xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 97e6f135)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
4  * Copyright (c) 2020, Intel Corporation. All rights reserved.
5  */
6 
7 #include <linux/debugfs.h>
8 #include <linux/highmem.h>
9 #include <linux/module.h>
10 #include <linux/init.h>
11 #include <linux/errno.h>
12 #include <linux/pci.h>
13 #include <linux/dma-mapping.h>
14 #include <linux/slab.h>
15 #include <linux/bitmap.h>
16 #include <linux/sched.h>
17 #include <linux/sched/mm.h>
18 #include <linux/sched/task.h>
19 #include <linux/delay.h>
20 #include <rdma/ib_user_verbs.h>
21 #include <rdma/ib_addr.h>
22 #include <rdma/ib_cache.h>
23 #include <linux/mlx5/port.h>
24 #include <linux/mlx5/vport.h>
25 #include <linux/mlx5/fs.h>
26 #include <linux/mlx5/eswitch.h>
27 #include <linux/list.h>
28 #include <rdma/ib_smi.h>
29 #include <rdma/ib_umem_odp.h>
30 #include <rdma/lag.h>
31 #include <linux/in.h>
32 #include <linux/etherdevice.h>
33 #include "mlx5_ib.h"
34 #include "ib_rep.h"
35 #include "cmd.h"
36 #include "devx.h"
37 #include "dm.h"
38 #include "fs.h"
39 #include "srq.h"
40 #include "qp.h"
41 #include "wr.h"
42 #include "restrack.h"
43 #include "counters.h"
44 #include "umr.h"
45 #include <rdma/uverbs_std_types.h>
46 #include <rdma/uverbs_ioctl.h>
47 #include <rdma/mlx5_user_ioctl_verbs.h>
48 #include <rdma/mlx5_user_ioctl_cmds.h>
49 
50 #define UVERBS_MODULE_NAME mlx5_ib
51 #include <rdma/uverbs_named_ioctl.h>
52 
53 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
54 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
55 MODULE_LICENSE("Dual BSD/GPL");
56 
57 struct mlx5_ib_event_work {
58 	struct work_struct	work;
59 	union {
60 		struct mlx5_ib_dev	      *dev;
61 		struct mlx5_ib_multiport_info *mpi;
62 	};
63 	bool			is_slave;
64 	unsigned int		event;
65 	void			*param;
66 };
67 
68 enum {
69 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
70 };
71 
72 static struct workqueue_struct *mlx5_ib_event_wq;
73 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
74 static LIST_HEAD(mlx5_ib_dev_list);
75 /*
76  * This mutex should be held when accessing either of the above lists
77  */
78 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
79 
80 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
81 {
82 	struct mlx5_ib_dev *dev;
83 
84 	mutex_lock(&mlx5_ib_multiport_mutex);
85 	dev = mpi->ibdev;
86 	mutex_unlock(&mlx5_ib_multiport_mutex);
87 	return dev;
88 }
89 
90 static enum rdma_link_layer
91 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
92 {
93 	switch (port_type_cap) {
94 	case MLX5_CAP_PORT_TYPE_IB:
95 		return IB_LINK_LAYER_INFINIBAND;
96 	case MLX5_CAP_PORT_TYPE_ETH:
97 		return IB_LINK_LAYER_ETHERNET;
98 	default:
99 		return IB_LINK_LAYER_UNSPECIFIED;
100 	}
101 }
102 
103 static enum rdma_link_layer
104 mlx5_ib_port_link_layer(struct ib_device *device, u32 port_num)
105 {
106 	struct mlx5_ib_dev *dev = to_mdev(device);
107 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
108 
109 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
110 }
111 
112 static int get_port_state(struct ib_device *ibdev,
113 			  u32 port_num,
114 			  enum ib_port_state *state)
115 {
116 	struct ib_port_attr attr;
117 	int ret;
118 
119 	memset(&attr, 0, sizeof(attr));
120 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
121 	if (!ret)
122 		*state = attr.state;
123 	return ret;
124 }
125 
126 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
127 					   struct net_device *ndev,
128 					   struct net_device *upper,
129 					   u32 *port_num)
130 {
131 	struct net_device *rep_ndev;
132 	struct mlx5_ib_port *port;
133 	int i;
134 
135 	for (i = 0; i < dev->num_ports; i++) {
136 		port  = &dev->port[i];
137 		if (!port->rep)
138 			continue;
139 
140 		if (upper == ndev && port->rep->vport == MLX5_VPORT_UPLINK) {
141 			*port_num = i + 1;
142 			return &port->roce;
143 		}
144 
145 		if (upper && port->rep->vport == MLX5_VPORT_UPLINK)
146 			continue;
147 
148 		read_lock(&port->roce.netdev_lock);
149 		rep_ndev = mlx5_ib_get_rep_netdev(port->rep->esw,
150 						  port->rep->vport);
151 		if (rep_ndev == ndev) {
152 			read_unlock(&port->roce.netdev_lock);
153 			*port_num = i + 1;
154 			return &port->roce;
155 		}
156 		read_unlock(&port->roce.netdev_lock);
157 	}
158 
159 	return NULL;
160 }
161 
162 static int mlx5_netdev_event(struct notifier_block *this,
163 			     unsigned long event, void *ptr)
164 {
165 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
166 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
167 	u32 port_num = roce->native_port_num;
168 	struct mlx5_core_dev *mdev;
169 	struct mlx5_ib_dev *ibdev;
170 
171 	ibdev = roce->dev;
172 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
173 	if (!mdev)
174 		return NOTIFY_DONE;
175 
176 	switch (event) {
177 	case NETDEV_REGISTER:
178 		/* Should already be registered during the load */
179 		if (ibdev->is_rep)
180 			break;
181 		write_lock(&roce->netdev_lock);
182 		if (ndev->dev.parent == mdev->device)
183 			roce->netdev = ndev;
184 		write_unlock(&roce->netdev_lock);
185 		break;
186 
187 	case NETDEV_UNREGISTER:
188 		/* In case of reps, ib device goes away before the netdevs */
189 		write_lock(&roce->netdev_lock);
190 		if (roce->netdev == ndev)
191 			roce->netdev = NULL;
192 		write_unlock(&roce->netdev_lock);
193 		break;
194 
195 	case NETDEV_CHANGE:
196 	case NETDEV_UP:
197 	case NETDEV_DOWN: {
198 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
199 		struct net_device *upper = NULL;
200 
201 		if (lag_ndev) {
202 			upper = netdev_master_upper_dev_get(lag_ndev);
203 			dev_put(lag_ndev);
204 		}
205 
206 		if (ibdev->is_rep)
207 			roce = mlx5_get_rep_roce(ibdev, ndev, upper, &port_num);
208 		if (!roce)
209 			return NOTIFY_DONE;
210 		if ((upper == ndev ||
211 		     ((!upper || ibdev->is_rep) && ndev == roce->netdev)) &&
212 		    ibdev->ib_active) {
213 			struct ib_event ibev = { };
214 			enum ib_port_state port_state;
215 
216 			if (get_port_state(&ibdev->ib_dev, port_num,
217 					   &port_state))
218 				goto done;
219 
220 			if (roce->last_port_state == port_state)
221 				goto done;
222 
223 			roce->last_port_state = port_state;
224 			ibev.device = &ibdev->ib_dev;
225 			if (port_state == IB_PORT_DOWN)
226 				ibev.event = IB_EVENT_PORT_ERR;
227 			else if (port_state == IB_PORT_ACTIVE)
228 				ibev.event = IB_EVENT_PORT_ACTIVE;
229 			else
230 				goto done;
231 
232 			ibev.element.port_num = port_num;
233 			ib_dispatch_event(&ibev);
234 		}
235 		break;
236 	}
237 
238 	default:
239 		break;
240 	}
241 done:
242 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
243 	return NOTIFY_DONE;
244 }
245 
246 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
247 					     u32 port_num)
248 {
249 	struct mlx5_ib_dev *ibdev = to_mdev(device);
250 	struct net_device *ndev;
251 	struct mlx5_core_dev *mdev;
252 
253 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
254 	if (!mdev)
255 		return NULL;
256 
257 	ndev = mlx5_lag_get_roce_netdev(mdev);
258 	if (ndev)
259 		goto out;
260 
261 	/* Ensure ndev does not disappear before we invoke dev_hold()
262 	 */
263 	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
264 	ndev = ibdev->port[port_num - 1].roce.netdev;
265 	if (ndev)
266 		dev_hold(ndev);
267 	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
268 
269 out:
270 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
271 	return ndev;
272 }
273 
274 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
275 						   u32 ib_port_num,
276 						   u32 *native_port_num)
277 {
278 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
279 							  ib_port_num);
280 	struct mlx5_core_dev *mdev = NULL;
281 	struct mlx5_ib_multiport_info *mpi;
282 	struct mlx5_ib_port *port;
283 
284 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
285 	    ll != IB_LINK_LAYER_ETHERNET) {
286 		if (native_port_num)
287 			*native_port_num = ib_port_num;
288 		return ibdev->mdev;
289 	}
290 
291 	if (native_port_num)
292 		*native_port_num = 1;
293 
294 	port = &ibdev->port[ib_port_num - 1];
295 	spin_lock(&port->mp.mpi_lock);
296 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
297 	if (mpi && !mpi->unaffiliate) {
298 		mdev = mpi->mdev;
299 		/* If it's the master no need to refcount, it'll exist
300 		 * as long as the ib_dev exists.
301 		 */
302 		if (!mpi->is_master)
303 			mpi->mdev_refcnt++;
304 	}
305 	spin_unlock(&port->mp.mpi_lock);
306 
307 	return mdev;
308 }
309 
310 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 port_num)
311 {
312 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
313 							  port_num);
314 	struct mlx5_ib_multiport_info *mpi;
315 	struct mlx5_ib_port *port;
316 
317 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
318 		return;
319 
320 	port = &ibdev->port[port_num - 1];
321 
322 	spin_lock(&port->mp.mpi_lock);
323 	mpi = ibdev->port[port_num - 1].mp.mpi;
324 	if (mpi->is_master)
325 		goto out;
326 
327 	mpi->mdev_refcnt--;
328 	if (mpi->unaffiliate)
329 		complete(&mpi->unref_comp);
330 out:
331 	spin_unlock(&port->mp.mpi_lock);
332 }
333 
334 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper,
335 					   u16 *active_speed, u8 *active_width)
336 {
337 	switch (eth_proto_oper) {
338 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
339 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
340 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
341 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
342 		*active_width = IB_WIDTH_1X;
343 		*active_speed = IB_SPEED_SDR;
344 		break;
345 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
346 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
347 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
348 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
349 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
350 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
351 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
352 		*active_width = IB_WIDTH_1X;
353 		*active_speed = IB_SPEED_QDR;
354 		break;
355 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
356 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
357 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
358 		*active_width = IB_WIDTH_1X;
359 		*active_speed = IB_SPEED_EDR;
360 		break;
361 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
362 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
363 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
364 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
365 		*active_width = IB_WIDTH_4X;
366 		*active_speed = IB_SPEED_QDR;
367 		break;
368 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
369 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
370 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
371 		*active_width = IB_WIDTH_1X;
372 		*active_speed = IB_SPEED_HDR;
373 		break;
374 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
375 		*active_width = IB_WIDTH_4X;
376 		*active_speed = IB_SPEED_FDR;
377 		break;
378 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
379 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
380 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
381 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
382 		*active_width = IB_WIDTH_4X;
383 		*active_speed = IB_SPEED_EDR;
384 		break;
385 	default:
386 		return -EINVAL;
387 	}
388 
389 	return 0;
390 }
391 
392 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed,
393 					u8 *active_width)
394 {
395 	switch (eth_proto_oper) {
396 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
397 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
398 		*active_width = IB_WIDTH_1X;
399 		*active_speed = IB_SPEED_SDR;
400 		break;
401 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
402 		*active_width = IB_WIDTH_1X;
403 		*active_speed = IB_SPEED_DDR;
404 		break;
405 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
406 		*active_width = IB_WIDTH_1X;
407 		*active_speed = IB_SPEED_QDR;
408 		break;
409 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
410 		*active_width = IB_WIDTH_4X;
411 		*active_speed = IB_SPEED_QDR;
412 		break;
413 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
414 		*active_width = IB_WIDTH_1X;
415 		*active_speed = IB_SPEED_EDR;
416 		break;
417 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
418 		*active_width = IB_WIDTH_2X;
419 		*active_speed = IB_SPEED_EDR;
420 		break;
421 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
422 		*active_width = IB_WIDTH_1X;
423 		*active_speed = IB_SPEED_HDR;
424 		break;
425 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
426 		*active_width = IB_WIDTH_4X;
427 		*active_speed = IB_SPEED_EDR;
428 		break;
429 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
430 		*active_width = IB_WIDTH_2X;
431 		*active_speed = IB_SPEED_HDR;
432 		break;
433 	case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR):
434 		*active_width = IB_WIDTH_1X;
435 		*active_speed = IB_SPEED_NDR;
436 		break;
437 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
438 		*active_width = IB_WIDTH_4X;
439 		*active_speed = IB_SPEED_HDR;
440 		break;
441 	case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2):
442 		*active_width = IB_WIDTH_2X;
443 		*active_speed = IB_SPEED_NDR;
444 		break;
445 	case MLX5E_PROT_MASK(MLX5E_400GAUI_8):
446 		*active_width = IB_WIDTH_8X;
447 		*active_speed = IB_SPEED_HDR;
448 		break;
449 	case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4):
450 		*active_width = IB_WIDTH_4X;
451 		*active_speed = IB_SPEED_NDR;
452 		break;
453 	default:
454 		return -EINVAL;
455 	}
456 
457 	return 0;
458 }
459 
460 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
461 				    u8 *active_width, bool ext)
462 {
463 	return ext ?
464 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
465 					     active_width) :
466 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
467 						active_width);
468 }
469 
470 static int mlx5_query_port_roce(struct ib_device *device, u32 port_num,
471 				struct ib_port_attr *props)
472 {
473 	struct mlx5_ib_dev *dev = to_mdev(device);
474 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
475 	struct mlx5_core_dev *mdev;
476 	struct net_device *ndev, *upper;
477 	enum ib_mtu ndev_ib_mtu;
478 	bool put_mdev = true;
479 	u32 eth_prot_oper;
480 	u32 mdev_port_num;
481 	bool ext;
482 	int err;
483 
484 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
485 	if (!mdev) {
486 		/* This means the port isn't affiliated yet. Get the
487 		 * info for the master port instead.
488 		 */
489 		put_mdev = false;
490 		mdev = dev->mdev;
491 		mdev_port_num = 1;
492 		port_num = 1;
493 	}
494 
495 	/* Possible bad flows are checked before filling out props so in case
496 	 * of an error it will still be zeroed out.
497 	 * Use native port in case of reps
498 	 */
499 	if (dev->is_rep)
500 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
501 					   1);
502 	else
503 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
504 					   mdev_port_num);
505 	if (err)
506 		goto out;
507 	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
508 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
509 
510 	props->active_width     = IB_WIDTH_4X;
511 	props->active_speed     = IB_SPEED_QDR;
512 
513 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
514 				 &props->active_width, ext);
515 
516 	if (!dev->is_rep && dev->mdev->roce.roce_en) {
517 		u16 qkey_viol_cntr;
518 
519 		props->port_cap_flags |= IB_PORT_CM_SUP;
520 		props->ip_gids = true;
521 		props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
522 						   roce_address_table_size);
523 		mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
524 		props->qkey_viol_cntr = qkey_viol_cntr;
525 	}
526 	props->max_mtu          = IB_MTU_4096;
527 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
528 	props->pkey_tbl_len     = 1;
529 	props->state            = IB_PORT_DOWN;
530 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
531 
532 	/* If this is a stub query for an unaffiliated port stop here */
533 	if (!put_mdev)
534 		goto out;
535 
536 	ndev = mlx5_ib_get_netdev(device, port_num);
537 	if (!ndev)
538 		goto out;
539 
540 	if (dev->lag_active) {
541 		rcu_read_lock();
542 		upper = netdev_master_upper_dev_get_rcu(ndev);
543 		if (upper) {
544 			dev_put(ndev);
545 			ndev = upper;
546 			dev_hold(ndev);
547 		}
548 		rcu_read_unlock();
549 	}
550 
551 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
552 		props->state      = IB_PORT_ACTIVE;
553 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
554 	}
555 
556 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
557 
558 	dev_put(ndev);
559 
560 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
561 out:
562 	if (put_mdev)
563 		mlx5_ib_put_native_port_mdev(dev, port_num);
564 	return err;
565 }
566 
567 static int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num,
568 			 unsigned int index, const union ib_gid *gid,
569 			 const struct ib_gid_attr *attr)
570 {
571 	enum ib_gid_type gid_type;
572 	u16 vlan_id = 0xffff;
573 	u8 roce_version = 0;
574 	u8 roce_l3_type = 0;
575 	u8 mac[ETH_ALEN];
576 	int ret;
577 
578 	gid_type = attr->gid_type;
579 	if (gid) {
580 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
581 		if (ret)
582 			return ret;
583 	}
584 
585 	switch (gid_type) {
586 	case IB_GID_TYPE_ROCE:
587 		roce_version = MLX5_ROCE_VERSION_1;
588 		break;
589 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
590 		roce_version = MLX5_ROCE_VERSION_2;
591 		if (gid && ipv6_addr_v4mapped((void *)gid))
592 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
593 		else
594 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
595 		break;
596 
597 	default:
598 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
599 	}
600 
601 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
602 				      roce_l3_type, gid->raw, mac,
603 				      vlan_id < VLAN_CFI_MASK, vlan_id,
604 				      port_num);
605 }
606 
607 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
608 			   __always_unused void **context)
609 {
610 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
611 			     attr->index, &attr->gid, attr);
612 }
613 
614 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
615 			   __always_unused void **context)
616 {
617 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
618 			     attr->index, NULL, attr);
619 }
620 
621 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
622 				   const struct ib_gid_attr *attr)
623 {
624 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
625 		return 0;
626 
627 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
628 }
629 
630 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
631 {
632 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
633 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
634 	return 0;
635 }
636 
637 enum {
638 	MLX5_VPORT_ACCESS_METHOD_MAD,
639 	MLX5_VPORT_ACCESS_METHOD_HCA,
640 	MLX5_VPORT_ACCESS_METHOD_NIC,
641 };
642 
643 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
644 {
645 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
646 		return MLX5_VPORT_ACCESS_METHOD_MAD;
647 
648 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
649 	    IB_LINK_LAYER_ETHERNET)
650 		return MLX5_VPORT_ACCESS_METHOD_NIC;
651 
652 	return MLX5_VPORT_ACCESS_METHOD_HCA;
653 }
654 
655 static void get_atomic_caps(struct mlx5_ib_dev *dev,
656 			    u8 atomic_size_qp,
657 			    struct ib_device_attr *props)
658 {
659 	u8 tmp;
660 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
661 	u8 atomic_req_8B_endianness_mode =
662 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
663 
664 	/* Check if HW supports 8 bytes standard atomic operations and capable
665 	 * of host endianness respond
666 	 */
667 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
668 	if (((atomic_operations & tmp) == tmp) &&
669 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
670 	    (atomic_req_8B_endianness_mode)) {
671 		props->atomic_cap = IB_ATOMIC_HCA;
672 	} else {
673 		props->atomic_cap = IB_ATOMIC_NONE;
674 	}
675 }
676 
677 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
678 			       struct ib_device_attr *props)
679 {
680 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
681 
682 	get_atomic_caps(dev, atomic_size_qp, props);
683 }
684 
685 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
686 					__be64 *sys_image_guid)
687 {
688 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
689 	struct mlx5_core_dev *mdev = dev->mdev;
690 	u64 tmp;
691 	int err;
692 
693 	switch (mlx5_get_vport_access_method(ibdev)) {
694 	case MLX5_VPORT_ACCESS_METHOD_MAD:
695 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
696 							    sys_image_guid);
697 
698 	case MLX5_VPORT_ACCESS_METHOD_HCA:
699 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
700 		break;
701 
702 	case MLX5_VPORT_ACCESS_METHOD_NIC:
703 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
704 		break;
705 
706 	default:
707 		return -EINVAL;
708 	}
709 
710 	if (!err)
711 		*sys_image_guid = cpu_to_be64(tmp);
712 
713 	return err;
714 
715 }
716 
717 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
718 				u16 *max_pkeys)
719 {
720 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
721 	struct mlx5_core_dev *mdev = dev->mdev;
722 
723 	switch (mlx5_get_vport_access_method(ibdev)) {
724 	case MLX5_VPORT_ACCESS_METHOD_MAD:
725 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
726 
727 	case MLX5_VPORT_ACCESS_METHOD_HCA:
728 	case MLX5_VPORT_ACCESS_METHOD_NIC:
729 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
730 						pkey_table_size));
731 		return 0;
732 
733 	default:
734 		return -EINVAL;
735 	}
736 }
737 
738 static int mlx5_query_vendor_id(struct ib_device *ibdev,
739 				u32 *vendor_id)
740 {
741 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
742 
743 	switch (mlx5_get_vport_access_method(ibdev)) {
744 	case MLX5_VPORT_ACCESS_METHOD_MAD:
745 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
746 
747 	case MLX5_VPORT_ACCESS_METHOD_HCA:
748 	case MLX5_VPORT_ACCESS_METHOD_NIC:
749 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
750 
751 	default:
752 		return -EINVAL;
753 	}
754 }
755 
756 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
757 				__be64 *node_guid)
758 {
759 	u64 tmp;
760 	int err;
761 
762 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
763 	case MLX5_VPORT_ACCESS_METHOD_MAD:
764 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
765 
766 	case MLX5_VPORT_ACCESS_METHOD_HCA:
767 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
768 		break;
769 
770 	case MLX5_VPORT_ACCESS_METHOD_NIC:
771 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
772 		break;
773 
774 	default:
775 		return -EINVAL;
776 	}
777 
778 	if (!err)
779 		*node_guid = cpu_to_be64(tmp);
780 
781 	return err;
782 }
783 
784 struct mlx5_reg_node_desc {
785 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
786 };
787 
788 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
789 {
790 	struct mlx5_reg_node_desc in;
791 
792 	if (mlx5_use_mad_ifc(dev))
793 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
794 
795 	memset(&in, 0, sizeof(in));
796 
797 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
798 				    sizeof(struct mlx5_reg_node_desc),
799 				    MLX5_REG_NODE_DESC, 0, 0);
800 }
801 
802 static int mlx5_ib_query_device(struct ib_device *ibdev,
803 				struct ib_device_attr *props,
804 				struct ib_udata *uhw)
805 {
806 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
807 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
808 	struct mlx5_core_dev *mdev = dev->mdev;
809 	int err = -ENOMEM;
810 	int max_sq_desc;
811 	int max_rq_sg;
812 	int max_sq_sg;
813 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
814 	bool raw_support = !mlx5_core_mp_enabled(mdev);
815 	struct mlx5_ib_query_device_resp resp = {};
816 	size_t resp_len;
817 	u64 max_tso;
818 
819 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
820 	if (uhw_outlen && uhw_outlen < resp_len)
821 		return -EINVAL;
822 
823 	resp.response_length = resp_len;
824 
825 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
826 		return -EINVAL;
827 
828 	memset(props, 0, sizeof(*props));
829 	err = mlx5_query_system_image_guid(ibdev,
830 					   &props->sys_image_guid);
831 	if (err)
832 		return err;
833 
834 	props->max_pkeys = dev->pkey_table_len;
835 
836 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
837 	if (err)
838 		return err;
839 
840 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
841 		(fw_rev_min(dev->mdev) << 16) |
842 		fw_rev_sub(dev->mdev);
843 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
844 		IB_DEVICE_PORT_ACTIVE_EVENT		|
845 		IB_DEVICE_SYS_IMAGE_GUID		|
846 		IB_DEVICE_RC_RNR_NAK_GEN;
847 
848 	if (MLX5_CAP_GEN(mdev, pkv))
849 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
850 	if (MLX5_CAP_GEN(mdev, qkv))
851 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
852 	if (MLX5_CAP_GEN(mdev, apm))
853 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
854 	if (MLX5_CAP_GEN(mdev, xrc))
855 		props->device_cap_flags |= IB_DEVICE_XRC;
856 	if (MLX5_CAP_GEN(mdev, imaicl)) {
857 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
858 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
859 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
860 		/* We support 'Gappy' memory registration too */
861 		props->kernel_cap_flags |= IBK_SG_GAPS_REG;
862 	}
863 	/* IB_WR_REG_MR always requires changing the entity size with UMR */
864 	if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
865 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
866 	if (MLX5_CAP_GEN(mdev, sho)) {
867 		props->kernel_cap_flags |= IBK_INTEGRITY_HANDOVER;
868 		/* At this stage no support for signature handover */
869 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
870 				      IB_PROT_T10DIF_TYPE_2 |
871 				      IB_PROT_T10DIF_TYPE_3;
872 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
873 				       IB_GUARD_T10DIF_CSUM;
874 	}
875 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
876 		props->kernel_cap_flags |= IBK_BLOCK_MULTICAST_LOOPBACK;
877 
878 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
879 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
880 			/* Legacy bit to support old userspace libraries */
881 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
882 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
883 		}
884 
885 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
886 			props->raw_packet_caps |=
887 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
888 
889 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
890 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
891 			if (max_tso) {
892 				resp.tso_caps.max_tso = 1 << max_tso;
893 				resp.tso_caps.supported_qpts |=
894 					1 << IB_QPT_RAW_PACKET;
895 				resp.response_length += sizeof(resp.tso_caps);
896 			}
897 		}
898 
899 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
900 			resp.rss_caps.rx_hash_function =
901 						MLX5_RX_HASH_FUNC_TOEPLITZ;
902 			resp.rss_caps.rx_hash_fields_mask =
903 						MLX5_RX_HASH_SRC_IPV4 |
904 						MLX5_RX_HASH_DST_IPV4 |
905 						MLX5_RX_HASH_SRC_IPV6 |
906 						MLX5_RX_HASH_DST_IPV6 |
907 						MLX5_RX_HASH_SRC_PORT_TCP |
908 						MLX5_RX_HASH_DST_PORT_TCP |
909 						MLX5_RX_HASH_SRC_PORT_UDP |
910 						MLX5_RX_HASH_DST_PORT_UDP |
911 						MLX5_RX_HASH_INNER;
912 			resp.response_length += sizeof(resp.rss_caps);
913 		}
914 	} else {
915 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
916 			resp.response_length += sizeof(resp.tso_caps);
917 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
918 			resp.response_length += sizeof(resp.rss_caps);
919 	}
920 
921 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
922 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
923 		props->kernel_cap_flags |= IBK_UD_TSO;
924 	}
925 
926 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
927 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
928 	    raw_support)
929 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
930 
931 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
932 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
933 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
934 
935 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
936 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
937 	    raw_support) {
938 		/* Legacy bit to support old userspace libraries */
939 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
940 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
941 	}
942 
943 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
944 		props->max_dm_size =
945 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
946 	}
947 
948 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
949 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
950 
951 	if (MLX5_CAP_GEN(mdev, end_pad))
952 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
953 
954 	props->vendor_part_id	   = mdev->pdev->device;
955 	props->hw_ver		   = mdev->pdev->revision;
956 
957 	props->max_mr_size	   = ~0ull;
958 	props->page_size_cap	   = ~(min_page_size - 1);
959 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
960 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
961 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
962 		     sizeof(struct mlx5_wqe_data_seg);
963 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
964 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
965 		     sizeof(struct mlx5_wqe_raddr_seg)) /
966 		sizeof(struct mlx5_wqe_data_seg);
967 	props->max_send_sge = max_sq_sg;
968 	props->max_recv_sge = max_rq_sg;
969 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
970 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
971 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
972 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
973 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
974 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
975 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
976 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
977 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
978 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
979 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
980 	props->max_srq_sge	   = max_rq_sg - 1;
981 	props->max_fast_reg_page_list_len =
982 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
983 	props->max_pi_fast_reg_page_list_len =
984 		props->max_fast_reg_page_list_len / 2;
985 	props->max_sgl_rd =
986 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
987 	get_atomic_caps_qp(dev, props);
988 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
989 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
990 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
991 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
992 					   props->max_mcast_grp;
993 	props->max_ah = INT_MAX;
994 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
995 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
996 
997 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
998 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
999 			props->kernel_cap_flags |= IBK_ON_DEMAND_PAGING;
1000 		props->odp_caps = dev->odp_caps;
1001 		if (!uhw) {
1002 			/* ODP for kernel QPs is not implemented for receive
1003 			 * WQEs and SRQ WQEs
1004 			 */
1005 			props->odp_caps.per_transport_caps.rc_odp_caps &=
1006 				~(IB_ODP_SUPPORT_READ |
1007 				  IB_ODP_SUPPORT_SRQ_RECV);
1008 			props->odp_caps.per_transport_caps.uc_odp_caps &=
1009 				~(IB_ODP_SUPPORT_READ |
1010 				  IB_ODP_SUPPORT_SRQ_RECV);
1011 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1012 				~(IB_ODP_SUPPORT_READ |
1013 				  IB_ODP_SUPPORT_SRQ_RECV);
1014 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1015 				~(IB_ODP_SUPPORT_READ |
1016 				  IB_ODP_SUPPORT_SRQ_RECV);
1017 		}
1018 	}
1019 
1020 	if (mlx5_core_is_vf(mdev))
1021 		props->kernel_cap_flags |= IBK_VIRTUAL_FUNCTION;
1022 
1023 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1024 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1025 		props->rss_caps.max_rwq_indirection_tables =
1026 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1027 		props->rss_caps.max_rwq_indirection_table_size =
1028 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1029 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1030 		props->max_wq_type_rq =
1031 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1032 	}
1033 
1034 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1035 		props->tm_caps.max_num_tags =
1036 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1037 		props->tm_caps.max_ops =
1038 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1039 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1040 	}
1041 
1042 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1043 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1044 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1045 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1046 	}
1047 
1048 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1049 		props->cq_caps.max_cq_moderation_count =
1050 						MLX5_MAX_CQ_COUNT;
1051 		props->cq_caps.max_cq_moderation_period =
1052 						MLX5_MAX_CQ_PERIOD;
1053 	}
1054 
1055 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1056 		resp.response_length += sizeof(resp.cqe_comp_caps);
1057 
1058 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1059 			resp.cqe_comp_caps.max_num =
1060 				MLX5_CAP_GEN(dev->mdev,
1061 					     cqe_compression_max_num);
1062 
1063 			resp.cqe_comp_caps.supported_format =
1064 				MLX5_IB_CQE_RES_FORMAT_HASH |
1065 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1066 
1067 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1068 				resp.cqe_comp_caps.supported_format |=
1069 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1070 		}
1071 	}
1072 
1073 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1074 	    raw_support) {
1075 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1076 		    MLX5_CAP_GEN(mdev, qos)) {
1077 			resp.packet_pacing_caps.qp_rate_limit_max =
1078 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1079 			resp.packet_pacing_caps.qp_rate_limit_min =
1080 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1081 			resp.packet_pacing_caps.supported_qpts |=
1082 				1 << IB_QPT_RAW_PACKET;
1083 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1084 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1085 				resp.packet_pacing_caps.cap_flags |=
1086 					MLX5_IB_PP_SUPPORT_BURST;
1087 		}
1088 		resp.response_length += sizeof(resp.packet_pacing_caps);
1089 	}
1090 
1091 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1092 	    uhw_outlen) {
1093 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1094 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1095 				MLX5_IB_ALLOW_MPW;
1096 
1097 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1098 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1099 				MLX5_IB_SUPPORT_EMPW;
1100 
1101 		resp.response_length +=
1102 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1103 	}
1104 
1105 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1106 		resp.response_length += sizeof(resp.flags);
1107 
1108 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1109 			resp.flags |=
1110 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1111 
1112 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1113 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1114 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1115 			resp.flags |=
1116 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1117 
1118 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1119 	}
1120 
1121 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1122 		resp.response_length += sizeof(resp.sw_parsing_caps);
1123 		if (MLX5_CAP_ETH(mdev, swp)) {
1124 			resp.sw_parsing_caps.sw_parsing_offloads |=
1125 				MLX5_IB_SW_PARSING;
1126 
1127 			if (MLX5_CAP_ETH(mdev, swp_csum))
1128 				resp.sw_parsing_caps.sw_parsing_offloads |=
1129 					MLX5_IB_SW_PARSING_CSUM;
1130 
1131 			if (MLX5_CAP_ETH(mdev, swp_lso))
1132 				resp.sw_parsing_caps.sw_parsing_offloads |=
1133 					MLX5_IB_SW_PARSING_LSO;
1134 
1135 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1136 				resp.sw_parsing_caps.supported_qpts =
1137 					BIT(IB_QPT_RAW_PACKET);
1138 		}
1139 	}
1140 
1141 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1142 	    raw_support) {
1143 		resp.response_length += sizeof(resp.striding_rq_caps);
1144 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1145 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1146 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1147 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1148 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1149 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1150 				resp.striding_rq_caps
1151 					.min_single_wqe_log_num_of_strides =
1152 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1153 			else
1154 				resp.striding_rq_caps
1155 					.min_single_wqe_log_num_of_strides =
1156 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1157 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1158 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1159 			resp.striding_rq_caps.supported_qpts =
1160 				BIT(IB_QPT_RAW_PACKET);
1161 		}
1162 	}
1163 
1164 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1165 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1166 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1167 			resp.tunnel_offloads_caps |=
1168 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1169 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1170 			resp.tunnel_offloads_caps |=
1171 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1172 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1173 			resp.tunnel_offloads_caps |=
1174 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1175 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1176 			resp.tunnel_offloads_caps |=
1177 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1178 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1179 			resp.tunnel_offloads_caps |=
1180 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1181 	}
1182 
1183 	if (offsetofend(typeof(resp), dci_streams_caps) <= uhw_outlen) {
1184 		resp.response_length += sizeof(resp.dci_streams_caps);
1185 
1186 		resp.dci_streams_caps.max_log_num_concurent =
1187 			MLX5_CAP_GEN(mdev, log_max_dci_stream_channels);
1188 
1189 		resp.dci_streams_caps.max_log_num_errored =
1190 			MLX5_CAP_GEN(mdev, log_max_dci_errored_streams);
1191 	}
1192 
1193 	if (uhw_outlen) {
1194 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1195 
1196 		if (err)
1197 			return err;
1198 	}
1199 
1200 	return 0;
1201 }
1202 
1203 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1204 				   u8 *ib_width)
1205 {
1206 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1207 
1208 	if (active_width & MLX5_PTYS_WIDTH_1X)
1209 		*ib_width = IB_WIDTH_1X;
1210 	else if (active_width & MLX5_PTYS_WIDTH_2X)
1211 		*ib_width = IB_WIDTH_2X;
1212 	else if (active_width & MLX5_PTYS_WIDTH_4X)
1213 		*ib_width = IB_WIDTH_4X;
1214 	else if (active_width & MLX5_PTYS_WIDTH_8X)
1215 		*ib_width = IB_WIDTH_8X;
1216 	else if (active_width & MLX5_PTYS_WIDTH_12X)
1217 		*ib_width = IB_WIDTH_12X;
1218 	else {
1219 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1220 			    active_width);
1221 		*ib_width = IB_WIDTH_4X;
1222 	}
1223 
1224 	return;
1225 }
1226 
1227 static int mlx5_mtu_to_ib_mtu(int mtu)
1228 {
1229 	switch (mtu) {
1230 	case 256: return 1;
1231 	case 512: return 2;
1232 	case 1024: return 3;
1233 	case 2048: return 4;
1234 	case 4096: return 5;
1235 	default:
1236 		pr_warn("invalid mtu\n");
1237 		return -1;
1238 	}
1239 }
1240 
1241 enum ib_max_vl_num {
1242 	__IB_MAX_VL_0		= 1,
1243 	__IB_MAX_VL_0_1		= 2,
1244 	__IB_MAX_VL_0_3		= 3,
1245 	__IB_MAX_VL_0_7		= 4,
1246 	__IB_MAX_VL_0_14	= 5,
1247 };
1248 
1249 enum mlx5_vl_hw_cap {
1250 	MLX5_VL_HW_0	= 1,
1251 	MLX5_VL_HW_0_1	= 2,
1252 	MLX5_VL_HW_0_2	= 3,
1253 	MLX5_VL_HW_0_3	= 4,
1254 	MLX5_VL_HW_0_4	= 5,
1255 	MLX5_VL_HW_0_5	= 6,
1256 	MLX5_VL_HW_0_6	= 7,
1257 	MLX5_VL_HW_0_7	= 8,
1258 	MLX5_VL_HW_0_14	= 15
1259 };
1260 
1261 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1262 				u8 *max_vl_num)
1263 {
1264 	switch (vl_hw_cap) {
1265 	case MLX5_VL_HW_0:
1266 		*max_vl_num = __IB_MAX_VL_0;
1267 		break;
1268 	case MLX5_VL_HW_0_1:
1269 		*max_vl_num = __IB_MAX_VL_0_1;
1270 		break;
1271 	case MLX5_VL_HW_0_3:
1272 		*max_vl_num = __IB_MAX_VL_0_3;
1273 		break;
1274 	case MLX5_VL_HW_0_7:
1275 		*max_vl_num = __IB_MAX_VL_0_7;
1276 		break;
1277 	case MLX5_VL_HW_0_14:
1278 		*max_vl_num = __IB_MAX_VL_0_14;
1279 		break;
1280 
1281 	default:
1282 		return -EINVAL;
1283 	}
1284 
1285 	return 0;
1286 }
1287 
1288 static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port,
1289 			       struct ib_port_attr *props)
1290 {
1291 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1292 	struct mlx5_core_dev *mdev = dev->mdev;
1293 	struct mlx5_hca_vport_context *rep;
1294 	u16 max_mtu;
1295 	u16 oper_mtu;
1296 	int err;
1297 	u16 ib_link_width_oper;
1298 	u8 vl_hw_cap;
1299 
1300 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1301 	if (!rep) {
1302 		err = -ENOMEM;
1303 		goto out;
1304 	}
1305 
1306 	/* props being zeroed by the caller, avoid zeroing it here */
1307 
1308 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1309 	if (err)
1310 		goto out;
1311 
1312 	props->lid		= rep->lid;
1313 	props->lmc		= rep->lmc;
1314 	props->sm_lid		= rep->sm_lid;
1315 	props->sm_sl		= rep->sm_sl;
1316 	props->state		= rep->vport_state;
1317 	props->phys_state	= rep->port_physical_state;
1318 	props->port_cap_flags	= rep->cap_mask1;
1319 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1320 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1321 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1322 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1323 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1324 	props->subnet_timeout	= rep->subnet_timeout;
1325 	props->init_type_reply	= rep->init_type_reply;
1326 
1327 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1328 		props->port_cap_flags2 = rep->cap_mask2;
1329 
1330 	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1331 				      &props->active_speed, port);
1332 	if (err)
1333 		goto out;
1334 
1335 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1336 
1337 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1338 
1339 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1340 
1341 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1342 
1343 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1344 
1345 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1346 	if (err)
1347 		goto out;
1348 
1349 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1350 				   &props->max_vl_num);
1351 out:
1352 	kfree(rep);
1353 	return err;
1354 }
1355 
1356 int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1357 		       struct ib_port_attr *props)
1358 {
1359 	unsigned int count;
1360 	int ret;
1361 
1362 	switch (mlx5_get_vport_access_method(ibdev)) {
1363 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1364 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1365 		break;
1366 
1367 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1368 		ret = mlx5_query_hca_port(ibdev, port, props);
1369 		break;
1370 
1371 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1372 		ret = mlx5_query_port_roce(ibdev, port, props);
1373 		break;
1374 
1375 	default:
1376 		ret = -EINVAL;
1377 	}
1378 
1379 	if (!ret && props) {
1380 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1381 		struct mlx5_core_dev *mdev;
1382 		bool put_mdev = true;
1383 
1384 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1385 		if (!mdev) {
1386 			/* If the port isn't affiliated yet query the master.
1387 			 * The master and slave will have the same values.
1388 			 */
1389 			mdev = dev->mdev;
1390 			port = 1;
1391 			put_mdev = false;
1392 		}
1393 		count = mlx5_core_reserved_gids_count(mdev);
1394 		if (put_mdev)
1395 			mlx5_ib_put_native_port_mdev(dev, port);
1396 		props->gid_tbl_len -= count;
1397 	}
1398 	return ret;
1399 }
1400 
1401 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port,
1402 				  struct ib_port_attr *props)
1403 {
1404 	return mlx5_query_port_roce(ibdev, port, props);
1405 }
1406 
1407 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1408 				  u16 *pkey)
1409 {
1410 	/* Default special Pkey for representor device port as per the
1411 	 * IB specification 1.3 section 10.9.1.2.
1412 	 */
1413 	*pkey = 0xffff;
1414 	return 0;
1415 }
1416 
1417 static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
1418 			     union ib_gid *gid)
1419 {
1420 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1421 	struct mlx5_core_dev *mdev = dev->mdev;
1422 
1423 	switch (mlx5_get_vport_access_method(ibdev)) {
1424 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1425 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1426 
1427 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1428 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1429 
1430 	default:
1431 		return -EINVAL;
1432 	}
1433 
1434 }
1435 
1436 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port,
1437 				   u16 index, u16 *pkey)
1438 {
1439 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1440 	struct mlx5_core_dev *mdev;
1441 	bool put_mdev = true;
1442 	u32 mdev_port_num;
1443 	int err;
1444 
1445 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1446 	if (!mdev) {
1447 		/* The port isn't affiliated yet, get the PKey from the master
1448 		 * port. For RoCE the PKey tables will be the same.
1449 		 */
1450 		put_mdev = false;
1451 		mdev = dev->mdev;
1452 		mdev_port_num = 1;
1453 	}
1454 
1455 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1456 					index, pkey);
1457 	if (put_mdev)
1458 		mlx5_ib_put_native_port_mdev(dev, port);
1459 
1460 	return err;
1461 }
1462 
1463 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1464 			      u16 *pkey)
1465 {
1466 	switch (mlx5_get_vport_access_method(ibdev)) {
1467 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1468 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1469 
1470 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1471 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1472 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1473 	default:
1474 		return -EINVAL;
1475 	}
1476 }
1477 
1478 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1479 				 struct ib_device_modify *props)
1480 {
1481 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1482 	struct mlx5_reg_node_desc in;
1483 	struct mlx5_reg_node_desc out;
1484 	int err;
1485 
1486 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1487 		return -EOPNOTSUPP;
1488 
1489 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1490 		return 0;
1491 
1492 	/*
1493 	 * If possible, pass node desc to FW, so it can generate
1494 	 * a 144 trap.  If cmd fails, just ignore.
1495 	 */
1496 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1497 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1498 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1499 	if (err)
1500 		return err;
1501 
1502 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1503 
1504 	return err;
1505 }
1506 
1507 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask,
1508 				u32 value)
1509 {
1510 	struct mlx5_hca_vport_context ctx = {};
1511 	struct mlx5_core_dev *mdev;
1512 	u32 mdev_port_num;
1513 	int err;
1514 
1515 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1516 	if (!mdev)
1517 		return -ENODEV;
1518 
1519 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1520 	if (err)
1521 		goto out;
1522 
1523 	if (~ctx.cap_mask1_perm & mask) {
1524 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1525 			     mask, ctx.cap_mask1_perm);
1526 		err = -EINVAL;
1527 		goto out;
1528 	}
1529 
1530 	ctx.cap_mask1 = value;
1531 	ctx.cap_mask1_perm = mask;
1532 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1533 						 0, &ctx);
1534 
1535 out:
1536 	mlx5_ib_put_native_port_mdev(dev, port_num);
1537 
1538 	return err;
1539 }
1540 
1541 static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask,
1542 			       struct ib_port_modify *props)
1543 {
1544 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1545 	struct ib_port_attr attr;
1546 	u32 tmp;
1547 	int err;
1548 	u32 change_mask;
1549 	u32 value;
1550 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1551 		      IB_LINK_LAYER_INFINIBAND);
1552 
1553 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1554 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1555 	 */
1556 	if (!is_ib)
1557 		return 0;
1558 
1559 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1560 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1561 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1562 		return set_port_caps_atomic(dev, port, change_mask, value);
1563 	}
1564 
1565 	mutex_lock(&dev->cap_mask_mutex);
1566 
1567 	err = ib_query_port(ibdev, port, &attr);
1568 	if (err)
1569 		goto out;
1570 
1571 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1572 		~props->clr_port_cap_mask;
1573 
1574 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1575 
1576 out:
1577 	mutex_unlock(&dev->cap_mask_mutex);
1578 	return err;
1579 }
1580 
1581 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1582 {
1583 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1584 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1585 }
1586 
1587 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1588 {
1589 	/* Large page with non 4k uar support might limit the dynamic size */
1590 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1591 		return MLX5_MIN_DYN_BFREGS;
1592 
1593 	return MLX5_MAX_DYN_BFREGS;
1594 }
1595 
1596 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1597 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1598 			     struct mlx5_bfreg_info *bfregi)
1599 {
1600 	int uars_per_sys_page;
1601 	int bfregs_per_sys_page;
1602 	int ref_bfregs = req->total_num_bfregs;
1603 
1604 	if (req->total_num_bfregs == 0)
1605 		return -EINVAL;
1606 
1607 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1608 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1609 
1610 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1611 		return -ENOMEM;
1612 
1613 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1614 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1615 	/* This holds the required static allocation asked by the user */
1616 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1617 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1618 		return -EINVAL;
1619 
1620 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1621 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1622 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1623 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1624 
1625 	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",
1626 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1627 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1628 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1629 		    bfregi->num_sys_pages);
1630 
1631 	return 0;
1632 }
1633 
1634 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1635 {
1636 	struct mlx5_bfreg_info *bfregi;
1637 	int err;
1638 	int i;
1639 
1640 	bfregi = &context->bfregi;
1641 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1642 		err = mlx5_cmd_uar_alloc(dev->mdev, &bfregi->sys_pages[i],
1643 					 context->devx_uid);
1644 		if (err)
1645 			goto error;
1646 
1647 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1648 	}
1649 
1650 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1651 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1652 
1653 	return 0;
1654 
1655 error:
1656 	for (--i; i >= 0; i--)
1657 		if (mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1658 					 context->devx_uid))
1659 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1660 
1661 	return err;
1662 }
1663 
1664 static void deallocate_uars(struct mlx5_ib_dev *dev,
1665 			    struct mlx5_ib_ucontext *context)
1666 {
1667 	struct mlx5_bfreg_info *bfregi;
1668 	int i;
1669 
1670 	bfregi = &context->bfregi;
1671 	for (i = 0; i < bfregi->num_sys_pages; i++)
1672 		if (i < bfregi->num_static_sys_pages ||
1673 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1674 			mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1675 					     context->devx_uid);
1676 }
1677 
1678 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1679 {
1680 	int err = 0;
1681 
1682 	mutex_lock(&dev->lb.mutex);
1683 	if (td)
1684 		dev->lb.user_td++;
1685 	if (qp)
1686 		dev->lb.qps++;
1687 
1688 	if (dev->lb.user_td == 2 ||
1689 	    dev->lb.qps == 1) {
1690 		if (!dev->lb.enabled) {
1691 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1692 			dev->lb.enabled = true;
1693 		}
1694 	}
1695 
1696 	mutex_unlock(&dev->lb.mutex);
1697 
1698 	return err;
1699 }
1700 
1701 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1702 {
1703 	mutex_lock(&dev->lb.mutex);
1704 	if (td)
1705 		dev->lb.user_td--;
1706 	if (qp)
1707 		dev->lb.qps--;
1708 
1709 	if (dev->lb.user_td == 1 &&
1710 	    dev->lb.qps == 0) {
1711 		if (dev->lb.enabled) {
1712 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1713 			dev->lb.enabled = false;
1714 		}
1715 	}
1716 
1717 	mutex_unlock(&dev->lb.mutex);
1718 }
1719 
1720 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1721 					  u16 uid)
1722 {
1723 	int err;
1724 
1725 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1726 		return 0;
1727 
1728 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1729 	if (err)
1730 		return err;
1731 
1732 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1733 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1734 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1735 		return err;
1736 
1737 	return mlx5_ib_enable_lb(dev, true, false);
1738 }
1739 
1740 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1741 					     u16 uid)
1742 {
1743 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1744 		return;
1745 
1746 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1747 
1748 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1749 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1750 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1751 		return;
1752 
1753 	mlx5_ib_disable_lb(dev, true, false);
1754 }
1755 
1756 static int set_ucontext_resp(struct ib_ucontext *uctx,
1757 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1758 {
1759 	struct ib_device *ibdev = uctx->device;
1760 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1761 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1762 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1763 
1764 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1765 		resp->dump_fill_mkey = dev->mkeys.dump_fill_mkey;
1766 		resp->comp_mask |=
1767 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1768 	}
1769 
1770 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1771 	if (dev->wc_support)
1772 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1773 						      log_bf_reg_size);
1774 	resp->cache_line_size = cache_line_size();
1775 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1776 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1777 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1778 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1779 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1780 	resp->cqe_version = context->cqe_version;
1781 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1782 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1783 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1784 					MLX5_CAP_GEN(dev->mdev,
1785 						     num_of_uars_per_page) : 1;
1786 	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1787 			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1788 	resp->num_ports = dev->num_ports;
1789 	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1790 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1791 
1792 	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1793 		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1794 		resp->eth_min_inline++;
1795 	}
1796 
1797 	if (dev->mdev->clock_info)
1798 		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1799 
1800 	/*
1801 	 * We don't want to expose information from the PCI bar that is located
1802 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1803 	 * pretend we don't support reading the HCA's core clock. This is also
1804 	 * forced by mmap function.
1805 	 */
1806 	if (PAGE_SIZE <= 4096) {
1807 		resp->comp_mask |=
1808 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1809 		resp->hca_core_clock_offset =
1810 			offsetof(struct mlx5_init_seg,
1811 				 internal_timer_h) % PAGE_SIZE;
1812 	}
1813 
1814 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1815 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1816 
1817 	if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) &&
1818 	    rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) &&
1819 	    rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format)))
1820 		resp->comp_mask |=
1821 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS;
1822 
1823 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1824 
1825 	if (MLX5_CAP_GEN(dev->mdev, drain_sigerr))
1826 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS;
1827 
1828 	resp->comp_mask |=
1829 		MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_MKEY_UPDATE_TAG;
1830 
1831 	return 0;
1832 }
1833 
1834 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1835 				  struct ib_udata *udata)
1836 {
1837 	struct ib_device *ibdev = uctx->device;
1838 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1839 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1840 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1841 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1842 	struct mlx5_bfreg_info *bfregi;
1843 	int ver;
1844 	int err;
1845 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1846 				     max_cqe_version);
1847 	bool lib_uar_4k;
1848 	bool lib_uar_dyn;
1849 
1850 	if (!dev->ib_active)
1851 		return -EAGAIN;
1852 
1853 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1854 		ver = 0;
1855 	else if (udata->inlen >= min_req_v2)
1856 		ver = 2;
1857 	else
1858 		return -EINVAL;
1859 
1860 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1861 	if (err)
1862 		return err;
1863 
1864 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1865 		return -EOPNOTSUPP;
1866 
1867 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1868 		return -EOPNOTSUPP;
1869 
1870 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1871 				    MLX5_NON_FP_BFREGS_PER_UAR);
1872 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1873 		return -EINVAL;
1874 
1875 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1876 		err = mlx5_ib_devx_create(dev, true);
1877 		if (err < 0)
1878 			goto out_ctx;
1879 		context->devx_uid = err;
1880 	}
1881 
1882 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1883 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1884 	bfregi = &context->bfregi;
1885 
1886 	if (lib_uar_dyn) {
1887 		bfregi->lib_uar_dyn = lib_uar_dyn;
1888 		goto uar_done;
1889 	}
1890 
1891 	/* updates req->total_num_bfregs */
1892 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1893 	if (err)
1894 		goto out_devx;
1895 
1896 	mutex_init(&bfregi->lock);
1897 	bfregi->lib_uar_4k = lib_uar_4k;
1898 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1899 				GFP_KERNEL);
1900 	if (!bfregi->count) {
1901 		err = -ENOMEM;
1902 		goto out_devx;
1903 	}
1904 
1905 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1906 				    sizeof(*bfregi->sys_pages),
1907 				    GFP_KERNEL);
1908 	if (!bfregi->sys_pages) {
1909 		err = -ENOMEM;
1910 		goto out_count;
1911 	}
1912 
1913 	err = allocate_uars(dev, context);
1914 	if (err)
1915 		goto out_sys_pages;
1916 
1917 uar_done:
1918 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1919 					     context->devx_uid);
1920 	if (err)
1921 		goto out_uars;
1922 
1923 	INIT_LIST_HEAD(&context->db_page_list);
1924 	mutex_init(&context->db_page_mutex);
1925 
1926 	context->cqe_version = min_t(__u8,
1927 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1928 				 req.max_cqe_version);
1929 
1930 	err = set_ucontext_resp(uctx, &resp);
1931 	if (err)
1932 		goto out_mdev;
1933 
1934 	resp.response_length = min(udata->outlen, sizeof(resp));
1935 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1936 	if (err)
1937 		goto out_mdev;
1938 
1939 	bfregi->ver = ver;
1940 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1941 	context->lib_caps = req.lib_caps;
1942 	print_lib_caps(dev, context->lib_caps);
1943 
1944 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1945 		u32 port = mlx5_core_native_port_num(dev->mdev) - 1;
1946 
1947 		atomic_set(&context->tx_port_affinity,
1948 			   atomic_add_return(
1949 				   1, &dev->port[port].roce.tx_port_affinity));
1950 	}
1951 
1952 	return 0;
1953 
1954 out_mdev:
1955 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1956 
1957 out_uars:
1958 	deallocate_uars(dev, context);
1959 
1960 out_sys_pages:
1961 	kfree(bfregi->sys_pages);
1962 
1963 out_count:
1964 	kfree(bfregi->count);
1965 
1966 out_devx:
1967 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1968 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1969 
1970 out_ctx:
1971 	return err;
1972 }
1973 
1974 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1975 				  struct uverbs_attr_bundle *attrs)
1976 {
1977 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1978 	int ret;
1979 
1980 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
1981 	if (ret)
1982 		return ret;
1983 
1984 	uctx_resp.response_length =
1985 		min_t(size_t,
1986 		      uverbs_attr_get_len(attrs,
1987 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
1988 		      sizeof(uctx_resp));
1989 
1990 	ret = uverbs_copy_to_struct_or_zero(attrs,
1991 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
1992 					&uctx_resp,
1993 					sizeof(uctx_resp));
1994 	return ret;
1995 }
1996 
1997 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1998 {
1999 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2000 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2001 	struct mlx5_bfreg_info *bfregi;
2002 
2003 	bfregi = &context->bfregi;
2004 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2005 
2006 	deallocate_uars(dev, context);
2007 	kfree(bfregi->sys_pages);
2008 	kfree(bfregi->count);
2009 
2010 	if (context->devx_uid)
2011 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2012 }
2013 
2014 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2015 				 int uar_idx)
2016 {
2017 	int fw_uars_per_page;
2018 
2019 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2020 
2021 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2022 }
2023 
2024 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2025 				 int uar_idx)
2026 {
2027 	unsigned int fw_uars_per_page;
2028 
2029 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2030 				MLX5_UARS_IN_PAGE : 1;
2031 
2032 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2033 }
2034 
2035 static int get_command(unsigned long offset)
2036 {
2037 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2038 }
2039 
2040 static int get_arg(unsigned long offset)
2041 {
2042 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2043 }
2044 
2045 static int get_index(unsigned long offset)
2046 {
2047 	return get_arg(offset);
2048 }
2049 
2050 /* Index resides in an extra byte to enable larger values than 255 */
2051 static int get_extended_index(unsigned long offset)
2052 {
2053 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2054 }
2055 
2056 
2057 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2058 {
2059 }
2060 
2061 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2062 {
2063 	switch (cmd) {
2064 	case MLX5_IB_MMAP_WC_PAGE:
2065 		return "WC";
2066 	case MLX5_IB_MMAP_REGULAR_PAGE:
2067 		return "best effort WC";
2068 	case MLX5_IB_MMAP_NC_PAGE:
2069 		return "NC";
2070 	case MLX5_IB_MMAP_DEVICE_MEM:
2071 		return "Device Memory";
2072 	default:
2073 		return NULL;
2074 	}
2075 }
2076 
2077 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2078 					struct vm_area_struct *vma,
2079 					struct mlx5_ib_ucontext *context)
2080 {
2081 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2082 	    !(vma->vm_flags & VM_SHARED))
2083 		return -EINVAL;
2084 
2085 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2086 		return -EOPNOTSUPP;
2087 
2088 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2089 		return -EPERM;
2090 	vm_flags_clear(vma, VM_MAYWRITE);
2091 
2092 	if (!dev->mdev->clock_info)
2093 		return -EOPNOTSUPP;
2094 
2095 	return vm_insert_page(vma, vma->vm_start,
2096 			      virt_to_page(dev->mdev->clock_info));
2097 }
2098 
2099 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2100 {
2101 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2102 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2103 	struct mlx5_var_table *var_table = &dev->var_table;
2104 	struct mlx5_ib_ucontext *context = to_mucontext(entry->ucontext);
2105 
2106 	switch (mentry->mmap_flag) {
2107 	case MLX5_IB_MMAP_TYPE_MEMIC:
2108 	case MLX5_IB_MMAP_TYPE_MEMIC_OP:
2109 		mlx5_ib_dm_mmap_free(dev, mentry);
2110 		break;
2111 	case MLX5_IB_MMAP_TYPE_VAR:
2112 		mutex_lock(&var_table->bitmap_lock);
2113 		clear_bit(mentry->page_idx, var_table->bitmap);
2114 		mutex_unlock(&var_table->bitmap_lock);
2115 		kfree(mentry);
2116 		break;
2117 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2118 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2119 		mlx5_cmd_uar_dealloc(dev->mdev, mentry->page_idx,
2120 				     context->devx_uid);
2121 		kfree(mentry);
2122 		break;
2123 	default:
2124 		WARN_ON(true);
2125 	}
2126 }
2127 
2128 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2129 		    struct vm_area_struct *vma,
2130 		    struct mlx5_ib_ucontext *context)
2131 {
2132 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2133 	int err;
2134 	unsigned long idx;
2135 	phys_addr_t pfn;
2136 	pgprot_t prot;
2137 	u32 bfreg_dyn_idx = 0;
2138 	u32 uar_index;
2139 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2140 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2141 				bfregi->num_static_sys_pages;
2142 
2143 	if (bfregi->lib_uar_dyn)
2144 		return -EINVAL;
2145 
2146 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2147 		return -EINVAL;
2148 
2149 	if (dyn_uar)
2150 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2151 	else
2152 		idx = get_index(vma->vm_pgoff);
2153 
2154 	if (idx >= max_valid_idx) {
2155 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2156 			     idx, max_valid_idx);
2157 		return -EINVAL;
2158 	}
2159 
2160 	switch (cmd) {
2161 	case MLX5_IB_MMAP_WC_PAGE:
2162 	case MLX5_IB_MMAP_ALLOC_WC:
2163 	case MLX5_IB_MMAP_REGULAR_PAGE:
2164 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2165 		prot = pgprot_writecombine(vma->vm_page_prot);
2166 		break;
2167 	case MLX5_IB_MMAP_NC_PAGE:
2168 		prot = pgprot_noncached(vma->vm_page_prot);
2169 		break;
2170 	default:
2171 		return -EINVAL;
2172 	}
2173 
2174 	if (dyn_uar) {
2175 		int uars_per_page;
2176 
2177 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2178 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2179 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2180 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2181 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2182 			return -EINVAL;
2183 		}
2184 
2185 		mutex_lock(&bfregi->lock);
2186 		/* Fail if uar already allocated, first bfreg index of each
2187 		 * page holds its count.
2188 		 */
2189 		if (bfregi->count[bfreg_dyn_idx]) {
2190 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2191 			mutex_unlock(&bfregi->lock);
2192 			return -EINVAL;
2193 		}
2194 
2195 		bfregi->count[bfreg_dyn_idx]++;
2196 		mutex_unlock(&bfregi->lock);
2197 
2198 		err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index,
2199 					 context->devx_uid);
2200 		if (err) {
2201 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2202 			goto free_bfreg;
2203 		}
2204 	} else {
2205 		uar_index = bfregi->sys_pages[idx];
2206 	}
2207 
2208 	pfn = uar_index2pfn(dev, uar_index);
2209 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2210 
2211 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2212 				prot, NULL);
2213 	if (err) {
2214 		mlx5_ib_err(dev,
2215 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2216 			    err, mmap_cmd2str(cmd));
2217 		goto err;
2218 	}
2219 
2220 	if (dyn_uar)
2221 		bfregi->sys_pages[idx] = uar_index;
2222 	return 0;
2223 
2224 err:
2225 	if (!dyn_uar)
2226 		return err;
2227 
2228 	mlx5_cmd_uar_dealloc(dev->mdev, idx, context->devx_uid);
2229 
2230 free_bfreg:
2231 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2232 
2233 	return err;
2234 }
2235 
2236 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2237 {
2238 	unsigned long idx;
2239 	u8 command;
2240 
2241 	command = get_command(vma->vm_pgoff);
2242 	idx = get_extended_index(vma->vm_pgoff);
2243 
2244 	return (command << 16 | idx);
2245 }
2246 
2247 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2248 			       struct vm_area_struct *vma,
2249 			       struct ib_ucontext *ucontext)
2250 {
2251 	struct mlx5_user_mmap_entry *mentry;
2252 	struct rdma_user_mmap_entry *entry;
2253 	unsigned long pgoff;
2254 	pgprot_t prot;
2255 	phys_addr_t pfn;
2256 	int ret;
2257 
2258 	pgoff = mlx5_vma_to_pgoff(vma);
2259 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2260 	if (!entry)
2261 		return -EINVAL;
2262 
2263 	mentry = to_mmmap(entry);
2264 	pfn = (mentry->address >> PAGE_SHIFT);
2265 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2266 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2267 		prot = pgprot_noncached(vma->vm_page_prot);
2268 	else
2269 		prot = pgprot_writecombine(vma->vm_page_prot);
2270 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2271 				entry->npages * PAGE_SIZE,
2272 				prot,
2273 				entry);
2274 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2275 	return ret;
2276 }
2277 
2278 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2279 {
2280 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2281 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2282 
2283 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2284 		(index & 0xFF)) << PAGE_SHIFT;
2285 }
2286 
2287 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2288 {
2289 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2290 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2291 	unsigned long command;
2292 	phys_addr_t pfn;
2293 
2294 	command = get_command(vma->vm_pgoff);
2295 	switch (command) {
2296 	case MLX5_IB_MMAP_WC_PAGE:
2297 	case MLX5_IB_MMAP_ALLOC_WC:
2298 		if (!dev->wc_support)
2299 			return -EPERM;
2300 		fallthrough;
2301 	case MLX5_IB_MMAP_NC_PAGE:
2302 	case MLX5_IB_MMAP_REGULAR_PAGE:
2303 		return uar_mmap(dev, command, vma, context);
2304 
2305 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2306 		return -ENOSYS;
2307 
2308 	case MLX5_IB_MMAP_CORE_CLOCK:
2309 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2310 			return -EINVAL;
2311 
2312 		if (vma->vm_flags & VM_WRITE)
2313 			return -EPERM;
2314 		vm_flags_clear(vma, VM_MAYWRITE);
2315 
2316 		/* Don't expose to user-space information it shouldn't have */
2317 		if (PAGE_SIZE > 4096)
2318 			return -EOPNOTSUPP;
2319 
2320 		pfn = (dev->mdev->iseg_base +
2321 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2322 			PAGE_SHIFT;
2323 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2324 					 PAGE_SIZE,
2325 					 pgprot_noncached(vma->vm_page_prot),
2326 					 NULL);
2327 	case MLX5_IB_MMAP_CLOCK_INFO:
2328 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2329 
2330 	default:
2331 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2332 	}
2333 
2334 	return 0;
2335 }
2336 
2337 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2338 {
2339 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2340 	struct ib_device *ibdev = ibpd->device;
2341 	struct mlx5_ib_alloc_pd_resp resp;
2342 	int err;
2343 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2344 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2345 	u16 uid = 0;
2346 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2347 		udata, struct mlx5_ib_ucontext, ibucontext);
2348 
2349 	uid = context ? context->devx_uid : 0;
2350 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2351 	MLX5_SET(alloc_pd_in, in, uid, uid);
2352 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2353 	if (err)
2354 		return err;
2355 
2356 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2357 	pd->uid = uid;
2358 	if (udata) {
2359 		resp.pdn = pd->pdn;
2360 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2361 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2362 			return -EFAULT;
2363 		}
2364 	}
2365 
2366 	return 0;
2367 }
2368 
2369 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2370 {
2371 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2372 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2373 
2374 	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2375 }
2376 
2377 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2378 {
2379 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2380 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2381 	int err;
2382 	u16 uid;
2383 
2384 	uid = ibqp->pd ?
2385 		to_mpd(ibqp->pd)->uid : 0;
2386 
2387 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2388 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2389 		return -EOPNOTSUPP;
2390 	}
2391 
2392 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2393 	if (err)
2394 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2395 			     ibqp->qp_num, gid->raw);
2396 
2397 	return err;
2398 }
2399 
2400 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2401 {
2402 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2403 	int err;
2404 	u16 uid;
2405 
2406 	uid = ibqp->pd ?
2407 		to_mpd(ibqp->pd)->uid : 0;
2408 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2409 	if (err)
2410 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2411 			     ibqp->qp_num, gid->raw);
2412 
2413 	return err;
2414 }
2415 
2416 static int init_node_data(struct mlx5_ib_dev *dev)
2417 {
2418 	int err;
2419 
2420 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2421 	if (err)
2422 		return err;
2423 
2424 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2425 
2426 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2427 }
2428 
2429 static ssize_t fw_pages_show(struct device *device,
2430 			     struct device_attribute *attr, char *buf)
2431 {
2432 	struct mlx5_ib_dev *dev =
2433 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2434 
2435 	return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2436 }
2437 static DEVICE_ATTR_RO(fw_pages);
2438 
2439 static ssize_t reg_pages_show(struct device *device,
2440 			      struct device_attribute *attr, char *buf)
2441 {
2442 	struct mlx5_ib_dev *dev =
2443 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2444 
2445 	return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2446 }
2447 static DEVICE_ATTR_RO(reg_pages);
2448 
2449 static ssize_t hca_type_show(struct device *device,
2450 			     struct device_attribute *attr, char *buf)
2451 {
2452 	struct mlx5_ib_dev *dev =
2453 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2454 
2455 	return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device);
2456 }
2457 static DEVICE_ATTR_RO(hca_type);
2458 
2459 static ssize_t hw_rev_show(struct device *device,
2460 			   struct device_attribute *attr, char *buf)
2461 {
2462 	struct mlx5_ib_dev *dev =
2463 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2464 
2465 	return sysfs_emit(buf, "%x\n", dev->mdev->rev_id);
2466 }
2467 static DEVICE_ATTR_RO(hw_rev);
2468 
2469 static ssize_t board_id_show(struct device *device,
2470 			     struct device_attribute *attr, char *buf)
2471 {
2472 	struct mlx5_ib_dev *dev =
2473 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2474 
2475 	return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2476 			  dev->mdev->board_id);
2477 }
2478 static DEVICE_ATTR_RO(board_id);
2479 
2480 static struct attribute *mlx5_class_attributes[] = {
2481 	&dev_attr_hw_rev.attr,
2482 	&dev_attr_hca_type.attr,
2483 	&dev_attr_board_id.attr,
2484 	&dev_attr_fw_pages.attr,
2485 	&dev_attr_reg_pages.attr,
2486 	NULL,
2487 };
2488 
2489 static const struct attribute_group mlx5_attr_group = {
2490 	.attrs = mlx5_class_attributes,
2491 };
2492 
2493 static void pkey_change_handler(struct work_struct *work)
2494 {
2495 	struct mlx5_ib_port_resources *ports =
2496 		container_of(work, struct mlx5_ib_port_resources,
2497 			     pkey_change_work);
2498 
2499 	if (!ports->gsi)
2500 		/*
2501 		 * We got this event before device was fully configured
2502 		 * and MAD registration code wasn't called/finished yet.
2503 		 */
2504 		return;
2505 
2506 	mlx5_ib_gsi_pkey_change(ports->gsi);
2507 }
2508 
2509 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2510 {
2511 	struct mlx5_ib_qp *mqp;
2512 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2513 	struct mlx5_core_cq *mcq;
2514 	struct list_head cq_armed_list;
2515 	unsigned long flags_qp;
2516 	unsigned long flags_cq;
2517 	unsigned long flags;
2518 
2519 	INIT_LIST_HEAD(&cq_armed_list);
2520 
2521 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2522 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2523 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2524 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2525 		if (mqp->sq.tail != mqp->sq.head) {
2526 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2527 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2528 			if (send_mcq->mcq.comp &&
2529 			    mqp->ibqp.send_cq->comp_handler) {
2530 				if (!send_mcq->mcq.reset_notify_added) {
2531 					send_mcq->mcq.reset_notify_added = 1;
2532 					list_add_tail(&send_mcq->mcq.reset_notify,
2533 						      &cq_armed_list);
2534 				}
2535 			}
2536 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2537 		}
2538 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2539 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2540 		/* no handling is needed for SRQ */
2541 		if (!mqp->ibqp.srq) {
2542 			if (mqp->rq.tail != mqp->rq.head) {
2543 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2544 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2545 				if (recv_mcq->mcq.comp &&
2546 				    mqp->ibqp.recv_cq->comp_handler) {
2547 					if (!recv_mcq->mcq.reset_notify_added) {
2548 						recv_mcq->mcq.reset_notify_added = 1;
2549 						list_add_tail(&recv_mcq->mcq.reset_notify,
2550 							      &cq_armed_list);
2551 					}
2552 				}
2553 				spin_unlock_irqrestore(&recv_mcq->lock,
2554 						       flags_cq);
2555 			}
2556 		}
2557 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2558 	}
2559 	/*At that point all inflight post send were put to be executed as of we
2560 	 * lock/unlock above locks Now need to arm all involved CQs.
2561 	 */
2562 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2563 		mcq->comp(mcq, NULL);
2564 	}
2565 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2566 }
2567 
2568 static void delay_drop_handler(struct work_struct *work)
2569 {
2570 	int err;
2571 	struct mlx5_ib_delay_drop *delay_drop =
2572 		container_of(work, struct mlx5_ib_delay_drop,
2573 			     delay_drop_work);
2574 
2575 	atomic_inc(&delay_drop->events_cnt);
2576 
2577 	mutex_lock(&delay_drop->lock);
2578 	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2579 	if (err) {
2580 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2581 			     delay_drop->timeout);
2582 		delay_drop->activate = false;
2583 	}
2584 	mutex_unlock(&delay_drop->lock);
2585 }
2586 
2587 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2588 				 struct ib_event *ibev)
2589 {
2590 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2591 
2592 	switch (eqe->sub_type) {
2593 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2594 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2595 					    IB_LINK_LAYER_ETHERNET)
2596 			schedule_work(&ibdev->delay_drop.delay_drop_work);
2597 		break;
2598 	default: /* do nothing */
2599 		return;
2600 	}
2601 }
2602 
2603 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2604 			      struct ib_event *ibev)
2605 {
2606 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2607 
2608 	ibev->element.port_num = port;
2609 
2610 	switch (eqe->sub_type) {
2611 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2612 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2613 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2614 		/* In RoCE, port up/down events are handled in
2615 		 * mlx5_netdev_event().
2616 		 */
2617 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2618 					    IB_LINK_LAYER_ETHERNET)
2619 			return -EINVAL;
2620 
2621 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2622 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2623 		break;
2624 
2625 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2626 		ibev->event = IB_EVENT_LID_CHANGE;
2627 		break;
2628 
2629 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2630 		ibev->event = IB_EVENT_PKEY_CHANGE;
2631 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2632 		break;
2633 
2634 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2635 		ibev->event = IB_EVENT_GID_CHANGE;
2636 		break;
2637 
2638 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2639 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2640 		break;
2641 	default:
2642 		return -EINVAL;
2643 	}
2644 
2645 	return 0;
2646 }
2647 
2648 static void mlx5_ib_handle_event(struct work_struct *_work)
2649 {
2650 	struct mlx5_ib_event_work *work =
2651 		container_of(_work, struct mlx5_ib_event_work, work);
2652 	struct mlx5_ib_dev *ibdev;
2653 	struct ib_event ibev;
2654 	bool fatal = false;
2655 
2656 	if (work->is_slave) {
2657 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2658 		if (!ibdev)
2659 			goto out;
2660 	} else {
2661 		ibdev = work->dev;
2662 	}
2663 
2664 	switch (work->event) {
2665 	case MLX5_DEV_EVENT_SYS_ERROR:
2666 		ibev.event = IB_EVENT_DEVICE_FATAL;
2667 		mlx5_ib_handle_internal_error(ibdev);
2668 		ibev.element.port_num  = (u8)(unsigned long)work->param;
2669 		fatal = true;
2670 		break;
2671 	case MLX5_EVENT_TYPE_PORT_CHANGE:
2672 		if (handle_port_change(ibdev, work->param, &ibev))
2673 			goto out;
2674 		break;
2675 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2676 		handle_general_event(ibdev, work->param, &ibev);
2677 		fallthrough;
2678 	default:
2679 		goto out;
2680 	}
2681 
2682 	ibev.device = &ibdev->ib_dev;
2683 
2684 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2685 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2686 		goto out;
2687 	}
2688 
2689 	if (ibdev->ib_active)
2690 		ib_dispatch_event(&ibev);
2691 
2692 	if (fatal)
2693 		ibdev->ib_active = false;
2694 out:
2695 	kfree(work);
2696 }
2697 
2698 static int mlx5_ib_event(struct notifier_block *nb,
2699 			 unsigned long event, void *param)
2700 {
2701 	struct mlx5_ib_event_work *work;
2702 
2703 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2704 	if (!work)
2705 		return NOTIFY_DONE;
2706 
2707 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2708 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2709 	work->is_slave = false;
2710 	work->param = param;
2711 	work->event = event;
2712 
2713 	queue_work(mlx5_ib_event_wq, &work->work);
2714 
2715 	return NOTIFY_OK;
2716 }
2717 
2718 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2719 				    unsigned long event, void *param)
2720 {
2721 	struct mlx5_ib_event_work *work;
2722 
2723 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2724 	if (!work)
2725 		return NOTIFY_DONE;
2726 
2727 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2728 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2729 	work->is_slave = true;
2730 	work->param = param;
2731 	work->event = event;
2732 	queue_work(mlx5_ib_event_wq, &work->work);
2733 
2734 	return NOTIFY_OK;
2735 }
2736 
2737 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2738 {
2739 	struct mlx5_hca_vport_context vport_ctx;
2740 	int err;
2741 	int port;
2742 
2743 	if (MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_IB)
2744 		return 0;
2745 
2746 	for (port = 1; port <= dev->num_ports; port++) {
2747 		if (!MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2748 			dev->port_caps[port - 1].has_smi = true;
2749 			continue;
2750 		}
2751 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port, 0,
2752 						   &vport_ctx);
2753 		if (err) {
2754 			mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2755 				    port, err);
2756 			return err;
2757 		}
2758 		dev->port_caps[port - 1].has_smi = vport_ctx.has_smi;
2759 	}
2760 
2761 	return 0;
2762 }
2763 
2764 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2765 {
2766 	unsigned int port;
2767 
2768 	rdma_for_each_port (&dev->ib_dev, port)
2769 		mlx5_query_ext_port_caps(dev, port);
2770 }
2771 
2772 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2773 {
2774 	switch (umr_fence_cap) {
2775 	case MLX5_CAP_UMR_FENCE_NONE:
2776 		return MLX5_FENCE_MODE_NONE;
2777 	case MLX5_CAP_UMR_FENCE_SMALL:
2778 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2779 	default:
2780 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2781 	}
2782 }
2783 
2784 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
2785 {
2786 	struct mlx5_ib_resources *devr = &dev->devr;
2787 	struct ib_srq_init_attr attr;
2788 	struct ib_device *ibdev;
2789 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2790 	int port;
2791 	int ret = 0;
2792 
2793 	ibdev = &dev->ib_dev;
2794 
2795 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
2796 		return -EOPNOTSUPP;
2797 
2798 	devr->p0 = ib_alloc_pd(ibdev, 0);
2799 	if (IS_ERR(devr->p0))
2800 		return PTR_ERR(devr->p0);
2801 
2802 	devr->c0 = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr);
2803 	if (IS_ERR(devr->c0)) {
2804 		ret = PTR_ERR(devr->c0);
2805 		goto error1;
2806 	}
2807 
2808 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
2809 	if (ret)
2810 		goto error2;
2811 
2812 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
2813 	if (ret)
2814 		goto error3;
2815 
2816 	memset(&attr, 0, sizeof(attr));
2817 	attr.attr.max_sge = 1;
2818 	attr.attr.max_wr = 1;
2819 	attr.srq_type = IB_SRQT_XRC;
2820 	attr.ext.cq = devr->c0;
2821 
2822 	devr->s0 = ib_create_srq(devr->p0, &attr);
2823 	if (IS_ERR(devr->s0)) {
2824 		ret = PTR_ERR(devr->s0);
2825 		goto err_create;
2826 	}
2827 
2828 	memset(&attr, 0, sizeof(attr));
2829 	attr.attr.max_sge = 1;
2830 	attr.attr.max_wr = 1;
2831 	attr.srq_type = IB_SRQT_BASIC;
2832 
2833 	devr->s1 = ib_create_srq(devr->p0, &attr);
2834 	if (IS_ERR(devr->s1)) {
2835 		ret = PTR_ERR(devr->s1);
2836 		goto error6;
2837 	}
2838 
2839 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2840 		INIT_WORK(&devr->ports[port].pkey_change_work,
2841 			  pkey_change_handler);
2842 
2843 	return 0;
2844 
2845 error6:
2846 	ib_destroy_srq(devr->s0);
2847 err_create:
2848 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2849 error3:
2850 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2851 error2:
2852 	ib_destroy_cq(devr->c0);
2853 error1:
2854 	ib_dealloc_pd(devr->p0);
2855 	return ret;
2856 }
2857 
2858 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
2859 {
2860 	struct mlx5_ib_resources *devr = &dev->devr;
2861 	int port;
2862 
2863 	/*
2864 	 * Make sure no change P_Key work items are still executing.
2865 	 *
2866 	 * At this stage, the mlx5_ib_event should be unregistered
2867 	 * and it ensures that no new works are added.
2868 	 */
2869 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2870 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2871 
2872 	ib_destroy_srq(devr->s1);
2873 	ib_destroy_srq(devr->s0);
2874 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2875 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2876 	ib_destroy_cq(devr->c0);
2877 	ib_dealloc_pd(devr->p0);
2878 }
2879 
2880 static u32 get_core_cap_flags(struct ib_device *ibdev,
2881 			      struct mlx5_hca_vport_context *rep)
2882 {
2883 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2884 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2885 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2886 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2887 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
2888 	u32 ret = 0;
2889 
2890 	if (rep->grh_required)
2891 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
2892 
2893 	if (ll == IB_LINK_LAYER_INFINIBAND)
2894 		return ret | RDMA_CORE_PORT_IBA_IB;
2895 
2896 	if (raw_support)
2897 		ret |= RDMA_CORE_PORT_RAW_PACKET;
2898 
2899 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2900 		return ret;
2901 
2902 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2903 		return ret;
2904 
2905 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2906 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2907 
2908 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2909 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2910 
2911 	return ret;
2912 }
2913 
2914 static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num,
2915 			       struct ib_port_immutable *immutable)
2916 {
2917 	struct ib_port_attr attr;
2918 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2919 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
2920 	struct mlx5_hca_vport_context rep = {0};
2921 	int err;
2922 
2923 	err = ib_query_port(ibdev, port_num, &attr);
2924 	if (err)
2925 		return err;
2926 
2927 	if (ll == IB_LINK_LAYER_INFINIBAND) {
2928 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
2929 						   &rep);
2930 		if (err)
2931 			return err;
2932 	}
2933 
2934 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2935 	immutable->gid_tbl_len = attr.gid_tbl_len;
2936 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
2937 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2938 
2939 	return 0;
2940 }
2941 
2942 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num,
2943 				   struct ib_port_immutable *immutable)
2944 {
2945 	struct ib_port_attr attr;
2946 	int err;
2947 
2948 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2949 
2950 	err = ib_query_port(ibdev, port_num, &attr);
2951 	if (err)
2952 		return err;
2953 
2954 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2955 	immutable->gid_tbl_len = attr.gid_tbl_len;
2956 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2957 
2958 	return 0;
2959 }
2960 
2961 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
2962 {
2963 	struct mlx5_ib_dev *dev =
2964 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2965 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
2966 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
2967 		 fw_rev_sub(dev->mdev));
2968 }
2969 
2970 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
2971 {
2972 	struct mlx5_core_dev *mdev = dev->mdev;
2973 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
2974 								 MLX5_FLOW_NAMESPACE_LAG);
2975 	struct mlx5_flow_table *ft;
2976 	int err;
2977 
2978 	if (!ns || !mlx5_lag_is_active(mdev))
2979 		return 0;
2980 
2981 	err = mlx5_cmd_create_vport_lag(mdev);
2982 	if (err)
2983 		return err;
2984 
2985 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
2986 	if (IS_ERR(ft)) {
2987 		err = PTR_ERR(ft);
2988 		goto err_destroy_vport_lag;
2989 	}
2990 
2991 	dev->flow_db->lag_demux_ft = ft;
2992 	dev->lag_ports = mlx5_lag_get_num_ports(mdev);
2993 	dev->lag_active = true;
2994 	return 0;
2995 
2996 err_destroy_vport_lag:
2997 	mlx5_cmd_destroy_vport_lag(mdev);
2998 	return err;
2999 }
3000 
3001 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3002 {
3003 	struct mlx5_core_dev *mdev = dev->mdev;
3004 
3005 	if (dev->lag_active) {
3006 		dev->lag_active = false;
3007 
3008 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3009 		dev->flow_db->lag_demux_ft = NULL;
3010 
3011 		mlx5_cmd_destroy_vport_lag(mdev);
3012 	}
3013 }
3014 
3015 static void mlx5_netdev_notifier_register(struct mlx5_roce *roce,
3016 					  struct net_device *netdev)
3017 {
3018 	int err;
3019 
3020 	if (roce->tracking_netdev)
3021 		return;
3022 	roce->tracking_netdev = netdev;
3023 	roce->nb.notifier_call = mlx5_netdev_event;
3024 	err = register_netdevice_notifier_dev_net(netdev, &roce->nb, &roce->nn);
3025 	WARN_ON(err);
3026 }
3027 
3028 static void mlx5_netdev_notifier_unregister(struct mlx5_roce *roce)
3029 {
3030 	if (!roce->tracking_netdev)
3031 		return;
3032 	unregister_netdevice_notifier_dev_net(roce->tracking_netdev, &roce->nb,
3033 					      &roce->nn);
3034 	roce->tracking_netdev = NULL;
3035 }
3036 
3037 static int mlx5e_mdev_notifier_event(struct notifier_block *nb,
3038 				     unsigned long event, void *data)
3039 {
3040 	struct mlx5_roce *roce = container_of(nb, struct mlx5_roce, mdev_nb);
3041 	struct net_device *netdev = data;
3042 
3043 	switch (event) {
3044 	case MLX5_DRIVER_EVENT_UPLINK_NETDEV:
3045 		if (netdev)
3046 			mlx5_netdev_notifier_register(roce, netdev);
3047 		else
3048 			mlx5_netdev_notifier_unregister(roce);
3049 		break;
3050 	default:
3051 		return NOTIFY_DONE;
3052 	}
3053 
3054 	return NOTIFY_OK;
3055 }
3056 
3057 static void mlx5_mdev_netdev_track(struct mlx5_ib_dev *dev, u32 port_num)
3058 {
3059 	struct mlx5_roce *roce = &dev->port[port_num].roce;
3060 
3061 	roce->mdev_nb.notifier_call = mlx5e_mdev_notifier_event;
3062 	mlx5_blocking_notifier_register(dev->mdev, &roce->mdev_nb);
3063 	mlx5_core_uplink_netdev_event_replay(dev->mdev);
3064 }
3065 
3066 static void mlx5_mdev_netdev_untrack(struct mlx5_ib_dev *dev, u32 port_num)
3067 {
3068 	struct mlx5_roce *roce = &dev->port[port_num].roce;
3069 
3070 	mlx5_blocking_notifier_unregister(dev->mdev, &roce->mdev_nb);
3071 	mlx5_netdev_notifier_unregister(roce);
3072 }
3073 
3074 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3075 {
3076 	int err;
3077 
3078 	if (!dev->is_rep && dev->profile != &raw_eth_profile) {
3079 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3080 		if (err)
3081 			return err;
3082 	}
3083 
3084 	err = mlx5_eth_lag_init(dev);
3085 	if (err)
3086 		goto err_disable_roce;
3087 
3088 	return 0;
3089 
3090 err_disable_roce:
3091 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3092 		mlx5_nic_vport_disable_roce(dev->mdev);
3093 
3094 	return err;
3095 }
3096 
3097 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3098 {
3099 	mlx5_eth_lag_cleanup(dev);
3100 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3101 		mlx5_nic_vport_disable_roce(dev->mdev);
3102 }
3103 
3104 static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num,
3105 				 enum rdma_netdev_t type,
3106 				 struct rdma_netdev_alloc_params *params)
3107 {
3108 	if (type != RDMA_NETDEV_IPOIB)
3109 		return -EOPNOTSUPP;
3110 
3111 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3112 }
3113 
3114 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3115 				       size_t count, loff_t *pos)
3116 {
3117 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3118 	char lbuf[20];
3119 	int len;
3120 
3121 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3122 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3123 }
3124 
3125 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3126 					size_t count, loff_t *pos)
3127 {
3128 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3129 	u32 timeout;
3130 	u32 var;
3131 
3132 	if (kstrtouint_from_user(buf, count, 0, &var))
3133 		return -EFAULT;
3134 
3135 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3136 			1000);
3137 	if (timeout != var)
3138 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3139 			    timeout);
3140 
3141 	delay_drop->timeout = timeout;
3142 
3143 	return count;
3144 }
3145 
3146 static const struct file_operations fops_delay_drop_timeout = {
3147 	.owner	= THIS_MODULE,
3148 	.open	= simple_open,
3149 	.write	= delay_drop_timeout_write,
3150 	.read	= delay_drop_timeout_read,
3151 };
3152 
3153 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3154 				      struct mlx5_ib_multiport_info *mpi)
3155 {
3156 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3157 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3158 	int comps;
3159 	int err;
3160 	int i;
3161 
3162 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3163 
3164 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3165 
3166 	spin_lock(&port->mp.mpi_lock);
3167 	if (!mpi->ibdev) {
3168 		spin_unlock(&port->mp.mpi_lock);
3169 		return;
3170 	}
3171 
3172 	mpi->ibdev = NULL;
3173 
3174 	spin_unlock(&port->mp.mpi_lock);
3175 	if (mpi->mdev_events.notifier_call)
3176 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3177 	mpi->mdev_events.notifier_call = NULL;
3178 	mlx5_mdev_netdev_untrack(ibdev, port_num);
3179 	spin_lock(&port->mp.mpi_lock);
3180 
3181 	comps = mpi->mdev_refcnt;
3182 	if (comps) {
3183 		mpi->unaffiliate = true;
3184 		init_completion(&mpi->unref_comp);
3185 		spin_unlock(&port->mp.mpi_lock);
3186 
3187 		for (i = 0; i < comps; i++)
3188 			wait_for_completion(&mpi->unref_comp);
3189 
3190 		spin_lock(&port->mp.mpi_lock);
3191 		mpi->unaffiliate = false;
3192 	}
3193 
3194 	port->mp.mpi = NULL;
3195 
3196 	spin_unlock(&port->mp.mpi_lock);
3197 
3198 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3199 
3200 	mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1);
3201 	/* Log an error, still needed to cleanup the pointers and add
3202 	 * it back to the list.
3203 	 */
3204 	if (err)
3205 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3206 			    port_num + 1);
3207 
3208 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3209 }
3210 
3211 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3212 				    struct mlx5_ib_multiport_info *mpi)
3213 {
3214 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3215 	int err;
3216 
3217 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3218 
3219 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3220 	if (ibdev->port[port_num].mp.mpi) {
3221 		mlx5_ib_dbg(ibdev, "port %u already affiliated.\n",
3222 			    port_num + 1);
3223 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3224 		return false;
3225 	}
3226 
3227 	ibdev->port[port_num].mp.mpi = mpi;
3228 	mpi->ibdev = ibdev;
3229 	mpi->mdev_events.notifier_call = NULL;
3230 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3231 
3232 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3233 	if (err)
3234 		goto unbind;
3235 
3236 	mlx5_mdev_netdev_track(ibdev, port_num);
3237 
3238 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3239 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3240 
3241 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3242 
3243 	return true;
3244 
3245 unbind:
3246 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3247 	return false;
3248 }
3249 
3250 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3251 {
3252 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3253 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3254 							  port_num + 1);
3255 	struct mlx5_ib_multiport_info *mpi;
3256 	int err;
3257 	u32 i;
3258 
3259 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3260 		return 0;
3261 
3262 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3263 						     &dev->sys_image_guid);
3264 	if (err)
3265 		return err;
3266 
3267 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3268 	if (err)
3269 		return err;
3270 
3271 	mutex_lock(&mlx5_ib_multiport_mutex);
3272 	for (i = 0; i < dev->num_ports; i++) {
3273 		bool bound = false;
3274 
3275 		/* build a stub multiport info struct for the native port. */
3276 		if (i == port_num) {
3277 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3278 			if (!mpi) {
3279 				mutex_unlock(&mlx5_ib_multiport_mutex);
3280 				mlx5_nic_vport_disable_roce(dev->mdev);
3281 				return -ENOMEM;
3282 			}
3283 
3284 			mpi->is_master = true;
3285 			mpi->mdev = dev->mdev;
3286 			mpi->sys_image_guid = dev->sys_image_guid;
3287 			dev->port[i].mp.mpi = mpi;
3288 			mpi->ibdev = dev;
3289 			mpi = NULL;
3290 			continue;
3291 		}
3292 
3293 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3294 				    list) {
3295 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3296 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3297 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3298 			}
3299 
3300 			if (bound) {
3301 				dev_dbg(mpi->mdev->device,
3302 					"removing port from unaffiliated list.\n");
3303 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3304 				list_del(&mpi->list);
3305 				break;
3306 			}
3307 		}
3308 		if (!bound)
3309 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3310 				    i + 1);
3311 	}
3312 
3313 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3314 	mutex_unlock(&mlx5_ib_multiport_mutex);
3315 	return err;
3316 }
3317 
3318 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3319 {
3320 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3321 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3322 							  port_num + 1);
3323 	u32 i;
3324 
3325 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3326 		return;
3327 
3328 	mutex_lock(&mlx5_ib_multiport_mutex);
3329 	for (i = 0; i < dev->num_ports; i++) {
3330 		if (dev->port[i].mp.mpi) {
3331 			/* Destroy the native port stub */
3332 			if (i == port_num) {
3333 				kfree(dev->port[i].mp.mpi);
3334 				dev->port[i].mp.mpi = NULL;
3335 			} else {
3336 				mlx5_ib_dbg(dev, "unbinding port_num: %u\n",
3337 					    i + 1);
3338 				list_add_tail(&dev->port[i].mp.mpi->list,
3339 					      &mlx5_ib_unaffiliated_port_list);
3340 				mlx5_ib_unbind_slave_port(dev,
3341 							  dev->port[i].mp.mpi);
3342 			}
3343 		}
3344 	}
3345 
3346 	mlx5_ib_dbg(dev, "removing from devlist\n");
3347 	list_del(&dev->ib_dev_list);
3348 	mutex_unlock(&mlx5_ib_multiport_mutex);
3349 
3350 	mlx5_nic_vport_disable_roce(dev->mdev);
3351 }
3352 
3353 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3354 			    enum rdma_remove_reason why,
3355 			    struct uverbs_attr_bundle *attrs)
3356 {
3357 	struct mlx5_user_mmap_entry *obj = uobject->object;
3358 
3359 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3360 	return 0;
3361 }
3362 
3363 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3364 					    struct mlx5_user_mmap_entry *entry,
3365 					    size_t length)
3366 {
3367 	return rdma_user_mmap_entry_insert_range(
3368 		&c->ibucontext, &entry->rdma_entry, length,
3369 		(MLX5_IB_MMAP_OFFSET_START << 16),
3370 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3371 }
3372 
3373 static struct mlx5_user_mmap_entry *
3374 alloc_var_entry(struct mlx5_ib_ucontext *c)
3375 {
3376 	struct mlx5_user_mmap_entry *entry;
3377 	struct mlx5_var_table *var_table;
3378 	u32 page_idx;
3379 	int err;
3380 
3381 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3382 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3383 	if (!entry)
3384 		return ERR_PTR(-ENOMEM);
3385 
3386 	mutex_lock(&var_table->bitmap_lock);
3387 	page_idx = find_first_zero_bit(var_table->bitmap,
3388 				       var_table->num_var_hw_entries);
3389 	if (page_idx >= var_table->num_var_hw_entries) {
3390 		err = -ENOSPC;
3391 		mutex_unlock(&var_table->bitmap_lock);
3392 		goto end;
3393 	}
3394 
3395 	set_bit(page_idx, var_table->bitmap);
3396 	mutex_unlock(&var_table->bitmap_lock);
3397 
3398 	entry->address = var_table->hw_start_addr +
3399 				(page_idx * var_table->stride_size);
3400 	entry->page_idx = page_idx;
3401 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3402 
3403 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3404 					       var_table->stride_size);
3405 	if (err)
3406 		goto err_insert;
3407 
3408 	return entry;
3409 
3410 err_insert:
3411 	mutex_lock(&var_table->bitmap_lock);
3412 	clear_bit(page_idx, var_table->bitmap);
3413 	mutex_unlock(&var_table->bitmap_lock);
3414 end:
3415 	kfree(entry);
3416 	return ERR_PTR(err);
3417 }
3418 
3419 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3420 	struct uverbs_attr_bundle *attrs)
3421 {
3422 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3423 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3424 	struct mlx5_ib_ucontext *c;
3425 	struct mlx5_user_mmap_entry *entry;
3426 	u64 mmap_offset;
3427 	u32 length;
3428 	int err;
3429 
3430 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3431 	if (IS_ERR(c))
3432 		return PTR_ERR(c);
3433 
3434 	entry = alloc_var_entry(c);
3435 	if (IS_ERR(entry))
3436 		return PTR_ERR(entry);
3437 
3438 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3439 	length = entry->rdma_entry.npages * PAGE_SIZE;
3440 	uobj->object = entry;
3441 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3442 
3443 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3444 			     &mmap_offset, sizeof(mmap_offset));
3445 	if (err)
3446 		return err;
3447 
3448 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3449 			     &entry->page_idx, sizeof(entry->page_idx));
3450 	if (err)
3451 		return err;
3452 
3453 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3454 			     &length, sizeof(length));
3455 	return err;
3456 }
3457 
3458 DECLARE_UVERBS_NAMED_METHOD(
3459 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3460 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3461 			MLX5_IB_OBJECT_VAR,
3462 			UVERBS_ACCESS_NEW,
3463 			UA_MANDATORY),
3464 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3465 			   UVERBS_ATTR_TYPE(u32),
3466 			   UA_MANDATORY),
3467 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3468 			   UVERBS_ATTR_TYPE(u32),
3469 			   UA_MANDATORY),
3470 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3471 			    UVERBS_ATTR_TYPE(u64),
3472 			    UA_MANDATORY));
3473 
3474 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3475 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3476 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3477 			MLX5_IB_OBJECT_VAR,
3478 			UVERBS_ACCESS_DESTROY,
3479 			UA_MANDATORY));
3480 
3481 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3482 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3483 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3484 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3485 
3486 static bool var_is_supported(struct ib_device *device)
3487 {
3488 	struct mlx5_ib_dev *dev = to_mdev(device);
3489 
3490 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3491 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3492 }
3493 
3494 static struct mlx5_user_mmap_entry *
3495 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3496 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3497 {
3498 	struct mlx5_user_mmap_entry *entry;
3499 	struct mlx5_ib_dev *dev;
3500 	u32 uar_index;
3501 	int err;
3502 
3503 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3504 	if (!entry)
3505 		return ERR_PTR(-ENOMEM);
3506 
3507 	dev = to_mdev(c->ibucontext.device);
3508 	err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index, c->devx_uid);
3509 	if (err)
3510 		goto end;
3511 
3512 	entry->page_idx = uar_index;
3513 	entry->address = uar_index2paddress(dev, uar_index);
3514 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3515 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3516 	else
3517 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3518 
3519 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3520 	if (err)
3521 		goto err_insert;
3522 
3523 	return entry;
3524 
3525 err_insert:
3526 	mlx5_cmd_uar_dealloc(dev->mdev, uar_index, c->devx_uid);
3527 end:
3528 	kfree(entry);
3529 	return ERR_PTR(err);
3530 }
3531 
3532 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3533 	struct uverbs_attr_bundle *attrs)
3534 {
3535 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3536 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3537 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3538 	struct mlx5_ib_ucontext *c;
3539 	struct mlx5_user_mmap_entry *entry;
3540 	u64 mmap_offset;
3541 	u32 length;
3542 	int err;
3543 
3544 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3545 	if (IS_ERR(c))
3546 		return PTR_ERR(c);
3547 
3548 	err = uverbs_get_const(&alloc_type, attrs,
3549 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3550 	if (err)
3551 		return err;
3552 
3553 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3554 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3555 		return -EOPNOTSUPP;
3556 
3557 	if (!to_mdev(c->ibucontext.device)->wc_support &&
3558 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3559 		return -EOPNOTSUPP;
3560 
3561 	entry = alloc_uar_entry(c, alloc_type);
3562 	if (IS_ERR(entry))
3563 		return PTR_ERR(entry);
3564 
3565 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3566 	length = entry->rdma_entry.npages * PAGE_SIZE;
3567 	uobj->object = entry;
3568 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3569 
3570 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3571 			     &mmap_offset, sizeof(mmap_offset));
3572 	if (err)
3573 		return err;
3574 
3575 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3576 			     &entry->page_idx, sizeof(entry->page_idx));
3577 	if (err)
3578 		return err;
3579 
3580 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3581 			     &length, sizeof(length));
3582 	return err;
3583 }
3584 
3585 DECLARE_UVERBS_NAMED_METHOD(
3586 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3587 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3588 			MLX5_IB_OBJECT_UAR,
3589 			UVERBS_ACCESS_NEW,
3590 			UA_MANDATORY),
3591 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3592 			     enum mlx5_ib_uapi_uar_alloc_type,
3593 			     UA_MANDATORY),
3594 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3595 			   UVERBS_ATTR_TYPE(u32),
3596 			   UA_MANDATORY),
3597 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3598 			   UVERBS_ATTR_TYPE(u32),
3599 			   UA_MANDATORY),
3600 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3601 			    UVERBS_ATTR_TYPE(u64),
3602 			    UA_MANDATORY));
3603 
3604 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3605 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3606 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3607 			MLX5_IB_OBJECT_UAR,
3608 			UVERBS_ACCESS_DESTROY,
3609 			UA_MANDATORY));
3610 
3611 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3612 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3613 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3614 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3615 
3616 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3617 	mlx5_ib_query_context,
3618 	UVERBS_OBJECT_DEVICE,
3619 	UVERBS_METHOD_QUERY_CONTEXT,
3620 	UVERBS_ATTR_PTR_OUT(
3621 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3622 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3623 				   dump_fill_mkey),
3624 		UA_MANDATORY));
3625 
3626 static const struct uapi_definition mlx5_ib_defs[] = {
3627 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3628 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3629 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3630 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3631 	UAPI_DEF_CHAIN(mlx5_ib_dm_defs),
3632 
3633 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3634 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3635 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3636 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3637 	{}
3638 };
3639 
3640 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3641 {
3642 	mlx5_ib_cleanup_multiport_master(dev);
3643 	WARN_ON(!xa_empty(&dev->odp_mkeys));
3644 	mutex_destroy(&dev->cap_mask_mutex);
3645 	WARN_ON(!xa_empty(&dev->sig_mrs));
3646 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3647 }
3648 
3649 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3650 {
3651 	struct mlx5_core_dev *mdev = dev->mdev;
3652 	int err;
3653 	int i;
3654 
3655 	dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3656 	dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3657 	dev->ib_dev.phys_port_cnt = dev->num_ports;
3658 	dev->ib_dev.dev.parent = mdev->device;
3659 	dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3660 
3661 	for (i = 0; i < dev->num_ports; i++) {
3662 		spin_lock_init(&dev->port[i].mp.mpi_lock);
3663 		rwlock_init(&dev->port[i].roce.netdev_lock);
3664 		dev->port[i].roce.dev = dev;
3665 		dev->port[i].roce.native_port_num = i + 1;
3666 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3667 	}
3668 
3669 	err = mlx5r_cmd_query_special_mkeys(dev);
3670 	if (err)
3671 		return err;
3672 
3673 	err = mlx5_ib_init_multiport_master(dev);
3674 	if (err)
3675 		return err;
3676 
3677 	err = set_has_smi_cap(dev);
3678 	if (err)
3679 		goto err_mp;
3680 
3681 	err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
3682 	if (err)
3683 		goto err_mp;
3684 
3685 	if (mlx5_use_mad_ifc(dev))
3686 		get_ext_port_caps(dev);
3687 
3688 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
3689 
3690 	mutex_init(&dev->cap_mask_mutex);
3691 	INIT_LIST_HEAD(&dev->qp_list);
3692 	spin_lock_init(&dev->reset_flow_resource_lock);
3693 	xa_init(&dev->odp_mkeys);
3694 	xa_init(&dev->sig_mrs);
3695 	atomic_set(&dev->mkey_var, 0);
3696 
3697 	spin_lock_init(&dev->dm.lock);
3698 	dev->dm.dev = mdev;
3699 	return 0;
3700 
3701 err_mp:
3702 	mlx5_ib_cleanup_multiport_master(dev);
3703 	return err;
3704 }
3705 
3706 static int mlx5_ib_enable_driver(struct ib_device *dev)
3707 {
3708 	struct mlx5_ib_dev *mdev = to_mdev(dev);
3709 	int ret;
3710 
3711 	ret = mlx5_ib_test_wc(mdev);
3712 	mlx5_ib_dbg(mdev, "Write-Combining %s",
3713 		    mdev->wc_support ? "supported" : "not supported");
3714 
3715 	return ret;
3716 }
3717 
3718 static const struct ib_device_ops mlx5_ib_dev_ops = {
3719 	.owner = THIS_MODULE,
3720 	.driver_id = RDMA_DRIVER_MLX5,
3721 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
3722 
3723 	.add_gid = mlx5_ib_add_gid,
3724 	.alloc_mr = mlx5_ib_alloc_mr,
3725 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
3726 	.alloc_pd = mlx5_ib_alloc_pd,
3727 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
3728 	.attach_mcast = mlx5_ib_mcg_attach,
3729 	.check_mr_status = mlx5_ib_check_mr_status,
3730 	.create_ah = mlx5_ib_create_ah,
3731 	.create_cq = mlx5_ib_create_cq,
3732 	.create_qp = mlx5_ib_create_qp,
3733 	.create_srq = mlx5_ib_create_srq,
3734 	.create_user_ah = mlx5_ib_create_ah,
3735 	.dealloc_pd = mlx5_ib_dealloc_pd,
3736 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
3737 	.del_gid = mlx5_ib_del_gid,
3738 	.dereg_mr = mlx5_ib_dereg_mr,
3739 	.destroy_ah = mlx5_ib_destroy_ah,
3740 	.destroy_cq = mlx5_ib_destroy_cq,
3741 	.destroy_qp = mlx5_ib_destroy_qp,
3742 	.destroy_srq = mlx5_ib_destroy_srq,
3743 	.detach_mcast = mlx5_ib_mcg_detach,
3744 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
3745 	.drain_rq = mlx5_ib_drain_rq,
3746 	.drain_sq = mlx5_ib_drain_sq,
3747 	.device_group = &mlx5_attr_group,
3748 	.enable_driver = mlx5_ib_enable_driver,
3749 	.get_dev_fw_str = get_dev_fw_str,
3750 	.get_dma_mr = mlx5_ib_get_dma_mr,
3751 	.get_link_layer = mlx5_ib_port_link_layer,
3752 	.map_mr_sg = mlx5_ib_map_mr_sg,
3753 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
3754 	.mmap = mlx5_ib_mmap,
3755 	.mmap_free = mlx5_ib_mmap_free,
3756 	.modify_cq = mlx5_ib_modify_cq,
3757 	.modify_device = mlx5_ib_modify_device,
3758 	.modify_port = mlx5_ib_modify_port,
3759 	.modify_qp = mlx5_ib_modify_qp,
3760 	.modify_srq = mlx5_ib_modify_srq,
3761 	.poll_cq = mlx5_ib_poll_cq,
3762 	.post_recv = mlx5_ib_post_recv_nodrain,
3763 	.post_send = mlx5_ib_post_send_nodrain,
3764 	.post_srq_recv = mlx5_ib_post_srq_recv,
3765 	.process_mad = mlx5_ib_process_mad,
3766 	.query_ah = mlx5_ib_query_ah,
3767 	.query_device = mlx5_ib_query_device,
3768 	.query_gid = mlx5_ib_query_gid,
3769 	.query_pkey = mlx5_ib_query_pkey,
3770 	.query_qp = mlx5_ib_query_qp,
3771 	.query_srq = mlx5_ib_query_srq,
3772 	.query_ucontext = mlx5_ib_query_ucontext,
3773 	.reg_user_mr = mlx5_ib_reg_user_mr,
3774 	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
3775 	.req_notify_cq = mlx5_ib_arm_cq,
3776 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
3777 	.resize_cq = mlx5_ib_resize_cq,
3778 
3779 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
3780 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
3781 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
3782 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
3783 	INIT_RDMA_OBJ_SIZE(ib_qp, mlx5_ib_qp, ibqp),
3784 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
3785 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
3786 };
3787 
3788 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
3789 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
3790 };
3791 
3792 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
3793 	.get_vf_config = mlx5_ib_get_vf_config,
3794 	.get_vf_guid = mlx5_ib_get_vf_guid,
3795 	.get_vf_stats = mlx5_ib_get_vf_stats,
3796 	.set_vf_guid = mlx5_ib_set_vf_guid,
3797 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
3798 };
3799 
3800 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
3801 	.alloc_mw = mlx5_ib_alloc_mw,
3802 	.dealloc_mw = mlx5_ib_dealloc_mw,
3803 
3804 	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
3805 };
3806 
3807 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
3808 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
3809 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
3810 
3811 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
3812 };
3813 
3814 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
3815 {
3816 	struct mlx5_core_dev *mdev = dev->mdev;
3817 	struct mlx5_var_table *var_table = &dev->var_table;
3818 	u8 log_doorbell_bar_size;
3819 	u8 log_doorbell_stride;
3820 	u64 bar_size;
3821 
3822 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3823 					log_doorbell_bar_size);
3824 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3825 					log_doorbell_stride);
3826 	var_table->hw_start_addr = dev->mdev->bar_addr +
3827 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
3828 					doorbell_bar_offset);
3829 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
3830 	var_table->stride_size = 1ULL << log_doorbell_stride;
3831 	var_table->num_var_hw_entries = div_u64(bar_size,
3832 						var_table->stride_size);
3833 	mutex_init(&var_table->bitmap_lock);
3834 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
3835 					  GFP_KERNEL);
3836 	return (var_table->bitmap) ? 0 : -ENOMEM;
3837 }
3838 
3839 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
3840 {
3841 	bitmap_free(dev->var_table.bitmap);
3842 }
3843 
3844 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
3845 {
3846 	struct mlx5_core_dev *mdev = dev->mdev;
3847 	int err;
3848 
3849 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
3850 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
3851 		ib_set_device_ops(&dev->ib_dev,
3852 				  &mlx5_ib_dev_ipoib_enhanced_ops);
3853 
3854 	if (mlx5_core_is_pf(mdev))
3855 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
3856 
3857 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3858 
3859 	if (MLX5_CAP_GEN(mdev, imaicl))
3860 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
3861 
3862 	if (MLX5_CAP_GEN(mdev, xrc))
3863 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
3864 
3865 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
3866 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3867 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
3868 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
3869 
3870 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
3871 
3872 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
3873 		dev->ib_dev.driver_def = mlx5_ib_defs;
3874 
3875 	err = init_node_data(dev);
3876 	if (err)
3877 		return err;
3878 
3879 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
3880 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
3881 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
3882 		mutex_init(&dev->lb.mutex);
3883 
3884 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3885 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
3886 		err = mlx5_ib_init_var_table(dev);
3887 		if (err)
3888 			return err;
3889 	}
3890 
3891 	dev->ib_dev.use_cq_dim = true;
3892 
3893 	return 0;
3894 }
3895 
3896 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
3897 	.get_port_immutable = mlx5_port_immutable,
3898 	.query_port = mlx5_ib_query_port,
3899 };
3900 
3901 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
3902 {
3903 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
3904 	return 0;
3905 }
3906 
3907 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
3908 	.get_port_immutable = mlx5_port_rep_immutable,
3909 	.query_port = mlx5_ib_rep_query_port,
3910 	.query_pkey = mlx5_ib_rep_query_pkey,
3911 };
3912 
3913 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
3914 {
3915 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
3916 	return 0;
3917 }
3918 
3919 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
3920 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
3921 	.create_wq = mlx5_ib_create_wq,
3922 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
3923 	.destroy_wq = mlx5_ib_destroy_wq,
3924 	.get_netdev = mlx5_ib_get_netdev,
3925 	.modify_wq = mlx5_ib_modify_wq,
3926 
3927 	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
3928 			   ib_rwq_ind_tbl),
3929 };
3930 
3931 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
3932 {
3933 	struct mlx5_core_dev *mdev = dev->mdev;
3934 	enum rdma_link_layer ll;
3935 	int port_type_cap;
3936 	u32 port_num = 0;
3937 	int err;
3938 
3939 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3940 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3941 
3942 	if (ll == IB_LINK_LAYER_ETHERNET) {
3943 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
3944 
3945 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3946 
3947 		/* Register only for native ports */
3948 		mlx5_mdev_netdev_track(dev, port_num);
3949 
3950 		err = mlx5_enable_eth(dev);
3951 		if (err)
3952 			goto cleanup;
3953 	}
3954 
3955 	return 0;
3956 cleanup:
3957 	mlx5_mdev_netdev_untrack(dev, port_num);
3958 	return err;
3959 }
3960 
3961 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
3962 {
3963 	struct mlx5_core_dev *mdev = dev->mdev;
3964 	enum rdma_link_layer ll;
3965 	int port_type_cap;
3966 	u32 port_num;
3967 
3968 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3969 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3970 
3971 	if (ll == IB_LINK_LAYER_ETHERNET) {
3972 		mlx5_disable_eth(dev);
3973 
3974 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3975 		mlx5_mdev_netdev_untrack(dev, port_num);
3976 	}
3977 }
3978 
3979 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
3980 {
3981 	mlx5_ib_init_cong_debugfs(dev,
3982 				  mlx5_core_native_port_num(dev->mdev) - 1);
3983 	return 0;
3984 }
3985 
3986 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
3987 {
3988 	mlx5_ib_cleanup_cong_debugfs(dev,
3989 				     mlx5_core_native_port_num(dev->mdev) - 1);
3990 }
3991 
3992 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
3993 {
3994 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
3995 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
3996 }
3997 
3998 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
3999 {
4000 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4001 }
4002 
4003 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4004 {
4005 	int err;
4006 
4007 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4008 	if (err)
4009 		return err;
4010 
4011 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4012 	if (err)
4013 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4014 
4015 	return err;
4016 }
4017 
4018 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4019 {
4020 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4021 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4022 }
4023 
4024 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4025 {
4026 	const char *name;
4027 
4028 	if (!mlx5_lag_is_active(dev->mdev))
4029 		name = "mlx5_%d";
4030 	else
4031 		name = "mlx5_bond_%d";
4032 	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4033 }
4034 
4035 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4036 {
4037 	mlx5_mkey_cache_cleanup(dev);
4038 	mlx5r_umr_resource_cleanup(dev);
4039 }
4040 
4041 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4042 {
4043 	ib_unregister_device(&dev->ib_dev);
4044 }
4045 
4046 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4047 {
4048 	int ret;
4049 
4050 	ret = mlx5r_umr_resource_init(dev);
4051 	if (ret)
4052 		return ret;
4053 
4054 	ret = mlx5_mkey_cache_init(dev);
4055 	if (ret) {
4056 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4057 		mlx5r_umr_resource_cleanup(dev);
4058 	}
4059 	return ret;
4060 }
4061 
4062 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4063 {
4064 	struct dentry *root;
4065 
4066 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4067 		return 0;
4068 
4069 	mutex_init(&dev->delay_drop.lock);
4070 	dev->delay_drop.dev = dev;
4071 	dev->delay_drop.activate = false;
4072 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4073 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4074 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4075 	atomic_set(&dev->delay_drop.events_cnt, 0);
4076 
4077 	if (!mlx5_debugfs_root)
4078 		return 0;
4079 
4080 	root = debugfs_create_dir("delay_drop", mlx5_debugfs_get_dev_root(dev->mdev));
4081 	dev->delay_drop.dir_debugfs = root;
4082 
4083 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4084 				&dev->delay_drop.events_cnt);
4085 	debugfs_create_atomic_t("num_rqs", 0400, root,
4086 				&dev->delay_drop.rqs_cnt);
4087 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4088 			    &fops_delay_drop_timeout);
4089 	return 0;
4090 }
4091 
4092 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4093 {
4094 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4095 		return;
4096 
4097 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4098 	if (!dev->delay_drop.dir_debugfs)
4099 		return;
4100 
4101 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4102 	dev->delay_drop.dir_debugfs = NULL;
4103 }
4104 
4105 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4106 {
4107 	dev->mdev_events.notifier_call = mlx5_ib_event;
4108 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4109 	return 0;
4110 }
4111 
4112 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4113 {
4114 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4115 }
4116 
4117 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4118 		      const struct mlx5_ib_profile *profile,
4119 		      int stage)
4120 {
4121 	dev->ib_active = false;
4122 
4123 	/* Number of stages to cleanup */
4124 	while (stage) {
4125 		stage--;
4126 		if (profile->stage[stage].cleanup)
4127 			profile->stage[stage].cleanup(dev);
4128 	}
4129 
4130 	kfree(dev->port);
4131 	ib_dealloc_device(&dev->ib_dev);
4132 }
4133 
4134 int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4135 		  const struct mlx5_ib_profile *profile)
4136 {
4137 	int err;
4138 	int i;
4139 
4140 	dev->profile = profile;
4141 
4142 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4143 		if (profile->stage[i].init) {
4144 			err = profile->stage[i].init(dev);
4145 			if (err)
4146 				goto err_out;
4147 		}
4148 	}
4149 
4150 	dev->ib_active = true;
4151 	return 0;
4152 
4153 err_out:
4154 	/* Clean up stages which were initialized */
4155 	while (i) {
4156 		i--;
4157 		if (profile->stage[i].cleanup)
4158 			profile->stage[i].cleanup(dev);
4159 	}
4160 	return -ENOMEM;
4161 }
4162 
4163 static const struct mlx5_ib_profile pf_profile = {
4164 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4165 		     mlx5_ib_stage_init_init,
4166 		     mlx5_ib_stage_init_cleanup),
4167 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4168 		     mlx5_ib_fs_init,
4169 		     mlx5_ib_fs_cleanup),
4170 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4171 		     mlx5_ib_stage_caps_init,
4172 		     mlx5_ib_stage_caps_cleanup),
4173 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4174 		     mlx5_ib_stage_non_default_cb,
4175 		     NULL),
4176 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4177 		     mlx5_ib_roce_init,
4178 		     mlx5_ib_roce_cleanup),
4179 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4180 		     mlx5_init_qp_table,
4181 		     mlx5_cleanup_qp_table),
4182 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4183 		     mlx5_init_srq_table,
4184 		     mlx5_cleanup_srq_table),
4185 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4186 		     mlx5_ib_dev_res_init,
4187 		     mlx5_ib_dev_res_cleanup),
4188 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4189 		     mlx5_ib_stage_dev_notifier_init,
4190 		     mlx5_ib_stage_dev_notifier_cleanup),
4191 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4192 		     mlx5_ib_odp_init_one,
4193 		     mlx5_ib_odp_cleanup_one),
4194 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4195 		     mlx5_ib_counters_init,
4196 		     mlx5_ib_counters_cleanup),
4197 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4198 		     mlx5_ib_stage_cong_debugfs_init,
4199 		     mlx5_ib_stage_cong_debugfs_cleanup),
4200 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4201 		     mlx5_ib_stage_uar_init,
4202 		     mlx5_ib_stage_uar_cleanup),
4203 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4204 		     mlx5_ib_stage_bfrag_init,
4205 		     mlx5_ib_stage_bfrag_cleanup),
4206 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4207 		     NULL,
4208 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4209 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4210 		     mlx5_ib_devx_init,
4211 		     mlx5_ib_devx_cleanup),
4212 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4213 		     mlx5_ib_stage_ib_reg_init,
4214 		     mlx5_ib_stage_ib_reg_cleanup),
4215 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4216 		     mlx5_ib_stage_post_ib_reg_umr_init,
4217 		     NULL),
4218 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4219 		     mlx5_ib_stage_delay_drop_init,
4220 		     mlx5_ib_stage_delay_drop_cleanup),
4221 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4222 		     mlx5_ib_restrack_init,
4223 		     NULL),
4224 };
4225 
4226 const struct mlx5_ib_profile raw_eth_profile = {
4227 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4228 		     mlx5_ib_stage_init_init,
4229 		     mlx5_ib_stage_init_cleanup),
4230 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4231 		     mlx5_ib_fs_init,
4232 		     mlx5_ib_fs_cleanup),
4233 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4234 		     mlx5_ib_stage_caps_init,
4235 		     mlx5_ib_stage_caps_cleanup),
4236 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4237 		     mlx5_ib_stage_raw_eth_non_default_cb,
4238 		     NULL),
4239 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4240 		     mlx5_ib_roce_init,
4241 		     mlx5_ib_roce_cleanup),
4242 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4243 		     mlx5_init_qp_table,
4244 		     mlx5_cleanup_qp_table),
4245 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4246 		     mlx5_init_srq_table,
4247 		     mlx5_cleanup_srq_table),
4248 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4249 		     mlx5_ib_dev_res_init,
4250 		     mlx5_ib_dev_res_cleanup),
4251 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4252 		     mlx5_ib_stage_dev_notifier_init,
4253 		     mlx5_ib_stage_dev_notifier_cleanup),
4254 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4255 		     mlx5_ib_counters_init,
4256 		     mlx5_ib_counters_cleanup),
4257 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4258 		     mlx5_ib_stage_cong_debugfs_init,
4259 		     mlx5_ib_stage_cong_debugfs_cleanup),
4260 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4261 		     mlx5_ib_stage_uar_init,
4262 		     mlx5_ib_stage_uar_cleanup),
4263 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4264 		     mlx5_ib_stage_bfrag_init,
4265 		     mlx5_ib_stage_bfrag_cleanup),
4266 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4267 		     NULL,
4268 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4269 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4270 		     mlx5_ib_devx_init,
4271 		     mlx5_ib_devx_cleanup),
4272 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4273 		     mlx5_ib_stage_ib_reg_init,
4274 		     mlx5_ib_stage_ib_reg_cleanup),
4275 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4276 		     mlx5_ib_stage_post_ib_reg_umr_init,
4277 		     NULL),
4278 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4279 		     mlx5_ib_stage_delay_drop_init,
4280 		     mlx5_ib_stage_delay_drop_cleanup),
4281 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4282 		     mlx5_ib_restrack_init,
4283 		     NULL),
4284 };
4285 
4286 static int mlx5r_mp_probe(struct auxiliary_device *adev,
4287 			  const struct auxiliary_device_id *id)
4288 {
4289 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4290 	struct mlx5_core_dev *mdev = idev->mdev;
4291 	struct mlx5_ib_multiport_info *mpi;
4292 	struct mlx5_ib_dev *dev;
4293 	bool bound = false;
4294 	int err;
4295 
4296 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4297 	if (!mpi)
4298 		return -ENOMEM;
4299 
4300 	mpi->mdev = mdev;
4301 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4302 						     &mpi->sys_image_guid);
4303 	if (err) {
4304 		kfree(mpi);
4305 		return err;
4306 	}
4307 
4308 	mutex_lock(&mlx5_ib_multiport_mutex);
4309 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4310 		if (dev->sys_image_guid == mpi->sys_image_guid)
4311 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4312 
4313 		if (bound) {
4314 			rdma_roce_rescan_device(&dev->ib_dev);
4315 			mpi->ibdev->ib_active = true;
4316 			break;
4317 		}
4318 	}
4319 
4320 	if (!bound) {
4321 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4322 		dev_dbg(mdev->device,
4323 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4324 	}
4325 	mutex_unlock(&mlx5_ib_multiport_mutex);
4326 
4327 	auxiliary_set_drvdata(adev, mpi);
4328 	return 0;
4329 }
4330 
4331 static void mlx5r_mp_remove(struct auxiliary_device *adev)
4332 {
4333 	struct mlx5_ib_multiport_info *mpi;
4334 
4335 	mpi = auxiliary_get_drvdata(adev);
4336 	mutex_lock(&mlx5_ib_multiport_mutex);
4337 	if (mpi->ibdev)
4338 		mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4339 	else
4340 		list_del(&mpi->list);
4341 	mutex_unlock(&mlx5_ib_multiport_mutex);
4342 	kfree(mpi);
4343 }
4344 
4345 static int mlx5r_probe(struct auxiliary_device *adev,
4346 		       const struct auxiliary_device_id *id)
4347 {
4348 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4349 	struct mlx5_core_dev *mdev = idev->mdev;
4350 	const struct mlx5_ib_profile *profile;
4351 	int port_type_cap, num_ports, ret;
4352 	enum rdma_link_layer ll;
4353 	struct mlx5_ib_dev *dev;
4354 
4355 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4356 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4357 
4358 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4359 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4360 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4361 	if (!dev)
4362 		return -ENOMEM;
4363 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4364 			     GFP_KERNEL);
4365 	if (!dev->port) {
4366 		ib_dealloc_device(&dev->ib_dev);
4367 		return -ENOMEM;
4368 	}
4369 
4370 	dev->mdev = mdev;
4371 	dev->num_ports = num_ports;
4372 
4373 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_get_roce_state(mdev))
4374 		profile = &raw_eth_profile;
4375 	else
4376 		profile = &pf_profile;
4377 
4378 	ret = __mlx5_ib_add(dev, profile);
4379 	if (ret) {
4380 		kfree(dev->port);
4381 		ib_dealloc_device(&dev->ib_dev);
4382 		return ret;
4383 	}
4384 
4385 	auxiliary_set_drvdata(adev, dev);
4386 	return 0;
4387 }
4388 
4389 static void mlx5r_remove(struct auxiliary_device *adev)
4390 {
4391 	struct mlx5_ib_dev *dev;
4392 
4393 	dev = auxiliary_get_drvdata(adev);
4394 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4395 }
4396 
4397 static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4398 	{ .name = MLX5_ADEV_NAME ".multiport", },
4399 	{},
4400 };
4401 
4402 static const struct auxiliary_device_id mlx5r_id_table[] = {
4403 	{ .name = MLX5_ADEV_NAME ".rdma", },
4404 	{},
4405 };
4406 
4407 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4408 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4409 
4410 static struct auxiliary_driver mlx5r_mp_driver = {
4411 	.name = "multiport",
4412 	.probe = mlx5r_mp_probe,
4413 	.remove = mlx5r_mp_remove,
4414 	.id_table = mlx5r_mp_id_table,
4415 };
4416 
4417 static struct auxiliary_driver mlx5r_driver = {
4418 	.name = "rdma",
4419 	.probe = mlx5r_probe,
4420 	.remove = mlx5r_remove,
4421 	.id_table = mlx5r_id_table,
4422 };
4423 
4424 static int __init mlx5_ib_init(void)
4425 {
4426 	int ret;
4427 
4428 	xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4429 	if (!xlt_emergency_page)
4430 		return -ENOMEM;
4431 
4432 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4433 	if (!mlx5_ib_event_wq) {
4434 		free_page((unsigned long)xlt_emergency_page);
4435 		return -ENOMEM;
4436 	}
4437 
4438 	ret = mlx5_ib_qp_event_init();
4439 	if (ret)
4440 		goto qp_event_err;
4441 
4442 	mlx5_ib_odp_init();
4443 	ret = mlx5r_rep_init();
4444 	if (ret)
4445 		goto rep_err;
4446 	ret = auxiliary_driver_register(&mlx5r_mp_driver);
4447 	if (ret)
4448 		goto mp_err;
4449 	ret = auxiliary_driver_register(&mlx5r_driver);
4450 	if (ret)
4451 		goto drv_err;
4452 	return 0;
4453 
4454 drv_err:
4455 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4456 mp_err:
4457 	mlx5r_rep_cleanup();
4458 rep_err:
4459 	mlx5_ib_qp_event_cleanup();
4460 qp_event_err:
4461 	destroy_workqueue(mlx5_ib_event_wq);
4462 	free_page((unsigned long)xlt_emergency_page);
4463 	return ret;
4464 }
4465 
4466 static void __exit mlx5_ib_cleanup(void)
4467 {
4468 	auxiliary_driver_unregister(&mlx5r_driver);
4469 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4470 	mlx5r_rep_cleanup();
4471 
4472 	mlx5_ib_qp_event_cleanup();
4473 	destroy_workqueue(mlx5_ib_event_wq);
4474 	free_page((unsigned long)xlt_emergency_page);
4475 }
4476 
4477 module_init(mlx5_ib_init);
4478 module_exit(mlx5_ib_cleanup);
4479