xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 82e6fdd6)
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/debugfs.h>
34 #include <linux/highmem.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/errno.h>
38 #include <linux/pci.h>
39 #include <linux/dma-mapping.h>
40 #include <linux/slab.h>
41 #if defined(CONFIG_X86)
42 #include <asm/pat.h>
43 #endif
44 #include <linux/sched.h>
45 #include <linux/sched/mm.h>
46 #include <linux/sched/task.h>
47 #include <linux/delay.h>
48 #include <rdma/ib_user_verbs.h>
49 #include <rdma/ib_addr.h>
50 #include <rdma/ib_cache.h>
51 #include <linux/mlx5/port.h>
52 #include <linux/mlx5/vport.h>
53 #include <linux/mlx5/fs.h>
54 #include <linux/list.h>
55 #include <rdma/ib_smi.h>
56 #include <rdma/ib_umem.h>
57 #include <linux/in.h>
58 #include <linux/etherdevice.h>
59 #include "mlx5_ib.h"
60 #include "cmd.h"
61 
62 #define DRIVER_NAME "mlx5_ib"
63 #define DRIVER_VERSION "5.0-0"
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 
69 static char mlx5_version[] =
70 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
71 	DRIVER_VERSION "\n";
72 
73 struct mlx5_ib_event_work {
74 	struct work_struct	work;
75 	struct mlx5_core_dev	*dev;
76 	void			*context;
77 	enum mlx5_dev_event	event;
78 	unsigned long		param;
79 };
80 
81 enum {
82 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
83 };
84 
85 static struct workqueue_struct *mlx5_ib_event_wq;
86 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
87 static LIST_HEAD(mlx5_ib_dev_list);
88 /*
89  * This mutex should be held when accessing either of the above lists
90  */
91 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
92 
93 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
94 {
95 	struct mlx5_ib_dev *dev;
96 
97 	mutex_lock(&mlx5_ib_multiport_mutex);
98 	dev = mpi->ibdev;
99 	mutex_unlock(&mlx5_ib_multiport_mutex);
100 	return dev;
101 }
102 
103 static enum rdma_link_layer
104 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
105 {
106 	switch (port_type_cap) {
107 	case MLX5_CAP_PORT_TYPE_IB:
108 		return IB_LINK_LAYER_INFINIBAND;
109 	case MLX5_CAP_PORT_TYPE_ETH:
110 		return IB_LINK_LAYER_ETHERNET;
111 	default:
112 		return IB_LINK_LAYER_UNSPECIFIED;
113 	}
114 }
115 
116 static enum rdma_link_layer
117 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
118 {
119 	struct mlx5_ib_dev *dev = to_mdev(device);
120 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
121 
122 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
123 }
124 
125 static int get_port_state(struct ib_device *ibdev,
126 			  u8 port_num,
127 			  enum ib_port_state *state)
128 {
129 	struct ib_port_attr attr;
130 	int ret;
131 
132 	memset(&attr, 0, sizeof(attr));
133 	ret = mlx5_ib_query_port(ibdev, port_num, &attr);
134 	if (!ret)
135 		*state = attr.state;
136 	return ret;
137 }
138 
139 static int mlx5_netdev_event(struct notifier_block *this,
140 			     unsigned long event, void *ptr)
141 {
142 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
143 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
144 	u8 port_num = roce->native_port_num;
145 	struct mlx5_core_dev *mdev;
146 	struct mlx5_ib_dev *ibdev;
147 
148 	ibdev = roce->dev;
149 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
150 	if (!mdev)
151 		return NOTIFY_DONE;
152 
153 	switch (event) {
154 	case NETDEV_REGISTER:
155 	case NETDEV_UNREGISTER:
156 		write_lock(&roce->netdev_lock);
157 
158 		if (ndev->dev.parent == &mdev->pdev->dev)
159 			roce->netdev = (event == NETDEV_UNREGISTER) ?
160 					NULL : ndev;
161 		write_unlock(&roce->netdev_lock);
162 		break;
163 
164 	case NETDEV_CHANGE:
165 	case NETDEV_UP:
166 	case NETDEV_DOWN: {
167 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
168 		struct net_device *upper = NULL;
169 
170 		if (lag_ndev) {
171 			upper = netdev_master_upper_dev_get(lag_ndev);
172 			dev_put(lag_ndev);
173 		}
174 
175 		if ((upper == ndev || (!upper && ndev == roce->netdev))
176 		    && ibdev->ib_active) {
177 			struct ib_event ibev = { };
178 			enum ib_port_state port_state;
179 
180 			if (get_port_state(&ibdev->ib_dev, port_num,
181 					   &port_state))
182 				goto done;
183 
184 			if (roce->last_port_state == port_state)
185 				goto done;
186 
187 			roce->last_port_state = port_state;
188 			ibev.device = &ibdev->ib_dev;
189 			if (port_state == IB_PORT_DOWN)
190 				ibev.event = IB_EVENT_PORT_ERR;
191 			else if (port_state == IB_PORT_ACTIVE)
192 				ibev.event = IB_EVENT_PORT_ACTIVE;
193 			else
194 				goto done;
195 
196 			ibev.element.port_num = port_num;
197 			ib_dispatch_event(&ibev);
198 		}
199 		break;
200 	}
201 
202 	default:
203 		break;
204 	}
205 done:
206 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
207 	return NOTIFY_DONE;
208 }
209 
210 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
211 					     u8 port_num)
212 {
213 	struct mlx5_ib_dev *ibdev = to_mdev(device);
214 	struct net_device *ndev;
215 	struct mlx5_core_dev *mdev;
216 
217 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
218 	if (!mdev)
219 		return NULL;
220 
221 	ndev = mlx5_lag_get_roce_netdev(mdev);
222 	if (ndev)
223 		goto out;
224 
225 	/* Ensure ndev does not disappear before we invoke dev_hold()
226 	 */
227 	read_lock(&ibdev->roce[port_num - 1].netdev_lock);
228 	ndev = ibdev->roce[port_num - 1].netdev;
229 	if (ndev)
230 		dev_hold(ndev);
231 	read_unlock(&ibdev->roce[port_num - 1].netdev_lock);
232 
233 out:
234 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
235 	return ndev;
236 }
237 
238 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
239 						   u8 ib_port_num,
240 						   u8 *native_port_num)
241 {
242 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
243 							  ib_port_num);
244 	struct mlx5_core_dev *mdev = NULL;
245 	struct mlx5_ib_multiport_info *mpi;
246 	struct mlx5_ib_port *port;
247 
248 	if (native_port_num)
249 		*native_port_num = 1;
250 
251 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
252 		return ibdev->mdev;
253 
254 	port = &ibdev->port[ib_port_num - 1];
255 	if (!port)
256 		return NULL;
257 
258 	spin_lock(&port->mp.mpi_lock);
259 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
260 	if (mpi && !mpi->unaffiliate) {
261 		mdev = mpi->mdev;
262 		/* If it's the master no need to refcount, it'll exist
263 		 * as long as the ib_dev exists.
264 		 */
265 		if (!mpi->is_master)
266 			mpi->mdev_refcnt++;
267 	}
268 	spin_unlock(&port->mp.mpi_lock);
269 
270 	return mdev;
271 }
272 
273 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u8 port_num)
274 {
275 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
276 							  port_num);
277 	struct mlx5_ib_multiport_info *mpi;
278 	struct mlx5_ib_port *port;
279 
280 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
281 		return;
282 
283 	port = &ibdev->port[port_num - 1];
284 
285 	spin_lock(&port->mp.mpi_lock);
286 	mpi = ibdev->port[port_num - 1].mp.mpi;
287 	if (mpi->is_master)
288 		goto out;
289 
290 	mpi->mdev_refcnt--;
291 	if (mpi->unaffiliate)
292 		complete(&mpi->unref_comp);
293 out:
294 	spin_unlock(&port->mp.mpi_lock);
295 }
296 
297 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
298 				    u8 *active_width)
299 {
300 	switch (eth_proto_oper) {
301 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
302 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
303 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
304 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
305 		*active_width = IB_WIDTH_1X;
306 		*active_speed = IB_SPEED_SDR;
307 		break;
308 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
309 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
310 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
311 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
312 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
313 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
314 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
315 		*active_width = IB_WIDTH_1X;
316 		*active_speed = IB_SPEED_QDR;
317 		break;
318 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
319 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
320 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
321 		*active_width = IB_WIDTH_1X;
322 		*active_speed = IB_SPEED_EDR;
323 		break;
324 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
325 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
326 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
327 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
328 		*active_width = IB_WIDTH_4X;
329 		*active_speed = IB_SPEED_QDR;
330 		break;
331 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
332 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
333 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
334 		*active_width = IB_WIDTH_1X;
335 		*active_speed = IB_SPEED_HDR;
336 		break;
337 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
338 		*active_width = IB_WIDTH_4X;
339 		*active_speed = IB_SPEED_FDR;
340 		break;
341 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
342 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
343 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
344 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
345 		*active_width = IB_WIDTH_4X;
346 		*active_speed = IB_SPEED_EDR;
347 		break;
348 	default:
349 		return -EINVAL;
350 	}
351 
352 	return 0;
353 }
354 
355 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
356 				struct ib_port_attr *props)
357 {
358 	struct mlx5_ib_dev *dev = to_mdev(device);
359 	struct mlx5_core_dev *mdev;
360 	struct net_device *ndev, *upper;
361 	enum ib_mtu ndev_ib_mtu;
362 	bool put_mdev = true;
363 	u16 qkey_viol_cntr;
364 	u32 eth_prot_oper;
365 	u8 mdev_port_num;
366 	int err;
367 
368 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
369 	if (!mdev) {
370 		/* This means the port isn't affiliated yet. Get the
371 		 * info for the master port instead.
372 		 */
373 		put_mdev = false;
374 		mdev = dev->mdev;
375 		mdev_port_num = 1;
376 		port_num = 1;
377 	}
378 
379 	/* Possible bad flows are checked before filling out props so in case
380 	 * of an error it will still be zeroed out.
381 	 */
382 	err = mlx5_query_port_eth_proto_oper(mdev, &eth_prot_oper,
383 					     mdev_port_num);
384 	if (err)
385 		goto out;
386 
387 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
388 				 &props->active_width);
389 
390 	props->port_cap_flags  |= IB_PORT_CM_SUP;
391 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
392 
393 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
394 						roce_address_table_size);
395 	props->max_mtu          = IB_MTU_4096;
396 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
397 	props->pkey_tbl_len     = 1;
398 	props->state            = IB_PORT_DOWN;
399 	props->phys_state       = 3;
400 
401 	mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
402 	props->qkey_viol_cntr = qkey_viol_cntr;
403 
404 	/* If this is a stub query for an unaffiliated port stop here */
405 	if (!put_mdev)
406 		goto out;
407 
408 	ndev = mlx5_ib_get_netdev(device, port_num);
409 	if (!ndev)
410 		goto out;
411 
412 	if (mlx5_lag_is_active(dev->mdev)) {
413 		rcu_read_lock();
414 		upper = netdev_master_upper_dev_get_rcu(ndev);
415 		if (upper) {
416 			dev_put(ndev);
417 			ndev = upper;
418 			dev_hold(ndev);
419 		}
420 		rcu_read_unlock();
421 	}
422 
423 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
424 		props->state      = IB_PORT_ACTIVE;
425 		props->phys_state = 5;
426 	}
427 
428 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
429 
430 	dev_put(ndev);
431 
432 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
433 out:
434 	if (put_mdev)
435 		mlx5_ib_put_native_port_mdev(dev, port_num);
436 	return err;
437 }
438 
439 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
440 			 unsigned int index, const union ib_gid *gid,
441 			 const struct ib_gid_attr *attr)
442 {
443 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
444 	u8 roce_version = 0;
445 	u8 roce_l3_type = 0;
446 	bool vlan = false;
447 	u8 mac[ETH_ALEN];
448 	u16 vlan_id = 0;
449 
450 	if (gid) {
451 		gid_type = attr->gid_type;
452 		ether_addr_copy(mac, attr->ndev->dev_addr);
453 
454 		if (is_vlan_dev(attr->ndev)) {
455 			vlan = true;
456 			vlan_id = vlan_dev_vlan_id(attr->ndev);
457 		}
458 	}
459 
460 	switch (gid_type) {
461 	case IB_GID_TYPE_IB:
462 		roce_version = MLX5_ROCE_VERSION_1;
463 		break;
464 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
465 		roce_version = MLX5_ROCE_VERSION_2;
466 		if (ipv6_addr_v4mapped((void *)gid))
467 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
468 		else
469 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
470 		break;
471 
472 	default:
473 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
474 	}
475 
476 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
477 				      roce_l3_type, gid->raw, mac, vlan,
478 				      vlan_id, port_num);
479 }
480 
481 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
482 			   unsigned int index, const union ib_gid *gid,
483 			   const struct ib_gid_attr *attr,
484 			   __always_unused void **context)
485 {
486 	return set_roce_addr(to_mdev(device), port_num, index, gid, attr);
487 }
488 
489 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
490 			   unsigned int index, __always_unused void **context)
491 {
492 	return set_roce_addr(to_mdev(device), port_num, index, NULL, NULL);
493 }
494 
495 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
496 			       int index)
497 {
498 	struct ib_gid_attr attr;
499 	union ib_gid gid;
500 
501 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
502 		return 0;
503 
504 	if (!attr.ndev)
505 		return 0;
506 
507 	dev_put(attr.ndev);
508 
509 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
510 		return 0;
511 
512 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
513 }
514 
515 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
516 			   int index, enum ib_gid_type *gid_type)
517 {
518 	struct ib_gid_attr attr;
519 	union ib_gid gid;
520 	int ret;
521 
522 	ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
523 	if (ret)
524 		return ret;
525 
526 	if (!attr.ndev)
527 		return -ENODEV;
528 
529 	dev_put(attr.ndev);
530 
531 	*gid_type = attr.gid_type;
532 
533 	return 0;
534 }
535 
536 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
537 {
538 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
539 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
540 	return 0;
541 }
542 
543 enum {
544 	MLX5_VPORT_ACCESS_METHOD_MAD,
545 	MLX5_VPORT_ACCESS_METHOD_HCA,
546 	MLX5_VPORT_ACCESS_METHOD_NIC,
547 };
548 
549 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
550 {
551 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
552 		return MLX5_VPORT_ACCESS_METHOD_MAD;
553 
554 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
555 	    IB_LINK_LAYER_ETHERNET)
556 		return MLX5_VPORT_ACCESS_METHOD_NIC;
557 
558 	return MLX5_VPORT_ACCESS_METHOD_HCA;
559 }
560 
561 static void get_atomic_caps(struct mlx5_ib_dev *dev,
562 			    u8 atomic_size_qp,
563 			    struct ib_device_attr *props)
564 {
565 	u8 tmp;
566 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
567 	u8 atomic_req_8B_endianness_mode =
568 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
569 
570 	/* Check if HW supports 8 bytes standard atomic operations and capable
571 	 * of host endianness respond
572 	 */
573 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
574 	if (((atomic_operations & tmp) == tmp) &&
575 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
576 	    (atomic_req_8B_endianness_mode)) {
577 		props->atomic_cap = IB_ATOMIC_HCA;
578 	} else {
579 		props->atomic_cap = IB_ATOMIC_NONE;
580 	}
581 }
582 
583 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
584 			       struct ib_device_attr *props)
585 {
586 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
587 
588 	get_atomic_caps(dev, atomic_size_qp, props);
589 }
590 
591 static void get_atomic_caps_dc(struct mlx5_ib_dev *dev,
592 			       struct ib_device_attr *props)
593 {
594 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
595 
596 	get_atomic_caps(dev, atomic_size_qp, props);
597 }
598 
599 bool mlx5_ib_dc_atomic_is_supported(struct mlx5_ib_dev *dev)
600 {
601 	struct ib_device_attr props = {};
602 
603 	get_atomic_caps_dc(dev, &props);
604 	return (props.atomic_cap == IB_ATOMIC_HCA) ? true : false;
605 }
606 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
607 					__be64 *sys_image_guid)
608 {
609 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
610 	struct mlx5_core_dev *mdev = dev->mdev;
611 	u64 tmp;
612 	int err;
613 
614 	switch (mlx5_get_vport_access_method(ibdev)) {
615 	case MLX5_VPORT_ACCESS_METHOD_MAD:
616 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
617 							    sys_image_guid);
618 
619 	case MLX5_VPORT_ACCESS_METHOD_HCA:
620 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
621 		break;
622 
623 	case MLX5_VPORT_ACCESS_METHOD_NIC:
624 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
625 		break;
626 
627 	default:
628 		return -EINVAL;
629 	}
630 
631 	if (!err)
632 		*sys_image_guid = cpu_to_be64(tmp);
633 
634 	return err;
635 
636 }
637 
638 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
639 				u16 *max_pkeys)
640 {
641 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
642 	struct mlx5_core_dev *mdev = dev->mdev;
643 
644 	switch (mlx5_get_vport_access_method(ibdev)) {
645 	case MLX5_VPORT_ACCESS_METHOD_MAD:
646 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
647 
648 	case MLX5_VPORT_ACCESS_METHOD_HCA:
649 	case MLX5_VPORT_ACCESS_METHOD_NIC:
650 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
651 						pkey_table_size));
652 		return 0;
653 
654 	default:
655 		return -EINVAL;
656 	}
657 }
658 
659 static int mlx5_query_vendor_id(struct ib_device *ibdev,
660 				u32 *vendor_id)
661 {
662 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
663 
664 	switch (mlx5_get_vport_access_method(ibdev)) {
665 	case MLX5_VPORT_ACCESS_METHOD_MAD:
666 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
667 
668 	case MLX5_VPORT_ACCESS_METHOD_HCA:
669 	case MLX5_VPORT_ACCESS_METHOD_NIC:
670 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
671 
672 	default:
673 		return -EINVAL;
674 	}
675 }
676 
677 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
678 				__be64 *node_guid)
679 {
680 	u64 tmp;
681 	int err;
682 
683 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
684 	case MLX5_VPORT_ACCESS_METHOD_MAD:
685 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
686 
687 	case MLX5_VPORT_ACCESS_METHOD_HCA:
688 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
689 		break;
690 
691 	case MLX5_VPORT_ACCESS_METHOD_NIC:
692 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
693 		break;
694 
695 	default:
696 		return -EINVAL;
697 	}
698 
699 	if (!err)
700 		*node_guid = cpu_to_be64(tmp);
701 
702 	return err;
703 }
704 
705 struct mlx5_reg_node_desc {
706 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
707 };
708 
709 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
710 {
711 	struct mlx5_reg_node_desc in;
712 
713 	if (mlx5_use_mad_ifc(dev))
714 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
715 
716 	memset(&in, 0, sizeof(in));
717 
718 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
719 				    sizeof(struct mlx5_reg_node_desc),
720 				    MLX5_REG_NODE_DESC, 0, 0);
721 }
722 
723 static int mlx5_ib_query_device(struct ib_device *ibdev,
724 				struct ib_device_attr *props,
725 				struct ib_udata *uhw)
726 {
727 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
728 	struct mlx5_core_dev *mdev = dev->mdev;
729 	int err = -ENOMEM;
730 	int max_sq_desc;
731 	int max_rq_sg;
732 	int max_sq_sg;
733 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
734 	bool raw_support = !mlx5_core_mp_enabled(mdev);
735 	struct mlx5_ib_query_device_resp resp = {};
736 	size_t resp_len;
737 	u64 max_tso;
738 
739 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
740 	if (uhw->outlen && uhw->outlen < resp_len)
741 		return -EINVAL;
742 	else
743 		resp.response_length = resp_len;
744 
745 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
746 		return -EINVAL;
747 
748 	memset(props, 0, sizeof(*props));
749 	err = mlx5_query_system_image_guid(ibdev,
750 					   &props->sys_image_guid);
751 	if (err)
752 		return err;
753 
754 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
755 	if (err)
756 		return err;
757 
758 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
759 	if (err)
760 		return err;
761 
762 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
763 		(fw_rev_min(dev->mdev) << 16) |
764 		fw_rev_sub(dev->mdev);
765 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
766 		IB_DEVICE_PORT_ACTIVE_EVENT		|
767 		IB_DEVICE_SYS_IMAGE_GUID		|
768 		IB_DEVICE_RC_RNR_NAK_GEN;
769 
770 	if (MLX5_CAP_GEN(mdev, pkv))
771 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
772 	if (MLX5_CAP_GEN(mdev, qkv))
773 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
774 	if (MLX5_CAP_GEN(mdev, apm))
775 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
776 	if (MLX5_CAP_GEN(mdev, xrc))
777 		props->device_cap_flags |= IB_DEVICE_XRC;
778 	if (MLX5_CAP_GEN(mdev, imaicl)) {
779 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
780 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
781 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
782 		/* We support 'Gappy' memory registration too */
783 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
784 	}
785 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
786 	if (MLX5_CAP_GEN(mdev, sho)) {
787 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
788 		/* At this stage no support for signature handover */
789 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
790 				      IB_PROT_T10DIF_TYPE_2 |
791 				      IB_PROT_T10DIF_TYPE_3;
792 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
793 				       IB_GUARD_T10DIF_CSUM;
794 	}
795 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
796 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
797 
798 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
799 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
800 			/* Legacy bit to support old userspace libraries */
801 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
802 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
803 		}
804 
805 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
806 			props->raw_packet_caps |=
807 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
808 
809 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
810 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
811 			if (max_tso) {
812 				resp.tso_caps.max_tso = 1 << max_tso;
813 				resp.tso_caps.supported_qpts |=
814 					1 << IB_QPT_RAW_PACKET;
815 				resp.response_length += sizeof(resp.tso_caps);
816 			}
817 		}
818 
819 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
820 			resp.rss_caps.rx_hash_function =
821 						MLX5_RX_HASH_FUNC_TOEPLITZ;
822 			resp.rss_caps.rx_hash_fields_mask =
823 						MLX5_RX_HASH_SRC_IPV4 |
824 						MLX5_RX_HASH_DST_IPV4 |
825 						MLX5_RX_HASH_SRC_IPV6 |
826 						MLX5_RX_HASH_DST_IPV6 |
827 						MLX5_RX_HASH_SRC_PORT_TCP |
828 						MLX5_RX_HASH_DST_PORT_TCP |
829 						MLX5_RX_HASH_SRC_PORT_UDP |
830 						MLX5_RX_HASH_DST_PORT_UDP |
831 						MLX5_RX_HASH_INNER;
832 			resp.response_length += sizeof(resp.rss_caps);
833 		}
834 	} else {
835 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
836 			resp.response_length += sizeof(resp.tso_caps);
837 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
838 			resp.response_length += sizeof(resp.rss_caps);
839 	}
840 
841 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
842 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
843 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
844 	}
845 
846 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
847 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
848 	    raw_support)
849 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
850 
851 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
852 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
853 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
854 
855 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
856 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
857 	    raw_support) {
858 		/* Legacy bit to support old userspace libraries */
859 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
860 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
861 	}
862 
863 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
864 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
865 
866 	if (MLX5_CAP_GEN(mdev, end_pad))
867 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
868 
869 	props->vendor_part_id	   = mdev->pdev->device;
870 	props->hw_ver		   = mdev->pdev->revision;
871 
872 	props->max_mr_size	   = ~0ull;
873 	props->page_size_cap	   = ~(min_page_size - 1);
874 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
875 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
876 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
877 		     sizeof(struct mlx5_wqe_data_seg);
878 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
879 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
880 		     sizeof(struct mlx5_wqe_raddr_seg)) /
881 		sizeof(struct mlx5_wqe_data_seg);
882 	props->max_sge = min(max_rq_sg, max_sq_sg);
883 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
884 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
885 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
886 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
887 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
888 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
889 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
890 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
891 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
892 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
893 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
894 	props->max_srq_sge	   = max_rq_sg - 1;
895 	props->max_fast_reg_page_list_len =
896 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
897 	get_atomic_caps_qp(dev, props);
898 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
899 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
900 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
901 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
902 					   props->max_mcast_grp;
903 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
904 	props->max_ah = INT_MAX;
905 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
906 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
907 
908 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
909 	if (MLX5_CAP_GEN(mdev, pg))
910 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
911 	props->odp_caps = dev->odp_caps;
912 #endif
913 
914 	if (MLX5_CAP_GEN(mdev, cd))
915 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
916 
917 	if (!mlx5_core_is_pf(mdev))
918 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
919 
920 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
921 	    IB_LINK_LAYER_ETHERNET && raw_support) {
922 		props->rss_caps.max_rwq_indirection_tables =
923 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
924 		props->rss_caps.max_rwq_indirection_table_size =
925 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
926 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
927 		props->max_wq_type_rq =
928 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
929 	}
930 
931 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
932 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
933 		props->tm_caps.max_num_tags =
934 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
935 		props->tm_caps.flags = IB_TM_CAP_RC;
936 		props->tm_caps.max_ops =
937 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
938 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
939 	}
940 
941 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
942 		props->cq_caps.max_cq_moderation_count =
943 						MLX5_MAX_CQ_COUNT;
944 		props->cq_caps.max_cq_moderation_period =
945 						MLX5_MAX_CQ_PERIOD;
946 	}
947 
948 	if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
949 		resp.cqe_comp_caps.max_num =
950 			MLX5_CAP_GEN(dev->mdev, cqe_compression) ?
951 			MLX5_CAP_GEN(dev->mdev, cqe_compression_max_num) : 0;
952 		resp.cqe_comp_caps.supported_format =
953 			MLX5_IB_CQE_RES_FORMAT_HASH |
954 			MLX5_IB_CQE_RES_FORMAT_CSUM;
955 		resp.response_length += sizeof(resp.cqe_comp_caps);
956 	}
957 
958 	if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen) &&
959 	    raw_support) {
960 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
961 		    MLX5_CAP_GEN(mdev, qos)) {
962 			resp.packet_pacing_caps.qp_rate_limit_max =
963 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
964 			resp.packet_pacing_caps.qp_rate_limit_min =
965 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
966 			resp.packet_pacing_caps.supported_qpts |=
967 				1 << IB_QPT_RAW_PACKET;
968 		}
969 		resp.response_length += sizeof(resp.packet_pacing_caps);
970 	}
971 
972 	if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
973 			uhw->outlen)) {
974 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
975 			resp.mlx5_ib_support_multi_pkt_send_wqes =
976 				MLX5_IB_ALLOW_MPW;
977 
978 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
979 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
980 				MLX5_IB_SUPPORT_EMPW;
981 
982 		resp.response_length +=
983 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
984 	}
985 
986 	if (field_avail(typeof(resp), flags, uhw->outlen)) {
987 		resp.response_length += sizeof(resp.flags);
988 
989 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
990 			resp.flags |=
991 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
992 
993 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
994 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
995 	}
996 
997 	if (field_avail(typeof(resp), sw_parsing_caps,
998 			uhw->outlen)) {
999 		resp.response_length += sizeof(resp.sw_parsing_caps);
1000 		if (MLX5_CAP_ETH(mdev, swp)) {
1001 			resp.sw_parsing_caps.sw_parsing_offloads |=
1002 				MLX5_IB_SW_PARSING;
1003 
1004 			if (MLX5_CAP_ETH(mdev, swp_csum))
1005 				resp.sw_parsing_caps.sw_parsing_offloads |=
1006 					MLX5_IB_SW_PARSING_CSUM;
1007 
1008 			if (MLX5_CAP_ETH(mdev, swp_lso))
1009 				resp.sw_parsing_caps.sw_parsing_offloads |=
1010 					MLX5_IB_SW_PARSING_LSO;
1011 
1012 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1013 				resp.sw_parsing_caps.supported_qpts =
1014 					BIT(IB_QPT_RAW_PACKET);
1015 		}
1016 	}
1017 
1018 	if (field_avail(typeof(resp), striding_rq_caps, uhw->outlen) &&
1019 	    raw_support) {
1020 		resp.response_length += sizeof(resp.striding_rq_caps);
1021 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1022 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1023 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1024 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1025 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1026 			resp.striding_rq_caps.min_single_wqe_log_num_of_strides =
1027 				MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1028 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1029 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1030 			resp.striding_rq_caps.supported_qpts =
1031 				BIT(IB_QPT_RAW_PACKET);
1032 		}
1033 	}
1034 
1035 	if (field_avail(typeof(resp), tunnel_offloads_caps,
1036 			uhw->outlen)) {
1037 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1038 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1039 			resp.tunnel_offloads_caps |=
1040 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1041 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1042 			resp.tunnel_offloads_caps |=
1043 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1044 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1045 			resp.tunnel_offloads_caps |=
1046 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1047 	}
1048 
1049 	if (uhw->outlen) {
1050 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1051 
1052 		if (err)
1053 			return err;
1054 	}
1055 
1056 	return 0;
1057 }
1058 
1059 enum mlx5_ib_width {
1060 	MLX5_IB_WIDTH_1X	= 1 << 0,
1061 	MLX5_IB_WIDTH_2X	= 1 << 1,
1062 	MLX5_IB_WIDTH_4X	= 1 << 2,
1063 	MLX5_IB_WIDTH_8X	= 1 << 3,
1064 	MLX5_IB_WIDTH_12X	= 1 << 4
1065 };
1066 
1067 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
1068 				  u8 *ib_width)
1069 {
1070 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1071 	int err = 0;
1072 
1073 	if (active_width & MLX5_IB_WIDTH_1X) {
1074 		*ib_width = IB_WIDTH_1X;
1075 	} else if (active_width & MLX5_IB_WIDTH_2X) {
1076 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
1077 			    (int)active_width);
1078 		err = -EINVAL;
1079 	} else if (active_width & MLX5_IB_WIDTH_4X) {
1080 		*ib_width = IB_WIDTH_4X;
1081 	} else if (active_width & MLX5_IB_WIDTH_8X) {
1082 		*ib_width = IB_WIDTH_8X;
1083 	} else if (active_width & MLX5_IB_WIDTH_12X) {
1084 		*ib_width = IB_WIDTH_12X;
1085 	} else {
1086 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
1087 			    (int)active_width);
1088 		err = -EINVAL;
1089 	}
1090 
1091 	return err;
1092 }
1093 
1094 static int mlx5_mtu_to_ib_mtu(int mtu)
1095 {
1096 	switch (mtu) {
1097 	case 256: return 1;
1098 	case 512: return 2;
1099 	case 1024: return 3;
1100 	case 2048: return 4;
1101 	case 4096: return 5;
1102 	default:
1103 		pr_warn("invalid mtu\n");
1104 		return -1;
1105 	}
1106 }
1107 
1108 enum ib_max_vl_num {
1109 	__IB_MAX_VL_0		= 1,
1110 	__IB_MAX_VL_0_1		= 2,
1111 	__IB_MAX_VL_0_3		= 3,
1112 	__IB_MAX_VL_0_7		= 4,
1113 	__IB_MAX_VL_0_14	= 5,
1114 };
1115 
1116 enum mlx5_vl_hw_cap {
1117 	MLX5_VL_HW_0	= 1,
1118 	MLX5_VL_HW_0_1	= 2,
1119 	MLX5_VL_HW_0_2	= 3,
1120 	MLX5_VL_HW_0_3	= 4,
1121 	MLX5_VL_HW_0_4	= 5,
1122 	MLX5_VL_HW_0_5	= 6,
1123 	MLX5_VL_HW_0_6	= 7,
1124 	MLX5_VL_HW_0_7	= 8,
1125 	MLX5_VL_HW_0_14	= 15
1126 };
1127 
1128 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1129 				u8 *max_vl_num)
1130 {
1131 	switch (vl_hw_cap) {
1132 	case MLX5_VL_HW_0:
1133 		*max_vl_num = __IB_MAX_VL_0;
1134 		break;
1135 	case MLX5_VL_HW_0_1:
1136 		*max_vl_num = __IB_MAX_VL_0_1;
1137 		break;
1138 	case MLX5_VL_HW_0_3:
1139 		*max_vl_num = __IB_MAX_VL_0_3;
1140 		break;
1141 	case MLX5_VL_HW_0_7:
1142 		*max_vl_num = __IB_MAX_VL_0_7;
1143 		break;
1144 	case MLX5_VL_HW_0_14:
1145 		*max_vl_num = __IB_MAX_VL_0_14;
1146 		break;
1147 
1148 	default:
1149 		return -EINVAL;
1150 	}
1151 
1152 	return 0;
1153 }
1154 
1155 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1156 			       struct ib_port_attr *props)
1157 {
1158 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1159 	struct mlx5_core_dev *mdev = dev->mdev;
1160 	struct mlx5_hca_vport_context *rep;
1161 	u16 max_mtu;
1162 	u16 oper_mtu;
1163 	int err;
1164 	u8 ib_link_width_oper;
1165 	u8 vl_hw_cap;
1166 
1167 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1168 	if (!rep) {
1169 		err = -ENOMEM;
1170 		goto out;
1171 	}
1172 
1173 	/* props being zeroed by the caller, avoid zeroing it here */
1174 
1175 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1176 	if (err)
1177 		goto out;
1178 
1179 	props->lid		= rep->lid;
1180 	props->lmc		= rep->lmc;
1181 	props->sm_lid		= rep->sm_lid;
1182 	props->sm_sl		= rep->sm_sl;
1183 	props->state		= rep->vport_state;
1184 	props->phys_state	= rep->port_physical_state;
1185 	props->port_cap_flags	= rep->cap_mask1;
1186 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1187 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1188 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1189 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1190 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1191 	props->subnet_timeout	= rep->subnet_timeout;
1192 	props->init_type_reply	= rep->init_type_reply;
1193 	props->grh_required	= rep->grh_required;
1194 
1195 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
1196 	if (err)
1197 		goto out;
1198 
1199 	err = translate_active_width(ibdev, ib_link_width_oper,
1200 				     &props->active_width);
1201 	if (err)
1202 		goto out;
1203 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
1204 	if (err)
1205 		goto out;
1206 
1207 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1208 
1209 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1210 
1211 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1212 
1213 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1214 
1215 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1216 	if (err)
1217 		goto out;
1218 
1219 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1220 				   &props->max_vl_num);
1221 out:
1222 	kfree(rep);
1223 	return err;
1224 }
1225 
1226 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1227 		       struct ib_port_attr *props)
1228 {
1229 	unsigned int count;
1230 	int ret;
1231 
1232 	switch (mlx5_get_vport_access_method(ibdev)) {
1233 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1234 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1235 		break;
1236 
1237 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1238 		ret = mlx5_query_hca_port(ibdev, port, props);
1239 		break;
1240 
1241 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1242 		ret = mlx5_query_port_roce(ibdev, port, props);
1243 		break;
1244 
1245 	default:
1246 		ret = -EINVAL;
1247 	}
1248 
1249 	if (!ret && props) {
1250 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1251 		struct mlx5_core_dev *mdev;
1252 		bool put_mdev = true;
1253 
1254 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1255 		if (!mdev) {
1256 			/* If the port isn't affiliated yet query the master.
1257 			 * The master and slave will have the same values.
1258 			 */
1259 			mdev = dev->mdev;
1260 			port = 1;
1261 			put_mdev = false;
1262 		}
1263 		count = mlx5_core_reserved_gids_count(mdev);
1264 		if (put_mdev)
1265 			mlx5_ib_put_native_port_mdev(dev, port);
1266 		props->gid_tbl_len -= count;
1267 	}
1268 	return ret;
1269 }
1270 
1271 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1272 			     union ib_gid *gid)
1273 {
1274 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1275 	struct mlx5_core_dev *mdev = dev->mdev;
1276 
1277 	switch (mlx5_get_vport_access_method(ibdev)) {
1278 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1279 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1280 
1281 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1282 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1283 
1284 	default:
1285 		return -EINVAL;
1286 	}
1287 
1288 }
1289 
1290 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1291 				   u16 index, u16 *pkey)
1292 {
1293 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1294 	struct mlx5_core_dev *mdev;
1295 	bool put_mdev = true;
1296 	u8 mdev_port_num;
1297 	int err;
1298 
1299 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1300 	if (!mdev) {
1301 		/* The port isn't affiliated yet, get the PKey from the master
1302 		 * port. For RoCE the PKey tables will be the same.
1303 		 */
1304 		put_mdev = false;
1305 		mdev = dev->mdev;
1306 		mdev_port_num = 1;
1307 	}
1308 
1309 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1310 					index, pkey);
1311 	if (put_mdev)
1312 		mlx5_ib_put_native_port_mdev(dev, port);
1313 
1314 	return err;
1315 }
1316 
1317 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1318 			      u16 *pkey)
1319 {
1320 	switch (mlx5_get_vport_access_method(ibdev)) {
1321 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1322 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1323 
1324 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1325 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1326 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1327 	default:
1328 		return -EINVAL;
1329 	}
1330 }
1331 
1332 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1333 				 struct ib_device_modify *props)
1334 {
1335 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1336 	struct mlx5_reg_node_desc in;
1337 	struct mlx5_reg_node_desc out;
1338 	int err;
1339 
1340 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1341 		return -EOPNOTSUPP;
1342 
1343 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1344 		return 0;
1345 
1346 	/*
1347 	 * If possible, pass node desc to FW, so it can generate
1348 	 * a 144 trap.  If cmd fails, just ignore.
1349 	 */
1350 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1351 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1352 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1353 	if (err)
1354 		return err;
1355 
1356 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1357 
1358 	return err;
1359 }
1360 
1361 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1362 				u32 value)
1363 {
1364 	struct mlx5_hca_vport_context ctx = {};
1365 	struct mlx5_core_dev *mdev;
1366 	u8 mdev_port_num;
1367 	int err;
1368 
1369 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1370 	if (!mdev)
1371 		return -ENODEV;
1372 
1373 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1374 	if (err)
1375 		goto out;
1376 
1377 	if (~ctx.cap_mask1_perm & mask) {
1378 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1379 			     mask, ctx.cap_mask1_perm);
1380 		err = -EINVAL;
1381 		goto out;
1382 	}
1383 
1384 	ctx.cap_mask1 = value;
1385 	ctx.cap_mask1_perm = mask;
1386 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1387 						 0, &ctx);
1388 
1389 out:
1390 	mlx5_ib_put_native_port_mdev(dev, port_num);
1391 
1392 	return err;
1393 }
1394 
1395 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1396 			       struct ib_port_modify *props)
1397 {
1398 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1399 	struct ib_port_attr attr;
1400 	u32 tmp;
1401 	int err;
1402 	u32 change_mask;
1403 	u32 value;
1404 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1405 		      IB_LINK_LAYER_INFINIBAND);
1406 
1407 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1408 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1409 	 */
1410 	if (!is_ib)
1411 		return 0;
1412 
1413 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1414 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1415 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1416 		return set_port_caps_atomic(dev, port, change_mask, value);
1417 	}
1418 
1419 	mutex_lock(&dev->cap_mask_mutex);
1420 
1421 	err = ib_query_port(ibdev, port, &attr);
1422 	if (err)
1423 		goto out;
1424 
1425 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1426 		~props->clr_port_cap_mask;
1427 
1428 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1429 
1430 out:
1431 	mutex_unlock(&dev->cap_mask_mutex);
1432 	return err;
1433 }
1434 
1435 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1436 {
1437 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1438 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1439 }
1440 
1441 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1442 {
1443 	/* Large page with non 4k uar support might limit the dynamic size */
1444 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1445 		return MLX5_MIN_DYN_BFREGS;
1446 
1447 	return MLX5_MAX_DYN_BFREGS;
1448 }
1449 
1450 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1451 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1452 			     struct mlx5_bfreg_info *bfregi)
1453 {
1454 	int uars_per_sys_page;
1455 	int bfregs_per_sys_page;
1456 	int ref_bfregs = req->total_num_bfregs;
1457 
1458 	if (req->total_num_bfregs == 0)
1459 		return -EINVAL;
1460 
1461 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1462 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1463 
1464 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1465 		return -ENOMEM;
1466 
1467 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1468 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1469 	/* This holds the required static allocation asked by the user */
1470 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1471 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1472 		return -EINVAL;
1473 
1474 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1475 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1476 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1477 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1478 
1479 	mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, allocated %d, total bfregs %d, using %d sys pages\n",
1480 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1481 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1482 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1483 		    bfregi->num_sys_pages);
1484 
1485 	return 0;
1486 }
1487 
1488 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1489 {
1490 	struct mlx5_bfreg_info *bfregi;
1491 	int err;
1492 	int i;
1493 
1494 	bfregi = &context->bfregi;
1495 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1496 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1497 		if (err)
1498 			goto error;
1499 
1500 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1501 	}
1502 
1503 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1504 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1505 
1506 	return 0;
1507 
1508 error:
1509 	for (--i; i >= 0; i--)
1510 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1511 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1512 
1513 	return err;
1514 }
1515 
1516 static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1517 {
1518 	struct mlx5_bfreg_info *bfregi;
1519 	int err;
1520 	int i;
1521 
1522 	bfregi = &context->bfregi;
1523 	for (i = 0; i < bfregi->num_sys_pages; i++) {
1524 		if (i < bfregi->num_static_sys_pages ||
1525 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX) {
1526 			err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1527 			if (err) {
1528 				mlx5_ib_warn(dev, "failed to free uar %d, err=%d\n", i, err);
1529 				return err;
1530 			}
1531 		}
1532 	}
1533 
1534 	return 0;
1535 }
1536 
1537 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn)
1538 {
1539 	int err;
1540 
1541 	err = mlx5_core_alloc_transport_domain(dev->mdev, tdn);
1542 	if (err)
1543 		return err;
1544 
1545 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1546 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1547 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1548 		return err;
1549 
1550 	mutex_lock(&dev->lb_mutex);
1551 	dev->user_td++;
1552 
1553 	if (dev->user_td == 2)
1554 		err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1555 
1556 	mutex_unlock(&dev->lb_mutex);
1557 	return err;
1558 }
1559 
1560 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn)
1561 {
1562 	mlx5_core_dealloc_transport_domain(dev->mdev, tdn);
1563 
1564 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1565 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1566 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1567 		return;
1568 
1569 	mutex_lock(&dev->lb_mutex);
1570 	dev->user_td--;
1571 
1572 	if (dev->user_td < 2)
1573 		mlx5_nic_vport_update_local_lb(dev->mdev, false);
1574 
1575 	mutex_unlock(&dev->lb_mutex);
1576 }
1577 
1578 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1579 						  struct ib_udata *udata)
1580 {
1581 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1582 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1583 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1584 	struct mlx5_core_dev *mdev = dev->mdev;
1585 	struct mlx5_ib_ucontext *context;
1586 	struct mlx5_bfreg_info *bfregi;
1587 	int ver;
1588 	int err;
1589 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1590 				     max_cqe_version);
1591 	bool lib_uar_4k;
1592 
1593 	if (!dev->ib_active)
1594 		return ERR_PTR(-EAGAIN);
1595 
1596 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1597 		ver = 0;
1598 	else if (udata->inlen >= min_req_v2)
1599 		ver = 2;
1600 	else
1601 		return ERR_PTR(-EINVAL);
1602 
1603 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1604 	if (err)
1605 		return ERR_PTR(err);
1606 
1607 	if (req.flags)
1608 		return ERR_PTR(-EINVAL);
1609 
1610 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1611 		return ERR_PTR(-EOPNOTSUPP);
1612 
1613 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1614 				    MLX5_NON_FP_BFREGS_PER_UAR);
1615 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1616 		return ERR_PTR(-EINVAL);
1617 
1618 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1619 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1620 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1621 	resp.cache_line_size = cache_line_size();
1622 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1623 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1624 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1625 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1626 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1627 	resp.cqe_version = min_t(__u8,
1628 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1629 				 req.max_cqe_version);
1630 	resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1631 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1632 	resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1633 					MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1634 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1635 				   sizeof(resp.response_length), udata->outlen);
1636 
1637 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1638 	if (!context)
1639 		return ERR_PTR(-ENOMEM);
1640 
1641 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1642 	bfregi = &context->bfregi;
1643 
1644 	/* updates req->total_num_bfregs */
1645 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1646 	if (err)
1647 		goto out_ctx;
1648 
1649 	mutex_init(&bfregi->lock);
1650 	bfregi->lib_uar_4k = lib_uar_4k;
1651 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1652 				GFP_KERNEL);
1653 	if (!bfregi->count) {
1654 		err = -ENOMEM;
1655 		goto out_ctx;
1656 	}
1657 
1658 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1659 				    sizeof(*bfregi->sys_pages),
1660 				    GFP_KERNEL);
1661 	if (!bfregi->sys_pages) {
1662 		err = -ENOMEM;
1663 		goto out_count;
1664 	}
1665 
1666 	err = allocate_uars(dev, context);
1667 	if (err)
1668 		goto out_sys_pages;
1669 
1670 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1671 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1672 #endif
1673 
1674 	context->upd_xlt_page = __get_free_page(GFP_KERNEL);
1675 	if (!context->upd_xlt_page) {
1676 		err = -ENOMEM;
1677 		goto out_uars;
1678 	}
1679 	mutex_init(&context->upd_xlt_page_mutex);
1680 
1681 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1682 		err = mlx5_ib_alloc_transport_domain(dev, &context->tdn);
1683 		if (err)
1684 			goto out_page;
1685 	}
1686 
1687 	INIT_LIST_HEAD(&context->vma_private_list);
1688 	mutex_init(&context->vma_private_list_mutex);
1689 	INIT_LIST_HEAD(&context->db_page_list);
1690 	mutex_init(&context->db_page_mutex);
1691 
1692 	resp.tot_bfregs = req.total_num_bfregs;
1693 	resp.num_ports = dev->num_ports;
1694 
1695 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1696 		resp.response_length += sizeof(resp.cqe_version);
1697 
1698 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1699 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1700 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1701 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1702 	}
1703 
1704 	if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1705 		if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1706 			mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1707 			resp.eth_min_inline++;
1708 		}
1709 		resp.response_length += sizeof(resp.eth_min_inline);
1710 	}
1711 
1712 	if (field_avail(typeof(resp), clock_info_versions, udata->outlen)) {
1713 		if (mdev->clock_info)
1714 			resp.clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1715 		resp.response_length += sizeof(resp.clock_info_versions);
1716 	}
1717 
1718 	/*
1719 	 * We don't want to expose information from the PCI bar that is located
1720 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1721 	 * pretend we don't support reading the HCA's core clock. This is also
1722 	 * forced by mmap function.
1723 	 */
1724 	if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1725 		if (PAGE_SIZE <= 4096) {
1726 			resp.comp_mask |=
1727 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1728 			resp.hca_core_clock_offset =
1729 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1730 		}
1731 		resp.response_length += sizeof(resp.hca_core_clock_offset);
1732 	}
1733 
1734 	if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1735 		resp.response_length += sizeof(resp.log_uar_size);
1736 
1737 	if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1738 		resp.response_length += sizeof(resp.num_uars_per_page);
1739 
1740 	if (field_avail(typeof(resp), num_dyn_bfregs, udata->outlen)) {
1741 		resp.num_dyn_bfregs = bfregi->num_dyn_bfregs;
1742 		resp.response_length += sizeof(resp.num_dyn_bfregs);
1743 	}
1744 
1745 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1746 	if (err)
1747 		goto out_td;
1748 
1749 	bfregi->ver = ver;
1750 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1751 	context->cqe_version = resp.cqe_version;
1752 	context->lib_caps = req.lib_caps;
1753 	print_lib_caps(dev, context->lib_caps);
1754 
1755 	return &context->ibucontext;
1756 
1757 out_td:
1758 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1759 		mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1760 
1761 out_page:
1762 	free_page(context->upd_xlt_page);
1763 
1764 out_uars:
1765 	deallocate_uars(dev, context);
1766 
1767 out_sys_pages:
1768 	kfree(bfregi->sys_pages);
1769 
1770 out_count:
1771 	kfree(bfregi->count);
1772 
1773 out_ctx:
1774 	kfree(context);
1775 
1776 	return ERR_PTR(err);
1777 }
1778 
1779 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1780 {
1781 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1782 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1783 	struct mlx5_bfreg_info *bfregi;
1784 
1785 	bfregi = &context->bfregi;
1786 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1787 		mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1788 
1789 	free_page(context->upd_xlt_page);
1790 	deallocate_uars(dev, context);
1791 	kfree(bfregi->sys_pages);
1792 	kfree(bfregi->count);
1793 	kfree(context);
1794 
1795 	return 0;
1796 }
1797 
1798 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1799 				 int uar_idx)
1800 {
1801 	int fw_uars_per_page;
1802 
1803 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1804 
1805 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
1806 }
1807 
1808 static int get_command(unsigned long offset)
1809 {
1810 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1811 }
1812 
1813 static int get_arg(unsigned long offset)
1814 {
1815 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1816 }
1817 
1818 static int get_index(unsigned long offset)
1819 {
1820 	return get_arg(offset);
1821 }
1822 
1823 /* Index resides in an extra byte to enable larger values than 255 */
1824 static int get_extended_index(unsigned long offset)
1825 {
1826 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
1827 }
1828 
1829 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1830 {
1831 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1832 	 * is done through either mremap flow or split_vma (usually due to
1833 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1834 	 * as this VMA is strongly hardware related.  Therefore we set the
1835 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1836 	 * calling us again and trying to do incorrect actions.  We assume that
1837 	 * the original VMA size is exactly a single page, and therefore all
1838 	 * "splitting" operation will not happen to it.
1839 	 */
1840 	area->vm_ops = NULL;
1841 }
1842 
1843 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1844 {
1845 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1846 
1847 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1848 	 * file itself is closed, therefore no sync is needed with the regular
1849 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1850 	 * However need a sync with accessing the vma as part of
1851 	 * mlx5_ib_disassociate_ucontext.
1852 	 * The close operation is usually called under mm->mmap_sem except when
1853 	 * process is exiting.
1854 	 * The exiting case is handled explicitly as part of
1855 	 * mlx5_ib_disassociate_ucontext.
1856 	 */
1857 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1858 
1859 	/* setting the vma context pointer to null in the mlx5_ib driver's
1860 	 * private data, to protect a race condition in
1861 	 * mlx5_ib_disassociate_ucontext().
1862 	 */
1863 	mlx5_ib_vma_priv_data->vma = NULL;
1864 	mutex_lock(mlx5_ib_vma_priv_data->vma_private_list_mutex);
1865 	list_del(&mlx5_ib_vma_priv_data->list);
1866 	mutex_unlock(mlx5_ib_vma_priv_data->vma_private_list_mutex);
1867 	kfree(mlx5_ib_vma_priv_data);
1868 }
1869 
1870 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1871 	.open = mlx5_ib_vma_open,
1872 	.close = mlx5_ib_vma_close
1873 };
1874 
1875 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1876 				struct mlx5_ib_ucontext *ctx)
1877 {
1878 	struct mlx5_ib_vma_private_data *vma_prv;
1879 	struct list_head *vma_head = &ctx->vma_private_list;
1880 
1881 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1882 	if (!vma_prv)
1883 		return -ENOMEM;
1884 
1885 	vma_prv->vma = vma;
1886 	vma_prv->vma_private_list_mutex = &ctx->vma_private_list_mutex;
1887 	vma->vm_private_data = vma_prv;
1888 	vma->vm_ops =  &mlx5_ib_vm_ops;
1889 
1890 	mutex_lock(&ctx->vma_private_list_mutex);
1891 	list_add(&vma_prv->list, vma_head);
1892 	mutex_unlock(&ctx->vma_private_list_mutex);
1893 
1894 	return 0;
1895 }
1896 
1897 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1898 {
1899 	int ret;
1900 	struct vm_area_struct *vma;
1901 	struct mlx5_ib_vma_private_data *vma_private, *n;
1902 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1903 	struct task_struct *owning_process  = NULL;
1904 	struct mm_struct   *owning_mm       = NULL;
1905 
1906 	owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1907 	if (!owning_process)
1908 		return;
1909 
1910 	owning_mm = get_task_mm(owning_process);
1911 	if (!owning_mm) {
1912 		pr_info("no mm, disassociate ucontext is pending task termination\n");
1913 		while (1) {
1914 			put_task_struct(owning_process);
1915 			usleep_range(1000, 2000);
1916 			owning_process = get_pid_task(ibcontext->tgid,
1917 						      PIDTYPE_PID);
1918 			if (!owning_process ||
1919 			    owning_process->state == TASK_DEAD) {
1920 				pr_info("disassociate ucontext done, task was terminated\n");
1921 				/* in case task was dead need to release the
1922 				 * task struct.
1923 				 */
1924 				if (owning_process)
1925 					put_task_struct(owning_process);
1926 				return;
1927 			}
1928 		}
1929 	}
1930 
1931 	/* need to protect from a race on closing the vma as part of
1932 	 * mlx5_ib_vma_close.
1933 	 */
1934 	down_write(&owning_mm->mmap_sem);
1935 	mutex_lock(&context->vma_private_list_mutex);
1936 	list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1937 				 list) {
1938 		vma = vma_private->vma;
1939 		ret = zap_vma_ptes(vma, vma->vm_start,
1940 				   PAGE_SIZE);
1941 		WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1942 		/* context going to be destroyed, should
1943 		 * not access ops any more.
1944 		 */
1945 		vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
1946 		vma->vm_ops = NULL;
1947 		list_del(&vma_private->list);
1948 		kfree(vma_private);
1949 	}
1950 	mutex_unlock(&context->vma_private_list_mutex);
1951 	up_write(&owning_mm->mmap_sem);
1952 	mmput(owning_mm);
1953 	put_task_struct(owning_process);
1954 }
1955 
1956 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1957 {
1958 	switch (cmd) {
1959 	case MLX5_IB_MMAP_WC_PAGE:
1960 		return "WC";
1961 	case MLX5_IB_MMAP_REGULAR_PAGE:
1962 		return "best effort WC";
1963 	case MLX5_IB_MMAP_NC_PAGE:
1964 		return "NC";
1965 	default:
1966 		return NULL;
1967 	}
1968 }
1969 
1970 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
1971 					struct vm_area_struct *vma,
1972 					struct mlx5_ib_ucontext *context)
1973 {
1974 	phys_addr_t pfn;
1975 	int err;
1976 
1977 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1978 		return -EINVAL;
1979 
1980 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
1981 		return -EOPNOTSUPP;
1982 
1983 	if (vma->vm_flags & VM_WRITE)
1984 		return -EPERM;
1985 
1986 	if (!dev->mdev->clock_info_page)
1987 		return -EOPNOTSUPP;
1988 
1989 	pfn = page_to_pfn(dev->mdev->clock_info_page);
1990 	err = remap_pfn_range(vma, vma->vm_start, pfn, PAGE_SIZE,
1991 			      vma->vm_page_prot);
1992 	if (err)
1993 		return err;
1994 
1995 	mlx5_ib_dbg(dev, "mapped clock info at 0x%lx, PA 0x%llx\n",
1996 		    vma->vm_start,
1997 		    (unsigned long long)pfn << PAGE_SHIFT);
1998 
1999 	return mlx5_ib_set_vma_data(vma, context);
2000 }
2001 
2002 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2003 		    struct vm_area_struct *vma,
2004 		    struct mlx5_ib_ucontext *context)
2005 {
2006 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2007 	int err;
2008 	unsigned long idx;
2009 	phys_addr_t pfn, pa;
2010 	pgprot_t prot;
2011 	u32 bfreg_dyn_idx = 0;
2012 	u32 uar_index;
2013 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2014 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2015 				bfregi->num_static_sys_pages;
2016 
2017 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2018 		return -EINVAL;
2019 
2020 	if (dyn_uar)
2021 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2022 	else
2023 		idx = get_index(vma->vm_pgoff);
2024 
2025 	if (idx >= max_valid_idx) {
2026 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2027 			     idx, max_valid_idx);
2028 		return -EINVAL;
2029 	}
2030 
2031 	switch (cmd) {
2032 	case MLX5_IB_MMAP_WC_PAGE:
2033 	case MLX5_IB_MMAP_ALLOC_WC:
2034 /* Some architectures don't support WC memory */
2035 #if defined(CONFIG_X86)
2036 		if (!pat_enabled())
2037 			return -EPERM;
2038 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
2039 			return -EPERM;
2040 #endif
2041 	/* fall through */
2042 	case MLX5_IB_MMAP_REGULAR_PAGE:
2043 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2044 		prot = pgprot_writecombine(vma->vm_page_prot);
2045 		break;
2046 	case MLX5_IB_MMAP_NC_PAGE:
2047 		prot = pgprot_noncached(vma->vm_page_prot);
2048 		break;
2049 	default:
2050 		return -EINVAL;
2051 	}
2052 
2053 	if (dyn_uar) {
2054 		int uars_per_page;
2055 
2056 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2057 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2058 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2059 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2060 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2061 			return -EINVAL;
2062 		}
2063 
2064 		mutex_lock(&bfregi->lock);
2065 		/* Fail if uar already allocated, first bfreg index of each
2066 		 * page holds its count.
2067 		 */
2068 		if (bfregi->count[bfreg_dyn_idx]) {
2069 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2070 			mutex_unlock(&bfregi->lock);
2071 			return -EINVAL;
2072 		}
2073 
2074 		bfregi->count[bfreg_dyn_idx]++;
2075 		mutex_unlock(&bfregi->lock);
2076 
2077 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2078 		if (err) {
2079 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2080 			goto free_bfreg;
2081 		}
2082 	} else {
2083 		uar_index = bfregi->sys_pages[idx];
2084 	}
2085 
2086 	pfn = uar_index2pfn(dev, uar_index);
2087 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2088 
2089 	vma->vm_page_prot = prot;
2090 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
2091 				 PAGE_SIZE, vma->vm_page_prot);
2092 	if (err) {
2093 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
2094 			    err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
2095 		err = -EAGAIN;
2096 		goto err;
2097 	}
2098 
2099 	pa = pfn << PAGE_SHIFT;
2100 	mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
2101 		    vma->vm_start, &pa);
2102 
2103 	err = mlx5_ib_set_vma_data(vma, context);
2104 	if (err)
2105 		goto err;
2106 
2107 	if (dyn_uar)
2108 		bfregi->sys_pages[idx] = uar_index;
2109 	return 0;
2110 
2111 err:
2112 	if (!dyn_uar)
2113 		return err;
2114 
2115 	mlx5_cmd_free_uar(dev->mdev, idx);
2116 
2117 free_bfreg:
2118 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2119 
2120 	return err;
2121 }
2122 
2123 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2124 {
2125 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2126 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2127 	unsigned long command;
2128 	phys_addr_t pfn;
2129 
2130 	command = get_command(vma->vm_pgoff);
2131 	switch (command) {
2132 	case MLX5_IB_MMAP_WC_PAGE:
2133 	case MLX5_IB_MMAP_NC_PAGE:
2134 	case MLX5_IB_MMAP_REGULAR_PAGE:
2135 	case MLX5_IB_MMAP_ALLOC_WC:
2136 		return uar_mmap(dev, command, vma, context);
2137 
2138 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2139 		return -ENOSYS;
2140 
2141 	case MLX5_IB_MMAP_CORE_CLOCK:
2142 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2143 			return -EINVAL;
2144 
2145 		if (vma->vm_flags & VM_WRITE)
2146 			return -EPERM;
2147 
2148 		/* Don't expose to user-space information it shouldn't have */
2149 		if (PAGE_SIZE > 4096)
2150 			return -EOPNOTSUPP;
2151 
2152 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
2153 		pfn = (dev->mdev->iseg_base +
2154 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2155 			PAGE_SHIFT;
2156 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
2157 				       PAGE_SIZE, vma->vm_page_prot))
2158 			return -EAGAIN;
2159 
2160 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
2161 			    vma->vm_start,
2162 			    (unsigned long long)pfn << PAGE_SHIFT);
2163 		break;
2164 	case MLX5_IB_MMAP_CLOCK_INFO:
2165 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2166 
2167 	default:
2168 		return -EINVAL;
2169 	}
2170 
2171 	return 0;
2172 }
2173 
2174 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
2175 				      struct ib_ucontext *context,
2176 				      struct ib_udata *udata)
2177 {
2178 	struct mlx5_ib_alloc_pd_resp resp;
2179 	struct mlx5_ib_pd *pd;
2180 	int err;
2181 
2182 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
2183 	if (!pd)
2184 		return ERR_PTR(-ENOMEM);
2185 
2186 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
2187 	if (err) {
2188 		kfree(pd);
2189 		return ERR_PTR(err);
2190 	}
2191 
2192 	if (context) {
2193 		resp.pdn = pd->pdn;
2194 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2195 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
2196 			kfree(pd);
2197 			return ERR_PTR(-EFAULT);
2198 		}
2199 	}
2200 
2201 	return &pd->ibpd;
2202 }
2203 
2204 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
2205 {
2206 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2207 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2208 
2209 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
2210 	kfree(mpd);
2211 
2212 	return 0;
2213 }
2214 
2215 enum {
2216 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
2217 	MATCH_CRITERIA_ENABLE_MISC_BIT,
2218 	MATCH_CRITERIA_ENABLE_INNER_BIT
2219 };
2220 
2221 #define HEADER_IS_ZERO(match_criteria, headers)			           \
2222 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
2223 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
2224 
2225 static u8 get_match_criteria_enable(u32 *match_criteria)
2226 {
2227 	u8 match_criteria_enable;
2228 
2229 	match_criteria_enable =
2230 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
2231 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
2232 	match_criteria_enable |=
2233 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
2234 		MATCH_CRITERIA_ENABLE_MISC_BIT;
2235 	match_criteria_enable |=
2236 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
2237 		MATCH_CRITERIA_ENABLE_INNER_BIT;
2238 
2239 	return match_criteria_enable;
2240 }
2241 
2242 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
2243 {
2244 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
2245 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
2246 }
2247 
2248 static void set_flow_label(void *misc_c, void *misc_v, u8 mask, u8 val,
2249 			   bool inner)
2250 {
2251 	if (inner) {
2252 		MLX5_SET(fte_match_set_misc,
2253 			 misc_c, inner_ipv6_flow_label, mask);
2254 		MLX5_SET(fte_match_set_misc,
2255 			 misc_v, inner_ipv6_flow_label, val);
2256 	} else {
2257 		MLX5_SET(fte_match_set_misc,
2258 			 misc_c, outer_ipv6_flow_label, mask);
2259 		MLX5_SET(fte_match_set_misc,
2260 			 misc_v, outer_ipv6_flow_label, val);
2261 	}
2262 }
2263 
2264 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
2265 {
2266 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
2267 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
2268 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
2269 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
2270 }
2271 
2272 #define LAST_ETH_FIELD vlan_tag
2273 #define LAST_IB_FIELD sl
2274 #define LAST_IPV4_FIELD tos
2275 #define LAST_IPV6_FIELD traffic_class
2276 #define LAST_TCP_UDP_FIELD src_port
2277 #define LAST_TUNNEL_FIELD tunnel_id
2278 #define LAST_FLOW_TAG_FIELD tag_id
2279 #define LAST_DROP_FIELD size
2280 
2281 /* Field is the last supported field */
2282 #define FIELDS_NOT_SUPPORTED(filter, field)\
2283 	memchr_inv((void *)&filter.field  +\
2284 		   sizeof(filter.field), 0,\
2285 		   sizeof(filter) -\
2286 		   offsetof(typeof(filter), field) -\
2287 		   sizeof(filter.field))
2288 
2289 #define IPV4_VERSION 4
2290 #define IPV6_VERSION 6
2291 static int parse_flow_attr(struct mlx5_core_dev *mdev, u32 *match_c,
2292 			   u32 *match_v, const union ib_flow_spec *ib_spec,
2293 			   u32 *tag_id, bool *is_drop)
2294 {
2295 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
2296 					   misc_parameters);
2297 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
2298 					   misc_parameters);
2299 	void *headers_c;
2300 	void *headers_v;
2301 	int match_ipv;
2302 
2303 	if (ib_spec->type & IB_FLOW_SPEC_INNER) {
2304 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2305 					 inner_headers);
2306 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2307 					 inner_headers);
2308 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2309 					ft_field_support.inner_ip_version);
2310 	} else {
2311 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2312 					 outer_headers);
2313 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2314 					 outer_headers);
2315 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2316 					ft_field_support.outer_ip_version);
2317 	}
2318 
2319 	switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2320 	case IB_FLOW_SPEC_ETH:
2321 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
2322 			return -EOPNOTSUPP;
2323 
2324 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2325 					     dmac_47_16),
2326 				ib_spec->eth.mask.dst_mac);
2327 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2328 					     dmac_47_16),
2329 				ib_spec->eth.val.dst_mac);
2330 
2331 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2332 					     smac_47_16),
2333 				ib_spec->eth.mask.src_mac);
2334 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2335 					     smac_47_16),
2336 				ib_spec->eth.val.src_mac);
2337 
2338 		if (ib_spec->eth.mask.vlan_tag) {
2339 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2340 				 cvlan_tag, 1);
2341 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2342 				 cvlan_tag, 1);
2343 
2344 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2345 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
2346 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2347 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
2348 
2349 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2350 				 first_cfi,
2351 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
2352 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2353 				 first_cfi,
2354 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
2355 
2356 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2357 				 first_prio,
2358 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
2359 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2360 				 first_prio,
2361 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
2362 		}
2363 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2364 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
2365 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2366 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
2367 		break;
2368 	case IB_FLOW_SPEC_IPV4:
2369 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
2370 			return -EOPNOTSUPP;
2371 
2372 		if (match_ipv) {
2373 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2374 				 ip_version, 0xf);
2375 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2376 				 ip_version, IPV4_VERSION);
2377 		} else {
2378 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2379 				 ethertype, 0xffff);
2380 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2381 				 ethertype, ETH_P_IP);
2382 		}
2383 
2384 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2385 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2386 		       &ib_spec->ipv4.mask.src_ip,
2387 		       sizeof(ib_spec->ipv4.mask.src_ip));
2388 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2389 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2390 		       &ib_spec->ipv4.val.src_ip,
2391 		       sizeof(ib_spec->ipv4.val.src_ip));
2392 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2393 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2394 		       &ib_spec->ipv4.mask.dst_ip,
2395 		       sizeof(ib_spec->ipv4.mask.dst_ip));
2396 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2397 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2398 		       &ib_spec->ipv4.val.dst_ip,
2399 		       sizeof(ib_spec->ipv4.val.dst_ip));
2400 
2401 		set_tos(headers_c, headers_v,
2402 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
2403 
2404 		set_proto(headers_c, headers_v,
2405 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
2406 		break;
2407 	case IB_FLOW_SPEC_IPV6:
2408 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
2409 			return -EOPNOTSUPP;
2410 
2411 		if (match_ipv) {
2412 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2413 				 ip_version, 0xf);
2414 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2415 				 ip_version, IPV6_VERSION);
2416 		} else {
2417 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2418 				 ethertype, 0xffff);
2419 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2420 				 ethertype, ETH_P_IPV6);
2421 		}
2422 
2423 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2424 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2425 		       &ib_spec->ipv6.mask.src_ip,
2426 		       sizeof(ib_spec->ipv6.mask.src_ip));
2427 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2428 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2429 		       &ib_spec->ipv6.val.src_ip,
2430 		       sizeof(ib_spec->ipv6.val.src_ip));
2431 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2432 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2433 		       &ib_spec->ipv6.mask.dst_ip,
2434 		       sizeof(ib_spec->ipv6.mask.dst_ip));
2435 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2436 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2437 		       &ib_spec->ipv6.val.dst_ip,
2438 		       sizeof(ib_spec->ipv6.val.dst_ip));
2439 
2440 		set_tos(headers_c, headers_v,
2441 			ib_spec->ipv6.mask.traffic_class,
2442 			ib_spec->ipv6.val.traffic_class);
2443 
2444 		set_proto(headers_c, headers_v,
2445 			  ib_spec->ipv6.mask.next_hdr,
2446 			  ib_spec->ipv6.val.next_hdr);
2447 
2448 		set_flow_label(misc_params_c, misc_params_v,
2449 			       ntohl(ib_spec->ipv6.mask.flow_label),
2450 			       ntohl(ib_spec->ipv6.val.flow_label),
2451 			       ib_spec->type & IB_FLOW_SPEC_INNER);
2452 
2453 		break;
2454 	case IB_FLOW_SPEC_TCP:
2455 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2456 					 LAST_TCP_UDP_FIELD))
2457 			return -EOPNOTSUPP;
2458 
2459 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2460 			 0xff);
2461 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2462 			 IPPROTO_TCP);
2463 
2464 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
2465 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2466 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
2467 			 ntohs(ib_spec->tcp_udp.val.src_port));
2468 
2469 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
2470 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2471 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
2472 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2473 		break;
2474 	case IB_FLOW_SPEC_UDP:
2475 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2476 					 LAST_TCP_UDP_FIELD))
2477 			return -EOPNOTSUPP;
2478 
2479 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2480 			 0xff);
2481 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2482 			 IPPROTO_UDP);
2483 
2484 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
2485 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2486 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
2487 			 ntohs(ib_spec->tcp_udp.val.src_port));
2488 
2489 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
2490 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2491 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
2492 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2493 		break;
2494 	case IB_FLOW_SPEC_VXLAN_TUNNEL:
2495 		if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
2496 					 LAST_TUNNEL_FIELD))
2497 			return -EOPNOTSUPP;
2498 
2499 		MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
2500 			 ntohl(ib_spec->tunnel.mask.tunnel_id));
2501 		MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
2502 			 ntohl(ib_spec->tunnel.val.tunnel_id));
2503 		break;
2504 	case IB_FLOW_SPEC_ACTION_TAG:
2505 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
2506 					 LAST_FLOW_TAG_FIELD))
2507 			return -EOPNOTSUPP;
2508 		if (ib_spec->flow_tag.tag_id >= BIT(24))
2509 			return -EINVAL;
2510 
2511 		*tag_id = ib_spec->flow_tag.tag_id;
2512 		break;
2513 	case IB_FLOW_SPEC_ACTION_DROP:
2514 		if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
2515 					 LAST_DROP_FIELD))
2516 			return -EOPNOTSUPP;
2517 		*is_drop = true;
2518 		break;
2519 	default:
2520 		return -EINVAL;
2521 	}
2522 
2523 	return 0;
2524 }
2525 
2526 /* If a flow could catch both multicast and unicast packets,
2527  * it won't fall into the multicast flow steering table and this rule
2528  * could steal other multicast packets.
2529  */
2530 static bool flow_is_multicast_only(const struct ib_flow_attr *ib_attr)
2531 {
2532 	union ib_flow_spec *flow_spec;
2533 
2534 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
2535 	    ib_attr->num_of_specs < 1)
2536 		return false;
2537 
2538 	flow_spec = (union ib_flow_spec *)(ib_attr + 1);
2539 	if (flow_spec->type == IB_FLOW_SPEC_IPV4) {
2540 		struct ib_flow_spec_ipv4 *ipv4_spec;
2541 
2542 		ipv4_spec = (struct ib_flow_spec_ipv4 *)flow_spec;
2543 		if (ipv4_is_multicast(ipv4_spec->val.dst_ip))
2544 			return true;
2545 
2546 		return false;
2547 	}
2548 
2549 	if (flow_spec->type == IB_FLOW_SPEC_ETH) {
2550 		struct ib_flow_spec_eth *eth_spec;
2551 
2552 		eth_spec = (struct ib_flow_spec_eth *)flow_spec;
2553 		return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
2554 		       is_multicast_ether_addr(eth_spec->val.dst_mac);
2555 	}
2556 
2557 	return false;
2558 }
2559 
2560 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
2561 			       const struct ib_flow_attr *flow_attr,
2562 			       bool check_inner)
2563 {
2564 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
2565 	int match_ipv = check_inner ?
2566 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2567 					ft_field_support.inner_ip_version) :
2568 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2569 					ft_field_support.outer_ip_version);
2570 	int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
2571 	bool ipv4_spec_valid, ipv6_spec_valid;
2572 	unsigned int ip_spec_type = 0;
2573 	bool has_ethertype = false;
2574 	unsigned int spec_index;
2575 	bool mask_valid = true;
2576 	u16 eth_type = 0;
2577 	bool type_valid;
2578 
2579 	/* Validate that ethertype is correct */
2580 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2581 		if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
2582 		    ib_spec->eth.mask.ether_type) {
2583 			mask_valid = (ib_spec->eth.mask.ether_type ==
2584 				      htons(0xffff));
2585 			has_ethertype = true;
2586 			eth_type = ntohs(ib_spec->eth.val.ether_type);
2587 		} else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
2588 			   (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
2589 			ip_spec_type = ib_spec->type;
2590 		}
2591 		ib_spec = (void *)ib_spec + ib_spec->size;
2592 	}
2593 
2594 	type_valid = (!has_ethertype) || (!ip_spec_type);
2595 	if (!type_valid && mask_valid) {
2596 		ipv4_spec_valid = (eth_type == ETH_P_IP) &&
2597 			(ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
2598 		ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
2599 			(ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
2600 
2601 		type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
2602 			     (((eth_type == ETH_P_MPLS_UC) ||
2603 			       (eth_type == ETH_P_MPLS_MC)) && match_ipv);
2604 	}
2605 
2606 	return type_valid;
2607 }
2608 
2609 static bool is_valid_attr(struct mlx5_core_dev *mdev,
2610 			  const struct ib_flow_attr *flow_attr)
2611 {
2612 	return is_valid_ethertype(mdev, flow_attr, false) &&
2613 	       is_valid_ethertype(mdev, flow_attr, true);
2614 }
2615 
2616 static void put_flow_table(struct mlx5_ib_dev *dev,
2617 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
2618 {
2619 	prio->refcount -= !!ft_added;
2620 	if (!prio->refcount) {
2621 		mlx5_destroy_flow_table(prio->flow_table);
2622 		prio->flow_table = NULL;
2623 	}
2624 }
2625 
2626 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
2627 {
2628 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
2629 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
2630 							  struct mlx5_ib_flow_handler,
2631 							  ibflow);
2632 	struct mlx5_ib_flow_handler *iter, *tmp;
2633 
2634 	mutex_lock(&dev->flow_db.lock);
2635 
2636 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
2637 		mlx5_del_flow_rules(iter->rule);
2638 		put_flow_table(dev, iter->prio, true);
2639 		list_del(&iter->list);
2640 		kfree(iter);
2641 	}
2642 
2643 	mlx5_del_flow_rules(handler->rule);
2644 	put_flow_table(dev, handler->prio, true);
2645 	mutex_unlock(&dev->flow_db.lock);
2646 
2647 	kfree(handler);
2648 
2649 	return 0;
2650 }
2651 
2652 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
2653 {
2654 	priority *= 2;
2655 	if (!dont_trap)
2656 		priority++;
2657 	return priority;
2658 }
2659 
2660 enum flow_table_type {
2661 	MLX5_IB_FT_RX,
2662 	MLX5_IB_FT_TX
2663 };
2664 
2665 #define MLX5_FS_MAX_TYPES	 6
2666 #define MLX5_FS_MAX_ENTRIES	 BIT(16)
2667 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
2668 						struct ib_flow_attr *flow_attr,
2669 						enum flow_table_type ft_type)
2670 {
2671 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
2672 	struct mlx5_flow_namespace *ns = NULL;
2673 	struct mlx5_ib_flow_prio *prio;
2674 	struct mlx5_flow_table *ft;
2675 	int max_table_size;
2676 	int num_entries;
2677 	int num_groups;
2678 	int priority;
2679 	int err = 0;
2680 
2681 	max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
2682 						       log_max_ft_size));
2683 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2684 		if (flow_is_multicast_only(flow_attr) &&
2685 		    !dont_trap)
2686 			priority = MLX5_IB_FLOW_MCAST_PRIO;
2687 		else
2688 			priority = ib_prio_to_core_prio(flow_attr->priority,
2689 							dont_trap);
2690 		ns = mlx5_get_flow_namespace(dev->mdev,
2691 					     MLX5_FLOW_NAMESPACE_BYPASS);
2692 		num_entries = MLX5_FS_MAX_ENTRIES;
2693 		num_groups = MLX5_FS_MAX_TYPES;
2694 		prio = &dev->flow_db.prios[priority];
2695 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2696 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2697 		ns = mlx5_get_flow_namespace(dev->mdev,
2698 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
2699 		build_leftovers_ft_param(&priority,
2700 					 &num_entries,
2701 					 &num_groups);
2702 		prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
2703 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2704 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
2705 					allow_sniffer_and_nic_rx_shared_tir))
2706 			return ERR_PTR(-ENOTSUPP);
2707 
2708 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
2709 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2710 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2711 
2712 		prio = &dev->flow_db.sniffer[ft_type];
2713 		priority = 0;
2714 		num_entries = 1;
2715 		num_groups = 1;
2716 	}
2717 
2718 	if (!ns)
2719 		return ERR_PTR(-ENOTSUPP);
2720 
2721 	if (num_entries > max_table_size)
2722 		return ERR_PTR(-ENOMEM);
2723 
2724 	ft = prio->flow_table;
2725 	if (!ft) {
2726 		ft = mlx5_create_auto_grouped_flow_table(ns, priority,
2727 							 num_entries,
2728 							 num_groups,
2729 							 0, 0);
2730 
2731 		if (!IS_ERR(ft)) {
2732 			prio->refcount = 0;
2733 			prio->flow_table = ft;
2734 		} else {
2735 			err = PTR_ERR(ft);
2736 		}
2737 	}
2738 
2739 	return err ? ERR_PTR(err) : prio;
2740 }
2741 
2742 static void set_underlay_qp(struct mlx5_ib_dev *dev,
2743 			    struct mlx5_flow_spec *spec,
2744 			    u32 underlay_qpn)
2745 {
2746 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param,
2747 					   spec->match_criteria,
2748 					   misc_parameters);
2749 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, spec->match_value,
2750 					   misc_parameters);
2751 
2752 	if (underlay_qpn &&
2753 	    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
2754 				      ft_field_support.bth_dst_qp)) {
2755 		MLX5_SET(fte_match_set_misc,
2756 			 misc_params_v, bth_dst_qp, underlay_qpn);
2757 		MLX5_SET(fte_match_set_misc,
2758 			 misc_params_c, bth_dst_qp, 0xffffff);
2759 	}
2760 }
2761 
2762 static struct mlx5_ib_flow_handler *_create_flow_rule(struct mlx5_ib_dev *dev,
2763 						      struct mlx5_ib_flow_prio *ft_prio,
2764 						      const struct ib_flow_attr *flow_attr,
2765 						      struct mlx5_flow_destination *dst,
2766 						      u32 underlay_qpn)
2767 {
2768 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
2769 	struct mlx5_ib_flow_handler *handler;
2770 	struct mlx5_flow_act flow_act = {0};
2771 	struct mlx5_flow_spec *spec;
2772 	struct mlx5_flow_destination *rule_dst = dst;
2773 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
2774 	unsigned int spec_index;
2775 	u32 flow_tag = MLX5_FS_DEFAULT_FLOW_TAG;
2776 	bool is_drop = false;
2777 	int err = 0;
2778 	int dest_num = 1;
2779 
2780 	if (!is_valid_attr(dev->mdev, flow_attr))
2781 		return ERR_PTR(-EINVAL);
2782 
2783 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
2784 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
2785 	if (!handler || !spec) {
2786 		err = -ENOMEM;
2787 		goto free;
2788 	}
2789 
2790 	INIT_LIST_HEAD(&handler->list);
2791 
2792 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2793 		err = parse_flow_attr(dev->mdev, spec->match_criteria,
2794 				      spec->match_value,
2795 				      ib_flow, &flow_tag, &is_drop);
2796 		if (err < 0)
2797 			goto free;
2798 
2799 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
2800 	}
2801 
2802 	if (!flow_is_multicast_only(flow_attr))
2803 		set_underlay_qp(dev, spec, underlay_qpn);
2804 
2805 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
2806 	if (is_drop) {
2807 		flow_act.action = MLX5_FLOW_CONTEXT_ACTION_DROP;
2808 		rule_dst = NULL;
2809 		dest_num = 0;
2810 	} else {
2811 		flow_act.action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
2812 		    MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
2813 	}
2814 
2815 	if (flow_tag != MLX5_FS_DEFAULT_FLOW_TAG &&
2816 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2817 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
2818 		mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
2819 			     flow_tag, flow_attr->type);
2820 		err = -EINVAL;
2821 		goto free;
2822 	}
2823 	flow_act.flow_tag = flow_tag;
2824 	handler->rule = mlx5_add_flow_rules(ft, spec,
2825 					    &flow_act,
2826 					    rule_dst, dest_num);
2827 
2828 	if (IS_ERR(handler->rule)) {
2829 		err = PTR_ERR(handler->rule);
2830 		goto free;
2831 	}
2832 
2833 	ft_prio->refcount++;
2834 	handler->prio = ft_prio;
2835 
2836 	ft_prio->flow_table = ft;
2837 free:
2838 	if (err)
2839 		kfree(handler);
2840 	kvfree(spec);
2841 	return err ? ERR_PTR(err) : handler;
2842 }
2843 
2844 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
2845 						     struct mlx5_ib_flow_prio *ft_prio,
2846 						     const struct ib_flow_attr *flow_attr,
2847 						     struct mlx5_flow_destination *dst)
2848 {
2849 	return _create_flow_rule(dev, ft_prio, flow_attr, dst, 0);
2850 }
2851 
2852 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
2853 							  struct mlx5_ib_flow_prio *ft_prio,
2854 							  struct ib_flow_attr *flow_attr,
2855 							  struct mlx5_flow_destination *dst)
2856 {
2857 	struct mlx5_ib_flow_handler *handler_dst = NULL;
2858 	struct mlx5_ib_flow_handler *handler = NULL;
2859 
2860 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
2861 	if (!IS_ERR(handler)) {
2862 		handler_dst = create_flow_rule(dev, ft_prio,
2863 					       flow_attr, dst);
2864 		if (IS_ERR(handler_dst)) {
2865 			mlx5_del_flow_rules(handler->rule);
2866 			ft_prio->refcount--;
2867 			kfree(handler);
2868 			handler = handler_dst;
2869 		} else {
2870 			list_add(&handler_dst->list, &handler->list);
2871 		}
2872 	}
2873 
2874 	return handler;
2875 }
2876 enum {
2877 	LEFTOVERS_MC,
2878 	LEFTOVERS_UC,
2879 };
2880 
2881 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
2882 							  struct mlx5_ib_flow_prio *ft_prio,
2883 							  struct ib_flow_attr *flow_attr,
2884 							  struct mlx5_flow_destination *dst)
2885 {
2886 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
2887 	struct mlx5_ib_flow_handler *handler = NULL;
2888 
2889 	static struct {
2890 		struct ib_flow_attr	flow_attr;
2891 		struct ib_flow_spec_eth eth_flow;
2892 	} leftovers_specs[] = {
2893 		[LEFTOVERS_MC] = {
2894 			.flow_attr = {
2895 				.num_of_specs = 1,
2896 				.size = sizeof(leftovers_specs[0])
2897 			},
2898 			.eth_flow = {
2899 				.type = IB_FLOW_SPEC_ETH,
2900 				.size = sizeof(struct ib_flow_spec_eth),
2901 				.mask = {.dst_mac = {0x1} },
2902 				.val =  {.dst_mac = {0x1} }
2903 			}
2904 		},
2905 		[LEFTOVERS_UC] = {
2906 			.flow_attr = {
2907 				.num_of_specs = 1,
2908 				.size = sizeof(leftovers_specs[0])
2909 			},
2910 			.eth_flow = {
2911 				.type = IB_FLOW_SPEC_ETH,
2912 				.size = sizeof(struct ib_flow_spec_eth),
2913 				.mask = {.dst_mac = {0x1} },
2914 				.val = {.dst_mac = {} }
2915 			}
2916 		}
2917 	};
2918 
2919 	handler = create_flow_rule(dev, ft_prio,
2920 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
2921 				   dst);
2922 	if (!IS_ERR(handler) &&
2923 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2924 		handler_ucast = create_flow_rule(dev, ft_prio,
2925 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
2926 						 dst);
2927 		if (IS_ERR(handler_ucast)) {
2928 			mlx5_del_flow_rules(handler->rule);
2929 			ft_prio->refcount--;
2930 			kfree(handler);
2931 			handler = handler_ucast;
2932 		} else {
2933 			list_add(&handler_ucast->list, &handler->list);
2934 		}
2935 	}
2936 
2937 	return handler;
2938 }
2939 
2940 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2941 							struct mlx5_ib_flow_prio *ft_rx,
2942 							struct mlx5_ib_flow_prio *ft_tx,
2943 							struct mlx5_flow_destination *dst)
2944 {
2945 	struct mlx5_ib_flow_handler *handler_rx;
2946 	struct mlx5_ib_flow_handler *handler_tx;
2947 	int err;
2948 	static const struct ib_flow_attr flow_attr  = {
2949 		.num_of_specs = 0,
2950 		.size = sizeof(flow_attr)
2951 	};
2952 
2953 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2954 	if (IS_ERR(handler_rx)) {
2955 		err = PTR_ERR(handler_rx);
2956 		goto err;
2957 	}
2958 
2959 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2960 	if (IS_ERR(handler_tx)) {
2961 		err = PTR_ERR(handler_tx);
2962 		goto err_tx;
2963 	}
2964 
2965 	list_add(&handler_tx->list, &handler_rx->list);
2966 
2967 	return handler_rx;
2968 
2969 err_tx:
2970 	mlx5_del_flow_rules(handler_rx->rule);
2971 	ft_rx->refcount--;
2972 	kfree(handler_rx);
2973 err:
2974 	return ERR_PTR(err);
2975 }
2976 
2977 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2978 					   struct ib_flow_attr *flow_attr,
2979 					   int domain)
2980 {
2981 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2982 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2983 	struct mlx5_ib_flow_handler *handler = NULL;
2984 	struct mlx5_flow_destination *dst = NULL;
2985 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2986 	struct mlx5_ib_flow_prio *ft_prio;
2987 	int err;
2988 	int underlay_qpn;
2989 
2990 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2991 		return ERR_PTR(-ENOMEM);
2992 
2993 	if (domain != IB_FLOW_DOMAIN_USER ||
2994 	    flow_attr->port > dev->num_ports ||
2995 	    (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2996 		return ERR_PTR(-EINVAL);
2997 
2998 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2999 	if (!dst)
3000 		return ERR_PTR(-ENOMEM);
3001 
3002 	mutex_lock(&dev->flow_db.lock);
3003 
3004 	ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
3005 	if (IS_ERR(ft_prio)) {
3006 		err = PTR_ERR(ft_prio);
3007 		goto unlock;
3008 	}
3009 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3010 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
3011 		if (IS_ERR(ft_prio_tx)) {
3012 			err = PTR_ERR(ft_prio_tx);
3013 			ft_prio_tx = NULL;
3014 			goto destroy_ft;
3015 		}
3016 	}
3017 
3018 	dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
3019 	if (mqp->flags & MLX5_IB_QP_RSS)
3020 		dst->tir_num = mqp->rss_qp.tirn;
3021 	else
3022 		dst->tir_num = mqp->raw_packet_qp.rq.tirn;
3023 
3024 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3025 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
3026 			handler = create_dont_trap_rule(dev, ft_prio,
3027 							flow_attr, dst);
3028 		} else {
3029 			underlay_qpn = (mqp->flags & MLX5_IB_QP_UNDERLAY) ?
3030 					mqp->underlay_qpn : 0;
3031 			handler = _create_flow_rule(dev, ft_prio, flow_attr,
3032 						    dst, underlay_qpn);
3033 		}
3034 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3035 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3036 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
3037 						dst);
3038 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3039 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
3040 	} else {
3041 		err = -EINVAL;
3042 		goto destroy_ft;
3043 	}
3044 
3045 	if (IS_ERR(handler)) {
3046 		err = PTR_ERR(handler);
3047 		handler = NULL;
3048 		goto destroy_ft;
3049 	}
3050 
3051 	mutex_unlock(&dev->flow_db.lock);
3052 	kfree(dst);
3053 
3054 	return &handler->ibflow;
3055 
3056 destroy_ft:
3057 	put_flow_table(dev, ft_prio, false);
3058 	if (ft_prio_tx)
3059 		put_flow_table(dev, ft_prio_tx, false);
3060 unlock:
3061 	mutex_unlock(&dev->flow_db.lock);
3062 	kfree(dst);
3063 	kfree(handler);
3064 	return ERR_PTR(err);
3065 }
3066 
3067 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
3068 {
3069 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3070 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
3071 	int err;
3072 
3073 	if (mqp->flags & MLX5_IB_QP_UNDERLAY) {
3074 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
3075 		return -EOPNOTSUPP;
3076 	}
3077 
3078 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
3079 	if (err)
3080 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
3081 			     ibqp->qp_num, gid->raw);
3082 
3083 	return err;
3084 }
3085 
3086 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
3087 {
3088 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3089 	int err;
3090 
3091 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
3092 	if (err)
3093 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
3094 			     ibqp->qp_num, gid->raw);
3095 
3096 	return err;
3097 }
3098 
3099 static int init_node_data(struct mlx5_ib_dev *dev)
3100 {
3101 	int err;
3102 
3103 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
3104 	if (err)
3105 		return err;
3106 
3107 	dev->mdev->rev_id = dev->mdev->pdev->revision;
3108 
3109 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
3110 }
3111 
3112 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
3113 			     char *buf)
3114 {
3115 	struct mlx5_ib_dev *dev =
3116 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3117 
3118 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
3119 }
3120 
3121 static ssize_t show_reg_pages(struct device *device,
3122 			      struct device_attribute *attr, char *buf)
3123 {
3124 	struct mlx5_ib_dev *dev =
3125 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3126 
3127 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
3128 }
3129 
3130 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
3131 			char *buf)
3132 {
3133 	struct mlx5_ib_dev *dev =
3134 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3135 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
3136 }
3137 
3138 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
3139 			char *buf)
3140 {
3141 	struct mlx5_ib_dev *dev =
3142 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3143 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
3144 }
3145 
3146 static ssize_t show_board(struct device *device, struct device_attribute *attr,
3147 			  char *buf)
3148 {
3149 	struct mlx5_ib_dev *dev =
3150 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3151 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
3152 		       dev->mdev->board_id);
3153 }
3154 
3155 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
3156 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
3157 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
3158 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
3159 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
3160 
3161 static struct device_attribute *mlx5_class_attributes[] = {
3162 	&dev_attr_hw_rev,
3163 	&dev_attr_hca_type,
3164 	&dev_attr_board_id,
3165 	&dev_attr_fw_pages,
3166 	&dev_attr_reg_pages,
3167 };
3168 
3169 static void pkey_change_handler(struct work_struct *work)
3170 {
3171 	struct mlx5_ib_port_resources *ports =
3172 		container_of(work, struct mlx5_ib_port_resources,
3173 			     pkey_change_work);
3174 
3175 	mutex_lock(&ports->devr->mutex);
3176 	mlx5_ib_gsi_pkey_change(ports->gsi);
3177 	mutex_unlock(&ports->devr->mutex);
3178 }
3179 
3180 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
3181 {
3182 	struct mlx5_ib_qp *mqp;
3183 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
3184 	struct mlx5_core_cq *mcq;
3185 	struct list_head cq_armed_list;
3186 	unsigned long flags_qp;
3187 	unsigned long flags_cq;
3188 	unsigned long flags;
3189 
3190 	INIT_LIST_HEAD(&cq_armed_list);
3191 
3192 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
3193 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
3194 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
3195 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
3196 		if (mqp->sq.tail != mqp->sq.head) {
3197 			send_mcq = to_mcq(mqp->ibqp.send_cq);
3198 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
3199 			if (send_mcq->mcq.comp &&
3200 			    mqp->ibqp.send_cq->comp_handler) {
3201 				if (!send_mcq->mcq.reset_notify_added) {
3202 					send_mcq->mcq.reset_notify_added = 1;
3203 					list_add_tail(&send_mcq->mcq.reset_notify,
3204 						      &cq_armed_list);
3205 				}
3206 			}
3207 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
3208 		}
3209 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
3210 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
3211 		/* no handling is needed for SRQ */
3212 		if (!mqp->ibqp.srq) {
3213 			if (mqp->rq.tail != mqp->rq.head) {
3214 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
3215 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
3216 				if (recv_mcq->mcq.comp &&
3217 				    mqp->ibqp.recv_cq->comp_handler) {
3218 					if (!recv_mcq->mcq.reset_notify_added) {
3219 						recv_mcq->mcq.reset_notify_added = 1;
3220 						list_add_tail(&recv_mcq->mcq.reset_notify,
3221 							      &cq_armed_list);
3222 					}
3223 				}
3224 				spin_unlock_irqrestore(&recv_mcq->lock,
3225 						       flags_cq);
3226 			}
3227 		}
3228 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
3229 	}
3230 	/*At that point all inflight post send were put to be executed as of we
3231 	 * lock/unlock above locks Now need to arm all involved CQs.
3232 	 */
3233 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
3234 		mcq->comp(mcq);
3235 	}
3236 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
3237 }
3238 
3239 static void delay_drop_handler(struct work_struct *work)
3240 {
3241 	int err;
3242 	struct mlx5_ib_delay_drop *delay_drop =
3243 		container_of(work, struct mlx5_ib_delay_drop,
3244 			     delay_drop_work);
3245 
3246 	atomic_inc(&delay_drop->events_cnt);
3247 
3248 	mutex_lock(&delay_drop->lock);
3249 	err = mlx5_core_set_delay_drop(delay_drop->dev->mdev,
3250 				       delay_drop->timeout);
3251 	if (err) {
3252 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
3253 			     delay_drop->timeout);
3254 		delay_drop->activate = false;
3255 	}
3256 	mutex_unlock(&delay_drop->lock);
3257 }
3258 
3259 static void mlx5_ib_handle_event(struct work_struct *_work)
3260 {
3261 	struct mlx5_ib_event_work *work =
3262 		container_of(_work, struct mlx5_ib_event_work, work);
3263 	struct mlx5_ib_dev *ibdev;
3264 	struct ib_event ibev;
3265 	bool fatal = false;
3266 	u8 port = 0;
3267 
3268 	if (mlx5_core_is_mp_slave(work->dev)) {
3269 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->context);
3270 		if (!ibdev)
3271 			goto out;
3272 	} else {
3273 		ibdev = work->context;
3274 	}
3275 
3276 	switch (work->event) {
3277 	case MLX5_DEV_EVENT_SYS_ERROR:
3278 		ibev.event = IB_EVENT_DEVICE_FATAL;
3279 		mlx5_ib_handle_internal_error(ibdev);
3280 		fatal = true;
3281 		break;
3282 
3283 	case MLX5_DEV_EVENT_PORT_UP:
3284 	case MLX5_DEV_EVENT_PORT_DOWN:
3285 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
3286 		port = (u8)work->param;
3287 
3288 		/* In RoCE, port up/down events are handled in
3289 		 * mlx5_netdev_event().
3290 		 */
3291 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
3292 			IB_LINK_LAYER_ETHERNET)
3293 			goto out;
3294 
3295 		ibev.event = (work->event == MLX5_DEV_EVENT_PORT_UP) ?
3296 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
3297 		break;
3298 
3299 	case MLX5_DEV_EVENT_LID_CHANGE:
3300 		ibev.event = IB_EVENT_LID_CHANGE;
3301 		port = (u8)work->param;
3302 		break;
3303 
3304 	case MLX5_DEV_EVENT_PKEY_CHANGE:
3305 		ibev.event = IB_EVENT_PKEY_CHANGE;
3306 		port = (u8)work->param;
3307 
3308 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
3309 		break;
3310 
3311 	case MLX5_DEV_EVENT_GUID_CHANGE:
3312 		ibev.event = IB_EVENT_GID_CHANGE;
3313 		port = (u8)work->param;
3314 		break;
3315 
3316 	case MLX5_DEV_EVENT_CLIENT_REREG:
3317 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
3318 		port = (u8)work->param;
3319 		break;
3320 	case MLX5_DEV_EVENT_DELAY_DROP_TIMEOUT:
3321 		schedule_work(&ibdev->delay_drop.delay_drop_work);
3322 		goto out;
3323 	default:
3324 		goto out;
3325 	}
3326 
3327 	ibev.device	      = &ibdev->ib_dev;
3328 	ibev.element.port_num = port;
3329 
3330 	if (port < 1 || port > ibdev->num_ports) {
3331 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
3332 		goto out;
3333 	}
3334 
3335 	if (ibdev->ib_active)
3336 		ib_dispatch_event(&ibev);
3337 
3338 	if (fatal)
3339 		ibdev->ib_active = false;
3340 out:
3341 	kfree(work);
3342 }
3343 
3344 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
3345 			  enum mlx5_dev_event event, unsigned long param)
3346 {
3347 	struct mlx5_ib_event_work *work;
3348 
3349 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
3350 	if (!work)
3351 		return;
3352 
3353 	INIT_WORK(&work->work, mlx5_ib_handle_event);
3354 	work->dev = dev;
3355 	work->param = param;
3356 	work->context = context;
3357 	work->event = event;
3358 
3359 	queue_work(mlx5_ib_event_wq, &work->work);
3360 }
3361 
3362 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
3363 {
3364 	struct mlx5_hca_vport_context vport_ctx;
3365 	int err;
3366 	int port;
3367 
3368 	for (port = 1; port <= dev->num_ports; port++) {
3369 		dev->mdev->port_caps[port - 1].has_smi = false;
3370 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
3371 		    MLX5_CAP_PORT_TYPE_IB) {
3372 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
3373 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
3374 								   port, 0,
3375 								   &vport_ctx);
3376 				if (err) {
3377 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
3378 						    port, err);
3379 					return err;
3380 				}
3381 				dev->mdev->port_caps[port - 1].has_smi =
3382 					vport_ctx.has_smi;
3383 			} else {
3384 				dev->mdev->port_caps[port - 1].has_smi = true;
3385 			}
3386 		}
3387 	}
3388 	return 0;
3389 }
3390 
3391 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
3392 {
3393 	int port;
3394 
3395 	for (port = 1; port <= dev->num_ports; port++)
3396 		mlx5_query_ext_port_caps(dev, port);
3397 }
3398 
3399 static int get_port_caps(struct mlx5_ib_dev *dev, u8 port)
3400 {
3401 	struct ib_device_attr *dprops = NULL;
3402 	struct ib_port_attr *pprops = NULL;
3403 	int err = -ENOMEM;
3404 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
3405 
3406 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
3407 	if (!pprops)
3408 		goto out;
3409 
3410 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
3411 	if (!dprops)
3412 		goto out;
3413 
3414 	err = set_has_smi_cap(dev);
3415 	if (err)
3416 		goto out;
3417 
3418 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
3419 	if (err) {
3420 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
3421 		goto out;
3422 	}
3423 
3424 	memset(pprops, 0, sizeof(*pprops));
3425 	err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
3426 	if (err) {
3427 		mlx5_ib_warn(dev, "query_port %d failed %d\n",
3428 			     port, err);
3429 		goto out;
3430 	}
3431 
3432 	dev->mdev->port_caps[port - 1].pkey_table_len =
3433 					dprops->max_pkeys;
3434 	dev->mdev->port_caps[port - 1].gid_table_len =
3435 					pprops->gid_tbl_len;
3436 	mlx5_ib_dbg(dev, "port %d: pkey_table_len %d, gid_table_len %d\n",
3437 		    port, dprops->max_pkeys, pprops->gid_tbl_len);
3438 
3439 out:
3440 	kfree(pprops);
3441 	kfree(dprops);
3442 
3443 	return err;
3444 }
3445 
3446 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
3447 {
3448 	int err;
3449 
3450 	err = mlx5_mr_cache_cleanup(dev);
3451 	if (err)
3452 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
3453 
3454 	mlx5_ib_destroy_qp(dev->umrc.qp);
3455 	ib_free_cq(dev->umrc.cq);
3456 	ib_dealloc_pd(dev->umrc.pd);
3457 }
3458 
3459 enum {
3460 	MAX_UMR_WR = 128,
3461 };
3462 
3463 static int create_umr_res(struct mlx5_ib_dev *dev)
3464 {
3465 	struct ib_qp_init_attr *init_attr = NULL;
3466 	struct ib_qp_attr *attr = NULL;
3467 	struct ib_pd *pd;
3468 	struct ib_cq *cq;
3469 	struct ib_qp *qp;
3470 	int ret;
3471 
3472 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
3473 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
3474 	if (!attr || !init_attr) {
3475 		ret = -ENOMEM;
3476 		goto error_0;
3477 	}
3478 
3479 	pd = ib_alloc_pd(&dev->ib_dev, 0);
3480 	if (IS_ERR(pd)) {
3481 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
3482 		ret = PTR_ERR(pd);
3483 		goto error_0;
3484 	}
3485 
3486 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
3487 	if (IS_ERR(cq)) {
3488 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
3489 		ret = PTR_ERR(cq);
3490 		goto error_2;
3491 	}
3492 
3493 	init_attr->send_cq = cq;
3494 	init_attr->recv_cq = cq;
3495 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
3496 	init_attr->cap.max_send_wr = MAX_UMR_WR;
3497 	init_attr->cap.max_send_sge = 1;
3498 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
3499 	init_attr->port_num = 1;
3500 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
3501 	if (IS_ERR(qp)) {
3502 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
3503 		ret = PTR_ERR(qp);
3504 		goto error_3;
3505 	}
3506 	qp->device     = &dev->ib_dev;
3507 	qp->real_qp    = qp;
3508 	qp->uobject    = NULL;
3509 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
3510 	qp->send_cq    = init_attr->send_cq;
3511 	qp->recv_cq    = init_attr->recv_cq;
3512 
3513 	attr->qp_state = IB_QPS_INIT;
3514 	attr->port_num = 1;
3515 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
3516 				IB_QP_PORT, NULL);
3517 	if (ret) {
3518 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
3519 		goto error_4;
3520 	}
3521 
3522 	memset(attr, 0, sizeof(*attr));
3523 	attr->qp_state = IB_QPS_RTR;
3524 	attr->path_mtu = IB_MTU_256;
3525 
3526 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
3527 	if (ret) {
3528 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
3529 		goto error_4;
3530 	}
3531 
3532 	memset(attr, 0, sizeof(*attr));
3533 	attr->qp_state = IB_QPS_RTS;
3534 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
3535 	if (ret) {
3536 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
3537 		goto error_4;
3538 	}
3539 
3540 	dev->umrc.qp = qp;
3541 	dev->umrc.cq = cq;
3542 	dev->umrc.pd = pd;
3543 
3544 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
3545 	ret = mlx5_mr_cache_init(dev);
3546 	if (ret) {
3547 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
3548 		goto error_4;
3549 	}
3550 
3551 	kfree(attr);
3552 	kfree(init_attr);
3553 
3554 	return 0;
3555 
3556 error_4:
3557 	mlx5_ib_destroy_qp(qp);
3558 
3559 error_3:
3560 	ib_free_cq(cq);
3561 
3562 error_2:
3563 	ib_dealloc_pd(pd);
3564 
3565 error_0:
3566 	kfree(attr);
3567 	kfree(init_attr);
3568 	return ret;
3569 }
3570 
3571 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
3572 {
3573 	switch (umr_fence_cap) {
3574 	case MLX5_CAP_UMR_FENCE_NONE:
3575 		return MLX5_FENCE_MODE_NONE;
3576 	case MLX5_CAP_UMR_FENCE_SMALL:
3577 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
3578 	default:
3579 		return MLX5_FENCE_MODE_STRONG_ORDERING;
3580 	}
3581 }
3582 
3583 static int create_dev_resources(struct mlx5_ib_resources *devr)
3584 {
3585 	struct ib_srq_init_attr attr;
3586 	struct mlx5_ib_dev *dev;
3587 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
3588 	int port;
3589 	int ret = 0;
3590 
3591 	dev = container_of(devr, struct mlx5_ib_dev, devr);
3592 
3593 	mutex_init(&devr->mutex);
3594 
3595 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
3596 	if (IS_ERR(devr->p0)) {
3597 		ret = PTR_ERR(devr->p0);
3598 		goto error0;
3599 	}
3600 	devr->p0->device  = &dev->ib_dev;
3601 	devr->p0->uobject = NULL;
3602 	atomic_set(&devr->p0->usecnt, 0);
3603 
3604 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
3605 	if (IS_ERR(devr->c0)) {
3606 		ret = PTR_ERR(devr->c0);
3607 		goto error1;
3608 	}
3609 	devr->c0->device        = &dev->ib_dev;
3610 	devr->c0->uobject       = NULL;
3611 	devr->c0->comp_handler  = NULL;
3612 	devr->c0->event_handler = NULL;
3613 	devr->c0->cq_context    = NULL;
3614 	atomic_set(&devr->c0->usecnt, 0);
3615 
3616 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
3617 	if (IS_ERR(devr->x0)) {
3618 		ret = PTR_ERR(devr->x0);
3619 		goto error2;
3620 	}
3621 	devr->x0->device = &dev->ib_dev;
3622 	devr->x0->inode = NULL;
3623 	atomic_set(&devr->x0->usecnt, 0);
3624 	mutex_init(&devr->x0->tgt_qp_mutex);
3625 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
3626 
3627 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
3628 	if (IS_ERR(devr->x1)) {
3629 		ret = PTR_ERR(devr->x1);
3630 		goto error3;
3631 	}
3632 	devr->x1->device = &dev->ib_dev;
3633 	devr->x1->inode = NULL;
3634 	atomic_set(&devr->x1->usecnt, 0);
3635 	mutex_init(&devr->x1->tgt_qp_mutex);
3636 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
3637 
3638 	memset(&attr, 0, sizeof(attr));
3639 	attr.attr.max_sge = 1;
3640 	attr.attr.max_wr = 1;
3641 	attr.srq_type = IB_SRQT_XRC;
3642 	attr.ext.cq = devr->c0;
3643 	attr.ext.xrc.xrcd = devr->x0;
3644 
3645 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
3646 	if (IS_ERR(devr->s0)) {
3647 		ret = PTR_ERR(devr->s0);
3648 		goto error4;
3649 	}
3650 	devr->s0->device	= &dev->ib_dev;
3651 	devr->s0->pd		= devr->p0;
3652 	devr->s0->uobject       = NULL;
3653 	devr->s0->event_handler = NULL;
3654 	devr->s0->srq_context   = NULL;
3655 	devr->s0->srq_type      = IB_SRQT_XRC;
3656 	devr->s0->ext.xrc.xrcd	= devr->x0;
3657 	devr->s0->ext.cq	= devr->c0;
3658 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
3659 	atomic_inc(&devr->s0->ext.cq->usecnt);
3660 	atomic_inc(&devr->p0->usecnt);
3661 	atomic_set(&devr->s0->usecnt, 0);
3662 
3663 	memset(&attr, 0, sizeof(attr));
3664 	attr.attr.max_sge = 1;
3665 	attr.attr.max_wr = 1;
3666 	attr.srq_type = IB_SRQT_BASIC;
3667 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
3668 	if (IS_ERR(devr->s1)) {
3669 		ret = PTR_ERR(devr->s1);
3670 		goto error5;
3671 	}
3672 	devr->s1->device	= &dev->ib_dev;
3673 	devr->s1->pd		= devr->p0;
3674 	devr->s1->uobject       = NULL;
3675 	devr->s1->event_handler = NULL;
3676 	devr->s1->srq_context   = NULL;
3677 	devr->s1->srq_type      = IB_SRQT_BASIC;
3678 	devr->s1->ext.cq	= devr->c0;
3679 	atomic_inc(&devr->p0->usecnt);
3680 	atomic_set(&devr->s1->usecnt, 0);
3681 
3682 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
3683 		INIT_WORK(&devr->ports[port].pkey_change_work,
3684 			  pkey_change_handler);
3685 		devr->ports[port].devr = devr;
3686 	}
3687 
3688 	return 0;
3689 
3690 error5:
3691 	mlx5_ib_destroy_srq(devr->s0);
3692 error4:
3693 	mlx5_ib_dealloc_xrcd(devr->x1);
3694 error3:
3695 	mlx5_ib_dealloc_xrcd(devr->x0);
3696 error2:
3697 	mlx5_ib_destroy_cq(devr->c0);
3698 error1:
3699 	mlx5_ib_dealloc_pd(devr->p0);
3700 error0:
3701 	return ret;
3702 }
3703 
3704 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
3705 {
3706 	struct mlx5_ib_dev *dev =
3707 		container_of(devr, struct mlx5_ib_dev, devr);
3708 	int port;
3709 
3710 	mlx5_ib_destroy_srq(devr->s1);
3711 	mlx5_ib_destroy_srq(devr->s0);
3712 	mlx5_ib_dealloc_xrcd(devr->x0);
3713 	mlx5_ib_dealloc_xrcd(devr->x1);
3714 	mlx5_ib_destroy_cq(devr->c0);
3715 	mlx5_ib_dealloc_pd(devr->p0);
3716 
3717 	/* Make sure no change P_Key work items are still executing */
3718 	for (port = 0; port < dev->num_ports; ++port)
3719 		cancel_work_sync(&devr->ports[port].pkey_change_work);
3720 }
3721 
3722 static u32 get_core_cap_flags(struct ib_device *ibdev)
3723 {
3724 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3725 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
3726 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
3727 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
3728 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
3729 	u32 ret = 0;
3730 
3731 	if (ll == IB_LINK_LAYER_INFINIBAND)
3732 		return RDMA_CORE_PORT_IBA_IB;
3733 
3734 	if (raw_support)
3735 		ret = RDMA_CORE_PORT_RAW_PACKET;
3736 
3737 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
3738 		return ret;
3739 
3740 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
3741 		return ret;
3742 
3743 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
3744 		ret |= RDMA_CORE_PORT_IBA_ROCE;
3745 
3746 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
3747 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3748 
3749 	return ret;
3750 }
3751 
3752 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
3753 			       struct ib_port_immutable *immutable)
3754 {
3755 	struct ib_port_attr attr;
3756 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3757 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
3758 	int err;
3759 
3760 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
3761 
3762 	err = ib_query_port(ibdev, port_num, &attr);
3763 	if (err)
3764 		return err;
3765 
3766 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3767 	immutable->gid_tbl_len = attr.gid_tbl_len;
3768 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
3769 	if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
3770 		immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3771 
3772 	return 0;
3773 }
3774 
3775 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
3776 {
3777 	struct mlx5_ib_dev *dev =
3778 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3779 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
3780 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
3781 		 fw_rev_sub(dev->mdev));
3782 }
3783 
3784 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3785 {
3786 	struct mlx5_core_dev *mdev = dev->mdev;
3787 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3788 								 MLX5_FLOW_NAMESPACE_LAG);
3789 	struct mlx5_flow_table *ft;
3790 	int err;
3791 
3792 	if (!ns || !mlx5_lag_is_active(mdev))
3793 		return 0;
3794 
3795 	err = mlx5_cmd_create_vport_lag(mdev);
3796 	if (err)
3797 		return err;
3798 
3799 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3800 	if (IS_ERR(ft)) {
3801 		err = PTR_ERR(ft);
3802 		goto err_destroy_vport_lag;
3803 	}
3804 
3805 	dev->flow_db.lag_demux_ft = ft;
3806 	return 0;
3807 
3808 err_destroy_vport_lag:
3809 	mlx5_cmd_destroy_vport_lag(mdev);
3810 	return err;
3811 }
3812 
3813 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3814 {
3815 	struct mlx5_core_dev *mdev = dev->mdev;
3816 
3817 	if (dev->flow_db.lag_demux_ft) {
3818 		mlx5_destroy_flow_table(dev->flow_db.lag_demux_ft);
3819 		dev->flow_db.lag_demux_ft = NULL;
3820 
3821 		mlx5_cmd_destroy_vport_lag(mdev);
3822 	}
3823 }
3824 
3825 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
3826 {
3827 	int err;
3828 
3829 	dev->roce[port_num].nb.notifier_call = mlx5_netdev_event;
3830 	err = register_netdevice_notifier(&dev->roce[port_num].nb);
3831 	if (err) {
3832 		dev->roce[port_num].nb.notifier_call = NULL;
3833 		return err;
3834 	}
3835 
3836 	return 0;
3837 }
3838 
3839 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
3840 {
3841 	if (dev->roce[port_num].nb.notifier_call) {
3842 		unregister_netdevice_notifier(&dev->roce[port_num].nb);
3843 		dev->roce[port_num].nb.notifier_call = NULL;
3844 	}
3845 }
3846 
3847 static int mlx5_enable_eth(struct mlx5_ib_dev *dev, u8 port_num)
3848 {
3849 	int err;
3850 
3851 	err = mlx5_add_netdev_notifier(dev, port_num);
3852 	if (err)
3853 		return err;
3854 
3855 	if (MLX5_CAP_GEN(dev->mdev, roce)) {
3856 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3857 		if (err)
3858 			goto err_unregister_netdevice_notifier;
3859 	}
3860 
3861 	err = mlx5_eth_lag_init(dev);
3862 	if (err)
3863 		goto err_disable_roce;
3864 
3865 	return 0;
3866 
3867 err_disable_roce:
3868 	if (MLX5_CAP_GEN(dev->mdev, roce))
3869 		mlx5_nic_vport_disable_roce(dev->mdev);
3870 
3871 err_unregister_netdevice_notifier:
3872 	mlx5_remove_netdev_notifier(dev, port_num);
3873 	return err;
3874 }
3875 
3876 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3877 {
3878 	mlx5_eth_lag_cleanup(dev);
3879 	if (MLX5_CAP_GEN(dev->mdev, roce))
3880 		mlx5_nic_vport_disable_roce(dev->mdev);
3881 }
3882 
3883 struct mlx5_ib_counter {
3884 	const char *name;
3885 	size_t offset;
3886 };
3887 
3888 #define INIT_Q_COUNTER(_name)		\
3889 	{ .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
3890 
3891 static const struct mlx5_ib_counter basic_q_cnts[] = {
3892 	INIT_Q_COUNTER(rx_write_requests),
3893 	INIT_Q_COUNTER(rx_read_requests),
3894 	INIT_Q_COUNTER(rx_atomic_requests),
3895 	INIT_Q_COUNTER(out_of_buffer),
3896 };
3897 
3898 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
3899 	INIT_Q_COUNTER(out_of_sequence),
3900 };
3901 
3902 static const struct mlx5_ib_counter retrans_q_cnts[] = {
3903 	INIT_Q_COUNTER(duplicate_request),
3904 	INIT_Q_COUNTER(rnr_nak_retry_err),
3905 	INIT_Q_COUNTER(packet_seq_err),
3906 	INIT_Q_COUNTER(implied_nak_seq_err),
3907 	INIT_Q_COUNTER(local_ack_timeout_err),
3908 };
3909 
3910 #define INIT_CONG_COUNTER(_name)		\
3911 	{ .name = #_name, .offset =	\
3912 		MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
3913 
3914 static const struct mlx5_ib_counter cong_cnts[] = {
3915 	INIT_CONG_COUNTER(rp_cnp_ignored),
3916 	INIT_CONG_COUNTER(rp_cnp_handled),
3917 	INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
3918 	INIT_CONG_COUNTER(np_cnp_sent),
3919 };
3920 
3921 static const struct mlx5_ib_counter extended_err_cnts[] = {
3922 	INIT_Q_COUNTER(resp_local_length_error),
3923 	INIT_Q_COUNTER(resp_cqe_error),
3924 	INIT_Q_COUNTER(req_cqe_error),
3925 	INIT_Q_COUNTER(req_remote_invalid_request),
3926 	INIT_Q_COUNTER(req_remote_access_errors),
3927 	INIT_Q_COUNTER(resp_remote_access_errors),
3928 	INIT_Q_COUNTER(resp_cqe_flush_error),
3929 	INIT_Q_COUNTER(req_cqe_flush_error),
3930 };
3931 
3932 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
3933 {
3934 	int i;
3935 
3936 	for (i = 0; i < dev->num_ports; i++) {
3937 		if (dev->port[i].cnts.set_id)
3938 			mlx5_core_dealloc_q_counter(dev->mdev,
3939 						    dev->port[i].cnts.set_id);
3940 		kfree(dev->port[i].cnts.names);
3941 		kfree(dev->port[i].cnts.offsets);
3942 	}
3943 }
3944 
3945 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
3946 				    struct mlx5_ib_counters *cnts)
3947 {
3948 	u32 num_counters;
3949 
3950 	num_counters = ARRAY_SIZE(basic_q_cnts);
3951 
3952 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
3953 		num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
3954 
3955 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
3956 		num_counters += ARRAY_SIZE(retrans_q_cnts);
3957 
3958 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters))
3959 		num_counters += ARRAY_SIZE(extended_err_cnts);
3960 
3961 	cnts->num_q_counters = num_counters;
3962 
3963 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
3964 		cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
3965 		num_counters += ARRAY_SIZE(cong_cnts);
3966 	}
3967 
3968 	cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
3969 	if (!cnts->names)
3970 		return -ENOMEM;
3971 
3972 	cnts->offsets = kcalloc(num_counters,
3973 				sizeof(cnts->offsets), GFP_KERNEL);
3974 	if (!cnts->offsets)
3975 		goto err_names;
3976 
3977 	return 0;
3978 
3979 err_names:
3980 	kfree(cnts->names);
3981 	cnts->names = NULL;
3982 	return -ENOMEM;
3983 }
3984 
3985 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
3986 				  const char **names,
3987 				  size_t *offsets)
3988 {
3989 	int i;
3990 	int j = 0;
3991 
3992 	for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
3993 		names[j] = basic_q_cnts[i].name;
3994 		offsets[j] = basic_q_cnts[i].offset;
3995 	}
3996 
3997 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
3998 		for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
3999 			names[j] = out_of_seq_q_cnts[i].name;
4000 			offsets[j] = out_of_seq_q_cnts[i].offset;
4001 		}
4002 	}
4003 
4004 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
4005 		for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
4006 			names[j] = retrans_q_cnts[i].name;
4007 			offsets[j] = retrans_q_cnts[i].offset;
4008 		}
4009 	}
4010 
4011 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters)) {
4012 		for (i = 0; i < ARRAY_SIZE(extended_err_cnts); i++, j++) {
4013 			names[j] = extended_err_cnts[i].name;
4014 			offsets[j] = extended_err_cnts[i].offset;
4015 		}
4016 	}
4017 
4018 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4019 		for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
4020 			names[j] = cong_cnts[i].name;
4021 			offsets[j] = cong_cnts[i].offset;
4022 		}
4023 	}
4024 }
4025 
4026 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
4027 {
4028 	int err = 0;
4029 	int i;
4030 
4031 	for (i = 0; i < dev->num_ports; i++) {
4032 		err = __mlx5_ib_alloc_counters(dev, &dev->port[i].cnts);
4033 		if (err)
4034 			goto err_alloc;
4035 
4036 		mlx5_ib_fill_counters(dev, dev->port[i].cnts.names,
4037 				      dev->port[i].cnts.offsets);
4038 
4039 		err = mlx5_core_alloc_q_counter(dev->mdev,
4040 						&dev->port[i].cnts.set_id);
4041 		if (err) {
4042 			mlx5_ib_warn(dev,
4043 				     "couldn't allocate queue counter for port %d, err %d\n",
4044 				     i + 1, err);
4045 			goto err_alloc;
4046 		}
4047 		dev->port[i].cnts.set_id_valid = true;
4048 	}
4049 
4050 	return 0;
4051 
4052 err_alloc:
4053 	mlx5_ib_dealloc_counters(dev);
4054 	return err;
4055 }
4056 
4057 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
4058 						    u8 port_num)
4059 {
4060 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4061 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
4062 
4063 	/* We support only per port stats */
4064 	if (port_num == 0)
4065 		return NULL;
4066 
4067 	return rdma_alloc_hw_stats_struct(port->cnts.names,
4068 					  port->cnts.num_q_counters +
4069 					  port->cnts.num_cong_counters,
4070 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
4071 }
4072 
4073 static int mlx5_ib_query_q_counters(struct mlx5_core_dev *mdev,
4074 				    struct mlx5_ib_port *port,
4075 				    struct rdma_hw_stats *stats)
4076 {
4077 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
4078 	void *out;
4079 	__be32 val;
4080 	int ret, i;
4081 
4082 	out = kvzalloc(outlen, GFP_KERNEL);
4083 	if (!out)
4084 		return -ENOMEM;
4085 
4086 	ret = mlx5_core_query_q_counter(mdev,
4087 					port->cnts.set_id, 0,
4088 					out, outlen);
4089 	if (ret)
4090 		goto free;
4091 
4092 	for (i = 0; i < port->cnts.num_q_counters; i++) {
4093 		val = *(__be32 *)(out + port->cnts.offsets[i]);
4094 		stats->value[i] = (u64)be32_to_cpu(val);
4095 	}
4096 
4097 free:
4098 	kvfree(out);
4099 	return ret;
4100 }
4101 
4102 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
4103 				struct rdma_hw_stats *stats,
4104 				u8 port_num, int index)
4105 {
4106 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4107 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
4108 	struct mlx5_core_dev *mdev;
4109 	int ret, num_counters;
4110 	u8 mdev_port_num;
4111 
4112 	if (!stats)
4113 		return -EINVAL;
4114 
4115 	num_counters = port->cnts.num_q_counters + port->cnts.num_cong_counters;
4116 
4117 	/* q_counters are per IB device, query the master mdev */
4118 	ret = mlx5_ib_query_q_counters(dev->mdev, port, stats);
4119 	if (ret)
4120 		return ret;
4121 
4122 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4123 		mdev = mlx5_ib_get_native_port_mdev(dev, port_num,
4124 						    &mdev_port_num);
4125 		if (!mdev) {
4126 			/* If port is not affiliated yet, its in down state
4127 			 * which doesn't have any counters yet, so it would be
4128 			 * zero. So no need to read from the HCA.
4129 			 */
4130 			goto done;
4131 		}
4132 		ret = mlx5_lag_query_cong_counters(dev->mdev,
4133 						   stats->value +
4134 						   port->cnts.num_q_counters,
4135 						   port->cnts.num_cong_counters,
4136 						   port->cnts.offsets +
4137 						   port->cnts.num_q_counters);
4138 
4139 		mlx5_ib_put_native_port_mdev(dev, port_num);
4140 		if (ret)
4141 			return ret;
4142 	}
4143 
4144 done:
4145 	return num_counters;
4146 }
4147 
4148 static void mlx5_ib_free_rdma_netdev(struct net_device *netdev)
4149 {
4150 	return mlx5_rdma_netdev_free(netdev);
4151 }
4152 
4153 static struct net_device*
4154 mlx5_ib_alloc_rdma_netdev(struct ib_device *hca,
4155 			  u8 port_num,
4156 			  enum rdma_netdev_t type,
4157 			  const char *name,
4158 			  unsigned char name_assign_type,
4159 			  void (*setup)(struct net_device *))
4160 {
4161 	struct net_device *netdev;
4162 	struct rdma_netdev *rn;
4163 
4164 	if (type != RDMA_NETDEV_IPOIB)
4165 		return ERR_PTR(-EOPNOTSUPP);
4166 
4167 	netdev = mlx5_rdma_netdev_alloc(to_mdev(hca)->mdev, hca,
4168 					name, setup);
4169 	if (likely(!IS_ERR_OR_NULL(netdev))) {
4170 		rn = netdev_priv(netdev);
4171 		rn->free_rdma_netdev = mlx5_ib_free_rdma_netdev;
4172 	}
4173 	return netdev;
4174 }
4175 
4176 static void delay_drop_debugfs_cleanup(struct mlx5_ib_dev *dev)
4177 {
4178 	if (!dev->delay_drop.dbg)
4179 		return;
4180 	debugfs_remove_recursive(dev->delay_drop.dbg->dir_debugfs);
4181 	kfree(dev->delay_drop.dbg);
4182 	dev->delay_drop.dbg = NULL;
4183 }
4184 
4185 static void cancel_delay_drop(struct mlx5_ib_dev *dev)
4186 {
4187 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4188 		return;
4189 
4190 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4191 	delay_drop_debugfs_cleanup(dev);
4192 }
4193 
4194 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
4195 				       size_t count, loff_t *pos)
4196 {
4197 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
4198 	char lbuf[20];
4199 	int len;
4200 
4201 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
4202 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
4203 }
4204 
4205 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
4206 					size_t count, loff_t *pos)
4207 {
4208 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
4209 	u32 timeout;
4210 	u32 var;
4211 
4212 	if (kstrtouint_from_user(buf, count, 0, &var))
4213 		return -EFAULT;
4214 
4215 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
4216 			1000);
4217 	if (timeout != var)
4218 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
4219 			    timeout);
4220 
4221 	delay_drop->timeout = timeout;
4222 
4223 	return count;
4224 }
4225 
4226 static const struct file_operations fops_delay_drop_timeout = {
4227 	.owner	= THIS_MODULE,
4228 	.open	= simple_open,
4229 	.write	= delay_drop_timeout_write,
4230 	.read	= delay_drop_timeout_read,
4231 };
4232 
4233 static int delay_drop_debugfs_init(struct mlx5_ib_dev *dev)
4234 {
4235 	struct mlx5_ib_dbg_delay_drop *dbg;
4236 
4237 	if (!mlx5_debugfs_root)
4238 		return 0;
4239 
4240 	dbg = kzalloc(sizeof(*dbg), GFP_KERNEL);
4241 	if (!dbg)
4242 		return -ENOMEM;
4243 
4244 	dev->delay_drop.dbg = dbg;
4245 
4246 	dbg->dir_debugfs =
4247 		debugfs_create_dir("delay_drop",
4248 				   dev->mdev->priv.dbg_root);
4249 	if (!dbg->dir_debugfs)
4250 		goto out_debugfs;
4251 
4252 	dbg->events_cnt_debugfs =
4253 		debugfs_create_atomic_t("num_timeout_events", 0400,
4254 					dbg->dir_debugfs,
4255 					&dev->delay_drop.events_cnt);
4256 	if (!dbg->events_cnt_debugfs)
4257 		goto out_debugfs;
4258 
4259 	dbg->rqs_cnt_debugfs =
4260 		debugfs_create_atomic_t("num_rqs", 0400,
4261 					dbg->dir_debugfs,
4262 					&dev->delay_drop.rqs_cnt);
4263 	if (!dbg->rqs_cnt_debugfs)
4264 		goto out_debugfs;
4265 
4266 	dbg->timeout_debugfs =
4267 		debugfs_create_file("timeout", 0600,
4268 				    dbg->dir_debugfs,
4269 				    &dev->delay_drop,
4270 				    &fops_delay_drop_timeout);
4271 	if (!dbg->timeout_debugfs)
4272 		goto out_debugfs;
4273 
4274 	return 0;
4275 
4276 out_debugfs:
4277 	delay_drop_debugfs_cleanup(dev);
4278 	return -ENOMEM;
4279 }
4280 
4281 static void init_delay_drop(struct mlx5_ib_dev *dev)
4282 {
4283 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4284 		return;
4285 
4286 	mutex_init(&dev->delay_drop.lock);
4287 	dev->delay_drop.dev = dev;
4288 	dev->delay_drop.activate = false;
4289 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4290 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4291 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4292 	atomic_set(&dev->delay_drop.events_cnt, 0);
4293 
4294 	if (delay_drop_debugfs_init(dev))
4295 		mlx5_ib_warn(dev, "Failed to init delay drop debugfs\n");
4296 }
4297 
4298 static const struct cpumask *
4299 mlx5_ib_get_vector_affinity(struct ib_device *ibdev, int comp_vector)
4300 {
4301 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4302 
4303 	return mlx5_get_vector_affinity(dev->mdev, comp_vector);
4304 }
4305 
4306 /* The mlx5_ib_multiport_mutex should be held when calling this function */
4307 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
4308 				      struct mlx5_ib_multiport_info *mpi)
4309 {
4310 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
4311 	struct mlx5_ib_port *port = &ibdev->port[port_num];
4312 	int comps;
4313 	int err;
4314 	int i;
4315 
4316 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
4317 
4318 	spin_lock(&port->mp.mpi_lock);
4319 	if (!mpi->ibdev) {
4320 		spin_unlock(&port->mp.mpi_lock);
4321 		return;
4322 	}
4323 	mpi->ibdev = NULL;
4324 
4325 	spin_unlock(&port->mp.mpi_lock);
4326 	mlx5_remove_netdev_notifier(ibdev, port_num);
4327 	spin_lock(&port->mp.mpi_lock);
4328 
4329 	comps = mpi->mdev_refcnt;
4330 	if (comps) {
4331 		mpi->unaffiliate = true;
4332 		init_completion(&mpi->unref_comp);
4333 		spin_unlock(&port->mp.mpi_lock);
4334 
4335 		for (i = 0; i < comps; i++)
4336 			wait_for_completion(&mpi->unref_comp);
4337 
4338 		spin_lock(&port->mp.mpi_lock);
4339 		mpi->unaffiliate = false;
4340 	}
4341 
4342 	port->mp.mpi = NULL;
4343 
4344 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4345 
4346 	spin_unlock(&port->mp.mpi_lock);
4347 
4348 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
4349 
4350 	mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
4351 	/* Log an error, still needed to cleanup the pointers and add
4352 	 * it back to the list.
4353 	 */
4354 	if (err)
4355 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
4356 			    port_num + 1);
4357 
4358 	ibdev->roce[port_num].last_port_state = IB_PORT_DOWN;
4359 }
4360 
4361 /* The mlx5_ib_multiport_mutex should be held when calling this function */
4362 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
4363 				    struct mlx5_ib_multiport_info *mpi)
4364 {
4365 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
4366 	int err;
4367 
4368 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
4369 	if (ibdev->port[port_num].mp.mpi) {
4370 		mlx5_ib_warn(ibdev, "port %d already affiliated.\n",
4371 			     port_num + 1);
4372 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
4373 		return false;
4374 	}
4375 
4376 	ibdev->port[port_num].mp.mpi = mpi;
4377 	mpi->ibdev = ibdev;
4378 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
4379 
4380 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
4381 	if (err)
4382 		goto unbind;
4383 
4384 	err = get_port_caps(ibdev, mlx5_core_native_port_num(mpi->mdev));
4385 	if (err)
4386 		goto unbind;
4387 
4388 	err = mlx5_add_netdev_notifier(ibdev, port_num);
4389 	if (err) {
4390 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
4391 			    port_num + 1);
4392 		goto unbind;
4393 	}
4394 
4395 	err = mlx5_ib_init_cong_debugfs(ibdev, port_num);
4396 	if (err)
4397 		goto unbind;
4398 
4399 	return true;
4400 
4401 unbind:
4402 	mlx5_ib_unbind_slave_port(ibdev, mpi);
4403 	return false;
4404 }
4405 
4406 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
4407 {
4408 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4409 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
4410 							  port_num + 1);
4411 	struct mlx5_ib_multiport_info *mpi;
4412 	int err;
4413 	int i;
4414 
4415 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
4416 		return 0;
4417 
4418 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
4419 						     &dev->sys_image_guid);
4420 	if (err)
4421 		return err;
4422 
4423 	err = mlx5_nic_vport_enable_roce(dev->mdev);
4424 	if (err)
4425 		return err;
4426 
4427 	mutex_lock(&mlx5_ib_multiport_mutex);
4428 	for (i = 0; i < dev->num_ports; i++) {
4429 		bool bound = false;
4430 
4431 		/* build a stub multiport info struct for the native port. */
4432 		if (i == port_num) {
4433 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4434 			if (!mpi) {
4435 				mutex_unlock(&mlx5_ib_multiport_mutex);
4436 				mlx5_nic_vport_disable_roce(dev->mdev);
4437 				return -ENOMEM;
4438 			}
4439 
4440 			mpi->is_master = true;
4441 			mpi->mdev = dev->mdev;
4442 			mpi->sys_image_guid = dev->sys_image_guid;
4443 			dev->port[i].mp.mpi = mpi;
4444 			mpi->ibdev = dev;
4445 			mpi = NULL;
4446 			continue;
4447 		}
4448 
4449 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
4450 				    list) {
4451 			if (dev->sys_image_guid == mpi->sys_image_guid &&
4452 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
4453 				bound = mlx5_ib_bind_slave_port(dev, mpi);
4454 			}
4455 
4456 			if (bound) {
4457 				dev_dbg(&mpi->mdev->pdev->dev, "removing port from unaffiliated list.\n");
4458 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
4459 				list_del(&mpi->list);
4460 				break;
4461 			}
4462 		}
4463 		if (!bound) {
4464 			get_port_caps(dev, i + 1);
4465 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
4466 				    i + 1);
4467 		}
4468 	}
4469 
4470 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
4471 	mutex_unlock(&mlx5_ib_multiport_mutex);
4472 	return err;
4473 }
4474 
4475 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
4476 {
4477 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4478 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
4479 							  port_num + 1);
4480 	int i;
4481 
4482 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
4483 		return;
4484 
4485 	mutex_lock(&mlx5_ib_multiport_mutex);
4486 	for (i = 0; i < dev->num_ports; i++) {
4487 		if (dev->port[i].mp.mpi) {
4488 			/* Destroy the native port stub */
4489 			if (i == port_num) {
4490 				kfree(dev->port[i].mp.mpi);
4491 				dev->port[i].mp.mpi = NULL;
4492 			} else {
4493 				mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
4494 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
4495 			}
4496 		}
4497 	}
4498 
4499 	mlx5_ib_dbg(dev, "removing from devlist\n");
4500 	list_del(&dev->ib_dev_list);
4501 	mutex_unlock(&mlx5_ib_multiport_mutex);
4502 
4503 	mlx5_nic_vport_disable_roce(dev->mdev);
4504 }
4505 
4506 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
4507 {
4508 	mlx5_ib_cleanup_multiport_master(dev);
4509 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4510 	cleanup_srcu_struct(&dev->mr_srcu);
4511 #endif
4512 	kfree(dev->port);
4513 }
4514 
4515 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
4516 {
4517 	struct mlx5_core_dev *mdev = dev->mdev;
4518 	const char *name;
4519 	int err;
4520 	int i;
4521 
4522 	dev->port = kcalloc(dev->num_ports, sizeof(*dev->port),
4523 			    GFP_KERNEL);
4524 	if (!dev->port)
4525 		return -ENOMEM;
4526 
4527 	for (i = 0; i < dev->num_ports; i++) {
4528 		spin_lock_init(&dev->port[i].mp.mpi_lock);
4529 		rwlock_init(&dev->roce[i].netdev_lock);
4530 	}
4531 
4532 	err = mlx5_ib_init_multiport_master(dev);
4533 	if (err)
4534 		goto err_free_port;
4535 
4536 	if (!mlx5_core_mp_enabled(mdev)) {
4537 		int i;
4538 
4539 		for (i = 1; i <= dev->num_ports; i++) {
4540 			err = get_port_caps(dev, i);
4541 			if (err)
4542 				break;
4543 		}
4544 	} else {
4545 		err = get_port_caps(dev, mlx5_core_native_port_num(mdev));
4546 	}
4547 	if (err)
4548 		goto err_mp;
4549 
4550 	if (mlx5_use_mad_ifc(dev))
4551 		get_ext_port_caps(dev);
4552 
4553 	if (!mlx5_lag_is_active(mdev))
4554 		name = "mlx5_%d";
4555 	else
4556 		name = "mlx5_bond_%d";
4557 
4558 	strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
4559 	dev->ib_dev.owner		= THIS_MODULE;
4560 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
4561 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
4562 	dev->ib_dev.phys_port_cnt	= dev->num_ports;
4563 	dev->ib_dev.num_comp_vectors    =
4564 		dev->mdev->priv.eq_table.num_comp_vectors;
4565 	dev->ib_dev.dev.parent		= &mdev->pdev->dev;
4566 
4567 	mutex_init(&dev->flow_db.lock);
4568 	mutex_init(&dev->cap_mask_mutex);
4569 	INIT_LIST_HEAD(&dev->qp_list);
4570 	spin_lock_init(&dev->reset_flow_resource_lock);
4571 
4572 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4573 	err = init_srcu_struct(&dev->mr_srcu);
4574 	if (err)
4575 		goto err_free_port;
4576 #endif
4577 
4578 	return 0;
4579 err_mp:
4580 	mlx5_ib_cleanup_multiport_master(dev);
4581 
4582 err_free_port:
4583 	kfree(dev->port);
4584 
4585 	return -ENOMEM;
4586 }
4587 
4588 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
4589 {
4590 	struct mlx5_core_dev *mdev = dev->mdev;
4591 	int err;
4592 
4593 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
4594 	dev->ib_dev.uverbs_cmd_mask	=
4595 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
4596 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
4597 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
4598 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
4599 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
4600 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
4601 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
4602 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
4603 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
4604 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
4605 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
4606 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
4607 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
4608 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
4609 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
4610 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
4611 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
4612 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
4613 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
4614 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
4615 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
4616 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
4617 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
4618 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
4619 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
4620 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
4621 	dev->ib_dev.uverbs_ex_cmd_mask =
4622 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
4623 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
4624 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
4625 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP)	|
4626 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ);
4627 
4628 	dev->ib_dev.query_device	= mlx5_ib_query_device;
4629 	dev->ib_dev.query_port		= mlx5_ib_query_port;
4630 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
4631 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
4632 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
4633 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
4634 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
4635 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
4636 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
4637 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
4638 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
4639 	dev->ib_dev.mmap		= mlx5_ib_mmap;
4640 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
4641 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
4642 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
4643 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
4644 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
4645 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
4646 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
4647 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
4648 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
4649 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
4650 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
4651 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
4652 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
4653 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
4654 	dev->ib_dev.post_send		= mlx5_ib_post_send;
4655 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
4656 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
4657 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
4658 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
4659 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
4660 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
4661 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
4662 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
4663 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
4664 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
4665 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
4666 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
4667 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
4668 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
4669 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
4670 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
4671 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
4672 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
4673 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
4674 	dev->ib_dev.get_vector_affinity	= mlx5_ib_get_vector_affinity;
4675 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads))
4676 		dev->ib_dev.alloc_rdma_netdev	= mlx5_ib_alloc_rdma_netdev;
4677 
4678 	if (mlx5_core_is_pf(mdev)) {
4679 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
4680 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
4681 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
4682 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
4683 	}
4684 
4685 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
4686 
4687 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
4688 
4689 	if (MLX5_CAP_GEN(mdev, imaicl)) {
4690 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
4691 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
4692 		dev->ib_dev.uverbs_cmd_mask |=
4693 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
4694 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
4695 	}
4696 
4697 	if (MLX5_CAP_GEN(mdev, xrc)) {
4698 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
4699 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
4700 		dev->ib_dev.uverbs_cmd_mask |=
4701 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
4702 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
4703 	}
4704 
4705 	dev->ib_dev.create_flow	= mlx5_ib_create_flow;
4706 	dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
4707 	dev->ib_dev.uverbs_ex_cmd_mask |=
4708 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
4709 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
4710 
4711 	err = init_node_data(dev);
4712 	if (err)
4713 		return err;
4714 
4715 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
4716 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
4717 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
4718 		mutex_init(&dev->lb_mutex);
4719 
4720 	return 0;
4721 }
4722 
4723 static int mlx5_ib_stage_roce_init(struct mlx5_ib_dev *dev)
4724 {
4725 	struct mlx5_core_dev *mdev = dev->mdev;
4726 	enum rdma_link_layer ll;
4727 	int port_type_cap;
4728 	u8 port_num;
4729 	int err;
4730 	int i;
4731 
4732 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4733 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4734 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4735 
4736 	if (ll == IB_LINK_LAYER_ETHERNET) {
4737 		for (i = 0; i < dev->num_ports; i++) {
4738 			dev->roce[i].dev = dev;
4739 			dev->roce[i].native_port_num = i + 1;
4740 			dev->roce[i].last_port_state = IB_PORT_DOWN;
4741 		}
4742 
4743 		dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
4744 		dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
4745 		dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
4746 		dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
4747 		dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
4748 		dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
4749 		dev->ib_dev.uverbs_ex_cmd_mask |=
4750 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
4751 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
4752 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
4753 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
4754 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
4755 		err = mlx5_enable_eth(dev, port_num);
4756 		if (err)
4757 			return err;
4758 	}
4759 
4760 	return 0;
4761 }
4762 
4763 static void mlx5_ib_stage_roce_cleanup(struct mlx5_ib_dev *dev)
4764 {
4765 	struct mlx5_core_dev *mdev = dev->mdev;
4766 	enum rdma_link_layer ll;
4767 	int port_type_cap;
4768 	u8 port_num;
4769 
4770 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4771 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4772 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4773 
4774 	if (ll == IB_LINK_LAYER_ETHERNET) {
4775 		mlx5_disable_eth(dev);
4776 		mlx5_remove_netdev_notifier(dev, port_num);
4777 	}
4778 }
4779 
4780 static int mlx5_ib_stage_dev_res_init(struct mlx5_ib_dev *dev)
4781 {
4782 	return create_dev_resources(&dev->devr);
4783 }
4784 
4785 static void mlx5_ib_stage_dev_res_cleanup(struct mlx5_ib_dev *dev)
4786 {
4787 	destroy_dev_resources(&dev->devr);
4788 }
4789 
4790 static int mlx5_ib_stage_odp_init(struct mlx5_ib_dev *dev)
4791 {
4792 	mlx5_ib_internal_fill_odp_caps(dev);
4793 
4794 	return mlx5_ib_odp_init_one(dev);
4795 }
4796 
4797 static int mlx5_ib_stage_counters_init(struct mlx5_ib_dev *dev)
4798 {
4799 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
4800 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
4801 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
4802 
4803 		return mlx5_ib_alloc_counters(dev);
4804 	}
4805 
4806 	return 0;
4807 }
4808 
4809 static void mlx5_ib_stage_counters_cleanup(struct mlx5_ib_dev *dev)
4810 {
4811 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
4812 		mlx5_ib_dealloc_counters(dev);
4813 }
4814 
4815 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
4816 {
4817 	return mlx5_ib_init_cong_debugfs(dev,
4818 					 mlx5_core_native_port_num(dev->mdev) - 1);
4819 }
4820 
4821 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
4822 {
4823 	mlx5_ib_cleanup_cong_debugfs(dev,
4824 				     mlx5_core_native_port_num(dev->mdev) - 1);
4825 }
4826 
4827 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
4828 {
4829 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4830 	if (!dev->mdev->priv.uar)
4831 		return -ENOMEM;
4832 	return 0;
4833 }
4834 
4835 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
4836 {
4837 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4838 }
4839 
4840 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4841 {
4842 	int err;
4843 
4844 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4845 	if (err)
4846 		return err;
4847 
4848 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4849 	if (err)
4850 		mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4851 
4852 	return err;
4853 }
4854 
4855 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4856 {
4857 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4858 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4859 }
4860 
4861 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4862 {
4863 	return ib_register_device(&dev->ib_dev, NULL);
4864 }
4865 
4866 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4867 {
4868 	ib_unregister_device(&dev->ib_dev);
4869 }
4870 
4871 static int mlx5_ib_stage_umr_res_init(struct mlx5_ib_dev *dev)
4872 {
4873 	return create_umr_res(dev);
4874 }
4875 
4876 static void mlx5_ib_stage_umr_res_cleanup(struct mlx5_ib_dev *dev)
4877 {
4878 	destroy_umrc_res(dev);
4879 }
4880 
4881 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4882 {
4883 	init_delay_drop(dev);
4884 
4885 	return 0;
4886 }
4887 
4888 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4889 {
4890 	cancel_delay_drop(dev);
4891 }
4892 
4893 static int mlx5_ib_stage_class_attr_init(struct mlx5_ib_dev *dev)
4894 {
4895 	int err;
4896 	int i;
4897 
4898 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
4899 		err = device_create_file(&dev->ib_dev.dev,
4900 					 mlx5_class_attributes[i]);
4901 		if (err)
4902 			return err;
4903 	}
4904 
4905 	return 0;
4906 }
4907 
4908 static void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4909 			     const struct mlx5_ib_profile *profile,
4910 			     int stage)
4911 {
4912 	/* Number of stages to cleanup */
4913 	while (stage) {
4914 		stage--;
4915 		if (profile->stage[stage].cleanup)
4916 			profile->stage[stage].cleanup(dev);
4917 	}
4918 
4919 	ib_dealloc_device((struct ib_device *)dev);
4920 }
4921 
4922 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev, u8 port_num);
4923 
4924 static void *__mlx5_ib_add(struct mlx5_core_dev *mdev,
4925 			   const struct mlx5_ib_profile *profile)
4926 {
4927 	struct mlx5_ib_dev *dev;
4928 	int err;
4929 	int i;
4930 
4931 	printk_once(KERN_INFO "%s", mlx5_version);
4932 
4933 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
4934 	if (!dev)
4935 		return NULL;
4936 
4937 	dev->mdev = mdev;
4938 	dev->num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4939 			     MLX5_CAP_GEN(mdev, num_vhca_ports));
4940 
4941 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4942 		if (profile->stage[i].init) {
4943 			err = profile->stage[i].init(dev);
4944 			if (err)
4945 				goto err_out;
4946 		}
4947 	}
4948 
4949 	dev->profile = profile;
4950 	dev->ib_active = true;
4951 
4952 	return dev;
4953 
4954 err_out:
4955 	__mlx5_ib_remove(dev, profile, i);
4956 
4957 	return NULL;
4958 }
4959 
4960 static const struct mlx5_ib_profile pf_profile = {
4961 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4962 		     mlx5_ib_stage_init_init,
4963 		     mlx5_ib_stage_init_cleanup),
4964 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4965 		     mlx5_ib_stage_caps_init,
4966 		     NULL),
4967 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4968 		     mlx5_ib_stage_roce_init,
4969 		     mlx5_ib_stage_roce_cleanup),
4970 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4971 		     mlx5_ib_stage_dev_res_init,
4972 		     mlx5_ib_stage_dev_res_cleanup),
4973 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4974 		     mlx5_ib_stage_odp_init,
4975 		     NULL),
4976 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4977 		     mlx5_ib_stage_counters_init,
4978 		     mlx5_ib_stage_counters_cleanup),
4979 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4980 		     mlx5_ib_stage_cong_debugfs_init,
4981 		     mlx5_ib_stage_cong_debugfs_cleanup),
4982 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4983 		     mlx5_ib_stage_uar_init,
4984 		     mlx5_ib_stage_uar_cleanup),
4985 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4986 		     mlx5_ib_stage_bfrag_init,
4987 		     mlx5_ib_stage_bfrag_cleanup),
4988 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4989 		     mlx5_ib_stage_ib_reg_init,
4990 		     mlx5_ib_stage_ib_reg_cleanup),
4991 	STAGE_CREATE(MLX5_IB_STAGE_UMR_RESOURCES,
4992 		     mlx5_ib_stage_umr_res_init,
4993 		     mlx5_ib_stage_umr_res_cleanup),
4994 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4995 		     mlx5_ib_stage_delay_drop_init,
4996 		     mlx5_ib_stage_delay_drop_cleanup),
4997 	STAGE_CREATE(MLX5_IB_STAGE_CLASS_ATTR,
4998 		     mlx5_ib_stage_class_attr_init,
4999 		     NULL),
5000 };
5001 
5002 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev, u8 port_num)
5003 {
5004 	struct mlx5_ib_multiport_info *mpi;
5005 	struct mlx5_ib_dev *dev;
5006 	bool bound = false;
5007 	int err;
5008 
5009 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
5010 	if (!mpi)
5011 		return NULL;
5012 
5013 	mpi->mdev = mdev;
5014 
5015 	err = mlx5_query_nic_vport_system_image_guid(mdev,
5016 						     &mpi->sys_image_guid);
5017 	if (err) {
5018 		kfree(mpi);
5019 		return NULL;
5020 	}
5021 
5022 	mutex_lock(&mlx5_ib_multiport_mutex);
5023 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
5024 		if (dev->sys_image_guid == mpi->sys_image_guid)
5025 			bound = mlx5_ib_bind_slave_port(dev, mpi);
5026 
5027 		if (bound) {
5028 			rdma_roce_rescan_device(&dev->ib_dev);
5029 			break;
5030 		}
5031 	}
5032 
5033 	if (!bound) {
5034 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
5035 		dev_dbg(&mdev->pdev->dev, "no suitable IB device found to bind to, added to unaffiliated list.\n");
5036 	} else {
5037 		mlx5_ib_dbg(dev, "bound port %u\n", port_num + 1);
5038 	}
5039 	mutex_unlock(&mlx5_ib_multiport_mutex);
5040 
5041 	return mpi;
5042 }
5043 
5044 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
5045 {
5046 	enum rdma_link_layer ll;
5047 	int port_type_cap;
5048 
5049 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5050 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5051 
5052 	if (mlx5_core_is_mp_slave(mdev) && ll == IB_LINK_LAYER_ETHERNET) {
5053 		u8 port_num = mlx5_core_native_port_num(mdev) - 1;
5054 
5055 		return mlx5_ib_add_slave_port(mdev, port_num);
5056 	}
5057 
5058 	return __mlx5_ib_add(mdev, &pf_profile);
5059 }
5060 
5061 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
5062 {
5063 	struct mlx5_ib_multiport_info *mpi;
5064 	struct mlx5_ib_dev *dev;
5065 
5066 	if (mlx5_core_is_mp_slave(mdev)) {
5067 		mpi = context;
5068 		mutex_lock(&mlx5_ib_multiport_mutex);
5069 		if (mpi->ibdev)
5070 			mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
5071 		list_del(&mpi->list);
5072 		mutex_unlock(&mlx5_ib_multiport_mutex);
5073 		return;
5074 	}
5075 
5076 	dev = context;
5077 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
5078 }
5079 
5080 static struct mlx5_interface mlx5_ib_interface = {
5081 	.add            = mlx5_ib_add,
5082 	.remove         = mlx5_ib_remove,
5083 	.event          = mlx5_ib_event,
5084 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5085 	.pfault		= mlx5_ib_pfault,
5086 #endif
5087 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
5088 };
5089 
5090 static int __init mlx5_ib_init(void)
5091 {
5092 	int err;
5093 
5094 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
5095 	if (!mlx5_ib_event_wq)
5096 		return -ENOMEM;
5097 
5098 	mlx5_ib_odp_init();
5099 
5100 	err = mlx5_register_interface(&mlx5_ib_interface);
5101 
5102 	return err;
5103 }
5104 
5105 static void __exit mlx5_ib_cleanup(void)
5106 {
5107 	mlx5_unregister_interface(&mlx5_ib_interface);
5108 	destroy_workqueue(mlx5_ib_event_wq);
5109 }
5110 
5111 module_init(mlx5_ib_init);
5112 module_exit(mlx5_ib_cleanup);
5113