xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_main.c (revision 77013d11e6483b970af25e13c9b892075742f7e5)
1 /*-
2  * Copyright (c) 2013-2021, 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 <rdma/uverbs_ioctl.h>
49 #include <linux/in.h>
50 #include <linux/etherdevice.h>
51 #include <dev/mlx5/fs.h>
52 #include "mlx5_ib.h"
53 
54 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
55 MODULE_LICENSE("Dual BSD/GPL");
56 MODULE_DEPEND(mlx5ib, linuxkpi, 1, 1, 1);
57 MODULE_DEPEND(mlx5ib, mlx5, 1, 1, 1);
58 MODULE_DEPEND(mlx5ib, ibcore, 1, 1, 1);
59 MODULE_VERSION(mlx5ib, 1);
60 
61 enum {
62 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
63 };
64 
65 static enum rdma_link_layer
66 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
67 {
68 	switch (port_type_cap) {
69 	case MLX5_CAP_PORT_TYPE_IB:
70 		return IB_LINK_LAYER_INFINIBAND;
71 	case MLX5_CAP_PORT_TYPE_ETH:
72 		return IB_LINK_LAYER_ETHERNET;
73 	default:
74 		return IB_LINK_LAYER_UNSPECIFIED;
75 	}
76 }
77 
78 static enum rdma_link_layer
79 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
80 {
81 	struct mlx5_ib_dev *dev = to_mdev(device);
82 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
83 
84 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
85 }
86 
87 static bool mlx5_netdev_match(struct ifnet *ndev,
88 			      struct mlx5_core_dev *mdev,
89 			      const char *dname)
90 {
91 	return ndev->if_type == IFT_ETHER &&
92 	  ndev->if_dname != NULL &&
93 	  strcmp(ndev->if_dname, dname) == 0 &&
94 	  ndev->if_softc != NULL &&
95 	  *(struct mlx5_core_dev **)ndev->if_softc == mdev;
96 }
97 
98 static int mlx5_netdev_event(struct notifier_block *this,
99 			     unsigned long event, void *ptr)
100 {
101 	struct ifnet *ndev = netdev_notifier_info_to_ifp(ptr);
102 	struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
103 						 roce.nb);
104 
105 	switch (event) {
106 	case NETDEV_REGISTER:
107 	case NETDEV_UNREGISTER:
108 		write_lock(&ibdev->roce.netdev_lock);
109 		/* check if network interface belongs to mlx5en */
110 		if (mlx5_netdev_match(ndev, ibdev->mdev, "mce"))
111 			ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
112 					     NULL : ndev;
113 		write_unlock(&ibdev->roce.netdev_lock);
114 		break;
115 
116 	case NETDEV_UP:
117 	case NETDEV_DOWN: {
118 		struct ifnet *upper = NULL;
119 
120 		if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
121 		    && ibdev->ib_active) {
122 			struct ib_event ibev = {0};
123 
124 			ibev.device = &ibdev->ib_dev;
125 			ibev.event = (event == NETDEV_UP) ?
126 				     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
127 			ibev.element.port_num = 1;
128 			ib_dispatch_event(&ibev);
129 		}
130 		break;
131 	}
132 
133 	default:
134 		break;
135 	}
136 
137 	return NOTIFY_DONE;
138 }
139 
140 static struct ifnet *mlx5_ib_get_netdev(struct ib_device *device,
141 					     u8 port_num)
142 {
143 	struct mlx5_ib_dev *ibdev = to_mdev(device);
144 	struct ifnet *ndev;
145 
146 	/* Ensure ndev does not disappear before we invoke if_ref()
147 	 */
148 	read_lock(&ibdev->roce.netdev_lock);
149 	ndev = ibdev->roce.netdev;
150 	if (ndev)
151 		if_ref(ndev);
152 	read_unlock(&ibdev->roce.netdev_lock);
153 
154 	return ndev;
155 }
156 
157 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
158 				    u8 *active_width)
159 {
160 	switch (eth_proto_oper) {
161 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
162 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
163 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
164 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
165 		*active_width = IB_WIDTH_1X;
166 		*active_speed = IB_SPEED_SDR;
167 		break;
168 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
169 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
170 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
171 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
172 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
173 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
174 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER_LR):
175 		*active_width = IB_WIDTH_1X;
176 		*active_speed = IB_SPEED_QDR;
177 		break;
178 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
179 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
180 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
181 		*active_width = IB_WIDTH_1X;
182 		*active_speed = IB_SPEED_EDR;
183 		break;
184 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
185 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
186 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
187 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4_ER4):
188 		*active_width = IB_WIDTH_4X;
189 		*active_speed = IB_SPEED_QDR;
190 		break;
191 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
192 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
193 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR4):
194 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
195 		*active_width = IB_WIDTH_1X;
196 		*active_speed = IB_SPEED_HDR;
197 		break;
198 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
199 		*active_width = IB_WIDTH_4X;
200 		*active_speed = IB_SPEED_FDR;
201 		break;
202 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
203 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
204 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
205 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
206 		*active_width = IB_WIDTH_4X;
207 		*active_speed = IB_SPEED_EDR;
208 		break;
209 	default:
210 		*active_width = IB_WIDTH_4X;
211 		*active_speed = IB_SPEED_QDR;
212 		return -EINVAL;
213 	}
214 
215 	return 0;
216 }
217 
218 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u8 *active_speed,
219 					u8 *active_width)
220 {
221 	switch (eth_proto_oper) {
222 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
223 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
224 		*active_width = IB_WIDTH_1X;
225 		*active_speed = IB_SPEED_SDR;
226 		break;
227 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
228 		*active_width = IB_WIDTH_1X;
229 		*active_speed = IB_SPEED_DDR;
230 		break;
231 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
232 		*active_width = IB_WIDTH_1X;
233 		*active_speed = IB_SPEED_QDR;
234 		break;
235 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
236 		*active_width = IB_WIDTH_4X;
237 		*active_speed = IB_SPEED_QDR;
238 		break;
239 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
240 		*active_width = IB_WIDTH_1X;
241 		*active_speed = IB_SPEED_EDR;
242 		break;
243 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
244 		*active_width = IB_WIDTH_2X;
245 		*active_speed = IB_SPEED_EDR;
246 		break;
247 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
248 		*active_width = IB_WIDTH_1X;
249 		*active_speed = IB_SPEED_HDR;
250 		break;
251 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
252 		*active_width = IB_WIDTH_4X;
253 		*active_speed = IB_SPEED_EDR;
254 		break;
255 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
256 		*active_width = IB_WIDTH_2X;
257 		*active_speed = IB_SPEED_HDR;
258 		break;
259 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
260 		*active_width = IB_WIDTH_4X;
261 		*active_speed = IB_SPEED_HDR;
262 		break;
263 	default:
264 		*active_width = IB_WIDTH_4X;
265 		*active_speed = IB_SPEED_QDR;
266 		return -EINVAL;
267 	}
268 
269 	return 0;
270 }
271 
272 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
273 				struct ib_port_attr *props)
274 {
275 	struct mlx5_ib_dev *dev = to_mdev(device);
276 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {};
277 	struct ifnet *ndev;
278 	enum ib_mtu ndev_ib_mtu;
279 	u16 qkey_viol_cntr;
280 	u32 eth_prot_oper;
281 	bool ext;
282 	int err;
283 
284 	memset(props, 0, sizeof(*props));
285 
286 	/* Possible bad flows are checked before filling out props so in case
287 	 * of an error it will still be zeroed out.
288 	 */
289 	err = mlx5_query_port_ptys(dev->mdev, out, sizeof(out), MLX5_PTYS_EN,
290 	    port_num);
291 	if (err)
292 		return err;
293 
294 	ext = MLX5_CAP_PCAM_FEATURE(dev->mdev, ptys_extended_ethernet);
295 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
296 
297 	if (ext)
298 		translate_eth_ext_proto_oper(eth_prot_oper, &props->active_speed,
299 		    &props->active_width);
300 	else
301 		translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
302 		    &props->active_width);
303 
304 	props->port_cap_flags  |= IB_PORT_CM_SUP;
305 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
306 
307 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
308 						roce_address_table_size);
309 	props->max_mtu          = IB_MTU_4096;
310 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
311 	props->pkey_tbl_len     = 1;
312 	props->state            = IB_PORT_DOWN;
313 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
314 
315 	mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
316 	props->qkey_viol_cntr = qkey_viol_cntr;
317 
318 	ndev = mlx5_ib_get_netdev(device, port_num);
319 	if (!ndev)
320 		return 0;
321 
322 	if (ndev->if_drv_flags & IFF_DRV_RUNNING &&
323 	    ndev->if_link_state == LINK_STATE_UP) {
324 		props->state      = IB_PORT_ACTIVE;
325 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
326 	}
327 
328 	ndev_ib_mtu = iboe_get_mtu(ndev->if_mtu);
329 
330 	if_rele(ndev);
331 
332 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
333 	return 0;
334 }
335 
336 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
337 				     const struct ib_gid_attr *attr,
338 				     void *mlx5_addr)
339 {
340 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
341 	char *mlx5_addr_l3_addr	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
342 					       source_l3_address);
343 	void *mlx5_addr_mac	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
344 					       source_mac_47_32);
345 	u16 vlan_id;
346 
347 	if (!gid)
348 		return;
349 	ether_addr_copy(mlx5_addr_mac, IF_LLADDR(attr->ndev));
350 
351 	vlan_id = rdma_vlan_dev_vlan_id(attr->ndev);
352 	if (vlan_id != 0xffff) {
353 		MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
354 		MLX5_SET_RA(mlx5_addr, vlan_id, vlan_id);
355 	}
356 
357 	switch (attr->gid_type) {
358 	case IB_GID_TYPE_IB:
359 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
360 		break;
361 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
362 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
363 		break;
364 
365 	default:
366 		WARN_ON(true);
367 	}
368 
369 	if (attr->gid_type != IB_GID_TYPE_IB) {
370 		if (ipv6_addr_v4mapped((void *)gid))
371 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
372 				    MLX5_ROCE_L3_TYPE_IPV4);
373 		else
374 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
375 				    MLX5_ROCE_L3_TYPE_IPV6);
376 	}
377 
378 	if ((attr->gid_type == IB_GID_TYPE_IB) ||
379 	    !ipv6_addr_v4mapped((void *)gid))
380 		memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
381 	else
382 		memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
383 }
384 
385 static int set_roce_addr(struct ib_device *device, u8 port_num,
386 			 unsigned int index,
387 			 const union ib_gid *gid,
388 			 const struct ib_gid_attr *attr)
389 {
390 	struct mlx5_ib_dev *dev = to_mdev(device);
391 	u32  in[MLX5_ST_SZ_DW(set_roce_address_in)]  = {0};
392 	u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0};
393 	void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
394 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
395 
396 	if (ll != IB_LINK_LAYER_ETHERNET)
397 		return -EINVAL;
398 
399 	ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
400 
401 	MLX5_SET(set_roce_address_in, in, roce_address_index, index);
402 	MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
403 	return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
404 }
405 
406 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
407 			   unsigned int index, const union ib_gid *gid,
408 			   const struct ib_gid_attr *attr,
409 			   __always_unused void **context)
410 {
411 	return set_roce_addr(device, port_num, index, gid, attr);
412 }
413 
414 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
415 			   unsigned int index, __always_unused void **context)
416 {
417 	return set_roce_addr(device, port_num, index, NULL, NULL);
418 }
419 
420 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
421 			       int index)
422 {
423 	struct ib_gid_attr attr;
424 	union ib_gid gid;
425 
426 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
427 		return 0;
428 
429 	if (!attr.ndev)
430 		return 0;
431 
432 	if_rele(attr.ndev);
433 
434 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
435 		return 0;
436 
437 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
438 }
439 
440 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
441 			   int index, enum ib_gid_type *gid_type)
442 {
443 	struct ib_gid_attr attr;
444 	union ib_gid gid;
445 	int ret;
446 
447 	ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
448 	if (ret)
449 		return ret;
450 
451 	if (!attr.ndev)
452 		return -ENODEV;
453 
454 	if_rele(attr.ndev);
455 
456 	*gid_type = attr.gid_type;
457 
458 	return 0;
459 }
460 
461 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
462 {
463 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
464 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
465 	return 0;
466 }
467 
468 enum {
469 	MLX5_VPORT_ACCESS_METHOD_MAD,
470 	MLX5_VPORT_ACCESS_METHOD_HCA,
471 	MLX5_VPORT_ACCESS_METHOD_NIC,
472 };
473 
474 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
475 {
476 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
477 		return MLX5_VPORT_ACCESS_METHOD_MAD;
478 
479 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
480 	    IB_LINK_LAYER_ETHERNET)
481 		return MLX5_VPORT_ACCESS_METHOD_NIC;
482 
483 	return MLX5_VPORT_ACCESS_METHOD_HCA;
484 }
485 
486 static void get_atomic_caps(struct mlx5_ib_dev *dev,
487 			    struct ib_device_attr *props)
488 {
489 	u8 tmp;
490 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
491 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
492 	u8 atomic_req_8B_endianness_mode =
493 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
494 
495 	/* Check if HW supports 8 bytes standard atomic operations and capable
496 	 * of host endianness respond
497 	 */
498 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
499 	if (((atomic_operations & tmp) == tmp) &&
500 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
501 	    (atomic_req_8B_endianness_mode)) {
502 		props->atomic_cap = IB_ATOMIC_HCA;
503 	} else {
504 		props->atomic_cap = IB_ATOMIC_NONE;
505 	}
506 }
507 
508 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
509 					__be64 *sys_image_guid)
510 {
511 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
512 	struct mlx5_core_dev *mdev = dev->mdev;
513 	u64 tmp;
514 	int err;
515 
516 	switch (mlx5_get_vport_access_method(ibdev)) {
517 	case MLX5_VPORT_ACCESS_METHOD_MAD:
518 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
519 							    sys_image_guid);
520 
521 	case MLX5_VPORT_ACCESS_METHOD_HCA:
522 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
523 		break;
524 
525 	case MLX5_VPORT_ACCESS_METHOD_NIC:
526 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
527 		break;
528 
529 	default:
530 		return -EINVAL;
531 	}
532 
533 	if (!err)
534 		*sys_image_guid = cpu_to_be64(tmp);
535 
536 	return err;
537 
538 }
539 
540 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
541 				u16 *max_pkeys)
542 {
543 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
544 	struct mlx5_core_dev *mdev = dev->mdev;
545 
546 	switch (mlx5_get_vport_access_method(ibdev)) {
547 	case MLX5_VPORT_ACCESS_METHOD_MAD:
548 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
549 
550 	case MLX5_VPORT_ACCESS_METHOD_HCA:
551 	case MLX5_VPORT_ACCESS_METHOD_NIC:
552 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
553 						pkey_table_size));
554 		return 0;
555 
556 	default:
557 		return -EINVAL;
558 	}
559 }
560 
561 static int mlx5_query_vendor_id(struct ib_device *ibdev,
562 				u32 *vendor_id)
563 {
564 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
565 
566 	switch (mlx5_get_vport_access_method(ibdev)) {
567 	case MLX5_VPORT_ACCESS_METHOD_MAD:
568 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
569 
570 	case MLX5_VPORT_ACCESS_METHOD_HCA:
571 	case MLX5_VPORT_ACCESS_METHOD_NIC:
572 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
573 
574 	default:
575 		return -EINVAL;
576 	}
577 }
578 
579 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
580 				__be64 *node_guid)
581 {
582 	u64 tmp;
583 	int err;
584 
585 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
586 	case MLX5_VPORT_ACCESS_METHOD_MAD:
587 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
588 
589 	case MLX5_VPORT_ACCESS_METHOD_HCA:
590 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
591 		break;
592 
593 	case MLX5_VPORT_ACCESS_METHOD_NIC:
594 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
595 		break;
596 
597 	default:
598 		return -EINVAL;
599 	}
600 
601 	if (!err)
602 		*node_guid = cpu_to_be64(tmp);
603 
604 	return err;
605 }
606 
607 struct mlx5_reg_node_desc {
608 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
609 };
610 
611 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
612 {
613 	struct mlx5_reg_node_desc in;
614 
615 	if (mlx5_use_mad_ifc(dev))
616 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
617 
618 	memset(&in, 0, sizeof(in));
619 
620 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
621 				    sizeof(struct mlx5_reg_node_desc),
622 				    MLX5_REG_NODE_DESC, 0, 0);
623 }
624 
625 static int mlx5_ib_query_device(struct ib_device *ibdev,
626 				struct ib_device_attr *props,
627 				struct ib_udata *uhw)
628 {
629 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
630 	struct mlx5_core_dev *mdev = dev->mdev;
631 	int err = -ENOMEM;
632 	int max_sq_desc;
633 	int max_rq_sg;
634 	int max_sq_sg;
635 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
636 	struct mlx5_ib_query_device_resp resp = {};
637 	size_t resp_len;
638 	u64 max_tso;
639 
640 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
641 	if (uhw->outlen && uhw->outlen < resp_len)
642 		return -EINVAL;
643 	else
644 		resp.response_length = resp_len;
645 
646 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
647 		return -EINVAL;
648 
649 	memset(props, 0, sizeof(*props));
650 	err = mlx5_query_system_image_guid(ibdev,
651 					   &props->sys_image_guid);
652 	if (err)
653 		return err;
654 
655 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
656 	if (err)
657 		return err;
658 
659 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
660 	if (err)
661 		return err;
662 
663 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
664 		((u32)fw_rev_min(dev->mdev) << 16) |
665 		fw_rev_sub(dev->mdev);
666 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
667 		IB_DEVICE_PORT_ACTIVE_EVENT		|
668 		IB_DEVICE_SYS_IMAGE_GUID		|
669 		IB_DEVICE_RC_RNR_NAK_GEN;
670 
671 	if (MLX5_CAP_GEN(mdev, pkv))
672 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
673 	if (MLX5_CAP_GEN(mdev, qkv))
674 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
675 	if (MLX5_CAP_GEN(mdev, apm))
676 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
677 	if (MLX5_CAP_GEN(mdev, xrc))
678 		props->device_cap_flags |= IB_DEVICE_XRC;
679 	if (MLX5_CAP_GEN(mdev, imaicl)) {
680 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
681 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
682 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
683 		/* We support 'Gappy' memory registration too */
684 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
685 	}
686 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
687 	if (MLX5_CAP_GEN(mdev, sho)) {
688 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
689 		/* At this stage no support for signature handover */
690 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
691 				      IB_PROT_T10DIF_TYPE_2 |
692 				      IB_PROT_T10DIF_TYPE_3;
693 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
694 				       IB_GUARD_T10DIF_CSUM;
695 	}
696 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
697 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
698 
699 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
700 		if (MLX5_CAP_ETH(mdev, csum_cap))
701 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
702 
703 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
704 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
705 			if (max_tso) {
706 				resp.tso_caps.max_tso = 1 << max_tso;
707 				resp.tso_caps.supported_qpts |=
708 					1 << IB_QPT_RAW_PACKET;
709 				resp.response_length += sizeof(resp.tso_caps);
710 			}
711 		}
712 
713 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
714 			resp.rss_caps.rx_hash_function =
715 						MLX5_RX_HASH_FUNC_TOEPLITZ;
716 			resp.rss_caps.rx_hash_fields_mask =
717 						MLX5_RX_HASH_SRC_IPV4 |
718 						MLX5_RX_HASH_DST_IPV4 |
719 						MLX5_RX_HASH_SRC_IPV6 |
720 						MLX5_RX_HASH_DST_IPV6 |
721 						MLX5_RX_HASH_SRC_PORT_TCP |
722 						MLX5_RX_HASH_DST_PORT_TCP |
723 						MLX5_RX_HASH_SRC_PORT_UDP |
724 						MLX5_RX_HASH_DST_PORT_UDP;
725 			resp.response_length += sizeof(resp.rss_caps);
726 		}
727 	} else {
728 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
729 			resp.response_length += sizeof(resp.tso_caps);
730 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
731 			resp.response_length += sizeof(resp.rss_caps);
732 	}
733 
734 	if (MLX5_CAP_GEN(mdev, ipoib_ipoib_offloads)) {
735 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
736 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
737 	}
738 
739 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
740 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs))
741 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
742 
743 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
744 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
745 
746 	props->vendor_part_id	   = mdev->pdev->device;
747 	props->hw_ver		   = mdev->pdev->revision;
748 
749 	props->max_mr_size	   = ~0ull;
750 	props->page_size_cap	   = ~(min_page_size - 1);
751 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
752 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
753 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
754 		     sizeof(struct mlx5_wqe_data_seg);
755 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
756 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
757 		     sizeof(struct mlx5_wqe_raddr_seg)) /
758 		sizeof(struct mlx5_wqe_data_seg);
759 	props->max_sge = min(max_rq_sg, max_sq_sg);
760 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
761 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
762 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
763 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
764 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
765 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
766 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
767 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
768 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
769 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
770 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
771 	props->max_srq_sge	   = max_rq_sg - 1;
772 	props->max_fast_reg_page_list_len =
773 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
774 	get_atomic_caps(dev, props);
775 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
776 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
777 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
778 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
779 					   props->max_mcast_grp;
780 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
781 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
782 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
783 
784 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
785 	if (MLX5_CAP_GEN(mdev, pg))
786 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
787 	props->odp_caps = dev->odp_caps;
788 #endif
789 
790 	if (MLX5_CAP_GEN(mdev, cd))
791 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
792 
793 	if (!mlx5_core_is_pf(mdev))
794 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
795 
796 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
797 	    IB_LINK_LAYER_ETHERNET) {
798 		props->rss_caps.max_rwq_indirection_tables =
799 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
800 		props->rss_caps.max_rwq_indirection_table_size =
801 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
802 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
803 		props->max_wq_type_rq =
804 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
805 	}
806 
807 	if (uhw->outlen) {
808 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
809 
810 		if (err)
811 			return err;
812 	}
813 
814 	return 0;
815 }
816 
817 enum mlx5_ib_width {
818 	MLX5_IB_WIDTH_1X	= 1 << 0,
819 	MLX5_IB_WIDTH_2X	= 1 << 1,
820 	MLX5_IB_WIDTH_4X	= 1 << 2,
821 	MLX5_IB_WIDTH_8X	= 1 << 3,
822 	MLX5_IB_WIDTH_12X	= 1 << 4
823 };
824 
825 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
826 				  u8 *ib_width)
827 {
828 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
829 	int err = 0;
830 
831 	if (active_width & MLX5_IB_WIDTH_1X) {
832 		*ib_width = IB_WIDTH_1X;
833 	} else if (active_width & MLX5_IB_WIDTH_2X) {
834 		*ib_width = IB_WIDTH_2X;
835 	} else if (active_width & MLX5_IB_WIDTH_4X) {
836 		*ib_width = IB_WIDTH_4X;
837 	} else if (active_width & MLX5_IB_WIDTH_8X) {
838 		*ib_width = IB_WIDTH_8X;
839 	} else if (active_width & MLX5_IB_WIDTH_12X) {
840 		*ib_width = IB_WIDTH_12X;
841 	} else {
842 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
843 			    (int)active_width);
844 		err = -EINVAL;
845 	}
846 
847 	return err;
848 }
849 
850 enum ib_max_vl_num {
851 	__IB_MAX_VL_0		= 1,
852 	__IB_MAX_VL_0_1		= 2,
853 	__IB_MAX_VL_0_3		= 3,
854 	__IB_MAX_VL_0_7		= 4,
855 	__IB_MAX_VL_0_14	= 5,
856 };
857 
858 enum mlx5_vl_hw_cap {
859 	MLX5_VL_HW_0	= 1,
860 	MLX5_VL_HW_0_1	= 2,
861 	MLX5_VL_HW_0_2	= 3,
862 	MLX5_VL_HW_0_3	= 4,
863 	MLX5_VL_HW_0_4	= 5,
864 	MLX5_VL_HW_0_5	= 6,
865 	MLX5_VL_HW_0_6	= 7,
866 	MLX5_VL_HW_0_7	= 8,
867 	MLX5_VL_HW_0_14	= 15
868 };
869 
870 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
871 				u8 *max_vl_num)
872 {
873 	switch (vl_hw_cap) {
874 	case MLX5_VL_HW_0:
875 		*max_vl_num = __IB_MAX_VL_0;
876 		break;
877 	case MLX5_VL_HW_0_1:
878 		*max_vl_num = __IB_MAX_VL_0_1;
879 		break;
880 	case MLX5_VL_HW_0_3:
881 		*max_vl_num = __IB_MAX_VL_0_3;
882 		break;
883 	case MLX5_VL_HW_0_7:
884 		*max_vl_num = __IB_MAX_VL_0_7;
885 		break;
886 	case MLX5_VL_HW_0_14:
887 		*max_vl_num = __IB_MAX_VL_0_14;
888 		break;
889 
890 	default:
891 		return -EINVAL;
892 	}
893 
894 	return 0;
895 }
896 
897 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
898 			       struct ib_port_attr *props)
899 {
900 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
901 	struct mlx5_core_dev *mdev = dev->mdev;
902 	u32 *rep;
903 	int replen = MLX5_ST_SZ_BYTES(query_hca_vport_context_out);
904 	struct mlx5_ptys_reg *ptys;
905 	struct mlx5_pmtu_reg *pmtu;
906 	struct mlx5_pvlc_reg pvlc;
907 	void *ctx;
908 	int err;
909 
910 	rep = mlx5_vzalloc(replen);
911 	ptys = kzalloc(sizeof(*ptys), GFP_KERNEL);
912 	pmtu = kzalloc(sizeof(*pmtu), GFP_KERNEL);
913 	if (!rep || !ptys || !pmtu) {
914 		err = -ENOMEM;
915 		goto out;
916 	}
917 
918 	memset(props, 0, sizeof(*props));
919 
920 	err = mlx5_query_hca_vport_context(mdev, port, 0, rep, replen);
921 	if (err)
922 		goto out;
923 
924 	ctx = MLX5_ADDR_OF(query_hca_vport_context_out, rep, hca_vport_context);
925 
926 	props->lid		= MLX5_GET(hca_vport_context, ctx, lid);
927 	props->lmc		= MLX5_GET(hca_vport_context, ctx, lmc);
928 	props->sm_lid		= MLX5_GET(hca_vport_context, ctx, sm_lid);
929 	props->sm_sl		= MLX5_GET(hca_vport_context, ctx, sm_sl);
930 	props->state		= MLX5_GET(hca_vport_context, ctx, vport_state);
931 	props->phys_state	= MLX5_GET(hca_vport_context, ctx,
932 					port_physical_state);
933 	props->port_cap_flags	= MLX5_GET(hca_vport_context, ctx, cap_mask1);
934 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
935 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
936 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
937 	props->bad_pkey_cntr	= MLX5_GET(hca_vport_context, ctx,
938 					pkey_violation_counter);
939 	props->qkey_viol_cntr	= MLX5_GET(hca_vport_context, ctx,
940 					qkey_violation_counter);
941 	props->subnet_timeout	= MLX5_GET(hca_vport_context, ctx,
942 					subnet_timeout);
943 	props->init_type_reply	= MLX5_GET(hca_vport_context, ctx,
944 					init_type_reply);
945 	props->grh_required	= MLX5_GET(hca_vport_context, ctx, grh_required);
946 
947 	ptys->proto_mask |= MLX5_PTYS_IB;
948 	ptys->local_port = port;
949 	err = mlx5_core_access_ptys(mdev, ptys, 0);
950 	if (err)
951 		goto out;
952 
953 	err = translate_active_width(ibdev, ptys->ib_link_width_oper,
954 				     &props->active_width);
955 	if (err)
956 		goto out;
957 
958 	props->active_speed	= (u8)ptys->ib_proto_oper;
959 
960 	pmtu->local_port = port;
961 	err = mlx5_core_access_pmtu(mdev, pmtu, 0);
962 	if (err)
963 		goto out;
964 
965 	props->max_mtu		= pmtu->max_mtu;
966 	props->active_mtu	= pmtu->oper_mtu;
967 
968 	memset(&pvlc, 0, sizeof(pvlc));
969 	pvlc.local_port = port;
970 	err = mlx5_core_access_pvlc(mdev, &pvlc, 0);
971 	if (err)
972 		goto out;
973 
974 	err = translate_max_vl_num(ibdev, pvlc.vl_hw_cap,
975 				   &props->max_vl_num);
976 out:
977 	kvfree(rep);
978 	kfree(ptys);
979 	kfree(pmtu);
980 	return err;
981 }
982 
983 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
984 		       struct ib_port_attr *props)
985 {
986 	switch (mlx5_get_vport_access_method(ibdev)) {
987 	case MLX5_VPORT_ACCESS_METHOD_MAD:
988 		return mlx5_query_mad_ifc_port(ibdev, port, props);
989 
990 	case MLX5_VPORT_ACCESS_METHOD_HCA:
991 		return mlx5_query_hca_port(ibdev, port, props);
992 
993 	case MLX5_VPORT_ACCESS_METHOD_NIC:
994 		return mlx5_query_port_roce(ibdev, port, props);
995 
996 	default:
997 		return -EINVAL;
998 	}
999 }
1000 
1001 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1002 			     union ib_gid *gid)
1003 {
1004 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1005 	struct mlx5_core_dev *mdev = dev->mdev;
1006 
1007 	switch (mlx5_get_vport_access_method(ibdev)) {
1008 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1009 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1010 
1011 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1012 		return mlx5_query_hca_vport_gid(mdev, port, 0, index, gid);
1013 
1014 	default:
1015 		return -EINVAL;
1016 	}
1017 
1018 }
1019 
1020 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1021 			      u16 *pkey)
1022 {
1023 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1024 	struct mlx5_core_dev *mdev = dev->mdev;
1025 
1026 	switch (mlx5_get_vport_access_method(ibdev)) {
1027 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1028 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1029 
1030 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1031 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1032 		return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
1033 						 pkey);
1034 	default:
1035 		return -EINVAL;
1036 	}
1037 }
1038 
1039 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1040 				 struct ib_device_modify *props)
1041 {
1042 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1043 	struct mlx5_reg_node_desc in;
1044 	struct mlx5_reg_node_desc out;
1045 	int err;
1046 
1047 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1048 		return -EOPNOTSUPP;
1049 
1050 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1051 		return 0;
1052 
1053 	/*
1054 	 * If possible, pass node desc to FW, so it can generate
1055 	 * a 144 trap.  If cmd fails, just ignore.
1056 	 */
1057 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1058 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1059 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1060 	if (err)
1061 		return err;
1062 
1063 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1064 
1065 	return err;
1066 }
1067 
1068 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1069 			       struct ib_port_modify *props)
1070 {
1071 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1072 	struct ib_port_attr attr;
1073 	u32 tmp;
1074 	int err;
1075 
1076 	/*
1077 	 * CM layer calls ib_modify_port() regardless of the link
1078 	 * layer. For Ethernet ports, qkey violation and Port
1079 	 * capabilities are meaningless.
1080 	 */
1081 	if (mlx5_ib_port_link_layer(ibdev, port) == IB_LINK_LAYER_ETHERNET)
1082 		return 0;
1083 
1084 	mutex_lock(&dev->cap_mask_mutex);
1085 
1086 	err = mlx5_ib_query_port(ibdev, port, &attr);
1087 	if (err)
1088 		goto out;
1089 
1090 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1091 		~props->clr_port_cap_mask;
1092 
1093 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1094 
1095 out:
1096 	mutex_unlock(&dev->cap_mask_mutex);
1097 	return err;
1098 }
1099 
1100 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1101 {
1102 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1103 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1104 }
1105 
1106 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1107 {
1108 	/* Large page with non 4k uar support might limit the dynamic size */
1109 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1110 		return MLX5_MIN_DYN_BFREGS;
1111 
1112 	return MLX5_MAX_DYN_BFREGS;
1113 }
1114 
1115 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1116 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1117 			     struct mlx5_bfreg_info *bfregi)
1118 {
1119 	int uars_per_sys_page;
1120 	int bfregs_per_sys_page;
1121 	int ref_bfregs = req->total_num_bfregs;
1122 
1123 	if (req->total_num_bfregs == 0)
1124 		return -EINVAL;
1125 
1126 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1127 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1128 
1129 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1130 		return -ENOMEM;
1131 
1132 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1133 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1134 	/* This holds the required static allocation asked by the user */
1135 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1136 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1137 		return -EINVAL;
1138 
1139 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1140 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1141 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1142 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1143 
1144 	mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, allocated %d, total bfregs %d, using %d sys pages\n",
1145 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1146 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1147 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1148 		    bfregi->num_sys_pages);
1149 
1150 	return 0;
1151 }
1152 
1153 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1154 {
1155 	struct mlx5_bfreg_info *bfregi;
1156 	int err;
1157 	int i;
1158 
1159 	bfregi = &context->bfregi;
1160 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1161 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1162 		if (err)
1163 			goto error;
1164 
1165 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1166 	}
1167 
1168 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1169 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1170 
1171 	return 0;
1172 
1173 error:
1174 	for (--i; i >= 0; i--)
1175 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1176 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1177 
1178 	return err;
1179 }
1180 
1181 static void deallocate_uars(struct mlx5_ib_dev *dev,
1182 			    struct mlx5_ib_ucontext *context)
1183 {
1184 	struct mlx5_bfreg_info *bfregi;
1185 	int i;
1186 
1187 	bfregi = &context->bfregi;
1188 	for (i = 0; i < bfregi->num_sys_pages; i++)
1189 		if (i < bfregi->num_static_sys_pages ||
1190 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1191 			mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1192 }
1193 
1194 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1195 					  u16 uid)
1196 {
1197 	int err;
1198 
1199 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1200 		return 0;
1201 
1202 	err = mlx5_alloc_transport_domain(dev->mdev, tdn, uid);
1203 	if (err)
1204 		return err;
1205 
1206 	return 0;
1207 }
1208 
1209 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1210 					     u16 uid)
1211 {
1212 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1213 		return;
1214 
1215 	mlx5_dealloc_transport_domain(dev->mdev, tdn, uid);
1216 }
1217 
1218 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1219 				  struct ib_udata *udata)
1220 {
1221 	struct ib_device *ibdev = uctx->device;
1222 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1223 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1224 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1225 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1226 	struct mlx5_bfreg_info *bfregi;
1227 	int ver;
1228 	int err;
1229 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1230 				     max_cqe_version);
1231 	bool lib_uar_4k;
1232 	bool lib_uar_dyn;
1233 
1234 	if (!dev->ib_active)
1235 		return -EAGAIN;
1236 
1237 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1238 		ver = 0;
1239 	else if (udata->inlen >= min_req_v2)
1240 		ver = 2;
1241 	else
1242 		return -EINVAL;
1243 
1244 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1245 	if (err)
1246 		return err;
1247 
1248 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1249 		return -EOPNOTSUPP;
1250 
1251 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1252 		return -EOPNOTSUPP;
1253 
1254 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1255 				    MLX5_NON_FP_BFREGS_PER_UAR);
1256 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1257 		return -EINVAL;
1258 
1259 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1260 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1261 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1262 	resp.cache_line_size = cache_line_size();
1263 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1264 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1265 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1266 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1267 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1268 	resp.cqe_version = min_t(__u8,
1269 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1270 				 req.max_cqe_version);
1271 	resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1272 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1273 	resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1274 					MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1275 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1276 				   sizeof(resp.response_length), udata->outlen);
1277 
1278 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1279 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1280 	bfregi = &context->bfregi;
1281 
1282 	if (lib_uar_dyn) {
1283 		bfregi->lib_uar_dyn = lib_uar_dyn;
1284 		goto uar_done;
1285 	}
1286 
1287 	/* updates req->total_num_bfregs */
1288 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1289 	if (err)
1290 		goto out_ctx;
1291 
1292 	mutex_init(&bfregi->lock);
1293 	bfregi->lib_uar_4k = lib_uar_4k;
1294 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1295 				GFP_KERNEL);
1296 	if (!bfregi->count) {
1297 		err = -ENOMEM;
1298 		goto out_ctx;
1299 	}
1300 
1301 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1302 				    sizeof(*bfregi->sys_pages),
1303 				    GFP_KERNEL);
1304 	if (!bfregi->sys_pages) {
1305 		err = -ENOMEM;
1306 		goto out_count;
1307 	}
1308 
1309 	err = allocate_uars(dev, context);
1310 	if (err)
1311 		goto out_sys_pages;
1312 
1313 uar_done:
1314 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1315 		err = mlx5_ib_devx_create(dev, true);
1316 		if (err < 0)
1317 			goto out_uars;
1318 		context->devx_uid = err;
1319 	}
1320 
1321 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1322 					     context->devx_uid);
1323 	if (err)
1324 		goto out_devx;
1325 
1326 	INIT_LIST_HEAD(&context->db_page_list);
1327 	mutex_init(&context->db_page_mutex);
1328 
1329 	resp.tot_bfregs = lib_uar_dyn ? 0 : req.total_num_bfregs;
1330 	resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1331 
1332 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1333 		resp.response_length += sizeof(resp.cqe_version);
1334 
1335 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1336 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1337 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1338 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1339 	}
1340 
1341 	/*
1342 	 * We don't want to expose information from the PCI bar that is located
1343 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1344 	 * pretend we don't support reading the HCA's core clock. This is also
1345 	 * forced by mmap function.
1346 	 */
1347 	if (offsetofend(typeof(resp), hca_core_clock_offset) <= udata->outlen) {
1348 		if (PAGE_SIZE <= 4096) {
1349 			resp.comp_mask |=
1350 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1351 			resp.hca_core_clock_offset =
1352 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1353 		}
1354 		resp.response_length += sizeof(resp.hca_core_clock_offset);
1355 	}
1356 
1357 	if (offsetofend(typeof(resp), log_uar_size) <= udata->outlen)
1358 		resp.response_length += sizeof(resp.log_uar_size);
1359 
1360 	if (offsetofend(typeof(resp), num_uars_per_page) <= udata->outlen)
1361 		resp.response_length += sizeof(resp.num_uars_per_page);
1362 
1363 	if (offsetofend(typeof(resp), num_dyn_bfregs) <= udata->outlen) {
1364 		resp.num_dyn_bfregs = bfregi->num_dyn_bfregs;
1365 		resp.response_length += sizeof(resp.num_dyn_bfregs);
1366 	}
1367 
1368 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1369 	if (err)
1370 		goto out_mdev;
1371 
1372 	bfregi->ver = ver;
1373 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1374 	context->cqe_version = resp.cqe_version;
1375 	context->lib_caps = req.lib_caps;
1376 	print_lib_caps(dev, context->lib_caps);
1377 
1378 	return 0;
1379 
1380 out_mdev:
1381 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1382 out_devx:
1383 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1384 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1385 
1386 out_uars:
1387 	deallocate_uars(dev, context);
1388 
1389 out_sys_pages:
1390 	kfree(bfregi->sys_pages);
1391 
1392 out_count:
1393 	kfree(bfregi->count);
1394 
1395 out_ctx:
1396 	return err;
1397 }
1398 
1399 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1400 {
1401 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1402 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1403 	struct mlx5_bfreg_info *bfregi;
1404 
1405 	bfregi = &context->bfregi;
1406 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1407 
1408 	if (context->devx_uid)
1409 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1410 
1411 	deallocate_uars(dev, context);
1412 	kfree(bfregi->sys_pages);
1413 	kfree(bfregi->count);
1414 }
1415 
1416 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1417 				 int uar_idx)
1418 {
1419 	int fw_uars_per_page;
1420 
1421 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1422 
1423 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
1424 }
1425 
1426 static int get_command(unsigned long offset)
1427 {
1428 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1429 }
1430 
1431 static int get_arg(unsigned long offset)
1432 {
1433 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1434 }
1435 
1436 static int get_index(unsigned long offset)
1437 {
1438 	return get_arg(offset);
1439 }
1440 
1441 /* Index resides in an extra byte to enable larger values than 255 */
1442 static int get_extended_index(unsigned long offset)
1443 {
1444 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
1445 }
1446 
1447 
1448 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1449 {
1450 }
1451 
1452 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1453 {
1454 	switch (cmd) {
1455 	case MLX5_IB_MMAP_WC_PAGE:
1456 		return "WC";
1457 	case MLX5_IB_MMAP_REGULAR_PAGE:
1458 		return "best effort WC";
1459 	case MLX5_IB_MMAP_NC_PAGE:
1460 		return "NC";
1461 	default:
1462 		return NULL;
1463 	}
1464 }
1465 
1466 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
1467 					struct vm_area_struct *vma,
1468 					struct mlx5_ib_ucontext *context)
1469 {
1470 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
1471 	    !(vma->vm_flags & VM_SHARED))
1472 		return -EINVAL;
1473 
1474 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
1475 		return -EOPNOTSUPP;
1476 
1477 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1478 		return -EPERM;
1479 
1480 	return -EOPNOTSUPP;
1481 }
1482 
1483 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
1484 {
1485 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
1486 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
1487 
1488 	switch (mentry->mmap_flag) {
1489 	case MLX5_IB_MMAP_TYPE_UAR_WC:
1490 	case MLX5_IB_MMAP_TYPE_UAR_NC:
1491 		mlx5_cmd_free_uar(dev->mdev, mentry->page_idx);
1492 		kfree(mentry);
1493 		break;
1494 	default:
1495 		WARN_ON(true);
1496 	}
1497 }
1498 
1499 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1500 		    struct vm_area_struct *vma,
1501 		    struct mlx5_ib_ucontext *context)
1502 {
1503 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1504 	int err;
1505 	unsigned long idx;
1506 	phys_addr_t pfn;
1507 	pgprot_t prot;
1508 	u32 bfreg_dyn_idx = 0;
1509 	u32 uar_index;
1510 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
1511 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
1512 				bfregi->num_static_sys_pages;
1513 
1514 	if (bfregi->lib_uar_dyn)
1515 		return -EINVAL;
1516 
1517 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1518 		return -EINVAL;
1519 
1520 	if (dyn_uar)
1521 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
1522 	else
1523 		idx = get_index(vma->vm_pgoff);
1524 
1525 	if (idx >= max_valid_idx) {
1526 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
1527 			     idx, max_valid_idx);
1528 		return -EINVAL;
1529 	}
1530 
1531 	switch (cmd) {
1532 	case MLX5_IB_MMAP_WC_PAGE:
1533 	case MLX5_IB_MMAP_ALLOC_WC:
1534 	case MLX5_IB_MMAP_REGULAR_PAGE:
1535 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1536 		prot = pgprot_writecombine(vma->vm_page_prot);
1537 		break;
1538 	case MLX5_IB_MMAP_NC_PAGE:
1539 		prot = pgprot_noncached(vma->vm_page_prot);
1540 		break;
1541 	default:
1542 		return -EINVAL;
1543 	}
1544 
1545 	if (dyn_uar) {
1546 		int uars_per_page;
1547 
1548 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
1549 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
1550 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
1551 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
1552 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
1553 			return -EINVAL;
1554 		}
1555 
1556 		mutex_lock(&bfregi->lock);
1557 		/* Fail if uar already allocated, first bfreg index of each
1558 		 * page holds its count.
1559 		 */
1560 		if (bfregi->count[bfreg_dyn_idx]) {
1561 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
1562 			mutex_unlock(&bfregi->lock);
1563 			return -EINVAL;
1564 		}
1565 
1566 		bfregi->count[bfreg_dyn_idx]++;
1567 		mutex_unlock(&bfregi->lock);
1568 
1569 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
1570 		if (err) {
1571 			mlx5_ib_warn(dev, "UAR alloc failed\n");
1572 			goto free_bfreg;
1573 		}
1574 	} else {
1575 		uar_index = bfregi->sys_pages[idx];
1576 	}
1577 
1578 	pfn = uar_index2pfn(dev, uar_index);
1579 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1580 
1581 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
1582 				prot, NULL);
1583 	if (err) {
1584 		mlx5_ib_err(dev,
1585 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
1586 			    err, mmap_cmd2str(cmd));
1587 		goto err;
1588 	}
1589 
1590 	if (dyn_uar)
1591 		bfregi->sys_pages[idx] = uar_index;
1592 	return 0;
1593 
1594 err:
1595 	if (!dyn_uar)
1596 		return err;
1597 
1598 	mlx5_cmd_free_uar(dev->mdev, idx);
1599 
1600 free_bfreg:
1601 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
1602 
1603 	return err;
1604 }
1605 
1606 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
1607 {
1608 	unsigned long idx;
1609 	u8 command;
1610 
1611 	command = get_command(vma->vm_pgoff);
1612 	idx = get_extended_index(vma->vm_pgoff);
1613 
1614 	return (command << 16 | idx);
1615 }
1616 
1617 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
1618 			       struct vm_area_struct *vma,
1619 			       struct ib_ucontext *ucontext)
1620 {
1621 	struct mlx5_user_mmap_entry *mentry;
1622 	struct rdma_user_mmap_entry *entry;
1623 	unsigned long pgoff;
1624 	pgprot_t prot;
1625 	phys_addr_t pfn;
1626 	int ret;
1627 
1628 	pgoff = mlx5_vma_to_pgoff(vma);
1629 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
1630 	if (!entry)
1631 		return -EINVAL;
1632 
1633 	mentry = to_mmmap(entry);
1634 	pfn = (mentry->address >> PAGE_SHIFT);
1635 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
1636 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
1637 		prot = pgprot_noncached(vma->vm_page_prot);
1638 	else
1639 		prot = pgprot_writecombine(vma->vm_page_prot);
1640 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
1641 				entry->npages * PAGE_SIZE,
1642 				prot,
1643 				entry);
1644 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
1645 	return ret;
1646 }
1647 
1648 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1649 {
1650 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1651 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1652 	unsigned long command;
1653 	phys_addr_t pfn;
1654 
1655 	command = get_command(vma->vm_pgoff);
1656 	switch (command) {
1657 	case MLX5_IB_MMAP_WC_PAGE:
1658 	case MLX5_IB_MMAP_ALLOC_WC:
1659 		if (!dev->wc_support)
1660 			return -EPERM;
1661 		/* FALLTHROUGH */
1662 	case MLX5_IB_MMAP_NC_PAGE:
1663 	case MLX5_IB_MMAP_REGULAR_PAGE:
1664 		return uar_mmap(dev, command, vma, context);
1665 
1666 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1667 		return -ENOSYS;
1668 
1669 	case MLX5_IB_MMAP_CORE_CLOCK:
1670 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1671 			return -EINVAL;
1672 
1673 		if (vma->vm_flags & VM_WRITE)
1674 			return -EPERM;
1675 
1676 		/* Don't expose to user-space information it shouldn't have */
1677 		if (PAGE_SIZE > 4096)
1678 			return -EOPNOTSUPP;
1679 
1680 		pfn = (dev->mdev->iseg_base +
1681 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1682 			PAGE_SHIFT;
1683 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
1684 					 PAGE_SIZE,
1685 					 pgprot_noncached(vma->vm_page_prot),
1686 					 NULL);
1687 	case MLX5_IB_MMAP_CLOCK_INFO:
1688 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
1689 
1690 	default:
1691 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
1692 	}
1693 
1694 	return 0;
1695 }
1696 
1697 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
1698 {
1699 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
1700 	struct ib_device *ibdev = ibpd->device;
1701 	struct mlx5_ib_alloc_pd_resp resp;
1702 	int err;
1703 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
1704 		udata, struct mlx5_ib_ucontext, ibucontext);
1705 	u16 uid = context ? context->devx_uid : 0;
1706 
1707 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn, uid);
1708 	if (err)
1709 		return (err);
1710 
1711 	pd->uid = uid;
1712 	if (udata) {
1713 		resp.pdn = pd->pdn;
1714 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1715 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
1716 			return -EFAULT;
1717 		}
1718 	}
1719 
1720 	return 0;
1721 }
1722 
1723 static void mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
1724 {
1725 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1726 	struct mlx5_ib_pd *mpd = to_mpd(pd);
1727 
1728 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
1729 }
1730 
1731 enum {
1732 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
1733 	MATCH_CRITERIA_ENABLE_MISC_BIT,
1734 	MATCH_CRITERIA_ENABLE_INNER_BIT
1735 };
1736 
1737 #define HEADER_IS_ZERO(match_criteria, headers)			           \
1738 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1739 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
1740 
1741 static u8 get_match_criteria_enable(u32 *match_criteria)
1742 {
1743 	u8 match_criteria_enable;
1744 
1745 	match_criteria_enable =
1746 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1747 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
1748 	match_criteria_enable |=
1749 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1750 		MATCH_CRITERIA_ENABLE_MISC_BIT;
1751 	match_criteria_enable |=
1752 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1753 		MATCH_CRITERIA_ENABLE_INNER_BIT;
1754 
1755 	return match_criteria_enable;
1756 }
1757 
1758 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1759 {
1760 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1761 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1762 }
1763 
1764 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1765 {
1766 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1767 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1768 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1769 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1770 }
1771 
1772 #define LAST_ETH_FIELD vlan_tag
1773 #define LAST_IB_FIELD sl
1774 #define LAST_IPV4_FIELD tos
1775 #define LAST_IPV6_FIELD traffic_class
1776 #define LAST_TCP_UDP_FIELD src_port
1777 
1778 /* Field is the last supported field */
1779 #define FIELDS_NOT_SUPPORTED(filter, field)\
1780 	memchr_inv((void *)&filter.field  +\
1781 		   sizeof(filter.field), 0,\
1782 		   sizeof(filter) -\
1783 		   offsetof(typeof(filter), field) -\
1784 		   sizeof(filter.field))
1785 
1786 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1787 			   const union ib_flow_spec *ib_spec)
1788 {
1789 	void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1790 					     outer_headers);
1791 	void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1792 					     outer_headers);
1793 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1794 					   misc_parameters);
1795 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1796 					   misc_parameters);
1797 
1798 	switch (ib_spec->type) {
1799 	case IB_FLOW_SPEC_ETH:
1800 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1801 			return -ENOTSUPP;
1802 
1803 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1804 					     dmac_47_16),
1805 				ib_spec->eth.mask.dst_mac);
1806 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1807 					     dmac_47_16),
1808 				ib_spec->eth.val.dst_mac);
1809 
1810 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1811 					     smac_47_16),
1812 				ib_spec->eth.mask.src_mac);
1813 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1814 					     smac_47_16),
1815 				ib_spec->eth.val.src_mac);
1816 
1817 		if (ib_spec->eth.mask.vlan_tag) {
1818 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1819 				 cvlan_tag, 1);
1820 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1821 				 cvlan_tag, 1);
1822 
1823 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1824 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1825 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1826 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1827 
1828 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1829 				 first_cfi,
1830 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1831 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1832 				 first_cfi,
1833 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1834 
1835 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1836 				 first_prio,
1837 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1838 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1839 				 first_prio,
1840 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1841 		}
1842 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1843 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1844 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1845 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
1846 		break;
1847 	case IB_FLOW_SPEC_IPV4:
1848 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1849 			return -ENOTSUPP;
1850 
1851 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1852 			 ethertype, 0xffff);
1853 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1854 			 ethertype, ETH_P_IP);
1855 
1856 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1857 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1858 		       &ib_spec->ipv4.mask.src_ip,
1859 		       sizeof(ib_spec->ipv4.mask.src_ip));
1860 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1861 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1862 		       &ib_spec->ipv4.val.src_ip,
1863 		       sizeof(ib_spec->ipv4.val.src_ip));
1864 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1865 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1866 		       &ib_spec->ipv4.mask.dst_ip,
1867 		       sizeof(ib_spec->ipv4.mask.dst_ip));
1868 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1869 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1870 		       &ib_spec->ipv4.val.dst_ip,
1871 		       sizeof(ib_spec->ipv4.val.dst_ip));
1872 
1873 		set_tos(outer_headers_c, outer_headers_v,
1874 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
1875 
1876 		set_proto(outer_headers_c, outer_headers_v,
1877 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
1878 		break;
1879 	case IB_FLOW_SPEC_IPV6:
1880 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1881 			return -ENOTSUPP;
1882 
1883 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1884 			 ethertype, 0xffff);
1885 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1886 			 ethertype, IPPROTO_IPV6);
1887 
1888 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1889 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1890 		       &ib_spec->ipv6.mask.src_ip,
1891 		       sizeof(ib_spec->ipv6.mask.src_ip));
1892 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1893 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1894 		       &ib_spec->ipv6.val.src_ip,
1895 		       sizeof(ib_spec->ipv6.val.src_ip));
1896 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1897 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1898 		       &ib_spec->ipv6.mask.dst_ip,
1899 		       sizeof(ib_spec->ipv6.mask.dst_ip));
1900 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1901 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1902 		       &ib_spec->ipv6.val.dst_ip,
1903 		       sizeof(ib_spec->ipv6.val.dst_ip));
1904 
1905 		set_tos(outer_headers_c, outer_headers_v,
1906 			ib_spec->ipv6.mask.traffic_class,
1907 			ib_spec->ipv6.val.traffic_class);
1908 
1909 		set_proto(outer_headers_c, outer_headers_v,
1910 			  ib_spec->ipv6.mask.next_hdr,
1911 			  ib_spec->ipv6.val.next_hdr);
1912 
1913 		MLX5_SET(fte_match_set_misc, misc_params_c,
1914 			 outer_ipv6_flow_label,
1915 			 ntohl(ib_spec->ipv6.mask.flow_label));
1916 		MLX5_SET(fte_match_set_misc, misc_params_v,
1917 			 outer_ipv6_flow_label,
1918 			 ntohl(ib_spec->ipv6.val.flow_label));
1919 		break;
1920 	case IB_FLOW_SPEC_TCP:
1921 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1922 					 LAST_TCP_UDP_FIELD))
1923 			return -ENOTSUPP;
1924 
1925 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1926 			 0xff);
1927 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1928 			 IPPROTO_TCP);
1929 
1930 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport,
1931 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1932 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport,
1933 			 ntohs(ib_spec->tcp_udp.val.src_port));
1934 
1935 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport,
1936 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1937 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport,
1938 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1939 		break;
1940 	case IB_FLOW_SPEC_UDP:
1941 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1942 					 LAST_TCP_UDP_FIELD))
1943 			return -ENOTSUPP;
1944 
1945 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1946 			 0xff);
1947 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1948 			 IPPROTO_UDP);
1949 
1950 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport,
1951 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1952 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport,
1953 			 ntohs(ib_spec->tcp_udp.val.src_port));
1954 
1955 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport,
1956 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1957 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport,
1958 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1959 		break;
1960 	default:
1961 		return -EINVAL;
1962 	}
1963 
1964 	return 0;
1965 }
1966 
1967 /* If a flow could catch both multicast and unicast packets,
1968  * it won't fall into the multicast flow steering table and this rule
1969  * could steal other multicast packets.
1970  */
1971 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1972 {
1973 	struct ib_flow_spec_eth *eth_spec;
1974 
1975 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1976 	    ib_attr->size < sizeof(struct ib_flow_attr) +
1977 	    sizeof(struct ib_flow_spec_eth) ||
1978 	    ib_attr->num_of_specs < 1)
1979 		return false;
1980 
1981 	eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1982 	if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1983 	    eth_spec->size != sizeof(*eth_spec))
1984 		return false;
1985 
1986 	return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1987 	       is_multicast_ether_addr(eth_spec->val.dst_mac);
1988 }
1989 
1990 static bool is_valid_attr(const struct ib_flow_attr *flow_attr)
1991 {
1992 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1993 	bool has_ipv4_spec = false;
1994 	bool eth_type_ipv4 = true;
1995 	unsigned int spec_index;
1996 
1997 	/* Validate that ethertype is correct */
1998 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1999 		if (ib_spec->type == IB_FLOW_SPEC_ETH &&
2000 		    ib_spec->eth.mask.ether_type) {
2001 			if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
2002 			      ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
2003 				eth_type_ipv4 = false;
2004 		} else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
2005 			has_ipv4_spec = true;
2006 		}
2007 		ib_spec = (void *)ib_spec + ib_spec->size;
2008 	}
2009 	return !has_ipv4_spec || eth_type_ipv4;
2010 }
2011 
2012 static void put_flow_table(struct mlx5_ib_dev *dev,
2013 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
2014 {
2015 	prio->refcount -= !!ft_added;
2016 	if (!prio->refcount) {
2017 		mlx5_destroy_flow_table(prio->flow_table);
2018 		prio->flow_table = NULL;
2019 	}
2020 }
2021 
2022 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
2023 {
2024 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
2025 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
2026 							  struct mlx5_ib_flow_handler,
2027 							  ibflow);
2028 	struct mlx5_ib_flow_handler *iter, *tmp;
2029 
2030 	mutex_lock(&dev->flow_db.lock);
2031 
2032 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
2033 		mlx5_del_flow_rule(iter->rule);
2034 		put_flow_table(dev, iter->prio, true);
2035 		list_del(&iter->list);
2036 		kfree(iter);
2037 	}
2038 
2039 	mlx5_del_flow_rule(handler->rule);
2040 	put_flow_table(dev, handler->prio, true);
2041 	mutex_unlock(&dev->flow_db.lock);
2042 
2043 	kfree(handler);
2044 
2045 	return 0;
2046 }
2047 
2048 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
2049 {
2050 	priority *= 2;
2051 	if (!dont_trap)
2052 		priority++;
2053 	return priority;
2054 }
2055 
2056 enum flow_table_type {
2057 	MLX5_IB_FT_RX,
2058 	MLX5_IB_FT_TX
2059 };
2060 
2061 #define MLX5_FS_MAX_TYPES	 10
2062 #define MLX5_FS_MAX_ENTRIES	 32000UL
2063 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
2064 						struct ib_flow_attr *flow_attr,
2065 						enum flow_table_type ft_type)
2066 {
2067 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
2068 	struct mlx5_flow_namespace *ns = NULL;
2069 	struct mlx5_ib_flow_prio *prio;
2070 	struct mlx5_flow_table *ft;
2071 	int num_entries;
2072 	int num_groups;
2073 	int priority;
2074 	int err = 0;
2075 
2076 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2077 		if (flow_is_multicast_only(flow_attr) &&
2078 		    !dont_trap)
2079 			priority = MLX5_IB_FLOW_MCAST_PRIO;
2080 		else
2081 			priority = ib_prio_to_core_prio(flow_attr->priority,
2082 							dont_trap);
2083 		ns = mlx5_get_flow_namespace(dev->mdev,
2084 					     MLX5_FLOW_NAMESPACE_BYPASS);
2085 		num_entries = MLX5_FS_MAX_ENTRIES;
2086 		num_groups = MLX5_FS_MAX_TYPES;
2087 		prio = &dev->flow_db.prios[priority];
2088 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2089 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2090 		ns = mlx5_get_flow_namespace(dev->mdev,
2091 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
2092 		build_leftovers_ft_param("bypass", &priority,
2093 					 &num_entries,
2094 					 &num_groups);
2095 		prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
2096 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2097 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
2098 					allow_sniffer_and_nic_rx_shared_tir))
2099 			return ERR_PTR(-ENOTSUPP);
2100 
2101 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
2102 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2103 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2104 
2105 		prio = &dev->flow_db.sniffer[ft_type];
2106 		priority = 0;
2107 		num_entries = 1;
2108 		num_groups = 1;
2109 	}
2110 
2111 	if (!ns)
2112 		return ERR_PTR(-ENOTSUPP);
2113 
2114 	ft = prio->flow_table;
2115 	if (!ft) {
2116 		ft = mlx5_create_auto_grouped_flow_table(ns, priority, "bypass",
2117 							 num_entries,
2118 							 num_groups);
2119 
2120 		if (!IS_ERR(ft)) {
2121 			prio->refcount = 0;
2122 			prio->flow_table = ft;
2123 		} else {
2124 			err = PTR_ERR(ft);
2125 		}
2126 	}
2127 
2128 	return err ? ERR_PTR(err) : prio;
2129 }
2130 
2131 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
2132 						     struct mlx5_ib_flow_prio *ft_prio,
2133 						     const struct ib_flow_attr *flow_attr,
2134 						     struct mlx5_flow_destination *dst)
2135 {
2136 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
2137 	struct mlx5_ib_flow_handler *handler;
2138 	struct mlx5_flow_spec *spec;
2139 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
2140 	unsigned int spec_index;
2141 	u32 action;
2142 	int err = 0;
2143 
2144 	if (!is_valid_attr(flow_attr))
2145 		return ERR_PTR(-EINVAL);
2146 
2147 	spec = mlx5_vzalloc(sizeof(*spec));
2148 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
2149 	if (!handler || !spec) {
2150 		err = -ENOMEM;
2151 		goto free;
2152 	}
2153 
2154 	INIT_LIST_HEAD(&handler->list);
2155 
2156 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2157 		err = parse_flow_attr(spec->match_criteria,
2158 				      spec->match_value, ib_flow);
2159 		if (err < 0)
2160 			goto free;
2161 
2162 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
2163 	}
2164 
2165 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
2166 	action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
2167 		MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
2168 	handler->rule = mlx5_add_flow_rule(ft, spec->match_criteria_enable,
2169 					   spec->match_criteria,
2170 					   spec->match_value,
2171 					   action,
2172 					   MLX5_FS_DEFAULT_FLOW_TAG,
2173 					   dst);
2174 
2175 	if (IS_ERR(handler->rule)) {
2176 		err = PTR_ERR(handler->rule);
2177 		goto free;
2178 	}
2179 
2180 	ft_prio->refcount++;
2181 	handler->prio = ft_prio;
2182 
2183 	ft_prio->flow_table = ft;
2184 free:
2185 	if (err)
2186 		kfree(handler);
2187 	kvfree(spec);
2188 	return err ? ERR_PTR(err) : handler;
2189 }
2190 
2191 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
2192 							  struct mlx5_ib_flow_prio *ft_prio,
2193 							  struct ib_flow_attr *flow_attr,
2194 							  struct mlx5_flow_destination *dst)
2195 {
2196 	struct mlx5_ib_flow_handler *handler_dst = NULL;
2197 	struct mlx5_ib_flow_handler *handler = NULL;
2198 
2199 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
2200 	if (!IS_ERR(handler)) {
2201 		handler_dst = create_flow_rule(dev, ft_prio,
2202 					       flow_attr, dst);
2203 		if (IS_ERR(handler_dst)) {
2204 			mlx5_del_flow_rule(handler->rule);
2205 			ft_prio->refcount--;
2206 			kfree(handler);
2207 			handler = handler_dst;
2208 		} else {
2209 			list_add(&handler_dst->list, &handler->list);
2210 		}
2211 	}
2212 
2213 	return handler;
2214 }
2215 enum {
2216 	LEFTOVERS_MC,
2217 	LEFTOVERS_UC,
2218 };
2219 
2220 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
2221 							  struct mlx5_ib_flow_prio *ft_prio,
2222 							  struct ib_flow_attr *flow_attr,
2223 							  struct mlx5_flow_destination *dst)
2224 {
2225 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
2226 	struct mlx5_ib_flow_handler *handler = NULL;
2227 
2228 	static struct {
2229 		struct ib_flow_attr	flow_attr;
2230 		struct ib_flow_spec_eth eth_flow;
2231 	} leftovers_specs[] = {
2232 		[LEFTOVERS_MC] = {
2233 			.flow_attr = {
2234 				.num_of_specs = 1,
2235 				.size = sizeof(leftovers_specs[0])
2236 			},
2237 			.eth_flow = {
2238 				.type = IB_FLOW_SPEC_ETH,
2239 				.size = sizeof(struct ib_flow_spec_eth),
2240 				.mask = {.dst_mac = {0x1} },
2241 				.val =  {.dst_mac = {0x1} }
2242 			}
2243 		},
2244 		[LEFTOVERS_UC] = {
2245 			.flow_attr = {
2246 				.num_of_specs = 1,
2247 				.size = sizeof(leftovers_specs[0])
2248 			},
2249 			.eth_flow = {
2250 				.type = IB_FLOW_SPEC_ETH,
2251 				.size = sizeof(struct ib_flow_spec_eth),
2252 				.mask = {.dst_mac = {0x1} },
2253 				.val = {.dst_mac = {} }
2254 			}
2255 		}
2256 	};
2257 
2258 	handler = create_flow_rule(dev, ft_prio,
2259 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
2260 				   dst);
2261 	if (!IS_ERR(handler) &&
2262 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2263 		handler_ucast = create_flow_rule(dev, ft_prio,
2264 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
2265 						 dst);
2266 		if (IS_ERR(handler_ucast)) {
2267 			mlx5_del_flow_rule(handler->rule);
2268 			ft_prio->refcount--;
2269 			kfree(handler);
2270 			handler = handler_ucast;
2271 		} else {
2272 			list_add(&handler_ucast->list, &handler->list);
2273 		}
2274 	}
2275 
2276 	return handler;
2277 }
2278 
2279 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2280 							struct mlx5_ib_flow_prio *ft_rx,
2281 							struct mlx5_ib_flow_prio *ft_tx,
2282 							struct mlx5_flow_destination *dst)
2283 {
2284 	struct mlx5_ib_flow_handler *handler_rx;
2285 	struct mlx5_ib_flow_handler *handler_tx;
2286 	int err;
2287 	static const struct ib_flow_attr flow_attr  = {
2288 		.num_of_specs = 0,
2289 		.size = sizeof(flow_attr)
2290 	};
2291 
2292 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2293 	if (IS_ERR(handler_rx)) {
2294 		err = PTR_ERR(handler_rx);
2295 		goto err;
2296 	}
2297 
2298 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2299 	if (IS_ERR(handler_tx)) {
2300 		err = PTR_ERR(handler_tx);
2301 		goto err_tx;
2302 	}
2303 
2304 	list_add(&handler_tx->list, &handler_rx->list);
2305 
2306 	return handler_rx;
2307 
2308 err_tx:
2309 	mlx5_del_flow_rule(handler_rx->rule);
2310 	ft_rx->refcount--;
2311 	kfree(handler_rx);
2312 err:
2313 	return ERR_PTR(err);
2314 }
2315 
2316 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2317 					   struct ib_flow_attr *flow_attr,
2318 					   int domain,
2319 					   struct ib_udata *udata)
2320 {
2321 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2322 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2323 	struct mlx5_ib_flow_handler *handler = NULL;
2324 	struct mlx5_flow_destination *dst = NULL;
2325 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2326 	struct mlx5_ib_flow_prio *ft_prio;
2327 	int err;
2328 
2329 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2330 		return ERR_PTR(-ENOSPC);
2331 
2332 	if (domain != IB_FLOW_DOMAIN_USER ||
2333 	    udata != NULL ||
2334 	    flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2335 	    (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2336 		return ERR_PTR(-EINVAL);
2337 
2338 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2339 	if (!dst)
2340 		return ERR_PTR(-ENOMEM);
2341 
2342 	mutex_lock(&dev->flow_db.lock);
2343 
2344 	ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2345 	if (IS_ERR(ft_prio)) {
2346 		err = PTR_ERR(ft_prio);
2347 		goto unlock;
2348 	}
2349 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2350 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2351 		if (IS_ERR(ft_prio_tx)) {
2352 			err = PTR_ERR(ft_prio_tx);
2353 			ft_prio_tx = NULL;
2354 			goto destroy_ft;
2355 		}
2356 	}
2357 
2358 	dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2359 	if (mqp->flags & MLX5_IB_QP_RSS)
2360 		dst->tir_num = mqp->rss_qp.tirn;
2361 	else
2362 		dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2363 
2364 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2365 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
2366 			handler = create_dont_trap_rule(dev, ft_prio,
2367 							flow_attr, dst);
2368 		} else {
2369 			handler = create_flow_rule(dev, ft_prio, flow_attr,
2370 						   dst);
2371 		}
2372 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2373 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2374 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2375 						dst);
2376 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2377 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2378 	} else {
2379 		err = -EINVAL;
2380 		goto destroy_ft;
2381 	}
2382 
2383 	if (IS_ERR(handler)) {
2384 		err = PTR_ERR(handler);
2385 		handler = NULL;
2386 		goto destroy_ft;
2387 	}
2388 
2389 	mutex_unlock(&dev->flow_db.lock);
2390 	kfree(dst);
2391 
2392 	return &handler->ibflow;
2393 
2394 destroy_ft:
2395 	put_flow_table(dev, ft_prio, false);
2396 	if (ft_prio_tx)
2397 		put_flow_table(dev, ft_prio_tx, false);
2398 unlock:
2399 	mutex_unlock(&dev->flow_db.lock);
2400 	kfree(dst);
2401 	kfree(handler);
2402 	return ERR_PTR(err);
2403 }
2404 
2405 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2406 {
2407 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2408 	int err;
2409 
2410 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2411 	if (err)
2412 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2413 			     ibqp->qp_num, gid->raw);
2414 
2415 	return err;
2416 }
2417 
2418 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2419 {
2420 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2421 	int err;
2422 
2423 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2424 	if (err)
2425 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2426 			     ibqp->qp_num, gid->raw);
2427 
2428 	return err;
2429 }
2430 
2431 static int init_node_data(struct mlx5_ib_dev *dev)
2432 {
2433 	int err;
2434 
2435 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2436 	if (err)
2437 		return err;
2438 
2439 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2440 }
2441 
2442 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2443 			     char *buf)
2444 {
2445 	struct mlx5_ib_dev *dev =
2446 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2447 
2448 	return sprintf(buf, "%lld\n", (long long)dev->mdev->priv.fw_pages);
2449 }
2450 
2451 static ssize_t show_reg_pages(struct device *device,
2452 			      struct device_attribute *attr, char *buf)
2453 {
2454 	struct mlx5_ib_dev *dev =
2455 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2456 
2457 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2458 }
2459 
2460 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2461 			char *buf)
2462 {
2463 	struct mlx5_ib_dev *dev =
2464 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2465 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2466 }
2467 
2468 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2469 			char *buf)
2470 {
2471 	struct mlx5_ib_dev *dev =
2472 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2473 	return sprintf(buf, "%x\n", dev->mdev->pdev->revision);
2474 }
2475 
2476 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2477 			  char *buf)
2478 {
2479 	struct mlx5_ib_dev *dev =
2480 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2481 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2482 		       dev->mdev->board_id);
2483 }
2484 
2485 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2486 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2487 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2488 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2489 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2490 
2491 static struct device_attribute *mlx5_class_attributes[] = {
2492 	&dev_attr_hw_rev,
2493 	&dev_attr_hca_type,
2494 	&dev_attr_board_id,
2495 	&dev_attr_fw_pages,
2496 	&dev_attr_reg_pages,
2497 };
2498 
2499 static void pkey_change_handler(struct work_struct *work)
2500 {
2501 	struct mlx5_ib_port_resources *ports =
2502 		container_of(work, struct mlx5_ib_port_resources,
2503 			     pkey_change_work);
2504 
2505 	mutex_lock(&ports->devr->mutex);
2506 	mlx5_ib_gsi_pkey_change(ports->gsi);
2507 	mutex_unlock(&ports->devr->mutex);
2508 }
2509 
2510 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2511 {
2512 	struct mlx5_ib_qp *mqp;
2513 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2514 	struct mlx5_core_cq *mcq;
2515 	struct list_head cq_armed_list;
2516 	unsigned long flags_qp;
2517 	unsigned long flags_cq;
2518 	unsigned long flags;
2519 
2520 	INIT_LIST_HEAD(&cq_armed_list);
2521 
2522 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2523 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2524 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2525 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2526 		if (mqp->sq.tail != mqp->sq.head) {
2527 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2528 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2529 			if (send_mcq->mcq.comp &&
2530 			    mqp->ibqp.send_cq->comp_handler) {
2531 				if (!send_mcq->mcq.reset_notify_added) {
2532 					send_mcq->mcq.reset_notify_added = 1;
2533 					list_add_tail(&send_mcq->mcq.reset_notify,
2534 						      &cq_armed_list);
2535 				}
2536 			}
2537 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2538 		}
2539 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2540 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2541 		/* no handling is needed for SRQ */
2542 		if (!mqp->ibqp.srq) {
2543 			if (mqp->rq.tail != mqp->rq.head) {
2544 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2545 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2546 				if (recv_mcq->mcq.comp &&
2547 				    mqp->ibqp.recv_cq->comp_handler) {
2548 					if (!recv_mcq->mcq.reset_notify_added) {
2549 						recv_mcq->mcq.reset_notify_added = 1;
2550 						list_add_tail(&recv_mcq->mcq.reset_notify,
2551 							      &cq_armed_list);
2552 					}
2553 				}
2554 				spin_unlock_irqrestore(&recv_mcq->lock,
2555 						       flags_cq);
2556 			}
2557 		}
2558 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2559 	}
2560 	/*At that point all inflight post send were put to be executed as of we
2561 	 * lock/unlock above locks Now need to arm all involved CQs.
2562 	 */
2563 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2564 		mcq->comp(mcq, NULL);
2565 	}
2566 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2567 }
2568 
2569 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2570 			  enum mlx5_dev_event event, unsigned long param)
2571 {
2572 	struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2573 	struct ib_event ibev;
2574 	bool fatal = false;
2575 	u8 port = (u8)param;
2576 
2577 	switch (event) {
2578 	case MLX5_DEV_EVENT_SYS_ERROR:
2579 		ibev.event = IB_EVENT_DEVICE_FATAL;
2580 		mlx5_ib_handle_internal_error(ibdev);
2581 		fatal = true;
2582 		break;
2583 
2584 	case MLX5_DEV_EVENT_PORT_UP:
2585 	case MLX5_DEV_EVENT_PORT_DOWN:
2586 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
2587 		/* In RoCE, port up/down events are handled in
2588 		 * mlx5_netdev_event().
2589 		 */
2590 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2591 			IB_LINK_LAYER_ETHERNET)
2592 			return;
2593 
2594 		ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2595 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2596 		break;
2597 
2598 	case MLX5_DEV_EVENT_LID_CHANGE:
2599 		ibev.event = IB_EVENT_LID_CHANGE;
2600 		break;
2601 
2602 	case MLX5_DEV_EVENT_PKEY_CHANGE:
2603 		ibev.event = IB_EVENT_PKEY_CHANGE;
2604 
2605 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2606 		break;
2607 
2608 	case MLX5_DEV_EVENT_GUID_CHANGE:
2609 		ibev.event = IB_EVENT_GID_CHANGE;
2610 		break;
2611 
2612 	case MLX5_DEV_EVENT_CLIENT_REREG:
2613 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
2614 		break;
2615 
2616 	default:
2617 		/* unsupported event */
2618 		return;
2619 	}
2620 
2621 	ibev.device	      = &ibdev->ib_dev;
2622 	ibev.element.port_num = port;
2623 
2624 	if (!rdma_is_port_valid(&ibdev->ib_dev, port)) {
2625 		mlx5_ib_warn(ibdev, "warning: event(%d) on port %d\n", event, port);
2626 		return;
2627 	}
2628 
2629 	if (ibdev->ib_active)
2630 		ib_dispatch_event(&ibev);
2631 
2632 	if (fatal)
2633 		ibdev->ib_active = false;
2634 }
2635 
2636 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2637 {
2638 	int port;
2639 
2640 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2641 		mlx5_query_ext_port_caps(dev, port);
2642 }
2643 
2644 static int get_port_caps(struct mlx5_ib_dev *dev)
2645 {
2646 	struct ib_device_attr *dprops = NULL;
2647 	struct ib_port_attr *pprops = NULL;
2648 	int err = -ENOMEM;
2649 	int port;
2650 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2651 
2652 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2653 	if (!pprops)
2654 		goto out;
2655 
2656 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2657 	if (!dprops)
2658 		goto out;
2659 
2660 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2661 	if (err) {
2662 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2663 		goto out;
2664 	}
2665 
2666 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2667 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2668 		if (err) {
2669 			mlx5_ib_warn(dev, "query_port %d failed %d\n",
2670 				     port, err);
2671 			break;
2672 		}
2673 		dev->mdev->port_caps[port - 1].pkey_table_len =
2674 						dprops->max_pkeys;
2675 		dev->mdev->port_caps[port - 1].gid_table_len =
2676 						pprops->gid_tbl_len;
2677 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2678 			    dprops->max_pkeys, pprops->gid_tbl_len);
2679 	}
2680 
2681 out:
2682 	kfree(pprops);
2683 	kfree(dprops);
2684 
2685 	return err;
2686 }
2687 
2688 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2689 {
2690 	int err;
2691 
2692 	err = mlx5_mr_cache_cleanup(dev);
2693 	if (err)
2694 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2695 
2696 	if (dev->umrc.qp)
2697 		mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
2698 	if (dev->umrc.cq)
2699 		ib_free_cq(dev->umrc.cq);
2700 	if (dev->umrc.pd)
2701 		ib_dealloc_pd(dev->umrc.pd);
2702 }
2703 
2704 enum {
2705 	MAX_UMR_WR = 128,
2706 };
2707 
2708 static int create_umr_res(struct mlx5_ib_dev *dev)
2709 {
2710 	struct ib_qp_init_attr *init_attr = NULL;
2711 	struct ib_qp_attr *attr = NULL;
2712 	struct ib_pd *pd;
2713 	struct ib_cq *cq;
2714 	struct ib_qp *qp;
2715 	int ret;
2716 
2717 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2718 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2719 	if (!attr || !init_attr) {
2720 		ret = -ENOMEM;
2721 		goto error_0;
2722 	}
2723 
2724 	pd = ib_alloc_pd(&dev->ib_dev, 0);
2725 	if (IS_ERR(pd)) {
2726 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2727 		ret = PTR_ERR(pd);
2728 		goto error_0;
2729 	}
2730 
2731 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2732 	if (IS_ERR(cq)) {
2733 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2734 		ret = PTR_ERR(cq);
2735 		goto error_2;
2736 	}
2737 
2738 	init_attr->send_cq = cq;
2739 	init_attr->recv_cq = cq;
2740 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2741 	init_attr->cap.max_send_wr = MAX_UMR_WR;
2742 	init_attr->cap.max_send_sge = 1;
2743 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2744 	init_attr->port_num = 1;
2745 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2746 	if (IS_ERR(qp)) {
2747 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2748 		ret = PTR_ERR(qp);
2749 		goto error_3;
2750 	}
2751 	qp->device     = &dev->ib_dev;
2752 	qp->real_qp    = qp;
2753 	qp->uobject    = NULL;
2754 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2755 
2756 	attr->qp_state = IB_QPS_INIT;
2757 	attr->port_num = 1;
2758 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2759 				IB_QP_PORT, NULL);
2760 	if (ret) {
2761 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2762 		goto error_4;
2763 	}
2764 
2765 	memset(attr, 0, sizeof(*attr));
2766 	attr->qp_state = IB_QPS_RTR;
2767 	attr->path_mtu = IB_MTU_256;
2768 
2769 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2770 	if (ret) {
2771 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2772 		goto error_4;
2773 	}
2774 
2775 	memset(attr, 0, sizeof(*attr));
2776 	attr->qp_state = IB_QPS_RTS;
2777 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2778 	if (ret) {
2779 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2780 		goto error_4;
2781 	}
2782 
2783 	dev->umrc.qp = qp;
2784 	dev->umrc.cq = cq;
2785 	dev->umrc.pd = pd;
2786 
2787 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
2788 	ret = mlx5_mr_cache_init(dev);
2789 	if (ret) {
2790 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2791 		goto error_4;
2792 	}
2793 
2794 	kfree(attr);
2795 	kfree(init_attr);
2796 
2797 	return 0;
2798 
2799 error_4:
2800 	mlx5_ib_destroy_qp(qp, NULL);
2801 	dev->umrc.qp = NULL;
2802 
2803 error_3:
2804 	ib_free_cq(cq);
2805 	dev->umrc.cq = NULL;
2806 
2807 error_2:
2808 	ib_dealloc_pd(pd);
2809 	dev->umrc.pd = NULL;
2810 
2811 error_0:
2812 	kfree(attr);
2813 	kfree(init_attr);
2814 	return ret;
2815 }
2816 
2817 static int create_dev_resources(struct mlx5_ib_resources *devr)
2818 {
2819 	struct ib_srq_init_attr attr;
2820 	struct mlx5_ib_dev *dev;
2821 	struct ib_device *ibdev;
2822 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2823 	int port;
2824 	int ret = 0;
2825 
2826 	dev = container_of(devr, struct mlx5_ib_dev, devr);
2827 	ibdev = &dev->ib_dev;
2828 
2829 	mutex_init(&devr->mutex);
2830 
2831 	devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
2832 	if (!devr->p0)
2833 		return -ENOMEM;
2834 
2835 	devr->p0->device  = ibdev;
2836 	devr->p0->uobject = NULL;
2837 	atomic_set(&devr->p0->usecnt, 0);
2838 
2839 	ret = mlx5_ib_alloc_pd(devr->p0, NULL);
2840 	if (ret)
2841 		goto error0;
2842 
2843 	devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
2844 	if (!devr->c0) {
2845 		ret = -ENOMEM;
2846 		goto error1;
2847 	}
2848 
2849 	devr->c0->device = &dev->ib_dev;
2850 	atomic_set(&devr->c0->usecnt, 0);
2851 
2852 	ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
2853 	if (ret)
2854 		goto err_create_cq;
2855 
2856 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL);
2857 	if (IS_ERR(devr->x0)) {
2858 		ret = PTR_ERR(devr->x0);
2859 		goto error2;
2860 	}
2861 	devr->x0->device = &dev->ib_dev;
2862 	devr->x0->inode = NULL;
2863 	atomic_set(&devr->x0->usecnt, 0);
2864 	mutex_init(&devr->x0->tgt_qp_mutex);
2865 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2866 
2867 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL);
2868 	if (IS_ERR(devr->x1)) {
2869 		ret = PTR_ERR(devr->x1);
2870 		goto error3;
2871 	}
2872 	devr->x1->device = &dev->ib_dev;
2873 	devr->x1->inode = NULL;
2874 	atomic_set(&devr->x1->usecnt, 0);
2875 	mutex_init(&devr->x1->tgt_qp_mutex);
2876 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2877 
2878 	memset(&attr, 0, sizeof(attr));
2879 	attr.attr.max_sge = 1;
2880 	attr.attr.max_wr = 1;
2881 	attr.srq_type = IB_SRQT_XRC;
2882 	attr.ext.cq = devr->c0;
2883 	attr.ext.xrc.xrcd = devr->x0;
2884 
2885 	devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
2886 	if (!devr->s0) {
2887 		ret = -ENOMEM;
2888 		goto error4;
2889 	}
2890 
2891 	devr->s0->device	= &dev->ib_dev;
2892 	devr->s0->pd		= devr->p0;
2893 	devr->s0->srq_type      = IB_SRQT_XRC;
2894 	devr->s0->ext.xrc.xrcd	= devr->x0;
2895 	devr->s0->ext.cq	= devr->c0;
2896 	ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
2897 	if (ret)
2898 		goto err_create;
2899 
2900 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2901 	atomic_inc(&devr->s0->ext.cq->usecnt);
2902 	atomic_inc(&devr->p0->usecnt);
2903 	atomic_set(&devr->s0->usecnt, 0);
2904 
2905 	memset(&attr, 0, sizeof(attr));
2906 	attr.attr.max_sge = 1;
2907 	attr.attr.max_wr = 1;
2908 	attr.srq_type = IB_SRQT_BASIC;
2909 	devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
2910 	if (!devr->s1) {
2911 		ret = -ENOMEM;
2912 		goto error5;
2913 	}
2914 
2915 	devr->s1->device	= &dev->ib_dev;
2916 	devr->s1->pd		= devr->p0;
2917 	devr->s1->srq_type      = IB_SRQT_BASIC;
2918 	devr->s1->ext.cq	= devr->c0;
2919 
2920 	ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
2921 	if (ret)
2922 		goto error6;
2923 
2924 	atomic_inc(&devr->p0->usecnt);
2925 	atomic_set(&devr->s1->usecnt, 0);
2926 
2927 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2928 		INIT_WORK(&devr->ports[port].pkey_change_work,
2929 			  pkey_change_handler);
2930 		devr->ports[port].devr = devr;
2931 	}
2932 
2933 	return 0;
2934 
2935 error6:
2936 	kfree(devr->s1);
2937 error5:
2938 	mlx5_ib_destroy_srq(devr->s0, NULL);
2939 err_create:
2940 	kfree(devr->s0);
2941 error4:
2942 	mlx5_ib_dealloc_xrcd(devr->x1, NULL);
2943 error3:
2944 	mlx5_ib_dealloc_xrcd(devr->x0, NULL);
2945 error2:
2946 	mlx5_ib_destroy_cq(devr->c0, NULL);
2947 err_create_cq:
2948 	kfree(devr->c0);
2949 error1:
2950 	mlx5_ib_dealloc_pd(devr->p0, NULL);
2951 error0:
2952 	kfree(devr->p0);
2953 	return ret;
2954 }
2955 
2956 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2957 {
2958 	int port;
2959 
2960 	mlx5_ib_destroy_srq(devr->s1, NULL);
2961 	kfree(devr->s1);
2962 	mlx5_ib_destroy_srq(devr->s0, NULL);
2963 	kfree(devr->s0);
2964 	mlx5_ib_dealloc_xrcd(devr->x0, NULL);
2965 	mlx5_ib_dealloc_xrcd(devr->x1, NULL);
2966 	mlx5_ib_destroy_cq(devr->c0, NULL);
2967 	kfree(devr->c0);
2968 	mlx5_ib_dealloc_pd(devr->p0, NULL);
2969 	kfree(devr->p0);
2970 
2971 	/* Make sure no change P_Key work items are still executing */
2972 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2973 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2974 }
2975 
2976 static u32 get_core_cap_flags(struct ib_device *ibdev)
2977 {
2978 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2979 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2980 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2981 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2982 	u32 ret = 0;
2983 
2984 	if (ll == IB_LINK_LAYER_INFINIBAND)
2985 		return RDMA_CORE_PORT_IBA_IB;
2986 
2987 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2988 		return 0;
2989 
2990 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2991 		return 0;
2992 
2993 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2994 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2995 
2996 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2997 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2998 
2999 	return ret;
3000 }
3001 
3002 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
3003 			       struct ib_port_immutable *immutable)
3004 {
3005 	struct ib_port_attr attr;
3006 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3007 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
3008 	int err;
3009 
3010 	err = mlx5_ib_query_port(ibdev, port_num, &attr);
3011 	if (err)
3012 		return err;
3013 
3014 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3015 	immutable->gid_tbl_len = attr.gid_tbl_len;
3016 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
3017 	if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
3018 		immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3019 
3020 	return 0;
3021 }
3022 
3023 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
3024 			   size_t str_len)
3025 {
3026 	struct mlx5_ib_dev *dev =
3027 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3028 	snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
3029 		       fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
3030 }
3031 
3032 static int mlx5_roce_lag_init(struct mlx5_ib_dev *dev)
3033 {
3034 	return 0;
3035 }
3036 
3037 static void mlx5_roce_lag_cleanup(struct mlx5_ib_dev *dev)
3038 {
3039 }
3040 
3041 static void mlx5_remove_roce_notifier(struct mlx5_ib_dev *dev)
3042 {
3043 	if (dev->roce.nb.notifier_call) {
3044 		unregister_netdevice_notifier(&dev->roce.nb);
3045 		dev->roce.nb.notifier_call = NULL;
3046 	}
3047 }
3048 
3049 static int mlx5_enable_roce(struct mlx5_ib_dev *dev)
3050 {
3051 	VNET_ITERATOR_DECL(vnet_iter);
3052 	struct ifnet *idev;
3053 	int err;
3054 
3055 	/* Check if mlx5en net device already exists */
3056 	VNET_LIST_RLOCK();
3057 	VNET_FOREACH(vnet_iter) {
3058 		IFNET_RLOCK();
3059 		CURVNET_SET_QUIET(vnet_iter);
3060 		CK_STAILQ_FOREACH(idev, &V_ifnet, if_link) {
3061 			/* check if network interface belongs to mlx5en */
3062 			if (!mlx5_netdev_match(idev, dev->mdev, "mce"))
3063 				continue;
3064 			write_lock(&dev->roce.netdev_lock);
3065 			dev->roce.netdev = idev;
3066 			write_unlock(&dev->roce.netdev_lock);
3067 		}
3068 		CURVNET_RESTORE();
3069 		IFNET_RUNLOCK();
3070 	}
3071 	VNET_LIST_RUNLOCK();
3072 
3073 	dev->roce.nb.notifier_call = mlx5_netdev_event;
3074 	err = register_netdevice_notifier(&dev->roce.nb);
3075 	if (err) {
3076 		dev->roce.nb.notifier_call = NULL;
3077 		return err;
3078 	}
3079 
3080 	if (MLX5_CAP_GEN(dev->mdev, roce)) {
3081 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3082 		if (err)
3083 			goto err_unregister_netdevice_notifier;
3084 	}
3085 
3086 	err = mlx5_roce_lag_init(dev);
3087 	if (err)
3088 		goto err_disable_roce;
3089 
3090 	return 0;
3091 
3092 err_disable_roce:
3093 	if (MLX5_CAP_GEN(dev->mdev, roce))
3094 		mlx5_nic_vport_disable_roce(dev->mdev);
3095 
3096 err_unregister_netdevice_notifier:
3097 	mlx5_remove_roce_notifier(dev);
3098 	return err;
3099 }
3100 
3101 static void mlx5_disable_roce(struct mlx5_ib_dev *dev)
3102 {
3103 	mlx5_roce_lag_cleanup(dev);
3104 	if (MLX5_CAP_GEN(dev->mdev, roce))
3105 		mlx5_nic_vport_disable_roce(dev->mdev);
3106 }
3107 
3108 static void mlx5_ib_dealloc_q_port_counter(struct mlx5_ib_dev *dev, u8 port_num)
3109 {
3110 	mlx5_vport_dealloc_q_counter(dev->mdev,
3111 				     MLX5_INTERFACE_PROTOCOL_IB,
3112 				     dev->port[port_num].q_cnt_id);
3113 	dev->port[port_num].q_cnt_id = 0;
3114 }
3115 
3116 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
3117 {
3118 	unsigned int i;
3119 
3120 	for (i = 0; i < dev->num_ports; i++)
3121 		mlx5_ib_dealloc_q_port_counter(dev, i);
3122 }
3123 
3124 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
3125 {
3126 	int i;
3127 	int ret;
3128 
3129 	for (i = 0; i < dev->num_ports; i++) {
3130 		ret = mlx5_vport_alloc_q_counter(dev->mdev,
3131 						 MLX5_INTERFACE_PROTOCOL_IB,
3132 						 &dev->port[i].q_cnt_id);
3133 		if (ret) {
3134 			mlx5_ib_warn(dev,
3135 				     "couldn't allocate queue counter for port %d, err %d\n",
3136 				     i + 1, ret);
3137 			goto dealloc_counters;
3138 		}
3139 	}
3140 
3141 	return 0;
3142 
3143 dealloc_counters:
3144 	while (--i >= 0)
3145 		mlx5_ib_dealloc_q_port_counter(dev, i);
3146 
3147 	return ret;
3148 }
3149 
3150 static const char * const names[] = {
3151 	"rx_write_requests",
3152 	"rx_read_requests",
3153 	"rx_atomic_requests",
3154 	"out_of_buffer",
3155 	"out_of_sequence",
3156 	"duplicate_request",
3157 	"rnr_nak_retry_err",
3158 	"packet_seq_err",
3159 	"implied_nak_seq_err",
3160 	"local_ack_timeout_err",
3161 };
3162 
3163 static const size_t stats_offsets[] = {
3164 	MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests),
3165 	MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests),
3166 	MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests),
3167 	MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer),
3168 	MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence),
3169 	MLX5_BYTE_OFF(query_q_counter_out, duplicate_request),
3170 	MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err),
3171 	MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err),
3172 	MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err),
3173 	MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err),
3174 };
3175 
3176 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
3177 						    u8 port_num)
3178 {
3179 	BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets));
3180 
3181 	/* We support only per port stats */
3182 	if (port_num == 0)
3183 		return NULL;
3184 
3185 	return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names),
3186 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
3187 }
3188 
3189 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
3190 				struct rdma_hw_stats *stats,
3191 				u8 port, int index)
3192 {
3193 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3194 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
3195 	void *out;
3196 	__be32 val;
3197 	int ret;
3198 	int i;
3199 
3200 	if (!port || !stats)
3201 		return -ENOSYS;
3202 
3203 	out = mlx5_vzalloc(outlen);
3204 	if (!out)
3205 		return -ENOMEM;
3206 
3207 	ret = mlx5_vport_query_q_counter(dev->mdev,
3208 					dev->port[port - 1].q_cnt_id, 0,
3209 					out, outlen);
3210 	if (ret)
3211 		goto free;
3212 
3213 	for (i = 0; i < ARRAY_SIZE(names); i++) {
3214 		val = *(__be32 *)(out + stats_offsets[i]);
3215 		stats->value[i] = (u64)be32_to_cpu(val);
3216 	}
3217 free:
3218 	kvfree(out);
3219 	return ARRAY_SIZE(names);
3220 }
3221 
3222 static int mlx5_ib_stage_bfreg_init(struct mlx5_ib_dev *dev)
3223 {
3224 	int err;
3225 
3226 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
3227 	if (err)
3228 		return err;
3229 
3230 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
3231 	if (err) {
3232 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3233 		return err;
3234 	}
3235 
3236 	err = mlx5_alloc_bfreg(dev->mdev, &dev->wc_bfreg, true, false);
3237 	if (err) {
3238 		mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3239 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3240 	}
3241 
3242 	return err;
3243 }
3244 
3245 static void mlx5_ib_stage_bfreg_cleanup(struct mlx5_ib_dev *dev)
3246 {
3247 	mlx5_free_bfreg(dev->mdev, &dev->wc_bfreg);
3248 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3249 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3250 }
3251 
3252 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
3253 {
3254 	struct mlx5_ib_dev *dev;
3255 	enum rdma_link_layer ll;
3256 	int port_type_cap;
3257 	int err;
3258 	int i;
3259 
3260 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3261 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3262 
3263 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
3264 	if (!dev)
3265 		return NULL;
3266 
3267 	dev->mdev = mdev;
3268 
3269 	dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
3270 			    GFP_KERNEL);
3271 	if (!dev->port)
3272 		goto err_dealloc;
3273 
3274 	rwlock_init(&dev->roce.netdev_lock);
3275 	err = get_port_caps(dev);
3276 	if (err)
3277 		goto err_free_port;
3278 
3279 	if (mlx5_use_mad_ifc(dev))
3280 		get_ext_port_caps(dev);
3281 
3282 	MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
3283 
3284 	INIT_IB_DEVICE_OPS(&dev->ib_dev.ops, mlx5, MLX5);
3285 	snprintf(dev->ib_dev.name, IB_DEVICE_NAME_MAX, "mlx5_%d", device_get_unit(mdev->pdev->dev.bsddev));
3286 	dev->ib_dev.owner		= THIS_MODULE;
3287 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
3288 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
3289 	dev->num_ports		= MLX5_CAP_GEN(mdev, num_ports);
3290 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
3291 	dev->ib_dev.num_comp_vectors    =
3292 		dev->mdev->priv.eq_table.num_comp_vectors;
3293 	dev->ib_dev.dma_device	= &mdev->pdev->dev;
3294 
3295 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
3296 	dev->ib_dev.uverbs_cmd_mask	=
3297 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
3298 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
3299 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
3300 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
3301 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
3302 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
3303 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
3304 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
3305 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
3306 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
3307 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
3308 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
3309 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
3310 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
3311 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
3312 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
3313 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
3314 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
3315 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
3316 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
3317 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
3318 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
3319 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
3320 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
3321 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
3322 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
3323 	dev->ib_dev.uverbs_ex_cmd_mask =
3324 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
3325 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
3326 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
3327 
3328 	dev->ib_dev.query_device	= mlx5_ib_query_device;
3329 	dev->ib_dev.query_port		= mlx5_ib_query_port;
3330 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
3331 	if (ll == IB_LINK_LAYER_ETHERNET)
3332 		dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
3333 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
3334 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
3335 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
3336 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
3337 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
3338 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
3339 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
3340 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
3341 	dev->ib_dev.mmap		= mlx5_ib_mmap;
3342 	dev->ib_dev.mmap_free		= mlx5_ib_mmap_free;
3343 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
3344 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
3345 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
3346 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
3347 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
3348 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
3349 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
3350 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
3351 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
3352 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
3353 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
3354 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
3355 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
3356 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
3357 	dev->ib_dev.post_send		= mlx5_ib_post_send;
3358 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
3359 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
3360 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
3361 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
3362 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
3363 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
3364 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
3365 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
3366 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
3367 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
3368 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
3369 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
3370 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
3371 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
3372 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
3373 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
3374 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
3375 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
3376 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
3377 	if (mlx5_core_is_pf(mdev)) {
3378 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
3379 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
3380 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
3381 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
3382 	}
3383 
3384 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
3385 
3386 	mlx5_ib_internal_fill_odp_caps(dev);
3387 
3388 	if (MLX5_CAP_GEN(mdev, imaicl)) {
3389 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
3390 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
3391 		dev->ib_dev.uverbs_cmd_mask |=
3392 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
3393 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
3394 	}
3395 
3396 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) &&
3397 	    MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3398 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
3399 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
3400 	}
3401 
3402 	if (MLX5_CAP_GEN(mdev, xrc)) {
3403 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
3404 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
3405 		dev->ib_dev.uverbs_cmd_mask |=
3406 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
3407 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
3408 	}
3409 
3410 	if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
3411 	    IB_LINK_LAYER_ETHERNET) {
3412 		dev->ib_dev.create_flow	= mlx5_ib_create_flow;
3413 		dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
3414 		dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
3415 		dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
3416 		dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
3417 		dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
3418 		dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
3419 		dev->ib_dev.uverbs_ex_cmd_mask |=
3420 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
3421 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
3422 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
3423 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
3424 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
3425 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
3426 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
3427 	}
3428 	err = init_node_data(dev);
3429 	if (err)
3430 		goto err_free_port;
3431 
3432 	mutex_init(&dev->flow_db.lock);
3433 	mutex_init(&dev->cap_mask_mutex);
3434 	INIT_LIST_HEAD(&dev->qp_list);
3435 	spin_lock_init(&dev->reset_flow_resource_lock);
3436 
3437 	if (ll == IB_LINK_LAYER_ETHERNET) {
3438 		err = mlx5_enable_roce(dev);
3439 		if (err)
3440 			goto err_free_port;
3441 	}
3442 
3443 	err = create_dev_resources(&dev->devr);
3444 	if (err)
3445 		goto err_disable_roce;
3446 
3447 	err = mlx5_ib_odp_init_one(dev);
3448 	if (err)
3449 		goto err_rsrc;
3450 
3451 	err = mlx5_ib_alloc_q_counters(dev);
3452 	if (err)
3453 		goto err_odp;
3454 
3455 	err = mlx5_ib_stage_bfreg_init(dev);
3456 	if (err)
3457 		goto err_q_cnt;
3458 
3459 	err = ib_register_device(&dev->ib_dev, NULL);
3460 	if (err)
3461 		goto err_bfreg;
3462 
3463 	err = create_umr_res(dev);
3464 	if (err)
3465 		goto err_dev;
3466 
3467 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
3468 		err = device_create_file(&dev->ib_dev.dev,
3469 					 mlx5_class_attributes[i]);
3470 		if (err)
3471 			goto err_umrc;
3472 	}
3473 
3474 	err = mlx5_ib_init_congestion(dev);
3475 	if (err)
3476 		goto err_umrc;
3477 
3478 	dev->ib_active = true;
3479 
3480 	return dev;
3481 
3482 err_umrc:
3483 	destroy_umrc_res(dev);
3484 
3485 err_dev:
3486 	ib_unregister_device(&dev->ib_dev);
3487 
3488 err_bfreg:
3489 	mlx5_ib_stage_bfreg_cleanup(dev);
3490 
3491 err_q_cnt:
3492 	mlx5_ib_dealloc_q_counters(dev);
3493 
3494 err_odp:
3495 	mlx5_ib_odp_remove_one(dev);
3496 
3497 err_rsrc:
3498 	destroy_dev_resources(&dev->devr);
3499 
3500 err_disable_roce:
3501 	if (ll == IB_LINK_LAYER_ETHERNET) {
3502 		mlx5_disable_roce(dev);
3503 		mlx5_remove_roce_notifier(dev);
3504 	}
3505 
3506 err_free_port:
3507 	kfree(dev->port);
3508 
3509 err_dealloc:
3510 	ib_dealloc_device((struct ib_device *)dev);
3511 
3512 	return NULL;
3513 }
3514 
3515 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3516 {
3517 	struct mlx5_ib_dev *dev = context;
3518 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3519 
3520 	mlx5_ib_cleanup_congestion(dev);
3521 	mlx5_remove_roce_notifier(dev);
3522 	ib_unregister_device(&dev->ib_dev);
3523 	mlx5_ib_stage_bfreg_cleanup(dev);
3524 	mlx5_ib_dealloc_q_counters(dev);
3525 	destroy_umrc_res(dev);
3526 	mlx5_ib_odp_remove_one(dev);
3527 	destroy_dev_resources(&dev->devr);
3528 	if (ll == IB_LINK_LAYER_ETHERNET)
3529 		mlx5_disable_roce(dev);
3530 	kfree(dev->port);
3531 	ib_dealloc_device(&dev->ib_dev);
3532 }
3533 
3534 static struct mlx5_interface mlx5_ib_interface = {
3535 	.add            = mlx5_ib_add,
3536 	.remove         = mlx5_ib_remove,
3537 	.event          = mlx5_ib_event,
3538 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
3539 };
3540 
3541 static int __init mlx5_ib_init(void)
3542 {
3543 	int err;
3544 
3545 	err = mlx5_ib_odp_init();
3546 	if (err)
3547 		return err;
3548 
3549 	err = mlx5_register_interface(&mlx5_ib_interface);
3550 	if (err)
3551 		goto clean_odp;
3552 
3553 	return err;
3554 
3555 clean_odp:
3556 	mlx5_ib_odp_cleanup();
3557 	return err;
3558 }
3559 
3560 static void __exit mlx5_ib_cleanup(void)
3561 {
3562 	mlx5_unregister_interface(&mlx5_ib_interface);
3563 	mlx5_ib_odp_cleanup();
3564 }
3565 
3566 module_init_order(mlx5_ib_init, SI_ORDER_SEVENTH);
3567 module_exit_order(mlx5_ib_cleanup, SI_ORDER_SEVENTH);
3568