xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_main.c (revision 2284664ef9fcb0baaf59f1ef7df877c0b0f2b187)
1 /*-
2  * Copyright (c) 2013-2015, Mellanox Technologies, Ltd.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  * $FreeBSD$
26  */
27 
28 #include <linux/module.h>
29 #include <linux/errno.h>
30 #include <linux/pci.h>
31 #include <linux/dma-mapping.h>
32 #include <linux/slab.h>
33 #if defined(CONFIG_X86)
34 #include <asm/pat.h>
35 #endif
36 #include <linux/sched.h>
37 #include <linux/delay.h>
38 #include <linux/fs.h>
39 #undef inode
40 #include <rdma/ib_user_verbs.h>
41 #include <rdma/ib_addr.h>
42 #include <rdma/ib_cache.h>
43 #include <dev/mlx5/port.h>
44 #include <dev/mlx5/vport.h>
45 #include <linux/list.h>
46 #include <rdma/ib_smi.h>
47 #include <rdma/ib_umem.h>
48 #include <linux/in.h>
49 #include <linux/etherdevice.h>
50 #include <dev/mlx5/fs.h>
51 #include "mlx5_ib.h"
52 
53 #define DRIVER_NAME "mlx5_ib"
54 #ifndef DRIVER_VERSION
55 #define DRIVER_VERSION "3.4.2"
56 #endif
57 #define DRIVER_RELDATE	"July 2018"
58 
59 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
60 MODULE_LICENSE("Dual BSD/GPL");
61 MODULE_DEPEND(mlx5ib, linuxkpi, 1, 1, 1);
62 MODULE_DEPEND(mlx5ib, mlx5, 1, 1, 1);
63 MODULE_DEPEND(mlx5ib, ibcore, 1, 1, 1);
64 MODULE_VERSION(mlx5ib, 1);
65 
66 static int deprecated_prof_sel = 2;
67 module_param_named(prof_sel, deprecated_prof_sel, int, 0444);
68 MODULE_PARM_DESC(prof_sel, "profile selector. Deprecated here. Moved to module mlx5_core");
69 
70 static const char mlx5_version[] =
71 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver "
72 	DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
73 
74 enum {
75 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
76 };
77 
78 static enum rdma_link_layer
79 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
80 {
81 	switch (port_type_cap) {
82 	case MLX5_CAP_PORT_TYPE_IB:
83 		return IB_LINK_LAYER_INFINIBAND;
84 	case MLX5_CAP_PORT_TYPE_ETH:
85 		return IB_LINK_LAYER_ETHERNET;
86 	default:
87 		return IB_LINK_LAYER_UNSPECIFIED;
88 	}
89 }
90 
91 static enum rdma_link_layer
92 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
93 {
94 	struct mlx5_ib_dev *dev = to_mdev(device);
95 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
96 
97 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
98 }
99 
100 static bool mlx5_netdev_match(struct net_device *ndev,
101 			      struct mlx5_core_dev *mdev,
102 			      const char *dname)
103 {
104 	return ndev->if_type == IFT_ETHER &&
105 	  ndev->if_dname != NULL &&
106 	  strcmp(ndev->if_dname, dname) == 0 &&
107 	  ndev->if_softc != NULL &&
108 	  *(struct mlx5_core_dev **)ndev->if_softc == mdev;
109 }
110 
111 static int mlx5_netdev_event(struct notifier_block *this,
112 			     unsigned long event, void *ptr)
113 {
114 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
115 	struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
116 						 roce.nb);
117 
118 	switch (event) {
119 	case NETDEV_REGISTER:
120 	case NETDEV_UNREGISTER:
121 		write_lock(&ibdev->roce.netdev_lock);
122 		/* check if network interface belongs to mlx5en */
123 		if (mlx5_netdev_match(ndev, ibdev->mdev, "mce"))
124 			ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
125 					     NULL : ndev;
126 		write_unlock(&ibdev->roce.netdev_lock);
127 		break;
128 
129 	case NETDEV_UP:
130 	case NETDEV_DOWN: {
131 		struct net_device *upper = NULL;
132 
133 		if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
134 		    && ibdev->ib_active) {
135 			struct ib_event ibev = {0};
136 
137 			ibev.device = &ibdev->ib_dev;
138 			ibev.event = (event == NETDEV_UP) ?
139 				     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
140 			ibev.element.port_num = 1;
141 			ib_dispatch_event(&ibev);
142 		}
143 		break;
144 	}
145 
146 	default:
147 		break;
148 	}
149 
150 	return NOTIFY_DONE;
151 }
152 
153 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
154 					     u8 port_num)
155 {
156 	struct mlx5_ib_dev *ibdev = to_mdev(device);
157 	struct net_device *ndev;
158 
159 	/* Ensure ndev does not disappear before we invoke dev_hold()
160 	 */
161 	read_lock(&ibdev->roce.netdev_lock);
162 	ndev = ibdev->roce.netdev;
163 	if (ndev)
164 		dev_hold(ndev);
165 	read_unlock(&ibdev->roce.netdev_lock);
166 
167 	return ndev;
168 }
169 
170 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
171 				    u8 *active_width)
172 {
173 	switch (eth_proto_oper) {
174 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
175 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
176 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
177 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
178 		*active_width = IB_WIDTH_1X;
179 		*active_speed = IB_SPEED_SDR;
180 		break;
181 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
182 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
183 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
184 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
185 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
186 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
187 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
188 		*active_width = IB_WIDTH_1X;
189 		*active_speed = IB_SPEED_QDR;
190 		break;
191 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
192 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
193 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
194 		*active_width = IB_WIDTH_1X;
195 		*active_speed = IB_SPEED_EDR;
196 		break;
197 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
198 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
199 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
200 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
201 		*active_width = IB_WIDTH_4X;
202 		*active_speed = IB_SPEED_QDR;
203 		break;
204 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
205 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
206 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
207 		*active_width = IB_WIDTH_1X;
208 		*active_speed = IB_SPEED_HDR;
209 		break;
210 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
211 		*active_width = IB_WIDTH_4X;
212 		*active_speed = IB_SPEED_FDR;
213 		break;
214 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
215 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
216 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
217 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
218 		*active_width = IB_WIDTH_4X;
219 		*active_speed = IB_SPEED_EDR;
220 		break;
221 	default:
222 		return -EINVAL;
223 	}
224 
225 	return 0;
226 }
227 
228 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
229 				struct ib_port_attr *props)
230 {
231 	struct mlx5_ib_dev *dev = to_mdev(device);
232 	struct net_device *ndev;
233 	enum ib_mtu ndev_ib_mtu;
234 	u16 qkey_viol_cntr;
235 	u32 eth_prot_oper;
236 	int err;
237 
238 	memset(props, 0, sizeof(*props));
239 
240 	/* Possible bad flows are checked before filling out props so in case
241 	 * of an error it will still be zeroed out.
242 	 */
243 	err = mlx5_query_port_eth_proto_oper(dev->mdev, &eth_prot_oper, port_num);
244 	if (err)
245 		return err;
246 
247 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
248 				 &props->active_width);
249 
250 	props->port_cap_flags  |= IB_PORT_CM_SUP;
251 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
252 
253 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
254 						roce_address_table_size);
255 	props->max_mtu          = IB_MTU_4096;
256 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
257 	props->pkey_tbl_len     = 1;
258 	props->state            = IB_PORT_DOWN;
259 	props->phys_state       = 3;
260 
261 	mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
262 	props->qkey_viol_cntr = qkey_viol_cntr;
263 
264 	ndev = mlx5_ib_get_netdev(device, port_num);
265 	if (!ndev)
266 		return 0;
267 
268 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
269 		props->state      = IB_PORT_ACTIVE;
270 		props->phys_state = 5;
271 	}
272 
273 	ndev_ib_mtu = iboe_get_mtu(ndev->if_mtu);
274 
275 	dev_put(ndev);
276 
277 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
278 	return 0;
279 }
280 
281 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
282 				     const struct ib_gid_attr *attr,
283 				     void *mlx5_addr)
284 {
285 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
286 	char *mlx5_addr_l3_addr	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
287 					       source_l3_address);
288 	void *mlx5_addr_mac	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
289 					       source_mac_47_32);
290 	u16 vlan_id;
291 
292 	if (!gid)
293 		return;
294 	ether_addr_copy(mlx5_addr_mac, IF_LLADDR(attr->ndev));
295 
296 	vlan_id = rdma_vlan_dev_vlan_id(attr->ndev);
297 	if (vlan_id != 0xffff) {
298 		MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
299 		MLX5_SET_RA(mlx5_addr, vlan_id, vlan_id);
300 	}
301 
302 	switch (attr->gid_type) {
303 	case IB_GID_TYPE_IB:
304 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
305 		break;
306 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
307 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
308 		break;
309 
310 	default:
311 		WARN_ON(true);
312 	}
313 
314 	if (attr->gid_type != IB_GID_TYPE_IB) {
315 		if (ipv6_addr_v4mapped((void *)gid))
316 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
317 				    MLX5_ROCE_L3_TYPE_IPV4);
318 		else
319 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
320 				    MLX5_ROCE_L3_TYPE_IPV6);
321 	}
322 
323 	if ((attr->gid_type == IB_GID_TYPE_IB) ||
324 	    !ipv6_addr_v4mapped((void *)gid))
325 		memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
326 	else
327 		memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
328 }
329 
330 static int set_roce_addr(struct ib_device *device, u8 port_num,
331 			 unsigned int index,
332 			 const union ib_gid *gid,
333 			 const struct ib_gid_attr *attr)
334 {
335 	struct mlx5_ib_dev *dev = to_mdev(device);
336 	u32  in[MLX5_ST_SZ_DW(set_roce_address_in)]  = {0};
337 	u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0};
338 	void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
339 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
340 
341 	if (ll != IB_LINK_LAYER_ETHERNET)
342 		return -EINVAL;
343 
344 	ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
345 
346 	MLX5_SET(set_roce_address_in, in, roce_address_index, index);
347 	MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
348 	return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
349 }
350 
351 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
352 			   unsigned int index, const union ib_gid *gid,
353 			   const struct ib_gid_attr *attr,
354 			   __always_unused void **context)
355 {
356 	return set_roce_addr(device, port_num, index, gid, attr);
357 }
358 
359 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
360 			   unsigned int index, __always_unused void **context)
361 {
362 	return set_roce_addr(device, port_num, index, NULL, NULL);
363 }
364 
365 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
366 			       int index)
367 {
368 	struct ib_gid_attr attr;
369 	union ib_gid gid;
370 
371 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
372 		return 0;
373 
374 	if (!attr.ndev)
375 		return 0;
376 
377 	dev_put(attr.ndev);
378 
379 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
380 		return 0;
381 
382 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
383 }
384 
385 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
386 			   int index, enum ib_gid_type *gid_type)
387 {
388 	struct ib_gid_attr attr;
389 	union ib_gid gid;
390 	int ret;
391 
392 	ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
393 	if (ret)
394 		return ret;
395 
396 	if (!attr.ndev)
397 		return -ENODEV;
398 
399 	dev_put(attr.ndev);
400 
401 	*gid_type = attr.gid_type;
402 
403 	return 0;
404 }
405 
406 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
407 {
408 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
409 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
410 	return 0;
411 }
412 
413 enum {
414 	MLX5_VPORT_ACCESS_METHOD_MAD,
415 	MLX5_VPORT_ACCESS_METHOD_HCA,
416 	MLX5_VPORT_ACCESS_METHOD_NIC,
417 };
418 
419 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
420 {
421 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
422 		return MLX5_VPORT_ACCESS_METHOD_MAD;
423 
424 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
425 	    IB_LINK_LAYER_ETHERNET)
426 		return MLX5_VPORT_ACCESS_METHOD_NIC;
427 
428 	return MLX5_VPORT_ACCESS_METHOD_HCA;
429 }
430 
431 static void get_atomic_caps(struct mlx5_ib_dev *dev,
432 			    struct ib_device_attr *props)
433 {
434 	u8 tmp;
435 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
436 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
437 	u8 atomic_req_8B_endianness_mode =
438 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
439 
440 	/* Check if HW supports 8 bytes standard atomic operations and capable
441 	 * of host endianness respond
442 	 */
443 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
444 	if (((atomic_operations & tmp) == tmp) &&
445 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
446 	    (atomic_req_8B_endianness_mode)) {
447 		props->atomic_cap = IB_ATOMIC_HCA;
448 	} else {
449 		props->atomic_cap = IB_ATOMIC_NONE;
450 	}
451 }
452 
453 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
454 					__be64 *sys_image_guid)
455 {
456 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
457 	struct mlx5_core_dev *mdev = dev->mdev;
458 	u64 tmp;
459 	int err;
460 
461 	switch (mlx5_get_vport_access_method(ibdev)) {
462 	case MLX5_VPORT_ACCESS_METHOD_MAD:
463 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
464 							    sys_image_guid);
465 
466 	case MLX5_VPORT_ACCESS_METHOD_HCA:
467 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
468 		break;
469 
470 	case MLX5_VPORT_ACCESS_METHOD_NIC:
471 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
472 		break;
473 
474 	default:
475 		return -EINVAL;
476 	}
477 
478 	if (!err)
479 		*sys_image_guid = cpu_to_be64(tmp);
480 
481 	return err;
482 
483 }
484 
485 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
486 				u16 *max_pkeys)
487 {
488 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
489 	struct mlx5_core_dev *mdev = dev->mdev;
490 
491 	switch (mlx5_get_vport_access_method(ibdev)) {
492 	case MLX5_VPORT_ACCESS_METHOD_MAD:
493 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
494 
495 	case MLX5_VPORT_ACCESS_METHOD_HCA:
496 	case MLX5_VPORT_ACCESS_METHOD_NIC:
497 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
498 						pkey_table_size));
499 		return 0;
500 
501 	default:
502 		return -EINVAL;
503 	}
504 }
505 
506 static int mlx5_query_vendor_id(struct ib_device *ibdev,
507 				u32 *vendor_id)
508 {
509 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
510 
511 	switch (mlx5_get_vport_access_method(ibdev)) {
512 	case MLX5_VPORT_ACCESS_METHOD_MAD:
513 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
514 
515 	case MLX5_VPORT_ACCESS_METHOD_HCA:
516 	case MLX5_VPORT_ACCESS_METHOD_NIC:
517 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
518 
519 	default:
520 		return -EINVAL;
521 	}
522 }
523 
524 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
525 				__be64 *node_guid)
526 {
527 	u64 tmp;
528 	int err;
529 
530 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
531 	case MLX5_VPORT_ACCESS_METHOD_MAD:
532 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
533 
534 	case MLX5_VPORT_ACCESS_METHOD_HCA:
535 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
536 		break;
537 
538 	case MLX5_VPORT_ACCESS_METHOD_NIC:
539 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
540 		break;
541 
542 	default:
543 		return -EINVAL;
544 	}
545 
546 	if (!err)
547 		*node_guid = cpu_to_be64(tmp);
548 
549 	return err;
550 }
551 
552 struct mlx5_reg_node_desc {
553 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
554 };
555 
556 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
557 {
558 	struct mlx5_reg_node_desc in;
559 
560 	if (mlx5_use_mad_ifc(dev))
561 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
562 
563 	memset(&in, 0, sizeof(in));
564 
565 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
566 				    sizeof(struct mlx5_reg_node_desc),
567 				    MLX5_REG_NODE_DESC, 0, 0);
568 }
569 
570 static int mlx5_ib_query_device(struct ib_device *ibdev,
571 				struct ib_device_attr *props,
572 				struct ib_udata *uhw)
573 {
574 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
575 	struct mlx5_core_dev *mdev = dev->mdev;
576 	int err = -ENOMEM;
577 	int max_rq_sg;
578 	int max_sq_sg;
579 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
580 	struct mlx5_ib_query_device_resp resp = {};
581 	size_t resp_len;
582 	u64 max_tso;
583 
584 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
585 	if (uhw->outlen && uhw->outlen < resp_len)
586 		return -EINVAL;
587 	else
588 		resp.response_length = resp_len;
589 
590 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
591 		return -EINVAL;
592 
593 	memset(props, 0, sizeof(*props));
594 	err = mlx5_query_system_image_guid(ibdev,
595 					   &props->sys_image_guid);
596 	if (err)
597 		return err;
598 
599 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
600 	if (err)
601 		return err;
602 
603 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
604 	if (err)
605 		return err;
606 
607 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
608 		(fw_rev_min(dev->mdev) << 16) |
609 		fw_rev_sub(dev->mdev);
610 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
611 		IB_DEVICE_PORT_ACTIVE_EVENT		|
612 		IB_DEVICE_SYS_IMAGE_GUID		|
613 		IB_DEVICE_RC_RNR_NAK_GEN;
614 
615 	if (MLX5_CAP_GEN(mdev, pkv))
616 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
617 	if (MLX5_CAP_GEN(mdev, qkv))
618 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
619 	if (MLX5_CAP_GEN(mdev, apm))
620 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
621 	if (MLX5_CAP_GEN(mdev, xrc))
622 		props->device_cap_flags |= IB_DEVICE_XRC;
623 	if (MLX5_CAP_GEN(mdev, imaicl)) {
624 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
625 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
626 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
627 		/* We support 'Gappy' memory registration too */
628 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
629 	}
630 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
631 	if (MLX5_CAP_GEN(mdev, sho)) {
632 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
633 		/* At this stage no support for signature handover */
634 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
635 				      IB_PROT_T10DIF_TYPE_2 |
636 				      IB_PROT_T10DIF_TYPE_3;
637 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
638 				       IB_GUARD_T10DIF_CSUM;
639 	}
640 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
641 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
642 
643 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
644 		if (MLX5_CAP_ETH(mdev, csum_cap))
645 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
646 
647 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
648 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
649 			if (max_tso) {
650 				resp.tso_caps.max_tso = 1 << max_tso;
651 				resp.tso_caps.supported_qpts |=
652 					1 << IB_QPT_RAW_PACKET;
653 				resp.response_length += sizeof(resp.tso_caps);
654 			}
655 		}
656 
657 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
658 			resp.rss_caps.rx_hash_function =
659 						MLX5_RX_HASH_FUNC_TOEPLITZ;
660 			resp.rss_caps.rx_hash_fields_mask =
661 						MLX5_RX_HASH_SRC_IPV4 |
662 						MLX5_RX_HASH_DST_IPV4 |
663 						MLX5_RX_HASH_SRC_IPV6 |
664 						MLX5_RX_HASH_DST_IPV6 |
665 						MLX5_RX_HASH_SRC_PORT_TCP |
666 						MLX5_RX_HASH_DST_PORT_TCP |
667 						MLX5_RX_HASH_SRC_PORT_UDP |
668 						MLX5_RX_HASH_DST_PORT_UDP;
669 			resp.response_length += sizeof(resp.rss_caps);
670 		}
671 	} else {
672 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
673 			resp.response_length += sizeof(resp.tso_caps);
674 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
675 			resp.response_length += sizeof(resp.rss_caps);
676 	}
677 
678 	if (MLX5_CAP_GEN(mdev, ipoib_ipoib_offloads)) {
679 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
680 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
681 	}
682 
683 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
684 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs))
685 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
686 
687 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
688 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
689 
690 	props->vendor_part_id	   = mdev->pdev->device;
691 	props->hw_ver		   = mdev->pdev->revision;
692 
693 	props->max_mr_size	   = ~0ull;
694 	props->page_size_cap	   = ~(min_page_size - 1);
695 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
696 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
697 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
698 		     sizeof(struct mlx5_wqe_data_seg);
699 	max_sq_sg = (MLX5_CAP_GEN(mdev, max_wqe_sz_sq) -
700 		     sizeof(struct mlx5_wqe_ctrl_seg)) /
701 		     sizeof(struct mlx5_wqe_data_seg);
702 	props->max_sge = min(max_rq_sg, max_sq_sg);
703 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
704 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
705 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
706 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
707 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
708 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
709 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
710 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
711 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
712 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
713 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
714 	props->max_srq_sge	   = max_rq_sg - 1;
715 	props->max_fast_reg_page_list_len =
716 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
717 	get_atomic_caps(dev, props);
718 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
719 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
720 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
721 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
722 					   props->max_mcast_grp;
723 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
724 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
725 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
726 
727 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
728 	if (MLX5_CAP_GEN(mdev, pg))
729 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
730 	props->odp_caps = dev->odp_caps;
731 #endif
732 
733 	if (MLX5_CAP_GEN(mdev, cd))
734 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
735 
736 	if (!mlx5_core_is_pf(mdev))
737 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
738 
739 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
740 	    IB_LINK_LAYER_ETHERNET) {
741 		props->rss_caps.max_rwq_indirection_tables =
742 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
743 		props->rss_caps.max_rwq_indirection_table_size =
744 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
745 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
746 		props->max_wq_type_rq =
747 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
748 	}
749 
750 	if (uhw->outlen) {
751 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
752 
753 		if (err)
754 			return err;
755 	}
756 
757 	return 0;
758 }
759 
760 enum mlx5_ib_width {
761 	MLX5_IB_WIDTH_1X	= 1 << 0,
762 	MLX5_IB_WIDTH_2X	= 1 << 1,
763 	MLX5_IB_WIDTH_4X	= 1 << 2,
764 	MLX5_IB_WIDTH_8X	= 1 << 3,
765 	MLX5_IB_WIDTH_12X	= 1 << 4
766 };
767 
768 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
769 				  u8 *ib_width)
770 {
771 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
772 	int err = 0;
773 
774 	if (active_width & MLX5_IB_WIDTH_1X) {
775 		*ib_width = IB_WIDTH_1X;
776 	} else if (active_width & MLX5_IB_WIDTH_2X) {
777 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
778 			    (int)active_width);
779 		err = -EINVAL;
780 	} else if (active_width & MLX5_IB_WIDTH_4X) {
781 		*ib_width = IB_WIDTH_4X;
782 	} else if (active_width & MLX5_IB_WIDTH_8X) {
783 		*ib_width = IB_WIDTH_8X;
784 	} else if (active_width & MLX5_IB_WIDTH_12X) {
785 		*ib_width = IB_WIDTH_12X;
786 	} else {
787 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
788 			    (int)active_width);
789 		err = -EINVAL;
790 	}
791 
792 	return err;
793 }
794 
795 enum ib_max_vl_num {
796 	__IB_MAX_VL_0		= 1,
797 	__IB_MAX_VL_0_1		= 2,
798 	__IB_MAX_VL_0_3		= 3,
799 	__IB_MAX_VL_0_7		= 4,
800 	__IB_MAX_VL_0_14	= 5,
801 };
802 
803 enum mlx5_vl_hw_cap {
804 	MLX5_VL_HW_0	= 1,
805 	MLX5_VL_HW_0_1	= 2,
806 	MLX5_VL_HW_0_2	= 3,
807 	MLX5_VL_HW_0_3	= 4,
808 	MLX5_VL_HW_0_4	= 5,
809 	MLX5_VL_HW_0_5	= 6,
810 	MLX5_VL_HW_0_6	= 7,
811 	MLX5_VL_HW_0_7	= 8,
812 	MLX5_VL_HW_0_14	= 15
813 };
814 
815 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
816 				u8 *max_vl_num)
817 {
818 	switch (vl_hw_cap) {
819 	case MLX5_VL_HW_0:
820 		*max_vl_num = __IB_MAX_VL_0;
821 		break;
822 	case MLX5_VL_HW_0_1:
823 		*max_vl_num = __IB_MAX_VL_0_1;
824 		break;
825 	case MLX5_VL_HW_0_3:
826 		*max_vl_num = __IB_MAX_VL_0_3;
827 		break;
828 	case MLX5_VL_HW_0_7:
829 		*max_vl_num = __IB_MAX_VL_0_7;
830 		break;
831 	case MLX5_VL_HW_0_14:
832 		*max_vl_num = __IB_MAX_VL_0_14;
833 		break;
834 
835 	default:
836 		return -EINVAL;
837 	}
838 
839 	return 0;
840 }
841 
842 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
843 			       struct ib_port_attr *props)
844 {
845 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
846 	struct mlx5_core_dev *mdev = dev->mdev;
847 	u32 *rep;
848 	int replen = MLX5_ST_SZ_BYTES(query_hca_vport_context_out);
849 	struct mlx5_ptys_reg *ptys;
850 	struct mlx5_pmtu_reg *pmtu;
851 	struct mlx5_pvlc_reg pvlc;
852 	void *ctx;
853 	int err;
854 
855 	rep = mlx5_vzalloc(replen);
856 	ptys = kzalloc(sizeof(*ptys), GFP_KERNEL);
857 	pmtu = kzalloc(sizeof(*pmtu), GFP_KERNEL);
858 	if (!rep || !ptys || !pmtu) {
859 		err = -ENOMEM;
860 		goto out;
861 	}
862 
863 	memset(props, 0, sizeof(*props));
864 
865 	err = mlx5_query_hca_vport_context(mdev, port, 0, rep, replen);
866 	if (err)
867 		goto out;
868 
869 	ctx = MLX5_ADDR_OF(query_hca_vport_context_out, rep, hca_vport_context);
870 
871 	props->lid		= MLX5_GET(hca_vport_context, ctx, lid);
872 	props->lmc		= MLX5_GET(hca_vport_context, ctx, lmc);
873 	props->sm_lid		= MLX5_GET(hca_vport_context, ctx, sm_lid);
874 	props->sm_sl		= MLX5_GET(hca_vport_context, ctx, sm_sl);
875 	props->state		= MLX5_GET(hca_vport_context, ctx, vport_state);
876 	props->phys_state	= MLX5_GET(hca_vport_context, ctx,
877 					port_physical_state);
878 	props->port_cap_flags	= MLX5_GET(hca_vport_context, ctx, cap_mask1);
879 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
880 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
881 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
882 	props->bad_pkey_cntr	= MLX5_GET(hca_vport_context, ctx,
883 					pkey_violation_counter);
884 	props->qkey_viol_cntr	= MLX5_GET(hca_vport_context, ctx,
885 					qkey_violation_counter);
886 	props->subnet_timeout	= MLX5_GET(hca_vport_context, ctx,
887 					subnet_timeout);
888 	props->init_type_reply	= MLX5_GET(hca_vport_context, ctx,
889 					init_type_reply);
890 	props->grh_required	= MLX5_GET(hca_vport_context, ctx, grh_required);
891 
892 	ptys->proto_mask |= MLX5_PTYS_IB;
893 	ptys->local_port = port;
894 	err = mlx5_core_access_ptys(mdev, ptys, 0);
895 	if (err)
896 		goto out;
897 
898 	err = translate_active_width(ibdev, ptys->ib_link_width_oper,
899 				     &props->active_width);
900 	if (err)
901 		goto out;
902 
903 	props->active_speed	= (u8)ptys->ib_proto_oper;
904 
905 	pmtu->local_port = port;
906 	err = mlx5_core_access_pmtu(mdev, pmtu, 0);
907 	if (err)
908 		goto out;
909 
910 	props->max_mtu		= pmtu->max_mtu;
911 	props->active_mtu	= pmtu->oper_mtu;
912 
913 	memset(&pvlc, 0, sizeof(pvlc));
914 	pvlc.local_port = port;
915 	err = mlx5_core_access_pvlc(mdev, &pvlc, 0);
916 	if (err)
917 		goto out;
918 
919 	err = translate_max_vl_num(ibdev, pvlc.vl_hw_cap,
920 				   &props->max_vl_num);
921 out:
922 	kvfree(rep);
923 	kfree(ptys);
924 	kfree(pmtu);
925 	return err;
926 }
927 
928 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
929 		       struct ib_port_attr *props)
930 {
931 	switch (mlx5_get_vport_access_method(ibdev)) {
932 	case MLX5_VPORT_ACCESS_METHOD_MAD:
933 		return mlx5_query_mad_ifc_port(ibdev, port, props);
934 
935 	case MLX5_VPORT_ACCESS_METHOD_HCA:
936 		return mlx5_query_hca_port(ibdev, port, props);
937 
938 	case MLX5_VPORT_ACCESS_METHOD_NIC:
939 		return mlx5_query_port_roce(ibdev, port, props);
940 
941 	default:
942 		return -EINVAL;
943 	}
944 }
945 
946 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
947 			     union ib_gid *gid)
948 {
949 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
950 	struct mlx5_core_dev *mdev = dev->mdev;
951 
952 	switch (mlx5_get_vport_access_method(ibdev)) {
953 	case MLX5_VPORT_ACCESS_METHOD_MAD:
954 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
955 
956 	case MLX5_VPORT_ACCESS_METHOD_HCA:
957 		return mlx5_query_hca_vport_gid(mdev, port, 0, index, gid);
958 
959 	default:
960 		return -EINVAL;
961 	}
962 
963 }
964 
965 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
966 			      u16 *pkey)
967 {
968 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
969 	struct mlx5_core_dev *mdev = dev->mdev;
970 
971 	switch (mlx5_get_vport_access_method(ibdev)) {
972 	case MLX5_VPORT_ACCESS_METHOD_MAD:
973 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
974 
975 	case MLX5_VPORT_ACCESS_METHOD_HCA:
976 	case MLX5_VPORT_ACCESS_METHOD_NIC:
977 		return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
978 						 pkey);
979 	default:
980 		return -EINVAL;
981 	}
982 }
983 
984 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
985 				 struct ib_device_modify *props)
986 {
987 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
988 	struct mlx5_reg_node_desc in;
989 	struct mlx5_reg_node_desc out;
990 	int err;
991 
992 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
993 		return -EOPNOTSUPP;
994 
995 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
996 		return 0;
997 
998 	/*
999 	 * If possible, pass node desc to FW, so it can generate
1000 	 * a 144 trap.  If cmd fails, just ignore.
1001 	 */
1002 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1003 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1004 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1005 	if (err)
1006 		return err;
1007 
1008 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1009 
1010 	return err;
1011 }
1012 
1013 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1014 			       struct ib_port_modify *props)
1015 {
1016 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1017 	struct ib_port_attr attr;
1018 	u32 tmp;
1019 	int err;
1020 
1021 	mutex_lock(&dev->cap_mask_mutex);
1022 
1023 	err = mlx5_ib_query_port(ibdev, port, &attr);
1024 	if (err)
1025 		goto out;
1026 
1027 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1028 		~props->clr_port_cap_mask;
1029 
1030 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1031 
1032 out:
1033 	mutex_unlock(&dev->cap_mask_mutex);
1034 	return err;
1035 }
1036 
1037 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1038 						  struct ib_udata *udata)
1039 {
1040 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1041 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1042 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1043 	struct mlx5_ib_ucontext *context;
1044 	struct mlx5_uuar_info *uuari;
1045 	struct mlx5_uar *uars;
1046 	int gross_uuars;
1047 	int num_uars;
1048 	int ver;
1049 	int uuarn;
1050 	int err;
1051 	int i;
1052 	size_t reqlen;
1053 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1054 				     max_cqe_version);
1055 
1056 	if (!dev->ib_active)
1057 		return ERR_PTR(-EAGAIN);
1058 
1059 	if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
1060 		return ERR_PTR(-EINVAL);
1061 
1062 	reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
1063 	if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1064 		ver = 0;
1065 	else if (reqlen >= min_req_v2)
1066 		ver = 2;
1067 	else
1068 		return ERR_PTR(-EINVAL);
1069 
1070 	err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
1071 	if (err)
1072 		return ERR_PTR(err);
1073 
1074 	if (req.flags)
1075 		return ERR_PTR(-EINVAL);
1076 
1077 	if (req.total_num_uuars > MLX5_MAX_UUARS)
1078 		return ERR_PTR(-ENOMEM);
1079 
1080 	if (req.total_num_uuars == 0)
1081 		return ERR_PTR(-EINVAL);
1082 
1083 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1084 		return ERR_PTR(-EOPNOTSUPP);
1085 
1086 	if (reqlen > sizeof(req) &&
1087 	    !ib_is_udata_cleared(udata, sizeof(req),
1088 				 reqlen - sizeof(req)))
1089 		return ERR_PTR(-EOPNOTSUPP);
1090 
1091 	req.total_num_uuars = ALIGN(req.total_num_uuars,
1092 				    MLX5_NON_FP_BF_REGS_PER_PAGE);
1093 	if (req.num_low_latency_uuars > req.total_num_uuars - 1)
1094 		return ERR_PTR(-EINVAL);
1095 
1096 	num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
1097 	gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
1098 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1099 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1100 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1101 	resp.cache_line_size = cache_line_size();
1102 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1103 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1104 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1105 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1106 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1107 	resp.cqe_version = min_t(__u8,
1108 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1109 				 req.max_cqe_version);
1110 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1111 				   sizeof(resp.response_length), udata->outlen);
1112 
1113 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1114 	if (!context)
1115 		return ERR_PTR(-ENOMEM);
1116 
1117 	uuari = &context->uuari;
1118 	mutex_init(&uuari->lock);
1119 	uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
1120 	if (!uars) {
1121 		err = -ENOMEM;
1122 		goto out_ctx;
1123 	}
1124 
1125 	uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
1126 				sizeof(*uuari->bitmap),
1127 				GFP_KERNEL);
1128 	if (!uuari->bitmap) {
1129 		err = -ENOMEM;
1130 		goto out_uar_ctx;
1131 	}
1132 	/*
1133 	 * clear all fast path uuars
1134 	 */
1135 	for (i = 0; i < gross_uuars; i++) {
1136 		uuarn = i & 3;
1137 		if (uuarn == 2 || uuarn == 3)
1138 			set_bit(i, uuari->bitmap);
1139 	}
1140 
1141 	uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
1142 	if (!uuari->count) {
1143 		err = -ENOMEM;
1144 		goto out_bitmap;
1145 	}
1146 
1147 	for (i = 0; i < num_uars; i++) {
1148 		err = mlx5_cmd_alloc_uar(dev->mdev, &uars[i].index);
1149 		if (err)
1150 			goto out_count;
1151 	}
1152 
1153 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1154 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1155 #endif
1156 
1157 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1158 		err = mlx5_alloc_transport_domain(dev->mdev,
1159 						       &context->tdn);
1160 		if (err)
1161 			goto out_uars;
1162 	}
1163 
1164 	INIT_LIST_HEAD(&context->vma_private_list);
1165 	INIT_LIST_HEAD(&context->db_page_list);
1166 	mutex_init(&context->db_page_mutex);
1167 
1168 	resp.tot_uuars = req.total_num_uuars;
1169 	resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1170 
1171 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1172 		resp.response_length += sizeof(resp.cqe_version);
1173 
1174 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1175 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1176 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1177 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1178 	}
1179 
1180 	/*
1181 	 * We don't want to expose information from the PCI bar that is located
1182 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1183 	 * pretend we don't support reading the HCA's core clock. This is also
1184 	 * forced by mmap function.
1185 	 */
1186 	if (PAGE_SIZE <= 4096 &&
1187 	    field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1188 		resp.comp_mask |=
1189 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1190 		resp.hca_core_clock_offset =
1191 			offsetof(struct mlx5_init_seg, internal_timer_h) %
1192 			PAGE_SIZE;
1193 		resp.response_length += sizeof(resp.hca_core_clock_offset) +
1194 					sizeof(resp.reserved2);
1195 	}
1196 
1197 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1198 	if (err)
1199 		goto out_td;
1200 
1201 	uuari->ver = ver;
1202 	uuari->num_low_latency_uuars = req.num_low_latency_uuars;
1203 	uuari->uars = uars;
1204 	uuari->num_uars = num_uars;
1205 	context->cqe_version = resp.cqe_version;
1206 
1207 	return &context->ibucontext;
1208 
1209 out_td:
1210 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1211 		mlx5_dealloc_transport_domain(dev->mdev, context->tdn);
1212 
1213 out_uars:
1214 	for (i--; i >= 0; i--)
1215 		mlx5_cmd_free_uar(dev->mdev, uars[i].index);
1216 out_count:
1217 	kfree(uuari->count);
1218 
1219 out_bitmap:
1220 	kfree(uuari->bitmap);
1221 
1222 out_uar_ctx:
1223 	kfree(uars);
1224 
1225 out_ctx:
1226 	kfree(context);
1227 	return ERR_PTR(err);
1228 }
1229 
1230 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1231 {
1232 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1233 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1234 	struct mlx5_uuar_info *uuari = &context->uuari;
1235 	int i;
1236 
1237 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1238 		mlx5_dealloc_transport_domain(dev->mdev, context->tdn);
1239 
1240 	for (i = 0; i < uuari->num_uars; i++) {
1241 		if (mlx5_cmd_free_uar(dev->mdev, uuari->uars[i].index))
1242 			mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
1243 	}
1244 
1245 	kfree(uuari->count);
1246 	kfree(uuari->bitmap);
1247 	kfree(uuari->uars);
1248 	kfree(context);
1249 
1250 	return 0;
1251 }
1252 
1253 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
1254 {
1255 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + index;
1256 }
1257 
1258 static int get_command(unsigned long offset)
1259 {
1260 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1261 }
1262 
1263 static int get_arg(unsigned long offset)
1264 {
1265 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1266 }
1267 
1268 static int get_index(unsigned long offset)
1269 {
1270 	return get_arg(offset);
1271 }
1272 
1273 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1274 {
1275 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1276 	 * is done through either mremap flow or split_vma (usually due to
1277 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1278 	 * as this VMA is strongly hardware related.  Therefore we set the
1279 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1280 	 * calling us again and trying to do incorrect actions.  We assume that
1281 	 * the original VMA size is exactly a single page, and therefore all
1282 	 * "splitting" operation will not happen to it.
1283 	 */
1284 	area->vm_ops = NULL;
1285 }
1286 
1287 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1288 {
1289 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1290 
1291 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1292 	 * file itself is closed, therefore no sync is needed with the regular
1293 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1294 	 * However need a sync with accessing the vma as part of
1295 	 * mlx5_ib_disassociate_ucontext.
1296 	 * The close operation is usually called under mm->mmap_sem except when
1297 	 * process is exiting.
1298 	 * The exiting case is handled explicitly as part of
1299 	 * mlx5_ib_disassociate_ucontext.
1300 	 */
1301 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1302 
1303 	/* setting the vma context pointer to null in the mlx5_ib driver's
1304 	 * private data, to protect a race condition in
1305 	 * mlx5_ib_disassociate_ucontext().
1306 	 */
1307 	mlx5_ib_vma_priv_data->vma = NULL;
1308 	list_del(&mlx5_ib_vma_priv_data->list);
1309 	kfree(mlx5_ib_vma_priv_data);
1310 }
1311 
1312 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1313 	.open = mlx5_ib_vma_open,
1314 	.close = mlx5_ib_vma_close
1315 };
1316 
1317 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1318 				struct mlx5_ib_ucontext *ctx)
1319 {
1320 	struct mlx5_ib_vma_private_data *vma_prv;
1321 	struct list_head *vma_head = &ctx->vma_private_list;
1322 
1323 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1324 	if (!vma_prv)
1325 		return -ENOMEM;
1326 
1327 	vma_prv->vma = vma;
1328 	vma->vm_private_data = vma_prv;
1329 	vma->vm_ops =  &mlx5_ib_vm_ops;
1330 
1331 	list_add(&vma_prv->list, vma_head);
1332 
1333 	return 0;
1334 }
1335 
1336 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1337 {
1338 	switch (cmd) {
1339 	case MLX5_IB_MMAP_WC_PAGE:
1340 		return "WC";
1341 	case MLX5_IB_MMAP_REGULAR_PAGE:
1342 		return "best effort WC";
1343 	case MLX5_IB_MMAP_NC_PAGE:
1344 		return "NC";
1345 	default:
1346 		return NULL;
1347 	}
1348 }
1349 
1350 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1351 		    struct vm_area_struct *vma,
1352 		    struct mlx5_ib_ucontext *context)
1353 {
1354 	struct mlx5_uuar_info *uuari = &context->uuari;
1355 	int err;
1356 	unsigned long idx;
1357 	phys_addr_t pfn, pa;
1358 	pgprot_t prot;
1359 
1360 	switch (cmd) {
1361 	case MLX5_IB_MMAP_WC_PAGE:
1362 /* Some architectures don't support WC memory */
1363 #if defined(CONFIG_X86)
1364 		if (!pat_enabled())
1365 			return -EPERM;
1366 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1367 			return -EPERM;
1368 #endif
1369 	/* fall through */
1370 	case MLX5_IB_MMAP_REGULAR_PAGE:
1371 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1372 		prot = pgprot_writecombine(vma->vm_page_prot);
1373 		break;
1374 	case MLX5_IB_MMAP_NC_PAGE:
1375 		prot = pgprot_noncached(vma->vm_page_prot);
1376 		break;
1377 	default:
1378 		return -EINVAL;
1379 	}
1380 
1381 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1382 		return -EINVAL;
1383 
1384 	idx = get_index(vma->vm_pgoff);
1385 	if (idx >= uuari->num_uars)
1386 		return -EINVAL;
1387 
1388 	pfn = uar_index2pfn(dev, uuari->uars[idx].index);
1389 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1390 
1391 	vma->vm_page_prot = prot;
1392 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1393 				 PAGE_SIZE, vma->vm_page_prot);
1394 	if (err) {
1395 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%llx, pfn=%pa, mmap_cmd=%s\n",
1396 			    err, (unsigned long long)vma->vm_start, &pfn, mmap_cmd2str(cmd));
1397 		return -EAGAIN;
1398 	}
1399 
1400 	pa = pfn << PAGE_SHIFT;
1401 	mlx5_ib_dbg(dev, "mapped %s at 0x%llx, PA %pa\n", mmap_cmd2str(cmd),
1402 		    (unsigned long long)vma->vm_start, &pa);
1403 
1404 	return mlx5_ib_set_vma_data(vma, context);
1405 }
1406 
1407 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1408 {
1409 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1410 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1411 	unsigned long command;
1412 	phys_addr_t pfn;
1413 
1414 	command = get_command(vma->vm_pgoff);
1415 	switch (command) {
1416 	case MLX5_IB_MMAP_WC_PAGE:
1417 	case MLX5_IB_MMAP_NC_PAGE:
1418 	case MLX5_IB_MMAP_REGULAR_PAGE:
1419 		return uar_mmap(dev, command, vma, context);
1420 
1421 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1422 		return -ENOSYS;
1423 
1424 	case MLX5_IB_MMAP_CORE_CLOCK:
1425 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1426 			return -EINVAL;
1427 
1428 		if (vma->vm_flags & VM_WRITE)
1429 			return -EPERM;
1430 
1431 		/* Don't expose to user-space information it shouldn't have */
1432 		if (PAGE_SIZE > 4096)
1433 			return -EOPNOTSUPP;
1434 
1435 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1436 		pfn = (dev->mdev->iseg_base +
1437 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1438 			PAGE_SHIFT;
1439 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1440 				       PAGE_SIZE, vma->vm_page_prot))
1441 			return -EAGAIN;
1442 
1443 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%llx, PA 0x%llx\n",
1444 			    (unsigned long long)vma->vm_start,
1445 			    (unsigned long long)pfn << PAGE_SHIFT);
1446 		break;
1447 
1448 	default:
1449 		return -EINVAL;
1450 	}
1451 
1452 	return 0;
1453 }
1454 
1455 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1456 				      struct ib_ucontext *context,
1457 				      struct ib_udata *udata)
1458 {
1459 	struct mlx5_ib_alloc_pd_resp resp;
1460 	struct mlx5_ib_pd *pd;
1461 	int err;
1462 
1463 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1464 	if (!pd)
1465 		return ERR_PTR(-ENOMEM);
1466 
1467 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1468 	if (err) {
1469 		kfree(pd);
1470 		return ERR_PTR(err);
1471 	}
1472 
1473 	if (context) {
1474 		resp.pdn = pd->pdn;
1475 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1476 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1477 			kfree(pd);
1478 			return ERR_PTR(-EFAULT);
1479 		}
1480 	}
1481 
1482 	return &pd->ibpd;
1483 }
1484 
1485 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1486 {
1487 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1488 	struct mlx5_ib_pd *mpd = to_mpd(pd);
1489 
1490 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1491 	kfree(mpd);
1492 
1493 	return 0;
1494 }
1495 
1496 enum {
1497 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
1498 	MATCH_CRITERIA_ENABLE_MISC_BIT,
1499 	MATCH_CRITERIA_ENABLE_INNER_BIT
1500 };
1501 
1502 #define HEADER_IS_ZERO(match_criteria, headers)			           \
1503 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1504 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
1505 
1506 static u8 get_match_criteria_enable(u32 *match_criteria)
1507 {
1508 	u8 match_criteria_enable;
1509 
1510 	match_criteria_enable =
1511 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1512 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
1513 	match_criteria_enable |=
1514 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1515 		MATCH_CRITERIA_ENABLE_MISC_BIT;
1516 	match_criteria_enable |=
1517 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1518 		MATCH_CRITERIA_ENABLE_INNER_BIT;
1519 
1520 	return match_criteria_enable;
1521 }
1522 
1523 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1524 {
1525 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1526 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1527 }
1528 
1529 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1530 {
1531 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1532 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1533 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1534 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1535 }
1536 
1537 #define LAST_ETH_FIELD vlan_tag
1538 #define LAST_IB_FIELD sl
1539 #define LAST_IPV4_FIELD tos
1540 #define LAST_IPV6_FIELD traffic_class
1541 #define LAST_TCP_UDP_FIELD src_port
1542 
1543 /* Field is the last supported field */
1544 #define FIELDS_NOT_SUPPORTED(filter, field)\
1545 	memchr_inv((void *)&filter.field  +\
1546 		   sizeof(filter.field), 0,\
1547 		   sizeof(filter) -\
1548 		   offsetof(typeof(filter), field) -\
1549 		   sizeof(filter.field))
1550 
1551 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1552 			   const union ib_flow_spec *ib_spec)
1553 {
1554 	void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1555 					     outer_headers);
1556 	void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1557 					     outer_headers);
1558 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1559 					   misc_parameters);
1560 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1561 					   misc_parameters);
1562 
1563 	switch (ib_spec->type) {
1564 	case IB_FLOW_SPEC_ETH:
1565 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1566 			return -ENOTSUPP;
1567 
1568 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1569 					     dmac_47_16),
1570 				ib_spec->eth.mask.dst_mac);
1571 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1572 					     dmac_47_16),
1573 				ib_spec->eth.val.dst_mac);
1574 
1575 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1576 					     smac_47_16),
1577 				ib_spec->eth.mask.src_mac);
1578 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1579 					     smac_47_16),
1580 				ib_spec->eth.val.src_mac);
1581 
1582 		if (ib_spec->eth.mask.vlan_tag) {
1583 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1584 				 cvlan_tag, 1);
1585 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1586 				 cvlan_tag, 1);
1587 
1588 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1589 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1590 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1591 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1592 
1593 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1594 				 first_cfi,
1595 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1596 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1597 				 first_cfi,
1598 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1599 
1600 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1601 				 first_prio,
1602 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1603 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1604 				 first_prio,
1605 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1606 		}
1607 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1608 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1609 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1610 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
1611 		break;
1612 	case IB_FLOW_SPEC_IPV4:
1613 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1614 			return -ENOTSUPP;
1615 
1616 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1617 			 ethertype, 0xffff);
1618 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1619 			 ethertype, ETH_P_IP);
1620 
1621 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1622 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1623 		       &ib_spec->ipv4.mask.src_ip,
1624 		       sizeof(ib_spec->ipv4.mask.src_ip));
1625 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1626 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1627 		       &ib_spec->ipv4.val.src_ip,
1628 		       sizeof(ib_spec->ipv4.val.src_ip));
1629 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1630 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1631 		       &ib_spec->ipv4.mask.dst_ip,
1632 		       sizeof(ib_spec->ipv4.mask.dst_ip));
1633 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1634 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1635 		       &ib_spec->ipv4.val.dst_ip,
1636 		       sizeof(ib_spec->ipv4.val.dst_ip));
1637 
1638 		set_tos(outer_headers_c, outer_headers_v,
1639 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
1640 
1641 		set_proto(outer_headers_c, outer_headers_v,
1642 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
1643 		break;
1644 	case IB_FLOW_SPEC_IPV6:
1645 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1646 			return -ENOTSUPP;
1647 
1648 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1649 			 ethertype, 0xffff);
1650 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1651 			 ethertype, IPPROTO_IPV6);
1652 
1653 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1654 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1655 		       &ib_spec->ipv6.mask.src_ip,
1656 		       sizeof(ib_spec->ipv6.mask.src_ip));
1657 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1658 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1659 		       &ib_spec->ipv6.val.src_ip,
1660 		       sizeof(ib_spec->ipv6.val.src_ip));
1661 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1662 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1663 		       &ib_spec->ipv6.mask.dst_ip,
1664 		       sizeof(ib_spec->ipv6.mask.dst_ip));
1665 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1666 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1667 		       &ib_spec->ipv6.val.dst_ip,
1668 		       sizeof(ib_spec->ipv6.val.dst_ip));
1669 
1670 		set_tos(outer_headers_c, outer_headers_v,
1671 			ib_spec->ipv6.mask.traffic_class,
1672 			ib_spec->ipv6.val.traffic_class);
1673 
1674 		set_proto(outer_headers_c, outer_headers_v,
1675 			  ib_spec->ipv6.mask.next_hdr,
1676 			  ib_spec->ipv6.val.next_hdr);
1677 
1678 		MLX5_SET(fte_match_set_misc, misc_params_c,
1679 			 outer_ipv6_flow_label,
1680 			 ntohl(ib_spec->ipv6.mask.flow_label));
1681 		MLX5_SET(fte_match_set_misc, misc_params_v,
1682 			 outer_ipv6_flow_label,
1683 			 ntohl(ib_spec->ipv6.val.flow_label));
1684 		break;
1685 	case IB_FLOW_SPEC_TCP:
1686 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1687 					 LAST_TCP_UDP_FIELD))
1688 			return -ENOTSUPP;
1689 
1690 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1691 			 0xff);
1692 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1693 			 IPPROTO_TCP);
1694 
1695 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport,
1696 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1697 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport,
1698 			 ntohs(ib_spec->tcp_udp.val.src_port));
1699 
1700 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport,
1701 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1702 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport,
1703 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1704 		break;
1705 	case IB_FLOW_SPEC_UDP:
1706 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1707 					 LAST_TCP_UDP_FIELD))
1708 			return -ENOTSUPP;
1709 
1710 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1711 			 0xff);
1712 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1713 			 IPPROTO_UDP);
1714 
1715 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport,
1716 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1717 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport,
1718 			 ntohs(ib_spec->tcp_udp.val.src_port));
1719 
1720 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport,
1721 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1722 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport,
1723 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1724 		break;
1725 	default:
1726 		return -EINVAL;
1727 	}
1728 
1729 	return 0;
1730 }
1731 
1732 /* If a flow could catch both multicast and unicast packets,
1733  * it won't fall into the multicast flow steering table and this rule
1734  * could steal other multicast packets.
1735  */
1736 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1737 {
1738 	struct ib_flow_spec_eth *eth_spec;
1739 
1740 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1741 	    ib_attr->size < sizeof(struct ib_flow_attr) +
1742 	    sizeof(struct ib_flow_spec_eth) ||
1743 	    ib_attr->num_of_specs < 1)
1744 		return false;
1745 
1746 	eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1747 	if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1748 	    eth_spec->size != sizeof(*eth_spec))
1749 		return false;
1750 
1751 	return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1752 	       is_multicast_ether_addr(eth_spec->val.dst_mac);
1753 }
1754 
1755 static bool is_valid_attr(const struct ib_flow_attr *flow_attr)
1756 {
1757 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1758 	bool has_ipv4_spec = false;
1759 	bool eth_type_ipv4 = true;
1760 	unsigned int spec_index;
1761 
1762 	/* Validate that ethertype is correct */
1763 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1764 		if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1765 		    ib_spec->eth.mask.ether_type) {
1766 			if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1767 			      ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1768 				eth_type_ipv4 = false;
1769 		} else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1770 			has_ipv4_spec = true;
1771 		}
1772 		ib_spec = (void *)ib_spec + ib_spec->size;
1773 	}
1774 	return !has_ipv4_spec || eth_type_ipv4;
1775 }
1776 
1777 static void put_flow_table(struct mlx5_ib_dev *dev,
1778 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
1779 {
1780 	prio->refcount -= !!ft_added;
1781 	if (!prio->refcount) {
1782 		mlx5_destroy_flow_table(prio->flow_table);
1783 		prio->flow_table = NULL;
1784 	}
1785 }
1786 
1787 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
1788 {
1789 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
1790 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
1791 							  struct mlx5_ib_flow_handler,
1792 							  ibflow);
1793 	struct mlx5_ib_flow_handler *iter, *tmp;
1794 
1795 	mutex_lock(&dev->flow_db.lock);
1796 
1797 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
1798 		mlx5_del_flow_rule(iter->rule);
1799 		put_flow_table(dev, iter->prio, true);
1800 		list_del(&iter->list);
1801 		kfree(iter);
1802 	}
1803 
1804 	mlx5_del_flow_rule(handler->rule);
1805 	put_flow_table(dev, handler->prio, true);
1806 	mutex_unlock(&dev->flow_db.lock);
1807 
1808 	kfree(handler);
1809 
1810 	return 0;
1811 }
1812 
1813 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
1814 {
1815 	priority *= 2;
1816 	if (!dont_trap)
1817 		priority++;
1818 	return priority;
1819 }
1820 
1821 enum flow_table_type {
1822 	MLX5_IB_FT_RX,
1823 	MLX5_IB_FT_TX
1824 };
1825 
1826 #define MLX5_FS_MAX_TYPES	 10
1827 #define MLX5_FS_MAX_ENTRIES	 32000UL
1828 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
1829 						struct ib_flow_attr *flow_attr,
1830 						enum flow_table_type ft_type)
1831 {
1832 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
1833 	struct mlx5_flow_namespace *ns = NULL;
1834 	struct mlx5_ib_flow_prio *prio;
1835 	struct mlx5_flow_table *ft;
1836 	int num_entries;
1837 	int num_groups;
1838 	int priority;
1839 	int err = 0;
1840 
1841 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1842 		if (flow_is_multicast_only(flow_attr) &&
1843 		    !dont_trap)
1844 			priority = MLX5_IB_FLOW_MCAST_PRIO;
1845 		else
1846 			priority = ib_prio_to_core_prio(flow_attr->priority,
1847 							dont_trap);
1848 		ns = mlx5_get_flow_namespace(dev->mdev,
1849 					     MLX5_FLOW_NAMESPACE_BYPASS);
1850 		num_entries = MLX5_FS_MAX_ENTRIES;
1851 		num_groups = MLX5_FS_MAX_TYPES;
1852 		prio = &dev->flow_db.prios[priority];
1853 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1854 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1855 		ns = mlx5_get_flow_namespace(dev->mdev,
1856 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
1857 		build_leftovers_ft_param("bypass", &priority,
1858 					 &num_entries,
1859 					 &num_groups);
1860 		prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
1861 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
1862 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
1863 					allow_sniffer_and_nic_rx_shared_tir))
1864 			return ERR_PTR(-ENOTSUPP);
1865 
1866 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
1867 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
1868 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
1869 
1870 		prio = &dev->flow_db.sniffer[ft_type];
1871 		priority = 0;
1872 		num_entries = 1;
1873 		num_groups = 1;
1874 	}
1875 
1876 	if (!ns)
1877 		return ERR_PTR(-ENOTSUPP);
1878 
1879 	ft = prio->flow_table;
1880 	if (!ft) {
1881 		ft = mlx5_create_auto_grouped_flow_table(ns, priority, "bypass",
1882 							 num_entries,
1883 							 num_groups);
1884 
1885 		if (!IS_ERR(ft)) {
1886 			prio->refcount = 0;
1887 			prio->flow_table = ft;
1888 		} else {
1889 			err = PTR_ERR(ft);
1890 		}
1891 	}
1892 
1893 	return err ? ERR_PTR(err) : prio;
1894 }
1895 
1896 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
1897 						     struct mlx5_ib_flow_prio *ft_prio,
1898 						     const struct ib_flow_attr *flow_attr,
1899 						     struct mlx5_flow_destination *dst)
1900 {
1901 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
1902 	struct mlx5_ib_flow_handler *handler;
1903 	struct mlx5_flow_spec *spec;
1904 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
1905 	unsigned int spec_index;
1906 	u32 action;
1907 	int err = 0;
1908 
1909 	if (!is_valid_attr(flow_attr))
1910 		return ERR_PTR(-EINVAL);
1911 
1912 	spec = mlx5_vzalloc(sizeof(*spec));
1913 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
1914 	if (!handler || !spec) {
1915 		err = -ENOMEM;
1916 		goto free;
1917 	}
1918 
1919 	INIT_LIST_HEAD(&handler->list);
1920 
1921 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1922 		err = parse_flow_attr(spec->match_criteria,
1923 				      spec->match_value, ib_flow);
1924 		if (err < 0)
1925 			goto free;
1926 
1927 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
1928 	}
1929 
1930 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
1931 	action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
1932 		MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
1933 	handler->rule = mlx5_add_flow_rule(ft, spec->match_criteria_enable,
1934 					   spec->match_criteria,
1935 					   spec->match_value,
1936 					   action,
1937 					   MLX5_FS_DEFAULT_FLOW_TAG,
1938 					   dst);
1939 
1940 	if (IS_ERR(handler->rule)) {
1941 		err = PTR_ERR(handler->rule);
1942 		goto free;
1943 	}
1944 
1945 	ft_prio->refcount++;
1946 	handler->prio = ft_prio;
1947 
1948 	ft_prio->flow_table = ft;
1949 free:
1950 	if (err)
1951 		kfree(handler);
1952 	kvfree(spec);
1953 	return err ? ERR_PTR(err) : handler;
1954 }
1955 
1956 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
1957 							  struct mlx5_ib_flow_prio *ft_prio,
1958 							  struct ib_flow_attr *flow_attr,
1959 							  struct mlx5_flow_destination *dst)
1960 {
1961 	struct mlx5_ib_flow_handler *handler_dst = NULL;
1962 	struct mlx5_ib_flow_handler *handler = NULL;
1963 
1964 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
1965 	if (!IS_ERR(handler)) {
1966 		handler_dst = create_flow_rule(dev, ft_prio,
1967 					       flow_attr, dst);
1968 		if (IS_ERR(handler_dst)) {
1969 			mlx5_del_flow_rule(handler->rule);
1970 			ft_prio->refcount--;
1971 			kfree(handler);
1972 			handler = handler_dst;
1973 		} else {
1974 			list_add(&handler_dst->list, &handler->list);
1975 		}
1976 	}
1977 
1978 	return handler;
1979 }
1980 enum {
1981 	LEFTOVERS_MC,
1982 	LEFTOVERS_UC,
1983 };
1984 
1985 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
1986 							  struct mlx5_ib_flow_prio *ft_prio,
1987 							  struct ib_flow_attr *flow_attr,
1988 							  struct mlx5_flow_destination *dst)
1989 {
1990 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
1991 	struct mlx5_ib_flow_handler *handler = NULL;
1992 
1993 	static struct {
1994 		struct ib_flow_attr	flow_attr;
1995 		struct ib_flow_spec_eth eth_flow;
1996 	} leftovers_specs[] = {
1997 		[LEFTOVERS_MC] = {
1998 			.flow_attr = {
1999 				.num_of_specs = 1,
2000 				.size = sizeof(leftovers_specs[0])
2001 			},
2002 			.eth_flow = {
2003 				.type = IB_FLOW_SPEC_ETH,
2004 				.size = sizeof(struct ib_flow_spec_eth),
2005 				.mask = {.dst_mac = {0x1} },
2006 				.val =  {.dst_mac = {0x1} }
2007 			}
2008 		},
2009 		[LEFTOVERS_UC] = {
2010 			.flow_attr = {
2011 				.num_of_specs = 1,
2012 				.size = sizeof(leftovers_specs[0])
2013 			},
2014 			.eth_flow = {
2015 				.type = IB_FLOW_SPEC_ETH,
2016 				.size = sizeof(struct ib_flow_spec_eth),
2017 				.mask = {.dst_mac = {0x1} },
2018 				.val = {.dst_mac = {} }
2019 			}
2020 		}
2021 	};
2022 
2023 	handler = create_flow_rule(dev, ft_prio,
2024 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
2025 				   dst);
2026 	if (!IS_ERR(handler) &&
2027 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2028 		handler_ucast = create_flow_rule(dev, ft_prio,
2029 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
2030 						 dst);
2031 		if (IS_ERR(handler_ucast)) {
2032 			mlx5_del_flow_rule(handler->rule);
2033 			ft_prio->refcount--;
2034 			kfree(handler);
2035 			handler = handler_ucast;
2036 		} else {
2037 			list_add(&handler_ucast->list, &handler->list);
2038 		}
2039 	}
2040 
2041 	return handler;
2042 }
2043 
2044 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2045 							struct mlx5_ib_flow_prio *ft_rx,
2046 							struct mlx5_ib_flow_prio *ft_tx,
2047 							struct mlx5_flow_destination *dst)
2048 {
2049 	struct mlx5_ib_flow_handler *handler_rx;
2050 	struct mlx5_ib_flow_handler *handler_tx;
2051 	int err;
2052 	static const struct ib_flow_attr flow_attr  = {
2053 		.num_of_specs = 0,
2054 		.size = sizeof(flow_attr)
2055 	};
2056 
2057 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2058 	if (IS_ERR(handler_rx)) {
2059 		err = PTR_ERR(handler_rx);
2060 		goto err;
2061 	}
2062 
2063 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2064 	if (IS_ERR(handler_tx)) {
2065 		err = PTR_ERR(handler_tx);
2066 		goto err_tx;
2067 	}
2068 
2069 	list_add(&handler_tx->list, &handler_rx->list);
2070 
2071 	return handler_rx;
2072 
2073 err_tx:
2074 	mlx5_del_flow_rule(handler_rx->rule);
2075 	ft_rx->refcount--;
2076 	kfree(handler_rx);
2077 err:
2078 	return ERR_PTR(err);
2079 }
2080 
2081 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2082 					   struct ib_flow_attr *flow_attr,
2083 					   int domain)
2084 {
2085 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2086 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2087 	struct mlx5_ib_flow_handler *handler = NULL;
2088 	struct mlx5_flow_destination *dst = NULL;
2089 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2090 	struct mlx5_ib_flow_prio *ft_prio;
2091 	int err;
2092 
2093 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2094 		return ERR_PTR(-ENOSPC);
2095 
2096 	if (domain != IB_FLOW_DOMAIN_USER ||
2097 	    flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2098 	    (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2099 		return ERR_PTR(-EINVAL);
2100 
2101 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2102 	if (!dst)
2103 		return ERR_PTR(-ENOMEM);
2104 
2105 	mutex_lock(&dev->flow_db.lock);
2106 
2107 	ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2108 	if (IS_ERR(ft_prio)) {
2109 		err = PTR_ERR(ft_prio);
2110 		goto unlock;
2111 	}
2112 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2113 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2114 		if (IS_ERR(ft_prio_tx)) {
2115 			err = PTR_ERR(ft_prio_tx);
2116 			ft_prio_tx = NULL;
2117 			goto destroy_ft;
2118 		}
2119 	}
2120 
2121 	dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2122 	if (mqp->flags & MLX5_IB_QP_RSS)
2123 		dst->tir_num = mqp->rss_qp.tirn;
2124 	else
2125 		dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2126 
2127 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2128 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
2129 			handler = create_dont_trap_rule(dev, ft_prio,
2130 							flow_attr, dst);
2131 		} else {
2132 			handler = create_flow_rule(dev, ft_prio, flow_attr,
2133 						   dst);
2134 		}
2135 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2136 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2137 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2138 						dst);
2139 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2140 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2141 	} else {
2142 		err = -EINVAL;
2143 		goto destroy_ft;
2144 	}
2145 
2146 	if (IS_ERR(handler)) {
2147 		err = PTR_ERR(handler);
2148 		handler = NULL;
2149 		goto destroy_ft;
2150 	}
2151 
2152 	mutex_unlock(&dev->flow_db.lock);
2153 	kfree(dst);
2154 
2155 	return &handler->ibflow;
2156 
2157 destroy_ft:
2158 	put_flow_table(dev, ft_prio, false);
2159 	if (ft_prio_tx)
2160 		put_flow_table(dev, ft_prio_tx, false);
2161 unlock:
2162 	mutex_unlock(&dev->flow_db.lock);
2163 	kfree(dst);
2164 	kfree(handler);
2165 	return ERR_PTR(err);
2166 }
2167 
2168 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2169 {
2170 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2171 	int err;
2172 
2173 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2174 	if (err)
2175 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2176 			     ibqp->qp_num, gid->raw);
2177 
2178 	return err;
2179 }
2180 
2181 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2182 {
2183 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2184 	int err;
2185 
2186 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2187 	if (err)
2188 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2189 			     ibqp->qp_num, gid->raw);
2190 
2191 	return err;
2192 }
2193 
2194 static int init_node_data(struct mlx5_ib_dev *dev)
2195 {
2196 	int err;
2197 
2198 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2199 	if (err)
2200 		return err;
2201 
2202 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2203 }
2204 
2205 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2206 			     char *buf)
2207 {
2208 	struct mlx5_ib_dev *dev =
2209 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2210 
2211 	return sprintf(buf, "%lld\n", (long long)dev->mdev->priv.fw_pages);
2212 }
2213 
2214 static ssize_t show_reg_pages(struct device *device,
2215 			      struct device_attribute *attr, char *buf)
2216 {
2217 	struct mlx5_ib_dev *dev =
2218 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2219 
2220 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2221 }
2222 
2223 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2224 			char *buf)
2225 {
2226 	struct mlx5_ib_dev *dev =
2227 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2228 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2229 }
2230 
2231 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2232 			char *buf)
2233 {
2234 	struct mlx5_ib_dev *dev =
2235 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2236 	return sprintf(buf, "%x\n", dev->mdev->pdev->revision);
2237 }
2238 
2239 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2240 			  char *buf)
2241 {
2242 	struct mlx5_ib_dev *dev =
2243 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2244 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2245 		       dev->mdev->board_id);
2246 }
2247 
2248 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2249 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2250 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2251 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2252 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2253 
2254 static struct device_attribute *mlx5_class_attributes[] = {
2255 	&dev_attr_hw_rev,
2256 	&dev_attr_hca_type,
2257 	&dev_attr_board_id,
2258 	&dev_attr_fw_pages,
2259 	&dev_attr_reg_pages,
2260 };
2261 
2262 static void pkey_change_handler(struct work_struct *work)
2263 {
2264 	struct mlx5_ib_port_resources *ports =
2265 		container_of(work, struct mlx5_ib_port_resources,
2266 			     pkey_change_work);
2267 
2268 	mutex_lock(&ports->devr->mutex);
2269 	mlx5_ib_gsi_pkey_change(ports->gsi);
2270 	mutex_unlock(&ports->devr->mutex);
2271 }
2272 
2273 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2274 {
2275 	struct mlx5_ib_qp *mqp;
2276 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2277 	struct mlx5_core_cq *mcq;
2278 	struct list_head cq_armed_list;
2279 	unsigned long flags_qp;
2280 	unsigned long flags_cq;
2281 	unsigned long flags;
2282 
2283 	INIT_LIST_HEAD(&cq_armed_list);
2284 
2285 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2286 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2287 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2288 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2289 		if (mqp->sq.tail != mqp->sq.head) {
2290 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2291 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2292 			if (send_mcq->mcq.comp &&
2293 			    mqp->ibqp.send_cq->comp_handler) {
2294 				if (!send_mcq->mcq.reset_notify_added) {
2295 					send_mcq->mcq.reset_notify_added = 1;
2296 					list_add_tail(&send_mcq->mcq.reset_notify,
2297 						      &cq_armed_list);
2298 				}
2299 			}
2300 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2301 		}
2302 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2303 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2304 		/* no handling is needed for SRQ */
2305 		if (!mqp->ibqp.srq) {
2306 			if (mqp->rq.tail != mqp->rq.head) {
2307 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2308 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2309 				if (recv_mcq->mcq.comp &&
2310 				    mqp->ibqp.recv_cq->comp_handler) {
2311 					if (!recv_mcq->mcq.reset_notify_added) {
2312 						recv_mcq->mcq.reset_notify_added = 1;
2313 						list_add_tail(&recv_mcq->mcq.reset_notify,
2314 							      &cq_armed_list);
2315 					}
2316 				}
2317 				spin_unlock_irqrestore(&recv_mcq->lock,
2318 						       flags_cq);
2319 			}
2320 		}
2321 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2322 	}
2323 	/*At that point all inflight post send were put to be executed as of we
2324 	 * lock/unlock above locks Now need to arm all involved CQs.
2325 	 */
2326 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2327 		mcq->comp(mcq);
2328 	}
2329 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2330 }
2331 
2332 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2333 			  enum mlx5_dev_event event, unsigned long param)
2334 {
2335 	struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2336 	struct ib_event ibev;
2337 	bool fatal = false;
2338 	u8 port = 0;
2339 
2340 	switch (event) {
2341 	case MLX5_DEV_EVENT_SYS_ERROR:
2342 		ibev.event = IB_EVENT_DEVICE_FATAL;
2343 		mlx5_ib_handle_internal_error(ibdev);
2344 		fatal = true;
2345 		break;
2346 
2347 	case MLX5_DEV_EVENT_PORT_UP:
2348 	case MLX5_DEV_EVENT_PORT_DOWN:
2349 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
2350 		port = (u8)param;
2351 
2352 		/* In RoCE, port up/down events are handled in
2353 		 * mlx5_netdev_event().
2354 		 */
2355 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2356 			IB_LINK_LAYER_ETHERNET)
2357 			return;
2358 
2359 		ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2360 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2361 		break;
2362 
2363 	case MLX5_DEV_EVENT_LID_CHANGE:
2364 		ibev.event = IB_EVENT_LID_CHANGE;
2365 		port = (u8)param;
2366 		break;
2367 
2368 	case MLX5_DEV_EVENT_PKEY_CHANGE:
2369 		ibev.event = IB_EVENT_PKEY_CHANGE;
2370 		port = (u8)param;
2371 
2372 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2373 		break;
2374 
2375 	case MLX5_DEV_EVENT_GUID_CHANGE:
2376 		ibev.event = IB_EVENT_GID_CHANGE;
2377 		port = (u8)param;
2378 		break;
2379 
2380 	case MLX5_DEV_EVENT_CLIENT_REREG:
2381 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
2382 		port = (u8)param;
2383 		break;
2384 
2385 	default:
2386 		/* unsupported event */
2387 		return;
2388 	}
2389 
2390 	ibev.device	      = &ibdev->ib_dev;
2391 	ibev.element.port_num = port;
2392 
2393 	if (port < 1 || port > ibdev->num_ports) {
2394 		mlx5_ib_warn(ibdev, "warning: event(%d) on port %d\n", event, port);
2395 		return;
2396 	}
2397 
2398 	if (ibdev->ib_active)
2399 		ib_dispatch_event(&ibev);
2400 
2401 	if (fatal)
2402 		ibdev->ib_active = false;
2403 }
2404 
2405 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2406 {
2407 	int port;
2408 
2409 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2410 		mlx5_query_ext_port_caps(dev, port);
2411 }
2412 
2413 static int get_port_caps(struct mlx5_ib_dev *dev)
2414 {
2415 	struct ib_device_attr *dprops = NULL;
2416 	struct ib_port_attr *pprops = NULL;
2417 	int err = -ENOMEM;
2418 	int port;
2419 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2420 
2421 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2422 	if (!pprops)
2423 		goto out;
2424 
2425 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2426 	if (!dprops)
2427 		goto out;
2428 
2429 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2430 	if (err) {
2431 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2432 		goto out;
2433 	}
2434 
2435 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2436 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2437 		if (err) {
2438 			mlx5_ib_warn(dev, "query_port %d failed %d\n",
2439 				     port, err);
2440 			break;
2441 		}
2442 		dev->mdev->port_caps[port - 1].pkey_table_len =
2443 						dprops->max_pkeys;
2444 		dev->mdev->port_caps[port - 1].gid_table_len =
2445 						pprops->gid_tbl_len;
2446 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2447 			    dprops->max_pkeys, pprops->gid_tbl_len);
2448 	}
2449 
2450 out:
2451 	kfree(pprops);
2452 	kfree(dprops);
2453 
2454 	return err;
2455 }
2456 
2457 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2458 {
2459 	int err;
2460 
2461 	err = mlx5_mr_cache_cleanup(dev);
2462 	if (err)
2463 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2464 
2465 	mlx5_ib_destroy_qp(dev->umrc.qp);
2466 	ib_free_cq(dev->umrc.cq);
2467 	ib_dealloc_pd(dev->umrc.pd);
2468 }
2469 
2470 enum {
2471 	MAX_UMR_WR = 128,
2472 };
2473 
2474 static int create_umr_res(struct mlx5_ib_dev *dev)
2475 {
2476 	struct ib_qp_init_attr *init_attr = NULL;
2477 	struct ib_qp_attr *attr = NULL;
2478 	struct ib_pd *pd;
2479 	struct ib_cq *cq;
2480 	struct ib_qp *qp;
2481 	int ret;
2482 
2483 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2484 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2485 	if (!attr || !init_attr) {
2486 		ret = -ENOMEM;
2487 		goto error_0;
2488 	}
2489 
2490 	pd = ib_alloc_pd(&dev->ib_dev, 0);
2491 	if (IS_ERR(pd)) {
2492 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2493 		ret = PTR_ERR(pd);
2494 		goto error_0;
2495 	}
2496 
2497 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2498 	if (IS_ERR(cq)) {
2499 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2500 		ret = PTR_ERR(cq);
2501 		goto error_2;
2502 	}
2503 
2504 	init_attr->send_cq = cq;
2505 	init_attr->recv_cq = cq;
2506 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2507 	init_attr->cap.max_send_wr = MAX_UMR_WR;
2508 	init_attr->cap.max_send_sge = 1;
2509 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2510 	init_attr->port_num = 1;
2511 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2512 	if (IS_ERR(qp)) {
2513 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2514 		ret = PTR_ERR(qp);
2515 		goto error_3;
2516 	}
2517 	qp->device     = &dev->ib_dev;
2518 	qp->real_qp    = qp;
2519 	qp->uobject    = NULL;
2520 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2521 
2522 	attr->qp_state = IB_QPS_INIT;
2523 	attr->port_num = 1;
2524 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2525 				IB_QP_PORT, NULL);
2526 	if (ret) {
2527 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2528 		goto error_4;
2529 	}
2530 
2531 	memset(attr, 0, sizeof(*attr));
2532 	attr->qp_state = IB_QPS_RTR;
2533 	attr->path_mtu = IB_MTU_256;
2534 
2535 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2536 	if (ret) {
2537 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2538 		goto error_4;
2539 	}
2540 
2541 	memset(attr, 0, sizeof(*attr));
2542 	attr->qp_state = IB_QPS_RTS;
2543 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2544 	if (ret) {
2545 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2546 		goto error_4;
2547 	}
2548 
2549 	dev->umrc.qp = qp;
2550 	dev->umrc.cq = cq;
2551 	dev->umrc.pd = pd;
2552 
2553 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
2554 	ret = mlx5_mr_cache_init(dev);
2555 	if (ret) {
2556 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2557 		goto error_4;
2558 	}
2559 
2560 	kfree(attr);
2561 	kfree(init_attr);
2562 
2563 	return 0;
2564 
2565 error_4:
2566 	mlx5_ib_destroy_qp(qp);
2567 
2568 error_3:
2569 	ib_free_cq(cq);
2570 
2571 error_2:
2572 	ib_dealloc_pd(pd);
2573 
2574 error_0:
2575 	kfree(attr);
2576 	kfree(init_attr);
2577 	return ret;
2578 }
2579 
2580 static int create_dev_resources(struct mlx5_ib_resources *devr)
2581 {
2582 	struct ib_srq_init_attr attr;
2583 	struct mlx5_ib_dev *dev;
2584 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2585 	int port;
2586 	int ret = 0;
2587 
2588 	dev = container_of(devr, struct mlx5_ib_dev, devr);
2589 
2590 	mutex_init(&devr->mutex);
2591 
2592 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2593 	if (IS_ERR(devr->p0)) {
2594 		ret = PTR_ERR(devr->p0);
2595 		goto error0;
2596 	}
2597 	devr->p0->device  = &dev->ib_dev;
2598 	devr->p0->uobject = NULL;
2599 	atomic_set(&devr->p0->usecnt, 0);
2600 
2601 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
2602 	if (IS_ERR(devr->c0)) {
2603 		ret = PTR_ERR(devr->c0);
2604 		goto error1;
2605 	}
2606 	devr->c0->device        = &dev->ib_dev;
2607 	devr->c0->uobject       = NULL;
2608 	devr->c0->comp_handler  = NULL;
2609 	devr->c0->event_handler = NULL;
2610 	devr->c0->cq_context    = NULL;
2611 	atomic_set(&devr->c0->usecnt, 0);
2612 
2613 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2614 	if (IS_ERR(devr->x0)) {
2615 		ret = PTR_ERR(devr->x0);
2616 		goto error2;
2617 	}
2618 	devr->x0->device = &dev->ib_dev;
2619 	devr->x0->inode = NULL;
2620 	atomic_set(&devr->x0->usecnt, 0);
2621 	mutex_init(&devr->x0->tgt_qp_mutex);
2622 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2623 
2624 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2625 	if (IS_ERR(devr->x1)) {
2626 		ret = PTR_ERR(devr->x1);
2627 		goto error3;
2628 	}
2629 	devr->x1->device = &dev->ib_dev;
2630 	devr->x1->inode = NULL;
2631 	atomic_set(&devr->x1->usecnt, 0);
2632 	mutex_init(&devr->x1->tgt_qp_mutex);
2633 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2634 
2635 	memset(&attr, 0, sizeof(attr));
2636 	attr.attr.max_sge = 1;
2637 	attr.attr.max_wr = 1;
2638 	attr.srq_type = IB_SRQT_XRC;
2639 	attr.ext.xrc.cq = devr->c0;
2640 	attr.ext.xrc.xrcd = devr->x0;
2641 
2642 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2643 	if (IS_ERR(devr->s0)) {
2644 		ret = PTR_ERR(devr->s0);
2645 		goto error4;
2646 	}
2647 	devr->s0->device	= &dev->ib_dev;
2648 	devr->s0->pd		= devr->p0;
2649 	devr->s0->uobject       = NULL;
2650 	devr->s0->event_handler = NULL;
2651 	devr->s0->srq_context   = NULL;
2652 	devr->s0->srq_type      = IB_SRQT_XRC;
2653 	devr->s0->ext.xrc.xrcd	= devr->x0;
2654 	devr->s0->ext.xrc.cq	= devr->c0;
2655 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2656 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2657 	atomic_inc(&devr->p0->usecnt);
2658 	atomic_set(&devr->s0->usecnt, 0);
2659 
2660 	memset(&attr, 0, sizeof(attr));
2661 	attr.attr.max_sge = 1;
2662 	attr.attr.max_wr = 1;
2663 	attr.srq_type = IB_SRQT_BASIC;
2664 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2665 	if (IS_ERR(devr->s1)) {
2666 		ret = PTR_ERR(devr->s1);
2667 		goto error5;
2668 	}
2669 	devr->s1->device	= &dev->ib_dev;
2670 	devr->s1->pd		= devr->p0;
2671 	devr->s1->uobject       = NULL;
2672 	devr->s1->event_handler = NULL;
2673 	devr->s1->srq_context   = NULL;
2674 	devr->s1->srq_type      = IB_SRQT_BASIC;
2675 	devr->s1->ext.xrc.cq	= devr->c0;
2676 	atomic_inc(&devr->p0->usecnt);
2677 	atomic_set(&devr->s0->usecnt, 0);
2678 
2679 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2680 		INIT_WORK(&devr->ports[port].pkey_change_work,
2681 			  pkey_change_handler);
2682 		devr->ports[port].devr = devr;
2683 	}
2684 
2685 	return 0;
2686 
2687 error5:
2688 	mlx5_ib_destroy_srq(devr->s0);
2689 error4:
2690 	mlx5_ib_dealloc_xrcd(devr->x1);
2691 error3:
2692 	mlx5_ib_dealloc_xrcd(devr->x0);
2693 error2:
2694 	mlx5_ib_destroy_cq(devr->c0);
2695 error1:
2696 	mlx5_ib_dealloc_pd(devr->p0);
2697 error0:
2698 	return ret;
2699 }
2700 
2701 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2702 {
2703 	struct mlx5_ib_dev *dev =
2704 		container_of(devr, struct mlx5_ib_dev, devr);
2705 	int port;
2706 
2707 	mlx5_ib_destroy_srq(devr->s1);
2708 	mlx5_ib_destroy_srq(devr->s0);
2709 	mlx5_ib_dealloc_xrcd(devr->x0);
2710 	mlx5_ib_dealloc_xrcd(devr->x1);
2711 	mlx5_ib_destroy_cq(devr->c0);
2712 	mlx5_ib_dealloc_pd(devr->p0);
2713 
2714 	/* Make sure no change P_Key work items are still executing */
2715 	for (port = 0; port < dev->num_ports; ++port)
2716 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2717 }
2718 
2719 static u32 get_core_cap_flags(struct ib_device *ibdev)
2720 {
2721 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2722 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2723 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2724 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2725 	u32 ret = 0;
2726 
2727 	if (ll == IB_LINK_LAYER_INFINIBAND)
2728 		return RDMA_CORE_PORT_IBA_IB;
2729 
2730 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2731 		return 0;
2732 
2733 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2734 		return 0;
2735 
2736 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2737 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2738 
2739 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2740 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2741 
2742 	return ret;
2743 }
2744 
2745 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
2746 			       struct ib_port_immutable *immutable)
2747 {
2748 	struct ib_port_attr attr;
2749 	int err;
2750 
2751 	err = mlx5_ib_query_port(ibdev, port_num, &attr);
2752 	if (err)
2753 		return err;
2754 
2755 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2756 	immutable->gid_tbl_len = attr.gid_tbl_len;
2757 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
2758 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2759 
2760 	return 0;
2761 }
2762 
2763 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
2764 			   size_t str_len)
2765 {
2766 	struct mlx5_ib_dev *dev =
2767 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2768 	snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
2769 		       fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
2770 }
2771 
2772 static int mlx5_roce_lag_init(struct mlx5_ib_dev *dev)
2773 {
2774 	return 0;
2775 }
2776 
2777 static void mlx5_roce_lag_cleanup(struct mlx5_ib_dev *dev)
2778 {
2779 }
2780 
2781 static void mlx5_remove_roce_notifier(struct mlx5_ib_dev *dev)
2782 {
2783 	if (dev->roce.nb.notifier_call) {
2784 		unregister_netdevice_notifier(&dev->roce.nb);
2785 		dev->roce.nb.notifier_call = NULL;
2786 	}
2787 }
2788 
2789 static int mlx5_enable_roce(struct mlx5_ib_dev *dev)
2790 {
2791 	VNET_ITERATOR_DECL(vnet_iter);
2792 	struct net_device *idev;
2793 	int err;
2794 
2795 	/* Check if mlx5en net device already exists */
2796 	VNET_LIST_RLOCK();
2797 	VNET_FOREACH(vnet_iter) {
2798 		IFNET_RLOCK();
2799 		CURVNET_SET_QUIET(vnet_iter);
2800 		CK_STAILQ_FOREACH(idev, &V_ifnet, if_link) {
2801 			/* check if network interface belongs to mlx5en */
2802 			if (!mlx5_netdev_match(idev, dev->mdev, "mce"))
2803 				continue;
2804 			write_lock(&dev->roce.netdev_lock);
2805 			dev->roce.netdev = idev;
2806 			write_unlock(&dev->roce.netdev_lock);
2807 		}
2808 		CURVNET_RESTORE();
2809 		IFNET_RUNLOCK();
2810 	}
2811 	VNET_LIST_RUNLOCK();
2812 
2813 	dev->roce.nb.notifier_call = mlx5_netdev_event;
2814 	err = register_netdevice_notifier(&dev->roce.nb);
2815 	if (err) {
2816 		dev->roce.nb.notifier_call = NULL;
2817 		return err;
2818 	}
2819 
2820 	err = mlx5_nic_vport_enable_roce(dev->mdev);
2821 	if (err)
2822 		goto err_unregister_netdevice_notifier;
2823 
2824 	err = mlx5_roce_lag_init(dev);
2825 	if (err)
2826 		goto err_disable_roce;
2827 
2828 	return 0;
2829 
2830 err_disable_roce:
2831 	mlx5_nic_vport_disable_roce(dev->mdev);
2832 
2833 err_unregister_netdevice_notifier:
2834 	mlx5_remove_roce_notifier(dev);
2835 	return err;
2836 }
2837 
2838 static void mlx5_disable_roce(struct mlx5_ib_dev *dev)
2839 {
2840 	mlx5_roce_lag_cleanup(dev);
2841 	mlx5_nic_vport_disable_roce(dev->mdev);
2842 }
2843 
2844 static void mlx5_ib_dealloc_q_port_counter(struct mlx5_ib_dev *dev, u8 port_num)
2845 {
2846 	mlx5_vport_dealloc_q_counter(dev->mdev,
2847 				     MLX5_INTERFACE_PROTOCOL_IB,
2848 				     dev->port[port_num].q_cnt_id);
2849 	dev->port[port_num].q_cnt_id = 0;
2850 }
2851 
2852 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
2853 {
2854 	unsigned int i;
2855 
2856 	for (i = 0; i < dev->num_ports; i++)
2857 		mlx5_ib_dealloc_q_port_counter(dev, i);
2858 }
2859 
2860 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
2861 {
2862 	int i;
2863 	int ret;
2864 
2865 	for (i = 0; i < dev->num_ports; i++) {
2866 		ret = mlx5_vport_alloc_q_counter(dev->mdev,
2867 						 MLX5_INTERFACE_PROTOCOL_IB,
2868 						 &dev->port[i].q_cnt_id);
2869 		if (ret) {
2870 			mlx5_ib_warn(dev,
2871 				     "couldn't allocate queue counter for port %d, err %d\n",
2872 				     i + 1, ret);
2873 			goto dealloc_counters;
2874 		}
2875 	}
2876 
2877 	return 0;
2878 
2879 dealloc_counters:
2880 	while (--i >= 0)
2881 		mlx5_ib_dealloc_q_port_counter(dev, i);
2882 
2883 	return ret;
2884 }
2885 
2886 static const char * const names[] = {
2887 	"rx_write_requests",
2888 	"rx_read_requests",
2889 	"rx_atomic_requests",
2890 	"out_of_buffer",
2891 	"out_of_sequence",
2892 	"duplicate_request",
2893 	"rnr_nak_retry_err",
2894 	"packet_seq_err",
2895 	"implied_nak_seq_err",
2896 	"local_ack_timeout_err",
2897 };
2898 
2899 static const size_t stats_offsets[] = {
2900 	MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests),
2901 	MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests),
2902 	MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests),
2903 	MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer),
2904 	MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence),
2905 	MLX5_BYTE_OFF(query_q_counter_out, duplicate_request),
2906 	MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err),
2907 	MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err),
2908 	MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err),
2909 	MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err),
2910 };
2911 
2912 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
2913 						    u8 port_num)
2914 {
2915 	BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets));
2916 
2917 	/* We support only per port stats */
2918 	if (port_num == 0)
2919 		return NULL;
2920 
2921 	return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names),
2922 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
2923 }
2924 
2925 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
2926 				struct rdma_hw_stats *stats,
2927 				u8 port, int index)
2928 {
2929 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2930 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
2931 	void *out;
2932 	__be32 val;
2933 	int ret;
2934 	int i;
2935 
2936 	if (!port || !stats)
2937 		return -ENOSYS;
2938 
2939 	out = mlx5_vzalloc(outlen);
2940 	if (!out)
2941 		return -ENOMEM;
2942 
2943 	ret = mlx5_vport_query_q_counter(dev->mdev,
2944 					dev->port[port - 1].q_cnt_id, 0,
2945 					out, outlen);
2946 	if (ret)
2947 		goto free;
2948 
2949 	for (i = 0; i < ARRAY_SIZE(names); i++) {
2950 		val = *(__be32 *)(out + stats_offsets[i]);
2951 		stats->value[i] = (u64)be32_to_cpu(val);
2952 	}
2953 free:
2954 	kvfree(out);
2955 	return ARRAY_SIZE(names);
2956 }
2957 
2958 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
2959 {
2960 	struct mlx5_ib_dev *dev;
2961 	enum rdma_link_layer ll;
2962 	int port_type_cap;
2963 	int err;
2964 	int i;
2965 
2966 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
2967 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
2968 
2969 	if ((ll == IB_LINK_LAYER_ETHERNET) && !MLX5_CAP_GEN(mdev, roce))
2970 		return NULL;
2971 
2972 	printk_once(KERN_INFO "%s", mlx5_version);
2973 
2974 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
2975 	if (!dev)
2976 		return NULL;
2977 
2978 	dev->mdev = mdev;
2979 
2980 	dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
2981 			    GFP_KERNEL);
2982 	if (!dev->port)
2983 		goto err_dealloc;
2984 
2985 	rwlock_init(&dev->roce.netdev_lock);
2986 	err = get_port_caps(dev);
2987 	if (err)
2988 		goto err_free_port;
2989 
2990 	if (mlx5_use_mad_ifc(dev))
2991 		get_ext_port_caps(dev);
2992 
2993 	MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
2994 
2995 	snprintf(dev->ib_dev.name, IB_DEVICE_NAME_MAX, "mlx5_%d", device_get_unit(mdev->pdev->dev.bsddev));
2996 	dev->ib_dev.owner		= THIS_MODULE;
2997 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
2998 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
2999 	dev->num_ports		= MLX5_CAP_GEN(mdev, num_ports);
3000 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
3001 	dev->ib_dev.num_comp_vectors    =
3002 		dev->mdev->priv.eq_table.num_comp_vectors;
3003 	dev->ib_dev.dma_device	= &mdev->pdev->dev;
3004 
3005 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
3006 	dev->ib_dev.uverbs_cmd_mask	=
3007 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
3008 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
3009 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
3010 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
3011 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
3012 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
3013 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
3014 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
3015 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
3016 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
3017 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
3018 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
3019 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
3020 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
3021 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
3022 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
3023 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
3024 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
3025 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
3026 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
3027 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
3028 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
3029 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
3030 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
3031 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
3032 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
3033 	dev->ib_dev.uverbs_ex_cmd_mask =
3034 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
3035 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
3036 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
3037 
3038 	dev->ib_dev.query_device	= mlx5_ib_query_device;
3039 	dev->ib_dev.query_port		= mlx5_ib_query_port;
3040 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
3041 	if (ll == IB_LINK_LAYER_ETHERNET)
3042 		dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
3043 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
3044 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
3045 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
3046 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
3047 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
3048 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
3049 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
3050 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
3051 	dev->ib_dev.mmap		= mlx5_ib_mmap;
3052 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
3053 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
3054 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
3055 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
3056 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
3057 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
3058 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
3059 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
3060 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
3061 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
3062 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
3063 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
3064 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
3065 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
3066 	dev->ib_dev.post_send		= mlx5_ib_post_send;
3067 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
3068 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
3069 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
3070 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
3071 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
3072 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
3073 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
3074 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
3075 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
3076 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
3077 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
3078 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
3079 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
3080 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
3081 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
3082 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
3083 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
3084 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
3085 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
3086 	if (mlx5_core_is_pf(mdev)) {
3087 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
3088 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
3089 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
3090 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
3091 	}
3092 
3093 	mlx5_ib_internal_fill_odp_caps(dev);
3094 
3095 	if (MLX5_CAP_GEN(mdev, imaicl)) {
3096 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
3097 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
3098 		dev->ib_dev.uverbs_cmd_mask |=
3099 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
3100 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
3101 	}
3102 
3103 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) &&
3104 	    MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3105 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
3106 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
3107 	}
3108 
3109 	if (MLX5_CAP_GEN(mdev, xrc)) {
3110 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
3111 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
3112 		dev->ib_dev.uverbs_cmd_mask |=
3113 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
3114 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
3115 	}
3116 
3117 	if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
3118 	    IB_LINK_LAYER_ETHERNET) {
3119 		dev->ib_dev.create_flow	= mlx5_ib_create_flow;
3120 		dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
3121 		dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
3122 		dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
3123 		dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
3124 		dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
3125 		dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
3126 		dev->ib_dev.uverbs_ex_cmd_mask |=
3127 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
3128 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
3129 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
3130 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
3131 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
3132 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
3133 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
3134 	}
3135 	err = init_node_data(dev);
3136 	if (err)
3137 		goto err_free_port;
3138 
3139 	mutex_init(&dev->flow_db.lock);
3140 	mutex_init(&dev->cap_mask_mutex);
3141 	INIT_LIST_HEAD(&dev->qp_list);
3142 	spin_lock_init(&dev->reset_flow_resource_lock);
3143 
3144 	if (ll == IB_LINK_LAYER_ETHERNET) {
3145 		err = mlx5_enable_roce(dev);
3146 		if (err)
3147 			goto err_free_port;
3148 	}
3149 
3150 	err = create_dev_resources(&dev->devr);
3151 	if (err)
3152 		goto err_disable_roce;
3153 
3154 	err = mlx5_ib_odp_init_one(dev);
3155 	if (err)
3156 		goto err_rsrc;
3157 
3158 	err = mlx5_ib_alloc_q_counters(dev);
3159 	if (err)
3160 		goto err_odp;
3161 
3162 	err = ib_register_device(&dev->ib_dev, NULL);
3163 	if (err)
3164 		goto err_q_cnt;
3165 
3166 	err = create_umr_res(dev);
3167 	if (err)
3168 		goto err_dev;
3169 
3170 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
3171 		err = device_create_file(&dev->ib_dev.dev,
3172 					 mlx5_class_attributes[i]);
3173 		if (err)
3174 			goto err_umrc;
3175 	}
3176 
3177 	err = mlx5_ib_init_congestion(dev);
3178 	if (err)
3179 		goto err_umrc;
3180 
3181 	dev->ib_active = true;
3182 
3183 	return dev;
3184 
3185 err_umrc:
3186 	destroy_umrc_res(dev);
3187 
3188 err_dev:
3189 	ib_unregister_device(&dev->ib_dev);
3190 
3191 err_q_cnt:
3192 	mlx5_ib_dealloc_q_counters(dev);
3193 
3194 err_odp:
3195 	mlx5_ib_odp_remove_one(dev);
3196 
3197 err_rsrc:
3198 	destroy_dev_resources(&dev->devr);
3199 
3200 err_disable_roce:
3201 	if (ll == IB_LINK_LAYER_ETHERNET) {
3202 		mlx5_disable_roce(dev);
3203 		mlx5_remove_roce_notifier(dev);
3204 	}
3205 
3206 err_free_port:
3207 	kfree(dev->port);
3208 
3209 err_dealloc:
3210 	ib_dealloc_device((struct ib_device *)dev);
3211 
3212 	return NULL;
3213 }
3214 
3215 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3216 {
3217 	struct mlx5_ib_dev *dev = context;
3218 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3219 
3220 	mlx5_ib_cleanup_congestion(dev);
3221 	mlx5_remove_roce_notifier(dev);
3222 	ib_unregister_device(&dev->ib_dev);
3223 	mlx5_ib_dealloc_q_counters(dev);
3224 	destroy_umrc_res(dev);
3225 	mlx5_ib_odp_remove_one(dev);
3226 	destroy_dev_resources(&dev->devr);
3227 	if (ll == IB_LINK_LAYER_ETHERNET)
3228 		mlx5_disable_roce(dev);
3229 	kfree(dev->port);
3230 	ib_dealloc_device(&dev->ib_dev);
3231 }
3232 
3233 static struct mlx5_interface mlx5_ib_interface = {
3234 	.add            = mlx5_ib_add,
3235 	.remove         = mlx5_ib_remove,
3236 	.event          = mlx5_ib_event,
3237 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
3238 };
3239 
3240 static int __init mlx5_ib_init(void)
3241 {
3242 	int err;
3243 
3244 	if (deprecated_prof_sel != 2)
3245 		pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
3246 
3247 	err = mlx5_ib_odp_init();
3248 	if (err)
3249 		return err;
3250 
3251 	err = mlx5_register_interface(&mlx5_ib_interface);
3252 	if (err)
3253 		goto clean_odp;
3254 
3255 	return err;
3256 
3257 clean_odp:
3258 	mlx5_ib_odp_cleanup();
3259 	return err;
3260 }
3261 
3262 static void __exit mlx5_ib_cleanup(void)
3263 {
3264 	mlx5_unregister_interface(&mlx5_ib_interface);
3265 	mlx5_ib_odp_cleanup();
3266 }
3267 
3268 module_init_order(mlx5_ib_init, SI_ORDER_THIRD);
3269 module_exit_order(mlx5_ib_cleanup, SI_ORDER_THIRD);
3270