xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_main.c (revision c5fda9bac0325eb8c5b447717862d279006f318f)
1 /*-
2  * Copyright (c) 2013-2015, Mellanox Technologies, Ltd.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  * $FreeBSD$
26  */
27 
28 #include <linux/module.h>
29 #include <linux/errno.h>
30 #include <linux/pci.h>
31 #include <linux/dma-mapping.h>
32 #include <linux/slab.h>
33 #if defined(CONFIG_X86)
34 #include <asm/pat.h>
35 #endif
36 #include <linux/sched.h>
37 #include <linux/delay.h>
38 #include <linux/fs.h>
39 #undef inode
40 #include <rdma/ib_user_verbs.h>
41 #include <rdma/ib_addr.h>
42 #include <rdma/ib_cache.h>
43 #include <dev/mlx5/port.h>
44 #include <dev/mlx5/vport.h>
45 #include <linux/list.h>
46 #include <rdma/ib_smi.h>
47 #include <rdma/ib_umem.h>
48 #include <linux/in.h>
49 #include <linux/etherdevice.h>
50 #include <dev/mlx5/fs.h>
51 #include "mlx5_ib.h"
52 
53 #define DRIVER_NAME "mlx5_ib"
54 #define DRIVER_VERSION "3.4.1-BETA"
55 #define DRIVER_RELDATE	"October 2017"
56 
57 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
58 MODULE_LICENSE("Dual BSD/GPL");
59 MODULE_DEPEND(mlx5ib, linuxkpi, 1, 1, 1);
60 MODULE_DEPEND(mlx5ib, mlx5, 1, 1, 1);
61 MODULE_DEPEND(mlx5ib, ibcore, 1, 1, 1);
62 MODULE_VERSION(mlx5ib, 1);
63 
64 static int deprecated_prof_sel = 2;
65 module_param_named(prof_sel, deprecated_prof_sel, int, 0444);
66 MODULE_PARM_DESC(prof_sel, "profile selector. Deprecated here. Moved to module mlx5_core");
67 
68 static char mlx5_version[] =
69 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
70 	DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
71 
72 enum {
73 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
74 };
75 
76 static enum rdma_link_layer
77 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
78 {
79 	switch (port_type_cap) {
80 	case MLX5_CAP_PORT_TYPE_IB:
81 		return IB_LINK_LAYER_INFINIBAND;
82 	case MLX5_CAP_PORT_TYPE_ETH:
83 		return IB_LINK_LAYER_ETHERNET;
84 	default:
85 		return IB_LINK_LAYER_UNSPECIFIED;
86 	}
87 }
88 
89 static enum rdma_link_layer
90 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
91 {
92 	struct mlx5_ib_dev *dev = to_mdev(device);
93 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
94 
95 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
96 }
97 
98 static bool mlx5_netdev_match(struct net_device *ndev,
99 			      struct mlx5_core_dev *mdev,
100 			      const char *dname)
101 {
102 	return ndev->if_type == IFT_ETHER &&
103 	  ndev->if_dname != NULL &&
104 	  strcmp(ndev->if_dname, dname) == 0 &&
105 	  ndev->if_softc != NULL &&
106 	  *(struct mlx5_core_dev **)ndev->if_softc == mdev;
107 }
108 
109 static int mlx5_netdev_event(struct notifier_block *this,
110 			     unsigned long event, void *ptr)
111 {
112 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
113 	struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
114 						 roce.nb);
115 
116 	switch (event) {
117 	case NETDEV_REGISTER:
118 	case NETDEV_UNREGISTER:
119 		write_lock(&ibdev->roce.netdev_lock);
120 		/* check if network interface belongs to mlx5en */
121 		if (mlx5_netdev_match(ndev, ibdev->mdev, "mce"))
122 			ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
123 					     NULL : ndev;
124 		write_unlock(&ibdev->roce.netdev_lock);
125 		break;
126 
127 	case NETDEV_UP:
128 	case NETDEV_DOWN: {
129 		struct net_device *upper = NULL;
130 
131 		if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
132 		    && ibdev->ib_active) {
133 			struct ib_event ibev = {0};
134 
135 			ibev.device = &ibdev->ib_dev;
136 			ibev.event = (event == NETDEV_UP) ?
137 				     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
138 			ibev.element.port_num = 1;
139 			ib_dispatch_event(&ibev);
140 		}
141 		break;
142 	}
143 
144 	default:
145 		break;
146 	}
147 
148 	return NOTIFY_DONE;
149 }
150 
151 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
152 					     u8 port_num)
153 {
154 	struct mlx5_ib_dev *ibdev = to_mdev(device);
155 	struct net_device *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 net_device *ndev;
231 	enum ib_mtu ndev_ib_mtu;
232 	u16 qkey_viol_cntr;
233 	u32 eth_prot_oper;
234 	int err;
235 
236 	memset(props, 0, sizeof(*props));
237 
238 	/* Possible bad flows are checked before filling out props so in case
239 	 * of an error it will still be zeroed out.
240 	 */
241 	err = mlx5_query_port_eth_proto_oper(dev->mdev, &eth_prot_oper, port_num);
242 	if (err)
243 		return err;
244 
245 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
246 				 &props->active_width);
247 
248 	props->port_cap_flags  |= IB_PORT_CM_SUP;
249 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
250 
251 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
252 						roce_address_table_size);
253 	props->max_mtu          = IB_MTU_4096;
254 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
255 	props->pkey_tbl_len     = 1;
256 	props->state            = IB_PORT_DOWN;
257 	props->phys_state       = 3;
258 
259 	mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
260 	props->qkey_viol_cntr = qkey_viol_cntr;
261 
262 	ndev = mlx5_ib_get_netdev(device, port_num);
263 	if (!ndev)
264 		return 0;
265 
266 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
267 		props->state      = IB_PORT_ACTIVE;
268 		props->phys_state = 5;
269 	}
270 
271 	ndev_ib_mtu = iboe_get_mtu(ndev->if_mtu);
272 
273 	dev_put(ndev);
274 
275 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
276 	return 0;
277 }
278 
279 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
280 				     const struct ib_gid_attr *attr,
281 				     void *mlx5_addr)
282 {
283 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
284 	char *mlx5_addr_l3_addr	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
285 					       source_l3_address);
286 	void *mlx5_addr_mac	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
287 					       source_mac_47_32);
288 
289 	if (!gid)
290 		return;
291 	ether_addr_copy(mlx5_addr_mac, IF_LLADDR(attr->ndev));
292 
293 	if (is_vlan_dev(attr->ndev)) {
294 		MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
295 		MLX5_SET_RA(mlx5_addr, vlan_id, vlan_dev_vlan_id(attr->ndev));
296 	}
297 
298 	switch (attr->gid_type) {
299 	case IB_GID_TYPE_IB:
300 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
301 		break;
302 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
303 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
304 		break;
305 
306 	default:
307 		WARN_ON(true);
308 	}
309 
310 	if (attr->gid_type != IB_GID_TYPE_IB) {
311 		if (ipv6_addr_v4mapped((void *)gid))
312 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
313 				    MLX5_ROCE_L3_TYPE_IPV4);
314 		else
315 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
316 				    MLX5_ROCE_L3_TYPE_IPV6);
317 	}
318 
319 	if ((attr->gid_type == IB_GID_TYPE_IB) ||
320 	    !ipv6_addr_v4mapped((void *)gid))
321 		memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
322 	else
323 		memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
324 }
325 
326 static int set_roce_addr(struct ib_device *device, u8 port_num,
327 			 unsigned int index,
328 			 const union ib_gid *gid,
329 			 const struct ib_gid_attr *attr)
330 {
331 	struct mlx5_ib_dev *dev = to_mdev(device);
332 	u32  in[MLX5_ST_SZ_DW(set_roce_address_in)]  = {0};
333 	u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0};
334 	void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
335 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
336 
337 	if (ll != IB_LINK_LAYER_ETHERNET)
338 		return -EINVAL;
339 
340 	ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
341 
342 	MLX5_SET(set_roce_address_in, in, roce_address_index, index);
343 	MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
344 	return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
345 }
346 
347 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
348 			   unsigned int index, const union ib_gid *gid,
349 			   const struct ib_gid_attr *attr,
350 			   __always_unused void **context)
351 {
352 	return set_roce_addr(device, port_num, index, gid, attr);
353 }
354 
355 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
356 			   unsigned int index, __always_unused void **context)
357 {
358 	return set_roce_addr(device, port_num, index, NULL, NULL);
359 }
360 
361 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
362 			       int index)
363 {
364 	struct ib_gid_attr attr;
365 	union ib_gid gid;
366 
367 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
368 		return 0;
369 
370 	if (!attr.ndev)
371 		return 0;
372 
373 	dev_put(attr.ndev);
374 
375 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
376 		return 0;
377 
378 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
379 }
380 
381 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
382 {
383 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
384 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
385 	return 0;
386 }
387 
388 enum {
389 	MLX5_VPORT_ACCESS_METHOD_MAD,
390 	MLX5_VPORT_ACCESS_METHOD_HCA,
391 	MLX5_VPORT_ACCESS_METHOD_NIC,
392 };
393 
394 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
395 {
396 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
397 		return MLX5_VPORT_ACCESS_METHOD_MAD;
398 
399 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
400 	    IB_LINK_LAYER_ETHERNET)
401 		return MLX5_VPORT_ACCESS_METHOD_NIC;
402 
403 	return MLX5_VPORT_ACCESS_METHOD_HCA;
404 }
405 
406 static void get_atomic_caps(struct mlx5_ib_dev *dev,
407 			    struct ib_device_attr *props)
408 {
409 	u8 tmp;
410 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
411 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
412 	u8 atomic_req_8B_endianness_mode =
413 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
414 
415 	/* Check if HW supports 8 bytes standard atomic operations and capable
416 	 * of host endianness respond
417 	 */
418 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
419 	if (((atomic_operations & tmp) == tmp) &&
420 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
421 	    (atomic_req_8B_endianness_mode)) {
422 		props->atomic_cap = IB_ATOMIC_HCA;
423 	} else {
424 		props->atomic_cap = IB_ATOMIC_NONE;
425 	}
426 }
427 
428 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
429 					__be64 *sys_image_guid)
430 {
431 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
432 	struct mlx5_core_dev *mdev = dev->mdev;
433 	u64 tmp;
434 	int err;
435 
436 	switch (mlx5_get_vport_access_method(ibdev)) {
437 	case MLX5_VPORT_ACCESS_METHOD_MAD:
438 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
439 							    sys_image_guid);
440 
441 	case MLX5_VPORT_ACCESS_METHOD_HCA:
442 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
443 		break;
444 
445 	case MLX5_VPORT_ACCESS_METHOD_NIC:
446 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
447 		break;
448 
449 	default:
450 		return -EINVAL;
451 	}
452 
453 	if (!err)
454 		*sys_image_guid = cpu_to_be64(tmp);
455 
456 	return err;
457 
458 }
459 
460 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
461 				u16 *max_pkeys)
462 {
463 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
464 	struct mlx5_core_dev *mdev = dev->mdev;
465 
466 	switch (mlx5_get_vport_access_method(ibdev)) {
467 	case MLX5_VPORT_ACCESS_METHOD_MAD:
468 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
469 
470 	case MLX5_VPORT_ACCESS_METHOD_HCA:
471 	case MLX5_VPORT_ACCESS_METHOD_NIC:
472 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
473 						pkey_table_size));
474 		return 0;
475 
476 	default:
477 		return -EINVAL;
478 	}
479 }
480 
481 static int mlx5_query_vendor_id(struct ib_device *ibdev,
482 				u32 *vendor_id)
483 {
484 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
485 
486 	switch (mlx5_get_vport_access_method(ibdev)) {
487 	case MLX5_VPORT_ACCESS_METHOD_MAD:
488 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
489 
490 	case MLX5_VPORT_ACCESS_METHOD_HCA:
491 	case MLX5_VPORT_ACCESS_METHOD_NIC:
492 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
493 
494 	default:
495 		return -EINVAL;
496 	}
497 }
498 
499 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
500 				__be64 *node_guid)
501 {
502 	u64 tmp;
503 	int err;
504 
505 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
506 	case MLX5_VPORT_ACCESS_METHOD_MAD:
507 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
508 
509 	case MLX5_VPORT_ACCESS_METHOD_HCA:
510 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
511 		break;
512 
513 	case MLX5_VPORT_ACCESS_METHOD_NIC:
514 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
515 		break;
516 
517 	default:
518 		return -EINVAL;
519 	}
520 
521 	if (!err)
522 		*node_guid = cpu_to_be64(tmp);
523 
524 	return err;
525 }
526 
527 struct mlx5_reg_node_desc {
528 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
529 };
530 
531 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
532 {
533 	struct mlx5_reg_node_desc in;
534 
535 	if (mlx5_use_mad_ifc(dev))
536 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
537 
538 	memset(&in, 0, sizeof(in));
539 
540 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
541 				    sizeof(struct mlx5_reg_node_desc),
542 				    MLX5_REG_NODE_DESC, 0, 0);
543 }
544 
545 static int mlx5_ib_query_device(struct ib_device *ibdev,
546 				struct ib_device_attr *props,
547 				struct ib_udata *uhw)
548 {
549 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
550 	struct mlx5_core_dev *mdev = dev->mdev;
551 	int err = -ENOMEM;
552 	int max_rq_sg;
553 	int max_sq_sg;
554 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
555 	struct mlx5_ib_query_device_resp resp = {};
556 	size_t resp_len;
557 	u64 max_tso;
558 
559 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
560 	if (uhw->outlen && uhw->outlen < resp_len)
561 		return -EINVAL;
562 	else
563 		resp.response_length = resp_len;
564 
565 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
566 		return -EINVAL;
567 
568 	memset(props, 0, sizeof(*props));
569 	err = mlx5_query_system_image_guid(ibdev,
570 					   &props->sys_image_guid);
571 	if (err)
572 		return err;
573 
574 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
575 	if (err)
576 		return err;
577 
578 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
579 	if (err)
580 		return err;
581 
582 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
583 		(fw_rev_min(dev->mdev) << 16) |
584 		fw_rev_sub(dev->mdev);
585 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
586 		IB_DEVICE_PORT_ACTIVE_EVENT		|
587 		IB_DEVICE_SYS_IMAGE_GUID		|
588 		IB_DEVICE_RC_RNR_NAK_GEN;
589 
590 	if (MLX5_CAP_GEN(mdev, pkv))
591 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
592 	if (MLX5_CAP_GEN(mdev, qkv))
593 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
594 	if (MLX5_CAP_GEN(mdev, apm))
595 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
596 	if (MLX5_CAP_GEN(mdev, xrc))
597 		props->device_cap_flags |= IB_DEVICE_XRC;
598 	if (MLX5_CAP_GEN(mdev, imaicl)) {
599 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
600 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
601 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
602 		/* We support 'Gappy' memory registration too */
603 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
604 	}
605 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
606 	if (MLX5_CAP_GEN(mdev, sho)) {
607 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
608 		/* At this stage no support for signature handover */
609 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
610 				      IB_PROT_T10DIF_TYPE_2 |
611 				      IB_PROT_T10DIF_TYPE_3;
612 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
613 				       IB_GUARD_T10DIF_CSUM;
614 	}
615 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
616 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
617 
618 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
619 		if (MLX5_CAP_ETH(mdev, csum_cap))
620 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
621 
622 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
623 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
624 			if (max_tso) {
625 				resp.tso_caps.max_tso = 1 << max_tso;
626 				resp.tso_caps.supported_qpts |=
627 					1 << IB_QPT_RAW_PACKET;
628 				resp.response_length += sizeof(resp.tso_caps);
629 			}
630 		}
631 
632 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
633 			resp.rss_caps.rx_hash_function =
634 						MLX5_RX_HASH_FUNC_TOEPLITZ;
635 			resp.rss_caps.rx_hash_fields_mask =
636 						MLX5_RX_HASH_SRC_IPV4 |
637 						MLX5_RX_HASH_DST_IPV4 |
638 						MLX5_RX_HASH_SRC_IPV6 |
639 						MLX5_RX_HASH_DST_IPV6 |
640 						MLX5_RX_HASH_SRC_PORT_TCP |
641 						MLX5_RX_HASH_DST_PORT_TCP |
642 						MLX5_RX_HASH_SRC_PORT_UDP |
643 						MLX5_RX_HASH_DST_PORT_UDP;
644 			resp.response_length += sizeof(resp.rss_caps);
645 		}
646 	} else {
647 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
648 			resp.response_length += sizeof(resp.tso_caps);
649 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
650 			resp.response_length += sizeof(resp.rss_caps);
651 	}
652 
653 	if (MLX5_CAP_GEN(mdev, ipoib_ipoib_offloads)) {
654 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
655 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
656 	}
657 
658 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
659 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs))
660 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
661 
662 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
663 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
664 
665 	props->vendor_part_id	   = mdev->pdev->device;
666 	props->hw_ver		   = mdev->pdev->revision;
667 
668 	props->max_mr_size	   = ~0ull;
669 	props->page_size_cap	   = ~(min_page_size - 1);
670 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
671 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
672 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
673 		     sizeof(struct mlx5_wqe_data_seg);
674 	max_sq_sg = (MLX5_CAP_GEN(mdev, max_wqe_sz_sq) -
675 		     sizeof(struct mlx5_wqe_ctrl_seg)) /
676 		     sizeof(struct mlx5_wqe_data_seg);
677 	props->max_sge = min(max_rq_sg, max_sq_sg);
678 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
679 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
680 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
681 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
682 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
683 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
684 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
685 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
686 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
687 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
688 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
689 	props->max_srq_sge	   = max_rq_sg - 1;
690 	props->max_fast_reg_page_list_len =
691 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
692 	get_atomic_caps(dev, props);
693 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
694 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
695 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
696 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
697 					   props->max_mcast_grp;
698 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
699 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
700 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
701 
702 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
703 	if (MLX5_CAP_GEN(mdev, pg))
704 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
705 	props->odp_caps = dev->odp_caps;
706 #endif
707 
708 	if (MLX5_CAP_GEN(mdev, cd))
709 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
710 
711 	if (!mlx5_core_is_pf(mdev))
712 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
713 
714 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
715 	    IB_LINK_LAYER_ETHERNET) {
716 		props->rss_caps.max_rwq_indirection_tables =
717 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
718 		props->rss_caps.max_rwq_indirection_table_size =
719 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
720 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
721 		props->max_wq_type_rq =
722 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
723 	}
724 
725 	if (uhw->outlen) {
726 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
727 
728 		if (err)
729 			return err;
730 	}
731 
732 	return 0;
733 }
734 
735 enum mlx5_ib_width {
736 	MLX5_IB_WIDTH_1X	= 1 << 0,
737 	MLX5_IB_WIDTH_2X	= 1 << 1,
738 	MLX5_IB_WIDTH_4X	= 1 << 2,
739 	MLX5_IB_WIDTH_8X	= 1 << 3,
740 	MLX5_IB_WIDTH_12X	= 1 << 4
741 };
742 
743 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
744 				  u8 *ib_width)
745 {
746 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
747 	int err = 0;
748 
749 	if (active_width & MLX5_IB_WIDTH_1X) {
750 		*ib_width = IB_WIDTH_1X;
751 	} else if (active_width & MLX5_IB_WIDTH_2X) {
752 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
753 			    (int)active_width);
754 		err = -EINVAL;
755 	} else if (active_width & MLX5_IB_WIDTH_4X) {
756 		*ib_width = IB_WIDTH_4X;
757 	} else if (active_width & MLX5_IB_WIDTH_8X) {
758 		*ib_width = IB_WIDTH_8X;
759 	} else if (active_width & MLX5_IB_WIDTH_12X) {
760 		*ib_width = IB_WIDTH_12X;
761 	} else {
762 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
763 			    (int)active_width);
764 		err = -EINVAL;
765 	}
766 
767 	return err;
768 }
769 
770 enum ib_max_vl_num {
771 	__IB_MAX_VL_0		= 1,
772 	__IB_MAX_VL_0_1		= 2,
773 	__IB_MAX_VL_0_3		= 3,
774 	__IB_MAX_VL_0_7		= 4,
775 	__IB_MAX_VL_0_14	= 5,
776 };
777 
778 enum mlx5_vl_hw_cap {
779 	MLX5_VL_HW_0	= 1,
780 	MLX5_VL_HW_0_1	= 2,
781 	MLX5_VL_HW_0_2	= 3,
782 	MLX5_VL_HW_0_3	= 4,
783 	MLX5_VL_HW_0_4	= 5,
784 	MLX5_VL_HW_0_5	= 6,
785 	MLX5_VL_HW_0_6	= 7,
786 	MLX5_VL_HW_0_7	= 8,
787 	MLX5_VL_HW_0_14	= 15
788 };
789 
790 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
791 				u8 *max_vl_num)
792 {
793 	switch (vl_hw_cap) {
794 	case MLX5_VL_HW_0:
795 		*max_vl_num = __IB_MAX_VL_0;
796 		break;
797 	case MLX5_VL_HW_0_1:
798 		*max_vl_num = __IB_MAX_VL_0_1;
799 		break;
800 	case MLX5_VL_HW_0_3:
801 		*max_vl_num = __IB_MAX_VL_0_3;
802 		break;
803 	case MLX5_VL_HW_0_7:
804 		*max_vl_num = __IB_MAX_VL_0_7;
805 		break;
806 	case MLX5_VL_HW_0_14:
807 		*max_vl_num = __IB_MAX_VL_0_14;
808 		break;
809 
810 	default:
811 		return -EINVAL;
812 	}
813 
814 	return 0;
815 }
816 
817 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
818 			       struct ib_port_attr *props)
819 {
820 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
821 	struct mlx5_core_dev *mdev = dev->mdev;
822 	u32 *rep;
823 	int replen = MLX5_ST_SZ_BYTES(query_hca_vport_context_out);
824 	struct mlx5_ptys_reg *ptys;
825 	struct mlx5_pmtu_reg *pmtu;
826 	struct mlx5_pvlc_reg pvlc;
827 	void *ctx;
828 	int err;
829 
830 	rep = mlx5_vzalloc(replen);
831 	ptys = kzalloc(sizeof(*ptys), GFP_KERNEL);
832 	pmtu = kzalloc(sizeof(*pmtu), GFP_KERNEL);
833 	if (!rep || !ptys || !pmtu) {
834 		err = -ENOMEM;
835 		goto out;
836 	}
837 
838 	memset(props, 0, sizeof(*props));
839 
840 	err = mlx5_query_hca_vport_context(mdev, port, 0, rep, replen);
841 	if (err)
842 		goto out;
843 
844 	ctx = MLX5_ADDR_OF(query_hca_vport_context_out, rep, hca_vport_context);
845 
846 	props->lid		= MLX5_GET(hca_vport_context, ctx, lid);
847 	props->lmc		= MLX5_GET(hca_vport_context, ctx, lmc);
848 	props->sm_lid		= MLX5_GET(hca_vport_context, ctx, sm_lid);
849 	props->sm_sl		= MLX5_GET(hca_vport_context, ctx, sm_sl);
850 	props->state		= MLX5_GET(hca_vport_context, ctx, vport_state);
851 	props->phys_state	= MLX5_GET(hca_vport_context, ctx,
852 					port_physical_state);
853 	props->port_cap_flags	= MLX5_GET(hca_vport_context, ctx, cap_mask1);
854 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
855 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
856 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
857 	props->bad_pkey_cntr	= MLX5_GET(hca_vport_context, ctx,
858 					pkey_violation_counter);
859 	props->qkey_viol_cntr	= MLX5_GET(hca_vport_context, ctx,
860 					qkey_violation_counter);
861 	props->subnet_timeout	= MLX5_GET(hca_vport_context, ctx,
862 					subnet_timeout);
863 	props->init_type_reply	= MLX5_GET(hca_vport_context, ctx,
864 					init_type_reply);
865 	props->grh_required	= MLX5_GET(hca_vport_context, ctx, grh_required);
866 
867 	ptys->proto_mask |= MLX5_PTYS_IB;
868 	ptys->local_port = port;
869 	err = mlx5_core_access_ptys(mdev, ptys, 0);
870 	if (err)
871 		goto out;
872 
873 	err = translate_active_width(ibdev, ptys->ib_link_width_oper,
874 				     &props->active_width);
875 	if (err)
876 		goto out;
877 
878 	props->active_speed	= (u8)ptys->ib_proto_oper;
879 
880 	pmtu->local_port = port;
881 	err = mlx5_core_access_pmtu(mdev, pmtu, 0);
882 	if (err)
883 		goto out;
884 
885 	props->max_mtu		= pmtu->max_mtu;
886 	props->active_mtu	= pmtu->oper_mtu;
887 
888 	memset(&pvlc, 0, sizeof(pvlc));
889 	pvlc.local_port = port;
890 	err = mlx5_core_access_pvlc(mdev, &pvlc, 0);
891 	if (err)
892 		goto out;
893 
894 	err = translate_max_vl_num(ibdev, pvlc.vl_hw_cap,
895 				   &props->max_vl_num);
896 out:
897 	kvfree(rep);
898 	kfree(ptys);
899 	kfree(pmtu);
900 	return err;
901 }
902 
903 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
904 		       struct ib_port_attr *props)
905 {
906 	switch (mlx5_get_vport_access_method(ibdev)) {
907 	case MLX5_VPORT_ACCESS_METHOD_MAD:
908 		return mlx5_query_mad_ifc_port(ibdev, port, props);
909 
910 	case MLX5_VPORT_ACCESS_METHOD_HCA:
911 		return mlx5_query_hca_port(ibdev, port, props);
912 
913 	case MLX5_VPORT_ACCESS_METHOD_NIC:
914 		return mlx5_query_port_roce(ibdev, port, props);
915 
916 	default:
917 		return -EINVAL;
918 	}
919 }
920 
921 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
922 			     union ib_gid *gid)
923 {
924 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
925 	struct mlx5_core_dev *mdev = dev->mdev;
926 
927 	switch (mlx5_get_vport_access_method(ibdev)) {
928 	case MLX5_VPORT_ACCESS_METHOD_MAD:
929 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
930 
931 	case MLX5_VPORT_ACCESS_METHOD_HCA:
932 		return mlx5_query_hca_vport_gid(mdev, port, 0, index, gid);
933 
934 	default:
935 		return -EINVAL;
936 	}
937 
938 }
939 
940 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
941 			      u16 *pkey)
942 {
943 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
944 	struct mlx5_core_dev *mdev = dev->mdev;
945 
946 	switch (mlx5_get_vport_access_method(ibdev)) {
947 	case MLX5_VPORT_ACCESS_METHOD_MAD:
948 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
949 
950 	case MLX5_VPORT_ACCESS_METHOD_HCA:
951 	case MLX5_VPORT_ACCESS_METHOD_NIC:
952 		return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
953 						 pkey);
954 	default:
955 		return -EINVAL;
956 	}
957 }
958 
959 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
960 				 struct ib_device_modify *props)
961 {
962 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
963 	struct mlx5_reg_node_desc in;
964 	struct mlx5_reg_node_desc out;
965 	int err;
966 
967 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
968 		return -EOPNOTSUPP;
969 
970 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
971 		return 0;
972 
973 	/*
974 	 * If possible, pass node desc to FW, so it can generate
975 	 * a 144 trap.  If cmd fails, just ignore.
976 	 */
977 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
978 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
979 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
980 	if (err)
981 		return err;
982 
983 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
984 
985 	return err;
986 }
987 
988 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
989 			       struct ib_port_modify *props)
990 {
991 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
992 	struct ib_port_attr attr;
993 	u32 tmp;
994 	int err;
995 
996 	mutex_lock(&dev->cap_mask_mutex);
997 
998 	err = mlx5_ib_query_port(ibdev, port, &attr);
999 	if (err)
1000 		goto out;
1001 
1002 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1003 		~props->clr_port_cap_mask;
1004 
1005 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1006 
1007 out:
1008 	mutex_unlock(&dev->cap_mask_mutex);
1009 	return err;
1010 }
1011 
1012 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1013 						  struct ib_udata *udata)
1014 {
1015 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1016 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1017 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1018 	struct mlx5_ib_ucontext *context;
1019 	struct mlx5_uuar_info *uuari;
1020 	struct mlx5_uar *uars;
1021 	int gross_uuars;
1022 	int num_uars;
1023 	int ver;
1024 	int uuarn;
1025 	int err;
1026 	int i;
1027 	size_t reqlen;
1028 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1029 				     max_cqe_version);
1030 
1031 	if (!dev->ib_active)
1032 		return ERR_PTR(-EAGAIN);
1033 
1034 	if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
1035 		return ERR_PTR(-EINVAL);
1036 
1037 	reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
1038 	if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1039 		ver = 0;
1040 	else if (reqlen >= min_req_v2)
1041 		ver = 2;
1042 	else
1043 		return ERR_PTR(-EINVAL);
1044 
1045 	err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
1046 	if (err)
1047 		return ERR_PTR(err);
1048 
1049 	if (req.flags)
1050 		return ERR_PTR(-EINVAL);
1051 
1052 	if (req.total_num_uuars > MLX5_MAX_UUARS)
1053 		return ERR_PTR(-ENOMEM);
1054 
1055 	if (req.total_num_uuars == 0)
1056 		return ERR_PTR(-EINVAL);
1057 
1058 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1059 		return ERR_PTR(-EOPNOTSUPP);
1060 
1061 	if (reqlen > sizeof(req) &&
1062 	    !ib_is_udata_cleared(udata, sizeof(req),
1063 				 reqlen - sizeof(req)))
1064 		return ERR_PTR(-EOPNOTSUPP);
1065 
1066 	req.total_num_uuars = ALIGN(req.total_num_uuars,
1067 				    MLX5_NON_FP_BF_REGS_PER_PAGE);
1068 	if (req.num_low_latency_uuars > req.total_num_uuars - 1)
1069 		return ERR_PTR(-EINVAL);
1070 
1071 	num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
1072 	gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
1073 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1074 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1075 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1076 	resp.cache_line_size = cache_line_size();
1077 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1078 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1079 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1080 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1081 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1082 	resp.cqe_version = min_t(__u8,
1083 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1084 				 req.max_cqe_version);
1085 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1086 				   sizeof(resp.response_length), udata->outlen);
1087 
1088 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1089 	if (!context)
1090 		return ERR_PTR(-ENOMEM);
1091 
1092 	uuari = &context->uuari;
1093 	mutex_init(&uuari->lock);
1094 	uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
1095 	if (!uars) {
1096 		err = -ENOMEM;
1097 		goto out_ctx;
1098 	}
1099 
1100 	uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
1101 				sizeof(*uuari->bitmap),
1102 				GFP_KERNEL);
1103 	if (!uuari->bitmap) {
1104 		err = -ENOMEM;
1105 		goto out_uar_ctx;
1106 	}
1107 	/*
1108 	 * clear all fast path uuars
1109 	 */
1110 	for (i = 0; i < gross_uuars; i++) {
1111 		uuarn = i & 3;
1112 		if (uuarn == 2 || uuarn == 3)
1113 			set_bit(i, uuari->bitmap);
1114 	}
1115 
1116 	uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
1117 	if (!uuari->count) {
1118 		err = -ENOMEM;
1119 		goto out_bitmap;
1120 	}
1121 
1122 	for (i = 0; i < num_uars; i++) {
1123 		err = mlx5_cmd_alloc_uar(dev->mdev, &uars[i].index);
1124 		if (err)
1125 			goto out_count;
1126 	}
1127 
1128 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1129 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1130 #endif
1131 
1132 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1133 		err = mlx5_alloc_transport_domain(dev->mdev,
1134 						       &context->tdn);
1135 		if (err)
1136 			goto out_uars;
1137 	}
1138 
1139 	INIT_LIST_HEAD(&context->vma_private_list);
1140 	INIT_LIST_HEAD(&context->db_page_list);
1141 	mutex_init(&context->db_page_mutex);
1142 
1143 	resp.tot_uuars = req.total_num_uuars;
1144 	resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1145 
1146 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1147 		resp.response_length += sizeof(resp.cqe_version);
1148 
1149 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1150 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1151 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1152 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1153 	}
1154 
1155 	/*
1156 	 * We don't want to expose information from the PCI bar that is located
1157 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1158 	 * pretend we don't support reading the HCA's core clock. This is also
1159 	 * forced by mmap function.
1160 	 */
1161 	if (PAGE_SIZE <= 4096 &&
1162 	    field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1163 		resp.comp_mask |=
1164 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1165 		resp.hca_core_clock_offset =
1166 			offsetof(struct mlx5_init_seg, internal_timer_h) %
1167 			PAGE_SIZE;
1168 		resp.response_length += sizeof(resp.hca_core_clock_offset) +
1169 					sizeof(resp.reserved2);
1170 	}
1171 
1172 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1173 	if (err)
1174 		goto out_td;
1175 
1176 	uuari->ver = ver;
1177 	uuari->num_low_latency_uuars = req.num_low_latency_uuars;
1178 	uuari->uars = uars;
1179 	uuari->num_uars = num_uars;
1180 	context->cqe_version = resp.cqe_version;
1181 
1182 	return &context->ibucontext;
1183 
1184 out_td:
1185 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1186 		mlx5_dealloc_transport_domain(dev->mdev, context->tdn);
1187 
1188 out_uars:
1189 	for (i--; i >= 0; i--)
1190 		mlx5_cmd_free_uar(dev->mdev, uars[i].index);
1191 out_count:
1192 	kfree(uuari->count);
1193 
1194 out_bitmap:
1195 	kfree(uuari->bitmap);
1196 
1197 out_uar_ctx:
1198 	kfree(uars);
1199 
1200 out_ctx:
1201 	kfree(context);
1202 	return ERR_PTR(err);
1203 }
1204 
1205 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1206 {
1207 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1208 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1209 	struct mlx5_uuar_info *uuari = &context->uuari;
1210 	int i;
1211 
1212 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1213 		mlx5_dealloc_transport_domain(dev->mdev, context->tdn);
1214 
1215 	for (i = 0; i < uuari->num_uars; i++) {
1216 		if (mlx5_cmd_free_uar(dev->mdev, uuari->uars[i].index))
1217 			mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
1218 	}
1219 
1220 	kfree(uuari->count);
1221 	kfree(uuari->bitmap);
1222 	kfree(uuari->uars);
1223 	kfree(context);
1224 
1225 	return 0;
1226 }
1227 
1228 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
1229 {
1230 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + index;
1231 }
1232 
1233 static int get_command(unsigned long offset)
1234 {
1235 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1236 }
1237 
1238 static int get_arg(unsigned long offset)
1239 {
1240 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1241 }
1242 
1243 static int get_index(unsigned long offset)
1244 {
1245 	return get_arg(offset);
1246 }
1247 
1248 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1249 {
1250 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1251 	 * is done through either mremap flow or split_vma (usually due to
1252 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1253 	 * as this VMA is strongly hardware related.  Therefore we set the
1254 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1255 	 * calling us again and trying to do incorrect actions.  We assume that
1256 	 * the original VMA size is exactly a single page, and therefore all
1257 	 * "splitting" operation will not happen to it.
1258 	 */
1259 	area->vm_ops = NULL;
1260 }
1261 
1262 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1263 {
1264 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1265 
1266 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1267 	 * file itself is closed, therefore no sync is needed with the regular
1268 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1269 	 * However need a sync with accessing the vma as part of
1270 	 * mlx5_ib_disassociate_ucontext.
1271 	 * The close operation is usually called under mm->mmap_sem except when
1272 	 * process is exiting.
1273 	 * The exiting case is handled explicitly as part of
1274 	 * mlx5_ib_disassociate_ucontext.
1275 	 */
1276 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1277 
1278 	/* setting the vma context pointer to null in the mlx5_ib driver's
1279 	 * private data, to protect a race condition in
1280 	 * mlx5_ib_disassociate_ucontext().
1281 	 */
1282 	mlx5_ib_vma_priv_data->vma = NULL;
1283 	list_del(&mlx5_ib_vma_priv_data->list);
1284 	kfree(mlx5_ib_vma_priv_data);
1285 }
1286 
1287 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1288 	.open = mlx5_ib_vma_open,
1289 	.close = mlx5_ib_vma_close
1290 };
1291 
1292 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1293 				struct mlx5_ib_ucontext *ctx)
1294 {
1295 	struct mlx5_ib_vma_private_data *vma_prv;
1296 	struct list_head *vma_head = &ctx->vma_private_list;
1297 
1298 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1299 	if (!vma_prv)
1300 		return -ENOMEM;
1301 
1302 	vma_prv->vma = vma;
1303 	vma->vm_private_data = vma_prv;
1304 	vma->vm_ops =  &mlx5_ib_vm_ops;
1305 
1306 	list_add(&vma_prv->list, vma_head);
1307 
1308 	return 0;
1309 }
1310 
1311 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1312 {
1313 	switch (cmd) {
1314 	case MLX5_IB_MMAP_WC_PAGE:
1315 		return "WC";
1316 	case MLX5_IB_MMAP_REGULAR_PAGE:
1317 		return "best effort WC";
1318 	case MLX5_IB_MMAP_NC_PAGE:
1319 		return "NC";
1320 	default:
1321 		return NULL;
1322 	}
1323 }
1324 
1325 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1326 		    struct vm_area_struct *vma,
1327 		    struct mlx5_ib_ucontext *context)
1328 {
1329 	struct mlx5_uuar_info *uuari = &context->uuari;
1330 	int err;
1331 	unsigned long idx;
1332 	phys_addr_t pfn, pa;
1333 	pgprot_t prot;
1334 
1335 	switch (cmd) {
1336 	case MLX5_IB_MMAP_WC_PAGE:
1337 /* Some architectures don't support WC memory */
1338 #if defined(CONFIG_X86)
1339 		if (!pat_enabled())
1340 			return -EPERM;
1341 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1342 			return -EPERM;
1343 #endif
1344 	/* fall through */
1345 	case MLX5_IB_MMAP_REGULAR_PAGE:
1346 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1347 		prot = pgprot_writecombine(vma->vm_page_prot);
1348 		break;
1349 	case MLX5_IB_MMAP_NC_PAGE:
1350 		prot = pgprot_noncached(vma->vm_page_prot);
1351 		break;
1352 	default:
1353 		return -EINVAL;
1354 	}
1355 
1356 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1357 		return -EINVAL;
1358 
1359 	idx = get_index(vma->vm_pgoff);
1360 	if (idx >= uuari->num_uars)
1361 		return -EINVAL;
1362 
1363 	pfn = uar_index2pfn(dev, uuari->uars[idx].index);
1364 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1365 
1366 	vma->vm_page_prot = prot;
1367 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1368 				 PAGE_SIZE, vma->vm_page_prot);
1369 	if (err) {
1370 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%llx, pfn=%pa, mmap_cmd=%s\n",
1371 			    err, (unsigned long long)vma->vm_start, &pfn, mmap_cmd2str(cmd));
1372 		return -EAGAIN;
1373 	}
1374 
1375 	pa = pfn << PAGE_SHIFT;
1376 	mlx5_ib_dbg(dev, "mapped %s at 0x%llx, PA %pa\n", mmap_cmd2str(cmd),
1377 		    (unsigned long long)vma->vm_start, &pa);
1378 
1379 	return mlx5_ib_set_vma_data(vma, context);
1380 }
1381 
1382 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1383 {
1384 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1385 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1386 	unsigned long command;
1387 	phys_addr_t pfn;
1388 
1389 	command = get_command(vma->vm_pgoff);
1390 	switch (command) {
1391 	case MLX5_IB_MMAP_WC_PAGE:
1392 	case MLX5_IB_MMAP_NC_PAGE:
1393 	case MLX5_IB_MMAP_REGULAR_PAGE:
1394 		return uar_mmap(dev, command, vma, context);
1395 
1396 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1397 		return -ENOSYS;
1398 
1399 	case MLX5_IB_MMAP_CORE_CLOCK:
1400 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1401 			return -EINVAL;
1402 
1403 		if (vma->vm_flags & VM_WRITE)
1404 			return -EPERM;
1405 
1406 		/* Don't expose to user-space information it shouldn't have */
1407 		if (PAGE_SIZE > 4096)
1408 			return -EOPNOTSUPP;
1409 
1410 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1411 		pfn = (dev->mdev->iseg_base +
1412 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1413 			PAGE_SHIFT;
1414 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1415 				       PAGE_SIZE, vma->vm_page_prot))
1416 			return -EAGAIN;
1417 
1418 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%llx, PA 0x%llx\n",
1419 			    (unsigned long long)vma->vm_start,
1420 			    (unsigned long long)pfn << PAGE_SHIFT);
1421 		break;
1422 
1423 	default:
1424 		return -EINVAL;
1425 	}
1426 
1427 	return 0;
1428 }
1429 
1430 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1431 				      struct ib_ucontext *context,
1432 				      struct ib_udata *udata)
1433 {
1434 	struct mlx5_ib_alloc_pd_resp resp;
1435 	struct mlx5_ib_pd *pd;
1436 	int err;
1437 
1438 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1439 	if (!pd)
1440 		return ERR_PTR(-ENOMEM);
1441 
1442 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1443 	if (err) {
1444 		kfree(pd);
1445 		return ERR_PTR(err);
1446 	}
1447 
1448 	if (context) {
1449 		resp.pdn = pd->pdn;
1450 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1451 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1452 			kfree(pd);
1453 			return ERR_PTR(-EFAULT);
1454 		}
1455 	}
1456 
1457 	return &pd->ibpd;
1458 }
1459 
1460 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1461 {
1462 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1463 	struct mlx5_ib_pd *mpd = to_mpd(pd);
1464 
1465 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1466 	kfree(mpd);
1467 
1468 	return 0;
1469 }
1470 
1471 enum {
1472 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
1473 	MATCH_CRITERIA_ENABLE_MISC_BIT,
1474 	MATCH_CRITERIA_ENABLE_INNER_BIT
1475 };
1476 
1477 #define HEADER_IS_ZERO(match_criteria, headers)			           \
1478 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1479 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
1480 
1481 static u8 get_match_criteria_enable(u32 *match_criteria)
1482 {
1483 	u8 match_criteria_enable;
1484 
1485 	match_criteria_enable =
1486 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1487 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
1488 	match_criteria_enable |=
1489 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1490 		MATCH_CRITERIA_ENABLE_MISC_BIT;
1491 	match_criteria_enable |=
1492 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1493 		MATCH_CRITERIA_ENABLE_INNER_BIT;
1494 
1495 	return match_criteria_enable;
1496 }
1497 
1498 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1499 {
1500 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1501 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1502 }
1503 
1504 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1505 {
1506 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1507 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1508 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1509 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1510 }
1511 
1512 #define LAST_ETH_FIELD vlan_tag
1513 #define LAST_IB_FIELD sl
1514 #define LAST_IPV4_FIELD tos
1515 #define LAST_IPV6_FIELD traffic_class
1516 #define LAST_TCP_UDP_FIELD src_port
1517 
1518 /* Field is the last supported field */
1519 #define FIELDS_NOT_SUPPORTED(filter, field)\
1520 	memchr_inv((void *)&filter.field  +\
1521 		   sizeof(filter.field), 0,\
1522 		   sizeof(filter) -\
1523 		   offsetof(typeof(filter), field) -\
1524 		   sizeof(filter.field))
1525 
1526 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1527 			   const union ib_flow_spec *ib_spec)
1528 {
1529 	void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1530 					     outer_headers);
1531 	void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1532 					     outer_headers);
1533 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1534 					   misc_parameters);
1535 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1536 					   misc_parameters);
1537 
1538 	switch (ib_spec->type) {
1539 	case IB_FLOW_SPEC_ETH:
1540 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1541 			return -ENOTSUPP;
1542 
1543 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1544 					     dmac_47_16),
1545 				ib_spec->eth.mask.dst_mac);
1546 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1547 					     dmac_47_16),
1548 				ib_spec->eth.val.dst_mac);
1549 
1550 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1551 					     smac_47_16),
1552 				ib_spec->eth.mask.src_mac);
1553 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1554 					     smac_47_16),
1555 				ib_spec->eth.val.src_mac);
1556 
1557 		if (ib_spec->eth.mask.vlan_tag) {
1558 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1559 				 cvlan_tag, 1);
1560 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1561 				 cvlan_tag, 1);
1562 
1563 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1564 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1565 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1566 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1567 
1568 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1569 				 first_cfi,
1570 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1571 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1572 				 first_cfi,
1573 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1574 
1575 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1576 				 first_prio,
1577 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1578 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1579 				 first_prio,
1580 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1581 		}
1582 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1583 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1584 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1585 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
1586 		break;
1587 	case IB_FLOW_SPEC_IPV4:
1588 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1589 			return -ENOTSUPP;
1590 
1591 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1592 			 ethertype, 0xffff);
1593 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1594 			 ethertype, ETH_P_IP);
1595 
1596 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1597 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1598 		       &ib_spec->ipv4.mask.src_ip,
1599 		       sizeof(ib_spec->ipv4.mask.src_ip));
1600 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1601 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1602 		       &ib_spec->ipv4.val.src_ip,
1603 		       sizeof(ib_spec->ipv4.val.src_ip));
1604 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1605 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1606 		       &ib_spec->ipv4.mask.dst_ip,
1607 		       sizeof(ib_spec->ipv4.mask.dst_ip));
1608 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1609 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1610 		       &ib_spec->ipv4.val.dst_ip,
1611 		       sizeof(ib_spec->ipv4.val.dst_ip));
1612 
1613 		set_tos(outer_headers_c, outer_headers_v,
1614 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
1615 
1616 		set_proto(outer_headers_c, outer_headers_v,
1617 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
1618 		break;
1619 	case IB_FLOW_SPEC_IPV6:
1620 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1621 			return -ENOTSUPP;
1622 
1623 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1624 			 ethertype, 0xffff);
1625 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1626 			 ethertype, IPPROTO_IPV6);
1627 
1628 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1629 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1630 		       &ib_spec->ipv6.mask.src_ip,
1631 		       sizeof(ib_spec->ipv6.mask.src_ip));
1632 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1633 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1634 		       &ib_spec->ipv6.val.src_ip,
1635 		       sizeof(ib_spec->ipv6.val.src_ip));
1636 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1637 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1638 		       &ib_spec->ipv6.mask.dst_ip,
1639 		       sizeof(ib_spec->ipv6.mask.dst_ip));
1640 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1641 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1642 		       &ib_spec->ipv6.val.dst_ip,
1643 		       sizeof(ib_spec->ipv6.val.dst_ip));
1644 
1645 		set_tos(outer_headers_c, outer_headers_v,
1646 			ib_spec->ipv6.mask.traffic_class,
1647 			ib_spec->ipv6.val.traffic_class);
1648 
1649 		set_proto(outer_headers_c, outer_headers_v,
1650 			  ib_spec->ipv6.mask.next_hdr,
1651 			  ib_spec->ipv6.val.next_hdr);
1652 
1653 		MLX5_SET(fte_match_set_misc, misc_params_c,
1654 			 outer_ipv6_flow_label,
1655 			 ntohl(ib_spec->ipv6.mask.flow_label));
1656 		MLX5_SET(fte_match_set_misc, misc_params_v,
1657 			 outer_ipv6_flow_label,
1658 			 ntohl(ib_spec->ipv6.val.flow_label));
1659 		break;
1660 	case IB_FLOW_SPEC_TCP:
1661 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1662 					 LAST_TCP_UDP_FIELD))
1663 			return -ENOTSUPP;
1664 
1665 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1666 			 0xff);
1667 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1668 			 IPPROTO_TCP);
1669 
1670 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport,
1671 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1672 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport,
1673 			 ntohs(ib_spec->tcp_udp.val.src_port));
1674 
1675 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport,
1676 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1677 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport,
1678 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1679 		break;
1680 	case IB_FLOW_SPEC_UDP:
1681 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1682 					 LAST_TCP_UDP_FIELD))
1683 			return -ENOTSUPP;
1684 
1685 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1686 			 0xff);
1687 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1688 			 IPPROTO_UDP);
1689 
1690 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport,
1691 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1692 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport,
1693 			 ntohs(ib_spec->tcp_udp.val.src_port));
1694 
1695 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport,
1696 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1697 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport,
1698 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1699 		break;
1700 	default:
1701 		return -EINVAL;
1702 	}
1703 
1704 	return 0;
1705 }
1706 
1707 /* If a flow could catch both multicast and unicast packets,
1708  * it won't fall into the multicast flow steering table and this rule
1709  * could steal other multicast packets.
1710  */
1711 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1712 {
1713 	struct ib_flow_spec_eth *eth_spec;
1714 
1715 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1716 	    ib_attr->size < sizeof(struct ib_flow_attr) +
1717 	    sizeof(struct ib_flow_spec_eth) ||
1718 	    ib_attr->num_of_specs < 1)
1719 		return false;
1720 
1721 	eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1722 	if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1723 	    eth_spec->size != sizeof(*eth_spec))
1724 		return false;
1725 
1726 	return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1727 	       is_multicast_ether_addr(eth_spec->val.dst_mac);
1728 }
1729 
1730 static bool is_valid_attr(const struct ib_flow_attr *flow_attr)
1731 {
1732 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1733 	bool has_ipv4_spec = false;
1734 	bool eth_type_ipv4 = true;
1735 	unsigned int spec_index;
1736 
1737 	/* Validate that ethertype is correct */
1738 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1739 		if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1740 		    ib_spec->eth.mask.ether_type) {
1741 			if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1742 			      ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1743 				eth_type_ipv4 = false;
1744 		} else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1745 			has_ipv4_spec = true;
1746 		}
1747 		ib_spec = (void *)ib_spec + ib_spec->size;
1748 	}
1749 	return !has_ipv4_spec || eth_type_ipv4;
1750 }
1751 
1752 static void put_flow_table(struct mlx5_ib_dev *dev,
1753 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
1754 {
1755 	prio->refcount -= !!ft_added;
1756 	if (!prio->refcount) {
1757 		mlx5_destroy_flow_table(prio->flow_table);
1758 		prio->flow_table = NULL;
1759 	}
1760 }
1761 
1762 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
1763 {
1764 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
1765 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
1766 							  struct mlx5_ib_flow_handler,
1767 							  ibflow);
1768 	struct mlx5_ib_flow_handler *iter, *tmp;
1769 
1770 	mutex_lock(&dev->flow_db.lock);
1771 
1772 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
1773 		mlx5_del_flow_rule(iter->rule);
1774 		put_flow_table(dev, iter->prio, true);
1775 		list_del(&iter->list);
1776 		kfree(iter);
1777 	}
1778 
1779 	mlx5_del_flow_rule(handler->rule);
1780 	put_flow_table(dev, handler->prio, true);
1781 	mutex_unlock(&dev->flow_db.lock);
1782 
1783 	kfree(handler);
1784 
1785 	return 0;
1786 }
1787 
1788 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
1789 {
1790 	priority *= 2;
1791 	if (!dont_trap)
1792 		priority++;
1793 	return priority;
1794 }
1795 
1796 enum flow_table_type {
1797 	MLX5_IB_FT_RX,
1798 	MLX5_IB_FT_TX
1799 };
1800 
1801 #define MLX5_FS_MAX_TYPES	 10
1802 #define MLX5_FS_MAX_ENTRIES	 32000UL
1803 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
1804 						struct ib_flow_attr *flow_attr,
1805 						enum flow_table_type ft_type)
1806 {
1807 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
1808 	struct mlx5_flow_namespace *ns = NULL;
1809 	struct mlx5_ib_flow_prio *prio;
1810 	struct mlx5_flow_table *ft;
1811 	int num_entries;
1812 	int num_groups;
1813 	int priority;
1814 	int err = 0;
1815 
1816 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1817 		if (flow_is_multicast_only(flow_attr) &&
1818 		    !dont_trap)
1819 			priority = MLX5_IB_FLOW_MCAST_PRIO;
1820 		else
1821 			priority = ib_prio_to_core_prio(flow_attr->priority,
1822 							dont_trap);
1823 		ns = mlx5_get_flow_namespace(dev->mdev,
1824 					     MLX5_FLOW_NAMESPACE_BYPASS);
1825 		num_entries = MLX5_FS_MAX_ENTRIES;
1826 		num_groups = MLX5_FS_MAX_TYPES;
1827 		prio = &dev->flow_db.prios[priority];
1828 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1829 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1830 		ns = mlx5_get_flow_namespace(dev->mdev,
1831 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
1832 		build_leftovers_ft_param("bypass", &priority,
1833 					 &num_entries,
1834 					 &num_groups);
1835 		prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
1836 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
1837 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
1838 					allow_sniffer_and_nic_rx_shared_tir))
1839 			return ERR_PTR(-ENOTSUPP);
1840 
1841 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
1842 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
1843 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
1844 
1845 		prio = &dev->flow_db.sniffer[ft_type];
1846 		priority = 0;
1847 		num_entries = 1;
1848 		num_groups = 1;
1849 	}
1850 
1851 	if (!ns)
1852 		return ERR_PTR(-ENOTSUPP);
1853 
1854 	ft = prio->flow_table;
1855 	if (!ft) {
1856 		ft = mlx5_create_auto_grouped_flow_table(ns, priority, "bypass",
1857 							 num_entries,
1858 							 num_groups);
1859 
1860 		if (!IS_ERR(ft)) {
1861 			prio->refcount = 0;
1862 			prio->flow_table = ft;
1863 		} else {
1864 			err = PTR_ERR(ft);
1865 		}
1866 	}
1867 
1868 	return err ? ERR_PTR(err) : prio;
1869 }
1870 
1871 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
1872 						     struct mlx5_ib_flow_prio *ft_prio,
1873 						     const struct ib_flow_attr *flow_attr,
1874 						     struct mlx5_flow_destination *dst)
1875 {
1876 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
1877 	struct mlx5_ib_flow_handler *handler;
1878 	struct mlx5_flow_spec *spec;
1879 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
1880 	unsigned int spec_index;
1881 	u32 action;
1882 	int err = 0;
1883 
1884 	if (!is_valid_attr(flow_attr))
1885 		return ERR_PTR(-EINVAL);
1886 
1887 	spec = mlx5_vzalloc(sizeof(*spec));
1888 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
1889 	if (!handler || !spec) {
1890 		err = -ENOMEM;
1891 		goto free;
1892 	}
1893 
1894 	INIT_LIST_HEAD(&handler->list);
1895 
1896 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1897 		err = parse_flow_attr(spec->match_criteria,
1898 				      spec->match_value, ib_flow);
1899 		if (err < 0)
1900 			goto free;
1901 
1902 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
1903 	}
1904 
1905 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
1906 	action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
1907 		MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
1908 	handler->rule = mlx5_add_flow_rule(ft, spec->match_criteria_enable,
1909 					   spec->match_criteria,
1910 					   spec->match_value,
1911 					   action,
1912 					   MLX5_FS_DEFAULT_FLOW_TAG,
1913 					   dst);
1914 
1915 	if (IS_ERR(handler->rule)) {
1916 		err = PTR_ERR(handler->rule);
1917 		goto free;
1918 	}
1919 
1920 	ft_prio->refcount++;
1921 	handler->prio = ft_prio;
1922 
1923 	ft_prio->flow_table = ft;
1924 free:
1925 	if (err)
1926 		kfree(handler);
1927 	kvfree(spec);
1928 	return err ? ERR_PTR(err) : handler;
1929 }
1930 
1931 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
1932 							  struct mlx5_ib_flow_prio *ft_prio,
1933 							  struct ib_flow_attr *flow_attr,
1934 							  struct mlx5_flow_destination *dst)
1935 {
1936 	struct mlx5_ib_flow_handler *handler_dst = NULL;
1937 	struct mlx5_ib_flow_handler *handler = NULL;
1938 
1939 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
1940 	if (!IS_ERR(handler)) {
1941 		handler_dst = create_flow_rule(dev, ft_prio,
1942 					       flow_attr, dst);
1943 		if (IS_ERR(handler_dst)) {
1944 			mlx5_del_flow_rule(handler->rule);
1945 			ft_prio->refcount--;
1946 			kfree(handler);
1947 			handler = handler_dst;
1948 		} else {
1949 			list_add(&handler_dst->list, &handler->list);
1950 		}
1951 	}
1952 
1953 	return handler;
1954 }
1955 enum {
1956 	LEFTOVERS_MC,
1957 	LEFTOVERS_UC,
1958 };
1959 
1960 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
1961 							  struct mlx5_ib_flow_prio *ft_prio,
1962 							  struct ib_flow_attr *flow_attr,
1963 							  struct mlx5_flow_destination *dst)
1964 {
1965 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
1966 	struct mlx5_ib_flow_handler *handler = NULL;
1967 
1968 	static struct {
1969 		struct ib_flow_attr	flow_attr;
1970 		struct ib_flow_spec_eth eth_flow;
1971 	} leftovers_specs[] = {
1972 		[LEFTOVERS_MC] = {
1973 			.flow_attr = {
1974 				.num_of_specs = 1,
1975 				.size = sizeof(leftovers_specs[0])
1976 			},
1977 			.eth_flow = {
1978 				.type = IB_FLOW_SPEC_ETH,
1979 				.size = sizeof(struct ib_flow_spec_eth),
1980 				.mask = {.dst_mac = {0x1} },
1981 				.val =  {.dst_mac = {0x1} }
1982 			}
1983 		},
1984 		[LEFTOVERS_UC] = {
1985 			.flow_attr = {
1986 				.num_of_specs = 1,
1987 				.size = sizeof(leftovers_specs[0])
1988 			},
1989 			.eth_flow = {
1990 				.type = IB_FLOW_SPEC_ETH,
1991 				.size = sizeof(struct ib_flow_spec_eth),
1992 				.mask = {.dst_mac = {0x1} },
1993 				.val = {.dst_mac = {} }
1994 			}
1995 		}
1996 	};
1997 
1998 	handler = create_flow_rule(dev, ft_prio,
1999 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
2000 				   dst);
2001 	if (!IS_ERR(handler) &&
2002 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2003 		handler_ucast = create_flow_rule(dev, ft_prio,
2004 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
2005 						 dst);
2006 		if (IS_ERR(handler_ucast)) {
2007 			mlx5_del_flow_rule(handler->rule);
2008 			ft_prio->refcount--;
2009 			kfree(handler);
2010 			handler = handler_ucast;
2011 		} else {
2012 			list_add(&handler_ucast->list, &handler->list);
2013 		}
2014 	}
2015 
2016 	return handler;
2017 }
2018 
2019 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2020 							struct mlx5_ib_flow_prio *ft_rx,
2021 							struct mlx5_ib_flow_prio *ft_tx,
2022 							struct mlx5_flow_destination *dst)
2023 {
2024 	struct mlx5_ib_flow_handler *handler_rx;
2025 	struct mlx5_ib_flow_handler *handler_tx;
2026 	int err;
2027 	static const struct ib_flow_attr flow_attr  = {
2028 		.num_of_specs = 0,
2029 		.size = sizeof(flow_attr)
2030 	};
2031 
2032 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2033 	if (IS_ERR(handler_rx)) {
2034 		err = PTR_ERR(handler_rx);
2035 		goto err;
2036 	}
2037 
2038 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2039 	if (IS_ERR(handler_tx)) {
2040 		err = PTR_ERR(handler_tx);
2041 		goto err_tx;
2042 	}
2043 
2044 	list_add(&handler_tx->list, &handler_rx->list);
2045 
2046 	return handler_rx;
2047 
2048 err_tx:
2049 	mlx5_del_flow_rule(handler_rx->rule);
2050 	ft_rx->refcount--;
2051 	kfree(handler_rx);
2052 err:
2053 	return ERR_PTR(err);
2054 }
2055 
2056 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2057 					   struct ib_flow_attr *flow_attr,
2058 					   int domain)
2059 {
2060 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2061 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2062 	struct mlx5_ib_flow_handler *handler = NULL;
2063 	struct mlx5_flow_destination *dst = NULL;
2064 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2065 	struct mlx5_ib_flow_prio *ft_prio;
2066 	int err;
2067 
2068 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2069 		return ERR_PTR(-ENOSPC);
2070 
2071 	if (domain != IB_FLOW_DOMAIN_USER ||
2072 	    flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2073 	    (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2074 		return ERR_PTR(-EINVAL);
2075 
2076 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2077 	if (!dst)
2078 		return ERR_PTR(-ENOMEM);
2079 
2080 	mutex_lock(&dev->flow_db.lock);
2081 
2082 	ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2083 	if (IS_ERR(ft_prio)) {
2084 		err = PTR_ERR(ft_prio);
2085 		goto unlock;
2086 	}
2087 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2088 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2089 		if (IS_ERR(ft_prio_tx)) {
2090 			err = PTR_ERR(ft_prio_tx);
2091 			ft_prio_tx = NULL;
2092 			goto destroy_ft;
2093 		}
2094 	}
2095 
2096 	dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2097 	if (mqp->flags & MLX5_IB_QP_RSS)
2098 		dst->tir_num = mqp->rss_qp.tirn;
2099 	else
2100 		dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2101 
2102 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2103 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
2104 			handler = create_dont_trap_rule(dev, ft_prio,
2105 							flow_attr, dst);
2106 		} else {
2107 			handler = create_flow_rule(dev, ft_prio, flow_attr,
2108 						   dst);
2109 		}
2110 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2111 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2112 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2113 						dst);
2114 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2115 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2116 	} else {
2117 		err = -EINVAL;
2118 		goto destroy_ft;
2119 	}
2120 
2121 	if (IS_ERR(handler)) {
2122 		err = PTR_ERR(handler);
2123 		handler = NULL;
2124 		goto destroy_ft;
2125 	}
2126 
2127 	mutex_unlock(&dev->flow_db.lock);
2128 	kfree(dst);
2129 
2130 	return &handler->ibflow;
2131 
2132 destroy_ft:
2133 	put_flow_table(dev, ft_prio, false);
2134 	if (ft_prio_tx)
2135 		put_flow_table(dev, ft_prio_tx, false);
2136 unlock:
2137 	mutex_unlock(&dev->flow_db.lock);
2138 	kfree(dst);
2139 	kfree(handler);
2140 	return ERR_PTR(err);
2141 }
2142 
2143 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2144 {
2145 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2146 	int err;
2147 
2148 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2149 	if (err)
2150 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2151 			     ibqp->qp_num, gid->raw);
2152 
2153 	return err;
2154 }
2155 
2156 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2157 {
2158 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2159 	int err;
2160 
2161 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2162 	if (err)
2163 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2164 			     ibqp->qp_num, gid->raw);
2165 
2166 	return err;
2167 }
2168 
2169 static int init_node_data(struct mlx5_ib_dev *dev)
2170 {
2171 	int err;
2172 
2173 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2174 	if (err)
2175 		return err;
2176 
2177 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2178 }
2179 
2180 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2181 			     char *buf)
2182 {
2183 	struct mlx5_ib_dev *dev =
2184 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2185 
2186 	return sprintf(buf, "%lld\n", (long long)dev->mdev->priv.fw_pages);
2187 }
2188 
2189 static ssize_t show_reg_pages(struct device *device,
2190 			      struct device_attribute *attr, char *buf)
2191 {
2192 	struct mlx5_ib_dev *dev =
2193 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2194 
2195 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2196 }
2197 
2198 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2199 			char *buf)
2200 {
2201 	struct mlx5_ib_dev *dev =
2202 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2203 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2204 }
2205 
2206 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2207 			char *buf)
2208 {
2209 	struct mlx5_ib_dev *dev =
2210 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2211 	return sprintf(buf, "%x\n", dev->mdev->pdev->revision);
2212 }
2213 
2214 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2215 			  char *buf)
2216 {
2217 	struct mlx5_ib_dev *dev =
2218 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2219 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2220 		       dev->mdev->board_id);
2221 }
2222 
2223 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2224 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2225 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2226 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2227 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2228 
2229 static struct device_attribute *mlx5_class_attributes[] = {
2230 	&dev_attr_hw_rev,
2231 	&dev_attr_hca_type,
2232 	&dev_attr_board_id,
2233 	&dev_attr_fw_pages,
2234 	&dev_attr_reg_pages,
2235 };
2236 
2237 static void pkey_change_handler(struct work_struct *work)
2238 {
2239 	struct mlx5_ib_port_resources *ports =
2240 		container_of(work, struct mlx5_ib_port_resources,
2241 			     pkey_change_work);
2242 
2243 	mutex_lock(&ports->devr->mutex);
2244 	mlx5_ib_gsi_pkey_change(ports->gsi);
2245 	mutex_unlock(&ports->devr->mutex);
2246 }
2247 
2248 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2249 {
2250 	struct mlx5_ib_qp *mqp;
2251 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2252 	struct mlx5_core_cq *mcq;
2253 	struct list_head cq_armed_list;
2254 	unsigned long flags_qp;
2255 	unsigned long flags_cq;
2256 	unsigned long flags;
2257 
2258 	INIT_LIST_HEAD(&cq_armed_list);
2259 
2260 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2261 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2262 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2263 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2264 		if (mqp->sq.tail != mqp->sq.head) {
2265 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2266 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2267 			if (send_mcq->mcq.comp &&
2268 			    mqp->ibqp.send_cq->comp_handler) {
2269 				if (!send_mcq->mcq.reset_notify_added) {
2270 					send_mcq->mcq.reset_notify_added = 1;
2271 					list_add_tail(&send_mcq->mcq.reset_notify,
2272 						      &cq_armed_list);
2273 				}
2274 			}
2275 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2276 		}
2277 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2278 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2279 		/* no handling is needed for SRQ */
2280 		if (!mqp->ibqp.srq) {
2281 			if (mqp->rq.tail != mqp->rq.head) {
2282 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2283 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2284 				if (recv_mcq->mcq.comp &&
2285 				    mqp->ibqp.recv_cq->comp_handler) {
2286 					if (!recv_mcq->mcq.reset_notify_added) {
2287 						recv_mcq->mcq.reset_notify_added = 1;
2288 						list_add_tail(&recv_mcq->mcq.reset_notify,
2289 							      &cq_armed_list);
2290 					}
2291 				}
2292 				spin_unlock_irqrestore(&recv_mcq->lock,
2293 						       flags_cq);
2294 			}
2295 		}
2296 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2297 	}
2298 	/*At that point all inflight post send were put to be executed as of we
2299 	 * lock/unlock above locks Now need to arm all involved CQs.
2300 	 */
2301 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2302 		mcq->comp(mcq);
2303 	}
2304 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2305 }
2306 
2307 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2308 			  enum mlx5_dev_event event, unsigned long param)
2309 {
2310 	struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2311 	struct ib_event ibev;
2312 	bool fatal = false;
2313 	u8 port = 0;
2314 
2315 	switch (event) {
2316 	case MLX5_DEV_EVENT_SYS_ERROR:
2317 		ibev.event = IB_EVENT_DEVICE_FATAL;
2318 		mlx5_ib_handle_internal_error(ibdev);
2319 		fatal = true;
2320 		break;
2321 
2322 	case MLX5_DEV_EVENT_PORT_UP:
2323 	case MLX5_DEV_EVENT_PORT_DOWN:
2324 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
2325 		port = (u8)param;
2326 
2327 		/* In RoCE, port up/down events are handled in
2328 		 * mlx5_netdev_event().
2329 		 */
2330 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2331 			IB_LINK_LAYER_ETHERNET)
2332 			return;
2333 
2334 		ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2335 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2336 		break;
2337 
2338 	case MLX5_DEV_EVENT_LID_CHANGE:
2339 		ibev.event = IB_EVENT_LID_CHANGE;
2340 		port = (u8)param;
2341 		break;
2342 
2343 	case MLX5_DEV_EVENT_PKEY_CHANGE:
2344 		ibev.event = IB_EVENT_PKEY_CHANGE;
2345 		port = (u8)param;
2346 
2347 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2348 		break;
2349 
2350 	case MLX5_DEV_EVENT_GUID_CHANGE:
2351 		ibev.event = IB_EVENT_GID_CHANGE;
2352 		port = (u8)param;
2353 		break;
2354 
2355 	case MLX5_DEV_EVENT_CLIENT_REREG:
2356 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
2357 		port = (u8)param;
2358 		break;
2359 
2360 	default:
2361 		break;
2362 	}
2363 
2364 	ibev.device	      = &ibdev->ib_dev;
2365 	ibev.element.port_num = port;
2366 
2367 	if (port < 1 || port > ibdev->num_ports) {
2368 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2369 		return;
2370 	}
2371 
2372 	if (ibdev->ib_active)
2373 		ib_dispatch_event(&ibev);
2374 
2375 	if (fatal)
2376 		ibdev->ib_active = false;
2377 }
2378 
2379 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2380 {
2381 	int port;
2382 
2383 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2384 		mlx5_query_ext_port_caps(dev, port);
2385 }
2386 
2387 static int get_port_caps(struct mlx5_ib_dev *dev)
2388 {
2389 	struct ib_device_attr *dprops = NULL;
2390 	struct ib_port_attr *pprops = NULL;
2391 	int err = -ENOMEM;
2392 	int port;
2393 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2394 
2395 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2396 	if (!pprops)
2397 		goto out;
2398 
2399 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2400 	if (!dprops)
2401 		goto out;
2402 
2403 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2404 	if (err) {
2405 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2406 		goto out;
2407 	}
2408 
2409 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2410 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2411 		if (err) {
2412 			mlx5_ib_warn(dev, "query_port %d failed %d\n",
2413 				     port, err);
2414 			break;
2415 		}
2416 		dev->mdev->port_caps[port - 1].pkey_table_len =
2417 						dprops->max_pkeys;
2418 		dev->mdev->port_caps[port - 1].gid_table_len =
2419 						pprops->gid_tbl_len;
2420 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2421 			    dprops->max_pkeys, pprops->gid_tbl_len);
2422 	}
2423 
2424 out:
2425 	kfree(pprops);
2426 	kfree(dprops);
2427 
2428 	return err;
2429 }
2430 
2431 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2432 {
2433 	int err;
2434 
2435 	err = mlx5_mr_cache_cleanup(dev);
2436 	if (err)
2437 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2438 
2439 	mlx5_ib_destroy_qp(dev->umrc.qp);
2440 	ib_free_cq(dev->umrc.cq);
2441 	ib_dealloc_pd(dev->umrc.pd);
2442 }
2443 
2444 enum {
2445 	MAX_UMR_WR = 128,
2446 };
2447 
2448 static int create_umr_res(struct mlx5_ib_dev *dev)
2449 {
2450 	struct ib_qp_init_attr *init_attr = NULL;
2451 	struct ib_qp_attr *attr = NULL;
2452 	struct ib_pd *pd;
2453 	struct ib_cq *cq;
2454 	struct ib_qp *qp;
2455 	int ret;
2456 
2457 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2458 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2459 	if (!attr || !init_attr) {
2460 		ret = -ENOMEM;
2461 		goto error_0;
2462 	}
2463 
2464 	pd = ib_alloc_pd(&dev->ib_dev, 0);
2465 	if (IS_ERR(pd)) {
2466 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2467 		ret = PTR_ERR(pd);
2468 		goto error_0;
2469 	}
2470 
2471 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2472 	if (IS_ERR(cq)) {
2473 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2474 		ret = PTR_ERR(cq);
2475 		goto error_2;
2476 	}
2477 
2478 	init_attr->send_cq = cq;
2479 	init_attr->recv_cq = cq;
2480 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2481 	init_attr->cap.max_send_wr = MAX_UMR_WR;
2482 	init_attr->cap.max_send_sge = 1;
2483 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2484 	init_attr->port_num = 1;
2485 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2486 	if (IS_ERR(qp)) {
2487 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2488 		ret = PTR_ERR(qp);
2489 		goto error_3;
2490 	}
2491 	qp->device     = &dev->ib_dev;
2492 	qp->real_qp    = qp;
2493 	qp->uobject    = NULL;
2494 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2495 
2496 	attr->qp_state = IB_QPS_INIT;
2497 	attr->port_num = 1;
2498 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2499 				IB_QP_PORT, NULL);
2500 	if (ret) {
2501 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2502 		goto error_4;
2503 	}
2504 
2505 	memset(attr, 0, sizeof(*attr));
2506 	attr->qp_state = IB_QPS_RTR;
2507 	attr->path_mtu = IB_MTU_256;
2508 
2509 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2510 	if (ret) {
2511 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2512 		goto error_4;
2513 	}
2514 
2515 	memset(attr, 0, sizeof(*attr));
2516 	attr->qp_state = IB_QPS_RTS;
2517 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2518 	if (ret) {
2519 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2520 		goto error_4;
2521 	}
2522 
2523 	dev->umrc.qp = qp;
2524 	dev->umrc.cq = cq;
2525 	dev->umrc.pd = pd;
2526 
2527 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
2528 	ret = mlx5_mr_cache_init(dev);
2529 	if (ret) {
2530 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2531 		goto error_4;
2532 	}
2533 
2534 	kfree(attr);
2535 	kfree(init_attr);
2536 
2537 	return 0;
2538 
2539 error_4:
2540 	mlx5_ib_destroy_qp(qp);
2541 
2542 error_3:
2543 	ib_free_cq(cq);
2544 
2545 error_2:
2546 	ib_dealloc_pd(pd);
2547 
2548 error_0:
2549 	kfree(attr);
2550 	kfree(init_attr);
2551 	return ret;
2552 }
2553 
2554 static int create_dev_resources(struct mlx5_ib_resources *devr)
2555 {
2556 	struct ib_srq_init_attr attr;
2557 	struct mlx5_ib_dev *dev;
2558 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2559 	int port;
2560 	int ret = 0;
2561 
2562 	dev = container_of(devr, struct mlx5_ib_dev, devr);
2563 
2564 	mutex_init(&devr->mutex);
2565 
2566 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2567 	if (IS_ERR(devr->p0)) {
2568 		ret = PTR_ERR(devr->p0);
2569 		goto error0;
2570 	}
2571 	devr->p0->device  = &dev->ib_dev;
2572 	devr->p0->uobject = NULL;
2573 	atomic_set(&devr->p0->usecnt, 0);
2574 
2575 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
2576 	if (IS_ERR(devr->c0)) {
2577 		ret = PTR_ERR(devr->c0);
2578 		goto error1;
2579 	}
2580 	devr->c0->device        = &dev->ib_dev;
2581 	devr->c0->uobject       = NULL;
2582 	devr->c0->comp_handler  = NULL;
2583 	devr->c0->event_handler = NULL;
2584 	devr->c0->cq_context    = NULL;
2585 	atomic_set(&devr->c0->usecnt, 0);
2586 
2587 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2588 	if (IS_ERR(devr->x0)) {
2589 		ret = PTR_ERR(devr->x0);
2590 		goto error2;
2591 	}
2592 	devr->x0->device = &dev->ib_dev;
2593 	devr->x0->inode = NULL;
2594 	atomic_set(&devr->x0->usecnt, 0);
2595 	mutex_init(&devr->x0->tgt_qp_mutex);
2596 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2597 
2598 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2599 	if (IS_ERR(devr->x1)) {
2600 		ret = PTR_ERR(devr->x1);
2601 		goto error3;
2602 	}
2603 	devr->x1->device = &dev->ib_dev;
2604 	devr->x1->inode = NULL;
2605 	atomic_set(&devr->x1->usecnt, 0);
2606 	mutex_init(&devr->x1->tgt_qp_mutex);
2607 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2608 
2609 	memset(&attr, 0, sizeof(attr));
2610 	attr.attr.max_sge = 1;
2611 	attr.attr.max_wr = 1;
2612 	attr.srq_type = IB_SRQT_XRC;
2613 	attr.ext.xrc.cq = devr->c0;
2614 	attr.ext.xrc.xrcd = devr->x0;
2615 
2616 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2617 	if (IS_ERR(devr->s0)) {
2618 		ret = PTR_ERR(devr->s0);
2619 		goto error4;
2620 	}
2621 	devr->s0->device	= &dev->ib_dev;
2622 	devr->s0->pd		= devr->p0;
2623 	devr->s0->uobject       = NULL;
2624 	devr->s0->event_handler = NULL;
2625 	devr->s0->srq_context   = NULL;
2626 	devr->s0->srq_type      = IB_SRQT_XRC;
2627 	devr->s0->ext.xrc.xrcd	= devr->x0;
2628 	devr->s0->ext.xrc.cq	= devr->c0;
2629 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2630 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2631 	atomic_inc(&devr->p0->usecnt);
2632 	atomic_set(&devr->s0->usecnt, 0);
2633 
2634 	memset(&attr, 0, sizeof(attr));
2635 	attr.attr.max_sge = 1;
2636 	attr.attr.max_wr = 1;
2637 	attr.srq_type = IB_SRQT_BASIC;
2638 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2639 	if (IS_ERR(devr->s1)) {
2640 		ret = PTR_ERR(devr->s1);
2641 		goto error5;
2642 	}
2643 	devr->s1->device	= &dev->ib_dev;
2644 	devr->s1->pd		= devr->p0;
2645 	devr->s1->uobject       = NULL;
2646 	devr->s1->event_handler = NULL;
2647 	devr->s1->srq_context   = NULL;
2648 	devr->s1->srq_type      = IB_SRQT_BASIC;
2649 	devr->s1->ext.xrc.cq	= devr->c0;
2650 	atomic_inc(&devr->p0->usecnt);
2651 	atomic_set(&devr->s0->usecnt, 0);
2652 
2653 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2654 		INIT_WORK(&devr->ports[port].pkey_change_work,
2655 			  pkey_change_handler);
2656 		devr->ports[port].devr = devr;
2657 	}
2658 
2659 	return 0;
2660 
2661 error5:
2662 	mlx5_ib_destroy_srq(devr->s0);
2663 error4:
2664 	mlx5_ib_dealloc_xrcd(devr->x1);
2665 error3:
2666 	mlx5_ib_dealloc_xrcd(devr->x0);
2667 error2:
2668 	mlx5_ib_destroy_cq(devr->c0);
2669 error1:
2670 	mlx5_ib_dealloc_pd(devr->p0);
2671 error0:
2672 	return ret;
2673 }
2674 
2675 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2676 {
2677 	struct mlx5_ib_dev *dev =
2678 		container_of(devr, struct mlx5_ib_dev, devr);
2679 	int port;
2680 
2681 	mlx5_ib_destroy_srq(devr->s1);
2682 	mlx5_ib_destroy_srq(devr->s0);
2683 	mlx5_ib_dealloc_xrcd(devr->x0);
2684 	mlx5_ib_dealloc_xrcd(devr->x1);
2685 	mlx5_ib_destroy_cq(devr->c0);
2686 	mlx5_ib_dealloc_pd(devr->p0);
2687 
2688 	/* Make sure no change P_Key work items are still executing */
2689 	for (port = 0; port < dev->num_ports; ++port)
2690 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2691 }
2692 
2693 static u32 get_core_cap_flags(struct ib_device *ibdev)
2694 {
2695 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2696 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2697 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2698 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2699 	u32 ret = 0;
2700 
2701 	if (ll == IB_LINK_LAYER_INFINIBAND)
2702 		return RDMA_CORE_PORT_IBA_IB;
2703 
2704 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2705 		return 0;
2706 
2707 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2708 		return 0;
2709 
2710 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2711 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2712 
2713 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2714 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2715 
2716 	return ret;
2717 }
2718 
2719 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
2720 			       struct ib_port_immutable *immutable)
2721 {
2722 	struct ib_port_attr attr;
2723 	int err;
2724 
2725 	err = mlx5_ib_query_port(ibdev, port_num, &attr);
2726 	if (err)
2727 		return err;
2728 
2729 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2730 	immutable->gid_tbl_len = attr.gid_tbl_len;
2731 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
2732 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2733 
2734 	return 0;
2735 }
2736 
2737 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
2738 			   size_t str_len)
2739 {
2740 	struct mlx5_ib_dev *dev =
2741 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2742 	snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
2743 		       fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
2744 }
2745 
2746 static int mlx5_roce_lag_init(struct mlx5_ib_dev *dev)
2747 {
2748 	return 0;
2749 }
2750 
2751 static void mlx5_roce_lag_cleanup(struct mlx5_ib_dev *dev)
2752 {
2753 }
2754 
2755 static void mlx5_remove_roce_notifier(struct mlx5_ib_dev *dev)
2756 {
2757 	if (dev->roce.nb.notifier_call) {
2758 		unregister_netdevice_notifier(&dev->roce.nb);
2759 		dev->roce.nb.notifier_call = NULL;
2760 	}
2761 }
2762 
2763 static int mlx5_enable_roce(struct mlx5_ib_dev *dev)
2764 {
2765 	VNET_ITERATOR_DECL(vnet_iter);
2766 	struct net_device *idev;
2767 	int err;
2768 
2769 	/* Check if mlx5en net device already exists */
2770 	VNET_LIST_RLOCK();
2771 	VNET_FOREACH(vnet_iter) {
2772 		IFNET_RLOCK();
2773 		CURVNET_SET_QUIET(vnet_iter);
2774 		CK_STAILQ_FOREACH(idev, &V_ifnet, if_link) {
2775 			/* check if network interface belongs to mlx5en */
2776 			if (!mlx5_netdev_match(idev, dev->mdev, "mce"))
2777 				continue;
2778 			write_lock(&dev->roce.netdev_lock);
2779 			dev->roce.netdev = idev;
2780 			write_unlock(&dev->roce.netdev_lock);
2781 		}
2782 		CURVNET_RESTORE();
2783 		IFNET_RUNLOCK();
2784 	}
2785 	VNET_LIST_RUNLOCK();
2786 
2787 	dev->roce.nb.notifier_call = mlx5_netdev_event;
2788 	err = register_netdevice_notifier(&dev->roce.nb);
2789 	if (err) {
2790 		dev->roce.nb.notifier_call = NULL;
2791 		return err;
2792 	}
2793 
2794 	err = mlx5_nic_vport_enable_roce(dev->mdev);
2795 	if (err)
2796 		goto err_unregister_netdevice_notifier;
2797 
2798 	err = mlx5_roce_lag_init(dev);
2799 	if (err)
2800 		goto err_disable_roce;
2801 
2802 	return 0;
2803 
2804 err_disable_roce:
2805 	mlx5_nic_vport_disable_roce(dev->mdev);
2806 
2807 err_unregister_netdevice_notifier:
2808 	mlx5_remove_roce_notifier(dev);
2809 	return err;
2810 }
2811 
2812 static void mlx5_disable_roce(struct mlx5_ib_dev *dev)
2813 {
2814 	mlx5_roce_lag_cleanup(dev);
2815 	mlx5_nic_vport_disable_roce(dev->mdev);
2816 }
2817 
2818 static void mlx5_ib_dealloc_q_port_counter(struct mlx5_ib_dev *dev, u8 port_num)
2819 {
2820 	mlx5_vport_dealloc_q_counter(dev->mdev,
2821 				     MLX5_INTERFACE_PROTOCOL_IB,
2822 				     dev->port[port_num].q_cnt_id);
2823 	dev->port[port_num].q_cnt_id = 0;
2824 }
2825 
2826 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
2827 {
2828 	unsigned int i;
2829 
2830 	for (i = 0; i < dev->num_ports; i++)
2831 		mlx5_ib_dealloc_q_port_counter(dev, i);
2832 }
2833 
2834 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
2835 {
2836 	int i;
2837 	int ret;
2838 
2839 	for (i = 0; i < dev->num_ports; i++) {
2840 		ret = mlx5_vport_alloc_q_counter(dev->mdev,
2841 						 MLX5_INTERFACE_PROTOCOL_IB,
2842 						 &dev->port[i].q_cnt_id);
2843 		if (ret) {
2844 			mlx5_ib_warn(dev,
2845 				     "couldn't allocate queue counter for port %d, err %d\n",
2846 				     i + 1, ret);
2847 			goto dealloc_counters;
2848 		}
2849 	}
2850 
2851 	return 0;
2852 
2853 dealloc_counters:
2854 	while (--i >= 0)
2855 		mlx5_ib_dealloc_q_port_counter(dev, i);
2856 
2857 	return ret;
2858 }
2859 
2860 static const char * const names[] = {
2861 	"rx_write_requests",
2862 	"rx_read_requests",
2863 	"rx_atomic_requests",
2864 	"out_of_buffer",
2865 	"out_of_sequence",
2866 	"duplicate_request",
2867 	"rnr_nak_retry_err",
2868 	"packet_seq_err",
2869 	"implied_nak_seq_err",
2870 	"local_ack_timeout_err",
2871 };
2872 
2873 static const size_t stats_offsets[] = {
2874 	MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests),
2875 	MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests),
2876 	MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests),
2877 	MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer),
2878 	MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence),
2879 	MLX5_BYTE_OFF(query_q_counter_out, duplicate_request),
2880 	MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err),
2881 	MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err),
2882 	MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err),
2883 	MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err),
2884 };
2885 
2886 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
2887 						    u8 port_num)
2888 {
2889 	BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets));
2890 
2891 	/* We support only per port stats */
2892 	if (port_num == 0)
2893 		return NULL;
2894 
2895 	return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names),
2896 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
2897 }
2898 
2899 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
2900 				struct rdma_hw_stats *stats,
2901 				u8 port, int index)
2902 {
2903 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2904 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
2905 	void *out;
2906 	__be32 val;
2907 	int ret;
2908 	int i;
2909 
2910 	if (!port || !stats)
2911 		return -ENOSYS;
2912 
2913 	out = mlx5_vzalloc(outlen);
2914 	if (!out)
2915 		return -ENOMEM;
2916 
2917 	ret = mlx5_vport_query_q_counter(dev->mdev,
2918 					dev->port[port - 1].q_cnt_id, 0,
2919 					out, outlen);
2920 	if (ret)
2921 		goto free;
2922 
2923 	for (i = 0; i < ARRAY_SIZE(names); i++) {
2924 		val = *(__be32 *)(out + stats_offsets[i]);
2925 		stats->value[i] = (u64)be32_to_cpu(val);
2926 	}
2927 free:
2928 	kvfree(out);
2929 	return ARRAY_SIZE(names);
2930 }
2931 
2932 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
2933 {
2934 	struct mlx5_ib_dev *dev;
2935 	enum rdma_link_layer ll;
2936 	int port_type_cap;
2937 	const char *name;
2938 	int err;
2939 	int i;
2940 
2941 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
2942 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
2943 
2944 	if ((ll == IB_LINK_LAYER_ETHERNET) && !MLX5_CAP_GEN(mdev, roce))
2945 		return NULL;
2946 
2947 	printk_once(KERN_INFO "%s", mlx5_version);
2948 
2949 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
2950 	if (!dev)
2951 		return NULL;
2952 
2953 	dev->mdev = mdev;
2954 
2955 	dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
2956 			    GFP_KERNEL);
2957 	if (!dev->port)
2958 		goto err_dealloc;
2959 
2960 	rwlock_init(&dev->roce.netdev_lock);
2961 	err = get_port_caps(dev);
2962 	if (err)
2963 		goto err_free_port;
2964 
2965 	if (mlx5_use_mad_ifc(dev))
2966 		get_ext_port_caps(dev);
2967 
2968 	MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
2969 
2970 	name = "mlx5_%d";
2971 
2972 	strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
2973 	dev->ib_dev.owner		= THIS_MODULE;
2974 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
2975 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
2976 	dev->num_ports		= MLX5_CAP_GEN(mdev, num_ports);
2977 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
2978 	dev->ib_dev.num_comp_vectors    =
2979 		dev->mdev->priv.eq_table.num_comp_vectors;
2980 	dev->ib_dev.dma_device	= &mdev->pdev->dev;
2981 
2982 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
2983 	dev->ib_dev.uverbs_cmd_mask	=
2984 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
2985 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
2986 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
2987 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
2988 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
2989 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
2990 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
2991 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
2992 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
2993 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
2994 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
2995 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
2996 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
2997 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
2998 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
2999 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
3000 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
3001 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
3002 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
3003 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
3004 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
3005 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
3006 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
3007 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
3008 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
3009 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
3010 	dev->ib_dev.uverbs_ex_cmd_mask =
3011 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
3012 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
3013 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
3014 
3015 	dev->ib_dev.query_device	= mlx5_ib_query_device;
3016 	dev->ib_dev.query_port		= mlx5_ib_query_port;
3017 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
3018 	if (ll == IB_LINK_LAYER_ETHERNET)
3019 		dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
3020 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
3021 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
3022 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
3023 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
3024 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
3025 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
3026 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
3027 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
3028 	dev->ib_dev.mmap		= mlx5_ib_mmap;
3029 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
3030 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
3031 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
3032 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
3033 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
3034 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
3035 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
3036 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
3037 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
3038 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
3039 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
3040 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
3041 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
3042 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
3043 	dev->ib_dev.post_send		= mlx5_ib_post_send;
3044 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
3045 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
3046 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
3047 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
3048 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
3049 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
3050 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
3051 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
3052 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
3053 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
3054 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
3055 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
3056 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
3057 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
3058 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
3059 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
3060 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
3061 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
3062 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
3063 	if (mlx5_core_is_pf(mdev)) {
3064 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
3065 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
3066 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
3067 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
3068 	}
3069 
3070 	mlx5_ib_internal_fill_odp_caps(dev);
3071 
3072 	if (MLX5_CAP_GEN(mdev, imaicl)) {
3073 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
3074 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
3075 		dev->ib_dev.uverbs_cmd_mask |=
3076 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
3077 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
3078 	}
3079 
3080 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) &&
3081 	    MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3082 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
3083 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
3084 	}
3085 
3086 	if (MLX5_CAP_GEN(mdev, xrc)) {
3087 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
3088 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
3089 		dev->ib_dev.uverbs_cmd_mask |=
3090 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
3091 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
3092 	}
3093 
3094 	if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
3095 	    IB_LINK_LAYER_ETHERNET) {
3096 		dev->ib_dev.create_flow	= mlx5_ib_create_flow;
3097 		dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
3098 		dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
3099 		dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
3100 		dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
3101 		dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
3102 		dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
3103 		dev->ib_dev.uverbs_ex_cmd_mask |=
3104 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
3105 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
3106 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
3107 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
3108 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
3109 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
3110 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
3111 	}
3112 	err = init_node_data(dev);
3113 	if (err)
3114 		goto err_free_port;
3115 
3116 	mutex_init(&dev->flow_db.lock);
3117 	mutex_init(&dev->cap_mask_mutex);
3118 	INIT_LIST_HEAD(&dev->qp_list);
3119 	spin_lock_init(&dev->reset_flow_resource_lock);
3120 
3121 	if (ll == IB_LINK_LAYER_ETHERNET) {
3122 		err = mlx5_enable_roce(dev);
3123 		if (err)
3124 			goto err_free_port;
3125 	}
3126 
3127 	err = create_dev_resources(&dev->devr);
3128 	if (err)
3129 		goto err_disable_roce;
3130 
3131 	err = mlx5_ib_odp_init_one(dev);
3132 	if (err)
3133 		goto err_rsrc;
3134 
3135 	err = mlx5_ib_alloc_q_counters(dev);
3136 	if (err)
3137 		goto err_odp;
3138 
3139 	err = ib_register_device(&dev->ib_dev, NULL);
3140 	if (err)
3141 		goto err_q_cnt;
3142 
3143 	err = create_umr_res(dev);
3144 	if (err)
3145 		goto err_dev;
3146 
3147 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
3148 		err = device_create_file(&dev->ib_dev.dev,
3149 					 mlx5_class_attributes[i]);
3150 		if (err)
3151 			goto err_umrc;
3152 	}
3153 
3154 	err = mlx5_ib_init_congestion(dev);
3155 	if (err)
3156 		goto err_umrc;
3157 
3158 	dev->ib_active = true;
3159 
3160 	return dev;
3161 
3162 err_umrc:
3163 	destroy_umrc_res(dev);
3164 
3165 err_dev:
3166 	ib_unregister_device(&dev->ib_dev);
3167 
3168 err_q_cnt:
3169 	mlx5_ib_dealloc_q_counters(dev);
3170 
3171 err_odp:
3172 	mlx5_ib_odp_remove_one(dev);
3173 
3174 err_rsrc:
3175 	destroy_dev_resources(&dev->devr);
3176 
3177 err_disable_roce:
3178 	if (ll == IB_LINK_LAYER_ETHERNET) {
3179 		mlx5_disable_roce(dev);
3180 		mlx5_remove_roce_notifier(dev);
3181 	}
3182 
3183 err_free_port:
3184 	kfree(dev->port);
3185 
3186 err_dealloc:
3187 	ib_dealloc_device((struct ib_device *)dev);
3188 
3189 	return NULL;
3190 }
3191 
3192 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3193 {
3194 	struct mlx5_ib_dev *dev = context;
3195 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3196 
3197 	mlx5_ib_cleanup_congestion(dev);
3198 	mlx5_remove_roce_notifier(dev);
3199 	ib_unregister_device(&dev->ib_dev);
3200 	mlx5_ib_dealloc_q_counters(dev);
3201 	destroy_umrc_res(dev);
3202 	mlx5_ib_odp_remove_one(dev);
3203 	destroy_dev_resources(&dev->devr);
3204 	if (ll == IB_LINK_LAYER_ETHERNET)
3205 		mlx5_disable_roce(dev);
3206 	kfree(dev->port);
3207 	ib_dealloc_device(&dev->ib_dev);
3208 }
3209 
3210 static struct mlx5_interface mlx5_ib_interface = {
3211 	.add            = mlx5_ib_add,
3212 	.remove         = mlx5_ib_remove,
3213 	.event          = mlx5_ib_event,
3214 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
3215 };
3216 
3217 static int __init mlx5_ib_init(void)
3218 {
3219 	int err;
3220 
3221 	if (deprecated_prof_sel != 2)
3222 		pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
3223 
3224 	err = mlx5_ib_odp_init();
3225 	if (err)
3226 		return err;
3227 
3228 	err = mlx5_register_interface(&mlx5_ib_interface);
3229 	if (err)
3230 		goto clean_odp;
3231 
3232 	return err;
3233 
3234 clean_odp:
3235 	mlx5_ib_odp_cleanup();
3236 	return err;
3237 }
3238 
3239 static void __exit mlx5_ib_cleanup(void)
3240 {
3241 	mlx5_unregister_interface(&mlx5_ib_interface);
3242 	mlx5_ib_odp_cleanup();
3243 }
3244 
3245 module_init_order(mlx5_ib_init, SI_ORDER_THIRD);
3246 module_exit_order(mlx5_ib_cleanup, SI_ORDER_THIRD);
3247