xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_main.c (revision d8a0fe102c0cfdfcd5b818f850eff09d8536c9bc)
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/vport.h>
44 #include <linux/list.h>
45 #include <rdma/ib_smi.h>
46 #include <rdma/ib_umem.h>
47 #include <linux/in.h>
48 #include <linux/etherdevice.h>
49 #include <dev/mlx5/fs.h>
50 #include "mlx5_ib.h"
51 
52 #define DRIVER_NAME "mlx5_ib"
53 #define DRIVER_VERSION "3.4.1-BETA"
54 #define DRIVER_RELDATE	"October 2017"
55 
56 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
57 MODULE_LICENSE("Dual BSD/GPL");
58 MODULE_DEPEND(mlx5ib, linuxkpi, 1, 1, 1);
59 MODULE_DEPEND(mlx5ib, mlx5, 1, 1, 1);
60 MODULE_DEPEND(mlx5ib, ibcore, 1, 1, 1);
61 MODULE_VERSION(mlx5ib, 1);
62 
63 static int deprecated_prof_sel = 2;
64 module_param_named(prof_sel, deprecated_prof_sel, int, 0444);
65 MODULE_PARM_DESC(prof_sel, "profile selector. Deprecated here. Moved to module mlx5_core");
66 
67 static char mlx5_version[] =
68 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
69 	DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
70 
71 enum {
72 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
73 };
74 
75 static enum rdma_link_layer
76 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
77 {
78 	switch (port_type_cap) {
79 	case MLX5_CAP_PORT_TYPE_IB:
80 		return IB_LINK_LAYER_INFINIBAND;
81 	case MLX5_CAP_PORT_TYPE_ETH:
82 		return IB_LINK_LAYER_ETHERNET;
83 	default:
84 		return IB_LINK_LAYER_UNSPECIFIED;
85 	}
86 }
87 
88 static enum rdma_link_layer
89 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
90 {
91 	struct mlx5_ib_dev *dev = to_mdev(device);
92 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
93 
94 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
95 }
96 
97 static bool mlx5_netdev_match(struct net_device *ndev,
98 			      struct mlx5_core_dev *mdev,
99 			      const char *dname)
100 {
101 	return ndev->if_type == IFT_ETHER &&
102 	  ndev->if_dname != NULL &&
103 	  strcmp(ndev->if_dname, dname) == 0 &&
104 	  ndev->if_softc != NULL &&
105 	  *(struct mlx5_core_dev **)ndev->if_softc == mdev;
106 }
107 
108 static int mlx5_netdev_event(struct notifier_block *this,
109 			     unsigned long event, void *ptr)
110 {
111 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
112 	struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
113 						 roce.nb);
114 
115 	switch (event) {
116 	case NETDEV_REGISTER:
117 	case NETDEV_UNREGISTER:
118 		write_lock(&ibdev->roce.netdev_lock);
119 		/* check if network interface belongs to mlx5en */
120 		if (mlx5_netdev_match(ndev, ibdev->mdev, "mce"))
121 			ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
122 					     NULL : ndev;
123 		write_unlock(&ibdev->roce.netdev_lock);
124 		break;
125 
126 	case NETDEV_UP:
127 	case NETDEV_DOWN: {
128 		struct net_device *upper = NULL;
129 
130 		if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
131 		    && ibdev->ib_active) {
132 			struct ib_event ibev = {0};
133 
134 			ibev.device = &ibdev->ib_dev;
135 			ibev.event = (event == NETDEV_UP) ?
136 				     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
137 			ibev.element.port_num = 1;
138 			ib_dispatch_event(&ibev);
139 		}
140 		break;
141 	}
142 
143 	default:
144 		break;
145 	}
146 
147 	return NOTIFY_DONE;
148 }
149 
150 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
151 					     u8 port_num)
152 {
153 	struct mlx5_ib_dev *ibdev = to_mdev(device);
154 	struct net_device *ndev;
155 
156 	/* Ensure ndev does not disappear before we invoke dev_hold()
157 	 */
158 	read_lock(&ibdev->roce.netdev_lock);
159 	ndev = ibdev->roce.netdev;
160 	if (ndev)
161 		dev_hold(ndev);
162 	read_unlock(&ibdev->roce.netdev_lock);
163 
164 	return ndev;
165 }
166 
167 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
168 				struct ib_port_attr *props)
169 {
170 	struct mlx5_ib_dev *dev = to_mdev(device);
171 	struct net_device *ndev;
172 	enum ib_mtu ndev_ib_mtu;
173 	u16 qkey_viol_cntr;
174 
175 	memset(props, 0, sizeof(*props));
176 
177 	props->port_cap_flags  |= IB_PORT_CM_SUP;
178 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
179 
180 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
181 						roce_address_table_size);
182 	props->max_mtu          = IB_MTU_4096;
183 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
184 	props->pkey_tbl_len     = 1;
185 	props->state            = IB_PORT_DOWN;
186 	props->phys_state       = 3;
187 
188 	mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
189 	props->qkey_viol_cntr = qkey_viol_cntr;
190 
191 	ndev = mlx5_ib_get_netdev(device, port_num);
192 	if (!ndev)
193 		return 0;
194 
195 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
196 		props->state      = IB_PORT_ACTIVE;
197 		props->phys_state = 5;
198 	}
199 
200 	ndev_ib_mtu = iboe_get_mtu(ndev->if_mtu);
201 
202 	dev_put(ndev);
203 
204 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
205 
206 	props->active_width	= IB_WIDTH_4X;  /* TODO */
207 	props->active_speed	= IB_SPEED_QDR; /* TODO */
208 
209 	return 0;
210 }
211 
212 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
213 				     const struct ib_gid_attr *attr,
214 				     void *mlx5_addr)
215 {
216 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
217 	char *mlx5_addr_l3_addr	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
218 					       source_l3_address);
219 	void *mlx5_addr_mac	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
220 					       source_mac_47_32);
221 
222 	if (!gid)
223 		return;
224 	ether_addr_copy(mlx5_addr_mac, IF_LLADDR(attr->ndev));
225 
226 	if (is_vlan_dev(attr->ndev)) {
227 		MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
228 		MLX5_SET_RA(mlx5_addr, vlan_id, vlan_dev_vlan_id(attr->ndev));
229 	}
230 
231 	switch (attr->gid_type) {
232 	case IB_GID_TYPE_IB:
233 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
234 		break;
235 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
236 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
237 		break;
238 
239 	default:
240 		WARN_ON(true);
241 	}
242 
243 	if (attr->gid_type != IB_GID_TYPE_IB) {
244 		if (ipv6_addr_v4mapped((void *)gid))
245 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
246 				    MLX5_ROCE_L3_TYPE_IPV4);
247 		else
248 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
249 				    MLX5_ROCE_L3_TYPE_IPV6);
250 	}
251 
252 	if ((attr->gid_type == IB_GID_TYPE_IB) ||
253 	    !ipv6_addr_v4mapped((void *)gid))
254 		memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
255 	else
256 		memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
257 }
258 
259 static int set_roce_addr(struct ib_device *device, u8 port_num,
260 			 unsigned int index,
261 			 const union ib_gid *gid,
262 			 const struct ib_gid_attr *attr)
263 {
264 	struct mlx5_ib_dev *dev = to_mdev(device);
265 	u32  in[MLX5_ST_SZ_DW(set_roce_address_in)]  = {0};
266 	u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0};
267 	void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
268 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
269 
270 	if (ll != IB_LINK_LAYER_ETHERNET)
271 		return -EINVAL;
272 
273 	ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
274 
275 	MLX5_SET(set_roce_address_in, in, roce_address_index, index);
276 	MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
277 	return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
278 }
279 
280 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
281 			   unsigned int index, const union ib_gid *gid,
282 			   const struct ib_gid_attr *attr,
283 			   __always_unused void **context)
284 {
285 	return set_roce_addr(device, port_num, index, gid, attr);
286 }
287 
288 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
289 			   unsigned int index, __always_unused void **context)
290 {
291 	return set_roce_addr(device, port_num, index, NULL, NULL);
292 }
293 
294 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
295 			       int index)
296 {
297 	struct ib_gid_attr attr;
298 	union ib_gid gid;
299 
300 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
301 		return 0;
302 
303 	if (!attr.ndev)
304 		return 0;
305 
306 	dev_put(attr.ndev);
307 
308 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
309 		return 0;
310 
311 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
312 }
313 
314 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
315 {
316 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
317 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
318 	return 0;
319 }
320 
321 enum {
322 	MLX5_VPORT_ACCESS_METHOD_MAD,
323 	MLX5_VPORT_ACCESS_METHOD_HCA,
324 	MLX5_VPORT_ACCESS_METHOD_NIC,
325 };
326 
327 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
328 {
329 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
330 		return MLX5_VPORT_ACCESS_METHOD_MAD;
331 
332 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
333 	    IB_LINK_LAYER_ETHERNET)
334 		return MLX5_VPORT_ACCESS_METHOD_NIC;
335 
336 	return MLX5_VPORT_ACCESS_METHOD_HCA;
337 }
338 
339 static void get_atomic_caps(struct mlx5_ib_dev *dev,
340 			    struct ib_device_attr *props)
341 {
342 	u8 tmp;
343 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
344 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
345 	u8 atomic_req_8B_endianness_mode =
346 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
347 
348 	/* Check if HW supports 8 bytes standard atomic operations and capable
349 	 * of host endianness respond
350 	 */
351 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
352 	if (((atomic_operations & tmp) == tmp) &&
353 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
354 	    (atomic_req_8B_endianness_mode)) {
355 		props->atomic_cap = IB_ATOMIC_HCA;
356 	} else {
357 		props->atomic_cap = IB_ATOMIC_NONE;
358 	}
359 }
360 
361 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
362 					__be64 *sys_image_guid)
363 {
364 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
365 	struct mlx5_core_dev *mdev = dev->mdev;
366 	u64 tmp;
367 	int err;
368 
369 	switch (mlx5_get_vport_access_method(ibdev)) {
370 	case MLX5_VPORT_ACCESS_METHOD_MAD:
371 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
372 							    sys_image_guid);
373 
374 	case MLX5_VPORT_ACCESS_METHOD_HCA:
375 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
376 		break;
377 
378 	case MLX5_VPORT_ACCESS_METHOD_NIC:
379 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
380 		break;
381 
382 	default:
383 		return -EINVAL;
384 	}
385 
386 	if (!err)
387 		*sys_image_guid = cpu_to_be64(tmp);
388 
389 	return err;
390 
391 }
392 
393 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
394 				u16 *max_pkeys)
395 {
396 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
397 	struct mlx5_core_dev *mdev = dev->mdev;
398 
399 	switch (mlx5_get_vport_access_method(ibdev)) {
400 	case MLX5_VPORT_ACCESS_METHOD_MAD:
401 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
402 
403 	case MLX5_VPORT_ACCESS_METHOD_HCA:
404 	case MLX5_VPORT_ACCESS_METHOD_NIC:
405 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
406 						pkey_table_size));
407 		return 0;
408 
409 	default:
410 		return -EINVAL;
411 	}
412 }
413 
414 static int mlx5_query_vendor_id(struct ib_device *ibdev,
415 				u32 *vendor_id)
416 {
417 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
418 
419 	switch (mlx5_get_vport_access_method(ibdev)) {
420 	case MLX5_VPORT_ACCESS_METHOD_MAD:
421 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
422 
423 	case MLX5_VPORT_ACCESS_METHOD_HCA:
424 	case MLX5_VPORT_ACCESS_METHOD_NIC:
425 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
426 
427 	default:
428 		return -EINVAL;
429 	}
430 }
431 
432 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
433 				__be64 *node_guid)
434 {
435 	u64 tmp;
436 	int err;
437 
438 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
439 	case MLX5_VPORT_ACCESS_METHOD_MAD:
440 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
441 
442 	case MLX5_VPORT_ACCESS_METHOD_HCA:
443 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
444 		break;
445 
446 	case MLX5_VPORT_ACCESS_METHOD_NIC:
447 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
448 		break;
449 
450 	default:
451 		return -EINVAL;
452 	}
453 
454 	if (!err)
455 		*node_guid = cpu_to_be64(tmp);
456 
457 	return err;
458 }
459 
460 struct mlx5_reg_node_desc {
461 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
462 };
463 
464 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
465 {
466 	struct mlx5_reg_node_desc in;
467 
468 	if (mlx5_use_mad_ifc(dev))
469 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
470 
471 	memset(&in, 0, sizeof(in));
472 
473 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
474 				    sizeof(struct mlx5_reg_node_desc),
475 				    MLX5_REG_NODE_DESC, 0, 0);
476 }
477 
478 static int mlx5_ib_query_device(struct ib_device *ibdev,
479 				struct ib_device_attr *props,
480 				struct ib_udata *uhw)
481 {
482 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
483 	struct mlx5_core_dev *mdev = dev->mdev;
484 	int err = -ENOMEM;
485 	int max_rq_sg;
486 	int max_sq_sg;
487 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
488 	struct mlx5_ib_query_device_resp resp = {};
489 	size_t resp_len;
490 	u64 max_tso;
491 
492 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
493 	if (uhw->outlen && uhw->outlen < resp_len)
494 		return -EINVAL;
495 	else
496 		resp.response_length = resp_len;
497 
498 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
499 		return -EINVAL;
500 
501 	memset(props, 0, sizeof(*props));
502 	err = mlx5_query_system_image_guid(ibdev,
503 					   &props->sys_image_guid);
504 	if (err)
505 		return err;
506 
507 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
508 	if (err)
509 		return err;
510 
511 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
512 	if (err)
513 		return err;
514 
515 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
516 		(fw_rev_min(dev->mdev) << 16) |
517 		fw_rev_sub(dev->mdev);
518 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
519 		IB_DEVICE_PORT_ACTIVE_EVENT		|
520 		IB_DEVICE_SYS_IMAGE_GUID		|
521 		IB_DEVICE_RC_RNR_NAK_GEN;
522 
523 	if (MLX5_CAP_GEN(mdev, pkv))
524 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
525 	if (MLX5_CAP_GEN(mdev, qkv))
526 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
527 	if (MLX5_CAP_GEN(mdev, apm))
528 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
529 	if (MLX5_CAP_GEN(mdev, xrc))
530 		props->device_cap_flags |= IB_DEVICE_XRC;
531 	if (MLX5_CAP_GEN(mdev, imaicl)) {
532 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
533 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
534 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
535 		/* We support 'Gappy' memory registration too */
536 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
537 	}
538 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
539 	if (MLX5_CAP_GEN(mdev, sho)) {
540 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
541 		/* At this stage no support for signature handover */
542 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
543 				      IB_PROT_T10DIF_TYPE_2 |
544 				      IB_PROT_T10DIF_TYPE_3;
545 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
546 				       IB_GUARD_T10DIF_CSUM;
547 	}
548 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
549 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
550 
551 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
552 		if (MLX5_CAP_ETH(mdev, csum_cap))
553 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
554 
555 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
556 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
557 			if (max_tso) {
558 				resp.tso_caps.max_tso = 1 << max_tso;
559 				resp.tso_caps.supported_qpts |=
560 					1 << IB_QPT_RAW_PACKET;
561 				resp.response_length += sizeof(resp.tso_caps);
562 			}
563 		}
564 
565 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
566 			resp.rss_caps.rx_hash_function =
567 						MLX5_RX_HASH_FUNC_TOEPLITZ;
568 			resp.rss_caps.rx_hash_fields_mask =
569 						MLX5_RX_HASH_SRC_IPV4 |
570 						MLX5_RX_HASH_DST_IPV4 |
571 						MLX5_RX_HASH_SRC_IPV6 |
572 						MLX5_RX_HASH_DST_IPV6 |
573 						MLX5_RX_HASH_SRC_PORT_TCP |
574 						MLX5_RX_HASH_DST_PORT_TCP |
575 						MLX5_RX_HASH_SRC_PORT_UDP |
576 						MLX5_RX_HASH_DST_PORT_UDP;
577 			resp.response_length += sizeof(resp.rss_caps);
578 		}
579 	} else {
580 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
581 			resp.response_length += sizeof(resp.tso_caps);
582 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
583 			resp.response_length += sizeof(resp.rss_caps);
584 	}
585 
586 	if (MLX5_CAP_GEN(mdev, ipoib_ipoib_offloads)) {
587 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
588 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
589 	}
590 
591 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
592 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs))
593 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
594 
595 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
596 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
597 
598 	props->vendor_part_id	   = mdev->pdev->device;
599 	props->hw_ver		   = mdev->pdev->revision;
600 
601 	props->max_mr_size	   = ~0ull;
602 	props->page_size_cap	   = ~(min_page_size - 1);
603 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
604 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
605 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
606 		     sizeof(struct mlx5_wqe_data_seg);
607 	max_sq_sg = (MLX5_CAP_GEN(mdev, max_wqe_sz_sq) -
608 		     sizeof(struct mlx5_wqe_ctrl_seg)) /
609 		     sizeof(struct mlx5_wqe_data_seg);
610 	props->max_sge = min(max_rq_sg, max_sq_sg);
611 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
612 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
613 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
614 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
615 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
616 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
617 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
618 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
619 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
620 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
621 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
622 	props->max_srq_sge	   = max_rq_sg - 1;
623 	props->max_fast_reg_page_list_len =
624 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
625 	get_atomic_caps(dev, props);
626 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
627 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
628 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
629 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
630 					   props->max_mcast_grp;
631 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
632 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
633 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
634 
635 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
636 	if (MLX5_CAP_GEN(mdev, pg))
637 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
638 	props->odp_caps = dev->odp_caps;
639 #endif
640 
641 	if (MLX5_CAP_GEN(mdev, cd))
642 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
643 
644 	if (!mlx5_core_is_pf(mdev))
645 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
646 
647 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
648 	    IB_LINK_LAYER_ETHERNET) {
649 		props->rss_caps.max_rwq_indirection_tables =
650 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
651 		props->rss_caps.max_rwq_indirection_table_size =
652 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
653 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
654 		props->max_wq_type_rq =
655 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
656 	}
657 
658 	if (uhw->outlen) {
659 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
660 
661 		if (err)
662 			return err;
663 	}
664 
665 	return 0;
666 }
667 
668 enum mlx5_ib_width {
669 	MLX5_IB_WIDTH_1X	= 1 << 0,
670 	MLX5_IB_WIDTH_2X	= 1 << 1,
671 	MLX5_IB_WIDTH_4X	= 1 << 2,
672 	MLX5_IB_WIDTH_8X	= 1 << 3,
673 	MLX5_IB_WIDTH_12X	= 1 << 4
674 };
675 
676 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
677 				  u8 *ib_width)
678 {
679 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
680 	int err = 0;
681 
682 	if (active_width & MLX5_IB_WIDTH_1X) {
683 		*ib_width = IB_WIDTH_1X;
684 	} else if (active_width & MLX5_IB_WIDTH_2X) {
685 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
686 			    (int)active_width);
687 		err = -EINVAL;
688 	} else if (active_width & MLX5_IB_WIDTH_4X) {
689 		*ib_width = IB_WIDTH_4X;
690 	} else if (active_width & MLX5_IB_WIDTH_8X) {
691 		*ib_width = IB_WIDTH_8X;
692 	} else if (active_width & MLX5_IB_WIDTH_12X) {
693 		*ib_width = IB_WIDTH_12X;
694 	} else {
695 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
696 			    (int)active_width);
697 		err = -EINVAL;
698 	}
699 
700 	return err;
701 }
702 
703 enum ib_max_vl_num {
704 	__IB_MAX_VL_0		= 1,
705 	__IB_MAX_VL_0_1		= 2,
706 	__IB_MAX_VL_0_3		= 3,
707 	__IB_MAX_VL_0_7		= 4,
708 	__IB_MAX_VL_0_14	= 5,
709 };
710 
711 enum mlx5_vl_hw_cap {
712 	MLX5_VL_HW_0	= 1,
713 	MLX5_VL_HW_0_1	= 2,
714 	MLX5_VL_HW_0_2	= 3,
715 	MLX5_VL_HW_0_3	= 4,
716 	MLX5_VL_HW_0_4	= 5,
717 	MLX5_VL_HW_0_5	= 6,
718 	MLX5_VL_HW_0_6	= 7,
719 	MLX5_VL_HW_0_7	= 8,
720 	MLX5_VL_HW_0_14	= 15
721 };
722 
723 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
724 				u8 *max_vl_num)
725 {
726 	switch (vl_hw_cap) {
727 	case MLX5_VL_HW_0:
728 		*max_vl_num = __IB_MAX_VL_0;
729 		break;
730 	case MLX5_VL_HW_0_1:
731 		*max_vl_num = __IB_MAX_VL_0_1;
732 		break;
733 	case MLX5_VL_HW_0_3:
734 		*max_vl_num = __IB_MAX_VL_0_3;
735 		break;
736 	case MLX5_VL_HW_0_7:
737 		*max_vl_num = __IB_MAX_VL_0_7;
738 		break;
739 	case MLX5_VL_HW_0_14:
740 		*max_vl_num = __IB_MAX_VL_0_14;
741 		break;
742 
743 	default:
744 		return -EINVAL;
745 	}
746 
747 	return 0;
748 }
749 
750 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
751 			       struct ib_port_attr *props)
752 {
753 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
754 	struct mlx5_core_dev *mdev = dev->mdev;
755 	u32 *rep;
756 	int replen = MLX5_ST_SZ_BYTES(query_hca_vport_context_out);
757 	struct mlx5_ptys_reg *ptys;
758 	struct mlx5_pmtu_reg *pmtu;
759 	struct mlx5_pvlc_reg pvlc;
760 	void *ctx;
761 	int err;
762 
763 	rep = mlx5_vzalloc(replen);
764 	ptys = kzalloc(sizeof(*ptys), GFP_KERNEL);
765 	pmtu = kzalloc(sizeof(*pmtu), GFP_KERNEL);
766 	if (!rep || !ptys || !pmtu) {
767 		err = -ENOMEM;
768 		goto out;
769 	}
770 
771 	memset(props, 0, sizeof(*props));
772 
773 	err = mlx5_query_hca_vport_context(mdev, port, 0, rep, replen);
774 	if (err)
775 		goto out;
776 
777 	ctx = MLX5_ADDR_OF(query_hca_vport_context_out, rep, hca_vport_context);
778 
779 	props->lid		= MLX5_GET(hca_vport_context, ctx, lid);
780 	props->lmc		= MLX5_GET(hca_vport_context, ctx, lmc);
781 	props->sm_lid		= MLX5_GET(hca_vport_context, ctx, sm_lid);
782 	props->sm_sl		= MLX5_GET(hca_vport_context, ctx, sm_sl);
783 	props->state		= MLX5_GET(hca_vport_context, ctx, vport_state);
784 	props->phys_state	= MLX5_GET(hca_vport_context, ctx,
785 					port_physical_state);
786 	props->port_cap_flags	= MLX5_GET(hca_vport_context, ctx, cap_mask1);
787 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
788 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
789 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
790 	props->bad_pkey_cntr	= MLX5_GET(hca_vport_context, ctx,
791 					pkey_violation_counter);
792 	props->qkey_viol_cntr	= MLX5_GET(hca_vport_context, ctx,
793 					qkey_violation_counter);
794 	props->subnet_timeout	= MLX5_GET(hca_vport_context, ctx,
795 					subnet_timeout);
796 	props->init_type_reply	= MLX5_GET(hca_vport_context, ctx,
797 					init_type_reply);
798 	props->grh_required	= MLX5_GET(hca_vport_context, ctx, grh_required);
799 
800 	ptys->proto_mask |= MLX5_PTYS_IB;
801 	ptys->local_port = port;
802 	err = mlx5_core_access_ptys(mdev, ptys, 0);
803 	if (err)
804 		goto out;
805 
806 	err = translate_active_width(ibdev, ptys->ib_link_width_oper,
807 				     &props->active_width);
808 	if (err)
809 		goto out;
810 
811 	props->active_speed	= (u8)ptys->ib_proto_oper;
812 
813 	pmtu->local_port = port;
814 	err = mlx5_core_access_pmtu(mdev, pmtu, 0);
815 	if (err)
816 		goto out;
817 
818 	props->max_mtu		= pmtu->max_mtu;
819 	props->active_mtu	= pmtu->oper_mtu;
820 
821 	memset(&pvlc, 0, sizeof(pvlc));
822 	pvlc.local_port = port;
823 	err = mlx5_core_access_pvlc(mdev, &pvlc, 0);
824 	if (err)
825 		goto out;
826 
827 	err = translate_max_vl_num(ibdev, pvlc.vl_hw_cap,
828 				   &props->max_vl_num);
829 out:
830 	kvfree(rep);
831 	kfree(ptys);
832 	kfree(pmtu);
833 	return err;
834 }
835 
836 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
837 		       struct ib_port_attr *props)
838 {
839 	switch (mlx5_get_vport_access_method(ibdev)) {
840 	case MLX5_VPORT_ACCESS_METHOD_MAD:
841 		return mlx5_query_mad_ifc_port(ibdev, port, props);
842 
843 	case MLX5_VPORT_ACCESS_METHOD_HCA:
844 		return mlx5_query_hca_port(ibdev, port, props);
845 
846 	case MLX5_VPORT_ACCESS_METHOD_NIC:
847 		return mlx5_query_port_roce(ibdev, port, props);
848 
849 	default:
850 		return -EINVAL;
851 	}
852 }
853 
854 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
855 			     union ib_gid *gid)
856 {
857 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
858 	struct mlx5_core_dev *mdev = dev->mdev;
859 
860 	switch (mlx5_get_vport_access_method(ibdev)) {
861 	case MLX5_VPORT_ACCESS_METHOD_MAD:
862 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
863 
864 	case MLX5_VPORT_ACCESS_METHOD_HCA:
865 		return mlx5_query_hca_vport_gid(mdev, port, 0, index, gid);
866 
867 	default:
868 		return -EINVAL;
869 	}
870 
871 }
872 
873 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
874 			      u16 *pkey)
875 {
876 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
877 	struct mlx5_core_dev *mdev = dev->mdev;
878 
879 	switch (mlx5_get_vport_access_method(ibdev)) {
880 	case MLX5_VPORT_ACCESS_METHOD_MAD:
881 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
882 
883 	case MLX5_VPORT_ACCESS_METHOD_HCA:
884 	case MLX5_VPORT_ACCESS_METHOD_NIC:
885 		return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
886 						 pkey);
887 	default:
888 		return -EINVAL;
889 	}
890 }
891 
892 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
893 				 struct ib_device_modify *props)
894 {
895 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
896 	struct mlx5_reg_node_desc in;
897 	struct mlx5_reg_node_desc out;
898 	int err;
899 
900 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
901 		return -EOPNOTSUPP;
902 
903 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
904 		return 0;
905 
906 	/*
907 	 * If possible, pass node desc to FW, so it can generate
908 	 * a 144 trap.  If cmd fails, just ignore.
909 	 */
910 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
911 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
912 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
913 	if (err)
914 		return err;
915 
916 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
917 
918 	return err;
919 }
920 
921 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
922 			       struct ib_port_modify *props)
923 {
924 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
925 	struct ib_port_attr attr;
926 	u32 tmp;
927 	int err;
928 
929 	mutex_lock(&dev->cap_mask_mutex);
930 
931 	err = mlx5_ib_query_port(ibdev, port, &attr);
932 	if (err)
933 		goto out;
934 
935 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
936 		~props->clr_port_cap_mask;
937 
938 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
939 
940 out:
941 	mutex_unlock(&dev->cap_mask_mutex);
942 	return err;
943 }
944 
945 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
946 						  struct ib_udata *udata)
947 {
948 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
949 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
950 	struct mlx5_ib_alloc_ucontext_resp resp = {};
951 	struct mlx5_ib_ucontext *context;
952 	struct mlx5_uuar_info *uuari;
953 	struct mlx5_uar *uars;
954 	int gross_uuars;
955 	int num_uars;
956 	int ver;
957 	int uuarn;
958 	int err;
959 	int i;
960 	size_t reqlen;
961 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
962 				     max_cqe_version);
963 
964 	if (!dev->ib_active)
965 		return ERR_PTR(-EAGAIN);
966 
967 	if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
968 		return ERR_PTR(-EINVAL);
969 
970 	reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
971 	if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
972 		ver = 0;
973 	else if (reqlen >= min_req_v2)
974 		ver = 2;
975 	else
976 		return ERR_PTR(-EINVAL);
977 
978 	err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
979 	if (err)
980 		return ERR_PTR(err);
981 
982 	if (req.flags)
983 		return ERR_PTR(-EINVAL);
984 
985 	if (req.total_num_uuars > MLX5_MAX_UUARS)
986 		return ERR_PTR(-ENOMEM);
987 
988 	if (req.total_num_uuars == 0)
989 		return ERR_PTR(-EINVAL);
990 
991 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
992 		return ERR_PTR(-EOPNOTSUPP);
993 
994 	if (reqlen > sizeof(req) &&
995 	    !ib_is_udata_cleared(udata, sizeof(req),
996 				 reqlen - sizeof(req)))
997 		return ERR_PTR(-EOPNOTSUPP);
998 
999 	req.total_num_uuars = ALIGN(req.total_num_uuars,
1000 				    MLX5_NON_FP_BF_REGS_PER_PAGE);
1001 	if (req.num_low_latency_uuars > req.total_num_uuars - 1)
1002 		return ERR_PTR(-EINVAL);
1003 
1004 	num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
1005 	gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
1006 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1007 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1008 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1009 	resp.cache_line_size = cache_line_size();
1010 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1011 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1012 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1013 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1014 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1015 	resp.cqe_version = min_t(__u8,
1016 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1017 				 req.max_cqe_version);
1018 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1019 				   sizeof(resp.response_length), udata->outlen);
1020 
1021 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1022 	if (!context)
1023 		return ERR_PTR(-ENOMEM);
1024 
1025 	uuari = &context->uuari;
1026 	mutex_init(&uuari->lock);
1027 	uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
1028 	if (!uars) {
1029 		err = -ENOMEM;
1030 		goto out_ctx;
1031 	}
1032 
1033 	uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
1034 				sizeof(*uuari->bitmap),
1035 				GFP_KERNEL);
1036 	if (!uuari->bitmap) {
1037 		err = -ENOMEM;
1038 		goto out_uar_ctx;
1039 	}
1040 	/*
1041 	 * clear all fast path uuars
1042 	 */
1043 	for (i = 0; i < gross_uuars; i++) {
1044 		uuarn = i & 3;
1045 		if (uuarn == 2 || uuarn == 3)
1046 			set_bit(i, uuari->bitmap);
1047 	}
1048 
1049 	uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
1050 	if (!uuari->count) {
1051 		err = -ENOMEM;
1052 		goto out_bitmap;
1053 	}
1054 
1055 	for (i = 0; i < num_uars; i++) {
1056 		err = mlx5_cmd_alloc_uar(dev->mdev, &uars[i].index);
1057 		if (err)
1058 			goto out_count;
1059 	}
1060 
1061 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1062 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1063 #endif
1064 
1065 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1066 		err = mlx5_alloc_transport_domain(dev->mdev,
1067 						       &context->tdn);
1068 		if (err)
1069 			goto out_uars;
1070 	}
1071 
1072 	INIT_LIST_HEAD(&context->vma_private_list);
1073 	INIT_LIST_HEAD(&context->db_page_list);
1074 	mutex_init(&context->db_page_mutex);
1075 
1076 	resp.tot_uuars = req.total_num_uuars;
1077 	resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1078 
1079 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1080 		resp.response_length += sizeof(resp.cqe_version);
1081 
1082 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1083 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE;
1084 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1085 	}
1086 
1087 	/*
1088 	 * We don't want to expose information from the PCI bar that is located
1089 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1090 	 * pretend we don't support reading the HCA's core clock. This is also
1091 	 * forced by mmap function.
1092 	 */
1093 	if (PAGE_SIZE <= 4096 &&
1094 	    field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1095 		resp.comp_mask |=
1096 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1097 		resp.hca_core_clock_offset =
1098 			offsetof(struct mlx5_init_seg, internal_timer_h) %
1099 			PAGE_SIZE;
1100 		resp.response_length += sizeof(resp.hca_core_clock_offset) +
1101 					sizeof(resp.reserved2);
1102 	}
1103 
1104 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1105 	if (err)
1106 		goto out_td;
1107 
1108 	uuari->ver = ver;
1109 	uuari->num_low_latency_uuars = req.num_low_latency_uuars;
1110 	uuari->uars = uars;
1111 	uuari->num_uars = num_uars;
1112 	context->cqe_version = resp.cqe_version;
1113 
1114 	return &context->ibucontext;
1115 
1116 out_td:
1117 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1118 		mlx5_dealloc_transport_domain(dev->mdev, context->tdn);
1119 
1120 out_uars:
1121 	for (i--; i >= 0; i--)
1122 		mlx5_cmd_free_uar(dev->mdev, uars[i].index);
1123 out_count:
1124 	kfree(uuari->count);
1125 
1126 out_bitmap:
1127 	kfree(uuari->bitmap);
1128 
1129 out_uar_ctx:
1130 	kfree(uars);
1131 
1132 out_ctx:
1133 	kfree(context);
1134 	return ERR_PTR(err);
1135 }
1136 
1137 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1138 {
1139 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1140 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1141 	struct mlx5_uuar_info *uuari = &context->uuari;
1142 	int i;
1143 
1144 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1145 		mlx5_dealloc_transport_domain(dev->mdev, context->tdn);
1146 
1147 	for (i = 0; i < uuari->num_uars; i++) {
1148 		if (mlx5_cmd_free_uar(dev->mdev, uuari->uars[i].index))
1149 			mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
1150 	}
1151 
1152 	kfree(uuari->count);
1153 	kfree(uuari->bitmap);
1154 	kfree(uuari->uars);
1155 	kfree(context);
1156 
1157 	return 0;
1158 }
1159 
1160 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
1161 {
1162 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + index;
1163 }
1164 
1165 static int get_command(unsigned long offset)
1166 {
1167 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1168 }
1169 
1170 static int get_arg(unsigned long offset)
1171 {
1172 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1173 }
1174 
1175 static int get_index(unsigned long offset)
1176 {
1177 	return get_arg(offset);
1178 }
1179 
1180 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1181 {
1182 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1183 	 * is done through either mremap flow or split_vma (usually due to
1184 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1185 	 * as this VMA is strongly hardware related.  Therefore we set the
1186 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1187 	 * calling us again and trying to do incorrect actions.  We assume that
1188 	 * the original VMA size is exactly a single page, and therefore all
1189 	 * "splitting" operation will not happen to it.
1190 	 */
1191 	area->vm_ops = NULL;
1192 }
1193 
1194 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1195 {
1196 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1197 
1198 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1199 	 * file itself is closed, therefore no sync is needed with the regular
1200 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1201 	 * However need a sync with accessing the vma as part of
1202 	 * mlx5_ib_disassociate_ucontext.
1203 	 * The close operation is usually called under mm->mmap_sem except when
1204 	 * process is exiting.
1205 	 * The exiting case is handled explicitly as part of
1206 	 * mlx5_ib_disassociate_ucontext.
1207 	 */
1208 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1209 
1210 	/* setting the vma context pointer to null in the mlx5_ib driver's
1211 	 * private data, to protect a race condition in
1212 	 * mlx5_ib_disassociate_ucontext().
1213 	 */
1214 	mlx5_ib_vma_priv_data->vma = NULL;
1215 	list_del(&mlx5_ib_vma_priv_data->list);
1216 	kfree(mlx5_ib_vma_priv_data);
1217 }
1218 
1219 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1220 	.open = mlx5_ib_vma_open,
1221 	.close = mlx5_ib_vma_close
1222 };
1223 
1224 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1225 				struct mlx5_ib_ucontext *ctx)
1226 {
1227 	struct mlx5_ib_vma_private_data *vma_prv;
1228 	struct list_head *vma_head = &ctx->vma_private_list;
1229 
1230 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1231 	if (!vma_prv)
1232 		return -ENOMEM;
1233 
1234 	vma_prv->vma = vma;
1235 	vma->vm_private_data = vma_prv;
1236 	vma->vm_ops =  &mlx5_ib_vm_ops;
1237 
1238 	list_add(&vma_prv->list, vma_head);
1239 
1240 	return 0;
1241 }
1242 
1243 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1244 {
1245 	int ret;
1246 	struct vm_area_struct *vma;
1247 	struct mlx5_ib_vma_private_data *vma_private, *n;
1248 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1249 	struct task_struct *owning_process  = NULL;
1250 	struct mm_struct   *owning_mm       = NULL;
1251 
1252 	owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1253 	if (!owning_process)
1254 		return;
1255 
1256 	owning_mm = get_task_mm(owning_process);
1257 	if (!owning_mm) {
1258 		pr_info("no mm, disassociate ucontext is pending task termination\n");
1259 		while (1) {
1260 			put_task_struct(owning_process);
1261 			usleep_range(1000, 2000);
1262 			owning_process = get_pid_task(ibcontext->tgid,
1263 						      PIDTYPE_PID);
1264 			if (!owning_process /* ||
1265 			    owning_process->state == TASK_DEAD */) {
1266 				pr_info("disassociate ucontext done, task was terminated\n");
1267 				/* in case task was dead need to release the
1268 				 * task struct.
1269 				 */
1270 				if (owning_process)
1271 					put_task_struct(owning_process);
1272 				return;
1273 			}
1274 		}
1275 	}
1276 
1277 	/* need to protect from a race on closing the vma as part of
1278 	 * mlx5_ib_vma_close.
1279 	 */
1280 	down_read(&owning_mm->mmap_sem);
1281 	list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1282 				 list) {
1283 		vma = vma_private->vma;
1284 		ret = zap_vma_ptes(vma, vma->vm_start,
1285 				   PAGE_SIZE);
1286 		WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1287 		/* context going to be destroyed, should
1288 		 * not access ops any more.
1289 		 */
1290 		vma->vm_ops = NULL;
1291 		list_del(&vma_private->list);
1292 		kfree(vma_private);
1293 	}
1294 	up_read(&owning_mm->mmap_sem);
1295 	mmput(owning_mm);
1296 	put_task_struct(owning_process);
1297 }
1298 
1299 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1300 {
1301 	switch (cmd) {
1302 	case MLX5_IB_MMAP_WC_PAGE:
1303 		return "WC";
1304 	case MLX5_IB_MMAP_REGULAR_PAGE:
1305 		return "best effort WC";
1306 	case MLX5_IB_MMAP_NC_PAGE:
1307 		return "NC";
1308 	default:
1309 		return NULL;
1310 	}
1311 }
1312 
1313 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1314 		    struct vm_area_struct *vma,
1315 		    struct mlx5_ib_ucontext *context)
1316 {
1317 	struct mlx5_uuar_info *uuari = &context->uuari;
1318 	int err;
1319 	unsigned long idx;
1320 	phys_addr_t pfn, pa;
1321 	pgprot_t prot;
1322 
1323 	switch (cmd) {
1324 	case MLX5_IB_MMAP_WC_PAGE:
1325 /* Some architectures don't support WC memory */
1326 #if defined(CONFIG_X86)
1327 		if (!pat_enabled())
1328 			return -EPERM;
1329 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1330 			return -EPERM;
1331 #endif
1332 	/* fall through */
1333 	case MLX5_IB_MMAP_REGULAR_PAGE:
1334 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1335 		prot = pgprot_writecombine(vma->vm_page_prot);
1336 		break;
1337 	case MLX5_IB_MMAP_NC_PAGE:
1338 		prot = pgprot_noncached(vma->vm_page_prot);
1339 		break;
1340 	default:
1341 		return -EINVAL;
1342 	}
1343 
1344 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1345 		return -EINVAL;
1346 
1347 	idx = get_index(vma->vm_pgoff);
1348 	if (idx >= uuari->num_uars)
1349 		return -EINVAL;
1350 
1351 	pfn = uar_index2pfn(dev, uuari->uars[idx].index);
1352 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1353 
1354 	vma->vm_page_prot = prot;
1355 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1356 				 PAGE_SIZE, vma->vm_page_prot);
1357 	if (err) {
1358 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%llx, pfn=%pa, mmap_cmd=%s\n",
1359 			    err, (unsigned long long)vma->vm_start, &pfn, mmap_cmd2str(cmd));
1360 		return -EAGAIN;
1361 	}
1362 
1363 	pa = pfn << PAGE_SHIFT;
1364 	mlx5_ib_dbg(dev, "mapped %s at 0x%llx, PA %pa\n", mmap_cmd2str(cmd),
1365 		    (unsigned long long)vma->vm_start, &pa);
1366 
1367 	return mlx5_ib_set_vma_data(vma, context);
1368 }
1369 
1370 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1371 {
1372 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1373 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1374 	unsigned long command;
1375 	phys_addr_t pfn;
1376 
1377 	command = get_command(vma->vm_pgoff);
1378 	switch (command) {
1379 	case MLX5_IB_MMAP_WC_PAGE:
1380 	case MLX5_IB_MMAP_NC_PAGE:
1381 	case MLX5_IB_MMAP_REGULAR_PAGE:
1382 		return uar_mmap(dev, command, vma, context);
1383 
1384 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1385 		return -ENOSYS;
1386 
1387 	case MLX5_IB_MMAP_CORE_CLOCK:
1388 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1389 			return -EINVAL;
1390 
1391 		if (vma->vm_flags & VM_WRITE)
1392 			return -EPERM;
1393 
1394 		/* Don't expose to user-space information it shouldn't have */
1395 		if (PAGE_SIZE > 4096)
1396 			return -EOPNOTSUPP;
1397 
1398 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1399 		pfn = (dev->mdev->iseg_base +
1400 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1401 			PAGE_SHIFT;
1402 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1403 				       PAGE_SIZE, vma->vm_page_prot))
1404 			return -EAGAIN;
1405 
1406 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%llx, PA 0x%llx\n",
1407 			    (unsigned long long)vma->vm_start,
1408 			    (unsigned long long)pfn << PAGE_SHIFT);
1409 		break;
1410 
1411 	default:
1412 		return -EINVAL;
1413 	}
1414 
1415 	return 0;
1416 }
1417 
1418 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1419 				      struct ib_ucontext *context,
1420 				      struct ib_udata *udata)
1421 {
1422 	struct mlx5_ib_alloc_pd_resp resp;
1423 	struct mlx5_ib_pd *pd;
1424 	int err;
1425 
1426 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1427 	if (!pd)
1428 		return ERR_PTR(-ENOMEM);
1429 
1430 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1431 	if (err) {
1432 		kfree(pd);
1433 		return ERR_PTR(err);
1434 	}
1435 
1436 	if (context) {
1437 		resp.pdn = pd->pdn;
1438 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1439 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1440 			kfree(pd);
1441 			return ERR_PTR(-EFAULT);
1442 		}
1443 	}
1444 
1445 	return &pd->ibpd;
1446 }
1447 
1448 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1449 {
1450 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1451 	struct mlx5_ib_pd *mpd = to_mpd(pd);
1452 
1453 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1454 	kfree(mpd);
1455 
1456 	return 0;
1457 }
1458 
1459 enum {
1460 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
1461 	MATCH_CRITERIA_ENABLE_MISC_BIT,
1462 	MATCH_CRITERIA_ENABLE_INNER_BIT
1463 };
1464 
1465 #define HEADER_IS_ZERO(match_criteria, headers)			           \
1466 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1467 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
1468 
1469 static u8 get_match_criteria_enable(u32 *match_criteria)
1470 {
1471 	u8 match_criteria_enable;
1472 
1473 	match_criteria_enable =
1474 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1475 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
1476 	match_criteria_enable |=
1477 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1478 		MATCH_CRITERIA_ENABLE_MISC_BIT;
1479 	match_criteria_enable |=
1480 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1481 		MATCH_CRITERIA_ENABLE_INNER_BIT;
1482 
1483 	return match_criteria_enable;
1484 }
1485 
1486 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1487 {
1488 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1489 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1490 }
1491 
1492 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1493 {
1494 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1495 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1496 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1497 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1498 }
1499 
1500 #define LAST_ETH_FIELD vlan_tag
1501 #define LAST_IB_FIELD sl
1502 #define LAST_IPV4_FIELD tos
1503 #define LAST_IPV6_FIELD traffic_class
1504 #define LAST_TCP_UDP_FIELD src_port
1505 
1506 /* Field is the last supported field */
1507 #define FIELDS_NOT_SUPPORTED(filter, field)\
1508 	memchr_inv((void *)&filter.field  +\
1509 		   sizeof(filter.field), 0,\
1510 		   sizeof(filter) -\
1511 		   offsetof(typeof(filter), field) -\
1512 		   sizeof(filter.field))
1513 
1514 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1515 			   const union ib_flow_spec *ib_spec)
1516 {
1517 	void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1518 					     outer_headers);
1519 	void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1520 					     outer_headers);
1521 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1522 					   misc_parameters);
1523 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1524 					   misc_parameters);
1525 
1526 	switch (ib_spec->type) {
1527 	case IB_FLOW_SPEC_ETH:
1528 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1529 			return -ENOTSUPP;
1530 
1531 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1532 					     dmac_47_16),
1533 				ib_spec->eth.mask.dst_mac);
1534 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1535 					     dmac_47_16),
1536 				ib_spec->eth.val.dst_mac);
1537 
1538 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1539 					     smac_47_16),
1540 				ib_spec->eth.mask.src_mac);
1541 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1542 					     smac_47_16),
1543 				ib_spec->eth.val.src_mac);
1544 
1545 		if (ib_spec->eth.mask.vlan_tag) {
1546 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1547 				 cvlan_tag, 1);
1548 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1549 				 cvlan_tag, 1);
1550 
1551 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1552 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1553 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1554 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1555 
1556 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1557 				 first_cfi,
1558 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1559 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1560 				 first_cfi,
1561 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1562 
1563 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1564 				 first_prio,
1565 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1566 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1567 				 first_prio,
1568 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1569 		}
1570 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1571 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1572 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1573 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
1574 		break;
1575 	case IB_FLOW_SPEC_IPV4:
1576 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1577 			return -ENOTSUPP;
1578 
1579 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1580 			 ethertype, 0xffff);
1581 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1582 			 ethertype, ETH_P_IP);
1583 
1584 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1585 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1586 		       &ib_spec->ipv4.mask.src_ip,
1587 		       sizeof(ib_spec->ipv4.mask.src_ip));
1588 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1589 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1590 		       &ib_spec->ipv4.val.src_ip,
1591 		       sizeof(ib_spec->ipv4.val.src_ip));
1592 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1593 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1594 		       &ib_spec->ipv4.mask.dst_ip,
1595 		       sizeof(ib_spec->ipv4.mask.dst_ip));
1596 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1597 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1598 		       &ib_spec->ipv4.val.dst_ip,
1599 		       sizeof(ib_spec->ipv4.val.dst_ip));
1600 
1601 		set_tos(outer_headers_c, outer_headers_v,
1602 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
1603 
1604 		set_proto(outer_headers_c, outer_headers_v,
1605 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
1606 		break;
1607 	case IB_FLOW_SPEC_IPV6:
1608 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1609 			return -ENOTSUPP;
1610 
1611 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1612 			 ethertype, 0xffff);
1613 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1614 			 ethertype, IPPROTO_IPV6);
1615 
1616 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1617 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1618 		       &ib_spec->ipv6.mask.src_ip,
1619 		       sizeof(ib_spec->ipv6.mask.src_ip));
1620 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1621 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1622 		       &ib_spec->ipv6.val.src_ip,
1623 		       sizeof(ib_spec->ipv6.val.src_ip));
1624 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1625 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1626 		       &ib_spec->ipv6.mask.dst_ip,
1627 		       sizeof(ib_spec->ipv6.mask.dst_ip));
1628 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1629 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1630 		       &ib_spec->ipv6.val.dst_ip,
1631 		       sizeof(ib_spec->ipv6.val.dst_ip));
1632 
1633 		set_tos(outer_headers_c, outer_headers_v,
1634 			ib_spec->ipv6.mask.traffic_class,
1635 			ib_spec->ipv6.val.traffic_class);
1636 
1637 		set_proto(outer_headers_c, outer_headers_v,
1638 			  ib_spec->ipv6.mask.next_hdr,
1639 			  ib_spec->ipv6.val.next_hdr);
1640 
1641 		MLX5_SET(fte_match_set_misc, misc_params_c,
1642 			 outer_ipv6_flow_label,
1643 			 ntohl(ib_spec->ipv6.mask.flow_label));
1644 		MLX5_SET(fte_match_set_misc, misc_params_v,
1645 			 outer_ipv6_flow_label,
1646 			 ntohl(ib_spec->ipv6.val.flow_label));
1647 		break;
1648 	case IB_FLOW_SPEC_TCP:
1649 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1650 					 LAST_TCP_UDP_FIELD))
1651 			return -ENOTSUPP;
1652 
1653 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1654 			 0xff);
1655 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1656 			 IPPROTO_TCP);
1657 
1658 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport,
1659 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1660 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport,
1661 			 ntohs(ib_spec->tcp_udp.val.src_port));
1662 
1663 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport,
1664 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1665 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport,
1666 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1667 		break;
1668 	case IB_FLOW_SPEC_UDP:
1669 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1670 					 LAST_TCP_UDP_FIELD))
1671 			return -ENOTSUPP;
1672 
1673 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1674 			 0xff);
1675 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1676 			 IPPROTO_UDP);
1677 
1678 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport,
1679 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1680 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport,
1681 			 ntohs(ib_spec->tcp_udp.val.src_port));
1682 
1683 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport,
1684 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1685 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport,
1686 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1687 		break;
1688 	default:
1689 		return -EINVAL;
1690 	}
1691 
1692 	return 0;
1693 }
1694 
1695 /* If a flow could catch both multicast and unicast packets,
1696  * it won't fall into the multicast flow steering table and this rule
1697  * could steal other multicast packets.
1698  */
1699 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1700 {
1701 	struct ib_flow_spec_eth *eth_spec;
1702 
1703 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1704 	    ib_attr->size < sizeof(struct ib_flow_attr) +
1705 	    sizeof(struct ib_flow_spec_eth) ||
1706 	    ib_attr->num_of_specs < 1)
1707 		return false;
1708 
1709 	eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1710 	if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1711 	    eth_spec->size != sizeof(*eth_spec))
1712 		return false;
1713 
1714 	return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1715 	       is_multicast_ether_addr(eth_spec->val.dst_mac);
1716 }
1717 
1718 static bool is_valid_attr(const struct ib_flow_attr *flow_attr)
1719 {
1720 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1721 	bool has_ipv4_spec = false;
1722 	bool eth_type_ipv4 = true;
1723 	unsigned int spec_index;
1724 
1725 	/* Validate that ethertype is correct */
1726 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1727 		if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1728 		    ib_spec->eth.mask.ether_type) {
1729 			if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1730 			      ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1731 				eth_type_ipv4 = false;
1732 		} else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1733 			has_ipv4_spec = true;
1734 		}
1735 		ib_spec = (void *)ib_spec + ib_spec->size;
1736 	}
1737 	return !has_ipv4_spec || eth_type_ipv4;
1738 }
1739 
1740 static void put_flow_table(struct mlx5_ib_dev *dev,
1741 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
1742 {
1743 	prio->refcount -= !!ft_added;
1744 	if (!prio->refcount) {
1745 		mlx5_destroy_flow_table(prio->flow_table);
1746 		prio->flow_table = NULL;
1747 	}
1748 }
1749 
1750 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
1751 {
1752 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
1753 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
1754 							  struct mlx5_ib_flow_handler,
1755 							  ibflow);
1756 	struct mlx5_ib_flow_handler *iter, *tmp;
1757 
1758 	mutex_lock(&dev->flow_db.lock);
1759 
1760 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
1761 		mlx5_del_flow_rule(iter->rule);
1762 		put_flow_table(dev, iter->prio, true);
1763 		list_del(&iter->list);
1764 		kfree(iter);
1765 	}
1766 
1767 	mlx5_del_flow_rule(handler->rule);
1768 	put_flow_table(dev, handler->prio, true);
1769 	mutex_unlock(&dev->flow_db.lock);
1770 
1771 	kfree(handler);
1772 
1773 	return 0;
1774 }
1775 
1776 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
1777 {
1778 	priority *= 2;
1779 	if (!dont_trap)
1780 		priority++;
1781 	return priority;
1782 }
1783 
1784 enum flow_table_type {
1785 	MLX5_IB_FT_RX,
1786 	MLX5_IB_FT_TX
1787 };
1788 
1789 #define MLX5_FS_MAX_TYPES	 10
1790 #define MLX5_FS_MAX_ENTRIES	 32000UL
1791 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
1792 						struct ib_flow_attr *flow_attr,
1793 						enum flow_table_type ft_type)
1794 {
1795 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
1796 	struct mlx5_flow_namespace *ns = NULL;
1797 	struct mlx5_ib_flow_prio *prio;
1798 	struct mlx5_flow_table *ft;
1799 	int num_entries;
1800 	int num_groups;
1801 	int priority;
1802 	int err = 0;
1803 
1804 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1805 		if (flow_is_multicast_only(flow_attr) &&
1806 		    !dont_trap)
1807 			priority = MLX5_IB_FLOW_MCAST_PRIO;
1808 		else
1809 			priority = ib_prio_to_core_prio(flow_attr->priority,
1810 							dont_trap);
1811 		ns = mlx5_get_flow_namespace(dev->mdev,
1812 					     MLX5_FLOW_NAMESPACE_BYPASS);
1813 		num_entries = MLX5_FS_MAX_ENTRIES;
1814 		num_groups = MLX5_FS_MAX_TYPES;
1815 		prio = &dev->flow_db.prios[priority];
1816 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1817 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1818 		ns = mlx5_get_flow_namespace(dev->mdev,
1819 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
1820 		build_leftovers_ft_param("bypass", &priority,
1821 					 &num_entries,
1822 					 &num_groups);
1823 		prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
1824 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
1825 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
1826 					allow_sniffer_and_nic_rx_shared_tir))
1827 			return ERR_PTR(-ENOTSUPP);
1828 
1829 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
1830 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
1831 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
1832 
1833 		prio = &dev->flow_db.sniffer[ft_type];
1834 		priority = 0;
1835 		num_entries = 1;
1836 		num_groups = 1;
1837 	}
1838 
1839 	if (!ns)
1840 		return ERR_PTR(-ENOTSUPP);
1841 
1842 	ft = prio->flow_table;
1843 	if (!ft) {
1844 		ft = mlx5_create_auto_grouped_flow_table(ns, priority, "bypass",
1845 							 num_entries,
1846 							 num_groups);
1847 
1848 		if (!IS_ERR(ft)) {
1849 			prio->refcount = 0;
1850 			prio->flow_table = ft;
1851 		} else {
1852 			err = PTR_ERR(ft);
1853 		}
1854 	}
1855 
1856 	return err ? ERR_PTR(err) : prio;
1857 }
1858 
1859 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
1860 						     struct mlx5_ib_flow_prio *ft_prio,
1861 						     const struct ib_flow_attr *flow_attr,
1862 						     struct mlx5_flow_destination *dst)
1863 {
1864 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
1865 	struct mlx5_ib_flow_handler *handler;
1866 	struct mlx5_flow_spec *spec;
1867 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
1868 	unsigned int spec_index;
1869 	u32 action;
1870 	int err = 0;
1871 
1872 	if (!is_valid_attr(flow_attr))
1873 		return ERR_PTR(-EINVAL);
1874 
1875 	spec = mlx5_vzalloc(sizeof(*spec));
1876 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
1877 	if (!handler || !spec) {
1878 		err = -ENOMEM;
1879 		goto free;
1880 	}
1881 
1882 	INIT_LIST_HEAD(&handler->list);
1883 
1884 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1885 		err = parse_flow_attr(spec->match_criteria,
1886 				      spec->match_value, ib_flow);
1887 		if (err < 0)
1888 			goto free;
1889 
1890 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
1891 	}
1892 
1893 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
1894 	action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
1895 		MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
1896 	handler->rule = mlx5_add_flow_rule(ft, spec->match_criteria_enable,
1897 					   spec->match_criteria,
1898 					   spec->match_value,
1899 					   action,
1900 					   MLX5_FS_DEFAULT_FLOW_TAG,
1901 					   dst);
1902 
1903 	if (IS_ERR(handler->rule)) {
1904 		err = PTR_ERR(handler->rule);
1905 		goto free;
1906 	}
1907 
1908 	ft_prio->refcount++;
1909 	handler->prio = ft_prio;
1910 
1911 	ft_prio->flow_table = ft;
1912 free:
1913 	if (err)
1914 		kfree(handler);
1915 	kvfree(spec);
1916 	return err ? ERR_PTR(err) : handler;
1917 }
1918 
1919 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
1920 							  struct mlx5_ib_flow_prio *ft_prio,
1921 							  struct ib_flow_attr *flow_attr,
1922 							  struct mlx5_flow_destination *dst)
1923 {
1924 	struct mlx5_ib_flow_handler *handler_dst = NULL;
1925 	struct mlx5_ib_flow_handler *handler = NULL;
1926 
1927 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
1928 	if (!IS_ERR(handler)) {
1929 		handler_dst = create_flow_rule(dev, ft_prio,
1930 					       flow_attr, dst);
1931 		if (IS_ERR(handler_dst)) {
1932 			mlx5_del_flow_rule(handler->rule);
1933 			ft_prio->refcount--;
1934 			kfree(handler);
1935 			handler = handler_dst;
1936 		} else {
1937 			list_add(&handler_dst->list, &handler->list);
1938 		}
1939 	}
1940 
1941 	return handler;
1942 }
1943 enum {
1944 	LEFTOVERS_MC,
1945 	LEFTOVERS_UC,
1946 };
1947 
1948 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
1949 							  struct mlx5_ib_flow_prio *ft_prio,
1950 							  struct ib_flow_attr *flow_attr,
1951 							  struct mlx5_flow_destination *dst)
1952 {
1953 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
1954 	struct mlx5_ib_flow_handler *handler = NULL;
1955 
1956 	static struct {
1957 		struct ib_flow_attr	flow_attr;
1958 		struct ib_flow_spec_eth eth_flow;
1959 	} leftovers_specs[] = {
1960 		[LEFTOVERS_MC] = {
1961 			.flow_attr = {
1962 				.num_of_specs = 1,
1963 				.size = sizeof(leftovers_specs[0])
1964 			},
1965 			.eth_flow = {
1966 				.type = IB_FLOW_SPEC_ETH,
1967 				.size = sizeof(struct ib_flow_spec_eth),
1968 				.mask = {.dst_mac = {0x1} },
1969 				.val =  {.dst_mac = {0x1} }
1970 			}
1971 		},
1972 		[LEFTOVERS_UC] = {
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 = {} }
1982 			}
1983 		}
1984 	};
1985 
1986 	handler = create_flow_rule(dev, ft_prio,
1987 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
1988 				   dst);
1989 	if (!IS_ERR(handler) &&
1990 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
1991 		handler_ucast = create_flow_rule(dev, ft_prio,
1992 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
1993 						 dst);
1994 		if (IS_ERR(handler_ucast)) {
1995 			mlx5_del_flow_rule(handler->rule);
1996 			ft_prio->refcount--;
1997 			kfree(handler);
1998 			handler = handler_ucast;
1999 		} else {
2000 			list_add(&handler_ucast->list, &handler->list);
2001 		}
2002 	}
2003 
2004 	return handler;
2005 }
2006 
2007 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2008 							struct mlx5_ib_flow_prio *ft_rx,
2009 							struct mlx5_ib_flow_prio *ft_tx,
2010 							struct mlx5_flow_destination *dst)
2011 {
2012 	struct mlx5_ib_flow_handler *handler_rx;
2013 	struct mlx5_ib_flow_handler *handler_tx;
2014 	int err;
2015 	static const struct ib_flow_attr flow_attr  = {
2016 		.num_of_specs = 0,
2017 		.size = sizeof(flow_attr)
2018 	};
2019 
2020 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2021 	if (IS_ERR(handler_rx)) {
2022 		err = PTR_ERR(handler_rx);
2023 		goto err;
2024 	}
2025 
2026 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2027 	if (IS_ERR(handler_tx)) {
2028 		err = PTR_ERR(handler_tx);
2029 		goto err_tx;
2030 	}
2031 
2032 	list_add(&handler_tx->list, &handler_rx->list);
2033 
2034 	return handler_rx;
2035 
2036 err_tx:
2037 	mlx5_del_flow_rule(handler_rx->rule);
2038 	ft_rx->refcount--;
2039 	kfree(handler_rx);
2040 err:
2041 	return ERR_PTR(err);
2042 }
2043 
2044 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2045 					   struct ib_flow_attr *flow_attr,
2046 					   int domain)
2047 {
2048 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2049 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2050 	struct mlx5_ib_flow_handler *handler = NULL;
2051 	struct mlx5_flow_destination *dst = NULL;
2052 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2053 	struct mlx5_ib_flow_prio *ft_prio;
2054 	int err;
2055 
2056 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2057 		return ERR_PTR(-ENOSPC);
2058 
2059 	if (domain != IB_FLOW_DOMAIN_USER ||
2060 	    flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2061 	    (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2062 		return ERR_PTR(-EINVAL);
2063 
2064 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2065 	if (!dst)
2066 		return ERR_PTR(-ENOMEM);
2067 
2068 	mutex_lock(&dev->flow_db.lock);
2069 
2070 	ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2071 	if (IS_ERR(ft_prio)) {
2072 		err = PTR_ERR(ft_prio);
2073 		goto unlock;
2074 	}
2075 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2076 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2077 		if (IS_ERR(ft_prio_tx)) {
2078 			err = PTR_ERR(ft_prio_tx);
2079 			ft_prio_tx = NULL;
2080 			goto destroy_ft;
2081 		}
2082 	}
2083 
2084 	dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2085 	if (mqp->flags & MLX5_IB_QP_RSS)
2086 		dst->tir_num = mqp->rss_qp.tirn;
2087 	else
2088 		dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2089 
2090 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2091 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
2092 			handler = create_dont_trap_rule(dev, ft_prio,
2093 							flow_attr, dst);
2094 		} else {
2095 			handler = create_flow_rule(dev, ft_prio, flow_attr,
2096 						   dst);
2097 		}
2098 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2099 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2100 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2101 						dst);
2102 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2103 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2104 	} else {
2105 		err = -EINVAL;
2106 		goto destroy_ft;
2107 	}
2108 
2109 	if (IS_ERR(handler)) {
2110 		err = PTR_ERR(handler);
2111 		handler = NULL;
2112 		goto destroy_ft;
2113 	}
2114 
2115 	mutex_unlock(&dev->flow_db.lock);
2116 	kfree(dst);
2117 
2118 	return &handler->ibflow;
2119 
2120 destroy_ft:
2121 	put_flow_table(dev, ft_prio, false);
2122 	if (ft_prio_tx)
2123 		put_flow_table(dev, ft_prio_tx, false);
2124 unlock:
2125 	mutex_unlock(&dev->flow_db.lock);
2126 	kfree(dst);
2127 	kfree(handler);
2128 	return ERR_PTR(err);
2129 }
2130 
2131 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2132 {
2133 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2134 	int err;
2135 
2136 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2137 	if (err)
2138 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2139 			     ibqp->qp_num, gid->raw);
2140 
2141 	return err;
2142 }
2143 
2144 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2145 {
2146 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2147 	int err;
2148 
2149 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2150 	if (err)
2151 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2152 			     ibqp->qp_num, gid->raw);
2153 
2154 	return err;
2155 }
2156 
2157 static int init_node_data(struct mlx5_ib_dev *dev)
2158 {
2159 	int err;
2160 
2161 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2162 	if (err)
2163 		return err;
2164 
2165 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2166 }
2167 
2168 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2169 			     char *buf)
2170 {
2171 	struct mlx5_ib_dev *dev =
2172 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2173 
2174 	return sprintf(buf, "%lld\n", (long long)dev->mdev->priv.fw_pages);
2175 }
2176 
2177 static ssize_t show_reg_pages(struct device *device,
2178 			      struct device_attribute *attr, char *buf)
2179 {
2180 	struct mlx5_ib_dev *dev =
2181 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2182 
2183 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2184 }
2185 
2186 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2187 			char *buf)
2188 {
2189 	struct mlx5_ib_dev *dev =
2190 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2191 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2192 }
2193 
2194 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2195 			char *buf)
2196 {
2197 	struct mlx5_ib_dev *dev =
2198 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2199 	return sprintf(buf, "%x\n", dev->mdev->pdev->revision);
2200 }
2201 
2202 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2203 			  char *buf)
2204 {
2205 	struct mlx5_ib_dev *dev =
2206 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2207 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2208 		       dev->mdev->board_id);
2209 }
2210 
2211 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2212 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2213 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2214 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2215 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2216 
2217 static struct device_attribute *mlx5_class_attributes[] = {
2218 	&dev_attr_hw_rev,
2219 	&dev_attr_hca_type,
2220 	&dev_attr_board_id,
2221 	&dev_attr_fw_pages,
2222 	&dev_attr_reg_pages,
2223 };
2224 
2225 static void pkey_change_handler(struct work_struct *work)
2226 {
2227 	struct mlx5_ib_port_resources *ports =
2228 		container_of(work, struct mlx5_ib_port_resources,
2229 			     pkey_change_work);
2230 
2231 	mutex_lock(&ports->devr->mutex);
2232 	mlx5_ib_gsi_pkey_change(ports->gsi);
2233 	mutex_unlock(&ports->devr->mutex);
2234 }
2235 
2236 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2237 {
2238 	struct mlx5_ib_qp *mqp;
2239 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2240 	struct mlx5_core_cq *mcq;
2241 	struct list_head cq_armed_list;
2242 	unsigned long flags_qp;
2243 	unsigned long flags_cq;
2244 	unsigned long flags;
2245 
2246 	INIT_LIST_HEAD(&cq_armed_list);
2247 
2248 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2249 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2250 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2251 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2252 		if (mqp->sq.tail != mqp->sq.head) {
2253 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2254 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2255 			if (send_mcq->mcq.comp &&
2256 			    mqp->ibqp.send_cq->comp_handler) {
2257 				if (!send_mcq->mcq.reset_notify_added) {
2258 					send_mcq->mcq.reset_notify_added = 1;
2259 					list_add_tail(&send_mcq->mcq.reset_notify,
2260 						      &cq_armed_list);
2261 				}
2262 			}
2263 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2264 		}
2265 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2266 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2267 		/* no handling is needed for SRQ */
2268 		if (!mqp->ibqp.srq) {
2269 			if (mqp->rq.tail != mqp->rq.head) {
2270 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2271 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2272 				if (recv_mcq->mcq.comp &&
2273 				    mqp->ibqp.recv_cq->comp_handler) {
2274 					if (!recv_mcq->mcq.reset_notify_added) {
2275 						recv_mcq->mcq.reset_notify_added = 1;
2276 						list_add_tail(&recv_mcq->mcq.reset_notify,
2277 							      &cq_armed_list);
2278 					}
2279 				}
2280 				spin_unlock_irqrestore(&recv_mcq->lock,
2281 						       flags_cq);
2282 			}
2283 		}
2284 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2285 	}
2286 	/*At that point all inflight post send were put to be executed as of we
2287 	 * lock/unlock above locks Now need to arm all involved CQs.
2288 	 */
2289 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2290 		mcq->comp(mcq);
2291 	}
2292 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2293 }
2294 
2295 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2296 			  enum mlx5_dev_event event, unsigned long param)
2297 {
2298 	struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2299 	struct ib_event ibev;
2300 	bool fatal = false;
2301 	u8 port = 0;
2302 
2303 	switch (event) {
2304 	case MLX5_DEV_EVENT_SYS_ERROR:
2305 		ibev.event = IB_EVENT_DEVICE_FATAL;
2306 		mlx5_ib_handle_internal_error(ibdev);
2307 		fatal = true;
2308 		break;
2309 
2310 	case MLX5_DEV_EVENT_PORT_UP:
2311 	case MLX5_DEV_EVENT_PORT_DOWN:
2312 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
2313 		port = (u8)param;
2314 
2315 		/* In RoCE, port up/down events are handled in
2316 		 * mlx5_netdev_event().
2317 		 */
2318 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2319 			IB_LINK_LAYER_ETHERNET)
2320 			return;
2321 
2322 		ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2323 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2324 		break;
2325 
2326 	case MLX5_DEV_EVENT_LID_CHANGE:
2327 		ibev.event = IB_EVENT_LID_CHANGE;
2328 		port = (u8)param;
2329 		break;
2330 
2331 	case MLX5_DEV_EVENT_PKEY_CHANGE:
2332 		ibev.event = IB_EVENT_PKEY_CHANGE;
2333 		port = (u8)param;
2334 
2335 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2336 		break;
2337 
2338 	case MLX5_DEV_EVENT_GUID_CHANGE:
2339 		ibev.event = IB_EVENT_GID_CHANGE;
2340 		port = (u8)param;
2341 		break;
2342 
2343 	case MLX5_DEV_EVENT_CLIENT_REREG:
2344 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
2345 		port = (u8)param;
2346 		break;
2347 
2348 	default:
2349 		break;
2350 	}
2351 
2352 	ibev.device	      = &ibdev->ib_dev;
2353 	ibev.element.port_num = port;
2354 
2355 	if (port < 1 || port > ibdev->num_ports) {
2356 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2357 		return;
2358 	}
2359 
2360 	if (ibdev->ib_active)
2361 		ib_dispatch_event(&ibev);
2362 
2363 	if (fatal)
2364 		ibdev->ib_active = false;
2365 }
2366 
2367 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2368 {
2369 	int port;
2370 
2371 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2372 		mlx5_query_ext_port_caps(dev, port);
2373 }
2374 
2375 static int get_port_caps(struct mlx5_ib_dev *dev)
2376 {
2377 	struct ib_device_attr *dprops = NULL;
2378 	struct ib_port_attr *pprops = NULL;
2379 	int err = -ENOMEM;
2380 	int port;
2381 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2382 
2383 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2384 	if (!pprops)
2385 		goto out;
2386 
2387 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2388 	if (!dprops)
2389 		goto out;
2390 
2391 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2392 	if (err) {
2393 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2394 		goto out;
2395 	}
2396 
2397 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2398 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2399 		if (err) {
2400 			mlx5_ib_warn(dev, "query_port %d failed %d\n",
2401 				     port, err);
2402 			break;
2403 		}
2404 		dev->mdev->port_caps[port - 1].pkey_table_len =
2405 						dprops->max_pkeys;
2406 		dev->mdev->port_caps[port - 1].gid_table_len =
2407 						pprops->gid_tbl_len;
2408 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2409 			    dprops->max_pkeys, pprops->gid_tbl_len);
2410 	}
2411 
2412 out:
2413 	kfree(pprops);
2414 	kfree(dprops);
2415 
2416 	return err;
2417 }
2418 
2419 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2420 {
2421 	int err;
2422 
2423 	err = mlx5_mr_cache_cleanup(dev);
2424 	if (err)
2425 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2426 
2427 	mlx5_ib_destroy_qp(dev->umrc.qp);
2428 	ib_free_cq(dev->umrc.cq);
2429 	ib_dealloc_pd(dev->umrc.pd);
2430 }
2431 
2432 enum {
2433 	MAX_UMR_WR = 128,
2434 };
2435 
2436 static int create_umr_res(struct mlx5_ib_dev *dev)
2437 {
2438 	struct ib_qp_init_attr *init_attr = NULL;
2439 	struct ib_qp_attr *attr = NULL;
2440 	struct ib_pd *pd;
2441 	struct ib_cq *cq;
2442 	struct ib_qp *qp;
2443 	int ret;
2444 
2445 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2446 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2447 	if (!attr || !init_attr) {
2448 		ret = -ENOMEM;
2449 		goto error_0;
2450 	}
2451 
2452 	pd = ib_alloc_pd(&dev->ib_dev, 0);
2453 	if (IS_ERR(pd)) {
2454 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2455 		ret = PTR_ERR(pd);
2456 		goto error_0;
2457 	}
2458 
2459 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2460 	if (IS_ERR(cq)) {
2461 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2462 		ret = PTR_ERR(cq);
2463 		goto error_2;
2464 	}
2465 
2466 	init_attr->send_cq = cq;
2467 	init_attr->recv_cq = cq;
2468 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2469 	init_attr->cap.max_send_wr = MAX_UMR_WR;
2470 	init_attr->cap.max_send_sge = 1;
2471 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2472 	init_attr->port_num = 1;
2473 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2474 	if (IS_ERR(qp)) {
2475 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2476 		ret = PTR_ERR(qp);
2477 		goto error_3;
2478 	}
2479 	qp->device     = &dev->ib_dev;
2480 	qp->real_qp    = qp;
2481 	qp->uobject    = NULL;
2482 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2483 
2484 	attr->qp_state = IB_QPS_INIT;
2485 	attr->port_num = 1;
2486 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2487 				IB_QP_PORT, NULL);
2488 	if (ret) {
2489 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2490 		goto error_4;
2491 	}
2492 
2493 	memset(attr, 0, sizeof(*attr));
2494 	attr->qp_state = IB_QPS_RTR;
2495 	attr->path_mtu = IB_MTU_256;
2496 
2497 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2498 	if (ret) {
2499 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2500 		goto error_4;
2501 	}
2502 
2503 	memset(attr, 0, sizeof(*attr));
2504 	attr->qp_state = IB_QPS_RTS;
2505 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2506 	if (ret) {
2507 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2508 		goto error_4;
2509 	}
2510 
2511 	dev->umrc.qp = qp;
2512 	dev->umrc.cq = cq;
2513 	dev->umrc.pd = pd;
2514 
2515 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
2516 	ret = mlx5_mr_cache_init(dev);
2517 	if (ret) {
2518 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2519 		goto error_4;
2520 	}
2521 
2522 	kfree(attr);
2523 	kfree(init_attr);
2524 
2525 	return 0;
2526 
2527 error_4:
2528 	mlx5_ib_destroy_qp(qp);
2529 
2530 error_3:
2531 	ib_free_cq(cq);
2532 
2533 error_2:
2534 	ib_dealloc_pd(pd);
2535 
2536 error_0:
2537 	kfree(attr);
2538 	kfree(init_attr);
2539 	return ret;
2540 }
2541 
2542 static int create_dev_resources(struct mlx5_ib_resources *devr)
2543 {
2544 	struct ib_srq_init_attr attr;
2545 	struct mlx5_ib_dev *dev;
2546 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2547 	int port;
2548 	int ret = 0;
2549 
2550 	dev = container_of(devr, struct mlx5_ib_dev, devr);
2551 
2552 	mutex_init(&devr->mutex);
2553 
2554 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2555 	if (IS_ERR(devr->p0)) {
2556 		ret = PTR_ERR(devr->p0);
2557 		goto error0;
2558 	}
2559 	devr->p0->device  = &dev->ib_dev;
2560 	devr->p0->uobject = NULL;
2561 	atomic_set(&devr->p0->usecnt, 0);
2562 
2563 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
2564 	if (IS_ERR(devr->c0)) {
2565 		ret = PTR_ERR(devr->c0);
2566 		goto error1;
2567 	}
2568 	devr->c0->device        = &dev->ib_dev;
2569 	devr->c0->uobject       = NULL;
2570 	devr->c0->comp_handler  = NULL;
2571 	devr->c0->event_handler = NULL;
2572 	devr->c0->cq_context    = NULL;
2573 	atomic_set(&devr->c0->usecnt, 0);
2574 
2575 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2576 	if (IS_ERR(devr->x0)) {
2577 		ret = PTR_ERR(devr->x0);
2578 		goto error2;
2579 	}
2580 	devr->x0->device = &dev->ib_dev;
2581 	devr->x0->inode = NULL;
2582 	atomic_set(&devr->x0->usecnt, 0);
2583 	mutex_init(&devr->x0->tgt_qp_mutex);
2584 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2585 
2586 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2587 	if (IS_ERR(devr->x1)) {
2588 		ret = PTR_ERR(devr->x1);
2589 		goto error3;
2590 	}
2591 	devr->x1->device = &dev->ib_dev;
2592 	devr->x1->inode = NULL;
2593 	atomic_set(&devr->x1->usecnt, 0);
2594 	mutex_init(&devr->x1->tgt_qp_mutex);
2595 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2596 
2597 	memset(&attr, 0, sizeof(attr));
2598 	attr.attr.max_sge = 1;
2599 	attr.attr.max_wr = 1;
2600 	attr.srq_type = IB_SRQT_XRC;
2601 	attr.ext.xrc.cq = devr->c0;
2602 	attr.ext.xrc.xrcd = devr->x0;
2603 
2604 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2605 	if (IS_ERR(devr->s0)) {
2606 		ret = PTR_ERR(devr->s0);
2607 		goto error4;
2608 	}
2609 	devr->s0->device	= &dev->ib_dev;
2610 	devr->s0->pd		= devr->p0;
2611 	devr->s0->uobject       = NULL;
2612 	devr->s0->event_handler = NULL;
2613 	devr->s0->srq_context   = NULL;
2614 	devr->s0->srq_type      = IB_SRQT_XRC;
2615 	devr->s0->ext.xrc.xrcd	= devr->x0;
2616 	devr->s0->ext.xrc.cq	= devr->c0;
2617 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2618 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2619 	atomic_inc(&devr->p0->usecnt);
2620 	atomic_set(&devr->s0->usecnt, 0);
2621 
2622 	memset(&attr, 0, sizeof(attr));
2623 	attr.attr.max_sge = 1;
2624 	attr.attr.max_wr = 1;
2625 	attr.srq_type = IB_SRQT_BASIC;
2626 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2627 	if (IS_ERR(devr->s1)) {
2628 		ret = PTR_ERR(devr->s1);
2629 		goto error5;
2630 	}
2631 	devr->s1->device	= &dev->ib_dev;
2632 	devr->s1->pd		= devr->p0;
2633 	devr->s1->uobject       = NULL;
2634 	devr->s1->event_handler = NULL;
2635 	devr->s1->srq_context   = NULL;
2636 	devr->s1->srq_type      = IB_SRQT_BASIC;
2637 	devr->s1->ext.xrc.cq	= devr->c0;
2638 	atomic_inc(&devr->p0->usecnt);
2639 	atomic_set(&devr->s0->usecnt, 0);
2640 
2641 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2642 		INIT_WORK(&devr->ports[port].pkey_change_work,
2643 			  pkey_change_handler);
2644 		devr->ports[port].devr = devr;
2645 	}
2646 
2647 	return 0;
2648 
2649 error5:
2650 	mlx5_ib_destroy_srq(devr->s0);
2651 error4:
2652 	mlx5_ib_dealloc_xrcd(devr->x1);
2653 error3:
2654 	mlx5_ib_dealloc_xrcd(devr->x0);
2655 error2:
2656 	mlx5_ib_destroy_cq(devr->c0);
2657 error1:
2658 	mlx5_ib_dealloc_pd(devr->p0);
2659 error0:
2660 	return ret;
2661 }
2662 
2663 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2664 {
2665 	struct mlx5_ib_dev *dev =
2666 		container_of(devr, struct mlx5_ib_dev, devr);
2667 	int port;
2668 
2669 	mlx5_ib_destroy_srq(devr->s1);
2670 	mlx5_ib_destroy_srq(devr->s0);
2671 	mlx5_ib_dealloc_xrcd(devr->x0);
2672 	mlx5_ib_dealloc_xrcd(devr->x1);
2673 	mlx5_ib_destroy_cq(devr->c0);
2674 	mlx5_ib_dealloc_pd(devr->p0);
2675 
2676 	/* Make sure no change P_Key work items are still executing */
2677 	for (port = 0; port < dev->num_ports; ++port)
2678 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2679 }
2680 
2681 static u32 get_core_cap_flags(struct ib_device *ibdev)
2682 {
2683 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2684 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2685 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2686 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2687 	u32 ret = 0;
2688 
2689 	if (ll == IB_LINK_LAYER_INFINIBAND)
2690 		return RDMA_CORE_PORT_IBA_IB;
2691 
2692 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2693 		return 0;
2694 
2695 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2696 		return 0;
2697 
2698 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2699 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2700 
2701 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2702 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2703 
2704 	return ret;
2705 }
2706 
2707 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
2708 			       struct ib_port_immutable *immutable)
2709 {
2710 	struct ib_port_attr attr;
2711 	int err;
2712 
2713 	err = mlx5_ib_query_port(ibdev, port_num, &attr);
2714 	if (err)
2715 		return err;
2716 
2717 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2718 	immutable->gid_tbl_len = attr.gid_tbl_len;
2719 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
2720 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2721 
2722 	return 0;
2723 }
2724 
2725 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
2726 			   size_t str_len)
2727 {
2728 	struct mlx5_ib_dev *dev =
2729 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2730 	snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
2731 		       fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
2732 }
2733 
2734 static int mlx5_roce_lag_init(struct mlx5_ib_dev *dev)
2735 {
2736 	return 0;
2737 }
2738 
2739 static void mlx5_roce_lag_cleanup(struct mlx5_ib_dev *dev)
2740 {
2741 }
2742 
2743 static void mlx5_remove_roce_notifier(struct mlx5_ib_dev *dev)
2744 {
2745 	if (dev->roce.nb.notifier_call) {
2746 		unregister_netdevice_notifier(&dev->roce.nb);
2747 		dev->roce.nb.notifier_call = NULL;
2748 	}
2749 }
2750 
2751 static int mlx5_enable_roce(struct mlx5_ib_dev *dev)
2752 {
2753 	VNET_ITERATOR_DECL(vnet_iter);
2754 	struct net_device *idev;
2755 	int err;
2756 
2757 	/* Check if mlx5en net device already exists */
2758 	VNET_LIST_RLOCK();
2759 	VNET_FOREACH(vnet_iter) {
2760 		IFNET_RLOCK();
2761 		CURVNET_SET_QUIET(vnet_iter);
2762 		TAILQ_FOREACH(idev, &V_ifnet, if_link) {
2763 			/* check if network interface belongs to mlx5en */
2764 			if (!mlx5_netdev_match(idev, dev->mdev, "mce"))
2765 				continue;
2766 			write_lock(&dev->roce.netdev_lock);
2767 			dev->roce.netdev = idev;
2768 			write_unlock(&dev->roce.netdev_lock);
2769 		}
2770 		CURVNET_RESTORE();
2771 		IFNET_RUNLOCK();
2772 	}
2773 	VNET_LIST_RUNLOCK();
2774 
2775 	dev->roce.nb.notifier_call = mlx5_netdev_event;
2776 	err = register_netdevice_notifier(&dev->roce.nb);
2777 	if (err) {
2778 		dev->roce.nb.notifier_call = NULL;
2779 		return err;
2780 	}
2781 
2782 	err = mlx5_nic_vport_enable_roce(dev->mdev);
2783 	if (err)
2784 		goto err_unregister_netdevice_notifier;
2785 
2786 	err = mlx5_roce_lag_init(dev);
2787 	if (err)
2788 		goto err_disable_roce;
2789 
2790 	return 0;
2791 
2792 err_disable_roce:
2793 	mlx5_nic_vport_disable_roce(dev->mdev);
2794 
2795 err_unregister_netdevice_notifier:
2796 	mlx5_remove_roce_notifier(dev);
2797 	return err;
2798 }
2799 
2800 static void mlx5_disable_roce(struct mlx5_ib_dev *dev)
2801 {
2802 	mlx5_roce_lag_cleanup(dev);
2803 	mlx5_nic_vport_disable_roce(dev->mdev);
2804 }
2805 
2806 static void mlx5_ib_dealloc_q_port_counter(struct mlx5_ib_dev *dev, u8 port_num)
2807 {
2808 	mlx5_vport_dealloc_q_counter(dev->mdev,
2809 				     MLX5_INTERFACE_PROTOCOL_IB,
2810 				     dev->port[port_num].q_cnt_id);
2811 	dev->port[port_num].q_cnt_id = 0;
2812 }
2813 
2814 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
2815 {
2816 	unsigned int i;
2817 
2818 	for (i = 0; i < dev->num_ports; i++)
2819 		mlx5_ib_dealloc_q_port_counter(dev, i);
2820 }
2821 
2822 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
2823 {
2824 	int i;
2825 	int ret;
2826 
2827 	for (i = 0; i < dev->num_ports; i++) {
2828 		ret = mlx5_vport_alloc_q_counter(dev->mdev,
2829 						 MLX5_INTERFACE_PROTOCOL_IB,
2830 						 &dev->port[i].q_cnt_id);
2831 		if (ret) {
2832 			mlx5_ib_warn(dev,
2833 				     "couldn't allocate queue counter for port %d, err %d\n",
2834 				     i + 1, ret);
2835 			goto dealloc_counters;
2836 		}
2837 	}
2838 
2839 	return 0;
2840 
2841 dealloc_counters:
2842 	while (--i >= 0)
2843 		mlx5_ib_dealloc_q_port_counter(dev, i);
2844 
2845 	return ret;
2846 }
2847 
2848 static const char * const names[] = {
2849 	"rx_write_requests",
2850 	"rx_read_requests",
2851 	"rx_atomic_requests",
2852 	"out_of_buffer",
2853 	"out_of_sequence",
2854 	"duplicate_request",
2855 	"rnr_nak_retry_err",
2856 	"packet_seq_err",
2857 	"implied_nak_seq_err",
2858 	"local_ack_timeout_err",
2859 };
2860 
2861 static const size_t stats_offsets[] = {
2862 	MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests),
2863 	MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests),
2864 	MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests),
2865 	MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer),
2866 	MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence),
2867 	MLX5_BYTE_OFF(query_q_counter_out, duplicate_request),
2868 	MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err),
2869 	MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err),
2870 	MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err),
2871 	MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err),
2872 };
2873 
2874 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
2875 						    u8 port_num)
2876 {
2877 	BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets));
2878 
2879 	/* We support only per port stats */
2880 	if (port_num == 0)
2881 		return NULL;
2882 
2883 	return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names),
2884 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
2885 }
2886 
2887 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
2888 				struct rdma_hw_stats *stats,
2889 				u8 port, int index)
2890 {
2891 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2892 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
2893 	void *out;
2894 	__be32 val;
2895 	int ret;
2896 	int i;
2897 
2898 	if (!port || !stats)
2899 		return -ENOSYS;
2900 
2901 	out = mlx5_vzalloc(outlen);
2902 	if (!out)
2903 		return -ENOMEM;
2904 
2905 	ret = mlx5_vport_query_q_counter(dev->mdev,
2906 					dev->port[port - 1].q_cnt_id, 0,
2907 					out, outlen);
2908 	if (ret)
2909 		goto free;
2910 
2911 	for (i = 0; i < ARRAY_SIZE(names); i++) {
2912 		val = *(__be32 *)(out + stats_offsets[i]);
2913 		stats->value[i] = (u64)be32_to_cpu(val);
2914 	}
2915 free:
2916 	kvfree(out);
2917 	return ARRAY_SIZE(names);
2918 }
2919 
2920 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
2921 {
2922 	struct mlx5_ib_dev *dev;
2923 	enum rdma_link_layer ll;
2924 	int port_type_cap;
2925 	const char *name;
2926 	int err;
2927 	int i;
2928 
2929 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
2930 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
2931 
2932 	if ((ll == IB_LINK_LAYER_ETHERNET) && !MLX5_CAP_GEN(mdev, roce))
2933 		return NULL;
2934 
2935 	printk_once(KERN_INFO "%s", mlx5_version);
2936 
2937 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
2938 	if (!dev)
2939 		return NULL;
2940 
2941 	dev->mdev = mdev;
2942 
2943 	dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
2944 			    GFP_KERNEL);
2945 	if (!dev->port)
2946 		goto err_dealloc;
2947 
2948 	rwlock_init(&dev->roce.netdev_lock);
2949 	err = get_port_caps(dev);
2950 	if (err)
2951 		goto err_free_port;
2952 
2953 	if (mlx5_use_mad_ifc(dev))
2954 		get_ext_port_caps(dev);
2955 
2956 	MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
2957 
2958 	name = "mlx5_%d";
2959 
2960 	strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
2961 	dev->ib_dev.owner		= THIS_MODULE;
2962 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
2963 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
2964 	dev->num_ports		= MLX5_CAP_GEN(mdev, num_ports);
2965 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
2966 	dev->ib_dev.num_comp_vectors    =
2967 		dev->mdev->priv.eq_table.num_comp_vectors;
2968 	dev->ib_dev.dma_device	= &mdev->pdev->dev;
2969 
2970 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
2971 	dev->ib_dev.uverbs_cmd_mask	=
2972 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
2973 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
2974 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
2975 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
2976 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
2977 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
2978 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
2979 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
2980 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
2981 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
2982 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
2983 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
2984 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
2985 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
2986 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
2987 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
2988 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
2989 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
2990 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
2991 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
2992 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
2993 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
2994 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
2995 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
2996 	dev->ib_dev.uverbs_ex_cmd_mask =
2997 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
2998 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
2999 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
3000 
3001 	dev->ib_dev.query_device	= mlx5_ib_query_device;
3002 	dev->ib_dev.query_port		= mlx5_ib_query_port;
3003 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
3004 	if (ll == IB_LINK_LAYER_ETHERNET)
3005 		dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
3006 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
3007 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
3008 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
3009 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
3010 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
3011 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
3012 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
3013 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
3014 	dev->ib_dev.mmap		= mlx5_ib_mmap;
3015 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
3016 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
3017 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
3018 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
3019 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
3020 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
3021 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
3022 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
3023 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
3024 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
3025 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
3026 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
3027 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
3028 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
3029 	dev->ib_dev.post_send		= mlx5_ib_post_send;
3030 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
3031 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
3032 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
3033 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
3034 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
3035 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
3036 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
3037 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
3038 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
3039 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
3040 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
3041 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
3042 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
3043 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
3044 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
3045 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
3046 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
3047 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
3048 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
3049 	if (mlx5_core_is_pf(mdev)) {
3050 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
3051 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
3052 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
3053 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
3054 	}
3055 
3056 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
3057 
3058 	mlx5_ib_internal_fill_odp_caps(dev);
3059 
3060 	if (MLX5_CAP_GEN(mdev, imaicl)) {
3061 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
3062 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
3063 		dev->ib_dev.uverbs_cmd_mask |=
3064 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
3065 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
3066 	}
3067 
3068 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) &&
3069 	    MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3070 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
3071 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
3072 	}
3073 
3074 	if (MLX5_CAP_GEN(mdev, xrc)) {
3075 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
3076 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
3077 		dev->ib_dev.uverbs_cmd_mask |=
3078 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
3079 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
3080 	}
3081 
3082 	if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
3083 	    IB_LINK_LAYER_ETHERNET) {
3084 		dev->ib_dev.create_flow	= mlx5_ib_create_flow;
3085 		dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
3086 		dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
3087 		dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
3088 		dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
3089 		dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
3090 		dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
3091 		dev->ib_dev.uverbs_ex_cmd_mask |=
3092 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
3093 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
3094 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
3095 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
3096 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
3097 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
3098 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
3099 	}
3100 	err = init_node_data(dev);
3101 	if (err)
3102 		goto err_free_port;
3103 
3104 	mutex_init(&dev->flow_db.lock);
3105 	mutex_init(&dev->cap_mask_mutex);
3106 	INIT_LIST_HEAD(&dev->qp_list);
3107 	spin_lock_init(&dev->reset_flow_resource_lock);
3108 
3109 	if (ll == IB_LINK_LAYER_ETHERNET) {
3110 		err = mlx5_enable_roce(dev);
3111 		if (err)
3112 			goto err_free_port;
3113 	}
3114 
3115 	err = create_dev_resources(&dev->devr);
3116 	if (err)
3117 		goto err_disable_roce;
3118 
3119 	err = mlx5_ib_odp_init_one(dev);
3120 	if (err)
3121 		goto err_rsrc;
3122 
3123 	err = mlx5_ib_alloc_q_counters(dev);
3124 	if (err)
3125 		goto err_odp;
3126 
3127 	err = ib_register_device(&dev->ib_dev, NULL);
3128 	if (err)
3129 		goto err_q_cnt;
3130 
3131 	err = create_umr_res(dev);
3132 	if (err)
3133 		goto err_dev;
3134 
3135 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
3136 		err = device_create_file(&dev->ib_dev.dev,
3137 					 mlx5_class_attributes[i]);
3138 		if (err)
3139 			goto err_umrc;
3140 	}
3141 
3142 	dev->ib_active = true;
3143 
3144 	return dev;
3145 
3146 err_umrc:
3147 	destroy_umrc_res(dev);
3148 
3149 err_dev:
3150 	ib_unregister_device(&dev->ib_dev);
3151 
3152 err_q_cnt:
3153 	mlx5_ib_dealloc_q_counters(dev);
3154 
3155 err_odp:
3156 	mlx5_ib_odp_remove_one(dev);
3157 
3158 err_rsrc:
3159 	destroy_dev_resources(&dev->devr);
3160 
3161 err_disable_roce:
3162 	if (ll == IB_LINK_LAYER_ETHERNET) {
3163 		mlx5_disable_roce(dev);
3164 		mlx5_remove_roce_notifier(dev);
3165 	}
3166 
3167 err_free_port:
3168 	kfree(dev->port);
3169 
3170 err_dealloc:
3171 	ib_dealloc_device((struct ib_device *)dev);
3172 
3173 	return NULL;
3174 }
3175 
3176 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3177 {
3178 	struct mlx5_ib_dev *dev = context;
3179 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3180 
3181 	mlx5_remove_roce_notifier(dev);
3182 	ib_unregister_device(&dev->ib_dev);
3183 	mlx5_ib_dealloc_q_counters(dev);
3184 	destroy_umrc_res(dev);
3185 	mlx5_ib_odp_remove_one(dev);
3186 	destroy_dev_resources(&dev->devr);
3187 	if (ll == IB_LINK_LAYER_ETHERNET)
3188 		mlx5_disable_roce(dev);
3189 	kfree(dev->port);
3190 	ib_dealloc_device(&dev->ib_dev);
3191 }
3192 
3193 static struct mlx5_interface mlx5_ib_interface = {
3194 	.add            = mlx5_ib_add,
3195 	.remove         = mlx5_ib_remove,
3196 	.event          = mlx5_ib_event,
3197 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
3198 };
3199 
3200 static int __init mlx5_ib_init(void)
3201 {
3202 	int err;
3203 
3204 	if (deprecated_prof_sel != 2)
3205 		pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
3206 
3207 	err = mlx5_ib_odp_init();
3208 	if (err)
3209 		return err;
3210 
3211 	err = mlx5_register_interface(&mlx5_ib_interface);
3212 	if (err)
3213 		goto clean_odp;
3214 
3215 	return err;
3216 
3217 clean_odp:
3218 	mlx5_ib_odp_cleanup();
3219 	return err;
3220 }
3221 
3222 static void __exit mlx5_ib_cleanup(void)
3223 {
3224 	mlx5_unregister_interface(&mlx5_ib_interface);
3225 	mlx5_ib_odp_cleanup();
3226 }
3227 
3228 module_init_order(mlx5_ib_init, SI_ORDER_THIRD);
3229 module_exit_order(mlx5_ib_cleanup, SI_ORDER_THIRD);
3230