2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/debugfs.h>
34 #include <linux/highmem.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/errno.h>
38 #include <linux/pci.h>
39 #include <linux/dma-mapping.h>
40 #include <linux/slab.h>
41 #include <linux/bitmap.h>
42 #if defined(CONFIG_X86)
45 #include <linux/sched.h>
46 #include <linux/sched/mm.h>
47 #include <linux/sched/task.h>
48 #include <linux/delay.h>
49 #include <rdma/ib_user_verbs.h>
50 #include <rdma/ib_addr.h>
51 #include <rdma/ib_cache.h>
52 #include <linux/mlx5/port.h>
53 #include <linux/mlx5/vport.h>
54 #include <linux/mlx5/fs.h>
55 #include <linux/mlx5/fs_helpers.h>
56 #include <linux/list.h>
57 #include <rdma/ib_smi.h>
58 #include <rdma/ib_umem.h>
60 #include <linux/etherdevice.h>
64 #include <linux/mlx5/fs_helpers.h>
65 #include <linux/mlx5/accel.h>
66 #include <rdma/uverbs_std_types.h>
67 #include <rdma/mlx5_user_ioctl_verbs.h>
68 #include <rdma/mlx5_user_ioctl_cmds.h>
70 #define UVERBS_MODULE_NAME mlx5_ib
71 #include <rdma/uverbs_named_ioctl.h>
73 #define DRIVER_NAME "mlx5_ib"
74 #define DRIVER_VERSION "5.0-0"
76 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
77 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
78 MODULE_LICENSE("Dual BSD/GPL");
80 static char mlx5_version[] =
81 DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
84 struct mlx5_ib_event_work {
85 struct work_struct work;
86 struct mlx5_core_dev *dev;
88 enum mlx5_dev_event event;
93 MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
96 static struct workqueue_struct *mlx5_ib_event_wq;
97 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
98 static LIST_HEAD(mlx5_ib_dev_list);
100 * This mutex should be held when accessing either of the above lists
102 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
104 /* We can't use an array for xlt_emergency_page because dma_map_single
105 * doesn't work on kernel modules memory
107 static unsigned long xlt_emergency_page;
108 static struct mutex xlt_emergency_page_mutex;
110 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
112 struct mlx5_ib_dev *dev;
114 mutex_lock(&mlx5_ib_multiport_mutex);
116 mutex_unlock(&mlx5_ib_multiport_mutex);
120 static enum rdma_link_layer
121 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
123 switch (port_type_cap) {
124 case MLX5_CAP_PORT_TYPE_IB:
125 return IB_LINK_LAYER_INFINIBAND;
126 case MLX5_CAP_PORT_TYPE_ETH:
127 return IB_LINK_LAYER_ETHERNET;
129 return IB_LINK_LAYER_UNSPECIFIED;
133 static enum rdma_link_layer
134 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
136 struct mlx5_ib_dev *dev = to_mdev(device);
137 int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
139 return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
142 static int get_port_state(struct ib_device *ibdev,
144 enum ib_port_state *state)
146 struct ib_port_attr attr;
149 memset(&attr, 0, sizeof(attr));
150 ret = ibdev->query_port(ibdev, port_num, &attr);
156 static int mlx5_netdev_event(struct notifier_block *this,
157 unsigned long event, void *ptr)
159 struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
160 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
161 u8 port_num = roce->native_port_num;
162 struct mlx5_core_dev *mdev;
163 struct mlx5_ib_dev *ibdev;
166 mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
171 case NETDEV_REGISTER:
172 case NETDEV_UNREGISTER:
173 write_lock(&roce->netdev_lock);
175 struct mlx5_eswitch *esw = ibdev->mdev->priv.eswitch;
176 struct net_device *rep_ndev;
178 rep_ndev = mlx5_ib_get_rep_netdev(esw,
180 if (rep_ndev == ndev)
181 roce->netdev = (event == NETDEV_UNREGISTER) ?
183 } else if (ndev->dev.parent == &ibdev->mdev->pdev->dev) {
184 roce->netdev = (event == NETDEV_UNREGISTER) ?
187 write_unlock(&roce->netdev_lock);
193 struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
194 struct net_device *upper = NULL;
197 upper = netdev_master_upper_dev_get(lag_ndev);
201 if ((upper == ndev || (!upper && ndev == roce->netdev))
202 && ibdev->ib_active) {
203 struct ib_event ibev = { };
204 enum ib_port_state port_state;
206 if (get_port_state(&ibdev->ib_dev, port_num,
210 if (roce->last_port_state == port_state)
213 roce->last_port_state = port_state;
214 ibev.device = &ibdev->ib_dev;
215 if (port_state == IB_PORT_DOWN)
216 ibev.event = IB_EVENT_PORT_ERR;
217 else if (port_state == IB_PORT_ACTIVE)
218 ibev.event = IB_EVENT_PORT_ACTIVE;
222 ibev.element.port_num = port_num;
223 ib_dispatch_event(&ibev);
232 mlx5_ib_put_native_port_mdev(ibdev, port_num);
236 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
239 struct mlx5_ib_dev *ibdev = to_mdev(device);
240 struct net_device *ndev;
241 struct mlx5_core_dev *mdev;
243 mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
247 ndev = mlx5_lag_get_roce_netdev(mdev);
251 /* Ensure ndev does not disappear before we invoke dev_hold()
253 read_lock(&ibdev->roce[port_num - 1].netdev_lock);
254 ndev = ibdev->roce[port_num - 1].netdev;
257 read_unlock(&ibdev->roce[port_num - 1].netdev_lock);
260 mlx5_ib_put_native_port_mdev(ibdev, port_num);
264 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
268 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
270 struct mlx5_core_dev *mdev = NULL;
271 struct mlx5_ib_multiport_info *mpi;
272 struct mlx5_ib_port *port;
274 if (!mlx5_core_mp_enabled(ibdev->mdev) ||
275 ll != IB_LINK_LAYER_ETHERNET) {
277 *native_port_num = ib_port_num;
282 *native_port_num = 1;
284 port = &ibdev->port[ib_port_num - 1];
288 spin_lock(&port->mp.mpi_lock);
289 mpi = ibdev->port[ib_port_num - 1].mp.mpi;
290 if (mpi && !mpi->unaffiliate) {
292 /* If it's the master no need to refcount, it'll exist
293 * as long as the ib_dev exists.
298 spin_unlock(&port->mp.mpi_lock);
303 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u8 port_num)
305 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
307 struct mlx5_ib_multiport_info *mpi;
308 struct mlx5_ib_port *port;
310 if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
313 port = &ibdev->port[port_num - 1];
315 spin_lock(&port->mp.mpi_lock);
316 mpi = ibdev->port[port_num - 1].mp.mpi;
321 if (mpi->unaffiliate)
322 complete(&mpi->unref_comp);
324 spin_unlock(&port->mp.mpi_lock);
327 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
330 switch (eth_proto_oper) {
331 case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
332 case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
333 case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
334 case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
335 *active_width = IB_WIDTH_1X;
336 *active_speed = IB_SPEED_SDR;
338 case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
339 case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
340 case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
341 case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
342 case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
343 case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
344 case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
345 *active_width = IB_WIDTH_1X;
346 *active_speed = IB_SPEED_QDR;
348 case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
349 case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
350 case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
351 *active_width = IB_WIDTH_1X;
352 *active_speed = IB_SPEED_EDR;
354 case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
355 case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
356 case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
357 case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
358 *active_width = IB_WIDTH_4X;
359 *active_speed = IB_SPEED_QDR;
361 case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
362 case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
363 case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
364 *active_width = IB_WIDTH_1X;
365 *active_speed = IB_SPEED_HDR;
367 case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
368 *active_width = IB_WIDTH_4X;
369 *active_speed = IB_SPEED_FDR;
371 case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
372 case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
373 case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
374 case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
375 *active_width = IB_WIDTH_4X;
376 *active_speed = IB_SPEED_EDR;
385 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
386 struct ib_port_attr *props)
388 struct mlx5_ib_dev *dev = to_mdev(device);
389 struct mlx5_core_dev *mdev;
390 struct net_device *ndev, *upper;
391 enum ib_mtu ndev_ib_mtu;
392 bool put_mdev = true;
398 mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
400 /* This means the port isn't affiliated yet. Get the
401 * info for the master port instead.
409 /* Possible bad flows are checked before filling out props so in case
410 * of an error it will still be zeroed out.
412 err = mlx5_query_port_eth_proto_oper(mdev, ð_prot_oper,
417 props->active_width = IB_WIDTH_4X;
418 props->active_speed = IB_SPEED_QDR;
420 translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
421 &props->active_width);
423 props->port_cap_flags |= IB_PORT_CM_SUP;
424 props->port_cap_flags |= IB_PORT_IP_BASED_GIDS;
426 props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
427 roce_address_table_size);
428 props->max_mtu = IB_MTU_4096;
429 props->max_msg_sz = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
430 props->pkey_tbl_len = 1;
431 props->state = IB_PORT_DOWN;
432 props->phys_state = 3;
434 mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
435 props->qkey_viol_cntr = qkey_viol_cntr;
437 /* If this is a stub query for an unaffiliated port stop here */
441 ndev = mlx5_ib_get_netdev(device, port_num);
445 if (mlx5_lag_is_active(dev->mdev)) {
447 upper = netdev_master_upper_dev_get_rcu(ndev);
456 if (netif_running(ndev) && netif_carrier_ok(ndev)) {
457 props->state = IB_PORT_ACTIVE;
458 props->phys_state = 5;
461 ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
465 props->active_mtu = min(props->max_mtu, ndev_ib_mtu);
468 mlx5_ib_put_native_port_mdev(dev, port_num);
472 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
473 unsigned int index, const union ib_gid *gid,
474 const struct ib_gid_attr *attr)
476 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
484 gid_type = attr->gid_type;
485 ether_addr_copy(mac, attr->ndev->dev_addr);
487 if (is_vlan_dev(attr->ndev)) {
489 vlan_id = vlan_dev_vlan_id(attr->ndev);
495 roce_version = MLX5_ROCE_VERSION_1;
497 case IB_GID_TYPE_ROCE_UDP_ENCAP:
498 roce_version = MLX5_ROCE_VERSION_2;
499 if (ipv6_addr_v4mapped((void *)gid))
500 roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
502 roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
506 mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
509 return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
510 roce_l3_type, gid->raw, mac, vlan,
514 static int mlx5_ib_add_gid(const union ib_gid *gid,
515 const struct ib_gid_attr *attr,
516 __always_unused void **context)
518 return set_roce_addr(to_mdev(attr->device), attr->port_num,
519 attr->index, gid, attr);
522 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
523 __always_unused void **context)
525 return set_roce_addr(to_mdev(attr->device), attr->port_num,
526 attr->index, NULL, NULL);
529 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
532 struct ib_gid_attr attr;
535 if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
540 if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
543 return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
546 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
547 int index, enum ib_gid_type *gid_type)
549 struct ib_gid_attr attr;
553 ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
559 *gid_type = attr.gid_type;
564 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
566 if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
567 return !MLX5_CAP_GEN(dev->mdev, ib_virt);
572 MLX5_VPORT_ACCESS_METHOD_MAD,
573 MLX5_VPORT_ACCESS_METHOD_HCA,
574 MLX5_VPORT_ACCESS_METHOD_NIC,
577 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
579 if (mlx5_use_mad_ifc(to_mdev(ibdev)))
580 return MLX5_VPORT_ACCESS_METHOD_MAD;
582 if (mlx5_ib_port_link_layer(ibdev, 1) ==
583 IB_LINK_LAYER_ETHERNET)
584 return MLX5_VPORT_ACCESS_METHOD_NIC;
586 return MLX5_VPORT_ACCESS_METHOD_HCA;
589 static void get_atomic_caps(struct mlx5_ib_dev *dev,
591 struct ib_device_attr *props)
594 u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
595 u8 atomic_req_8B_endianness_mode =
596 MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
598 /* Check if HW supports 8 bytes standard atomic operations and capable
599 * of host endianness respond
601 tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
602 if (((atomic_operations & tmp) == tmp) &&
603 (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
604 (atomic_req_8B_endianness_mode)) {
605 props->atomic_cap = IB_ATOMIC_HCA;
607 props->atomic_cap = IB_ATOMIC_NONE;
611 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
612 struct ib_device_attr *props)
614 u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
616 get_atomic_caps(dev, atomic_size_qp, props);
619 static void get_atomic_caps_dc(struct mlx5_ib_dev *dev,
620 struct ib_device_attr *props)
622 u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
624 get_atomic_caps(dev, atomic_size_qp, props);
627 bool mlx5_ib_dc_atomic_is_supported(struct mlx5_ib_dev *dev)
629 struct ib_device_attr props = {};
631 get_atomic_caps_dc(dev, &props);
632 return (props.atomic_cap == IB_ATOMIC_HCA) ? true : false;
634 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
635 __be64 *sys_image_guid)
637 struct mlx5_ib_dev *dev = to_mdev(ibdev);
638 struct mlx5_core_dev *mdev = dev->mdev;
642 switch (mlx5_get_vport_access_method(ibdev)) {
643 case MLX5_VPORT_ACCESS_METHOD_MAD:
644 return mlx5_query_mad_ifc_system_image_guid(ibdev,
647 case MLX5_VPORT_ACCESS_METHOD_HCA:
648 err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
651 case MLX5_VPORT_ACCESS_METHOD_NIC:
652 err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
660 *sys_image_guid = cpu_to_be64(tmp);
666 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
669 struct mlx5_ib_dev *dev = to_mdev(ibdev);
670 struct mlx5_core_dev *mdev = dev->mdev;
672 switch (mlx5_get_vport_access_method(ibdev)) {
673 case MLX5_VPORT_ACCESS_METHOD_MAD:
674 return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
676 case MLX5_VPORT_ACCESS_METHOD_HCA:
677 case MLX5_VPORT_ACCESS_METHOD_NIC:
678 *max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
687 static int mlx5_query_vendor_id(struct ib_device *ibdev,
690 struct mlx5_ib_dev *dev = to_mdev(ibdev);
692 switch (mlx5_get_vport_access_method(ibdev)) {
693 case MLX5_VPORT_ACCESS_METHOD_MAD:
694 return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
696 case MLX5_VPORT_ACCESS_METHOD_HCA:
697 case MLX5_VPORT_ACCESS_METHOD_NIC:
698 return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
705 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
711 switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
712 case MLX5_VPORT_ACCESS_METHOD_MAD:
713 return mlx5_query_mad_ifc_node_guid(dev, node_guid);
715 case MLX5_VPORT_ACCESS_METHOD_HCA:
716 err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
719 case MLX5_VPORT_ACCESS_METHOD_NIC:
720 err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
728 *node_guid = cpu_to_be64(tmp);
733 struct mlx5_reg_node_desc {
734 u8 desc[IB_DEVICE_NODE_DESC_MAX];
737 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
739 struct mlx5_reg_node_desc in;
741 if (mlx5_use_mad_ifc(dev))
742 return mlx5_query_mad_ifc_node_desc(dev, node_desc);
744 memset(&in, 0, sizeof(in));
746 return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
747 sizeof(struct mlx5_reg_node_desc),
748 MLX5_REG_NODE_DESC, 0, 0);
751 static int mlx5_ib_query_device(struct ib_device *ibdev,
752 struct ib_device_attr *props,
753 struct ib_udata *uhw)
755 struct mlx5_ib_dev *dev = to_mdev(ibdev);
756 struct mlx5_core_dev *mdev = dev->mdev;
761 u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
762 bool raw_support = !mlx5_core_mp_enabled(mdev);
763 struct mlx5_ib_query_device_resp resp = {};
767 resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
768 if (uhw->outlen && uhw->outlen < resp_len)
771 resp.response_length = resp_len;
773 if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
776 memset(props, 0, sizeof(*props));
777 err = mlx5_query_system_image_guid(ibdev,
778 &props->sys_image_guid);
782 err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
786 err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
790 props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
791 (fw_rev_min(dev->mdev) << 16) |
792 fw_rev_sub(dev->mdev);
793 props->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT |
794 IB_DEVICE_PORT_ACTIVE_EVENT |
795 IB_DEVICE_SYS_IMAGE_GUID |
796 IB_DEVICE_RC_RNR_NAK_GEN;
798 if (MLX5_CAP_GEN(mdev, pkv))
799 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
800 if (MLX5_CAP_GEN(mdev, qkv))
801 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
802 if (MLX5_CAP_GEN(mdev, apm))
803 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
804 if (MLX5_CAP_GEN(mdev, xrc))
805 props->device_cap_flags |= IB_DEVICE_XRC;
806 if (MLX5_CAP_GEN(mdev, imaicl)) {
807 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
808 IB_DEVICE_MEM_WINDOW_TYPE_2B;
809 props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
810 /* We support 'Gappy' memory registration too */
811 props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
813 props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
814 if (MLX5_CAP_GEN(mdev, sho)) {
815 props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
816 /* At this stage no support for signature handover */
817 props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
818 IB_PROT_T10DIF_TYPE_2 |
819 IB_PROT_T10DIF_TYPE_3;
820 props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
821 IB_GUARD_T10DIF_CSUM;
823 if (MLX5_CAP_GEN(mdev, block_lb_mc))
824 props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
826 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
827 if (MLX5_CAP_ETH(mdev, csum_cap)) {
828 /* Legacy bit to support old userspace libraries */
829 props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
830 props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
833 if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
834 props->raw_packet_caps |=
835 IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
837 if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
838 max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
840 resp.tso_caps.max_tso = 1 << max_tso;
841 resp.tso_caps.supported_qpts |=
842 1 << IB_QPT_RAW_PACKET;
843 resp.response_length += sizeof(resp.tso_caps);
847 if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
848 resp.rss_caps.rx_hash_function =
849 MLX5_RX_HASH_FUNC_TOEPLITZ;
850 resp.rss_caps.rx_hash_fields_mask =
851 MLX5_RX_HASH_SRC_IPV4 |
852 MLX5_RX_HASH_DST_IPV4 |
853 MLX5_RX_HASH_SRC_IPV6 |
854 MLX5_RX_HASH_DST_IPV6 |
855 MLX5_RX_HASH_SRC_PORT_TCP |
856 MLX5_RX_HASH_DST_PORT_TCP |
857 MLX5_RX_HASH_SRC_PORT_UDP |
858 MLX5_RX_HASH_DST_PORT_UDP |
860 if (mlx5_accel_ipsec_device_caps(dev->mdev) &
861 MLX5_ACCEL_IPSEC_CAP_DEVICE)
862 resp.rss_caps.rx_hash_fields_mask |=
863 MLX5_RX_HASH_IPSEC_SPI;
864 resp.response_length += sizeof(resp.rss_caps);
867 if (field_avail(typeof(resp), tso_caps, uhw->outlen))
868 resp.response_length += sizeof(resp.tso_caps);
869 if (field_avail(typeof(resp), rss_caps, uhw->outlen))
870 resp.response_length += sizeof(resp.rss_caps);
873 if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
874 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
875 props->device_cap_flags |= IB_DEVICE_UD_TSO;
878 if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
879 MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
881 props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
883 if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
884 MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
885 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
887 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
888 MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
890 /* Legacy bit to support old userspace libraries */
891 props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
892 props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
895 if (MLX5_CAP_DEV_MEM(mdev, memic)) {
897 MLX5_CAP_DEV_MEM(mdev, max_memic_size);
900 if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
901 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
903 if (MLX5_CAP_GEN(mdev, end_pad))
904 props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
906 props->vendor_part_id = mdev->pdev->device;
907 props->hw_ver = mdev->pdev->revision;
909 props->max_mr_size = ~0ull;
910 props->page_size_cap = ~(min_page_size - 1);
911 props->max_qp = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
912 props->max_qp_wr = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
913 max_rq_sg = MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
914 sizeof(struct mlx5_wqe_data_seg);
915 max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
916 max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
917 sizeof(struct mlx5_wqe_raddr_seg)) /
918 sizeof(struct mlx5_wqe_data_seg);
919 props->max_sge = min(max_rq_sg, max_sq_sg);
920 props->max_sge_rd = MLX5_MAX_SGE_RD;
921 props->max_cq = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
922 props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
923 props->max_mr = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
924 props->max_pd = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
925 props->max_qp_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
926 props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
927 props->max_srq = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
928 props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
929 props->local_ca_ack_delay = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
930 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp;
931 props->max_srq_sge = max_rq_sg - 1;
932 props->max_fast_reg_page_list_len =
933 1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
934 get_atomic_caps_qp(dev, props);
935 props->masked_atomic_cap = IB_ATOMIC_NONE;
936 props->max_mcast_grp = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
937 props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
938 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
939 props->max_mcast_grp;
940 props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
941 props->max_ah = INT_MAX;
942 props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
943 props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
945 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
946 if (MLX5_CAP_GEN(mdev, pg))
947 props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
948 props->odp_caps = dev->odp_caps;
951 if (MLX5_CAP_GEN(mdev, cd))
952 props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
954 if (!mlx5_core_is_pf(mdev))
955 props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
957 if (mlx5_ib_port_link_layer(ibdev, 1) ==
958 IB_LINK_LAYER_ETHERNET && raw_support) {
959 props->rss_caps.max_rwq_indirection_tables =
960 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
961 props->rss_caps.max_rwq_indirection_table_size =
962 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
963 props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
964 props->max_wq_type_rq =
965 1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
968 if (MLX5_CAP_GEN(mdev, tag_matching)) {
969 props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
970 props->tm_caps.max_num_tags =
971 (1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
972 props->tm_caps.flags = IB_TM_CAP_RC;
973 props->tm_caps.max_ops =
974 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
975 props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
978 if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
979 props->cq_caps.max_cq_moderation_count =
981 props->cq_caps.max_cq_moderation_period =
985 if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
986 resp.cqe_comp_caps.max_num =
987 MLX5_CAP_GEN(dev->mdev, cqe_compression) ?
988 MLX5_CAP_GEN(dev->mdev, cqe_compression_max_num) : 0;
989 resp.cqe_comp_caps.supported_format =
990 MLX5_IB_CQE_RES_FORMAT_HASH |
991 MLX5_IB_CQE_RES_FORMAT_CSUM;
992 resp.response_length += sizeof(resp.cqe_comp_caps);
995 if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen) &&
997 if (MLX5_CAP_QOS(mdev, packet_pacing) &&
998 MLX5_CAP_GEN(mdev, qos)) {
999 resp.packet_pacing_caps.qp_rate_limit_max =
1000 MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1001 resp.packet_pacing_caps.qp_rate_limit_min =
1002 MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1003 resp.packet_pacing_caps.supported_qpts |=
1004 1 << IB_QPT_RAW_PACKET;
1005 if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1006 MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1007 resp.packet_pacing_caps.cap_flags |=
1008 MLX5_IB_PP_SUPPORT_BURST;
1010 resp.response_length += sizeof(resp.packet_pacing_caps);
1013 if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
1015 if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1016 resp.mlx5_ib_support_multi_pkt_send_wqes =
1019 if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1020 resp.mlx5_ib_support_multi_pkt_send_wqes |=
1021 MLX5_IB_SUPPORT_EMPW;
1023 resp.response_length +=
1024 sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1027 if (field_avail(typeof(resp), flags, uhw->outlen)) {
1028 resp.response_length += sizeof(resp.flags);
1030 if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1032 MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1034 if (MLX5_CAP_GEN(mdev, cqe_128_always))
1035 resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1038 if (field_avail(typeof(resp), sw_parsing_caps,
1040 resp.response_length += sizeof(resp.sw_parsing_caps);
1041 if (MLX5_CAP_ETH(mdev, swp)) {
1042 resp.sw_parsing_caps.sw_parsing_offloads |=
1045 if (MLX5_CAP_ETH(mdev, swp_csum))
1046 resp.sw_parsing_caps.sw_parsing_offloads |=
1047 MLX5_IB_SW_PARSING_CSUM;
1049 if (MLX5_CAP_ETH(mdev, swp_lso))
1050 resp.sw_parsing_caps.sw_parsing_offloads |=
1051 MLX5_IB_SW_PARSING_LSO;
1053 if (resp.sw_parsing_caps.sw_parsing_offloads)
1054 resp.sw_parsing_caps.supported_qpts =
1055 BIT(IB_QPT_RAW_PACKET);
1059 if (field_avail(typeof(resp), striding_rq_caps, uhw->outlen) &&
1061 resp.response_length += sizeof(resp.striding_rq_caps);
1062 if (MLX5_CAP_GEN(mdev, striding_rq)) {
1063 resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1064 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1065 resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1066 MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1067 resp.striding_rq_caps.min_single_wqe_log_num_of_strides =
1068 MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1069 resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1070 MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1071 resp.striding_rq_caps.supported_qpts =
1072 BIT(IB_QPT_RAW_PACKET);
1076 if (field_avail(typeof(resp), tunnel_offloads_caps,
1078 resp.response_length += sizeof(resp.tunnel_offloads_caps);
1079 if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1080 resp.tunnel_offloads_caps |=
1081 MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1082 if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1083 resp.tunnel_offloads_caps |=
1084 MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1085 if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1086 resp.tunnel_offloads_caps |=
1087 MLX5_IB_TUNNELED_OFFLOADS_GRE;
1091 err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1100 enum mlx5_ib_width {
1101 MLX5_IB_WIDTH_1X = 1 << 0,
1102 MLX5_IB_WIDTH_2X = 1 << 1,
1103 MLX5_IB_WIDTH_4X = 1 << 2,
1104 MLX5_IB_WIDTH_8X = 1 << 3,
1105 MLX5_IB_WIDTH_12X = 1 << 4
1108 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
1111 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1114 if (active_width & MLX5_IB_WIDTH_1X) {
1115 *ib_width = IB_WIDTH_1X;
1116 } else if (active_width & MLX5_IB_WIDTH_2X) {
1117 mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
1120 } else if (active_width & MLX5_IB_WIDTH_4X) {
1121 *ib_width = IB_WIDTH_4X;
1122 } else if (active_width & MLX5_IB_WIDTH_8X) {
1123 *ib_width = IB_WIDTH_8X;
1124 } else if (active_width & MLX5_IB_WIDTH_12X) {
1125 *ib_width = IB_WIDTH_12X;
1127 mlx5_ib_dbg(dev, "Invalid active_width %d\n",
1135 static int mlx5_mtu_to_ib_mtu(int mtu)
1140 case 1024: return 3;
1141 case 2048: return 4;
1142 case 4096: return 5;
1144 pr_warn("invalid mtu\n");
1149 enum ib_max_vl_num {
1151 __IB_MAX_VL_0_1 = 2,
1152 __IB_MAX_VL_0_3 = 3,
1153 __IB_MAX_VL_0_7 = 4,
1154 __IB_MAX_VL_0_14 = 5,
1157 enum mlx5_vl_hw_cap {
1166 MLX5_VL_HW_0_14 = 15
1169 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1172 switch (vl_hw_cap) {
1174 *max_vl_num = __IB_MAX_VL_0;
1176 case MLX5_VL_HW_0_1:
1177 *max_vl_num = __IB_MAX_VL_0_1;
1179 case MLX5_VL_HW_0_3:
1180 *max_vl_num = __IB_MAX_VL_0_3;
1182 case MLX5_VL_HW_0_7:
1183 *max_vl_num = __IB_MAX_VL_0_7;
1185 case MLX5_VL_HW_0_14:
1186 *max_vl_num = __IB_MAX_VL_0_14;
1196 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1197 struct ib_port_attr *props)
1199 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1200 struct mlx5_core_dev *mdev = dev->mdev;
1201 struct mlx5_hca_vport_context *rep;
1205 u8 ib_link_width_oper;
1208 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1214 /* props being zeroed by the caller, avoid zeroing it here */
1216 err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1220 props->lid = rep->lid;
1221 props->lmc = rep->lmc;
1222 props->sm_lid = rep->sm_lid;
1223 props->sm_sl = rep->sm_sl;
1224 props->state = rep->vport_state;
1225 props->phys_state = rep->port_physical_state;
1226 props->port_cap_flags = rep->cap_mask1;
1227 props->gid_tbl_len = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1228 props->max_msg_sz = 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1229 props->pkey_tbl_len = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1230 props->bad_pkey_cntr = rep->pkey_violation_counter;
1231 props->qkey_viol_cntr = rep->qkey_violation_counter;
1232 props->subnet_timeout = rep->subnet_timeout;
1233 props->init_type_reply = rep->init_type_reply;
1234 props->grh_required = rep->grh_required;
1236 err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
1240 err = translate_active_width(ibdev, ib_link_width_oper,
1241 &props->active_width);
1244 err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
1248 mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1250 props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1252 mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1254 props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1256 err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1260 err = translate_max_vl_num(ibdev, vl_hw_cap,
1261 &props->max_vl_num);
1267 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1268 struct ib_port_attr *props)
1273 switch (mlx5_get_vport_access_method(ibdev)) {
1274 case MLX5_VPORT_ACCESS_METHOD_MAD:
1275 ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1278 case MLX5_VPORT_ACCESS_METHOD_HCA:
1279 ret = mlx5_query_hca_port(ibdev, port, props);
1282 case MLX5_VPORT_ACCESS_METHOD_NIC:
1283 ret = mlx5_query_port_roce(ibdev, port, props);
1290 if (!ret && props) {
1291 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1292 struct mlx5_core_dev *mdev;
1293 bool put_mdev = true;
1295 mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1297 /* If the port isn't affiliated yet query the master.
1298 * The master and slave will have the same values.
1304 count = mlx5_core_reserved_gids_count(mdev);
1306 mlx5_ib_put_native_port_mdev(dev, port);
1307 props->gid_tbl_len -= count;
1312 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u8 port,
1313 struct ib_port_attr *props)
1317 /* Only link layer == ethernet is valid for representors */
1318 ret = mlx5_query_port_roce(ibdev, port, props);
1322 /* We don't support GIDS */
1323 props->gid_tbl_len = 0;
1328 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1331 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1332 struct mlx5_core_dev *mdev = dev->mdev;
1334 switch (mlx5_get_vport_access_method(ibdev)) {
1335 case MLX5_VPORT_ACCESS_METHOD_MAD:
1336 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1338 case MLX5_VPORT_ACCESS_METHOD_HCA:
1339 return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1347 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1348 u16 index, u16 *pkey)
1350 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1351 struct mlx5_core_dev *mdev;
1352 bool put_mdev = true;
1356 mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1358 /* The port isn't affiliated yet, get the PKey from the master
1359 * port. For RoCE the PKey tables will be the same.
1366 err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1369 mlx5_ib_put_native_port_mdev(dev, port);
1374 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1377 switch (mlx5_get_vport_access_method(ibdev)) {
1378 case MLX5_VPORT_ACCESS_METHOD_MAD:
1379 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1381 case MLX5_VPORT_ACCESS_METHOD_HCA:
1382 case MLX5_VPORT_ACCESS_METHOD_NIC:
1383 return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1389 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1390 struct ib_device_modify *props)
1392 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1393 struct mlx5_reg_node_desc in;
1394 struct mlx5_reg_node_desc out;
1397 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1400 if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1404 * If possible, pass node desc to FW, so it can generate
1405 * a 144 trap. If cmd fails, just ignore.
1407 memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1408 err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1409 sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1413 memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1418 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1421 struct mlx5_hca_vport_context ctx = {};
1422 struct mlx5_core_dev *mdev;
1426 mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1430 err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1434 if (~ctx.cap_mask1_perm & mask) {
1435 mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1436 mask, ctx.cap_mask1_perm);
1441 ctx.cap_mask1 = value;
1442 ctx.cap_mask1_perm = mask;
1443 err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1447 mlx5_ib_put_native_port_mdev(dev, port_num);
1452 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1453 struct ib_port_modify *props)
1455 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1456 struct ib_port_attr attr;
1461 bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1462 IB_LINK_LAYER_INFINIBAND);
1464 /* CM layer calls ib_modify_port() regardless of the link layer. For
1465 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1470 if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1471 change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1472 value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1473 return set_port_caps_atomic(dev, port, change_mask, value);
1476 mutex_lock(&dev->cap_mask_mutex);
1478 err = ib_query_port(ibdev, port, &attr);
1482 tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1483 ~props->clr_port_cap_mask;
1485 err = mlx5_set_port_caps(dev->mdev, port, tmp);
1488 mutex_unlock(&dev->cap_mask_mutex);
1492 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1494 mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1495 caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1498 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1500 /* Large page with non 4k uar support might limit the dynamic size */
1501 if (uars_per_sys_page == 1 && PAGE_SIZE > 4096)
1502 return MLX5_MIN_DYN_BFREGS;
1504 return MLX5_MAX_DYN_BFREGS;
1507 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1508 struct mlx5_ib_alloc_ucontext_req_v2 *req,
1509 struct mlx5_bfreg_info *bfregi)
1511 int uars_per_sys_page;
1512 int bfregs_per_sys_page;
1513 int ref_bfregs = req->total_num_bfregs;
1515 if (req->total_num_bfregs == 0)
1518 BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1519 BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1521 if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1524 uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1525 bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1526 /* This holds the required static allocation asked by the user */
1527 req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1528 if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1531 bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1532 bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1533 bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1534 bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1536 mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, allocated %d, total bfregs %d, using %d sys pages\n",
1537 MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1538 lib_uar_4k ? "yes" : "no", ref_bfregs,
1539 req->total_num_bfregs, bfregi->total_num_bfregs,
1540 bfregi->num_sys_pages);
1545 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1547 struct mlx5_bfreg_info *bfregi;
1551 bfregi = &context->bfregi;
1552 for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1553 err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1557 mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1560 for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1561 bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1566 for (--i; i >= 0; i--)
1567 if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1568 mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1573 static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1575 struct mlx5_bfreg_info *bfregi;
1579 bfregi = &context->bfregi;
1580 for (i = 0; i < bfregi->num_sys_pages; i++) {
1581 if (i < bfregi->num_static_sys_pages ||
1582 bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX) {
1583 err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1585 mlx5_ib_warn(dev, "failed to free uar %d, err=%d\n", i, err);
1594 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn)
1598 err = mlx5_core_alloc_transport_domain(dev->mdev, tdn);
1602 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1603 (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1604 !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1607 mutex_lock(&dev->lb_mutex);
1610 if (dev->user_td == 2)
1611 err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1613 mutex_unlock(&dev->lb_mutex);
1617 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn)
1619 mlx5_core_dealloc_transport_domain(dev->mdev, tdn);
1621 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1622 (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1623 !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1626 mutex_lock(&dev->lb_mutex);
1629 if (dev->user_td < 2)
1630 mlx5_nic_vport_update_local_lb(dev->mdev, false);
1632 mutex_unlock(&dev->lb_mutex);
1635 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1636 struct ib_udata *udata)
1638 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1639 struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1640 struct mlx5_ib_alloc_ucontext_resp resp = {};
1641 struct mlx5_core_dev *mdev = dev->mdev;
1642 struct mlx5_ib_ucontext *context;
1643 struct mlx5_bfreg_info *bfregi;
1646 size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1650 if (!dev->ib_active)
1651 return ERR_PTR(-EAGAIN);
1653 if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1655 else if (udata->inlen >= min_req_v2)
1658 return ERR_PTR(-EINVAL);
1660 err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1662 return ERR_PTR(err);
1665 return ERR_PTR(-EINVAL);
1667 if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1668 return ERR_PTR(-EOPNOTSUPP);
1670 req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1671 MLX5_NON_FP_BFREGS_PER_UAR);
1672 if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1673 return ERR_PTR(-EINVAL);
1675 resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1676 if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1677 resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1678 resp.cache_line_size = cache_line_size();
1679 resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1680 resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1681 resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1682 resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1683 resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1684 resp.cqe_version = min_t(__u8,
1685 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1686 req.max_cqe_version);
1687 resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1688 MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1689 resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1690 MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1691 resp.response_length = min(offsetof(typeof(resp), response_length) +
1692 sizeof(resp.response_length), udata->outlen);
1694 if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1695 if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_EGRESS))
1696 resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1697 if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1698 resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1699 if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1700 resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1701 if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1702 resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1703 /* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1706 context = kzalloc(sizeof(*context), GFP_KERNEL);
1708 return ERR_PTR(-ENOMEM);
1710 lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1711 bfregi = &context->bfregi;
1713 /* updates req->total_num_bfregs */
1714 err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1718 mutex_init(&bfregi->lock);
1719 bfregi->lib_uar_4k = lib_uar_4k;
1720 bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1722 if (!bfregi->count) {
1727 bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1728 sizeof(*bfregi->sys_pages),
1730 if (!bfregi->sys_pages) {
1735 err = allocate_uars(dev, context);
1739 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1740 context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1743 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1744 err = mlx5_ib_alloc_transport_domain(dev, &context->tdn);
1749 INIT_LIST_HEAD(&context->vma_private_list);
1750 mutex_init(&context->vma_private_list_mutex);
1751 INIT_LIST_HEAD(&context->db_page_list);
1752 mutex_init(&context->db_page_mutex);
1754 resp.tot_bfregs = req.total_num_bfregs;
1755 resp.num_ports = dev->num_ports;
1757 if (field_avail(typeof(resp), cqe_version, udata->outlen))
1758 resp.response_length += sizeof(resp.cqe_version);
1760 if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1761 resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1762 MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1763 resp.response_length += sizeof(resp.cmds_supp_uhw);
1766 if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1767 if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1768 mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1769 resp.eth_min_inline++;
1771 resp.response_length += sizeof(resp.eth_min_inline);
1774 if (field_avail(typeof(resp), clock_info_versions, udata->outlen)) {
1775 if (mdev->clock_info)
1776 resp.clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1777 resp.response_length += sizeof(resp.clock_info_versions);
1781 * We don't want to expose information from the PCI bar that is located
1782 * after 4096 bytes, so if the arch only supports larger pages, let's
1783 * pretend we don't support reading the HCA's core clock. This is also
1784 * forced by mmap function.
1786 if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1787 if (PAGE_SIZE <= 4096) {
1789 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1790 resp.hca_core_clock_offset =
1791 offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1793 resp.response_length += sizeof(resp.hca_core_clock_offset);
1796 if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1797 resp.response_length += sizeof(resp.log_uar_size);
1799 if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1800 resp.response_length += sizeof(resp.num_uars_per_page);
1802 if (field_avail(typeof(resp), num_dyn_bfregs, udata->outlen)) {
1803 resp.num_dyn_bfregs = bfregi->num_dyn_bfregs;
1804 resp.response_length += sizeof(resp.num_dyn_bfregs);
1807 err = ib_copy_to_udata(udata, &resp, resp.response_length);
1812 bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1813 context->cqe_version = resp.cqe_version;
1814 context->lib_caps = req.lib_caps;
1815 print_lib_caps(dev, context->lib_caps);
1817 return &context->ibucontext;
1820 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1821 mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1824 deallocate_uars(dev, context);
1827 kfree(bfregi->sys_pages);
1830 kfree(bfregi->count);
1835 return ERR_PTR(err);
1838 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1840 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1841 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1842 struct mlx5_bfreg_info *bfregi;
1844 bfregi = &context->bfregi;
1845 if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1846 mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1848 deallocate_uars(dev, context);
1849 kfree(bfregi->sys_pages);
1850 kfree(bfregi->count);
1856 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1859 int fw_uars_per_page;
1861 fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1863 return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
1866 static int get_command(unsigned long offset)
1868 return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1871 static int get_arg(unsigned long offset)
1873 return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1876 static int get_index(unsigned long offset)
1878 return get_arg(offset);
1881 /* Index resides in an extra byte to enable larger values than 255 */
1882 static int get_extended_index(unsigned long offset)
1884 return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
1887 static void mlx5_ib_vma_open(struct vm_area_struct *area)
1889 /* vma_open is called when a new VMA is created on top of our VMA. This
1890 * is done through either mremap flow or split_vma (usually due to
1891 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1892 * as this VMA is strongly hardware related. Therefore we set the
1893 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1894 * calling us again and trying to do incorrect actions. We assume that
1895 * the original VMA size is exactly a single page, and therefore all
1896 * "splitting" operation will not happen to it.
1898 area->vm_ops = NULL;
1901 static void mlx5_ib_vma_close(struct vm_area_struct *area)
1903 struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1905 /* It's guaranteed that all VMAs opened on a FD are closed before the
1906 * file itself is closed, therefore no sync is needed with the regular
1907 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1908 * However need a sync with accessing the vma as part of
1909 * mlx5_ib_disassociate_ucontext.
1910 * The close operation is usually called under mm->mmap_sem except when
1911 * process is exiting.
1912 * The exiting case is handled explicitly as part of
1913 * mlx5_ib_disassociate_ucontext.
1915 mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1917 /* setting the vma context pointer to null in the mlx5_ib driver's
1918 * private data, to protect a race condition in
1919 * mlx5_ib_disassociate_ucontext().
1921 mlx5_ib_vma_priv_data->vma = NULL;
1922 mutex_lock(mlx5_ib_vma_priv_data->vma_private_list_mutex);
1923 list_del(&mlx5_ib_vma_priv_data->list);
1924 mutex_unlock(mlx5_ib_vma_priv_data->vma_private_list_mutex);
1925 kfree(mlx5_ib_vma_priv_data);
1928 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1929 .open = mlx5_ib_vma_open,
1930 .close = mlx5_ib_vma_close
1933 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1934 struct mlx5_ib_ucontext *ctx)
1936 struct mlx5_ib_vma_private_data *vma_prv;
1937 struct list_head *vma_head = &ctx->vma_private_list;
1939 vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1944 vma_prv->vma_private_list_mutex = &ctx->vma_private_list_mutex;
1945 vma->vm_private_data = vma_prv;
1946 vma->vm_ops = &mlx5_ib_vm_ops;
1948 mutex_lock(&ctx->vma_private_list_mutex);
1949 list_add(&vma_prv->list, vma_head);
1950 mutex_unlock(&ctx->vma_private_list_mutex);
1955 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1958 struct vm_area_struct *vma;
1959 struct mlx5_ib_vma_private_data *vma_private, *n;
1960 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1961 struct task_struct *owning_process = NULL;
1962 struct mm_struct *owning_mm = NULL;
1964 owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1965 if (!owning_process)
1968 owning_mm = get_task_mm(owning_process);
1970 pr_info("no mm, disassociate ucontext is pending task termination\n");
1972 put_task_struct(owning_process);
1973 usleep_range(1000, 2000);
1974 owning_process = get_pid_task(ibcontext->tgid,
1976 if (!owning_process ||
1977 owning_process->state == TASK_DEAD) {
1978 pr_info("disassociate ucontext done, task was terminated\n");
1979 /* in case task was dead need to release the
1983 put_task_struct(owning_process);
1989 /* need to protect from a race on closing the vma as part of
1990 * mlx5_ib_vma_close.
1992 down_write(&owning_mm->mmap_sem);
1993 mutex_lock(&context->vma_private_list_mutex);
1994 list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1996 vma = vma_private->vma;
1997 ret = zap_vma_ptes(vma, vma->vm_start,
1999 WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
2000 /* context going to be destroyed, should
2001 * not access ops any more.
2003 vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
2005 list_del(&vma_private->list);
2008 mutex_unlock(&context->vma_private_list_mutex);
2009 up_write(&owning_mm->mmap_sem);
2011 put_task_struct(owning_process);
2014 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2017 case MLX5_IB_MMAP_WC_PAGE:
2019 case MLX5_IB_MMAP_REGULAR_PAGE:
2020 return "best effort WC";
2021 case MLX5_IB_MMAP_NC_PAGE:
2023 case MLX5_IB_MMAP_DEVICE_MEM:
2024 return "Device Memory";
2030 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2031 struct vm_area_struct *vma,
2032 struct mlx5_ib_ucontext *context)
2037 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2040 if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2043 if (vma->vm_flags & VM_WRITE)
2046 if (!dev->mdev->clock_info_page)
2049 pfn = page_to_pfn(dev->mdev->clock_info_page);
2050 err = remap_pfn_range(vma, vma->vm_start, pfn, PAGE_SIZE,
2055 mlx5_ib_dbg(dev, "mapped clock info at 0x%lx, PA 0x%llx\n",
2057 (unsigned long long)pfn << PAGE_SHIFT);
2059 return mlx5_ib_set_vma_data(vma, context);
2062 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2063 struct vm_area_struct *vma,
2064 struct mlx5_ib_ucontext *context)
2066 struct mlx5_bfreg_info *bfregi = &context->bfregi;
2069 phys_addr_t pfn, pa;
2071 u32 bfreg_dyn_idx = 0;
2073 int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2074 int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2075 bfregi->num_static_sys_pages;
2077 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2081 idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2083 idx = get_index(vma->vm_pgoff);
2085 if (idx >= max_valid_idx) {
2086 mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2087 idx, max_valid_idx);
2092 case MLX5_IB_MMAP_WC_PAGE:
2093 case MLX5_IB_MMAP_ALLOC_WC:
2094 /* Some architectures don't support WC memory */
2095 #if defined(CONFIG_X86)
2098 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
2102 case MLX5_IB_MMAP_REGULAR_PAGE:
2103 /* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2104 prot = pgprot_writecombine(vma->vm_page_prot);
2106 case MLX5_IB_MMAP_NC_PAGE:
2107 prot = pgprot_noncached(vma->vm_page_prot);
2116 uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2117 bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2118 if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2119 mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2120 bfreg_dyn_idx, bfregi->total_num_bfregs);
2124 mutex_lock(&bfregi->lock);
2125 /* Fail if uar already allocated, first bfreg index of each
2126 * page holds its count.
2128 if (bfregi->count[bfreg_dyn_idx]) {
2129 mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2130 mutex_unlock(&bfregi->lock);
2134 bfregi->count[bfreg_dyn_idx]++;
2135 mutex_unlock(&bfregi->lock);
2137 err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2139 mlx5_ib_warn(dev, "UAR alloc failed\n");
2143 uar_index = bfregi->sys_pages[idx];
2146 pfn = uar_index2pfn(dev, uar_index);
2147 mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2149 vma->vm_page_prot = prot;
2150 err = io_remap_pfn_range(vma, vma->vm_start, pfn,
2151 PAGE_SIZE, vma->vm_page_prot);
2153 mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
2154 err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
2159 pa = pfn << PAGE_SHIFT;
2160 mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
2161 vma->vm_start, &pa);
2163 err = mlx5_ib_set_vma_data(vma, context);
2168 bfregi->sys_pages[idx] = uar_index;
2175 mlx5_cmd_free_uar(dev->mdev, idx);
2178 mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2183 static int dm_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
2185 struct mlx5_ib_ucontext *mctx = to_mucontext(context);
2186 struct mlx5_ib_dev *dev = to_mdev(context->device);
2187 u16 page_idx = get_extended_index(vma->vm_pgoff);
2188 size_t map_size = vma->vm_end - vma->vm_start;
2189 u32 npages = map_size >> PAGE_SHIFT;
2193 if (find_next_zero_bit(mctx->dm_pages, page_idx + npages, page_idx) !=
2197 pfn = ((pci_resource_start(dev->mdev->pdev, 0) +
2198 MLX5_CAP64_DEV_MEM(dev->mdev, memic_bar_start_addr)) >>
2201 prot = pgprot_writecombine(vma->vm_page_prot);
2202 vma->vm_page_prot = prot;
2204 if (io_remap_pfn_range(vma, vma->vm_start, pfn, map_size,
2208 return mlx5_ib_set_vma_data(vma, mctx);
2211 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2213 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2214 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2215 unsigned long command;
2218 command = get_command(vma->vm_pgoff);
2220 case MLX5_IB_MMAP_WC_PAGE:
2221 case MLX5_IB_MMAP_NC_PAGE:
2222 case MLX5_IB_MMAP_REGULAR_PAGE:
2223 case MLX5_IB_MMAP_ALLOC_WC:
2224 return uar_mmap(dev, command, vma, context);
2226 case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2229 case MLX5_IB_MMAP_CORE_CLOCK:
2230 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2233 if (vma->vm_flags & VM_WRITE)
2236 /* Don't expose to user-space information it shouldn't have */
2237 if (PAGE_SIZE > 4096)
2240 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
2241 pfn = (dev->mdev->iseg_base +
2242 offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2244 if (io_remap_pfn_range(vma, vma->vm_start, pfn,
2245 PAGE_SIZE, vma->vm_page_prot))
2248 mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
2250 (unsigned long long)pfn << PAGE_SHIFT);
2252 case MLX5_IB_MMAP_CLOCK_INFO:
2253 return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2255 case MLX5_IB_MMAP_DEVICE_MEM:
2256 return dm_mmap(ibcontext, vma);
2265 struct ib_dm *mlx5_ib_alloc_dm(struct ib_device *ibdev,
2266 struct ib_ucontext *context,
2267 struct ib_dm_alloc_attr *attr,
2268 struct uverbs_attr_bundle *attrs)
2270 u64 act_size = roundup(attr->length, MLX5_MEMIC_BASE_SIZE);
2271 struct mlx5_memic *memic = &to_mdev(ibdev)->memic;
2272 phys_addr_t memic_addr;
2273 struct mlx5_ib_dm *dm;
2278 dm = kzalloc(sizeof(*dm), GFP_KERNEL);
2280 return ERR_PTR(-ENOMEM);
2282 mlx5_ib_dbg(to_mdev(ibdev), "alloc_memic req: user_length=0x%llx act_length=0x%llx log_alignment=%d\n",
2283 attr->length, act_size, attr->alignment);
2285 err = mlx5_cmd_alloc_memic(memic, &memic_addr,
2286 act_size, attr->alignment);
2290 start_offset = memic_addr & ~PAGE_MASK;
2291 page_idx = (memic_addr - pci_resource_start(memic->dev->pdev, 0) -
2292 MLX5_CAP64_DEV_MEM(memic->dev, memic_bar_start_addr)) >>
2295 err = uverbs_copy_to(attrs,
2296 MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2297 &start_offset, sizeof(start_offset));
2301 err = uverbs_copy_to(attrs,
2302 MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
2303 &page_idx, sizeof(page_idx));
2307 bitmap_set(to_mucontext(context)->dm_pages, page_idx,
2308 DIV_ROUND_UP(act_size, PAGE_SIZE));
2310 dm->dev_addr = memic_addr;
2315 mlx5_cmd_dealloc_memic(memic, memic_addr,
2319 return ERR_PTR(err);
2322 int mlx5_ib_dealloc_dm(struct ib_dm *ibdm)
2324 struct mlx5_memic *memic = &to_mdev(ibdm->device)->memic;
2325 struct mlx5_ib_dm *dm = to_mdm(ibdm);
2326 u64 act_size = roundup(dm->ibdm.length, MLX5_MEMIC_BASE_SIZE);
2330 ret = mlx5_cmd_dealloc_memic(memic, dm->dev_addr, act_size);
2334 page_idx = (dm->dev_addr - pci_resource_start(memic->dev->pdev, 0) -
2335 MLX5_CAP64_DEV_MEM(memic->dev, memic_bar_start_addr)) >>
2337 bitmap_clear(to_mucontext(ibdm->uobject->context)->dm_pages,
2339 DIV_ROUND_UP(act_size, PAGE_SIZE));
2346 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
2347 struct ib_ucontext *context,
2348 struct ib_udata *udata)
2350 struct mlx5_ib_alloc_pd_resp resp;
2351 struct mlx5_ib_pd *pd;
2354 pd = kmalloc(sizeof(*pd), GFP_KERNEL);
2356 return ERR_PTR(-ENOMEM);
2358 err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
2361 return ERR_PTR(err);
2366 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2367 mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
2369 return ERR_PTR(-EFAULT);
2376 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
2378 struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2379 struct mlx5_ib_pd *mpd = to_mpd(pd);
2381 mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
2388 MATCH_CRITERIA_ENABLE_OUTER_BIT,
2389 MATCH_CRITERIA_ENABLE_MISC_BIT,
2390 MATCH_CRITERIA_ENABLE_INNER_BIT
2393 #define HEADER_IS_ZERO(match_criteria, headers) \
2394 !(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
2395 0, MLX5_FLD_SZ_BYTES(fte_match_param, headers))) \
2397 static u8 get_match_criteria_enable(u32 *match_criteria)
2399 u8 match_criteria_enable;
2401 match_criteria_enable =
2402 (!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
2403 MATCH_CRITERIA_ENABLE_OUTER_BIT;
2404 match_criteria_enable |=
2405 (!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
2406 MATCH_CRITERIA_ENABLE_MISC_BIT;
2407 match_criteria_enable |=
2408 (!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
2409 MATCH_CRITERIA_ENABLE_INNER_BIT;
2411 return match_criteria_enable;
2414 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
2416 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
2417 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
2420 static void set_flow_label(void *misc_c, void *misc_v, u8 mask, u8 val,
2424 MLX5_SET(fte_match_set_misc,
2425 misc_c, inner_ipv6_flow_label, mask);
2426 MLX5_SET(fte_match_set_misc,
2427 misc_v, inner_ipv6_flow_label, val);
2429 MLX5_SET(fte_match_set_misc,
2430 misc_c, outer_ipv6_flow_label, mask);
2431 MLX5_SET(fte_match_set_misc,
2432 misc_v, outer_ipv6_flow_label, val);
2436 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
2438 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
2439 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
2440 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
2441 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
2444 #define LAST_ETH_FIELD vlan_tag
2445 #define LAST_IB_FIELD sl
2446 #define LAST_IPV4_FIELD tos
2447 #define LAST_IPV6_FIELD traffic_class
2448 #define LAST_TCP_UDP_FIELD src_port
2449 #define LAST_TUNNEL_FIELD tunnel_id
2450 #define LAST_FLOW_TAG_FIELD tag_id
2451 #define LAST_DROP_FIELD size
2452 #define LAST_COUNTERS_FIELD counters
2454 /* Field is the last supported field */
2455 #define FIELDS_NOT_SUPPORTED(filter, field)\
2456 memchr_inv((void *)&filter.field +\
2457 sizeof(filter.field), 0,\
2459 offsetof(typeof(filter), field) -\
2460 sizeof(filter.field))
2462 static int parse_flow_flow_action(const union ib_flow_spec *ib_spec,
2463 const struct ib_flow_attr *flow_attr,
2464 struct mlx5_flow_act *action)
2466 struct mlx5_ib_flow_action *maction = to_mflow_act(ib_spec->action.act);
2468 switch (maction->ib_action.type) {
2469 case IB_FLOW_ACTION_ESP:
2470 /* Currently only AES_GCM keymat is supported by the driver */
2471 action->esp_id = (uintptr_t)maction->esp_aes_gcm.ctx;
2472 action->action |= flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS ?
2473 MLX5_FLOW_CONTEXT_ACTION_ENCRYPT :
2474 MLX5_FLOW_CONTEXT_ACTION_DECRYPT;
2481 static int parse_flow_attr(struct mlx5_core_dev *mdev, u32 *match_c,
2482 u32 *match_v, const union ib_flow_spec *ib_spec,
2483 const struct ib_flow_attr *flow_attr,
2484 struct mlx5_flow_act *action)
2486 void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
2488 void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
2495 if (ib_spec->type & IB_FLOW_SPEC_INNER) {
2496 headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2498 headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2500 match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2501 ft_field_support.inner_ip_version);
2503 headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2505 headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2507 match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2508 ft_field_support.outer_ip_version);
2511 switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2512 case IB_FLOW_SPEC_ETH:
2513 if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
2516 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2518 ib_spec->eth.mask.dst_mac);
2519 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2521 ib_spec->eth.val.dst_mac);
2523 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2525 ib_spec->eth.mask.src_mac);
2526 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2528 ib_spec->eth.val.src_mac);
2530 if (ib_spec->eth.mask.vlan_tag) {
2531 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2533 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2536 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2537 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
2538 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2539 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
2541 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2543 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
2544 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2546 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
2548 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2550 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
2551 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2553 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
2555 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2556 ethertype, ntohs(ib_spec->eth.mask.ether_type));
2557 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2558 ethertype, ntohs(ib_spec->eth.val.ether_type));
2560 case IB_FLOW_SPEC_IPV4:
2561 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
2565 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2567 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2568 ip_version, MLX5_FS_IPV4_VERSION);
2570 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2572 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2573 ethertype, ETH_P_IP);
2576 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2577 src_ipv4_src_ipv6.ipv4_layout.ipv4),
2578 &ib_spec->ipv4.mask.src_ip,
2579 sizeof(ib_spec->ipv4.mask.src_ip));
2580 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2581 src_ipv4_src_ipv6.ipv4_layout.ipv4),
2582 &ib_spec->ipv4.val.src_ip,
2583 sizeof(ib_spec->ipv4.val.src_ip));
2584 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2585 dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2586 &ib_spec->ipv4.mask.dst_ip,
2587 sizeof(ib_spec->ipv4.mask.dst_ip));
2588 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2589 dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2590 &ib_spec->ipv4.val.dst_ip,
2591 sizeof(ib_spec->ipv4.val.dst_ip));
2593 set_tos(headers_c, headers_v,
2594 ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
2596 set_proto(headers_c, headers_v,
2597 ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
2599 case IB_FLOW_SPEC_IPV6:
2600 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
2604 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2606 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2607 ip_version, MLX5_FS_IPV6_VERSION);
2609 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2611 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2612 ethertype, ETH_P_IPV6);
2615 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2616 src_ipv4_src_ipv6.ipv6_layout.ipv6),
2617 &ib_spec->ipv6.mask.src_ip,
2618 sizeof(ib_spec->ipv6.mask.src_ip));
2619 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2620 src_ipv4_src_ipv6.ipv6_layout.ipv6),
2621 &ib_spec->ipv6.val.src_ip,
2622 sizeof(ib_spec->ipv6.val.src_ip));
2623 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2624 dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2625 &ib_spec->ipv6.mask.dst_ip,
2626 sizeof(ib_spec->ipv6.mask.dst_ip));
2627 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2628 dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2629 &ib_spec->ipv6.val.dst_ip,
2630 sizeof(ib_spec->ipv6.val.dst_ip));
2632 set_tos(headers_c, headers_v,
2633 ib_spec->ipv6.mask.traffic_class,
2634 ib_spec->ipv6.val.traffic_class);
2636 set_proto(headers_c, headers_v,
2637 ib_spec->ipv6.mask.next_hdr,
2638 ib_spec->ipv6.val.next_hdr);
2640 set_flow_label(misc_params_c, misc_params_v,
2641 ntohl(ib_spec->ipv6.mask.flow_label),
2642 ntohl(ib_spec->ipv6.val.flow_label),
2643 ib_spec->type & IB_FLOW_SPEC_INNER);
2645 case IB_FLOW_SPEC_ESP:
2646 if (ib_spec->esp.mask.seq)
2649 MLX5_SET(fte_match_set_misc, misc_params_c, outer_esp_spi,
2650 ntohl(ib_spec->esp.mask.spi));
2651 MLX5_SET(fte_match_set_misc, misc_params_v, outer_esp_spi,
2652 ntohl(ib_spec->esp.val.spi));
2654 case IB_FLOW_SPEC_TCP:
2655 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2656 LAST_TCP_UDP_FIELD))
2659 MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2661 MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2664 MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
2665 ntohs(ib_spec->tcp_udp.mask.src_port));
2666 MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
2667 ntohs(ib_spec->tcp_udp.val.src_port));
2669 MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
2670 ntohs(ib_spec->tcp_udp.mask.dst_port));
2671 MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
2672 ntohs(ib_spec->tcp_udp.val.dst_port));
2674 case IB_FLOW_SPEC_UDP:
2675 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2676 LAST_TCP_UDP_FIELD))
2679 MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2681 MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2684 MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
2685 ntohs(ib_spec->tcp_udp.mask.src_port));
2686 MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
2687 ntohs(ib_spec->tcp_udp.val.src_port));
2689 MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
2690 ntohs(ib_spec->tcp_udp.mask.dst_port));
2691 MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
2692 ntohs(ib_spec->tcp_udp.val.dst_port));
2694 case IB_FLOW_SPEC_VXLAN_TUNNEL:
2695 if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
2699 MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
2700 ntohl(ib_spec->tunnel.mask.tunnel_id));
2701 MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
2702 ntohl(ib_spec->tunnel.val.tunnel_id));
2704 case IB_FLOW_SPEC_ACTION_TAG:
2705 if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
2706 LAST_FLOW_TAG_FIELD))
2708 if (ib_spec->flow_tag.tag_id >= BIT(24))
2711 action->flow_tag = ib_spec->flow_tag.tag_id;
2712 action->has_flow_tag = true;
2714 case IB_FLOW_SPEC_ACTION_DROP:
2715 if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
2718 action->action |= MLX5_FLOW_CONTEXT_ACTION_DROP;
2720 case IB_FLOW_SPEC_ACTION_HANDLE:
2721 ret = parse_flow_flow_action(ib_spec, flow_attr, action);
2725 case IB_FLOW_SPEC_ACTION_COUNT:
2726 if (FIELDS_NOT_SUPPORTED(ib_spec->flow_count,
2727 LAST_COUNTERS_FIELD))
2730 /* for now support only one counters spec per flow */
2731 if (action->action & MLX5_FLOW_CONTEXT_ACTION_COUNT)
2734 action->counters = ib_spec->flow_count.counters;
2735 action->action |= MLX5_FLOW_CONTEXT_ACTION_COUNT;
2744 /* If a flow could catch both multicast and unicast packets,
2745 * it won't fall into the multicast flow steering table and this rule
2746 * could steal other multicast packets.
2748 static bool flow_is_multicast_only(const struct ib_flow_attr *ib_attr)
2750 union ib_flow_spec *flow_spec;
2752 if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
2753 ib_attr->num_of_specs < 1)
2756 flow_spec = (union ib_flow_spec *)(ib_attr + 1);
2757 if (flow_spec->type == IB_FLOW_SPEC_IPV4) {
2758 struct ib_flow_spec_ipv4 *ipv4_spec;
2760 ipv4_spec = (struct ib_flow_spec_ipv4 *)flow_spec;
2761 if (ipv4_is_multicast(ipv4_spec->val.dst_ip))
2767 if (flow_spec->type == IB_FLOW_SPEC_ETH) {
2768 struct ib_flow_spec_eth *eth_spec;
2770 eth_spec = (struct ib_flow_spec_eth *)flow_spec;
2771 return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
2772 is_multicast_ether_addr(eth_spec->val.dst_mac);
2784 static enum valid_spec
2785 is_valid_esp_aes_gcm(struct mlx5_core_dev *mdev,
2786 const struct mlx5_flow_spec *spec,
2787 const struct mlx5_flow_act *flow_act,
2790 const u32 *match_c = spec->match_criteria;
2792 (flow_act->action & (MLX5_FLOW_CONTEXT_ACTION_ENCRYPT |
2793 MLX5_FLOW_CONTEXT_ACTION_DECRYPT));
2794 bool is_ipsec = mlx5_fs_is_ipsec_flow(match_c);
2795 bool is_drop = flow_act->action & MLX5_FLOW_CONTEXT_ACTION_DROP;
2798 * Currently only crypto is supported in egress, when regular egress
2799 * rules would be supported, always return VALID_SPEC_NA.
2802 return egress ? VALID_SPEC_INVALID : VALID_SPEC_NA;
2804 return is_crypto && is_ipsec &&
2805 (!egress || (!is_drop && !flow_act->has_flow_tag)) ?
2806 VALID_SPEC_VALID : VALID_SPEC_INVALID;
2809 static bool is_valid_spec(struct mlx5_core_dev *mdev,
2810 const struct mlx5_flow_spec *spec,
2811 const struct mlx5_flow_act *flow_act,
2814 /* We curretly only support ipsec egress flow */
2815 return is_valid_esp_aes_gcm(mdev, spec, flow_act, egress) != VALID_SPEC_INVALID;
2818 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
2819 const struct ib_flow_attr *flow_attr,
2822 union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
2823 int match_ipv = check_inner ?
2824 MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2825 ft_field_support.inner_ip_version) :
2826 MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2827 ft_field_support.outer_ip_version);
2828 int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
2829 bool ipv4_spec_valid, ipv6_spec_valid;
2830 unsigned int ip_spec_type = 0;
2831 bool has_ethertype = false;
2832 unsigned int spec_index;
2833 bool mask_valid = true;
2837 /* Validate that ethertype is correct */
2838 for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2839 if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
2840 ib_spec->eth.mask.ether_type) {
2841 mask_valid = (ib_spec->eth.mask.ether_type ==
2843 has_ethertype = true;
2844 eth_type = ntohs(ib_spec->eth.val.ether_type);
2845 } else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
2846 (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
2847 ip_spec_type = ib_spec->type;
2849 ib_spec = (void *)ib_spec + ib_spec->size;
2852 type_valid = (!has_ethertype) || (!ip_spec_type);
2853 if (!type_valid && mask_valid) {
2854 ipv4_spec_valid = (eth_type == ETH_P_IP) &&
2855 (ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
2856 ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
2857 (ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
2859 type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
2860 (((eth_type == ETH_P_MPLS_UC) ||
2861 (eth_type == ETH_P_MPLS_MC)) && match_ipv);
2867 static bool is_valid_attr(struct mlx5_core_dev *mdev,
2868 const struct ib_flow_attr *flow_attr)
2870 return is_valid_ethertype(mdev, flow_attr, false) &&
2871 is_valid_ethertype(mdev, flow_attr, true);
2874 static void put_flow_table(struct mlx5_ib_dev *dev,
2875 struct mlx5_ib_flow_prio *prio, bool ft_added)
2877 prio->refcount -= !!ft_added;
2878 if (!prio->refcount) {
2879 mlx5_destroy_flow_table(prio->flow_table);
2880 prio->flow_table = NULL;
2884 static void counters_clear_description(struct ib_counters *counters)
2886 struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
2888 mutex_lock(&mcounters->mcntrs_mutex);
2889 kfree(mcounters->counters_data);
2890 mcounters->counters_data = NULL;
2891 mcounters->cntrs_max_index = 0;
2892 mutex_unlock(&mcounters->mcntrs_mutex);
2895 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
2897 struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
2898 struct mlx5_ib_flow_handler *handler = container_of(flow_id,
2899 struct mlx5_ib_flow_handler,
2901 struct mlx5_ib_flow_handler *iter, *tmp;
2903 mutex_lock(&dev->flow_db->lock);
2905 list_for_each_entry_safe(iter, tmp, &handler->list, list) {
2906 mlx5_del_flow_rules(iter->rule);
2907 put_flow_table(dev, iter->prio, true);
2908 list_del(&iter->list);
2912 mlx5_del_flow_rules(handler->rule);
2913 put_flow_table(dev, handler->prio, true);
2914 if (handler->ibcounters &&
2915 atomic_read(&handler->ibcounters->usecnt) == 1)
2916 counters_clear_description(handler->ibcounters);
2918 mutex_unlock(&dev->flow_db->lock);
2924 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
2932 enum flow_table_type {
2937 #define MLX5_FS_MAX_TYPES 6
2938 #define MLX5_FS_MAX_ENTRIES BIT(16)
2939 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
2940 struct ib_flow_attr *flow_attr,
2941 enum flow_table_type ft_type)
2943 bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
2944 struct mlx5_flow_namespace *ns = NULL;
2945 struct mlx5_ib_flow_prio *prio;
2946 struct mlx5_flow_table *ft;
2953 max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
2955 if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2956 if (ft_type == MLX5_IB_FT_TX)
2958 else if (flow_is_multicast_only(flow_attr) &&
2960 priority = MLX5_IB_FLOW_MCAST_PRIO;
2962 priority = ib_prio_to_core_prio(flow_attr->priority,
2964 ns = mlx5_get_flow_namespace(dev->mdev,
2965 ft_type == MLX5_IB_FT_TX ?
2966 MLX5_FLOW_NAMESPACE_EGRESS :
2967 MLX5_FLOW_NAMESPACE_BYPASS);
2968 num_entries = MLX5_FS_MAX_ENTRIES;
2969 num_groups = MLX5_FS_MAX_TYPES;
2970 prio = &dev->flow_db->prios[priority];
2971 } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2972 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2973 ns = mlx5_get_flow_namespace(dev->mdev,
2974 MLX5_FLOW_NAMESPACE_LEFTOVERS);
2975 build_leftovers_ft_param(&priority,
2978 prio = &dev->flow_db->prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
2979 } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2980 if (!MLX5_CAP_FLOWTABLE(dev->mdev,
2981 allow_sniffer_and_nic_rx_shared_tir))
2982 return ERR_PTR(-ENOTSUPP);
2984 ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
2985 MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2986 MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2988 prio = &dev->flow_db->sniffer[ft_type];
2995 return ERR_PTR(-ENOTSUPP);
2997 if (num_entries > max_table_size)
2998 return ERR_PTR(-ENOMEM);
3000 ft = prio->flow_table;
3002 ft = mlx5_create_auto_grouped_flow_table(ns, priority,
3009 prio->flow_table = ft;
3015 return err ? ERR_PTR(err) : prio;
3018 static void set_underlay_qp(struct mlx5_ib_dev *dev,
3019 struct mlx5_flow_spec *spec,
3022 void *misc_params_c = MLX5_ADDR_OF(fte_match_param,
3023 spec->match_criteria,
3025 void *misc_params_v = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3029 MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3030 ft_field_support.bth_dst_qp)) {
3031 MLX5_SET(fte_match_set_misc,
3032 misc_params_v, bth_dst_qp, underlay_qpn);
3033 MLX5_SET(fte_match_set_misc,
3034 misc_params_c, bth_dst_qp, 0xffffff);
3038 static int counters_set_description(struct ib_counters *counters,
3039 enum mlx5_ib_counters_type counters_type,
3040 struct mlx5_ib_flow_counters_desc *desc_data,
3043 struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
3044 u32 cntrs_max_index = 0;
3047 if (counters_type != MLX5_IB_COUNTERS_FLOW)
3050 /* init the fields for the object */
3051 mcounters->type = counters_type;
3052 mcounters->ncounters = ncounters;
3053 /* each counter entry have both description and index pair */
3054 for (i = 0; i < ncounters; i++) {
3055 if (desc_data[i].description > IB_COUNTER_BYTES)
3058 if (cntrs_max_index <= desc_data[i].index)
3059 cntrs_max_index = desc_data[i].index + 1;
3062 mutex_lock(&mcounters->mcntrs_mutex);
3063 mcounters->counters_data = desc_data;
3064 mcounters->cntrs_max_index = cntrs_max_index;
3065 mutex_unlock(&mcounters->mcntrs_mutex);
3070 #define MAX_COUNTERS_NUM (USHRT_MAX / (sizeof(u32) * 2))
3071 static int flow_counters_set_data(struct ib_counters *ibcounters,
3072 struct mlx5_ib_create_flow *ucmd)
3074 struct mlx5_ib_mcounters *mcounters = to_mcounters(ibcounters);
3075 struct mlx5_ib_flow_counters_data *cntrs_data = NULL;
3076 struct mlx5_ib_flow_counters_desc *desc_data = NULL;
3077 bool hw_hndl = false;
3080 if (ucmd && ucmd->ncounters_data != 0) {
3081 cntrs_data = ucmd->data;
3082 if (cntrs_data->ncounters > MAX_COUNTERS_NUM)
3085 desc_data = kcalloc(cntrs_data->ncounters,
3091 if (copy_from_user(desc_data,
3092 u64_to_user_ptr(cntrs_data->counters_data),
3093 sizeof(*desc_data) * cntrs_data->ncounters)) {
3099 if (!mcounters->hw_cntrs_hndl) {
3100 mcounters->hw_cntrs_hndl = mlx5_fc_create(
3101 to_mdev(ibcounters->device)->mdev, false);
3102 if (!mcounters->hw_cntrs_hndl) {
3110 /* counters already bound to at least one flow */
3111 if (mcounters->cntrs_max_index) {
3116 ret = counters_set_description(ibcounters,
3117 MLX5_IB_COUNTERS_FLOW,
3119 cntrs_data->ncounters);
3123 } else if (!mcounters->cntrs_max_index) {
3124 /* counters not bound yet, must have udata passed */
3133 mlx5_fc_destroy(to_mdev(ibcounters->device)->mdev,
3134 mcounters->hw_cntrs_hndl);
3135 mcounters->hw_cntrs_hndl = NULL;
3142 static struct mlx5_ib_flow_handler *_create_flow_rule(struct mlx5_ib_dev *dev,
3143 struct mlx5_ib_flow_prio *ft_prio,
3144 const struct ib_flow_attr *flow_attr,
3145 struct mlx5_flow_destination *dst,
3147 struct mlx5_ib_create_flow *ucmd)
3149 struct mlx5_flow_table *ft = ft_prio->flow_table;
3150 struct mlx5_ib_flow_handler *handler;
3151 struct mlx5_flow_act flow_act = {.flow_tag = MLX5_FS_DEFAULT_FLOW_TAG};
3152 struct mlx5_flow_spec *spec;
3153 struct mlx5_flow_destination dest_arr[2] = {};
3154 struct mlx5_flow_destination *rule_dst = dest_arr;
3155 const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
3156 unsigned int spec_index;
3159 bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3161 if (!is_valid_attr(dev->mdev, flow_attr))
3162 return ERR_PTR(-EINVAL);
3164 spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
3165 handler = kzalloc(sizeof(*handler), GFP_KERNEL);
3166 if (!handler || !spec) {
3171 INIT_LIST_HEAD(&handler->list);
3173 memcpy(&dest_arr[0], dst, sizeof(*dst));
3177 for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
3178 err = parse_flow_attr(dev->mdev, spec->match_criteria,
3180 ib_flow, flow_attr, &flow_act);
3184 ib_flow += ((union ib_flow_spec *)ib_flow)->size;
3187 if (!flow_is_multicast_only(flow_attr))
3188 set_underlay_qp(dev, spec, underlay_qpn);
3193 misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3195 MLX5_SET(fte_match_set_misc, misc, source_port,
3197 misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
3199 MLX5_SET_TO_ONES(fte_match_set_misc, misc, source_port);
3202 spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
3205 !is_valid_spec(dev->mdev, spec, &flow_act, is_egress)) {
3210 if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_COUNT) {
3211 err = flow_counters_set_data(flow_act.counters, ucmd);
3215 handler->ibcounters = flow_act.counters;
3216 dest_arr[dest_num].type =
3217 MLX5_FLOW_DESTINATION_TYPE_COUNTER;
3218 dest_arr[dest_num].counter =
3219 to_mcounters(flow_act.counters)->hw_cntrs_hndl;
3223 if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_DROP) {
3224 if (!(flow_act.action & MLX5_FLOW_CONTEXT_ACTION_COUNT)) {
3230 flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_ALLOW;
3233 dest_num ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
3234 MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
3237 if (flow_act.has_flow_tag &&
3238 (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3239 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
3240 mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
3241 flow_act.flow_tag, flow_attr->type);
3245 handler->rule = mlx5_add_flow_rules(ft, spec,
3247 rule_dst, dest_num);
3249 if (IS_ERR(handler->rule)) {
3250 err = PTR_ERR(handler->rule);
3254 ft_prio->refcount++;
3255 handler->prio = ft_prio;
3257 ft_prio->flow_table = ft;
3259 if (err && handler) {
3260 if (handler->ibcounters &&
3261 atomic_read(&handler->ibcounters->usecnt) == 1)
3262 counters_clear_description(handler->ibcounters);
3266 return err ? ERR_PTR(err) : handler;
3269 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
3270 struct mlx5_ib_flow_prio *ft_prio,
3271 const struct ib_flow_attr *flow_attr,
3272 struct mlx5_flow_destination *dst)
3274 return _create_flow_rule(dev, ft_prio, flow_attr, dst, 0, NULL);
3277 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
3278 struct mlx5_ib_flow_prio *ft_prio,
3279 struct ib_flow_attr *flow_attr,
3280 struct mlx5_flow_destination *dst)
3282 struct mlx5_ib_flow_handler *handler_dst = NULL;
3283 struct mlx5_ib_flow_handler *handler = NULL;
3285 handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
3286 if (!IS_ERR(handler)) {
3287 handler_dst = create_flow_rule(dev, ft_prio,
3289 if (IS_ERR(handler_dst)) {
3290 mlx5_del_flow_rules(handler->rule);
3291 ft_prio->refcount--;
3293 handler = handler_dst;
3295 list_add(&handler_dst->list, &handler->list);
3306 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
3307 struct mlx5_ib_flow_prio *ft_prio,
3308 struct ib_flow_attr *flow_attr,
3309 struct mlx5_flow_destination *dst)
3311 struct mlx5_ib_flow_handler *handler_ucast = NULL;
3312 struct mlx5_ib_flow_handler *handler = NULL;
3315 struct ib_flow_attr flow_attr;
3316 struct ib_flow_spec_eth eth_flow;
3317 } leftovers_specs[] = {
3321 .size = sizeof(leftovers_specs[0])
3324 .type = IB_FLOW_SPEC_ETH,
3325 .size = sizeof(struct ib_flow_spec_eth),
3326 .mask = {.dst_mac = {0x1} },
3327 .val = {.dst_mac = {0x1} }
3333 .size = sizeof(leftovers_specs[0])
3336 .type = IB_FLOW_SPEC_ETH,
3337 .size = sizeof(struct ib_flow_spec_eth),
3338 .mask = {.dst_mac = {0x1} },
3339 .val = {.dst_mac = {} }
3344 handler = create_flow_rule(dev, ft_prio,
3345 &leftovers_specs[LEFTOVERS_MC].flow_attr,
3347 if (!IS_ERR(handler) &&
3348 flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
3349 handler_ucast = create_flow_rule(dev, ft_prio,
3350 &leftovers_specs[LEFTOVERS_UC].flow_attr,
3352 if (IS_ERR(handler_ucast)) {
3353 mlx5_del_flow_rules(handler->rule);
3354 ft_prio->refcount--;
3356 handler = handler_ucast;
3358 list_add(&handler_ucast->list, &handler->list);
3365 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
3366 struct mlx5_ib_flow_prio *ft_rx,
3367 struct mlx5_ib_flow_prio *ft_tx,
3368 struct mlx5_flow_destination *dst)
3370 struct mlx5_ib_flow_handler *handler_rx;
3371 struct mlx5_ib_flow_handler *handler_tx;
3373 static const struct ib_flow_attr flow_attr = {
3375 .size = sizeof(flow_attr)
3378 handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
3379 if (IS_ERR(handler_rx)) {
3380 err = PTR_ERR(handler_rx);
3384 handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
3385 if (IS_ERR(handler_tx)) {
3386 err = PTR_ERR(handler_tx);
3390 list_add(&handler_tx->list, &handler_rx->list);
3395 mlx5_del_flow_rules(handler_rx->rule);
3399 return ERR_PTR(err);
3402 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
3403 struct ib_flow_attr *flow_attr,
3405 struct ib_udata *udata)
3407 struct mlx5_ib_dev *dev = to_mdev(qp->device);
3408 struct mlx5_ib_qp *mqp = to_mqp(qp);
3409 struct mlx5_ib_flow_handler *handler = NULL;
3410 struct mlx5_flow_destination *dst = NULL;
3411 struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
3412 struct mlx5_ib_flow_prio *ft_prio;
3413 bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3414 struct mlx5_ib_create_flow *ucmd = NULL, ucmd_hdr;
3415 size_t min_ucmd_sz, required_ucmd_sz;
3419 if (udata && udata->inlen) {
3420 min_ucmd_sz = offsetof(typeof(ucmd_hdr), reserved) +
3421 sizeof(ucmd_hdr.reserved);
3422 if (udata->inlen < min_ucmd_sz)
3423 return ERR_PTR(-EOPNOTSUPP);
3425 err = ib_copy_from_udata(&ucmd_hdr, udata, min_ucmd_sz);
3427 return ERR_PTR(err);
3429 /* currently supports only one counters data */
3430 if (ucmd_hdr.ncounters_data > 1)
3431 return ERR_PTR(-EINVAL);
3433 required_ucmd_sz = min_ucmd_sz +
3434 sizeof(struct mlx5_ib_flow_counters_data) *
3435 ucmd_hdr.ncounters_data;
3436 if (udata->inlen > required_ucmd_sz &&
3437 !ib_is_udata_cleared(udata, required_ucmd_sz,
3438 udata->inlen - required_ucmd_sz))
3439 return ERR_PTR(-EOPNOTSUPP);
3441 ucmd = kzalloc(required_ucmd_sz, GFP_KERNEL);
3443 return ERR_PTR(-ENOMEM);
3445 err = ib_copy_from_udata(ucmd, udata, required_ucmd_sz);
3448 return ERR_PTR(err);
3452 if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
3453 return ERR_PTR(-ENOMEM);
3455 if (domain != IB_FLOW_DOMAIN_USER ||
3456 flow_attr->port > dev->num_ports ||
3457 (flow_attr->flags & ~(IB_FLOW_ATTR_FLAGS_DONT_TRAP |
3458 IB_FLOW_ATTR_FLAGS_EGRESS)))
3459 return ERR_PTR(-EINVAL);
3462 (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3463 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT))
3464 return ERR_PTR(-EINVAL);
3466 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
3468 return ERR_PTR(-ENOMEM);
3470 mutex_lock(&dev->flow_db->lock);
3472 ft_prio = get_flow_table(dev, flow_attr,
3473 is_egress ? MLX5_IB_FT_TX : MLX5_IB_FT_RX);
3474 if (IS_ERR(ft_prio)) {
3475 err = PTR_ERR(ft_prio);
3478 if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3479 ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
3480 if (IS_ERR(ft_prio_tx)) {
3481 err = PTR_ERR(ft_prio_tx);
3488 dst->type = MLX5_FLOW_DESTINATION_TYPE_PORT;
3490 dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
3491 if (mqp->flags & MLX5_IB_QP_RSS)
3492 dst->tir_num = mqp->rss_qp.tirn;
3494 dst->tir_num = mqp->raw_packet_qp.rq.tirn;
3497 if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3498 if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) {
3499 handler = create_dont_trap_rule(dev, ft_prio,
3502 underlay_qpn = (mqp->flags & MLX5_IB_QP_UNDERLAY) ?
3503 mqp->underlay_qpn : 0;
3504 handler = _create_flow_rule(dev, ft_prio, flow_attr,
3505 dst, underlay_qpn, ucmd);
3507 } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3508 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3509 handler = create_leftovers_rule(dev, ft_prio, flow_attr,
3511 } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3512 handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
3518 if (IS_ERR(handler)) {
3519 err = PTR_ERR(handler);
3524 mutex_unlock(&dev->flow_db->lock);
3528 return &handler->ibflow;
3531 put_flow_table(dev, ft_prio, false);
3533 put_flow_table(dev, ft_prio_tx, false);
3535 mutex_unlock(&dev->flow_db->lock);
3539 return ERR_PTR(err);
3542 static u32 mlx5_ib_flow_action_flags_to_accel_xfrm_flags(u32 mlx5_flags)
3546 if (mlx5_flags & MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA)
3547 flags |= MLX5_ACCEL_XFRM_FLAG_REQUIRE_METADATA;
3552 #define MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA
3553 static struct ib_flow_action *
3554 mlx5_ib_create_flow_action_esp(struct ib_device *device,
3555 const struct ib_flow_action_attrs_esp *attr,
3556 struct uverbs_attr_bundle *attrs)
3558 struct mlx5_ib_dev *mdev = to_mdev(device);
3559 struct ib_uverbs_flow_action_esp_keymat_aes_gcm *aes_gcm;
3560 struct mlx5_accel_esp_xfrm_attrs accel_attrs = {};
3561 struct mlx5_ib_flow_action *action;
3566 if (IS_UVERBS_COPY_ERR(uverbs_copy_from(&action_flags, attrs,
3567 MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS)))
3568 return ERR_PTR(-EFAULT);
3570 if (action_flags >= (MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED << 1))
3571 return ERR_PTR(-EOPNOTSUPP);
3573 flags = mlx5_ib_flow_action_flags_to_accel_xfrm_flags(action_flags);
3575 /* We current only support a subset of the standard features. Only a
3576 * keymat of type AES_GCM, with icv_len == 16, iv_algo == SEQ and esn
3577 * (with overlap). Full offload mode isn't supported.
3579 if (!attr->keymat || attr->replay || attr->encap ||
3580 attr->spi || attr->seq || attr->tfc_pad ||
3581 attr->hard_limit_pkts ||
3582 (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3583 IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)))
3584 return ERR_PTR(-EOPNOTSUPP);
3586 if (attr->keymat->protocol !=
3587 IB_UVERBS_FLOW_ACTION_ESP_KEYMAT_AES_GCM)
3588 return ERR_PTR(-EOPNOTSUPP);
3590 aes_gcm = &attr->keymat->keymat.aes_gcm;
3592 if (aes_gcm->icv_len != 16 ||
3593 aes_gcm->iv_algo != IB_UVERBS_FLOW_ACTION_IV_ALGO_SEQ)
3594 return ERR_PTR(-EOPNOTSUPP);
3596 action = kmalloc(sizeof(*action), GFP_KERNEL);
3598 return ERR_PTR(-ENOMEM);
3600 action->esp_aes_gcm.ib_flags = attr->flags;
3601 memcpy(&accel_attrs.keymat.aes_gcm.aes_key, &aes_gcm->aes_key,
3602 sizeof(accel_attrs.keymat.aes_gcm.aes_key));
3603 accel_attrs.keymat.aes_gcm.key_len = aes_gcm->key_len * 8;
3604 memcpy(&accel_attrs.keymat.aes_gcm.salt, &aes_gcm->salt,
3605 sizeof(accel_attrs.keymat.aes_gcm.salt));
3606 memcpy(&accel_attrs.keymat.aes_gcm.seq_iv, &aes_gcm->iv,
3607 sizeof(accel_attrs.keymat.aes_gcm.seq_iv));
3608 accel_attrs.keymat.aes_gcm.icv_len = aes_gcm->icv_len * 8;
3609 accel_attrs.keymat.aes_gcm.iv_algo = MLX5_ACCEL_ESP_AES_GCM_IV_ALGO_SEQ;
3610 accel_attrs.keymat_type = MLX5_ACCEL_ESP_KEYMAT_AES_GCM;
3612 accel_attrs.esn = attr->esn;
3613 if (attr->flags & IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED)
3614 accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_TRIGGERED;
3615 if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
3616 accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3618 if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)
3619 accel_attrs.action |= MLX5_ACCEL_ESP_ACTION_ENCRYPT;
3621 action->esp_aes_gcm.ctx =
3622 mlx5_accel_esp_create_xfrm(mdev->mdev, &accel_attrs, flags);
3623 if (IS_ERR(action->esp_aes_gcm.ctx)) {
3624 err = PTR_ERR(action->esp_aes_gcm.ctx);
3628 action->esp_aes_gcm.ib_flags = attr->flags;
3630 return &action->ib_action;
3634 return ERR_PTR(err);
3638 mlx5_ib_modify_flow_action_esp(struct ib_flow_action *action,
3639 const struct ib_flow_action_attrs_esp *attr,
3640 struct uverbs_attr_bundle *attrs)
3642 struct mlx5_ib_flow_action *maction = to_mflow_act(action);
3643 struct mlx5_accel_esp_xfrm_attrs accel_attrs;
3646 if (attr->keymat || attr->replay || attr->encap ||
3647 attr->spi || attr->seq || attr->tfc_pad ||
3648 attr->hard_limit_pkts ||
3649 (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3650 IB_FLOW_ACTION_ESP_FLAGS_MOD_ESP_ATTRS |
3651 IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)))
3654 /* Only the ESN value or the MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP can
3657 if (!(maction->esp_aes_gcm.ib_flags &
3658 IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED) &&
3659 attr->flags & (IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3660 IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW))
3663 memcpy(&accel_attrs, &maction->esp_aes_gcm.ctx->attrs,
3664 sizeof(accel_attrs));
3666 accel_attrs.esn = attr->esn;
3667 if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
3668 accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3670 accel_attrs.flags &= ~MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3672 err = mlx5_accel_esp_modify_xfrm(maction->esp_aes_gcm.ctx,
3677 maction->esp_aes_gcm.ib_flags &=
3678 ~IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
3679 maction->esp_aes_gcm.ib_flags |=
3680 attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
3685 static int mlx5_ib_destroy_flow_action(struct ib_flow_action *action)
3687 struct mlx5_ib_flow_action *maction = to_mflow_act(action);
3689 switch (action->type) {
3690 case IB_FLOW_ACTION_ESP:
3692 * We only support aes_gcm by now, so we implicitly know this is
3693 * the underline crypto.
3695 mlx5_accel_esp_destroy_xfrm(maction->esp_aes_gcm.ctx);
3706 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
3708 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3709 struct mlx5_ib_qp *mqp = to_mqp(ibqp);
3712 if (mqp->flags & MLX5_IB_QP_UNDERLAY) {
3713 mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
3717 err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
3719 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
3720 ibqp->qp_num, gid->raw);
3725 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
3727 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3730 err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
3732 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
3733 ibqp->qp_num, gid->raw);
3738 static int init_node_data(struct mlx5_ib_dev *dev)
3742 err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
3746 dev->mdev->rev_id = dev->mdev->pdev->revision;
3748 return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
3751 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
3754 struct mlx5_ib_dev *dev =
3755 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3757 return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
3760 static ssize_t show_reg_pages(struct device *device,
3761 struct device_attribute *attr, char *buf)
3763 struct mlx5_ib_dev *dev =
3764 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3766 return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
3769 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
3772 struct mlx5_ib_dev *dev =
3773 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3774 return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
3777 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
3780 struct mlx5_ib_dev *dev =
3781 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3782 return sprintf(buf, "%x\n", dev->mdev->rev_id);
3785 static ssize_t show_board(struct device *device, struct device_attribute *attr,
3788 struct mlx5_ib_dev *dev =
3789 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3790 return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
3791 dev->mdev->board_id);
3794 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
3795 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
3796 static DEVICE_ATTR(board_id, S_IRUGO, show_board, NULL);
3797 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
3798 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
3800 static struct device_attribute *mlx5_class_attributes[] = {
3805 &dev_attr_reg_pages,
3808 static void pkey_change_handler(struct work_struct *work)
3810 struct mlx5_ib_port_resources *ports =
3811 container_of(work, struct mlx5_ib_port_resources,
3814 mutex_lock(&ports->devr->mutex);
3815 mlx5_ib_gsi_pkey_change(ports->gsi);
3816 mutex_unlock(&ports->devr->mutex);
3819 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
3821 struct mlx5_ib_qp *mqp;
3822 struct mlx5_ib_cq *send_mcq, *recv_mcq;
3823 struct mlx5_core_cq *mcq;
3824 struct list_head cq_armed_list;
3825 unsigned long flags_qp;
3826 unsigned long flags_cq;
3827 unsigned long flags;
3829 INIT_LIST_HEAD(&cq_armed_list);
3831 /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
3832 spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
3833 list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
3834 spin_lock_irqsave(&mqp->sq.lock, flags_qp);
3835 if (mqp->sq.tail != mqp->sq.head) {
3836 send_mcq = to_mcq(mqp->ibqp.send_cq);
3837 spin_lock_irqsave(&send_mcq->lock, flags_cq);
3838 if (send_mcq->mcq.comp &&
3839 mqp->ibqp.send_cq->comp_handler) {
3840 if (!send_mcq->mcq.reset_notify_added) {
3841 send_mcq->mcq.reset_notify_added = 1;
3842 list_add_tail(&send_mcq->mcq.reset_notify,
3846 spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
3848 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
3849 spin_lock_irqsave(&mqp->rq.lock, flags_qp);
3850 /* no handling is needed for SRQ */
3851 if (!mqp->ibqp.srq) {
3852 if (mqp->rq.tail != mqp->rq.head) {
3853 recv_mcq = to_mcq(mqp->ibqp.recv_cq);
3854 spin_lock_irqsave(&recv_mcq->lock, flags_cq);
3855 if (recv_mcq->mcq.comp &&
3856 mqp->ibqp.recv_cq->comp_handler) {
3857 if (!recv_mcq->mcq.reset_notify_added) {
3858 recv_mcq->mcq.reset_notify_added = 1;
3859 list_add_tail(&recv_mcq->mcq.reset_notify,
3863 spin_unlock_irqrestore(&recv_mcq->lock,
3867 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
3869 /*At that point all inflight post send were put to be executed as of we
3870 * lock/unlock above locks Now need to arm all involved CQs.
3872 list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
3875 spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
3878 static void delay_drop_handler(struct work_struct *work)
3881 struct mlx5_ib_delay_drop *delay_drop =
3882 container_of(work, struct mlx5_ib_delay_drop,
3885 atomic_inc(&delay_drop->events_cnt);
3887 mutex_lock(&delay_drop->lock);
3888 err = mlx5_core_set_delay_drop(delay_drop->dev->mdev,
3889 delay_drop->timeout);
3891 mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
3892 delay_drop->timeout);
3893 delay_drop->activate = false;
3895 mutex_unlock(&delay_drop->lock);
3898 static void mlx5_ib_handle_event(struct work_struct *_work)
3900 struct mlx5_ib_event_work *work =
3901 container_of(_work, struct mlx5_ib_event_work, work);
3902 struct mlx5_ib_dev *ibdev;
3903 struct ib_event ibev;
3905 u8 port = (u8)work->param;
3907 if (mlx5_core_is_mp_slave(work->dev)) {
3908 ibdev = mlx5_ib_get_ibdev_from_mpi(work->context);
3912 ibdev = work->context;
3915 switch (work->event) {
3916 case MLX5_DEV_EVENT_SYS_ERROR:
3917 ibev.event = IB_EVENT_DEVICE_FATAL;
3918 mlx5_ib_handle_internal_error(ibdev);
3922 case MLX5_DEV_EVENT_PORT_UP:
3923 case MLX5_DEV_EVENT_PORT_DOWN:
3924 case MLX5_DEV_EVENT_PORT_INITIALIZED:
3925 /* In RoCE, port up/down events are handled in
3926 * mlx5_netdev_event().
3928 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
3929 IB_LINK_LAYER_ETHERNET)
3932 ibev.event = (work->event == MLX5_DEV_EVENT_PORT_UP) ?
3933 IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
3936 case MLX5_DEV_EVENT_LID_CHANGE:
3937 ibev.event = IB_EVENT_LID_CHANGE;
3940 case MLX5_DEV_EVENT_PKEY_CHANGE:
3941 ibev.event = IB_EVENT_PKEY_CHANGE;
3942 schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
3945 case MLX5_DEV_EVENT_GUID_CHANGE:
3946 ibev.event = IB_EVENT_GID_CHANGE;
3949 case MLX5_DEV_EVENT_CLIENT_REREG:
3950 ibev.event = IB_EVENT_CLIENT_REREGISTER;
3952 case MLX5_DEV_EVENT_DELAY_DROP_TIMEOUT:
3953 schedule_work(&ibdev->delay_drop.delay_drop_work);
3959 ibev.device = &ibdev->ib_dev;
3960 ibev.element.port_num = port;
3962 if (!rdma_is_port_valid(&ibdev->ib_dev, port)) {
3963 mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
3967 if (ibdev->ib_active)
3968 ib_dispatch_event(&ibev);
3971 ibdev->ib_active = false;
3976 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
3977 enum mlx5_dev_event event, unsigned long param)
3979 struct mlx5_ib_event_work *work;
3981 work = kmalloc(sizeof(*work), GFP_ATOMIC);
3985 INIT_WORK(&work->work, mlx5_ib_handle_event);
3987 work->param = param;
3988 work->context = context;
3989 work->event = event;
3991 queue_work(mlx5_ib_event_wq, &work->work);
3994 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
3996 struct mlx5_hca_vport_context vport_ctx;
4000 for (port = 1; port <= dev->num_ports; port++) {
4001 dev->mdev->port_caps[port - 1].has_smi = false;
4002 if (MLX5_CAP_GEN(dev->mdev, port_type) ==
4003 MLX5_CAP_PORT_TYPE_IB) {
4004 if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
4005 err = mlx5_query_hca_vport_context(dev->mdev, 0,
4009 mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
4013 dev->mdev->port_caps[port - 1].has_smi =
4016 dev->mdev->port_caps[port - 1].has_smi = true;
4023 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
4027 for (port = 1; port <= dev->num_ports; port++)
4028 mlx5_query_ext_port_caps(dev, port);
4031 static int get_port_caps(struct mlx5_ib_dev *dev, u8 port)
4033 struct ib_device_attr *dprops = NULL;
4034 struct ib_port_attr *pprops = NULL;
4036 struct ib_udata uhw = {.inlen = 0, .outlen = 0};
4038 pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
4042 dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
4046 err = set_has_smi_cap(dev);
4050 err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
4052 mlx5_ib_warn(dev, "query_device failed %d\n", err);
4056 memset(pprops, 0, sizeof(*pprops));
4057 err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
4059 mlx5_ib_warn(dev, "query_port %d failed %d\n",
4064 dev->mdev->port_caps[port - 1].pkey_table_len =
4066 dev->mdev->port_caps[port - 1].gid_table_len =
4067 pprops->gid_tbl_len;
4068 mlx5_ib_dbg(dev, "port %d: pkey_table_len %d, gid_table_len %d\n",
4069 port, dprops->max_pkeys, pprops->gid_tbl_len);
4078 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
4082 err = mlx5_mr_cache_cleanup(dev);
4084 mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4087 mlx5_ib_destroy_qp(dev->umrc.qp);
4089 ib_free_cq(dev->umrc.cq);
4091 ib_dealloc_pd(dev->umrc.pd);
4098 static int create_umr_res(struct mlx5_ib_dev *dev)
4100 struct ib_qp_init_attr *init_attr = NULL;
4101 struct ib_qp_attr *attr = NULL;
4107 attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4108 init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4109 if (!attr || !init_attr) {
4114 pd = ib_alloc_pd(&dev->ib_dev, 0);
4116 mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4121 cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4123 mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4128 init_attr->send_cq = cq;
4129 init_attr->recv_cq = cq;
4130 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4131 init_attr->cap.max_send_wr = MAX_UMR_WR;
4132 init_attr->cap.max_send_sge = 1;
4133 init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4134 init_attr->port_num = 1;
4135 qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4137 mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4141 qp->device = &dev->ib_dev;
4144 qp->qp_type = MLX5_IB_QPT_REG_UMR;
4145 qp->send_cq = init_attr->send_cq;
4146 qp->recv_cq = init_attr->recv_cq;
4148 attr->qp_state = IB_QPS_INIT;
4150 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4153 mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4157 memset(attr, 0, sizeof(*attr));
4158 attr->qp_state = IB_QPS_RTR;
4159 attr->path_mtu = IB_MTU_256;
4161 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4163 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4167 memset(attr, 0, sizeof(*attr));
4168 attr->qp_state = IB_QPS_RTS;
4169 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4171 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4179 sema_init(&dev->umrc.sem, MAX_UMR_WR);
4180 ret = mlx5_mr_cache_init(dev);
4182 mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4192 mlx5_ib_destroy_qp(qp);
4193 dev->umrc.qp = NULL;
4197 dev->umrc.cq = NULL;
4201 dev->umrc.pd = NULL;
4209 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
4211 switch (umr_fence_cap) {
4212 case MLX5_CAP_UMR_FENCE_NONE:
4213 return MLX5_FENCE_MODE_NONE;
4214 case MLX5_CAP_UMR_FENCE_SMALL:
4215 return MLX5_FENCE_MODE_INITIATOR_SMALL;
4217 return MLX5_FENCE_MODE_STRONG_ORDERING;
4221 static int create_dev_resources(struct mlx5_ib_resources *devr)
4223 struct ib_srq_init_attr attr;
4224 struct mlx5_ib_dev *dev;
4225 struct ib_cq_init_attr cq_attr = {.cqe = 1};
4229 dev = container_of(devr, struct mlx5_ib_dev, devr);
4231 mutex_init(&devr->mutex);
4233 devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
4234 if (IS_ERR(devr->p0)) {
4235 ret = PTR_ERR(devr->p0);
4238 devr->p0->device = &dev->ib_dev;
4239 devr->p0->uobject = NULL;
4240 atomic_set(&devr->p0->usecnt, 0);
4242 devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
4243 if (IS_ERR(devr->c0)) {
4244 ret = PTR_ERR(devr->c0);
4247 devr->c0->device = &dev->ib_dev;
4248 devr->c0->uobject = NULL;
4249 devr->c0->comp_handler = NULL;
4250 devr->c0->event_handler = NULL;
4251 devr->c0->cq_context = NULL;
4252 atomic_set(&devr->c0->usecnt, 0);
4254 devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
4255 if (IS_ERR(devr->x0)) {
4256 ret = PTR_ERR(devr->x0);
4259 devr->x0->device = &dev->ib_dev;
4260 devr->x0->inode = NULL;
4261 atomic_set(&devr->x0->usecnt, 0);
4262 mutex_init(&devr->x0->tgt_qp_mutex);
4263 INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
4265 devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
4266 if (IS_ERR(devr->x1)) {
4267 ret = PTR_ERR(devr->x1);
4270 devr->x1->device = &dev->ib_dev;
4271 devr->x1->inode = NULL;
4272 atomic_set(&devr->x1->usecnt, 0);
4273 mutex_init(&devr->x1->tgt_qp_mutex);
4274 INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
4276 memset(&attr, 0, sizeof(attr));
4277 attr.attr.max_sge = 1;
4278 attr.attr.max_wr = 1;
4279 attr.srq_type = IB_SRQT_XRC;
4280 attr.ext.cq = devr->c0;
4281 attr.ext.xrc.xrcd = devr->x0;
4283 devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
4284 if (IS_ERR(devr->s0)) {
4285 ret = PTR_ERR(devr->s0);
4288 devr->s0->device = &dev->ib_dev;
4289 devr->s0->pd = devr->p0;
4290 devr->s0->uobject = NULL;
4291 devr->s0->event_handler = NULL;
4292 devr->s0->srq_context = NULL;
4293 devr->s0->srq_type = IB_SRQT_XRC;
4294 devr->s0->ext.xrc.xrcd = devr->x0;
4295 devr->s0->ext.cq = devr->c0;
4296 atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
4297 atomic_inc(&devr->s0->ext.cq->usecnt);
4298 atomic_inc(&devr->p0->usecnt);
4299 atomic_set(&devr->s0->usecnt, 0);
4301 memset(&attr, 0, sizeof(attr));
4302 attr.attr.max_sge = 1;
4303 attr.attr.max_wr = 1;
4304 attr.srq_type = IB_SRQT_BASIC;
4305 devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
4306 if (IS_ERR(devr->s1)) {
4307 ret = PTR_ERR(devr->s1);
4310 devr->s1->device = &dev->ib_dev;
4311 devr->s1->pd = devr->p0;
4312 devr->s1->uobject = NULL;
4313 devr->s1->event_handler = NULL;
4314 devr->s1->srq_context = NULL;
4315 devr->s1->srq_type = IB_SRQT_BASIC;
4316 devr->s1->ext.cq = devr->c0;
4317 atomic_inc(&devr->p0->usecnt);
4318 atomic_set(&devr->s1->usecnt, 0);
4320 for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
4321 INIT_WORK(&devr->ports[port].pkey_change_work,
4322 pkey_change_handler);
4323 devr->ports[port].devr = devr;
4329 mlx5_ib_destroy_srq(devr->s0);
4331 mlx5_ib_dealloc_xrcd(devr->x1);
4333 mlx5_ib_dealloc_xrcd(devr->x0);
4335 mlx5_ib_destroy_cq(devr->c0);
4337 mlx5_ib_dealloc_pd(devr->p0);
4342 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
4344 struct mlx5_ib_dev *dev =
4345 container_of(devr, struct mlx5_ib_dev, devr);
4348 mlx5_ib_destroy_srq(devr->s1);
4349 mlx5_ib_destroy_srq(devr->s0);
4350 mlx5_ib_dealloc_xrcd(devr->x0);
4351 mlx5_ib_dealloc_xrcd(devr->x1);
4352 mlx5_ib_destroy_cq(devr->c0);
4353 mlx5_ib_dealloc_pd(devr->p0);
4355 /* Make sure no change P_Key work items are still executing */
4356 for (port = 0; port < dev->num_ports; ++port)
4357 cancel_work_sync(&devr->ports[port].pkey_change_work);
4360 static u32 get_core_cap_flags(struct ib_device *ibdev)
4362 struct mlx5_ib_dev *dev = to_mdev(ibdev);
4363 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
4364 u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
4365 u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
4366 bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
4369 if (ll == IB_LINK_LAYER_INFINIBAND)
4370 return RDMA_CORE_PORT_IBA_IB;
4373 ret = RDMA_CORE_PORT_RAW_PACKET;
4375 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
4378 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
4381 if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
4382 ret |= RDMA_CORE_PORT_IBA_ROCE;
4384 if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
4385 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
4390 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
4391 struct ib_port_immutable *immutable)
4393 struct ib_port_attr attr;
4394 struct mlx5_ib_dev *dev = to_mdev(ibdev);
4395 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
4398 immutable->core_cap_flags = get_core_cap_flags(ibdev);
4400 err = ib_query_port(ibdev, port_num, &attr);
4404 immutable->pkey_tbl_len = attr.pkey_tbl_len;
4405 immutable->gid_tbl_len = attr.gid_tbl_len;
4406 immutable->core_cap_flags = get_core_cap_flags(ibdev);
4407 if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
4408 immutable->max_mad_size = IB_MGMT_MAD_SIZE;
4413 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u8 port_num,
4414 struct ib_port_immutable *immutable)
4416 struct ib_port_attr attr;
4419 immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
4421 err = ib_query_port(ibdev, port_num, &attr);
4425 immutable->pkey_tbl_len = attr.pkey_tbl_len;
4426 immutable->gid_tbl_len = attr.gid_tbl_len;
4427 immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
4432 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
4434 struct mlx5_ib_dev *dev =
4435 container_of(ibdev, struct mlx5_ib_dev, ib_dev);
4436 snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
4437 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
4438 fw_rev_sub(dev->mdev));
4441 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
4443 struct mlx5_core_dev *mdev = dev->mdev;
4444 struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
4445 MLX5_FLOW_NAMESPACE_LAG);
4446 struct mlx5_flow_table *ft;
4449 if (!ns || !mlx5_lag_is_active(mdev))
4452 err = mlx5_cmd_create_vport_lag(mdev);
4456 ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
4459 goto err_destroy_vport_lag;
4462 dev->flow_db->lag_demux_ft = ft;
4465 err_destroy_vport_lag:
4466 mlx5_cmd_destroy_vport_lag(mdev);
4470 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
4472 struct mlx5_core_dev *mdev = dev->mdev;
4474 if (dev->flow_db->lag_demux_ft) {
4475 mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
4476 dev->flow_db->lag_demux_ft = NULL;
4478 mlx5_cmd_destroy_vport_lag(mdev);
4482 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
4486 dev->roce[port_num].nb.notifier_call = mlx5_netdev_event;
4487 err = register_netdevice_notifier(&dev->roce[port_num].nb);
4489 dev->roce[port_num].nb.notifier_call = NULL;
4496 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
4498 if (dev->roce[port_num].nb.notifier_call) {
4499 unregister_netdevice_notifier(&dev->roce[port_num].nb);
4500 dev->roce[port_num].nb.notifier_call = NULL;
4504 static int mlx5_enable_eth(struct mlx5_ib_dev *dev, u8 port_num)
4508 if (MLX5_CAP_GEN(dev->mdev, roce)) {
4509 err = mlx5_nic_vport_enable_roce(dev->mdev);
4514 err = mlx5_eth_lag_init(dev);
4516 goto err_disable_roce;
4521 if (MLX5_CAP_GEN(dev->mdev, roce))
4522 mlx5_nic_vport_disable_roce(dev->mdev);
4527 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
4529 mlx5_eth_lag_cleanup(dev);
4530 if (MLX5_CAP_GEN(dev->mdev, roce))
4531 mlx5_nic_vport_disable_roce(dev->mdev);
4534 struct mlx5_ib_counter {
4539 #define INIT_Q_COUNTER(_name) \
4540 { .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
4542 static const struct mlx5_ib_counter basic_q_cnts[] = {
4543 INIT_Q_COUNTER(rx_write_requests),
4544 INIT_Q_COUNTER(rx_read_requests),
4545 INIT_Q_COUNTER(rx_atomic_requests),
4546 INIT_Q_COUNTER(out_of_buffer),
4549 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
4550 INIT_Q_COUNTER(out_of_sequence),
4553 static const struct mlx5_ib_counter retrans_q_cnts[] = {
4554 INIT_Q_COUNTER(duplicate_request),
4555 INIT_Q_COUNTER(rnr_nak_retry_err),
4556 INIT_Q_COUNTER(packet_seq_err),
4557 INIT_Q_COUNTER(implied_nak_seq_err),
4558 INIT_Q_COUNTER(local_ack_timeout_err),
4561 #define INIT_CONG_COUNTER(_name) \
4562 { .name = #_name, .offset = \
4563 MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
4565 static const struct mlx5_ib_counter cong_cnts[] = {
4566 INIT_CONG_COUNTER(rp_cnp_ignored),
4567 INIT_CONG_COUNTER(rp_cnp_handled),
4568 INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
4569 INIT_CONG_COUNTER(np_cnp_sent),
4572 static const struct mlx5_ib_counter extended_err_cnts[] = {
4573 INIT_Q_COUNTER(resp_local_length_error),
4574 INIT_Q_COUNTER(resp_cqe_error),
4575 INIT_Q_COUNTER(req_cqe_error),
4576 INIT_Q_COUNTER(req_remote_invalid_request),
4577 INIT_Q_COUNTER(req_remote_access_errors),
4578 INIT_Q_COUNTER(resp_remote_access_errors),
4579 INIT_Q_COUNTER(resp_cqe_flush_error),
4580 INIT_Q_COUNTER(req_cqe_flush_error),
4583 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
4587 for (i = 0; i < dev->num_ports; i++) {
4588 if (dev->port[i].cnts.set_id)
4589 mlx5_core_dealloc_q_counter(dev->mdev,
4590 dev->port[i].cnts.set_id);
4591 kfree(dev->port[i].cnts.names);
4592 kfree(dev->port[i].cnts.offsets);
4596 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
4597 struct mlx5_ib_counters *cnts)
4601 num_counters = ARRAY_SIZE(basic_q_cnts);
4603 if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
4604 num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
4606 if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
4607 num_counters += ARRAY_SIZE(retrans_q_cnts);
4609 if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters))
4610 num_counters += ARRAY_SIZE(extended_err_cnts);
4612 cnts->num_q_counters = num_counters;
4614 if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4615 cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
4616 num_counters += ARRAY_SIZE(cong_cnts);
4619 cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
4623 cnts->offsets = kcalloc(num_counters,
4624 sizeof(cnts->offsets), GFP_KERNEL);
4636 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
4643 for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
4644 names[j] = basic_q_cnts[i].name;
4645 offsets[j] = basic_q_cnts[i].offset;
4648 if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
4649 for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
4650 names[j] = out_of_seq_q_cnts[i].name;
4651 offsets[j] = out_of_seq_q_cnts[i].offset;
4655 if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
4656 for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
4657 names[j] = retrans_q_cnts[i].name;
4658 offsets[j] = retrans_q_cnts[i].offset;
4662 if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters)) {
4663 for (i = 0; i < ARRAY_SIZE(extended_err_cnts); i++, j++) {
4664 names[j] = extended_err_cnts[i].name;
4665 offsets[j] = extended_err_cnts[i].offset;
4669 if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4670 for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
4671 names[j] = cong_cnts[i].name;
4672 offsets[j] = cong_cnts[i].offset;
4677 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
4682 for (i = 0; i < dev->num_ports; i++) {
4683 err = __mlx5_ib_alloc_counters(dev, &dev->port[i].cnts);
4687 mlx5_ib_fill_counters(dev, dev->port[i].cnts.names,
4688 dev->port[i].cnts.offsets);
4690 err = mlx5_core_alloc_q_counter(dev->mdev,
4691 &dev->port[i].cnts.set_id);
4694 "couldn't allocate queue counter for port %d, err %d\n",
4698 dev->port[i].cnts.set_id_valid = true;
4704 mlx5_ib_dealloc_counters(dev);
4708 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
4711 struct mlx5_ib_dev *dev = to_mdev(ibdev);
4712 struct mlx5_ib_port *port = &dev->port[port_num - 1];
4714 /* We support only per port stats */
4718 return rdma_alloc_hw_stats_struct(port->cnts.names,
4719 port->cnts.num_q_counters +
4720 port->cnts.num_cong_counters,
4721 RDMA_HW_STATS_DEFAULT_LIFESPAN);
4724 static int mlx5_ib_query_q_counters(struct mlx5_core_dev *mdev,
4725 struct mlx5_ib_port *port,
4726 struct rdma_hw_stats *stats)
4728 int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
4733 out = kvzalloc(outlen, GFP_KERNEL);
4737 ret = mlx5_core_query_q_counter(mdev,
4738 port->cnts.set_id, 0,
4743 for (i = 0; i < port->cnts.num_q_counters; i++) {
4744 val = *(__be32 *)(out + port->cnts.offsets[i]);
4745 stats->value[i] = (u64)be32_to_cpu(val);
4753 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
4754 struct rdma_hw_stats *stats,
4755 u8 port_num, int index)
4757 struct mlx5_ib_dev *dev = to_mdev(ibdev);
4758 struct mlx5_ib_port *port = &dev->port[port_num - 1];
4759 struct mlx5_core_dev *mdev;
4760 int ret, num_counters;
4766 num_counters = port->cnts.num_q_counters + port->cnts.num_cong_counters;
4768 /* q_counters are per IB device, query the master mdev */
4769 ret = mlx5_ib_query_q_counters(dev->mdev, port, stats);
4773 if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4774 mdev = mlx5_ib_get_native_port_mdev(dev, port_num,
4777 /* If port is not affiliated yet, its in down state
4778 * which doesn't have any counters yet, so it would be
4779 * zero. So no need to read from the HCA.
4783 ret = mlx5_lag_query_cong_counters(dev->mdev,
4785 port->cnts.num_q_counters,
4786 port->cnts.num_cong_counters,
4787 port->cnts.offsets +
4788 port->cnts.num_q_counters);
4790 mlx5_ib_put_native_port_mdev(dev, port_num);
4796 return num_counters;
4799 static void mlx5_ib_free_rdma_netdev(struct net_device *netdev)
4801 return mlx5_rdma_netdev_free(netdev);
4804 static struct net_device*
4805 mlx5_ib_alloc_rdma_netdev(struct ib_device *hca,
4807 enum rdma_netdev_t type,
4809 unsigned char name_assign_type,
4810 void (*setup)(struct net_device *))
4812 struct net_device *netdev;
4813 struct rdma_netdev *rn;
4815 if (type != RDMA_NETDEV_IPOIB)
4816 return ERR_PTR(-EOPNOTSUPP);
4818 netdev = mlx5_rdma_netdev_alloc(to_mdev(hca)->mdev, hca,
4820 if (likely(!IS_ERR_OR_NULL(netdev))) {
4821 rn = netdev_priv(netdev);
4822 rn->free_rdma_netdev = mlx5_ib_free_rdma_netdev;
4827 static void delay_drop_debugfs_cleanup(struct mlx5_ib_dev *dev)
4829 if (!dev->delay_drop.dbg)
4831 debugfs_remove_recursive(dev->delay_drop.dbg->dir_debugfs);
4832 kfree(dev->delay_drop.dbg);
4833 dev->delay_drop.dbg = NULL;
4836 static void cancel_delay_drop(struct mlx5_ib_dev *dev)
4838 if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4841 cancel_work_sync(&dev->delay_drop.delay_drop_work);
4842 delay_drop_debugfs_cleanup(dev);
4845 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
4846 size_t count, loff_t *pos)
4848 struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
4852 len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
4853 return simple_read_from_buffer(buf, count, pos, lbuf, len);
4856 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
4857 size_t count, loff_t *pos)
4859 struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
4863 if (kstrtouint_from_user(buf, count, 0, &var))
4866 timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
4869 mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
4872 delay_drop->timeout = timeout;
4877 static const struct file_operations fops_delay_drop_timeout = {
4878 .owner = THIS_MODULE,
4879 .open = simple_open,
4880 .write = delay_drop_timeout_write,
4881 .read = delay_drop_timeout_read,
4884 static int delay_drop_debugfs_init(struct mlx5_ib_dev *dev)
4886 struct mlx5_ib_dbg_delay_drop *dbg;
4888 if (!mlx5_debugfs_root)
4891 dbg = kzalloc(sizeof(*dbg), GFP_KERNEL);
4895 dev->delay_drop.dbg = dbg;
4898 debugfs_create_dir("delay_drop",
4899 dev->mdev->priv.dbg_root);
4900 if (!dbg->dir_debugfs)
4903 dbg->events_cnt_debugfs =
4904 debugfs_create_atomic_t("num_timeout_events", 0400,
4906 &dev->delay_drop.events_cnt);
4907 if (!dbg->events_cnt_debugfs)
4910 dbg->rqs_cnt_debugfs =
4911 debugfs_create_atomic_t("num_rqs", 0400,
4913 &dev->delay_drop.rqs_cnt);
4914 if (!dbg->rqs_cnt_debugfs)
4917 dbg->timeout_debugfs =
4918 debugfs_create_file("timeout", 0600,
4921 &fops_delay_drop_timeout);
4922 if (!dbg->timeout_debugfs)
4928 delay_drop_debugfs_cleanup(dev);
4932 static void init_delay_drop(struct mlx5_ib_dev *dev)
4934 if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4937 mutex_init(&dev->delay_drop.lock);
4938 dev->delay_drop.dev = dev;
4939 dev->delay_drop.activate = false;
4940 dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4941 INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4942 atomic_set(&dev->delay_drop.rqs_cnt, 0);
4943 atomic_set(&dev->delay_drop.events_cnt, 0);
4945 if (delay_drop_debugfs_init(dev))
4946 mlx5_ib_warn(dev, "Failed to init delay drop debugfs\n");
4949 static const struct cpumask *
4950 mlx5_ib_get_vector_affinity(struct ib_device *ibdev, int comp_vector)
4952 struct mlx5_ib_dev *dev = to_mdev(ibdev);
4954 return mlx5_get_vector_affinity(dev->mdev, comp_vector);
4957 /* The mlx5_ib_multiport_mutex should be held when calling this function */
4958 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
4959 struct mlx5_ib_multiport_info *mpi)
4961 u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
4962 struct mlx5_ib_port *port = &ibdev->port[port_num];
4967 mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
4969 spin_lock(&port->mp.mpi_lock);
4971 spin_unlock(&port->mp.mpi_lock);
4976 spin_unlock(&port->mp.mpi_lock);
4977 mlx5_remove_netdev_notifier(ibdev, port_num);
4978 spin_lock(&port->mp.mpi_lock);
4980 comps = mpi->mdev_refcnt;
4982 mpi->unaffiliate = true;
4983 init_completion(&mpi->unref_comp);
4984 spin_unlock(&port->mp.mpi_lock);
4986 for (i = 0; i < comps; i++)
4987 wait_for_completion(&mpi->unref_comp);
4989 spin_lock(&port->mp.mpi_lock);
4990 mpi->unaffiliate = false;
4993 port->mp.mpi = NULL;
4995 list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4997 spin_unlock(&port->mp.mpi_lock);
4999 err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
5001 mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
5002 /* Log an error, still needed to cleanup the pointers and add
5003 * it back to the list.
5006 mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
5009 ibdev->roce[port_num].last_port_state = IB_PORT_DOWN;
5012 /* The mlx5_ib_multiport_mutex should be held when calling this function */
5013 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
5014 struct mlx5_ib_multiport_info *mpi)
5016 u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
5019 spin_lock(&ibdev->port[port_num].mp.mpi_lock);
5020 if (ibdev->port[port_num].mp.mpi) {
5021 mlx5_ib_warn(ibdev, "port %d already affiliated.\n",
5023 spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
5027 ibdev->port[port_num].mp.mpi = mpi;
5029 spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
5031 err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
5035 err = get_port_caps(ibdev, mlx5_core_native_port_num(mpi->mdev));
5039 err = mlx5_add_netdev_notifier(ibdev, port_num);
5041 mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
5046 err = mlx5_ib_init_cong_debugfs(ibdev, port_num);
5053 mlx5_ib_unbind_slave_port(ibdev, mpi);
5057 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
5059 int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5060 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
5062 struct mlx5_ib_multiport_info *mpi;
5066 if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
5069 err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
5070 &dev->sys_image_guid);
5074 err = mlx5_nic_vport_enable_roce(dev->mdev);
5078 mutex_lock(&mlx5_ib_multiport_mutex);
5079 for (i = 0; i < dev->num_ports; i++) {
5082 /* build a stub multiport info struct for the native port. */
5083 if (i == port_num) {
5084 mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
5086 mutex_unlock(&mlx5_ib_multiport_mutex);
5087 mlx5_nic_vport_disable_roce(dev->mdev);
5091 mpi->is_master = true;
5092 mpi->mdev = dev->mdev;
5093 mpi->sys_image_guid = dev->sys_image_guid;
5094 dev->port[i].mp.mpi = mpi;
5100 list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
5102 if (dev->sys_image_guid == mpi->sys_image_guid &&
5103 (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
5104 bound = mlx5_ib_bind_slave_port(dev, mpi);
5108 dev_dbg(&mpi->mdev->pdev->dev, "removing port from unaffiliated list.\n");
5109 mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
5110 list_del(&mpi->list);
5115 get_port_caps(dev, i + 1);
5116 mlx5_ib_dbg(dev, "no free port found for port %d\n",
5121 list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
5122 mutex_unlock(&mlx5_ib_multiport_mutex);
5126 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
5128 int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5129 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
5133 if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
5136 mutex_lock(&mlx5_ib_multiport_mutex);
5137 for (i = 0; i < dev->num_ports; i++) {
5138 if (dev->port[i].mp.mpi) {
5139 /* Destroy the native port stub */
5140 if (i == port_num) {
5141 kfree(dev->port[i].mp.mpi);
5142 dev->port[i].mp.mpi = NULL;
5144 mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
5145 mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
5150 mlx5_ib_dbg(dev, "removing from devlist\n");
5151 list_del(&dev->ib_dev_list);
5152 mutex_unlock(&mlx5_ib_multiport_mutex);
5154 mlx5_nic_vport_disable_roce(dev->mdev);
5157 ADD_UVERBS_ATTRIBUTES_SIMPLE(mlx5_ib_dm, UVERBS_OBJECT_DM,
5158 UVERBS_METHOD_DM_ALLOC,
5159 &UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
5160 UVERBS_ATTR_TYPE(u64),
5161 UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)),
5162 &UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
5163 UVERBS_ATTR_TYPE(u16),
5164 UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)));
5166 ADD_UVERBS_ATTRIBUTES_SIMPLE(mlx5_ib_flow_action, UVERBS_OBJECT_FLOW_ACTION,
5167 UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
5168 &UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
5169 UVERBS_ATTR_TYPE(u64),
5170 UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)));
5173 static int populate_specs_root(struct mlx5_ib_dev *dev)
5175 const struct uverbs_object_tree_def *default_root[NUM_TREES + 1] = {
5176 uverbs_default_get_objects()};
5177 size_t num_trees = 1;
5179 if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE &&
5180 !WARN_ON(num_trees >= ARRAY_SIZE(default_root)))
5181 default_root[num_trees++] = &mlx5_ib_flow_action;
5183 if (MLX5_CAP_DEV_MEM(dev->mdev, memic) &&
5184 !WARN_ON(num_trees >= ARRAY_SIZE(default_root)))
5185 default_root[num_trees++] = &mlx5_ib_dm;
5187 dev->ib_dev.specs_root =
5188 uverbs_alloc_spec_tree(num_trees, default_root);
5190 return PTR_ERR_OR_ZERO(dev->ib_dev.specs_root);
5193 static void depopulate_specs_root(struct mlx5_ib_dev *dev)
5195 uverbs_free_spec_tree(dev->ib_dev.specs_root);
5198 static int mlx5_ib_destroy_counters(struct ib_counters *counters)
5200 struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
5202 counters_clear_description(counters);
5203 if (mcounters->hw_cntrs_hndl)
5204 mlx5_fc_destroy(to_mdev(counters->device)->mdev,
5205 mcounters->hw_cntrs_hndl);
5212 static struct ib_counters *mlx5_ib_create_counters(struct ib_device *device,
5213 struct uverbs_attr_bundle *attrs)
5215 struct mlx5_ib_mcounters *mcounters;
5217 mcounters = kzalloc(sizeof(*mcounters), GFP_KERNEL);
5219 return ERR_PTR(-ENOMEM);
5221 mutex_init(&mcounters->mcntrs_mutex);
5223 return &mcounters->ibcntrs;
5226 void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
5228 mlx5_ib_cleanup_multiport_master(dev);
5229 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5230 cleanup_srcu_struct(&dev->mr_srcu);
5235 int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
5237 struct mlx5_core_dev *mdev = dev->mdev;
5242 dev->port = kcalloc(dev->num_ports, sizeof(*dev->port),
5247 for (i = 0; i < dev->num_ports; i++) {
5248 spin_lock_init(&dev->port[i].mp.mpi_lock);
5249 rwlock_init(&dev->roce[i].netdev_lock);
5252 err = mlx5_ib_init_multiport_master(dev);
5256 if (!mlx5_core_mp_enabled(mdev)) {
5257 for (i = 1; i <= dev->num_ports; i++) {
5258 err = get_port_caps(dev, i);
5263 err = get_port_caps(dev, mlx5_core_native_port_num(mdev));
5268 if (mlx5_use_mad_ifc(dev))
5269 get_ext_port_caps(dev);
5271 if (!mlx5_lag_is_active(mdev))
5274 name = "mlx5_bond_%d";
5276 strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
5277 dev->ib_dev.owner = THIS_MODULE;
5278 dev->ib_dev.node_type = RDMA_NODE_IB_CA;
5279 dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
5280 dev->ib_dev.phys_port_cnt = dev->num_ports;
5281 dev->ib_dev.num_comp_vectors =
5282 dev->mdev->priv.eq_table.num_comp_vectors;
5283 dev->ib_dev.dev.parent = &mdev->pdev->dev;
5285 mutex_init(&dev->cap_mask_mutex);
5286 INIT_LIST_HEAD(&dev->qp_list);
5287 spin_lock_init(&dev->reset_flow_resource_lock);
5289 spin_lock_init(&dev->memic.memic_lock);
5290 dev->memic.dev = mdev;
5292 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5293 err = init_srcu_struct(&dev->mr_srcu);
5300 mlx5_ib_cleanup_multiport_master(dev);
5308 static int mlx5_ib_stage_flow_db_init(struct mlx5_ib_dev *dev)
5310 dev->flow_db = kzalloc(sizeof(*dev->flow_db), GFP_KERNEL);
5315 mutex_init(&dev->flow_db->lock);
5320 int mlx5_ib_stage_rep_flow_db_init(struct mlx5_ib_dev *dev)
5322 struct mlx5_ib_dev *nic_dev;
5324 nic_dev = mlx5_ib_get_uplink_ibdev(dev->mdev->priv.eswitch);
5329 dev->flow_db = nic_dev->flow_db;
5334 static void mlx5_ib_stage_flow_db_cleanup(struct mlx5_ib_dev *dev)
5336 kfree(dev->flow_db);
5339 int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
5341 struct mlx5_core_dev *mdev = dev->mdev;
5344 dev->ib_dev.uverbs_abi_ver = MLX5_IB_UVERBS_ABI_VERSION;
5345 dev->ib_dev.uverbs_cmd_mask =
5346 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
5347 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
5348 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
5349 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
5350 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
5351 (1ull << IB_USER_VERBS_CMD_CREATE_AH) |
5352 (1ull << IB_USER_VERBS_CMD_DESTROY_AH) |
5353 (1ull << IB_USER_VERBS_CMD_REG_MR) |
5354 (1ull << IB_USER_VERBS_CMD_REREG_MR) |
5355 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
5356 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
5357 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
5358 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
5359 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
5360 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
5361 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
5362 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
5363 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
5364 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
5365 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) |
5366 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
5367 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
5368 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
5369 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) |
5370 (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ) |
5371 (1ull << IB_USER_VERBS_CMD_OPEN_QP);
5372 dev->ib_dev.uverbs_ex_cmd_mask =
5373 (1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE) |
5374 (1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ) |
5375 (1ull << IB_USER_VERBS_EX_CMD_CREATE_QP) |
5376 (1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP) |
5377 (1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ);
5379 dev->ib_dev.query_device = mlx5_ib_query_device;
5380 dev->ib_dev.get_link_layer = mlx5_ib_port_link_layer;
5381 dev->ib_dev.query_gid = mlx5_ib_query_gid;
5382 dev->ib_dev.add_gid = mlx5_ib_add_gid;
5383 dev->ib_dev.del_gid = mlx5_ib_del_gid;
5384 dev->ib_dev.query_pkey = mlx5_ib_query_pkey;
5385 dev->ib_dev.modify_device = mlx5_ib_modify_device;
5386 dev->ib_dev.modify_port = mlx5_ib_modify_port;
5387 dev->ib_dev.alloc_ucontext = mlx5_ib_alloc_ucontext;
5388 dev->ib_dev.dealloc_ucontext = mlx5_ib_dealloc_ucontext;
5389 dev->ib_dev.mmap = mlx5_ib_mmap;
5390 dev->ib_dev.alloc_pd = mlx5_ib_alloc_pd;
5391 dev->ib_dev.dealloc_pd = mlx5_ib_dealloc_pd;
5392 dev->ib_dev.create_ah = mlx5_ib_create_ah;
5393 dev->ib_dev.query_ah = mlx5_ib_query_ah;
5394 dev->ib_dev.destroy_ah = mlx5_ib_destroy_ah;
5395 dev->ib_dev.create_srq = mlx5_ib_create_srq;
5396 dev->ib_dev.modify_srq = mlx5_ib_modify_srq;
5397 dev->ib_dev.query_srq = mlx5_ib_query_srq;
5398 dev->ib_dev.destroy_srq = mlx5_ib_destroy_srq;
5399 dev->ib_dev.post_srq_recv = mlx5_ib_post_srq_recv;
5400 dev->ib_dev.create_qp = mlx5_ib_create_qp;
5401 dev->ib_dev.modify_qp = mlx5_ib_modify_qp;
5402 dev->ib_dev.query_qp = mlx5_ib_query_qp;
5403 dev->ib_dev.destroy_qp = mlx5_ib_destroy_qp;
5404 dev->ib_dev.post_send = mlx5_ib_post_send;
5405 dev->ib_dev.post_recv = mlx5_ib_post_recv;
5406 dev->ib_dev.create_cq = mlx5_ib_create_cq;
5407 dev->ib_dev.modify_cq = mlx5_ib_modify_cq;
5408 dev->ib_dev.resize_cq = mlx5_ib_resize_cq;
5409 dev->ib_dev.destroy_cq = mlx5_ib_destroy_cq;
5410 dev->ib_dev.poll_cq = mlx5_ib_poll_cq;
5411 dev->ib_dev.req_notify_cq = mlx5_ib_arm_cq;
5412 dev->ib_dev.get_dma_mr = mlx5_ib_get_dma_mr;
5413 dev->ib_dev.reg_user_mr = mlx5_ib_reg_user_mr;
5414 dev->ib_dev.rereg_user_mr = mlx5_ib_rereg_user_mr;
5415 dev->ib_dev.dereg_mr = mlx5_ib_dereg_mr;
5416 dev->ib_dev.attach_mcast = mlx5_ib_mcg_attach;
5417 dev->ib_dev.detach_mcast = mlx5_ib_mcg_detach;
5418 dev->ib_dev.process_mad = mlx5_ib_process_mad;
5419 dev->ib_dev.alloc_mr = mlx5_ib_alloc_mr;
5420 dev->ib_dev.map_mr_sg = mlx5_ib_map_mr_sg;
5421 dev->ib_dev.check_mr_status = mlx5_ib_check_mr_status;
5422 dev->ib_dev.get_dev_fw_str = get_dev_fw_str;
5423 dev->ib_dev.get_vector_affinity = mlx5_ib_get_vector_affinity;
5424 if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads))
5425 dev->ib_dev.alloc_rdma_netdev = mlx5_ib_alloc_rdma_netdev;
5427 if (mlx5_core_is_pf(mdev)) {
5428 dev->ib_dev.get_vf_config = mlx5_ib_get_vf_config;
5429 dev->ib_dev.set_vf_link_state = mlx5_ib_set_vf_link_state;
5430 dev->ib_dev.get_vf_stats = mlx5_ib_get_vf_stats;
5431 dev->ib_dev.set_vf_guid = mlx5_ib_set_vf_guid;
5434 dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
5436 dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
5438 if (MLX5_CAP_GEN(mdev, imaicl)) {
5439 dev->ib_dev.alloc_mw = mlx5_ib_alloc_mw;
5440 dev->ib_dev.dealloc_mw = mlx5_ib_dealloc_mw;
5441 dev->ib_dev.uverbs_cmd_mask |=
5442 (1ull << IB_USER_VERBS_CMD_ALLOC_MW) |
5443 (1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
5446 if (MLX5_CAP_GEN(mdev, xrc)) {
5447 dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
5448 dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
5449 dev->ib_dev.uverbs_cmd_mask |=
5450 (1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
5451 (1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
5454 if (MLX5_CAP_DEV_MEM(mdev, memic)) {
5455 dev->ib_dev.alloc_dm = mlx5_ib_alloc_dm;
5456 dev->ib_dev.dealloc_dm = mlx5_ib_dealloc_dm;
5457 dev->ib_dev.reg_dm_mr = mlx5_ib_reg_dm_mr;
5460 dev->ib_dev.create_flow = mlx5_ib_create_flow;
5461 dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
5462 dev->ib_dev.uverbs_ex_cmd_mask |=
5463 (1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
5464 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
5465 dev->ib_dev.create_flow_action_esp = mlx5_ib_create_flow_action_esp;
5466 dev->ib_dev.destroy_flow_action = mlx5_ib_destroy_flow_action;
5467 dev->ib_dev.modify_flow_action_esp = mlx5_ib_modify_flow_action_esp;
5468 dev->ib_dev.driver_id = RDMA_DRIVER_MLX5;
5469 dev->ib_dev.create_counters = mlx5_ib_create_counters;
5470 dev->ib_dev.destroy_counters = mlx5_ib_destroy_counters;
5472 err = init_node_data(dev);
5476 if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
5477 (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
5478 MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
5479 mutex_init(&dev->lb_mutex);
5484 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
5486 dev->ib_dev.get_port_immutable = mlx5_port_immutable;
5487 dev->ib_dev.query_port = mlx5_ib_query_port;
5492 int mlx5_ib_stage_rep_non_default_cb(struct mlx5_ib_dev *dev)
5494 dev->ib_dev.get_port_immutable = mlx5_port_rep_immutable;
5495 dev->ib_dev.query_port = mlx5_ib_rep_query_port;
5500 static int mlx5_ib_stage_common_roce_init(struct mlx5_ib_dev *dev,
5505 for (i = 0; i < dev->num_ports; i++) {
5506 dev->roce[i].dev = dev;
5507 dev->roce[i].native_port_num = i + 1;
5508 dev->roce[i].last_port_state = IB_PORT_DOWN;
5511 dev->ib_dev.get_netdev = mlx5_ib_get_netdev;
5512 dev->ib_dev.create_wq = mlx5_ib_create_wq;
5513 dev->ib_dev.modify_wq = mlx5_ib_modify_wq;
5514 dev->ib_dev.destroy_wq = mlx5_ib_destroy_wq;
5515 dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
5516 dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
5518 dev->ib_dev.uverbs_ex_cmd_mask |=
5519 (1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
5520 (1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
5521 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
5522 (1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
5523 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
5525 return mlx5_add_netdev_notifier(dev, port_num);
5528 static void mlx5_ib_stage_common_roce_cleanup(struct mlx5_ib_dev *dev)
5530 u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5532 mlx5_remove_netdev_notifier(dev, port_num);
5535 int mlx5_ib_stage_rep_roce_init(struct mlx5_ib_dev *dev)
5537 struct mlx5_core_dev *mdev = dev->mdev;
5538 enum rdma_link_layer ll;
5543 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5544 port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5545 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5547 if (ll == IB_LINK_LAYER_ETHERNET)
5548 err = mlx5_ib_stage_common_roce_init(dev, port_num);
5553 void mlx5_ib_stage_rep_roce_cleanup(struct mlx5_ib_dev *dev)
5555 mlx5_ib_stage_common_roce_cleanup(dev);
5558 static int mlx5_ib_stage_roce_init(struct mlx5_ib_dev *dev)
5560 struct mlx5_core_dev *mdev = dev->mdev;
5561 enum rdma_link_layer ll;
5566 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5567 port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5568 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5570 if (ll == IB_LINK_LAYER_ETHERNET) {
5571 err = mlx5_ib_stage_common_roce_init(dev, port_num);
5575 err = mlx5_enable_eth(dev, port_num);
5582 mlx5_ib_stage_common_roce_cleanup(dev);
5587 static void mlx5_ib_stage_roce_cleanup(struct mlx5_ib_dev *dev)
5589 struct mlx5_core_dev *mdev = dev->mdev;
5590 enum rdma_link_layer ll;
5594 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5595 port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5596 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5598 if (ll == IB_LINK_LAYER_ETHERNET) {
5599 mlx5_disable_eth(dev);
5600 mlx5_ib_stage_common_roce_cleanup(dev);
5604 int mlx5_ib_stage_dev_res_init(struct mlx5_ib_dev *dev)
5606 return create_dev_resources(&dev->devr);
5609 void mlx5_ib_stage_dev_res_cleanup(struct mlx5_ib_dev *dev)
5611 destroy_dev_resources(&dev->devr);
5614 static int mlx5_ib_stage_odp_init(struct mlx5_ib_dev *dev)
5616 mlx5_ib_internal_fill_odp_caps(dev);
5618 return mlx5_ib_odp_init_one(dev);
5621 int mlx5_ib_stage_counters_init(struct mlx5_ib_dev *dev)
5623 if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
5624 dev->ib_dev.get_hw_stats = mlx5_ib_get_hw_stats;
5625 dev->ib_dev.alloc_hw_stats = mlx5_ib_alloc_hw_stats;
5627 return mlx5_ib_alloc_counters(dev);
5633 void mlx5_ib_stage_counters_cleanup(struct mlx5_ib_dev *dev)
5635 if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
5636 mlx5_ib_dealloc_counters(dev);
5639 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
5641 return mlx5_ib_init_cong_debugfs(dev,
5642 mlx5_core_native_port_num(dev->mdev) - 1);
5645 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
5647 mlx5_ib_cleanup_cong_debugfs(dev,
5648 mlx5_core_native_port_num(dev->mdev) - 1);
5651 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
5653 dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
5654 if (!dev->mdev->priv.uar)
5659 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
5661 mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
5664 int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
5668 err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
5672 err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
5674 mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
5679 void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
5681 mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
5682 mlx5_free_bfreg(dev->mdev, &dev->bfreg);
5685 static int mlx5_ib_stage_populate_specs(struct mlx5_ib_dev *dev)
5687 return populate_specs_root(dev);
5690 int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
5692 return ib_register_device(&dev->ib_dev, NULL);
5695 static void mlx5_ib_stage_depopulate_specs(struct mlx5_ib_dev *dev)
5697 depopulate_specs_root(dev);
5700 void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
5702 destroy_umrc_res(dev);
5705 void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
5707 ib_unregister_device(&dev->ib_dev);
5710 int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
5712 return create_umr_res(dev);
5715 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
5717 init_delay_drop(dev);
5722 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
5724 cancel_delay_drop(dev);
5727 int mlx5_ib_stage_class_attr_init(struct mlx5_ib_dev *dev)
5732 for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
5733 err = device_create_file(&dev->ib_dev.dev,
5734 mlx5_class_attributes[i]);
5742 static int mlx5_ib_stage_rep_reg_init(struct mlx5_ib_dev *dev)
5744 mlx5_ib_register_vport_reps(dev);
5749 static void mlx5_ib_stage_rep_reg_cleanup(struct mlx5_ib_dev *dev)
5751 mlx5_ib_unregister_vport_reps(dev);
5754 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
5755 const struct mlx5_ib_profile *profile,
5758 /* Number of stages to cleanup */
5761 if (profile->stage[stage].cleanup)
5762 profile->stage[stage].cleanup(dev);
5765 ib_dealloc_device((struct ib_device *)dev);
5768 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev, u8 port_num);
5770 void *__mlx5_ib_add(struct mlx5_ib_dev *dev,
5771 const struct mlx5_ib_profile *profile)
5776 printk_once(KERN_INFO "%s", mlx5_version);
5778 for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
5779 if (profile->stage[i].init) {
5780 err = profile->stage[i].init(dev);
5786 dev->profile = profile;
5787 dev->ib_active = true;
5792 __mlx5_ib_remove(dev, profile, i);
5797 static const struct mlx5_ib_profile pf_profile = {
5798 STAGE_CREATE(MLX5_IB_STAGE_INIT,
5799 mlx5_ib_stage_init_init,
5800 mlx5_ib_stage_init_cleanup),
5801 STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
5802 mlx5_ib_stage_flow_db_init,
5803 mlx5_ib_stage_flow_db_cleanup),
5804 STAGE_CREATE(MLX5_IB_STAGE_CAPS,
5805 mlx5_ib_stage_caps_init,
5807 STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
5808 mlx5_ib_stage_non_default_cb,
5810 STAGE_CREATE(MLX5_IB_STAGE_ROCE,
5811 mlx5_ib_stage_roce_init,
5812 mlx5_ib_stage_roce_cleanup),
5813 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
5814 mlx5_ib_stage_dev_res_init,
5815 mlx5_ib_stage_dev_res_cleanup),
5816 STAGE_CREATE(MLX5_IB_STAGE_ODP,
5817 mlx5_ib_stage_odp_init,
5819 STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
5820 mlx5_ib_stage_counters_init,
5821 mlx5_ib_stage_counters_cleanup),
5822 STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
5823 mlx5_ib_stage_cong_debugfs_init,
5824 mlx5_ib_stage_cong_debugfs_cleanup),
5825 STAGE_CREATE(MLX5_IB_STAGE_UAR,
5826 mlx5_ib_stage_uar_init,
5827 mlx5_ib_stage_uar_cleanup),
5828 STAGE_CREATE(MLX5_IB_STAGE_BFREG,
5829 mlx5_ib_stage_bfrag_init,
5830 mlx5_ib_stage_bfrag_cleanup),
5831 STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
5833 mlx5_ib_stage_pre_ib_reg_umr_cleanup),
5834 STAGE_CREATE(MLX5_IB_STAGE_SPECS,
5835 mlx5_ib_stage_populate_specs,
5836 mlx5_ib_stage_depopulate_specs),
5837 STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
5838 mlx5_ib_stage_ib_reg_init,
5839 mlx5_ib_stage_ib_reg_cleanup),
5840 STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
5841 mlx5_ib_stage_post_ib_reg_umr_init,
5843 STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
5844 mlx5_ib_stage_delay_drop_init,
5845 mlx5_ib_stage_delay_drop_cleanup),
5846 STAGE_CREATE(MLX5_IB_STAGE_CLASS_ATTR,
5847 mlx5_ib_stage_class_attr_init,
5851 static const struct mlx5_ib_profile nic_rep_profile = {
5852 STAGE_CREATE(MLX5_IB_STAGE_INIT,
5853 mlx5_ib_stage_init_init,
5854 mlx5_ib_stage_init_cleanup),
5855 STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
5856 mlx5_ib_stage_flow_db_init,
5857 mlx5_ib_stage_flow_db_cleanup),
5858 STAGE_CREATE(MLX5_IB_STAGE_CAPS,
5859 mlx5_ib_stage_caps_init,
5861 STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
5862 mlx5_ib_stage_rep_non_default_cb,
5864 STAGE_CREATE(MLX5_IB_STAGE_ROCE,
5865 mlx5_ib_stage_rep_roce_init,
5866 mlx5_ib_stage_rep_roce_cleanup),
5867 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
5868 mlx5_ib_stage_dev_res_init,
5869 mlx5_ib_stage_dev_res_cleanup),
5870 STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
5871 mlx5_ib_stage_counters_init,
5872 mlx5_ib_stage_counters_cleanup),
5873 STAGE_CREATE(MLX5_IB_STAGE_UAR,
5874 mlx5_ib_stage_uar_init,
5875 mlx5_ib_stage_uar_cleanup),
5876 STAGE_CREATE(MLX5_IB_STAGE_BFREG,
5877 mlx5_ib_stage_bfrag_init,
5878 mlx5_ib_stage_bfrag_cleanup),
5879 STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
5881 mlx5_ib_stage_pre_ib_reg_umr_cleanup),
5882 STAGE_CREATE(MLX5_IB_STAGE_SPECS,
5883 mlx5_ib_stage_populate_specs,
5884 mlx5_ib_stage_depopulate_specs),
5885 STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
5886 mlx5_ib_stage_ib_reg_init,
5887 mlx5_ib_stage_ib_reg_cleanup),
5888 STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
5889 mlx5_ib_stage_post_ib_reg_umr_init,
5891 STAGE_CREATE(MLX5_IB_STAGE_CLASS_ATTR,
5892 mlx5_ib_stage_class_attr_init,
5894 STAGE_CREATE(MLX5_IB_STAGE_REP_REG,
5895 mlx5_ib_stage_rep_reg_init,
5896 mlx5_ib_stage_rep_reg_cleanup),
5899 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev, u8 port_num)
5901 struct mlx5_ib_multiport_info *mpi;
5902 struct mlx5_ib_dev *dev;
5906 mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
5912 err = mlx5_query_nic_vport_system_image_guid(mdev,
5913 &mpi->sys_image_guid);
5919 mutex_lock(&mlx5_ib_multiport_mutex);
5920 list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
5921 if (dev->sys_image_guid == mpi->sys_image_guid)
5922 bound = mlx5_ib_bind_slave_port(dev, mpi);
5925 rdma_roce_rescan_device(&dev->ib_dev);
5931 list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
5932 dev_dbg(&mdev->pdev->dev, "no suitable IB device found to bind to, added to unaffiliated list.\n");
5934 mlx5_ib_dbg(dev, "bound port %u\n", port_num + 1);
5936 mutex_unlock(&mlx5_ib_multiport_mutex);
5941 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
5943 enum rdma_link_layer ll;
5944 struct mlx5_ib_dev *dev;
5947 printk_once(KERN_INFO "%s", mlx5_version);
5949 port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5950 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5952 if (mlx5_core_is_mp_slave(mdev) && ll == IB_LINK_LAYER_ETHERNET) {
5953 u8 port_num = mlx5_core_native_port_num(mdev) - 1;
5955 return mlx5_ib_add_slave_port(mdev, port_num);
5958 dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
5963 dev->num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
5964 MLX5_CAP_GEN(mdev, num_vhca_ports));
5966 if (MLX5_VPORT_MANAGER(mdev) &&
5967 mlx5_ib_eswitch_mode(mdev->priv.eswitch) == SRIOV_OFFLOADS) {
5968 dev->rep = mlx5_ib_vport_rep(mdev->priv.eswitch, 0);
5970 return __mlx5_ib_add(dev, &nic_rep_profile);
5973 return __mlx5_ib_add(dev, &pf_profile);
5976 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
5978 struct mlx5_ib_multiport_info *mpi;
5979 struct mlx5_ib_dev *dev;
5981 if (mlx5_core_is_mp_slave(mdev)) {
5983 mutex_lock(&mlx5_ib_multiport_mutex);
5985 mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
5986 list_del(&mpi->list);
5987 mutex_unlock(&mlx5_ib_multiport_mutex);
5992 __mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
5995 static struct mlx5_interface mlx5_ib_interface = {
5997 .remove = mlx5_ib_remove,
5998 .event = mlx5_ib_event,
5999 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
6000 .pfault = mlx5_ib_pfault,
6002 .protocol = MLX5_INTERFACE_PROTOCOL_IB,
6005 unsigned long mlx5_ib_get_xlt_emergency_page(void)
6007 mutex_lock(&xlt_emergency_page_mutex);
6008 return xlt_emergency_page;
6011 void mlx5_ib_put_xlt_emergency_page(void)
6013 mutex_unlock(&xlt_emergency_page_mutex);
6016 static int __init mlx5_ib_init(void)
6020 xlt_emergency_page = __get_free_page(GFP_KERNEL);
6021 if (!xlt_emergency_page)
6024 mutex_init(&xlt_emergency_page_mutex);
6026 mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
6027 if (!mlx5_ib_event_wq) {
6028 free_page(xlt_emergency_page);
6034 err = mlx5_register_interface(&mlx5_ib_interface);
6039 static void __exit mlx5_ib_cleanup(void)
6041 mlx5_unregister_interface(&mlx5_ib_interface);
6042 destroy_workqueue(mlx5_ib_event_wq);
6043 mutex_destroy(&xlt_emergency_page_mutex);
6044 free_page(xlt_emergency_page);
6047 module_init(mlx5_ib_init);
6048 module_exit(mlx5_ib_cleanup);