1901 return match_criteria_enable;
1904 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1906 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1907 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1910 static void set_flow_label(void *misc_c, void *misc_v, u8 mask, u8 val,
1914 MLX5_SET(fte_match_set_misc,
1915 misc_c, inner_ipv6_flow_label, mask);
1916 MLX5_SET(fte_match_set_misc,
1917 misc_v, inner_ipv6_flow_label, val);
1919 MLX5_SET(fte_match_set_misc,
1920 misc_c, outer_ipv6_flow_label, mask);
1921 MLX5_SET(fte_match_set_misc,
1922 misc_v, outer_ipv6_flow_label, val);
1926 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1928 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1929 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1930 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1931 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1934 #define LAST_ETH_FIELD vlan_tag
1935 #define LAST_IB_FIELD sl
1936 #define LAST_IPV4_FIELD tos
1937 #define LAST_IPV6_FIELD traffic_class
1938 #define LAST_TCP_UDP_FIELD src_port
1939 #define LAST_TUNNEL_FIELD tunnel_id
1940 #define LAST_FLOW_TAG_FIELD tag_id
1941 #define LAST_DROP_FIELD size
1943 /* Field is the last supported field */
1944 #define FIELDS_NOT_SUPPORTED(filter, field)\
1945 memchr_inv((void *)&filter.field +\
1946 sizeof(filter.field), 0,\
1948 offsetof(typeof(filter), field) -\
1949 sizeof(filter.field))
1951 #define IPV4_VERSION 4
1952 #define IPV6_VERSION 6
1953 static int parse_flow_attr(struct mlx5_core_dev *mdev, u32 *match_c,
1954 u32 *match_v, const union ib_flow_spec *ib_spec,
1955 u32 *tag_id, bool *is_drop)
1957 void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1959 void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1965 if (ib_spec->type & IB_FLOW_SPEC_INNER) {
1966 headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1968 headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1970 match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
1971 ft_field_support.inner_ip_version);
1973 headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1975 headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1977 match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
1978 ft_field_support.outer_ip_version);
1981 switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
1982 case IB_FLOW_SPEC_ETH:
1983 if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1986 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1988 ib_spec->eth.mask.dst_mac);
1989 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1991 ib_spec->eth.val.dst_mac);
1993 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1995 ib_spec->eth.mask.src_mac);
1996 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1998 ib_spec->eth.val.src_mac);
2000 if (ib_spec->eth.mask.vlan_tag) {
2001 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2003 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2006 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2007 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
2008 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2009 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
2011 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2013 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
2014 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2016 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
2018 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2020 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
2021 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2023 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
2025 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2026 ethertype, ntohs(ib_spec->eth.mask.ether_type));
2027 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2028 ethertype, ntohs(ib_spec->eth.val.ether_type));
2030 case IB_FLOW_SPEC_IPV4:
2031 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
2035 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2037 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2038 ip_version, IPV4_VERSION);
2040 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2042 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2043 ethertype, ETH_P_IP);
2046 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2047 src_ipv4_src_ipv6.ipv4_layout.ipv4),
2048 &ib_spec->ipv4.mask.src_ip,
2049 sizeof(ib_spec->ipv4.mask.src_ip));
2050 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2051 src_ipv4_src_ipv6.ipv4_layout.ipv4),
2052 &ib_spec->ipv4.val.src_ip,
2053 sizeof(ib_spec->ipv4.val.src_ip));
2054 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2055 dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2056 &ib_spec->ipv4.mask.dst_ip,
2057 sizeof(ib_spec->ipv4.mask.dst_ip));
2058 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2059 dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2060 &ib_spec->ipv4.val.dst_ip,
2061 sizeof(ib_spec->ipv4.val.dst_ip));
2063 set_tos(headers_c, headers_v,
2064 ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
2066 set_proto(headers_c, headers_v,
2067 ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
2069 case IB_FLOW_SPEC_IPV6:
2070 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
2074 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2076 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2077 ip_version, IPV6_VERSION);
2079 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2081 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2082 ethertype, ETH_P_IPV6);
2085 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2086 src_ipv4_src_ipv6.ipv6_layout.ipv6),
2087 &ib_spec->ipv6.mask.src_ip,
2088 sizeof(ib_spec->ipv6.mask.src_ip));
2089 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2090 src_ipv4_src_ipv6.ipv6_layout.ipv6),
2091 &ib_spec->ipv6.val.src_ip,
2092 sizeof(ib_spec->ipv6.val.src_ip));
2093 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2094 dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2095 &ib_spec->ipv6.mask.dst_ip,
2096 sizeof(ib_spec->ipv6.mask.dst_ip));
2097 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2098 dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2099 &ib_spec->ipv6.val.dst_ip,
2100 sizeof(ib_spec->ipv6.val.dst_ip));
2102 set_tos(headers_c, headers_v,
2103 ib_spec->ipv6.mask.traffic_class,
2104 ib_spec->ipv6.val.traffic_class);
2106 set_proto(headers_c, headers_v,
2107 ib_spec->ipv6.mask.next_hdr,
2108 ib_spec->ipv6.val.next_hdr);
2110 set_flow_label(misc_params_c, misc_params_v,
2111 ntohl(ib_spec->ipv6.mask.flow_label),
2112 ntohl(ib_spec->ipv6.val.flow_label),
2113 ib_spec->type & IB_FLOW_SPEC_INNER);
2116 case IB_FLOW_SPEC_TCP:
2117 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2118 LAST_TCP_UDP_FIELD))
2121 MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2123 MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2126 MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
2127 ntohs(ib_spec->tcp_udp.mask.src_port));
2128 MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
2129 ntohs(ib_spec->tcp_udp.val.src_port));
2131 MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
2132 ntohs(ib_spec->tcp_udp.mask.dst_port));
2133 MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
2134 ntohs(ib_spec->tcp_udp.val.dst_port));
2136 case IB_FLOW_SPEC_UDP:
2137 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2138 LAST_TCP_UDP_FIELD))
2141 MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2143 MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2146 MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
2147 ntohs(ib_spec->tcp_udp.mask.src_port));
2148 MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
2149 ntohs(ib_spec->tcp_udp.val.src_port));
2151 MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
2152 ntohs(ib_spec->tcp_udp.mask.dst_port));
2153 MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
2154 ntohs(ib_spec->tcp_udp.val.dst_port));
2156 case IB_FLOW_SPEC_VXLAN_TUNNEL:
2157 if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
2161 MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
2162 ntohl(ib_spec->tunnel.mask.tunnel_id));
2163 MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
2164 ntohl(ib_spec->tunnel.val.tunnel_id));
2166 case IB_FLOW_SPEC_ACTION_TAG:
2167 if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
2168 LAST_FLOW_TAG_FIELD))
2170 if (ib_spec->flow_tag.tag_id >= BIT(24))
2173 *tag_id = ib_spec->flow_tag.tag_id;
2175 case IB_FLOW_SPEC_ACTION_DROP:
2176 if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
2188 /* If a flow could catch both multicast and unicast packets,
2189 * it won't fall into the multicast flow steering table and this rule
2190 * could steal other multicast packets.
2192 static bool flow_is_multicast_only(const struct ib_flow_attr *ib_attr)
2194 union ib_flow_spec *flow_spec;
2196 if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
2197 ib_attr->num_of_specs < 1)
2200 flow_spec = (union ib_flow_spec *)(ib_attr + 1);
2201 if (flow_spec->type == IB_FLOW_SPEC_IPV4) {
2202 struct ib_flow_spec_ipv4 *ipv4_spec;
2204 ipv4_spec = (struct ib_flow_spec_ipv4 *)flow_spec;
2205 if (ipv4_is_multicast(ipv4_spec->val.dst_ip))
2211 if (flow_spec->type == IB_FLOW_SPEC_ETH) {
2212 struct ib_flow_spec_eth *eth_spec;
2214 eth_spec = (struct ib_flow_spec_eth *)flow_spec;
2215 return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
2216 is_multicast_ether_addr(eth_spec->val.dst_mac);
2222 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
2223 const struct ib_flow_attr *flow_attr,
2226 union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
2227 int match_ipv = check_inner ?
2228 MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2229 ft_field_support.inner_ip_version) :
2230 MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2231 ft_field_support.outer_ip_version);
2232 int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
2233 bool ipv4_spec_valid, ipv6_spec_valid;
2234 unsigned int ip_spec_type = 0;
2235 bool has_ethertype = false;
2236 unsigned int spec_index;
2237 bool mask_valid = true;
2241 /* Validate that ethertype is correct */
2242 for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2243 if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
2244 ib_spec->eth.mask.ether_type) {
2245 mask_valid = (ib_spec->eth.mask.ether_type ==
2247 has_ethertype = true;
2248 eth_type = ntohs(ib_spec->eth.val.ether_type);
2249 } else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
2250 (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
2251 ip_spec_type = ib_spec->type;
2253 ib_spec = (void *)ib_spec + ib_spec->size;
2256 type_valid = (!has_ethertype) || (!ip_spec_type);
2257 if (!type_valid && mask_valid) {
2258 ipv4_spec_valid = (eth_type == ETH_P_IP) &&
2259 (ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
2260 ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
2261 (ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
2263 type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
2264 (((eth_type == ETH_P_MPLS_UC) ||
2265 (eth_type == ETH_P_MPLS_MC)) && match_ipv);
2271 static bool is_valid_attr(struct mlx5_core_dev *mdev,
2272 const struct ib_flow_attr *flow_attr)
2274 return is_valid_ethertype(mdev, flow_attr, false) &&
2275 is_valid_ethertype(mdev, flow_attr, true);
2278 static void put_flow_table(struct mlx5_ib_dev *dev,
2279 struct mlx5_ib_flow_prio *prio, bool ft_added)
2281 prio->refcount -= !!ft_added;
2282 if (!prio->refcount) {
2283 mlx5_destroy_flow_table(prio->flow_table);
2284 prio->flow_table = NULL;
2288 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
2290 struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
2291 struct mlx5_ib_flow_handler *handler = container_of(flow_id,
2292 struct mlx5_ib_flow_handler,
2294 struct mlx5_ib_flow_handler *iter, *tmp;
2296 mutex_lock(&dev->flow_db.lock);
2298 list_for_each_entry_safe(iter, tmp, &handler->list, list) {
2299 mlx5_del_flow_rules(iter->rule);
2300 put_flow_table(dev, iter->prio, true);
2301 list_del(&iter->list);
2305 mlx5_del_flow_rules(handler->rule);
2306 put_flow_table(dev, handler->prio, true);
2307 mutex_unlock(&dev->flow_db.lock);
2314 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
2322 enum flow_table_type {
2327 #define MLX5_FS_MAX_TYPES 6
2328 #define MLX5_FS_MAX_ENTRIES BIT(16)
2329 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
2330 struct ib_flow_attr *flow_attr,
2331 enum flow_table_type ft_type)
2333 bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
2334 struct mlx5_flow_namespace *ns = NULL;
2335 struct mlx5_ib_flow_prio *prio;
2336 struct mlx5_flow_table *ft;
2343 max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
2345 if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2346 if (flow_is_multicast_only(flow_attr) &&
2348 priority = MLX5_IB_FLOW_MCAST_PRIO;
2350 priority = ib_prio_to_core_prio(flow_attr->priority,
2352 ns = mlx5_get_flow_namespace(dev->mdev,
2353 MLX5_FLOW_NAMESPACE_BYPASS);
2354 num_entries = MLX5_FS_MAX_ENTRIES;
2355 num_groups = MLX5_FS_MAX_TYPES;
2356 prio = &dev->flow_db.prios[priority];
2357 } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2358 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2359 ns = mlx5_get_flow_namespace(dev->mdev,
2360 MLX5_FLOW_NAMESPACE_LEFTOVERS);
2361 build_leftovers_ft_param(&priority,
2364 prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
2365 } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2366 if (!MLX5_CAP_FLOWTABLE(dev->mdev,
2367 allow_sniffer_and_nic_rx_shared_tir))
2368 return ERR_PTR(-ENOTSUPP);
2370 ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
2371 MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2372 MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2374 prio = &dev->flow_db.sniffer[ft_type];
2381 return ERR_PTR(-ENOTSUPP);
2383 if (num_entries > max_table_size)
2384 return ERR_PTR(-ENOMEM);
2386 ft = prio->flow_table;
2388 ft = mlx5_create_auto_grouped_flow_table(ns, priority,
2395 prio->flow_table = ft;
2401 return err ? ERR_PTR(err) : prio;
2404 static void set_underlay_qp(struct mlx5_ib_dev *dev,
2405 struct mlx5_flow_spec *spec,
2408 void *misc_params_c = MLX5_ADDR_OF(fte_match_param,
2409 spec->match_criteria,
2411 void *misc_params_v = MLX5_ADDR_OF(fte_match_param, spec->match_value,
2415 MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
2416 ft_field_support.bth_dst_qp)) {
2417 MLX5_SET(fte_match_set_misc,
2418 misc_params_v, bth_dst_qp, underlay_qpn);
2419 MLX5_SET(fte_match_set_misc,
2420 misc_params_c, bth_dst_qp, 0xffffff);
2424 static struct mlx5_ib_flow_handler *_create_flow_rule(struct mlx5_ib_dev *dev,
2425 struct mlx5_ib_flow_prio *ft_prio,
2426 const struct ib_flow_attr *flow_attr,
2427 struct mlx5_flow_destination *dst,
2430 struct mlx5_flow_table *ft = ft_prio->flow_table;
2431 struct mlx5_ib_flow_handler *handler;
2432 struct mlx5_flow_act flow_act = {0};
2433 struct mlx5_flow_spec *spec;
2434 struct mlx5_flow_destination *rule_dst = dst;
2435 const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
2436 unsigned int spec_index;
2437 u32 flow_tag = MLX5_FS_DEFAULT_FLOW_TAG;
2438 bool is_drop = false;
2442 if (!is_valid_attr(dev->mdev, flow_attr))
2443 return ERR_PTR(-EINVAL);
2445 spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
2446 handler = kzalloc(sizeof(*handler), GFP_KERNEL);
2447 if (!handler || !spec) {
2452 INIT_LIST_HEAD(&handler->list);
2454 for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2455 err = parse_flow_attr(dev->mdev, spec->match_criteria,
2457 ib_flow, &flow_tag, &is_drop);
2461 ib_flow += ((union ib_flow_spec *)ib_flow)->size;
2464 if (!flow_is_multicast_only(flow_attr))
2465 set_underlay_qp(dev, spec, underlay_qpn);
2467 spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
2469 flow_act.action = MLX5_FLOW_CONTEXT_ACTION_DROP;
2473 flow_act.action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
2474 MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
2477 if (flow_tag != MLX5_FS_DEFAULT_FLOW_TAG &&
2478 (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2479 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
2480 mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
2481 flow_tag, flow_attr->type);
2485 flow_act.flow_tag = flow_tag;
2486 handler->rule = mlx5_add_flow_rules(ft, spec,
2488 rule_dst, dest_num);
2490 if (IS_ERR(handler->rule)) {
2491 err = PTR_ERR(handler->rule);
2495 ft_prio->refcount++;
2496 handler->prio = ft_prio;
2498 ft_prio->flow_table = ft;
2503 return err ? ERR_PTR(err) : handler;
2506 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
2507 struct mlx5_ib_flow_prio *ft_prio,
2508 const struct ib_flow_attr *flow_attr,
2509 struct mlx5_flow_destination *dst)
2511 return _create_flow_rule(dev, ft_prio, flow_attr, dst, 0);
2514 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
2515 struct mlx5_ib_flow_prio *ft_prio,
2516 struct ib_flow_attr *flow_attr,
2517 struct mlx5_flow_destination *dst)
2519 struct mlx5_ib_flow_handler *handler_dst = NULL;
2520 struct mlx5_ib_flow_handler *handler = NULL;
2522 handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
2523 if (!IS_ERR(handler)) {
2524 handler_dst = create_flow_rule(dev, ft_prio,
2526 if (IS_ERR(handler_dst)) {
2527 mlx5_del_flow_rules(handler->rule);
2528 ft_prio->refcount--;
2530 handler = handler_dst;
2532 list_add(&handler_dst->list, &handler->list);
2543 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
2544 struct mlx5_ib_flow_prio *ft_prio,
2545 struct ib_flow_attr *flow_attr,
2546 struct mlx5_flow_destination *dst)
2548 struct mlx5_ib_flow_handler *handler_ucast = NULL;
2549 struct mlx5_ib_flow_handler *handler = NULL;
2552 struct ib_flow_attr flow_attr;
2553 struct ib_flow_spec_eth eth_flow;
2554 } leftovers_specs[] = {
2558 .size = sizeof(leftovers_specs[0])
2561 .type = IB_FLOW_SPEC_ETH,
2562 .size = sizeof(struct ib_flow_spec_eth),
2563 .mask = {.dst_mac = {0x1} },
2564 .val = {.dst_mac = {0x1} }
2570 .size = sizeof(leftovers_specs[0])
2573 .type = IB_FLOW_SPEC_ETH,
2574 .size = sizeof(struct ib_flow_spec_eth),
2575 .mask = {.dst_mac = {0x1} },
2576 .val = {.dst_mac = {} }
2581 handler = create_flow_rule(dev, ft_prio,
2582 &leftovers_specs[LEFTOVERS_MC].flow_attr,
2584 if (!IS_ERR(handler) &&
2585 flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2586 handler_ucast = create_flow_rule(dev, ft_prio,
2587 &leftovers_specs[LEFTOVERS_UC].flow_attr,
2589 if (IS_ERR(handler_ucast)) {
2590 mlx5_del_flow_rules(handler->rule);
2591 ft_prio->refcount--;
2593 handler = handler_ucast;
2595 list_add(&handler_ucast->list, &handler->list);
2602 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2603 struct mlx5_ib_flow_prio *ft_rx,
2604 struct mlx5_ib_flow_prio *ft_tx,
2605 struct mlx5_flow_destination *dst)
2607 struct mlx5_ib_flow_handler *handler_rx;
2608 struct mlx5_ib_flow_handler *handler_tx;
2610 static const struct ib_flow_attr flow_attr = {
2612 .size = sizeof(flow_attr)
2615 handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2616 if (IS_ERR(handler_rx)) {
2617 err = PTR_ERR(handler_rx);
2621 handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2622 if (IS_ERR(handler_tx)) {
2623 err = PTR_ERR(handler_tx);
2627 list_add(&handler_tx->list, &handler_rx->list);
2632 mlx5_del_flow_rules(handler_rx->rule);
2636 return ERR_PTR(err);
2639 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2640 struct ib_flow_attr *flow_attr,
2643 struct mlx5_ib_dev *dev = to_mdev(qp->device);
2644 struct mlx5_ib_qp *mqp = to_mqp(qp);
2645 struct mlx5_ib_flow_handler *handler = NULL;
2646 struct mlx5_flow_destination *dst = NULL;
2647 struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2648 struct mlx5_ib_flow_prio *ft_prio;
2652 if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2653 return ERR_PTR(-ENOMEM);
2655 if (domain != IB_FLOW_DOMAIN_USER ||
2656 flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2657 (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2658 return ERR_PTR(-EINVAL);
2660 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2662 return ERR_PTR(-ENOMEM);
2664 mutex_lock(&dev->flow_db.lock);
2666 ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2667 if (IS_ERR(ft_prio)) {
2668 err = PTR_ERR(ft_prio);
2671 if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2672 ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2673 if (IS_ERR(ft_prio_tx)) {
2674 err = PTR_ERR(ft_prio_tx);
2680 dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2681 if (mqp->flags & MLX5_IB_QP_RSS)
2682 dst->tir_num = mqp->rss_qp.tirn;
2684 dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2686 if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2687 if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) {
2688 handler = create_dont_trap_rule(dev, ft_prio,
2691 underlay_qpn = (mqp->flags & MLX5_IB_QP_UNDERLAY) ?
2692 mqp->underlay_qpn : 0;
2693 handler = _create_flow_rule(dev, ft_prio, flow_attr,
2696 } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2697 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2698 handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2700 } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2701 handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2707 if (IS_ERR(handler)) {
2708 err = PTR_ERR(handler);
2713 mutex_unlock(&dev->flow_db.lock);
2716 return &handler->ibflow;
2719 put_flow_table(dev, ft_prio, false);
2721 put_flow_table(dev, ft_prio_tx, false);
2723 mutex_unlock(&dev->flow_db.lock);
2726 return ERR_PTR(err);
2729 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2731 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2732 struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2735 if (mqp->flags & MLX5_IB_QP_UNDERLAY) {
2736 mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2740 err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2742 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2743 ibqp->qp_num, gid->raw);
2748 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2750 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2753 err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2755 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2756 ibqp->qp_num, gid->raw);
2761 static int init_node_data(struct mlx5_ib_dev *dev)
2765 err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2769 dev->mdev->rev_id = dev->mdev->pdev->revision;
2771 return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2774 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2777 struct mlx5_ib_dev *dev =
2778 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2780 return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
2783 static ssize_t show_reg_pages(struct device *device,
2784 struct device_attribute *attr, char *buf)
2786 struct mlx5_ib_dev *dev =
2787 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2789 return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2792 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2795 struct mlx5_ib_dev *dev =
2796 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2797 return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2800 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2803 struct mlx5_ib_dev *dev =
2804 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2805 return sprintf(buf, "%x\n", dev->mdev->rev_id);
2808 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2811 struct mlx5_ib_dev *dev =
2812 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2813 return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2814 dev->mdev->board_id);
2817 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
2818 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
2819 static DEVICE_ATTR(board_id, S_IRUGO, show_board, NULL);
2820 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2821 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2823 static struct device_attribute *mlx5_class_attributes[] = {
2828 &dev_attr_reg_pages,
2831 static void pkey_change_handler(struct work_struct *work)
2833 struct mlx5_ib_port_resources *ports =
2834 container_of(work, struct mlx5_ib_port_resources,
2837 mutex_lock(&ports->devr->mutex);
2838 mlx5_ib_gsi_pkey_change(ports->gsi);
2839 mutex_unlock(&ports->devr->mutex);
2842 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2844 struct mlx5_ib_qp *mqp;
2845 struct mlx5_ib_cq *send_mcq, *recv_mcq;
2846 struct mlx5_core_cq *mcq;
2847 struct list_head cq_armed_list;
2848 unsigned long flags_qp;
2849 unsigned long flags_cq;
2850 unsigned long flags;
2852 INIT_LIST_HEAD(&cq_armed_list);
2854 /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2855 spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2856 list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2857 spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2858 if (mqp->sq.tail != mqp->sq.head) {
2859 send_mcq = to_mcq(mqp->ibqp.send_cq);
2860 spin_lock_irqsave(&send_mcq->lock, flags_cq);
2861 if (send_mcq->mcq.comp &&
2862 mqp->ibqp.send_cq->comp_handler) {
2863 if (!send_mcq->mcq.reset_notify_added) {
2864 send_mcq->mcq.reset_notify_added = 1;
2865 list_add_tail(&send_mcq->mcq.reset_notify,
2869 spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2871 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2872 spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2873 /* no handling is needed for SRQ */
2874 if (!mqp->ibqp.srq) {
2875 if (mqp->rq.tail != mqp->rq.head) {
2876 recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2877 spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2878 if (recv_mcq->mcq.comp &&
2879 mqp->ibqp.recv_cq->comp_handler) {
2880 if (!recv_mcq->mcq.reset_notify_added) {
2881 recv_mcq->mcq.reset_notify_added = 1;
2882 list_add_tail(&recv_mcq->mcq.reset_notify,
2886 spin_unlock_irqrestore(&recv_mcq->lock,
2890 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2892 /*At that point all inflight post send were put to be executed as of we
2893 * lock/unlock above locks Now need to arm all involved CQs.
2895 list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2898 spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2901 static void delay_drop_handler(struct work_struct *work)
2904 struct mlx5_ib_delay_drop *delay_drop =
2905 container_of(work, struct mlx5_ib_delay_drop,
2908 atomic_inc(&delay_drop->events_cnt);
2910 mutex_lock(&delay_drop->lock);
2911 err = mlx5_core_set_delay_drop(delay_drop->dev->mdev,
2912 delay_drop->timeout);
2914 mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2915 delay_drop->timeout);
2916 delay_drop->activate = false;
2918 mutex_unlock(&delay_drop->lock);
2921 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2922 enum mlx5_dev_event event, unsigned long param)
2924 struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2925 struct ib_event ibev;
2930 case MLX5_DEV_EVENT_SYS_ERROR:
2931 ibev.event = IB_EVENT_DEVICE_FATAL;
2932 mlx5_ib_handle_internal_error(ibdev);
2936 case MLX5_DEV_EVENT_PORT_UP:
2937 case MLX5_DEV_EVENT_PORT_DOWN:
2938 case MLX5_DEV_EVENT_PORT_INITIALIZED:
2941 /* In RoCE, port up/down events are handled in
2942 * mlx5_netdev_event().
2944 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2945 IB_LINK_LAYER_ETHERNET)
2948 ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2949 IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2952 case MLX5_DEV_EVENT_LID_CHANGE:
2953 ibev.event = IB_EVENT_LID_CHANGE;
2957 case MLX5_DEV_EVENT_PKEY_CHANGE:
2958 ibev.event = IB_EVENT_PKEY_CHANGE;
2961 schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2964 case MLX5_DEV_EVENT_GUID_CHANGE:
2965 ibev.event = IB_EVENT_GID_CHANGE;
2969 case MLX5_DEV_EVENT_CLIENT_REREG:
2970 ibev.event = IB_EVENT_CLIENT_REREGISTER;
2973 case MLX5_DEV_EVENT_DELAY_DROP_TIMEOUT:
2974 schedule_work(&ibdev->delay_drop.delay_drop_work);
2980 ibev.device = &ibdev->ib_dev;
2981 ibev.element.port_num = port;
2983 if (port < 1 || port > ibdev->num_ports) {
2984 mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2988 if (ibdev->ib_active)
2989 ib_dispatch_event(&ibev);
2992 ibdev->ib_active = false;
2998 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
3000 struct mlx5_hca_vport_context vport_ctx;
3004 for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
3005 dev->mdev->port_caps[port - 1].has_smi = false;
3006 if (MLX5_CAP_GEN(dev->mdev, port_type) ==
3007 MLX5_CAP_PORT_TYPE_IB) {
3008 if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
3009 err = mlx5_query_hca_vport_context(dev->mdev, 0,
3013 mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
3017 dev->mdev->port_caps[port - 1].has_smi =
3020 dev->mdev->port_caps[port - 1].has_smi = true;
3027 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
3031 for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
3032 mlx5_query_ext_port_caps(dev, port);
3035 static int get_port_caps(struct mlx5_ib_dev *dev)
3037 struct ib_device_attr *dprops = NULL;
3038 struct ib_port_attr *pprops = NULL;
3041 struct ib_udata uhw = {.inlen = 0, .outlen = 0};
3043 pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
3047 dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
3051 err = set_has_smi_cap(dev);
3055 err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
3057 mlx5_ib_warn(dev, "query_device failed %d\n", err);
3061 for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
3062 memset(pprops, 0, sizeof(*pprops));
3063 err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
3065 mlx5_ib_warn(dev, "query_port %d failed %d\n",
3069 dev->mdev->port_caps[port - 1].pkey_table_len =
3071 dev->mdev->port_caps[port - 1].gid_table_len =
3072 pprops->gid_tbl_len;
3073 mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
3074 dprops->max_pkeys, pprops->gid_tbl_len);
3084 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
3088 err = mlx5_mr_cache_cleanup(dev);
3090 mlx5_ib_warn(dev, "mr cache cleanup failed\n");
3092 mlx5_ib_destroy_qp(dev->umrc.qp);
3093 ib_free_cq(dev->umrc.cq);
3094 ib_dealloc_pd(dev->umrc.pd);
3101 static int create_umr_res(struct mlx5_ib_dev *dev)
3103 struct ib_qp_init_attr *init_attr = NULL;
3104 struct ib_qp_attr *attr = NULL;
3110 attr = kzalloc(sizeof(*attr), GFP_KERNEL);
3111 init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
3112 if (!attr || !init_attr) {
3117 pd = ib_alloc_pd(&dev->ib_dev, 0);
3119 mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
3124 cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
3126 mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
3131 init_attr->send_cq = cq;
3132 init_attr->recv_cq = cq;
3133 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
3134 init_attr->cap.max_send_wr = MAX_UMR_WR;
3135 init_attr->cap.max_send_sge = 1;
3136 init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
3137 init_attr->port_num = 1;
3138 qp = mlx5_ib_create_qp(pd, init_attr, NULL);
3140 mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
3144 qp->device = &dev->ib_dev;
3147 qp->qp_type = MLX5_IB_QPT_REG_UMR;
3148 qp->send_cq = init_attr->send_cq;
3149 qp->recv_cq = init_attr->recv_cq;
3151 attr->qp_state = IB_QPS_INIT;
3153 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
3156 mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
3160 memset(attr, 0, sizeof(*attr));
3161 attr->qp_state = IB_QPS_RTR;
3162 attr->path_mtu = IB_MTU_256;
3164 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
3166 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
3170 memset(attr, 0, sizeof(*attr));
3171 attr->qp_state = IB_QPS_RTS;
3172 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
3174 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
3182 sema_init(&dev->umrc.sem, MAX_UMR_WR);
3183 ret = mlx5_mr_cache_init(dev);
3185 mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
3195 mlx5_ib_destroy_qp(qp);
3209 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
3211 switch (umr_fence_cap) {
3212 case MLX5_CAP_UMR_FENCE_NONE:
3213 return MLX5_FENCE_MODE_NONE;
3214 case MLX5_CAP_UMR_FENCE_SMALL:
3215 return MLX5_FENCE_MODE_INITIATOR_SMALL;
3217 return MLX5_FENCE_MODE_STRONG_ORDERING;
3221 static int create_dev_resources(struct mlx5_ib_resources *devr)
3223 struct ib_srq_init_attr attr;
3224 struct mlx5_ib_dev *dev;
3225 struct ib_cq_init_attr cq_attr = {.cqe = 1};
3229 dev = container_of(devr, struct mlx5_ib_dev, devr);
3231 mutex_init(&devr->mutex);
3233 devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
3234 if (IS_ERR(devr->p0)) {
3235 ret = PTR_ERR(devr->p0);
3238 devr->p0->device = &dev->ib_dev;
3239 devr->p0->uobject = NULL;
3240 atomic_set(&devr->p0->usecnt, 0);
3242 devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
3243 if (IS_ERR(devr->c0)) {
3244 ret = PTR_ERR(devr->c0);
3247 devr->c0->device = &dev->ib_dev;
3248 devr->c0->uobject = NULL;
3249 devr->c0->comp_handler = NULL;
3250 devr->c0->event_handler = NULL;
3251 devr->c0->cq_context = NULL;
3252 atomic_set(&devr->c0->usecnt, 0);
3254 devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
3255 if (IS_ERR(devr->x0)) {
3256 ret = PTR_ERR(devr->x0);
3259 devr->x0->device = &dev->ib_dev;
3260 devr->x0->inode = NULL;
3261 atomic_set(&devr->x0->usecnt, 0);
3262 mutex_init(&devr->x0->tgt_qp_mutex);
3263 INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
3265 devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
3266 if (IS_ERR(devr->x1)) {
3267 ret = PTR_ERR(devr->x1);
3270 devr->x1->device = &dev->ib_dev;
3271 devr->x1->inode = NULL;
3272 atomic_set(&devr->x1->usecnt, 0);
3273 mutex_init(&devr->x1->tgt_qp_mutex);
3274 INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
3276 memset(&attr, 0, sizeof(attr));
3277 attr.attr.max_sge = 1;
3278 attr.attr.max_wr = 1;
3279 attr.srq_type = IB_SRQT_XRC;
3280 attr.ext.cq = devr->c0;
3281 attr.ext.xrc.xrcd = devr->x0;
3283 devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
3284 if (IS_ERR(devr->s0)) {
3285 ret = PTR_ERR(devr->s0);
3288 devr->s0->device = &dev->ib_dev;
3289 devr->s0->pd = devr->p0;
3290 devr->s0->uobject = NULL;
3291 devr->s0->event_handler = NULL;
3292 devr->s0->srq_context = NULL;
3293 devr->s0->srq_type = IB_SRQT_XRC;
3294 devr->s0->ext.xrc.xrcd = devr->x0;
3295 devr->s0->ext.cq = devr->c0;
3296 atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
3297 atomic_inc(&devr->s0->ext.cq->usecnt);
3298 atomic_inc(&devr->p0->usecnt);
3299 atomic_set(&devr->s0->usecnt, 0);
3301 memset(&attr, 0, sizeof(attr));
3302 attr.attr.max_sge = 1;
3303 attr.attr.max_wr = 1;
3304 attr.srq_type = IB_SRQT_BASIC;
3305 devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
3306 if (IS_ERR(devr->s1)) {
3307 ret = PTR_ERR(devr->s1);
3310 devr->s1->device = &dev->ib_dev;
3311 devr->s1->pd = devr->p0;
3312 devr->s1->uobject = NULL;
3313 devr->s1->event_handler = NULL;
3314 devr->s1->srq_context = NULL;
3315 devr->s1->srq_type = IB_SRQT_BASIC;
3316 devr->s1->ext.cq = devr->c0;
3317 atomic_inc(&devr->p0->usecnt);
3318 atomic_set(&devr->s1->usecnt, 0);
3320 for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
3321 INIT_WORK(&devr->ports[port].pkey_change_work,
3322 pkey_change_handler);
3323 devr->ports[port].devr = devr;
3329 mlx5_ib_destroy_srq(devr->s0);
3331 mlx5_ib_dealloc_xrcd(devr->x1);
3333 mlx5_ib_dealloc_xrcd(devr->x0);
3335 mlx5_ib_destroy_cq(devr->c0);
3337 mlx5_ib_dealloc_pd(devr->p0);
3342 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
3344 struct mlx5_ib_dev *dev =
3345 container_of(devr, struct mlx5_ib_dev, devr);
3348 mlx5_ib_destroy_srq(devr->s1);
3349 mlx5_ib_destroy_srq(devr->s0);
3350 mlx5_ib_dealloc_xrcd(devr->x0);
3351 mlx5_ib_dealloc_xrcd(devr->x1);
3352 mlx5_ib_destroy_cq(devr->c0);
3353 mlx5_ib_dealloc_pd(devr->p0);
3355 /* Make sure no change P_Key work items are still executing */
3356 for (port = 0; port < dev->num_ports; ++port)
3357 cancel_work_sync(&devr->ports[port].pkey_change_work);
3360 static u32 get_core_cap_flags(struct ib_device *ibdev)
3362 struct mlx5_ib_dev *dev = to_mdev(ibdev);
3363 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
3364 u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
3365 u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
3368 if (ll == IB_LINK_LAYER_INFINIBAND)
3369 return RDMA_CORE_PORT_IBA_IB;
3371 ret = RDMA_CORE_PORT_RAW_PACKET;
3373 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
3376 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
3379 if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
3380 ret |= RDMA_CORE_PORT_IBA_ROCE;
3382 if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
3383 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3388 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
3389 struct ib_port_immutable *immutable)
3391 struct ib_port_attr attr;
3392 struct mlx5_ib_dev *dev = to_mdev(ibdev);
3393 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
3396 immutable->core_cap_flags = get_core_cap_flags(ibdev);
3398 err = ib_query_port(ibdev, port_num, &attr);
3402 immutable->pkey_tbl_len = attr.pkey_tbl_len;
3403 immutable->gid_tbl_len = attr.gid_tbl_len;
3404 immutable->core_cap_flags = get_core_cap_flags(ibdev);
3405 if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
3406 immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3411 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
3413 struct mlx5_ib_dev *dev =
3414 container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3415 snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
3416 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
3417 fw_rev_sub(dev->mdev));
3420 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3422 struct mlx5_core_dev *mdev = dev->mdev;
3423 struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3424 MLX5_FLOW_NAMESPACE_LAG);
3425 struct mlx5_flow_table *ft;
3428 if (!ns || !mlx5_lag_is_active(mdev))
3431 err = mlx5_cmd_create_vport_lag(mdev);
3435 ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3438 goto err_destroy_vport_lag;
3441 dev->flow_db.lag_demux_ft = ft;
3444 err_destroy_vport_lag:
3445 mlx5_cmd_destroy_vport_lag(mdev);
3449 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3451 struct mlx5_core_dev *mdev = dev->mdev;
3453 if (dev->flow_db.lag_demux_ft) {
3454 mlx5_destroy_flow_table(dev->flow_db.lag_demux_ft);
3455 dev->flow_db.lag_demux_ft = NULL;
3457 mlx5_cmd_destroy_vport_lag(mdev);
3461 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev)
3465 dev->roce.nb.notifier_call = mlx5_netdev_event;
3466 err = register_netdevice_notifier(&dev->roce.nb);
3468 dev->roce.nb.notifier_call = NULL;
3475 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev)
3477 if (dev->roce.nb.notifier_call) {
3478 unregister_netdevice_notifier(&dev->roce.nb);
3479 dev->roce.nb.notifier_call = NULL;
3483 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3487 err = mlx5_add_netdev_notifier(dev);
3491 if (MLX5_CAP_GEN(dev->mdev, roce)) {
3492 err = mlx5_nic_vport_enable_roce(dev->mdev);
3494 goto err_unregister_netdevice_notifier;
3497 err = mlx5_eth_lag_init(dev);
3499 goto err_disable_roce;
3504 if (MLX5_CAP_GEN(dev->mdev, roce))
3505 mlx5_nic_vport_disable_roce(dev->mdev);
3507 err_unregister_netdevice_notifier:
3508 mlx5_remove_netdev_notifier(dev);
3512 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3514 mlx5_eth_lag_cleanup(dev);
3515 if (MLX5_CAP_GEN(dev->mdev, roce))
3516 mlx5_nic_vport_disable_roce(dev->mdev);
3519 struct mlx5_ib_counter {
3524 #define INIT_Q_COUNTER(_name) \
3525 { .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
3527 static const struct mlx5_ib_counter basic_q_cnts[] = {
3528 INIT_Q_COUNTER(rx_write_requests),
3529 INIT_Q_COUNTER(rx_read_requests),
3530 INIT_Q_COUNTER(rx_atomic_requests),
3531 INIT_Q_COUNTER(out_of_buffer),
3534 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
3535 INIT_Q_COUNTER(out_of_sequence),
3538 static const struct mlx5_ib_counter retrans_q_cnts[] = {
3539 INIT_Q_COUNTER(duplicate_request),
3540 INIT_Q_COUNTER(rnr_nak_retry_err),
3541 INIT_Q_COUNTER(packet_seq_err),
3542 INIT_Q_COUNTER(implied_nak_seq_err),
3543 INIT_Q_COUNTER(local_ack_timeout_err),
3546 #define INIT_CONG_COUNTER(_name) \
3547 { .name = #_name, .offset = \
3548 MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
3550 static const struct mlx5_ib_counter cong_cnts[] = {
3551 INIT_CONG_COUNTER(rp_cnp_ignored),
3552 INIT_CONG_COUNTER(rp_cnp_handled),
3553 INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
3554 INIT_CONG_COUNTER(np_cnp_sent),
3557 static const struct mlx5_ib_counter extended_err_cnts[] = {
3558 INIT_Q_COUNTER(resp_local_length_error),
3559 INIT_Q_COUNTER(resp_cqe_error),
3560 INIT_Q_COUNTER(req_cqe_error),
3561 INIT_Q_COUNTER(req_remote_invalid_request),
3562 INIT_Q_COUNTER(req_remote_access_errors),
3563 INIT_Q_COUNTER(resp_remote_access_errors),
3564 INIT_Q_COUNTER(resp_cqe_flush_error),
3565 INIT_Q_COUNTER(req_cqe_flush_error),
3568 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
3572 for (i = 0; i < dev->num_ports; i++) {
3573 mlx5_core_dealloc_q_counter(dev->mdev,
3574 dev->port[i].cnts.set_id);
3575 kfree(dev->port[i].cnts.names);
3576 kfree(dev->port[i].cnts.offsets);
3580 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
3581 struct mlx5_ib_counters *cnts)
3585 num_counters = ARRAY_SIZE(basic_q_cnts);
3587 if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
3588 num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
3590 if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
3591 num_counters += ARRAY_SIZE(retrans_q_cnts);
3593 if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters))
3594 num_counters += ARRAY_SIZE(extended_err_cnts);
3596 cnts->num_q_counters = num_counters;
3598 if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
3599 cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
3600 num_counters += ARRAY_SIZE(cong_cnts);
3603 cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
3607 cnts->offsets = kcalloc(num_counters,
3608 sizeof(cnts->offsets), GFP_KERNEL);
3619 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
3626 for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
3627 names[j] = basic_q_cnts[i].name;
3628 offsets[j] = basic_q_cnts[i].offset;
3631 if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
3632 for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
3633 names[j] = out_of_seq_q_cnts[i].name;
3634 offsets[j] = out_of_seq_q_cnts[i].offset;
3638 if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3639 for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
3640 names[j] = retrans_q_cnts[i].name;
3641 offsets[j] = retrans_q_cnts[i].offset;
3645 if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters)) {
3646 for (i = 0; i < ARRAY_SIZE(extended_err_cnts); i++, j++) {
3647 names[j] = extended_err_cnts[i].name;
3648 offsets[j] = extended_err_cnts[i].offset;
3652 if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
3653 for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
3654 names[j] = cong_cnts[i].name;
3655 offsets[j] = cong_cnts[i].offset;
3660 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
3665 for (i = 0; i < dev->num_ports; i++) {
3666 struct mlx5_ib_port *port = &dev->port[i];
3668 ret = mlx5_core_alloc_q_counter(dev->mdev,
3669 &port->cnts.set_id);
3672 "couldn't allocate queue counter for port %d, err %d\n",
3674 goto dealloc_counters;
3677 ret = __mlx5_ib_alloc_counters(dev, &port->cnts);
3679 goto dealloc_counters;
3681 mlx5_ib_fill_counters(dev, port->cnts.names,
3682 port->cnts.offsets);
3689 mlx5_core_dealloc_q_counter(dev->mdev,
3690 dev->port[i].cnts.set_id);
3695 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
3698 struct mlx5_ib_dev *dev = to_mdev(ibdev);
3699 struct mlx5_ib_port *port = &dev->port[port_num - 1];
3701 /* We support only per port stats */
3705 return rdma_alloc_hw_stats_struct(port->cnts.names,
3706 port->cnts.num_q_counters +
3707 port->cnts.num_cong_counters,
3708 RDMA_HW_STATS_DEFAULT_LIFESPAN);
3711 static int mlx5_ib_query_q_counters(struct mlx5_ib_dev *dev,
3712 struct mlx5_ib_port *port,
3713 struct rdma_hw_stats *stats)
3715 int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
3720 out = kvzalloc(outlen, GFP_KERNEL);
3724 ret = mlx5_core_query_q_counter(dev->mdev,
3725 port->cnts.set_id, 0,
3730 for (i = 0; i < port->cnts.num_q_counters; i++) {
3731 val = *(__be32 *)(out + port->cnts.offsets[i]);
3732 stats->value[i] = (u64)be32_to_cpu(val);
3740 static int mlx5_ib_query_cong_counters(struct mlx5_ib_dev *dev,
3741 struct mlx5_ib_port *port,
3742 struct rdma_hw_stats *stats)
3744 int outlen = MLX5_ST_SZ_BYTES(query_cong_statistics_out);
3747 int offset = port->cnts.num_q_counters;
3749 out = kvzalloc(outlen, GFP_KERNEL);
3753 ret = mlx5_cmd_query_cong_counter(dev->mdev, false, out, outlen);
3757 for (i = 0; i < port->cnts.num_cong_counters; i++) {
3758 stats->value[i + offset] =
3759 be64_to_cpup((__be64 *)(out +
3760 port->cnts.offsets[i + offset]));
3768 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
3769 struct rdma_hw_stats *stats,
3770 u8 port_num, int index)
3772 struct mlx5_ib_dev *dev = to_mdev(ibdev);
3773 struct mlx5_ib_port *port = &dev->port[port_num - 1];
3774 int ret, num_counters;
3779 ret = mlx5_ib_query_q_counters(dev, port, stats);
3782 num_counters = port->cnts.num_q_counters;
3784 if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
3785 ret = mlx5_ib_query_cong_counters(dev, port, stats);
3788 num_counters += port->cnts.num_cong_counters;
3791 return num_counters;
3794 static void mlx5_ib_free_rdma_netdev(struct net_device *netdev)
3796 return mlx5_rdma_netdev_free(netdev);
3799 static struct net_device*
3800 mlx5_ib_alloc_rdma_netdev(struct ib_device *hca,
3802 enum rdma_netdev_t type,
3804 unsigned char name_assign_type,
3805 void (*setup)(struct net_device *))
3807 struct net_device *netdev;
3808 struct rdma_netdev *rn;
3810 if (type != RDMA_NETDEV_IPOIB)
3811 return ERR_PTR(-EOPNOTSUPP);
3813 netdev = mlx5_rdma_netdev_alloc(to_mdev(hca)->mdev, hca,
3815 if (likely(!IS_ERR_OR_NULL(netdev))) {
3816 rn = netdev_priv(netdev);
3817 rn->free_rdma_netdev = mlx5_ib_free_rdma_netdev;
3822 static void delay_drop_debugfs_cleanup(struct mlx5_ib_dev *dev)
3824 if (!dev->delay_drop.dbg)
3826 debugfs_remove_recursive(dev->delay_drop.dbg->dir_debugfs);
3827 kfree(dev->delay_drop.dbg);
3828 dev->delay_drop.dbg = NULL;
3831 static void cancel_delay_drop(struct mlx5_ib_dev *dev)
3833 if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
3836 cancel_work_sync(&dev->delay_drop.delay_drop_work);
3837 delay_drop_debugfs_cleanup(dev);
3840 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3841 size_t count, loff_t *pos)
3843 struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3847 len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3848 return simple_read_from_buffer(buf, count, pos, lbuf, len);
3851 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3852 size_t count, loff_t *pos)
3854 struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3858 if (kstrtouint_from_user(buf, count, 0, &var))
3861 timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3864 mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3867 delay_drop->timeout = timeout;
3872 static const struct file_operations fops_delay_drop_timeout = {
3873 .owner = THIS_MODULE,
3874 .open = simple_open,
3875 .write = delay_drop_timeout_write,
3876 .read = delay_drop_timeout_read,
3879 static int delay_drop_debugfs_init(struct mlx5_ib_dev *dev)
3881 struct mlx5_ib_dbg_delay_drop *dbg;
3883 if (!mlx5_debugfs_root)
3886 dbg = kzalloc(sizeof(*dbg), GFP_KERNEL);
3890 dev->delay_drop.dbg = dbg;
3893 debugfs_create_dir("delay_drop",
3894 dev->mdev->priv.dbg_root);
3895 if (!dbg->dir_debugfs)
3898 dbg->events_cnt_debugfs =
3899 debugfs_create_atomic_t("num_timeout_events", 0400,
3901 &dev->delay_drop.events_cnt);
3902 if (!dbg->events_cnt_debugfs)
3905 dbg->rqs_cnt_debugfs =
3906 debugfs_create_atomic_t("num_rqs", 0400,
3908 &dev->delay_drop.rqs_cnt);
3909 if (!dbg->rqs_cnt_debugfs)
3912 dbg->timeout_debugfs =
3913 debugfs_create_file("timeout", 0600,
3916 &fops_delay_drop_timeout);
3917 if (!dbg->timeout_debugfs)
3923 delay_drop_debugfs_cleanup(dev);
3927 static void init_delay_drop(struct mlx5_ib_dev *dev)
3929 if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
3932 mutex_init(&dev->delay_drop.lock);
3933 dev->delay_drop.dev = dev;
3934 dev->delay_drop.activate = false;
3935 dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
3936 INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
3937 atomic_set(&dev->delay_drop.rqs_cnt, 0);
3938 atomic_set(&dev->delay_drop.events_cnt, 0);
3940 if (delay_drop_debugfs_init(dev))
3941 mlx5_ib_warn(dev, "Failed to init delay drop debugfs\n");
3944 static const struct cpumask *
3945 mlx5_ib_get_vector_affinity(struct ib_device *ibdev, int comp_vector)
3947 struct mlx5_ib_dev *dev = to_mdev(ibdev);
3949 return mlx5_get_vector_affinity(dev->mdev, comp_vector);
3952 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
3954 struct mlx5_ib_dev *dev;
3955 enum rdma_link_layer ll;
3961 port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3962 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3964 printk_once(KERN_INFO "%s", mlx5_version);
3966 dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
3972 dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
3977 rwlock_init(&dev->roce.netdev_lock);
3978 err = get_port_caps(dev);
3982 if (mlx5_use_mad_ifc(dev))
3983 get_ext_port_caps(dev);
3985 if (!mlx5_lag_is_active(mdev))
3988 name = "mlx5_bond_%d";
3990 strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
3991 dev->ib_dev.owner = THIS_MODULE;
3992 dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3993 dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3994 dev->num_ports = MLX5_CAP_GEN(mdev, num_ports);
3995 dev->ib_dev.phys_port_cnt = dev->num_ports;
3996 dev->ib_dev.num_comp_vectors =
3997 dev->mdev->priv.eq_table.num_comp_vectors;
3998 dev->ib_dev.dev.parent = &mdev->pdev->dev;
4000 dev->ib_dev.uverbs_abi_ver = MLX5_IB_UVERBS_ABI_VERSION;
4001 dev->ib_dev.uverbs_cmd_mask =
4002 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
4003 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
4004 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
4005 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
4006 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
4007 (1ull << IB_USER_VERBS_CMD_CREATE_AH) |
4008 (1ull << IB_USER_VERBS_CMD_DESTROY_AH) |
4009 (1ull << IB_USER_VERBS_CMD_REG_MR) |
4010 (1ull << IB_USER_VERBS_CMD_REREG_MR) |
4011 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
4012 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
4013 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
4014 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
4015 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
4016 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
4017 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
4018 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
4019 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
4020 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
4021 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) |
4022 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
4023 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
4024 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
4025 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) |
4026 (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ) |
4027 (1ull << IB_USER_VERBS_CMD_OPEN_QP);
4028 dev->ib_dev.uverbs_ex_cmd_mask =
4029 (1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE) |
4030 (1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ) |
4031 (1ull << IB_USER_VERBS_EX_CMD_CREATE_QP) |
4032 (1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP) |
4033 (1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ);
4035 dev->ib_dev.query_device = mlx5_ib_query_device;
4036 dev->ib_dev.query_port = mlx5_ib_query_port;
4037 dev->ib_dev.get_link_layer = mlx5_ib_port_link_layer;
4038 if (ll == IB_LINK_LAYER_ETHERNET)
4039 dev->ib_dev.get_netdev = mlx5_ib_get_netdev;
4040 dev->ib_dev.query_gid = mlx5_ib_query_gid;
4041 dev->ib_dev.add_gid = mlx5_ib_add_gid;
4042 dev->ib_dev.del_gid = mlx5_ib_del_gid;
4043 dev->ib_dev.query_pkey = mlx5_ib_query_pkey;
4044 dev->ib_dev.modify_device = mlx5_ib_modify_device;
4045 dev->ib_dev.modify_port = mlx5_ib_modify_port;
4046 dev->ib_dev.alloc_ucontext = mlx5_ib_alloc_ucontext;
4047 dev->ib_dev.dealloc_ucontext = mlx5_ib_dealloc_ucontext;
4048 dev->ib_dev.mmap = mlx5_ib_mmap;
4049 dev->ib_dev.alloc_pd = mlx5_ib_alloc_pd;
4050 dev->ib_dev.dealloc_pd = mlx5_ib_dealloc_pd;
4051 dev->ib_dev.create_ah = mlx5_ib_create_ah;
4052 dev->ib_dev.query_ah = mlx5_ib_query_ah;
4053 dev->ib_dev.destroy_ah = mlx5_ib_destroy_ah;
4054 dev->ib_dev.create_srq = mlx5_ib_create_srq;
4055 dev->ib_dev.modify_srq = mlx5_ib_modify_srq;
4056 dev->ib_dev.query_srq = mlx5_ib_query_srq;
4057 dev->ib_dev.destroy_srq = mlx5_ib_destroy_srq;
4058 dev->ib_dev.post_srq_recv = mlx5_ib_post_srq_recv;
4059 dev->ib_dev.create_qp = mlx5_ib_create_qp;
4060 dev->ib_dev.modify_qp = mlx5_ib_modify_qp;
4061 dev->ib_dev.query_qp = mlx5_ib_query_qp;
4062 dev->ib_dev.destroy_qp = mlx5_ib_destroy_qp;
4063 dev->ib_dev.post_send = mlx5_ib_post_send;
4064 dev->ib_dev.post_recv = mlx5_ib_post_recv;
4065 dev->ib_dev.create_cq = mlx5_ib_create_cq;
4066 dev->ib_dev.modify_cq = mlx5_ib_modify_cq;
4067 dev->ib_dev.resize_cq = mlx5_ib_resize_cq;
4068 dev->ib_dev.destroy_cq = mlx5_ib_destroy_cq;
4069 dev->ib_dev.poll_cq = mlx5_ib_poll_cq;
4070 dev->ib_dev.req_notify_cq = mlx5_ib_arm_cq;
4071 dev->ib_dev.get_dma_mr = mlx5_ib_get_dma_mr;
4072 dev->ib_dev.reg_user_mr = mlx5_ib_reg_user_mr;
4073 dev->ib_dev.rereg_user_mr = mlx5_ib_rereg_user_mr;
4074 dev->ib_dev.dereg_mr = mlx5_ib_dereg_mr;
4075 dev->ib_dev.attach_mcast = mlx5_ib_mcg_attach;
4076 dev->ib_dev.detach_mcast = mlx5_ib_mcg_detach;
4077 dev->ib_dev.process_mad = mlx5_ib_process_mad;
4078 dev->ib_dev.alloc_mr = mlx5_ib_alloc_mr;
4079 dev->ib_dev.map_mr_sg = mlx5_ib_map_mr_sg;
4080 dev->ib_dev.check_mr_status = mlx5_ib_check_mr_status;
4081 dev->ib_dev.get_port_immutable = mlx5_port_immutable;
4082 dev->ib_dev.get_dev_fw_str = get_dev_fw_str;
4083 dev->ib_dev.get_vector_affinity = mlx5_ib_get_vector_affinity;
4084 if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads))
4085 dev->ib_dev.alloc_rdma_netdev = mlx5_ib_alloc_rdma_netdev;
4087 if (mlx5_core_is_pf(mdev)) {
4088 dev->ib_dev.get_vf_config = mlx5_ib_get_vf_config;
4089 dev->ib_dev.set_vf_link_state = mlx5_ib_set_vf_link_state;
4090 dev->ib_dev.get_vf_stats = mlx5_ib_get_vf_stats;
4091 dev->ib_dev.set_vf_guid = mlx5_ib_set_vf_guid;
4094 dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
4096 mlx5_ib_internal_fill_odp_caps(dev);
4098 dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
4100 if (MLX5_CAP_GEN(mdev, imaicl)) {
4101 dev->ib_dev.alloc_mw = mlx5_ib_alloc_mw;
4102 dev->ib_dev.dealloc_mw = mlx5_ib_dealloc_mw;
4103 dev->ib_dev.uverbs_cmd_mask |=
4104 (1ull << IB_USER_VERBS_CMD_ALLOC_MW) |
4105 (1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
4108 if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
4109 dev->ib_dev.get_hw_stats = mlx5_ib_get_hw_stats;
4110 dev->ib_dev.alloc_hw_stats = mlx5_ib_alloc_hw_stats;
4113 if (MLX5_CAP_GEN(mdev, xrc)) {
4114 dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
4115 dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
4116 dev->ib_dev.uverbs_cmd_mask |=
4117 (1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
4118 (1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
4121 dev->ib_dev.create_flow = mlx5_ib_create_flow;
4122 dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
4123 dev->ib_dev.uverbs_ex_cmd_mask |=
4124 (1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
4125 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
4127 if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
4128 IB_LINK_LAYER_ETHERNET) {
4129 dev->ib_dev.create_wq = mlx5_ib_create_wq;
4130 dev->ib_dev.modify_wq = mlx5_ib_modify_wq;
4131 dev->ib_dev.destroy_wq = mlx5_ib_destroy_wq;
4132 dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
4133 dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
4134 dev->ib_dev.uverbs_ex_cmd_mask |=
4135 (1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
4136 (1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
4137 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
4138 (1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
4139 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
4141 err = init_node_data(dev);
4145 mutex_init(&dev->flow_db.lock);
4146 mutex_init(&dev->cap_mask_mutex);
4147 INIT_LIST_HEAD(&dev->qp_list);
4148 spin_lock_init(&dev->reset_flow_resource_lock);
4150 if (ll == IB_LINK_LAYER_ETHERNET) {
4151 err = mlx5_enable_eth(dev);
4154 dev->roce.last_port_state = IB_PORT_DOWN;
4157 err = create_dev_resources(&dev->devr);
4159 goto err_disable_eth;
4161 err = mlx5_ib_odp_init_one(dev);
4165 if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
4166 err = mlx5_ib_alloc_counters(dev);
4171 err = mlx5_ib_init_cong_debugfs(dev);
4175 dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4176 if (!dev->mdev->priv.uar)
4179 err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4183 err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4187 err = ib_register_device(&dev->ib_dev, NULL);
4191 err = create_umr_res(dev);
4195 init_delay_drop(dev);
4197 for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
4198 err = device_create_file(&dev->ib_dev.dev,
4199 mlx5_class_attributes[i]);
4201 goto err_delay_drop;
4204 if ((MLX5_CAP_GEN(mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
4205 MLX5_CAP_GEN(mdev, disable_local_lb))
4206 mutex_init(&dev->lb_mutex);
4208 dev->ib_active = true;
4213 cancel_delay_drop(dev);
4214 destroy_umrc_res(dev);
4217 ib_unregister_device(&dev->ib_dev);
4220 mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4223 mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4226 mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4229 mlx5_ib_cleanup_cong_debugfs(dev);
4231 if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
4232 mlx5_ib_dealloc_counters(dev);
4235 mlx5_ib_odp_remove_one(dev);
4238 destroy_dev_resources(&dev->devr);
4241 if (ll == IB_LINK_LAYER_ETHERNET) {
4242 mlx5_disable_eth(dev);
4243 mlx5_remove_netdev_notifier(dev);
4250 ib_dealloc_device((struct ib_device *)dev);
4255 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
4257 struct mlx5_ib_dev *dev = context;
4258 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
4260 cancel_delay_drop(dev);
4261 mlx5_remove_netdev_notifier(dev);
4262 ib_unregister_device(&dev->ib_dev);
4263 mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4264 mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4265 mlx5_put_uars_page(dev->mdev, mdev->priv.uar);
4266 mlx5_ib_cleanup_cong_debugfs(dev);
4267 if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
4268 mlx5_ib_dealloc_counters(dev);
4269 destroy_umrc_res(dev);
4270 mlx5_ib_odp_remove_one(dev);
4271 destroy_dev_resources(&dev->devr);
4272 if (ll == IB_LINK_LAYER_ETHERNET)
4273 mlx5_disable_eth(dev);
4275 ib_dealloc_device(&dev->ib_dev);
4278 static struct mlx5_interface mlx5_ib_interface = {
4280 .remove = mlx5_ib_remove,
4281 .event = mlx5_ib_event,
4282 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4283 .pfault = mlx5_ib_pfault,
4285 .protocol = MLX5_INTERFACE_PROTOCOL_IB,
4288 static int __init mlx5_ib_init(void)
4294 err = mlx5_register_interface(&mlx5_ib_interface);
4299 static void __exit mlx5_ib_cleanup(void)
4301 mlx5_unregister_interface(&mlx5_ib_interface);
4304 module_init(mlx5_ib_init);
4305 module_exit(mlx5_ib_cleanup);