xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 1c2dd16a)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/highmem.h>
34 #include <linux/module.h>
35 #include <linux/init.h>
36 #include <linux/errno.h>
37 #include <linux/pci.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/slab.h>
40 #if defined(CONFIG_X86)
41 #include <asm/pat.h>
42 #endif
43 #include <linux/sched.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/task.h>
46 #include <linux/delay.h>
47 #include <rdma/ib_user_verbs.h>
48 #include <rdma/ib_addr.h>
49 #include <rdma/ib_cache.h>
50 #include <linux/mlx5/port.h>
51 #include <linux/mlx5/vport.h>
52 #include <linux/list.h>
53 #include <rdma/ib_smi.h>
54 #include <rdma/ib_umem.h>
55 #include <linux/in.h>
56 #include <linux/etherdevice.h>
57 #include <linux/mlx5/fs.h>
58 #include <linux/mlx5/vport.h>
59 #include "mlx5_ib.h"
60 
61 #define DRIVER_NAME "mlx5_ib"
62 #define DRIVER_VERSION "2.2-1"
63 #define DRIVER_RELDATE	"Feb 2014"
64 
65 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
66 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
67 MODULE_LICENSE("Dual BSD/GPL");
68 MODULE_VERSION(DRIVER_VERSION);
69 
70 static char mlx5_version[] =
71 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
72 	DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
73 
74 enum {
75 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
76 };
77 
78 static enum rdma_link_layer
79 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
80 {
81 	switch (port_type_cap) {
82 	case MLX5_CAP_PORT_TYPE_IB:
83 		return IB_LINK_LAYER_INFINIBAND;
84 	case MLX5_CAP_PORT_TYPE_ETH:
85 		return IB_LINK_LAYER_ETHERNET;
86 	default:
87 		return IB_LINK_LAYER_UNSPECIFIED;
88 	}
89 }
90 
91 static enum rdma_link_layer
92 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
93 {
94 	struct mlx5_ib_dev *dev = to_mdev(device);
95 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
96 
97 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
98 }
99 
100 static int mlx5_netdev_event(struct notifier_block *this,
101 			     unsigned long event, void *ptr)
102 {
103 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
104 	struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
105 						 roce.nb);
106 
107 	switch (event) {
108 	case NETDEV_REGISTER:
109 	case NETDEV_UNREGISTER:
110 		write_lock(&ibdev->roce.netdev_lock);
111 		if (ndev->dev.parent == &ibdev->mdev->pdev->dev)
112 			ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
113 					     NULL : ndev;
114 		write_unlock(&ibdev->roce.netdev_lock);
115 		break;
116 
117 	case NETDEV_UP:
118 	case NETDEV_DOWN: {
119 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
120 		struct net_device *upper = NULL;
121 
122 		if (lag_ndev) {
123 			upper = netdev_master_upper_dev_get(lag_ndev);
124 			dev_put(lag_ndev);
125 		}
126 
127 		if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
128 		    && ibdev->ib_active) {
129 			struct ib_event ibev = { };
130 
131 			ibev.device = &ibdev->ib_dev;
132 			ibev.event = (event == NETDEV_UP) ?
133 				     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
134 			ibev.element.port_num = 1;
135 			ib_dispatch_event(&ibev);
136 		}
137 		break;
138 	}
139 
140 	default:
141 		break;
142 	}
143 
144 	return NOTIFY_DONE;
145 }
146 
147 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
148 					     u8 port_num)
149 {
150 	struct mlx5_ib_dev *ibdev = to_mdev(device);
151 	struct net_device *ndev;
152 
153 	ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
154 	if (ndev)
155 		return ndev;
156 
157 	/* Ensure ndev does not disappear before we invoke dev_hold()
158 	 */
159 	read_lock(&ibdev->roce.netdev_lock);
160 	ndev = ibdev->roce.netdev;
161 	if (ndev)
162 		dev_hold(ndev);
163 	read_unlock(&ibdev->roce.netdev_lock);
164 
165 	return ndev;
166 }
167 
168 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
169 				struct ib_port_attr *props)
170 {
171 	struct mlx5_ib_dev *dev = to_mdev(device);
172 	struct net_device *ndev, *upper;
173 	enum ib_mtu ndev_ib_mtu;
174 	u16 qkey_viol_cntr;
175 
176 	/* props being zeroed by the caller, avoid zeroing it here */
177 
178 	props->port_cap_flags  |= IB_PORT_CM_SUP;
179 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
180 
181 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
182 						roce_address_table_size);
183 	props->max_mtu          = IB_MTU_4096;
184 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
185 	props->pkey_tbl_len     = 1;
186 	props->state            = IB_PORT_DOWN;
187 	props->phys_state       = 3;
188 
189 	mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
190 	props->qkey_viol_cntr = qkey_viol_cntr;
191 
192 	ndev = mlx5_ib_get_netdev(device, port_num);
193 	if (!ndev)
194 		return 0;
195 
196 	if (mlx5_lag_is_active(dev->mdev)) {
197 		rcu_read_lock();
198 		upper = netdev_master_upper_dev_get_rcu(ndev);
199 		if (upper) {
200 			dev_put(ndev);
201 			ndev = upper;
202 			dev_hold(ndev);
203 		}
204 		rcu_read_unlock();
205 	}
206 
207 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
208 		props->state      = IB_PORT_ACTIVE;
209 		props->phys_state = 5;
210 	}
211 
212 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
213 
214 	dev_put(ndev);
215 
216 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
217 
218 	props->active_width	= IB_WIDTH_4X;  /* TODO */
219 	props->active_speed	= IB_SPEED_QDR; /* TODO */
220 
221 	return 0;
222 }
223 
224 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
225 				     const struct ib_gid_attr *attr,
226 				     void *mlx5_addr)
227 {
228 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
229 	char *mlx5_addr_l3_addr	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
230 					       source_l3_address);
231 	void *mlx5_addr_mac	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
232 					       source_mac_47_32);
233 
234 	if (!gid)
235 		return;
236 
237 	ether_addr_copy(mlx5_addr_mac, attr->ndev->dev_addr);
238 
239 	if (is_vlan_dev(attr->ndev)) {
240 		MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
241 		MLX5_SET_RA(mlx5_addr, vlan_id, vlan_dev_vlan_id(attr->ndev));
242 	}
243 
244 	switch (attr->gid_type) {
245 	case IB_GID_TYPE_IB:
246 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
247 		break;
248 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
249 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
250 		break;
251 
252 	default:
253 		WARN_ON(true);
254 	}
255 
256 	if (attr->gid_type != IB_GID_TYPE_IB) {
257 		if (ipv6_addr_v4mapped((void *)gid))
258 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
259 				    MLX5_ROCE_L3_TYPE_IPV4);
260 		else
261 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
262 				    MLX5_ROCE_L3_TYPE_IPV6);
263 	}
264 
265 	if ((attr->gid_type == IB_GID_TYPE_IB) ||
266 	    !ipv6_addr_v4mapped((void *)gid))
267 		memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
268 	else
269 		memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
270 }
271 
272 static int set_roce_addr(struct ib_device *device, u8 port_num,
273 			 unsigned int index,
274 			 const union ib_gid *gid,
275 			 const struct ib_gid_attr *attr)
276 {
277 	struct mlx5_ib_dev *dev = to_mdev(device);
278 	u32  in[MLX5_ST_SZ_DW(set_roce_address_in)]  = {0};
279 	u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0};
280 	void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
281 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
282 
283 	if (ll != IB_LINK_LAYER_ETHERNET)
284 		return -EINVAL;
285 
286 	ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
287 
288 	MLX5_SET(set_roce_address_in, in, roce_address_index, index);
289 	MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
290 	return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
291 }
292 
293 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
294 			   unsigned int index, const union ib_gid *gid,
295 			   const struct ib_gid_attr *attr,
296 			   __always_unused void **context)
297 {
298 	return set_roce_addr(device, port_num, index, gid, attr);
299 }
300 
301 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
302 			   unsigned int index, __always_unused void **context)
303 {
304 	return set_roce_addr(device, port_num, index, NULL, NULL);
305 }
306 
307 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
308 			       int index)
309 {
310 	struct ib_gid_attr attr;
311 	union ib_gid gid;
312 
313 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
314 		return 0;
315 
316 	if (!attr.ndev)
317 		return 0;
318 
319 	dev_put(attr.ndev);
320 
321 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
322 		return 0;
323 
324 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
325 }
326 
327 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
328 			   int index, enum ib_gid_type *gid_type)
329 {
330 	struct ib_gid_attr attr;
331 	union ib_gid gid;
332 	int ret;
333 
334 	ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
335 	if (ret)
336 		return ret;
337 
338 	if (!attr.ndev)
339 		return -ENODEV;
340 
341 	dev_put(attr.ndev);
342 
343 	*gid_type = attr.gid_type;
344 
345 	return 0;
346 }
347 
348 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
349 {
350 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
351 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
352 	return 0;
353 }
354 
355 enum {
356 	MLX5_VPORT_ACCESS_METHOD_MAD,
357 	MLX5_VPORT_ACCESS_METHOD_HCA,
358 	MLX5_VPORT_ACCESS_METHOD_NIC,
359 };
360 
361 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
362 {
363 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
364 		return MLX5_VPORT_ACCESS_METHOD_MAD;
365 
366 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
367 	    IB_LINK_LAYER_ETHERNET)
368 		return MLX5_VPORT_ACCESS_METHOD_NIC;
369 
370 	return MLX5_VPORT_ACCESS_METHOD_HCA;
371 }
372 
373 static void get_atomic_caps(struct mlx5_ib_dev *dev,
374 			    struct ib_device_attr *props)
375 {
376 	u8 tmp;
377 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
378 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
379 	u8 atomic_req_8B_endianness_mode =
380 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
381 
382 	/* Check if HW supports 8 bytes standard atomic operations and capable
383 	 * of host endianness respond
384 	 */
385 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
386 	if (((atomic_operations & tmp) == tmp) &&
387 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
388 	    (atomic_req_8B_endianness_mode)) {
389 		props->atomic_cap = IB_ATOMIC_HCA;
390 	} else {
391 		props->atomic_cap = IB_ATOMIC_NONE;
392 	}
393 }
394 
395 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
396 					__be64 *sys_image_guid)
397 {
398 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
399 	struct mlx5_core_dev *mdev = dev->mdev;
400 	u64 tmp;
401 	int err;
402 
403 	switch (mlx5_get_vport_access_method(ibdev)) {
404 	case MLX5_VPORT_ACCESS_METHOD_MAD:
405 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
406 							    sys_image_guid);
407 
408 	case MLX5_VPORT_ACCESS_METHOD_HCA:
409 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
410 		break;
411 
412 	case MLX5_VPORT_ACCESS_METHOD_NIC:
413 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
414 		break;
415 
416 	default:
417 		return -EINVAL;
418 	}
419 
420 	if (!err)
421 		*sys_image_guid = cpu_to_be64(tmp);
422 
423 	return err;
424 
425 }
426 
427 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
428 				u16 *max_pkeys)
429 {
430 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
431 	struct mlx5_core_dev *mdev = dev->mdev;
432 
433 	switch (mlx5_get_vport_access_method(ibdev)) {
434 	case MLX5_VPORT_ACCESS_METHOD_MAD:
435 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
436 
437 	case MLX5_VPORT_ACCESS_METHOD_HCA:
438 	case MLX5_VPORT_ACCESS_METHOD_NIC:
439 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
440 						pkey_table_size));
441 		return 0;
442 
443 	default:
444 		return -EINVAL;
445 	}
446 }
447 
448 static int mlx5_query_vendor_id(struct ib_device *ibdev,
449 				u32 *vendor_id)
450 {
451 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
452 
453 	switch (mlx5_get_vport_access_method(ibdev)) {
454 	case MLX5_VPORT_ACCESS_METHOD_MAD:
455 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
456 
457 	case MLX5_VPORT_ACCESS_METHOD_HCA:
458 	case MLX5_VPORT_ACCESS_METHOD_NIC:
459 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
460 
461 	default:
462 		return -EINVAL;
463 	}
464 }
465 
466 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
467 				__be64 *node_guid)
468 {
469 	u64 tmp;
470 	int err;
471 
472 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
473 	case MLX5_VPORT_ACCESS_METHOD_MAD:
474 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
475 
476 	case MLX5_VPORT_ACCESS_METHOD_HCA:
477 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
478 		break;
479 
480 	case MLX5_VPORT_ACCESS_METHOD_NIC:
481 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
482 		break;
483 
484 	default:
485 		return -EINVAL;
486 	}
487 
488 	if (!err)
489 		*node_guid = cpu_to_be64(tmp);
490 
491 	return err;
492 }
493 
494 struct mlx5_reg_node_desc {
495 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
496 };
497 
498 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
499 {
500 	struct mlx5_reg_node_desc in;
501 
502 	if (mlx5_use_mad_ifc(dev))
503 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
504 
505 	memset(&in, 0, sizeof(in));
506 
507 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
508 				    sizeof(struct mlx5_reg_node_desc),
509 				    MLX5_REG_NODE_DESC, 0, 0);
510 }
511 
512 static int mlx5_ib_query_device(struct ib_device *ibdev,
513 				struct ib_device_attr *props,
514 				struct ib_udata *uhw)
515 {
516 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
517 	struct mlx5_core_dev *mdev = dev->mdev;
518 	int err = -ENOMEM;
519 	int max_sq_desc;
520 	int max_rq_sg;
521 	int max_sq_sg;
522 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
523 	struct mlx5_ib_query_device_resp resp = {};
524 	size_t resp_len;
525 	u64 max_tso;
526 
527 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
528 	if (uhw->outlen && uhw->outlen < resp_len)
529 		return -EINVAL;
530 	else
531 		resp.response_length = resp_len;
532 
533 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
534 		return -EINVAL;
535 
536 	memset(props, 0, sizeof(*props));
537 	err = mlx5_query_system_image_guid(ibdev,
538 					   &props->sys_image_guid);
539 	if (err)
540 		return err;
541 
542 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
543 	if (err)
544 		return err;
545 
546 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
547 	if (err)
548 		return err;
549 
550 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
551 		(fw_rev_min(dev->mdev) << 16) |
552 		fw_rev_sub(dev->mdev);
553 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
554 		IB_DEVICE_PORT_ACTIVE_EVENT		|
555 		IB_DEVICE_SYS_IMAGE_GUID		|
556 		IB_DEVICE_RC_RNR_NAK_GEN;
557 
558 	if (MLX5_CAP_GEN(mdev, pkv))
559 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
560 	if (MLX5_CAP_GEN(mdev, qkv))
561 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
562 	if (MLX5_CAP_GEN(mdev, apm))
563 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
564 	if (MLX5_CAP_GEN(mdev, xrc))
565 		props->device_cap_flags |= IB_DEVICE_XRC;
566 	if (MLX5_CAP_GEN(mdev, imaicl)) {
567 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
568 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
569 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
570 		/* We support 'Gappy' memory registration too */
571 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
572 	}
573 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
574 	if (MLX5_CAP_GEN(mdev, sho)) {
575 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
576 		/* At this stage no support for signature handover */
577 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
578 				      IB_PROT_T10DIF_TYPE_2 |
579 				      IB_PROT_T10DIF_TYPE_3;
580 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
581 				       IB_GUARD_T10DIF_CSUM;
582 	}
583 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
584 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
585 
586 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
587 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
588 			/* Legacy bit to support old userspace libraries */
589 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
590 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
591 		}
592 
593 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
594 			props->raw_packet_caps |=
595 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
596 
597 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
598 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
599 			if (max_tso) {
600 				resp.tso_caps.max_tso = 1 << max_tso;
601 				resp.tso_caps.supported_qpts |=
602 					1 << IB_QPT_RAW_PACKET;
603 				resp.response_length += sizeof(resp.tso_caps);
604 			}
605 		}
606 
607 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
608 			resp.rss_caps.rx_hash_function =
609 						MLX5_RX_HASH_FUNC_TOEPLITZ;
610 			resp.rss_caps.rx_hash_fields_mask =
611 						MLX5_RX_HASH_SRC_IPV4 |
612 						MLX5_RX_HASH_DST_IPV4 |
613 						MLX5_RX_HASH_SRC_IPV6 |
614 						MLX5_RX_HASH_DST_IPV6 |
615 						MLX5_RX_HASH_SRC_PORT_TCP |
616 						MLX5_RX_HASH_DST_PORT_TCP |
617 						MLX5_RX_HASH_SRC_PORT_UDP |
618 						MLX5_RX_HASH_DST_PORT_UDP;
619 			resp.response_length += sizeof(resp.rss_caps);
620 		}
621 	} else {
622 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
623 			resp.response_length += sizeof(resp.tso_caps);
624 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
625 			resp.response_length += sizeof(resp.rss_caps);
626 	}
627 
628 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
629 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
630 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
631 	}
632 
633 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
634 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
635 		/* Legacy bit to support old userspace libraries */
636 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
637 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
638 	}
639 
640 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
641 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
642 
643 	props->vendor_part_id	   = mdev->pdev->device;
644 	props->hw_ver		   = mdev->pdev->revision;
645 
646 	props->max_mr_size	   = ~0ull;
647 	props->page_size_cap	   = ~(min_page_size - 1);
648 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
649 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
650 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
651 		     sizeof(struct mlx5_wqe_data_seg);
652 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
653 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
654 		     sizeof(struct mlx5_wqe_raddr_seg)) /
655 		sizeof(struct mlx5_wqe_data_seg);
656 	props->max_sge = min(max_rq_sg, max_sq_sg);
657 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
658 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
659 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
660 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
661 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
662 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
663 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
664 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
665 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
666 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
667 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
668 	props->max_srq_sge	   = max_rq_sg - 1;
669 	props->max_fast_reg_page_list_len =
670 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
671 	get_atomic_caps(dev, props);
672 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
673 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
674 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
675 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
676 					   props->max_mcast_grp;
677 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
678 	props->max_ah = INT_MAX;
679 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
680 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
681 
682 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
683 	if (MLX5_CAP_GEN(mdev, pg))
684 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
685 	props->odp_caps = dev->odp_caps;
686 #endif
687 
688 	if (MLX5_CAP_GEN(mdev, cd))
689 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
690 
691 	if (!mlx5_core_is_pf(mdev))
692 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
693 
694 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
695 	    IB_LINK_LAYER_ETHERNET) {
696 		props->rss_caps.max_rwq_indirection_tables =
697 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
698 		props->rss_caps.max_rwq_indirection_table_size =
699 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
700 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
701 		props->max_wq_type_rq =
702 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
703 	}
704 
705 	if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
706 		resp.cqe_comp_caps.max_num =
707 			MLX5_CAP_GEN(dev->mdev, cqe_compression) ?
708 			MLX5_CAP_GEN(dev->mdev, cqe_compression_max_num) : 0;
709 		resp.cqe_comp_caps.supported_format =
710 			MLX5_IB_CQE_RES_FORMAT_HASH |
711 			MLX5_IB_CQE_RES_FORMAT_CSUM;
712 		resp.response_length += sizeof(resp.cqe_comp_caps);
713 	}
714 
715 	if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen)) {
716 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
717 		    MLX5_CAP_GEN(mdev, qos)) {
718 			resp.packet_pacing_caps.qp_rate_limit_max =
719 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
720 			resp.packet_pacing_caps.qp_rate_limit_min =
721 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
722 			resp.packet_pacing_caps.supported_qpts |=
723 				1 << IB_QPT_RAW_PACKET;
724 		}
725 		resp.response_length += sizeof(resp.packet_pacing_caps);
726 	}
727 
728 	if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
729 			uhw->outlen)) {
730 		resp.mlx5_ib_support_multi_pkt_send_wqes =
731 			MLX5_CAP_ETH(mdev, multi_pkt_send_wqe);
732 		resp.response_length +=
733 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
734 	}
735 
736 	if (field_avail(typeof(resp), reserved, uhw->outlen))
737 		resp.response_length += sizeof(resp.reserved);
738 
739 	if (uhw->outlen) {
740 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
741 
742 		if (err)
743 			return err;
744 	}
745 
746 	return 0;
747 }
748 
749 enum mlx5_ib_width {
750 	MLX5_IB_WIDTH_1X	= 1 << 0,
751 	MLX5_IB_WIDTH_2X	= 1 << 1,
752 	MLX5_IB_WIDTH_4X	= 1 << 2,
753 	MLX5_IB_WIDTH_8X	= 1 << 3,
754 	MLX5_IB_WIDTH_12X	= 1 << 4
755 };
756 
757 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
758 				  u8 *ib_width)
759 {
760 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
761 	int err = 0;
762 
763 	if (active_width & MLX5_IB_WIDTH_1X) {
764 		*ib_width = IB_WIDTH_1X;
765 	} else if (active_width & MLX5_IB_WIDTH_2X) {
766 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
767 			    (int)active_width);
768 		err = -EINVAL;
769 	} else if (active_width & MLX5_IB_WIDTH_4X) {
770 		*ib_width = IB_WIDTH_4X;
771 	} else if (active_width & MLX5_IB_WIDTH_8X) {
772 		*ib_width = IB_WIDTH_8X;
773 	} else if (active_width & MLX5_IB_WIDTH_12X) {
774 		*ib_width = IB_WIDTH_12X;
775 	} else {
776 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
777 			    (int)active_width);
778 		err = -EINVAL;
779 	}
780 
781 	return err;
782 }
783 
784 static int mlx5_mtu_to_ib_mtu(int mtu)
785 {
786 	switch (mtu) {
787 	case 256: return 1;
788 	case 512: return 2;
789 	case 1024: return 3;
790 	case 2048: return 4;
791 	case 4096: return 5;
792 	default:
793 		pr_warn("invalid mtu\n");
794 		return -1;
795 	}
796 }
797 
798 enum ib_max_vl_num {
799 	__IB_MAX_VL_0		= 1,
800 	__IB_MAX_VL_0_1		= 2,
801 	__IB_MAX_VL_0_3		= 3,
802 	__IB_MAX_VL_0_7		= 4,
803 	__IB_MAX_VL_0_14	= 5,
804 };
805 
806 enum mlx5_vl_hw_cap {
807 	MLX5_VL_HW_0	= 1,
808 	MLX5_VL_HW_0_1	= 2,
809 	MLX5_VL_HW_0_2	= 3,
810 	MLX5_VL_HW_0_3	= 4,
811 	MLX5_VL_HW_0_4	= 5,
812 	MLX5_VL_HW_0_5	= 6,
813 	MLX5_VL_HW_0_6	= 7,
814 	MLX5_VL_HW_0_7	= 8,
815 	MLX5_VL_HW_0_14	= 15
816 };
817 
818 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
819 				u8 *max_vl_num)
820 {
821 	switch (vl_hw_cap) {
822 	case MLX5_VL_HW_0:
823 		*max_vl_num = __IB_MAX_VL_0;
824 		break;
825 	case MLX5_VL_HW_0_1:
826 		*max_vl_num = __IB_MAX_VL_0_1;
827 		break;
828 	case MLX5_VL_HW_0_3:
829 		*max_vl_num = __IB_MAX_VL_0_3;
830 		break;
831 	case MLX5_VL_HW_0_7:
832 		*max_vl_num = __IB_MAX_VL_0_7;
833 		break;
834 	case MLX5_VL_HW_0_14:
835 		*max_vl_num = __IB_MAX_VL_0_14;
836 		break;
837 
838 	default:
839 		return -EINVAL;
840 	}
841 
842 	return 0;
843 }
844 
845 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
846 			       struct ib_port_attr *props)
847 {
848 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
849 	struct mlx5_core_dev *mdev = dev->mdev;
850 	struct mlx5_hca_vport_context *rep;
851 	u16 max_mtu;
852 	u16 oper_mtu;
853 	int err;
854 	u8 ib_link_width_oper;
855 	u8 vl_hw_cap;
856 
857 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
858 	if (!rep) {
859 		err = -ENOMEM;
860 		goto out;
861 	}
862 
863 	/* props being zeroed by the caller, avoid zeroing it here */
864 
865 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
866 	if (err)
867 		goto out;
868 
869 	props->lid		= rep->lid;
870 	props->lmc		= rep->lmc;
871 	props->sm_lid		= rep->sm_lid;
872 	props->sm_sl		= rep->sm_sl;
873 	props->state		= rep->vport_state;
874 	props->phys_state	= rep->port_physical_state;
875 	props->port_cap_flags	= rep->cap_mask1;
876 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
877 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
878 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
879 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
880 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
881 	props->subnet_timeout	= rep->subnet_timeout;
882 	props->init_type_reply	= rep->init_type_reply;
883 	props->grh_required	= rep->grh_required;
884 
885 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
886 	if (err)
887 		goto out;
888 
889 	err = translate_active_width(ibdev, ib_link_width_oper,
890 				     &props->active_width);
891 	if (err)
892 		goto out;
893 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
894 	if (err)
895 		goto out;
896 
897 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
898 
899 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
900 
901 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
902 
903 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
904 
905 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
906 	if (err)
907 		goto out;
908 
909 	err = translate_max_vl_num(ibdev, vl_hw_cap,
910 				   &props->max_vl_num);
911 out:
912 	kfree(rep);
913 	return err;
914 }
915 
916 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
917 		       struct ib_port_attr *props)
918 {
919 	switch (mlx5_get_vport_access_method(ibdev)) {
920 	case MLX5_VPORT_ACCESS_METHOD_MAD:
921 		return mlx5_query_mad_ifc_port(ibdev, port, props);
922 
923 	case MLX5_VPORT_ACCESS_METHOD_HCA:
924 		return mlx5_query_hca_port(ibdev, port, props);
925 
926 	case MLX5_VPORT_ACCESS_METHOD_NIC:
927 		return mlx5_query_port_roce(ibdev, port, props);
928 
929 	default:
930 		return -EINVAL;
931 	}
932 }
933 
934 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
935 			     union ib_gid *gid)
936 {
937 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
938 	struct mlx5_core_dev *mdev = dev->mdev;
939 
940 	switch (mlx5_get_vport_access_method(ibdev)) {
941 	case MLX5_VPORT_ACCESS_METHOD_MAD:
942 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
943 
944 	case MLX5_VPORT_ACCESS_METHOD_HCA:
945 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
946 
947 	default:
948 		return -EINVAL;
949 	}
950 
951 }
952 
953 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
954 			      u16 *pkey)
955 {
956 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
957 	struct mlx5_core_dev *mdev = dev->mdev;
958 
959 	switch (mlx5_get_vport_access_method(ibdev)) {
960 	case MLX5_VPORT_ACCESS_METHOD_MAD:
961 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
962 
963 	case MLX5_VPORT_ACCESS_METHOD_HCA:
964 	case MLX5_VPORT_ACCESS_METHOD_NIC:
965 		return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
966 						 pkey);
967 	default:
968 		return -EINVAL;
969 	}
970 }
971 
972 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
973 				 struct ib_device_modify *props)
974 {
975 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
976 	struct mlx5_reg_node_desc in;
977 	struct mlx5_reg_node_desc out;
978 	int err;
979 
980 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
981 		return -EOPNOTSUPP;
982 
983 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
984 		return 0;
985 
986 	/*
987 	 * If possible, pass node desc to FW, so it can generate
988 	 * a 144 trap.  If cmd fails, just ignore.
989 	 */
990 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
991 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
992 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
993 	if (err)
994 		return err;
995 
996 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
997 
998 	return err;
999 }
1000 
1001 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1002 				u32 value)
1003 {
1004 	struct mlx5_hca_vport_context ctx = {};
1005 	int err;
1006 
1007 	err = mlx5_query_hca_vport_context(dev->mdev, 0,
1008 					   port_num, 0, &ctx);
1009 	if (err)
1010 		return err;
1011 
1012 	if (~ctx.cap_mask1_perm & mask) {
1013 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1014 			     mask, ctx.cap_mask1_perm);
1015 		return -EINVAL;
1016 	}
1017 
1018 	ctx.cap_mask1 = value;
1019 	ctx.cap_mask1_perm = mask;
1020 	err = mlx5_core_modify_hca_vport_context(dev->mdev, 0,
1021 						 port_num, 0, &ctx);
1022 
1023 	return err;
1024 }
1025 
1026 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1027 			       struct ib_port_modify *props)
1028 {
1029 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1030 	struct ib_port_attr attr;
1031 	u32 tmp;
1032 	int err;
1033 	u32 change_mask;
1034 	u32 value;
1035 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1036 		      IB_LINK_LAYER_INFINIBAND);
1037 
1038 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1039 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1040 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1041 		return set_port_caps_atomic(dev, port, change_mask, value);
1042 	}
1043 
1044 	mutex_lock(&dev->cap_mask_mutex);
1045 
1046 	err = ib_query_port(ibdev, port, &attr);
1047 	if (err)
1048 		goto out;
1049 
1050 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1051 		~props->clr_port_cap_mask;
1052 
1053 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1054 
1055 out:
1056 	mutex_unlock(&dev->cap_mask_mutex);
1057 	return err;
1058 }
1059 
1060 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1061 {
1062 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1063 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1064 }
1065 
1066 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1067 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1068 			     u32 *num_sys_pages)
1069 {
1070 	int uars_per_sys_page;
1071 	int bfregs_per_sys_page;
1072 	int ref_bfregs = req->total_num_bfregs;
1073 
1074 	if (req->total_num_bfregs == 0)
1075 		return -EINVAL;
1076 
1077 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1078 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1079 
1080 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1081 		return -ENOMEM;
1082 
1083 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1084 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1085 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1086 	*num_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1087 
1088 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1089 		return -EINVAL;
1090 
1091 	mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, alloated %d, using %d sys pages\n",
1092 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1093 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1094 		    req->total_num_bfregs, *num_sys_pages);
1095 
1096 	return 0;
1097 }
1098 
1099 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1100 {
1101 	struct mlx5_bfreg_info *bfregi;
1102 	int err;
1103 	int i;
1104 
1105 	bfregi = &context->bfregi;
1106 	for (i = 0; i < bfregi->num_sys_pages; i++) {
1107 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1108 		if (err)
1109 			goto error;
1110 
1111 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1112 	}
1113 	return 0;
1114 
1115 error:
1116 	for (--i; i >= 0; i--)
1117 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1118 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1119 
1120 	return err;
1121 }
1122 
1123 static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1124 {
1125 	struct mlx5_bfreg_info *bfregi;
1126 	int err;
1127 	int i;
1128 
1129 	bfregi = &context->bfregi;
1130 	for (i = 0; i < bfregi->num_sys_pages; i++) {
1131 		err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1132 		if (err) {
1133 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1134 			return err;
1135 		}
1136 	}
1137 	return 0;
1138 }
1139 
1140 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1141 						  struct ib_udata *udata)
1142 {
1143 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1144 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1145 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1146 	struct mlx5_ib_ucontext *context;
1147 	struct mlx5_bfreg_info *bfregi;
1148 	int ver;
1149 	int err;
1150 	size_t reqlen;
1151 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1152 				     max_cqe_version);
1153 	bool lib_uar_4k;
1154 
1155 	if (!dev->ib_active)
1156 		return ERR_PTR(-EAGAIN);
1157 
1158 	if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
1159 		return ERR_PTR(-EINVAL);
1160 
1161 	reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
1162 	if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1163 		ver = 0;
1164 	else if (reqlen >= min_req_v2)
1165 		ver = 2;
1166 	else
1167 		return ERR_PTR(-EINVAL);
1168 
1169 	err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
1170 	if (err)
1171 		return ERR_PTR(err);
1172 
1173 	if (req.flags)
1174 		return ERR_PTR(-EINVAL);
1175 
1176 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1177 		return ERR_PTR(-EOPNOTSUPP);
1178 
1179 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1180 				    MLX5_NON_FP_BFREGS_PER_UAR);
1181 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1182 		return ERR_PTR(-EINVAL);
1183 
1184 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1185 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1186 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1187 	resp.cache_line_size = cache_line_size();
1188 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1189 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1190 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1191 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1192 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1193 	resp.cqe_version = min_t(__u8,
1194 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1195 				 req.max_cqe_version);
1196 	resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1197 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1198 	resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1199 					MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1200 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1201 				   sizeof(resp.response_length), udata->outlen);
1202 
1203 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1204 	if (!context)
1205 		return ERR_PTR(-ENOMEM);
1206 
1207 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1208 	bfregi = &context->bfregi;
1209 
1210 	/* updates req->total_num_bfregs */
1211 	err = calc_total_bfregs(dev, lib_uar_4k, &req, &bfregi->num_sys_pages);
1212 	if (err)
1213 		goto out_ctx;
1214 
1215 	mutex_init(&bfregi->lock);
1216 	bfregi->lib_uar_4k = lib_uar_4k;
1217 	bfregi->count = kcalloc(req.total_num_bfregs, sizeof(*bfregi->count),
1218 				GFP_KERNEL);
1219 	if (!bfregi->count) {
1220 		err = -ENOMEM;
1221 		goto out_ctx;
1222 	}
1223 
1224 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1225 				    sizeof(*bfregi->sys_pages),
1226 				    GFP_KERNEL);
1227 	if (!bfregi->sys_pages) {
1228 		err = -ENOMEM;
1229 		goto out_count;
1230 	}
1231 
1232 	err = allocate_uars(dev, context);
1233 	if (err)
1234 		goto out_sys_pages;
1235 
1236 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1237 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1238 #endif
1239 
1240 	context->upd_xlt_page = __get_free_page(GFP_KERNEL);
1241 	if (!context->upd_xlt_page) {
1242 		err = -ENOMEM;
1243 		goto out_uars;
1244 	}
1245 	mutex_init(&context->upd_xlt_page_mutex);
1246 
1247 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1248 		err = mlx5_core_alloc_transport_domain(dev->mdev,
1249 						       &context->tdn);
1250 		if (err)
1251 			goto out_page;
1252 	}
1253 
1254 	INIT_LIST_HEAD(&context->vma_private_list);
1255 	INIT_LIST_HEAD(&context->db_page_list);
1256 	mutex_init(&context->db_page_mutex);
1257 
1258 	resp.tot_bfregs = req.total_num_bfregs;
1259 	resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1260 
1261 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1262 		resp.response_length += sizeof(resp.cqe_version);
1263 
1264 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1265 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1266 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1267 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1268 	}
1269 
1270 	if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1271 		if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1272 			mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1273 			resp.eth_min_inline++;
1274 		}
1275 		resp.response_length += sizeof(resp.eth_min_inline);
1276 	}
1277 
1278 	/*
1279 	 * We don't want to expose information from the PCI bar that is located
1280 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1281 	 * pretend we don't support reading the HCA's core clock. This is also
1282 	 * forced by mmap function.
1283 	 */
1284 	if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1285 		if (PAGE_SIZE <= 4096) {
1286 			resp.comp_mask |=
1287 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1288 			resp.hca_core_clock_offset =
1289 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1290 		}
1291 		resp.response_length += sizeof(resp.hca_core_clock_offset) +
1292 					sizeof(resp.reserved2);
1293 	}
1294 
1295 	if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1296 		resp.response_length += sizeof(resp.log_uar_size);
1297 
1298 	if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1299 		resp.response_length += sizeof(resp.num_uars_per_page);
1300 
1301 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1302 	if (err)
1303 		goto out_td;
1304 
1305 	bfregi->ver = ver;
1306 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1307 	context->cqe_version = resp.cqe_version;
1308 	context->lib_caps = req.lib_caps;
1309 	print_lib_caps(dev, context->lib_caps);
1310 
1311 	return &context->ibucontext;
1312 
1313 out_td:
1314 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1315 		mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1316 
1317 out_page:
1318 	free_page(context->upd_xlt_page);
1319 
1320 out_uars:
1321 	deallocate_uars(dev, context);
1322 
1323 out_sys_pages:
1324 	kfree(bfregi->sys_pages);
1325 
1326 out_count:
1327 	kfree(bfregi->count);
1328 
1329 out_ctx:
1330 	kfree(context);
1331 
1332 	return ERR_PTR(err);
1333 }
1334 
1335 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1336 {
1337 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1338 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1339 	struct mlx5_bfreg_info *bfregi;
1340 
1341 	bfregi = &context->bfregi;
1342 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1343 		mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1344 
1345 	free_page(context->upd_xlt_page);
1346 	deallocate_uars(dev, context);
1347 	kfree(bfregi->sys_pages);
1348 	kfree(bfregi->count);
1349 	kfree(context);
1350 
1351 	return 0;
1352 }
1353 
1354 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1355 				 struct mlx5_bfreg_info *bfregi,
1356 				 int idx)
1357 {
1358 	int fw_uars_per_page;
1359 
1360 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1361 
1362 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) +
1363 			bfregi->sys_pages[idx] / fw_uars_per_page;
1364 }
1365 
1366 static int get_command(unsigned long offset)
1367 {
1368 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1369 }
1370 
1371 static int get_arg(unsigned long offset)
1372 {
1373 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1374 }
1375 
1376 static int get_index(unsigned long offset)
1377 {
1378 	return get_arg(offset);
1379 }
1380 
1381 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1382 {
1383 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1384 	 * is done through either mremap flow or split_vma (usually due to
1385 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1386 	 * as this VMA is strongly hardware related.  Therefore we set the
1387 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1388 	 * calling us again and trying to do incorrect actions.  We assume that
1389 	 * the original VMA size is exactly a single page, and therefore all
1390 	 * "splitting" operation will not happen to it.
1391 	 */
1392 	area->vm_ops = NULL;
1393 }
1394 
1395 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1396 {
1397 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1398 
1399 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1400 	 * file itself is closed, therefore no sync is needed with the regular
1401 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1402 	 * However need a sync with accessing the vma as part of
1403 	 * mlx5_ib_disassociate_ucontext.
1404 	 * The close operation is usually called under mm->mmap_sem except when
1405 	 * process is exiting.
1406 	 * The exiting case is handled explicitly as part of
1407 	 * mlx5_ib_disassociate_ucontext.
1408 	 */
1409 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1410 
1411 	/* setting the vma context pointer to null in the mlx5_ib driver's
1412 	 * private data, to protect a race condition in
1413 	 * mlx5_ib_disassociate_ucontext().
1414 	 */
1415 	mlx5_ib_vma_priv_data->vma = NULL;
1416 	list_del(&mlx5_ib_vma_priv_data->list);
1417 	kfree(mlx5_ib_vma_priv_data);
1418 }
1419 
1420 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1421 	.open = mlx5_ib_vma_open,
1422 	.close = mlx5_ib_vma_close
1423 };
1424 
1425 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1426 				struct mlx5_ib_ucontext *ctx)
1427 {
1428 	struct mlx5_ib_vma_private_data *vma_prv;
1429 	struct list_head *vma_head = &ctx->vma_private_list;
1430 
1431 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1432 	if (!vma_prv)
1433 		return -ENOMEM;
1434 
1435 	vma_prv->vma = vma;
1436 	vma->vm_private_data = vma_prv;
1437 	vma->vm_ops =  &mlx5_ib_vm_ops;
1438 
1439 	list_add(&vma_prv->list, vma_head);
1440 
1441 	return 0;
1442 }
1443 
1444 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1445 {
1446 	int ret;
1447 	struct vm_area_struct *vma;
1448 	struct mlx5_ib_vma_private_data *vma_private, *n;
1449 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1450 	struct task_struct *owning_process  = NULL;
1451 	struct mm_struct   *owning_mm       = NULL;
1452 
1453 	owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1454 	if (!owning_process)
1455 		return;
1456 
1457 	owning_mm = get_task_mm(owning_process);
1458 	if (!owning_mm) {
1459 		pr_info("no mm, disassociate ucontext is pending task termination\n");
1460 		while (1) {
1461 			put_task_struct(owning_process);
1462 			usleep_range(1000, 2000);
1463 			owning_process = get_pid_task(ibcontext->tgid,
1464 						      PIDTYPE_PID);
1465 			if (!owning_process ||
1466 			    owning_process->state == TASK_DEAD) {
1467 				pr_info("disassociate ucontext done, task was terminated\n");
1468 				/* in case task was dead need to release the
1469 				 * task struct.
1470 				 */
1471 				if (owning_process)
1472 					put_task_struct(owning_process);
1473 				return;
1474 			}
1475 		}
1476 	}
1477 
1478 	/* need to protect from a race on closing the vma as part of
1479 	 * mlx5_ib_vma_close.
1480 	 */
1481 	down_read(&owning_mm->mmap_sem);
1482 	list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1483 				 list) {
1484 		vma = vma_private->vma;
1485 		ret = zap_vma_ptes(vma, vma->vm_start,
1486 				   PAGE_SIZE);
1487 		WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1488 		/* context going to be destroyed, should
1489 		 * not access ops any more.
1490 		 */
1491 		vma->vm_ops = NULL;
1492 		list_del(&vma_private->list);
1493 		kfree(vma_private);
1494 	}
1495 	up_read(&owning_mm->mmap_sem);
1496 	mmput(owning_mm);
1497 	put_task_struct(owning_process);
1498 }
1499 
1500 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1501 {
1502 	switch (cmd) {
1503 	case MLX5_IB_MMAP_WC_PAGE:
1504 		return "WC";
1505 	case MLX5_IB_MMAP_REGULAR_PAGE:
1506 		return "best effort WC";
1507 	case MLX5_IB_MMAP_NC_PAGE:
1508 		return "NC";
1509 	default:
1510 		return NULL;
1511 	}
1512 }
1513 
1514 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1515 		    struct vm_area_struct *vma,
1516 		    struct mlx5_ib_ucontext *context)
1517 {
1518 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1519 	int err;
1520 	unsigned long idx;
1521 	phys_addr_t pfn, pa;
1522 	pgprot_t prot;
1523 	int uars_per_page;
1524 
1525 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1526 		return -EINVAL;
1527 
1528 	uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
1529 	idx = get_index(vma->vm_pgoff);
1530 	if (idx % uars_per_page ||
1531 	    idx * uars_per_page >= bfregi->num_sys_pages) {
1532 		mlx5_ib_warn(dev, "invalid uar index %lu\n", idx);
1533 		return -EINVAL;
1534 	}
1535 
1536 	switch (cmd) {
1537 	case MLX5_IB_MMAP_WC_PAGE:
1538 /* Some architectures don't support WC memory */
1539 #if defined(CONFIG_X86)
1540 		if (!pat_enabled())
1541 			return -EPERM;
1542 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1543 			return -EPERM;
1544 #endif
1545 	/* fall through */
1546 	case MLX5_IB_MMAP_REGULAR_PAGE:
1547 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1548 		prot = pgprot_writecombine(vma->vm_page_prot);
1549 		break;
1550 	case MLX5_IB_MMAP_NC_PAGE:
1551 		prot = pgprot_noncached(vma->vm_page_prot);
1552 		break;
1553 	default:
1554 		return -EINVAL;
1555 	}
1556 
1557 	pfn = uar_index2pfn(dev, bfregi, idx);
1558 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1559 
1560 	vma->vm_page_prot = prot;
1561 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1562 				 PAGE_SIZE, vma->vm_page_prot);
1563 	if (err) {
1564 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
1565 			    err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
1566 		return -EAGAIN;
1567 	}
1568 
1569 	pa = pfn << PAGE_SHIFT;
1570 	mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
1571 		    vma->vm_start, &pa);
1572 
1573 	return mlx5_ib_set_vma_data(vma, context);
1574 }
1575 
1576 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1577 {
1578 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1579 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1580 	unsigned long command;
1581 	phys_addr_t pfn;
1582 
1583 	command = get_command(vma->vm_pgoff);
1584 	switch (command) {
1585 	case MLX5_IB_MMAP_WC_PAGE:
1586 	case MLX5_IB_MMAP_NC_PAGE:
1587 	case MLX5_IB_MMAP_REGULAR_PAGE:
1588 		return uar_mmap(dev, command, vma, context);
1589 
1590 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1591 		return -ENOSYS;
1592 
1593 	case MLX5_IB_MMAP_CORE_CLOCK:
1594 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1595 			return -EINVAL;
1596 
1597 		if (vma->vm_flags & VM_WRITE)
1598 			return -EPERM;
1599 
1600 		/* Don't expose to user-space information it shouldn't have */
1601 		if (PAGE_SIZE > 4096)
1602 			return -EOPNOTSUPP;
1603 
1604 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1605 		pfn = (dev->mdev->iseg_base +
1606 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1607 			PAGE_SHIFT;
1608 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1609 				       PAGE_SIZE, vma->vm_page_prot))
1610 			return -EAGAIN;
1611 
1612 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
1613 			    vma->vm_start,
1614 			    (unsigned long long)pfn << PAGE_SHIFT);
1615 		break;
1616 
1617 	default:
1618 		return -EINVAL;
1619 	}
1620 
1621 	return 0;
1622 }
1623 
1624 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1625 				      struct ib_ucontext *context,
1626 				      struct ib_udata *udata)
1627 {
1628 	struct mlx5_ib_alloc_pd_resp resp;
1629 	struct mlx5_ib_pd *pd;
1630 	int err;
1631 
1632 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1633 	if (!pd)
1634 		return ERR_PTR(-ENOMEM);
1635 
1636 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1637 	if (err) {
1638 		kfree(pd);
1639 		return ERR_PTR(err);
1640 	}
1641 
1642 	if (context) {
1643 		resp.pdn = pd->pdn;
1644 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1645 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1646 			kfree(pd);
1647 			return ERR_PTR(-EFAULT);
1648 		}
1649 	}
1650 
1651 	return &pd->ibpd;
1652 }
1653 
1654 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1655 {
1656 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1657 	struct mlx5_ib_pd *mpd = to_mpd(pd);
1658 
1659 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1660 	kfree(mpd);
1661 
1662 	return 0;
1663 }
1664 
1665 enum {
1666 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
1667 	MATCH_CRITERIA_ENABLE_MISC_BIT,
1668 	MATCH_CRITERIA_ENABLE_INNER_BIT
1669 };
1670 
1671 #define HEADER_IS_ZERO(match_criteria, headers)			           \
1672 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1673 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
1674 
1675 static u8 get_match_criteria_enable(u32 *match_criteria)
1676 {
1677 	u8 match_criteria_enable;
1678 
1679 	match_criteria_enable =
1680 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1681 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
1682 	match_criteria_enable |=
1683 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1684 		MATCH_CRITERIA_ENABLE_MISC_BIT;
1685 	match_criteria_enable |=
1686 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1687 		MATCH_CRITERIA_ENABLE_INNER_BIT;
1688 
1689 	return match_criteria_enable;
1690 }
1691 
1692 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1693 {
1694 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1695 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1696 }
1697 
1698 static void set_flow_label(void *misc_c, void *misc_v, u8 mask, u8 val,
1699 			   bool inner)
1700 {
1701 	if (inner) {
1702 		MLX5_SET(fte_match_set_misc,
1703 			 misc_c, inner_ipv6_flow_label, mask);
1704 		MLX5_SET(fte_match_set_misc,
1705 			 misc_v, inner_ipv6_flow_label, val);
1706 	} else {
1707 		MLX5_SET(fte_match_set_misc,
1708 			 misc_c, outer_ipv6_flow_label, mask);
1709 		MLX5_SET(fte_match_set_misc,
1710 			 misc_v, outer_ipv6_flow_label, val);
1711 	}
1712 }
1713 
1714 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1715 {
1716 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1717 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1718 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1719 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1720 }
1721 
1722 #define LAST_ETH_FIELD vlan_tag
1723 #define LAST_IB_FIELD sl
1724 #define LAST_IPV4_FIELD tos
1725 #define LAST_IPV6_FIELD traffic_class
1726 #define LAST_TCP_UDP_FIELD src_port
1727 #define LAST_TUNNEL_FIELD tunnel_id
1728 #define LAST_FLOW_TAG_FIELD tag_id
1729 
1730 /* Field is the last supported field */
1731 #define FIELDS_NOT_SUPPORTED(filter, field)\
1732 	memchr_inv((void *)&filter.field  +\
1733 		   sizeof(filter.field), 0,\
1734 		   sizeof(filter) -\
1735 		   offsetof(typeof(filter), field) -\
1736 		   sizeof(filter.field))
1737 
1738 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1739 			   const union ib_flow_spec *ib_spec, u32 *tag_id)
1740 {
1741 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1742 					   misc_parameters);
1743 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1744 					   misc_parameters);
1745 	void *headers_c;
1746 	void *headers_v;
1747 
1748 	if (ib_spec->type & IB_FLOW_SPEC_INNER) {
1749 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1750 					 inner_headers);
1751 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1752 					 inner_headers);
1753 	} else {
1754 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1755 					 outer_headers);
1756 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1757 					 outer_headers);
1758 	}
1759 
1760 	switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
1761 	case IB_FLOW_SPEC_ETH:
1762 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1763 			return -EOPNOTSUPP;
1764 
1765 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1766 					     dmac_47_16),
1767 				ib_spec->eth.mask.dst_mac);
1768 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1769 					     dmac_47_16),
1770 				ib_spec->eth.val.dst_mac);
1771 
1772 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1773 					     smac_47_16),
1774 				ib_spec->eth.mask.src_mac);
1775 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1776 					     smac_47_16),
1777 				ib_spec->eth.val.src_mac);
1778 
1779 		if (ib_spec->eth.mask.vlan_tag) {
1780 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1781 				 cvlan_tag, 1);
1782 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1783 				 cvlan_tag, 1);
1784 
1785 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1786 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1787 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1788 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1789 
1790 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1791 				 first_cfi,
1792 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1793 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1794 				 first_cfi,
1795 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1796 
1797 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1798 				 first_prio,
1799 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1800 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1801 				 first_prio,
1802 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1803 		}
1804 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1805 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1806 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1807 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
1808 		break;
1809 	case IB_FLOW_SPEC_IPV4:
1810 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1811 			return -EOPNOTSUPP;
1812 
1813 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1814 			 ethertype, 0xffff);
1815 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1816 			 ethertype, ETH_P_IP);
1817 
1818 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1819 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1820 		       &ib_spec->ipv4.mask.src_ip,
1821 		       sizeof(ib_spec->ipv4.mask.src_ip));
1822 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1823 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1824 		       &ib_spec->ipv4.val.src_ip,
1825 		       sizeof(ib_spec->ipv4.val.src_ip));
1826 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1827 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1828 		       &ib_spec->ipv4.mask.dst_ip,
1829 		       sizeof(ib_spec->ipv4.mask.dst_ip));
1830 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1831 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1832 		       &ib_spec->ipv4.val.dst_ip,
1833 		       sizeof(ib_spec->ipv4.val.dst_ip));
1834 
1835 		set_tos(headers_c, headers_v,
1836 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
1837 
1838 		set_proto(headers_c, headers_v,
1839 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
1840 		break;
1841 	case IB_FLOW_SPEC_IPV6:
1842 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1843 			return -EOPNOTSUPP;
1844 
1845 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1846 			 ethertype, 0xffff);
1847 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1848 			 ethertype, ETH_P_IPV6);
1849 
1850 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1851 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1852 		       &ib_spec->ipv6.mask.src_ip,
1853 		       sizeof(ib_spec->ipv6.mask.src_ip));
1854 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1855 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1856 		       &ib_spec->ipv6.val.src_ip,
1857 		       sizeof(ib_spec->ipv6.val.src_ip));
1858 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1859 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1860 		       &ib_spec->ipv6.mask.dst_ip,
1861 		       sizeof(ib_spec->ipv6.mask.dst_ip));
1862 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1863 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1864 		       &ib_spec->ipv6.val.dst_ip,
1865 		       sizeof(ib_spec->ipv6.val.dst_ip));
1866 
1867 		set_tos(headers_c, headers_v,
1868 			ib_spec->ipv6.mask.traffic_class,
1869 			ib_spec->ipv6.val.traffic_class);
1870 
1871 		set_proto(headers_c, headers_v,
1872 			  ib_spec->ipv6.mask.next_hdr,
1873 			  ib_spec->ipv6.val.next_hdr);
1874 
1875 		set_flow_label(misc_params_c, misc_params_v,
1876 			       ntohl(ib_spec->ipv6.mask.flow_label),
1877 			       ntohl(ib_spec->ipv6.val.flow_label),
1878 			       ib_spec->type & IB_FLOW_SPEC_INNER);
1879 
1880 		break;
1881 	case IB_FLOW_SPEC_TCP:
1882 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1883 					 LAST_TCP_UDP_FIELD))
1884 			return -EOPNOTSUPP;
1885 
1886 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
1887 			 0xff);
1888 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
1889 			 IPPROTO_TCP);
1890 
1891 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
1892 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1893 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
1894 			 ntohs(ib_spec->tcp_udp.val.src_port));
1895 
1896 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
1897 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1898 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
1899 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1900 		break;
1901 	case IB_FLOW_SPEC_UDP:
1902 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1903 					 LAST_TCP_UDP_FIELD))
1904 			return -EOPNOTSUPP;
1905 
1906 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
1907 			 0xff);
1908 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
1909 			 IPPROTO_UDP);
1910 
1911 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
1912 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1913 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
1914 			 ntohs(ib_spec->tcp_udp.val.src_port));
1915 
1916 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
1917 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1918 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
1919 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1920 		break;
1921 	case IB_FLOW_SPEC_VXLAN_TUNNEL:
1922 		if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
1923 					 LAST_TUNNEL_FIELD))
1924 			return -EOPNOTSUPP;
1925 
1926 		MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
1927 			 ntohl(ib_spec->tunnel.mask.tunnel_id));
1928 		MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
1929 			 ntohl(ib_spec->tunnel.val.tunnel_id));
1930 		break;
1931 	case IB_FLOW_SPEC_ACTION_TAG:
1932 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
1933 					 LAST_FLOW_TAG_FIELD))
1934 			return -EOPNOTSUPP;
1935 		if (ib_spec->flow_tag.tag_id >= BIT(24))
1936 			return -EINVAL;
1937 
1938 		*tag_id = ib_spec->flow_tag.tag_id;
1939 		break;
1940 	default:
1941 		return -EINVAL;
1942 	}
1943 
1944 	return 0;
1945 }
1946 
1947 /* If a flow could catch both multicast and unicast packets,
1948  * it won't fall into the multicast flow steering table and this rule
1949  * could steal other multicast packets.
1950  */
1951 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1952 {
1953 	struct ib_flow_spec_eth *eth_spec;
1954 
1955 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1956 	    ib_attr->size < sizeof(struct ib_flow_attr) +
1957 	    sizeof(struct ib_flow_spec_eth) ||
1958 	    ib_attr->num_of_specs < 1)
1959 		return false;
1960 
1961 	eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1962 	if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1963 	    eth_spec->size != sizeof(*eth_spec))
1964 		return false;
1965 
1966 	return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1967 	       is_multicast_ether_addr(eth_spec->val.dst_mac);
1968 }
1969 
1970 static bool is_valid_attr(const struct ib_flow_attr *flow_attr)
1971 {
1972 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1973 	bool has_ipv4_spec = false;
1974 	bool eth_type_ipv4 = true;
1975 	unsigned int spec_index;
1976 
1977 	/* Validate that ethertype is correct */
1978 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1979 		if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1980 		    ib_spec->eth.mask.ether_type) {
1981 			if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1982 			      ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1983 				eth_type_ipv4 = false;
1984 		} else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1985 			has_ipv4_spec = true;
1986 		}
1987 		ib_spec = (void *)ib_spec + ib_spec->size;
1988 	}
1989 	return !has_ipv4_spec || eth_type_ipv4;
1990 }
1991 
1992 static void put_flow_table(struct mlx5_ib_dev *dev,
1993 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
1994 {
1995 	prio->refcount -= !!ft_added;
1996 	if (!prio->refcount) {
1997 		mlx5_destroy_flow_table(prio->flow_table);
1998 		prio->flow_table = NULL;
1999 	}
2000 }
2001 
2002 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
2003 {
2004 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
2005 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
2006 							  struct mlx5_ib_flow_handler,
2007 							  ibflow);
2008 	struct mlx5_ib_flow_handler *iter, *tmp;
2009 
2010 	mutex_lock(&dev->flow_db.lock);
2011 
2012 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
2013 		mlx5_del_flow_rules(iter->rule);
2014 		put_flow_table(dev, iter->prio, true);
2015 		list_del(&iter->list);
2016 		kfree(iter);
2017 	}
2018 
2019 	mlx5_del_flow_rules(handler->rule);
2020 	put_flow_table(dev, handler->prio, true);
2021 	mutex_unlock(&dev->flow_db.lock);
2022 
2023 	kfree(handler);
2024 
2025 	return 0;
2026 }
2027 
2028 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
2029 {
2030 	priority *= 2;
2031 	if (!dont_trap)
2032 		priority++;
2033 	return priority;
2034 }
2035 
2036 enum flow_table_type {
2037 	MLX5_IB_FT_RX,
2038 	MLX5_IB_FT_TX
2039 };
2040 
2041 #define MLX5_FS_MAX_TYPES	 10
2042 #define MLX5_FS_MAX_ENTRIES	 32000UL
2043 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
2044 						struct ib_flow_attr *flow_attr,
2045 						enum flow_table_type ft_type)
2046 {
2047 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
2048 	struct mlx5_flow_namespace *ns = NULL;
2049 	struct mlx5_ib_flow_prio *prio;
2050 	struct mlx5_flow_table *ft;
2051 	int num_entries;
2052 	int num_groups;
2053 	int priority;
2054 	int err = 0;
2055 
2056 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2057 		if (flow_is_multicast_only(flow_attr) &&
2058 		    !dont_trap)
2059 			priority = MLX5_IB_FLOW_MCAST_PRIO;
2060 		else
2061 			priority = ib_prio_to_core_prio(flow_attr->priority,
2062 							dont_trap);
2063 		ns = mlx5_get_flow_namespace(dev->mdev,
2064 					     MLX5_FLOW_NAMESPACE_BYPASS);
2065 		num_entries = MLX5_FS_MAX_ENTRIES;
2066 		num_groups = MLX5_FS_MAX_TYPES;
2067 		prio = &dev->flow_db.prios[priority];
2068 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2069 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2070 		ns = mlx5_get_flow_namespace(dev->mdev,
2071 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
2072 		build_leftovers_ft_param(&priority,
2073 					 &num_entries,
2074 					 &num_groups);
2075 		prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
2076 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2077 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
2078 					allow_sniffer_and_nic_rx_shared_tir))
2079 			return ERR_PTR(-ENOTSUPP);
2080 
2081 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
2082 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2083 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2084 
2085 		prio = &dev->flow_db.sniffer[ft_type];
2086 		priority = 0;
2087 		num_entries = 1;
2088 		num_groups = 1;
2089 	}
2090 
2091 	if (!ns)
2092 		return ERR_PTR(-ENOTSUPP);
2093 
2094 	ft = prio->flow_table;
2095 	if (!ft) {
2096 		ft = mlx5_create_auto_grouped_flow_table(ns, priority,
2097 							 num_entries,
2098 							 num_groups,
2099 							 0, 0);
2100 
2101 		if (!IS_ERR(ft)) {
2102 			prio->refcount = 0;
2103 			prio->flow_table = ft;
2104 		} else {
2105 			err = PTR_ERR(ft);
2106 		}
2107 	}
2108 
2109 	return err ? ERR_PTR(err) : prio;
2110 }
2111 
2112 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
2113 						     struct mlx5_ib_flow_prio *ft_prio,
2114 						     const struct ib_flow_attr *flow_attr,
2115 						     struct mlx5_flow_destination *dst)
2116 {
2117 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
2118 	struct mlx5_ib_flow_handler *handler;
2119 	struct mlx5_flow_act flow_act = {0};
2120 	struct mlx5_flow_spec *spec;
2121 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
2122 	unsigned int spec_index;
2123 	u32 flow_tag = MLX5_FS_DEFAULT_FLOW_TAG;
2124 	int err = 0;
2125 
2126 	if (!is_valid_attr(flow_attr))
2127 		return ERR_PTR(-EINVAL);
2128 
2129 	spec = mlx5_vzalloc(sizeof(*spec));
2130 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
2131 	if (!handler || !spec) {
2132 		err = -ENOMEM;
2133 		goto free;
2134 	}
2135 
2136 	INIT_LIST_HEAD(&handler->list);
2137 
2138 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2139 		err = parse_flow_attr(spec->match_criteria,
2140 				      spec->match_value, ib_flow, &flow_tag);
2141 		if (err < 0)
2142 			goto free;
2143 
2144 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
2145 	}
2146 
2147 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
2148 	flow_act.action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
2149 		MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
2150 
2151 	if (flow_tag != MLX5_FS_DEFAULT_FLOW_TAG &&
2152 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2153 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
2154 		mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
2155 			     flow_tag, flow_attr->type);
2156 		err = -EINVAL;
2157 		goto free;
2158 	}
2159 	flow_act.flow_tag = flow_tag;
2160 	handler->rule = mlx5_add_flow_rules(ft, spec,
2161 					    &flow_act,
2162 					    dst, 1);
2163 
2164 	if (IS_ERR(handler->rule)) {
2165 		err = PTR_ERR(handler->rule);
2166 		goto free;
2167 	}
2168 
2169 	ft_prio->refcount++;
2170 	handler->prio = ft_prio;
2171 
2172 	ft_prio->flow_table = ft;
2173 free:
2174 	if (err)
2175 		kfree(handler);
2176 	kvfree(spec);
2177 	return err ? ERR_PTR(err) : handler;
2178 }
2179 
2180 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
2181 							  struct mlx5_ib_flow_prio *ft_prio,
2182 							  struct ib_flow_attr *flow_attr,
2183 							  struct mlx5_flow_destination *dst)
2184 {
2185 	struct mlx5_ib_flow_handler *handler_dst = NULL;
2186 	struct mlx5_ib_flow_handler *handler = NULL;
2187 
2188 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
2189 	if (!IS_ERR(handler)) {
2190 		handler_dst = create_flow_rule(dev, ft_prio,
2191 					       flow_attr, dst);
2192 		if (IS_ERR(handler_dst)) {
2193 			mlx5_del_flow_rules(handler->rule);
2194 			ft_prio->refcount--;
2195 			kfree(handler);
2196 			handler = handler_dst;
2197 		} else {
2198 			list_add(&handler_dst->list, &handler->list);
2199 		}
2200 	}
2201 
2202 	return handler;
2203 }
2204 enum {
2205 	LEFTOVERS_MC,
2206 	LEFTOVERS_UC,
2207 };
2208 
2209 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
2210 							  struct mlx5_ib_flow_prio *ft_prio,
2211 							  struct ib_flow_attr *flow_attr,
2212 							  struct mlx5_flow_destination *dst)
2213 {
2214 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
2215 	struct mlx5_ib_flow_handler *handler = NULL;
2216 
2217 	static struct {
2218 		struct ib_flow_attr	flow_attr;
2219 		struct ib_flow_spec_eth eth_flow;
2220 	} leftovers_specs[] = {
2221 		[LEFTOVERS_MC] = {
2222 			.flow_attr = {
2223 				.num_of_specs = 1,
2224 				.size = sizeof(leftovers_specs[0])
2225 			},
2226 			.eth_flow = {
2227 				.type = IB_FLOW_SPEC_ETH,
2228 				.size = sizeof(struct ib_flow_spec_eth),
2229 				.mask = {.dst_mac = {0x1} },
2230 				.val =  {.dst_mac = {0x1} }
2231 			}
2232 		},
2233 		[LEFTOVERS_UC] = {
2234 			.flow_attr = {
2235 				.num_of_specs = 1,
2236 				.size = sizeof(leftovers_specs[0])
2237 			},
2238 			.eth_flow = {
2239 				.type = IB_FLOW_SPEC_ETH,
2240 				.size = sizeof(struct ib_flow_spec_eth),
2241 				.mask = {.dst_mac = {0x1} },
2242 				.val = {.dst_mac = {} }
2243 			}
2244 		}
2245 	};
2246 
2247 	handler = create_flow_rule(dev, ft_prio,
2248 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
2249 				   dst);
2250 	if (!IS_ERR(handler) &&
2251 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2252 		handler_ucast = create_flow_rule(dev, ft_prio,
2253 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
2254 						 dst);
2255 		if (IS_ERR(handler_ucast)) {
2256 			mlx5_del_flow_rules(handler->rule);
2257 			ft_prio->refcount--;
2258 			kfree(handler);
2259 			handler = handler_ucast;
2260 		} else {
2261 			list_add(&handler_ucast->list, &handler->list);
2262 		}
2263 	}
2264 
2265 	return handler;
2266 }
2267 
2268 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2269 							struct mlx5_ib_flow_prio *ft_rx,
2270 							struct mlx5_ib_flow_prio *ft_tx,
2271 							struct mlx5_flow_destination *dst)
2272 {
2273 	struct mlx5_ib_flow_handler *handler_rx;
2274 	struct mlx5_ib_flow_handler *handler_tx;
2275 	int err;
2276 	static const struct ib_flow_attr flow_attr  = {
2277 		.num_of_specs = 0,
2278 		.size = sizeof(flow_attr)
2279 	};
2280 
2281 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2282 	if (IS_ERR(handler_rx)) {
2283 		err = PTR_ERR(handler_rx);
2284 		goto err;
2285 	}
2286 
2287 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2288 	if (IS_ERR(handler_tx)) {
2289 		err = PTR_ERR(handler_tx);
2290 		goto err_tx;
2291 	}
2292 
2293 	list_add(&handler_tx->list, &handler_rx->list);
2294 
2295 	return handler_rx;
2296 
2297 err_tx:
2298 	mlx5_del_flow_rules(handler_rx->rule);
2299 	ft_rx->refcount--;
2300 	kfree(handler_rx);
2301 err:
2302 	return ERR_PTR(err);
2303 }
2304 
2305 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2306 					   struct ib_flow_attr *flow_attr,
2307 					   int domain)
2308 {
2309 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2310 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2311 	struct mlx5_ib_flow_handler *handler = NULL;
2312 	struct mlx5_flow_destination *dst = NULL;
2313 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2314 	struct mlx5_ib_flow_prio *ft_prio;
2315 	int err;
2316 
2317 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2318 		return ERR_PTR(-ENOSPC);
2319 
2320 	if (domain != IB_FLOW_DOMAIN_USER ||
2321 	    flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2322 	    (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2323 		return ERR_PTR(-EINVAL);
2324 
2325 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2326 	if (!dst)
2327 		return ERR_PTR(-ENOMEM);
2328 
2329 	mutex_lock(&dev->flow_db.lock);
2330 
2331 	ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2332 	if (IS_ERR(ft_prio)) {
2333 		err = PTR_ERR(ft_prio);
2334 		goto unlock;
2335 	}
2336 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2337 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2338 		if (IS_ERR(ft_prio_tx)) {
2339 			err = PTR_ERR(ft_prio_tx);
2340 			ft_prio_tx = NULL;
2341 			goto destroy_ft;
2342 		}
2343 	}
2344 
2345 	dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2346 	if (mqp->flags & MLX5_IB_QP_RSS)
2347 		dst->tir_num = mqp->rss_qp.tirn;
2348 	else
2349 		dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2350 
2351 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2352 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
2353 			handler = create_dont_trap_rule(dev, ft_prio,
2354 							flow_attr, dst);
2355 		} else {
2356 			handler = create_flow_rule(dev, ft_prio, flow_attr,
2357 						   dst);
2358 		}
2359 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2360 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2361 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2362 						dst);
2363 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2364 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2365 	} else {
2366 		err = -EINVAL;
2367 		goto destroy_ft;
2368 	}
2369 
2370 	if (IS_ERR(handler)) {
2371 		err = PTR_ERR(handler);
2372 		handler = NULL;
2373 		goto destroy_ft;
2374 	}
2375 
2376 	mutex_unlock(&dev->flow_db.lock);
2377 	kfree(dst);
2378 
2379 	return &handler->ibflow;
2380 
2381 destroy_ft:
2382 	put_flow_table(dev, ft_prio, false);
2383 	if (ft_prio_tx)
2384 		put_flow_table(dev, ft_prio_tx, false);
2385 unlock:
2386 	mutex_unlock(&dev->flow_db.lock);
2387 	kfree(dst);
2388 	kfree(handler);
2389 	return ERR_PTR(err);
2390 }
2391 
2392 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2393 {
2394 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2395 	int err;
2396 
2397 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2398 	if (err)
2399 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2400 			     ibqp->qp_num, gid->raw);
2401 
2402 	return err;
2403 }
2404 
2405 static int mlx5_ib_mcg_detach(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_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2411 	if (err)
2412 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2413 			     ibqp->qp_num, gid->raw);
2414 
2415 	return err;
2416 }
2417 
2418 static int init_node_data(struct mlx5_ib_dev *dev)
2419 {
2420 	int err;
2421 
2422 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2423 	if (err)
2424 		return err;
2425 
2426 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2427 
2428 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2429 }
2430 
2431 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2432 			     char *buf)
2433 {
2434 	struct mlx5_ib_dev *dev =
2435 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2436 
2437 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
2438 }
2439 
2440 static ssize_t show_reg_pages(struct device *device,
2441 			      struct device_attribute *attr, char *buf)
2442 {
2443 	struct mlx5_ib_dev *dev =
2444 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2445 
2446 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2447 }
2448 
2449 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2450 			char *buf)
2451 {
2452 	struct mlx5_ib_dev *dev =
2453 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2454 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2455 }
2456 
2457 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2458 			char *buf)
2459 {
2460 	struct mlx5_ib_dev *dev =
2461 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2462 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
2463 }
2464 
2465 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2466 			  char *buf)
2467 {
2468 	struct mlx5_ib_dev *dev =
2469 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2470 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2471 		       dev->mdev->board_id);
2472 }
2473 
2474 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2475 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2476 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2477 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2478 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2479 
2480 static struct device_attribute *mlx5_class_attributes[] = {
2481 	&dev_attr_hw_rev,
2482 	&dev_attr_hca_type,
2483 	&dev_attr_board_id,
2484 	&dev_attr_fw_pages,
2485 	&dev_attr_reg_pages,
2486 };
2487 
2488 static void pkey_change_handler(struct work_struct *work)
2489 {
2490 	struct mlx5_ib_port_resources *ports =
2491 		container_of(work, struct mlx5_ib_port_resources,
2492 			     pkey_change_work);
2493 
2494 	mutex_lock(&ports->devr->mutex);
2495 	mlx5_ib_gsi_pkey_change(ports->gsi);
2496 	mutex_unlock(&ports->devr->mutex);
2497 }
2498 
2499 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2500 {
2501 	struct mlx5_ib_qp *mqp;
2502 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2503 	struct mlx5_core_cq *mcq;
2504 	struct list_head cq_armed_list;
2505 	unsigned long flags_qp;
2506 	unsigned long flags_cq;
2507 	unsigned long flags;
2508 
2509 	INIT_LIST_HEAD(&cq_armed_list);
2510 
2511 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2512 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2513 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2514 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2515 		if (mqp->sq.tail != mqp->sq.head) {
2516 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2517 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2518 			if (send_mcq->mcq.comp &&
2519 			    mqp->ibqp.send_cq->comp_handler) {
2520 				if (!send_mcq->mcq.reset_notify_added) {
2521 					send_mcq->mcq.reset_notify_added = 1;
2522 					list_add_tail(&send_mcq->mcq.reset_notify,
2523 						      &cq_armed_list);
2524 				}
2525 			}
2526 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2527 		}
2528 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2529 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2530 		/* no handling is needed for SRQ */
2531 		if (!mqp->ibqp.srq) {
2532 			if (mqp->rq.tail != mqp->rq.head) {
2533 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2534 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2535 				if (recv_mcq->mcq.comp &&
2536 				    mqp->ibqp.recv_cq->comp_handler) {
2537 					if (!recv_mcq->mcq.reset_notify_added) {
2538 						recv_mcq->mcq.reset_notify_added = 1;
2539 						list_add_tail(&recv_mcq->mcq.reset_notify,
2540 							      &cq_armed_list);
2541 					}
2542 				}
2543 				spin_unlock_irqrestore(&recv_mcq->lock,
2544 						       flags_cq);
2545 			}
2546 		}
2547 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2548 	}
2549 	/*At that point all inflight post send were put to be executed as of we
2550 	 * lock/unlock above locks Now need to arm all involved CQs.
2551 	 */
2552 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2553 		mcq->comp(mcq);
2554 	}
2555 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2556 }
2557 
2558 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2559 			  enum mlx5_dev_event event, unsigned long param)
2560 {
2561 	struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2562 	struct ib_event ibev;
2563 	bool fatal = false;
2564 	u8 port = 0;
2565 
2566 	switch (event) {
2567 	case MLX5_DEV_EVENT_SYS_ERROR:
2568 		ibev.event = IB_EVENT_DEVICE_FATAL;
2569 		mlx5_ib_handle_internal_error(ibdev);
2570 		fatal = true;
2571 		break;
2572 
2573 	case MLX5_DEV_EVENT_PORT_UP:
2574 	case MLX5_DEV_EVENT_PORT_DOWN:
2575 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
2576 		port = (u8)param;
2577 
2578 		/* In RoCE, port up/down events are handled in
2579 		 * mlx5_netdev_event().
2580 		 */
2581 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2582 			IB_LINK_LAYER_ETHERNET)
2583 			return;
2584 
2585 		ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2586 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2587 		break;
2588 
2589 	case MLX5_DEV_EVENT_LID_CHANGE:
2590 		ibev.event = IB_EVENT_LID_CHANGE;
2591 		port = (u8)param;
2592 		break;
2593 
2594 	case MLX5_DEV_EVENT_PKEY_CHANGE:
2595 		ibev.event = IB_EVENT_PKEY_CHANGE;
2596 		port = (u8)param;
2597 
2598 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2599 		break;
2600 
2601 	case MLX5_DEV_EVENT_GUID_CHANGE:
2602 		ibev.event = IB_EVENT_GID_CHANGE;
2603 		port = (u8)param;
2604 		break;
2605 
2606 	case MLX5_DEV_EVENT_CLIENT_REREG:
2607 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
2608 		port = (u8)param;
2609 		break;
2610 	default:
2611 		return;
2612 	}
2613 
2614 	ibev.device	      = &ibdev->ib_dev;
2615 	ibev.element.port_num = port;
2616 
2617 	if (port < 1 || port > ibdev->num_ports) {
2618 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2619 		return;
2620 	}
2621 
2622 	if (ibdev->ib_active)
2623 		ib_dispatch_event(&ibev);
2624 
2625 	if (fatal)
2626 		ibdev->ib_active = false;
2627 }
2628 
2629 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2630 {
2631 	struct mlx5_hca_vport_context vport_ctx;
2632 	int err;
2633 	int port;
2634 
2635 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2636 		dev->mdev->port_caps[port - 1].has_smi = false;
2637 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2638 		    MLX5_CAP_PORT_TYPE_IB) {
2639 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2640 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
2641 								   port, 0,
2642 								   &vport_ctx);
2643 				if (err) {
2644 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2645 						    port, err);
2646 					return err;
2647 				}
2648 				dev->mdev->port_caps[port - 1].has_smi =
2649 					vport_ctx.has_smi;
2650 			} else {
2651 				dev->mdev->port_caps[port - 1].has_smi = true;
2652 			}
2653 		}
2654 	}
2655 	return 0;
2656 }
2657 
2658 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2659 {
2660 	int port;
2661 
2662 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2663 		mlx5_query_ext_port_caps(dev, port);
2664 }
2665 
2666 static int get_port_caps(struct mlx5_ib_dev *dev)
2667 {
2668 	struct ib_device_attr *dprops = NULL;
2669 	struct ib_port_attr *pprops = NULL;
2670 	int err = -ENOMEM;
2671 	int port;
2672 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2673 
2674 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2675 	if (!pprops)
2676 		goto out;
2677 
2678 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2679 	if (!dprops)
2680 		goto out;
2681 
2682 	err = set_has_smi_cap(dev);
2683 	if (err)
2684 		goto out;
2685 
2686 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2687 	if (err) {
2688 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2689 		goto out;
2690 	}
2691 
2692 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2693 		memset(pprops, 0, sizeof(*pprops));
2694 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2695 		if (err) {
2696 			mlx5_ib_warn(dev, "query_port %d failed %d\n",
2697 				     port, err);
2698 			break;
2699 		}
2700 		dev->mdev->port_caps[port - 1].pkey_table_len =
2701 						dprops->max_pkeys;
2702 		dev->mdev->port_caps[port - 1].gid_table_len =
2703 						pprops->gid_tbl_len;
2704 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2705 			    dprops->max_pkeys, pprops->gid_tbl_len);
2706 	}
2707 
2708 out:
2709 	kfree(pprops);
2710 	kfree(dprops);
2711 
2712 	return err;
2713 }
2714 
2715 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2716 {
2717 	int err;
2718 
2719 	err = mlx5_mr_cache_cleanup(dev);
2720 	if (err)
2721 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2722 
2723 	mlx5_ib_destroy_qp(dev->umrc.qp);
2724 	ib_free_cq(dev->umrc.cq);
2725 	ib_dealloc_pd(dev->umrc.pd);
2726 }
2727 
2728 enum {
2729 	MAX_UMR_WR = 128,
2730 };
2731 
2732 static int create_umr_res(struct mlx5_ib_dev *dev)
2733 {
2734 	struct ib_qp_init_attr *init_attr = NULL;
2735 	struct ib_qp_attr *attr = NULL;
2736 	struct ib_pd *pd;
2737 	struct ib_cq *cq;
2738 	struct ib_qp *qp;
2739 	int ret;
2740 
2741 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2742 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2743 	if (!attr || !init_attr) {
2744 		ret = -ENOMEM;
2745 		goto error_0;
2746 	}
2747 
2748 	pd = ib_alloc_pd(&dev->ib_dev, 0);
2749 	if (IS_ERR(pd)) {
2750 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2751 		ret = PTR_ERR(pd);
2752 		goto error_0;
2753 	}
2754 
2755 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2756 	if (IS_ERR(cq)) {
2757 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2758 		ret = PTR_ERR(cq);
2759 		goto error_2;
2760 	}
2761 
2762 	init_attr->send_cq = cq;
2763 	init_attr->recv_cq = cq;
2764 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2765 	init_attr->cap.max_send_wr = MAX_UMR_WR;
2766 	init_attr->cap.max_send_sge = 1;
2767 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2768 	init_attr->port_num = 1;
2769 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2770 	if (IS_ERR(qp)) {
2771 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2772 		ret = PTR_ERR(qp);
2773 		goto error_3;
2774 	}
2775 	qp->device     = &dev->ib_dev;
2776 	qp->real_qp    = qp;
2777 	qp->uobject    = NULL;
2778 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2779 
2780 	attr->qp_state = IB_QPS_INIT;
2781 	attr->port_num = 1;
2782 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2783 				IB_QP_PORT, NULL);
2784 	if (ret) {
2785 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2786 		goto error_4;
2787 	}
2788 
2789 	memset(attr, 0, sizeof(*attr));
2790 	attr->qp_state = IB_QPS_RTR;
2791 	attr->path_mtu = IB_MTU_256;
2792 
2793 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2794 	if (ret) {
2795 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2796 		goto error_4;
2797 	}
2798 
2799 	memset(attr, 0, sizeof(*attr));
2800 	attr->qp_state = IB_QPS_RTS;
2801 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2802 	if (ret) {
2803 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2804 		goto error_4;
2805 	}
2806 
2807 	dev->umrc.qp = qp;
2808 	dev->umrc.cq = cq;
2809 	dev->umrc.pd = pd;
2810 
2811 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
2812 	ret = mlx5_mr_cache_init(dev);
2813 	if (ret) {
2814 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2815 		goto error_4;
2816 	}
2817 
2818 	kfree(attr);
2819 	kfree(init_attr);
2820 
2821 	return 0;
2822 
2823 error_4:
2824 	mlx5_ib_destroy_qp(qp);
2825 
2826 error_3:
2827 	ib_free_cq(cq);
2828 
2829 error_2:
2830 	ib_dealloc_pd(pd);
2831 
2832 error_0:
2833 	kfree(attr);
2834 	kfree(init_attr);
2835 	return ret;
2836 }
2837 
2838 static int create_dev_resources(struct mlx5_ib_resources *devr)
2839 {
2840 	struct ib_srq_init_attr attr;
2841 	struct mlx5_ib_dev *dev;
2842 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2843 	int port;
2844 	int ret = 0;
2845 
2846 	dev = container_of(devr, struct mlx5_ib_dev, devr);
2847 
2848 	mutex_init(&devr->mutex);
2849 
2850 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2851 	if (IS_ERR(devr->p0)) {
2852 		ret = PTR_ERR(devr->p0);
2853 		goto error0;
2854 	}
2855 	devr->p0->device  = &dev->ib_dev;
2856 	devr->p0->uobject = NULL;
2857 	atomic_set(&devr->p0->usecnt, 0);
2858 
2859 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
2860 	if (IS_ERR(devr->c0)) {
2861 		ret = PTR_ERR(devr->c0);
2862 		goto error1;
2863 	}
2864 	devr->c0->device        = &dev->ib_dev;
2865 	devr->c0->uobject       = NULL;
2866 	devr->c0->comp_handler  = NULL;
2867 	devr->c0->event_handler = NULL;
2868 	devr->c0->cq_context    = NULL;
2869 	atomic_set(&devr->c0->usecnt, 0);
2870 
2871 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2872 	if (IS_ERR(devr->x0)) {
2873 		ret = PTR_ERR(devr->x0);
2874 		goto error2;
2875 	}
2876 	devr->x0->device = &dev->ib_dev;
2877 	devr->x0->inode = NULL;
2878 	atomic_set(&devr->x0->usecnt, 0);
2879 	mutex_init(&devr->x0->tgt_qp_mutex);
2880 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2881 
2882 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2883 	if (IS_ERR(devr->x1)) {
2884 		ret = PTR_ERR(devr->x1);
2885 		goto error3;
2886 	}
2887 	devr->x1->device = &dev->ib_dev;
2888 	devr->x1->inode = NULL;
2889 	atomic_set(&devr->x1->usecnt, 0);
2890 	mutex_init(&devr->x1->tgt_qp_mutex);
2891 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2892 
2893 	memset(&attr, 0, sizeof(attr));
2894 	attr.attr.max_sge = 1;
2895 	attr.attr.max_wr = 1;
2896 	attr.srq_type = IB_SRQT_XRC;
2897 	attr.ext.xrc.cq = devr->c0;
2898 	attr.ext.xrc.xrcd = devr->x0;
2899 
2900 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2901 	if (IS_ERR(devr->s0)) {
2902 		ret = PTR_ERR(devr->s0);
2903 		goto error4;
2904 	}
2905 	devr->s0->device	= &dev->ib_dev;
2906 	devr->s0->pd		= devr->p0;
2907 	devr->s0->uobject       = NULL;
2908 	devr->s0->event_handler = NULL;
2909 	devr->s0->srq_context   = NULL;
2910 	devr->s0->srq_type      = IB_SRQT_XRC;
2911 	devr->s0->ext.xrc.xrcd	= devr->x0;
2912 	devr->s0->ext.xrc.cq	= devr->c0;
2913 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2914 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2915 	atomic_inc(&devr->p0->usecnt);
2916 	atomic_set(&devr->s0->usecnt, 0);
2917 
2918 	memset(&attr, 0, sizeof(attr));
2919 	attr.attr.max_sge = 1;
2920 	attr.attr.max_wr = 1;
2921 	attr.srq_type = IB_SRQT_BASIC;
2922 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2923 	if (IS_ERR(devr->s1)) {
2924 		ret = PTR_ERR(devr->s1);
2925 		goto error5;
2926 	}
2927 	devr->s1->device	= &dev->ib_dev;
2928 	devr->s1->pd		= devr->p0;
2929 	devr->s1->uobject       = NULL;
2930 	devr->s1->event_handler = NULL;
2931 	devr->s1->srq_context   = NULL;
2932 	devr->s1->srq_type      = IB_SRQT_BASIC;
2933 	devr->s1->ext.xrc.cq	= devr->c0;
2934 	atomic_inc(&devr->p0->usecnt);
2935 	atomic_set(&devr->s0->usecnt, 0);
2936 
2937 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2938 		INIT_WORK(&devr->ports[port].pkey_change_work,
2939 			  pkey_change_handler);
2940 		devr->ports[port].devr = devr;
2941 	}
2942 
2943 	return 0;
2944 
2945 error5:
2946 	mlx5_ib_destroy_srq(devr->s0);
2947 error4:
2948 	mlx5_ib_dealloc_xrcd(devr->x1);
2949 error3:
2950 	mlx5_ib_dealloc_xrcd(devr->x0);
2951 error2:
2952 	mlx5_ib_destroy_cq(devr->c0);
2953 error1:
2954 	mlx5_ib_dealloc_pd(devr->p0);
2955 error0:
2956 	return ret;
2957 }
2958 
2959 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2960 {
2961 	struct mlx5_ib_dev *dev =
2962 		container_of(devr, struct mlx5_ib_dev, devr);
2963 	int port;
2964 
2965 	mlx5_ib_destroy_srq(devr->s1);
2966 	mlx5_ib_destroy_srq(devr->s0);
2967 	mlx5_ib_dealloc_xrcd(devr->x0);
2968 	mlx5_ib_dealloc_xrcd(devr->x1);
2969 	mlx5_ib_destroy_cq(devr->c0);
2970 	mlx5_ib_dealloc_pd(devr->p0);
2971 
2972 	/* Make sure no change P_Key work items are still executing */
2973 	for (port = 0; port < dev->num_ports; ++port)
2974 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2975 }
2976 
2977 static u32 get_core_cap_flags(struct ib_device *ibdev)
2978 {
2979 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2980 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2981 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2982 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2983 	u32 ret = 0;
2984 
2985 	if (ll == IB_LINK_LAYER_INFINIBAND)
2986 		return RDMA_CORE_PORT_IBA_IB;
2987 
2988 	ret = RDMA_CORE_PORT_RAW_PACKET;
2989 
2990 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2991 		return ret;
2992 
2993 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2994 		return ret;
2995 
2996 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2997 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2998 
2999 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
3000 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3001 
3002 	return ret;
3003 }
3004 
3005 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
3006 			       struct ib_port_immutable *immutable)
3007 {
3008 	struct ib_port_attr attr;
3009 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3010 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
3011 	int err;
3012 
3013 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
3014 
3015 	err = ib_query_port(ibdev, port_num, &attr);
3016 	if (err)
3017 		return err;
3018 
3019 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3020 	immutable->gid_tbl_len = attr.gid_tbl_len;
3021 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
3022 	if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
3023 		immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3024 
3025 	return 0;
3026 }
3027 
3028 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
3029 			   size_t str_len)
3030 {
3031 	struct mlx5_ib_dev *dev =
3032 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3033 	snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
3034 		       fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
3035 }
3036 
3037 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3038 {
3039 	struct mlx5_core_dev *mdev = dev->mdev;
3040 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3041 								 MLX5_FLOW_NAMESPACE_LAG);
3042 	struct mlx5_flow_table *ft;
3043 	int err;
3044 
3045 	if (!ns || !mlx5_lag_is_active(mdev))
3046 		return 0;
3047 
3048 	err = mlx5_cmd_create_vport_lag(mdev);
3049 	if (err)
3050 		return err;
3051 
3052 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3053 	if (IS_ERR(ft)) {
3054 		err = PTR_ERR(ft);
3055 		goto err_destroy_vport_lag;
3056 	}
3057 
3058 	dev->flow_db.lag_demux_ft = ft;
3059 	return 0;
3060 
3061 err_destroy_vport_lag:
3062 	mlx5_cmd_destroy_vport_lag(mdev);
3063 	return err;
3064 }
3065 
3066 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3067 {
3068 	struct mlx5_core_dev *mdev = dev->mdev;
3069 
3070 	if (dev->flow_db.lag_demux_ft) {
3071 		mlx5_destroy_flow_table(dev->flow_db.lag_demux_ft);
3072 		dev->flow_db.lag_demux_ft = NULL;
3073 
3074 		mlx5_cmd_destroy_vport_lag(mdev);
3075 	}
3076 }
3077 
3078 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev)
3079 {
3080 	int err;
3081 
3082 	dev->roce.nb.notifier_call = mlx5_netdev_event;
3083 	err = register_netdevice_notifier(&dev->roce.nb);
3084 	if (err) {
3085 		dev->roce.nb.notifier_call = NULL;
3086 		return err;
3087 	}
3088 
3089 	return 0;
3090 }
3091 
3092 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev)
3093 {
3094 	if (dev->roce.nb.notifier_call) {
3095 		unregister_netdevice_notifier(&dev->roce.nb);
3096 		dev->roce.nb.notifier_call = NULL;
3097 	}
3098 }
3099 
3100 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3101 {
3102 	int err;
3103 
3104 	err = mlx5_add_netdev_notifier(dev);
3105 	if (err)
3106 		return err;
3107 
3108 	if (MLX5_CAP_GEN(dev->mdev, roce)) {
3109 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3110 		if (err)
3111 			goto err_unregister_netdevice_notifier;
3112 	}
3113 
3114 	err = mlx5_eth_lag_init(dev);
3115 	if (err)
3116 		goto err_disable_roce;
3117 
3118 	return 0;
3119 
3120 err_disable_roce:
3121 	if (MLX5_CAP_GEN(dev->mdev, roce))
3122 		mlx5_nic_vport_disable_roce(dev->mdev);
3123 
3124 err_unregister_netdevice_notifier:
3125 	mlx5_remove_netdev_notifier(dev);
3126 	return err;
3127 }
3128 
3129 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3130 {
3131 	mlx5_eth_lag_cleanup(dev);
3132 	if (MLX5_CAP_GEN(dev->mdev, roce))
3133 		mlx5_nic_vport_disable_roce(dev->mdev);
3134 }
3135 
3136 struct mlx5_ib_q_counter {
3137 	const char *name;
3138 	size_t offset;
3139 };
3140 
3141 #define INIT_Q_COUNTER(_name)		\
3142 	{ .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
3143 
3144 static const struct mlx5_ib_q_counter basic_q_cnts[] = {
3145 	INIT_Q_COUNTER(rx_write_requests),
3146 	INIT_Q_COUNTER(rx_read_requests),
3147 	INIT_Q_COUNTER(rx_atomic_requests),
3148 	INIT_Q_COUNTER(out_of_buffer),
3149 };
3150 
3151 static const struct mlx5_ib_q_counter out_of_seq_q_cnts[] = {
3152 	INIT_Q_COUNTER(out_of_sequence),
3153 };
3154 
3155 static const struct mlx5_ib_q_counter retrans_q_cnts[] = {
3156 	INIT_Q_COUNTER(duplicate_request),
3157 	INIT_Q_COUNTER(rnr_nak_retry_err),
3158 	INIT_Q_COUNTER(packet_seq_err),
3159 	INIT_Q_COUNTER(implied_nak_seq_err),
3160 	INIT_Q_COUNTER(local_ack_timeout_err),
3161 };
3162 
3163 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
3164 {
3165 	unsigned int i;
3166 
3167 	for (i = 0; i < dev->num_ports; i++) {
3168 		mlx5_core_dealloc_q_counter(dev->mdev,
3169 					    dev->port[i].q_cnts.set_id);
3170 		kfree(dev->port[i].q_cnts.names);
3171 		kfree(dev->port[i].q_cnts.offsets);
3172 	}
3173 }
3174 
3175 static int __mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev,
3176 				      const char ***names,
3177 				      size_t **offsets,
3178 				      u32 *num)
3179 {
3180 	u32 num_counters;
3181 
3182 	num_counters = ARRAY_SIZE(basic_q_cnts);
3183 
3184 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
3185 		num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
3186 
3187 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
3188 		num_counters += ARRAY_SIZE(retrans_q_cnts);
3189 
3190 	*names = kcalloc(num_counters, sizeof(**names), GFP_KERNEL);
3191 	if (!*names)
3192 		return -ENOMEM;
3193 
3194 	*offsets = kcalloc(num_counters, sizeof(**offsets), GFP_KERNEL);
3195 	if (!*offsets)
3196 		goto err_names;
3197 
3198 	*num = num_counters;
3199 
3200 	return 0;
3201 
3202 err_names:
3203 	kfree(*names);
3204 	return -ENOMEM;
3205 }
3206 
3207 static void mlx5_ib_fill_q_counters(struct mlx5_ib_dev *dev,
3208 				    const char **names,
3209 				    size_t *offsets)
3210 {
3211 	int i;
3212 	int j = 0;
3213 
3214 	for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
3215 		names[j] = basic_q_cnts[i].name;
3216 		offsets[j] = basic_q_cnts[i].offset;
3217 	}
3218 
3219 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
3220 		for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
3221 			names[j] = out_of_seq_q_cnts[i].name;
3222 			offsets[j] = out_of_seq_q_cnts[i].offset;
3223 		}
3224 	}
3225 
3226 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3227 		for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
3228 			names[j] = retrans_q_cnts[i].name;
3229 			offsets[j] = retrans_q_cnts[i].offset;
3230 		}
3231 	}
3232 }
3233 
3234 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
3235 {
3236 	int i;
3237 	int ret;
3238 
3239 	for (i = 0; i < dev->num_ports; i++) {
3240 		struct mlx5_ib_port *port = &dev->port[i];
3241 
3242 		ret = mlx5_core_alloc_q_counter(dev->mdev,
3243 						&port->q_cnts.set_id);
3244 		if (ret) {
3245 			mlx5_ib_warn(dev,
3246 				     "couldn't allocate queue counter for port %d, err %d\n",
3247 				     i + 1, ret);
3248 			goto dealloc_counters;
3249 		}
3250 
3251 		ret = __mlx5_ib_alloc_q_counters(dev,
3252 						 &port->q_cnts.names,
3253 						 &port->q_cnts.offsets,
3254 						 &port->q_cnts.num_counters);
3255 		if (ret)
3256 			goto dealloc_counters;
3257 
3258 		mlx5_ib_fill_q_counters(dev, port->q_cnts.names,
3259 					port->q_cnts.offsets);
3260 	}
3261 
3262 	return 0;
3263 
3264 dealloc_counters:
3265 	while (--i >= 0)
3266 		mlx5_core_dealloc_q_counter(dev->mdev,
3267 					    dev->port[i].q_cnts.set_id);
3268 
3269 	return ret;
3270 }
3271 
3272 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
3273 						    u8 port_num)
3274 {
3275 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3276 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
3277 
3278 	/* We support only per port stats */
3279 	if (port_num == 0)
3280 		return NULL;
3281 
3282 	return rdma_alloc_hw_stats_struct(port->q_cnts.names,
3283 					  port->q_cnts.num_counters,
3284 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
3285 }
3286 
3287 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
3288 				struct rdma_hw_stats *stats,
3289 				u8 port_num, int index)
3290 {
3291 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3292 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
3293 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
3294 	void *out;
3295 	__be32 val;
3296 	int ret;
3297 	int i;
3298 
3299 	if (!stats)
3300 		return -ENOSYS;
3301 
3302 	out = mlx5_vzalloc(outlen);
3303 	if (!out)
3304 		return -ENOMEM;
3305 
3306 	ret = mlx5_core_query_q_counter(dev->mdev,
3307 					port->q_cnts.set_id, 0,
3308 					out, outlen);
3309 	if (ret)
3310 		goto free;
3311 
3312 	for (i = 0; i < port->q_cnts.num_counters; i++) {
3313 		val = *(__be32 *)(out + port->q_cnts.offsets[i]);
3314 		stats->value[i] = (u64)be32_to_cpu(val);
3315 	}
3316 
3317 free:
3318 	kvfree(out);
3319 	return port->q_cnts.num_counters;
3320 }
3321 
3322 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
3323 {
3324 	struct mlx5_ib_dev *dev;
3325 	enum rdma_link_layer ll;
3326 	int port_type_cap;
3327 	const char *name;
3328 	int err;
3329 	int i;
3330 
3331 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3332 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3333 
3334 	printk_once(KERN_INFO "%s", mlx5_version);
3335 
3336 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
3337 	if (!dev)
3338 		return NULL;
3339 
3340 	dev->mdev = mdev;
3341 
3342 	dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
3343 			    GFP_KERNEL);
3344 	if (!dev->port)
3345 		goto err_dealloc;
3346 
3347 	rwlock_init(&dev->roce.netdev_lock);
3348 	err = get_port_caps(dev);
3349 	if (err)
3350 		goto err_free_port;
3351 
3352 	if (mlx5_use_mad_ifc(dev))
3353 		get_ext_port_caps(dev);
3354 
3355 	if (!mlx5_lag_is_active(mdev))
3356 		name = "mlx5_%d";
3357 	else
3358 		name = "mlx5_bond_%d";
3359 
3360 	strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
3361 	dev->ib_dev.owner		= THIS_MODULE;
3362 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
3363 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
3364 	dev->num_ports		= MLX5_CAP_GEN(mdev, num_ports);
3365 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
3366 	dev->ib_dev.num_comp_vectors    =
3367 		dev->mdev->priv.eq_table.num_comp_vectors;
3368 	dev->ib_dev.dev.parent		= &mdev->pdev->dev;
3369 
3370 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
3371 	dev->ib_dev.uverbs_cmd_mask	=
3372 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
3373 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
3374 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
3375 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
3376 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
3377 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
3378 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
3379 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
3380 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
3381 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
3382 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
3383 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
3384 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
3385 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
3386 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
3387 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
3388 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
3389 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
3390 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
3391 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
3392 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
3393 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
3394 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
3395 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
3396 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
3397 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
3398 	dev->ib_dev.uverbs_ex_cmd_mask =
3399 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
3400 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
3401 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
3402 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP);
3403 
3404 	dev->ib_dev.query_device	= mlx5_ib_query_device;
3405 	dev->ib_dev.query_port		= mlx5_ib_query_port;
3406 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
3407 	if (ll == IB_LINK_LAYER_ETHERNET)
3408 		dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
3409 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
3410 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
3411 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
3412 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
3413 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
3414 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
3415 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
3416 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
3417 	dev->ib_dev.mmap		= mlx5_ib_mmap;
3418 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
3419 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
3420 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
3421 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
3422 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
3423 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
3424 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
3425 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
3426 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
3427 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
3428 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
3429 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
3430 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
3431 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
3432 	dev->ib_dev.post_send		= mlx5_ib_post_send;
3433 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
3434 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
3435 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
3436 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
3437 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
3438 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
3439 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
3440 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
3441 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
3442 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
3443 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
3444 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
3445 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
3446 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
3447 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
3448 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
3449 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
3450 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
3451 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
3452 	if (mlx5_core_is_pf(mdev)) {
3453 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
3454 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
3455 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
3456 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
3457 	}
3458 
3459 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
3460 
3461 	mlx5_ib_internal_fill_odp_caps(dev);
3462 
3463 	if (MLX5_CAP_GEN(mdev, imaicl)) {
3464 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
3465 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
3466 		dev->ib_dev.uverbs_cmd_mask |=
3467 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
3468 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
3469 	}
3470 
3471 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
3472 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
3473 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
3474 	}
3475 
3476 	if (MLX5_CAP_GEN(mdev, xrc)) {
3477 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
3478 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
3479 		dev->ib_dev.uverbs_cmd_mask |=
3480 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
3481 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
3482 	}
3483 
3484 	if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
3485 	    IB_LINK_LAYER_ETHERNET) {
3486 		dev->ib_dev.create_flow	= mlx5_ib_create_flow;
3487 		dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
3488 		dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
3489 		dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
3490 		dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
3491 		dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
3492 		dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
3493 		dev->ib_dev.uverbs_ex_cmd_mask |=
3494 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
3495 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
3496 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
3497 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
3498 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
3499 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
3500 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
3501 	}
3502 	err = init_node_data(dev);
3503 	if (err)
3504 		goto err_free_port;
3505 
3506 	mutex_init(&dev->flow_db.lock);
3507 	mutex_init(&dev->cap_mask_mutex);
3508 	INIT_LIST_HEAD(&dev->qp_list);
3509 	spin_lock_init(&dev->reset_flow_resource_lock);
3510 
3511 	if (ll == IB_LINK_LAYER_ETHERNET) {
3512 		err = mlx5_enable_eth(dev);
3513 		if (err)
3514 			goto err_free_port;
3515 	}
3516 
3517 	err = create_dev_resources(&dev->devr);
3518 	if (err)
3519 		goto err_disable_eth;
3520 
3521 	err = mlx5_ib_odp_init_one(dev);
3522 	if (err)
3523 		goto err_rsrc;
3524 
3525 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
3526 		err = mlx5_ib_alloc_q_counters(dev);
3527 		if (err)
3528 			goto err_odp;
3529 	}
3530 
3531 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
3532 	if (!dev->mdev->priv.uar)
3533 		goto err_q_cnt;
3534 
3535 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
3536 	if (err)
3537 		goto err_uar_page;
3538 
3539 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
3540 	if (err)
3541 		goto err_bfreg;
3542 
3543 	err = ib_register_device(&dev->ib_dev, NULL);
3544 	if (err)
3545 		goto err_fp_bfreg;
3546 
3547 	err = create_umr_res(dev);
3548 	if (err)
3549 		goto err_dev;
3550 
3551 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
3552 		err = device_create_file(&dev->ib_dev.dev,
3553 					 mlx5_class_attributes[i]);
3554 		if (err)
3555 			goto err_umrc;
3556 	}
3557 
3558 	dev->ib_active = true;
3559 
3560 	return dev;
3561 
3562 err_umrc:
3563 	destroy_umrc_res(dev);
3564 
3565 err_dev:
3566 	ib_unregister_device(&dev->ib_dev);
3567 
3568 err_fp_bfreg:
3569 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3570 
3571 err_bfreg:
3572 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3573 
3574 err_uar_page:
3575 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
3576 
3577 err_q_cnt:
3578 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
3579 		mlx5_ib_dealloc_q_counters(dev);
3580 
3581 err_odp:
3582 	mlx5_ib_odp_remove_one(dev);
3583 
3584 err_rsrc:
3585 	destroy_dev_resources(&dev->devr);
3586 
3587 err_disable_eth:
3588 	if (ll == IB_LINK_LAYER_ETHERNET) {
3589 		mlx5_disable_eth(dev);
3590 		mlx5_remove_netdev_notifier(dev);
3591 	}
3592 
3593 err_free_port:
3594 	kfree(dev->port);
3595 
3596 err_dealloc:
3597 	ib_dealloc_device((struct ib_device *)dev);
3598 
3599 	return NULL;
3600 }
3601 
3602 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3603 {
3604 	struct mlx5_ib_dev *dev = context;
3605 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3606 
3607 	mlx5_remove_netdev_notifier(dev);
3608 	ib_unregister_device(&dev->ib_dev);
3609 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3610 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3611 	mlx5_put_uars_page(dev->mdev, mdev->priv.uar);
3612 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
3613 		mlx5_ib_dealloc_q_counters(dev);
3614 	destroy_umrc_res(dev);
3615 	mlx5_ib_odp_remove_one(dev);
3616 	destroy_dev_resources(&dev->devr);
3617 	if (ll == IB_LINK_LAYER_ETHERNET)
3618 		mlx5_disable_eth(dev);
3619 	kfree(dev->port);
3620 	ib_dealloc_device(&dev->ib_dev);
3621 }
3622 
3623 static struct mlx5_interface mlx5_ib_interface = {
3624 	.add            = mlx5_ib_add,
3625 	.remove         = mlx5_ib_remove,
3626 	.event          = mlx5_ib_event,
3627 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
3628 	.pfault		= mlx5_ib_pfault,
3629 #endif
3630 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
3631 };
3632 
3633 static int __init mlx5_ib_init(void)
3634 {
3635 	int err;
3636 
3637 	mlx5_ib_odp_init();
3638 
3639 	err = mlx5_register_interface(&mlx5_ib_interface);
3640 
3641 	return err;
3642 }
3643 
3644 static void __exit mlx5_ib_cleanup(void)
3645 {
3646 	mlx5_unregister_interface(&mlx5_ib_interface);
3647 }
3648 
3649 module_init(mlx5_ib_init);
3650 module_exit(mlx5_ib_cleanup);
3651