xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision cd196fe2)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
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:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
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.
22  *
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
30  * SOFTWARE.
31  */
32 
33 #include <linux/highmem.h>
34 #include <linux/module.h>
35 #include <linux/init.h>
36 #include <linux/errno.h>
37 #include <linux/pci.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/slab.h>
40 #if defined(CONFIG_X86)
41 #include <asm/pat.h>
42 #endif
43 #include <linux/sched.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/task.h>
46 #include <linux/delay.h>
47 #include <rdma/ib_user_verbs.h>
48 #include <rdma/ib_addr.h>
49 #include <rdma/ib_cache.h>
50 #include <linux/mlx5/port.h>
51 #include <linux/mlx5/vport.h>
52 #include <linux/list.h>
53 #include <rdma/ib_smi.h>
54 #include <rdma/ib_umem.h>
55 #include <linux/in.h>
56 #include <linux/etherdevice.h>
57 #include <linux/mlx5/fs.h>
58 #include <linux/mlx5/vport.h>
59 #include "mlx5_ib.h"
60 #include "cmd.h"
61 
62 #define DRIVER_NAME "mlx5_ib"
63 #define DRIVER_VERSION "5.0-0"
64 
65 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
66 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
67 MODULE_LICENSE("Dual BSD/GPL");
68 MODULE_VERSION(DRIVER_VERSION);
69 
70 static char mlx5_version[] =
71 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
72 	DRIVER_VERSION "\n";
73 
74 enum {
75 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
76 };
77 
78 static enum rdma_link_layer
79 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
80 {
81 	switch (port_type_cap) {
82 	case MLX5_CAP_PORT_TYPE_IB:
83 		return IB_LINK_LAYER_INFINIBAND;
84 	case MLX5_CAP_PORT_TYPE_ETH:
85 		return IB_LINK_LAYER_ETHERNET;
86 	default:
87 		return IB_LINK_LAYER_UNSPECIFIED;
88 	}
89 }
90 
91 static enum rdma_link_layer
92 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
93 {
94 	struct mlx5_ib_dev *dev = to_mdev(device);
95 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
96 
97 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
98 }
99 
100 static int mlx5_netdev_event(struct notifier_block *this,
101 			     unsigned long event, void *ptr)
102 {
103 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
104 	struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
105 						 roce.nb);
106 
107 	switch (event) {
108 	case NETDEV_REGISTER:
109 	case NETDEV_UNREGISTER:
110 		write_lock(&ibdev->roce.netdev_lock);
111 		if (ndev->dev.parent == &ibdev->mdev->pdev->dev)
112 			ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
113 					     NULL : ndev;
114 		write_unlock(&ibdev->roce.netdev_lock);
115 		break;
116 
117 	case NETDEV_UP:
118 	case NETDEV_DOWN: {
119 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
120 		struct net_device *upper = NULL;
121 
122 		if (lag_ndev) {
123 			upper = netdev_master_upper_dev_get(lag_ndev);
124 			dev_put(lag_ndev);
125 		}
126 
127 		if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
128 		    && ibdev->ib_active) {
129 			struct ib_event ibev = { };
130 
131 			ibev.device = &ibdev->ib_dev;
132 			ibev.event = (event == NETDEV_UP) ?
133 				     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
134 			ibev.element.port_num = 1;
135 			ib_dispatch_event(&ibev);
136 		}
137 		break;
138 	}
139 
140 	default:
141 		break;
142 	}
143 
144 	return NOTIFY_DONE;
145 }
146 
147 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
148 					     u8 port_num)
149 {
150 	struct mlx5_ib_dev *ibdev = to_mdev(device);
151 	struct net_device *ndev;
152 
153 	ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
154 	if (ndev)
155 		return ndev;
156 
157 	/* Ensure ndev does not disappear before we invoke dev_hold()
158 	 */
159 	read_lock(&ibdev->roce.netdev_lock);
160 	ndev = ibdev->roce.netdev;
161 	if (ndev)
162 		dev_hold(ndev);
163 	read_unlock(&ibdev->roce.netdev_lock);
164 
165 	return ndev;
166 }
167 
168 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
169 				    u8 *active_width)
170 {
171 	switch (eth_proto_oper) {
172 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
173 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
174 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
175 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
176 		*active_width = IB_WIDTH_1X;
177 		*active_speed = IB_SPEED_SDR;
178 		break;
179 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
180 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
181 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
182 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
183 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
184 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
185 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
186 		*active_width = IB_WIDTH_1X;
187 		*active_speed = IB_SPEED_QDR;
188 		break;
189 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
190 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
191 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
192 		*active_width = IB_WIDTH_1X;
193 		*active_speed = IB_SPEED_EDR;
194 		break;
195 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
196 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
197 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
198 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
199 		*active_width = IB_WIDTH_4X;
200 		*active_speed = IB_SPEED_QDR;
201 		break;
202 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
203 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
204 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
205 		*active_width = IB_WIDTH_1X;
206 		*active_speed = IB_SPEED_HDR;
207 		break;
208 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
209 		*active_width = IB_WIDTH_4X;
210 		*active_speed = IB_SPEED_FDR;
211 		break;
212 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
213 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
214 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
215 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
216 		*active_width = IB_WIDTH_4X;
217 		*active_speed = IB_SPEED_EDR;
218 		break;
219 	default:
220 		return -EINVAL;
221 	}
222 
223 	return 0;
224 }
225 
226 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
227 				struct ib_port_attr *props)
228 {
229 	struct mlx5_ib_dev *dev = to_mdev(device);
230 	struct mlx5_core_dev *mdev = dev->mdev;
231 	struct net_device *ndev, *upper;
232 	enum ib_mtu ndev_ib_mtu;
233 	u16 qkey_viol_cntr;
234 	u32 eth_prot_oper;
235 	int err;
236 
237 	/* Possible bad flows are checked before filling out props so in case
238 	 * of an error it will still be zeroed out.
239 	 */
240 	err = mlx5_query_port_eth_proto_oper(mdev, &eth_prot_oper, port_num);
241 	if (err)
242 		return err;
243 
244 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
245 				 &props->active_width);
246 
247 	props->port_cap_flags  |= IB_PORT_CM_SUP;
248 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
249 
250 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
251 						roce_address_table_size);
252 	props->max_mtu          = IB_MTU_4096;
253 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
254 	props->pkey_tbl_len     = 1;
255 	props->state            = IB_PORT_DOWN;
256 	props->phys_state       = 3;
257 
258 	mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
259 	props->qkey_viol_cntr = qkey_viol_cntr;
260 
261 	ndev = mlx5_ib_get_netdev(device, port_num);
262 	if (!ndev)
263 		return 0;
264 
265 	if (mlx5_lag_is_active(dev->mdev)) {
266 		rcu_read_lock();
267 		upper = netdev_master_upper_dev_get_rcu(ndev);
268 		if (upper) {
269 			dev_put(ndev);
270 			ndev = upper;
271 			dev_hold(ndev);
272 		}
273 		rcu_read_unlock();
274 	}
275 
276 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
277 		props->state      = IB_PORT_ACTIVE;
278 		props->phys_state = 5;
279 	}
280 
281 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
282 
283 	dev_put(ndev);
284 
285 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
286 	return 0;
287 }
288 
289 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
290 			 unsigned int index, const union ib_gid *gid,
291 			 const struct ib_gid_attr *attr)
292 {
293 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
294 	u8 roce_version = 0;
295 	u8 roce_l3_type = 0;
296 	bool vlan = false;
297 	u8 mac[ETH_ALEN];
298 	u16 vlan_id = 0;
299 
300 	if (gid) {
301 		gid_type = attr->gid_type;
302 		ether_addr_copy(mac, attr->ndev->dev_addr);
303 
304 		if (is_vlan_dev(attr->ndev)) {
305 			vlan = true;
306 			vlan_id = vlan_dev_vlan_id(attr->ndev);
307 		}
308 	}
309 
310 	switch (gid_type) {
311 	case IB_GID_TYPE_IB:
312 		roce_version = MLX5_ROCE_VERSION_1;
313 		break;
314 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
315 		roce_version = MLX5_ROCE_VERSION_2;
316 		if (ipv6_addr_v4mapped((void *)gid))
317 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
318 		else
319 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
320 		break;
321 
322 	default:
323 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
324 	}
325 
326 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
327 				      roce_l3_type, gid->raw, mac, vlan,
328 				      vlan_id);
329 }
330 
331 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
332 			   unsigned int index, const union ib_gid *gid,
333 			   const struct ib_gid_attr *attr,
334 			   __always_unused void **context)
335 {
336 	return set_roce_addr(to_mdev(device), port_num, index, gid, attr);
337 }
338 
339 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
340 			   unsigned int index, __always_unused void **context)
341 {
342 	return set_roce_addr(to_mdev(device), port_num, index, NULL, NULL);
343 }
344 
345 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
346 			       int index)
347 {
348 	struct ib_gid_attr attr;
349 	union ib_gid gid;
350 
351 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
352 		return 0;
353 
354 	if (!attr.ndev)
355 		return 0;
356 
357 	dev_put(attr.ndev);
358 
359 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
360 		return 0;
361 
362 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
363 }
364 
365 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
366 			   int index, enum ib_gid_type *gid_type)
367 {
368 	struct ib_gid_attr attr;
369 	union ib_gid gid;
370 	int ret;
371 
372 	ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
373 	if (ret)
374 		return ret;
375 
376 	if (!attr.ndev)
377 		return -ENODEV;
378 
379 	dev_put(attr.ndev);
380 
381 	*gid_type = attr.gid_type;
382 
383 	return 0;
384 }
385 
386 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
387 {
388 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
389 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
390 	return 0;
391 }
392 
393 enum {
394 	MLX5_VPORT_ACCESS_METHOD_MAD,
395 	MLX5_VPORT_ACCESS_METHOD_HCA,
396 	MLX5_VPORT_ACCESS_METHOD_NIC,
397 };
398 
399 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
400 {
401 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
402 		return MLX5_VPORT_ACCESS_METHOD_MAD;
403 
404 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
405 	    IB_LINK_LAYER_ETHERNET)
406 		return MLX5_VPORT_ACCESS_METHOD_NIC;
407 
408 	return MLX5_VPORT_ACCESS_METHOD_HCA;
409 }
410 
411 static void get_atomic_caps(struct mlx5_ib_dev *dev,
412 			    struct ib_device_attr *props)
413 {
414 	u8 tmp;
415 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
416 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
417 	u8 atomic_req_8B_endianness_mode =
418 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
419 
420 	/* Check if HW supports 8 bytes standard atomic operations and capable
421 	 * of host endianness respond
422 	 */
423 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
424 	if (((atomic_operations & tmp) == tmp) &&
425 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
426 	    (atomic_req_8B_endianness_mode)) {
427 		props->atomic_cap = IB_ATOMIC_HCA;
428 	} else {
429 		props->atomic_cap = IB_ATOMIC_NONE;
430 	}
431 }
432 
433 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
434 					__be64 *sys_image_guid)
435 {
436 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
437 	struct mlx5_core_dev *mdev = dev->mdev;
438 	u64 tmp;
439 	int err;
440 
441 	switch (mlx5_get_vport_access_method(ibdev)) {
442 	case MLX5_VPORT_ACCESS_METHOD_MAD:
443 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
444 							    sys_image_guid);
445 
446 	case MLX5_VPORT_ACCESS_METHOD_HCA:
447 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
448 		break;
449 
450 	case MLX5_VPORT_ACCESS_METHOD_NIC:
451 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
452 		break;
453 
454 	default:
455 		return -EINVAL;
456 	}
457 
458 	if (!err)
459 		*sys_image_guid = cpu_to_be64(tmp);
460 
461 	return err;
462 
463 }
464 
465 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
466 				u16 *max_pkeys)
467 {
468 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
469 	struct mlx5_core_dev *mdev = dev->mdev;
470 
471 	switch (mlx5_get_vport_access_method(ibdev)) {
472 	case MLX5_VPORT_ACCESS_METHOD_MAD:
473 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
474 
475 	case MLX5_VPORT_ACCESS_METHOD_HCA:
476 	case MLX5_VPORT_ACCESS_METHOD_NIC:
477 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
478 						pkey_table_size));
479 		return 0;
480 
481 	default:
482 		return -EINVAL;
483 	}
484 }
485 
486 static int mlx5_query_vendor_id(struct ib_device *ibdev,
487 				u32 *vendor_id)
488 {
489 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
490 
491 	switch (mlx5_get_vport_access_method(ibdev)) {
492 	case MLX5_VPORT_ACCESS_METHOD_MAD:
493 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
494 
495 	case MLX5_VPORT_ACCESS_METHOD_HCA:
496 	case MLX5_VPORT_ACCESS_METHOD_NIC:
497 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
498 
499 	default:
500 		return -EINVAL;
501 	}
502 }
503 
504 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
505 				__be64 *node_guid)
506 {
507 	u64 tmp;
508 	int err;
509 
510 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
511 	case MLX5_VPORT_ACCESS_METHOD_MAD:
512 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
513 
514 	case MLX5_VPORT_ACCESS_METHOD_HCA:
515 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
516 		break;
517 
518 	case MLX5_VPORT_ACCESS_METHOD_NIC:
519 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
520 		break;
521 
522 	default:
523 		return -EINVAL;
524 	}
525 
526 	if (!err)
527 		*node_guid = cpu_to_be64(tmp);
528 
529 	return err;
530 }
531 
532 struct mlx5_reg_node_desc {
533 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
534 };
535 
536 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
537 {
538 	struct mlx5_reg_node_desc in;
539 
540 	if (mlx5_use_mad_ifc(dev))
541 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
542 
543 	memset(&in, 0, sizeof(in));
544 
545 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
546 				    sizeof(struct mlx5_reg_node_desc),
547 				    MLX5_REG_NODE_DESC, 0, 0);
548 }
549 
550 static int mlx5_ib_query_device(struct ib_device *ibdev,
551 				struct ib_device_attr *props,
552 				struct ib_udata *uhw)
553 {
554 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
555 	struct mlx5_core_dev *mdev = dev->mdev;
556 	int err = -ENOMEM;
557 	int max_sq_desc;
558 	int max_rq_sg;
559 	int max_sq_sg;
560 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
561 	struct mlx5_ib_query_device_resp resp = {};
562 	size_t resp_len;
563 	u64 max_tso;
564 
565 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
566 	if (uhw->outlen && uhw->outlen < resp_len)
567 		return -EINVAL;
568 	else
569 		resp.response_length = resp_len;
570 
571 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
572 		return -EINVAL;
573 
574 	memset(props, 0, sizeof(*props));
575 	err = mlx5_query_system_image_guid(ibdev,
576 					   &props->sys_image_guid);
577 	if (err)
578 		return err;
579 
580 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
581 	if (err)
582 		return err;
583 
584 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
585 	if (err)
586 		return err;
587 
588 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
589 		(fw_rev_min(dev->mdev) << 16) |
590 		fw_rev_sub(dev->mdev);
591 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
592 		IB_DEVICE_PORT_ACTIVE_EVENT		|
593 		IB_DEVICE_SYS_IMAGE_GUID		|
594 		IB_DEVICE_RC_RNR_NAK_GEN;
595 
596 	if (MLX5_CAP_GEN(mdev, pkv))
597 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
598 	if (MLX5_CAP_GEN(mdev, qkv))
599 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
600 	if (MLX5_CAP_GEN(mdev, apm))
601 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
602 	if (MLX5_CAP_GEN(mdev, xrc))
603 		props->device_cap_flags |= IB_DEVICE_XRC;
604 	if (MLX5_CAP_GEN(mdev, imaicl)) {
605 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
606 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
607 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
608 		/* We support 'Gappy' memory registration too */
609 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
610 	}
611 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
612 	if (MLX5_CAP_GEN(mdev, sho)) {
613 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
614 		/* At this stage no support for signature handover */
615 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
616 				      IB_PROT_T10DIF_TYPE_2 |
617 				      IB_PROT_T10DIF_TYPE_3;
618 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
619 				       IB_GUARD_T10DIF_CSUM;
620 	}
621 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
622 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
623 
624 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
625 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
626 			/* Legacy bit to support old userspace libraries */
627 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
628 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
629 		}
630 
631 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
632 			props->raw_packet_caps |=
633 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
634 
635 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
636 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
637 			if (max_tso) {
638 				resp.tso_caps.max_tso = 1 << max_tso;
639 				resp.tso_caps.supported_qpts |=
640 					1 << IB_QPT_RAW_PACKET;
641 				resp.response_length += sizeof(resp.tso_caps);
642 			}
643 		}
644 
645 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
646 			resp.rss_caps.rx_hash_function =
647 						MLX5_RX_HASH_FUNC_TOEPLITZ;
648 			resp.rss_caps.rx_hash_fields_mask =
649 						MLX5_RX_HASH_SRC_IPV4 |
650 						MLX5_RX_HASH_DST_IPV4 |
651 						MLX5_RX_HASH_SRC_IPV6 |
652 						MLX5_RX_HASH_DST_IPV6 |
653 						MLX5_RX_HASH_SRC_PORT_TCP |
654 						MLX5_RX_HASH_DST_PORT_TCP |
655 						MLX5_RX_HASH_SRC_PORT_UDP |
656 						MLX5_RX_HASH_DST_PORT_UDP;
657 			resp.response_length += sizeof(resp.rss_caps);
658 		}
659 	} else {
660 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
661 			resp.response_length += sizeof(resp.tso_caps);
662 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
663 			resp.response_length += sizeof(resp.rss_caps);
664 	}
665 
666 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
667 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
668 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
669 	}
670 
671 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
672 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
673 		/* Legacy bit to support old userspace libraries */
674 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
675 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
676 	}
677 
678 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
679 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
680 
681 	props->vendor_part_id	   = mdev->pdev->device;
682 	props->hw_ver		   = mdev->pdev->revision;
683 
684 	props->max_mr_size	   = ~0ull;
685 	props->page_size_cap	   = ~(min_page_size - 1);
686 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
687 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
688 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
689 		     sizeof(struct mlx5_wqe_data_seg);
690 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
691 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
692 		     sizeof(struct mlx5_wqe_raddr_seg)) /
693 		sizeof(struct mlx5_wqe_data_seg);
694 	props->max_sge = min(max_rq_sg, max_sq_sg);
695 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
696 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
697 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
698 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
699 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
700 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
701 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
702 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
703 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
704 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
705 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
706 	props->max_srq_sge	   = max_rq_sg - 1;
707 	props->max_fast_reg_page_list_len =
708 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
709 	get_atomic_caps(dev, props);
710 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
711 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
712 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
713 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
714 					   props->max_mcast_grp;
715 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
716 	props->max_ah = INT_MAX;
717 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
718 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
719 
720 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
721 	if (MLX5_CAP_GEN(mdev, pg))
722 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
723 	props->odp_caps = dev->odp_caps;
724 #endif
725 
726 	if (MLX5_CAP_GEN(mdev, cd))
727 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
728 
729 	if (!mlx5_core_is_pf(mdev))
730 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
731 
732 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
733 	    IB_LINK_LAYER_ETHERNET) {
734 		props->rss_caps.max_rwq_indirection_tables =
735 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
736 		props->rss_caps.max_rwq_indirection_table_size =
737 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
738 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
739 		props->max_wq_type_rq =
740 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
741 	}
742 
743 	if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
744 		resp.cqe_comp_caps.max_num =
745 			MLX5_CAP_GEN(dev->mdev, cqe_compression) ?
746 			MLX5_CAP_GEN(dev->mdev, cqe_compression_max_num) : 0;
747 		resp.cqe_comp_caps.supported_format =
748 			MLX5_IB_CQE_RES_FORMAT_HASH |
749 			MLX5_IB_CQE_RES_FORMAT_CSUM;
750 		resp.response_length += sizeof(resp.cqe_comp_caps);
751 	}
752 
753 	if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen)) {
754 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
755 		    MLX5_CAP_GEN(mdev, qos)) {
756 			resp.packet_pacing_caps.qp_rate_limit_max =
757 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
758 			resp.packet_pacing_caps.qp_rate_limit_min =
759 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
760 			resp.packet_pacing_caps.supported_qpts |=
761 				1 << IB_QPT_RAW_PACKET;
762 		}
763 		resp.response_length += sizeof(resp.packet_pacing_caps);
764 	}
765 
766 	if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
767 			uhw->outlen)) {
768 		resp.mlx5_ib_support_multi_pkt_send_wqes =
769 			MLX5_CAP_ETH(mdev, multi_pkt_send_wqe);
770 		resp.response_length +=
771 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
772 	}
773 
774 	if (field_avail(typeof(resp), reserved, uhw->outlen))
775 		resp.response_length += sizeof(resp.reserved);
776 
777 	if (uhw->outlen) {
778 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
779 
780 		if (err)
781 			return err;
782 	}
783 
784 	return 0;
785 }
786 
787 enum mlx5_ib_width {
788 	MLX5_IB_WIDTH_1X	= 1 << 0,
789 	MLX5_IB_WIDTH_2X	= 1 << 1,
790 	MLX5_IB_WIDTH_4X	= 1 << 2,
791 	MLX5_IB_WIDTH_8X	= 1 << 3,
792 	MLX5_IB_WIDTH_12X	= 1 << 4
793 };
794 
795 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
796 				  u8 *ib_width)
797 {
798 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
799 	int err = 0;
800 
801 	if (active_width & MLX5_IB_WIDTH_1X) {
802 		*ib_width = IB_WIDTH_1X;
803 	} else if (active_width & MLX5_IB_WIDTH_2X) {
804 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
805 			    (int)active_width);
806 		err = -EINVAL;
807 	} else if (active_width & MLX5_IB_WIDTH_4X) {
808 		*ib_width = IB_WIDTH_4X;
809 	} else if (active_width & MLX5_IB_WIDTH_8X) {
810 		*ib_width = IB_WIDTH_8X;
811 	} else if (active_width & MLX5_IB_WIDTH_12X) {
812 		*ib_width = IB_WIDTH_12X;
813 	} else {
814 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
815 			    (int)active_width);
816 		err = -EINVAL;
817 	}
818 
819 	return err;
820 }
821 
822 static int mlx5_mtu_to_ib_mtu(int mtu)
823 {
824 	switch (mtu) {
825 	case 256: return 1;
826 	case 512: return 2;
827 	case 1024: return 3;
828 	case 2048: return 4;
829 	case 4096: return 5;
830 	default:
831 		pr_warn("invalid mtu\n");
832 		return -1;
833 	}
834 }
835 
836 enum ib_max_vl_num {
837 	__IB_MAX_VL_0		= 1,
838 	__IB_MAX_VL_0_1		= 2,
839 	__IB_MAX_VL_0_3		= 3,
840 	__IB_MAX_VL_0_7		= 4,
841 	__IB_MAX_VL_0_14	= 5,
842 };
843 
844 enum mlx5_vl_hw_cap {
845 	MLX5_VL_HW_0	= 1,
846 	MLX5_VL_HW_0_1	= 2,
847 	MLX5_VL_HW_0_2	= 3,
848 	MLX5_VL_HW_0_3	= 4,
849 	MLX5_VL_HW_0_4	= 5,
850 	MLX5_VL_HW_0_5	= 6,
851 	MLX5_VL_HW_0_6	= 7,
852 	MLX5_VL_HW_0_7	= 8,
853 	MLX5_VL_HW_0_14	= 15
854 };
855 
856 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
857 				u8 *max_vl_num)
858 {
859 	switch (vl_hw_cap) {
860 	case MLX5_VL_HW_0:
861 		*max_vl_num = __IB_MAX_VL_0;
862 		break;
863 	case MLX5_VL_HW_0_1:
864 		*max_vl_num = __IB_MAX_VL_0_1;
865 		break;
866 	case MLX5_VL_HW_0_3:
867 		*max_vl_num = __IB_MAX_VL_0_3;
868 		break;
869 	case MLX5_VL_HW_0_7:
870 		*max_vl_num = __IB_MAX_VL_0_7;
871 		break;
872 	case MLX5_VL_HW_0_14:
873 		*max_vl_num = __IB_MAX_VL_0_14;
874 		break;
875 
876 	default:
877 		return -EINVAL;
878 	}
879 
880 	return 0;
881 }
882 
883 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
884 			       struct ib_port_attr *props)
885 {
886 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
887 	struct mlx5_core_dev *mdev = dev->mdev;
888 	struct mlx5_hca_vport_context *rep;
889 	u16 max_mtu;
890 	u16 oper_mtu;
891 	int err;
892 	u8 ib_link_width_oper;
893 	u8 vl_hw_cap;
894 
895 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
896 	if (!rep) {
897 		err = -ENOMEM;
898 		goto out;
899 	}
900 
901 	/* props being zeroed by the caller, avoid zeroing it here */
902 
903 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
904 	if (err)
905 		goto out;
906 
907 	props->lid		= rep->lid;
908 	props->lmc		= rep->lmc;
909 	props->sm_lid		= rep->sm_lid;
910 	props->sm_sl		= rep->sm_sl;
911 	props->state		= rep->vport_state;
912 	props->phys_state	= rep->port_physical_state;
913 	props->port_cap_flags	= rep->cap_mask1;
914 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
915 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
916 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
917 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
918 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
919 	props->subnet_timeout	= rep->subnet_timeout;
920 	props->init_type_reply	= rep->init_type_reply;
921 	props->grh_required	= rep->grh_required;
922 
923 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
924 	if (err)
925 		goto out;
926 
927 	err = translate_active_width(ibdev, ib_link_width_oper,
928 				     &props->active_width);
929 	if (err)
930 		goto out;
931 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
932 	if (err)
933 		goto out;
934 
935 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
936 
937 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
938 
939 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
940 
941 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
942 
943 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
944 	if (err)
945 		goto out;
946 
947 	err = translate_max_vl_num(ibdev, vl_hw_cap,
948 				   &props->max_vl_num);
949 out:
950 	kfree(rep);
951 	return err;
952 }
953 
954 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
955 		       struct ib_port_attr *props)
956 {
957 	unsigned int count;
958 	int ret;
959 
960 	switch (mlx5_get_vport_access_method(ibdev)) {
961 	case MLX5_VPORT_ACCESS_METHOD_MAD:
962 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
963 		break;
964 
965 	case MLX5_VPORT_ACCESS_METHOD_HCA:
966 		ret = mlx5_query_hca_port(ibdev, port, props);
967 		break;
968 
969 	case MLX5_VPORT_ACCESS_METHOD_NIC:
970 		ret = mlx5_query_port_roce(ibdev, port, props);
971 		break;
972 
973 	default:
974 		ret = -EINVAL;
975 	}
976 
977 	if (!ret && props) {
978 		count = mlx5_core_reserved_gids_count(to_mdev(ibdev)->mdev);
979 		props->gid_tbl_len -= count;
980 	}
981 	return ret;
982 }
983 
984 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
985 			     union ib_gid *gid)
986 {
987 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
988 	struct mlx5_core_dev *mdev = dev->mdev;
989 
990 	switch (mlx5_get_vport_access_method(ibdev)) {
991 	case MLX5_VPORT_ACCESS_METHOD_MAD:
992 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
993 
994 	case MLX5_VPORT_ACCESS_METHOD_HCA:
995 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
996 
997 	default:
998 		return -EINVAL;
999 	}
1000 
1001 }
1002 
1003 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1004 			      u16 *pkey)
1005 {
1006 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1007 	struct mlx5_core_dev *mdev = dev->mdev;
1008 
1009 	switch (mlx5_get_vport_access_method(ibdev)) {
1010 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1011 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1012 
1013 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1014 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1015 		return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
1016 						 pkey);
1017 	default:
1018 		return -EINVAL;
1019 	}
1020 }
1021 
1022 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1023 				 struct ib_device_modify *props)
1024 {
1025 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1026 	struct mlx5_reg_node_desc in;
1027 	struct mlx5_reg_node_desc out;
1028 	int err;
1029 
1030 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1031 		return -EOPNOTSUPP;
1032 
1033 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1034 		return 0;
1035 
1036 	/*
1037 	 * If possible, pass node desc to FW, so it can generate
1038 	 * a 144 trap.  If cmd fails, just ignore.
1039 	 */
1040 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1041 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1042 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1043 	if (err)
1044 		return err;
1045 
1046 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1047 
1048 	return err;
1049 }
1050 
1051 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1052 				u32 value)
1053 {
1054 	struct mlx5_hca_vport_context ctx = {};
1055 	int err;
1056 
1057 	err = mlx5_query_hca_vport_context(dev->mdev, 0,
1058 					   port_num, 0, &ctx);
1059 	if (err)
1060 		return err;
1061 
1062 	if (~ctx.cap_mask1_perm & mask) {
1063 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1064 			     mask, ctx.cap_mask1_perm);
1065 		return -EINVAL;
1066 	}
1067 
1068 	ctx.cap_mask1 = value;
1069 	ctx.cap_mask1_perm = mask;
1070 	err = mlx5_core_modify_hca_vport_context(dev->mdev, 0,
1071 						 port_num, 0, &ctx);
1072 
1073 	return err;
1074 }
1075 
1076 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1077 			       struct ib_port_modify *props)
1078 {
1079 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1080 	struct ib_port_attr attr;
1081 	u32 tmp;
1082 	int err;
1083 	u32 change_mask;
1084 	u32 value;
1085 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1086 		      IB_LINK_LAYER_INFINIBAND);
1087 
1088 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1089 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1090 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1091 		return set_port_caps_atomic(dev, port, change_mask, value);
1092 	}
1093 
1094 	mutex_lock(&dev->cap_mask_mutex);
1095 
1096 	err = ib_query_port(ibdev, port, &attr);
1097 	if (err)
1098 		goto out;
1099 
1100 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1101 		~props->clr_port_cap_mask;
1102 
1103 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1104 
1105 out:
1106 	mutex_unlock(&dev->cap_mask_mutex);
1107 	return err;
1108 }
1109 
1110 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1111 {
1112 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1113 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1114 }
1115 
1116 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1117 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1118 			     u32 *num_sys_pages)
1119 {
1120 	int uars_per_sys_page;
1121 	int bfregs_per_sys_page;
1122 	int ref_bfregs = req->total_num_bfregs;
1123 
1124 	if (req->total_num_bfregs == 0)
1125 		return -EINVAL;
1126 
1127 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1128 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1129 
1130 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1131 		return -ENOMEM;
1132 
1133 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1134 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1135 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1136 	*num_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1137 
1138 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1139 		return -EINVAL;
1140 
1141 	mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, alloated %d, using %d sys pages\n",
1142 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1143 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1144 		    req->total_num_bfregs, *num_sys_pages);
1145 
1146 	return 0;
1147 }
1148 
1149 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1150 {
1151 	struct mlx5_bfreg_info *bfregi;
1152 	int err;
1153 	int i;
1154 
1155 	bfregi = &context->bfregi;
1156 	for (i = 0; i < bfregi->num_sys_pages; i++) {
1157 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1158 		if (err)
1159 			goto error;
1160 
1161 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1162 	}
1163 	return 0;
1164 
1165 error:
1166 	for (--i; i >= 0; i--)
1167 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1168 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1169 
1170 	return err;
1171 }
1172 
1173 static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1174 {
1175 	struct mlx5_bfreg_info *bfregi;
1176 	int err;
1177 	int i;
1178 
1179 	bfregi = &context->bfregi;
1180 	for (i = 0; i < bfregi->num_sys_pages; i++) {
1181 		err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1182 		if (err) {
1183 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1184 			return err;
1185 		}
1186 	}
1187 	return 0;
1188 }
1189 
1190 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1191 						  struct ib_udata *udata)
1192 {
1193 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1194 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1195 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1196 	struct mlx5_ib_ucontext *context;
1197 	struct mlx5_bfreg_info *bfregi;
1198 	int ver;
1199 	int err;
1200 	size_t reqlen;
1201 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1202 				     max_cqe_version);
1203 	bool lib_uar_4k;
1204 
1205 	if (!dev->ib_active)
1206 		return ERR_PTR(-EAGAIN);
1207 
1208 	if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
1209 		return ERR_PTR(-EINVAL);
1210 
1211 	reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
1212 	if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1213 		ver = 0;
1214 	else if (reqlen >= min_req_v2)
1215 		ver = 2;
1216 	else
1217 		return ERR_PTR(-EINVAL);
1218 
1219 	err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
1220 	if (err)
1221 		return ERR_PTR(err);
1222 
1223 	if (req.flags)
1224 		return ERR_PTR(-EINVAL);
1225 
1226 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1227 		return ERR_PTR(-EOPNOTSUPP);
1228 
1229 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1230 				    MLX5_NON_FP_BFREGS_PER_UAR);
1231 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1232 		return ERR_PTR(-EINVAL);
1233 
1234 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1235 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1236 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1237 	resp.cache_line_size = cache_line_size();
1238 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1239 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1240 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1241 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1242 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1243 	resp.cqe_version = min_t(__u8,
1244 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1245 				 req.max_cqe_version);
1246 	resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1247 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1248 	resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1249 					MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1250 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1251 				   sizeof(resp.response_length), udata->outlen);
1252 
1253 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1254 	if (!context)
1255 		return ERR_PTR(-ENOMEM);
1256 
1257 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1258 	bfregi = &context->bfregi;
1259 
1260 	/* updates req->total_num_bfregs */
1261 	err = calc_total_bfregs(dev, lib_uar_4k, &req, &bfregi->num_sys_pages);
1262 	if (err)
1263 		goto out_ctx;
1264 
1265 	mutex_init(&bfregi->lock);
1266 	bfregi->lib_uar_4k = lib_uar_4k;
1267 	bfregi->count = kcalloc(req.total_num_bfregs, sizeof(*bfregi->count),
1268 				GFP_KERNEL);
1269 	if (!bfregi->count) {
1270 		err = -ENOMEM;
1271 		goto out_ctx;
1272 	}
1273 
1274 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1275 				    sizeof(*bfregi->sys_pages),
1276 				    GFP_KERNEL);
1277 	if (!bfregi->sys_pages) {
1278 		err = -ENOMEM;
1279 		goto out_count;
1280 	}
1281 
1282 	err = allocate_uars(dev, context);
1283 	if (err)
1284 		goto out_sys_pages;
1285 
1286 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1287 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1288 #endif
1289 
1290 	context->upd_xlt_page = __get_free_page(GFP_KERNEL);
1291 	if (!context->upd_xlt_page) {
1292 		err = -ENOMEM;
1293 		goto out_uars;
1294 	}
1295 	mutex_init(&context->upd_xlt_page_mutex);
1296 
1297 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1298 		err = mlx5_core_alloc_transport_domain(dev->mdev,
1299 						       &context->tdn);
1300 		if (err)
1301 			goto out_page;
1302 	}
1303 
1304 	INIT_LIST_HEAD(&context->vma_private_list);
1305 	INIT_LIST_HEAD(&context->db_page_list);
1306 	mutex_init(&context->db_page_mutex);
1307 
1308 	resp.tot_bfregs = req.total_num_bfregs;
1309 	resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1310 
1311 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1312 		resp.response_length += sizeof(resp.cqe_version);
1313 
1314 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1315 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1316 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1317 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1318 	}
1319 
1320 	if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1321 		if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1322 			mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1323 			resp.eth_min_inline++;
1324 		}
1325 		resp.response_length += sizeof(resp.eth_min_inline);
1326 	}
1327 
1328 	/*
1329 	 * We don't want to expose information from the PCI bar that is located
1330 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1331 	 * pretend we don't support reading the HCA's core clock. This is also
1332 	 * forced by mmap function.
1333 	 */
1334 	if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1335 		if (PAGE_SIZE <= 4096) {
1336 			resp.comp_mask |=
1337 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1338 			resp.hca_core_clock_offset =
1339 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1340 		}
1341 		resp.response_length += sizeof(resp.hca_core_clock_offset) +
1342 					sizeof(resp.reserved2);
1343 	}
1344 
1345 	if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1346 		resp.response_length += sizeof(resp.log_uar_size);
1347 
1348 	if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1349 		resp.response_length += sizeof(resp.num_uars_per_page);
1350 
1351 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1352 	if (err)
1353 		goto out_td;
1354 
1355 	bfregi->ver = ver;
1356 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1357 	context->cqe_version = resp.cqe_version;
1358 	context->lib_caps = req.lib_caps;
1359 	print_lib_caps(dev, context->lib_caps);
1360 
1361 	return &context->ibucontext;
1362 
1363 out_td:
1364 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1365 		mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1366 
1367 out_page:
1368 	free_page(context->upd_xlt_page);
1369 
1370 out_uars:
1371 	deallocate_uars(dev, context);
1372 
1373 out_sys_pages:
1374 	kfree(bfregi->sys_pages);
1375 
1376 out_count:
1377 	kfree(bfregi->count);
1378 
1379 out_ctx:
1380 	kfree(context);
1381 
1382 	return ERR_PTR(err);
1383 }
1384 
1385 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1386 {
1387 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1388 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1389 	struct mlx5_bfreg_info *bfregi;
1390 
1391 	bfregi = &context->bfregi;
1392 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1393 		mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1394 
1395 	free_page(context->upd_xlt_page);
1396 	deallocate_uars(dev, context);
1397 	kfree(bfregi->sys_pages);
1398 	kfree(bfregi->count);
1399 	kfree(context);
1400 
1401 	return 0;
1402 }
1403 
1404 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1405 				 struct mlx5_bfreg_info *bfregi,
1406 				 int idx)
1407 {
1408 	int fw_uars_per_page;
1409 
1410 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1411 
1412 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) +
1413 			bfregi->sys_pages[idx] / fw_uars_per_page;
1414 }
1415 
1416 static int get_command(unsigned long offset)
1417 {
1418 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1419 }
1420 
1421 static int get_arg(unsigned long offset)
1422 {
1423 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1424 }
1425 
1426 static int get_index(unsigned long offset)
1427 {
1428 	return get_arg(offset);
1429 }
1430 
1431 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1432 {
1433 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1434 	 * is done through either mremap flow or split_vma (usually due to
1435 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1436 	 * as this VMA is strongly hardware related.  Therefore we set the
1437 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1438 	 * calling us again and trying to do incorrect actions.  We assume that
1439 	 * the original VMA size is exactly a single page, and therefore all
1440 	 * "splitting" operation will not happen to it.
1441 	 */
1442 	area->vm_ops = NULL;
1443 }
1444 
1445 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1446 {
1447 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1448 
1449 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1450 	 * file itself is closed, therefore no sync is needed with the regular
1451 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1452 	 * However need a sync with accessing the vma as part of
1453 	 * mlx5_ib_disassociate_ucontext.
1454 	 * The close operation is usually called under mm->mmap_sem except when
1455 	 * process is exiting.
1456 	 * The exiting case is handled explicitly as part of
1457 	 * mlx5_ib_disassociate_ucontext.
1458 	 */
1459 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1460 
1461 	/* setting the vma context pointer to null in the mlx5_ib driver's
1462 	 * private data, to protect a race condition in
1463 	 * mlx5_ib_disassociate_ucontext().
1464 	 */
1465 	mlx5_ib_vma_priv_data->vma = NULL;
1466 	list_del(&mlx5_ib_vma_priv_data->list);
1467 	kfree(mlx5_ib_vma_priv_data);
1468 }
1469 
1470 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1471 	.open = mlx5_ib_vma_open,
1472 	.close = mlx5_ib_vma_close
1473 };
1474 
1475 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1476 				struct mlx5_ib_ucontext *ctx)
1477 {
1478 	struct mlx5_ib_vma_private_data *vma_prv;
1479 	struct list_head *vma_head = &ctx->vma_private_list;
1480 
1481 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1482 	if (!vma_prv)
1483 		return -ENOMEM;
1484 
1485 	vma_prv->vma = vma;
1486 	vma->vm_private_data = vma_prv;
1487 	vma->vm_ops =  &mlx5_ib_vm_ops;
1488 
1489 	list_add(&vma_prv->list, vma_head);
1490 
1491 	return 0;
1492 }
1493 
1494 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1495 {
1496 	int ret;
1497 	struct vm_area_struct *vma;
1498 	struct mlx5_ib_vma_private_data *vma_private, *n;
1499 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1500 	struct task_struct *owning_process  = NULL;
1501 	struct mm_struct   *owning_mm       = NULL;
1502 
1503 	owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1504 	if (!owning_process)
1505 		return;
1506 
1507 	owning_mm = get_task_mm(owning_process);
1508 	if (!owning_mm) {
1509 		pr_info("no mm, disassociate ucontext is pending task termination\n");
1510 		while (1) {
1511 			put_task_struct(owning_process);
1512 			usleep_range(1000, 2000);
1513 			owning_process = get_pid_task(ibcontext->tgid,
1514 						      PIDTYPE_PID);
1515 			if (!owning_process ||
1516 			    owning_process->state == TASK_DEAD) {
1517 				pr_info("disassociate ucontext done, task was terminated\n");
1518 				/* in case task was dead need to release the
1519 				 * task struct.
1520 				 */
1521 				if (owning_process)
1522 					put_task_struct(owning_process);
1523 				return;
1524 			}
1525 		}
1526 	}
1527 
1528 	/* need to protect from a race on closing the vma as part of
1529 	 * mlx5_ib_vma_close.
1530 	 */
1531 	down_write(&owning_mm->mmap_sem);
1532 	list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1533 				 list) {
1534 		vma = vma_private->vma;
1535 		ret = zap_vma_ptes(vma, vma->vm_start,
1536 				   PAGE_SIZE);
1537 		WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1538 		/* context going to be destroyed, should
1539 		 * not access ops any more.
1540 		 */
1541 		vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
1542 		vma->vm_ops = NULL;
1543 		list_del(&vma_private->list);
1544 		kfree(vma_private);
1545 	}
1546 	up_write(&owning_mm->mmap_sem);
1547 	mmput(owning_mm);
1548 	put_task_struct(owning_process);
1549 }
1550 
1551 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1552 {
1553 	switch (cmd) {
1554 	case MLX5_IB_MMAP_WC_PAGE:
1555 		return "WC";
1556 	case MLX5_IB_MMAP_REGULAR_PAGE:
1557 		return "best effort WC";
1558 	case MLX5_IB_MMAP_NC_PAGE:
1559 		return "NC";
1560 	default:
1561 		return NULL;
1562 	}
1563 }
1564 
1565 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1566 		    struct vm_area_struct *vma,
1567 		    struct mlx5_ib_ucontext *context)
1568 {
1569 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1570 	int err;
1571 	unsigned long idx;
1572 	phys_addr_t pfn, pa;
1573 	pgprot_t prot;
1574 	int uars_per_page;
1575 
1576 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1577 		return -EINVAL;
1578 
1579 	uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
1580 	idx = get_index(vma->vm_pgoff);
1581 	if (idx % uars_per_page ||
1582 	    idx * uars_per_page >= bfregi->num_sys_pages) {
1583 		mlx5_ib_warn(dev, "invalid uar index %lu\n", idx);
1584 		return -EINVAL;
1585 	}
1586 
1587 	switch (cmd) {
1588 	case MLX5_IB_MMAP_WC_PAGE:
1589 /* Some architectures don't support WC memory */
1590 #if defined(CONFIG_X86)
1591 		if (!pat_enabled())
1592 			return -EPERM;
1593 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1594 			return -EPERM;
1595 #endif
1596 	/* fall through */
1597 	case MLX5_IB_MMAP_REGULAR_PAGE:
1598 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1599 		prot = pgprot_writecombine(vma->vm_page_prot);
1600 		break;
1601 	case MLX5_IB_MMAP_NC_PAGE:
1602 		prot = pgprot_noncached(vma->vm_page_prot);
1603 		break;
1604 	default:
1605 		return -EINVAL;
1606 	}
1607 
1608 	pfn = uar_index2pfn(dev, bfregi, idx);
1609 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1610 
1611 	vma->vm_page_prot = prot;
1612 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1613 				 PAGE_SIZE, vma->vm_page_prot);
1614 	if (err) {
1615 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
1616 			    err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
1617 		return -EAGAIN;
1618 	}
1619 
1620 	pa = pfn << PAGE_SHIFT;
1621 	mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
1622 		    vma->vm_start, &pa);
1623 
1624 	return mlx5_ib_set_vma_data(vma, context);
1625 }
1626 
1627 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1628 {
1629 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1630 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1631 	unsigned long command;
1632 	phys_addr_t pfn;
1633 
1634 	command = get_command(vma->vm_pgoff);
1635 	switch (command) {
1636 	case MLX5_IB_MMAP_WC_PAGE:
1637 	case MLX5_IB_MMAP_NC_PAGE:
1638 	case MLX5_IB_MMAP_REGULAR_PAGE:
1639 		return uar_mmap(dev, command, vma, context);
1640 
1641 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1642 		return -ENOSYS;
1643 
1644 	case MLX5_IB_MMAP_CORE_CLOCK:
1645 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1646 			return -EINVAL;
1647 
1648 		if (vma->vm_flags & VM_WRITE)
1649 			return -EPERM;
1650 
1651 		/* Don't expose to user-space information it shouldn't have */
1652 		if (PAGE_SIZE > 4096)
1653 			return -EOPNOTSUPP;
1654 
1655 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1656 		pfn = (dev->mdev->iseg_base +
1657 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1658 			PAGE_SHIFT;
1659 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1660 				       PAGE_SIZE, vma->vm_page_prot))
1661 			return -EAGAIN;
1662 
1663 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
1664 			    vma->vm_start,
1665 			    (unsigned long long)pfn << PAGE_SHIFT);
1666 		break;
1667 
1668 	default:
1669 		return -EINVAL;
1670 	}
1671 
1672 	return 0;
1673 }
1674 
1675 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1676 				      struct ib_ucontext *context,
1677 				      struct ib_udata *udata)
1678 {
1679 	struct mlx5_ib_alloc_pd_resp resp;
1680 	struct mlx5_ib_pd *pd;
1681 	int err;
1682 
1683 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1684 	if (!pd)
1685 		return ERR_PTR(-ENOMEM);
1686 
1687 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1688 	if (err) {
1689 		kfree(pd);
1690 		return ERR_PTR(err);
1691 	}
1692 
1693 	if (context) {
1694 		resp.pdn = pd->pdn;
1695 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1696 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1697 			kfree(pd);
1698 			return ERR_PTR(-EFAULT);
1699 		}
1700 	}
1701 
1702 	return &pd->ibpd;
1703 }
1704 
1705 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1706 {
1707 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1708 	struct mlx5_ib_pd *mpd = to_mpd(pd);
1709 
1710 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1711 	kfree(mpd);
1712 
1713 	return 0;
1714 }
1715 
1716 enum {
1717 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
1718 	MATCH_CRITERIA_ENABLE_MISC_BIT,
1719 	MATCH_CRITERIA_ENABLE_INNER_BIT
1720 };
1721 
1722 #define HEADER_IS_ZERO(match_criteria, headers)			           \
1723 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1724 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
1725 
1726 static u8 get_match_criteria_enable(u32 *match_criteria)
1727 {
1728 	u8 match_criteria_enable;
1729 
1730 	match_criteria_enable =
1731 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1732 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
1733 	match_criteria_enable |=
1734 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1735 		MATCH_CRITERIA_ENABLE_MISC_BIT;
1736 	match_criteria_enable |=
1737 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1738 		MATCH_CRITERIA_ENABLE_INNER_BIT;
1739 
1740 	return match_criteria_enable;
1741 }
1742 
1743 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1744 {
1745 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1746 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1747 }
1748 
1749 static void set_flow_label(void *misc_c, void *misc_v, u8 mask, u8 val,
1750 			   bool inner)
1751 {
1752 	if (inner) {
1753 		MLX5_SET(fte_match_set_misc,
1754 			 misc_c, inner_ipv6_flow_label, mask);
1755 		MLX5_SET(fte_match_set_misc,
1756 			 misc_v, inner_ipv6_flow_label, val);
1757 	} else {
1758 		MLX5_SET(fte_match_set_misc,
1759 			 misc_c, outer_ipv6_flow_label, mask);
1760 		MLX5_SET(fte_match_set_misc,
1761 			 misc_v, outer_ipv6_flow_label, val);
1762 	}
1763 }
1764 
1765 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1766 {
1767 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1768 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1769 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1770 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1771 }
1772 
1773 #define LAST_ETH_FIELD vlan_tag
1774 #define LAST_IB_FIELD sl
1775 #define LAST_IPV4_FIELD tos
1776 #define LAST_IPV6_FIELD traffic_class
1777 #define LAST_TCP_UDP_FIELD src_port
1778 #define LAST_TUNNEL_FIELD tunnel_id
1779 #define LAST_FLOW_TAG_FIELD tag_id
1780 #define LAST_DROP_FIELD size
1781 
1782 /* Field is the last supported field */
1783 #define FIELDS_NOT_SUPPORTED(filter, field)\
1784 	memchr_inv((void *)&filter.field  +\
1785 		   sizeof(filter.field), 0,\
1786 		   sizeof(filter) -\
1787 		   offsetof(typeof(filter), field) -\
1788 		   sizeof(filter.field))
1789 
1790 #define IPV4_VERSION 4
1791 #define IPV6_VERSION 6
1792 static int parse_flow_attr(struct mlx5_core_dev *mdev, u32 *match_c,
1793 			   u32 *match_v, const union ib_flow_spec *ib_spec,
1794 			   u32 *tag_id, bool *is_drop)
1795 {
1796 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1797 					   misc_parameters);
1798 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1799 					   misc_parameters);
1800 	void *headers_c;
1801 	void *headers_v;
1802 	int match_ipv;
1803 
1804 	if (ib_spec->type & IB_FLOW_SPEC_INNER) {
1805 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1806 					 inner_headers);
1807 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1808 					 inner_headers);
1809 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
1810 					ft_field_support.inner_ip_version);
1811 	} else {
1812 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1813 					 outer_headers);
1814 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1815 					 outer_headers);
1816 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
1817 					ft_field_support.outer_ip_version);
1818 	}
1819 
1820 	switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
1821 	case IB_FLOW_SPEC_ETH:
1822 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1823 			return -EOPNOTSUPP;
1824 
1825 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1826 					     dmac_47_16),
1827 				ib_spec->eth.mask.dst_mac);
1828 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1829 					     dmac_47_16),
1830 				ib_spec->eth.val.dst_mac);
1831 
1832 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1833 					     smac_47_16),
1834 				ib_spec->eth.mask.src_mac);
1835 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1836 					     smac_47_16),
1837 				ib_spec->eth.val.src_mac);
1838 
1839 		if (ib_spec->eth.mask.vlan_tag) {
1840 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1841 				 cvlan_tag, 1);
1842 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1843 				 cvlan_tag, 1);
1844 
1845 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1846 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1847 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1848 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1849 
1850 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1851 				 first_cfi,
1852 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1853 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1854 				 first_cfi,
1855 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1856 
1857 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1858 				 first_prio,
1859 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1860 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1861 				 first_prio,
1862 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1863 		}
1864 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1865 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1866 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1867 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
1868 		break;
1869 	case IB_FLOW_SPEC_IPV4:
1870 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1871 			return -EOPNOTSUPP;
1872 
1873 		if (match_ipv) {
1874 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1875 				 ip_version, 0xf);
1876 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1877 				 ip_version, IPV4_VERSION);
1878 		} else {
1879 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1880 				 ethertype, 0xffff);
1881 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1882 				 ethertype, ETH_P_IP);
1883 		}
1884 
1885 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1886 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1887 		       &ib_spec->ipv4.mask.src_ip,
1888 		       sizeof(ib_spec->ipv4.mask.src_ip));
1889 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1890 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1891 		       &ib_spec->ipv4.val.src_ip,
1892 		       sizeof(ib_spec->ipv4.val.src_ip));
1893 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1894 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1895 		       &ib_spec->ipv4.mask.dst_ip,
1896 		       sizeof(ib_spec->ipv4.mask.dst_ip));
1897 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1898 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1899 		       &ib_spec->ipv4.val.dst_ip,
1900 		       sizeof(ib_spec->ipv4.val.dst_ip));
1901 
1902 		set_tos(headers_c, headers_v,
1903 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
1904 
1905 		set_proto(headers_c, headers_v,
1906 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
1907 		break;
1908 	case IB_FLOW_SPEC_IPV6:
1909 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1910 			return -EOPNOTSUPP;
1911 
1912 		if (match_ipv) {
1913 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1914 				 ip_version, 0xf);
1915 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1916 				 ip_version, IPV6_VERSION);
1917 		} else {
1918 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1919 				 ethertype, 0xffff);
1920 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1921 				 ethertype, ETH_P_IPV6);
1922 		}
1923 
1924 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1925 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1926 		       &ib_spec->ipv6.mask.src_ip,
1927 		       sizeof(ib_spec->ipv6.mask.src_ip));
1928 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1929 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1930 		       &ib_spec->ipv6.val.src_ip,
1931 		       sizeof(ib_spec->ipv6.val.src_ip));
1932 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1933 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1934 		       &ib_spec->ipv6.mask.dst_ip,
1935 		       sizeof(ib_spec->ipv6.mask.dst_ip));
1936 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1937 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1938 		       &ib_spec->ipv6.val.dst_ip,
1939 		       sizeof(ib_spec->ipv6.val.dst_ip));
1940 
1941 		set_tos(headers_c, headers_v,
1942 			ib_spec->ipv6.mask.traffic_class,
1943 			ib_spec->ipv6.val.traffic_class);
1944 
1945 		set_proto(headers_c, headers_v,
1946 			  ib_spec->ipv6.mask.next_hdr,
1947 			  ib_spec->ipv6.val.next_hdr);
1948 
1949 		set_flow_label(misc_params_c, misc_params_v,
1950 			       ntohl(ib_spec->ipv6.mask.flow_label),
1951 			       ntohl(ib_spec->ipv6.val.flow_label),
1952 			       ib_spec->type & IB_FLOW_SPEC_INNER);
1953 
1954 		break;
1955 	case IB_FLOW_SPEC_TCP:
1956 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1957 					 LAST_TCP_UDP_FIELD))
1958 			return -EOPNOTSUPP;
1959 
1960 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
1961 			 0xff);
1962 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
1963 			 IPPROTO_TCP);
1964 
1965 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
1966 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1967 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
1968 			 ntohs(ib_spec->tcp_udp.val.src_port));
1969 
1970 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
1971 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1972 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
1973 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1974 		break;
1975 	case IB_FLOW_SPEC_UDP:
1976 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1977 					 LAST_TCP_UDP_FIELD))
1978 			return -EOPNOTSUPP;
1979 
1980 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
1981 			 0xff);
1982 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
1983 			 IPPROTO_UDP);
1984 
1985 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
1986 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1987 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
1988 			 ntohs(ib_spec->tcp_udp.val.src_port));
1989 
1990 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
1991 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1992 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
1993 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1994 		break;
1995 	case IB_FLOW_SPEC_VXLAN_TUNNEL:
1996 		if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
1997 					 LAST_TUNNEL_FIELD))
1998 			return -EOPNOTSUPP;
1999 
2000 		MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
2001 			 ntohl(ib_spec->tunnel.mask.tunnel_id));
2002 		MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
2003 			 ntohl(ib_spec->tunnel.val.tunnel_id));
2004 		break;
2005 	case IB_FLOW_SPEC_ACTION_TAG:
2006 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
2007 					 LAST_FLOW_TAG_FIELD))
2008 			return -EOPNOTSUPP;
2009 		if (ib_spec->flow_tag.tag_id >= BIT(24))
2010 			return -EINVAL;
2011 
2012 		*tag_id = ib_spec->flow_tag.tag_id;
2013 		break;
2014 	case IB_FLOW_SPEC_ACTION_DROP:
2015 		if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
2016 					 LAST_DROP_FIELD))
2017 			return -EOPNOTSUPP;
2018 		*is_drop = true;
2019 		break;
2020 	default:
2021 		return -EINVAL;
2022 	}
2023 
2024 	return 0;
2025 }
2026 
2027 /* If a flow could catch both multicast and unicast packets,
2028  * it won't fall into the multicast flow steering table and this rule
2029  * could steal other multicast packets.
2030  */
2031 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
2032 {
2033 	struct ib_flow_spec_eth *eth_spec;
2034 
2035 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
2036 	    ib_attr->size < sizeof(struct ib_flow_attr) +
2037 	    sizeof(struct ib_flow_spec_eth) ||
2038 	    ib_attr->num_of_specs < 1)
2039 		return false;
2040 
2041 	eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
2042 	if (eth_spec->type != IB_FLOW_SPEC_ETH ||
2043 	    eth_spec->size != sizeof(*eth_spec))
2044 		return false;
2045 
2046 	return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
2047 	       is_multicast_ether_addr(eth_spec->val.dst_mac);
2048 }
2049 
2050 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
2051 			       const struct ib_flow_attr *flow_attr,
2052 			       bool check_inner)
2053 {
2054 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
2055 	int match_ipv = check_inner ?
2056 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2057 					ft_field_support.inner_ip_version) :
2058 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2059 					ft_field_support.outer_ip_version);
2060 	int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
2061 	bool ipv4_spec_valid, ipv6_spec_valid;
2062 	unsigned int ip_spec_type = 0;
2063 	bool has_ethertype = false;
2064 	unsigned int spec_index;
2065 	bool mask_valid = true;
2066 	u16 eth_type = 0;
2067 	bool type_valid;
2068 
2069 	/* Validate that ethertype is correct */
2070 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2071 		if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
2072 		    ib_spec->eth.mask.ether_type) {
2073 			mask_valid = (ib_spec->eth.mask.ether_type ==
2074 				      htons(0xffff));
2075 			has_ethertype = true;
2076 			eth_type = ntohs(ib_spec->eth.val.ether_type);
2077 		} else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
2078 			   (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
2079 			ip_spec_type = ib_spec->type;
2080 		}
2081 		ib_spec = (void *)ib_spec + ib_spec->size;
2082 	}
2083 
2084 	type_valid = (!has_ethertype) || (!ip_spec_type);
2085 	if (!type_valid && mask_valid) {
2086 		ipv4_spec_valid = (eth_type == ETH_P_IP) &&
2087 			(ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
2088 		ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
2089 			(ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
2090 
2091 		type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
2092 			     (((eth_type == ETH_P_MPLS_UC) ||
2093 			       (eth_type == ETH_P_MPLS_MC)) && match_ipv);
2094 	}
2095 
2096 	return type_valid;
2097 }
2098 
2099 static bool is_valid_attr(struct mlx5_core_dev *mdev,
2100 			  const struct ib_flow_attr *flow_attr)
2101 {
2102 	return is_valid_ethertype(mdev, flow_attr, false) &&
2103 	       is_valid_ethertype(mdev, flow_attr, true);
2104 }
2105 
2106 static void put_flow_table(struct mlx5_ib_dev *dev,
2107 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
2108 {
2109 	prio->refcount -= !!ft_added;
2110 	if (!prio->refcount) {
2111 		mlx5_destroy_flow_table(prio->flow_table);
2112 		prio->flow_table = NULL;
2113 	}
2114 }
2115 
2116 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
2117 {
2118 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
2119 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
2120 							  struct mlx5_ib_flow_handler,
2121 							  ibflow);
2122 	struct mlx5_ib_flow_handler *iter, *tmp;
2123 
2124 	mutex_lock(&dev->flow_db.lock);
2125 
2126 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
2127 		mlx5_del_flow_rules(iter->rule);
2128 		put_flow_table(dev, iter->prio, true);
2129 		list_del(&iter->list);
2130 		kfree(iter);
2131 	}
2132 
2133 	mlx5_del_flow_rules(handler->rule);
2134 	put_flow_table(dev, handler->prio, true);
2135 	mutex_unlock(&dev->flow_db.lock);
2136 
2137 	kfree(handler);
2138 
2139 	return 0;
2140 }
2141 
2142 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
2143 {
2144 	priority *= 2;
2145 	if (!dont_trap)
2146 		priority++;
2147 	return priority;
2148 }
2149 
2150 enum flow_table_type {
2151 	MLX5_IB_FT_RX,
2152 	MLX5_IB_FT_TX
2153 };
2154 
2155 #define MLX5_FS_MAX_TYPES	 6
2156 #define MLX5_FS_MAX_ENTRIES	 BIT(16)
2157 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
2158 						struct ib_flow_attr *flow_attr,
2159 						enum flow_table_type ft_type)
2160 {
2161 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
2162 	struct mlx5_flow_namespace *ns = NULL;
2163 	struct mlx5_ib_flow_prio *prio;
2164 	struct mlx5_flow_table *ft;
2165 	int max_table_size;
2166 	int num_entries;
2167 	int num_groups;
2168 	int priority;
2169 	int err = 0;
2170 
2171 	max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
2172 						       log_max_ft_size));
2173 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2174 		if (flow_is_multicast_only(flow_attr) &&
2175 		    !dont_trap)
2176 			priority = MLX5_IB_FLOW_MCAST_PRIO;
2177 		else
2178 			priority = ib_prio_to_core_prio(flow_attr->priority,
2179 							dont_trap);
2180 		ns = mlx5_get_flow_namespace(dev->mdev,
2181 					     MLX5_FLOW_NAMESPACE_BYPASS);
2182 		num_entries = MLX5_FS_MAX_ENTRIES;
2183 		num_groups = MLX5_FS_MAX_TYPES;
2184 		prio = &dev->flow_db.prios[priority];
2185 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2186 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2187 		ns = mlx5_get_flow_namespace(dev->mdev,
2188 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
2189 		build_leftovers_ft_param(&priority,
2190 					 &num_entries,
2191 					 &num_groups);
2192 		prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
2193 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2194 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
2195 					allow_sniffer_and_nic_rx_shared_tir))
2196 			return ERR_PTR(-ENOTSUPP);
2197 
2198 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
2199 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2200 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2201 
2202 		prio = &dev->flow_db.sniffer[ft_type];
2203 		priority = 0;
2204 		num_entries = 1;
2205 		num_groups = 1;
2206 	}
2207 
2208 	if (!ns)
2209 		return ERR_PTR(-ENOTSUPP);
2210 
2211 	if (num_entries > max_table_size)
2212 		return ERR_PTR(-ENOMEM);
2213 
2214 	ft = prio->flow_table;
2215 	if (!ft) {
2216 		ft = mlx5_create_auto_grouped_flow_table(ns, priority,
2217 							 num_entries,
2218 							 num_groups,
2219 							 0, 0);
2220 
2221 		if (!IS_ERR(ft)) {
2222 			prio->refcount = 0;
2223 			prio->flow_table = ft;
2224 		} else {
2225 			err = PTR_ERR(ft);
2226 		}
2227 	}
2228 
2229 	return err ? ERR_PTR(err) : prio;
2230 }
2231 
2232 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
2233 						     struct mlx5_ib_flow_prio *ft_prio,
2234 						     const struct ib_flow_attr *flow_attr,
2235 						     struct mlx5_flow_destination *dst)
2236 {
2237 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
2238 	struct mlx5_ib_flow_handler *handler;
2239 	struct mlx5_flow_act flow_act = {0};
2240 	struct mlx5_flow_spec *spec;
2241 	struct mlx5_flow_destination *rule_dst = dst;
2242 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
2243 	unsigned int spec_index;
2244 	u32 flow_tag = MLX5_FS_DEFAULT_FLOW_TAG;
2245 	bool is_drop = false;
2246 	int err = 0;
2247 	int dest_num = 1;
2248 
2249 	if (!is_valid_attr(dev->mdev, flow_attr))
2250 		return ERR_PTR(-EINVAL);
2251 
2252 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
2253 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
2254 	if (!handler || !spec) {
2255 		err = -ENOMEM;
2256 		goto free;
2257 	}
2258 
2259 	INIT_LIST_HEAD(&handler->list);
2260 
2261 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2262 		err = parse_flow_attr(dev->mdev, spec->match_criteria,
2263 				      spec->match_value,
2264 				      ib_flow, &flow_tag, &is_drop);
2265 		if (err < 0)
2266 			goto free;
2267 
2268 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
2269 	}
2270 
2271 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
2272 	if (is_drop) {
2273 		flow_act.action = MLX5_FLOW_CONTEXT_ACTION_DROP;
2274 		rule_dst = NULL;
2275 		dest_num = 0;
2276 	} else {
2277 		flow_act.action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
2278 		    MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
2279 	}
2280 
2281 	if (flow_tag != MLX5_FS_DEFAULT_FLOW_TAG &&
2282 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2283 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
2284 		mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
2285 			     flow_tag, flow_attr->type);
2286 		err = -EINVAL;
2287 		goto free;
2288 	}
2289 	flow_act.flow_tag = flow_tag;
2290 	handler->rule = mlx5_add_flow_rules(ft, spec,
2291 					    &flow_act,
2292 					    rule_dst, dest_num);
2293 
2294 	if (IS_ERR(handler->rule)) {
2295 		err = PTR_ERR(handler->rule);
2296 		goto free;
2297 	}
2298 
2299 	ft_prio->refcount++;
2300 	handler->prio = ft_prio;
2301 
2302 	ft_prio->flow_table = ft;
2303 free:
2304 	if (err)
2305 		kfree(handler);
2306 	kvfree(spec);
2307 	return err ? ERR_PTR(err) : handler;
2308 }
2309 
2310 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
2311 							  struct mlx5_ib_flow_prio *ft_prio,
2312 							  struct ib_flow_attr *flow_attr,
2313 							  struct mlx5_flow_destination *dst)
2314 {
2315 	struct mlx5_ib_flow_handler *handler_dst = NULL;
2316 	struct mlx5_ib_flow_handler *handler = NULL;
2317 
2318 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
2319 	if (!IS_ERR(handler)) {
2320 		handler_dst = create_flow_rule(dev, ft_prio,
2321 					       flow_attr, dst);
2322 		if (IS_ERR(handler_dst)) {
2323 			mlx5_del_flow_rules(handler->rule);
2324 			ft_prio->refcount--;
2325 			kfree(handler);
2326 			handler = handler_dst;
2327 		} else {
2328 			list_add(&handler_dst->list, &handler->list);
2329 		}
2330 	}
2331 
2332 	return handler;
2333 }
2334 enum {
2335 	LEFTOVERS_MC,
2336 	LEFTOVERS_UC,
2337 };
2338 
2339 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
2340 							  struct mlx5_ib_flow_prio *ft_prio,
2341 							  struct ib_flow_attr *flow_attr,
2342 							  struct mlx5_flow_destination *dst)
2343 {
2344 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
2345 	struct mlx5_ib_flow_handler *handler = NULL;
2346 
2347 	static struct {
2348 		struct ib_flow_attr	flow_attr;
2349 		struct ib_flow_spec_eth eth_flow;
2350 	} leftovers_specs[] = {
2351 		[LEFTOVERS_MC] = {
2352 			.flow_attr = {
2353 				.num_of_specs = 1,
2354 				.size = sizeof(leftovers_specs[0])
2355 			},
2356 			.eth_flow = {
2357 				.type = IB_FLOW_SPEC_ETH,
2358 				.size = sizeof(struct ib_flow_spec_eth),
2359 				.mask = {.dst_mac = {0x1} },
2360 				.val =  {.dst_mac = {0x1} }
2361 			}
2362 		},
2363 		[LEFTOVERS_UC] = {
2364 			.flow_attr = {
2365 				.num_of_specs = 1,
2366 				.size = sizeof(leftovers_specs[0])
2367 			},
2368 			.eth_flow = {
2369 				.type = IB_FLOW_SPEC_ETH,
2370 				.size = sizeof(struct ib_flow_spec_eth),
2371 				.mask = {.dst_mac = {0x1} },
2372 				.val = {.dst_mac = {} }
2373 			}
2374 		}
2375 	};
2376 
2377 	handler = create_flow_rule(dev, ft_prio,
2378 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
2379 				   dst);
2380 	if (!IS_ERR(handler) &&
2381 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2382 		handler_ucast = create_flow_rule(dev, ft_prio,
2383 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
2384 						 dst);
2385 		if (IS_ERR(handler_ucast)) {
2386 			mlx5_del_flow_rules(handler->rule);
2387 			ft_prio->refcount--;
2388 			kfree(handler);
2389 			handler = handler_ucast;
2390 		} else {
2391 			list_add(&handler_ucast->list, &handler->list);
2392 		}
2393 	}
2394 
2395 	return handler;
2396 }
2397 
2398 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2399 							struct mlx5_ib_flow_prio *ft_rx,
2400 							struct mlx5_ib_flow_prio *ft_tx,
2401 							struct mlx5_flow_destination *dst)
2402 {
2403 	struct mlx5_ib_flow_handler *handler_rx;
2404 	struct mlx5_ib_flow_handler *handler_tx;
2405 	int err;
2406 	static const struct ib_flow_attr flow_attr  = {
2407 		.num_of_specs = 0,
2408 		.size = sizeof(flow_attr)
2409 	};
2410 
2411 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2412 	if (IS_ERR(handler_rx)) {
2413 		err = PTR_ERR(handler_rx);
2414 		goto err;
2415 	}
2416 
2417 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2418 	if (IS_ERR(handler_tx)) {
2419 		err = PTR_ERR(handler_tx);
2420 		goto err_tx;
2421 	}
2422 
2423 	list_add(&handler_tx->list, &handler_rx->list);
2424 
2425 	return handler_rx;
2426 
2427 err_tx:
2428 	mlx5_del_flow_rules(handler_rx->rule);
2429 	ft_rx->refcount--;
2430 	kfree(handler_rx);
2431 err:
2432 	return ERR_PTR(err);
2433 }
2434 
2435 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2436 					   struct ib_flow_attr *flow_attr,
2437 					   int domain)
2438 {
2439 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2440 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2441 	struct mlx5_ib_flow_handler *handler = NULL;
2442 	struct mlx5_flow_destination *dst = NULL;
2443 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2444 	struct mlx5_ib_flow_prio *ft_prio;
2445 	int err;
2446 
2447 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2448 		return ERR_PTR(-ENOMEM);
2449 
2450 	if (domain != IB_FLOW_DOMAIN_USER ||
2451 	    flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2452 	    (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2453 		return ERR_PTR(-EINVAL);
2454 
2455 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2456 	if (!dst)
2457 		return ERR_PTR(-ENOMEM);
2458 
2459 	mutex_lock(&dev->flow_db.lock);
2460 
2461 	ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2462 	if (IS_ERR(ft_prio)) {
2463 		err = PTR_ERR(ft_prio);
2464 		goto unlock;
2465 	}
2466 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2467 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2468 		if (IS_ERR(ft_prio_tx)) {
2469 			err = PTR_ERR(ft_prio_tx);
2470 			ft_prio_tx = NULL;
2471 			goto destroy_ft;
2472 		}
2473 	}
2474 
2475 	dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2476 	if (mqp->flags & MLX5_IB_QP_RSS)
2477 		dst->tir_num = mqp->rss_qp.tirn;
2478 	else
2479 		dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2480 
2481 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2482 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
2483 			handler = create_dont_trap_rule(dev, ft_prio,
2484 							flow_attr, dst);
2485 		} else {
2486 			handler = create_flow_rule(dev, ft_prio, flow_attr,
2487 						   dst);
2488 		}
2489 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2490 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2491 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2492 						dst);
2493 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2494 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2495 	} else {
2496 		err = -EINVAL;
2497 		goto destroy_ft;
2498 	}
2499 
2500 	if (IS_ERR(handler)) {
2501 		err = PTR_ERR(handler);
2502 		handler = NULL;
2503 		goto destroy_ft;
2504 	}
2505 
2506 	mutex_unlock(&dev->flow_db.lock);
2507 	kfree(dst);
2508 
2509 	return &handler->ibflow;
2510 
2511 destroy_ft:
2512 	put_flow_table(dev, ft_prio, false);
2513 	if (ft_prio_tx)
2514 		put_flow_table(dev, ft_prio_tx, false);
2515 unlock:
2516 	mutex_unlock(&dev->flow_db.lock);
2517 	kfree(dst);
2518 	kfree(handler);
2519 	return ERR_PTR(err);
2520 }
2521 
2522 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2523 {
2524 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2525 	int err;
2526 
2527 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2528 	if (err)
2529 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2530 			     ibqp->qp_num, gid->raw);
2531 
2532 	return err;
2533 }
2534 
2535 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2536 {
2537 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2538 	int err;
2539 
2540 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2541 	if (err)
2542 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2543 			     ibqp->qp_num, gid->raw);
2544 
2545 	return err;
2546 }
2547 
2548 static int init_node_data(struct mlx5_ib_dev *dev)
2549 {
2550 	int err;
2551 
2552 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2553 	if (err)
2554 		return err;
2555 
2556 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2557 
2558 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2559 }
2560 
2561 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2562 			     char *buf)
2563 {
2564 	struct mlx5_ib_dev *dev =
2565 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2566 
2567 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
2568 }
2569 
2570 static ssize_t show_reg_pages(struct device *device,
2571 			      struct device_attribute *attr, char *buf)
2572 {
2573 	struct mlx5_ib_dev *dev =
2574 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2575 
2576 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2577 }
2578 
2579 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2580 			char *buf)
2581 {
2582 	struct mlx5_ib_dev *dev =
2583 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2584 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2585 }
2586 
2587 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2588 			char *buf)
2589 {
2590 	struct mlx5_ib_dev *dev =
2591 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2592 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
2593 }
2594 
2595 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2596 			  char *buf)
2597 {
2598 	struct mlx5_ib_dev *dev =
2599 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2600 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2601 		       dev->mdev->board_id);
2602 }
2603 
2604 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2605 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2606 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2607 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2608 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2609 
2610 static struct device_attribute *mlx5_class_attributes[] = {
2611 	&dev_attr_hw_rev,
2612 	&dev_attr_hca_type,
2613 	&dev_attr_board_id,
2614 	&dev_attr_fw_pages,
2615 	&dev_attr_reg_pages,
2616 };
2617 
2618 static void pkey_change_handler(struct work_struct *work)
2619 {
2620 	struct mlx5_ib_port_resources *ports =
2621 		container_of(work, struct mlx5_ib_port_resources,
2622 			     pkey_change_work);
2623 
2624 	mutex_lock(&ports->devr->mutex);
2625 	mlx5_ib_gsi_pkey_change(ports->gsi);
2626 	mutex_unlock(&ports->devr->mutex);
2627 }
2628 
2629 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2630 {
2631 	struct mlx5_ib_qp *mqp;
2632 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2633 	struct mlx5_core_cq *mcq;
2634 	struct list_head cq_armed_list;
2635 	unsigned long flags_qp;
2636 	unsigned long flags_cq;
2637 	unsigned long flags;
2638 
2639 	INIT_LIST_HEAD(&cq_armed_list);
2640 
2641 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2642 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2643 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2644 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2645 		if (mqp->sq.tail != mqp->sq.head) {
2646 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2647 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2648 			if (send_mcq->mcq.comp &&
2649 			    mqp->ibqp.send_cq->comp_handler) {
2650 				if (!send_mcq->mcq.reset_notify_added) {
2651 					send_mcq->mcq.reset_notify_added = 1;
2652 					list_add_tail(&send_mcq->mcq.reset_notify,
2653 						      &cq_armed_list);
2654 				}
2655 			}
2656 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2657 		}
2658 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2659 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2660 		/* no handling is needed for SRQ */
2661 		if (!mqp->ibqp.srq) {
2662 			if (mqp->rq.tail != mqp->rq.head) {
2663 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2664 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2665 				if (recv_mcq->mcq.comp &&
2666 				    mqp->ibqp.recv_cq->comp_handler) {
2667 					if (!recv_mcq->mcq.reset_notify_added) {
2668 						recv_mcq->mcq.reset_notify_added = 1;
2669 						list_add_tail(&recv_mcq->mcq.reset_notify,
2670 							      &cq_armed_list);
2671 					}
2672 				}
2673 				spin_unlock_irqrestore(&recv_mcq->lock,
2674 						       flags_cq);
2675 			}
2676 		}
2677 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2678 	}
2679 	/*At that point all inflight post send were put to be executed as of we
2680 	 * lock/unlock above locks Now need to arm all involved CQs.
2681 	 */
2682 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2683 		mcq->comp(mcq);
2684 	}
2685 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2686 }
2687 
2688 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2689 			  enum mlx5_dev_event event, unsigned long param)
2690 {
2691 	struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2692 	struct ib_event ibev;
2693 	bool fatal = false;
2694 	u8 port = 0;
2695 
2696 	switch (event) {
2697 	case MLX5_DEV_EVENT_SYS_ERROR:
2698 		ibev.event = IB_EVENT_DEVICE_FATAL;
2699 		mlx5_ib_handle_internal_error(ibdev);
2700 		fatal = true;
2701 		break;
2702 
2703 	case MLX5_DEV_EVENT_PORT_UP:
2704 	case MLX5_DEV_EVENT_PORT_DOWN:
2705 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
2706 		port = (u8)param;
2707 
2708 		/* In RoCE, port up/down events are handled in
2709 		 * mlx5_netdev_event().
2710 		 */
2711 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2712 			IB_LINK_LAYER_ETHERNET)
2713 			return;
2714 
2715 		ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2716 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2717 		break;
2718 
2719 	case MLX5_DEV_EVENT_LID_CHANGE:
2720 		ibev.event = IB_EVENT_LID_CHANGE;
2721 		port = (u8)param;
2722 		break;
2723 
2724 	case MLX5_DEV_EVENT_PKEY_CHANGE:
2725 		ibev.event = IB_EVENT_PKEY_CHANGE;
2726 		port = (u8)param;
2727 
2728 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2729 		break;
2730 
2731 	case MLX5_DEV_EVENT_GUID_CHANGE:
2732 		ibev.event = IB_EVENT_GID_CHANGE;
2733 		port = (u8)param;
2734 		break;
2735 
2736 	case MLX5_DEV_EVENT_CLIENT_REREG:
2737 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
2738 		port = (u8)param;
2739 		break;
2740 	default:
2741 		return;
2742 	}
2743 
2744 	ibev.device	      = &ibdev->ib_dev;
2745 	ibev.element.port_num = port;
2746 
2747 	if (port < 1 || port > ibdev->num_ports) {
2748 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2749 		return;
2750 	}
2751 
2752 	if (ibdev->ib_active)
2753 		ib_dispatch_event(&ibev);
2754 
2755 	if (fatal)
2756 		ibdev->ib_active = false;
2757 }
2758 
2759 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2760 {
2761 	struct mlx5_hca_vport_context vport_ctx;
2762 	int err;
2763 	int port;
2764 
2765 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2766 		dev->mdev->port_caps[port - 1].has_smi = false;
2767 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2768 		    MLX5_CAP_PORT_TYPE_IB) {
2769 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2770 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
2771 								   port, 0,
2772 								   &vport_ctx);
2773 				if (err) {
2774 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2775 						    port, err);
2776 					return err;
2777 				}
2778 				dev->mdev->port_caps[port - 1].has_smi =
2779 					vport_ctx.has_smi;
2780 			} else {
2781 				dev->mdev->port_caps[port - 1].has_smi = true;
2782 			}
2783 		}
2784 	}
2785 	return 0;
2786 }
2787 
2788 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2789 {
2790 	int port;
2791 
2792 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2793 		mlx5_query_ext_port_caps(dev, port);
2794 }
2795 
2796 static int get_port_caps(struct mlx5_ib_dev *dev)
2797 {
2798 	struct ib_device_attr *dprops = NULL;
2799 	struct ib_port_attr *pprops = NULL;
2800 	int err = -ENOMEM;
2801 	int port;
2802 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2803 
2804 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2805 	if (!pprops)
2806 		goto out;
2807 
2808 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2809 	if (!dprops)
2810 		goto out;
2811 
2812 	err = set_has_smi_cap(dev);
2813 	if (err)
2814 		goto out;
2815 
2816 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2817 	if (err) {
2818 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2819 		goto out;
2820 	}
2821 
2822 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2823 		memset(pprops, 0, sizeof(*pprops));
2824 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2825 		if (err) {
2826 			mlx5_ib_warn(dev, "query_port %d failed %d\n",
2827 				     port, err);
2828 			break;
2829 		}
2830 		dev->mdev->port_caps[port - 1].pkey_table_len =
2831 						dprops->max_pkeys;
2832 		dev->mdev->port_caps[port - 1].gid_table_len =
2833 						pprops->gid_tbl_len;
2834 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2835 			    dprops->max_pkeys, pprops->gid_tbl_len);
2836 	}
2837 
2838 out:
2839 	kfree(pprops);
2840 	kfree(dprops);
2841 
2842 	return err;
2843 }
2844 
2845 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2846 {
2847 	int err;
2848 
2849 	err = mlx5_mr_cache_cleanup(dev);
2850 	if (err)
2851 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2852 
2853 	mlx5_ib_destroy_qp(dev->umrc.qp);
2854 	ib_free_cq(dev->umrc.cq);
2855 	ib_dealloc_pd(dev->umrc.pd);
2856 }
2857 
2858 enum {
2859 	MAX_UMR_WR = 128,
2860 };
2861 
2862 static int create_umr_res(struct mlx5_ib_dev *dev)
2863 {
2864 	struct ib_qp_init_attr *init_attr = NULL;
2865 	struct ib_qp_attr *attr = NULL;
2866 	struct ib_pd *pd;
2867 	struct ib_cq *cq;
2868 	struct ib_qp *qp;
2869 	int ret;
2870 
2871 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2872 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2873 	if (!attr || !init_attr) {
2874 		ret = -ENOMEM;
2875 		goto error_0;
2876 	}
2877 
2878 	pd = ib_alloc_pd(&dev->ib_dev, 0);
2879 	if (IS_ERR(pd)) {
2880 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2881 		ret = PTR_ERR(pd);
2882 		goto error_0;
2883 	}
2884 
2885 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2886 	if (IS_ERR(cq)) {
2887 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2888 		ret = PTR_ERR(cq);
2889 		goto error_2;
2890 	}
2891 
2892 	init_attr->send_cq = cq;
2893 	init_attr->recv_cq = cq;
2894 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2895 	init_attr->cap.max_send_wr = MAX_UMR_WR;
2896 	init_attr->cap.max_send_sge = 1;
2897 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2898 	init_attr->port_num = 1;
2899 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2900 	if (IS_ERR(qp)) {
2901 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2902 		ret = PTR_ERR(qp);
2903 		goto error_3;
2904 	}
2905 	qp->device     = &dev->ib_dev;
2906 	qp->real_qp    = qp;
2907 	qp->uobject    = NULL;
2908 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2909 
2910 	attr->qp_state = IB_QPS_INIT;
2911 	attr->port_num = 1;
2912 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2913 				IB_QP_PORT, NULL);
2914 	if (ret) {
2915 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2916 		goto error_4;
2917 	}
2918 
2919 	memset(attr, 0, sizeof(*attr));
2920 	attr->qp_state = IB_QPS_RTR;
2921 	attr->path_mtu = IB_MTU_256;
2922 
2923 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2924 	if (ret) {
2925 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2926 		goto error_4;
2927 	}
2928 
2929 	memset(attr, 0, sizeof(*attr));
2930 	attr->qp_state = IB_QPS_RTS;
2931 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2932 	if (ret) {
2933 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2934 		goto error_4;
2935 	}
2936 
2937 	dev->umrc.qp = qp;
2938 	dev->umrc.cq = cq;
2939 	dev->umrc.pd = pd;
2940 
2941 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
2942 	ret = mlx5_mr_cache_init(dev);
2943 	if (ret) {
2944 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2945 		goto error_4;
2946 	}
2947 
2948 	kfree(attr);
2949 	kfree(init_attr);
2950 
2951 	return 0;
2952 
2953 error_4:
2954 	mlx5_ib_destroy_qp(qp);
2955 
2956 error_3:
2957 	ib_free_cq(cq);
2958 
2959 error_2:
2960 	ib_dealloc_pd(pd);
2961 
2962 error_0:
2963 	kfree(attr);
2964 	kfree(init_attr);
2965 	return ret;
2966 }
2967 
2968 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2969 {
2970 	switch (umr_fence_cap) {
2971 	case MLX5_CAP_UMR_FENCE_NONE:
2972 		return MLX5_FENCE_MODE_NONE;
2973 	case MLX5_CAP_UMR_FENCE_SMALL:
2974 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2975 	default:
2976 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2977 	}
2978 }
2979 
2980 static int create_dev_resources(struct mlx5_ib_resources *devr)
2981 {
2982 	struct ib_srq_init_attr attr;
2983 	struct mlx5_ib_dev *dev;
2984 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2985 	int port;
2986 	int ret = 0;
2987 
2988 	dev = container_of(devr, struct mlx5_ib_dev, devr);
2989 
2990 	mutex_init(&devr->mutex);
2991 
2992 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2993 	if (IS_ERR(devr->p0)) {
2994 		ret = PTR_ERR(devr->p0);
2995 		goto error0;
2996 	}
2997 	devr->p0->device  = &dev->ib_dev;
2998 	devr->p0->uobject = NULL;
2999 	atomic_set(&devr->p0->usecnt, 0);
3000 
3001 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
3002 	if (IS_ERR(devr->c0)) {
3003 		ret = PTR_ERR(devr->c0);
3004 		goto error1;
3005 	}
3006 	devr->c0->device        = &dev->ib_dev;
3007 	devr->c0->uobject       = NULL;
3008 	devr->c0->comp_handler  = NULL;
3009 	devr->c0->event_handler = NULL;
3010 	devr->c0->cq_context    = NULL;
3011 	atomic_set(&devr->c0->usecnt, 0);
3012 
3013 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
3014 	if (IS_ERR(devr->x0)) {
3015 		ret = PTR_ERR(devr->x0);
3016 		goto error2;
3017 	}
3018 	devr->x0->device = &dev->ib_dev;
3019 	devr->x0->inode = NULL;
3020 	atomic_set(&devr->x0->usecnt, 0);
3021 	mutex_init(&devr->x0->tgt_qp_mutex);
3022 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
3023 
3024 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
3025 	if (IS_ERR(devr->x1)) {
3026 		ret = PTR_ERR(devr->x1);
3027 		goto error3;
3028 	}
3029 	devr->x1->device = &dev->ib_dev;
3030 	devr->x1->inode = NULL;
3031 	atomic_set(&devr->x1->usecnt, 0);
3032 	mutex_init(&devr->x1->tgt_qp_mutex);
3033 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
3034 
3035 	memset(&attr, 0, sizeof(attr));
3036 	attr.attr.max_sge = 1;
3037 	attr.attr.max_wr = 1;
3038 	attr.srq_type = IB_SRQT_XRC;
3039 	attr.ext.xrc.cq = devr->c0;
3040 	attr.ext.xrc.xrcd = devr->x0;
3041 
3042 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
3043 	if (IS_ERR(devr->s0)) {
3044 		ret = PTR_ERR(devr->s0);
3045 		goto error4;
3046 	}
3047 	devr->s0->device	= &dev->ib_dev;
3048 	devr->s0->pd		= devr->p0;
3049 	devr->s0->uobject       = NULL;
3050 	devr->s0->event_handler = NULL;
3051 	devr->s0->srq_context   = NULL;
3052 	devr->s0->srq_type      = IB_SRQT_XRC;
3053 	devr->s0->ext.xrc.xrcd	= devr->x0;
3054 	devr->s0->ext.xrc.cq	= devr->c0;
3055 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
3056 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
3057 	atomic_inc(&devr->p0->usecnt);
3058 	atomic_set(&devr->s0->usecnt, 0);
3059 
3060 	memset(&attr, 0, sizeof(attr));
3061 	attr.attr.max_sge = 1;
3062 	attr.attr.max_wr = 1;
3063 	attr.srq_type = IB_SRQT_BASIC;
3064 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
3065 	if (IS_ERR(devr->s1)) {
3066 		ret = PTR_ERR(devr->s1);
3067 		goto error5;
3068 	}
3069 	devr->s1->device	= &dev->ib_dev;
3070 	devr->s1->pd		= devr->p0;
3071 	devr->s1->uobject       = NULL;
3072 	devr->s1->event_handler = NULL;
3073 	devr->s1->srq_context   = NULL;
3074 	devr->s1->srq_type      = IB_SRQT_BASIC;
3075 	devr->s1->ext.xrc.cq	= devr->c0;
3076 	atomic_inc(&devr->p0->usecnt);
3077 	atomic_set(&devr->s0->usecnt, 0);
3078 
3079 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
3080 		INIT_WORK(&devr->ports[port].pkey_change_work,
3081 			  pkey_change_handler);
3082 		devr->ports[port].devr = devr;
3083 	}
3084 
3085 	return 0;
3086 
3087 error5:
3088 	mlx5_ib_destroy_srq(devr->s0);
3089 error4:
3090 	mlx5_ib_dealloc_xrcd(devr->x1);
3091 error3:
3092 	mlx5_ib_dealloc_xrcd(devr->x0);
3093 error2:
3094 	mlx5_ib_destroy_cq(devr->c0);
3095 error1:
3096 	mlx5_ib_dealloc_pd(devr->p0);
3097 error0:
3098 	return ret;
3099 }
3100 
3101 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
3102 {
3103 	struct mlx5_ib_dev *dev =
3104 		container_of(devr, struct mlx5_ib_dev, devr);
3105 	int port;
3106 
3107 	mlx5_ib_destroy_srq(devr->s1);
3108 	mlx5_ib_destroy_srq(devr->s0);
3109 	mlx5_ib_dealloc_xrcd(devr->x0);
3110 	mlx5_ib_dealloc_xrcd(devr->x1);
3111 	mlx5_ib_destroy_cq(devr->c0);
3112 	mlx5_ib_dealloc_pd(devr->p0);
3113 
3114 	/* Make sure no change P_Key work items are still executing */
3115 	for (port = 0; port < dev->num_ports; ++port)
3116 		cancel_work_sync(&devr->ports[port].pkey_change_work);
3117 }
3118 
3119 static u32 get_core_cap_flags(struct ib_device *ibdev)
3120 {
3121 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3122 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
3123 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
3124 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
3125 	u32 ret = 0;
3126 
3127 	if (ll == IB_LINK_LAYER_INFINIBAND)
3128 		return RDMA_CORE_PORT_IBA_IB;
3129 
3130 	ret = RDMA_CORE_PORT_RAW_PACKET;
3131 
3132 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
3133 		return ret;
3134 
3135 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
3136 		return ret;
3137 
3138 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
3139 		ret |= RDMA_CORE_PORT_IBA_ROCE;
3140 
3141 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
3142 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3143 
3144 	return ret;
3145 }
3146 
3147 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
3148 			       struct ib_port_immutable *immutable)
3149 {
3150 	struct ib_port_attr attr;
3151 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3152 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
3153 	int err;
3154 
3155 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
3156 
3157 	err = ib_query_port(ibdev, port_num, &attr);
3158 	if (err)
3159 		return err;
3160 
3161 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3162 	immutable->gid_tbl_len = attr.gid_tbl_len;
3163 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
3164 	if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
3165 		immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3166 
3167 	return 0;
3168 }
3169 
3170 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
3171 			   size_t str_len)
3172 {
3173 	struct mlx5_ib_dev *dev =
3174 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3175 	snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
3176 		       fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
3177 }
3178 
3179 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3180 {
3181 	struct mlx5_core_dev *mdev = dev->mdev;
3182 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3183 								 MLX5_FLOW_NAMESPACE_LAG);
3184 	struct mlx5_flow_table *ft;
3185 	int err;
3186 
3187 	if (!ns || !mlx5_lag_is_active(mdev))
3188 		return 0;
3189 
3190 	err = mlx5_cmd_create_vport_lag(mdev);
3191 	if (err)
3192 		return err;
3193 
3194 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3195 	if (IS_ERR(ft)) {
3196 		err = PTR_ERR(ft);
3197 		goto err_destroy_vport_lag;
3198 	}
3199 
3200 	dev->flow_db.lag_demux_ft = ft;
3201 	return 0;
3202 
3203 err_destroy_vport_lag:
3204 	mlx5_cmd_destroy_vport_lag(mdev);
3205 	return err;
3206 }
3207 
3208 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3209 {
3210 	struct mlx5_core_dev *mdev = dev->mdev;
3211 
3212 	if (dev->flow_db.lag_demux_ft) {
3213 		mlx5_destroy_flow_table(dev->flow_db.lag_demux_ft);
3214 		dev->flow_db.lag_demux_ft = NULL;
3215 
3216 		mlx5_cmd_destroy_vport_lag(mdev);
3217 	}
3218 }
3219 
3220 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev)
3221 {
3222 	int err;
3223 
3224 	dev->roce.nb.notifier_call = mlx5_netdev_event;
3225 	err = register_netdevice_notifier(&dev->roce.nb);
3226 	if (err) {
3227 		dev->roce.nb.notifier_call = NULL;
3228 		return err;
3229 	}
3230 
3231 	return 0;
3232 }
3233 
3234 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev)
3235 {
3236 	if (dev->roce.nb.notifier_call) {
3237 		unregister_netdevice_notifier(&dev->roce.nb);
3238 		dev->roce.nb.notifier_call = NULL;
3239 	}
3240 }
3241 
3242 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3243 {
3244 	int err;
3245 
3246 	err = mlx5_add_netdev_notifier(dev);
3247 	if (err)
3248 		return err;
3249 
3250 	if (MLX5_CAP_GEN(dev->mdev, roce)) {
3251 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3252 		if (err)
3253 			goto err_unregister_netdevice_notifier;
3254 	}
3255 
3256 	err = mlx5_eth_lag_init(dev);
3257 	if (err)
3258 		goto err_disable_roce;
3259 
3260 	return 0;
3261 
3262 err_disable_roce:
3263 	if (MLX5_CAP_GEN(dev->mdev, roce))
3264 		mlx5_nic_vport_disable_roce(dev->mdev);
3265 
3266 err_unregister_netdevice_notifier:
3267 	mlx5_remove_netdev_notifier(dev);
3268 	return err;
3269 }
3270 
3271 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3272 {
3273 	mlx5_eth_lag_cleanup(dev);
3274 	if (MLX5_CAP_GEN(dev->mdev, roce))
3275 		mlx5_nic_vport_disable_roce(dev->mdev);
3276 }
3277 
3278 struct mlx5_ib_counter {
3279 	const char *name;
3280 	size_t offset;
3281 };
3282 
3283 #define INIT_Q_COUNTER(_name)		\
3284 	{ .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
3285 
3286 static const struct mlx5_ib_counter basic_q_cnts[] = {
3287 	INIT_Q_COUNTER(rx_write_requests),
3288 	INIT_Q_COUNTER(rx_read_requests),
3289 	INIT_Q_COUNTER(rx_atomic_requests),
3290 	INIT_Q_COUNTER(out_of_buffer),
3291 };
3292 
3293 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
3294 	INIT_Q_COUNTER(out_of_sequence),
3295 };
3296 
3297 static const struct mlx5_ib_counter retrans_q_cnts[] = {
3298 	INIT_Q_COUNTER(duplicate_request),
3299 	INIT_Q_COUNTER(rnr_nak_retry_err),
3300 	INIT_Q_COUNTER(packet_seq_err),
3301 	INIT_Q_COUNTER(implied_nak_seq_err),
3302 	INIT_Q_COUNTER(local_ack_timeout_err),
3303 };
3304 
3305 #define INIT_CONG_COUNTER(_name)		\
3306 	{ .name = #_name, .offset =	\
3307 		MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
3308 
3309 static const struct mlx5_ib_counter cong_cnts[] = {
3310 	INIT_CONG_COUNTER(rp_cnp_ignored),
3311 	INIT_CONG_COUNTER(rp_cnp_handled),
3312 	INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
3313 	INIT_CONG_COUNTER(np_cnp_sent),
3314 };
3315 
3316 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
3317 {
3318 	unsigned int i;
3319 
3320 	for (i = 0; i < dev->num_ports; i++) {
3321 		mlx5_core_dealloc_q_counter(dev->mdev,
3322 					    dev->port[i].cnts.set_id);
3323 		kfree(dev->port[i].cnts.names);
3324 		kfree(dev->port[i].cnts.offsets);
3325 	}
3326 }
3327 
3328 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
3329 				    struct mlx5_ib_counters *cnts)
3330 {
3331 	u32 num_counters;
3332 
3333 	num_counters = ARRAY_SIZE(basic_q_cnts);
3334 
3335 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
3336 		num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
3337 
3338 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
3339 		num_counters += ARRAY_SIZE(retrans_q_cnts);
3340 	cnts->num_q_counters = num_counters;
3341 
3342 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
3343 		cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
3344 		num_counters += ARRAY_SIZE(cong_cnts);
3345 	}
3346 
3347 	cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
3348 	if (!cnts->names)
3349 		return -ENOMEM;
3350 
3351 	cnts->offsets = kcalloc(num_counters,
3352 				sizeof(cnts->offsets), GFP_KERNEL);
3353 	if (!cnts->offsets)
3354 		goto err_names;
3355 
3356 	return 0;
3357 
3358 err_names:
3359 	kfree(cnts->names);
3360 	return -ENOMEM;
3361 }
3362 
3363 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
3364 				  const char **names,
3365 				  size_t *offsets)
3366 {
3367 	int i;
3368 	int j = 0;
3369 
3370 	for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
3371 		names[j] = basic_q_cnts[i].name;
3372 		offsets[j] = basic_q_cnts[i].offset;
3373 	}
3374 
3375 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
3376 		for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
3377 			names[j] = out_of_seq_q_cnts[i].name;
3378 			offsets[j] = out_of_seq_q_cnts[i].offset;
3379 		}
3380 	}
3381 
3382 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3383 		for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
3384 			names[j] = retrans_q_cnts[i].name;
3385 			offsets[j] = retrans_q_cnts[i].offset;
3386 		}
3387 	}
3388 
3389 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
3390 		for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
3391 			names[j] = cong_cnts[i].name;
3392 			offsets[j] = cong_cnts[i].offset;
3393 		}
3394 	}
3395 }
3396 
3397 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
3398 {
3399 	int i;
3400 	int ret;
3401 
3402 	for (i = 0; i < dev->num_ports; i++) {
3403 		struct mlx5_ib_port *port = &dev->port[i];
3404 
3405 		ret = mlx5_core_alloc_q_counter(dev->mdev,
3406 						&port->cnts.set_id);
3407 		if (ret) {
3408 			mlx5_ib_warn(dev,
3409 				     "couldn't allocate queue counter for port %d, err %d\n",
3410 				     i + 1, ret);
3411 			goto dealloc_counters;
3412 		}
3413 
3414 		ret = __mlx5_ib_alloc_counters(dev, &port->cnts);
3415 		if (ret)
3416 			goto dealloc_counters;
3417 
3418 		mlx5_ib_fill_counters(dev, port->cnts.names,
3419 				      port->cnts.offsets);
3420 	}
3421 
3422 	return 0;
3423 
3424 dealloc_counters:
3425 	while (--i >= 0)
3426 		mlx5_core_dealloc_q_counter(dev->mdev,
3427 					    dev->port[i].cnts.set_id);
3428 
3429 	return ret;
3430 }
3431 
3432 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
3433 						    u8 port_num)
3434 {
3435 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3436 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
3437 
3438 	/* We support only per port stats */
3439 	if (port_num == 0)
3440 		return NULL;
3441 
3442 	return rdma_alloc_hw_stats_struct(port->cnts.names,
3443 					  port->cnts.num_q_counters +
3444 					  port->cnts.num_cong_counters,
3445 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
3446 }
3447 
3448 static int mlx5_ib_query_q_counters(struct mlx5_ib_dev *dev,
3449 				    struct mlx5_ib_port *port,
3450 				    struct rdma_hw_stats *stats)
3451 {
3452 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
3453 	void *out;
3454 	__be32 val;
3455 	int ret, i;
3456 
3457 	out = kvzalloc(outlen, GFP_KERNEL);
3458 	if (!out)
3459 		return -ENOMEM;
3460 
3461 	ret = mlx5_core_query_q_counter(dev->mdev,
3462 					port->cnts.set_id, 0,
3463 					out, outlen);
3464 	if (ret)
3465 		goto free;
3466 
3467 	for (i = 0; i < port->cnts.num_q_counters; i++) {
3468 		val = *(__be32 *)(out + port->cnts.offsets[i]);
3469 		stats->value[i] = (u64)be32_to_cpu(val);
3470 	}
3471 
3472 free:
3473 	kvfree(out);
3474 	return ret;
3475 }
3476 
3477 static int mlx5_ib_query_cong_counters(struct mlx5_ib_dev *dev,
3478 				       struct mlx5_ib_port *port,
3479 				       struct rdma_hw_stats *stats)
3480 {
3481 	int outlen = MLX5_ST_SZ_BYTES(query_cong_statistics_out);
3482 	void *out;
3483 	int ret, i;
3484 	int offset = port->cnts.num_q_counters;
3485 
3486 	out = kvzalloc(outlen, GFP_KERNEL);
3487 	if (!out)
3488 		return -ENOMEM;
3489 
3490 	ret = mlx5_cmd_query_cong_counter(dev->mdev, false, out, outlen);
3491 	if (ret)
3492 		goto free;
3493 
3494 	for (i = 0; i < port->cnts.num_cong_counters; i++) {
3495 		stats->value[i + offset] =
3496 			be64_to_cpup((__be64 *)(out +
3497 				     port->cnts.offsets[i + offset]));
3498 	}
3499 
3500 free:
3501 	kvfree(out);
3502 	return ret;
3503 }
3504 
3505 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
3506 				struct rdma_hw_stats *stats,
3507 				u8 port_num, int index)
3508 {
3509 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3510 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
3511 	int ret, num_counters;
3512 
3513 	if (!stats)
3514 		return -EINVAL;
3515 
3516 	ret = mlx5_ib_query_q_counters(dev, port, stats);
3517 	if (ret)
3518 		return ret;
3519 	num_counters = port->cnts.num_q_counters;
3520 
3521 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
3522 		ret = mlx5_ib_query_cong_counters(dev, port, stats);
3523 		if (ret)
3524 			return ret;
3525 		num_counters += port->cnts.num_cong_counters;
3526 	}
3527 
3528 	return num_counters;
3529 }
3530 
3531 static void mlx5_ib_free_rdma_netdev(struct net_device *netdev)
3532 {
3533 	return mlx5_rdma_netdev_free(netdev);
3534 }
3535 
3536 static struct net_device*
3537 mlx5_ib_alloc_rdma_netdev(struct ib_device *hca,
3538 			  u8 port_num,
3539 			  enum rdma_netdev_t type,
3540 			  const char *name,
3541 			  unsigned char name_assign_type,
3542 			  void (*setup)(struct net_device *))
3543 {
3544 	struct net_device *netdev;
3545 	struct rdma_netdev *rn;
3546 
3547 	if (type != RDMA_NETDEV_IPOIB)
3548 		return ERR_PTR(-EOPNOTSUPP);
3549 
3550 	netdev = mlx5_rdma_netdev_alloc(to_mdev(hca)->mdev, hca,
3551 					name, setup);
3552 	if (likely(!IS_ERR_OR_NULL(netdev))) {
3553 		rn = netdev_priv(netdev);
3554 		rn->free_rdma_netdev = mlx5_ib_free_rdma_netdev;
3555 	}
3556 	return netdev;
3557 }
3558 
3559 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
3560 {
3561 	struct mlx5_ib_dev *dev;
3562 	enum rdma_link_layer ll;
3563 	int port_type_cap;
3564 	const char *name;
3565 	int err;
3566 	int i;
3567 
3568 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3569 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3570 
3571 	printk_once(KERN_INFO "%s", mlx5_version);
3572 
3573 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
3574 	if (!dev)
3575 		return NULL;
3576 
3577 	dev->mdev = mdev;
3578 
3579 	dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
3580 			    GFP_KERNEL);
3581 	if (!dev->port)
3582 		goto err_dealloc;
3583 
3584 	rwlock_init(&dev->roce.netdev_lock);
3585 	err = get_port_caps(dev);
3586 	if (err)
3587 		goto err_free_port;
3588 
3589 	if (mlx5_use_mad_ifc(dev))
3590 		get_ext_port_caps(dev);
3591 
3592 	if (!mlx5_lag_is_active(mdev))
3593 		name = "mlx5_%d";
3594 	else
3595 		name = "mlx5_bond_%d";
3596 
3597 	strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
3598 	dev->ib_dev.owner		= THIS_MODULE;
3599 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
3600 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
3601 	dev->num_ports		= MLX5_CAP_GEN(mdev, num_ports);
3602 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
3603 	dev->ib_dev.num_comp_vectors    =
3604 		dev->mdev->priv.eq_table.num_comp_vectors;
3605 	dev->ib_dev.dev.parent		= &mdev->pdev->dev;
3606 
3607 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
3608 	dev->ib_dev.uverbs_cmd_mask	=
3609 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
3610 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
3611 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
3612 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
3613 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
3614 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
3615 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
3616 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
3617 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
3618 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
3619 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
3620 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
3621 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
3622 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
3623 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
3624 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
3625 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
3626 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
3627 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
3628 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
3629 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
3630 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
3631 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
3632 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
3633 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
3634 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
3635 	dev->ib_dev.uverbs_ex_cmd_mask =
3636 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
3637 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
3638 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
3639 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP);
3640 
3641 	dev->ib_dev.query_device	= mlx5_ib_query_device;
3642 	dev->ib_dev.query_port		= mlx5_ib_query_port;
3643 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
3644 	if (ll == IB_LINK_LAYER_ETHERNET)
3645 		dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
3646 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
3647 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
3648 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
3649 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
3650 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
3651 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
3652 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
3653 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
3654 	dev->ib_dev.mmap		= mlx5_ib_mmap;
3655 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
3656 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
3657 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
3658 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
3659 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
3660 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
3661 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
3662 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
3663 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
3664 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
3665 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
3666 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
3667 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
3668 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
3669 	dev->ib_dev.post_send		= mlx5_ib_post_send;
3670 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
3671 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
3672 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
3673 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
3674 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
3675 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
3676 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
3677 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
3678 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
3679 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
3680 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
3681 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
3682 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
3683 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
3684 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
3685 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
3686 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
3687 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
3688 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
3689 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads))
3690 		dev->ib_dev.alloc_rdma_netdev	= mlx5_ib_alloc_rdma_netdev;
3691 
3692 	if (mlx5_core_is_pf(mdev)) {
3693 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
3694 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
3695 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
3696 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
3697 	}
3698 
3699 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
3700 
3701 	mlx5_ib_internal_fill_odp_caps(dev);
3702 
3703 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3704 
3705 	if (MLX5_CAP_GEN(mdev, imaicl)) {
3706 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
3707 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
3708 		dev->ib_dev.uverbs_cmd_mask |=
3709 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
3710 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
3711 	}
3712 
3713 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
3714 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
3715 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
3716 	}
3717 
3718 	if (MLX5_CAP_GEN(mdev, xrc)) {
3719 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
3720 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
3721 		dev->ib_dev.uverbs_cmd_mask |=
3722 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
3723 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
3724 	}
3725 
3726 	if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
3727 	    IB_LINK_LAYER_ETHERNET) {
3728 		dev->ib_dev.create_flow	= mlx5_ib_create_flow;
3729 		dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
3730 		dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
3731 		dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
3732 		dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
3733 		dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
3734 		dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
3735 		dev->ib_dev.uverbs_ex_cmd_mask |=
3736 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
3737 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
3738 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
3739 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
3740 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
3741 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
3742 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
3743 	}
3744 	err = init_node_data(dev);
3745 	if (err)
3746 		goto err_free_port;
3747 
3748 	mutex_init(&dev->flow_db.lock);
3749 	mutex_init(&dev->cap_mask_mutex);
3750 	INIT_LIST_HEAD(&dev->qp_list);
3751 	spin_lock_init(&dev->reset_flow_resource_lock);
3752 
3753 	if (ll == IB_LINK_LAYER_ETHERNET) {
3754 		err = mlx5_enable_eth(dev);
3755 		if (err)
3756 			goto err_free_port;
3757 	}
3758 
3759 	err = create_dev_resources(&dev->devr);
3760 	if (err)
3761 		goto err_disable_eth;
3762 
3763 	err = mlx5_ib_odp_init_one(dev);
3764 	if (err)
3765 		goto err_rsrc;
3766 
3767 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
3768 		err = mlx5_ib_alloc_counters(dev);
3769 		if (err)
3770 			goto err_odp;
3771 	}
3772 
3773 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
3774 	if (!dev->mdev->priv.uar)
3775 		goto err_cnt;
3776 
3777 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
3778 	if (err)
3779 		goto err_uar_page;
3780 
3781 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
3782 	if (err)
3783 		goto err_bfreg;
3784 
3785 	err = ib_register_device(&dev->ib_dev, NULL);
3786 	if (err)
3787 		goto err_fp_bfreg;
3788 
3789 	err = create_umr_res(dev);
3790 	if (err)
3791 		goto err_dev;
3792 
3793 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
3794 		err = device_create_file(&dev->ib_dev.dev,
3795 					 mlx5_class_attributes[i]);
3796 		if (err)
3797 			goto err_umrc;
3798 	}
3799 
3800 	dev->ib_active = true;
3801 
3802 	return dev;
3803 
3804 err_umrc:
3805 	destroy_umrc_res(dev);
3806 
3807 err_dev:
3808 	ib_unregister_device(&dev->ib_dev);
3809 
3810 err_fp_bfreg:
3811 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3812 
3813 err_bfreg:
3814 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3815 
3816 err_uar_page:
3817 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
3818 
3819 err_cnt:
3820 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
3821 		mlx5_ib_dealloc_counters(dev);
3822 
3823 err_odp:
3824 	mlx5_ib_odp_remove_one(dev);
3825 
3826 err_rsrc:
3827 	destroy_dev_resources(&dev->devr);
3828 
3829 err_disable_eth:
3830 	if (ll == IB_LINK_LAYER_ETHERNET) {
3831 		mlx5_disable_eth(dev);
3832 		mlx5_remove_netdev_notifier(dev);
3833 	}
3834 
3835 err_free_port:
3836 	kfree(dev->port);
3837 
3838 err_dealloc:
3839 	ib_dealloc_device((struct ib_device *)dev);
3840 
3841 	return NULL;
3842 }
3843 
3844 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3845 {
3846 	struct mlx5_ib_dev *dev = context;
3847 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3848 
3849 	mlx5_remove_netdev_notifier(dev);
3850 	ib_unregister_device(&dev->ib_dev);
3851 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3852 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3853 	mlx5_put_uars_page(dev->mdev, mdev->priv.uar);
3854 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
3855 		mlx5_ib_dealloc_counters(dev);
3856 	destroy_umrc_res(dev);
3857 	mlx5_ib_odp_remove_one(dev);
3858 	destroy_dev_resources(&dev->devr);
3859 	if (ll == IB_LINK_LAYER_ETHERNET)
3860 		mlx5_disable_eth(dev);
3861 	kfree(dev->port);
3862 	ib_dealloc_device(&dev->ib_dev);
3863 }
3864 
3865 static struct mlx5_interface mlx5_ib_interface = {
3866 	.add            = mlx5_ib_add,
3867 	.remove         = mlx5_ib_remove,
3868 	.event          = mlx5_ib_event,
3869 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
3870 	.pfault		= mlx5_ib_pfault,
3871 #endif
3872 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
3873 };
3874 
3875 static int __init mlx5_ib_init(void)
3876 {
3877 	int err;
3878 
3879 	mlx5_ib_odp_init();
3880 
3881 	err = mlx5_register_interface(&mlx5_ib_interface);
3882 
3883 	return err;
3884 }
3885 
3886 static void __exit mlx5_ib_cleanup(void)
3887 {
3888 	mlx5_unregister_interface(&mlx5_ib_interface);
3889 }
3890 
3891 module_init(mlx5_ib_init);
3892 module_exit(mlx5_ib_cleanup);
3893