xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_main.c (revision 0b57cec536236d46e3dba9bd041533462f33dbb7)
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 "mlx5ib"
54 #ifndef DRIVER_VERSION
55 #define DRIVER_VERSION "3.5.2"
56 #endif
57 #define DRIVER_RELDATE	"September 2019"
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 const char mlx5_version[] =
67 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver "
68 	DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
69 
70 enum {
71 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
72 };
73 
74 static enum rdma_link_layer
75 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
76 {
77 	switch (port_type_cap) {
78 	case MLX5_CAP_PORT_TYPE_IB:
79 		return IB_LINK_LAYER_INFINIBAND;
80 	case MLX5_CAP_PORT_TYPE_ETH:
81 		return IB_LINK_LAYER_ETHERNET;
82 	default:
83 		return IB_LINK_LAYER_UNSPECIFIED;
84 	}
85 }
86 
87 static enum rdma_link_layer
88 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
89 {
90 	struct mlx5_ib_dev *dev = to_mdev(device);
91 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
92 
93 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
94 }
95 
96 static bool mlx5_netdev_match(struct net_device *ndev,
97 			      struct mlx5_core_dev *mdev,
98 			      const char *dname)
99 {
100 	return ndev->if_type == IFT_ETHER &&
101 	  ndev->if_dname != NULL &&
102 	  strcmp(ndev->if_dname, dname) == 0 &&
103 	  ndev->if_softc != NULL &&
104 	  *(struct mlx5_core_dev **)ndev->if_softc == mdev;
105 }
106 
107 static int mlx5_netdev_event(struct notifier_block *this,
108 			     unsigned long event, void *ptr)
109 {
110 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
111 	struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
112 						 roce.nb);
113 
114 	switch (event) {
115 	case NETDEV_REGISTER:
116 	case NETDEV_UNREGISTER:
117 		write_lock(&ibdev->roce.netdev_lock);
118 		/* check if network interface belongs to mlx5en */
119 		if (mlx5_netdev_match(ndev, ibdev->mdev, "mce"))
120 			ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
121 					     NULL : ndev;
122 		write_unlock(&ibdev->roce.netdev_lock);
123 		break;
124 
125 	case NETDEV_UP:
126 	case NETDEV_DOWN: {
127 		struct net_device *upper = NULL;
128 
129 		if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
130 		    && ibdev->ib_active) {
131 			struct ib_event ibev = {0};
132 
133 			ibev.device = &ibdev->ib_dev;
134 			ibev.event = (event == NETDEV_UP) ?
135 				     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
136 			ibev.element.port_num = 1;
137 			ib_dispatch_event(&ibev);
138 		}
139 		break;
140 	}
141 
142 	default:
143 		break;
144 	}
145 
146 	return NOTIFY_DONE;
147 }
148 
149 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
150 					     u8 port_num)
151 {
152 	struct mlx5_ib_dev *ibdev = to_mdev(device);
153 	struct net_device *ndev;
154 
155 	/* Ensure ndev does not disappear before we invoke dev_hold()
156 	 */
157 	read_lock(&ibdev->roce.netdev_lock);
158 	ndev = ibdev->roce.netdev;
159 	if (ndev)
160 		dev_hold(ndev);
161 	read_unlock(&ibdev->roce.netdev_lock);
162 
163 	return ndev;
164 }
165 
166 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
167 				    u8 *active_width)
168 {
169 	switch (eth_proto_oper) {
170 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
171 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
172 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
173 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
174 		*active_width = IB_WIDTH_1X;
175 		*active_speed = IB_SPEED_SDR;
176 		break;
177 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
178 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
179 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
180 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
181 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
182 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
183 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER_LR):
184 		*active_width = IB_WIDTH_1X;
185 		*active_speed = IB_SPEED_QDR;
186 		break;
187 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
188 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
189 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
190 		*active_width = IB_WIDTH_1X;
191 		*active_speed = IB_SPEED_EDR;
192 		break;
193 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
194 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
195 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
196 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4_ER4):
197 		*active_width = IB_WIDTH_4X;
198 		*active_speed = IB_SPEED_QDR;
199 		break;
200 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
201 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
202 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
203 		*active_width = IB_WIDTH_1X;
204 		*active_speed = IB_SPEED_HDR;
205 		break;
206 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
207 		*active_width = IB_WIDTH_4X;
208 		*active_speed = IB_SPEED_FDR;
209 		break;
210 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
211 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
212 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
213 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
214 		*active_width = IB_WIDTH_4X;
215 		*active_speed = IB_SPEED_EDR;
216 		break;
217 	default:
218 		*active_width = IB_WIDTH_4X;
219 		*active_speed = IB_SPEED_QDR;
220 		return -EINVAL;
221 	}
222 
223 	return 0;
224 }
225 
226 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u8 *active_speed,
227 					u8 *active_width)
228 {
229 	switch (eth_proto_oper) {
230 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
231 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
232 		*active_width = IB_WIDTH_1X;
233 		*active_speed = IB_SPEED_SDR;
234 		break;
235 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
236 		*active_width = IB_WIDTH_1X;
237 		*active_speed = IB_SPEED_DDR;
238 		break;
239 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
240 		*active_width = IB_WIDTH_1X;
241 		*active_speed = IB_SPEED_QDR;
242 		break;
243 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
244 		*active_width = IB_WIDTH_4X;
245 		*active_speed = IB_SPEED_QDR;
246 		break;
247 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
248 		*active_width = IB_WIDTH_1X;
249 		*active_speed = IB_SPEED_EDR;
250 		break;
251 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
252 		*active_width = IB_WIDTH_2X;
253 		*active_speed = IB_SPEED_EDR;
254 		break;
255 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
256 		*active_width = IB_WIDTH_1X;
257 		*active_speed = IB_SPEED_HDR;
258 		break;
259 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
260 		*active_width = IB_WIDTH_4X;
261 		*active_speed = IB_SPEED_EDR;
262 		break;
263 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
264 		*active_width = IB_WIDTH_2X;
265 		*active_speed = IB_SPEED_HDR;
266 		break;
267 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
268 		*active_width = IB_WIDTH_4X;
269 		*active_speed = IB_SPEED_HDR;
270 		break;
271 	default:
272 		*active_width = IB_WIDTH_4X;
273 		*active_speed = IB_SPEED_QDR;
274 		return -EINVAL;
275 	}
276 
277 	return 0;
278 }
279 
280 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
281 				struct ib_port_attr *props)
282 {
283 	struct mlx5_ib_dev *dev = to_mdev(device);
284 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {};
285 	struct net_device *ndev;
286 	enum ib_mtu ndev_ib_mtu;
287 	u16 qkey_viol_cntr;
288 	u32 eth_prot_oper;
289 	bool ext;
290 	int err;
291 
292 	memset(props, 0, sizeof(*props));
293 
294 	/* Possible bad flows are checked before filling out props so in case
295 	 * of an error it will still be zeroed out.
296 	 */
297 	err = mlx5_query_port_ptys(dev->mdev, out, sizeof(out), MLX5_PTYS_EN,
298 	    port_num);
299 	if (err)
300 		return err;
301 
302 	ext = MLX5_CAP_PCAM_FEATURE(dev->mdev, ptys_extended_ethernet);
303 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
304 
305 	if (ext)
306 		translate_eth_ext_proto_oper(eth_prot_oper, &props->active_speed,
307 		    &props->active_width);
308 	else
309 		translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
310 		    &props->active_width);
311 
312 	props->port_cap_flags  |= IB_PORT_CM_SUP;
313 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
314 
315 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
316 						roce_address_table_size);
317 	props->max_mtu          = IB_MTU_4096;
318 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
319 	props->pkey_tbl_len     = 1;
320 	props->state            = IB_PORT_DOWN;
321 	props->phys_state       = 3;
322 
323 	mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
324 	props->qkey_viol_cntr = qkey_viol_cntr;
325 
326 	ndev = mlx5_ib_get_netdev(device, port_num);
327 	if (!ndev)
328 		return 0;
329 
330 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
331 		props->state      = IB_PORT_ACTIVE;
332 		props->phys_state = 5;
333 	}
334 
335 	ndev_ib_mtu = iboe_get_mtu(ndev->if_mtu);
336 
337 	dev_put(ndev);
338 
339 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
340 	return 0;
341 }
342 
343 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
344 				     const struct ib_gid_attr *attr,
345 				     void *mlx5_addr)
346 {
347 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
348 	char *mlx5_addr_l3_addr	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
349 					       source_l3_address);
350 	void *mlx5_addr_mac	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
351 					       source_mac_47_32);
352 	u16 vlan_id;
353 
354 	if (!gid)
355 		return;
356 	ether_addr_copy(mlx5_addr_mac, IF_LLADDR(attr->ndev));
357 
358 	vlan_id = rdma_vlan_dev_vlan_id(attr->ndev);
359 	if (vlan_id != 0xffff) {
360 		MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
361 		MLX5_SET_RA(mlx5_addr, vlan_id, vlan_id);
362 	}
363 
364 	switch (attr->gid_type) {
365 	case IB_GID_TYPE_IB:
366 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
367 		break;
368 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
369 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
370 		break;
371 
372 	default:
373 		WARN_ON(true);
374 	}
375 
376 	if (attr->gid_type != IB_GID_TYPE_IB) {
377 		if (ipv6_addr_v4mapped((void *)gid))
378 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
379 				    MLX5_ROCE_L3_TYPE_IPV4);
380 		else
381 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
382 				    MLX5_ROCE_L3_TYPE_IPV6);
383 	}
384 
385 	if ((attr->gid_type == IB_GID_TYPE_IB) ||
386 	    !ipv6_addr_v4mapped((void *)gid))
387 		memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
388 	else
389 		memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
390 }
391 
392 static int set_roce_addr(struct ib_device *device, u8 port_num,
393 			 unsigned int index,
394 			 const union ib_gid *gid,
395 			 const struct ib_gid_attr *attr)
396 {
397 	struct mlx5_ib_dev *dev = to_mdev(device);
398 	u32  in[MLX5_ST_SZ_DW(set_roce_address_in)]  = {0};
399 	u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0};
400 	void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
401 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
402 
403 	if (ll != IB_LINK_LAYER_ETHERNET)
404 		return -EINVAL;
405 
406 	ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
407 
408 	MLX5_SET(set_roce_address_in, in, roce_address_index, index);
409 	MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
410 	return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
411 }
412 
413 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
414 			   unsigned int index, const union ib_gid *gid,
415 			   const struct ib_gid_attr *attr,
416 			   __always_unused void **context)
417 {
418 	return set_roce_addr(device, port_num, index, gid, attr);
419 }
420 
421 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
422 			   unsigned int index, __always_unused void **context)
423 {
424 	return set_roce_addr(device, port_num, index, NULL, NULL);
425 }
426 
427 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
428 			       int index)
429 {
430 	struct ib_gid_attr attr;
431 	union ib_gid gid;
432 
433 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
434 		return 0;
435 
436 	if (!attr.ndev)
437 		return 0;
438 
439 	dev_put(attr.ndev);
440 
441 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
442 		return 0;
443 
444 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
445 }
446 
447 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
448 			   int index, enum ib_gid_type *gid_type)
449 {
450 	struct ib_gid_attr attr;
451 	union ib_gid gid;
452 	int ret;
453 
454 	ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
455 	if (ret)
456 		return ret;
457 
458 	if (!attr.ndev)
459 		return -ENODEV;
460 
461 	dev_put(attr.ndev);
462 
463 	*gid_type = attr.gid_type;
464 
465 	return 0;
466 }
467 
468 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
469 {
470 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
471 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
472 	return 0;
473 }
474 
475 enum {
476 	MLX5_VPORT_ACCESS_METHOD_MAD,
477 	MLX5_VPORT_ACCESS_METHOD_HCA,
478 	MLX5_VPORT_ACCESS_METHOD_NIC,
479 };
480 
481 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
482 {
483 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
484 		return MLX5_VPORT_ACCESS_METHOD_MAD;
485 
486 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
487 	    IB_LINK_LAYER_ETHERNET)
488 		return MLX5_VPORT_ACCESS_METHOD_NIC;
489 
490 	return MLX5_VPORT_ACCESS_METHOD_HCA;
491 }
492 
493 static void get_atomic_caps(struct mlx5_ib_dev *dev,
494 			    struct ib_device_attr *props)
495 {
496 	u8 tmp;
497 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
498 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
499 	u8 atomic_req_8B_endianness_mode =
500 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
501 
502 	/* Check if HW supports 8 bytes standard atomic operations and capable
503 	 * of host endianness respond
504 	 */
505 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
506 	if (((atomic_operations & tmp) == tmp) &&
507 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
508 	    (atomic_req_8B_endianness_mode)) {
509 		props->atomic_cap = IB_ATOMIC_HCA;
510 	} else {
511 		props->atomic_cap = IB_ATOMIC_NONE;
512 	}
513 }
514 
515 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
516 					__be64 *sys_image_guid)
517 {
518 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
519 	struct mlx5_core_dev *mdev = dev->mdev;
520 	u64 tmp;
521 	int err;
522 
523 	switch (mlx5_get_vport_access_method(ibdev)) {
524 	case MLX5_VPORT_ACCESS_METHOD_MAD:
525 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
526 							    sys_image_guid);
527 
528 	case MLX5_VPORT_ACCESS_METHOD_HCA:
529 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
530 		break;
531 
532 	case MLX5_VPORT_ACCESS_METHOD_NIC:
533 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
534 		break;
535 
536 	default:
537 		return -EINVAL;
538 	}
539 
540 	if (!err)
541 		*sys_image_guid = cpu_to_be64(tmp);
542 
543 	return err;
544 
545 }
546 
547 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
548 				u16 *max_pkeys)
549 {
550 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
551 	struct mlx5_core_dev *mdev = dev->mdev;
552 
553 	switch (mlx5_get_vport_access_method(ibdev)) {
554 	case MLX5_VPORT_ACCESS_METHOD_MAD:
555 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
556 
557 	case MLX5_VPORT_ACCESS_METHOD_HCA:
558 	case MLX5_VPORT_ACCESS_METHOD_NIC:
559 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
560 						pkey_table_size));
561 		return 0;
562 
563 	default:
564 		return -EINVAL;
565 	}
566 }
567 
568 static int mlx5_query_vendor_id(struct ib_device *ibdev,
569 				u32 *vendor_id)
570 {
571 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
572 
573 	switch (mlx5_get_vport_access_method(ibdev)) {
574 	case MLX5_VPORT_ACCESS_METHOD_MAD:
575 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
576 
577 	case MLX5_VPORT_ACCESS_METHOD_HCA:
578 	case MLX5_VPORT_ACCESS_METHOD_NIC:
579 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
580 
581 	default:
582 		return -EINVAL;
583 	}
584 }
585 
586 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
587 				__be64 *node_guid)
588 {
589 	u64 tmp;
590 	int err;
591 
592 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
593 	case MLX5_VPORT_ACCESS_METHOD_MAD:
594 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
595 
596 	case MLX5_VPORT_ACCESS_METHOD_HCA:
597 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
598 		break;
599 
600 	case MLX5_VPORT_ACCESS_METHOD_NIC:
601 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
602 		break;
603 
604 	default:
605 		return -EINVAL;
606 	}
607 
608 	if (!err)
609 		*node_guid = cpu_to_be64(tmp);
610 
611 	return err;
612 }
613 
614 struct mlx5_reg_node_desc {
615 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
616 };
617 
618 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
619 {
620 	struct mlx5_reg_node_desc in;
621 
622 	if (mlx5_use_mad_ifc(dev))
623 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
624 
625 	memset(&in, 0, sizeof(in));
626 
627 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
628 				    sizeof(struct mlx5_reg_node_desc),
629 				    MLX5_REG_NODE_DESC, 0, 0);
630 }
631 
632 static int mlx5_ib_query_device(struct ib_device *ibdev,
633 				struct ib_device_attr *props,
634 				struct ib_udata *uhw)
635 {
636 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
637 	struct mlx5_core_dev *mdev = dev->mdev;
638 	int err = -ENOMEM;
639 	int max_sq_desc;
640 	int max_rq_sg;
641 	int max_sq_sg;
642 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
643 	struct mlx5_ib_query_device_resp resp = {};
644 	size_t resp_len;
645 	u64 max_tso;
646 
647 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
648 	if (uhw->outlen && uhw->outlen < resp_len)
649 		return -EINVAL;
650 	else
651 		resp.response_length = resp_len;
652 
653 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
654 		return -EINVAL;
655 
656 	memset(props, 0, sizeof(*props));
657 	err = mlx5_query_system_image_guid(ibdev,
658 					   &props->sys_image_guid);
659 	if (err)
660 		return err;
661 
662 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
663 	if (err)
664 		return err;
665 
666 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
667 	if (err)
668 		return err;
669 
670 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
671 		((u32)fw_rev_min(dev->mdev) << 16) |
672 		fw_rev_sub(dev->mdev);
673 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
674 		IB_DEVICE_PORT_ACTIVE_EVENT		|
675 		IB_DEVICE_SYS_IMAGE_GUID		|
676 		IB_DEVICE_RC_RNR_NAK_GEN;
677 
678 	if (MLX5_CAP_GEN(mdev, pkv))
679 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
680 	if (MLX5_CAP_GEN(mdev, qkv))
681 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
682 	if (MLX5_CAP_GEN(mdev, apm))
683 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
684 	if (MLX5_CAP_GEN(mdev, xrc))
685 		props->device_cap_flags |= IB_DEVICE_XRC;
686 	if (MLX5_CAP_GEN(mdev, imaicl)) {
687 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
688 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
689 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
690 		/* We support 'Gappy' memory registration too */
691 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
692 	}
693 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
694 	if (MLX5_CAP_GEN(mdev, sho)) {
695 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
696 		/* At this stage no support for signature handover */
697 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
698 				      IB_PROT_T10DIF_TYPE_2 |
699 				      IB_PROT_T10DIF_TYPE_3;
700 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
701 				       IB_GUARD_T10DIF_CSUM;
702 	}
703 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
704 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
705 
706 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
707 		if (MLX5_CAP_ETH(mdev, csum_cap))
708 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
709 
710 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
711 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
712 			if (max_tso) {
713 				resp.tso_caps.max_tso = 1 << max_tso;
714 				resp.tso_caps.supported_qpts |=
715 					1 << IB_QPT_RAW_PACKET;
716 				resp.response_length += sizeof(resp.tso_caps);
717 			}
718 		}
719 
720 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
721 			resp.rss_caps.rx_hash_function =
722 						MLX5_RX_HASH_FUNC_TOEPLITZ;
723 			resp.rss_caps.rx_hash_fields_mask =
724 						MLX5_RX_HASH_SRC_IPV4 |
725 						MLX5_RX_HASH_DST_IPV4 |
726 						MLX5_RX_HASH_SRC_IPV6 |
727 						MLX5_RX_HASH_DST_IPV6 |
728 						MLX5_RX_HASH_SRC_PORT_TCP |
729 						MLX5_RX_HASH_DST_PORT_TCP |
730 						MLX5_RX_HASH_SRC_PORT_UDP |
731 						MLX5_RX_HASH_DST_PORT_UDP;
732 			resp.response_length += sizeof(resp.rss_caps);
733 		}
734 	} else {
735 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
736 			resp.response_length += sizeof(resp.tso_caps);
737 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
738 			resp.response_length += sizeof(resp.rss_caps);
739 	}
740 
741 	if (MLX5_CAP_GEN(mdev, ipoib_ipoib_offloads)) {
742 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
743 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
744 	}
745 
746 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
747 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs))
748 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
749 
750 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
751 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
752 
753 	props->vendor_part_id	   = mdev->pdev->device;
754 	props->hw_ver		   = mdev->pdev->revision;
755 
756 	props->max_mr_size	   = ~0ull;
757 	props->page_size_cap	   = ~(min_page_size - 1);
758 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
759 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
760 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
761 		     sizeof(struct mlx5_wqe_data_seg);
762 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
763 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
764 		     sizeof(struct mlx5_wqe_raddr_seg)) /
765 		sizeof(struct mlx5_wqe_data_seg);
766 	props->max_sge = min(max_rq_sg, max_sq_sg);
767 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
768 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
769 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
770 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
771 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
772 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
773 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
774 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
775 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
776 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
777 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
778 	props->max_srq_sge	   = max_rq_sg - 1;
779 	props->max_fast_reg_page_list_len =
780 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
781 	get_atomic_caps(dev, props);
782 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
783 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
784 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
785 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
786 					   props->max_mcast_grp;
787 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
788 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
789 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
790 
791 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
792 	if (MLX5_CAP_GEN(mdev, pg))
793 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
794 	props->odp_caps = dev->odp_caps;
795 #endif
796 
797 	if (MLX5_CAP_GEN(mdev, cd))
798 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
799 
800 	if (!mlx5_core_is_pf(mdev))
801 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
802 
803 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
804 	    IB_LINK_LAYER_ETHERNET) {
805 		props->rss_caps.max_rwq_indirection_tables =
806 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
807 		props->rss_caps.max_rwq_indirection_table_size =
808 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
809 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
810 		props->max_wq_type_rq =
811 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
812 	}
813 
814 	if (uhw->outlen) {
815 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
816 
817 		if (err)
818 			return err;
819 	}
820 
821 	return 0;
822 }
823 
824 enum mlx5_ib_width {
825 	MLX5_IB_WIDTH_1X	= 1 << 0,
826 	MLX5_IB_WIDTH_2X	= 1 << 1,
827 	MLX5_IB_WIDTH_4X	= 1 << 2,
828 	MLX5_IB_WIDTH_8X	= 1 << 3,
829 	MLX5_IB_WIDTH_12X	= 1 << 4
830 };
831 
832 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
833 				  u8 *ib_width)
834 {
835 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
836 	int err = 0;
837 
838 	if (active_width & MLX5_IB_WIDTH_1X) {
839 		*ib_width = IB_WIDTH_1X;
840 	} else if (active_width & MLX5_IB_WIDTH_2X) {
841 		*ib_width = IB_WIDTH_2X;
842 	} else if (active_width & MLX5_IB_WIDTH_4X) {
843 		*ib_width = IB_WIDTH_4X;
844 	} else if (active_width & MLX5_IB_WIDTH_8X) {
845 		*ib_width = IB_WIDTH_8X;
846 	} else if (active_width & MLX5_IB_WIDTH_12X) {
847 		*ib_width = IB_WIDTH_12X;
848 	} else {
849 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
850 			    (int)active_width);
851 		err = -EINVAL;
852 	}
853 
854 	return err;
855 }
856 
857 enum ib_max_vl_num {
858 	__IB_MAX_VL_0		= 1,
859 	__IB_MAX_VL_0_1		= 2,
860 	__IB_MAX_VL_0_3		= 3,
861 	__IB_MAX_VL_0_7		= 4,
862 	__IB_MAX_VL_0_14	= 5,
863 };
864 
865 enum mlx5_vl_hw_cap {
866 	MLX5_VL_HW_0	= 1,
867 	MLX5_VL_HW_0_1	= 2,
868 	MLX5_VL_HW_0_2	= 3,
869 	MLX5_VL_HW_0_3	= 4,
870 	MLX5_VL_HW_0_4	= 5,
871 	MLX5_VL_HW_0_5	= 6,
872 	MLX5_VL_HW_0_6	= 7,
873 	MLX5_VL_HW_0_7	= 8,
874 	MLX5_VL_HW_0_14	= 15
875 };
876 
877 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
878 				u8 *max_vl_num)
879 {
880 	switch (vl_hw_cap) {
881 	case MLX5_VL_HW_0:
882 		*max_vl_num = __IB_MAX_VL_0;
883 		break;
884 	case MLX5_VL_HW_0_1:
885 		*max_vl_num = __IB_MAX_VL_0_1;
886 		break;
887 	case MLX5_VL_HW_0_3:
888 		*max_vl_num = __IB_MAX_VL_0_3;
889 		break;
890 	case MLX5_VL_HW_0_7:
891 		*max_vl_num = __IB_MAX_VL_0_7;
892 		break;
893 	case MLX5_VL_HW_0_14:
894 		*max_vl_num = __IB_MAX_VL_0_14;
895 		break;
896 
897 	default:
898 		return -EINVAL;
899 	}
900 
901 	return 0;
902 }
903 
904 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
905 			       struct ib_port_attr *props)
906 {
907 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
908 	struct mlx5_core_dev *mdev = dev->mdev;
909 	u32 *rep;
910 	int replen = MLX5_ST_SZ_BYTES(query_hca_vport_context_out);
911 	struct mlx5_ptys_reg *ptys;
912 	struct mlx5_pmtu_reg *pmtu;
913 	struct mlx5_pvlc_reg pvlc;
914 	void *ctx;
915 	int err;
916 
917 	rep = mlx5_vzalloc(replen);
918 	ptys = kzalloc(sizeof(*ptys), GFP_KERNEL);
919 	pmtu = kzalloc(sizeof(*pmtu), GFP_KERNEL);
920 	if (!rep || !ptys || !pmtu) {
921 		err = -ENOMEM;
922 		goto out;
923 	}
924 
925 	memset(props, 0, sizeof(*props));
926 
927 	err = mlx5_query_hca_vport_context(mdev, port, 0, rep, replen);
928 	if (err)
929 		goto out;
930 
931 	ctx = MLX5_ADDR_OF(query_hca_vport_context_out, rep, hca_vport_context);
932 
933 	props->lid		= MLX5_GET(hca_vport_context, ctx, lid);
934 	props->lmc		= MLX5_GET(hca_vport_context, ctx, lmc);
935 	props->sm_lid		= MLX5_GET(hca_vport_context, ctx, sm_lid);
936 	props->sm_sl		= MLX5_GET(hca_vport_context, ctx, sm_sl);
937 	props->state		= MLX5_GET(hca_vport_context, ctx, vport_state);
938 	props->phys_state	= MLX5_GET(hca_vport_context, ctx,
939 					port_physical_state);
940 	props->port_cap_flags	= MLX5_GET(hca_vport_context, ctx, cap_mask1);
941 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
942 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
943 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
944 	props->bad_pkey_cntr	= MLX5_GET(hca_vport_context, ctx,
945 					pkey_violation_counter);
946 	props->qkey_viol_cntr	= MLX5_GET(hca_vport_context, ctx,
947 					qkey_violation_counter);
948 	props->subnet_timeout	= MLX5_GET(hca_vport_context, ctx,
949 					subnet_timeout);
950 	props->init_type_reply	= MLX5_GET(hca_vport_context, ctx,
951 					init_type_reply);
952 	props->grh_required	= MLX5_GET(hca_vport_context, ctx, grh_required);
953 
954 	ptys->proto_mask |= MLX5_PTYS_IB;
955 	ptys->local_port = port;
956 	err = mlx5_core_access_ptys(mdev, ptys, 0);
957 	if (err)
958 		goto out;
959 
960 	err = translate_active_width(ibdev, ptys->ib_link_width_oper,
961 				     &props->active_width);
962 	if (err)
963 		goto out;
964 
965 	props->active_speed	= (u8)ptys->ib_proto_oper;
966 
967 	pmtu->local_port = port;
968 	err = mlx5_core_access_pmtu(mdev, pmtu, 0);
969 	if (err)
970 		goto out;
971 
972 	props->max_mtu		= pmtu->max_mtu;
973 	props->active_mtu	= pmtu->oper_mtu;
974 
975 	memset(&pvlc, 0, sizeof(pvlc));
976 	pvlc.local_port = port;
977 	err = mlx5_core_access_pvlc(mdev, &pvlc, 0);
978 	if (err)
979 		goto out;
980 
981 	err = translate_max_vl_num(ibdev, pvlc.vl_hw_cap,
982 				   &props->max_vl_num);
983 out:
984 	kvfree(rep);
985 	kfree(ptys);
986 	kfree(pmtu);
987 	return err;
988 }
989 
990 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
991 		       struct ib_port_attr *props)
992 {
993 	switch (mlx5_get_vport_access_method(ibdev)) {
994 	case MLX5_VPORT_ACCESS_METHOD_MAD:
995 		return mlx5_query_mad_ifc_port(ibdev, port, props);
996 
997 	case MLX5_VPORT_ACCESS_METHOD_HCA:
998 		return mlx5_query_hca_port(ibdev, port, props);
999 
1000 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1001 		return mlx5_query_port_roce(ibdev, port, props);
1002 
1003 	default:
1004 		return -EINVAL;
1005 	}
1006 }
1007 
1008 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1009 			     union ib_gid *gid)
1010 {
1011 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1012 	struct mlx5_core_dev *mdev = dev->mdev;
1013 
1014 	switch (mlx5_get_vport_access_method(ibdev)) {
1015 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1016 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1017 
1018 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1019 		return mlx5_query_hca_vport_gid(mdev, port, 0, index, gid);
1020 
1021 	default:
1022 		return -EINVAL;
1023 	}
1024 
1025 }
1026 
1027 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1028 			      u16 *pkey)
1029 {
1030 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1031 	struct mlx5_core_dev *mdev = dev->mdev;
1032 
1033 	switch (mlx5_get_vport_access_method(ibdev)) {
1034 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1035 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1036 
1037 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1038 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1039 		return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
1040 						 pkey);
1041 	default:
1042 		return -EINVAL;
1043 	}
1044 }
1045 
1046 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1047 				 struct ib_device_modify *props)
1048 {
1049 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1050 	struct mlx5_reg_node_desc in;
1051 	struct mlx5_reg_node_desc out;
1052 	int err;
1053 
1054 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1055 		return -EOPNOTSUPP;
1056 
1057 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1058 		return 0;
1059 
1060 	/*
1061 	 * If possible, pass node desc to FW, so it can generate
1062 	 * a 144 trap.  If cmd fails, just ignore.
1063 	 */
1064 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1065 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1066 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1067 	if (err)
1068 		return err;
1069 
1070 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1071 
1072 	return err;
1073 }
1074 
1075 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1076 			       struct ib_port_modify *props)
1077 {
1078 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1079 	struct ib_port_attr attr;
1080 	u32 tmp;
1081 	int err;
1082 
1083 	/*
1084 	 * CM layer calls ib_modify_port() regardless of the link
1085 	 * layer. For Ethernet ports, qkey violation and Port
1086 	 * capabilities are meaningless.
1087 	 */
1088 	if (mlx5_ib_port_link_layer(ibdev, port) == IB_LINK_LAYER_ETHERNET)
1089 		return 0;
1090 
1091 	mutex_lock(&dev->cap_mask_mutex);
1092 
1093 	err = mlx5_ib_query_port(ibdev, port, &attr);
1094 	if (err)
1095 		goto out;
1096 
1097 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1098 		~props->clr_port_cap_mask;
1099 
1100 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1101 
1102 out:
1103 	mutex_unlock(&dev->cap_mask_mutex);
1104 	return err;
1105 }
1106 
1107 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1108 						  struct ib_udata *udata)
1109 {
1110 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1111 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1112 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1113 	struct mlx5_ib_ucontext *context;
1114 	struct mlx5_uuar_info *uuari;
1115 	struct mlx5_uar *uars;
1116 	int gross_uuars;
1117 	int num_uars;
1118 	int ver;
1119 	int uuarn;
1120 	int err;
1121 	int i;
1122 	size_t reqlen;
1123 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1124 				     max_cqe_version);
1125 
1126 	if (!dev->ib_active)
1127 		return ERR_PTR(-EAGAIN);
1128 
1129 	if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
1130 		return ERR_PTR(-EINVAL);
1131 
1132 	reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
1133 	if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1134 		ver = 0;
1135 	else if (reqlen >= min_req_v2)
1136 		ver = 2;
1137 	else
1138 		return ERR_PTR(-EINVAL);
1139 
1140 	err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
1141 	if (err)
1142 		return ERR_PTR(err);
1143 
1144 	if (req.flags)
1145 		return ERR_PTR(-EINVAL);
1146 
1147 	if (req.total_num_uuars > MLX5_MAX_UUARS)
1148 		return ERR_PTR(-ENOMEM);
1149 
1150 	if (req.total_num_uuars == 0)
1151 		return ERR_PTR(-EINVAL);
1152 
1153 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1154 		return ERR_PTR(-EOPNOTSUPP);
1155 
1156 	if (reqlen > sizeof(req) &&
1157 	    !ib_is_udata_cleared(udata, sizeof(req),
1158 				 reqlen - sizeof(req)))
1159 		return ERR_PTR(-EOPNOTSUPP);
1160 
1161 	req.total_num_uuars = ALIGN(req.total_num_uuars,
1162 				    MLX5_NON_FP_BF_REGS_PER_PAGE);
1163 	if (req.num_low_latency_uuars > req.total_num_uuars - 1)
1164 		return ERR_PTR(-EINVAL);
1165 
1166 	num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
1167 	gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
1168 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1169 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1170 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1171 	resp.cache_line_size = cache_line_size();
1172 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1173 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1174 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1175 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1176 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1177 	resp.cqe_version = min_t(__u8,
1178 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1179 				 req.max_cqe_version);
1180 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1181 				   sizeof(resp.response_length), udata->outlen);
1182 
1183 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1184 	if (!context)
1185 		return ERR_PTR(-ENOMEM);
1186 
1187 	uuari = &context->uuari;
1188 	mutex_init(&uuari->lock);
1189 	uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
1190 	if (!uars) {
1191 		err = -ENOMEM;
1192 		goto out_ctx;
1193 	}
1194 
1195 	uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
1196 				sizeof(*uuari->bitmap),
1197 				GFP_KERNEL);
1198 	if (!uuari->bitmap) {
1199 		err = -ENOMEM;
1200 		goto out_uar_ctx;
1201 	}
1202 	/*
1203 	 * clear all fast path uuars
1204 	 */
1205 	for (i = 0; i < gross_uuars; i++) {
1206 		uuarn = i & 3;
1207 		if (uuarn == 2 || uuarn == 3)
1208 			set_bit(i, uuari->bitmap);
1209 	}
1210 
1211 	uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
1212 	if (!uuari->count) {
1213 		err = -ENOMEM;
1214 		goto out_bitmap;
1215 	}
1216 
1217 	for (i = 0; i < num_uars; i++) {
1218 		err = mlx5_cmd_alloc_uar(dev->mdev, &uars[i].index);
1219 		if (err)
1220 			goto out_count;
1221 	}
1222 
1223 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1224 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1225 #endif
1226 
1227 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1228 		err = mlx5_alloc_transport_domain(dev->mdev,
1229 						       &context->tdn);
1230 		if (err)
1231 			goto out_uars;
1232 	}
1233 
1234 	INIT_LIST_HEAD(&context->vma_private_list);
1235 	INIT_LIST_HEAD(&context->db_page_list);
1236 	mutex_init(&context->db_page_mutex);
1237 
1238 	resp.tot_uuars = req.total_num_uuars;
1239 	resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1240 
1241 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1242 		resp.response_length += sizeof(resp.cqe_version);
1243 
1244 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1245 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1246 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1247 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1248 	}
1249 
1250 	/*
1251 	 * We don't want to expose information from the PCI bar that is located
1252 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1253 	 * pretend we don't support reading the HCA's core clock. This is also
1254 	 * forced by mmap function.
1255 	 */
1256 	if (PAGE_SIZE <= 4096 &&
1257 	    field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1258 		resp.comp_mask |=
1259 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1260 		resp.hca_core_clock_offset =
1261 			offsetof(struct mlx5_init_seg, internal_timer_h) %
1262 			PAGE_SIZE;
1263 		resp.response_length += sizeof(resp.hca_core_clock_offset) +
1264 					sizeof(resp.reserved2);
1265 	}
1266 
1267 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1268 	if (err)
1269 		goto out_td;
1270 
1271 	uuari->ver = ver;
1272 	uuari->num_low_latency_uuars = req.num_low_latency_uuars;
1273 	uuari->uars = uars;
1274 	uuari->num_uars = num_uars;
1275 	context->cqe_version = resp.cqe_version;
1276 
1277 	return &context->ibucontext;
1278 
1279 out_td:
1280 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1281 		mlx5_dealloc_transport_domain(dev->mdev, context->tdn);
1282 
1283 out_uars:
1284 	for (i--; i >= 0; i--)
1285 		mlx5_cmd_free_uar(dev->mdev, uars[i].index);
1286 out_count:
1287 	kfree(uuari->count);
1288 
1289 out_bitmap:
1290 	kfree(uuari->bitmap);
1291 
1292 out_uar_ctx:
1293 	kfree(uars);
1294 
1295 out_ctx:
1296 	kfree(context);
1297 	return ERR_PTR(err);
1298 }
1299 
1300 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1301 {
1302 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1303 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1304 	struct mlx5_uuar_info *uuari = &context->uuari;
1305 	int i;
1306 
1307 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1308 		mlx5_dealloc_transport_domain(dev->mdev, context->tdn);
1309 
1310 	for (i = 0; i < uuari->num_uars; i++) {
1311 		if (mlx5_cmd_free_uar(dev->mdev, uuari->uars[i].index))
1312 			mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
1313 	}
1314 
1315 	kfree(uuari->count);
1316 	kfree(uuari->bitmap);
1317 	kfree(uuari->uars);
1318 	kfree(context);
1319 
1320 	return 0;
1321 }
1322 
1323 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
1324 {
1325 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + index;
1326 }
1327 
1328 static int get_command(unsigned long offset)
1329 {
1330 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1331 }
1332 
1333 static int get_arg(unsigned long offset)
1334 {
1335 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1336 }
1337 
1338 static int get_index(unsigned long offset)
1339 {
1340 	return get_arg(offset);
1341 }
1342 
1343 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1344 {
1345 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1346 	 * is done through either mremap flow or split_vma (usually due to
1347 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1348 	 * as this VMA is strongly hardware related.  Therefore we set the
1349 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1350 	 * calling us again and trying to do incorrect actions.  We assume that
1351 	 * the original VMA size is exactly a single page, and therefore all
1352 	 * "splitting" operation will not happen to it.
1353 	 */
1354 	area->vm_ops = NULL;
1355 }
1356 
1357 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1358 {
1359 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1360 
1361 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1362 	 * file itself is closed, therefore no sync is needed with the regular
1363 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1364 	 * However need a sync with accessing the vma as part of
1365 	 * mlx5_ib_disassociate_ucontext.
1366 	 * The close operation is usually called under mm->mmap_sem except when
1367 	 * process is exiting.
1368 	 * The exiting case is handled explicitly as part of
1369 	 * mlx5_ib_disassociate_ucontext.
1370 	 */
1371 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1372 
1373 	/* setting the vma context pointer to null in the mlx5_ib driver's
1374 	 * private data, to protect a race condition in
1375 	 * mlx5_ib_disassociate_ucontext().
1376 	 */
1377 	mlx5_ib_vma_priv_data->vma = NULL;
1378 	list_del(&mlx5_ib_vma_priv_data->list);
1379 	kfree(mlx5_ib_vma_priv_data);
1380 }
1381 
1382 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1383 	.open = mlx5_ib_vma_open,
1384 	.close = mlx5_ib_vma_close
1385 };
1386 
1387 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1388 				struct mlx5_ib_ucontext *ctx)
1389 {
1390 	struct mlx5_ib_vma_private_data *vma_prv;
1391 	struct list_head *vma_head = &ctx->vma_private_list;
1392 
1393 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1394 	if (!vma_prv)
1395 		return -ENOMEM;
1396 
1397 	vma_prv->vma = vma;
1398 	vma->vm_private_data = vma_prv;
1399 	vma->vm_ops =  &mlx5_ib_vm_ops;
1400 
1401 	list_add(&vma_prv->list, vma_head);
1402 
1403 	return 0;
1404 }
1405 
1406 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1407 {
1408 	int ret;
1409 	struct vm_area_struct *vma;
1410 	struct mlx5_ib_vma_private_data *vma_private, *n;
1411 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1412 	struct task_struct *owning_process  = NULL;
1413 	struct mm_struct   *owning_mm       = NULL;
1414 
1415 	owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1416 	if (!owning_process)
1417 		return;
1418 
1419 	owning_mm = get_task_mm(owning_process);
1420 	if (!owning_mm) {
1421 		pr_info("no mm, disassociate ucontext is pending task termination\n");
1422 		while (1) {
1423 			put_task_struct(owning_process);
1424 			usleep_range(1000, 2000);
1425 			owning_process = get_pid_task(ibcontext->tgid,
1426 						      PIDTYPE_PID);
1427 			if (!owning_process || owning_process->task_thread->
1428 			    td_proc->p_state == PRS_ZOMBIE) {
1429 				pr_info("disassociate ucontext done, task was terminated\n");
1430 				/* in case task was dead need to release the
1431 				 * task struct.
1432 				 */
1433 				if (owning_process)
1434 					put_task_struct(owning_process);
1435 				return;
1436 			}
1437 		}
1438 	}
1439 
1440 	/* need to protect from a race on closing the vma as part of
1441 	 * mlx5_ib_vma_close.
1442 	 */
1443 	down_write(&owning_mm->mmap_sem);
1444 	list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1445 				 list) {
1446 		vma = vma_private->vma;
1447 		ret = zap_vma_ptes(vma, vma->vm_start,
1448 				   PAGE_SIZE);
1449 		if (ret == -ENOTSUP) {
1450 			if (bootverbose)
1451 				WARN_ONCE(
1452 	"%s: zap_vma_ptes not implemented for unmanaged mappings", __func__);
1453 		} else {
1454 			WARN(ret, "%s: zap_vma_ptes failed, error %d",
1455 			    __func__, -ret);
1456 		}
1457 		/* context going to be destroyed, should
1458 		 * not access ops any more.
1459 		 */
1460 		/* XXXKIB vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE); */
1461 		vma->vm_ops = NULL;
1462 		list_del(&vma_private->list);
1463 		kfree(vma_private);
1464 	}
1465 	up_write(&owning_mm->mmap_sem);
1466 	mmput(owning_mm);
1467 	put_task_struct(owning_process);
1468 }
1469 
1470 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1471 {
1472 	switch (cmd) {
1473 	case MLX5_IB_MMAP_WC_PAGE:
1474 		return "WC";
1475 	case MLX5_IB_MMAP_REGULAR_PAGE:
1476 		return "best effort WC";
1477 	case MLX5_IB_MMAP_NC_PAGE:
1478 		return "NC";
1479 	default:
1480 		return NULL;
1481 	}
1482 }
1483 
1484 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1485 		    struct vm_area_struct *vma,
1486 		    struct mlx5_ib_ucontext *context)
1487 {
1488 	struct mlx5_uuar_info *uuari = &context->uuari;
1489 	int err;
1490 	unsigned long idx;
1491 	phys_addr_t pfn, pa;
1492 	pgprot_t prot;
1493 
1494 	switch (cmd) {
1495 	case MLX5_IB_MMAP_WC_PAGE:
1496 /* Some architectures don't support WC memory */
1497 #if defined(CONFIG_X86)
1498 		if (!pat_enabled())
1499 			return -EPERM;
1500 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1501 			return -EPERM;
1502 #endif
1503 	/* fall through */
1504 	case MLX5_IB_MMAP_REGULAR_PAGE:
1505 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1506 		prot = pgprot_writecombine(vma->vm_page_prot);
1507 		break;
1508 	case MLX5_IB_MMAP_NC_PAGE:
1509 		prot = pgprot_noncached(vma->vm_page_prot);
1510 		break;
1511 	default:
1512 		return -EINVAL;
1513 	}
1514 
1515 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1516 		return -EINVAL;
1517 
1518 	idx = get_index(vma->vm_pgoff);
1519 	if (idx >= uuari->num_uars)
1520 		return -EINVAL;
1521 
1522 	pfn = uar_index2pfn(dev, uuari->uars[idx].index);
1523 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1524 
1525 	vma->vm_page_prot = prot;
1526 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1527 				 PAGE_SIZE, vma->vm_page_prot);
1528 	if (err) {
1529 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%llx, pfn=%pa, mmap_cmd=%s\n",
1530 			    err, (unsigned long long)vma->vm_start, &pfn, mmap_cmd2str(cmd));
1531 		return -EAGAIN;
1532 	}
1533 
1534 	pa = pfn << PAGE_SHIFT;
1535 	mlx5_ib_dbg(dev, "mapped %s at 0x%llx, PA %pa\n", mmap_cmd2str(cmd),
1536 		    (unsigned long long)vma->vm_start, &pa);
1537 
1538 	return mlx5_ib_set_vma_data(vma, context);
1539 }
1540 
1541 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1542 {
1543 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1544 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1545 	unsigned long command;
1546 	phys_addr_t pfn;
1547 
1548 	command = get_command(vma->vm_pgoff);
1549 	switch (command) {
1550 	case MLX5_IB_MMAP_WC_PAGE:
1551 	case MLX5_IB_MMAP_NC_PAGE:
1552 	case MLX5_IB_MMAP_REGULAR_PAGE:
1553 		return uar_mmap(dev, command, vma, context);
1554 
1555 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1556 		return -ENOSYS;
1557 
1558 	case MLX5_IB_MMAP_CORE_CLOCK:
1559 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1560 			return -EINVAL;
1561 
1562 		if (vma->vm_flags & VM_WRITE)
1563 			return -EPERM;
1564 
1565 		/* Don't expose to user-space information it shouldn't have */
1566 		if (PAGE_SIZE > 4096)
1567 			return -EOPNOTSUPP;
1568 
1569 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1570 		pfn = (dev->mdev->iseg_base +
1571 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1572 			PAGE_SHIFT;
1573 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1574 				       PAGE_SIZE, vma->vm_page_prot))
1575 			return -EAGAIN;
1576 
1577 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%llx, PA 0x%llx\n",
1578 			    (unsigned long long)vma->vm_start,
1579 			    (unsigned long long)pfn << PAGE_SHIFT);
1580 		break;
1581 
1582 	default:
1583 		return -EINVAL;
1584 	}
1585 
1586 	return 0;
1587 }
1588 
1589 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1590 				      struct ib_ucontext *context,
1591 				      struct ib_udata *udata)
1592 {
1593 	struct mlx5_ib_alloc_pd_resp resp;
1594 	struct mlx5_ib_pd *pd;
1595 	int err;
1596 
1597 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1598 	if (!pd)
1599 		return ERR_PTR(-ENOMEM);
1600 
1601 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1602 	if (err) {
1603 		kfree(pd);
1604 		return ERR_PTR(err);
1605 	}
1606 
1607 	if (context) {
1608 		resp.pdn = pd->pdn;
1609 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1610 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1611 			kfree(pd);
1612 			return ERR_PTR(-EFAULT);
1613 		}
1614 	}
1615 
1616 	return &pd->ibpd;
1617 }
1618 
1619 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1620 {
1621 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1622 	struct mlx5_ib_pd *mpd = to_mpd(pd);
1623 
1624 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1625 	kfree(mpd);
1626 
1627 	return 0;
1628 }
1629 
1630 enum {
1631 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
1632 	MATCH_CRITERIA_ENABLE_MISC_BIT,
1633 	MATCH_CRITERIA_ENABLE_INNER_BIT
1634 };
1635 
1636 #define HEADER_IS_ZERO(match_criteria, headers)			           \
1637 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1638 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
1639 
1640 static u8 get_match_criteria_enable(u32 *match_criteria)
1641 {
1642 	u8 match_criteria_enable;
1643 
1644 	match_criteria_enable =
1645 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1646 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
1647 	match_criteria_enable |=
1648 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1649 		MATCH_CRITERIA_ENABLE_MISC_BIT;
1650 	match_criteria_enable |=
1651 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1652 		MATCH_CRITERIA_ENABLE_INNER_BIT;
1653 
1654 	return match_criteria_enable;
1655 }
1656 
1657 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1658 {
1659 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1660 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1661 }
1662 
1663 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1664 {
1665 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1666 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1667 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1668 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1669 }
1670 
1671 #define LAST_ETH_FIELD vlan_tag
1672 #define LAST_IB_FIELD sl
1673 #define LAST_IPV4_FIELD tos
1674 #define LAST_IPV6_FIELD traffic_class
1675 #define LAST_TCP_UDP_FIELD src_port
1676 
1677 /* Field is the last supported field */
1678 #define FIELDS_NOT_SUPPORTED(filter, field)\
1679 	memchr_inv((void *)&filter.field  +\
1680 		   sizeof(filter.field), 0,\
1681 		   sizeof(filter) -\
1682 		   offsetof(typeof(filter), field) -\
1683 		   sizeof(filter.field))
1684 
1685 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1686 			   const union ib_flow_spec *ib_spec)
1687 {
1688 	void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1689 					     outer_headers);
1690 	void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1691 					     outer_headers);
1692 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1693 					   misc_parameters);
1694 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1695 					   misc_parameters);
1696 
1697 	switch (ib_spec->type) {
1698 	case IB_FLOW_SPEC_ETH:
1699 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1700 			return -ENOTSUPP;
1701 
1702 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1703 					     dmac_47_16),
1704 				ib_spec->eth.mask.dst_mac);
1705 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1706 					     dmac_47_16),
1707 				ib_spec->eth.val.dst_mac);
1708 
1709 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1710 					     smac_47_16),
1711 				ib_spec->eth.mask.src_mac);
1712 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1713 					     smac_47_16),
1714 				ib_spec->eth.val.src_mac);
1715 
1716 		if (ib_spec->eth.mask.vlan_tag) {
1717 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1718 				 cvlan_tag, 1);
1719 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1720 				 cvlan_tag, 1);
1721 
1722 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1723 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1724 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1725 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1726 
1727 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1728 				 first_cfi,
1729 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1730 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1731 				 first_cfi,
1732 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1733 
1734 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1735 				 first_prio,
1736 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1737 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1738 				 first_prio,
1739 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1740 		}
1741 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1742 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1743 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1744 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
1745 		break;
1746 	case IB_FLOW_SPEC_IPV4:
1747 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1748 			return -ENOTSUPP;
1749 
1750 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1751 			 ethertype, 0xffff);
1752 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1753 			 ethertype, ETH_P_IP);
1754 
1755 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1756 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1757 		       &ib_spec->ipv4.mask.src_ip,
1758 		       sizeof(ib_spec->ipv4.mask.src_ip));
1759 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1760 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1761 		       &ib_spec->ipv4.val.src_ip,
1762 		       sizeof(ib_spec->ipv4.val.src_ip));
1763 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1764 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1765 		       &ib_spec->ipv4.mask.dst_ip,
1766 		       sizeof(ib_spec->ipv4.mask.dst_ip));
1767 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1768 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1769 		       &ib_spec->ipv4.val.dst_ip,
1770 		       sizeof(ib_spec->ipv4.val.dst_ip));
1771 
1772 		set_tos(outer_headers_c, outer_headers_v,
1773 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
1774 
1775 		set_proto(outer_headers_c, outer_headers_v,
1776 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
1777 		break;
1778 	case IB_FLOW_SPEC_IPV6:
1779 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1780 			return -ENOTSUPP;
1781 
1782 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1783 			 ethertype, 0xffff);
1784 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1785 			 ethertype, IPPROTO_IPV6);
1786 
1787 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1788 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1789 		       &ib_spec->ipv6.mask.src_ip,
1790 		       sizeof(ib_spec->ipv6.mask.src_ip));
1791 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1792 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1793 		       &ib_spec->ipv6.val.src_ip,
1794 		       sizeof(ib_spec->ipv6.val.src_ip));
1795 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1796 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1797 		       &ib_spec->ipv6.mask.dst_ip,
1798 		       sizeof(ib_spec->ipv6.mask.dst_ip));
1799 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1800 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1801 		       &ib_spec->ipv6.val.dst_ip,
1802 		       sizeof(ib_spec->ipv6.val.dst_ip));
1803 
1804 		set_tos(outer_headers_c, outer_headers_v,
1805 			ib_spec->ipv6.mask.traffic_class,
1806 			ib_spec->ipv6.val.traffic_class);
1807 
1808 		set_proto(outer_headers_c, outer_headers_v,
1809 			  ib_spec->ipv6.mask.next_hdr,
1810 			  ib_spec->ipv6.val.next_hdr);
1811 
1812 		MLX5_SET(fte_match_set_misc, misc_params_c,
1813 			 outer_ipv6_flow_label,
1814 			 ntohl(ib_spec->ipv6.mask.flow_label));
1815 		MLX5_SET(fte_match_set_misc, misc_params_v,
1816 			 outer_ipv6_flow_label,
1817 			 ntohl(ib_spec->ipv6.val.flow_label));
1818 		break;
1819 	case IB_FLOW_SPEC_TCP:
1820 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1821 					 LAST_TCP_UDP_FIELD))
1822 			return -ENOTSUPP;
1823 
1824 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1825 			 0xff);
1826 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1827 			 IPPROTO_TCP);
1828 
1829 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport,
1830 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1831 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport,
1832 			 ntohs(ib_spec->tcp_udp.val.src_port));
1833 
1834 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport,
1835 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1836 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport,
1837 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1838 		break;
1839 	case IB_FLOW_SPEC_UDP:
1840 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1841 					 LAST_TCP_UDP_FIELD))
1842 			return -ENOTSUPP;
1843 
1844 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1845 			 0xff);
1846 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1847 			 IPPROTO_UDP);
1848 
1849 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport,
1850 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1851 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport,
1852 			 ntohs(ib_spec->tcp_udp.val.src_port));
1853 
1854 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport,
1855 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1856 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport,
1857 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1858 		break;
1859 	default:
1860 		return -EINVAL;
1861 	}
1862 
1863 	return 0;
1864 }
1865 
1866 /* If a flow could catch both multicast and unicast packets,
1867  * it won't fall into the multicast flow steering table and this rule
1868  * could steal other multicast packets.
1869  */
1870 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1871 {
1872 	struct ib_flow_spec_eth *eth_spec;
1873 
1874 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1875 	    ib_attr->size < sizeof(struct ib_flow_attr) +
1876 	    sizeof(struct ib_flow_spec_eth) ||
1877 	    ib_attr->num_of_specs < 1)
1878 		return false;
1879 
1880 	eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1881 	if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1882 	    eth_spec->size != sizeof(*eth_spec))
1883 		return false;
1884 
1885 	return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1886 	       is_multicast_ether_addr(eth_spec->val.dst_mac);
1887 }
1888 
1889 static bool is_valid_attr(const struct ib_flow_attr *flow_attr)
1890 {
1891 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1892 	bool has_ipv4_spec = false;
1893 	bool eth_type_ipv4 = true;
1894 	unsigned int spec_index;
1895 
1896 	/* Validate that ethertype is correct */
1897 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1898 		if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1899 		    ib_spec->eth.mask.ether_type) {
1900 			if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1901 			      ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1902 				eth_type_ipv4 = false;
1903 		} else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1904 			has_ipv4_spec = true;
1905 		}
1906 		ib_spec = (void *)ib_spec + ib_spec->size;
1907 	}
1908 	return !has_ipv4_spec || eth_type_ipv4;
1909 }
1910 
1911 static void put_flow_table(struct mlx5_ib_dev *dev,
1912 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
1913 {
1914 	prio->refcount -= !!ft_added;
1915 	if (!prio->refcount) {
1916 		mlx5_destroy_flow_table(prio->flow_table);
1917 		prio->flow_table = NULL;
1918 	}
1919 }
1920 
1921 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
1922 {
1923 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
1924 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
1925 							  struct mlx5_ib_flow_handler,
1926 							  ibflow);
1927 	struct mlx5_ib_flow_handler *iter, *tmp;
1928 
1929 	mutex_lock(&dev->flow_db.lock);
1930 
1931 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
1932 		mlx5_del_flow_rule(iter->rule);
1933 		put_flow_table(dev, iter->prio, true);
1934 		list_del(&iter->list);
1935 		kfree(iter);
1936 	}
1937 
1938 	mlx5_del_flow_rule(handler->rule);
1939 	put_flow_table(dev, handler->prio, true);
1940 	mutex_unlock(&dev->flow_db.lock);
1941 
1942 	kfree(handler);
1943 
1944 	return 0;
1945 }
1946 
1947 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
1948 {
1949 	priority *= 2;
1950 	if (!dont_trap)
1951 		priority++;
1952 	return priority;
1953 }
1954 
1955 enum flow_table_type {
1956 	MLX5_IB_FT_RX,
1957 	MLX5_IB_FT_TX
1958 };
1959 
1960 #define MLX5_FS_MAX_TYPES	 10
1961 #define MLX5_FS_MAX_ENTRIES	 32000UL
1962 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
1963 						struct ib_flow_attr *flow_attr,
1964 						enum flow_table_type ft_type)
1965 {
1966 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
1967 	struct mlx5_flow_namespace *ns = NULL;
1968 	struct mlx5_ib_flow_prio *prio;
1969 	struct mlx5_flow_table *ft;
1970 	int num_entries;
1971 	int num_groups;
1972 	int priority;
1973 	int err = 0;
1974 
1975 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1976 		if (flow_is_multicast_only(flow_attr) &&
1977 		    !dont_trap)
1978 			priority = MLX5_IB_FLOW_MCAST_PRIO;
1979 		else
1980 			priority = ib_prio_to_core_prio(flow_attr->priority,
1981 							dont_trap);
1982 		ns = mlx5_get_flow_namespace(dev->mdev,
1983 					     MLX5_FLOW_NAMESPACE_BYPASS);
1984 		num_entries = MLX5_FS_MAX_ENTRIES;
1985 		num_groups = MLX5_FS_MAX_TYPES;
1986 		prio = &dev->flow_db.prios[priority];
1987 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1988 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1989 		ns = mlx5_get_flow_namespace(dev->mdev,
1990 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
1991 		build_leftovers_ft_param("bypass", &priority,
1992 					 &num_entries,
1993 					 &num_groups);
1994 		prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
1995 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
1996 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
1997 					allow_sniffer_and_nic_rx_shared_tir))
1998 			return ERR_PTR(-ENOTSUPP);
1999 
2000 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
2001 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2002 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2003 
2004 		prio = &dev->flow_db.sniffer[ft_type];
2005 		priority = 0;
2006 		num_entries = 1;
2007 		num_groups = 1;
2008 	}
2009 
2010 	if (!ns)
2011 		return ERR_PTR(-ENOTSUPP);
2012 
2013 	ft = prio->flow_table;
2014 	if (!ft) {
2015 		ft = mlx5_create_auto_grouped_flow_table(ns, priority, "bypass",
2016 							 num_entries,
2017 							 num_groups);
2018 
2019 		if (!IS_ERR(ft)) {
2020 			prio->refcount = 0;
2021 			prio->flow_table = ft;
2022 		} else {
2023 			err = PTR_ERR(ft);
2024 		}
2025 	}
2026 
2027 	return err ? ERR_PTR(err) : prio;
2028 }
2029 
2030 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
2031 						     struct mlx5_ib_flow_prio *ft_prio,
2032 						     const struct ib_flow_attr *flow_attr,
2033 						     struct mlx5_flow_destination *dst)
2034 {
2035 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
2036 	struct mlx5_ib_flow_handler *handler;
2037 	struct mlx5_flow_spec *spec;
2038 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
2039 	unsigned int spec_index;
2040 	u32 action;
2041 	int err = 0;
2042 
2043 	if (!is_valid_attr(flow_attr))
2044 		return ERR_PTR(-EINVAL);
2045 
2046 	spec = mlx5_vzalloc(sizeof(*spec));
2047 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
2048 	if (!handler || !spec) {
2049 		err = -ENOMEM;
2050 		goto free;
2051 	}
2052 
2053 	INIT_LIST_HEAD(&handler->list);
2054 
2055 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2056 		err = parse_flow_attr(spec->match_criteria,
2057 				      spec->match_value, ib_flow);
2058 		if (err < 0)
2059 			goto free;
2060 
2061 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
2062 	}
2063 
2064 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
2065 	action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
2066 		MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
2067 	handler->rule = mlx5_add_flow_rule(ft, spec->match_criteria_enable,
2068 					   spec->match_criteria,
2069 					   spec->match_value,
2070 					   action,
2071 					   MLX5_FS_DEFAULT_FLOW_TAG,
2072 					   dst);
2073 
2074 	if (IS_ERR(handler->rule)) {
2075 		err = PTR_ERR(handler->rule);
2076 		goto free;
2077 	}
2078 
2079 	ft_prio->refcount++;
2080 	handler->prio = ft_prio;
2081 
2082 	ft_prio->flow_table = ft;
2083 free:
2084 	if (err)
2085 		kfree(handler);
2086 	kvfree(spec);
2087 	return err ? ERR_PTR(err) : handler;
2088 }
2089 
2090 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
2091 							  struct mlx5_ib_flow_prio *ft_prio,
2092 							  struct ib_flow_attr *flow_attr,
2093 							  struct mlx5_flow_destination *dst)
2094 {
2095 	struct mlx5_ib_flow_handler *handler_dst = NULL;
2096 	struct mlx5_ib_flow_handler *handler = NULL;
2097 
2098 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
2099 	if (!IS_ERR(handler)) {
2100 		handler_dst = create_flow_rule(dev, ft_prio,
2101 					       flow_attr, dst);
2102 		if (IS_ERR(handler_dst)) {
2103 			mlx5_del_flow_rule(handler->rule);
2104 			ft_prio->refcount--;
2105 			kfree(handler);
2106 			handler = handler_dst;
2107 		} else {
2108 			list_add(&handler_dst->list, &handler->list);
2109 		}
2110 	}
2111 
2112 	return handler;
2113 }
2114 enum {
2115 	LEFTOVERS_MC,
2116 	LEFTOVERS_UC,
2117 };
2118 
2119 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
2120 							  struct mlx5_ib_flow_prio *ft_prio,
2121 							  struct ib_flow_attr *flow_attr,
2122 							  struct mlx5_flow_destination *dst)
2123 {
2124 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
2125 	struct mlx5_ib_flow_handler *handler = NULL;
2126 
2127 	static struct {
2128 		struct ib_flow_attr	flow_attr;
2129 		struct ib_flow_spec_eth eth_flow;
2130 	} leftovers_specs[] = {
2131 		[LEFTOVERS_MC] = {
2132 			.flow_attr = {
2133 				.num_of_specs = 1,
2134 				.size = sizeof(leftovers_specs[0])
2135 			},
2136 			.eth_flow = {
2137 				.type = IB_FLOW_SPEC_ETH,
2138 				.size = sizeof(struct ib_flow_spec_eth),
2139 				.mask = {.dst_mac = {0x1} },
2140 				.val =  {.dst_mac = {0x1} }
2141 			}
2142 		},
2143 		[LEFTOVERS_UC] = {
2144 			.flow_attr = {
2145 				.num_of_specs = 1,
2146 				.size = sizeof(leftovers_specs[0])
2147 			},
2148 			.eth_flow = {
2149 				.type = IB_FLOW_SPEC_ETH,
2150 				.size = sizeof(struct ib_flow_spec_eth),
2151 				.mask = {.dst_mac = {0x1} },
2152 				.val = {.dst_mac = {} }
2153 			}
2154 		}
2155 	};
2156 
2157 	handler = create_flow_rule(dev, ft_prio,
2158 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
2159 				   dst);
2160 	if (!IS_ERR(handler) &&
2161 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2162 		handler_ucast = create_flow_rule(dev, ft_prio,
2163 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
2164 						 dst);
2165 		if (IS_ERR(handler_ucast)) {
2166 			mlx5_del_flow_rule(handler->rule);
2167 			ft_prio->refcount--;
2168 			kfree(handler);
2169 			handler = handler_ucast;
2170 		} else {
2171 			list_add(&handler_ucast->list, &handler->list);
2172 		}
2173 	}
2174 
2175 	return handler;
2176 }
2177 
2178 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2179 							struct mlx5_ib_flow_prio *ft_rx,
2180 							struct mlx5_ib_flow_prio *ft_tx,
2181 							struct mlx5_flow_destination *dst)
2182 {
2183 	struct mlx5_ib_flow_handler *handler_rx;
2184 	struct mlx5_ib_flow_handler *handler_tx;
2185 	int err;
2186 	static const struct ib_flow_attr flow_attr  = {
2187 		.num_of_specs = 0,
2188 		.size = sizeof(flow_attr)
2189 	};
2190 
2191 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2192 	if (IS_ERR(handler_rx)) {
2193 		err = PTR_ERR(handler_rx);
2194 		goto err;
2195 	}
2196 
2197 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2198 	if (IS_ERR(handler_tx)) {
2199 		err = PTR_ERR(handler_tx);
2200 		goto err_tx;
2201 	}
2202 
2203 	list_add(&handler_tx->list, &handler_rx->list);
2204 
2205 	return handler_rx;
2206 
2207 err_tx:
2208 	mlx5_del_flow_rule(handler_rx->rule);
2209 	ft_rx->refcount--;
2210 	kfree(handler_rx);
2211 err:
2212 	return ERR_PTR(err);
2213 }
2214 
2215 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2216 					   struct ib_flow_attr *flow_attr,
2217 					   int domain)
2218 {
2219 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2220 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2221 	struct mlx5_ib_flow_handler *handler = NULL;
2222 	struct mlx5_flow_destination *dst = NULL;
2223 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2224 	struct mlx5_ib_flow_prio *ft_prio;
2225 	int err;
2226 
2227 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2228 		return ERR_PTR(-ENOSPC);
2229 
2230 	if (domain != IB_FLOW_DOMAIN_USER ||
2231 	    flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2232 	    (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2233 		return ERR_PTR(-EINVAL);
2234 
2235 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2236 	if (!dst)
2237 		return ERR_PTR(-ENOMEM);
2238 
2239 	mutex_lock(&dev->flow_db.lock);
2240 
2241 	ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2242 	if (IS_ERR(ft_prio)) {
2243 		err = PTR_ERR(ft_prio);
2244 		goto unlock;
2245 	}
2246 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2247 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2248 		if (IS_ERR(ft_prio_tx)) {
2249 			err = PTR_ERR(ft_prio_tx);
2250 			ft_prio_tx = NULL;
2251 			goto destroy_ft;
2252 		}
2253 	}
2254 
2255 	dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2256 	if (mqp->flags & MLX5_IB_QP_RSS)
2257 		dst->tir_num = mqp->rss_qp.tirn;
2258 	else
2259 		dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2260 
2261 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2262 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
2263 			handler = create_dont_trap_rule(dev, ft_prio,
2264 							flow_attr, dst);
2265 		} else {
2266 			handler = create_flow_rule(dev, ft_prio, flow_attr,
2267 						   dst);
2268 		}
2269 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2270 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2271 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2272 						dst);
2273 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2274 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2275 	} else {
2276 		err = -EINVAL;
2277 		goto destroy_ft;
2278 	}
2279 
2280 	if (IS_ERR(handler)) {
2281 		err = PTR_ERR(handler);
2282 		handler = NULL;
2283 		goto destroy_ft;
2284 	}
2285 
2286 	mutex_unlock(&dev->flow_db.lock);
2287 	kfree(dst);
2288 
2289 	return &handler->ibflow;
2290 
2291 destroy_ft:
2292 	put_flow_table(dev, ft_prio, false);
2293 	if (ft_prio_tx)
2294 		put_flow_table(dev, ft_prio_tx, false);
2295 unlock:
2296 	mutex_unlock(&dev->flow_db.lock);
2297 	kfree(dst);
2298 	kfree(handler);
2299 	return ERR_PTR(err);
2300 }
2301 
2302 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2303 {
2304 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2305 	int err;
2306 
2307 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2308 	if (err)
2309 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2310 			     ibqp->qp_num, gid->raw);
2311 
2312 	return err;
2313 }
2314 
2315 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2316 {
2317 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2318 	int err;
2319 
2320 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2321 	if (err)
2322 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2323 			     ibqp->qp_num, gid->raw);
2324 
2325 	return err;
2326 }
2327 
2328 static int init_node_data(struct mlx5_ib_dev *dev)
2329 {
2330 	int err;
2331 
2332 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2333 	if (err)
2334 		return err;
2335 
2336 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2337 }
2338 
2339 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2340 			     char *buf)
2341 {
2342 	struct mlx5_ib_dev *dev =
2343 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2344 
2345 	return sprintf(buf, "%lld\n", (long long)dev->mdev->priv.fw_pages);
2346 }
2347 
2348 static ssize_t show_reg_pages(struct device *device,
2349 			      struct device_attribute *attr, char *buf)
2350 {
2351 	struct mlx5_ib_dev *dev =
2352 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2353 
2354 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2355 }
2356 
2357 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2358 			char *buf)
2359 {
2360 	struct mlx5_ib_dev *dev =
2361 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2362 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2363 }
2364 
2365 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2366 			char *buf)
2367 {
2368 	struct mlx5_ib_dev *dev =
2369 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2370 	return sprintf(buf, "%x\n", dev->mdev->pdev->revision);
2371 }
2372 
2373 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2374 			  char *buf)
2375 {
2376 	struct mlx5_ib_dev *dev =
2377 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2378 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2379 		       dev->mdev->board_id);
2380 }
2381 
2382 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2383 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2384 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2385 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2386 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2387 
2388 static struct device_attribute *mlx5_class_attributes[] = {
2389 	&dev_attr_hw_rev,
2390 	&dev_attr_hca_type,
2391 	&dev_attr_board_id,
2392 	&dev_attr_fw_pages,
2393 	&dev_attr_reg_pages,
2394 };
2395 
2396 static void pkey_change_handler(struct work_struct *work)
2397 {
2398 	struct mlx5_ib_port_resources *ports =
2399 		container_of(work, struct mlx5_ib_port_resources,
2400 			     pkey_change_work);
2401 
2402 	mutex_lock(&ports->devr->mutex);
2403 	mlx5_ib_gsi_pkey_change(ports->gsi);
2404 	mutex_unlock(&ports->devr->mutex);
2405 }
2406 
2407 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2408 {
2409 	struct mlx5_ib_qp *mqp;
2410 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2411 	struct mlx5_core_cq *mcq;
2412 	struct list_head cq_armed_list;
2413 	unsigned long flags_qp;
2414 	unsigned long flags_cq;
2415 	unsigned long flags;
2416 
2417 	INIT_LIST_HEAD(&cq_armed_list);
2418 
2419 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2420 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2421 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2422 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2423 		if (mqp->sq.tail != mqp->sq.head) {
2424 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2425 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2426 			if (send_mcq->mcq.comp &&
2427 			    mqp->ibqp.send_cq->comp_handler) {
2428 				if (!send_mcq->mcq.reset_notify_added) {
2429 					send_mcq->mcq.reset_notify_added = 1;
2430 					list_add_tail(&send_mcq->mcq.reset_notify,
2431 						      &cq_armed_list);
2432 				}
2433 			}
2434 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2435 		}
2436 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2437 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2438 		/* no handling is needed for SRQ */
2439 		if (!mqp->ibqp.srq) {
2440 			if (mqp->rq.tail != mqp->rq.head) {
2441 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2442 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2443 				if (recv_mcq->mcq.comp &&
2444 				    mqp->ibqp.recv_cq->comp_handler) {
2445 					if (!recv_mcq->mcq.reset_notify_added) {
2446 						recv_mcq->mcq.reset_notify_added = 1;
2447 						list_add_tail(&recv_mcq->mcq.reset_notify,
2448 							      &cq_armed_list);
2449 					}
2450 				}
2451 				spin_unlock_irqrestore(&recv_mcq->lock,
2452 						       flags_cq);
2453 			}
2454 		}
2455 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2456 	}
2457 	/*At that point all inflight post send were put to be executed as of we
2458 	 * lock/unlock above locks Now need to arm all involved CQs.
2459 	 */
2460 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2461 		mcq->comp(mcq);
2462 	}
2463 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2464 }
2465 
2466 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2467 			  enum mlx5_dev_event event, unsigned long param)
2468 {
2469 	struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2470 	struct ib_event ibev;
2471 	bool fatal = false;
2472 	u8 port = (u8)param;
2473 
2474 	switch (event) {
2475 	case MLX5_DEV_EVENT_SYS_ERROR:
2476 		ibev.event = IB_EVENT_DEVICE_FATAL;
2477 		mlx5_ib_handle_internal_error(ibdev);
2478 		fatal = true;
2479 		break;
2480 
2481 	case MLX5_DEV_EVENT_PORT_UP:
2482 	case MLX5_DEV_EVENT_PORT_DOWN:
2483 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
2484 		/* In RoCE, port up/down events are handled in
2485 		 * mlx5_netdev_event().
2486 		 */
2487 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2488 			IB_LINK_LAYER_ETHERNET)
2489 			return;
2490 
2491 		ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2492 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2493 		break;
2494 
2495 	case MLX5_DEV_EVENT_LID_CHANGE:
2496 		ibev.event = IB_EVENT_LID_CHANGE;
2497 		break;
2498 
2499 	case MLX5_DEV_EVENT_PKEY_CHANGE:
2500 		ibev.event = IB_EVENT_PKEY_CHANGE;
2501 
2502 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2503 		break;
2504 
2505 	case MLX5_DEV_EVENT_GUID_CHANGE:
2506 		ibev.event = IB_EVENT_GID_CHANGE;
2507 		break;
2508 
2509 	case MLX5_DEV_EVENT_CLIENT_REREG:
2510 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
2511 		break;
2512 
2513 	default:
2514 		/* unsupported event */
2515 		return;
2516 	}
2517 
2518 	ibev.device	      = &ibdev->ib_dev;
2519 	ibev.element.port_num = port;
2520 
2521 	if (!rdma_is_port_valid(&ibdev->ib_dev, port)) {
2522 		mlx5_ib_warn(ibdev, "warning: event(%d) on port %d\n", event, port);
2523 		return;
2524 	}
2525 
2526 	if (ibdev->ib_active)
2527 		ib_dispatch_event(&ibev);
2528 
2529 	if (fatal)
2530 		ibdev->ib_active = false;
2531 }
2532 
2533 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2534 {
2535 	int port;
2536 
2537 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2538 		mlx5_query_ext_port_caps(dev, port);
2539 }
2540 
2541 static int get_port_caps(struct mlx5_ib_dev *dev)
2542 {
2543 	struct ib_device_attr *dprops = NULL;
2544 	struct ib_port_attr *pprops = NULL;
2545 	int err = -ENOMEM;
2546 	int port;
2547 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2548 
2549 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2550 	if (!pprops)
2551 		goto out;
2552 
2553 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2554 	if (!dprops)
2555 		goto out;
2556 
2557 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2558 	if (err) {
2559 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2560 		goto out;
2561 	}
2562 
2563 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2564 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2565 		if (err) {
2566 			mlx5_ib_warn(dev, "query_port %d failed %d\n",
2567 				     port, err);
2568 			break;
2569 		}
2570 		dev->mdev->port_caps[port - 1].pkey_table_len =
2571 						dprops->max_pkeys;
2572 		dev->mdev->port_caps[port - 1].gid_table_len =
2573 						pprops->gid_tbl_len;
2574 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2575 			    dprops->max_pkeys, pprops->gid_tbl_len);
2576 	}
2577 
2578 out:
2579 	kfree(pprops);
2580 	kfree(dprops);
2581 
2582 	return err;
2583 }
2584 
2585 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2586 {
2587 	int err;
2588 
2589 	err = mlx5_mr_cache_cleanup(dev);
2590 	if (err)
2591 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2592 
2593 	mlx5_ib_destroy_qp(dev->umrc.qp);
2594 	ib_free_cq(dev->umrc.cq);
2595 	ib_dealloc_pd(dev->umrc.pd);
2596 }
2597 
2598 enum {
2599 	MAX_UMR_WR = 128,
2600 };
2601 
2602 static int create_umr_res(struct mlx5_ib_dev *dev)
2603 {
2604 	struct ib_qp_init_attr *init_attr = NULL;
2605 	struct ib_qp_attr *attr = NULL;
2606 	struct ib_pd *pd;
2607 	struct ib_cq *cq;
2608 	struct ib_qp *qp;
2609 	int ret;
2610 
2611 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2612 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2613 	if (!attr || !init_attr) {
2614 		ret = -ENOMEM;
2615 		goto error_0;
2616 	}
2617 
2618 	pd = ib_alloc_pd(&dev->ib_dev, 0);
2619 	if (IS_ERR(pd)) {
2620 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2621 		ret = PTR_ERR(pd);
2622 		goto error_0;
2623 	}
2624 
2625 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2626 	if (IS_ERR(cq)) {
2627 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2628 		ret = PTR_ERR(cq);
2629 		goto error_2;
2630 	}
2631 
2632 	init_attr->send_cq = cq;
2633 	init_attr->recv_cq = cq;
2634 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2635 	init_attr->cap.max_send_wr = MAX_UMR_WR;
2636 	init_attr->cap.max_send_sge = 1;
2637 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2638 	init_attr->port_num = 1;
2639 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2640 	if (IS_ERR(qp)) {
2641 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2642 		ret = PTR_ERR(qp);
2643 		goto error_3;
2644 	}
2645 	qp->device     = &dev->ib_dev;
2646 	qp->real_qp    = qp;
2647 	qp->uobject    = NULL;
2648 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2649 
2650 	attr->qp_state = IB_QPS_INIT;
2651 	attr->port_num = 1;
2652 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2653 				IB_QP_PORT, NULL);
2654 	if (ret) {
2655 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2656 		goto error_4;
2657 	}
2658 
2659 	memset(attr, 0, sizeof(*attr));
2660 	attr->qp_state = IB_QPS_RTR;
2661 	attr->path_mtu = IB_MTU_256;
2662 
2663 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2664 	if (ret) {
2665 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2666 		goto error_4;
2667 	}
2668 
2669 	memset(attr, 0, sizeof(*attr));
2670 	attr->qp_state = IB_QPS_RTS;
2671 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2672 	if (ret) {
2673 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2674 		goto error_4;
2675 	}
2676 
2677 	dev->umrc.qp = qp;
2678 	dev->umrc.cq = cq;
2679 	dev->umrc.pd = pd;
2680 
2681 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
2682 	ret = mlx5_mr_cache_init(dev);
2683 	if (ret) {
2684 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2685 		goto error_4;
2686 	}
2687 
2688 	kfree(attr);
2689 	kfree(init_attr);
2690 
2691 	return 0;
2692 
2693 error_4:
2694 	mlx5_ib_destroy_qp(qp);
2695 
2696 error_3:
2697 	ib_free_cq(cq);
2698 
2699 error_2:
2700 	ib_dealloc_pd(pd);
2701 
2702 error_0:
2703 	kfree(attr);
2704 	kfree(init_attr);
2705 	return ret;
2706 }
2707 
2708 static int create_dev_resources(struct mlx5_ib_resources *devr)
2709 {
2710 	struct ib_srq_init_attr attr;
2711 	struct mlx5_ib_dev *dev;
2712 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2713 	int port;
2714 	int ret = 0;
2715 
2716 	dev = container_of(devr, struct mlx5_ib_dev, devr);
2717 
2718 	mutex_init(&devr->mutex);
2719 
2720 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2721 	if (IS_ERR(devr->p0)) {
2722 		ret = PTR_ERR(devr->p0);
2723 		goto error0;
2724 	}
2725 	devr->p0->device  = &dev->ib_dev;
2726 	devr->p0->uobject = NULL;
2727 	atomic_set(&devr->p0->usecnt, 0);
2728 
2729 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
2730 	if (IS_ERR(devr->c0)) {
2731 		ret = PTR_ERR(devr->c0);
2732 		goto error1;
2733 	}
2734 	devr->c0->device        = &dev->ib_dev;
2735 	devr->c0->uobject       = NULL;
2736 	devr->c0->comp_handler  = NULL;
2737 	devr->c0->event_handler = NULL;
2738 	devr->c0->cq_context    = NULL;
2739 	atomic_set(&devr->c0->usecnt, 0);
2740 
2741 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2742 	if (IS_ERR(devr->x0)) {
2743 		ret = PTR_ERR(devr->x0);
2744 		goto error2;
2745 	}
2746 	devr->x0->device = &dev->ib_dev;
2747 	devr->x0->inode = NULL;
2748 	atomic_set(&devr->x0->usecnt, 0);
2749 	mutex_init(&devr->x0->tgt_qp_mutex);
2750 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2751 
2752 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2753 	if (IS_ERR(devr->x1)) {
2754 		ret = PTR_ERR(devr->x1);
2755 		goto error3;
2756 	}
2757 	devr->x1->device = &dev->ib_dev;
2758 	devr->x1->inode = NULL;
2759 	atomic_set(&devr->x1->usecnt, 0);
2760 	mutex_init(&devr->x1->tgt_qp_mutex);
2761 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2762 
2763 	memset(&attr, 0, sizeof(attr));
2764 	attr.attr.max_sge = 1;
2765 	attr.attr.max_wr = 1;
2766 	attr.srq_type = IB_SRQT_XRC;
2767 	attr.ext.xrc.cq = devr->c0;
2768 	attr.ext.xrc.xrcd = devr->x0;
2769 
2770 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2771 	if (IS_ERR(devr->s0)) {
2772 		ret = PTR_ERR(devr->s0);
2773 		goto error4;
2774 	}
2775 	devr->s0->device	= &dev->ib_dev;
2776 	devr->s0->pd		= devr->p0;
2777 	devr->s0->uobject       = NULL;
2778 	devr->s0->event_handler = NULL;
2779 	devr->s0->srq_context   = NULL;
2780 	devr->s0->srq_type      = IB_SRQT_XRC;
2781 	devr->s0->ext.xrc.xrcd	= devr->x0;
2782 	devr->s0->ext.xrc.cq	= devr->c0;
2783 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2784 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2785 	atomic_inc(&devr->p0->usecnt);
2786 	atomic_set(&devr->s0->usecnt, 0);
2787 
2788 	memset(&attr, 0, sizeof(attr));
2789 	attr.attr.max_sge = 1;
2790 	attr.attr.max_wr = 1;
2791 	attr.srq_type = IB_SRQT_BASIC;
2792 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2793 	if (IS_ERR(devr->s1)) {
2794 		ret = PTR_ERR(devr->s1);
2795 		goto error5;
2796 	}
2797 	devr->s1->device	= &dev->ib_dev;
2798 	devr->s1->pd		= devr->p0;
2799 	devr->s1->uobject       = NULL;
2800 	devr->s1->event_handler = NULL;
2801 	devr->s1->srq_context   = NULL;
2802 	devr->s1->srq_type      = IB_SRQT_BASIC;
2803 	devr->s1->ext.xrc.cq	= devr->c0;
2804 	atomic_inc(&devr->p0->usecnt);
2805 	atomic_set(&devr->s0->usecnt, 0);
2806 
2807 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2808 		INIT_WORK(&devr->ports[port].pkey_change_work,
2809 			  pkey_change_handler);
2810 		devr->ports[port].devr = devr;
2811 	}
2812 
2813 	return 0;
2814 
2815 error5:
2816 	mlx5_ib_destroy_srq(devr->s0);
2817 error4:
2818 	mlx5_ib_dealloc_xrcd(devr->x1);
2819 error3:
2820 	mlx5_ib_dealloc_xrcd(devr->x0);
2821 error2:
2822 	mlx5_ib_destroy_cq(devr->c0);
2823 error1:
2824 	mlx5_ib_dealloc_pd(devr->p0);
2825 error0:
2826 	return ret;
2827 }
2828 
2829 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2830 {
2831 	struct mlx5_ib_dev *dev =
2832 		container_of(devr, struct mlx5_ib_dev, devr);
2833 	int port;
2834 
2835 	mlx5_ib_destroy_srq(devr->s1);
2836 	mlx5_ib_destroy_srq(devr->s0);
2837 	mlx5_ib_dealloc_xrcd(devr->x0);
2838 	mlx5_ib_dealloc_xrcd(devr->x1);
2839 	mlx5_ib_destroy_cq(devr->c0);
2840 	mlx5_ib_dealloc_pd(devr->p0);
2841 
2842 	/* Make sure no change P_Key work items are still executing */
2843 	for (port = 0; port < dev->num_ports; ++port)
2844 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2845 }
2846 
2847 static u32 get_core_cap_flags(struct ib_device *ibdev)
2848 {
2849 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2850 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2851 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2852 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2853 	u32 ret = 0;
2854 
2855 	if (ll == IB_LINK_LAYER_INFINIBAND)
2856 		return RDMA_CORE_PORT_IBA_IB;
2857 
2858 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2859 		return 0;
2860 
2861 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2862 		return 0;
2863 
2864 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2865 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2866 
2867 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2868 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2869 
2870 	return ret;
2871 }
2872 
2873 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
2874 			       struct ib_port_immutable *immutable)
2875 {
2876 	struct ib_port_attr attr;
2877 	int err;
2878 
2879 	err = mlx5_ib_query_port(ibdev, port_num, &attr);
2880 	if (err)
2881 		return err;
2882 
2883 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2884 	immutable->gid_tbl_len = attr.gid_tbl_len;
2885 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
2886 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2887 
2888 	return 0;
2889 }
2890 
2891 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
2892 			   size_t str_len)
2893 {
2894 	struct mlx5_ib_dev *dev =
2895 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2896 	snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
2897 		       fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
2898 }
2899 
2900 static int mlx5_roce_lag_init(struct mlx5_ib_dev *dev)
2901 {
2902 	return 0;
2903 }
2904 
2905 static void mlx5_roce_lag_cleanup(struct mlx5_ib_dev *dev)
2906 {
2907 }
2908 
2909 static void mlx5_remove_roce_notifier(struct mlx5_ib_dev *dev)
2910 {
2911 	if (dev->roce.nb.notifier_call) {
2912 		unregister_netdevice_notifier(&dev->roce.nb);
2913 		dev->roce.nb.notifier_call = NULL;
2914 	}
2915 }
2916 
2917 static int mlx5_enable_roce(struct mlx5_ib_dev *dev)
2918 {
2919 	VNET_ITERATOR_DECL(vnet_iter);
2920 	struct net_device *idev;
2921 	int err;
2922 
2923 	/* Check if mlx5en net device already exists */
2924 	VNET_LIST_RLOCK();
2925 	VNET_FOREACH(vnet_iter) {
2926 		IFNET_RLOCK();
2927 		CURVNET_SET_QUIET(vnet_iter);
2928 		CK_STAILQ_FOREACH(idev, &V_ifnet, if_link) {
2929 			/* check if network interface belongs to mlx5en */
2930 			if (!mlx5_netdev_match(idev, dev->mdev, "mce"))
2931 				continue;
2932 			write_lock(&dev->roce.netdev_lock);
2933 			dev->roce.netdev = idev;
2934 			write_unlock(&dev->roce.netdev_lock);
2935 		}
2936 		CURVNET_RESTORE();
2937 		IFNET_RUNLOCK();
2938 	}
2939 	VNET_LIST_RUNLOCK();
2940 
2941 	dev->roce.nb.notifier_call = mlx5_netdev_event;
2942 	err = register_netdevice_notifier(&dev->roce.nb);
2943 	if (err) {
2944 		dev->roce.nb.notifier_call = NULL;
2945 		return err;
2946 	}
2947 
2948 	err = mlx5_nic_vport_enable_roce(dev->mdev);
2949 	if (err)
2950 		goto err_unregister_netdevice_notifier;
2951 
2952 	err = mlx5_roce_lag_init(dev);
2953 	if (err)
2954 		goto err_disable_roce;
2955 
2956 	return 0;
2957 
2958 err_disable_roce:
2959 	mlx5_nic_vport_disable_roce(dev->mdev);
2960 
2961 err_unregister_netdevice_notifier:
2962 	mlx5_remove_roce_notifier(dev);
2963 	return err;
2964 }
2965 
2966 static void mlx5_disable_roce(struct mlx5_ib_dev *dev)
2967 {
2968 	mlx5_roce_lag_cleanup(dev);
2969 	mlx5_nic_vport_disable_roce(dev->mdev);
2970 }
2971 
2972 static void mlx5_ib_dealloc_q_port_counter(struct mlx5_ib_dev *dev, u8 port_num)
2973 {
2974 	mlx5_vport_dealloc_q_counter(dev->mdev,
2975 				     MLX5_INTERFACE_PROTOCOL_IB,
2976 				     dev->port[port_num].q_cnt_id);
2977 	dev->port[port_num].q_cnt_id = 0;
2978 }
2979 
2980 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
2981 {
2982 	unsigned int i;
2983 
2984 	for (i = 0; i < dev->num_ports; i++)
2985 		mlx5_ib_dealloc_q_port_counter(dev, i);
2986 }
2987 
2988 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
2989 {
2990 	int i;
2991 	int ret;
2992 
2993 	for (i = 0; i < dev->num_ports; i++) {
2994 		ret = mlx5_vport_alloc_q_counter(dev->mdev,
2995 						 MLX5_INTERFACE_PROTOCOL_IB,
2996 						 &dev->port[i].q_cnt_id);
2997 		if (ret) {
2998 			mlx5_ib_warn(dev,
2999 				     "couldn't allocate queue counter for port %d, err %d\n",
3000 				     i + 1, ret);
3001 			goto dealloc_counters;
3002 		}
3003 	}
3004 
3005 	return 0;
3006 
3007 dealloc_counters:
3008 	while (--i >= 0)
3009 		mlx5_ib_dealloc_q_port_counter(dev, i);
3010 
3011 	return ret;
3012 }
3013 
3014 static const char * const names[] = {
3015 	"rx_write_requests",
3016 	"rx_read_requests",
3017 	"rx_atomic_requests",
3018 	"out_of_buffer",
3019 	"out_of_sequence",
3020 	"duplicate_request",
3021 	"rnr_nak_retry_err",
3022 	"packet_seq_err",
3023 	"implied_nak_seq_err",
3024 	"local_ack_timeout_err",
3025 };
3026 
3027 static const size_t stats_offsets[] = {
3028 	MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests),
3029 	MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests),
3030 	MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests),
3031 	MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer),
3032 	MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence),
3033 	MLX5_BYTE_OFF(query_q_counter_out, duplicate_request),
3034 	MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err),
3035 	MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err),
3036 	MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err),
3037 	MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err),
3038 };
3039 
3040 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
3041 						    u8 port_num)
3042 {
3043 	BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets));
3044 
3045 	/* We support only per port stats */
3046 	if (port_num == 0)
3047 		return NULL;
3048 
3049 	return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names),
3050 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
3051 }
3052 
3053 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
3054 				struct rdma_hw_stats *stats,
3055 				u8 port, int index)
3056 {
3057 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3058 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
3059 	void *out;
3060 	__be32 val;
3061 	int ret;
3062 	int i;
3063 
3064 	if (!port || !stats)
3065 		return -ENOSYS;
3066 
3067 	out = mlx5_vzalloc(outlen);
3068 	if (!out)
3069 		return -ENOMEM;
3070 
3071 	ret = mlx5_vport_query_q_counter(dev->mdev,
3072 					dev->port[port - 1].q_cnt_id, 0,
3073 					out, outlen);
3074 	if (ret)
3075 		goto free;
3076 
3077 	for (i = 0; i < ARRAY_SIZE(names); i++) {
3078 		val = *(__be32 *)(out + stats_offsets[i]);
3079 		stats->value[i] = (u64)be32_to_cpu(val);
3080 	}
3081 free:
3082 	kvfree(out);
3083 	return ARRAY_SIZE(names);
3084 }
3085 
3086 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
3087 {
3088 	struct mlx5_ib_dev *dev;
3089 	enum rdma_link_layer ll;
3090 	int port_type_cap;
3091 	int err;
3092 	int i;
3093 
3094 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3095 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3096 
3097 	if ((ll == IB_LINK_LAYER_ETHERNET) && !MLX5_CAP_GEN(mdev, roce))
3098 		return NULL;
3099 
3100 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
3101 	if (!dev)
3102 		return NULL;
3103 
3104 	dev->mdev = mdev;
3105 
3106 	dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
3107 			    GFP_KERNEL);
3108 	if (!dev->port)
3109 		goto err_dealloc;
3110 
3111 	rwlock_init(&dev->roce.netdev_lock);
3112 	err = get_port_caps(dev);
3113 	if (err)
3114 		goto err_free_port;
3115 
3116 	if (mlx5_use_mad_ifc(dev))
3117 		get_ext_port_caps(dev);
3118 
3119 	MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
3120 
3121 	snprintf(dev->ib_dev.name, IB_DEVICE_NAME_MAX, "mlx5_%d", device_get_unit(mdev->pdev->dev.bsddev));
3122 	dev->ib_dev.owner		= THIS_MODULE;
3123 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
3124 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
3125 	dev->num_ports		= MLX5_CAP_GEN(mdev, num_ports);
3126 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
3127 	dev->ib_dev.num_comp_vectors    =
3128 		dev->mdev->priv.eq_table.num_comp_vectors;
3129 	dev->ib_dev.dma_device	= &mdev->pdev->dev;
3130 
3131 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
3132 	dev->ib_dev.uverbs_cmd_mask	=
3133 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
3134 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
3135 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
3136 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
3137 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
3138 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
3139 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
3140 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
3141 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
3142 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
3143 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
3144 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
3145 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
3146 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
3147 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
3148 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
3149 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
3150 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
3151 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
3152 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
3153 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
3154 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
3155 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
3156 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
3157 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
3158 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
3159 	dev->ib_dev.uverbs_ex_cmd_mask =
3160 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
3161 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
3162 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
3163 
3164 	dev->ib_dev.query_device	= mlx5_ib_query_device;
3165 	dev->ib_dev.query_port		= mlx5_ib_query_port;
3166 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
3167 	if (ll == IB_LINK_LAYER_ETHERNET)
3168 		dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
3169 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
3170 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
3171 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
3172 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
3173 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
3174 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
3175 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
3176 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
3177 	dev->ib_dev.mmap		= mlx5_ib_mmap;
3178 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
3179 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
3180 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
3181 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
3182 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
3183 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
3184 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
3185 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
3186 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
3187 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
3188 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
3189 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
3190 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
3191 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
3192 	dev->ib_dev.post_send		= mlx5_ib_post_send;
3193 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
3194 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
3195 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
3196 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
3197 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
3198 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
3199 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
3200 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
3201 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
3202 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
3203 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
3204 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
3205 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
3206 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
3207 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
3208 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
3209 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
3210 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
3211 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
3212 	if (mlx5_core_is_pf(mdev)) {
3213 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
3214 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
3215 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
3216 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
3217 	}
3218 
3219 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
3220 
3221 	mlx5_ib_internal_fill_odp_caps(dev);
3222 
3223 	if (MLX5_CAP_GEN(mdev, imaicl)) {
3224 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
3225 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
3226 		dev->ib_dev.uverbs_cmd_mask |=
3227 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
3228 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
3229 	}
3230 
3231 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) &&
3232 	    MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3233 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
3234 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
3235 	}
3236 
3237 	if (MLX5_CAP_GEN(mdev, xrc)) {
3238 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
3239 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
3240 		dev->ib_dev.uverbs_cmd_mask |=
3241 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
3242 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
3243 	}
3244 
3245 	if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
3246 	    IB_LINK_LAYER_ETHERNET) {
3247 		dev->ib_dev.create_flow	= mlx5_ib_create_flow;
3248 		dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
3249 		dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
3250 		dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
3251 		dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
3252 		dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
3253 		dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
3254 		dev->ib_dev.uverbs_ex_cmd_mask |=
3255 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
3256 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
3257 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
3258 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
3259 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
3260 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
3261 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
3262 	}
3263 	err = init_node_data(dev);
3264 	if (err)
3265 		goto err_free_port;
3266 
3267 	mutex_init(&dev->flow_db.lock);
3268 	mutex_init(&dev->cap_mask_mutex);
3269 	INIT_LIST_HEAD(&dev->qp_list);
3270 	spin_lock_init(&dev->reset_flow_resource_lock);
3271 
3272 	if (ll == IB_LINK_LAYER_ETHERNET) {
3273 		err = mlx5_enable_roce(dev);
3274 		if (err)
3275 			goto err_free_port;
3276 	}
3277 
3278 	err = create_dev_resources(&dev->devr);
3279 	if (err)
3280 		goto err_disable_roce;
3281 
3282 	err = mlx5_ib_odp_init_one(dev);
3283 	if (err)
3284 		goto err_rsrc;
3285 
3286 	err = mlx5_ib_alloc_q_counters(dev);
3287 	if (err)
3288 		goto err_odp;
3289 
3290 	err = ib_register_device(&dev->ib_dev, NULL);
3291 	if (err)
3292 		goto err_q_cnt;
3293 
3294 	err = create_umr_res(dev);
3295 	if (err)
3296 		goto err_dev;
3297 
3298 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
3299 		err = device_create_file(&dev->ib_dev.dev,
3300 					 mlx5_class_attributes[i]);
3301 		if (err)
3302 			goto err_umrc;
3303 	}
3304 
3305 	err = mlx5_ib_init_congestion(dev);
3306 	if (err)
3307 		goto err_umrc;
3308 
3309 	dev->ib_active = true;
3310 
3311 	return dev;
3312 
3313 err_umrc:
3314 	destroy_umrc_res(dev);
3315 
3316 err_dev:
3317 	ib_unregister_device(&dev->ib_dev);
3318 
3319 err_q_cnt:
3320 	mlx5_ib_dealloc_q_counters(dev);
3321 
3322 err_odp:
3323 	mlx5_ib_odp_remove_one(dev);
3324 
3325 err_rsrc:
3326 	destroy_dev_resources(&dev->devr);
3327 
3328 err_disable_roce:
3329 	if (ll == IB_LINK_LAYER_ETHERNET) {
3330 		mlx5_disable_roce(dev);
3331 		mlx5_remove_roce_notifier(dev);
3332 	}
3333 
3334 err_free_port:
3335 	kfree(dev->port);
3336 
3337 err_dealloc:
3338 	ib_dealloc_device((struct ib_device *)dev);
3339 
3340 	return NULL;
3341 }
3342 
3343 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3344 {
3345 	struct mlx5_ib_dev *dev = context;
3346 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3347 
3348 	mlx5_ib_cleanup_congestion(dev);
3349 	mlx5_remove_roce_notifier(dev);
3350 	ib_unregister_device(&dev->ib_dev);
3351 	mlx5_ib_dealloc_q_counters(dev);
3352 	destroy_umrc_res(dev);
3353 	mlx5_ib_odp_remove_one(dev);
3354 	destroy_dev_resources(&dev->devr);
3355 	if (ll == IB_LINK_LAYER_ETHERNET)
3356 		mlx5_disable_roce(dev);
3357 	kfree(dev->port);
3358 	ib_dealloc_device(&dev->ib_dev);
3359 }
3360 
3361 static struct mlx5_interface mlx5_ib_interface = {
3362 	.add            = mlx5_ib_add,
3363 	.remove         = mlx5_ib_remove,
3364 	.event          = mlx5_ib_event,
3365 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
3366 };
3367 
3368 static int __init mlx5_ib_init(void)
3369 {
3370 	int err;
3371 
3372 	err = mlx5_ib_odp_init();
3373 	if (err)
3374 		return err;
3375 
3376 	err = mlx5_register_interface(&mlx5_ib_interface);
3377 	if (err)
3378 		goto clean_odp;
3379 
3380 	return err;
3381 
3382 clean_odp:
3383 	mlx5_ib_odp_cleanup();
3384 	return err;
3385 }
3386 
3387 static void __exit mlx5_ib_cleanup(void)
3388 {
3389 	mlx5_unregister_interface(&mlx5_ib_interface);
3390 	mlx5_ib_odp_cleanup();
3391 }
3392 
3393 static void
3394 mlx5_ib_show_version(void __unused *arg)
3395 {
3396 
3397 	printf("%s", mlx5_version);
3398 }
3399 SYSINIT(mlx5_ib_show_version, SI_SUB_DRIVERS, SI_ORDER_ANY, mlx5_ib_show_version, NULL);
3400 
3401 module_init_order(mlx5_ib_init, SI_ORDER_THIRD);
3402 module_exit_order(mlx5_ib_cleanup, SI_ORDER_THIRD);
3403