2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/highmem.h>
34 #include <linux/module.h>
35 #include <linux/init.h>
36 #include <linux/errno.h>
37 #include <linux/pci.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/slab.h>
40 #if defined(CONFIG_X86)
43 #include <linux/sched.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/task.h>
46 #include <linux/delay.h>
47 #include <rdma/ib_user_verbs.h>
48 #include <rdma/ib_addr.h>
49 #include <rdma/ib_cache.h>
50 #include <linux/mlx5/port.h>
51 #include <linux/mlx5/vport.h>
52 #include <linux/list.h>
53 #include <rdma/ib_smi.h>
54 #include <rdma/ib_umem.h>
56 #include <linux/etherdevice.h>
57 #include <linux/mlx5/fs.h>
58 #include <linux/mlx5/vport.h>
61 #include <linux/mlx5/vport.h>
63 #define DRIVER_NAME "mlx5_ib"
64 #define DRIVER_VERSION "5.0-0"
66 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
67 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
68 MODULE_LICENSE("Dual BSD/GPL");
69 MODULE_VERSION(DRIVER_VERSION);
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, eth_net_offloads) &&
701 MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
702 /* Legacy bit to support old userspace libraries */
703 props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
704 props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
707 if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
708 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
710 props->vendor_part_id = mdev->pdev->device;
711 props->hw_ver = mdev->pdev->revision;
713 props->max_mr_size = ~0ull;
714 props->page_size_cap = ~(min_page_size - 1);
715 props->max_qp = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
716 props->max_qp_wr = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
717 max_rq_sg = MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
718 sizeof(struct mlx5_wqe_data_seg);
719 max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
720 max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
721 sizeof(struct mlx5_wqe_raddr_seg)) /
722 sizeof(struct mlx5_wqe_data_seg);
723 props->max_sge = min(max_rq_sg, max_sq_sg);
724 props->max_sge_rd = MLX5_MAX_SGE_RD;
725 props->max_cq = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
726 props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
727 props->max_mr = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
728 props->max_pd = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
729 props->max_qp_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
730 props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
731 props->max_srq = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
732 props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
733 props->local_ca_ack_delay = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
734 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp;
735 props->max_srq_sge = max_rq_sg - 1;
736 props->max_fast_reg_page_list_len =
737 1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
738 get_atomic_caps(dev, props);
739 props->masked_atomic_cap = IB_ATOMIC_NONE;
740 props->max_mcast_grp = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
741 props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
742 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
743 props->max_mcast_grp;
744 props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
745 props->max_ah = INT_MAX;
746 props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
747 props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
749 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
750 if (MLX5_CAP_GEN(mdev, pg))
751 props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
752 props->odp_caps = dev->odp_caps;
755 if (MLX5_CAP_GEN(mdev, cd))
756 props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
758 if (!mlx5_core_is_pf(mdev))
759 props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
761 if (mlx5_ib_port_link_layer(ibdev, 1) ==
762 IB_LINK_LAYER_ETHERNET) {
763 props->rss_caps.max_rwq_indirection_tables =
764 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
765 props->rss_caps.max_rwq_indirection_table_size =
766 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
767 props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
768 props->max_wq_type_rq =
769 1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
772 if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
773 resp.cqe_comp_caps.max_num =
774 MLX5_CAP_GEN(dev->mdev, cqe_compression) ?
775 MLX5_CAP_GEN(dev->mdev, cqe_compression_max_num) : 0;
776 resp.cqe_comp_caps.supported_format =
777 MLX5_IB_CQE_RES_FORMAT_HASH |
778 MLX5_IB_CQE_RES_FORMAT_CSUM;
779 resp.response_length += sizeof(resp.cqe_comp_caps);
782 if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen)) {
783 if (MLX5_CAP_QOS(mdev, packet_pacing) &&
784 MLX5_CAP_GEN(mdev, qos)) {
785 resp.packet_pacing_caps.qp_rate_limit_max =
786 MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
787 resp.packet_pacing_caps.qp_rate_limit_min =
788 MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
789 resp.packet_pacing_caps.supported_qpts |=
790 1 << IB_QPT_RAW_PACKET;
792 resp.response_length += sizeof(resp.packet_pacing_caps);
795 if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
797 resp.mlx5_ib_support_multi_pkt_send_wqes =
798 MLX5_CAP_ETH(mdev, multi_pkt_send_wqe);
799 resp.response_length +=
800 sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
803 if (field_avail(typeof(resp), reserved, uhw->outlen))
804 resp.response_length += sizeof(resp.reserved);
807 err = ib_copy_to_udata(uhw, &resp, resp.response_length);
817 MLX5_IB_WIDTH_1X = 1 << 0,
818 MLX5_IB_WIDTH_2X = 1 << 1,
819 MLX5_IB_WIDTH_4X = 1 << 2,
820 MLX5_IB_WIDTH_8X = 1 << 3,
821 MLX5_IB_WIDTH_12X = 1 << 4
824 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
827 struct mlx5_ib_dev *dev = to_mdev(ibdev);
830 if (active_width & MLX5_IB_WIDTH_1X) {
831 *ib_width = IB_WIDTH_1X;
832 } else if (active_width & MLX5_IB_WIDTH_2X) {
833 mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
836 } else if (active_width & MLX5_IB_WIDTH_4X) {
837 *ib_width = IB_WIDTH_4X;
838 } else if (active_width & MLX5_IB_WIDTH_8X) {
839 *ib_width = IB_WIDTH_8X;
840 } else if (active_width & MLX5_IB_WIDTH_12X) {
841 *ib_width = IB_WIDTH_12X;
843 mlx5_ib_dbg(dev, "Invalid active_width %d\n",
851 static int mlx5_mtu_to_ib_mtu(int mtu)
860 pr_warn("invalid mtu\n");
870 __IB_MAX_VL_0_14 = 5,
873 enum mlx5_vl_hw_cap {
885 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
890 *max_vl_num = __IB_MAX_VL_0;
893 *max_vl_num = __IB_MAX_VL_0_1;
896 *max_vl_num = __IB_MAX_VL_0_3;
899 *max_vl_num = __IB_MAX_VL_0_7;
901 case MLX5_VL_HW_0_14:
902 *max_vl_num = __IB_MAX_VL_0_14;
912 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
913 struct ib_port_attr *props)
915 struct mlx5_ib_dev *dev = to_mdev(ibdev);
916 struct mlx5_core_dev *mdev = dev->mdev;
917 struct mlx5_hca_vport_context *rep;
921 u8 ib_link_width_oper;
924 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
930 /* props being zeroed by the caller, avoid zeroing it here */
932 err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
936 props->lid = rep->lid;
937 props->lmc = rep->lmc;
938 props->sm_lid = rep->sm_lid;
939 props->sm_sl = rep->sm_sl;
940 props->state = rep->vport_state;
941 props->phys_state = rep->port_physical_state;
942 props->port_cap_flags = rep->cap_mask1;
943 props->gid_tbl_len = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
944 props->max_msg_sz = 1 << MLX5_CAP_GEN(mdev, log_max_msg);
945 props->pkey_tbl_len = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
946 props->bad_pkey_cntr = rep->pkey_violation_counter;
947 props->qkey_viol_cntr = rep->qkey_violation_counter;
948 props->subnet_timeout = rep->subnet_timeout;
949 props->init_type_reply = rep->init_type_reply;
950 props->grh_required = rep->grh_required;
952 err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
956 err = translate_active_width(ibdev, ib_link_width_oper,
957 &props->active_width);
960 err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
964 mlx5_query_port_max_mtu(mdev, &max_mtu, port);
966 props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
968 mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
970 props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
972 err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
976 err = translate_max_vl_num(ibdev, vl_hw_cap,
983 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
984 struct ib_port_attr *props)
989 switch (mlx5_get_vport_access_method(ibdev)) {
990 case MLX5_VPORT_ACCESS_METHOD_MAD:
991 ret = mlx5_query_mad_ifc_port(ibdev, port, props);
994 case MLX5_VPORT_ACCESS_METHOD_HCA:
995 ret = mlx5_query_hca_port(ibdev, port, props);
998 case MLX5_VPORT_ACCESS_METHOD_NIC:
999 ret = mlx5_query_port_roce(ibdev, port, props);
1006 if (!ret && props) {
1007 count = mlx5_core_reserved_gids_count(to_mdev(ibdev)->mdev);
1008 props->gid_tbl_len -= count;
1013 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1016 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1017 struct mlx5_core_dev *mdev = dev->mdev;
1019 switch (mlx5_get_vport_access_method(ibdev)) {
1020 case MLX5_VPORT_ACCESS_METHOD_MAD:
1021 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1023 case MLX5_VPORT_ACCESS_METHOD_HCA:
1024 return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1032 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1035 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1036 struct mlx5_core_dev *mdev = dev->mdev;
1038 switch (mlx5_get_vport_access_method(ibdev)) {
1039 case MLX5_VPORT_ACCESS_METHOD_MAD:
1040 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1042 case MLX5_VPORT_ACCESS_METHOD_HCA:
1043 case MLX5_VPORT_ACCESS_METHOD_NIC:
1044 return mlx5_query_hca_vport_pkey(mdev, 0, port, 0, index,
1051 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1052 struct ib_device_modify *props)
1054 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1055 struct mlx5_reg_node_desc in;
1056 struct mlx5_reg_node_desc out;
1059 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1062 if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1066 * If possible, pass node desc to FW, so it can generate
1067 * a 144 trap. If cmd fails, just ignore.
1069 memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1070 err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1071 sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1075 memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1080 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1083 struct mlx5_hca_vport_context ctx = {};
1086 err = mlx5_query_hca_vport_context(dev->mdev, 0,
1091 if (~ctx.cap_mask1_perm & mask) {
1092 mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1093 mask, ctx.cap_mask1_perm);
1097 ctx.cap_mask1 = value;
1098 ctx.cap_mask1_perm = mask;
1099 err = mlx5_core_modify_hca_vport_context(dev->mdev, 0,
1105 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1106 struct ib_port_modify *props)
1108 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1109 struct ib_port_attr attr;
1114 bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1115 IB_LINK_LAYER_INFINIBAND);
1117 if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1118 change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1119 value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1120 return set_port_caps_atomic(dev, port, change_mask, value);
1123 mutex_lock(&dev->cap_mask_mutex);
1125 err = ib_query_port(ibdev, port, &attr);
1129 tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1130 ~props->clr_port_cap_mask;
1132 err = mlx5_set_port_caps(dev->mdev, port, tmp);
1135 mutex_unlock(&dev->cap_mask_mutex);
1139 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1141 mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1142 caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1145 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1146 struct mlx5_ib_alloc_ucontext_req_v2 *req,
1149 int uars_per_sys_page;
1150 int bfregs_per_sys_page;
1151 int ref_bfregs = req->total_num_bfregs;
1153 if (req->total_num_bfregs == 0)
1156 BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1157 BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1159 if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1162 uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1163 bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1164 req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1165 *num_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1167 if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1170 mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, alloated %d, using %d sys pages\n",
1171 MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1172 lib_uar_4k ? "yes" : "no", ref_bfregs,
1173 req->total_num_bfregs, *num_sys_pages);
1178 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1180 struct mlx5_bfreg_info *bfregi;
1184 bfregi = &context->bfregi;
1185 for (i = 0; i < bfregi->num_sys_pages; i++) {
1186 err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1190 mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1195 for (--i; i >= 0; i--)
1196 if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1197 mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1202 static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1204 struct mlx5_bfreg_info *bfregi;
1208 bfregi = &context->bfregi;
1209 for (i = 0; i < bfregi->num_sys_pages; i++) {
1210 err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1212 mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1219 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn)
1223 err = mlx5_core_alloc_transport_domain(dev->mdev, tdn);
1227 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1228 !MLX5_CAP_GEN(dev->mdev, disable_local_lb))
1231 mutex_lock(&dev->lb_mutex);
1234 if (dev->user_td == 2)
1235 err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1237 mutex_unlock(&dev->lb_mutex);
1241 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn)
1243 mlx5_core_dealloc_transport_domain(dev->mdev, tdn);
1245 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1246 !MLX5_CAP_GEN(dev->mdev, disable_local_lb))
1249 mutex_lock(&dev->lb_mutex);
1252 if (dev->user_td < 2)
1253 mlx5_nic_vport_update_local_lb(dev->mdev, false);
1255 mutex_unlock(&dev->lb_mutex);
1258 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1259 struct ib_udata *udata)
1261 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1262 struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1263 struct mlx5_ib_alloc_ucontext_resp resp = {};
1264 struct mlx5_ib_ucontext *context;
1265 struct mlx5_bfreg_info *bfregi;
1269 size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1273 if (!dev->ib_active)
1274 return ERR_PTR(-EAGAIN);
1276 if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
1277 return ERR_PTR(-EINVAL);
1279 reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
1280 if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1282 else if (reqlen >= min_req_v2)
1285 return ERR_PTR(-EINVAL);
1287 err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
1289 return ERR_PTR(err);
1292 return ERR_PTR(-EINVAL);
1294 if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1295 return ERR_PTR(-EOPNOTSUPP);
1297 req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1298 MLX5_NON_FP_BFREGS_PER_UAR);
1299 if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1300 return ERR_PTR(-EINVAL);
1302 resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1303 if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1304 resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1305 resp.cache_line_size = cache_line_size();
1306 resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1307 resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1308 resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1309 resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1310 resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1311 resp.cqe_version = min_t(__u8,
1312 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1313 req.max_cqe_version);
1314 resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1315 MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1316 resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1317 MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1318 resp.response_length = min(offsetof(typeof(resp), response_length) +
1319 sizeof(resp.response_length), udata->outlen);
1321 context = kzalloc(sizeof(*context), GFP_KERNEL);
1323 return ERR_PTR(-ENOMEM);
1325 lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1326 bfregi = &context->bfregi;
1328 /* updates req->total_num_bfregs */
1329 err = calc_total_bfregs(dev, lib_uar_4k, &req, &bfregi->num_sys_pages);
1333 mutex_init(&bfregi->lock);
1334 bfregi->lib_uar_4k = lib_uar_4k;
1335 bfregi->count = kcalloc(req.total_num_bfregs, sizeof(*bfregi->count),
1337 if (!bfregi->count) {
1342 bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1343 sizeof(*bfregi->sys_pages),
1345 if (!bfregi->sys_pages) {
1350 err = allocate_uars(dev, context);
1354 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1355 context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1358 context->upd_xlt_page = __get_free_page(GFP_KERNEL);
1359 if (!context->upd_xlt_page) {
1363 mutex_init(&context->upd_xlt_page_mutex);
1365 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1366 err = mlx5_ib_alloc_transport_domain(dev, &context->tdn);
1371 INIT_LIST_HEAD(&context->vma_private_list);
1372 INIT_LIST_HEAD(&context->db_page_list);
1373 mutex_init(&context->db_page_mutex);
1375 resp.tot_bfregs = req.total_num_bfregs;
1376 resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1378 if (field_avail(typeof(resp), cqe_version, udata->outlen))
1379 resp.response_length += sizeof(resp.cqe_version);
1381 if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1382 resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1383 MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1384 resp.response_length += sizeof(resp.cmds_supp_uhw);
1387 if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1388 if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1389 mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1390 resp.eth_min_inline++;
1392 resp.response_length += sizeof(resp.eth_min_inline);
1396 * We don't want to expose information from the PCI bar that is located
1397 * after 4096 bytes, so if the arch only supports larger pages, let's
1398 * pretend we don't support reading the HCA's core clock. This is also
1399 * forced by mmap function.
1401 if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1402 if (PAGE_SIZE <= 4096) {
1404 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1405 resp.hca_core_clock_offset =
1406 offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1408 resp.response_length += sizeof(resp.hca_core_clock_offset) +
1409 sizeof(resp.reserved2);
1412 if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1413 resp.response_length += sizeof(resp.log_uar_size);
1415 if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1416 resp.response_length += sizeof(resp.num_uars_per_page);
1418 err = ib_copy_to_udata(udata, &resp, resp.response_length);
1423 bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1424 context->cqe_version = resp.cqe_version;
1425 context->lib_caps = req.lib_caps;
1426 print_lib_caps(dev, context->lib_caps);
1428 return &context->ibucontext;
1431 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1432 mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1435 free_page(context->upd_xlt_page);
1438 deallocate_uars(dev, context);
1441 kfree(bfregi->sys_pages);
1444 kfree(bfregi->count);
1449 return ERR_PTR(err);
1452 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1454 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1455 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1456 struct mlx5_bfreg_info *bfregi;
1458 bfregi = &context->bfregi;
1459 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1460 mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1462 free_page(context->upd_xlt_page);
1463 deallocate_uars(dev, context);
1464 kfree(bfregi->sys_pages);
1465 kfree(bfregi->count);
1471 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1472 struct mlx5_bfreg_info *bfregi,
1475 int fw_uars_per_page;
1477 fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1479 return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) +
1480 bfregi->sys_pages[idx] / fw_uars_per_page;
1483 static int get_command(unsigned long offset)
1485 return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1488 static int get_arg(unsigned long offset)
1490 return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1493 static int get_index(unsigned long offset)
1495 return get_arg(offset);
1498 static void mlx5_ib_vma_open(struct vm_area_struct *area)
1500 /* vma_open is called when a new VMA is created on top of our VMA. This
1501 * is done through either mremap flow or split_vma (usually due to
1502 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1503 * as this VMA is strongly hardware related. Therefore we set the
1504 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1505 * calling us again and trying to do incorrect actions. We assume that
1506 * the original VMA size is exactly a single page, and therefore all
1507 * "splitting" operation will not happen to it.
1509 area->vm_ops = NULL;
1512 static void mlx5_ib_vma_close(struct vm_area_struct *area)
1514 struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1516 /* It's guaranteed that all VMAs opened on a FD are closed before the
1517 * file itself is closed, therefore no sync is needed with the regular
1518 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1519 * However need a sync with accessing the vma as part of
1520 * mlx5_ib_disassociate_ucontext.
1521 * The close operation is usually called under mm->mmap_sem except when
1522 * process is exiting.
1523 * The exiting case is handled explicitly as part of
1524 * mlx5_ib_disassociate_ucontext.
1526 mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1528 /* setting the vma context pointer to null in the mlx5_ib driver's
1529 * private data, to protect a race condition in
1530 * mlx5_ib_disassociate_ucontext().
1532 mlx5_ib_vma_priv_data->vma = NULL;
1533 list_del(&mlx5_ib_vma_priv_data->list);
1534 kfree(mlx5_ib_vma_priv_data);
1537 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1538 .open = mlx5_ib_vma_open,
1539 .close = mlx5_ib_vma_close
1542 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1543 struct mlx5_ib_ucontext *ctx)
1545 struct mlx5_ib_vma_private_data *vma_prv;
1546 struct list_head *vma_head = &ctx->vma_private_list;
1548 vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1553 vma->vm_private_data = vma_prv;
1554 vma->vm_ops = &mlx5_ib_vm_ops;
1556 list_add(&vma_prv->list, vma_head);
1561 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1564 struct vm_area_struct *vma;
1565 struct mlx5_ib_vma_private_data *vma_private, *n;
1566 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1567 struct task_struct *owning_process = NULL;
1568 struct mm_struct *owning_mm = NULL;
1570 owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1571 if (!owning_process)
1574 owning_mm = get_task_mm(owning_process);
1576 pr_info("no mm, disassociate ucontext is pending task termination\n");
1578 put_task_struct(owning_process);
1579 usleep_range(1000, 2000);
1580 owning_process = get_pid_task(ibcontext->tgid,
1582 if (!owning_process ||
1583 owning_process->state == TASK_DEAD) {
1584 pr_info("disassociate ucontext done, task was terminated\n");
1585 /* in case task was dead need to release the
1589 put_task_struct(owning_process);
1595 /* need to protect from a race on closing the vma as part of
1596 * mlx5_ib_vma_close.
1598 down_write(&owning_mm->mmap_sem);
1599 list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1601 vma = vma_private->vma;
1602 ret = zap_vma_ptes(vma, vma->vm_start,
1604 WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1605 /* context going to be destroyed, should
1606 * not access ops any more.
1608 vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
1610 list_del(&vma_private->list);
1613 up_write(&owning_mm->mmap_sem);
1615 put_task_struct(owning_process);
1618 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1621 case MLX5_IB_MMAP_WC_PAGE:
1623 case MLX5_IB_MMAP_REGULAR_PAGE:
1624 return "best effort WC";
1625 case MLX5_IB_MMAP_NC_PAGE:
1632 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1633 struct vm_area_struct *vma,
1634 struct mlx5_ib_ucontext *context)
1636 struct mlx5_bfreg_info *bfregi = &context->bfregi;
1639 phys_addr_t pfn, pa;
1643 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1646 uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
1647 idx = get_index(vma->vm_pgoff);
1648 if (idx % uars_per_page ||
1649 idx * uars_per_page >= bfregi->num_sys_pages) {
1650 mlx5_ib_warn(dev, "invalid uar index %lu\n", idx);
1655 case MLX5_IB_MMAP_WC_PAGE:
1656 /* Some architectures don't support WC memory */
1657 #if defined(CONFIG_X86)
1660 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1664 case MLX5_IB_MMAP_REGULAR_PAGE:
1665 /* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1666 prot = pgprot_writecombine(vma->vm_page_prot);
1668 case MLX5_IB_MMAP_NC_PAGE:
1669 prot = pgprot_noncached(vma->vm_page_prot);
1675 pfn = uar_index2pfn(dev, bfregi, idx);
1676 mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1678 vma->vm_page_prot = prot;
1679 err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1680 PAGE_SIZE, vma->vm_page_prot);
1682 mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
1683 err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
1687 pa = pfn << PAGE_SHIFT;
1688 mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
1689 vma->vm_start, &pa);
1691 return mlx5_ib_set_vma_data(vma, context);
1694 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1696 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1697 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1698 unsigned long command;
1701 command = get_command(vma->vm_pgoff);
1703 case MLX5_IB_MMAP_WC_PAGE:
1704 case MLX5_IB_MMAP_NC_PAGE:
1705 case MLX5_IB_MMAP_REGULAR_PAGE:
1706 return uar_mmap(dev, command, vma, context);
1708 case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1711 case MLX5_IB_MMAP_CORE_CLOCK:
1712 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1715 if (vma->vm_flags & VM_WRITE)
1718 /* Don't expose to user-space information it shouldn't have */
1719 if (PAGE_SIZE > 4096)
1722 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1723 pfn = (dev->mdev->iseg_base +
1724 offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1726 if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1727 PAGE_SIZE, vma->vm_page_prot))
1730 mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
1732 (unsigned long long)pfn << PAGE_SHIFT);
1742 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1743 struct ib_ucontext *context,
1744 struct ib_udata *udata)
1746 struct mlx5_ib_alloc_pd_resp resp;
1747 struct mlx5_ib_pd *pd;
1750 pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1752 return ERR_PTR(-ENOMEM);
1754 err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1757 return ERR_PTR(err);
1762 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1763 mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1765 return ERR_PTR(-EFAULT);
1772 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1774 struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1775 struct mlx5_ib_pd *mpd = to_mpd(pd);
1777 mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1784 MATCH_CRITERIA_ENABLE_OUTER_BIT,
1785 MATCH_CRITERIA_ENABLE_MISC_BIT,
1786 MATCH_CRITERIA_ENABLE_INNER_BIT
1789 #define HEADER_IS_ZERO(match_criteria, headers) \
1790 !(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1791 0, MLX5_FLD_SZ_BYTES(fte_match_param, headers))) \
1793 static u8 get_match_criteria_enable(u32 *match_criteria)
1795 u8 match_criteria_enable;
1797 match_criteria_enable =
1798 (!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1799 MATCH_CRITERIA_ENABLE_OUTER_BIT;
1800 match_criteria_enable |=
1801 (!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1802 MATCH_CRITERIA_ENABLE_MISC_BIT;
1803 match_criteria_enable |=
1804 (!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1805 MATCH_CRITERIA_ENABLE_INNER_BIT;
1807 return match_criteria_enable;
1810 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1812 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1813 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1816 static void set_flow_label(void *misc_c, void *misc_v, u8 mask, u8 val,
1820 MLX5_SET(fte_match_set_misc,
1821 misc_c, inner_ipv6_flow_label, mask);
1822 MLX5_SET(fte_match_set_misc,
1823 misc_v, inner_ipv6_flow_label, val);
1825 MLX5_SET(fte_match_set_misc,
1826 misc_c, outer_ipv6_flow_label, mask);
1827 MLX5_SET(fte_match_set_misc,
1828 misc_v, outer_ipv6_flow_label, val);
1832 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1834 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1835 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1836 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1837 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1840 #define LAST_ETH_FIELD vlan_tag
1841 #define LAST_IB_FIELD sl
1842 #define LAST_IPV4_FIELD tos
1843 #define LAST_IPV6_FIELD traffic_class
1844 #define LAST_TCP_UDP_FIELD src_port
1845 #define LAST_TUNNEL_FIELD tunnel_id
1846 #define LAST_FLOW_TAG_FIELD tag_id
1847 #define LAST_DROP_FIELD size
1849 /* Field is the last supported field */
1850 #define FIELDS_NOT_SUPPORTED(filter, field)\
1851 memchr_inv((void *)&filter.field +\
1852 sizeof(filter.field), 0,\
1854 offsetof(typeof(filter), field) -\
1855 sizeof(filter.field))
1857 #define IPV4_VERSION 4
1858 #define IPV6_VERSION 6
1859 static int parse_flow_attr(struct mlx5_core_dev *mdev, u32 *match_c,
1860 u32 *match_v, const union ib_flow_spec *ib_spec,
1861 u32 *tag_id, bool *is_drop)
1863 void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1865 void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1871 if (ib_spec->type & IB_FLOW_SPEC_INNER) {
1872 headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1874 headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1876 match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
1877 ft_field_support.inner_ip_version);
1879 headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1881 headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1883 match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
1884 ft_field_support.outer_ip_version);
1887 switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
1888 case IB_FLOW_SPEC_ETH:
1889 if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1892 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1894 ib_spec->eth.mask.dst_mac);
1895 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1897 ib_spec->eth.val.dst_mac);
1899 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1901 ib_spec->eth.mask.src_mac);
1902 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1904 ib_spec->eth.val.src_mac);
1906 if (ib_spec->eth.mask.vlan_tag) {
1907 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1909 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1912 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1913 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1914 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1915 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1917 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1919 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1920 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1922 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1924 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1926 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1927 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1929 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1931 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1932 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1933 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1934 ethertype, ntohs(ib_spec->eth.val.ether_type));
1936 case IB_FLOW_SPEC_IPV4:
1937 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1941 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1943 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1944 ip_version, IPV4_VERSION);
1946 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1948 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1949 ethertype, ETH_P_IP);
1952 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1953 src_ipv4_src_ipv6.ipv4_layout.ipv4),
1954 &ib_spec->ipv4.mask.src_ip,
1955 sizeof(ib_spec->ipv4.mask.src_ip));
1956 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1957 src_ipv4_src_ipv6.ipv4_layout.ipv4),
1958 &ib_spec->ipv4.val.src_ip,
1959 sizeof(ib_spec->ipv4.val.src_ip));
1960 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1961 dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1962 &ib_spec->ipv4.mask.dst_ip,
1963 sizeof(ib_spec->ipv4.mask.dst_ip));
1964 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1965 dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1966 &ib_spec->ipv4.val.dst_ip,
1967 sizeof(ib_spec->ipv4.val.dst_ip));
1969 set_tos(headers_c, headers_v,
1970 ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
1972 set_proto(headers_c, headers_v,
1973 ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
1975 case IB_FLOW_SPEC_IPV6:
1976 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1980 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1982 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1983 ip_version, IPV6_VERSION);
1985 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1987 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1988 ethertype, ETH_P_IPV6);
1991 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1992 src_ipv4_src_ipv6.ipv6_layout.ipv6),
1993 &ib_spec->ipv6.mask.src_ip,
1994 sizeof(ib_spec->ipv6.mask.src_ip));
1995 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1996 src_ipv4_src_ipv6.ipv6_layout.ipv6),
1997 &ib_spec->ipv6.val.src_ip,
1998 sizeof(ib_spec->ipv6.val.src_ip));
1999 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2000 dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2001 &ib_spec->ipv6.mask.dst_ip,
2002 sizeof(ib_spec->ipv6.mask.dst_ip));
2003 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2004 dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2005 &ib_spec->ipv6.val.dst_ip,
2006 sizeof(ib_spec->ipv6.val.dst_ip));
2008 set_tos(headers_c, headers_v,
2009 ib_spec->ipv6.mask.traffic_class,
2010 ib_spec->ipv6.val.traffic_class);
2012 set_proto(headers_c, headers_v,
2013 ib_spec->ipv6.mask.next_hdr,
2014 ib_spec->ipv6.val.next_hdr);
2016 set_flow_label(misc_params_c, misc_params_v,
2017 ntohl(ib_spec->ipv6.mask.flow_label),
2018 ntohl(ib_spec->ipv6.val.flow_label),
2019 ib_spec->type & IB_FLOW_SPEC_INNER);
2022 case IB_FLOW_SPEC_TCP:
2023 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2024 LAST_TCP_UDP_FIELD))
2027 MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2029 MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2032 MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
2033 ntohs(ib_spec->tcp_udp.mask.src_port));
2034 MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
2035 ntohs(ib_spec->tcp_udp.val.src_port));
2037 MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
2038 ntohs(ib_spec->tcp_udp.mask.dst_port));
2039 MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
2040 ntohs(ib_spec->tcp_udp.val.dst_port));
2042 case IB_FLOW_SPEC_UDP:
2043 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2044 LAST_TCP_UDP_FIELD))
2047 MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2049 MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2052 MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
2053 ntohs(ib_spec->tcp_udp.mask.src_port));
2054 MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
2055 ntohs(ib_spec->tcp_udp.val.src_port));
2057 MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
2058 ntohs(ib_spec->tcp_udp.mask.dst_port));
2059 MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
2060 ntohs(ib_spec->tcp_udp.val.dst_port));
2062 case IB_FLOW_SPEC_VXLAN_TUNNEL:
2063 if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
2067 MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
2068 ntohl(ib_spec->tunnel.mask.tunnel_id));
2069 MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
2070 ntohl(ib_spec->tunnel.val.tunnel_id));
2072 case IB_FLOW_SPEC_ACTION_TAG:
2073 if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
2074 LAST_FLOW_TAG_FIELD))
2076 if (ib_spec->flow_tag.tag_id >= BIT(24))
2079 *tag_id = ib_spec->flow_tag.tag_id;
2081 case IB_FLOW_SPEC_ACTION_DROP:
2082 if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
2094 /* If a flow could catch both multicast and unicast packets,
2095 * it won't fall into the multicast flow steering table and this rule
2096 * could steal other multicast packets.
2098 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
2100 struct ib_flow_spec_eth *eth_spec;
2102 if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
2103 ib_attr->size < sizeof(struct ib_flow_attr) +
2104 sizeof(struct ib_flow_spec_eth) ||
2105 ib_attr->num_of_specs < 1)
2108 eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
2109 if (eth_spec->type != IB_FLOW_SPEC_ETH ||
2110 eth_spec->size != sizeof(*eth_spec))
2113 return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
2114 is_multicast_ether_addr(eth_spec->val.dst_mac);
2117 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
2118 const struct ib_flow_attr *flow_attr,
2121 union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
2122 int match_ipv = check_inner ?
2123 MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2124 ft_field_support.inner_ip_version) :
2125 MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2126 ft_field_support.outer_ip_version);
2127 int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
2128 bool ipv4_spec_valid, ipv6_spec_valid;
2129 unsigned int ip_spec_type = 0;
2130 bool has_ethertype = false;
2131 unsigned int spec_index;
2132 bool mask_valid = true;
2136 /* Validate that ethertype is correct */
2137 for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2138 if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
2139 ib_spec->eth.mask.ether_type) {
2140 mask_valid = (ib_spec->eth.mask.ether_type ==
2142 has_ethertype = true;
2143 eth_type = ntohs(ib_spec->eth.val.ether_type);
2144 } else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
2145 (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
2146 ip_spec_type = ib_spec->type;
2148 ib_spec = (void *)ib_spec + ib_spec->size;
2151 type_valid = (!has_ethertype) || (!ip_spec_type);
2152 if (!type_valid && mask_valid) {
2153 ipv4_spec_valid = (eth_type == ETH_P_IP) &&
2154 (ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
2155 ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
2156 (ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
2158 type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
2159 (((eth_type == ETH_P_MPLS_UC) ||
2160 (eth_type == ETH_P_MPLS_MC)) && match_ipv);
2166 static bool is_valid_attr(struct mlx5_core_dev *mdev,
2167 const struct ib_flow_attr *flow_attr)
2169 return is_valid_ethertype(mdev, flow_attr, false) &&
2170 is_valid_ethertype(mdev, flow_attr, true);
2173 static void put_flow_table(struct mlx5_ib_dev *dev,
2174 struct mlx5_ib_flow_prio *prio, bool ft_added)
2176 prio->refcount -= !!ft_added;
2177 if (!prio->refcount) {
2178 mlx5_destroy_flow_table(prio->flow_table);
2179 prio->flow_table = NULL;
2183 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
2185 struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
2186 struct mlx5_ib_flow_handler *handler = container_of(flow_id,
2187 struct mlx5_ib_flow_handler,
2189 struct mlx5_ib_flow_handler *iter, *tmp;
2191 mutex_lock(&dev->flow_db.lock);
2193 list_for_each_entry_safe(iter, tmp, &handler->list, list) {
2194 mlx5_del_flow_rules(iter->rule);
2195 put_flow_table(dev, iter->prio, true);
2196 list_del(&iter->list);
2200 mlx5_del_flow_rules(handler->rule);
2201 put_flow_table(dev, handler->prio, true);
2202 mutex_unlock(&dev->flow_db.lock);
2209 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
2217 enum flow_table_type {
2222 #define MLX5_FS_MAX_TYPES 6
2223 #define MLX5_FS_MAX_ENTRIES BIT(16)
2224 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
2225 struct ib_flow_attr *flow_attr,
2226 enum flow_table_type ft_type)
2228 bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
2229 struct mlx5_flow_namespace *ns = NULL;
2230 struct mlx5_ib_flow_prio *prio;
2231 struct mlx5_flow_table *ft;
2238 max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
2240 if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2241 if (flow_is_multicast_only(flow_attr) &&
2243 priority = MLX5_IB_FLOW_MCAST_PRIO;
2245 priority = ib_prio_to_core_prio(flow_attr->priority,
2247 ns = mlx5_get_flow_namespace(dev->mdev,
2248 MLX5_FLOW_NAMESPACE_BYPASS);
2249 num_entries = MLX5_FS_MAX_ENTRIES;
2250 num_groups = MLX5_FS_MAX_TYPES;
2251 prio = &dev->flow_db.prios[priority];
2252 } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2253 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2254 ns = mlx5_get_flow_namespace(dev->mdev,
2255 MLX5_FLOW_NAMESPACE_LEFTOVERS);
2256 build_leftovers_ft_param(&priority,
2259 prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
2260 } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2261 if (!MLX5_CAP_FLOWTABLE(dev->mdev,
2262 allow_sniffer_and_nic_rx_shared_tir))
2263 return ERR_PTR(-ENOTSUPP);
2265 ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
2266 MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2267 MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2269 prio = &dev->flow_db.sniffer[ft_type];
2276 return ERR_PTR(-ENOTSUPP);
2278 if (num_entries > max_table_size)
2279 return ERR_PTR(-ENOMEM);
2281 ft = prio->flow_table;
2283 ft = mlx5_create_auto_grouped_flow_table(ns, priority,
2290 prio->flow_table = ft;
2296 return err ? ERR_PTR(err) : prio;
2299 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
2300 struct mlx5_ib_flow_prio *ft_prio,
2301 const struct ib_flow_attr *flow_attr,
2302 struct mlx5_flow_destination *dst)
2304 struct mlx5_flow_table *ft = ft_prio->flow_table;
2305 struct mlx5_ib_flow_handler *handler;
2306 struct mlx5_flow_act flow_act = {0};
2307 struct mlx5_flow_spec *spec;
2308 struct mlx5_flow_destination *rule_dst = dst;
2309 const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
2310 unsigned int spec_index;
2311 u32 flow_tag = MLX5_FS_DEFAULT_FLOW_TAG;
2312 bool is_drop = false;
2316 if (!is_valid_attr(dev->mdev, flow_attr))
2317 return ERR_PTR(-EINVAL);
2319 spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
2320 handler = kzalloc(sizeof(*handler), GFP_KERNEL);
2321 if (!handler || !spec) {
2326 INIT_LIST_HEAD(&handler->list);
2328 for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2329 err = parse_flow_attr(dev->mdev, spec->match_criteria,
2331 ib_flow, &flow_tag, &is_drop);
2335 ib_flow += ((union ib_flow_spec *)ib_flow)->size;
2338 spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
2340 flow_act.action = MLX5_FLOW_CONTEXT_ACTION_DROP;
2344 flow_act.action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
2345 MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
2348 if (flow_tag != MLX5_FS_DEFAULT_FLOW_TAG &&
2349 (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2350 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
2351 mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
2352 flow_tag, flow_attr->type);
2356 flow_act.flow_tag = flow_tag;
2357 handler->rule = mlx5_add_flow_rules(ft, spec,
2359 rule_dst, dest_num);
2361 if (IS_ERR(handler->rule)) {
2362 err = PTR_ERR(handler->rule);
2366 ft_prio->refcount++;
2367 handler->prio = ft_prio;
2369 ft_prio->flow_table = ft;
2374 return err ? ERR_PTR(err) : handler;
2377 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
2378 struct mlx5_ib_flow_prio *ft_prio,
2379 struct ib_flow_attr *flow_attr,
2380 struct mlx5_flow_destination *dst)
2382 struct mlx5_ib_flow_handler *handler_dst = NULL;
2383 struct mlx5_ib_flow_handler *handler = NULL;
2385 handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
2386 if (!IS_ERR(handler)) {
2387 handler_dst = create_flow_rule(dev, ft_prio,
2389 if (IS_ERR(handler_dst)) {
2390 mlx5_del_flow_rules(handler->rule);
2391 ft_prio->refcount--;
2393 handler = handler_dst;
2395 list_add(&handler_dst->list, &handler->list);
2406 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
2407 struct mlx5_ib_flow_prio *ft_prio,
2408 struct ib_flow_attr *flow_attr,
2409 struct mlx5_flow_destination *dst)
2411 struct mlx5_ib_flow_handler *handler_ucast = NULL;
2412 struct mlx5_ib_flow_handler *handler = NULL;
2415 struct ib_flow_attr flow_attr;
2416 struct ib_flow_spec_eth eth_flow;
2417 } leftovers_specs[] = {
2421 .size = sizeof(leftovers_specs[0])
2424 .type = IB_FLOW_SPEC_ETH,
2425 .size = sizeof(struct ib_flow_spec_eth),
2426 .mask = {.dst_mac = {0x1} },
2427 .val = {.dst_mac = {0x1} }
2433 .size = sizeof(leftovers_specs[0])
2436 .type = IB_FLOW_SPEC_ETH,
2437 .size = sizeof(struct ib_flow_spec_eth),
2438 .mask = {.dst_mac = {0x1} },
2439 .val = {.dst_mac = {} }
2444 handler = create_flow_rule(dev, ft_prio,
2445 &leftovers_specs[LEFTOVERS_MC].flow_attr,
2447 if (!IS_ERR(handler) &&
2448 flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2449 handler_ucast = create_flow_rule(dev, ft_prio,
2450 &leftovers_specs[LEFTOVERS_UC].flow_attr,
2452 if (IS_ERR(handler_ucast)) {
2453 mlx5_del_flow_rules(handler->rule);
2454 ft_prio->refcount--;
2456 handler = handler_ucast;
2458 list_add(&handler_ucast->list, &handler->list);
2465 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2466 struct mlx5_ib_flow_prio *ft_rx,
2467 struct mlx5_ib_flow_prio *ft_tx,
2468 struct mlx5_flow_destination *dst)
2470 struct mlx5_ib_flow_handler *handler_rx;
2471 struct mlx5_ib_flow_handler *handler_tx;
2473 static const struct ib_flow_attr flow_attr = {
2475 .size = sizeof(flow_attr)
2478 handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2479 if (IS_ERR(handler_rx)) {
2480 err = PTR_ERR(handler_rx);
2484 handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2485 if (IS_ERR(handler_tx)) {
2486 err = PTR_ERR(handler_tx);
2490 list_add(&handler_tx->list, &handler_rx->list);
2495 mlx5_del_flow_rules(handler_rx->rule);
2499 return ERR_PTR(err);
2502 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2503 struct ib_flow_attr *flow_attr,
2506 struct mlx5_ib_dev *dev = to_mdev(qp->device);
2507 struct mlx5_ib_qp *mqp = to_mqp(qp);
2508 struct mlx5_ib_flow_handler *handler = NULL;
2509 struct mlx5_flow_destination *dst = NULL;
2510 struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2511 struct mlx5_ib_flow_prio *ft_prio;
2514 if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2515 return ERR_PTR(-ENOMEM);
2517 if (domain != IB_FLOW_DOMAIN_USER ||
2518 flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2519 (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2520 return ERR_PTR(-EINVAL);
2522 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2524 return ERR_PTR(-ENOMEM);
2526 mutex_lock(&dev->flow_db.lock);
2528 ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2529 if (IS_ERR(ft_prio)) {
2530 err = PTR_ERR(ft_prio);
2533 if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2534 ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2535 if (IS_ERR(ft_prio_tx)) {
2536 err = PTR_ERR(ft_prio_tx);
2542 dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2543 if (mqp->flags & MLX5_IB_QP_RSS)
2544 dst->tir_num = mqp->rss_qp.tirn;
2546 dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2548 if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2549 if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) {
2550 handler = create_dont_trap_rule(dev, ft_prio,
2553 handler = create_flow_rule(dev, ft_prio, flow_attr,
2556 } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2557 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2558 handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2560 } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2561 handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2567 if (IS_ERR(handler)) {
2568 err = PTR_ERR(handler);
2573 mutex_unlock(&dev->flow_db.lock);
2576 return &handler->ibflow;
2579 put_flow_table(dev, ft_prio, false);
2581 put_flow_table(dev, ft_prio_tx, false);
2583 mutex_unlock(&dev->flow_db.lock);
2586 return ERR_PTR(err);
2589 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2591 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2594 err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2596 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2597 ibqp->qp_num, gid->raw);
2602 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2604 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2607 err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2609 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2610 ibqp->qp_num, gid->raw);
2615 static int init_node_data(struct mlx5_ib_dev *dev)
2619 err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2623 dev->mdev->rev_id = dev->mdev->pdev->revision;
2625 return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2628 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2631 struct mlx5_ib_dev *dev =
2632 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2634 return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
2637 static ssize_t show_reg_pages(struct device *device,
2638 struct device_attribute *attr, char *buf)
2640 struct mlx5_ib_dev *dev =
2641 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2643 return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2646 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2649 struct mlx5_ib_dev *dev =
2650 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2651 return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2654 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2657 struct mlx5_ib_dev *dev =
2658 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2659 return sprintf(buf, "%x\n", dev->mdev->rev_id);
2662 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2665 struct mlx5_ib_dev *dev =
2666 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2667 return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2668 dev->mdev->board_id);
2671 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
2672 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
2673 static DEVICE_ATTR(board_id, S_IRUGO, show_board, NULL);
2674 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2675 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2677 static struct device_attribute *mlx5_class_attributes[] = {
2682 &dev_attr_reg_pages,
2685 static void pkey_change_handler(struct work_struct *work)
2687 struct mlx5_ib_port_resources *ports =
2688 container_of(work, struct mlx5_ib_port_resources,
2691 mutex_lock(&ports->devr->mutex);
2692 mlx5_ib_gsi_pkey_change(ports->gsi);
2693 mutex_unlock(&ports->devr->mutex);
2696 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2698 struct mlx5_ib_qp *mqp;
2699 struct mlx5_ib_cq *send_mcq, *recv_mcq;
2700 struct mlx5_core_cq *mcq;
2701 struct list_head cq_armed_list;
2702 unsigned long flags_qp;
2703 unsigned long flags_cq;
2704 unsigned long flags;
2706 INIT_LIST_HEAD(&cq_armed_list);
2708 /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2709 spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2710 list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2711 spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2712 if (mqp->sq.tail != mqp->sq.head) {
2713 send_mcq = to_mcq(mqp->ibqp.send_cq);
2714 spin_lock_irqsave(&send_mcq->lock, flags_cq);
2715 if (send_mcq->mcq.comp &&
2716 mqp->ibqp.send_cq->comp_handler) {
2717 if (!send_mcq->mcq.reset_notify_added) {
2718 send_mcq->mcq.reset_notify_added = 1;
2719 list_add_tail(&send_mcq->mcq.reset_notify,
2723 spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2725 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2726 spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2727 /* no handling is needed for SRQ */
2728 if (!mqp->ibqp.srq) {
2729 if (mqp->rq.tail != mqp->rq.head) {
2730 recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2731 spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2732 if (recv_mcq->mcq.comp &&
2733 mqp->ibqp.recv_cq->comp_handler) {
2734 if (!recv_mcq->mcq.reset_notify_added) {
2735 recv_mcq->mcq.reset_notify_added = 1;
2736 list_add_tail(&recv_mcq->mcq.reset_notify,
2740 spin_unlock_irqrestore(&recv_mcq->lock,
2744 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2746 /*At that point all inflight post send were put to be executed as of we
2747 * lock/unlock above locks Now need to arm all involved CQs.
2749 list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2752 spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2755 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2756 enum mlx5_dev_event event, unsigned long param)
2758 struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2759 struct ib_event ibev;
2764 case MLX5_DEV_EVENT_SYS_ERROR:
2765 ibev.event = IB_EVENT_DEVICE_FATAL;
2766 mlx5_ib_handle_internal_error(ibdev);
2770 case MLX5_DEV_EVENT_PORT_UP:
2771 case MLX5_DEV_EVENT_PORT_DOWN:
2772 case MLX5_DEV_EVENT_PORT_INITIALIZED:
2775 /* In RoCE, port up/down events are handled in
2776 * mlx5_netdev_event().
2778 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2779 IB_LINK_LAYER_ETHERNET)
2782 ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2783 IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2786 case MLX5_DEV_EVENT_LID_CHANGE:
2787 ibev.event = IB_EVENT_LID_CHANGE;
2791 case MLX5_DEV_EVENT_PKEY_CHANGE:
2792 ibev.event = IB_EVENT_PKEY_CHANGE;
2795 schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2798 case MLX5_DEV_EVENT_GUID_CHANGE:
2799 ibev.event = IB_EVENT_GID_CHANGE;
2803 case MLX5_DEV_EVENT_CLIENT_REREG:
2804 ibev.event = IB_EVENT_CLIENT_REREGISTER;
2811 ibev.device = &ibdev->ib_dev;
2812 ibev.element.port_num = port;
2814 if (port < 1 || port > ibdev->num_ports) {
2815 mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2819 if (ibdev->ib_active)
2820 ib_dispatch_event(&ibev);
2823 ibdev->ib_active = false;
2826 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2828 struct mlx5_hca_vport_context vport_ctx;
2832 for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2833 dev->mdev->port_caps[port - 1].has_smi = false;
2834 if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2835 MLX5_CAP_PORT_TYPE_IB) {
2836 if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2837 err = mlx5_query_hca_vport_context(dev->mdev, 0,
2841 mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2845 dev->mdev->port_caps[port - 1].has_smi =
2848 dev->mdev->port_caps[port - 1].has_smi = true;
2855 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2859 for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2860 mlx5_query_ext_port_caps(dev, port);
2863 static int get_port_caps(struct mlx5_ib_dev *dev)
2865 struct ib_device_attr *dprops = NULL;
2866 struct ib_port_attr *pprops = NULL;
2869 struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2871 pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2875 dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2879 err = set_has_smi_cap(dev);
2883 err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2885 mlx5_ib_warn(dev, "query_device failed %d\n", err);
2889 for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2890 memset(pprops, 0, sizeof(*pprops));
2891 err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2893 mlx5_ib_warn(dev, "query_port %d failed %d\n",
2897 dev->mdev->port_caps[port - 1].pkey_table_len =
2899 dev->mdev->port_caps[port - 1].gid_table_len =
2900 pprops->gid_tbl_len;
2901 mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2902 dprops->max_pkeys, pprops->gid_tbl_len);
2912 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2916 err = mlx5_mr_cache_cleanup(dev);
2918 mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2920 mlx5_ib_destroy_qp(dev->umrc.qp);
2921 ib_free_cq(dev->umrc.cq);
2922 ib_dealloc_pd(dev->umrc.pd);
2929 static int create_umr_res(struct mlx5_ib_dev *dev)
2931 struct ib_qp_init_attr *init_attr = NULL;
2932 struct ib_qp_attr *attr = NULL;
2938 attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2939 init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2940 if (!attr || !init_attr) {
2945 pd = ib_alloc_pd(&dev->ib_dev, 0);
2947 mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2952 cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2954 mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2959 init_attr->send_cq = cq;
2960 init_attr->recv_cq = cq;
2961 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2962 init_attr->cap.max_send_wr = MAX_UMR_WR;
2963 init_attr->cap.max_send_sge = 1;
2964 init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2965 init_attr->port_num = 1;
2966 qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2968 mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2972 qp->device = &dev->ib_dev;
2975 qp->qp_type = MLX5_IB_QPT_REG_UMR;
2977 attr->qp_state = IB_QPS_INIT;
2979 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2982 mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2986 memset(attr, 0, sizeof(*attr));
2987 attr->qp_state = IB_QPS_RTR;
2988 attr->path_mtu = IB_MTU_256;
2990 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2992 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2996 memset(attr, 0, sizeof(*attr));
2997 attr->qp_state = IB_QPS_RTS;
2998 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
3000 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
3008 sema_init(&dev->umrc.sem, MAX_UMR_WR);
3009 ret = mlx5_mr_cache_init(dev);
3011 mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
3021 mlx5_ib_destroy_qp(qp);
3035 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
3037 switch (umr_fence_cap) {
3038 case MLX5_CAP_UMR_FENCE_NONE:
3039 return MLX5_FENCE_MODE_NONE;
3040 case MLX5_CAP_UMR_FENCE_SMALL:
3041 return MLX5_FENCE_MODE_INITIATOR_SMALL;
3043 return MLX5_FENCE_MODE_STRONG_ORDERING;
3047 static int create_dev_resources(struct mlx5_ib_resources *devr)
3049 struct ib_srq_init_attr attr;
3050 struct mlx5_ib_dev *dev;
3051 struct ib_cq_init_attr cq_attr = {.cqe = 1};
3055 dev = container_of(devr, struct mlx5_ib_dev, devr);
3057 mutex_init(&devr->mutex);
3059 devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
3060 if (IS_ERR(devr->p0)) {
3061 ret = PTR_ERR(devr->p0);
3064 devr->p0->device = &dev->ib_dev;
3065 devr->p0->uobject = NULL;
3066 atomic_set(&devr->p0->usecnt, 0);
3068 devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
3069 if (IS_ERR(devr->c0)) {
3070 ret = PTR_ERR(devr->c0);
3073 devr->c0->device = &dev->ib_dev;
3074 devr->c0->uobject = NULL;
3075 devr->c0->comp_handler = NULL;
3076 devr->c0->event_handler = NULL;
3077 devr->c0->cq_context = NULL;
3078 atomic_set(&devr->c0->usecnt, 0);
3080 devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
3081 if (IS_ERR(devr->x0)) {
3082 ret = PTR_ERR(devr->x0);
3085 devr->x0->device = &dev->ib_dev;
3086 devr->x0->inode = NULL;
3087 atomic_set(&devr->x0->usecnt, 0);
3088 mutex_init(&devr->x0->tgt_qp_mutex);
3089 INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
3091 devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
3092 if (IS_ERR(devr->x1)) {
3093 ret = PTR_ERR(devr->x1);
3096 devr->x1->device = &dev->ib_dev;
3097 devr->x1->inode = NULL;
3098 atomic_set(&devr->x1->usecnt, 0);
3099 mutex_init(&devr->x1->tgt_qp_mutex);
3100 INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
3102 memset(&attr, 0, sizeof(attr));
3103 attr.attr.max_sge = 1;
3104 attr.attr.max_wr = 1;
3105 attr.srq_type = IB_SRQT_XRC;
3106 attr.ext.xrc.cq = devr->c0;
3107 attr.ext.xrc.xrcd = devr->x0;
3109 devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
3110 if (IS_ERR(devr->s0)) {
3111 ret = PTR_ERR(devr->s0);
3114 devr->s0->device = &dev->ib_dev;
3115 devr->s0->pd = devr->p0;
3116 devr->s0->uobject = NULL;
3117 devr->s0->event_handler = NULL;
3118 devr->s0->srq_context = NULL;
3119 devr->s0->srq_type = IB_SRQT_XRC;
3120 devr->s0->ext.xrc.xrcd = devr->x0;
3121 devr->s0->ext.xrc.cq = devr->c0;
3122 atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
3123 atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
3124 atomic_inc(&devr->p0->usecnt);
3125 atomic_set(&devr->s0->usecnt, 0);
3127 memset(&attr, 0, sizeof(attr));
3128 attr.attr.max_sge = 1;
3129 attr.attr.max_wr = 1;
3130 attr.srq_type = IB_SRQT_BASIC;
3131 devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
3132 if (IS_ERR(devr->s1)) {
3133 ret = PTR_ERR(devr->s1);
3136 devr->s1->device = &dev->ib_dev;
3137 devr->s1->pd = devr->p0;
3138 devr->s1->uobject = NULL;
3139 devr->s1->event_handler = NULL;
3140 devr->s1->srq_context = NULL;
3141 devr->s1->srq_type = IB_SRQT_BASIC;
3142 devr->s1->ext.xrc.cq = devr->c0;
3143 atomic_inc(&devr->p0->usecnt);
3144 atomic_set(&devr->s0->usecnt, 0);
3146 for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
3147 INIT_WORK(&devr->ports[port].pkey_change_work,
3148 pkey_change_handler);
3149 devr->ports[port].devr = devr;
3155 mlx5_ib_destroy_srq(devr->s0);
3157 mlx5_ib_dealloc_xrcd(devr->x1);
3159 mlx5_ib_dealloc_xrcd(devr->x0);
3161 mlx5_ib_destroy_cq(devr->c0);
3163 mlx5_ib_dealloc_pd(devr->p0);
3168 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
3170 struct mlx5_ib_dev *dev =
3171 container_of(devr, struct mlx5_ib_dev, devr);
3174 mlx5_ib_destroy_srq(devr->s1);
3175 mlx5_ib_destroy_srq(devr->s0);
3176 mlx5_ib_dealloc_xrcd(devr->x0);
3177 mlx5_ib_dealloc_xrcd(devr->x1);
3178 mlx5_ib_destroy_cq(devr->c0);
3179 mlx5_ib_dealloc_pd(devr->p0);
3181 /* Make sure no change P_Key work items are still executing */
3182 for (port = 0; port < dev->num_ports; ++port)
3183 cancel_work_sync(&devr->ports[port].pkey_change_work);
3186 static u32 get_core_cap_flags(struct ib_device *ibdev)
3188 struct mlx5_ib_dev *dev = to_mdev(ibdev);
3189 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
3190 u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
3191 u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
3194 if (ll == IB_LINK_LAYER_INFINIBAND)
3195 return RDMA_CORE_PORT_IBA_IB;
3197 ret = RDMA_CORE_PORT_RAW_PACKET;
3199 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
3202 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
3205 if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
3206 ret |= RDMA_CORE_PORT_IBA_ROCE;
3208 if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
3209 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3214 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
3215 struct ib_port_immutable *immutable)
3217 struct ib_port_attr attr;
3218 struct mlx5_ib_dev *dev = to_mdev(ibdev);
3219 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
3222 immutable->core_cap_flags = get_core_cap_flags(ibdev);
3224 err = ib_query_port(ibdev, port_num, &attr);
3228 immutable->pkey_tbl_len = attr.pkey_tbl_len;
3229 immutable->gid_tbl_len = attr.gid_tbl_len;
3230 immutable->core_cap_flags = get_core_cap_flags(ibdev);
3231 if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
3232 immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3237 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
3240 struct mlx5_ib_dev *dev =
3241 container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3242 snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
3243 fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
3246 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3248 struct mlx5_core_dev *mdev = dev->mdev;
3249 struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3250 MLX5_FLOW_NAMESPACE_LAG);
3251 struct mlx5_flow_table *ft;
3254 if (!ns || !mlx5_lag_is_active(mdev))
3257 err = mlx5_cmd_create_vport_lag(mdev);
3261 ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3264 goto err_destroy_vport_lag;
3267 dev->flow_db.lag_demux_ft = ft;
3270 err_destroy_vport_lag:
3271 mlx5_cmd_destroy_vport_lag(mdev);
3275 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3277 struct mlx5_core_dev *mdev = dev->mdev;
3279 if (dev->flow_db.lag_demux_ft) {
3280 mlx5_destroy_flow_table(dev->flow_db.lag_demux_ft);
3281 dev->flow_db.lag_demux_ft = NULL;
3283 mlx5_cmd_destroy_vport_lag(mdev);
3287 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev)
3291 dev->roce.nb.notifier_call = mlx5_netdev_event;
3292 err = register_netdevice_notifier(&dev->roce.nb);
3294 dev->roce.nb.notifier_call = NULL;
3301 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev)
3303 if (dev->roce.nb.notifier_call) {
3304 unregister_netdevice_notifier(&dev->roce.nb);
3305 dev->roce.nb.notifier_call = NULL;
3309 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3313 err = mlx5_add_netdev_notifier(dev);
3317 if (MLX5_CAP_GEN(dev->mdev, roce)) {
3318 err = mlx5_nic_vport_enable_roce(dev->mdev);
3320 goto err_unregister_netdevice_notifier;
3323 err = mlx5_eth_lag_init(dev);
3325 goto err_disable_roce;
3330 if (MLX5_CAP_GEN(dev->mdev, roce))
3331 mlx5_nic_vport_disable_roce(dev->mdev);
3333 err_unregister_netdevice_notifier:
3334 mlx5_remove_netdev_notifier(dev);
3338 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3340 mlx5_eth_lag_cleanup(dev);
3341 if (MLX5_CAP_GEN(dev->mdev, roce))
3342 mlx5_nic_vport_disable_roce(dev->mdev);
3345 struct mlx5_ib_counter {
3350 #define INIT_Q_COUNTER(_name) \
3351 { .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
3353 static const struct mlx5_ib_counter basic_q_cnts[] = {
3354 INIT_Q_COUNTER(rx_write_requests),
3355 INIT_Q_COUNTER(rx_read_requests),
3356 INIT_Q_COUNTER(rx_atomic_requests),
3357 INIT_Q_COUNTER(out_of_buffer),
3360 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
3361 INIT_Q_COUNTER(out_of_sequence),
3364 static const struct mlx5_ib_counter retrans_q_cnts[] = {
3365 INIT_Q_COUNTER(duplicate_request),
3366 INIT_Q_COUNTER(rnr_nak_retry_err),
3367 INIT_Q_COUNTER(packet_seq_err),
3368 INIT_Q_COUNTER(implied_nak_seq_err),
3369 INIT_Q_COUNTER(local_ack_timeout_err),
3372 #define INIT_CONG_COUNTER(_name) \
3373 { .name = #_name, .offset = \
3374 MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
3376 static const struct mlx5_ib_counter cong_cnts[] = {
3377 INIT_CONG_COUNTER(rp_cnp_ignored),
3378 INIT_CONG_COUNTER(rp_cnp_handled),
3379 INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
3380 INIT_CONG_COUNTER(np_cnp_sent),
3383 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
3387 for (i = 0; i < dev->num_ports; i++) {
3388 mlx5_core_dealloc_q_counter(dev->mdev,
3389 dev->port[i].cnts.set_id);
3390 kfree(dev->port[i].cnts.names);
3391 kfree(dev->port[i].cnts.offsets);
3395 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
3396 struct mlx5_ib_counters *cnts)
3400 num_counters = ARRAY_SIZE(basic_q_cnts);
3402 if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
3403 num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
3405 if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
3406 num_counters += ARRAY_SIZE(retrans_q_cnts);
3407 cnts->num_q_counters = num_counters;
3409 if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
3410 cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
3411 num_counters += ARRAY_SIZE(cong_cnts);
3414 cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
3418 cnts->offsets = kcalloc(num_counters,
3419 sizeof(cnts->offsets), GFP_KERNEL);
3430 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
3437 for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
3438 names[j] = basic_q_cnts[i].name;
3439 offsets[j] = basic_q_cnts[i].offset;
3442 if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
3443 for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
3444 names[j] = out_of_seq_q_cnts[i].name;
3445 offsets[j] = out_of_seq_q_cnts[i].offset;
3449 if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3450 for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
3451 names[j] = retrans_q_cnts[i].name;
3452 offsets[j] = retrans_q_cnts[i].offset;
3456 if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
3457 for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
3458 names[j] = cong_cnts[i].name;
3459 offsets[j] = cong_cnts[i].offset;
3464 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
3469 for (i = 0; i < dev->num_ports; i++) {
3470 struct mlx5_ib_port *port = &dev->port[i];
3472 ret = mlx5_core_alloc_q_counter(dev->mdev,
3473 &port->cnts.set_id);
3476 "couldn't allocate queue counter for port %d, err %d\n",
3478 goto dealloc_counters;
3481 ret = __mlx5_ib_alloc_counters(dev, &port->cnts);
3483 goto dealloc_counters;
3485 mlx5_ib_fill_counters(dev, port->cnts.names,
3486 port->cnts.offsets);
3493 mlx5_core_dealloc_q_counter(dev->mdev,
3494 dev->port[i].cnts.set_id);
3499 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
3502 struct mlx5_ib_dev *dev = to_mdev(ibdev);
3503 struct mlx5_ib_port *port = &dev->port[port_num - 1];
3505 /* We support only per port stats */
3509 return rdma_alloc_hw_stats_struct(port->cnts.names,
3510 port->cnts.num_q_counters +
3511 port->cnts.num_cong_counters,
3512 RDMA_HW_STATS_DEFAULT_LIFESPAN);
3515 static int mlx5_ib_query_q_counters(struct mlx5_ib_dev *dev,
3516 struct mlx5_ib_port *port,
3517 struct rdma_hw_stats *stats)
3519 int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
3524 out = kvzalloc(outlen, GFP_KERNEL);
3528 ret = mlx5_core_query_q_counter(dev->mdev,
3529 port->cnts.set_id, 0,
3534 for (i = 0; i < port->cnts.num_q_counters; i++) {
3535 val = *(__be32 *)(out + port->cnts.offsets[i]);
3536 stats->value[i] = (u64)be32_to_cpu(val);
3544 static int mlx5_ib_query_cong_counters(struct mlx5_ib_dev *dev,
3545 struct mlx5_ib_port *port,
3546 struct rdma_hw_stats *stats)
3548 int outlen = MLX5_ST_SZ_BYTES(query_cong_statistics_out);
3551 int offset = port->cnts.num_q_counters;
3553 out = kvzalloc(outlen, GFP_KERNEL);
3557 ret = mlx5_cmd_query_cong_counter(dev->mdev, false, out, outlen);
3561 for (i = 0; i < port->cnts.num_cong_counters; i++) {
3562 stats->value[i + offset] =
3563 be64_to_cpup((__be64 *)(out +
3564 port->cnts.offsets[i + offset]));
3572 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
3573 struct rdma_hw_stats *stats,
3574 u8 port_num, int index)
3576 struct mlx5_ib_dev *dev = to_mdev(ibdev);
3577 struct mlx5_ib_port *port = &dev->port[port_num - 1];
3578 int ret, num_counters;
3583 ret = mlx5_ib_query_q_counters(dev, port, stats);
3586 num_counters = port->cnts.num_q_counters;
3588 if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
3589 ret = mlx5_ib_query_cong_counters(dev, port, stats);
3592 num_counters += port->cnts.num_cong_counters;
3595 return num_counters;
3598 static void mlx5_ib_free_rdma_netdev(struct net_device *netdev)
3600 return mlx5_rdma_netdev_free(netdev);
3603 static struct net_device*
3604 mlx5_ib_alloc_rdma_netdev(struct ib_device *hca,
3606 enum rdma_netdev_t type,
3608 unsigned char name_assign_type,
3609 void (*setup)(struct net_device *))
3611 struct net_device *netdev;
3612 struct rdma_netdev *rn;
3614 if (type != RDMA_NETDEV_IPOIB)
3615 return ERR_PTR(-EOPNOTSUPP);
3617 netdev = mlx5_rdma_netdev_alloc(to_mdev(hca)->mdev, hca,
3619 if (likely(!IS_ERR_OR_NULL(netdev))) {
3620 rn = netdev_priv(netdev);
3621 rn->free_rdma_netdev = mlx5_ib_free_rdma_netdev;
3626 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
3628 struct mlx5_ib_dev *dev;
3629 enum rdma_link_layer ll;
3635 port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3636 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3638 printk_once(KERN_INFO "%s", mlx5_version);
3640 dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
3646 dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
3651 rwlock_init(&dev->roce.netdev_lock);
3652 err = get_port_caps(dev);
3656 if (mlx5_use_mad_ifc(dev))
3657 get_ext_port_caps(dev);
3659 if (!mlx5_lag_is_active(mdev))
3662 name = "mlx5_bond_%d";
3664 strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
3665 dev->ib_dev.owner = THIS_MODULE;
3666 dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3667 dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3668 dev->num_ports = MLX5_CAP_GEN(mdev, num_ports);
3669 dev->ib_dev.phys_port_cnt = dev->num_ports;
3670 dev->ib_dev.num_comp_vectors =
3671 dev->mdev->priv.eq_table.num_comp_vectors;
3672 dev->ib_dev.dev.parent = &mdev->pdev->dev;
3674 dev->ib_dev.uverbs_abi_ver = MLX5_IB_UVERBS_ABI_VERSION;
3675 dev->ib_dev.uverbs_cmd_mask =
3676 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
3677 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
3678 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
3679 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
3680 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
3681 (1ull << IB_USER_VERBS_CMD_CREATE_AH) |
3682 (1ull << IB_USER_VERBS_CMD_DESTROY_AH) |
3683 (1ull << IB_USER_VERBS_CMD_REG_MR) |
3684 (1ull << IB_USER_VERBS_CMD_REREG_MR) |
3685 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
3686 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
3687 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
3688 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
3689 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
3690 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
3691 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
3692 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
3693 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
3694 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
3695 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) |
3696 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
3697 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
3698 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
3699 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) |
3700 (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ) |
3701 (1ull << IB_USER_VERBS_CMD_OPEN_QP);
3702 dev->ib_dev.uverbs_ex_cmd_mask =
3703 (1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE) |
3704 (1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ) |
3705 (1ull << IB_USER_VERBS_EX_CMD_CREATE_QP) |
3706 (1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP);
3708 dev->ib_dev.query_device = mlx5_ib_query_device;
3709 dev->ib_dev.query_port = mlx5_ib_query_port;
3710 dev->ib_dev.get_link_layer = mlx5_ib_port_link_layer;
3711 if (ll == IB_LINK_LAYER_ETHERNET)
3712 dev->ib_dev.get_netdev = mlx5_ib_get_netdev;
3713 dev->ib_dev.query_gid = mlx5_ib_query_gid;
3714 dev->ib_dev.add_gid = mlx5_ib_add_gid;
3715 dev->ib_dev.del_gid = mlx5_ib_del_gid;
3716 dev->ib_dev.query_pkey = mlx5_ib_query_pkey;
3717 dev->ib_dev.modify_device = mlx5_ib_modify_device;
3718 dev->ib_dev.modify_port = mlx5_ib_modify_port;
3719 dev->ib_dev.alloc_ucontext = mlx5_ib_alloc_ucontext;
3720 dev->ib_dev.dealloc_ucontext = mlx5_ib_dealloc_ucontext;
3721 dev->ib_dev.mmap = mlx5_ib_mmap;
3722 dev->ib_dev.alloc_pd = mlx5_ib_alloc_pd;
3723 dev->ib_dev.dealloc_pd = mlx5_ib_dealloc_pd;
3724 dev->ib_dev.create_ah = mlx5_ib_create_ah;
3725 dev->ib_dev.query_ah = mlx5_ib_query_ah;
3726 dev->ib_dev.destroy_ah = mlx5_ib_destroy_ah;
3727 dev->ib_dev.create_srq = mlx5_ib_create_srq;
3728 dev->ib_dev.modify_srq = mlx5_ib_modify_srq;
3729 dev->ib_dev.query_srq = mlx5_ib_query_srq;
3730 dev->ib_dev.destroy_srq = mlx5_ib_destroy_srq;
3731 dev->ib_dev.post_srq_recv = mlx5_ib_post_srq_recv;
3732 dev->ib_dev.create_qp = mlx5_ib_create_qp;
3733 dev->ib_dev.modify_qp = mlx5_ib_modify_qp;
3734 dev->ib_dev.query_qp = mlx5_ib_query_qp;
3735 dev->ib_dev.destroy_qp = mlx5_ib_destroy_qp;
3736 dev->ib_dev.post_send = mlx5_ib_post_send;
3737 dev->ib_dev.post_recv = mlx5_ib_post_recv;
3738 dev->ib_dev.create_cq = mlx5_ib_create_cq;
3739 dev->ib_dev.modify_cq = mlx5_ib_modify_cq;
3740 dev->ib_dev.resize_cq = mlx5_ib_resize_cq;
3741 dev->ib_dev.destroy_cq = mlx5_ib_destroy_cq;
3742 dev->ib_dev.poll_cq = mlx5_ib_poll_cq;
3743 dev->ib_dev.req_notify_cq = mlx5_ib_arm_cq;
3744 dev->ib_dev.get_dma_mr = mlx5_ib_get_dma_mr;
3745 dev->ib_dev.reg_user_mr = mlx5_ib_reg_user_mr;
3746 dev->ib_dev.rereg_user_mr = mlx5_ib_rereg_user_mr;
3747 dev->ib_dev.dereg_mr = mlx5_ib_dereg_mr;
3748 dev->ib_dev.attach_mcast = mlx5_ib_mcg_attach;
3749 dev->ib_dev.detach_mcast = mlx5_ib_mcg_detach;
3750 dev->ib_dev.process_mad = mlx5_ib_process_mad;
3751 dev->ib_dev.alloc_mr = mlx5_ib_alloc_mr;
3752 dev->ib_dev.map_mr_sg = mlx5_ib_map_mr_sg;
3753 dev->ib_dev.check_mr_status = mlx5_ib_check_mr_status;
3754 dev->ib_dev.get_port_immutable = mlx5_port_immutable;
3755 dev->ib_dev.get_dev_fw_str = get_dev_fw_str;
3756 if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads))
3757 dev->ib_dev.alloc_rdma_netdev = mlx5_ib_alloc_rdma_netdev;
3759 if (mlx5_core_is_pf(mdev)) {
3760 dev->ib_dev.get_vf_config = mlx5_ib_get_vf_config;
3761 dev->ib_dev.set_vf_link_state = mlx5_ib_set_vf_link_state;
3762 dev->ib_dev.get_vf_stats = mlx5_ib_get_vf_stats;
3763 dev->ib_dev.set_vf_guid = mlx5_ib_set_vf_guid;
3766 dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
3768 mlx5_ib_internal_fill_odp_caps(dev);
3770 dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3772 if (MLX5_CAP_GEN(mdev, imaicl)) {
3773 dev->ib_dev.alloc_mw = mlx5_ib_alloc_mw;
3774 dev->ib_dev.dealloc_mw = mlx5_ib_dealloc_mw;
3775 dev->ib_dev.uverbs_cmd_mask |=
3776 (1ull << IB_USER_VERBS_CMD_ALLOC_MW) |
3777 (1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
3780 if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
3781 dev->ib_dev.get_hw_stats = mlx5_ib_get_hw_stats;
3782 dev->ib_dev.alloc_hw_stats = mlx5_ib_alloc_hw_stats;
3785 if (MLX5_CAP_GEN(mdev, xrc)) {
3786 dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
3787 dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
3788 dev->ib_dev.uverbs_cmd_mask |=
3789 (1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
3790 (1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
3793 if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
3794 IB_LINK_LAYER_ETHERNET) {
3795 dev->ib_dev.create_flow = mlx5_ib_create_flow;
3796 dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
3797 dev->ib_dev.create_wq = mlx5_ib_create_wq;
3798 dev->ib_dev.modify_wq = mlx5_ib_modify_wq;
3799 dev->ib_dev.destroy_wq = mlx5_ib_destroy_wq;
3800 dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
3801 dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
3802 dev->ib_dev.uverbs_ex_cmd_mask |=
3803 (1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
3804 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
3805 (1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
3806 (1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
3807 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
3808 (1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
3809 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
3811 err = init_node_data(dev);
3815 mutex_init(&dev->flow_db.lock);
3816 mutex_init(&dev->cap_mask_mutex);
3817 INIT_LIST_HEAD(&dev->qp_list);
3818 spin_lock_init(&dev->reset_flow_resource_lock);
3820 if (ll == IB_LINK_LAYER_ETHERNET) {
3821 err = mlx5_enable_eth(dev);
3824 dev->roce.last_port_state = IB_PORT_DOWN;
3827 err = create_dev_resources(&dev->devr);
3829 goto err_disable_eth;
3831 err = mlx5_ib_odp_init_one(dev);
3835 if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
3836 err = mlx5_ib_alloc_counters(dev);
3841 err = mlx5_ib_init_cong_debugfs(dev);
3845 dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
3846 if (!dev->mdev->priv.uar)
3849 err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
3853 err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
3857 err = ib_register_device(&dev->ib_dev, NULL);
3861 err = create_umr_res(dev);
3865 for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
3866 err = device_create_file(&dev->ib_dev.dev,
3867 mlx5_class_attributes[i]);
3872 if ((MLX5_CAP_GEN(mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
3873 MLX5_CAP_GEN(mdev, disable_local_lb))
3874 mutex_init(&dev->lb_mutex);
3876 dev->ib_active = true;
3881 destroy_umrc_res(dev);
3884 ib_unregister_device(&dev->ib_dev);
3887 mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3890 mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3893 mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
3896 mlx5_ib_cleanup_cong_debugfs(dev);
3898 if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
3899 mlx5_ib_dealloc_counters(dev);
3902 mlx5_ib_odp_remove_one(dev);
3905 destroy_dev_resources(&dev->devr);
3908 if (ll == IB_LINK_LAYER_ETHERNET) {
3909 mlx5_disable_eth(dev);
3910 mlx5_remove_netdev_notifier(dev);
3917 ib_dealloc_device((struct ib_device *)dev);
3922 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3924 struct mlx5_ib_dev *dev = context;
3925 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3927 mlx5_remove_netdev_notifier(dev);
3928 ib_unregister_device(&dev->ib_dev);
3929 mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3930 mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3931 mlx5_put_uars_page(dev->mdev, mdev->priv.uar);
3932 mlx5_ib_cleanup_cong_debugfs(dev);
3933 if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
3934 mlx5_ib_dealloc_counters(dev);
3935 destroy_umrc_res(dev);
3936 mlx5_ib_odp_remove_one(dev);
3937 destroy_dev_resources(&dev->devr);
3938 if (ll == IB_LINK_LAYER_ETHERNET)
3939 mlx5_disable_eth(dev);
3941 ib_dealloc_device(&dev->ib_dev);
3944 static struct mlx5_interface mlx5_ib_interface = {
3946 .remove = mlx5_ib_remove,
3947 .event = mlx5_ib_event,
3948 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
3949 .pfault = mlx5_ib_pfault,
3951 .protocol = MLX5_INTERFACE_PROTOCOL_IB,
3954 static int __init mlx5_ib_init(void)
3960 err = mlx5_register_interface(&mlx5_ib_interface);
3965 static void __exit mlx5_ib_cleanup(void)
3967 mlx5_unregister_interface(&mlx5_ib_interface);
3970 module_init(mlx5_ib_init);
3971 module_exit(mlx5_ib_cleanup);